content
stringlengths 1
15.9M
|
|---|
\section{INTRODUCTION}
Neural networks have currently become one of the remarkable approaches in machine learning. These models can achieve excellent performance in many tasks due to the diversity in their architecture and their ability to automatically extract features. Automatic feature selection is crucial in our big data world because manually choosing the essential features for any data is either impossible or too costly. This ability is essential for neural networks to outperform other AI algorithms and even humans in some tasks. Deep Mind’s Alpha Go is one of the most well-known events in which AI surpasses human ability by the power of the neural network.
Observing the development of artificial neural networks, we can see that the number of parameters in these models is increasing. This trend will obviously exacerbate the information catastrophe \cite{vopson2020information}. In addition, the steady growth of hardware power in line with Moore Law’s predictions will continue to support this trend. In 2018, the BERT \cite{devlin2018bert} model on transformer architecture had 110 million parameters that can perform impressively in a wide range of tasks. Two years later, in 2020, OpenAI released the GPT-3 \cite{brown2020language} model with 175 billion parameters. This trend can be a valid idea supporting the belief that the more complex the model, the better the performance.
Along with the neural network, another effective direction for machine learning is swarm intelligence. Swarm studies appear in many different domains. This approach is much cheaper in terms of complexity than neural networks. The general idea is to use interactions between individuals and the environment to make decisions. Previous studies consider the swarm as an active object that can achieve the goal by distributed intelligence of its member.
Yang et al. \cite{yang2013swarm} indicated different reasons for the developing prominence of algorithms based on swarm intelligence, in particular the versatility and adaptability offered by the algorithms. The adaptability of algorithms to external factors, together with their simplicity and self-learning capacity, are the key highlights displayed. This has increased interest in developing applications in various areas. From the computational perspective, swarm intelligence models provide algorithms for solving distributed optimization calculations.
In addition to the active aspect, the passive aspect of a swarm is also worth studying. Swarm can be the feature providing essential hints that help to solve various problems. For instance, it seems impossible for humans with healthy eyes to spot and identify a single insect flying miles away. However, this becomes feasible if the insects fly in a swarm. A single isolated individual cannot carry the features of a swarm’s individuals. Figure \ref{fig:koi} shows approximately one shoal of koi fish; they can be identified more easily than a single fish. From this observation, the swarm filter \cite{nguyen2019swarm} was invented as a component in a neural network to extract swarm features automatically. In terms of computation, the model can extract swarm information of an individual as a data feature.
\begin{figure}[h]
\centering
\includegraphics[width=.6\linewidth]{figures/koi-fish.jpg}
\caption{Shoal of koi fish can be identified more easily than a single fish}
\label{fig:koi}
\end{figure}
In this paper, we want to show the efficiency of an architecture using a swarm filter, which is the SCNN. We compare the SCNN with the conventional architecture using basic elements of neural networks. We trained all models with a large amount of data from scratch without any pretraining or transfer learning techniques so that we can better verify the learning ability of the networks. In our experiments, the SCNN showed impressive results on both effectiveness and efficiency.
The main contributions of this paper are the introduction of a novel architecture for neural networks using the swarm filter and examining its performance and learning behavior. To convey the paper in a straightforward manner, we cover some background knowledge in Section \ref{sec:background}. The model as well as our detailed experiments and analysis are covered in \ref{sec:scnn}. Section \ref{sec:related_works} distinguishes our work from others. Finally, we present some conclusions and future directions for this model in Section \ref{sec:conclusion}.
\section{BACKGROUND}
\label{sec:background}
\subsection{SWARM INTELLIGENCE}
Swarm intelligence is a field of artificial intelligence that is motivated by the conduct of some social living creatures, for example, ants, termites, birds, and fish. For the most part, decentralized control and self-association are the essential highlights of systems based on swarm intelligence, which result in emergent behavior. Emergent behavior develops through local cooperation among the components of the system, and it cannot be accomplished by any of the individuals of the framework acting alone. Accordingly, swarm intelligence is decentralized, self-arranging and disseminated through a given environment. In the natural habitat of organisms, swarm intelligence systems are usually used to tackle issues such as the relocation of colonies, evasion of prey and foraging for food. Data are stored regularly through participating agents and stored in the surroundings.
Following Dorigo et al. \cite{dorigo2014swarm}, systems that are based on swarm intelligence have the following properties. First, they are made of numerous participants who work collectively with others. In addition, the participants are moderately homogeneous. Dorigo et al. stated that another property is that the connections between the participants depend on straightforward behavioral rules that utilize only the local data that the participants exchange using the environment or directly with others. Likewise, the overall result of the framework comes from the collaborations among the participants in the environment. In other words, the behavior of the group is self-organizing. Accordingly, the characterizing feature of the swarm intelligence framework is its capacity to act in a planned manner without the nearness of an organizer or an outside controller.
Ant colony optimization is one of the most common swarm intelligence algorithms. Lopez-Ibanez et al. \cite{lopez2012automatic} indicated that it is propelled by the social conduct and steering method of ants looking for food. Lim et al. \cite{lim2009innovations} explained that ants can create the shortest path from their colonies to their sources of food and back. While scanning for food, ants initially meander around their environment arbitrarily. After discovering food, they return to their colonies while laying a path of a synthetic substance called a “pheromone” along their way. This pheromone is utilized for communication whereby different ants will identify the trail of the pheromone. Lopez-Ibanez et al. stated that the richer the pheromones along a path, the more probable that different ants will identify and follow the route. Ants will, in general, pick the course set apart by the most potent concentration of pheromone fixation.
The concentration of pheromone vanishes after some time, decreasing its alluring quality. The pheromone will vanish faster if it takes the ants more time to go to the food and return. In this manner, the pheromone along a shorter way will be fortified faster on the grounds that when different ants follow the way, they continue adding their pheromone to the path before it dissipates. As a result, the shortest route between the colonies and the identified sources of food inevitably emerges. From the computational perspective, one of the benefits of the evaporation of the ants’ pheromone is to mitigate convergence to a locally optimal solution. Without a doubt, pheromone dissipation is a valuable mechanism for forcing a type of overlooking, henceforth permitting the solution space to be investigated in an all-inclusive way.
Another conventional swarm intelligence algorithm is known as particle swarm optimization. Garro et al. \cite{garro2015designing} articulated that this swarm intelligence is propelled by social conduct fish schooling as well as the flocking of birds. In this manner, the particle swarm optimization algorithm is a developmental calculation model that has underlying foundations in both natural and artificial life. It uses a populace of particles that fly through a multidimensional solution space with their velocities set \cite{lim2009innovations}. Every particle encodes a solitary convergence of all search measurements. The related position and speed of every particle are arbitrarily created. Algorithm \ref{alg:pso} demonstrates how particle swarm optimization works.
\begin{algorithm}[H]
\label{alg:pso}
\SetAlgoLined
Initialize each particle and its velocity V and location X\;
\Repeat{$t$=MAX or $f_{gbest}$=0}{
\For{each particle i}{
Calculate the fitness value $f_i$ of particle $i$\;
\If{$f_i$ \textless $f_{pbest_i}$}{
$pbest_i$ = $X_i$;\\
}
\If{$f_i$ \textless $f_{gbest_i}$}{
$gbest_i$ = $X_i$;\\
}
Update $X_i$ and $V_i$;
}
}
\caption{Particle Swarm Optimization}
\end{algorithm}
Bee colony optimization is yet another well-known swarm intelligence algorithm. Yuce et al. \cite{yuce2013honey} stated that the bee colony optimization algorithm was motivated by the search practices of honeybees. Foraging honeybees are conveyed from their colonies to scan for quality nourishment sources, for example, blossom patches. After discovering a decent nourishment source, Lim et al. \cite{lim2009innovations} said that a foraging honeybee comes back to the hive and educates its hive mates using waggle dance. The mystery behind the honeybee waggle dance was decoded by Von Frisch \cite{von1967dance}, who identified that the waggle dance is a specialized communication mechanism with different honeybees.
Accordingly, the waggle dance of the foraging honeybee passes on three bits of significant data to different honeybees, that is, the distance to the identified food source, the direction, and the quality of the nourishment source. Specifically, the foraging honeybee utilizes this waggle dance as a way to persuade different honeybees to be followers and return to the nourishment source. Subsequently, more honeybees are pulled into all of the more encouraging nourishment sources. Yuce et al. \cite{yuce2013honey} said that such a system is sent by honeybee colonies to acquire quality nourishment quickly and productively.
\subsection{ARTIFICIAL NEURAL NETWORKS}
These algorithms perceive essential features from data through a procedure that impersonates how the human cerebrum works. A neural network alludes to a system of neurons, which can either be artificial or natural, existing in nature. In an artificial neural network, a neuron is a mathematical function that collects and classifies data. The model is composed of thousands of neurons that are densely interconnected and incrementally calculates the abstraction of the data from the input layer to the output layer. There is a weighted connection between nodes in a neural network. Instead of biology signals as weights for connection, the artificial neural network uses a numerical value. When that value passes a threshold, the neuron fires a new signal to the next one.
There is no single general architecture for every neural network; each of them has a different design to perform best in their application problems. However, all of them are constructed using some basic components. The simplest elements of neural networks are perceptions \cite{rosenblatt1958perceptron}. More complex components are later developed based on them, such as the cells and gates of RNN networks and the convolution units of CNN networks. Each architecture fits specific tasks with its characteristics.
The first component in the history of artificial neural networks is the perceptron. It is a mathematical function that calculates $f(x)=w \cdot x + b$, in which $x$ is the input vector and $w$ and $b$ are weight and bias parameters of the model, respectively. This simple model can only work on linearly separable data.
Overcoming that weakness of perceptron, multilayered perceptron (MLP) was proposed. MLP contains perceptrons that are organized in layers that are inter connected, whereby the information layer gathers the patterns that are input. The layer used for output is composed of output signals as well as classifications to which patterns used during input may outline. Within neural networks, the margin of error is minimized through the hidden layers, which are used for fine-tuning the weightings that are input.
During the training process, the model needs to optimize the parameters so that it can fit with the training data. After being trained, the model can predict the outcome of unseen data in the future.
MLP encounters a challenge of overfitting. Another common neural network that overcame that challenge is known as the convolutional neural network (CNN). A CNN is composed of at least a number of convolutional layers. These layers can either be pooled or interconnected completely. Prior to passing the outcome to the following layer, the convolutional layer utilizes a convolutional function on the information \cite{kim2014convolutional}. Because of this convolutional function, the neural network can be enabled at a much deeper location. Because of this capacity, convolutional neural networks show exceptionally successful results in various applications.
Another common type of neural network is the recurrent neural network (RNN), in which the output of a specific layer is stored and subsequently fed back to the input node. Sundermeyer et al. \cite{sundermeyer2013comparison} stated that this is essential in predicting the results of the layer. Accordingly, following every data item stored and then fed back to the input, every node will recall some data that it had in the previous step. LSTM is an improved neural network based on RNN with a cell architecture that allows for the selection of important information for prediction steps. Every node is used partly as a memory cell while executing operations. The neural system starts with propagation at the front, yet it recalls the data it may require to use far ahead. If the prediction is not correct, then the framework begins the process of self-learning and progresses in the course of making the correct prediction during backpropagation.
\section{SCNN: SWARM CHARACTERISTIC NEURAL NETWORK}
\label{sec:scnn}
\subsection{SWARM FILTER AND SWARM FEATURE}
Swarm filters and swarm features are important concepts proposed by Nguyen et al. \cite{nguyen2019swarm}. These concepts are based on a simple observation of swarming that individuals in a swarm often show distinctive signs from singular individuals or individuals in another swarm. A swarm here is not limited to a natural swarm but rather a symbolic swarm, for example, a company has its own rules about working style, uniformity and communication style that employees must follow. Based on such attributes, we can recognize an individual as an employee of a certain company. As another example, texts are a regular set of words, and based on the features in the word arrangement, we can perceive the meanings to which the texts refer.
Swarm features are the unique properties of an instance that exists in a swarm. These features may exist explicitly or in latent dimensions. Therefore, they need to be extracted using learning methods such as neural networks. Nguyen et al. proposed the swarm filter as a part of a neural network that extracts these features. In other words, the swarm filter is a set of weights in the neural network. By calculating the outer product with input, it produces swarm features. Swarm filter layers can be stacked and combined with other elements in the neural network.
\subsection{PROPOSED ARCHITECTURE}
The swarm characteristic neural network (SCNN) is a neural network that makes use of the swarm filter \cite{nguyen2019swarm} to extract the features in the latent swarm to which the samples belong. The general architecture of the network is represented in Figure \ref{fig:general_architecture}. This neural network contains two swarm filters and one fully connected layer as its components.
\begin{figure}[h]
\centering
\includegraphics[width=.8\linewidth]{figures/general_architecture.png}
\caption{General architecture of SCNN}
\label{fig:general_architecture}
\end{figure}
The embedding layer maps the words in text into points in vector space. To determine the model’s learning capability, no pretrained embedding was used. After being trained, the points in the space corresponding to the words will reflect the semantic relationship between them. At its simplest, pairs of words with the same meaning will have a shorter distance than other pairs of words. This is the basis for the model to make a decision on the input sentence based on recognizing the difference in the semantics of words.
The input under the form of a concatenated tensor is fed into the neural network. The purpose of the swarm filter is to make the data reflect their swarm features. Let $x=(x_1,x_2,..,x_n)$ and $s=(s_1,s_2,..,s_m)$ be the data vector and the swarm filter vector, respectively. We can obtain the swarm feature vector $f$ by using Equation \ref{eq:scnn}.
\begin{equation}
\label{eq:scnn}
f = (f_j | f_j = \frac{1}{n}\sum_{i=1}^{n}{(x \otimes s)_{ij}})\\
\end{equation}
Since the model updates its parameters, the samples sharing the same swarm feature will make the filter value shift in the same direction. Consequently, the model can extract the latent swarm feature after the training process. Swarm filters transform the data to a more abstract level in a smaller dimensional space. In terms of computational efficiency, the swarm filter requires fewer parameters than other standard components in deep learning models. We only need s parameters of a swarm filter to transform the data to s-dimensional space. The fully connected layer converts the information of the final swarm filter layer into meaningful information for the purpose of the model. In other words, with the fully connected layer, the model can mix the signals needed to make the final decision. It can be seen that every element in the SCNN can be fine-tuned through backpropagation. As a result, it can be used as an end-to-end model in learning problems.
SCNN is quite straightforward in its idea. If we consider a single input combined with the weight in the swarm filter as a particle in a latent swarm, they facsimile the behavior that is described in swarm intelligence in general. Before the network is trained, each particle follows a random direction that is constrained by the shared weights in the swarm filter. The loss function assesses how good the random solutions are. By the process of backpropagation, every particle must follow the best solution with the minimized loss. The swarm filter in this context can be considered the mean for particles to communicate and share information about the reward. This observation complies with the statement of Dorigo et al. \cite{dorigo2014swarm} about particle communication.
\subsection{EXPERIMENTS AND ANALYSIS}
\label{sec:exp}
In this section, we conduct different experiments to learn how the proposed architecture works. First, we study the performance of the model and examine the model’s effectiveness and efficiency. After that, we analyze the learning curve on the experimental data of the model compared to other basic models in deep learning. Finally, the swarm feature visualization in the last experiment provides an intuitive understanding of how the data are processed in the information flow of the model.
\subsubsection{Task and Dataset}
\label{sec:task_and_data}
We run the experiment on a sentiment analysis task using SCNN and other deep learning architectures to compare their performance. Sentiment analysis is a tool in social media that focuses on the qualitative aspect of views, likes and retweets about a particular product. It came about through the eagerness of people to express and share their opinions. Sentiment analysis is an automated process that analyses text data and sorts them into positive, negative or neutral comments \cite{saif2012semantic}. Given a sentence, the model needs to predict whether its sentiment is positive or negative.
To verify the efficiency in terms of calculation, we compare the number of parameters of the models. We also analyze in detail the values of the swarm feature tensors attained from the trained model. From the point of view of deep learning, the better features the model can extract from the data, the better performance the model can achieve. The dataset for the experiment is Sentiment 140 \cite{go2009twitter}.
The most important reason we choose this dataset is the number of samples. After being preprocessed, there are a total of 1.6 million samples with a balanced distribution between positive and negative labels. This large dataset enables us to keep the experiment fair by training the model from scratch without pretraining or transfer learning. The neural network should be able to learn the hierarchy of abstract meaning for the concepts inside the sentence without any supplemental information. The test data contain 177 negative and 182 positive samples constructed manually. Figure \ref{fig:train_length_dis} and Figure \ref{fig:test_length_dis} show the distributions of the sample length of this dataset.
\begin{figure}[h]
\centering
\includegraphics[width=.6\linewidth]{figures/train_length_distribution.png}
\caption{Distribution of sample length in training set.}
\label{fig:train_length_dis}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=.6\linewidth]{figures/test_length_distribution.png}
\caption{Distribution of sample length in test set.}
\label{fig:test_length_dis}
\end{figure}
We preprocess the data before training. All of the nonstandard characters, such as emoji characters and hyperlinks, are removed from the text. Preprocessing is very important because many tweets contain nonstandard text, which makes the model difficult to predict. The word vectors are obtained at the word embedding layer and fed into the neural network with the architecture in Figure \ref{fig:general_architecture}. The final output is a 2-dimensional vector used for binary classification.
We train four deep learning models from scratch. The first model is a multilayer perceptron model with one hidden layer connecting the embedding layer and the output layer directly. The second model is a convolutional neural network with one convolutional layer of kernel size $100 \times 100$, one max-pooling layer with a kernel size of $20 \times 20$, and a fully connected layer. The third one is a bidirectional LSTM neural network with two stacked LSTM layers containing 128 neurons, followed by a fully connected layer. The last layer is an SCNN with two swarm filters of $300$ and $100$ neurons and a fully connected layer.
All models are trained and validated in the same setting. The total vocabulary used in this experiment contains the 100 K most popular English words. The embedding layer converts the words to 100-dimensional vectors. The max sequence length is $140$, the batch size is $32$, the learning rate is $0.001$, and the number of epochs is 10. We use $10\%$ of the original training set as the development set, and the remaining is used to train the model.
\subsubsection{Effectiveness and Efficiency}
The first criterion to evaluate deep learning models is effectiveness. Effectiveness tells us how well the model can achieve the goal in our defined task. The more accurate the value that the model can achieve, the better the model. On the other hand, efficiency shows how optimal in cost the model should be in order to obtain the result. From our perspective, the fewer parameters the model requires to learn and perform a task, the more efficient the model. Table \ref{tab:acc_para_compare} lists the models’ number of parameters and their performance in this task.
\begin{table}
\caption{Efficiency (number of parameters excluding embedding neurons) and effectiveness (accuracy on test set) of each model}
\label{tab:acc_para_compare}
\begin{center}
\begin{tabular}{lcc}
\hline
\textbf{Model}& \textbf{Number of Parameters}& \textbf{Accuracy}\\
\hline
MLP&28,002&81.1\\
CNN&200,142&83.3\\
LSTM&738,307&83.6\\
SCNN&\textbf{332}&\textbf{84.4}\\
\hline
\end{tabular}
\end{center}
\end{table}
Although SCNN did not produce a large gap compared to other models in terms of effectiveness, this model has a bold difference in terms of efficiency. The SCNN achieved state-of-the-art performance and used only 332 parameters. MLP achieved 81.1\% accuracy with approximately 28 K parameters. CNN and LSTM outperformed MLP with accuracy values of 83.3\% and 83.6\% but require many more parameters (200 K and 700 K, respectively).
The swarm filter is an efficient way to reduce the dimension of the data from layer to layer. Suppose we replace swarm filters with fully connected layers. To convert the information from the embedding layer to the second layer with a dimension of 300 and the third layer with a dimension of 10, it would take 4 million parameters. Such architecture is thousands of times costlier than SCNN.
\subsubsection{Learning Curve Analysis}
\begin{figure}
\centering
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/mlp_loss.png}
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/mlp_acc.png}
\caption{Learning curve of MLP in first training epoch}
\label{fig:mlp_learning_curve}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/cnn_loss.png}
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/cnn_acc.png}
\caption{Learning curve of CNN in first training epoch}
\label{fig:cnn_learning_curve}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/lstm_loss.png}
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/lstm_acc.png}
\caption{Learning curve of LSTM in first training epoch}
\label{fig:lstm_learning_curve}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/scnn_loss.png}
\includegraphics[width=.45\linewidth,height=100pt]{figures/learning_curves/scnn_acc.png}
\caption{Learning curve of SCNN in first training epoch}
\label{fig:scnn_learning_curve}
\end{figure}
We analyze the learning curve of 4 models on the training set to compare their ability to find features that represent the data. Observing the model trained with 1.6 million sentences, we found that most of the changes in accuracy and loss value occur in the first epoch. From the second epoch, these values are only adjusted by a small amount. Analyzing the learning curve in detail in the first epoch enables us to understand the learning behavior of these models as well as the level of their data hunger. Figures \ref{fig:mlp_learning_curve}-\ref{fig:scnn_learning_curve} show the learning curves of MLP, LSTM, CNN and LSTM. The accuracy and loss values were recorded after every 100 samples.
The MLP learning curve differs significantly from that of the other models. The loss increased and decreased sharply in the beginning samples. The accuracy curve in the right subfigure shows that the model then becomes stable and learns well in the subsequent sentences. The other three models had better regulation. As a result, the loss value did not swing drastically, as in MLP’s case. However, both CNN and LSTM have a struggling period before they become stable. The learning curve of the SCNN is the smoothest among all four models. This experiment reveals the learning ability of the models when compared to each other.
We made another interesting discovery when analyzing the learning curves of the models. Looking at the first 2000 iterations, we can see that the SCNN has achieved significant performance, while the rest are still struggling. After iteration 2000, SCNN performance increases but not significantly. This model is capable of identifying essential features with a few initial iterations and enhancing feature extraction if more data are available. With fewer than 2000 samples, the SCNN produces a large gap compared to the three other models.
With fewer parameters than other models, the SCNN still has good performance and efficient feature extraction and data generalization. The swarm filter was probably the origin of this result. We visualize the output of the swarm filter to better understand its behavior.
\subsubsection{Swarm Feature Visualization}
\begin{table}
\caption{Last swarm filter output tensor}
\label{tab:swarm_feature_examples}
\begin{center}
\begin{tabular}{lcc}
\hline
\textbf{Sentence}& \textbf{Sentiment}& \textbf{Tensor value}\\
\hline
I love it so much.&+& 0.14 0.15 0 0.15 0 0.14 0 0 0.21 0.15\\
I hate it.&-&0.34 0.37 0 0.36 0 0.34 0 0 0.50 0.35\\
I am mad at you.&-&0.25 0.27 0 0.26 0 0.25 0 0 0.37 0.26\\
It a terrible situation.&-&0.44 0.47 0 0.47 0 0.44 0 0 0.65 0.46\\
Today, I am very happy.&+&0.14 0.15 0 0.14 0 0.14 0 0 0.20 0.14\\
I am sad.&-&0.45 0.48 0 0.47 0 0.45 0 0 0.66 0.46\\
Today is a wonderful day.&+&0.11 0.12 0 0.12 0 0.11 0 0 0.16 0.12\\
I love it.&+&0.14 0.15 0 0.15 0 0.14 0 0 0.21 0.15\\
\hline
\end{tabular}
\end{center}
\end{table}
Calculating the features using Equation \ref{eq:scnn}, the swarm filter makes them follow the ratio of nodes in the filter as a simple pattern. For a better illustration of the generalization capability of SCNN, we visualize the features generated by swarm filters. Table \ref{tab:swarm_feature_examples} lists some example inputs and their sentiment class and the corresponding tensors outputted by the last swarm filter. All tensors are parallel to each other regardless of whether the data are positive or negative. Figure \ref{fig:i_love_it} and Figure \ref{fig:i_am_sad} show heatmaps of the tensors in the positive and negative cases; there is no difference between them.
\begin{figure}[h]
\centering
\includegraphics[width=.6\linewidth]{figures/plot_tensors/I_love_it.png}
\caption{Plot of heatmap of last swarm features generated from positive sentence, ``I love it.” Cells with greater values are brighter.}
\label{fig:i_love_it}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=.6\linewidth]{figures/plot_tensors/I_love_it.png}
\caption{Plot of heatmap of last swarm features generated from negative sentence, ``I am sad.” Cells with greater values are brighter.}
\label{fig:i_am_sad}
\end{figure}
The visualization experiment shows the ability of the SCNN to abstract the characteristic features of individuals belonging to a latent swarm. Passing the data through two swarm filters, the model views every sample from the same aspect. In other words, it simplifies the pattern of the data. The last swarm filter’s output tensor is passed into a dense layer to obtain the final decision. The SCNN collects and abstracts the features of individuals effectively before making the final prediction. Such an approach could be the reason that the SCNN requires fewer parameters but performs better than other architectures.
\section{RELATED WORKS}
\label{sec:related_works}
Compared to the ideas of SCNN in this paper, previous works have applied the idea of swarm intelligence to neural networks in different ways. Swarm optimization algorithms can help to find optimized parameters for deep learning. Singh et al. \cite{singh2018integration}, Ye et al. \cite{ye2017particle} and Kenny et al. \cite{kenny2017study} are some of the authors who made proposals using this approach.
Singh et al. proposed the follow-the-leader (FTL) algorithm to improve the effectiveness of the artificial neural network in forecast problems. The algorithm is inspired by the movement of sheep belonging to a herd. Each sheep should follow his leader, which is the sheep with the best moves. In detail, after the feedforward neural network calculates the error, the configuration is then optimized by FTL.
With the same purpose of optimizing the structure of deep neural networks, Ye et al. proposed a method using the PSO algorithm. In their proposal, each particle independently represents a configuration of the neural networks. Hyperparameters of the model are set to the optimal value by the algorithm.
In another work, autoencoder networks were enhanced by Kenny et al. using the PSO approach. They conducted a study investigating the effects of the method in a high-dimensional space. The network was divided into smaller subnetworks with a lower number of connection weights. After that, subnetworks were optimized using PSO.
Swarming is limited not only in swarm intelligence and the dynamic aspect but also in the swarm feature and the static aspect. Such information is not only surface information but also latent features that require extraction to be understood. From the above observation, the swarm filter \cite{nguyen2019swarm} was proposed as an innovative way of combining swarming and deep learning research.
\section{CONCLUSION}
\label{sec:conclusion}
By conducting detailed experiments, we verified the effectiveness and efficiency of the SCNN. This architecture outperformed models using standard deep learning components such as the multilayer perceptron, LSTM, and CNN. Furthermore, we showed that the SCNN requires fewer parameters and fewer training data.
There are some future directions for the SCNN and swarm filter. With its excellent ability to extract features, this architecture could contribute to more elaborate designs such as transformer-based models \cite{vaswani2017attention} for performance improvements. The lightweight property of this model can benefit a system with limited storage. In addition, the learning ability of this model is useful for domains in which training data are scarce or costly to obtain.
\bibliographystyle{plain}
|
\section{Introduction and Preliminaries}
\label{sec:introduction}
All graphs considered in this work are assumed to be finite and connected. We use $d_G(p, q)$ to indicate the distance between two vertices $p$ and $q$ in graph $G$ as the length of a shortest path between $p$ and $q$ in $G$. We also, use $L(G)$ to indicate the line graph of a graph $G$, as the vertex set of $L(G)$ is the edges of $G$ and two vertices of $L(G)$ are adjacent in $L(G)$ if and only if they are incident in $G$.
The cartesian product of two graphs $G$ and $H$, denoted by $G \Box H$, is the graph with vertex set $V(G) \times V(H)$ and with edge set
$E(G \times H)$ so that $(g_1,h_1)(g_2,h_2) \in E(G \Box H)$, whenever $h_1 = h_2$ and $g_1g_2 \in E(G)$, or $g_1 = g_2$ and $h_1h_2 \in E(H)$, see [1].
A graphical representation of a vertex $p$ of a connected graph $G$ relative to an arranged subset $Q = \{q_1, ..., q_k\}$ of vertices of $G$ is defined as the $k$-tuple $(d(p, q_1), ..., d(p, q_k ))$, and this $k$-tuple is denoted by $r(p|Q)$, where $d(p, q_i)$ is considered as the minimum length of a shortest path from $p$ to $q_i$ in graph $G$. If any vertices $p$ and $q$ that belong to $V(G)-Q$ have various representations with respect to the set $Q$, then $Q$ is called a resolving set for $G$ [2].
Slater [3]
considered the concept and notation of the metric dimension problem under the term locating set. Also, Harary and Melter [4]
considered these problems under the term metric dimension as follows.
A resolving set of vertices of graph $G$ with the minimum size or cardinality is called the metric dimension of $G$ and this minimum size denoted by $\beta(G)$. Resolving parameters in graphs have been studied in [5-11].
In 2007 C\'{a}ceres et al. [12]
considered the concept and notation of a doubly resolving set of graph $G$.
Two vertices $u$, $v$ in a graph $G$ are doubly resolved by $x, y \in V (G)$ if $d(u,x)-d(u,y)\neq d(v,x)-d(v,y)$,
and we can see that a subset $Q = \{q_1, q_2, ..., q_l\}$ of vertices of a graph $G$ is a doubly resolving set of $G$ if for any various vertices $x, y \in V(G)$ we have $r(x|Q)-r(y|Q)\neq\lambda I$, where $\lambda$ is an integer, and $I$ indicates the unit $l$- vector $(1,..., 1)$, see [13].
A doubly resolving set of vertices of graph $G$ with the minimum size, is denoted by $\psi(G)$.
In 2000 Chartrand et al. showed that for every connected graph $G$ and the path $P_2$, $\beta(G \Box P_2)\leq \beta(G)+1$, see Theorem 7 in [14].
In 2007 C\'{a}ceres et al. obtained an upper bound for the metric dimension of cartesian product of graphs $G$ and $H$. They showed that for all graphs $G$ and $H\neq K_1$, $\beta(G \Box H)\leq \beta(G)+\psi(H)-1$.
In particular, C\'{a}ceres et al.
showed that for every connected graph $G$ and the path $P_n$, $\beta(G \Box P_n)\leq \beta(G)+1$.
Doubly resolving sets have played a special role in the study of resolving sets.
Applications of above concepts and related parameters in graph theory and other sciences have a long history, in particular, if we consider a graph as a chemical compound then the determination of a doubly resolving set with the minimum size is very useful to analysis of chemical compound, and note that these problems are NP hard, see [15-19].
The notion of a strong metric dimension problem set of vertices of graph $G$ introduced by A. Seb\"{o} and E. Tannier [20],
indeed introduced a more restricted invariant than the metric dimension and this was further investigated by O. R. Oellermann and Peters-Fransen [21].
A set $Q\subseteq V(G)$ is called strong resolving set of $G$, if for any various vertices $p$ and $q$ of $G$ there is a vertex of $Q$, say $r$ so that $p$ belongs to a shortest $q - r$ path or $q$ belongs to a shortest $p - r$ path. A strong metric basis of $G$ is indicated by $sdim(G)$ defined as the minimum size of a strong resolving set of $G$, for more results see [22, 23].\\
Now, we use $C_n$ and $P_k$ to denote the cycle on $n \geq 3$ and the path on $k\geq 3$ vertices, respectively.
In this article, we will consider the computational study of some resolving sets with the minimum size for $(C_n\Box P_k)\Box P_m$.
Indeed, in Section 3.1, we define a graph isomorphic to the cartesian product $C_n\Box P_k$, and we will consider the determination of a doubly resolving set of vertices with the minimum size for the cartesian product $C_n\Box P_k$. In particular, in Section 3.1, we construct a graph so that this graph is isomorphic to $(C_n\Box P_k)\Box P_m$, and we compute some resolving parameters with the minimum size for
$(C_n\Box P_k)\Box P_m$. Moreover, in Section 3.2, we consider the determination of some resolving parameters for the graph $H(n)$, and we study the minimum size of some resolving sets for the line graph of the graph $H(n)$ is denoted by $L(n)$.\newline
\section{ Some Facts}
\begin{definition}\label{b.1}
Consider two graphs $G$ and $H$. If there is a bijection, $\theta: V(G)\rightarrow V(H)$ so that $u$ is adjacent to $v$ in $G$ if and only if $\theta(u)$ is adjacent to $\theta(v)$ in $H$, then we say that $G$ and $H$ are isomorphic.
\end{definition}
\begin{definition}\label{b.1.1}
A vertex $u$ of a graph $G$ is called maximally distant from a vertex $v$ of $G$, if for every $w\in N_G(u)$, we have $d(v,w) \leq d(v, u)$, where $N_G (u)$ to denote the set of neighbors that $u$ has in $G$. If $u$ is maximally distant from $v$ and $v$ is maximally distant from $u$, then $u$ and $v$ are said to be mutually maximally distant.
\end{definition}
\begin{remark} \label{b.4}
Suppose that $n$ is an even natural number greater than or equal to $4$ and $G$ is the cycle graph $C_n$ . Then $\beta(G) =2$, $\psi(G)=3$ and $sdim(G)=\lceil \frac{n}{2}\rceil$.
\end{remark}
\begin{remark} \label{b.5}
Suppose that $n$ is an odd natural number greater than or equal to $3$ and $G$ is the cycle graph $C_n$. Then $\beta(G) =2$, $\psi(G)=2$ and $sdim(G)=\lceil \frac{n}{2}\rceil$.
\end{remark}
\begin{remark} \label{b.5.1}
Consider the path $P_n$ for each $n\geq2$. Then $\beta(P_n) =1$, $\psi(P_n)=2$.
\end{remark}
\begin{theorem}\label{b.6}
Suppose that $n$ is an odd integer greater than or equal to $3$. Then the minimum size of a resolving set in the cartesian product
$C_n\Box P_k$ is $2$.
\end{theorem}
\begin{theorem}\label{b.7}
Suppose that $n$ is an even integer greater than or equal to $4$.
Then the minimum size of a resolving set in the cartesian product $C_n\Box P_k$ is $3$.
\end{theorem}
\begin{theorem}\label{b.8}
If $n$ is an even or odd integer is greater than or equal to $3$, then the minimum size of a strong resolving set in the cartesian product
$C_n\Box P_k$ is $n$.
\end{theorem}
\section{Main Results}
\noindent \textbf{3.1 \,\ Metric dimension, doubly resolving set and strong metric dimension for $(C_n\Box P_k)\Box P_m$ }\\
Some resolving parameters such as the minimum size of resolving sets and strong resolving sets calculated for
the cartesian product $C_n\Box P_k$, see [12, 24],
but in this section we will determine some resolving sets of vertices with the minimum size for $(C_n\Box P_k)\Box P_m$. Suppose $n$ and $k$ are natural numbers greater than or equal to $3$, and $[n]=\{1, ..., n\}$. Now, suppose that $G$ is a graph with vertex set $\{x_1, ..., x_{nk}\}$ on layers $V_1, V_2, ..., V_k$, where $V_p=\{x_{(p-1)n+1}, x_{(p-1)n+2}, ..., x_{(p-1)n+n}\}$ for $1\leq p\leq k$, and the edge set of graph $G$ is
$E(G)=\{x_ix_j\,|\, x_i, x_j\in V_p, 1\leq i<j\leq nk, j-i=1 \text{or} j-i=n-1 \}\cup \{x_ix_j\,|\, x_i\in V_q, x_j\in V_{q+1}, 1\leq i<j\leq nk, 1\leq q\leq k-1, j-i=n \}$.
We can see that this graph is isomorphic to the cartesian product $C_n\Box P_k$. So, we can assume throughout this article
$V(C_n\Box P_k)=\{x_1, ..., x_{nk}\}$. Now, in this section we give a more elaborate description of the cartesian product $C_n\Box P_k$,
that are required to prove of Theorems. We use $V_p$, $1\leq p\leq k$, to indicate a layer of the cartesian product $C_n\Box P_k$, where $V_p$, is defined already. Also, for $1\leq e<d\leq nk$, we say that two vertices $x_e$ and $x_d$ in $C_n\Box P_k$ are compatible, if $n|d-e$.
We can see that the degree of a vertex in the layers $V_1$ and $V_k$ is 3, also the degree of a vertex in the layer $V_p$, $1<p<k$ is 4, and hence
$C_n\Box P_k$ is not regular. We say that two layers of $C_n\Box P_k$ are congruous, if the degree of compatible vertices in two layers are identical.
Note that, if $n$ is an even natural number, then $C_n\Box P_k$ contains no cycles of odd length, and hence in this case $C_n\Box P_k$ is bipartite.
For more result of families of graphs with constant metric, see [25, 26].
The cartesian product $C_4\Box P_3$ is depicted in Figure 1.\newline
\begin{center}
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=15.0cm,y=15.0cm]
\clip(6.199538042905856,0.29983666911777535) rectangle (6.450634938365554,0.7288631527509081);
\draw [color=black] (6.298712432261516,0.6857056072380374)-- (6.248939348759354,0.6177720743499512);
\draw [color=black] (6.248939348759354,0.6177720743499512)-- (6.348485515763677,0.6197899020594984);
\draw [color=black] (6.348485515763677,0.6197899020594984)-- (6.398258599265839,0.6863782164745531);
\draw [color=black] (6.398258599265839,0.6863782164745531)-- (6.298712432261516,0.6857056072380374);
\draw [color=black] (6.3,0.55)-- (6.246921521049807,0.4859406639928731);
\draw [color=black] (6.246921521049807,0.4859406639928731)-- (6.350029890846933,0.4852492773545749);
\draw [color=black] (6.350029890846933,0.4852492773545749)-- (6.4,0.55);
\draw [color=black] (6.4,0.55)-- (6.3,0.55);
\draw [color=black] (6.298039823025,0.4328045343081325)-- (6.25025629000202,0.3680152963618003);
\draw [color=black] (6.25025629000202,0.3680152963618003)-- (6.350861337520641,0.3671838496880927);
\draw [color=black] (6.350861337520641,0.3671838496880927)-- (6.39960381773887,0.4328045343081325);
\draw [color=black] (6.39960381773887,0.4328045343081325)-- (6.298039823025,0.4328045343081325);
\draw [color=black] (6.248939348759354,0.6177720743499512)-- (6.25025629000202,0.3680152963618003);
\draw [color=black] (6.348485515763677,0.6197899020594984)-- (6.350861337520641,0.3671838496880927);
\draw [color=black] (6.398258599265839,0.6863782164745531)-- (6.39960381773887,0.4328045343081325);
\draw [color=black] (6.298712432261516,0.6857056072380374)-- (6.298039823025,0.4328045343081325);
\draw (6.204322963708524,0.34972346954023265) node[anchor=north west] {Figure 1. $C_4\Box P_3$};
\begin{scriptsize}
\draw [fill=black] (6.298712432261516,0.6857056072380374) circle (1.5pt);
\draw[color=black] (6.304300323793015,0.6972681791500184) node {$x_{12}$};
\draw [fill=black] (6.248939348759354,0.6177720743499512) circle (1.5pt);
\draw[color=black] (6.237784589896406,0.6207750851689172) node {$x_9$};
\draw [fill=black] (6.348485515763677,0.6197899020594984) circle (1.5pt);
\draw[color=black] (6.343378317457272,0.6324153386008239) node {$x_{10}$};
\draw [fill=black] (6.398258599265839,0.6863782164745531) circle (1.5pt);
\draw[color=black] (6.4040739246379275,0.698099625823726) node {$x_{11}$};
\draw [fill=black] (6.3,0.55) circle (1.5pt);
\draw[color=black] (6.31596321714043,0.5617423713356761) node {$x_8$};
\draw [fill=black] (6.246921521049807,0.4859406639928731) circle (1.5pt);
\draw[color=black] (6.235290249875283,0.49439519076535876) node {$x_5$};
\draw [fill=black] (6.350029890846933,0.4852492773545749) circle (1.5pt);
\draw[color=black] (6.337558190741319,0.4968895307864816) node {$x_6$};
\draw [fill=black] (6.4,0.55) circle (1.5pt);
\draw[color=black] (6.409636817985343,0.5617423713356761) node {$x_7$};
\draw [fill=black] (6.298039823025,0.4328045343081325) circle (1.5pt);
\draw[color=black] (6.3134688771193075,0.4445083903429015) node {$x_4$};
\draw [fill=black] (6.25025629000202,0.3680152963618003) circle (1.5pt);
\draw[color=black] (6.239447483243821,0.37217252973033843) node {$x_1$};
\draw [fill=black] (6.350861337520641,0.3671838496880927) circle (1.5pt);
\draw[color=black] (6.337558190741319,0.3788241031199994) node {$x_2$};
\draw [fill=black] (6.39960381773887,0.4328045343081325) circle (1.5pt);
\draw[color=black] (6.409062604680173,0.4386882636269481) node {$x_3$};
\end{scriptsize}
\end{tikzpicture}
\end{center}
Now, let $m\geq2$ be an integer. Suppose $1\leq i\leq m$ and consider $i^{th}$ copy of the cartesian product $C_n\Box P_k$ with the vertex set $\{x_1^{(i)}, ..., x_{nk}^{(i)}\}$ on the layers $V_1^{(i)}, V_2^{(i)}, ..., V_k^{(i)}$, where it can be defined $V_p^{(i)}$ as similar $V_p$ on the vertex set $\{x_1^{(i)}, ..., x_{nk}^{(i)}\}$. Also, suppose that $K$ is a graph with vertex set $\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}\cup ... \cup\{x_1^{(m)}, ..., x_{nk}^{(m)}\}$ so that for $1\leq t \leq nk$, the vertex $x_t^{(r)}$ is adjacent to the vertex $x_t^{(r+1)}$ in $K$, for $1\leq r\leq m-1$, then we can see that the graph $K$ is isomorphic to $(C_n\Box P_k)\Box P_m$.
For $1\leq e<d\leq nk$, we say that two vertices $x_e^{(i)}$ and $x_d^{(i)}$ in $i^{th}$ copy of the cartesian product $C_n\Box P_k$ are compatible, if $n|d-e$.
The graph $(C_4\Box P_3)\Box P_2$ is depicted in Figure 2. \newline
\begin{center}
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=3.5cm,y=3.5cm]
\clip(6.5645454339715394,1.0260516182926602) rectangle (8.674716775107852,3.529151416054353);
\draw (8.218853014490843,3.2967058297352008)-- (7.204885316407781,3.292391073573145);
\draw (8.218853014490843,3.2967058297352008)-- (8.2,2.6);
\draw (8.2,2.6)-- (7.2,2.6);
\draw (7.2,2.6)-- (7.204885316407781,3.292391073573145);
\draw (6.989147508305002,2.7357875286679763)-- (7.998800450226008,2.7530465533161985);
\draw (7.998800450226008,2.7530465533161985)-- (7.994485694063952,2.1403511783043068);
\draw (7.994485694063952,2.1403511783043068)-- (6.997777020629113,2.148980690628418);
\draw (6.997777020629113,2.148980690628418)-- (6.989147508305002,2.7357875286679763);
\draw (7.204885316407781,3.292391073573145)-- (6.989147508305002,2.7357875286679763);
\draw (8.218853014490843,3.2967058297352008)-- (7.998800450226008,2.7530465533161985);
\draw (8.2,2.6)-- (7.994485694063952,2.1403511783043068);
\draw (7.2,2.6)-- (6.997777020629113,2.148980690628418);
\draw (6.997777020629113,2.148980690628418)-- (7.,1.4);
\draw (7.,1.4)-- (8.,1.4);
\draw (8.2,2.6)-- (8.2,2.);
\draw (7.2,2.6)-- (7.2,2.);
\draw (8.2,2.)-- (8.,1.4);
\draw (8.2,2.)-- (7.2,2.);
\draw (7.2,2.)-- (7.,1.4);
\draw (7.994485694063952,2.1403511783043068)-- (8.,1.4);
\draw [color=black] (7.739915080502673,3.089597533956533)-- (7.398986742414686,3.09282282304291);
\draw [color=black] (7.398986742414686,3.09282282304291)-- (7.307807711449688,2.889423446274836);
\draw [color=black] (7.739915080502673,3.089597533956533)-- (7.657934713423617,2.8911187505019766);
\draw [color=black] (7.307807711449688,2.889423446274836)-- (7.657934713423617,2.8911187505019766);
\draw [color=black] (7.748544592826784,2.455328378134364)-- (7.399049343700282,2.455328378134364);
\draw [color=black] (7.399049343700282,2.455328378134364)-- (7.34295751359356,2.2568495946798075);
\draw [color=black] (7.748544592826784,2.455328378134364)-- (7.679508494233895,2.261164350841863);
\draw [color=black] (7.679508494233895,2.261164350841863)-- (7.34295751359356,2.2568495946798075);
\draw [color=black] (7.399049343700282,1.8512625154465834)-- (7.7528593489888395,1.855577271608639);
\draw [color=black] (7.7528593489888395,1.855577271608639)-- (7.692452762720062,1.6053214142094154);
\draw [color=black] (7.692452762720062,1.6053214142094154)-- (7.347272269755615,1.6010066580473599);
\draw [color=black] (7.399049343700282,1.8512625154465834)-- (7.347272269755615,1.6010066580473599);
\draw [color=black] (7.307807711449688,2.889423446274836)-- (7.34295751359356,2.2568495946798075);
\draw [color=black] (7.398986742414686,3.09282282304291)-- (7.399049343700282,2.455328378134364);
\draw [color=black] (7.739915080502673,3.089597533956533)-- (7.748544592826784,2.455328378134364);
\draw [color=black] (7.657934713423617,2.8911187505019766)-- (7.679508494233895,2.261164350841863);
\draw [color=black] (7.679508494233895,2.261164350841863)-- (7.692452762720062,1.6053214142094154);
\draw [color=black] (7.748544592826784,2.455328378134364)-- (7.7528593489888395,1.855577271608639);
\draw [color=black] (7.34295751359356,2.2568495946798075)-- (7.347272269755615,1.6010066580473599);
\draw [color=black] (7.399049343700282,2.455328378134364)-- (7.399049343700282,1.8512625154465834);
\draw [color=black] (7.398986742414686,3.09282282304291)-- (7.204885316407781,3.292391073573145);
\draw [color=black] (7.739915080502673,3.089597533956533)-- (8.218853014490843,3.2967058297352008);
\draw [color=black] (7.657934713423617,2.8911187505019766)-- (7.998800450226008,2.7530465533161985);
\draw [color=black] (7.307807711449688,2.889423446274836)-- (6.989147508305002,2.7357875286679763);
\draw [color=black] (7.748544592826784,2.455328378134364)-- (8.2,2.6);
\draw [color=black] (7.399049343700282,2.455328378134364)-- (7.2,2.6);
\draw [color=black] (7.34295751359356,2.2568495946798075)-- (6.997777020629113,2.148980690628418);
\draw [color=black] (7.679508494233895,2.261164350841863)-- (7.994485694063952,2.1403511783043068);
\draw [color=black] (7.7528593489888395,1.855577271608639)-- (8.2,2.);
\draw [color=black] (7.399049343700282,1.8512625154465834)-- (7.2,2.);
\draw [color=black] (7.347272269755615,1.6010066580473599)-- (7.,1.4);
\draw [color=black] (7.692452762720062,1.6053214142094154)-- (8.,1.4);
\draw (7.304061064625124,1.2297922995058213) node[anchor=north west] {Figure 2. $(C_4\Box P_3)\Box P_2$};
\begin{scriptsize}
\draw [fill=black] (8.218853014490843,3.2967058297352008) circle (1.5pt);
\draw[color=black] (8.25268250688059,3.3642184836436986) node {$x_{11}^{(2)}$};
\draw [fill=black] (7.204885316407781,3.292391073573145) circle (1.5pt);
\draw[color=black] (7.238830069415096,3.3593675150433855) node {$x_{12}^{(2)}$};
\draw [fill=black] (8.2,2.6) circle (1.5pt);
\draw[color=black] (8.25268250688059,2.631722224996382) node {$x_7^{(2)}$};
\draw [fill=black] (7.2,2.6) circle (1.5pt);
\draw[color=black] (7.136959728808516,2.636573193596695) node {$x_8^{(2)}$};
\draw [fill=black] (6.989147508305002,2.7357875286679763) circle (1.5pt);
\draw[color=black] (6.9235171103947275,2.7626983772048423) node {$x_9^{(2)}$};
\draw [fill=black] (7.998800450226008,2.7530465533161985) circle (1.5pt);
\draw[color=black] (8.078643762868055,2.7626983772048423) node {$x_{10}^{(2)}$};
\draw [fill=black] (7.994485694063952,2.1403511783043068) circle (1.5pt);
\draw[color=black] (8.063494731468367,2.146625364965046) node {$x_6^{(2)}$};
\draw [fill=black] (6.997777020629113,2.148980690628418) circle (1.5pt);
\draw[color=black] (6.913815173194101,2.1563273021656726) node {$x_5^{(2)}$};
\draw [fill=black] (7.,1.4) circle (1.5pt);
\draw[color=black] (6.942026796397861,1.3559174831139686) node {$x_1^{(2)}$};
\draw [fill=black] (8.,1.4) circle (1.5pt);
\draw[color=black] (8.044090857067115,1.3704703889149088) node {$x_2^{(2)}$};
\draw [fill=black] (8.2,2.) circle (1.5pt);
\draw[color=black] (8.257533475480901,2.030202118557525) node {$x_3^{(2)}$};
\draw [fill=black] (7.2,2.) circle (1.5pt);
\draw[color=black] (7.127257791607889,2.0447550243584653) node {$x_4^{(2)}$};
\draw [fill=black] (7.739915080502673,3.089597533956533) circle (1.5pt);
\draw[color=black] (7.7384798352473725,3.1653287710308513) node {$x_{11}^{(1)}$};
\draw [fill=black] (7.398986742414686,3.09282282304291) circle (1.5pt);
\draw[color=black] (7.432868813427631,3.1604778024305378) node {$x_{12}^{(1)}$};
\draw [fill=black] (7.307807711449688,2.889423446274836) circle (1.5pt);
\draw[color=black] (7.268532006615723,2.942184215416437) node {$x_9^{(1)}$};
\draw [fill=black] (7.657934713423617,2.8911187505019766) circle (1.5pt);
\draw[color=black] (7.626907557440165,2.9518861526170634) node {$x_{10}^{(1)}$};
\draw [fill=black] (7.748544592826784,2.455328378134364) circle (1.5pt);
\draw[color=black] (7.802138552650193,2.5250009157894877) node {$x_7^{(1)}$};
\draw [fill=black] (7.399049343700282,2.455328378134364) circle (1.5pt);
\draw[color=black] (7.452868813427631,2.5250009157894877) node {$x_8^{(1)}$};
\draw [fill=black] (7.34295751359356,2.2568495946798075) circle (1.5pt);
\draw[color=black] (7.277637818217603,2.3067073287753868) node {$x_5^{(1)}$};
\draw [fill=black] (7.679508494233895,2.261164350841863) circle (1.5pt);
\draw[color=black] (7.627205620239538,2.321260234576327) node {$x_6^{(1)}$};
\draw [fill=black] (7.399049343700282,1.8512625154465834) circle (1.5pt);
\draw[color=black] (7.452868813427631,1.9186298407503182) node {$x_4^{(1)}$};
\draw [fill=black] (7.7528593489888395,1.855577271608639) circle (1.5pt);
\draw[color=black] (7.806989521250506,1.9283317779509448) node {$x_3^{(1)}$};
\draw [fill=black] (7.692452762720062,1.6053214142094154) circle (1.5pt);
\draw[color=black] (7.757883709648626,1.6421246305324568) node {$x_2^{(1)}$};
\draw [fill=black] (7.347272269755615,1.6010066580473599) circle (1.5pt);
\draw[color=black] (7.2921907240185435,1.6518265677330835) node {$x_1^{(1)}$};
\end{scriptsize}
\end{tikzpicture}
\end{center}
\begin{theorem}\label{d.1}
Consider the cartesian product $C_n\Box P_k$. If $n\geq3$ is an odd integer, then the minimum size of a doubly resolving set of vertices for the cartesian product $C_n\Box P_k$ is $3$.
\end{theorem}
\begin{proof}
In the following cases we show that the minimum size of a doubly resolving set of vertices for the cartesian product $C_n\Box P_k$ is $3$.\\
Case 1. First, we show that the minimum size of a doubly resolving set of vertices in $C_n\Box P_k$ must be greater than $2$.
Consider the cartesian product $C_n\Box P_k$ with the vertex set $\{x_1, ..., x_{nk}\}$ on the layers
$V_1, V_2, ..., V_k$, is defined already. Based on Theorem \ref{b.6}, we know that $\beta(C_n\Box P_k)=2$.
We can show that if $n$ is an odd integer then all the elements of every minimum resolving set of vertices in $C_n\Box P_k$ must be lie in
exactly one of the congruous layers $V_1$ or $V_k$. Without lack of theory if we consider the layer $V_1$ of the cartesian product
$C_n\Box P_k$ then we can show that all the minimum resolving sets of vertices in the layer $V_1$ of $C_n\Box P_k$ are the sets as to form
$M_i=\{x_i, x_{\lceil\frac{n}{2}\rceil+i-1}\}$, $1\leq i\leq \lceil\frac{n}{2}\rceil$ and $N_j=\{x_j, x_{\lceil\frac{n}{2}\rceil+j}\}$, $1\leq j\leq \lfloor\frac{n}{2}\rfloor$.
On the other hand, we can see that the arranged subsets $M_i$ of vertices in $C_n\Box P_k$ cannot be doubly resolving sets for $C_n\Box P_k$ because for
$1\leq i\leq \lceil\frac{n}{2}\rceil$ and two compatible vertices $x_{i+n}$ and $x_{i+2n}$ with respect to $x_i$, we have
$r(x_{i+n}|M_i)-r(x_{i+2n}|M_i)=-I$, where $I$ indicates the unit $2$-vector $(1, 1)$. By applying the same argument we can show that
the arranged subsets $N_j$ of vertices in $C_n\Box P_k$ cannot be doubly resolving sets for $C_n\Box P_k$. Hence, the minimum size of a doubly resolving set in $C_n\Box P_k$
must be greater than $2$.\\
Case 2. Now, we show that the minimum size of a doubly resolving set of vertices in the cartesian product $C_n\Box P_k$ is $3$.
For $1\leq i\leq \lceil\frac{n}{2}\rceil$, let $x_i$ be a vertex in the layer $V_1$ of $C_n\Box P_k$ and $x_c$ be a compatible
vertex with respect to $x_i$, where $x_c$ lie in the layer $V_k$ of $C_n\Box P_k$, then we can show that
the arranged subsets $A_i=M_i\cup x_c=\{x_i, x_{\lceil\frac{n}{2}\rceil+i-1}, x_c\}$ of vertices in the cartesian product $C_n\Box P_k$ are
doubly resolving sets with the minimum size for the cartesian product $C_n\Box P_k$. It will be enough to show that for any compatible vertices $x_e$
and $x_d$ in $C_n\Box P_k$, $r(x_e|A_i)-r(x_d|A_i)\neq\lambda I$. Suppose $x_e\in V_p$ and $x_d\in V_q$ are compatible vertices in the
cartesian product $C_n\Box P_k$, $1\leq p<q\leq k$. Hence, $r(x_e|M_i)-r(x_d|M_i)=-\lambda I$, where $\lambda$ is a positive integer, and $I$
indicates the unit $2$-vector $(1, 1)$. Also, for the compatible vertex $x_c$ with respect to $x_i$, $r(x_e|x_c)-r(x_d|x_c)=\lambda$. So, $r(x_e|A_i)-r(x_d|A_i)\neq\lambda I$, where $I$ indicates the unit $3$-vector $(1, 1, 1)$. Especially, for $1\leq j\leq \lfloor\frac{n}{2}\rfloor$
if we consider the arranged subsets $B_j=N_j\cup x_c=\{x_j, x_{\lceil\frac{n}{2}\rceil+j}, x_c\}$ of vertices in the cartesian product $C_n\Box P_k$,
where $x_c$ lie in the layer $V_k$ of the cartesian product $C_n\Box P_k$ and $x_c$ is a compatible vertex with respect to $x_j$, then by applying
the same argument we can show that the arranged subsets $B_j=N_j\cup x_c=\{x_j, x_{\lceil\frac{n}{2}\rceil+j}, x_c\}$ of vertices in the cartesian product $C_n\Box P_k$ are doubly resolving sets with the minimum size
for the cartesian product $C_n\Box P_k$.\newline
\end{proof}
\begin{theorem}\label{f.1}
Consider the cartesian product $C_n\Box P_k$. If $n\geq3$ is an odd integer, then the minimum size of a resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is $3$.
\end{theorem}
\begin{proof}
Suppose $V((C_n\Box P_k)\Box P_2)=\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}$. Based on Theorem \ref{b.6}, we know that if $n\geq3$ is an odd integer, then the minimum size of a resolving set of vertices in $C_n\Box P_k$ is $2$. Also, by definition of $(C_n\Box P_k)\Box P_2$ we can verify that for $1\leq t \leq nk$, the vertex $x_t^{(1)}$ is adjacent to the vertex $x_t^{(2)}$ in $(C_n\Box P_k)\Box P_2$, and hence none of the minimal resolving sets of $C_n\Box P_k$ cannot be a resolving set for $(C_n\Box P_k)\Box P_2$. Therefore, the minimum size of a resolving set of vertices in $(C_n\Box P_k)\Box P_2$ must be greater than $2$.
Now, we show that the minimum size of a resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is $3$.
For $1\leq i\leq \lceil\frac{n}{2}\rceil$, let $x_i^{(1)}$ be a vertex in the layer $V_1^{(1)}$ of $(C_n\Box P_k)\Box P_2$ and $x_c^{(1)}$ be a compatible vertex with respect to $x_i^{(1)}$, where $x_c^{(1)}$ lie in the layer $V_k^{(1)}$ of $(C_n\Box P_k)\Box P_2$. Based on Theorem \ref{d.1}, we know that for $1\leq i\leq \lceil\frac{n}{2}\rceil$, $1^{th}$ copy of the arranged subsets
$A_i=\{x_i, x_{\lceil\frac{n}{2}\rceil+i-1}, x_c\}$, denoted by the sets
$A_i^{(1)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$ are doubly resolving sets for the arranged set $\{x_1^{(1)}, ..., x_{nk}^{(1)}\}$ of vertices of $(C_n\Box P_k)$, and hence the arranged sets
$A_i^{(1)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$ are resolving sets for $(C_n\Box P_k)\Box P_2$, because for each vertex in the set $\{x_1^{(2)}, ..., x_{nk}^{(2)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$, we have
$$r(x_t^{(2)}|A_i^{(1)})=( d(x_t^{(1)}, x_i^{(1)})+1, d(x_t^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)})+1, d(x_t^{(1)}, x_{c}^{(1)})+1),$$ so all the vertices of $(C_n\Box P_k)\Box P_2$ have various representations with respect to the sets $A_i^{(1)}$, and hence the minimum size of a resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is $3$.
In the same way for $1\leq j\leq \lfloor\frac{n}{2}\rfloor$, if we consider $1^{th}$ copy the arranged subsets $B_j=\{x_j, x_{\lceil\frac{n}{2}\rceil+j}, x_c\}$, denoted by the sets
$B_j^{(1)}=\{x_j^{(1)}, x_{\lceil\frac{n}{2}\rceil+j}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$, where $x_c^{(1)}$ lie in the layer $V_k^{(1)}$ of $(C_n\Box P_k)\Box P_2$ and $x_c^{(1)}$ is a compatible vertex with respect to $x_j^{(1)}$, then by applying the same argument
we can show that the arranged sets $B_j^{(1)}=\{x_j^{(1)}, x_{\lceil\frac{n}{2}\rceil+j}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$ are resolving sets for $(C_n\Box P_k)\Box P_2$.
\end{proof}
\begin{lemma}\label{f.2}
Consider the cartesian product $C_n\Box P_k$. If $n\geq3$ is an odd integer, then the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is greater than $3$.
\end{lemma}
\begin{proof}
Suppose $V((C_n\Box P_k)\Box P_2)=\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}$ and $1\leq t\leq nk$.
For $1\leq i\leq \lceil\frac{n}{2}\rceil$, let $x_i^{(1)}$ be a vertex in the layer $V_1^{(1)}$ of $(C_n\Box P_k)\Box P_2$ and $x_c^{(1)}$ be a compatible vertex with respect to $x_i^{(1)}$, where $x_c^{(1)}$ lie in the layer $V_k^{(1)}$ of $(C_n\Box P_k)\Box P_2$. Based on proof of Theorem \ref{f.1}, we know that the arranged sets $A_i^{(1)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$ cannot be doubly resolving sets for $(C_n\Box P_k)\Box P_2$, because
$$r(x_t^{(2)}|A_i^{(1)})=( d(x_t^{(1)}, x_i^{(1)})+1, d(x_t^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)})+1, d(x_t^{(1)}, x_{c}^{(1)})+1).$$
In the same way for $1\leq j\leq \lfloor\frac{n}{2}\rfloor$, if we consider the arranged sets $B_j^{(1)}=\{x_j^{(1)}, x_{\lceil\frac{n}{2}\rceil+j}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$, where $x_c^{(1)}$ lie in the layer $V_k^{(1)}$ of $(C_n\Box P_k)\Box P_2$ and $x_c^{(1)}$ is a compatible vertex with respect to $x_j^{(1)}$, then we can show that the arranged sets $B_j^{(1)}$
cannot be doubly resolving sets for $(C_n\Box P_k)\Box P_2$. Hence the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is greater than $3$.
\end{proof}
\begin{theorem}\label{f.4}
Consider the cartesian product $C_n\Box P_k$. If $n\geq3$ is an odd integer, then the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is $4$.
\end{theorem}
\begin{proof}
Suppose $V((C_n\Box P_k)\Box P_2)=\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}$ and $1\leq t\leq nk$.
Based on Theorem \ref{f.1}, we know that if $n\geq3$ is an odd integer, then $\beta((C_n\Box P_k)\Box P_2)=3$ and by Lemma \ref{f.2}, we know that the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is greater than $3$. In particular,
it is well known that $\beta((C_n\Box P_k)\Box P_2)\leq \psi((C_n\Box P_k)\Box P_2)$. Now, we show that if $n\geq3$ is an odd integer, then the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is $4$.
For $1\leq i\leq \lceil\frac{n}{2}\rceil$, let $x_i^{(1)}$ be a vertex in the layer $V_1^{(1)}$ of $(C_n\Box P_k)\Box P_2$ and $x_c^{(1)}$ be a compatible vertex with respect to $x_i^{(1)}$, where $x_c^{(1)}$ lie in the layer $V_k^{(1)}$ of $(C_n\Box P_k)\Box P_2$. Based on Lemma \ref{f.2}, we know that the arranged sets
$A_i^{(1)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_2$, defined already, cannot be
doubly resolving sets for $(C_n\Box P_k)\Box P_2$. Let, $C_i=A_i^{(1)}\cup x_c^{(2)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}, x_c^{(2)}\}$ be an arranged subset of vertices of $(C_n\Box P_k)\Box P_2$, where $x_c^{(2)}$ lie in the layer $V_k^{(2)}$ of $(C_n\Box P_k)\Box P_2$ and the vertex $x_c^{(2)}$ is adjacent to the vertex $x_c^{(1)}$. We show that the arranged subset $C_i$ is a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_2$.
It will be enough to show that for any adjacent vertices $x_t^{(1)}$ and $x_t^{(2)}$, $r(x_t^{(1)}|C_i)-r(x_t^{(2)}|C_i)\neq - I$, where $I$ indicates the unit $4$-vector $(1, ..., 1)$.
We can verify that, $r(x_t^{(1)}|A_i^{(1)})-r(x_t^{(2)}|A_i^{(1)})= - I$, where $I$ indicates the unit $3$-vector, and $r(x_t^{(1)}|x_c^{(2)})-r(x_t^{(2)}|x_c^{(2)})= 1$. Thus the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_2$ is $4$.
\end{proof}
\begin{conclusion}\label{f.4.1}
Consider the cartesian product $C_n\Box P_k$. If $n\geq3$ is an odd integer, then the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is $4$.
\end{conclusion}
\begin{proof}
Suppose $(C_n\Box P_k)\Box P_m$ is a graph with vertex set $\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}\cup ... \cup\{x_1^{(m)}, ..., x_{nk}^{(m)}\}$ so that for $1\leq t \leq nk$, the vertex $x_t^{(r)}$ is adjacent to $x_t^{(r+1)}$ in $(C_n\Box P_k)\Box P_m$, for $1\leq r\leq m-1$. On the other hand we know that for every connected graph $G$ and the path $P_m$, $\beta(G \Box P_m)\leq \beta(G)+1$. So, by considering $G=(C_n\Box P_k)$ we have $\beta(G \Box P_m)\leq \beta(G)+1=3$. Moreover it is not hard to see that
for $1\leq i\leq \lceil\frac{n}{2}\rceil$, the arranged sets
$A_i^{(1)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_m$, defined already,
are resolving sets with the minimum size for $(C_n\Box P_k)\Box P_m$, also by applying the same argument in Theorem \ref{f.4}, we can see that
the arranged sets
$A_i^{(1)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_m$, cannot be doubly resolving sets for $(C_n\Box P_k)\Box P_m$, and hence the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is greater than $3$.
Now, let $D_i=A_i^{(1)}\cup x_c^{(m)}=\{x_i^{(1)}, x_{\lceil\frac{n}{2}\rceil+i-1}^{(1)}, x_c^{(1)}, x_c^{(m)}\}$ be an arranged subset of vertices of $(C_n\Box P_k)\Box P_m$, where $x_c^{(m)}$ lie in the layer $V_k^{(m)}$ of $(C_n\Box P_k)\Box P_m$.
We show that the arranged subset $D_i$ is a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_m$.
It will be enough to show that for every two vertices $x_t^{(r)}$ and $x_t^{(s)}$, $1\leq t\leq nk$, $1\leq r<s\leq m$, $r(x_t^{(r)}|D_i)-r(x_t^{(s)}|D_i)\neq -\lambda I$, where $I$ indicates the unit $4$-vector $(1, ..., 1)$ and $\lambda$ is a positive integer.
For this purpose, let the distance between two the vertices $x_t^{(r)}$ and $x_t^{(s)}$ in $(C_n\Box P_k)\Box P_m$ is $\lambda$, then
we can verify that, $r(x_t^{(r)}|A_i^{(1)})-r(x_t^{(s)}|A_i^{(1)})= - \lambda I$, where $I$ indicates the unit $3$-vector, and $r(x_t^{(t)}|x_c^{(m)})-r(x_t^{(s)}|x_c^{(m)})= \lambda$. Thus the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is $4$.
\end{proof}
\begin{example}
Consider graph $(C_5\Box P_4)\Box P_4$ with vertex set $\{x_1^{(1)}, ..., x_{20}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{20}^{(2)}\}\cup\{x_1^{(3)}, ..., x_{20}^{(3)}\}\cup\{x_1^{(4)}, ..., x_{20}^{(4)}\}$, we can see that the set $D_1=\{x_1^{(1)}, x_3^{(1)}, x_{16}^{(1)}, x_{16}^{(4)} \}$ of vertices of $(C_5\Box P_4)\Box P_4$ is one of the minimum doubly resolving sets for $(C_5\Box P_4)\Box P_4$, and hence the minimum size of a doubly resolving set of vertices in $(C_5\Box P_4)\Box P_4$ is $4$.\\
\\
$r(x_1^{(1)}|D_1)=(0,2,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_1^{(2)}|D_1)=(1,3,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_1^{(3)}|D_1)=(2,4,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_1^{(4)}|D_1)=(3,5,6,3)$\\
$r(x_2^{(1)}|D_1)=(1,1,4,7),\,\,\,\,\,\,\,\,\,\,\,\,r(x_2^{(2)}|D_1)=(2,2,5,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_2^{(3)}|D_1)=(3,3,6,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_2^{(4)}|D_1)=(4,4,7,4)$\\
$r(x_3^{(1)}|D_1)=(2,0,5,8),\,\,\,\,\,\,\,\,\,\,\,\,r(x_3^{(2)}|D_1)=(3,1,6,7),\,\,\,\,\,\,\,\,\,\,\,\,r(x_3^{(3)}|D_1)=(4,2,7,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_3^{(4)}|D_1)=(5,3,8,5)$\\
$r(x_4^{(1)}|D_1)=(2,1,5,8),\,\,\,\,\,\,\,\,\,\,\,\,r(x_4^{(2)}|D_1)=(3,2,6,7),\,\,\,\,\,\,\,\,\,\,\,\,r(x_4^{(3)}|D_1)=(4,3,7,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_4^{(4)}|D_1)=(5,4,8,5)$\\
$r(x_5^{(1)}|D_1)=(1,2,4,7),\,\,\,\,\,\,\,\,\,\,\,\,r(x_5^{(2)}|D_1)=(2,3,5,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_5^{(3)}|D_1)=(3,4,6,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_5^{(4)}|D_1)=(4,5,7,4)$\\
\\
$r(x_6^{(1)}|D_1)=(1,3,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_6^{(2)}|D_1)=(2,4,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_6^{(3)}|D_1)=(3,5,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_6^{(4)}|D_1)=(4,6,5,2)$\\
$r(x_7^{(1)}|D_1)=(2,2,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_7^{(2)}|D_1)=(3,3,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_7^{(3)}|D_1)=(4,4,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_7^{(4)}|D_1)=(5,5,6,3)$\\
$r(x_8^{(1)}|D_1)=(3,1,4,7),\,\,\,\,\,\,\,\,\,\,\,\,r(x_8^{(2)}|D_1)=(4,2,5,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_8^{(3)}|D_1)=(5,3,6,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_8^{(4)}|D_1)=(6,4,7,4)$\\
$r(x_9^{(1)}|D_1)=(3,2,4,7),\,\,\,\,\,\,\,\,\,\,\,\,r(x_9^{(2)}|D_1)=(4,3,5,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_9^{(3)}|D_1)=(5,4,6,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_9^{(4)}|D_1)=(6,5,7,4)$\\
$r(x_{10}^{(1)}|D_1)=(2,3,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{10}^{(2)}|D_1)=(3,4,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{10}^{(3)}|D_1)=(4,5,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{10}^{(4)}|D_1)=(5,6,6,3)$\\
\\
$r(x_{11}^{(1)}|D_1)=(2,4,1,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{11}^{(2)}|D_1)=(3,5,2,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{11}^{(3)}|D_1)=(4,6,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{11}^{(4)}|D_1)=(5,7,4,3)$\\
$r(x_{12}^{(1)}|D_1)=(3,3,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{12}^{(2)}|D_1)=(4,4,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{12}^{(3)}|D_1)=(5,5,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{12}^{(4)}|D_1)=(6,6,5,2)$\\
$r(x_{13}^{(1)}|D_1)=(4,2,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{13}^{(2)}|D_1)=(5,3,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{13}^{(3)}|D_1)=(6,4,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{13}^{(4)}|D_1)=(7,5,6,3)$\\
$r(x_{14}^{(1)}|D_1)=(4,3,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{14}^{(2)}|D_1)=(5,4,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{14}^{(3)}|D_1)=(6,5,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{14}^{(4)}|D_1)=(7,6,6,3)$\\
$r(x_{15}^{(1)}|D_1)=(3,4,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{15}^{(2)}|D_1)=(4,5,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{15}^{(3)}|D_1)=(5,6,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{15}^{(4)}|D_1)=(6,7,5,2)$\\
\\
\\
$r(x_{16}^{(1)}|D_1)=(3,5,0,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{16}^{(2)}|D_1)=(4,6,1,2),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{16}^{(3)}|D_1)=(5,7,2,1),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{16}^{(4)}|D_1)=(6,8,3,0)$\\
$r(x_{17}^{(1)}|D_1)=(4,4,1,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{17}^{(2)}|D_1)=(5,5,2,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{17}^{(3)}|D_1)=(6,6,3,2),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{17}^{(4)}|D_1)=(7,7,4,1)$\\
$r(x_{18}^{(1)}|D_1)=(5,3,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{18}^{(2)}|D_1)=(6,4,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{18}^{(3)}|D_1)=(7,5,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{18}^{(4)}|D_1)=(8,6,5,2)$\\
$r(x_{19}^{(1)}|D_1)=(5,4,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{19}^{(2)}|D_1)=(6,5,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{19}^{(3)}|D_1)=(7,6,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{19}^{(4)}|D_1)=(8,7,5,2)$\\
$r(x_{20}^{(1)}|D_1)=(4,5,1,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{20}^{(2)}|D_1)=(5,6,2,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{20}^{(3)}|D_1)=(6,7,3,2),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{20}^{(4)}|D_1)=(7,8,4,1)$\\
\end{example}
\begin{remark}\label{u.1}
Consider the cartesian product $C_n\Box P_k$. If $n\geq4$ is an even integer,
then every pair of various vertices in $C_n\Box P_k$ cannot be a resolving set for $C_n\Box P_k$.
\end{remark}
\begin{lemma}\label{u.2}
Consider the cartesian product $C_n\Box P_k$. If $n\geq 4$ is an even integer, then the minimum size of a doubly resolving set of vertices for the cartesian product $C_n\Box P_k$ is $4$.
\end{lemma}
\begin{proof}
In the following cases we show that the minimum size of a doubly resolving set of vertices for the cartesian product $C_n\Box P_k$ is $4$.\\
Case 1. Contrary to Theorem \ref{d.1}, if $n\geq 4$ is an even integer then every minimum resolving set of vertices in $C_n\Box P_k$ may be lie in congruous layers $V_1$, $V_k$, or $V_1\cup V_k$.
Also, based on Theorem \ref{b.7}, we know that the minimum size of a resolving set in the cartesian product $C_n\Box P_k$ is $3$.
If $E_1$ is an arranged subset of vertices of $C_n\Box P_k$ so that $E_1$ is a minimal resolving set for $C_n\Box P_k$ and two elements of $E_1$ lie in the layer $V_1$ and one element of $E_1$ lie in the layer $V_k$, then without loss of generality we can consider $E_1=\{x_1, x_2, x_c \}$, where $x_c$ is a compatible vertex with respect to $x_1$ and $x_c$ lie in the layer $V_k$ of $C_n\Box P_k$.
Besides, the arranged subset $E_1=\{x_1, x_2, x_c \}$ of vertices of $C_n\Box P_k$ cannot be a doubly resolving set for $C_n\Box P_k$, because $n$ is an even integer and each layer $V_p$, $1\leq p\leq k$ of $C_n\Box P_k$ is isomorphic to the cycle $C_n$, and hence there are two adjacent vertices in each layer $V_p$ of $C_n\Box P_k$ say $x_i$ and $x_j$, $i<j$, so that $r(x_i|E_1)-r(x_j|E_1)=-I$, where $I$ indicates the unit $3$-vector $(1, 1, 1)$.
Thus the arranged subset $E_1=\{x_1, x_2, x_c \}$ cannot be a doubly resolving set for $C_n\Box P_k$.
If $E_2$ is an arranged subset of vertices of $C_n\Box P_k$ so that $E_2$ is a minimal resolving set for $C_n\Box P_k$ and all the elements of $E_2$ lie in exactly one of the congruous layers $V_1$ or $V_k$, then without loss of generality we can consider $E_2=\{x_1, x_{\frac{n}{2}}, x_{{\frac{n}{2}}+1} \}$. Besides, the arranged subset $E_2=\{x_1, x_{\frac{n}{2}}, x_{{\frac{n}{2}}+1} \}$ of vertices in the layer $V_1$ of the cartesian product $C_n\Box P_k$ cannot be a doubly resolving set for $C_n\Box P_k$, because if we consider two compatible vertices $x_e\in V_p$ and $x_d\in V_q$ in $C_n\Box P_k$, $1\leq p< q\leq k$, then there is a positive integer $\lambda$ so that $r(x_e|E_2)-r(x_d|E_2)=-\lambda I$, where $I$ indicates the unit $3$-vector $(1, 1, 1)$, and hence the arranged subset $E_2=\{x_1, x_{\frac{n}{2}}, x_{{\frac{n}{2}}+1} \}$ cannot be a doubly resolving set for $C_n\Box P_k$. Thus the minimum size of a doubly resolving set of vertices for the cartesian product $C_n\Box P_k$ is greater than $3$.\\
Case 2. Now, we show that the minimum size of a doubly resolving set of vertices in the cartesian product $C_n\Box P_k$ is $4$.
Let $x_c$ be a compatible vertex with respect to $x_1$, where $x_c$ lie in the layer $V_k$ of $C_n\Box P_k$ then we can show that
the arranged subset $E_3=E_2\cup x_c=\{x_1, x_{\frac{n}{2}}, x_{{\frac{n}{2}}+1}, x_c\}$ of vertices in the cartesian product $C_n\Box P_k$ is
one of the minimum doubly resolving sets for the cartesian product $C_n\Box P_k$. It will be enough to show that for any compatible vertices $x_e$
and $x_d$ in $C_n\Box P_k$, $r(x_e|E_3)-r(x_d|E_3)\neq\lambda I$. Suppose $x_e\in V_p$ and $x_d\in V_q$ are compatible vertices in the
cartesian product $C_n\Box P_k$, $1\leq p<q\leq k$. Hence, $r(x_e|E_2)-r(x_d|E_2)=-\lambda I$, where $\lambda$ is a positive integer, and $I$
indicates the unit $3$-vector $(1,1, 1)$. Also, for the compatible vertex $x_c$ with respect to $x_1$, $r(x_e|x_c)-r(x_d|x_c)=\lambda$.
So, $r(x_e|E_3)-r(x_d|E_3)\neq\lambda I$, where $I$ indicates the unit $4$-vector $(1,..., 1)$.\newline
\end{proof}
\begin{theorem}\label{u.3}
Consider the cartesian product $C_n\Box P_k$. If $n\geq4$ is an even integer,
then the minimum size of a resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is $4$.
\end{theorem}
\begin{proof}
Suppose $(C_n\Box P_k)\Box P_m$ is a graph with vertex set $\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}\cup ... \cup\{x_1^{(m)}, ..., x_{nk}^{(m)}\}$ so that for $1\leq t \leq nk$, the vertex $x_t^{(r)}$ is adjacent to $x_t^{(r+1)}$ in $(C_n\Box P_k)\Box P_m$, for $1\leq r\leq m-1$.
Hence, none of the minimal resolving sets of $C_n\Box P_k$ cannot be a resolving set for $(C_n\Box P_k)\Box P_m$. Therefore, the minimum size of a resolving set of vertices in $(C_n\Box P_k)\Box P_m$ must be greater than $3$.
Now, we show that the minimum size of a resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is $4$.
Let $x_1^{(1)}$ be a vertex in the layer $V_1^{(1)}$ of $(C_n\Box P_k)\Box P_m$ and $x_c^{(1)}$ be a compatible vertex with respect to $x_1^{(1)}$, where $x_c^{(1)}$ lie in the layer $V_k^{(1)}$ of $(C_n\Box P_k)\Box P_m$. Based on Lemma \ref{u.2}, we know that
$1^{th}$ copy of the arranged subset $E_3=\{x_1, x_{\frac{n}{2}}, x_{{\frac{n}{2}}+1}, x_c\}$, denoted by the set
$E_3^{(1)}=\{x_1^{(1)}, x_{\frac{n}{2}}^{(1)}, x_{{\frac{n}{2}}+1}^{(1)}, x_c^{(1)}\}$ is one of the minimum doubly resolving sets for the cartesian product $C_n\Box P_k$. Besides, the vertex $x_t^{(r)}$ is adjacent to the vertex $x_t^{(r+1)}$ in $(C_n\Box P_k)\Box P_m$,
and hence the arranged set
$E_3^{(1)}=\{x_1^{(1)}, x_{\frac{n}{2}}^{(1)}, x_{{\frac{n}{2}}+1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_m$ is one of the resolving sets for $(C_n\Box P_k)\Box P_m$, because for each vertex $x_t^{(i)}$ of $(C_n\Box P_k)\Box P_m$, we have
$$r(x_t^{(i)}|E_3^{(1)})=( d(x_t^{(1)}, x_1^{(1)})+i-1, d(x_t^{(1)}, x_{\frac{n}{2}}^{(1)})+i-1, d(x_t^{(1)}, x_{{\frac{n}{2}}+1}^{(1)})+i-1, d(x_t^{(1)}, x_{c}^{(1)})+i-1),$$ so all the vertices $\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup ...\cup \{x_1^{(m)}, ..., x_{nk}^{(m)}\}$ of $(C_n\Box P_k)\Box P_m$ have various representations with respect to the set $E_3^{(1)}$. Thus the minimum size of a resolving set of vertices in $C_n\Box P_k\Box P_m$ is $4$.
\end{proof}
\begin{theorem}\label{u.4}
Consider the cartesian product $C_n\Box P_k$. If $n\geq4$ is an even integer,
then the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is $5$.
\end{theorem}
\begin{proof}
Suppose $(C_n\Box P_k)\Box P_m$ is a graph with vertex set $\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}\cup ... \cup\{x_1^{(m)}, ..., x_{nk}^{(m)}\}$ so that for $1\leq t \leq nk$, the vertex $x_t^{(r)}$ is adjacent to $x_t^{(r+1)}$ in $(C_n\Box P_k)\Box P_m$, for $1\leq r\leq m-1$. Based on the previous Theorem, we know that
the arranged set $E_3^{(1)}=\{x_1^{(1)}, x_{\frac{n}{2}}^{(1)}, x_{{\frac{n}{2}}+1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_m$ is one of the resolving sets for $(C_n\Box P_k)\Box P_m$, so that the arranged set $E_3^{(1)}=\{x_1^{(1)}, x_{\frac{n}{2}}^{(1)}, x_{{\frac{n}{2}}+1}^{(1)}, x_c^{(1)}\}$ of vertices of $(C_n\Box P_k)\Box P_m$ cannot be a doubly resolving set for $(C_n\Box P_k)\Box P_m$, and hence the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is greater than $4$.
Now, let $E_4=E_3^{(1)}\cup x_c^{(m)}=\{x_1^{(1)}, x_{\frac{n}{2}}^{(1)}, x_{{\frac{n}{2}}+1}^{(1)}, x_c^{(1)}, x_c^{(m)}\}$
be an arranged subset of vertices of $(C_n\Box P_k)\Box P_m$, where $x_c^{(m)}$ lie in the layer $V_k^{(m)}$ of $(C_n\Box P_k)\Box P_m$.
It will be enough to show that for every two vertices $x_t^{(r)}$ and $x_t^{(s)}$, $1\leq t\leq nk$, $1\leq r<s\leq m$, $r(x_t^{(r)}|E_4)-r(x_t^{(s)}|E_4)\neq -\lambda I$, where $I$ indicates the unit $5$-vector $(1, ..., 1)$ and $\lambda$ is a positive integer.
For this purpose, let the distance between two the vertices $x_t^{(r)}$ and $x_t^{(s)}$ in $(C_n\Box P_k)\Box P_m$ is $\lambda$, then
we can verify that, $r(x_t^{(r)}|E_3^{(1)})-r(x_t^{(s)}|E_3^{(1)})= - \lambda I$, where $I$ indicates the unit $4$-vector, and $r(x_t^{(t)}|x_c^{(m)})-r(x_t^{(s)}|x_c^{(m)})= \lambda$. Therefore, the arranged subset $E_4$ is one of the minimum doubly resolving sets of vertices in $(C_n\Box P_k)\Box P_m$. Thus the minimum size of a doubly resolving set of vertices in $(C_n\Box P_k)\Box P_m$ is $5$.
\end{proof}
\begin{example}
Consider graph $(C_4\Box P_3)\Box P_4$ with vertex set $\{x_1^{(1)}, ..., x_{12}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{12}^{(2)}\}\cup\{x_1^{(3)}, ..., x_{12}^{(3)}\}\cup\{x_1^{(4)}, ..., x_{12}^{(4)}\}$, we can see that the set $E=\{x_1^{(1)}, x_2^{(1)} x_3^{(1)}, x_{9}^{(1)}, x_{9}^{(4)} \}$ of vertices of $(C_4\Box P_3)\Box P_4$ is one of the minimum doubly resolving sets for $(C_4\Box P_3)\Box P_4$, and hence the minimum size of a doubly resolving set of vertices in $(C_4\Box P_3)\Box P_4$ is $4$.\\
\\
$r(x_1^{(1)}|E)=(0,1,2,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_1^{(2)}|E)=(1,2,3,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_1^{(3)}|E)=(2,3,4,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_1^{(4)}|E)=(3,4,5,5,2)$\\
$r(x_2^{(1)}|E)=(1,0,1,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_2^{(2)}|E)=(2,1,2,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_2^{(3)}|E)=(3,2,3,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_2^{(4)}|E)=(4,3,4,6,3)$\\
$r(x_3^{(1)}|E)=(2,1,0,4,7),\,\,\,\,\,\,\,\,\,\,\,\,r(x_3^{(2)}|E)=(3,2,1,5,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_3^{(3)}|E)=(4,3,2,6,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_3^{(4)}|E)=(5,4,3,7,4)$\\
$r(x_4^{(1)}|E)=(1,2,1,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_4^{(2)}|E)=(2,3,2,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_4^{(3)}|E)=(3,4,3,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_4^{(4)}|E)=(4,5,4,6,3)$\\
\\
$r(x_5^{(1)}|E)=(1,2,3,1,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_5^{(2)}|E)=(2,3,4,2,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_5^{(3)}|E)=(3,4,5,3,2),\,\,\,\,\,\,\,\,\,\,\,\,r(x_5^{(4)}|E)=(4,5,6,4,1)$\\
$r(x_6^{(1)}|E)=(2,1,2,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_6^{(2)}|E)=(3,2,3,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_6^{(3)}|E)=(4,3,4,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_6^{(4)}|E)=(5,4,5,5,2)$\\
$r(x_7^{(1)}|E)=(3,2,1,3,6),\,\,\,\,\,\,\,\,\,\,\,\,r(x_7^{(2)}|E)=(4,3,2,4,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_7^{(3)}|E)=(5,4,3,5,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_7^{(4)}|E)=(6,5,4,6,3)$\\
$r(x_8^{(1)}|E)=(2,3,2,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_8^{(2)}|E)=(3,4,3,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_8^{(3)}|E)=(4,5,4,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_8^{(4)}|E)=(5,6,5,5,2)$\\
\\
$r(x_9^{(1)}|E)=(2,3,4,0,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_9^{(2)}|E)=(3,4,5,1,2),\,\,\,\,\,\,\,\,\,\,\,\,r(x_9^{(3)}|E)=(4,5,6,2,1),\,\,\,\,\,\,\,\,\,\,\,\,r(x_9^{(4)}|E)=(5,6,7,3,0)$\\
$r(x_{10}^{(1)}|E)=(3,2,3,1,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{10}^{(2)}|E)=(4,3,4,2,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{10}^{(3)}|E)=(5,4,5,3,2),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{10}^{(4)}|E)=(6,5,6,4,1)$\\
$r(x_{11}^{(1)}|E)=(4,3,2,2,5),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{11}^{(2)}|E)=(5,4,3,3,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{11}^{(3)}|E)=(6,5,4,4,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{11}^{(4)}|E)=(7,6,5,5,2)$\\
$r(x_{12}^{(1)}|E)=(3,4,3,1,4),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{12}^{(2)}|E)=(4,5,4,2,3),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{12}^{(3)}|E)=(5,6,5,3,2),\,\,\,\,\,\,\,\,\,\,\,\,r(x_{12}^{(4)}|E)=(6,7,6,4,1)$\\
\end{example}
\begin{theorem}\label{w.1}
If $n$ is an even or odd integer is greater than or equal to $3$, then the minimum size of a strong resolving set of vertices for the cartesian product $(C_n\Box P_k)\Box P_m$ is $2n$.
\end{theorem}
\begin{proof}
Suppose that $(C_n\Box P_k)\Box P_m$ is a graph with vertex set $\{x_1^{(1)}, ..., x_{nk}^{(1)}\}\cup \{x_1^{(2)}, ..., x_{nk}^{(2)}\}\cup ... \cup\{x_1^{(m)}, ..., x_{nk}^{(m)}\}$ so that for $1\leq t \leq nk$, the vertex $x_t^{(r)}$ is adjacent to $x_t^{(r+1)}$ in $(C_n\Box P_k)\Box P_m$, for $1\leq r\leq m-1$. We know that each vertex of the layer $V_1^{(1)}$ is maximally distant from a vertex of the layer $V_k^{(m)}$ and
each vertex of the layer $V_k^{(m)}$ is maximally distant from a vertex of the layer $V_1^{(1)}$. In particular, each vertex of the layer $V_1^{(m)}$ is maximally distant from a vertex of the layer $V_k^{(1)}$ and
each vertex of the layer $V_k^{(1)}$ is maximally distant from a vertex of the layer $V_1^{(m)}$, and hence the minimum size of a strong resolving set of vertices for the cartesian product $(C_n\Box P_k)\Box P_m$ is equal or greater than $2n$, because it is well known that for
every pair of mutually maximally distant vertices $u$ and $v$ of a connected graph $G$ and for every strong metric
basis $S$ of $G$, it follows that $u\in S$ or $v\in S$.
Suppose the set $\{x_1^{(1)}, ..., x_n^{(1)}\}$ is an arranged subset of vertices in the layer $V_1^{(1)}$ of the cartesian product $(C_n\Box P_k)\Box P_m$ and suppose that the set $\{x_1^{(m)}, ..., x_n^{(m)}\}$ is an arranged subset of vertices in the layer $V_1^{(m)}$ of the cartesian product $(C_n\Box P_k)\Box P_m$. Now, let $T=\{x_1^{(1)}, ..., x_n^{(1)}\}\cup \{x_1^{(m)}, ..., x_n^{(m)}\}$ be an arranged subset of vertices of the cartesian product $(C_n\Box P_k)\Box P_m$. In the following cases we show that the arranged set $T$, defined already, is one of the minimum strong resolving sets of vertices for the cartesian product $(C_n\Box P_k)\Box P_m$. For this purpose let $x_e^{(i)}$ and $x_d^{(j)}$ be two various vertices of $(C_n\Box P_k)\Box P_m$, $1\leq i,j\leq m$, $1\leq e,d\leq nk$ and $1\leq r\leq n$.\\
Case 1. If $i=j$ then $x_e^{(i)}$ and $x_d^{(i)}$ lie in $i^{th}$ copy of $(C_n\Box P_k)$ with vertex set $\{x_1^{(i)}, ..., x_{nk}^{(i)}\}$ so that $i^{th}$ copy of $(C_n\Box P_k)$ is a subgraph of $(C_n\Box P_k)\Box P_m$. Since $i=j$ then we can assume that $e<d$, because $x_e^{(i)}$ and $x_d^{(i)}$ are various vertices.\\
Case 1.1. If both vertices $x_e^{(i)}$ and $x_d^{(i)}$ are compatible in $i^{th}$ copy of $(C_n\Box P_k)$ relative to $x_r^{(i)}\in V_1^{(i)}$, then there is the vertex $x_r^{(1)}\in V_1^{(1)}\subset T$ so that $x_e^{(i)}$ belongs to shortest path $x_r^{(1)}-x_d^{(i)}$, say as $x_r^{(1)},..., x_r^{(i)},..., x_e^{(i)}, ..., x_d^{(i)}$.\\
Case 1.2. Suppose both vertices $x_e^{(i)}$ and $x_d^{(i)}$ are not compatible in $i^{th}$ copy of $(C_n\Box P_k)$, and lie
in various layers or lie in the same layer in $i^{th}$ copy of $(C_n\Box P_k)$, also let
$x_r^{(i)}\in V_1^{(i)}$, be a compatible vertex relative to $x_e^{(i)}$. Hence there is
the vertex $x_r^{(m)}\in V_1^{(m)}\subset T$ so that
$x_e^{(i)}$ belongs to shortest path $x_r^{(m)}-x_d^{(i)}$,
say as $x_r^{(m)},..., x_r^{(i)},..., x_e^{(i)}, ..., x_d^{(i)}$.\\
Case 2. If $i\neq j$ then $x_e^{(i)}$ lie in $i^{th}$ copy of $(C_n\Box P_k)$ with vertex set $\{x_1^{(i)}, ..., x_{nk}^{(i)}\}$, and $x_d^{(j)}$ lie in $j^{th}$ copy of $(C_n\Box P_k)$ with vertex set $\{x_1^{(j)}, ..., x_{nk}^{(j)}\}$. In this case we can assume that $i<j$.\\
Case 2.1. If $e=d$ and $x_r^{(j)}\in V_1^{(j)}$ is a compatible vertex relative to $x_d^{(j)}$, then there is the vertex $x_r^{(m)}\in V_1^{(m)}\subset T$ so that $x_d^{(j)}$ belongs to shortest path $x_r^{(m)}-x_e^{(i)}$, say as $x_r^{(m)},..., x_r^{(j)},..., x_d^{(j)}, ..., x_e^{(i)}$.\\
Case 2.2. If $e< d$, also $x_e^{(i)}$ and $x_d^{(j)}$ lie in various layers of $(C_n\Box P_k)\Box P_m$ or $x_e^{(i)}$ and $x_d^{(j)}$ lie in the same layer of $(C_n\Box P_k)\Box P_m$ and $x_r^{(i)}\in V_1^{(i)}$ is a compatible vertex relative to $x_e^{(i)}$, then there
is the vertex $x_r^{(1)}\in V_1^{(1)}\subset T$ so that $x_e^{(i)}$ belongs to shortest path $x_r^{(1)}-x_d^{(j)}$, say as $x_r^{(1)},..., x_r^{(i)},..., x_e^{(i)}, ..., x_d^{(j)}$.\\
Case 2.3. If $e> d$, also $x_e^{(i)}$ and $x_d^{(j)}$ lie in various layers of $(C_n\Box P_k)\Box P_m$ or $x_e^{(i)}$ and $x_d^{(j)}$ lie in the same layer of $(C_n\Box P_k)\Box P_m$ and $x_r^{(j)}\in V_1^{(j)}$ is a compatible vertex relative to $x_d^{(j)}$, then there
is the vertex $x_r^{(m)}\in V_1^{(m)}\subset T$ so that $x_d^{(j)}$ belongs to shortest path $x_r^{(m)}-x_e^{(i)}$, say as
$x_r^{(m)},..., x_r^{(j)},..., x_d^{(j)}, ..., x_e^{(i)}$.
\end{proof}
\noindent \textbf{3.2 \,\ Some resolving sets for the graph $H(n)$ and the Line graph of the graph $H(n)$}\\
Let $n$ be a natural number greater than or equal to $5$ and $[n]=\{1, 2, ..., n\}$. The graph $H(n)$ is a graph with vertex set $V = V_1\cup V_2$, where
$V_1=\{v_1,v_2, ..., v_n\}=\{v_r\,| \, r\in [n]\}$, $V_2=\{v_i v_j\,| \, i,j \in[n], i\neq j, i<j,1\leq i\leq n-1, 2\leq j\leq n\}$,
and the edge set of $H(n)$ is
$E = \{\{v_r, v_i v_j\}\,| \, v_r\in V_1, v_iv_j\in V_2 , v_r=v_i \,\,\ or \,\,\ v_r=v_j\}.$ Note that for simply we use refinement of the natural relabelling of the graph $H(n)$ which is defined in [27].
Now we undertake the necessary task of introducing some of the basic notation for this class of graphs.
Based on definition of the vertex set $V_2$ of $H(n)$, the vertex $v_iv_j\in V_2$ if $i<j$ and hence if $v_iv_j\in V_2$ then $v_jv_i\notin V_2$.
In particular, two vertices $v_iv_j$ and $v_pv_q$ are identical if and only if $i=p$ and $j=q$.
We say that two distinct vertices $v_iv_j$ and $v_pv_q$ from $V_2$ are left-invariant in the graph $H(n)$, if $v_i=v_p$. Also, we say that two distinct vertices $v_iv_j$ and $v_pv_q$ from $V_2$ are right-invariant in the graph $H(n)$ if $v_j=v_q$.
Now, suppose that $G$ is a graph with vertex set
$W_1\cup ...\cup W_n$, where for $1\leq r\leq n$ we take $W_r=\{\{v_r, v_iv_j\}\,\ | \,\ i,r,s \in [n], i\neq j, v_r=v_i \,\ \text{or} \,\ v_r=v_j \}$, and we say that two various vertices
$\{v_r, v_iv_j\}$ and $\{v_k, v_pv_q\}$ are adjacent in $G$ if and only if $v_r=v_k$ or $v_iv_j=v_pv_q$. It is not hard to see that this family of graphs is isomorphic with the line graph of the graph $H(n)$, and hence is indicated by $L(n)$, where $H(n)$, is defined above. We can see that $L(n)$ is a connected vertex transitive graph of valency $n-1$, with diameter 3, and the order $n(n-1)$. It is easy to see that every $W_r$ is a maximal clique of size $n-1$ in the graph $L(n)$. We also, undertake the necessary task of introducing some of the basic notation for this class of graphs. We say that two maximal cliques $W_r$ and $W_k$ are adjacent in $L(n)$, if there is a vertex in maximal clique $W_r$ so that this vertex is adjacent to exactly one vertex of maximal clique $W_k$, $r,k \in [n], r\neq k$. Also, for any maximal clique $W_r$ in $G=L(n)$ we use $N(W_r)=\bigcup_{w\in W_r}N_G(w)$ to indicate the vertices in the all maximal cliques $W_k$, say $w_k$,
$1\leq k\leq n$ and $k\neq r$ so that $w_k$ is adjacent one vertex of the maximal clique $W_r$. In this section we study some resolving sets for the graph $H(n)$ and the Line graph of the graph $H(n)$. For more results in resolving sets of line graph see [28, 29].
\begin{proposition}\label{h.1}
Suppose that $n$ is a natural number greater than or equal to $5$. Then each subset of $V_1$ of size $n-1$ in graph $H(n)$ is a doubly resolving set for $H(n)$.
\end{proposition}
\begin{proof}
Suppose that $V(H(n))=V_1 \cup V_2$, where $V_1=\{v_1, ..., v_n\}=\{v_r\,| \, r\in [n]\}$, $V_2=\{v_iv_j\,| \, i,j \in[n], i\neq j, i<j,1\leq i\leq n-1, 2\leq j\leq n\}$, is defined already.
It is straightforward to verify that the distance between two distinct vertices in $V_1$ is equal to $2$, and none of the subsets of $V_1$ of size at most $n-2$ cannot be a resolving set for $H(n)$. In particular, we can show that each subset of $V_1$ of size $n-1$ in graph $H(n)$ is a resolving set for $H(n)$. Now, let $R_1$ be an arranged subset of $V_1$ of size $n-1$.
Without loss of generality we may take $R_1=\{v_1,v_2, ..., v_{n-1}\}$.
We show that $R_1$ is a doubly resolving set for graph $H(n)$.
It will be enough to show that for two distinct vertices $u$ and $v$ from $V(H(n))- R_1$ there are elements $x$ and $y$ from $R_1$ so that
$d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$. Consider two distinct vertices $u$ and $v$ from $V(H(n))- R_1$, then we have the following:\newline
Case 1. Suppose, both vertices $u$ and $v$ belong to $V_2$, so that $u$ and $v$ are left-invariant. So we can assume that
$u=v_iv_j$ and $v=v_iv_q$, where $i,j, q \in[n]$, $j\neq q$ and $i<j,q$. In this case if we consider $x=v_j$ and $y=v_q$, then we have $d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.\newline
Case 2. Suppose, both vertices $u$ and $v$ belong to $V_2$, so that $u$ and $v$ are right-invariant. So we can assume that
$u=v_iv_j$ and $v=v_pv_j$, where $i,j, p \in[n]$, $i\neq p$ and $i, p<j$. In this case if we consider $x=v_i$ and $y=v_p$, then we have $d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.\newline
Case 3. Suppose, both vertices $u$ and $v$ belong to $V_2$, so that these vertices are not, left-invariant and right-invariant. So we can assume that $u=v_iv_j$ and $v=v_pv_q$, where $i,j,p, q \in[n]$, $i\neq p$ and $j\neq q$. In this case if we consider $x=v_i$ and $y=v_p$, then we have $d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.\newline
Case 4. Now, suppose that $u=v_n\in V_1$ and $ v=v_iv_j \in V_2$, where $i,j \in[n]$, and $i<j$. In this case, may be $j=n$ or $j\neq n$. If we consider $x=v_i$ and $y=v_p$, $p< i$, then we have
$d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.\newline
\end{proof}
\begin{proposition}\label{h.2}
Suppose that $n$ is a natural number greater than or equal to $5$. Then any subset of $V_2$ of size $n-2$ in graph $H(n)$ so that the distance between two distinct vertices in that set is equal $2$, cannot be a doubly resolving set for $H(n)$.
\end{proposition}
\begin{proof}
Suppose that $V(H(n))=V_1 \cup V_2$, where $V_1=\{v_1, ..., v_n\}=\{v_r\,| \, r\in [n]\}$, $V_2=\{v_iv_j\,| \, i,j \in[n], i\neq j, i<j,1\leq i\leq n-1, 2\leq j\leq n\}$, is defined already.
Also, it is straightforward to verify that the distance between two distinct vertices in $V_2$ is equal to $2$ or $4$.
Now, let $R_2$ be an arranged subset of $V_2$ of size $n-2$
so that the distance between two distinct vertices in $R_2$ is equal $2$.
Without loss of generality we may take $R_2=\{v_1v_2, v_1v_3,..., v_1v_{n-1}\}$.
In particular, we can show that the arranged subset $R_2$ of $V_2$ is a resolving set for $H(n)$, although this subset cannot be a doubly resolving set for $H(n)$. Because if we consider $u=v_1$ and $v=v_n$, then for any elements $x$ and $y$ in $R_2$ we have $0=1-1=d(u, x) - d(u, y) = d(v, x) - d(v, y)=3-3=0$. Therefore, the arranged subset $R_2$ cannot be a doubly resolving set for the graph $H(n)$, and so we can verify that any subset of $V_2$ of size $n-2$, so that the distance between two distinct vertices in such set is $2$, cannot be a doubly resolving set for $H(n)$.\newline
\end{proof}
\begin{theorem}\label{h.3}
Suppose that $n$ is a natural number greater than or equal to $6$. If $3|n$ then the minimum size of a resolving set for graph $H(n)$ is $n-\frac{n}{3}$.
\end{theorem}
\begin{proof}
Suppose that $V(H(n))=V_1 \cup V_2$, where $V_1=\{v_1, ..., v_n\}=\{v_r\,| \, r\in [n]\}$, $V_2=\{v_iv_j\,| \, i,j \in[n], i\neq j, i<j,1\leq i\leq n-1, 2\leq j\leq n\}$, is defined already.
For $1\leq i\leq n-1$, if we take $T_i= \cup_{j=i+1} ^n\{v_iv_j\}$ then we can see that $V_2=T_1\cup T_2\cup ...\cup T_{n-1}$, also it is not hard to see that $|T_i|=n-i$, in particular we can verify that $T_1$ and $T_2$ are resolving sets for $H(n)$. Now, for $1\leq i\leq n-2$ if we take
$P_i= \cup_{j=i+1} ^{i+2}\{v_iv_j\}$ then we can view that $P_i$ is a subset of $T_i$ of size $2$. Since $3|n$, this implies that there is an element $k\in \mathbb{N}$ such that $n=3k$, and so for
$1\leq t\leq k$ if we consider $i=3t-2$ and take $P=\cup_{t=1} ^{k}\{P_{3t-2}\}$, where $P_i$ is defined already, then not only we can verify that $P$ is a resolving set for $H(n)$, but also $P$ is a minimal resolving set for $H(n)$ because the cardinality of any $P_i$ is $2$, and none of subsets of $P$ of size less than $2k$ cannot be a resolving set for graph $H(n)$. Indeed, there are exactly
$k$ subsets $P_i$ of $T_i$ of size $2$, so that $P=\cup_{t=1} ^{k}\{P_{3t-2}\}$ is a minimal resolving set of size $n-\frac{n}{3}$ for $H(n)$.\newline
\end{proof}
\begin{example}\label{h.4}
Consider graph $H(12)$. We can verify that the subset
$$P_1\cup P_4\cup P_7 \cup P_{10}=\{v_1v_2, v_1v_3, v_4v_5, v_4v_6, v_7v_8, v_7v_9, v_{10}v_{11},v_{10}v_{12}\},$$ where $P_i$ is defined in the previous Theorem is a minimal resolving set for $H(12)$.
\end{example}
\begin{corollary}\label{h.5}
Suppose that $n$ is a natural number greater than or equal to $6$. If $n=3k$ then $2k<\beta(H(n+1))<\beta(H(n+2))\leq 2(k+1)$.
\end{corollary}
\begin{lemma}\label{l.1}
Consider graph $L(n)$ with vertex set $W_1\cup ...\cup W_n$ for $n\geq 5$. Then for $1\leq r\leq n$ each subset of $N(W_r)$ of size at least $n-2$ can be a resolving set for $L(n)$.
\end{lemma}
\begin{proof}
Suppose that $V(L(n))=W_1\cup ...\cup W_n$, where for $1\leq r\leq n$ the set $W_r=\{\{v_r, v_iv_j\}\,\ | \,\ i,r,s \in [n], i\neq j, v_r=v_i \,\ \text{or} \,\ v_r=v_j \}$ to indicate a maximal clique of size $n-1$ in the graph $L(n)$.
We know that $N(W_r)$ to indicate the vertices in the all maximal cliques $W_k$, say $w_k$, $1\leq k\leq n$ and $k\neq r$ so that $w_k$ is adjacent one vertex of the maximal clique $W_r$, also we can see that the cardinality of $N(W_r)$ is $n-1$.
Since $L(n)$ is a vertex transitive graph, then without loss of generality we may consider the maximal clique $W_1$. Hence
$N(W_1)= \{y_2, ... , y_{n-1}, y_n\}$, where for $2\leq k\leq n$ we have $y_k= \{v_k, v_1v_k\}\in W_k$. Based on the following cases
it will be enough to show that the arranged set $N(W_1)-\{y_{n-1}, y_n\}$ of size $n-3$ cannot be a resolving set for $L(n)$ and the arranged set of vertices $N(W_1)-y_n$ of size $n-2$ is a resolving set for $L(n)$.\newline
Case 1.
First, we show that any subset of $N(W_1)$ of size $n-3$ cannot be a resolving set for $L(n)$.
Without loss of generality if we consider $C_1=N(W_1)-\{y_{n-1}, y_n\}= \{y_2, ... , y_{n-2}\}$, then there are exactly two vertices
$\{v_1, v_1v_{n-1}\}, \{v_1, v_1v_n\}\in W_1$ so that
$r(\{v_1, v_1v_{n-1}\}|C_1)=r(\{v_1, v_1v_n\}|C_1)= (2,..., 2)$. Thus, $C_1$ cannot be a resolving set for $L(n)$, and so
any subset of $N(W_r)$ of size $n-3$ cannot be a resolving set for $L(n)$.\newline
Case 2.
Now, we take $C_2=N(W_1)-y_n=\{y_2, ... , y_{n-1}\}$ and show that all the vertices in $V(L(n))- C_2$ have different representations relative to $C_2$.
In this case the vertex $\{v_1, v_1v_{n-1}\}\in W_1$ is adjacent to the vertex $y_{n-1}\in W_{n-1}$, and hence
$r(\{v_1, v_1v_{n-1}\}|C_2)\neq r(\{v_1, v_1v_n\}|C_2)$. In particular every vertex $w$ in the maximal clique $ W_1$ is adjacent to exactly a vertex of each maximal clique $W_j$, $2\leq j\leq n$. So, all the vertices $ w\in W_1$ have various metric representations relative to the subset $C_2$.
Also, for every vertex $w\in W_r$, $2\leq r\leq n-1$ so that $w\notin N(W_1)$ and each $y_s\in C_2 $, $2\leq s\leq n-1$, if $w, y_s$ lie in a maximal clique $ W_s$, $2\leq s\leq n-1$, then we have $d(w, y_s)=1$; otherwise $d(w, y_s)\geq 2$.
In particular, all the vertices in the maximal clique $W_n$ have various metric representations relative to the subset $C_2$ because for every vertex $w$ in the maximal clique $W_n$ so that $w$ is not equal to the vertex $\{v_n, v_1v_n\}$ in the maximal clique $W_n$, there is exactly one element $y_s\in C_2$ such that $d(w, y_s)=2$; otherwise $d(w, y_s)>2$, $2\leq s\leq n-1$. In particular, for the vertex $y_n=\{v_n, v_1v_n\}$ in the maximal clique $W_n$ and every element $y_s\in C_2$ we have $d(w, y_s)=3$. Thus, the arranged subset $C_2= \{y_2, ... , y_{n-1}\}$ of vertices in $L(n)$ is a resolving set for $L(n)$ of size $n-2$, and so each subset of $N(W_r)$ of size $n-2$ is a resolving set for $L(n)$.\newline
\end{proof}
\begin{theorem}\label{l.2}
Suppose $n$ is a natural number greater than or equal to $5$, then the minimum size of a resolving set in graph $L(n)$ is $n-2$.
\end{theorem}
\begin{proof}
Suppose that $V(L(n))=W_1\cup ...\cup W_n$, where for $1\leq r\leq n$ the set $W_r=\{\{v_r, v_iv_j\}\,\ | \,\ i,r,s \in [n], i\neq j, v_r=v_i \,\ \text{or} \,\ v_r=v_j \}$ to indicate a maximal clique of size $n-1$ in the graph $L(n)$.
Let $D_1=\{W_1, W_2, ..., W_k\}$ be a subset of vertices of $L(n)$, consisting of some of the maximal cliques of $L(n)$, and let $D_2$ be a subset of vertices of $L(n)$, consisting of exactly three maximal cliques of $L(n)$ so that none of the vertices of $D_2$ belong to $D_1$.
Without loss of generality we may take $D_1=\{W_1, W_2, ..., W_{n-3}\}$ and $D_2=\{W_{n-2}, W_{n-1}, W_{n}\}$. Now, let $D_3$ be a subset of $D_2$, consisting of exactly one maximal clique of $D_2$, say $W_{n}$, and let $D_3=\{W_n\}$. Thus there are exactly two distinct vertices in $D_3=\{W_n\}$ say $x$ and $y$ so that $x$ is adjacent to a vertex of $W_{n-1}$ and $y$ is adjacent to a vertex of $W_{n-2}$, and hence the metric representations of two vertices $x$ and $y$ are identical relative to $D_1$. So if the arranged set $D_4=\{w_1, w_2, ..., w_l\}$ of vertices of graph $L(n)$ so that $w_r\in W_r$ is a resolving set for graph $L(n)$, then the cardinality of $D_2$ must be less than or equal $2$, or the cardinality of $D_4$ must be greater than or equal $n-2$.
In particular, based on the previous Lemma the arranged subset $C_2=N(W_1)-y_n=\{y_2, ... , y_{n-1}\}$ of vertices in $L(n)$ is a resolving set for $L(n)$ of size $n-2$, and so the minimum size of a resolving set in graph $L(n)$ is $n-2$.
\newline
\end{proof}
\begin{lemma} \label{l.3}
Consider the graph $L(n)$ with vertex set $W_1\cup ...\cup W_n$ for $n\geq 5$. Then any subset of $N(W_r)$ of size $n-2$ cannot be a doubly resolving set for $L(n)$.
\end{lemma}
\begin{proof}
Since $L(n)$ is a vertex transitive graph, then without loss of generality we may consider the maximal clique $W_1$.
Hence if we take $C_2=N(W_1)-y_n=\{y_2, ..., y_{n-1}\}$, where for $2\leq k\leq n$ we have $y_k= \{v_k, v_1v_k\}\in V_k$, then
Based on Lemma \ref{l.1}, and Theorem \ref{l.2}, the subset
$C_2=N(W_1)-y_n=\{y_2, ..., y_{n-1}\}$ of vertices in $L(n)$ is a minimum resolving set for $L(n)$ of size $n-2$.
Now, by considering the vertices $u=\{v_1, v_1v_n\}\in W_1$ and $y_n=\{v_n, v_1v_n\}\in W_n$, we see that
$d(u, r) - d(u, s) = d(y_n, r) - d(y_n, s)$ for elements $r, s \in C_2$,
because for each element $z\in C_2$ we have $d(u, z)=2$ and $d(y_n, z)=3$. Thus the subset
$C_2=N(W_1)-y_n=\{y_2, ..., y_{n-1}\}$ of vertices in $L(n)$ cannot be a doubly resolving set for $L(n)$, and so
any subset $N(W_r)$ of graph $L(n)$ of size $n-2$ cannot be a doubly resolving set for $L(n)$.
\end{proof}
\begin{theorem}\label{l.4}
Suppose $n$ is a natural number greater than or equal to $5$, then the minimum size of a doubly resolving set in graph $L(n)$ is $n-1$.
\end{theorem}
\begin{proof}
Suppose that $V(L(n))=W_1\cup ...\cup W_n$, where $W_i=\{\{v_i, v_iv_j\}\,\ | \,\ i,j \in [n], i\neq j \}$ for $1\leq i\leq n$.
Based on Lemma \ref{l.1}, and Theorem \ref{l.2}, the subset
$C_2=N(W_1)-y_n=\{y_2, ..., y_{n-1}\}$ of vertices in $L(n)$ is a minimum resolving set for $L(n)$ of size $n-2$, where $y_k= \{v_k, v_1v_k\}\in W_k$ for $2\leq k\leq n$.
Also, from the previous Lemma we know that the subset $C_2$ is not a doubly resolving set for $L(n)$, and hence the minimum size of a doubly resolving set in $L(n)$ must be greater than or equal to $n-1$.
Now, if we take
$C_3=N(W_1)= \{y_2, ... , y_{n-1}, y_n\}$, where for $2\leq k\leq n$ we have $y_k= \{v_k, v_1v_k\}\in W_k$, then Based on Lemma \ref{l.1},
we know that the subset
$C_3=N(W_1)=\{y_2, ..., y_{n-1}, y_n\}$ of vertices in $L(n)$ is a resolving set for $L(n)$ of size $n-1$. We show that $C_3$ is a doubly resolving set for $L(n)$. It will be enough to show that for any two various vertices $u$ and $v$ in $L(n)$ there exist elements $x$ and $y$ from $C_3$ so that $d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$. Consider two vertices $u$ and $v$ in $L(n)$.
Then the result can be deduced from the following cases: \newline
Case 1.
Suppose, both vertices $u$ and $v$ lie in the maximal clique $W_1$. Hence, there exists an element $x\in C_3$ so that $x\in W_r$ and $x$ is adjacent to $u$, also, there exists an element $y\in C_3$ so that $y\in W_k$ and $y$ is adjacent to $v$ for some $r,k\in [n]-1$, $r\neq k$; and hence
$-1=1-2=d(u, x) - d(u, y) \neq d(v, x) - d(v, y)=2-1=1$.\newline
Case 2.
Suppose, both vertices $u$ and $v$ lie in the maximal clique $W_r$, $r\in[n]-1$, so that $u, v\notin C_3$. Hence, there exists an element $x\in C_3$ so that $x\in W_r$ and $d(u,x)=d(v, x)=1$, also there exists an element $y\in C_3$ so that $y\in W_k$, $r\neq k$, and $d(u, y)=2$, $d(v, y)=3$ or $d(u, y)=3$, $d(v, y)=2$. Thus $d(u, x) - d(u, y) \neq d(v, x) - d(v, y).$\newline
Case 3.
Suppose that $u$ and $v$ are two distinct vertices in $L(n)$ so that $u\in W_1$ and $v\in W_r$, $r\in[n]-1$. Hence $d(u,v)=t$, for $1\leq t\leq 3$. \newline
Case 3.1. If $t=1$, then $v\in C_3$. So if we consider $x=v$ and $v\neq y\in C_3$, then we have $d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.\newline
Case 3.2. If $t=2$, then in this case may be $v\in C_3$ or $v\notin C_3$. If $v\in C_3$, then there exists an element $x\in C_3$ so that $x\in W_k$,
$k\in [n]-1$, $r\neq k$ and $d(u, x)=1, d(v, x)=3$. So if we consider $v=y$, then we have
$-1=1-2=d(u, x) - d(u, y) \neq d(v, x) - d(v, y)=3-0=3$.
If $v\notin C_3$, then there exists an element $x\in W_r$ so that $x\in C_3$ and $d(u, x)=d(v, x)=1$, also there exists an element $y\in C_3$ so that
$y\in W_k$, $k\in[n]-\{1, r\}$, and $d(u, y)=2$, $d(v, y)=3$ or $d(u, y)=3$, $d(v, y)=2$, and hence we have
$d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.\newline
Case 3.3. If $t=3$, then there exists an element $x\in W_r$ so that $x\in C_3$ and $d(u, x)=2$, $d(v, x)=1$, also there exists an element $y\in C_3$ so that $y\in W_k$, $k\in[n]-\{1, r\}$, and $d(u, y)=1$, $d(v, y)=3$, and hence we have
$d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.
\newline
Case 4.
Suppose that $u$ and $v$ are two distinct vertices in $L(n)$ so that $u\in W_r$ and $v\in W_k$, $r, k\in[n]-1$, $r\neq k$. If both two vertices
$u$ and $v$ lie in $C_3$ or exactly one of them vertices lie in $C_3$ then there is nothing to prove. Now suppose that both two vertices $u, v \notin C_3$. Hence there exist elements $x\in C_3$ and $y\in C_3$ so that $x\in W_r$ and $y\in W_k$,
and hence we have
$d(u, x) - d(u, y) \neq d(v, x) - d(v, y)$.\newline
\end{proof}
\begin{proposition}\label{l.5}
Consider graph $L(n)$ with vertex set $W_1\cup...\cup W_n$ for $n\geq 5$. Then for $1\leq r\leq n$, any set $N(W_r)$ of size $n-1$ cannot be a strong resolving set for $L(n)$.\newline
\end{proposition}
\begin{proof}
Suppose that $V(L(n))=W_1\cup...\cup W_n$, where for $1\leq r\leq n$ the set $W_r=\{\{v_r, v_iv_j\}\,\ | \,\ i,r,s \in [n], i\neq j, v_r=v_i \,\ \text{or} \,\ v_r=v_j \}$ to indicate a maximal clique of size $n-1$ in the graph $L(n)$.
Without loss of generality if we consider the maximal clique $W_1$ and we take
$C_3=N(W_1)= \{y_2, ... , y_{n-1}, y_n\}$, where for $2\leq k\leq n$ we have $y_k= \{v_k, v_1v_k\}\in W_k$, then
based on Lemma \ref{l.1}, we know that for the maximal clique $W_1$ in $L(n)$, the subset
$C_3=N(W_1)=\{y_2, ..., y_{n-1}, y_n\}$ of vertices in $L(n)$ is a resolving set for $L(n)$ of size $n-1$.
By considering various vertices $ w_1\in W_r$ and $w_2\in W_k$, $1 <r, k <n$, $r\neq k$, so that $d(w_1, w_2)=3$ and $w_1, w_2\notin C_3$, there is not a $y_r\in C_3$ so that $w_1$ belongs to a shortest $w_2 - y_r$ path or $w_2$ belongs to a shortest $w_1 - y_r$ path. Thus the set $C_3=N(W_1)=\{y_2, ..., y_{n-1}, y_n\}$ cannot be a strong resolving set for $L(n)$, and so
any set $N(W_r)$ of graph $L(n)$ of size $n-1$ cannot be a strong resolving set for $L(n)$.\newline
\end{proof}
\begin{theorem}\label{l.6}
Consider graph $L(n)$ with vertex set $W_1\cup ...\cup W_n$ for $n\geq 5$. Then the minimum size of a strong resolving set in graph $L(n)$ is $n(n-2)$.
\end{theorem}
\begin{proof}
Suppose that $V(L(n))=W_1\cup...\cup W_n$, where for $1\leq r\leq n$ the set $W_r=\{\{v_r, v_iv_j\}\,\ | \,\ i,r,s \in [n], i\neq j, v_r=v_i \,\ \text{or} \,\ v_r=v_j \}$ to indicate a maximal clique of size $n-1$ in graph $L(n)$.
Without loss of generality if we consider the vertex $\{v_1, v_1v_2\}$ in the maximal clique $W_1$,
then there are exactly $(n-2)$ vertices in any maximal cliques $W_3, W_4, ..., W_n$, so that the distance between the vertex $\{v_1, v_1v_2\}\in W_1$ and these vertices in any maximal cliques $W_3, W_4, ..., W_n$ is $3$, and hence these vertices must be lie in every minimal strong resolving set of $L(n)$. Note that the cardinality of these vertices is $(n-2)(n-2)$. On the other hand if we take $C_3=N(W_1)= \{y_2, ... , y_{n-1}, y_n\}$, where for $2\leq k\leq n$ we have $y_k= \{v_k, v_1v_k\}\in W_k$, then the distance between two distinct vertices of $N(W_1)$ is $3$, and so $n-2$ vertices of $N(W_1)$ must be lie in every minimal strong resolving set of $L(n)$, we may consider these vertices are $y_3, ... , y_{n-1}, y_n$. Now, if we consider the maximal cliques $W_1$ and $W_2$,
then there are exactly $(n-2)$ vertices in the maximal clique $W_1$, so that the distance these vertices from $(n-2)$ vertices in the maximal clique $W_2$ is $3$, and hence we may assume that $(n-2)$ vertices of the maximal clique $W_2$ so that the distance between these vertices from $(n-2)$ vertices of $W_1$ is $3$, must be lie in every minimal strong resolving set of $L(n)$. Thus the minimum size of a strong resolving set in the graph $L(n)$ is $n(n-2)$.
\end{proof}
\bigskip
{\footnotesize
\noindent \textbf{Authors' informations}\\
\noindent Jia-Bao Liu${}^a$
(\url{<EMAIL>;<EMAIL>})\\
Ali Zafari${}^{a}$(\textsc{Corresponding Author})
(\url{<EMAIL>}; \url{<EMAIL>})\\
\noindent ${}^{a}$ School of Mathematics and Physics, Anhui Jianzhu University, Hefei 230601, P.R. China.\\
${}^{a}$ Department of Mathematics, Faculty of Science,
Payame Noor University, P.O. Box 19395-4697, Tehran, Iran.
{\footnotesize
|
\section{Introduction}
Please follow the steps outlined below when submitting your manuscript to
the IEEE Computer Society Press. This style guide now has several
important modifications (for example, you are no longer warned against the
use of sticky tape to attach your artwork to the paper), so all authors
should read this new version.
\subsection{Language}
All manuscripts must be in English.
\subsection{Dual submission}
Please refer to the author guidelines on the ICCV 2021 web page for a
discussion of the policy on dual submissions.
\subsection{Paper length}
Papers, excluding the references section,
must be no longer than eight pages in length. The references section
will not be included in the page count, and there is no limit on the
length of the references section. For example, a paper of eight pages
with two pages of references would have a total length of 10 pages.
{\bf There will be no extra page charges for ICCV 2021.}
Overlength papers will simply not be reviewed. This includes papers
where the margins and formatting are deemed to have been significantly
altered from those laid down by this style guide. Note that this
\LaTeX\ guide already sets figure captions and references in a smaller font.
The reason such papers will not be reviewed is that there is no provision for
supervised revisions of manuscripts. The reviewing process cannot determine
the suitability of the paper for presentation in eight pages if it is
reviewed in eleven.
\subsection{The ruler}
The \LaTeX\ style defines a printed ruler which should be present in the
version submitted for review. The ruler is provided in order that
reviewers may comment on particular lines in the paper without
circumlocution. If you are preparing a document using a non-\LaTeX\
document preparation system, please arrange for an equivalent ruler to
appear on the final output pages. The presence or absence of the ruler
should not change the appearance of any other content on the page. The
camera ready copy should not contain a ruler. (\LaTeX\ users may uncomment
the \verb'\iccvfinalcopy' command in the document preamble.) Reviewers:
note that the ruler measurements do not align well with lines in the paper
--- this turns out to be very difficult to do well when the paper contains
many figures and equations, and, when done, looks ugly. Just use fractional
references (e.g.\ this line is $095.5$), although in most cases one would
expect that the approximate location will be adequate.
\subsection{Mathematics}
Please number all of your sections and displayed equations. It is
important for readers to be able to refer to any particular equation. Just
because you didn't refer to it in the text doesn't mean some future reader
might not need to refer to it. It is cumbersome to have to use
circumlocutions like ``the equation second from the top of page 3 column
1''. (Note that the ruler will not be present in the final copy, so is not
an alternative to equation numbers). All authors will benefit from reading
Mermin's description of how to write mathematics:
\url{http://www.pamitc.org/documents/mermin.pdf}.
\subsection{Blind review}
Many authors misunderstand the concept of anonymizing for blind
review. Blind review does not mean that one must remove
citations to one's own work---in fact it is often impossible to
review a paper unless the previous citations are known and
available.
Blind review means that you do not use the words ``my'' or ``our''
when citing previous work. That is all. (But see below for
tech reports.)
Saying ``this builds on the work of Lucy Smith [1]'' does not say
that you are Lucy Smith; it says that you are building on her
work. If you are Smith and Jones, do not say ``as we show in
[7]'', say ``as Smith and Jones show in [7]'' and at the end of the
paper, include reference 7 as you would any other cited work.
An example of a bad paper just asking to be rejected:
\begin{quote}
\begin{center}
An analysis of the frobnicatable foo filter.
\end{center}
In this paper we present a performance analysis of our
previous paper [1], and show it to be inferior to all
previously known methods. Why the previous paper was
accepted without this analysis is beyond me.
[1] Removed for blind review
\end{quote}
An example of an acceptable paper:
\begin{quote}
\begin{center}
An analysis of the frobnicatable foo filter.
\end{center}
In this paper we present a performance analysis of the
paper of Smith \etal [1], and show it to be inferior to
all previously known methods. Why the previous paper
was accepted without this analysis is beyond me.
[1] Smith, L and Jones, C. ``The frobnicatable foo
filter, a fundamental contribution to human knowledge''.
Nature 381(12), 1-213.
\end{quote}
If you are making a submission to another conference at the same time,
which covers similar or overlapping material, you may need to refer to that
submission in order to explain the differences, just as you would if you
had previously published related work. In such cases, include the
anonymized parallel submission~\cite{Authors14} as additional material and
cite it as
\begin{quote}
[1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324,
Supplied as additional material {\tt fg324.pdf}.
\end{quote}
Finally, you may feel you need to tell the reader that more details can be
found elsewhere, and refer them to a technical report. For conference
submissions, the paper must stand on its own, and not {\em require} the
reviewer to go to a tech report for further details. Thus, you may say in
the body of the paper ``further details may be found
in~\cite{Authors14b}''. Then submit the tech report as additional material.
Again, you may not assume the reviewers will read this material.
Sometimes your paper is about a problem which you tested using a tool which
is widely known to be restricted to a single institution. For example,
let's say it's 1969, you have solved a key problem on the Apollo lander,
and you believe that the ICCV70 audience would like to hear about your
solution. The work is a development of your celebrated 1968 paper entitled
``Zero-g frobnication: How being the only people in the world with access to
the Apollo lander source code makes us a wow at parties'', by Zeus \etal.
You can handle this paper like any other. Don't write ``We show how to
improve our previous work [Anonymous, 1968]. This time we tested the
algorithm on a lunar lander [name of lander removed for blind review]''.
That would be silly, and would immediately identify the authors. Instead
write the following:
\begin{quotation}
\noindent
We describe a system for zero-g frobnication. This
system is new because it handles the following cases:
A, B. Previous systems [Zeus et al. 1968] didn't
handle case B properly. Ours handles it by including
a foo term in the bar integral.
...
The proposed system was integrated with the Apollo
lunar lander, and went all the way to the moon, don't
you know. It displayed the following behaviours
which show how well we solved cases A and B: ...
\end{quotation}
As you can see, the above text follows standard scientific convention,
reads better than the first version, and does not explicitly name you as
the authors. A reviewer might think it likely that the new paper was
written by Zeus \etal, but cannot make any decision based on that guess.
He or she would have to be sure that no other authors could have been
contracted to solve problem B.
\medskip
\noindent
FAQ\medskip\\
{\bf Q:} Are acknowledgements OK?\\
{\bf A:} No. Leave them for the final copy.\medskip\\
{\bf Q:} How do I cite my results reported in open challenges?
{\bf A:} To conform with the double blind review policy, you can report results of other challenge participants together with your results in your paper. For your results, however, you should not identify yourself and should not mention your participation in the challenge. Instead present your results referring to the method proposed in your paper and draw conclusions based on the experimental comparison to other results.\medskip\\
\begin{figure}[t]
\begin{center}
\fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
\end{center}
\caption{Example of caption. It is set in Roman so that mathematics
(always set in Roman: $B \sin A = A \sin B$) may be included without an
ugly clash.}
\label{fig:long}
\label{fig:onecol}
\end{figure}
\subsection{Miscellaneous}
\noindent
Compare the following:\\
\begin{tabular}{ll}
\verb'$conf_a$' & $conf_a$ \\
\verb'$\mathit{conf}_a$' & $\mathit{conf}_a$
\end{tabular}\\
See The \TeX book, p165.
The space after \eg, meaning ``for example'', should not be a
sentence-ending space. So \eg is correct, {\em e.g.} is not. The provided
\verb'\eg' macro takes care of this.
When citing a multi-author paper, you may save space by using ``et alia'',
shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word.)
However, use it only when there are three or more authors. Thus, the
following is correct: ``
Frobnication has been trendy lately.
It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by
Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.''
This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...''
because reference~\cite{Alpher03} has just two authors. If you use the
\verb'\etal' macro provided, then you need not worry about double periods
when used at the end of a sentence as in Alpher \etal.
For this citation style, keep multiple citations in numerical (not
chronological) order, so prefer \cite{Alpher03,Alpher02,Authors14} to
\cite{Alpher02,Alpher03,Authors14}.
\begin{figure*}
\begin{center}
\fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}}
\end{center}
\caption{Example of a short caption, which should be centered.}
\label{fig:short}
\end{figure*}
\section{Formatting your paper}
All text must be in a two-column format. The total allowable width of the
text area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54
cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a
$\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the
first page) should begin 1.0 inch (2.54 cm) from the top edge of the
page. The second and following pages should begin 1.0 inch (2.54 cm) from
the top edge. On all pages, the bottom margin should be 1-1/8 inches (2.86
cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4
paper, approximately 1-5/8 inches (4.13 cm) from the bottom edge of the
page.
\subsection{Margins and page numbering}
All printed material, including text, illustrations, and charts, must be kept
within a print area 6-7/8 inches (17.5 cm) wide by 8-7/8 inches (22.54 cm)
high.
Page numbers should be included for review submissions but not for the
final paper. Review submissions papers should have page numbers in the
footer with numbers centered and .75 inches (1.905 cm) from the bottom
of the page and start on the first page with the number 1.
Page numbers will be added by the publisher to all camera-ready papers
prior to including them in the proceedings and before submitting the
papers to IEEE Xplore. As such, your camera-ready submission should
not include any page numbers. Page numbers should automatically be
removed by uncommenting (if it's not already) the line
\begin{verbatim}
\end{verbatim}
near the beginning of the .tex file.
\subsection{Type-style and fonts}
Wherever Times is specified, Times Roman may also be used. If neither is
available on your word processor, please use the font closest in
appearance to Times to which you have access.
MAIN TITLE. Center the title 1-3/8 inches (3.49 cm) from the top edge of
the first page. The title should be in Times 14-point, boldface type.
Capitalize the first letter of nouns, pronouns, verbs, adjectives, and
adverbs; do not capitalize articles, coordinate conjunctions, or
prepositions (unless the title begins with such a word). Leave two blank
lines after the title.
AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title
and printed in Times 12-point, non-boldface type. This information is to
be followed by two blank lines.
The ABSTRACT and MAIN TEXT are to be in a two-column format.
MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use
double-spacing. All paragraphs should be indented 1 pica (approx. 1/6
inch or 0.422 cm). Make sure your text is fully justified---that is,
flush left and flush right. Please do not place any additional blank
lines between paragraphs.
Figure and table captions should be 9-point Roman type as in
Figures~\ref{fig:onecol} and~\ref{fig:short}. Short captions should be centered.
\noindent Callouts should be 9-point Helvetica, non-boldface type.
Initially capitalize only the first word of section titles and first-,
second-, and third-order headings.
FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction})
should be Times 12-point boldface, initially capitalized, flush left,
with one blank line before, and one blank line after.
SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements})
should be Times 11-point boldface, initially capitalized, flush left,
with one blank line before, and one after. If you require a third-order
heading (we discourage it), use 10-point Times, boldface, initially
capitalized, flush left, preceded by one blank line, followed by a period
and your text on the same line.
\subsection{Footnotes}
Please use footnotes\footnote {This is what a footnote looks like. It
often distracts the reader from the main flow of the argument.} sparingly.
Indeed, try to avoid footnotes altogether and include necessary peripheral
observations in
the text (within parentheses, if you prefer, as in this sentence). If you
wish to use a footnote, place it at the bottom of the column on the page on
which it is referenced. Use Times 8-point type, single-spaced.
\subsection{References}
List and number all bibliographical references in 9-point Times,
single-spaced, at the end of your paper. When referenced in the text,
enclose the citation number in square brackets, for
example~\cite{Authors14}. Where appropriate, include the name(s) of
editors of referenced books.
\begin{table}
\begin{center}
\begin{tabular}{|l|c|}
\hline
Method & Frobnability \\
\hline\hline
Theirs & Frumpy \\
Yours & Frobbly \\
Ours & Makes one's heart Frob\\
\hline
\end{tabular}
\end{center}
\caption{Results. Ours is better.}
\end{table}
\subsection{Illustrations, graphs, and photographs}
All graphics should be centered. Please ensure that any point you wish to
make is resolvable in a printed copy of the paper. Resize fonts in figures
to match the font in the body text, and choose line widths which render
effectively in print. Many readers (and reviewers), even of an electronic
copy, will choose to print your paper in order to read it. You cannot
insist that they do otherwise, and therefore must not assume that they can
zoom in to see tiny details on a graphic.
When placing figures in \LaTeX, it's almost always best to use
\verb+\includegraphics+, and to specify the figure width as a multiple of
the line width as in the example below
{\small\begin{verbatim}
\usepackage[dvips]{graphicx} ...
\includegraphics[width=0.8\linewidth]
{myfile.eps}
\end{verbatim}
}
\subsection{Color}
Please refer to the author guidelines on the ICCV 2021 web page for a discussion
of the use of color in your document.
\section{Final copy}
You must include your signed IEEE copyright release form when you submit
your finished paper. We MUST have this form before your paper can be
published in the proceedings.
{\small
\bibliographystyle{ieee_fullname}
\section{Per-scene optimization.}
\boldstart{More details.} As described in the paper Sec.~3.4, we optimize the predicted neural encoding volume with the MLP decoder for each scene to achieve final high-quality fine-tuning results.
The neural encoding volume with the MLP is an effective neural representation of a radiance field. All fine-tuning results (and also NeRF's optimization comparison results) are generated using a single NVIDIA RTX 2080Ti GPU. On this hardware, our 15min fine-tuning corresponds to 10k training iterations and NeRF's 10.2h optimization corresponds to 200k training iterations.
Our neural encoding volume is reconstructed in the frustrum of the reference view; it thus only covers the scene content in the frustrum. As a result, for a large scene, artifacts may appear when some parts that are not located in the frustrum show up in the novel view.
Therefore, we extend the neural encoding volume by padding its boundary voxels when fine-tuning on some large scenes. This can address the out-of-frustrum artifacts, though the padding voxels are not well reconstructed initially by the network and may require longer fine-tuning to achieve high quality.
As described in the paper, we do not apply two-stage (coarse and fine) ray sampling as done in NeRF \cite{nerf}. We uniformly sample points (with per-step small randomness) along each marching ray. We find that 128 points are enough for most scenes and keep using 128 point for our across-scene training on the DTU dataset. When fine-tuning, we increase the number of points to 256 for some challenging scenes.
\boldstartspace{Optimization progress.}
We have demonstrated in the paper that our 15min fine-tuning results of DTU and LLFF dataset are comparable with the 10.2h optimization results of NeRF \cite{nerf}.
We now show comparisons of the per-scene optimization progress between our fine-tuning and NeRF's from-scratch optimization.
Note that, thanks to the strong initial reconstruction predicted by our network, our fine-tuning is consistently better than NeRF's optimization through 200k training iterations. As mentioned, our each 18min result corresponds to the result at the 12k-th training iteration, which is at the very early stage in the curves; however, as demonstrated, it can be already better than the NeRF's result after 48k iterations, corresponding to the 10.2h optimization result shown in the paper.
Moreover, while our 15min results are already very good,
our results can be further improved over more iterations, if continuing optimizing the radiance fields.
\begin{table}[ht]
\begin{tabular}{cccccc}
\toprule\hline
\textbf{Layer} & \textbf{k} & \textbf{s} & \textbf{d} & \textbf{chns} & input \\ \hline
CBR2D$_0$ & 3 & 1 & 1 & $3/8$ & $I$\\
CBR2D$_1$ & 3 & 1 & 1 & $8/8$ & CBR2D$_0$\\
CBR2D$_2$ & 5 & 2 & 2 & $8/16$ & CBR2D$_1$\\
CBR2D$_3$ & 3 & 1 & 1 & $16/16$ & CBR2D$_2$\\
CBR2D$_4$ & 3 & 1 & 1 & $16/16$ & CBR2D$_3$\\
CBR2D$_5$ & 5 & 2 & 2 & $16/32$ & CBR2D$_4$\\
CBR2D$_6$ & 3 & 1 & 1 & $32/32$ & CBR2D$_5$\\
$T$ & 3 & 1 & 1 & $32/32$ & CBR2D$_6$\\ \hline
CBR3D$_0$ & 3 & 1 & 1 & $32+9/8$ & $T, I$\\
CBR3D$_1$ & 3 & 2 & 1 & $8/16$ & CBR3D$_0$\\
CBR3D$_2$ & 3 & 1 & 1 & $16/16$ & CBR3D$_1$\\
CBR3D$_3$ & 3 & 2 & 1 & $16/32$ & CBR3D$_2$\\
CBR3D$_4$ & 3 & 1 & 1 & $32/32$ & CBR3D$_3$\\
CBR3D$_5$ & 3 & 2 & 1 & $32/64$ & CBR3D$_4$\\
CBR3D$_6$ & 3 & 1 & 1 & $64/64$ & CBR3D$_5$\\
CTB3D$_0$ & 3 & 2 & 1 & $64/32$ & CTB3D$_0$ + CBR3D$_4$\\
CTB3D$_1$ & 3 & 2 & 1 & 32/16 & CTB3D$_1$ + CBR3D$_2$\\
CTB3D$_2$ & 3 & 2 & 1 & $16/8$ & CTB3D$_2$ + CBR3D$_0$\\ \hline
PE$_0$ & - & - & - & 3/63 & $x$ \\
LR$_0$ & - & - & - & 8+12/256 & $f,c$ \\
LR$_1$ & - & - & - & 63/256 & PE\\
LR$_{i+1}$ & - & - & - & 256/256 & LR$_i$ $\cdot$ LR$_0$ \\
$\sigma$ & - & - & - & 256/1 & LR$_6$\\
PE$_1$ & - & - & - & 3/27 & $d$\\
LR$_7$ & - & - & - & 27+256/256 & PE$_1$, LR$_6$\\
$c$ & - & - & - & 256/3 & LR$_7$\\
\bottomrule
\end{tabular}
\caption{From top to bottom: 2D CNN based feature extraction model, 3D CNN based neural encoding volume prediction model and MLP based volume properties regression model $(i\in[1,...,5])$. \textbf{k} is the kernel size, \textbf{s} is the stride, \textbf{d} is the kernel dilation, and \textbf{chns} shows the number of input and output channels for each layer. We denote CBR2D/CBR3D/CTB3D/LR to be ConvBnReLU2D, ConvBnReLU3D, ConvTransposeBn3D and LinearRelu layer structure respectively. PE refers to the positional encoding as used in \cite{nerf}. }
\label{tbl:network}
\end{table}
\section{Network Architectures}
We show detailed network architecture specifications of our 2D CNN (that extracts 2D image features ), 3D CNN (that outputs a neural encoding volume), and MLP decoder (that regresses volume properties) in Tab ~\ref{tbl:network}.
\section{Limitations.}
Our approach generally achieves fast radiance field reconstruction for view synthesis on diverse real scenes. However, for highly challenging scenes with high glossiness/specularities, the strong view-dependent shading effects can be hard to directly recovered via network inference and a longer fine-tuning process can be required to fully reconstruct such effects.
Our radiance field representation is reconstructed within the frustrum of the reference view. As a result, only the scene content seen by the reference view is well reconstructed and initialized for the following fine-tuning stage.
Padding the volume (as discussed earlier) can incorporate content out of the original frustrum; however, the unseen parts (including those that are in the frustrucm but are occluded and invisible in the view) are not directly recovered by the network. Therefore, it is challenging to use a single neural encoding volume to achieve rendering in a wide viewing range around a scene (like 360$^\circ$ rendering). Note that, a long per-scene fine-tuning process with dense images covering around the scene can still achieve 360$^\circ$ rendering, though it can be as slow as training a standard NeRF \cite{nerf} (or Sparse Voxel Fields \cite{Neural_sparse_voxel_fields} that is similar to our representation) to recover those uninitialized regions in the encoding volume. Combining multiple neural encoding volumes at multiple views can be an interesting future direction to achieve fast radiance field reconstruction with larger viewing ranges.
\begin{table}[t]
\centering
\begin{tabular}{l|ccccc}
\multicolumn{6}{c}{\textbf{DTU Dataset}}\\ \hline
Scan & \#1 & \#8 & \#21 & \#103 & \#114 \\ \hline
\multicolumn{6}{c}{PSNR$\uparrow$}\\ \hline
PixelNeRF & 21.64& 23.70& 16.04& 16.76&18.40\\
IBRNet & 25.97& \textbf{27.45}& 20.94& 27.91&27.91\\
Ours & \textbf{26.96} & 27.43 &\textbf{ 21.55}&\textbf{29.25}&\textbf{27.99}\\ \hline
NeRF$_{10.2h}$ & 26.62& 28.33 & 23.24 & 30.40 &26.47\\
IBRNet$_{ft-1h}$ & \textbf{31.00}& \textbf{32.46}& \textbf{27.88}& \textbf{34.40}&\textbf{31.00}\\
Ours$_{ft-15min}$& 28.05 & 28.88 & 24.87 & 32.23 & 28.47 \\ \hline
\multicolumn{6}{c}{SSIM$\uparrow$}\\ \hline
PixelNeRF & 0.827& 0.829& 0.691& 0.836&0.763\\
IBRNet & 0.918& 0.903& 0.873& 0.950&0.943\\
Ours & \textbf{0.937} & \textbf{0.922} &\textbf{0.890} &\textbf{0.962}&\textbf{0.949}\\ \hline
NeRF$_{10.2h}$ & 0.902 & 0.876 & 0.874 & 0.944 & 0.913\\
IBRNet$_{ft-1h}$ & \textbf{0.955}& 0.945& \textbf{0.947}& 0.968&\textbf{0.964}\\
Ours$_{ft-15min}$& \textbf{0.934} & 0.900& \textbf{0.922}&\textbf{0.964} & 0.945\\\hline
\multicolumn{6}{c}{LPIPS $\downarrow$ }\\ \hline
PixelNeRF & 0.373& 0.384& 0.407& 0.376&0.372\\
IBRNet & 0.190& 0.252& 0.179& 0.195 & 0.136\\
Ours & \textbf{0.155} & \textbf{0.220} &\textbf{0.166}&\textbf{0.165}&\textbf{0.135}\\ \hline
NeRF$_{10.2h}$ & 0.265& 0.321&0.246& 0.256 & 0.225\\
IBRNet$_{ft-1h}$ & \textbf{0.129} &\textbf{0.170}&\textbf{0.104}&\textbf{0.156}&\textbf{0.099} \\
Ours$_{ft-15min}$ & 0.171 &0.261&0.142&0.170&0.153 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{Quantity comparison on five sample scenes in the DTU testing set}.}
\label{tb:ablation}
\end{table}
\section{Per-scene breakdown.}
We show the pre-scene breakdown analysis of the quantitative results presented in the main paper for the three dataset (\textit{Realistic Synthetic}, \textit{DTU} and \textit{LLFF}).
These results are consistent with the averaged results shown in the paper. In general, since the training set consists of DTU scenes, all three methods can work reasonably well on the DTU testing set.
Our approach can outperform PixelNeRF \cite{yu2020pixelnerf}, when using the same three-image input, and achieve higher PSNR and SSIM and lower LPIPS. Note that, as mentioned in the paper, the implementation of IBRNet~\cite{ibrnet} is trained and tested with 10 input images to achieve its best performance as used in their paper. Nonetheless, our results with three input images are still quantitatively comparable to the results of IBRNet with 10 input images on the DTU testing set; IBRNet often achieves better PSNRs while we often achieve better SSIMs and LPIPSs.
More importantly, as already shown in paper, when testing on novel datasets, our approach generalizes significantly better than PixelNeRF and IBRNet, leading to much better quantitative results on the Synthetic Data and the Forward-Facing dataset. We also provide detailed per-scene quantitative results for the three testing datasets in Tab.~3-10. Please also refer to the supplementary video for video comparisons.
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\hline
& Chair & Drums & Ficus & Hotdog & Lego & Materials & Mic & Ship \\ \hline
\multicolumn{9}{c}{PSNR$\uparrow$ }\\ \hline
PixelNeRF & 7.18 & 8.15 & 6.61 & 6.80 & 7.74 & 7.61 & 7.71& 7.30\\
IBRNet & \textbf{24.20} & 18.63 & 21.59 & 27.70 & 22.01 & \textbf{20.91} & 22.10& 22.36\\
Ours & 23.35 & \textbf{20.71} & \textbf{21.98} & \textbf{28.44} & \textbf{23.18}& 20.05 & \textbf{22.62} & \textbf{23.35}\\ \hline
NeRF & \textbf{31.07} & \textbf{25.46} & \textbf{29.73} & \textbf{34.63} & \textbf{32.66} &\textbf{30.22}& \textbf{31.81}& 29.49\\
IBRNet$_{ft-1h}$ & 28.18 & 21.93 & 25.01 & 31.48 & 25.34 & 24.27 & 27.29& 21.48\\
Ours$_{ft-15min}$ & 26.80 & 22.48 & 26.24 & 32.65 & 26.62 & 25.28 & 29.78 & 26.73 \\\hline
\multicolumn{9}{c}{SSIM$\uparrow$ }\\\hline
PixelNeRF & 0.624 & 0.670 & 0.669 & 0.669 & 0.671& 0.644 & 0.729& 0.584\\
IBRNet & 0.888& 0.836 & 0.881 & 0.923 & 0.874 & 0.872 & 0.927& 0.794\\
Ours & \textbf{0.876} & \textbf{0.886} & \textbf{0.898} & \textbf{0.962} & \textbf{0.902} & \textbf{0.893} & \textbf{0.923} & \textbf{0.886} \\ \hline
NeRF & \textbf{0.971}& \textbf{0.943}& \textbf{0.969}& \textbf{0.980}&\textbf{0.975}& \textbf{0.968}& \textbf{0.981} & \textbf{0.908}\\
IBRNet$_{ft-1h}$ & 0.955 & 0.913 & 0.940 & 0.978 & 0.940 & 0.937 & 0.974& 0.877\\
Ours$_{ft-15min}$ & 0.934 & 0.898 & 0.944 & 0.971 & 0.924 & 0.927 & 0.970&0.879 \\ \hline
\multicolumn{9}{c}{LPIPS $\downarrow$}\\ \hline
PixelNeRF & 0.386 & 0.421 & 0.335 & 0.433 & 0.427& 0.432 & 0.329 & 0.526\\
IBRNet & \textbf{0.144} & 0.241 & \textbf{0.159} & 0.175 &\textbf{0.202} & \textbf{0.164} & \textbf{0.103} & 0.369 \\
Ours & 0.282& \textbf{0.187} & 0.211 & \textbf{0.173} & 0.204 & 0.216 & 0.177 & \textbf{0.244} \\ \hline
NeRF & \textbf{0.055}& \textbf{0.101}&\textbf{0.047}& \textbf{0.089}&\textbf{0.054}&0.105& \textbf{0.033}& 0.263\\
IBRNet$_{ft-1h}$ & 0.079 & 0.133 & 0.082 & 0.093 & 0.105 & \textbf{0.093} & 0.040 & \textbf{0.257}\\
Ours$_{ft-15min}$ & 0.129& 0.197 & 0.171 & 0.094 & 0.176 & 0.167 & 0.117 & 0.294 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{Quantity comparison on the Realistic Synthetic dataset}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\hline
& Fern &Flower&Fortress & Horns & Leaves & Orchids & Room & Trex\\ \hline
\multicolumn{9}{c}{PSNR$\uparrow$ }\\ \hline
PixelNeRF & 12.40 & 10.00 &14.07&11.07& 9.85 &9.62&11.75& 10.55\\
IBRNet & 20.83 & 22.38 & \textbf{27.67} & 22.06 & \textbf{18.75} & 15.29 & \textbf{27.26}& 20.06\\
Ours & \textbf{21.15} & \textbf{24.74} & 26.03 &\textbf{23.57}&17.51& \textbf{17.85} &26.95&\textbf{23.20}\\\hline
NeRF$_{10.2h}$ & \textbf{23.87} & 26.84 &\textbf{31.37}&25.96& 21.21 &19.81&\textbf{33.54}& \textbf{25.19}\\
IBRNet$_{ft-1h}$ & 22.64 & 26.55 & 30.34 & 25.01 & \textbf{22.07} & 19.01 & 31.05 & 22.34\\
Ours$_{ft-15min}$ & 23.10 & \textbf{27.23} &30.43 & \textbf{26.35} & 21.54 & \textbf{20.51} & 30.12 & 24.32 \\\hline
\multicolumn{9}{c}{SSIM$\uparrow$}\\\hline
PixelNeRF & 0.531 & 0.433 &0.674&0.516& 0.268 & 0.317 &0.691&0.458\\
IBRNet & \textbf{0.710} & 0.854 &\textbf{0.894}&0.840&\textbf{0.705}&0.571&0.950& 0.768\\
Ours & 0.638 & \textbf{0.888} &0.872&\textbf{0.868}& 0.667 &\textbf{0.657}&\textbf{0.951} & \textbf{0.868}\\ \hline
NeRF$_{10.2h}$ & \textbf{0.828}& 0.897 &\textbf{0.945}&0.900& 0.792 &0.721&\textbf{0.978}&\textbf{0.899}\\
IBRNet$_{ft-1h}$ & 0.774 & 0.909 & 0.937 & 0.904 & \textbf{0.843} & 0.705 & 0.972 & 0.842\\
Ours$_{ft-15min}$ & 0.795 & \textbf{0.912}&0.943&\textbf{0.917} & 0.826 & \textbf{0.732}&0.966 &0.895\\\hline
\multicolumn{9}{c}{LPIPS $\downarrow$ }\\\hline
& Fern & Flower & Fortress & Horns & Leaves & Orchids & Room & Trex\\ \hline
PixelNeRF & 0.650 & 0.708 &0.608&0.705& 0.695 &0.721&0.611& 0.667\\
IBRNet &0.349 & 0.224 & \textbf{0.196} & 0.285& \textbf{0.292} &0.413&\textbf{0.161}& 0.314\\
Ours &\textbf{0.238} & \textbf{0.196}&0.208& \textbf{0.237}& 0.313 &\textbf{0.274}&0.172& \textbf{0.184}\\ \hline
NeRF$_{10.2h}$ & 0.291& 0.176&0.147& 0.247& 0.301& 0.321 & 0.157 &0.245\\
IBRNet$_{ft-1h}$ & 0.266 & 0.146 & \textbf{0.133} & 0.190 & \textbf{0.180} & 0.286 & \textbf{0.089} & 0.222\\
Ours$_{ft-15min}$ &\textbf{0.253}&\textbf{0.143}&0.134&\textbf{0.188}&0.222&\textbf{0.258}&0.149&\textbf{0.187} \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{Quantity comparison on the Forward Facing dataset}.}
\label{tb:ablation}
\end{table*}
\section{Update the quantitative results on DTU.}
We found an issue in our code when computing the PSNR/SSIM/LPIPS of the results (without fine-tuning) of DTU testing scenes. We therefore show the updated Table 1 of the main paper in Tab.~\ref{tb:rendering}. We apologize for the inconvenience. Note that, only the DTU numbers of the three methods with no per-scene optimization are
updated. We only consider foreground regions when computing the numbers of DTU results.
In general, since the training set consists of DTU scenes, all three methods can work reasonably well on the DTU testing set.
Our approach can outperform PixelNeRF \cite{yu2020pixelnerf}, when using the same three-image input, and achieve higher PSNR and SSIM and lower LPIPS. Note that, as mentioned in the paper, the implementation of IBRNet~\cite{ibrnet} is trained and tested with 10 input images to achieve its best performance as used in their paper. Nonetheless, our results with three input images are still quantitatively comparable to the results of IBRNet with 10 input images on the DTU testing set; IBRNet achieves better PSNR while we achieve better SSIM and LPIPS.
More importantly, as already shown in paper, when testing on novel datasets, our approach generalizes significantly better than PixelNeRF and IBRNet, leading to much better quantitative results on the Synthetic Data and the Forward-Facing dataset. We also provide detailed per-scene quantitative results for the three testing datasets in Tab.~3-10. Please also refer to the supplementary video for video comparisons.
\section{Per-scene optimization.}
\boldstart{More details.} As described in the paper Sec.~3.4, we optimize the predicted neural encoding volume with the MLP decoder for each scene to achieve final high-quality fine-tuning results.
The neural encoding volume and the MLP is an effective neural representation of a radiance field. All fine-tuning results (and also NeRF's optimization comparison results) are generated using a single NVIDIA RTX 2080Ti GPU. On this hardware, our 15min fine-tuning corresponds to 10k training iterations and NeRF's 9.5h optimization corresponds to 200k training iterations.
Our neural encoding volume is reconstructed in the frustrum of the reference view; it thus only covers the scene content in the frustrum. As a result, for a large scene, artifacts may appear when some parts that are not located in the frustrum show up in the novel view.
Therefore, we extend the neural encoding volume by padding its boundary voxels when fine-tuning on some large scenes. This can address the out-of-frustrum artifacts, though the padding voxels are not well reconstructed initially by the network and may require longer fine-tuning to achieve high quality.
As described in the paper, we do not apply two-stage (coarse and fine) ray sampling as done in NeRF \cite{nerf}. We uniformly sample points (with per-step small randomness) along each marching ray. We find that 128 points are enough for most scenes and keep using 128 point for our across-scene training on the DTU dataset. When fine-tuning, we increase the number of points to 256 for some challenging scenes.
\boldstartspace{Optimization progress.}
We have demonstrated in the paper that our 15min fine-tuning results can be better than the 9.5h optimization results of NeRF \cite{nerf}.
We now show comparisons of the per-scene optimization progress between our fine-tuning and NeRF's from-scratch optimization.
The comparisons of the training PSNRs on four testing scenes are shown in Fig~\ref{fig:curves}.
Note that, thanks to the strong initial reconstruction predicted by our network, our fine-tuning is consistently better than NeRF's optimization through 200k training iterations. As mentioned, our each 15min result corresponds to the result at the 10k-th training iteration, which is at the very early stage in the curves; however, as demonstrated, it can be already better than the NeRF's result after 200k iterations, corresponding to the 9.5h optimization result shown in the paper.
Moreover, while our 15min results are already very good,
our results can be further improved over more iterations, if continuing optimizing the radiance fields, as reflected by the curves.
\begin{table*}[ht]
\centering
\begin{tabular}{lcccccccc}
\toprule
\multirow{2}{*}{Method} & \multirow{2}{*}{Settings} & \multicolumn{3}{c}{Synthetic Data (NeRF ~\cite{nerf_in_the_wild})} & \multicolumn{3}{c}{Real Data (DTU ~\cite{dtu}, Forward-Facing ~\cite{llff})} \\
\cmidrule(lr){3-5}\cmidrule(lr){6-8}
&& \multicolumn{1}{c}{PSNR$\uparrow$} & \multicolumn{1}{c}{SSIM$\uparrow$} & \multicolumn{1}{c}{LPIPS$\downarrow$} &
\multicolumn{1}{c}{PSNR$\uparrow$} &
\multicolumn{1}{c}{SSIM$\uparrow$} &
\multicolumn{1}{c}{LPIPS$\downarrow$} \\
\midrule
PixelNeRF~\cite{yu2020pixelnerf} & \multirow{3}{*}{\shortstack{No per-scene \\ optimization}} & 4.36 & 0.46 & 0.44 & 24.14/11.266 & 0.887/0.388 & 0.224/0.757 \\
IBRNet~\cite{ibrnet} & & 19.43 & 0.841 & 0.231 & \textbf{25.84}/16.70 & 0.902/0.566 & 0.213/0.498 \\
Ours & & \textbf{22.67} & \textbf{0.90} & \textbf{0.21} & 24.38/\textbf{17.56} & \textbf{0.906}/\textbf{0.691} & \textbf{0.197}/\textbf{0.381} \\
\midrule
NeRF$_{9.5h}$~\cite{nerf} & \multirow{3}{*}{\shortstack{Per-scene \\ optimization}} & \textbf{26.95} & \textbf{0.939} & \textbf{0.136} & 23.70/\textbf{23.67} & 0.893/0.820& 0.247/ 0.339 \\
Ours$_{ft-15min}$ & & 25.17 & 0.919 & 0.257 & 27.72/23.03 & 0.939/0.852 & 0.135/0.210 \\
Ours$_{ft-75min}$ & & 25.95 & 0.932 & 0.230 & \textbf{28.70}/23.37 & \textbf{0.947}/\textbf{0.857} & \textbf{0.116}/\textbf{0.184} \\
\bottomrule
\end{tabular}
\rule{0pt}{0.01pt}
\caption{\textbf{Quantitative results of novel view synthesis.} We show averaged results of PSNRs, SSIMs and LPISs on three different datasets. They are a synthesis dataset \cite{nerf} (left) of 8 scenes and two real datasets of 16 DTU (5 of them used for fine tuning evaluation) and 8 Forward-Facing scenes. On the top, we compare our method with concurrent neural rendering methods \cite{yu2020pixelnerf,ibrnet} with direct network inference. On the bottom, we show our fine-tuning results with only 15min and 75min optimization and compare with NeRF's \cite{nerf} results with 9.5h optimization.}
\label{tb:rendering}
\end{table*}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=\linewidth]{LaTeX/fig/curves.pdf}
\end{center}
\caption{Optimization progress. We compare the optimization progress of our fine-tuning with NeRF's optimization from scratch. The PSNRs of 200k training iterations on four testing scenes are plotted. Because of the strong initial radiance field reconstruction from our network, our training is consistently much faster than NeRF through the iterations.}
\label{fig:curves}
\end{figure*}
\section{Network Architectures}
We show detailed network architecture specifications of our 2D CNN (that extracts 2D image features ), 3D CNN (that outputs a neural encoding volume), and MLP decoder (that regresses volume properties) in Tab ~\ref{tbl:network}.
\section{Limitations.}
Our approach generally achieves fast radiance field reconstruction for view synthesis on diverse real scenes. However, for highly challenging scenes with high glossiness/specularities, the strong view-dependent shading effects can be hard to directly recovered via network inference and a longer fine-tuning process can be required to fully reconstruct such effects.
Our radiance field representation is reconstructed within the frustrum of the reference view. As a result, only the scene content seen by the reference view is well reconstructed and initialized for the following fine-tuning stage.
Padding the volume (as discussed earlier) can incorporate content out of the original frustrum; however, the unseen parts (including those that are in the frustrucm but are occluded and invisible in the view) are not directly recovered by the network. Therefore, it is challenging to use a single neural encoding volume to achieve rendering in a wide viewing range around a scene (like 360$^\circ$ rendering). Note that, a long per-scene fine-tuning process with dense images covering around the scene can still achieve 360$^\circ$ rendering, though it can be as slow as training a standard NeRF \cite{nerf} (or Sparse Voxel Fields \cite{Neural_sparse_voxel_fields} that is similar to our representation) to recover those uninitialized regions in the encoding volume. Combining multiple neural encoding volumes at multiple views can be an interesting future direction to achieve fast radiance field reconstruction with larger viewing ranges.
\begin{table}[ht]
\begin{tabular}{cccccc}
\toprule\hline
\textbf{Layer} & \textbf{k} & \textbf{s} & \textbf{d} & \textbf{chns} & input \\ \hline
CBR2D$_0$ & 3 & 1 & 1 & $3/8$ & $I$\\
CBR2D$_1$ & 3 & 1 & 1 & $8/8$ & CBR2D$_0$\\
CBR2D$_2$ & 5 & 2 & 2 & $8/16$ & CBR2D$_1$\\
CBR2D$_3$ & 3 & 1 & 1 & $16/16$ & CBR2D$_2$\\
CBR2D$_4$ & 3 & 1 & 1 & $16/16$ & CBR2D$_3$\\
CBR2D$_5$ & 5 & 2 & 2 & $16/32$ & CBR2D$_4$\\
CBR2D$_6$ & 3 & 1 & 1 & $32/32$ & CBR2D$_5$\\
$T$ & 3 & 1 & 1 & $32/32$ & CBR2D$_6$\\ \hline
CBR3D$_0$ & 3 & 1 & 1 & $32+9/8$ & $T, I$\\
CBR3D$_1$ & 3 & 2 & 1 & $8/16$ & CBR3D$_0$\\
CBR3D$_2$ & 3 & 1 & 1 & $16/16$ & CBR3D$_1$\\
CBR3D$_3$ & 3 & 2 & 1 & $16/32$ & CBR3D$_2$\\
CBR3D$_4$ & 3 & 1 & 1 & $32/32$ & CBR3D$_3$\\
CBR3D$_5$ & 3 & 2 & 1 & $32/64$ & CBR3D$_4$\\
CBR3D$_6$ & 3 & 1 & 1 & $64/64$ & CBR3D$_5$\\
CTB3D$_0$ & 3 & 2 & 1 & $64/32$ & CBR3D$_6$\\
CTB3D$_1$ & 3 & 2 & 1 & $64/32$ & CTB3D$_0$ + CBR3D$_4$\\
CTB3D$_2$ & 3 & 2 & 1 & $64/32$ & CTB3D$_1$ + CBR3D$_2$\\
CTB3D$_3$ & 3 & 2 & 1 & $64/32$ & CTB3D$_2$ + CBR3D$_0$\\ \hline
PE$_0$ & - & - & - & 3/63 & $x$ \\
LR$_0$ & - & - & - & 8+9/256 & $f,c$ \\
LR$_1$ & - & - & - & 63/256 & PE\\
LR$_{i+1}$ & - & - & - & 256/256 & LR$_i$+LR$_0$ \\
$\sigma$ & - & - & - & 256/1 & LR$_6$\\
PE$_1$ & - & - & - & 3/27 & $d$\\
LR$_7$ & - & - & - & 27+256/256 & PE$_1$,LR$_6$\\
$c$ & - & - & - & 256/3 & LR$_7$\\
\bottomrule
\end{tabular}
\caption{From top to bottom: 2D CNN based feature extraction model, 3D CNN based neural encoding volume prediction model and MLP based volume properties regression model $(i\in[1,...,5])$. \textbf{k} is the kernel size, \textbf{s} is the stride, \textbf{d} is the kernel dilation, and \textbf{chns} shows the number of input and output channels for each layer. We denote CBR2D/CBR3D/CTB3D/LR to be ConvBnReLU2D, ConvBnReLU3D, ConvTransposeBn3D and LinearRelu layer structure respectively. PE refers to the positional encoding as used in \cite{nerf}. }
\label{tbl:network}
\end{table}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\multicolumn{9}{c}{PSNR$\uparrow$ on \textbf{Realistic Synthetic Dataset} }\\
& Chair & Drums & Ficus & Hotdog & Lego & Materials & Mic & Ship \\ \hline
PixelNeRF & 3.988 & 5.02 & 4.06 & 4.22 & 4.73 & 4.32 & 4.24& 4.33\\
IBRNet & 21.70 & 17.97 & 22.08 & 20.73 & 18.66 & 15.79 & 19.67& 18.85\\
Ours & 23.35 & 20.71 & 21.98 & 28.44 & 23.18& 20.05 & 22.62 & 23.35\\ \hline
NeRF & 27.21 & 21.91 &25.33 &33.04 & 28.50 &26.38& 26.79& 26.44\\
Ours$_{ft-75min}$ & 26.60 & 22.35 & 23.60 & 30.94 & 27.1 & 24.26 & 25.84 & 26.62 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{PSNR quantity comparision on NeRF dataset}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\multicolumn{9}{c}{SSIM$\uparrow$ on \textbf{Realistic Synthetic Dataset}}\\
& Chair & Drums & Ficus & Hotdog & Lego & Materials & Mic & Ship \\ \hline
PixelNeRF & 0.428 & 0.485 & 0.479 & 0.492 & 0.477& 0.440 & 0.491& 0.405\\
IBRNet & 0.852& 0.817 & 0.895 & 0.863 & 0.823 & 0.285 & 0.899& 0.841\\
Ours & 0.876 & 0.886 & 0.898 & 0.962 & 0.902 & 0.893 & 0.923 & 0.886 \\ \hline
NeRF & 0.945& 0.901& 0.938& 0.978&0.952& 0.948& 0.959 & 0.892\\
Ours$_{ft-75min}$ & 0.93 & 0.90 & 0.922 & 0.971 & 0.944 & 0.926 & 0.953&0.904 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{SSIM quantity comparision on NeRF dataset}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\multicolumn{9}{c}{LPIPS $\downarrow$ on \textbf{Realistic Synthetic Dataset}}\\
& Chair & Drums & Ficus & Hotdog & Lego & Materials & Mic & Ship \\ \hline
PixelNeRF & 0.437 & 0.423 & 0.429 & 0.419 & 0.480& 0.414 & 0.433 & 0.461\\
IBRNet & 0.214 & 0.255 & 0.142 & 0.238 & 0.251 & 0.285 & 0.174 & 0.285\\
Ours & 0.282& 0.187 & 0.211 & 0.173 & 0.204 & 0.216 & 0.177 & 0.244 \\ \hline
NeRF & 0.127& 0.188&0.092& 0.111&0.100&0.107& 0.081& 0.282\\
Ours$_{ft-75min}$ & 0.226 & 0.246 & 0.304 & 0.202 & 0.243 & 0.188 & 0.215 & 0.212 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{LPIPS quantity comparision on NeRF dataset}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\multicolumn{9}{c}{PSNR$\uparrow$ on \textbf{LLFF Dataset}}\\
& Fern & Flower &Fortress & Horns & Leaves & Orchids & Room & Trex\\ \hline
PixelNeRF & 10.00 & 12.23 &13.90&10.90& 10.43 &11.67&11.23& 10.73\\
IBRNet & 18.45 & 16.70 &19.94&14.51& 14.84 &12.34& 19.31& 17.51\\
Ours & 17.17 & 22.82 & 21.50 &18.73&16.76& 18.90 &15.76&14.12\\\hline
NeRF$_{9.5h}$ & 22.60 & 25.72 &28.13&23.22& 19.04 &20.25&29.86& 22.58\\
Ours$_{ft-75min}$ & 22.42 & 27.39 & 21.84 & 27.24 & 18.98 & 19.84 & 27.89 & 25.39 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{PSNR quantity comparision f per-scene optimization on LLFF datasett}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\multicolumn{9}{c}{SSIM$\uparrow$ on \textbf{LLFF Dataset}}\\
& Fern & Flower & Fortress & Horns & Leaves & Orchids & Room & Trex\\ \hline
PixelNeRF & 0.402 & 0.375 &0.590&0.387& 0.192 & 0.223 &0.575&0.350\\
IBRNet & 0.608 & 0.566 &0.637&0.514&0.444&0.363&0.784& 0.614\\
Ours & 0.638 & 0.856 &0.861&0.702& 0.668 &0.751&0.687& 0.526\\ \hline
NeRF$_{9.5h}$ & 0.782& 0.864 &0.921&0.821& 0.676 &0.737&0.966&0.837\\
Ours$_{ft-75min}$ & 0.797 & 0.926&0.904 & 0.914 & 0.752&0.775 &0.948&0.912\\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{SSIM quantity comparision of per-scene optimization on LLFF dataset}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccccccc}
\multicolumn{9}{c}{LPIPS $\downarrow$ on \textbf{LLFF Dataset}}\\
& Fern & Flower & Fortress & Horns & Leaves & Orchids & Room & Trex\\ \hline
PixelNeRF & 0.738 & 0.794 &0.725&0.743& 0.754 &0.797&0.782& 0.761\\
IBRNet &0.453 & 0.498&0.456& 0.576& 0.487 &0.604&0.452& 0.454\\
Ours &0.407 & 0.212&0.273& 0.377& 0.348 &0.274&0.493& 0.494\\ \hline
NeRF$_{9.5h}$ & 0.341& 0.265&0.216& 0.400& 0.433& 0.337 & 0.174 &0.339\\
Ours$_{ft-75min}$ & 0.220&0.107&0.150&0.179&0.239&0.205&0.133&0.159 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{LPIPS quantity comparision of per-scene optimization on LLFF dataset}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|ccccc|ccccc}
\multicolumn{6}{c}{PSNR$\uparrow$ }& \multicolumn{5}{c}{SSIM $\uparrow$ }\\
& Scan1 & Scan8 & Scan21 & Scan103 & Scan114 & Scan1 & Scan8 & Scan21 & Scan103 & Scan114 \\ \hline
PixelNeRF & 23.33& 23.76& 18.77& 22.68&21.03& 0.849& 0.877& 0.729& 0.870&0.763\\
IBRNet & 26.81& 26.65& 21.23& 27.36&26.06& 0.891& 0.898& 0.811& 0.958&0.902\\
Ours & 25.33 & 22.96 & 20.92&26.08&18.59 & 0.92 & 0.884 &0.848 &0.931&0.938\\ \hline
NeRF$_{9.5h}$ & 23.43& 22.35& 22.82& 24.54&25.36& 0.870& 0.892& 0.858& 0.927&0.917\\
Ours$_{ft-75min}$& 30.61 & 27.05 & 24.58 & 33.06 & 29.07 & 0.959 & 0.934& 0.925&0.972 & 0.958\\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{PSNR quantity comparision f per-scene optimization on DTU datasett}.}
\label{tb:ablation}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{l|ccccc}
\multicolumn{6}{c}{LPIPS $\downarrow$ on \textbf{DTU Dataset}}\\
& Scan1 & Scan8 & Scan21 & Scan103 & Scan114 \\ \hline
PixelNeRF & 0.310& 0.301& 0.323& 0.321&0.324\\
IBRNet & 0.268& 0.268& 0.285& 0.141 & 0.203\\
Ours & 0.16 & 0.171 &0.206&0.205&0.164\\ \hline
NeRF$_{9.5h}$ & 0.262& 0.165&0.278& 0.308 & 0.215\\
Ours$_{ft-75min}$ & 0.109 &0.121&0.130&0.110&0.106 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.05pt}
\caption{\textbf{LPIPS quantity comparision of per-scene optimization on DTU dataset}.}
\label{tb:ablation}
\end{table*}
\section{Implementation details}
\label{sec:impl}
\boldstart{Dataset.}
We train our framework on the DTU~\cite{dtu} dataset to learn a generalizable network.
We follow PixelNeRF \cite{yu2020pixelnerf} to partition the data to $88$ training scenes and $16$ testing scenes, and use an image resolution of $512 \times 640$.
We also test our model (merely trained on DTU) on the Realistic Synthetic NeRF data~\cite{nerf} and the Forward-Facing data \cite{llff}, which have different scene and view distributions from our training set.
For each testing scene, we select $20$ nearby views; we then select 3 center views as input, 13 as additional input for per-scene fine-tuning, and take the remaining 4 as testing views.
\boldstart{Network details.} We use $f=32$ channels for feature extraction, which is also the number of feature channels in the cost volume and neural encoding volume (before appending color channels). We adopt $D=128$ depth hypotheses uniformly sampled from near to far to specify the plane sweep volume. Our MLP decoder is similar to the MLP of NeRF ~\cite{nerf}, but more compact, consisting of $6$ layers. Unlike NeRF reconstructing two (coarse and fine) radiance fields as separate networks, we only reconstruct one single radiance field and can already achieve good results; an extension to coarse-to-fine radiance fields can be potentially achieved at fine-tuning, by optimizing two separate encoding volumes with the same initialization.
For ray marching, we sample $128$ shading points on each marching ray.
We show detailed network structure in the supplementary materials.
We train our network using one \textit{RTX 2080 Ti} GPU. For the across-scene training on DTU, we randomly sample $1024$ pixels from one novel viewpoints as a batch, and use Adam ~\cite{kingma2014adam} optimizer with an initial learning rate of $5e-4$.
\begin{table*}[ht]
\centering
\begin{tabular}{lcccccccc}
\toprule
\multirow{2}{*}{Method} & \multirow{2}{*}{Settings} & \multicolumn{3}{c}{Synthetic Data (NeRF ~\cite{nerf_in_the_wild})} & \multicolumn{3}{c}{Real Data (DTU ~\cite{dtu} / Forward-Facing ~\cite{llff})} \\
\cmidrule(lr){3-5}\cmidrule(lr){6-8}
&& \multicolumn{1}{c}{PSNR$\uparrow$} & \multicolumn{1}{c}{SSIM$\uparrow$} & \multicolumn{1}{c}{LPIPS$\downarrow$} &
\multicolumn{1}{c}{PSNR$\uparrow$} &
\multicolumn{1}{c}{SSIM$\uparrow$} &
\multicolumn{1}{c}{LPIPS$\downarrow$} \\
\midrule
PixelNeRF~\cite{yu2020pixelnerf} & \multirow{3}{*}{\shortstack{No per-scene \\ optimization}} & 7.39 & 0.658 & 0.411 & 19.31/11.24 & 0.789/0.486 & 0.382/0.671 \\
IBRNet~\cite{ibrnet} & & 22.44 & 0.874 & 0.195 & 26.04/21.79 & 0.917/0.786 & 0.190/0.279 \\
Ours & & \textbf{23.62} & \textbf{0.897} & \textbf{0.176} & \textbf{26.63}/\textbf{21.93} & \textbf{0.931}/\textbf{0.795} & \textbf{0.168}/\textbf{0.252} \\
\midrule
NeRF$_{10.2h}$~\cite{nerf} & \multirow{3}{*}{\shortstack{Per-scene \\ optimization}} & \textbf{30.63} & \textbf{0.962} & \textbf{0.093} & 27.01/\textbf{25.97} & 0.902/0.870& 0.263/ 0.236 \\
IBRNet$_{ft-1.0h}$~\cite{ibrnet} & & 25.62 & 0.939 & 0.110 & \textbf{31.35}/24.88 & \textbf{0.956}/0.861 & \textbf{0.131}/\textbf{0.189} \\
Ours$_{ft-15min}$ & & 27.07 & 0.931 & 0.168 & 28.50/25.45 & 0.933/\textbf{0.877} & 0.179/0.192 \\
\bottomrule
\end{tabular}
\rule{0pt}{0.01pt}
\caption{\textbf{Quantitative results of novel view synthesis.} We show averaged results of PSNRs, SSIMs and LPISs on three different datasets.
On the top, we compare our method with concurrent neural rendering methods \cite{yu2020pixelnerf,ibrnet} with direct network inference. On the bottom, we show our fine-tuning results with only 15min optimization ($10k$ iterations), IBRNet 1.0h optimization ($10k$ iterations) and compare with NeRF's \cite{nerf} results with 10.2h optimization ($200k$ iterations).}
\label{tb:rendering}
\end{table*}
\section{Experiments}
We now evaluate our method and show our results.
\boldstartspace{Comparisons on results with three-image input.}
We compare with two recent concurrent works, PixelNeRF\cite{yu2020pixelnerf} and IBRNet \cite{ibrnet} that also aim to achieve the generalization of radiance field reconstruction.
We use the released code and trained model of PixelNeRF and retrain IBRNet on the DTU data (see Sec.~\ref{sec:impl}); we train and test these methods using 3 input views as used in our paper.
We compare all methods on three datesets \cite{nerf,dtu,llff} with the same input views and use 4 additional images to test each scene.
We show the quantitative results in Tab.~\ref{tb:rendering} and visual comparisons in Fig.~\ref{fig:rendering}.
As shown in Fig.~\ref{fig:rendering}, our approach can achieve realistic view synthesis results using only three images as input across different datasets.
While our model is trained only on DTU, it can generalize well to the other two datesets that have highly different scene and view distributions.
In contrast, PixelNeRF \cite{yu2020pixelnerf} tends to overfit the training setting on DTU.
Although it works reasonably on the DTU testing scenes, it contains obvious artifacts on the Realistic Synthetic scenes and even completely fails on the Forward-Facing scenes.
IBRNet \cite{ibrnet} can do a better job than PixelNeRF when testing on other datasets, but flicker artifacts can still be observed and much more obvious than ours as shown in the appendix video.
These visual results clearly reflect the quantitative results shown in Tab.~\ref{tb:rendering}.
The three methods can all obtain reasonable PSNRs, SSIMs and LPIPs on the DTU testing set.
However, our approach consistently outperforms PixelNeRF and IBRNet with the same input for all three metrics.
More impressively, our results on the other two testing datasets are significantly better than the comparison methods, clearly demonstrating the good generalizability of our technique.
In general, the two comparison methods both directly aggregate across-view 2D image features at ray marching points for radiance field inference.
Our approach instead leverages MVS techniques for geometry-aware scene reasoning in plane-swept cost volumes, and reconstructs a localized radiance field representation as a neural encoding volume with explicit 3D structures.
This leads to the best generalizablity and the highest rendering quality of our results across different testing scenes.
\begin{table}[t]
\centering
\begin{tabular}{lccc}
\toprule
Method & \multicolumn{1}{c}{Abs err$\downarrow$} & \multicolumn{1}{c}{Acc (0.01)$\uparrow$} &
\multicolumn{1}{c}{Acc (0.05)$\uparrow$} \\
\midrule
MVSNet & $\textbf{0.018} / \ \ \ - \ \ \ $ & $0.603/ \ \ \ - \ \ \ $ & $\textbf{0.955}/\ \ \ - \ \ \ $ \\
PixelNeRF & $0.245/0.239$ & $0.037/0.039$ & $0.176/0.187$ \\
IBRNet & $ 1.69 / 1.62 $ & $ 0.000 /0.000$ & $ 0.000 /0.001$\\
Ours & $0.023/\textbf{0.035}$ & $\textbf{0.746}/\textbf{0.717}$ & $0.913/\textbf{0.866}$\\
\bottomrule
\end{tabular}
\rule{0pt}{1.0pt}
\caption{\textbf{Depth reconstruction.} We evaluate our unsupervised depth reconstruction on the DTU testing set and compare with other two neural rendering methods (also without depth supervision) PixelNeRF ~\cite{yu2020pixelnerf} and IBRNet ~\cite{ibrnet}, and a learning based MVS method MVSNet ~\cite{yao2018mvsnet} that is trained with groundtruth depth. Our method significantly outperforms other neural rendering methods (PixelNeRF and IBRNet) and achieve high depth accuracy comparable to MVSNet. The two numbers of each item refers to the depth at reference/novel views; we mark with "-" when one does not have a reference/novel view. }
\label{tb:geometric}
\end{table}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=\linewidth, height=0.8\linewidth]{fig/result.pdf}
\end{center}
\caption{Rendering quality comparison. On the left, we show rendering results of our method and concurrent neural rendering methods \cite{yu2020pixelnerf,ibrnet} by directly running the networks. We show our 15-min fine-tuning results and NeRF's \cite{nerf} 10.2h-optimization results on the right.}
\label{fig:rendering}
\end{figure*}
\boldstartspace{Per-scene fine-tuning results.}
We also show our per-scene optimization results using 16 additional input images in Tab.~\ref{tb:rendering} and Fig.~\ref{fig:rendering}, generated by fine-tuning the neural encoding volume (with the MLP) predicted by our network (Sec.~\ref{sec:perscene}).
Because of the strong initialization obtained from our network,
we only fine-tune our neural reconstruction for a short period of 15 minutes (10k iterations), which can already lead to photo-realistic results.
We compare our fast fine-tuning results with NeRF's \cite{nerf} results generated with substantially longer optimization time (as long as 10.2 hours).
Note that, our initial rendering results can be significantly boosted with even only 15min fine-tuning; this leads to high-quality results that are on par (Realistic Synthetic) or better (DTU and Forward-Facing) than NeRF's results with 30 times longer optimization time.
We also show results on one example scene that compare the optimization progresses of our method and NeRF with different optimization times in Fig.~\ref{fig:progress}, which clearly demonstrates the significantly faster convergence of our technique.
By taking our generic network to achieve strong initial radiance field, our approach enables highly practical per-scene radiance field reconstruction when dense images are available.
\boldstartspace{Depth reconstruction.}
Our approach reconstructs a radiance field that represents scene geometry as volume density. We evaluate our geometry reconstruction quality by comparing depth reconstruction results, generated from the volume density by a weighted sum of the depth values of the sampled points on marched rays (as is done in \cite{nerf}). We compare our approach with the two comparison radiance field methods \cite{yu2020pixelnerf,ibrnet} and also the classic deep MVS method MVSNet \cite{yao2018mvsnet} on the DTU testing set.
Thanks to our cost-volume based reconstruction, our approach achieves significantly more accurate depth than the other neural rendering methods \cite{yu2020pixelnerf,ibrnet}.
Note that, although our network is trained with only rendering supervision and no depth supervision, our approach can achieve high reconstruction accuracy comparable to the MVS method \cite{yao2018mvsnet} that has direct depth supervision.
This demonstrates the high quality of our geometry reconstruction, which is one critical factor that leads to our realistic rendering.
\section{Conclusion}
We present a novel generalizable approach for high-quality radiance field reconstruction and realistic neural rendering.
Our approach combines the main advantages of deep MVS and neural rendering,
successfully incorporating cost-volume based scene reasoning into physically based neural volumetric rendering.
Our approach enables high-quality radiance field reconstruction from only three input views and can achieve realistic view synthesis results from the reconstruction.
Our method generalizes well across diverse testing datasets and can significantly outperform concurrent works \cite{yu2020pixelnerf,ibrnet} on generalizable radiance field reconstruction.
Our neural reconstruction can also be fine-tined easily for per-scene optimization, when dense input images are available, allowing us to achieve photo-realistic renderings that are better than NeRF while using substantially less optimization time.
Our work offers practical neural rendering techniques using either few or dense images as input.
\section{Acknowledgements}
This work was supported by NSFC programs (61976138, 61977047); the National Key Research and Development Program (2018YFB2100500); STCSM (2015F0203-000-06) and SHMEC (2019-01-07-00-01-E00003); NSF grant IIS-1764078 and gift money from VIVO.
\section{Introduction}
Novel view synthesis is a long-standing problem in computer vision and graphics.
Recently, neural rendering approaches have significantly advanced the progress in this area. Neural radiance fields (NeRF) and its following works~\cite{nerf,nerf_in_the_wild,Neural_sparse_voxel_fields} can already produce photo-realistic novel view synthesis results.
However, one significant drawback of these prior works is that they require a very long per-scene optimization process to obtain high-quality radiance fields, which is expensive and highly limits the practicality.
Our goal is to make neural scene reconstruction and rendering more practical, by enabling \emph{highly efficient} radiance field estimation.
We propose MVSNeRF, a novel approach that \emph{generalizes well across scenes} for the task of reconstructing a radiance field from \emph{only several} (as few as three) unstructured multi-view input images.
With strong generalizability, we avoid the tedious per-scene optimization and can directly regress realistic images at novel viewpoints via fast network inference. If further optimized on more images with only a short period (5-15 min), our reconstructed radiance fields can even outperform NeRFs \cite{nerf} with hours of optimization (see Fig.~\ref{fig:teaser}).
We take advantage of the recent success on deep multi-view stereo (MVS) \cite{yao2018mvsnet,gu2020cascade,cheng2020deep}. This line of work can train generalizable neural networks for the task of 3D reconstruction by applying 3D convolutions on cost volumes.
Similar to \cite{yao2018mvsnet}, we build a cost volume at the input reference view by warping 2D image features (inferred by a 2D CNN) from nearby input views onto sweeping planes in the reference view's frustrum.
Unlike MVS methods \cite{yao2018mvsnet,cheng2020deep} that merely conduct depth inference on such a cost volume,
our network reasons about both scene geometry and appearance, and outputs a neural radiance field (see Fig.~\ref{fig:pipeline}), enabling view synthesis.
Specifically, leveraging 3D CNN, we reconstruct (from the cost volume) a \emph{neural scene encoding volume} that consists of per-voxel neural features that encode information about the local scene geometry and appearance.
Then, we make use of a multi-layer perceptron (MLP) to decode the volume density and radiance at arbitrary continuous locations using tri-linearly interpolated neural features inside the encoding volume.
In essence, the encoding volume is a localized neural representation of the radiance field; once estimated, this volume can be used directly (dropping the 3D CNN) for final rendering by differentiable ray marching (as in \cite{nerf}).
Our approach takes the best of the two worlds, learning-based MVS and neural rendering. Compared with existing MVS methods, we enable differentiable neural rendering that allows for training without 3D supervision and inference time optimization for further quality improvement. Compared with existing neural rendering works, our MVS-like architecture is natural to conduct cross-view correspondence reasoning, facilitating the generalization to unseen testing scenes and also leading to better neural scene reconstruction and rendering.
Our approach can, therefore, significantly outperform the recent concurrent generalizable NeRF work \cite{yu2020pixelnerf,ibrnet} that mainly considers 2D image features without explicit geometry-aware 3D structures (See Tab.~\ref{tb:rendering} and Fig.~\ref{fig:rendering}).
We demonstrate that, using only three input images, our network trained from the DTU dataset can synthesize photo-realistic images on testing DTU scenes, and can even generate reasonable results on other datasets that have very different scene distributions.
Moreover, our estimated three-image radiance field (the neural encoding volume) can be further easily optimized on novel testing scenes to improve the neural reconstruction if more images are captured, leading to photo-realistic results that are comparable or even better than a per-scene overfit NeRF, despite of ours taking substantially less optimization time than NeRF (see Fig.~\ref{fig:teaser}).
These experiments showcase that our technique can be used either as a strong reconstructor that can reconstruct a radiance field for realistic view synthesis when there are only few images captured, or as a strong initializor that significantly facilitates the per-scene radiance field optimization when dense images are available.
Our approach takes an important step towards making realistic neural rendering practical. We have released the code at \href{https://apchenstu.github.io/mvsnerf/}{mvsnerf.github.io}.
\section{MVSNeRF}
We now present our MVSNeRF. Unlike NeRF ~\cite{nerf} that reconstructs a radiance field via a per-scene "network memorization", our MVSNeRF learns a generic network for radiance field reconstruction.
Given $M$ input captured images $I_i$ ($i=1,..,M$) of a real scene and their known camera parameters $\Phi_i$, we present a novel network that can reconstruct a radiance feild as a neural encoding volume and use it to regress volume rendering properties (density and view-dependent radiance) at arbitrary scene locations for view synthesis. In general, our entire network can be seen as a function of a radiance field, expressed by:
\begin{equation}
\sigma, r = \mathrm{MVSNeRF}(x, d; I_i, \Phi_i)
\label{eq:mvsnerf}
\end{equation}
where $x$ represents a 3D location, $d$ is a viewing direction, $\sigma$ is the volume density at $x$, and $r$ is the output radiance (RGB color) at $x$ depending on the viewing direction $d$.
The output volume properties from our network can be directly used to synthesize a novel image $I_t$ at a novel target viewpoint $\Phi_t$ via differentiable ray marching.
In this paper, we consider a sparse set of nearby input views for efficient radiance field reconstruction. In practice we use $M=3$ views for our experiments, while our approach handles unstructured views and can easily support other numbers of inputs.
The overview of our MVSNeRF is shown in Fig.~\ref{fig:pipeline}.
We first build a cost volume at the reference view (we refer to the view $i=1$ as the reference view) by warping 2D neural features onto multiple sweeping planes (Sec.~\ref{sec:costvolume}).
We then leverage a 3D CNN to reconstruct the neural encoding volume, and use an MLP to regress volume rendering properties, expressing a radiance field (Sec.~\ref{sec:radiancefield}).
We leverage differentiable ray marching to regress images at novel viewpoints using the radiance field modeled by our network;
this enables end-to-end training of our entire framework with a rendering loss (Sec.~\ref{sec:training}).
Our framework achieves radiance field reconstruction from few images.
On the other hand, when dense images are captured, the reconstructed encoding volume and the MLP decoder can also be fast fine-tuned independently to further improve the rendering quality (Sec.~\ref{sec:perscene}).
\subsection{ Cost volume construction.}
\label{sec:costvolume}
Inspired by the recent deep MVS methods \cite{yao2018mvsnet},
we build a cost volume $P$ at the reference view ($i=1$), allowing for geometry-aware scene understanding.
This is achieved by warping 2D image features from the $m$ input images to a plane sweep volume on the reference view's frustrum.
\boldstartspace{Extracting image features.}
We use a deep 2D CNN $T$ to extract 2D image features at individual input views to effectively extract 2D neural features that represent local image appearance.
This sub-network consists of downsampling convolutional layers and convert an input image $I_i \in \mathbb{R}^{H_i \times W_i \times 3}$ into a 2D feature map $F_i \in \mathbb{R}^{H_i/4 \times W_i/4 \times C}$,
\begin{equation}
\begin{aligned}
F_i = T(I_i),
\label{eq:extract}
\end{aligned}
\end{equation}
where $H$ and $W$ are the image height and width, and $C$ is the number of resulting feature channels.
\boldstartspace{Warping feature maps.}
Given the camera intrinsic and extrinsic parameters $\Phi=[K, R, t]$, we consider the homographic warping
\begin{equation}
\begin{aligned}
\mathcal{H}_i(z) = K_i \cdot( R_i \cdot R_1^T + \frac{(t_1-t_i)\cdot n^{T}_1}{z}) \cdot K^{-1}_{1}
\label{eq:homograph}
\end{aligned}
\end{equation}
where $\mathcal{H}_i(z)$ is the matrix warping from the view $i$ to the reference view at depth $z$, $K$ is the intrinsic matrix, and $R$ and $t$ are the camera rotation and translation. Each feature map $F_i$ can be warped to the reference view by:
\begin{equation}
\begin{aligned}
F_{i,z}(u,v) = F_i(\mathcal{H}_i(z)[u,v,1]^T),
\label{eq:warping}
\end{aligned}
\end{equation}
where $F_{i,z}$ is the warped feature map at depth $z$, and $(u,v)$ represents a pixel location in the reference view.
In this work, we parameterize $(u,v,z)$ using the normalized device coordinate (NDC) at the reference view.
\boldstartspace{Cost volume.} The cost volume $P$ is constructed from the warped feature maps on the $D$ sweeping planes. We leverage the variance-based metric to compute the cost, which has been widely used in MVS \cite{yao2018mvsnet,cheng2020deep} for geometry reconstruction.
In particular, for each voxel in $P$ centered at $(u,v,z)$, its cost feature vector is computed by:
\begin{equation}
\begin{aligned}
P(u,v,z) = \mathrm{Var}(F_{i,z}(u,v)),
\label{eq:cost}
\end{aligned}
\end{equation}
where $\mathrm{Var}$ computes the variance across $M$ views.
This variance-based cost volume encodes the image appearance variations across different input views; this explains the appearance variations caused by both scene geometry and view-dependent shading effects.
While MVS work uses such a volume only for geometry reconstruction, we demonstrate that it can be used to also infer complete scene appearance and enable realistic neural rendering.
\subsection{Radiance field reconstruction.}
\label{sec:radiancefield}
We propose to use deep neural networks to effectively convert the built cost volume into a reconstruction of radiance field for realistic view synthesis.
We utilize a 3D CNN $B$ to reconstruct a neural encoding volume $S$ from the cost volume $P$ of raw 2D image feature costs; $S$ consists of per-voxel features that encode local scene geometry and appearance. An MLP decoder $A$ is used to regress volume rendering properties from this encoding volume.
\boldstartspace{Neural encoding volume.}
Previous MVS works \cite{yao2018mvsnet,gu2020cascade,cheng2020deep} usually predict depth probabilities directly from a cost volume, which express scene geometry only.
We aim to achieve high-quality rendering that necessitates inferring more appearance-aware information from the cost volume.
Therefore, we train a deep 3D CNN $B$ to transform the built image-feature cost volume into a new $C$-channel neural feature volume $S$, where the feature space is learned and discovered by the network itself for the following volume property regression.
This process is expressed by:
\begin{equation}
\begin{aligned}
S = B(P).
\label{eq:encoding}
\end{aligned}
\end{equation}
The 3D CNN $B$ is a 3D UNet with downsampling and upsampling convolutional layers and skip connections, which can effectively infer and propagate scene appearance information, leading to a meaningful scene encoding volume $S$.
Note that, this encoding volume is predicted in a unsupervised way and inferred in the end-to-end training with volume rendering (see Sec.~\ref{sec:training}). Our network can learn to encode meaningful scene geometry and appearance in the per-voxel neural features;
these features are later continuously interpolated and converted into volume density and view-dependent radiance.
The scene encoding volume is of relative low resolution because of the downsampling of 2D feature extraction; it is challenging to regress high-frequency appearance from this information alone.
We thus also incorporate the original image pixel data for the following volume regression stage, though we later show that this high-frequency can be also recovered in an augmented volume via a fast per-scene fine-tuning optimization (Sec.~\ref{sec:perscene}).
\boldstartspace{Regressing volume properties.}
Given an arbitrary 3D location $x$ and a viewing direction $d$, we use an MLP $A$ to regress the corresponding volume density $\sigma$ and view-dependent radiance $r$ from the neural encoding volume $S$. As mentioned, we also consider pixel colors $c=[I(u_i,v_i)]$ from the original images $I_i$ as additional input; here $(u_i,v_i)$ is the pixel location when projecting the 3D point $x$ onto view $i$, and $c$ concatenates the colors $I(u_i, v_i)$ from all views as a $3M$-channel vector. The MLP is expressed by:
\begin{equation}
\begin{aligned}
\sigma, r = A(x, d, f, c),\quad f=S(x),
\label{eq:decoding}
\end{aligned}
\end{equation}
where $f=S(x)$ is the neural feature trilinearly interpolated from the volume $S$ at the location $x$.
In particular, $x$ is parameterized in the reference view's NDC space and $d$ is represented by a unit vector at the reference view's coordinate.
Using NDC space can effectively normalize the scene scales across different data sources, contributing to the good generalizability of our method.
In addition, inspired by NeRF \cite{nerf}, we also apply positional encoding on the position and direction vectors ($x$ and $d$), which further enhance the high-frequency details in our results.
\boldstartspace{Radiance field.}
As a result, our entire framework models a neural radiance field, regressing volume density and view-dependent radiance in the scene from few (three) input images.
In addition, once the scene encoding volume $S$ is reconstructed, this volume combined with the MLP decoder $A$ can be used independently without the prepending 2D and 3D CNNs.
They can be seen as a standalone neural representation of the radiance field, outputting volume properties and thus supporting volume rendering.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{fig/room.pdf}
\end{center}
\caption{Optimization progress. We show results of our fine-tuning (top) and optimizing a NeRF \cite{nerf} (bottom) with different time periods. Our 0-min result refers to the initial output from our network inference. Note that our 18-min results are already much better than the 215-min NeRF results. PSNRs of the image crops are shown in the figure.
}
\label{fig:progress}
\end{figure*}
\subsection{Volume rendering and end-to-end training.}
\label{sec:training}
Our MVSNeRF reconstructs a neural encoding volume and regresses volume density and view-dependent radiance at arbitrary points in a scene. This enables applying differentiable volume rendering to regress images colors.
\boldstartspace{Volume rendering.} The physically based volume rendering equation can be numerically evaluated via differentiable ray marching (as is in NeRF \cite{nerf}) for view synthesis. In particular, a pixel's radiance value (color) is computed by marching a ray through the pixel and accumulating radiance at sampled shading points on the ray, given by:
\begin{equation}
\begin{aligned}
c_t &= \sum_k\tau_k (1-\exp (-\sigma_k)) r_k, \\
\tau_k &= \exp (-\sum_{j=1}^{k-1} \sigma_j ),
\label{eq:raymarching}
\end{aligned}
\end{equation}
where $c_t$ is the final pixel color output, and $\tau$ represents the volume transmittance. Our MVSNeRF as a radiance field function essentially provides the volume rendering properties $\sigma_k$ and $r_k$ for the ray marching.
\boldstartspace{End-to-end training.} This ray marching rendering is fully differentiable; it thus allows our framework to regress final pixel colors at novel viewpoints using the three input views from end to end.
We supervise our entire framework with the groundtruth pixel colors, using an L2 rendering loss:
\begin{equation}
\begin{aligned}
L = \|c_t - \tilde{c}_t \|^2_2,
\label{eq:loss}
\end{aligned}
\end{equation}
where $\tilde{c}_t$ is the groundtruth pixel color sampled from the target image $I_t$ at a novel viewpoint.
This is the only loss we use to supervise our entire system.
Thanks to the physically based volume rendering and end-to-end training, the rendering supervision can propagate the scene appearance and correspondence information through every network components and regularize them to make sense for final view synthesis.
Unlike previous NeRF works \cite{nerf,nerf_in_the_wild,Neural_sparse_voxel_fields} that mainly focus on per-scene training,
we train our entire network across different scenes on the DTU dataset.
Our MVSNeRF benefits from the geometric-aware scene reasoning in cost volume processing and can effectively learn a generic function that can reconstruct radiance fields as neural encoding volumes on novel testing scenes enabling high-quality view synthesis.
\subsection{Optimizing the neural encoding volume.}
\label{sec:perscene}
When training across scenes, our MVSNeRF can already learn a powerful generalizable function, reconstructing radiance fields across scenes from only three input images.
However, because of the limited input and the high diversity across different scenes and datasets, it is highly challenging to achieve perfect results on different scenes using such a generic solution.
On the other hand, NeRF avoids this hard generalization problem by performing per-scene optimization on dense input images;
this leads to photo-realistic results but is extremely expensive.
In contrast, we propose to fine-tune our neural encoding volume -- that can be instantly reconstructed by our network from only few images -- to achieve fast per-scene optimization when dense images are captured.
\boldstartspace{Appending colors.} As mentioned, our neural encoding volume is combined with pixel colors when sent to the MLP decoder (Eqn.~\ref{eq:decoding}). Retaining this design for fine-tuning still works but leads to a reconstruction that always depends on the three inputs.
We instead achieve an independent neural reconstruction by appending the per-view colors of voxel centers as additional channels to the encoding volume; these colors as features are also trainable in the per-scene optimization.
This simple appending initially introduces blurriness in the rendering, which however is addressed very quickly in the fine-tuning process.
\boldstartspace{Optimization.}
After appended with colors, the neural encoding volume with the MLP is a decent initial radiance field that can already synthesize reasonable images. We propose to further fine-tune the voxel features along with the MLP decoder to perform fast per-scene optimization when dense images are available.
Note that, we optimize only the encoding volume and the MLP, instead of our entire network.
This grants more flexibility to the neural optimization to adjust the per-voxel local neural features independently upon optimization; this is an easier task than trying to optimize shared
convolutional operations across voxels.
In addition, this fine-tuning avoids the expensive network processing of the 2D CNN, plane-sweep warping, and 3D CNN.
As a result, our optimization can therefore be very fast, taking substantially less time than optimizing a NeRF from scratch, as shown in Fig.~\ref{fig:progress}.
Our per-scene optimization leads to a clean neural reconstruction, independent of any input image data (thanks to appending color channels), similar to \cite{nerf,Neural_sparse_voxel_fields}; the dense input images can be therefore dropped after optimization.
In contrast, the concurrent works \cite{yu2020pixelnerf,ibrnet} require retaining the input images for rendering.
Our encoding volume is also similar to Sparse Voxel fields \cite{Neural_sparse_voxel_fields}; however ours is initially predicted by our network via fast inference, instead of the pure per-scene optimization in \cite{Neural_sparse_voxel_fields}.
On the other hand, we can (as future work) potentially subdivide our volume grid in the fine-tuning for better performance as is done in \cite{Neural_sparse_voxel_fields}.
\section{MVSNeRF}
Given neighbor source views $I_s$, our method attempts to synthesis high quality target view images $I_t$ without/with small amount feed-forward optimization. Unlike NeRF ~\cite{xxx} rendering free-viewpoint video (FVV) via a per-scene "network memory", our method borrow ideas from traditional image-based-rendering (IBR) methods ~\cite{xxx} that to learn a view interpolation rules from given nearby sampled view rendering, which enabling generalize our method to novel scenes. Our pipeline (Fig. ~\ref{fig:pipeline}) consistences three parts: 1) feature extraction and cost-volume construction, which extracting image feature from each source view and learns multi-view correspondences with 3D CNN based feature aggregation, and 2) rendering and regressing target images (RGB-$\sigma$) through a volumetric rendering scheme. Then, 3) we present how to rendering FVV from given any number of neighboring views in Sec. ~\ref{sec:Extension to novel scenes}.
In a nutshell, our complete novel view image rendering process can be formulate as:
\begin{equation}
I_t = \mathrm{MVSNeRF}(I_s, \mathcal{C}_s, \mathcal{C}_t) = \mathcal{R}(x, \mathbf{d}, f, \tilde{c})
\label{eq:mvsnerf}
\end{equation}
where $\mathcal R$ denotes the volumetric renderer which taking normalized device coordinates (NDC) $x \in \mathbb{R}^3 \ | \ x = z \mathbf{d}$ with depth $z$, ray directions $\mathbf{d}$ under referent view space, volume features $f$ and skip color connection $\widetilde{c}$ as input. More specifically,
\begin{equation}
f = \mathcal{G}(x, Cost(I_s, \mathcal{C})); \ \tilde{c} = \mathcal{G}(\mathcal{H}(x, \mathcal{C}), I_s)
\label{eq:mvsnerf_notation}
\end{equation}
$\mathcal{C} = [R,T,K]$ stands for the source and target camera poses modeling with three dimensional rotation $R$, translation $T$ and intrinsic $K$. $\mathcal{G}$ is a linear grid interpolation operation inside the convolved cost-volume and $\mathcal{H}$ is homograph wrapping from volume points to source image plan.
\subsection{Feature extraction and Cost-volume construction}
\textbf{Extracting image feature.} Instead of representing entry scenes, our method attempts to learn local radiance fields from fixing number (e.g., $3$ views in experiment) of source images $\textbf{I}^i_s \in \mathbb{R}^{H_i \times W_i \times 3}, i \in [1,2,3]$, similar to most image process tasks, we extract image feature $\textbf{F}_i \in \mathbb{R}^{H_i/4 \times W_i/4 \times C} $ with shared 2D CNN layers to capture multi-level information for the following regression phase.
\textbf{Learning multi-view correspondences: Cost-Volume}
Given referent views and their feature maps, we set up to build a cost volume that include epipolar geometric constant and resolve 3D geometric ambiguities . Unlike most neural scene representations learn geometric constant via Markov process ~\cite{NeRF like}, or point feature matching ~\cite{GRF, pixelnerf, ibrnet} (i.e., indexing and aggregating sample points' features via a projection scheme). The continuity of neighboring points aren't welly represented and explored, thus we
apply two volumes cost-volume and feature volume for surface radiance and geometric modeling, which are build on referent camera space, as shown in Fig.~\ref{fig:pipeline}.
We borrow the cost volume idea from multi-view stereo matching networks ~\cite{xxx} and further to extent it to a continue volume with a unsupervised volume rendering scheme. Similar to previous multi-view stereo matching networks, we utilize plane hypothesis to compute a cost variance volume from the extracted feature maps,
\begin{equation}
\begin{aligned}
\mathcal{H}_i(t) = K_i \cdot( R_i \cdot R_1^T + \frac{(t_1-t_i)\cdot n^{T}_1}{z}) \cdot K^{-1}_{1}
\label{eq:homograph}
\end{aligned}
\end{equation}
where $\mathcal{H}_i(z)$ refers to the homograph between the feature maps of the $i^{th}$ view and the reference view at depth $z$, $K_i, R_i, t_i$ denote the camera intrinsic/rotation/translation matrix respectively and $n_1$ is the look-at direction of the reference camera. Note that, we use referent view coordinate system instead of canonical coordinate frame in NeRF like representations, since no clear canonical coordinate system exists on scenes with multiple objects or real scenes.
Thereby, we build a perspective cost volume grid $PCV$ to modeling the scene (as shown in the middle of Fig.~\ref{xxx}). Feature $f_p$ of each grid point $p(u,v,z) \in \mathbb{R}^3$ is aggregated as,
\begin{equation}
\begin{aligned}
PCV(p) &= f_p = \frac{1}{n} <f^1_p, ..., f^m_p>\\
f^i_p &= \mathcal{G}( F_i, H_i(z) \cdot [u,v,1]^T)
\label{eq:cost volume}
\end{aligned}
\end{equation}
where $<,>$ denotes the inner product between source view features under the same grid point $p$, while $m$ is the number of the sources views and $m=3$ in our experiment.
Next, we adopt a 3D CNN on the cost volume to matching features and explicitly enforce neighboring volume points share contents. However, 3D CNN on the grid points couldn't guarantee volume continuity and usually result in the well-known edge-fattening issue in MVSNets, to address this issue, we propose to use a volume rendering scheme after the 3D CONV, which enable preserving the volume continuity.
\subsection{Rendering Images}
\textbf{Predicting RGB-$\sigma$}
Our method synchronously optimize geometric and radiance field with the classical volume rendering: accumulating colors and densities along rays to render image. Unlike NeRF-like methods ~\cite{nerf} modelling a continue scene as a $5D$ vector-valued functions, which taking $3D$ position $p$ and $2D$ viewing direction $(\theta,\phi)$ as inputs. Since geometric and radiance fields can be varying across scenes, coordinate-based function mapping lacking the generalizationbility, we extend the classical $5D$ coordinate inputs with convolved volume feature $f \in \mathbb{R}^{M}$ and color $c \in \mathbb{R}^{n \times 3}$ to handle scene specify information:
\begin{equation}
\mathcal{R}(\mathbf{r}) = \int_{t_n}^{t_f}T(t)\sigma(t)\textbf{c}(t)dt
\end{equation}
Where $\mathbf{r} = (x,\mathbf{d}, f, c)$ is point samples along the ray direction, $T(t) = exp(-\int_{t_n}^{t}\sigma(\textbf{r}(s))ds)$ is the "cumproded" density along the ray from $t_n$ to $t$, while the $\mathbf{c},\sigma$ is the RGB and density of the given spacial point $\mathbf{x}$ its direction $\mathbf{d}$ toward to viewpoint. Under this modeling, target views can be obtained by casting rays to the volume and thus we can easily regressive the scene via a differential rendering scheme.
\textbf{Training Objective.}
Since our cost volume construction and volume renderer are fully differentiable, our MVSNeRF can optimize end-to-end. For each iteration, We sample $4$ pair-wised neighboring images from training set and use $3$ of them to construct cost volume and the lefr one as target view. We sample rays emitting from the target view and minimize the mean squared error between the rendered colors and ground truth pixel colors for training:
\begin{equation}
\mathcal{L} = \sum_{\mathbf{r} \in R} || I_t^\mathbf{r} - \mathcal{R}(\mathbf{r}) ||^2
\end{equation}
\textbf{Soft volume fusion.} Rendering novel view only from single local metric inevitably lead to artifacts on such occlusion areas, thus it is naturally predict target images by fusing neighboring local volumes. Instead of directly fusing final rendering results obtained from nearby volumes or average pooling in feature space ~\cite{pixelnerf}. We propose an explicit weighting function that softly fuse nearby local volumes in feature space, and fine-tuning the network by fixing the weight function to ensure that the space is interpolable.
First, we build $m$ $PCV$ from $m$ source view images with Eq.~\ref{eq:cost volume}. Then, we present a fusion scheme analogous to unstructured lumigraph \cite{buehler2001unstructured}, that adaptively modify the weights according to visibility, the angular distance between the novel view and the source volumes. For each volume point $x$, the weight of $i_{th}$ volume as:
\begin{equation}
\begin{aligned}
w_i = \frac{ \tau_i \cdot e^{-\alpha \cdot (1.0 + n_i \cdot \mathbb{d})}}{\sum_{i=1}^{m} \tau_i \cdot e^{-\alpha \cdot (1.0 + n_i \cdot \mathbb{d})}}
\end{aligned}
\end{equation}
where $m$ denotes the number views, $n_i$ is the lookat direction of $i_{th}$ source view and $\tau_i \in [0.0,1.0]$ is the one-hot label that determine the sample point $x$ whether inside the referent volume $i$. $\alpha$ is a hyperparameter, we set $30$ in our experiment.
Finally, we fine-tuning the network by fixing the weighting function.
\subsection{Extension to any number of nearby views}
\label{sec:Extension to novel scenes}
Furthermore, it's naturally to extend our framework to allow for an arbitrary number of views at training or testing time, thus can prove additional information about the scene and enabling to render larger free viewpoint image baseline. To this end, we can render each target view with two ways: 1) volume interpolation between the source local volumes, or 2) fine tuning the network to better represent the specify scene from the pre-trained modules. Before the interpolation and fine-tuning, we need to select the referent views for each target view.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{LaTeX/fig/view_seletion.pdf}
\end{center}
\caption{xxx}
\label{fig:view_seletion}
\end{figure}
\textbf{Referent View selection.} Given multiply neighboring source views $I_s$ and the pose of target virtual view $[R_t,t]$, we select $3$ reference views base on its position and lookat direction. As mentioned in the above sections, our pipeline take $3$ neighboring images for the cost volume construction and radiance field rendering in training stage. Note that our cost volume construction is not limited by the input view number since the inner product operation in Eq.~\ref{eq:cost volume} is order and number free, thus we can use multiply input views for a better cost volume construction, and select three specific views for the color skipping (i.e., xx line of Fig. ~\ref{xx}).
To better recover target views, we propose a view triangulation scheme to make sure the pair-wised referent views are include as much radiance information as required. As shown in Fig~\ref{xxx}, we first select $V^i_p, i \in [1,M]$ and $V^l_l, i \in [1,N]$ nearest views via the position and lookat-angle distance respectively. Next, we used the first $K$ intersection views as the referent view candidates, i.e. $V_p \cap V_l$. Furthermore, we pair the view candidates condition on the target view is "inside" the triangles formed with referent view position, as (b) of Fig.~\ref{xxx}. We project the camera position of candidate views to the camera space of the target view with their extrinsic parameters and classify each camera into its' azimuth:
\begin{equation}
\begin{aligned}
C(p) &= f_p = \frac{1}{3} <f^1_p, f^2_p, f^3_p>\\
f^i_p &= \mathcal{G}( F_i, H_i(d) \cdot [x,y,1]^T)
\label{eq:view selection}
\end{aligned}
\end{equation}
Next, we group the cameras inside bin $0,2,4$ and $1,3,5$ to be the referent pairs that can make sure target view are "inside" Referent-View-Triangles and thus can significant avoid the artifacts cased by occlusion (demonstrated in Appendix).
\textbf{Fine-tuning.}
By combining the above Soft volume fusion and referent view selection, we can achieve large baseline or even $360^\circ$ free viewpoint rendering just from the pre-train models on large scale dataset. However, ghosting artifacts and blurry edges are generally observed when view changing due to the limited compatibility of the trained models. Thus, we include a small number of feed-forward optimization on the input views. Instead of using local ray direction and NDC position in the previous sections, we use global scene coordinate when optimizing on a specify scene. Fine-tuning only on a specific scene with scene coordinate can improve performance on view consistence, geometric and radiance.
\section{Overview}
\section{Related Work}
\boldstart{Multi-view stereo.}
Multi-view stereo (MVS) is a classical computer vision problem,
aiming to achieve dense geometry reconstruction using images captured from multiple viewpoints,
and has been extensively explored by various traditional methods \cite{de1999poxels,kutulakos2000theory,kolmogorov2002multi,esteban2004silhouette,seitz2006comparison,furukawa2010accurate,schonberger2016pixelwise}.
Recently, deep learning techniques have been introduced to address MVS problems \cite{yao2018mvsnet,im2018dpsnet}.
MVSNet \cite{yao2018mvsnet} applies a 3D CNN on a plane-swept cost volume at the reference view for depth estimation, achieving high-quality 3D reconstruction that outperforms classical traditional methods \cite{furukawa2010accurate,schonberger2016pixelwise}.
Following works have extended this technique with recurrent plane sweeping \cite{yao2019recurrent}, point-based densification \cite{chen2019point}, confidence-based aggregation \cite{luo2019p}, and multiple cost volumes \cite{cheng2020deep,gu2020cascade}, improving the reconstruction quality.
We propose to combine the cost-volume based deep MVS technique with differentiable volume rendering, enabling efficient reconstruction of radiance fields for neural rendering.
Unlike MVS approaches that use direct depth supervision,
we train our network with image loss only for novel view synthesis. This ensures the network to satisfy multi-view consistency, naturally allowing for high-quality geometry reconstruction.
As a side product, our MVSNeRF can achieve accurate depth reconstruction (despite of no direct depth supervision) comparable to the MVSNet \cite{yao2018mvsnet}.
This can potentially inspire future work on developing unsupervised geometry reconstruction methods.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{fig/pipeline.pdf}
\end{center}
\caption{Overview of MVSNeRF. Our framework first constructs a cost volume (a) by warping 2D image features onto a plane sweep. We then apply 3D CNN to reconstruct a neural encoding volume with per-voxel neural features (b). We use an MLP to regress volume density and RGB radiance at an arbitrary location using features interpolated from the encoding volume. These volume properties are used by differentiable ray marching for final rendering (c). }
\label{fig:pipeline}
\end{figure*}
\boldstartspace{View synthesis.}
View synthesis has been studied for decades with various approaches including light fields \cite{gortler1996lumigraph,levoy1996light,wood2000surface,kalantari2016learning,srinivasan2017learning,chen2018deepsurface}, image-based rendering \cite{debevec1996modeling,buehler2001unstructured,sinha2009piecewise,chaurasia2011silhouette,chaurasia2013depth}, and other recent deep learning based methods \cite{zhou2016view,zhou2018stereo,xu2019deep,flynn2016deepstereo}.
Plane sweep volumes have also been used for view synthesis \cite{penner2017soft,zhou2018stereo,flynn2016deepstereo,llff,xu2019deep}.
With deep learning, MPI based methods \cite{zhou2018stereo,extremeview,llff,srinivasan2019pushing} build plane sweep volumes at reference views, while other methods \cite{flynn2016deepstereo, xu2019deep} construct plane sweeps at novel viewpoints;
these prior works usually predict colors at the discrete sweeping planes and aggregate per-plane colors using alpha-blending or learned weights.
Instead of direct per-plane color prediction, our approach infers per-voxel neural features in the plane sweep as a scene encoding volume and can regress volume rendering properties from it at arbitrary 3D locations. This models a continuous neural radiance field, allowing for physically based volume rendering to achieve realistic view synthesis.
\boldstartspace{Neural rendering.}
Recently, various neural scene representations have been presented to achieve view synthesis and geometric reconstruction tasks \cite{zhou2018stereo,thies2019deferred,lombardi2019neural,bi2020deep,nerf}.
In particular, NeRF \cite{nerf} combines MLPs with differentiable volume rendering and achieves photo-realistic view synthesis.
Following works have tried to advance its performance on view synthesis \cite{nerf_in_the_wild,Neural_sparse_voxel_fields};
other relevant works extend it to support other neural rendering tasks like dynamic view synthesis ~\cite{li2020neural,pumarola2021d,sun2021HOI-FVV}, challenge scenes ~\cite{luo2021convolutional,yariv2021volume}, pose estimation ~\cite{meng2021gnerf}, real-time rendering ~\cite{yu2021plenoctrees}, relighting~\cite{rebain2020derf,bi2020neural,chen2020neural}, and editing ~\cite{xiang2021neutex,sofgan}.
We refer the readers to \cite{tewari2020state} for a comprehensive review of neural rendering.
However, most prior works still follow the original NeRF and require an expensive per-scene optimization process.
We instead leverage deep MVS techniques to achieve across-scene neural radiance field estimation for view synthesis using only few images as input.
Our approach utilizes a plane swept 3D cost volume for geometric-aware scene understanding, achieving significantly better performance than concurrent works \cite{yu2020pixelnerf,ibrnet} that only consider 2D image features for the generalization of radiance field reconstruction.
|
\section{Introduction}\label{sec:introduction}
Human image synthesis has seen fast progress from the realistic image-to-image translation of human characters, as milestones are being passed in a wide range of applications, including face synthesis, style transferring, and motion retargeting. Despite remarkable advances in synthesizing virtual moving actors with image-to-image translation approaches, creating plausible virtual actors from images of real actors still remains a key challenge. These approaches face the unique cognitive capability of humans for perceiving attributes of objects from images. Humans start learning early in their lives to recognize human forms and make sense of what emotions and meanings are being communicated by human movement. We are, by nature, specialists in the human form and movement analysis. Even small imperfections when synthesizing virtual actors might create a false appearance, especially body motion shaking effects when dealing with moving actors.
Many works in the Computer Vision and Computer Graphics communities have made great strides in capturing human geometry and motion through model-based and learning techniques. In particular, end-to-end learning approaches such as~\cite{2018-TOG-SFV},~\cite{kanazawaHMR18}, and~\cite{kolotouros2019spin} have achieved state of the art in capturing three-dimensional motion, shape, and appearance from videos and still images from real actors. Most recently, several methods have been proposed on body reenactment from source images~\citep{Lassner_GeneratingPeople,ZhaoWCLF17,ma2017pose,chan2018dance,Esser_2018_CVPR,lwb2019}. The ultimate goal of these methods is to create a video where the body of a target person is reenacted according to the motion extracted from the monocular video. The motion is estimated considering the set of poses of a source person. Despite the impressive results for several input conditions, there are instances where most of these methods perform poorly. For instance, the works of~\cite{chan2018dance} and~\cite{wang2018vid2vid}, only perform good reenacting of the appearance/style from one actor to another if a strict setup has complied, \textit{e}.\textit{g}., static backgrounds, a large set of motion data of the target person to train, and actors in the same distance from the camera~\citep{tewari2020cgf}. Furthermore, it is hard to gauge progress from these results in the field of retargeting, as most works only report the performance of their algorithms in their own set of images which, in general, is built considering the specificities in the training regime of their approaches.
In this paper, we tackle the problem of not only body reenactment but transferring a target actor from a video sequence to a new video. In the synthesized new video, the target must perform a similar motion to a source actor considering his/her physical motion constraints and the video's background where the source actor appears, \textit{i}.\textit{e}., the target actor is placed into a new context performing the source actor's movements.
To face this problem, we propose a shape-aware human retargeting approach and a new dataset with many annotations that provide a practical benchmark to assess the progress of both retargeting and body reenactment methods.
Our dataset comprises different videos with annotated human-to-object interactions and paired motions to evaluate the task of synthesizing videos of people. This dataset can be used as a common base to improve tracking the progress of the field by showing where current approaches might fail and measuring the quality of future proposals. All data were collected in order to test methods of retargeting motion in some common real conditions, more general, challenging scenarios, and not designed according to a specific training regime.
Our shape-aware human retargeting approach from monocular videos of people has several advantages. It might be applied in general scenes and conditions, it is compatible with existing rendering pipelines, and it is not actor specific, conversely to most end-to-end learning techniques. The retargeting method has four main components: motion estimation of the source actor, extraction of shape and appearance of the target actor, motion retargeting with spatio-temporal constraints, and image-based rendering and composition of the video. By imposing spatial and temporal constraints on the characters' joints, our method preserves features of the motion, such as feet touching the floor and hands touching a particular object in the transferring.
We performed experiments using our new dataset containing several types of motions and actors with different body shapes and heights. Our results show that an image-based rendering technique can still exhibit a competitive quality compared to the recent deep learning techniques in generic transferring tests. Our approach achieved better results compared with end-to-end learning methodologies such as the works of~\cite{wang2018vid2vid} and~\cite{Esser_2018_CVPR} in most scenarios for appearance metrics as structural similarity (SSIM), learned perceptual similarity (LPIPS), mean squared error (MSE), and Fr\'echet Video Distance (FVD). The experimental results extend the observation of~\cite{wang2020cvpr}, where they indicate that CNN-generated images are still surprisingly easy to spot. Our results suggest that retargeting strategies based on image-to-image learning still perform poorly to several motion types and transferring contexts.
Although remarkable advances have been made in neural human rendering approaches \citep{wang2018vid2vid,chan2018dance,Aberman_2018,sun2020human}, simultaneously transferring the appearance and motion of virtual characters is still a challenging problem. Most existing algorithms are restricted to perform self-transfer (or reenactment), suffering from poor generalization to changes in scale, camera viewpoint, and intrinsic parameters. Moreover, the lack of suitable datasets for performing and evaluating retargeting approaches on a common ground hampers to track the progress of the field. These limitations are the central motivation of our proposed approach. Therefore, the contribution of this paper is three-fold:
\begin{enumerate}
\item We present a novel video retargeting technique for human motion and appearance transferring. This technique is designed to leverage invariant information along the video transferring, such as the body shape of the person, while also taking into account physical constraints of the motion in 3D and the image domain;
\item We present a new benchmark dataset composed of $88$ videos, being $32$ paired videos with annotated human-to-object interactions and paired actor motions, eight videos where each actor rotates vertically while holding an A-pose, eight four-minute videos with the subject performing random moves, and eight $15$-seconds videos with actors dancing in pairs. The set participating actors comprises eight human subjects, carefully chosen with varying body shapes, clothing styles, heights, and gender. The designed dataset and evaluation protocol are suitable for both self-transfer and cross-transfer scenarios;
\item The dataset and presented results also propose to open discussions on common evaluation metrics for video retargeting schemes. In this context, our results suggest an image-based rendering technique can still exhibit a competitive quality when compared against recent end-to-end deep learning techniques in challenging and more general video-to-video human retargeting contexts.
\end{enumerate}
\section{Related Work}\label{sec:rel}
\noindent\textbf{Human appearance transfer and image-to-image translation.} The past five years have witnessed the explosion of generative adversarial networks (GANs) to new view synthesis. The synthesis is formulated as being an end-to-end learning problem~\citep{CNN_for_view_synthesis_0,CNN_for_view_synthesis_1,CNN_for_view_synthesis_2,Balakrishnan,Esser_2018_CVPR}, where a distribution is estimated to sample new views. A representative approach is the work of~\cite{ma2017pose}, where the authors proposed to transfer the appearance of a person to a given body pose. Similarly, \cite{Lassner_GeneratingPeople} proposed ClothNet to generate images of people with similar poses and shapes in different clothing styles. \cite{Esser_2018_CVPR} used a conditional U-Net to synthesize new images based on estimated edges and body joint locations.
Recent works such as~\cite{Aberman_2018} and~\cite{chan2018dance} start applying adversarial training to map 2D poses to the appearance of a target subject. Although these works employ a scale-and-translate step to handle the difference in the limb proportions between the source skeleton and the target, their synthesized views still have clear gaps in the test time compared with the training time. \cite{wang2018vid2vid} proposed a general video-to-video synthesis framework based on conditional GANs to generate high-resolution and temporally consistent videos of people. \cite{Shysheya_2019_CVPR} attempt to handle the poor generalization by training a model using different actors' point-of-views. Their approach is also actor-specific and requires training a model for each new character, including the full acquisition setup information and camera poses. \cite{mir20} proposed to leverage the information from DensePose to learn a model to perform texture transfer of garments. Although their method is texture agnostic and not actor specific, it is designed to deal with garments transference (shirts and pants) and does not address the problem of full-body transference neither handle the cross-transference. It also disregards the motion constraints and human-to-object interactions~\citep{Hassan2019}. In the same line,~\cite{NeverovaGK18} investigated a combination of surface-based pose estimation and deep generative models; however, their method only considers the layout locations and ignores the personalized shape and limb (joint) rotations.
Despite the impressive results for several inputs, end-to-end learning-based techniques still fail to synthesize the human body's details, such as face and hands. Furthermore, it is worth noting that these techniques focus on transferring style, which leads to undesired distortions when the characters have different morphologies (proportions or body parts' lengths). An alluring example is depicted in Figure~\ref{fig:ret_vid2vid}, where we perform the transfer between actors with differences in body shape (first row) and height (second row).
\begin{figure}[t!]
\centering
\includegraphics[width=1\linewidth]{shape_transfer_v2}
\caption{\small {\bf Motion and appearance transfer in different morphologies}. From left to right: target person, source motion video with a human of different body shape, vid2vid~\citep{wang2018vid2vid}, and our retargeting results. Note that vid2vid stretched, squeezed, and shranked the body forms whenever the transferring characters have different morphologies.}
\label{fig:ret_vid2vid}
\end{figure}
Another limitation of recent approaches such as~\cite{Aberman_2018},~\cite{chan2018dance}, and~\cite{wang2018vid2vid} is that they are dataset specific, \textit{i}.\textit{e}., they require training a different GAN for each video of the target person with different motions to perform the transferring. This training is computationally intensive and takes several days on a single GPU. In order to overcome these limitations,~\cite{lwb2019} proposed a 3D body mesh recovery module to disentangle the pose and shape; however, their performance significantly decreases when the source image comes from a different domain from their dataset, indicating that they are also affected by poor generalization to camera viewing changes.
Recently,~\cite{sun2020human} also proposed a dataset-specific method where projections of the reconstructed 3D human model are used to condition the GAN training, in order to maintain the structural integrity of the transfer to different poses. Nevertheless, all analyses were made in a strict setup where the person is standing parallel to the image plane, and the considered motions have reduced lateral translations.
In this paper, we point out that there is still a performance gap of recent end-to-end deep learning techniques against an image-based model when this comprises carefully designed steps for human shape and pose estimation and retargeting. These results extend the observation of the works of~\cite{bau2019seeing} and~\cite{wang2020cvpr}, where the authors observed that GANs still present limited generation capacity. While~\citeauthor{bau2019seeing}~showed that generative network models could ignore classes that are too hard at the same time producing outputs of high average visual quality,~\citeauthor{wang2020cvpr}~demonstrated that CNN-generated images are yet surprisingly easy to spot.
\vspace*{0.3cm}
\noindent\textbf{3D human shape and pose estimation.} Significant advances have been recently developed to estimate both the human skeleton and 3D body shape from images. \cite{Sigal} estimate shape by fitting a generative model, the SCAPE~\citep{Anguelov_2005}, to image silhouettes.~\cite{Bogo_2016} proposed the SMPLify method, which is a fully automated approach for estimating 3D body shape and pose from 2D joints in images. SMPLify uses a CNN to estimate 2D joint locations and then it fits an SMPL body model~\citep{Loper_2015} to these joints. \cite{Lassner_2017} used the curated results from SMPLify to train $91$ keypoint detectors. Similarly,~\cite{kanazawaHMR18} used unpaired 2D keypoint annotations and 3D scans to train an end-to-end network to infer the 3D mesh parameters and the camera pose.~\cite{kolotouros2019spin} combined an optimization method and a deep network to design a method less sensitive to the optimization initialization. Even though their method outperformed the works of \citeauthor{Bogo_2016}, \citeauthor{Lassner_2017}, and \citeauthor{kanazawaHMR18} regarding 3D joint error and runtime, their bounding box cropping strategy does not allow motion reconstruction from poses, since it frees three-dimensional pose regression from having to localize the person with scale and translation in image space. Moreover, they lack global information and temporal consistency in shape, pose, and human-to-object interactions, which are required in video retargeting with consistent motion transferring.
\vspace*{0.3cm}
\noindent\textbf{Retargeting motion.} \cite{Gleicher}'s seminal work of retargeting motion addressed the problem of transferring motion from one virtual actor to another with different morphologies. \cite{motion_retargetting_1} pushed further \citeauthor{Gleicher}'s method by presenting a real-time motion retargeting approach based on inverse rate control. Both~\citeauthor{Gleicher}'s and~\citeauthor{motion_retargetting_1}'s approaches require an iterative optimization with hand-designed activation constraints trajectories over time for several particular motions (like jumping or walking). Moreover, these constraints are sometimes hard to be designed if the 3D information of the environment is not explicitly available, such as in monocular videos.
\cite{Villegas_2018_CVPR} proposed a kinematic neural network with an adversarial cycle consistency to remove the manual step of detecting the motion constraints. In the same direction, the recent work of~\cite{2018-TOG-SFV} takes a step towards automatically transferring motion between humans and virtual humanoids. Similarly,~\cite{retargeting_2d} proposed a 2D motion retargeting using a high-level latent motion representation. Their method has the benefit of not explicitly reconstructing 3D poses and camera parameters, but it fails to transfer motions if the character walks towards the camera or with variations of the camera's point-of-view. Differently from~\citeauthor{Gleicher},~\citeauthor{motion_retargetting_1},~\citeauthor{Villegas_2018_CVPR},~and \citeauthor{2018-TOG-SFV}, we propose a retargeting formulation that does not assume that all the desired constraints should be inferred only from 3D end-effectors pose, but also from the 2D locations in the image where the body-environment interactions happen (please see Figure~\ref{fig:constraits_2D}). Furthermore, our spatio-temporal modeling allows a competitive runtime during the retargeting, conversely to \citeauthor{Gleicher}'s work whose optimization can last several hours, even for a small batch of images.
Thus, inspired by the promising ideas employed to adapt motion from one character to another, 3D shape and pose estimation, our proposed methodology brings forth the advantages of jointly modeling 3D motion, shape, and appearance information. However, differently from the works mentioned above, our approach does not require training a model for each actor, keeps visual details from the target actor, does not require strict setups, and preserves the features of the transferred motion as well as the human-object interactions in the video.
\vspace*{0.3cm}
\noindent\textbf{Existing datasets.}
Datasets are one cornerstone of recent advances in Computer Vision. While there are many large datasets available for human shape and pose estimation~\citep{MPII_dataset,COCO_dataset,AMASS:ICCV:2019}, existing motion retargeting datasets are yet rare, hampering progress on this area.~\cite{Villegas_2018_CVPR} provided human joint poses from synthetic paired motions, however, paired visual information is not available, limiting the appearance transferring.~\cite{chan2018dance} made available videos with random actor movements that can be used to learn the appearance of the target actor. However, the provided data is limited to their setup requirements, and it does not allow the analysis and comparison with other methods.~\cite{lwb2019} presented a set of videos with random actions of target subjects, as well as videos of the subjects performing an A-pose movement. This set enables methods focusing on modeling 3D human body estimation or using few keyframes to be executed using their data. On the other hand, the lack of paired motions limits motion and appearance retargeting results in quantitative terms, where the source and target actors are different subjects. Conversely, our proposed dataset, in addition to videos with random actions of the target subjects and videos of the subjects performing an A-pose video, also provides several carefully paired reconstructed 3D motions and annotated human-to-object interactions in the image and 3D space.
\section{Retargeting Approach}
\begin{figure*}[t!]
\includegraphics[width=1.0\linewidth]{overview_v4.pdf}
\caption{\textbf{Overview of our retargeting approach}. Our method is composed of four main components: human motion estimation in the source video (first component); we retarget this motion into a different target character (second component), considering the motion constraints (third component), and by last, we synthesize the appearance of the target character into the source video.}
\label{fig:method}
\end{figure*}
This section presents our human transferring method considering the importance of human motion, shape, and appearance in the retargeting. Unlike most techniques that transfer either appearance~\citep{Esser_2018_CVPR,Aberman_2018,chan2018dance,wang2018vid2vid} or motion independently~\citep{Villegas_2018_CVPR,2018-TOG-SFV}, we present a new method that simultaneously considers body shape, motion retargeting constraints, and human-to-object interactions over time, while retaining visual appearance quality.
Our method comprises four main components. We first estimate the motion of the character performing actions in the source video, where essential aspects of plausible movements, such as a shared coordinate system for all image frames and temporal motion smoothness are ensured. Second, we extract the~\text{body shape and texture} of the target character in the second video. We also extract the texture, refine body geometry and estimate the visibility information of the body parts for transferring the appearance. Then, the~\textit{retargeting} component adapts the estimated movement to the body shape of the target character while considering temporal motion consistency and the physical human interactions (constraints) with the environment. Finally, the~\textit{image-based rendering and composition} component combines classical geometry rendering and image-based rendering to render the texture (appearance extracted from the target character) into the background of the source video. Figure~\ref{fig:method} shows a schematic representation of the method.
\vspace*{0.3cm}
\noindent\textbf{Human body and motion representation.}
We represent the human motion by a set of translations and rotations over time of the joints that specify a human skeleton. This skeleton is attached to the actor's body and is defined as a $24$ linked joints hierarchy. The $i$-th joint pose $\mbf P^i \in \mathbb{SE}(3)$ is given by recursively rotating the joints of the skeleton tree, starting from the root joint and ending in its leaf joints, \textit{i}.\textit{e}., the forward kinematics denoted as FK. To represent the 3D shape of the human body, we adopted the SMPL model parametrization~\citep{Loper_2015}, which is composed of a learned human shape distribution $\mathcal{M}$, 3D joint angles ($\boldsymbol{\theta} \in \mathbb{R}^{72}$ defining 3D rotations of the skeleton joint tree), and shape coefficients $\boldsymbol{\beta} \in \mathbb{R}^{10}$ that model the proportions and dimensions of the human body.
\subsection{Shape-Aware Human Motion Estimation}\label{method:motion_estimation}
We start estimating the actor's motion in the source video. Our method builds upon the SPIN, a learning-based human model estimation framework of \cite{kolotouros2019spin}, where the human pose and body shape are predicted in the coordinate system of the person’s bounding box computed by OpenPose~\citep{openpose1,openpose2,openpose3}. SPIN leverages the advantages of regression and optimization-based human pose estimation frameworks to provide an efficient and accurate SMPL human model estimate. This bounding box normalizes the person in size and position. As also observed by~\cite{Mehta_2017}, the bounding box normalization frees 3D pose estimation from the burden of computing the scale factor (between the body shape to the camera distance) and the location in the image. However, this normalization incurs in a loss of temporal pose consistency required in the motion transfer. The loss of consistency also often leads to wrong body shape estimates for each frame, which should be constant through the video.
In order to overcome these issues, we map the initial pose estimation using virtual camera coordinates, following the motion reconstruction strategy presented by~\cite{Gomes_2020_WACV}. The loss in motion reconstruction is composed of two terms. While the first term enforces the pose projections of the joints to remain in the same locations into the common reference coordinate system, the second term favors maintaining the joints' angles configuration reinforcing the character shape to have averaged shape coefficients ($\boldsymbol{\beta^s}$) in the entire video. The pose of the human model in each frame is then obtained with the forward kinematics (FK) from the obtained joints configuration $\boldsymbol{\theta}_k^s$ of our shape-ware motion reconstruction:
\begin{equation}
\left[\mbf P^0_{k} ~ \mbf P^1_{k} ~ \ldots~ \mbf P^{23}_{k}\right] = \mbox{FK}(\boldsymbol{\beta}^s,\boldsymbol{\theta}_k^s),
\end{equation}
\noindent where $\mbf P_k^{i} = [\mbox{FK}(\boldsymbol{\beta}^s,\boldsymbol{\theta}_k^s)]_i$ is the pose of the $i$-th joint at frame $k$.
The raw actor motion is defined as $\mbf M(\boldsymbol{\beta}^s,\boldsymbol{\theta}^s) = [\mbf P_1 ~ \mbf P_2 ~ ... ~ \mbf P_n] \in \mathbb{R}^{24\times 4 \times 4 \times n}$, where $\boldsymbol{\beta} = [\boldsymbol{\beta}^s_1, \boldsymbol{\beta}^s_2, \ldots,\boldsymbol{\beta}^s_n] \in \mathbb{R}^{10\times n}\mbox{ and } \boldsymbol{\theta} = [\boldsymbol{\theta}_1^s, \boldsymbol{\theta}_2^s, \ldots, \boldsymbol{\theta}_n^s] \in \mathbb{R}^{72\times n}$ are composed of the stacked $\boldsymbol{\beta}^s,\boldsymbol{\theta}_k^s$ over time.
\subsubsection{Motion Regularization}
Since we estimate the character poses frame-by-frame, the resulting motion might present shaking motion with high-frequency artifacts in some short sections of the video. To alleviate these effects, we perform a regularization to seek a new set of joint angles $\widehat{\boldsymbol{\theta}^s}$ that creates a smoother motion. After applying a cubic-spline interpolation~\citep{splines} over the joints' motion $\mbf{M}(\boldsymbol{\beta}^s,\boldsymbol{\theta}^s)$, we remove the outlier joints from the interpolated spline. The final motion estimate is obtained by minimizing the cost:
\begin{equation}\label{eq:inversek}
\min\Big(||\widehat{\boldsymbol{\theta}^s} - \Theta||_2 + \gamma||\mbox{FK}(\boldsymbol{\beta}^s,\widehat{\boldsymbol{\theta}^s}) - \mbf P_{sp}||_2 \Big),
\end{equation}
\noindent where $\Theta$ is the subset of inlier joints, $\mbox{FK}$ is the forward kinematics, $\boldsymbol{\beta}^s$ defines the proportions and dimensions of the human body in the source video, $\mbf P_{sp}$ is the spline interpolated joint positions, and $\gamma$ is the scaling factor between the original joint angles and the interpolated positions. This strategy removes high-frequency artifacts of the joints' motion while retaining the movement features.
\subsection{Target Character Processing}\label{method:shape_appearance}
Most methods based on Generative Adversarial Networks, such as~\cite{wang2018vid2vid},~\cite{chan2018dance},~\cite{Aberman_2018}, and~\cite{sun2020human}, have emerged as effective approaches for human appearance synthesis. However, these methods still suffer in creating fine texture details, notably in some body parts as the face and hands. Besides, it is well known that these methods suffer from quality instability when applied in contexts slightly different from the original ones, \textit{i}.\textit{e}., a small difference in camera position, uncommon motions, pose translation, {\it etc}. These limitations motivate the proposal of our semantic-guided appearance retargeting method, which is designed to leverage visibility map information and semantic body parts to refine the initial target mesh model while keeping finer texture details in the transferring.
Thus, in order to create a more stable method and overcome the lack of details, we design a new semantic-guided image-based rendering approach that copies local patterns from input images to the correct position in the generated images. Our idea stems from using semantic information of the body (\textit{e}.\textit{g}., face, arms, torso locations, {\it etc.}) in the geometric rendering to encode patch positions and image-based rendering to copy pixels from the target images, and therefore maintaining texture details. This strategy estimates a generic target body model $\boldsymbol{\beta}^t$, comprising body geometry, texture, and visibility information at each frame that will be transferred to the source video in the retargeting step.
\subsubsection{Semantic-Guided Human Model Extraction}\label{method:geometry}
When extracting the appearance and geometry of the human body, the self-occlusion of body parts and the deformable nature of the human body (and of clothes) bring challenging conditions. In order to tackle these difficulties, we propose a semantic-guided image-based rendering of body parts that explores the global and local information of the human body into the body model estimation.
While we gathered the global geometric information from the pose and shape as discussed in Section~\ref{method:motion_estimation}, the local geometric information is extracted for each viewpoint and aligned with the contours of their semantic body parts in the image. To perform this alignment, we partitioned and computed the correspondence of the 3D body model into fourteen meaningful body part labels (face, left arm, {\it etc.}). The 2D semantic body labels are computed using the~\cite{Gong2018InstancelevelHP}'s body parsing model, which we fine-tuned using the people-snapshot dataset~\citep{alldieck2018video}.
After computing the semantic map of the body, for each contour point (red squares in Figure~\ref{fig:deformation}) in the map, we define the vertex from the body mesh with the same semantic and the smallest Euclidean distance to the contour point as a control point (blue circles in Figure~\ref{fig:deformation}). The Euclidean distance is computed between the contour point and the 2D projection of the vertex. Each control point will receive a target position given its correspondent contour point. These control points and their new locations guide the deformation of the mesh to fit the shape into the semantic map's contour. In the deformation, we seek a new body mesh $Q$ that is a locally rigid transformation of the source body mesh $P$, following the control points given by the semantic contours. The mesh deformation is solved efficiently with the local rigid deformation As-Rigid-As-Possible (ARAP)~\citep{Zohar2015}. The correspondence between each contour point and control point is represented in Figure~\ref{fig:deformation} with colored small lines. Notice that the desired motion of the control points guides the deformation to fit the body mesh into the contours of the semantic map.
\begin{figure}[t!]
\includegraphics[width=1\linewidth]{semantic_guidance_deform.pdf}
\caption{\textbf{Semantic guided deformation}. The contour points (red squares) in the semantic map indicate the target localization of the control points (blue circles) in the body mesh that it will guide the ARAP algorithm to fit the body mesh into the contours of the semantic map. The correspondence between each contour point and control point is illustrated with colored small lines.}
\label{fig:deformation}
\end{figure}
\subsubsection{Human Textures and Visibility Maps Extraction}\label{method:texture}
The geometric information allows rendering the human target character in a new viewpoint by applying the desired transformations and re-projecting them onto the image plane. In order to compare and merge the information from the human actor from different viewpoints, we map the views to a common UV texture map space. The mapping function is given by a parametric function ${\mathcal{S}}$ that maps a point in the mesh from frame $k$ to a point in the texture space with coordinates $(u, v)$, $\mathcal{S}:\mathbb{R}^3 \rightarrow \mathbb{R}^2$. Then, the accumulated texture map $UV(u,v)$ for all available images of the target character is done by the rendering with ${\mathcal{S}}$:
\begin{equation}\label{eq:texture_projection}
UV({\mathcal{S}}(x,y,z)) = \mathcal{I}({\Pi}((x,y,z),\mbf K)),
\end{equation}
\noindent where the ${\Pi}(.,\mbf K)$ operator performs the rendering taking a 3D mesh point $(x,y,z)$ and projecting it into the image plane given the camera parameters $\mbf K$, and $\mathcal{I}$ is the texture information in the image coordinates.
Finally, we assert which mesh points are visible by exploring the inverse map ${\mathcal{S}}^{-1}(u,v)$, as illustrated in Figure~\ref{fig:texture-vis-map}. Each visibility map indicates which parts of the body model are visible per frame. Then we select the closest viewpoint to the desired new viewpoint, for each part of the body model from the visibility maps.
\begin{figure}[t]
\includegraphics[width=1\linewidth]{target_processing.pdf}
\caption{\textbf{Rendering of the visibility maps and texture images.} {\it Top:} We project each target actor viewpoint in a common UV texture space using the estimated geometry and create a binary map of visibility body parts. {\it Bottom:} Given the goal pose (retargeted pose), we estimate its visibility body parts map, and then select the better matching visibility body parts created from the viewpoints from the target actor.}
\label{fig:texture-vis-map}
\end{figure}
\subsection{Motion Retargeting using Hybrid 2D/3D Spatio-Temporal Constraints}
After estimating the motion from the input video, \textit{i}.\textit{e}., $\mbf M$($\boldsymbol{\beta}^s,\boldsymbol{\theta}^s$), and 3D model $\boldsymbol{\beta}^t$ of the target human, we can proceed to the motion retargeting step. We assume that the target character has a homeomorphic skeleton structure to the source character, \textit{i}.\textit{e}., the geometric differences are in terms of bone lengths and body proportions. Our retargeting motion estimation loss is designed to guarantee the motion similarity and physical human-object interaction constraints over time. Similar to~\cite{Gleicher}, our first goal is to retain the joint configuration of the target as close as possible to the source joint configurations at instant $k$, $\boldsymbol{\theta}_k^t \approx \boldsymbol{\theta}_k^s$, \textit{i}.\textit{e}., to keep $\mbf e_k$ small such as: $\boldsymbol{\theta}_k^t = \boldsymbol{\theta}_k^s + \mbf e_k.$
We also aim to keep similar movement style and speed in the retargeted motion. Thus, we propose a one step speed prediction in 3D space defined as $\Delta \mbf M(\boldsymbol{\beta},\boldsymbol{\theta}_k) = \mbox{FK}(\boldsymbol{\beta},\boldsymbol{\theta}_{k+1}) - \mbox{FK}(\boldsymbol{\beta},\boldsymbol{\theta}_k)$ to maintain the motion style from the original joints' motion:
\begin{equation}\label{eq:pred}
\mathcal{L}_P(\mbf e) = \sum_{k=i+1}^{i+n}||\Delta \mbf M(\boldsymbol{\beta}^t, \boldsymbol{\theta}_{k}^s + \mbf e_k) - \Delta \mbf M(\boldsymbol{\beta}^s, \boldsymbol{\theta}_{k}^s) ||_1,
\end{equation}
\noindent where $\mbf e = [\mbf e_{i+1}, \ldots, \mbf e_{i+n}]^T$, and $n$ is the number of frames considered in the retargeting.
Rather than considering a loss for the total number of frames, we use only the frames belonging to a neighboring temporal window of $n$ frames equivalent to two seconds of video. This neighboring temporal window scheme allows us to track the local temporal motion style producing a motion that tends to be natural compared with a realistic-looking of the estimated source motion. The retargeting considering a local neighboring window of frames also results in a more efficient optimization.
\vspace*{0.3cm}
\noindent\textbf{2D/3D human-to-object interactions.}
The human-to-object interactions (\textit{i}.\textit{e}., motion constraints) are important to identify key features of the original motion that must be preserved in the retargeted motion. The specification of these interactions typically involves only a small amount of work in comparison with the task of creating new motions. Typical interactions are, for instance, that the target character's feet should be on the floor; holding hands while dancing or while grabbing/manipulating an object in the source video. Some examples of human-to-object motion constraints are shown in Figure~\ref{fig:constraits_2D}, where the actor is interacting with a box.
Going one step further than classic retargeting constraints defined in~\cite{Gleicher} and \cite{motion_retargetting_1}, where end-effectors must be at solely a desired 3D position at a given moment, we propose an extended hybrid constraint in the image domain, \textit{i}.\textit{e}., the joint of the character must also be projected at a specific location in the image. This type of motion constraint allows the user to exploit the visual knowledge of interactions of the actor in the scene. Some examples are shown in Figure~\ref{fig:constraits_2D}, where two types of constraints are defined: 3D interactions (blue dots) impose the feet and right hand to be in the same location after the retargeting, and 2D constraints (red dots) imposing the correct position to the left hand in the image.
\begin{figure}[t!]
\includegraphics[width=1\linewidth]{constraints.pdf}
\caption{\textbf{Example of hybrid 2D/3D constraints from human-to-object interactions.} {\it Top row:} Original video with 3D constraints (blue dots) and 2D constraints (red dots). {\it Middle row:} Motion retargeting to a new character using only 3D constraints. {\it Bottom row:} The results for the new character applying hybrid 2D/3D constraints using our retargeting approach. Please observe that the hands' positions are more consistent when adopting our hybrid strategy.}
\label{fig:constraits_2D}
\end{figure}
Our retargeting is capable of adapting to such situations by defining the motion retargeting constraints losses in respect to end-effectors' (hands, feet) 3D poses $\mbf P_{R3D}$ and 2D poses $\mbf P_{R2D}$ as:
\begin{align}
&\mathcal{L}_{R3D}(\mbf e_k) = ||\mbox{FK}(\boldsymbol{\beta}^t,\boldsymbol{\theta}_{k}^s + \mbf e_k) - \mbf P_{R3D}||_1,\label{eq:constr} \\
&\mathcal{L}_{R2D}(\mbf e_k) = ||\Pi(\mbox{FK}(\boldsymbol{\beta}^t,\boldsymbol{\theta}_{k}^s + \mbf e_k), \mbf K) - \mbf P_{R2D}||_1. \label{eq:constr_2d}
\end{align}
\vspace*{0.3cm}
\noindent\textbf{Space-time loss optimization.} The final motion retargeting loss $\mathcal{L}$ combines the source motion appearance with the different shape and constraints of the target character from Equations~\ref{eq:pred}, \ref{eq:constr}, and \ref{eq:constr_2d}:
\begin{equation}
\mathcal{L} = \mbf ||\mbf W_1\mbf e||_2 + \lambda_1\mathcal{L}_P(\mbf e) + \lambda_2\mathcal{L}_{R3D}(\mbf e) + \lambda_3 \mathcal{L}_{R2D}(\mbf e),
\end{equation}
\noindent where the joint parameters to be optimized are $\mbf e = [\mbf e_{i+1}, \ldots, \mbf e_{i+n}]^T$, $n$ is the number of frames considered in the retargeting window, $\lambda_1$, $\lambda_2$, and $\lambda_3$ are the contributions for the different error terms, and $\mbf W_1$ is a positive diagonal matrix of weights for the motion appearance for each body joint. This weight matrix is set to penalize more errors in joints that are closer to the root joint.
One representative example of the retargeting strategy considering these hybrid 2D/3D constraints is shown in Figure~\ref{fig:constraits_2D}. In this video sequence, the target actor is bigger and taller than the one in the source video (shown in the two first rows of the figure). Notice that the retargeting of the target actor (shown in the third row) results in more bent poses to maintain the human-to-object interactions, and thus the hands' positions are consistent when adopting our strategy.
\subsection{Model Rendering and Image Compositing}\label{method:render}
In our framework's last step, we combine the rendered target character and the source background to make a convincing final image. We first segment the source image into a background layer using, as a mask, the projection of the actor body model with a dilation, and then the background is inpainted.
Our experiments showed that different inpainting techniques, including end-to-end learning-based methods such as~\cite{yu2018free},~\cite{yu2018generative},~\cite{wang2018videoinp}, and~\cite{Xu_2019_CVPR}, displayed small differences in the quality of results, probably because most part of the filled region will be covered by the target person. Thus, we apply the method proposed by~\cite{based-inpainting} that presented the best overall results, especially when the background texture is homogeneous. To ensure temporal smoothness in the inpainting background, we compute the pixel color value as the median value between the neighboring $n$ frames.
In the sequence, we explored the visibility map of the retargeted body model (global geometric information discussed in Section~\ref{method:texture}) to select the human body parts that better matches the target parts. Since the transformation between the retargeted SMPL model and the estimated SMPL model to the images is known, we apply the same transformation used in the local geometric information to move them to the correct positions. Instead of directly applying 3D warping in the selected images, we use our pre-warping step (texture map) to improve the rendering speed of 3D warping. In order to fill the remaining part holes in the warped image, we explored the median of the accumulated texture $UV$ (see Figure~\ref{fig:texture-vis-map}). Finally, the background and the target character are combined in the retargeted frame.
\section{Human Retargeting Dataset and Evaluation}
Due to the lack of suitable benchmark datasets, recent works on neural view synthesis and body reenacting~\citep{chan2018dance,Esser_2018_CVPR,wang2018vid2vid} only analyze their results in qualitative terms or quantitative terms to self-transfer in which the source and target actors are the same subjects. The provided data is adapted to the requirements of their method setup, and it is hard to perform a comparison with other methods on this data. Cross-transfer is far more complicated than self-subject transfer. First, self-transfer is not affected by body appearance changes (from body shape, clothing). Second, the self-shape transfer does not account for human-to-object interactions or disregards the influence of existing human-environment physical interactions. Moreover, as previously discussed in Section~\ref{sec:rel}, existing video retargeting datasets are still rare.
\subsection{Human Retargeting Dataset}
\begin{figure*}[t!]
\centering
\includegraphics[width=1.0\linewidth]{dataset_v5.pdf}
\caption{\textbf{Human retargeting dataset.} {\it a)} The subjects participating in our dataset, their respective height and estimated SMPL body models. {\it b)} Overview of all motions available in our proposed dataset. {\it c)} Paired motions (upper and lower rows) with annotated motion constraints (3D constraints in blue and 2D constraints in red). {\it d)} The reconstructed 3D motions.}
\label{fig:dataset}
\end{figure*}
To evaluate the retargeting and appearance transfer with different actor motions, consistent reconstructed 3D motions, and with human-to-object interactions, we created a new dataset with {\it paired motion} sequences from different characters and {\it annotated motion retargeting constraints}. For each video sequence, we provide a refined 3D actor reconstructed motion and the actor body shape estimated with~\cite{alldieck2018video}. The refined reconstructed 3D motions and 2D-3D annotation of interactions were collected by manual annotation. The provided motions are not prone to typical motion artifacts such as bended knees, body shape variations, and camera-to-actor translation changes.
Our carefully designed dataset comprises $8$ subjects. To keep the dataset with diversity, we choose participants (subjects S0 to S7) with different gender, sizes, clothing styles, and body shapes. Figure~\ref{fig:dataset} shows the subjects, their respective body models, and labels. We also choose a set of movements that are representative of the problem of the retargeting, increasing the level of difficulty and with different motion constraints. These movements are: ``pick up a box'', ``spinning'', ``jump'', ``walk'', ``shake hands'', ``touch a cone'', ``pull down'', and ``fusion dance''.
Each actor performed all eight actions. We paired two actors to perform the same motion sequence, where the subjects were instructed to follow marks on the floor to perform the same action, resulting in four paired videos per action (a total of $32$ paired videos). Then, we define the combination of actors aiming at the most challenging configuration for the task of human retargeting motion. For instance, actors $S0$, $S2$, $S4$, and $S6$ were paired respectively with $S1$, $S3$, $S5$, and $S7$. We also provide, for each subject, three videos: one video where the subject is rotating and holding an A-pose, a four-minute video where the subject is performing different poses, and a $15$-second video where the subject is dancing. All videos were recorded with $1{,}000 \times 1{,}080$ of resolution at $30FPS$. This information allows training most existing approaches for evaluation. Figure~\ref{fig:dataset} shows some examples of frames from our dataset.
Additionally, we provide three publicly available videos acquired under uncontrolled conditions: {\it joao-pedro}, {\it tom-cruise}, and {\it bruno-mars} sequences. The {\it bruno-mars} sequence was also adopted in the experiments of~\cite{chan2018dance}. The provided videos present more general conditions to analyze the effectiveness of reenacting and retargeting methods in the wild with more complex motions, background, and varying illumination.
\subsection{Protocol}
The evaluation often employed by works on character reenacting/synthesis~\citep{Aberman_2018,chan2018dance,lwb2019,sun2020human} consists in setting the source character equal to the target character. However, we argue that this protocol is used because of the absence of paired motion sequences, as also noted in~\cite{lwb2019} and~\cite{sun2020human}. While this protocol might be appropriate to assess new synthesized view/pose in the same scene background, we state that it is not appropriate for evaluating the synthesis of new videos of people when taking into account motion constraints (\textit{e}.\textit{g}., human-to-object interactions), distinct shapes and heights, and transferring them in different backgrounds where they were initially recorded.
Therefore, we run our evaluation protocol as follows: when evaluating a new synthesized video, we place the target actor performing a similar motion to the source actor (all physical constraints and the human-to-object interactions are taken into account). Then, we move the target actor to a different scenario configuration to perform the retargeting. If the retargeting is successfully executed, the method will place the target actor moving as the source actor into the source actor's scenario.
\subsection{Evaluation Metrics}
We measure the quality of the synthesized frames in terms of the following metrics: i) the structural similarity (SSIM)~\citep{Wang04imagequality} that compares local patterns of pixel intensities normalized for luminance and contrast. SSIM assumes that human visual perception is highly adapted for extracting structural information from objects; ii) learned perceptual similarity (LPIPS)~\citep{zhang2018perceptual}, which provides a deep neural learned similarity distance closer to human visual perception; iii) the mean squared error (MSE) that is computed by averaging the squared intensity differences between the pixels; and iv) Fr\'echet video distance (FVD)~\citep{unterthiner2019accurate}, which was designed to capture the quality of the synthesized frames and their temporal coherence in videos. These are widely used perceptual distances to measure how similar the synthesized images are in a way that coincides with human judgment.
To properly evaluate the quality of retargeting, it is required a paired video where the target character is the same as the source video sequence. Collecting real paired motion sequences from different characters is a challenging task, even when the movement of the actors is carefully predefined and synchronized. For instance, each actor has a movement style that can result in videos with unsynchronized actions (as seen in the third column in Figure~\ref{fig:dataset}-c). Thus, to make possible the computation of quantitative metrics, we relax the assumption that the frame $k$ in the synthesized video must be the frame $k$ in the paired video by applying a small window of acceptance around $k$, \textit{i}.\textit{e}., we evaluate the quality of one synthesized frame as a better answer between the frame and a small window of frames in the paired video. The window of acceptance $w$ is estimated for each pair of videos according to:
\begin{equation}\label{eq:window_size}
w(V_1, V_2) = \max(15, 2\times( \abs{len(V_1) - len(V_2)} ),
\end{equation}
\noindent where $len(V)$ is the number of frames in the video $V$. Equation~\ref{eq:window_size} captures how much two videos are not synchronized allowing the synthesized frames to match with the paired video. The lower bound value of $15$ frames was empirically selected by a visual analysis of our videos.
Aside from the image quality metrics, we also propose to evaluate the approaches with fake image detectors. Since our dataset provides paired motion sequences, we can evaluate the retargeted frames' quality and realism with an image forgery detector. We applied the detector in the generated and real frames from the sequence where the same subject performs the retargeted motion. The retargeted frames were generated using motion extracted from source videos whose subject performing the motion is different from the target subject.
We adopted the image forgery detection algorithm presented by~\cite{marra2019e2e} to evaluate all methods in our experiments. The~\citeauthor{marra2019e2e}'s method evaluates images by detecting pixel artifacts, which is a common metric for detecting CNN-generated images. We remark that we tested different image forgery detection algorithms such as the method proposed by~\cite{Wang2019CNNgeneratedIA}, but the results were inconclusive, which might be because of the high resolution of our images. Furthermore, our images are not only generated by CNN methods, and as the authors stated, their method performs at chance in visual fakes produced by image-based rendering or classical commercial rendering engines. Finally, we define the forgery performance as the difference between the probability of the paired real image and its respective synthetic frame of being classified as fake. This difference indicates how far a synthetic frame is from being recognized as fake with a similar result of a real image.
\section{Experiments and Results}
\noindent\textbf{Parameter settings.} We selected a fixed set of parameters for all experiments. We applied a grid search over a portion of the $4$-minutes videos training examples of the dataset. The grid search $\gamma$ values ranged from $5$ to $100$ with a step of $5$, and $\lambda$ from $1$ to $10$, with a step of $1$. In the human motion estimation and retargeting steps, we used $\gamma = 10$, $\lambda_1 = 5$, $\lambda_2 = 1$, and $\lambda_3 = 1$. We minimize the retargeting loss function with Adam optimizer using $300$ iterations, learning rate $0.01$, $\beta_1=0.9$, and $\beta_2=0.99$. To deform the target body model with the semantic contour control points, we employed the default parameters proposed by \cite{Zohar2015}, and we fed the method with eight images taken from different viewpoints of the actor (see Figure~\ref{fig:method}). For a complete comparison, we also present qualitative results of the new frames from the state-of-art approaches by replacing the generated background regions from the methods by the source video background.
\begin{table}[t!]
\centering
\caption{{\bf Ablation study}. Comparison of mean MSE and FVD for different ablated versions of our method on all motion types of our dataset (best in bold).}
\label{table:ablation_results}
\resizebox{0.99\columnwidth}{!}{%
\begin{tabular}{@{}lrrrr@{}}
\toprule
{\bf Method} & \phantom{a} & {\centering MSE$\downarrow$}
& {\centering FVD$\downarrow$} \\ \midrule
No motion regularization & \phantom{a} &$275.47$ & $887.94$ \\
No semantic guidance & \phantom{a} &$275.88$ & $879.86$ \\
No 2D/3D constraints & \phantom{a} &$285.33$ & $859.67$ \\
No use of visibility mask & \phantom{a} &$273.45$ & $789.06$ \\
Full method & \phantom{a} &$\mathbf{259.75}$ & $\mathbf{738.96}$ \\
\bottomrule
\end{tabular}
}
\end{table}
\begin{table*}[t!]
\centering
\caption{{\bf Quantitative ablation analysis of the retargeting.} Table shows the error in pixels between the constraints position and the end-effectors location. This error indicates how far the end-effector is from the target position (better close to $0$).}
\label{table:ablation_retargeting}
\resizebox{0.99\linewidth}{!}{%
\begin{tabular}{@{}lccccccccccccccc@{}}
\toprule
\multirow{3}{*}{\bf Method} & \multicolumn{9}{c}{{\bf Motion type}} & \phantom{a} & \multicolumn{5}{c}{{\bf Pair of actors}}\\ \cmidrule{2-10} \cmidrule{12-16}
\multicolumn{1}{r}{}
& {\centering jump}
& {\centering walk}
& {\centering spinning}
& {\centering shake hands}
& {\centering cone}
& {\centering fusion dance}
& {\centering pull down}
& {\centering box}
& {\centering Avg.}
& \phantom{a}
& {\centering S0-S1}
& {\centering S2-S3}
& {\centering S4-S5}
& {\centering S6-S7}
& {\centering Avg.} \\ \midrule
Direct Transfer & $13.73$ & $14.49$ & $21.90$ & $18.22$ & $16.77$ & $24.67$ & $34.11$ & $22.49$ & $20.80$ & \phantom{a}
& $27.41$ & $22.84$ & $10.06$ & $22.88$ & $20.80$ \\
Retargeting & $2.44$ & $1.94$ & $2.04$ & $2.90$ & $2.76$ & $10.66$ & $3.14$ & $5.35$ & $\mathbf{3.90}$ & \phantom{a}
& $4.57$ & $3.19$ & $2.67$ & $5.19$ & $\mathbf{3.90}$ \\
\bottomrule \\
\end{tabular}
}
\end{table*}
\begin{figure*}[t!]
\centering
\includegraphics[width=0.8\linewidth]{retargeting_cone.pdf}
\caption{\textbf{Qualitative ablation analysis of the retargeting.} {\it Top row:} source video containing the actor S3. {\it Middle row:} transferring results to the actor S7 without the physical interactions. {\it Bottom row:} transferring results of our method with 2D-3D interactions.}
\label{result:ablation-cone}
\end{figure*}
\subsection{Ablation Study}
To verify that the space-time motion transfer optimization, motion regularization, semantic-guided deformation, and visibility maps contribute to our approach's success in producing more realistic frames, we conducted several ablation analysis using the motion sequences from our dataset.
Table~\ref{table:ablation_results} shows the results of our ablation study in terms of MSE and FVD of five ablated versions of our method. We draw the following observations. First, the best result is achieved when the full method is applied. Second, removing the 2D/3D constraints reduces the performance in terms of MSE. These constraints play a key role in the compliance of the poses to motion constraints. By removing them, large fragments from the background are computed as part of the retarget character body when computing MSE using the paired video, leading to the worst MSE value. Third, without the shape-aware regularization, which hinders the temporal coherence, the model presents the worst value of FVD. We can also see that after removing the semantic guidance, which decreases the quality of the texture applied onto the 3D model, the frames have more artifacts, and the model also performs poorly in terms of FVD.
The results of a more detailed performance assessment of the effects from motion constraints in the video retargeting are shown in Table~\ref{table:ablation_retargeting}. The error between the computed position of the end-effectors and the target positions (motion constraints from the human-to-object interactions) is significantly smaller for all motion sequences and pairs of actors when applying our retargeting strategy. Some frames are shown in Figure~\ref{result:ablation-cone} to illustrate the created visual artifacts when not considering the motion constraints. In this setup, the source actor (top row) is taller than the target (bottom row), and the target actor does not touch the floor nor touch with her hands the cone without the hybrid motion constraints (middle row). Conversely, these features are kept when considering the 2D/3D human-to-object losses in the retargeting. Another representative example of the retargeted motion trajectory over time, with one shorter actor interacting with a box, is shown in Figure~\ref{fig:grafico_retarget}. Please notice the smooth motion adaptation produced by the retargeting with the restrictions in frames $47$ and $138$ (green line) when the character's left hand is touching the box. Additionally, these results illustrate that our method is able to impose different space-time human-to-object interactions and motion constraints in the retargeting.
\begin{figure}[t]
\includegraphics[width=1\linewidth]{grafico_retargeting.pdf}
\caption{\textbf{Retargeted trajectory with motion constraints.} The curves show the left hand's trajectory on the y-axis when transferring the motion of {\it picking up a box} between two differently sized characters: original motion (blue line), a na\"ive transfer without constraints at the person's hand (red line), and with constraints (green line). Frames containing motion constraints are located between the red circles.}
\label{fig:grafico_retarget}
\end{figure}
Table~\ref{table:ablation_study} shows the forgery performance when synthesizing the frames after removing the visibility map extraction and the semantic-guided human model extraction. We can see that these two steps significantly enhance the quality of the results to a point where the detector returns for the ``shake hands" sequence a probability of $29.24$ higher when removing these two components.
\begin{table}[t!]
\centering
\caption{{\bf Visibility maps and semantic-guidance analysis}. Average forgery performance for each movement (better close to $0$).}
\label{table:ablation_study}
\scriptsize
\resizebox{\columnwidth}{!}{
\begin{tabular}{@{}lccr@{}}
\toprule
\multirow{2}[3]{*}{Motion} & \multicolumn{2}{c}{Method} \\
\cmidrule(l{2pt}r{2pt}){2-3} & \vtop{\hbox{\strut No Semantic-guidance}\hbox{\strut and Visibility Map}} & \vtop{\hbox{\strut Complete}\hbox{\strut Model}} \\ \midrule
\multicolumn{1}{l}{jump} & $46.97$ & $\mathbf{31.11}$\\
\multicolumn{1}{l}{walk} & $45.76$ & $\mathbf{23.35}$\\
\multicolumn{1}{l}{spinning} & $34.46$ & $\mathbf{21.09}$\\
\multicolumn{1}{l}{shake hands} & $34.91$ & $\mathbf{5.67}$\\
\multicolumn{1}{l}{cone} & $39.20$ & $\mathbf{24.80}$\\
\multicolumn{1}{l}{fusion dance} & $33.24$ & $\mathbf{15.97}$\\
\multicolumn{1}{l}{pull down} & $35.68$ & $\mathbf{17.85}$\\
\multicolumn{1}{l}{box} & $38.09$ & $\mathbf{17.63}$\\
\bottomrule
\end{tabular}
}
\end{table}
\subsection{Comparison with Previous Approaches}
\begin{table*}[t!]
\centering
\caption{{\bf Comparison with state of the art}. SSIM, LPIPS, MSE, and FVD comparison by motion types and pair of actors from our dataset (best in bold, second-best in italic).}
\label{table:metrics_result}
\resizebox{0.99\linewidth}{!}{%
\begin{tabular}{@{}clrrrrrrrrrrrrrrr@{}}
\toprule
\multirow{3}{*}{\bf Metric} & \multirow{3}{*}{\bf Method} & \multicolumn{9}{c}{{\bf Motion type}} & \phantom{a} & \multicolumn{5}{c}{{\bf Pair of actors}}\\ \cmidrule{3-11} \cmidrule{13-17}
& \multicolumn{1}{r}{}
& {\centering jump}
& {\centering walk}
& {\centering spinning}
& {\centering shake hands}
& {\centering cone}
& {\centering fusion dance}
& {\centering pull down}
& {\centering box}
& {\centering Avg.}
& \phantom{a}
& {\centering S0-S1}
& {\centering S2-S3}
& {\centering S4-S5}
& {\centering S6-S7}
& {\centering Avg.} \\ \midrule
\multirow{5}{*}{\rotatebox[origin=c]{90}{\parbox[c]{1.5cm}{\centering SSIM$\uparrow$}}}
& V-Unet & $0.870$ & $0.871$ & $0.843$ & $0.847$ & $0.862$ & $0.797$ & $0.847$ & $0.857$ & $0.849$ & \phantom{a}
& $0.855$ & $0.886$ & $0.830$ & $0.826$ & $0.849$ \\
& Vid2Vid & $0.880$ & $0.884$ & $0.856$ & $0.858$ & $0.878$ & $0.821$ & $0.859$ & $0.866$ & $\mathit{0.862}$ & \phantom{a}
& $0.868$ & $0.901$ & $0.848$ & $0.835$ & $0.862$ \\
& EBDN & $0.878$ & $0.880$ & $0.855$ & $0.859$ & $0.878$ & $0.820$ & $0.857$ & $0.858$ & $0.861$ & \phantom{a}
& $0.867$ & $0.898$ & $0.844$ & $0.834$ & $0.861$ \\
& iPER & $0.877$ & $0.880$ & $0.852$ & $0.859$ & $0.877$ & $0.816$ & $0.855$ & $0.856$ & $0.859$ & \phantom{a}
& $0.867$ & $0.896$ & $0.842$ & $0.831$ & $0.859$ \\
& Ours & $0.881$ & $0.885$ & $0.855$ & $0.860$ & $0.879$ & $0.820$ & $0.861$ & $0.869$ & $\mathbf{0.864}$ & \phantom{a}
& $0.872$ & $0.902$ & $0.846$ & $0.834$ & $\mathbf{0.864}$ \\
\midrule
\multirow{5}{*}{\rotatebox[origin=c]{90}{\parbox[c]{1.5cm}{\centering LPIPS$\downarrow$}}}
& V-Unet & $0.147$ & $0.132$ & $0.157$ & $0.161$ & $0.174$ & $0.243$ & $0.166$ & $0.158$ & $0.167$ & \phantom{a}
& $0.184$ & $0.160$ & $0.166$ & $0.158$ & $0.167$ \\
& Vid2Vid & $0.131$ & $0.105$ & $0.126$ & $0.136$ & $0.133$ & $0.203$ & $0.142$ & $0.133$ & $\mathit{0.138}$ & \phantom{a}
& $0.148$ & $0.131$ & $0.129$ & $0.147$ & $0.138$ \\
& EBDN & $0.141$ & $0.122$ & $0.139$ & $0.138$ & $0.143$ & $0.215$ & $0.151$ & $0.170$ & $0.153$ & \phantom{a}
& $0.159$ & $0.145$ & $0.147$ & $0.159$ & $0.153$ \\
& iPER & $0.151$ & $0.134$ & $0.151$ & $0.151$ & $0.155$ & $0.239$ & $0.168$ & $0.184$ & $0.167$ & \phantom{a}
& $0.161$ & $0.165$ & $0.170$ & $0.171$ & $0.167$ \\
& Ours & $0.125$ & $0.099$ & $0.130$ & $0.131$ & $0.128$ & $0.206$ & $0.131$ & $0.127$ & $\mathbf{0.135}$ & \phantom{a}
& $0.133$ & $0.129$ & $0.133$ & $0.143$ & $\mathbf{0.135}$ \\
\midrule
\multirow{5}{*}{\rotatebox[origin=c]{90}{\parbox[c]{1.5cm}{\centering MSE$\downarrow$}}}
& V-Unet & $295.04$ & $269.59$ & $354.33$ & $377.02$ & $328.68$ & $559.75$ & $417.00$ & $346.13$ & $368.44$ & \phantom{a}
& $381.77$ & $344.71$ & $362.63$ & $384.66$ & $368.44$ \\
& Vid2Vid & $257.33$ & $206.42$ & $286.18$ & $332.09$ & $253.04$ & $452.77$ & $349.28$ & $288.54$ & $\mathit{303.32}$ & \phantom{a}
& $312.40$ & $274.80$ & $269.81$ & $356.26$ & $\mathit{303.32}$ \\
& EBDN & $306.92$ & $269.58$ & $312.69$ & $312.12$ & $266.17$ & $463.79$ & $384.23$ & $361.57$ & $334.63$ & \phantom{a}
& $324.40$ & $314.10$ & $331.83$ & $368.19$ & $334.63$ \\
& iPER & $313.43$ & $275.22$ & $344.94$ & $314.92$ & $267.39$ & $504.00$ & $404.79$ & $377.16$ & $350.23$ & \phantom{a}
& $277.98$ & $328.07$ & $358.30$ & $436.57$ & $350.23$ \\
& Ours & $237.16$ & $178.57$ & $286.86$ & $270.25$ & $237.86$ & $434.64$ & $301.66$ & $245.86$ & $\mathbf{274.11}$ & \phantom{a}
& $243.95$ & $260.14$ & $294.88$ & $297.46$ & $\mathbf{274.11}$ \\
\midrule
\multirow{5}{*}{\rotatebox[origin=c]{90}{\parbox[c]{1.5cm}{\centering FVD$\downarrow$}}}
& V-Unet & $1,491.63$ & $845.44$ & $1,721.81$ & $1,257.20$ & $1,415.24$ & $1,712.93$ & $2,437.98$ & $1,816.94$ & $1,587.39$ & \phantom{a}
& $2,239.14$ & $1,352.10$ & $1,856.78$ & $1,108.34$ & $1,639.09$ \\
& Vid2Vid & $879.94$ & $266.6$ & $1,085.49$ & $396.31$ & $790.79$ & $997.42$ & $997.96$ & $1,069.85$ & $810.48$ & \phantom{a}
& $778.53$ & $719.80$ & $762.46$ & $574.08$ & $\mathit{708.72}$ \\
& EBDN & $887.56$ & $273.00$ & $918.94$ & $423.08$ & $725.49$ & $952.22$ & $1,113.46$ & $853.26$ & $\mathit{768.37}$ & \phantom{a}
& $791.98$ & $751.45$ & $560.27$ & $826.71$ & $732.60$ \\
& iPER & $1,770.31$ & $656.07$ & $1,531.64$ & $1,266.14$ & $1,051.42$ & $1,322.72$ & $1,440.94$ & $1,719.55$ & $1,344.84$ & \phantom{a}
& $1,270.41$ & $1,092.82$ & $1,395.81$ & $1,214.64$ & $1,243.42$ \\
& Ours & $1,119.50$ & $330.91$ & $674.99$ & $478.93$ & $767.68$ & $791.01$ & $988.35$ & $760.33$ & $\mathbf{738.96}$ & \phantom{a}
& $715.00$ & $653.30$ & $720.49$ & $515.61$ & $\mathbf{651.10}$ \\
\bottomrule
\end{tabular}
}
\end{table*}
We compare our method against four recent representative methods with different assumptions, including V-Unet~\citep{Esser_2018_CVPR}, vid2vid~\citep{wang2018vid2vid}, EBDN~\citep{chan2018dance}, and iPER~\citep{lwb2019}. V-Unet is a notorious representative of image-to-image translation methods using conditional variational autoencoders to generate images based only on a 2D skeleton and an image from the target actor. Similar to V-Unet, iPER is not dataset-specific, but it is a generative model trained in an adversarial manner. Vid2vid and EBDN methods, for their turn, are dataset-specific methods, \textit{i}.\textit{e}., they require training a GAN for several days over one video of the target subject in a large set of different poses.
\vspace*{0.3cm}
\noindent\textbf{Processing time.} Although vid2vid and EBDN required a few seconds to generate a new frame, the training step of vid2vid spent approximately $10$ days on an NVIDIA Titan XP GPU for each target subject, and to run the fine-tuning of the EBDN took approximately $4$ days to complete all stages for each subject. On the other hand, our retargeting approach does not need to be trained. The significant parts of our method's processing time are the retargeting optimization, the deformation, and the model rendering. On an Intel Core i7-7700 CPU and NVIDIA Titan XP GPU, the average run-time for one frame of retargeting optimization was about $1.2$ seconds, including I/O. The deformation took around $12$ min on $8$ frames. The model rendering, \textit{i}.\textit{e}., selecting the best texture map considering the visibility map, warping the selected parts, and filling all the holes in the texture, took about $30$ seconds per frame. Thus, the total processing time $t(N)$ in seconds to run our method on a video with $N$ frames with a resolution of $1{,}920 \times 1{,}080$ is approximately $t(N) = 1.2\times N + 720 + 30 \times N$.
\begin{figure*}[t!]
\includegraphics[width=1.0\linewidth]{result_dataset_v1.pdf}
\caption{\textbf{Motion analysis in the dataset sequences.} Transferring results considering the cases where the person is not standing parallel to the image plane or has the arms in front of the face. In each sequence: the first row shows the worst generated frame for each method and the second row presents the best generated frame for each method.}
\label{result:quali-dataset-0}
\end{figure*}
\subsubsection{Quantitative Analysis}
We performed the video retargeting in all sequences in our dataset, including several public dance videos also adopted in the works~\cite{chan2018dance} and~\cite{lwb2019}. Table~\ref{table:metrics_result} shows the comparison of our approach in the dataset considering the motion types and pair of actors. We can see that despite not being dataset-specific, V-Unet and iPER did not perform well when the reference image is not from their datasets. One can see that our method outperforms, on average, all methods in considering SSIM, LPIPS, MSE, and FVD metrics. Regarding the experiments considering the pairs of actors, our approach also achieved the best average results in all metrics. In particular, our method presented better results when the subjects have different heights (S0-S1 and S6-S7). We ascribe this performance to our method being aware of the shape and physical interactions, which allows it to correct the person's position when the source actor is taller or smaller than the target person.
We also evaluated our results (in terms of SSIM, LIPIPS, and MSE metrics) with the Wilcoxon Signed-Rank Test to check whether the difference between our results' central tendency and the baselines are statistically significant. For all baselines, except vid2vid, the test rejects the null hypothesis. In other words, the test indicates that the samples were drawn from a population with different distributions and the differences between the metrics indicate that our method performed better statistically. Moreover, it is noteworthy that the sequences ``spinning'' and ``fusion dance'' are challenging for all methods, including our methodology that was affected by wrong pose estimations. Our approach was slightly outperformed by vid2vid only in these two sequences.
\begin{table}[t!]
\centering
\caption{{\bf Forgery performance}. Quantitative metrics of the movement transfer approaches considering the forgery performance metric (better close to $0$).}
\label{table:results_deepfake}
\resizebox{\columnwidth}{!}{
\begin{tabular}{@{}crrrrrrrrrrrrc@{}}
\toprule
\multirow{2}[3]{*}{Motion} & \multicolumn{5}{c}{Method} \\
\cmidrule(l{3pt}r{3pt}){2-6}
& EBDN & iPER & vid2vid & V-Unet & Ours \\
\midrule
\multicolumn{0}{l}{jump} & $39.46$ & $46.56$ & $41.66$ & $47.41$ & $\mathbf{31.11}$ \\
\multicolumn{0}{l}{walk} & $41.00$ & $44.27$ & $30.79$ & $45.79$ & $\mathbf{23.35}$\\
\multicolumn{0}{l}{spinning} & $33.64$ & $33.56$ & $32.36$ & $34.51$ & $\mathbf{21.09}$ \\
\multicolumn{0}{l}{shake hands} & $31.71$ & $33.56$ & $19.39$ & $34.93$ & $\mathbf{5.67}$ \\
\multicolumn{0}{l}{cone} & $38.23$ & $37.84$ & $\mathbf{24.74}$ & $39.25$ & $24.80$ \\
\multicolumn{0}{l}{fusion dance} & $26.07$ & $32.64$ & $16.49$ & $33.38$ & $\mathbf{15.97}$ \\
\multicolumn{0}{l}{pull down} & $31.30$ & $34.65$ & $19.85$ & $35.67$ & $\mathbf{17.85}$ \\
\multicolumn{0}{l}{box} & $35.96$ & $37.11$ & $24.70$ & $38.12$ & $\mathbf{17.63}$ \\
\bottomrule
\end{tabular}
}
\end{table}
\vspace*{0.3cm}
Table~\ref{table:results_deepfake} shows the results for the experiments on image forgery detection.
These experiments indicate that the fake detector in the frames generated by our method has the closest performances to real images. For instance, in the ``shake hands'' sequence, the probability of a frame synthesized by our method to be fake is only $5.67\%$ higher than when applying the detector to the respective real frame.
\subsection{Qualitative Analysis}
We evaluated the capability of our method to transfer motion and appearance and retain interactions of the original motion despite the target actor having different proportions to the actor in the source video.
\begin{figure}[!t]
\includegraphics[width=0.95\linewidth]{result_bruno_v3.pdf}
\caption{\textbf{Qualitative evaluation to bruno-mars sequence.} {\it First row:} original video and target actor S5; {\it Second row:} Result of vid2vid and their compositing with the target background; {\it Third row:} Results of EBDN and their compositing with the target background; Fourth row: Our results for both backgrounds.}
\label{result:bruno}
\end{figure}
Some frames used to compute the metrics in Table~\ref{table:metrics_result} are shown in Figure~\ref{result:quali-dataset-0}. One can note the large discrepancy between the quality of the frames for the same method. As shown in Figures~\ref{result:quali-dataset-0} and~\ref{result:bruno}, the end-to-end learning techniques have impressive results when the person is standing parallel to the image plane and the arms are not in front of the face; however, these methods perform poorly when the person is out of these contexts, such as when bending. Our method, for its turn, retains the same quality for most poses.
\begin{figure}[!t]
\includegraphics[width=0.95\linewidth]{result_tom_v3.pdf}
\caption{\textbf{Qualitative evaluation to tom-cruise sequence.} {\it First row:} original video and target actor S2; {\it Second row:} Result of vid2vid and their compositing with the target background; {\it Third row:} Result of EBDN and their compositing with the target background; {\it Fourth row:} Our results for both backgrounds.}
\label{result:tom}
\end{figure}
We also analyzed the impact of the camera pose and the actor's scale in the quality of the resulting videos of the methods. We transferred two target persons from our dataset to two videos with different camera setups and image resolutions: ``bruno-mars'' and ``tom-cruise'' sequences. In the ``bruno-mars'' sequence, shown in Figure~\ref{result:bruno}, we ran vid2vid and EBDN using their respective solution to tackle with different image resolutions and actors with different proportions from the training data. Vid2vid's strategy (scaling to keep the aspect ratio and then cropping the image) results in an actor with different proportions compared with the training data, which leads to a degradation of quality. EBDN's strategy (pose normalization, scaling to keep the aspect ratio, and then cropping the image) keeps the similarity between the input video and training data, but body and face occlusions are still problems. The red squares highlight these issues on the right side in Figure~\ref{result:bruno}, where EBDN and vid2vid reconstructed poorly the target's face. These strategies also incur a loss of the relative position in the original image; thus, whenever the actor presents a large translation, he/she can stay out of the crop area. Figure~\ref{result:tom} depicts the results of ``tom-cruise'' sequence. In this sequence, the source actor has a large translation in all directions; thus, we include a margin in the image to allow EBDN and vid2vid to process it. Nevertheless, the difference in the actor’s scale results in low quality for vid2vid and EBDN, which suggests their lack of generalization of these methods to different poses, and changes in camera viewpoint and intrinsic parameters. On its turn, our method did not suffer from problems caused by the camera and image resolution, and provided the best scores and visual results.
\subsubsection{Results in the iPER Dataset}
Aside from our dataset, we also evaluated our approach using the iPER dataset~\citep{lwb2019}. Since the iPER dataset does not provide paired motions as our dataset, we evaluated our method according to~\cite{lwb2019}'s protocol, where SSIM and LPIPS are used to measure self-imitation transferring. We also include FVD in the evaluation, which was designed to capture the temporal coherence among videos and frame quality.
Table~\ref{table:iper_dataset} shows the transferring results in the iPER dataset. It can be seen that our method outperforms iPER in terms of SSIM and LPIPS metrics. This result also concurs with the visual results shown in Figure~\ref{result:iper}. We can also note that iPER achieved the best FVD value, since they are tested now in the same context where it was trained. However, it still performs poorly when the person is not standing up straight, as indicated in the facial zoom (red box in Figure~\ref{result:iper}). Our method retains the same quality for most poses.
\begin{table}[t!]
\centering
\caption{{\bf Comparison with iPER in their proposed dataset}. Our approach is able to provide the best values in terms of SSIM and LPIPS, while performing worse in terms of FVD (best in bold).}
\label{table:iper_dataset}
\begin{tabular}{@{}clcc@{}}
\toprule
& {\bf Metric}& {\bf iPER} & {\bf Ours} \\ \cmidrule{2-4}
& SSIM$\uparrow$ & $0.8410$ & $\mathbf{0.8936}$ \\
& LPIPS$\downarrow$ & $0.0848$ & $\mathbf{0.0722}$ \\
& FVD$\downarrow$ & $\mathbf{955}$ & $1269$ \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[!t]
\includegraphics[width=0.95\linewidth]{self-imitation-iPer.pdf}
\caption{
\textbf{Qualitative evaluation on a sequence from the iPER dataset.} {\it First row:} target actor and original video; {\it Second row:} Results of iPER network; {\it Third row:} Our results.}
\label{result:iper}
\end{figure}
\subsection{Applications and Limitations}
In our experiments, we observed that image-based rendering and 3D reasoning still have several advantages regarding end-to-end learning human view synthesis, particularly in terms of control and generalizing to furthest views. Our method enables the replacement of the background, which is crucial for many applications. Moreover, it also allows to create a deformed model from images that can be included in virtual scenes using existing rendering frameworks, such as Blender. Figure~\ref{result:model-to-virtual} illustrates an application of this capability.
Although achieving the best results in these more generic test conditions, the proposed approach also suffers from certain limitations. Ideally, the constrained optimization problem would maintain the main features of the source motion. However, there is no single solution to a set of constraints, which can result in undesired motions, as shown in Figure~\ref{result:limitations}, where the actor positions his hand down and curve his back instead of bending his knees. The textured avatars may exhibit artifacts in the presence of part segmentation estimation errors, which can also lead to wrong deformations, and errors in the deformation result in body parts with unreal shapes. Typical failure cases such as retargeting and segmentation errors are also depicted in Figure~\ref{result:limitations}.
Regarding our method's processing time, our current implementation is a modular Python code, but without any optimization, \textit{i}.\textit{e}., the code was not parallelized either and was not adapted to run fully on a GPU. However, we highlight the available room for speeding up the processing time with a parallel implementation of different parts of the approach as, for instance, in the deformation model steps, which currently require $30$ seconds per frame. They could be easily adapted to be executed in parallel with multiprocessing.
\begin{figure}
\includegraphics[width=0.95\linewidth]{3d_applicability.pdf}
\caption{\textbf{Model-to-virtual}. Our method is able to provide a deformed model from images that can be included in virtual scenes using existing graphic rendering tools, such as Blender. The partial occlusions between the scene and model are handled in a natural way (red squares).}
\label{result:model-to-virtual}
\end{figure}
\begin{figure}
\includegraphics[width=0.95\linewidth]{limitations.pdf}
\caption{\textbf{Limitations}. {\it Top:} Retargeting resulting in undesired motion where the actor positions his hand down and curves his back instead of bending his knees. {\it Bottom:} Typical failure cases in the avatar: artifacts in the texture and body parts with unreal shapes.}
\label{result:limitations}
\end{figure}
Finally, in the proposed dataset, all videos were selected and designed to stress the evaluation of the human motion and appearance transfer capabilities of the methods for several paired movements of different actors. We adopted a simpler background to avoid eventual errors that could be added during the evaluation from the foreground-background segmentation (which is done before inpainting). This choice is meant to allow a better assessment of each method's capabilities to transfer actors' motions and appearance with different morphologies. However, there is still a lack of paired videos with complex backgrounds to assess the quality of the human retargeting and the background synthesis together.
\section{Conclusions}
This paper proposes a new model-based retargeting methodology that incorporates different strategies to extract 3D shape, pose, and appearance to transfer motions between two real human characters using the information from monocular videos. Unlike retargeting methods that use either appearance information or motion information only, our approach simultaneously considers the critical factors to retargeting such as pose, shape, appearance, and motion features.
Furthermore, we have introduced a new dataset comprising different videos with annotated human-to-object interactions, actors with different body shapes, and paired motions to evaluate the task of retargeting humans in videos. We believe that these videos can be used as a common base to improve tracking the field's progress by showing where current approaches might fail and measuring the quality of future proposals.
Despite the impressive advances in image-to-image translation approaches, we presented several contexts where most end-to-end image translation methods perform poorly. Yet, our experiments show that a model-based rendering technique can still exhibit a competitive quality compared to recent end-to-end learning techniques, besides having several advantages in terms of control and ability to generalize to furthest views. Our results indicate that retargeting strategies based on image-to-image learning are still challenged to retarget motions while keeping the desired movement constraints, shape, and appearance simultaneously. These findings suggest the potential of hybrid strategies by leveraging the advantages provided by model-based retargeting into recent neural rendering approaches. \\
\noindent\textbf{Acknowledgements.} The authors thank CAPES, CNPq, and FAPEMIG for funding this work. We also thank NVIDIA for the donation of a Titan XP GPU used in this research.
{ \small
\bibliographystyle{spbasic}
|
\section{Introduction}
\label{sec:Intro}
\input{introduction}
\section{The Bayesian approach for solving inverse problems}
\label{sec:bayes}
\input{bayes}
\section{Evaluating the Posterior}
\label{sec:posterior}
\input{posterior}
\section{Numerical Experiments}
\label{sec:numerics}
\input{numerics}
\section{Conclusions}
\label{sec:conclusions}
\input{conclusions}
\vspace{5mm}
\noindent {\bf Acknowledgments.}
Anne Gelb's work is supported in part by the NSF grants DMS \#1502640 and DMS \#1912685, AFOSR grant \#FA9550-18-1-0316, and ONR MURI grant \#N00014-20-1-2595.
Theresa Scarnati's work is supported in part by AFOSR LRIR \#18RYCOR011.
\subsection{{Bayes' Theorem}}\label{sec:bayestheorem}
Let $x,y,e \in {\mathbb{R}^n}$ and $\lambda\in[0,1]$ be realizations of the random variables $\mathbf{X}$, $\mathbf{Y}$, $\mathbf{E}$ and $\bm{\lambda}$ respectively. Bayes' theorem is given by
\begin{equation}
\label{eq:full_posterior}
f_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x,\lambda|y) =
\frac{f_{\mathbf{Y|X}}(y|x){f}_{\mathbf{X}| \bm{\lambda}}(x| \lambda)f_{\bm{\lambda}}(\lambda)}{f_{\mathbf{Y}}(y)},
\end{equation}
where ${f}_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x,\lambda|y)$ is the {posterior density function} we seek to recover, $f_{\mathbf{Y}|\mathbf{X}}(y|x)$ is the {likelihood function}, and $\tilde{f}_{\mathbf{X},\bm{\lambda}}(x,\lambda) = {f}_{\mathbf{X}| \bm{\lambda}}(x| \lambda)f_{\bm{\lambda}}(\lambda)$ is the {prior probability density function}. The prior is broken into two parts, the conditional prior ${f}_{\mathbf{X}| \bm{\lambda}}(x| \lambda)$ and the hyper-prior $f_{\bm{\lambda}}(\lambda)$. Further, $f_{\mathbf{Y}}(y)$, often called the {evidence}, is given by the total law of probability as
\begin{equation}\label{eq:MarginalDensity}
f_{\mathbf{Y}}(y) =
\int_{\mathbb{R}^n}\int_{[0, 1]}f_{\mathbf{X}, \mathbf{Y}, \bm{\lambda}}(x, y, \lambda) d\lambda dx=
\int_{\mathbb{R}^n}\int_{[0, 1]} f_{\mathbf{Y|X}}(y|x){f}_{\mathbf{X}| \bm{\lambda}}(x| \lambda)f_{\bm{\lambda}}(\lambda)d\lambda dx.
\end{equation}
It is reasonable to assume that \eqref{eq:MarginalDensity} is nonzero because {otherwise} the observation $y \in \mathbb{R}^n$ has a probability of zero and is therefore irrelevant in {the applications we are considering. Moreover,} since calculating \eqref{eq:MarginalDensity} is often computationally intractable, particularly in high dimensional domains, a closed form posterior density is often unavailable. Thus sampling methods such as those described in {Section \ref{sec:IntervalEstimation}} are typically used to approximate the un-normalized version of the posterior \eqref{eq:full_posterior} given by
\begin{equation}
\label{eq:f_hat}
\hat{f}_{\mathbf{X}, \bm{\lambda}|\mathbf{Y}}(x, \lambda|y) =
f_{\mathbf{Y|X}}(y|x){f}_{\mathbf{X}| \bm{\lambda}}(x| \lambda)f_{\bm{\lambda}}(\lambda).
\end{equation}
{In our numerical experiments} we use the Metropolis-Hastings (MH) algorithm to draw samples of (\ref{eq:f_hat}).
In what follows, Section \ref{sec:likelihood} defines the likelihood $f_{\mathbf{Y|X}}(y|x)$ with respect to the proposed linear model (\ref{eq:ForwardModel}), while Section \ref{sec:hyper-prior} describes how we estimate the hyper-prior $\lambda$. Section \ref{sec:prior} focuses on the definition of the conditional prior ${f}_{\mathbf{X}|\bm{\lambda}}(x| \lambda)$. There we will first discuss {how} sparsity enforcing priors are often employed in practice and then introduce our novel support informed sparse prior that is {empirically} constructed by exploiting the joint sparsity across the multiple measurements in the sparse domain of the underlying signal. Finally, in Section \ref{sec:post} we define and explore the posterior $\hat{f}_{\mathbf{X}, \bm{\lambda}|\mathbf{Y}}(x, \lambda|y)$.
\subsection{Estimating the likelihood}
\label{sec:likelihood}
Returning to the model in \eqref{eq:ForwardModel}, observe that
\[
f_{\mathbf{Y}|\mathbf{X}, \mathbf{E}}(y|x, e) = \delta(y-Ax-e),
\]
where $\delta(\cdot)$ is the Dirac delta function. Marginalizing with respect to the noise $e$ yields
\begin{align}
f_{\mathbf{Y}|\mathbf{X}}(y|x) &= \int_{\mathbb{R}^n} f_{\mathbf{Y}|\mathbf{X}, \mathbf{E}}(y|x, e)f_{\mathbf{E}|\mathbf{X}}(e|x)de\nonumber\\
&= \int_{\mathbb{R}^n} \delta(y-Ax-e)f_{\mathbf{E}|\mathbf{X}}(e|x)de\nonumber\\
&= f_{\mathbf{E}|\mathbf{X}}(y-Ax|x)\nonumber\\
&= f_{\mathbf{E}}(y-Ax),\label{eq:errordensity}
\end{align}
with the last equality resulting from the independence between $\mathbf{X}$ and $\mathbf{E}$, \cite{kaipio2006statistical}. Hence we see that the likelihood function, $f_{\mathbf{Y}|\mathbf{X}}(y|x)$, depends on the noise variable $\mathbf{E}$. Since $\mathbf{E}$ is Gaussian with mean $0$ and {covariance} matrix $\Sigma = \sigma^2 \mathbb{I}_{{n}}$, the noise density is given by
\begin{equation}\label{eq:NoiseDensity}
f_{\mathbf{E}}(e) \propto \exp \{ - \frac{1}{2\sigma^2}(e-0)^T (e-0) \} = \exp \{ - \frac{||e||^2_2}{2 \sigma^2} \}.
\end{equation}
Thus from \eqref{eq:errordensity} and \eqref{eq:NoiseDensity} we obtain the likelihood density
\begin{equation}\label{eq:LikelihoodDensity}
f_{\mathbf{Y}|\mathbf{X}}(y|x)=f_{\mathbf{E}}(y-Ax) \propto \exp\{ - \frac{1}{2\sigma^2}||y-Ax||^2_2\}.
\end{equation}
\subsection{Estimating the hyper-prior} \label{sec:hyper-prior}
{We use $K$-fold cross validation to estimate the hyper-prior, \cite{murphy2012machine}. To summarize this procedure, we begin by approximating}
\begin{align}\label{eq:hyper_prior_posterior1}
f_{\bm{\lambda}}(\lambda) \approx f_{\bm{\lambda}|\mathbf{Y}}(\lambda|y) =
\frac{f_{\mathbf{Y}| \bm{\lambda}}(y | \lambda)\tilde{f}_{\bm \lambda}(\lambda)}{f(y)} \propto f_{\mathbf{Y}|\bm{\lambda}}(y|\lambda)\tilde{f}_{\bm{\lambda}}(\lambda),
\end{align}
where $\tilde{f}_{\bm \lambda}(\lambda)$ is the prior on the hyper-prior. By assuming $\lambda\sim U[0,1]$ so that $\tilde{f}_{\bm \lambda}(\lambda) = 1$, we have
\begin{align}
\label{eq:hyper_prior_posterior2}
f_{\bm{\lambda}}(\lambda) \propto f_{\mathbf{Y}|\bm{\lambda}}(y|\lambda).
\end{align}
The mode of the distribution (\ref{eq:hyper_prior_posterior2}) can be approximated using the MAP estimate as
\begin{equation}
\label{eq:lambda_hat_MLE}
\hat{\lambda}
= \argmax{\lambda} \{ f_{\mathbf{Y} |\bm{\lambda}}(y|\lambda) \}
= \argmax{\lambda} \int_{\mathbb{R}^n} f_{\mathbf{Y}|\mathbf{X}}(y|x)f_{\mathbf{X}|\bm{\lambda}}(x|\lambda) dx.
\end{equation}
Integration with respect to $x$ over the space $\mathbb{R}^{n}$ is not computationally feasible. We therefore instead employ $K$-fold cross-validation to approximate $\hat{\lambda}$ given the data $y_j \in \Omega$ for $j = 1,\dots,J$. The general process for $K$-fold cross-validation is to first calculate $M$ MAP estimates of the unknown from a subset of $M < J$ observable measurements using sample candidates of the hyper-prior $\hat{\lambda}$. These are then used to generate $M$ {\em training vectors}, which are in turn compared to the remaining $J-M$ observable measurements, or {\em testing vectors}. The whole process is then repeated for $K$ independent trials, and
$\hat{\lambda}$
is chosen to minimize the mean square error of (\ref{eq:ForwardModel}) between all training and testing vectors.
The procedure is summarized in Algorithm \ref{alg:Cross-Validation}.
\begin{algorithm}[H]
\caption{$K$-Fold Cross-Validation to Estimate $\hat{\lambda}$.}
\label{alg:Cross-Validation}
\begin{algorithmic}[1]
\STATE{\textbf{Input:} Set of multiple measurements (MMVs) ${\bf y} = [y_1, y_2, y_3, \ldots, y_J]$.}
\STATE{\textbf{Output:} Point Estimate of the hyper-prior $\hat{\lambda}$.}
\FOR{$k = 1$ to $K$}
\STATE{Randomly partition MMVs into $M$ training vectors and $J-M$ testing vectors.}
\FOR{$i=1$ to $M$}
\STATE{Sample candidate hyper-prior from $\tilde{\lambda}_{i,k}\sim U[0,1]$.}
\STATE{Calculate the MAP estimate of the unknown as
$$
\hat{x}_{i,k} = \argmax{x} \{f_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x,\tilde{\lambda}_{i,k}|y_i)\}.
$$}
\STATE{Compute training data according to $\hat{y}_{i,k} = A\hat{x}_{i,k}.$}
\STATE{Evaluate the training data by calculating
$$
E_{i,k} = \frac{1}{J-M}\sum_{l=M+1}^{J}\text{MSE}(\hat{y}_{i,k},y_l),
$$
where $\{y_l\}_{l = M+1}^J$ are the partitioned $J-M$ testing vectors.}
\ENDFOR
\ENDFOR
\STATE{Choose $(i^*,k^*)= \argmin{i,k} \hspace{1mm} E_{i,k}$ and set $\hat{\lambda} = \tilde{\lambda}_{i^*,k^*}$.}
\end{algorithmic}
\end{algorithm}
{The hyper-prior $\hat{\lambda}$ recovered from Algorithm \ref{alg:Cross-Validation} approximates the solution to (\ref{eq:lambda_hat_MLE}) such that
\begin{equation}
\label{eq:lambda_hat_delta}
f_{\bm{\lambda}}(\lambda) \approx \delta_{\hat{\lambda}}(\lambda).
\end{equation}
Intuitively, \eqref{eq:lambda_hat_delta} implies that we are simply concentrating the density (\ref{eq:hyper_prior_posterior2}) on the value $\hat{\lambda}$, which is our estimate of the mode of the hyper-prior distribution (\ref{eq:lambda_hat_MLE}). {Note that the $K$ fold validation process assumes that the distribution for $\lambda$ is unimodal. Previous efforts, see e.g. \cite{murphy2012machine}, have also demonstrated that (\ref{eq:lambda_hat_delta}) is a good approximation of (\ref{eq:lambda_hat_MLE}).}
\begin{remark}
The hyper-prior $\hat{\lambda}$ recovered in Algorithm \ref{alg:Cross-Validation} is consistent with how one ideally chooses the regularization parameter in the CS approximation given by \eqref{eq:l1reg}, or as we will again see in \eqref{eq:MAPEstimate}, where the {fidelity and regularization} parameters are combined to form a single parameter. Specifically this parameter should be chosen to offset the variance of the noise, and the $K$-fold cross validation method provides an empirical approach to calculating this variance.
\end{remark}
\subsection{A support informed sparsity prior}\label{sec:prior}
The main focus of this investigation is to construct a support informed sparsity prior $\tilde{f}_{\mathbf{X},\bm{\lambda}}(x,\lambda) = {f}_{\mathbf{X}| \bm{\lambda}}(x| \lambda)f_{\bm{\lambda}}(\lambda)$ in \eqref{eq:full_posterior}. As will be demonstrated in what follows, this will reduce the uncertainty of the posterior for models given by \eqref{eq:ForwardModel}.
Using the approximation in (\ref{eq:lambda_hat_delta}) we can make the approximation
\begin{equation*}
\tilde{f}_{\mathbf{X},\bm{\lambda}}(x,\lambda) = {f}_{\mathbf{X}| \bm{\lambda}}(x| \lambda)f_{\bm{\lambda}}(\lambda) \approx {f}_{\mathbf{X}| \bm{\lambda}}(x| \hat\lambda),
\end{equation*}
and it remains to define the prior through ${f}_{\mathbf{X}| \bm{\lambda}}(x| \hat\lambda)$.
To do so, we make two assumptions: (i) the prior enforces the sparsity of some transformation of $\mathbf{X}$ and (ii) there are multiple observations $y$ of $\mathbf{Y}$ for each realization $x$ of $\mathbf{X}$.\footnote{{Similarly to what was done in \cite{VBJS2, VBJSGelb}, it is possible to relax this second assumption and instead have a single measurement vector {\em processed} multiple times. For purposes of brevitity we exclude such considerations in this investigation.}}
{There are many applications for which it is reasonable to assume that the desired solution is sparse in some domain. In our investigation, the underlying signal is assumed to be piecewise smooth, implying that the edge domain is sparse. It is important to note that as a departure from the traditional sparse priors commonly used in formulating the posterior, our new support informed sparse prior considers both the {\em existence} of the sparse domain as well as the {\em locations} of the non-zero values within it, i.e., its support.}
\subsubsection{{Sparsity promoting priors}}
\label{sec:prior_smv}
Within the Bayesian formulation, sparsity {may be} enforced through the Laplace prior given by
\begin{equation}\label{eq:laplace_prior}
f_{\mathbf{X}| \bm{\lambda}}(x| \hat\lambda) \propto \exp\{ - \hat\lambda ||\mathcal{L}x||_1 \},
\end{equation}
where the operator $\mathcal{L}$ is a transform operator used to approximate the sparse domain. In this investigation we choose $\mathcal{L}$ to be the {\em polynomial annihilation} (PA) transform, \cite{ArchibaldGelb,VBJS}, {which is designed to approximate the corresponding edge function of a piecewise smooth function when acting on the unknown (given by $x$ in \eqref{eq:laplace_prior}) {at specified grid points} in the spatial domain.\footnote{A short description of the PA transform is provided in Appendix \ref{appendix:PA}.} We note that our new method is not defined by the choice of sparsifying transform (here ${\mathcal L}$), {nor by the sparsity domain (here the edge domain)}, and that other choices may be more appropriate for different applications.
Regardless of how both the hyper-parameter $\hat{\lambda}$ and sparsifying transform matrix $\mathcal{L}$ are determined, the Laplace prior given in (\ref{eq:laplace_prior}) may not provide the most information about what is known about the underlying signal. As discussed {already}, a better prior would make use of the support {\em locations} in the sparse domain.
Therefore, following the ideas in \cite{VBJSGelb,VBJS}, we develop a new support informed sparse prior. In so doing, we better capture the piecewise smoothness of the underlying signal. A couple of remarks are in order:
\begin{remark}
{Our numerical experiments confirm that using the PA transform with $m > 1$ is appropriate when constructing the support informed sparse prior for a piecewise smooth function. In particular using $m = 1$, which is equivalent to total variation (TV), does not perform as well. In general when the underlying function is {\em not} constant in smooth domains, a {\em higher degree} polynomial annihilation is needed to recover an accurate representation of the sparse domain. This issue is extensively discussed for the compressive sensing framework in \cite{ArchibaldGelb}, where higher order PA is used to avoid the ``staircasing'' effect caused by TV regularization. Similarly, the high order PA describes a better prior in our approach used here.}
\end{remark}
\begin{remark}
\label{rem:horseshoe}
{To be clear, there are a variety of sparsity promoting priors to choose from, including but not limited to the spike and slab prior and the horseshoe prior \cite{bhadra2019lasso, piironen2017sparsity}. {Hierarchical Gaussian priors, which provides the basic framework for the methods described in \cite{BardsleyUQ, tipping2001sparse}, may also be used.} {Regardless of the type of prior chosen, some transform operator ${\mathcal L}$ to the sparse domain is still needed.} This investigation does not intend to provide a complete survey of the advantages and drawbacks of using various sparsity promoting priors. Instead we demonstrate that our proposed prior is a departure from the aforementioned priors because it {\em explicitly} includes the support locations in the sparse domain. Indeed, such knowledge may be incorporated into some of the aforementioned choices. Even further, our approach may be used to form mixed priors. That is, because our spatially adaptive prior encourages the separation of scales in the sparse domain, different priors may be used in different regions of the domain. The Laplace prior in \eqref{eq:laplace_prior} is mentioned as a prototype, and is {a common} choice when interested in {constructing the MAP estimate of the posterior} (CS approach). As will be demonstrated by our numerical examples, our support informed sparse prior yields better accuracy and reduced uncertainty than other priors typically used. More information about sparse priors may be found in \cite{piironen2017sparsity}.
\end{remark}
\subsubsection{Joint sparsity}\label{subsec:jointsparse}
As noted in the introduction, many applications collect multiple measurement vectors (MMVs) of observable data. By exploiting the {\em joint sparsity} in the sparse domain of the underlying signal that can be obtained from the MMVs, a technique was developed in \cite{VBJSGelb, VBJS} to improve the robustness of the MAP estimate.\footnote{{We note that the method was derived from the CS perspective.}} For reasons that will become apparent later in this section, the technique was coined the Variance Based Joint Sparsity (VBJS) method. We now adapt VBJS to construct a new {\em support informed sparse prior}.
We then demonstrate that this new prior yields reduced uncertainty and improved efficiency in the {sampling of} the posterior in \eqref{eq:f_hat} when compared to other typical prior definitions, such as the Laplace prior in (\ref{eq:laplace_prior}).
As previously mentioned, the domain of the problem is $\mathbb{R}^n$ so that the unknown $x$ may be represented by a $n$ length vector. {A $d$-dimensional problem will then yield an $n^d$ length vector representation of the unknown, and as already noted our method is easily parallelizable since it samples the posterior for each pixel of interest.} Suppose that we have multiple vector realizations of $y_j \sim \mathbf{Y} \in \mathbb{R}^n$ to form $Y = [y_1, y_2, y_3, \ldots, y_J]$. In practice, the data vectors $y_j$ are often noisy {indirect} measurements of the same underlying unknown signal represented by a random variable $\mathbf{X}$. It is therefore realistic to suppose that they should share similar support in a chosen sparse domain, for instance in the edge domain. This is the {\em joint sparsity} assumption. As noted previously, other sparse domains may also be used for this purpose.
In order to make use of the joint sparsity concept, we first provide some definitions and notation, which may be found in \cite{VBJSGelb}. Let $p \in \mathbb{R}^n$. As is conventional, we write $\nm{\cdot}_0$ for the $\ell^0$-`norm', i.e.
\bes{
\nm{p}_0 = \left | \mathrm{supp}(p) \right |,
}
where $\mathrm{supp}(p)$ is the support of $p = (p_i)^{n}_{i=1}$ defined by
\bes{
\mathrm{supp}(p) = \{ i : p_i \neq 0 \}.
}
Given a vector $\bm{w} = (w_i)^{n}_{i=1}$ of positive weights, we define the weighted $\ell^1_{\bm{w}}$-norm as
\begin{equation}
\nm{p}_{1,\bm{w}} = \sum^{n}_{i=1} w_i | p_i |.
\label{eq:weightedl1}
\end{equation}
It is also possible to define weighted $\ell^q_{\bm{w}}$-norms for $q \neq 1$, but this is not needed for our investigation.
If $P = (p_{i,j})^{n,J}_{i,j=1} \in \mathbb{R}^{n \times J}$ is a matrix, we define the $\ell^{q_1,q_2}$-norms by
\bes{
\nm{P}_{q_1,q_2} = \left( \sum^{n}_{i=1} \left ( \sum^{J}_{j=1} | p_{i,j} |^{q_1} \right )^{q_2/q_1} \right )^{1/q_2}.
}
Again by convention the $\ell^{2,0}$-`norm' is defined as
\bes{
\nm{P}_{2,0} = \left | \left \{ i : \sum^{J}_{j=1} | p_{i,j} |^2 \neq 0 \right \} \right |.
}
We are now able to provide the following definition:
\begin{definition}
\label{def:joint_sparsity}
A vector $p \in \mathbb{R}^n$ is $s$-sparse for some $1 \leq s \leq N$ if
$$
\nm{p}_{0} = | \mathrm{supp}(p) | \leq s.
$$
A collections of vectors ${p}_1,\ldots,p_J \in \mathbb{R}^n$ is $s$-joint sparse if
$$
\nm{P}_{\alpha,0} = \left | \bigcup^{J}_{j=1} \mathrm{supp} \left (p_j \right ) \right | \leq s,\quad\quad{{\alpha > 0}}
$$
where $P = \left [ p_1 | \cdots | p_J \right ]$. {In typical applications $\alpha$ is chosen to be $2$.}
\end{definition}
Based on our discussion following (\ref{eq:laplace_prior}), the sparsifying operator $\mathcal{L}$ is the PA transform where $p_j = \mathcal{L}x_j$ approximates the solution $x$ in the edge domain for each $j = 1,\cdots,J$. It follows that $P = \left[\mathcal{L}x_1|\cdots| \mathcal{L}x_J\right]$ in Definition \ref{def:joint_sparsity}.
If the edges were explicitly known, the prior in \eqref{eq:laplace_prior} could be tailored to promote a posterior that not only reflects the existence of the sparse domain but also the regions of its support. Specifically, a more accurate prior would be dependent on $\mathrm{supp}(\mathcal{L} x)$. Although this information is not directly available in most applications, it is possible to extract it (approximately) from the observable data, as was demonstrated in \cite{VBJSGelb,VBJS, VBJS2} using the VBJS method. In the context of this investigation, the VBJS approach is an example of empirical Bayesian inference, and results in a support informed sparse prior.
\subsubsection{Variance based joint sparsity (VBJS)}\label{subsec:VBJS}
To use the VBJS approach, we first require the calculation of the sample variance $v \in \mathbb{R}^n$ across the rows of $P$, where $P = [\mathcal{L} x_1 | \dots |\mathcal{L}x_J]$ is the sparse matrix defined by Definition \ref{def:joint_sparsity}. The (spatial) entries of $v$ are given by
\begin{equation}
\label{eq:var_comp}
{v}_i = \frac{1}{J}\displaystyle\sum_{j=1}^J {P}_{i,j}^2 - \left(\frac{1}{J}\displaystyle\sum_{j=1}^J {P}_{i,j}\right)^2, \quad i = 1,\dots,n.
\end{equation}
We note that because we do not have the true solution vector, $x$, we first must approximate each $\mathcal{L}x_j$ from the observable MMV data $y_j$, $j = 1,\dots,J$. This may be done in a variety of ways, and in some cases does not require an initial approximation of the solution itself, \cite{VBJS2}.
The VBJS technique constructs a weighting matrix $W = diag(w_1,\dots, w_n)$ that weighs the importance of the prior for different regions of the domain based on the sparse representations of the data. {For the compressive sensing approach used in \cite{VBJSGelb,VBJS}, these spatially varying weights were scaled to essentially be inversely proportional to the reciprocal of \eqref{eq:var_comp}.} That is,
{
$$
w_{i} = \frac{1}{v_i + \epsilon},
$$
where $\epsilon > 0$ is some small parameter that offsets zero variance. As discussed in \cite{VBJS}, the weights may be further scaled to account for the different magnitudes of the non-zero values in the sparse domain. While such separation of scales is important for determining the spatially varying weighting parameter in the CS environment, it is not needed in the Bayesian framework for reasons that will become clear below. What is important is to note that} the VBJS technique's reliance on the variance of the measurements in the sparse domain is motivated by the provable convergence rates of edge detection methods near and away from jumps, \cite{archibald2005polynomial,GT1,GT06,VBJS2}. In particular, if the variance in the sparse domain is high in a region of $\mathbb{R}^n$, then the corresponding entries of $W$ are small. This follows the assumption that high variance values indicate regions where the measurements share support in the sparse domain. Conversely, low variance values suggest truly sparse regions, and in this case the corresponding entries of $W$ should be large to strongly enforce the sparse prior.
For the purpose of empirical Bayesian inference, and specifically to define the support informed sparse prior, we simply threshold the variance so that $W$ becomes a {diagonal} binary mask {$M = diag(m_1,\dots,m_n) \in \mathbb{R}^{n \times n}$} that enforces a prior {in regions that are truly sparse, but assumes no prior in regions of support.} More specifically, for threshold $\tau = \frac{1}{n}$ we define {the diagonal entries of $M$ as}
\begin{equation}
m_{i} = \begin{cases}
{1}, &\quad w_i \geq \tau \\
{0}, &\quad w_i < \tau,
\end{cases}
\label{eq:mask}
\end{equation}
where $w_i$, $i = 1,\dots, n$, {are provided above.}
Figure \ref{fig:mask_demo} demonstrates the process for obtaining (\ref{eq:mask}). Observe that because the weights $w_i$ exhibit a large separation of scale, our masking technique is not sensitive to the thresholding parameter $\tau$, {which was chosen simply to be consistent with the resolution of the data}.
\begin{figure}[htb]
\includegraphics[width = \textwidth]{mask_demo.pdf}
\caption{Demonstration of the calculation of the mask used in the proposed support informed sparsity prior. (top-left) True function. (top-middle) Jointly sparse vectors generated from indirect, noisy measurements $y_j$, $j = 1,\dots,10$, of $x$. (top-right) Spatial variance $v$ (\ref{eq:var_comp}) across joint sparsity vectors. (bottom-left) Weights calculated using the VBJS technique, \cite{VBJSGelb,VBJS}. (bottom-right) Mask (\ref{eq:mask}) generated for use in the support informed sparsity prior. The red crosses indicate locations {in the spatial domain $s$} where we calculate evaluation metrics described in Section \ref{sec:numerics} }
\label{fig:mask_demo}
\end{figure}
Given the {binary mask defined} in (\ref{eq:mask}), Algorithm \ref{alg:SISP} describes the VBJS process for constructing a support informed sparsity prior {as it is used for this investigation}.
\begin{algorithm}[H]
\caption{Support Informed Sparsity Prior}
\label{alg:SISP}
\begin{algorithmic}[1]
\STATE{\textbf{Input:} Set of measurement vectors, ${{ Y}} = [y_1, y_2, \dots, y_J]$.}
\STATE{\textbf{Output:} Support informed sparsity prior.}
\STATE{Estimate the hyper-parameter $\hat{\lambda}$ using Algorithm \ref{alg:Cross-Validation}.}
\STATE{{Approximate the joint sparsity matrix $P$ defined in Definition \ref{def:joint_sparsity} from the $J$ measurements.}}
\STATE{{Compute the spatial variance $v$ in \eqref{eq:var_comp} from the joint sparsity matrix $P$.}}
\STATE{Use the spatial variance $v$ to determine the weighting vector $w$.}
\STATE{{Use $w$ to determine the binary mask $M$ according to \eqref{eq:mask}.}}
\STATE{Define the support informed sparsity prior density as
\begin{equation}
\label{eq:SISP_prior}
\tilde{f}_{\mathbf{X}|\bm{\hat{\lambda}},M}(x|\hat{\lambda},M) = C \exp\{-{\hat{\lambda} }||{M}\mathcal{L}x||_1\},{\quad C > 0}.%
\end{equation}}
\end{algorithmic}
\end{algorithm}
A couple of remarks are in order:
\begin{remark}
In the original construction of the weighting matrix $W$ in \cite{VBJS}, the hyper-parameter $\hat{\lambda}$ {was intrinsically fixed}. Specifically, in the CS solution \eqref{eq:l1reg}, the regularization parameter $\lambda$ was set to $1$. {Therefore, to separate the scales of the non-zero components in the sparse domain, it was important that $W$ {\em not} be binary, that is, it had to vary in magnitude to accommodate the various scales of the non-zero components in the sparse domain.} Furthermore, although not derived this way, one may understand this scaling to be informed by the error estimate of the fidelity term, much like what is obtained through Algorithm \ref{alg:Cross-Validation}. By contrast, in the current Bayesian framework we consider the hyper-parameter in the prior to be independent from the support in the sparse domain. This may be advantageous in circumstances where it is difficult to determine regions of support. In such cases the resulting mask is close to the identity matrix, and using a hyper-parameter determined through methods such as cross validation may still reduce uncertainty when compared to standard priors that {have fixed hyper-parameters (which is analogous to choosing the best regularization parameter in the CS approach).} Hence we set the weighting matrix to be a {binary} mask matrix indicating the support regions and learn the hyper-parameter $\hat{\lambda}$ separately, rather than combining these parameters into one as was done in \cite{VBJS}.
\end{remark}
{
\begin{remark}
\label{rem:ell2}
The Laplace prior is commonly used to promote sparsity in the sparse domain of the underlying signal, especially when {constructing the MAP estimate of the posterior.} Analogously, $\ell_1$ regularization is chosen in the compressive sensing formulation since using $\ell_2$ does not effectively promote sparsity, \cite{CompressiveSensingCandes}. However, when the mask matrix $M$ is applied, the {sparse} regions of $\mathbf{X}$ that are to be regularized are identically zero. Hence there is no fundamental reason to choose the $\ell_1$ norm. Indeed, the $\ell_2$ norm was employed in the compressive sensing algorithms designed in \cite{churchill2018edge} and \cite{VBJS}, and in both cases, the resulting algorithms more efficient. Adopting the same rationale here, for comparative purposes, we also consider a support informed $\ell_2$ prior.
\end{remark}
\subsection{Defining the Posterior}
\label{sec:post}
Armed with both a likelihood (\ref{eq:LikelihoodDensity}) and the support informed sparsity prior (\ref{eq:SISP_prior}), we are able to define the full posterior (\ref{eq:f_hat}) for the model in (\ref{eq:ForwardModel}).
{Based on our previous discussion, we will consider two posterior estimates which both use \eqref{eq:LikelihoodDensity} and hyper-parameter $\hat{\lambda}$ calculated in Algorithm \ref{alg:Cross-Validation}:}
\begin{subequations}
\label{eq:final_posterior}
\begin{equation}
\label{eq:PosteriorDensitylaplace}
\hat{f}_{\mathbf{X}, \bm{\lambda}|\mathbf{Y}}(x, \hat{\lambda}|y)= C_L \exp \{ - {\hat{\lambda}} ||\mathcal{L}x||_1 - \frac{1}{2\sigma^2}||y-Ax||_2^2\},\quad\quad C_L> 0,
\end{equation}
\begin{equation}
\label{eq:PosteriorDensitymask}
\hat{f}_{\mathbf{X}, \bm{\lambda}|\mathbf{Y}}(x, \hat{\lambda}|y)= C_M \exp \{ - {\hat{\lambda}} ||M\mathcal{L}x||_1 - \frac{1}{2\sigma^2}||y-Ax||_2^2\},\quad\quad C_M > 0.
\end{equation}
\end{subequations}
The first uses the Laplace prior \eqref{eq:laplace_prior}, while the second uses the support informed sparse {$\ell_1$} prior \eqref{eq:SISP_prior} constructed in Algorithm \ref{alg:SISP}.
Although parameters $C_L$ and $C_M$ are not explicitly known, they are not needed in calculation. {Note that within this MMV framework, $\sigma{^2}$ can simply be estimated as the average sample variance across all MMVs in the data space}, {obtained as
\[{\sigma}^2 \approx \frac{1}{n}\left(\frac{1}{J}\displaystyle\sum_{j=1}^J {Y}_{i,j}^2 - \left(\frac{1}{J}\displaystyle\sum_{j=1}^J {Y}_{i,j}\right)^2\right),\]
where ${Y} = \left [ y_1 | \cdots | y_J \right ]$.
}
For comparative purposes, we will also consider the following posteriors,
\begin{subequations}
\label{eq:final_posteriorl2}
\begin{equation}
\label{eq:PosteriorDensityhierarchical}
\hat{f}_{\mathbf{X}, \bm{\lambda}|\mathbf{Y}}(x, \hat{\lambda}|y)= C_L \exp \{ - \frac{\hat{\lambda}}{2} ||\mathcal{L}x||_2^2 - \frac{1}{2\sigma^2}||y-Ax||_2^2\},\quad\quad C_L> 0,
\end{equation}
\begin{equation}
\label{eq:PosteriorDensitymaskl2}
\hat{f}_{\mathbf{X}, \bm{\lambda}|\mathbf{Y}}(x, \hat{\lambda}|y)= C_M \exp \{ - \frac{\hat{\lambda}}{2} ||M\mathcal{L}x||_2^2 - \frac{1}{2\sigma^2}||y-Ax||_2^2\},\quad\quad C_M > 0,
\end{equation}
\end{subequations}
{for which we use the same approximation of $\sigma^2$ as above.} As discussed in the introduction, the method in \cite{BardsleyUQ} uses a hierarchical Gaussian prior (HGP) with the scalar hyper-priors chosen to be Gamma distributed. Furthermore, the sparse transform operator in \cite{BardsleyUQ} is symmetric positive definite. Because of the resulting conjugacy relationships, the posteriors are known in closed form and the Gibbs sampling method can be efficiently applied, which is a primary objective in \cite{BardsleyUQ}. {It is important to note that employing the mask matrix $M$ to ${\mathcal L}$ (even if ${\mathcal L}$ is constructed to be SPD), would not yield an SPD transform.} Without considering {the convergence properties of the corresponding numerical implementations}, the posterior provided in \eqref{eq:PosteriorDensityhierarchical} is comparable to the one in \cite{BardsleyUQ}. Indeed it is suggested there that additional information {may be used to determine the hyper-prior. The $K$-fold cross validation algorithm in Algorithm \ref{alg:Cross-Validation} provides such information.} On the other hand}, {when the underlying signal is itself sparse so that ${\mathcal L} = \mathcal{I}$ the identity matrix,} \eqref{eq:PosteriorDensitymaskl2}, which uses the support informed $\ell_2$ prior, can be viewed as alternative approach to SBL (using EM), \cite{tipping2001sparse, wipf2004sparse}, since both our new algorithm and SBL {incorporate the support locations} in the sparse domain to inform the prior. Additional contrasts and comparisons are provided in the introduction.
Finally we note that while using Gibbs sampling is generally more efficient, it is not consistent with employing the support informed prior which reduces the uncertainty of the solution. Moreover, our method will allow downstream processing to be more efficient, as we can more efficiently probe regions of interest.
{Finally, we point out that we can incorporate the MMV framework directly into the likelihood estimate that appears in each of the posterior estimates, \eqref{eq:final_posterior} and \eqref{eq:final_posteriorl2}. However, as we are only considering one data source given by (\ref{eq:ForwardModel}), this information is somewhat redundant. Furthermore, the added MMV information is already accounted for in determining the support of the underlying signal. In our numerical experiments we use $y = \frac{1}{J}\sum_1^J y_j$, i.e. the mean of the observations, in each posterior estimate.}
\section{Introduction}
\label{sec:intro}
This file is documentation for the SIAM \LaTeX\ style, including how
to typeset the main document, the {\scshape Bib}\TeX\xspace\ file, and any supplementary
material. More information
about SIAM's editorial style can be found in the style manual, available
at \url{https://www.siam.org/journals/pdf/stylemanual.pdf}.
The major changes in the SIAM standard class are summarized in \cref{sec:changes}.
The SIAM \LaTeX\@ files can be found at
\url{https://www.siam.org/journals/auth-info.php}. The files that
are distributed for the standard macros are given below.
\begin{itemize}
\item \texttt{siamart190516.cls} (required): Main SIAM standard \LaTeX\ class file.
\item \texttt{siamplain.bst} (required): Bibliographic style file for
{\scshape Bib}\TeX\xspace.
\item \texttt{docsiamart.tex}: Produces this documentation.
\item \texttt{references.bib}: {\scshape Bib}\TeX\xspace\ database for this
documentation and examples.
\item \texttt{ex\_article.tex}: Template for article.
\item \texttt{ex\_supplement.tex}: Template for supplement.
\item \texttt{ex\_shared.tex}: Template for shared information for
article and supplement.
\end{itemize}
To use these files, put \texttt{siamart190516.cls} and
\texttt{siamplain.bst} in the directory with your
paper or, alternatively, into your \LaTeX\@ and {\scshape Bib}\TeX\xspace\@ paths,
respectively.
The outline of a SIAM \LaTeX\ article is shown in
\cref{ex:outline}. Templates are provided and discussed in more detail
in \cref{sec:template}.
\begin{example}[label={ex:outline},listing only,%
listing options={style=siamlatex,{morekeywords=[1]{maketitle},
morekeywords=[2]{siamart190516}},}]%
{Document outline}
\documentclass{siamart190516}
\begin{document}
\maketitle
\end{document}
\end{example}
\section{Class options}
\label{sec:class-options}
Class options can be included in the bracketed argument of the
command, separated by commas. The possible class options are:
\begin{itemize}
\item \code{review} --- Recommended for submitting your manuscript to
a SIAM journal. Adds line numbers as well as the statement ``This
manuscript is for review purposes only'' to the bottom of each page.
\item \code{final} --- Turns off the black boxes that help authors
identify lines that are too long. The final published version will
have this option on.
\item \code{supplement} --- Specifies that the file is a supplement
and not the main document, causing changes in the appearance of the
title and numbering; see \cref{sec:supplement} for details.
\item \code{hidelinks} --- Turns off colors on hyperlinks; see
\cref{sec:cr+hyp}. The hyperlinks still exist, but there is no color
to differentiate them.
The final published version will have this option on.
\end{itemize}
\section{Front matter}
\label{sec:front}
The title and author parts are formatted using the standard
\code{\title}, \code{\author}, and \code{\maketitle} commands as
described in Lamport \cite{La86}. The title and author should be
declared in the preamble. The title and author names are automatically
converted to uppercase in the document.
If there is more than one author, each additional author should be preceded by
the \code{\and} command.
The addresses and support acknowledgments are added via
\code{\thanks}. Each author's thanks should specify their address.
The support acknowledgment should be put in the title thanks,
unless specific support needs to be specified for individual authors,
in which case it should follow the author address.
The header for this file was produced by the code in \cref{ex:header}, including an
example of a shared footnote. Each thanks produces a footnote, so the
footnote of the second author is \#3.
The command
\code{\headers{title}{authors}} command, with
the title (possibly shortened to fit) and the authors' names, creates
the page headers, automatically converted to uppercase.
\examplefile[label={ex:header},listing only,%
listing options={style=siamlatex,%
deletetexcs={and,thanks,title,author},%
{moretexcs=[2]{and,thanks,title,author,maketitle,headers,email}}}
]{Title and authors in preamble}{tmp_\jobname_header.tex}
\newpage
Following the author and title is the abstract, key words listing, and AMS subject
classifications, designated using the \code{abstract}, \code{keywords}, and \code{AMS}
environments.
Authors are responsible for providing AMS numbers which can be found
on the AMS web site \cite{AMSMSC2010}. The abstract, keywords, and AMS subject classifications
for this document are specified in \cref{ex:abstract}.
\examplefile[label={ex:abstract},%
before upper={\preamble{\bs newcommand\{\bs BibTeX\}\{\{\bs scshape Bib\}\bs TeX\bs xspace\}}},
listing only,%
listing options={style=siamlatex,%
{morekeywords=[2]{abstract,keywords,AMS}}}
]{Abstract, keywords, and AMS classifications}{tmp_\jobname_abstract.tex}
A more complete example, including a PDF supplement, that uses the
included files \texttt{ex\_article.tex}, \texttt{ex\_supplement.tex},
and \texttt{ex\_shared.tex} is discussed in
\cref{sec:template}. The example files can be used as a starting point
for producing a document.
\section{Cross references and hyperlinks}
\label{sec:cr+hyp}
SIAM now supports cross references and hyperlinks via the
\texttt{cleveref} and \texttt{hyperef} packages, which are loaded by
the class file.
\subsection{Cleveref}
\label{sec:cleveref}
SIAM strongly recommends using the commands provided by
the \texttt{cleveref} package for cross referencing.
The package is automatically loaded and already customized
to adhere to SIAM's style guidelines.
To create a cross reference, use the command \code{\cref} (inside
sentence) or \code{\Cref} (beginning of a sentence) in place of the
object name and \code{\ref}.
The \texttt{cleveref}
package enhances \LaTeX's cross-referencing features, allowing
the format of cross references to be determined automatically
according to the ``type" of cross reference (equation, section, etc.)
and the context in which the cross reference is used.
So, the package
\emph{automatically} inserts the object name as well as the
appropriate hyperlink; see \cref{ex:cref}.
It may require two \LaTeX\@ compilations for the references to show up
correctly.
Additional examples are shown in the sections below
for equations, tables, figures, sections, etc.
\begin{example}[label=ex:cref,bicolor,listing options={style=siamlatex,%
{morekeywords=[2]{cref,ref}}}]{Advantage of using cleveref}
The normal way to get a cross reference with a hyperlink requires a
lot of typing: \hyperref[thm:mvt]{Theorem~\ref*{thm:mvt}}.
The \texttt{cleveref} package gets both the name and hyperlink
automatically using a single macro: \cref{thm:mvt}.
It also handles multiple references with the same macro, such as
\cref{thm:mvt,fig:pgfplots,fig:testfig}.
\end{example}
\subsection{Hyperef}
\label{sec:hyperef}
Hyperlinks are created with the \code{\href} and \code{\url} commands,
as shown in \cref{ex:href}.
SIAM has also defined the \code{\email} command, as shown in
\cref{ex:header}.
You can hide links (i.e., turn off link colors) with the \code{hidelinks} option.
\begin{example}[label={ex:href},bicolor,%
listing options={style=siamlatex,%
{morekeywords=[2]{href,url}}}]{Creating hyperlinks}
The \href{https://www.siam.org}{SIAM homepage} has general information.
Note that the colored text will \emph{not} appear in the print version
nor will the hyperlink be active, so the writer may want to specify
the location explicitly instead by using \url{https://www.siam.org}.
\end{example}
Note that homepage links via \code{\url} in the \code{\thanks} environment require
special formatting for the tilde (\string~) character. The formatting
is used in the template and shown in \cref{ex:shared}.
\section{Math and equations}
\label{sec:math}
Here we show some example equations, with numbering, and examples of
referencing the equations. SIAM now includes the package
\texttt{amsmath} by default, and we include some of its features as
well, although the reader should consult the package user manual for
further guidance \cite{amsmath,shortmath}.
Several of the example are adapted
from Mittlebach and Goossen's guide to \LaTeX~\cite{MiGo04}.
\Cref{ex:textmath} is a straightforward example of inline mathematics
equations that does not use any special packages or features.
\begin{example}[label={ex:textmath},bicolor]{Inline math}
The following shows an example of math in text:
Let $S=[s_{ij}]$ ($1\leq i,j\leq n$) be a $(0,1,-1)$-matrix of order $n$.
\end{example}
In \cref{ex:bbm}, we show the recommended method for getting
blackboard fonts using the \texttt{amsfonts} package. This is not
loaded by default and must be included in the preamble.
\begin{example}[label={ex:bbm},bicolor,before upper={\preamble{\bs
usepackage\{amsfonts\}}},%
listing options={style=siamlatex,%
{morekeywords=[2]{mathbb}}}]{Blackboard math}
Blackboard bold characters, such as $\mathbb{C}$ and $\mathbb{R}$,
should be created with the \texttt{amsfonts} package, although this
is not included by default.
\end{example}
\Cref{ex:smallmatrix} shows the \code{smallmatrix} environment for an
inline matrix from the \texttt{amsmath} package, which is included by
default.
\begin{example}[label={ex:smallmatrix},bicolor,%
listing options={style=siamlatex,%
{morekeywords=[2]{smallmatrix}}}]{Inline matrix}
Matrices of no more than two rows appearing in text can be created
as shown in the next example:
$B = \bigl[ \begin{smallmatrix} B_{11} & B_{12} \\
B_{21} & B_{22} \end{smallmatrix} \bigr]$.
\end{example}
Bigger matrices can be rendered with environments from
the \texttt{amsmath} package, such as \code{bmatrix} and
\code{pmatrix} used in
\cref{ex:matrices}.
\begin{example}[label={ex:matrices},bicolor,%
listing options={style=siamlatex,%
{morekeywords=[2]{bmatrix,pmatrix}}}]{Creating matrices}
Display matrices can be rendered using environments from \texttt{amsmath}:
\begin{equation}\label{eq:matrices}
S=\begin{bmatrix}1&0\\0&0\end{bmatrix}
\quad\text{and}\quad
C=\begin{pmatrix}1&1&0\\1&1&0\\0&0&0\end{pmatrix}.
\end{equation}
\Cref{eq:matrices} shows some example matrices.
\end{example}
\newpage
\Cref{ex:dmo} shows how to use the \code{\DeclareMathOperator} command
from the \texttt{amsopn} package to declare the \code{\Range} macro.
(This example also uses the \texttt{braket} package for the
\code{\set} macro, but this is not necessarily recommended by SIAM.)
\begin{example}[label={ex:dmo},%
before upper={\preamble{\bs usepackage\{braket,amsfonts,amsopn\}}\\
\noindent\preamble{\bs DeclareMathOperator\{\bs Range\}\{Range\}}},%
bicolor,%
listing options={style=siamlatex,%
{moretexcs=[2]{Range}}}
]{Declaring math operators}
An example of a math operator:
\begin{equation}\label{eq:range}
\Range(A) = \set{ y \in \mathbb{R}^n | y = Ax }.
\end{equation}
\end{example}
\Cref{ex:foo} shows how to use the \code{align} environment from
\texttt{amsmath} to easily align multiple equations.
\begin{example}[label={ex:foo},bicolor,%
listing options={style=siamlatex,%
{morekeywords=[2]{align}}}]{Aligned equations}
\Cref{eq:a,eq:b,eq:c} show three aligned equations.
\begin{align}
f &= g, \label{eq:a} \\
f' &= g', \quad\text{and} \label{eq:b} \\
\mathcal{L}f &= \mathcal{L}g \label{eq:c}.
\end{align}
\end{example}
Another way to number a set of equations is the
\code{subequations} environment from \texttt{amsmath}, as shown in \cref{ex:aligned}.
\begin{example}[label={ex:aligned},bicolor,%
listing options={style=siamlatex,%
{morekeywords=[2]{subequations}}}]{Subequations}
We calculate the Fr\'{e}chet derivative of $F$ as follows:
\begin{subequations}
\begin{align}
F'(U,V)(H,K)
&= \langle R(U,V),H\Sigma V^{T} + U\Sigma K^{T} -
P(H\Sigma V^{T} + U\Sigma K^{T})\rangle \label{eq:aa} \\
&= \langle R(U,V),H\Sigma V^{T} + U\Sigma K^{T}\rangle
\nonumber \\
&= \langle R(U,V)V\Sigma^{T},H\rangle +
\langle \Sigma^{T}U^{T}R(U,V),K^{T}\rangle. \label{eq:bb}
\end{align}
\end{subequations}
\Cref{eq:aa} is the first line, and \cref{eq:bb} is the last line.
\end{example}
~
For an equation split over multiple lines, \cref{ex:ml} shows the
usage of the \code{multline} environment provided by \texttt{amsmath}.
~
\begin{example}[label={ex:ml},bicolor,%
listing options={style=siamlatex,%
{morekeywords=[2]{multline}}}]{Equation split across lines}
We claim that the projection $g(U,V)$ is given by the pair of matrices:
\begin{multline} \label{eq:ml}
g(U,V) = \biggl( \frac{R(U,V)V\Sigma^{T}U^{T}
- U\Sigma V^{T}R(U,V)^{T}}{2}U,\\
\frac{R(U,V)^{T}U\Sigma V^{T}-V \Sigma^{T}U^{T}R(U,V)}{2}V \biggr).
\end{multline}
\end{example}
\section{Theorem-like environments}
\label{sec:thm}
SIAM loads \texttt{ntheorem} package and uses it to define the following
theorem-like environments:
\code{theorem},
\code{lemma},
\code{corollary},
\code{definition}, and
\code{proposition}.
SIAM also defines a \code{proof} environment that automatically
inserts the symbol ``$\,\proofbox\,$'' at the end of any proof, even if it ends in an
equation environment. \emph{Note that the document may need to be
compiled twice for the mark to appear.}
Some of the calculus examples were adapted from \cite{CalcI}.
\Cref{ex:theorem} shows usage of the \code{theorem} environment.
An optional argument can be used to name the theorem.
\Cref{ex:cor} illustrates a corollary, without a name, and the
proof environment.
~
\begin{example}[label=ex:theorem,bicolor,parbox=false,%
listing options={style=siamlatex,%
{morekeywords=[2]{theorem}}}]{Theorem}
\begin{theorem}[Mean Value Theorem]\label{thm:mvt}
Suppose $f$ is a function that is continuous on the closed interval
$[a,b]$. and differentiable on the open interval $(a,b)$.
Then there exists a number $c$ such that $a < c < b$ and
\begin{displaymath}
f'(c) = \frac{f(b)-f(a)}{b-a}.
\end{displaymath}
In other words, $f(b)-f(a) = f'(c)(b-a)$.
\end{theorem}
\end{example}
\begin{example}[label=ex:cor,bicolor,parbox=false,%
listing options={style=siamlatex,%
{morekeywords=[2]{corollary,proof}}}]%
{Corollary and proof}
\begin{corollary}
Let $f(x)$ be continuous and differentiable everywhere. If $f(x)$
has at least two roots, then $f'(x)$ must have at least one root.
\end{corollary}
\begin{proof}
Let $a$ and $b$ be two distinct roots of $f$.
By \cref{thm:mvt}, there exists a number $c$ such that
\begin{displaymath}
f'(c) = \frac{f(b)-f(a)}{b-a} = \frac{0-0}{b-a} = 0.
\end{displaymath}
\end{proof}
\end{example}
SIAM also defines commands to create your own theorem- and
remark-like environments:
\begin{itemize}
\item \code{newsiamthm} --- Small caps header, italized body.
\item \code{newsiamremark} --- Italics header, roman body.
\end{itemize}
Each command takes two arguments. The first is the environment name,
and the second is the name to show in the document. These commands should
be used instead of \code{\newtheorem}.
\Cref{ex:claim,ex:ref} shows how to use the commands above, including how to
specify the plural version for \texttt{cleveref} if it is unusual.
\begin{example}[label=ex:claim,bicolor,%
before upper={\preamble{\bs newsiamthm\{claim\}\{Claim\}}\\
\noindent\preamble{\bs newsiamremark\{hypothesis\}\{Hypothesis\}}\\
\noindent\preamble{\bs crefname\{hypothesis\}\{Hypothesis\}\{Hypotheses\}}},%
parbox=false,%
listing options={style=siamlatex,%
{morekeywords=[2]{claim,proof,hypothesis}}}]{New theorem-like environment}
\begin{claim}\label{cl:constant}
If $f'(x) = 0$ for all $x \in (a,b)$ then $f(x)$ is constant on $(a,b)$.
\end{claim}
\begin{hypothesis}\label{hyp1}
The function $f$ is continuously differentiable.
\end{hypothesis}
\begin{hypothesis}\label{hyp2}
The random variable is normally distributed.
\end{hypothesis}
\end{example}
\begin{example}[label=ex:ref,bicolor,listing options={style=siamlatex,%
{morekeywords=[2]{cref}}}]{References}
We can reference multiple types of objects with a single reference:
\cref{cl:constant,thm:mvt,hyp1,hyp2}.
\end{example}
\section{Tables}
\label{sec:tab}
Table captions should go above the tables.
\Cref{ex:simpletable} shows the code to generate a
\cref{tab:simpletable}.
A more complicated example is shown in
\cref{ex:table}, which generates \cref{tab:KoMa14}. This
example uses subfloats via the \texttt{subfig} package, as well as
special column options from the \texttt{array} package.
\begin{tcbverbatimwrite}{tmp_\jobname_simpletable.tex}
\begin{table}[tbhp]
{\footnotesize
\caption{Example table}\label{tab:simpletable}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
Species & \bf Mean & \bf Std.~Dev. \\ \hline
1 & 3.4 & 1.2 \\
2 & 5.4 & 0.6 \\ \hline
\end{tabular}
\end{center}
}
\end{table}
\end{tcbverbatimwrite}
\examplefile[label={ex:simpletable},%
listing only, listing options={style=siamlatex}]%
{Example table.}{tmp_\jobname_simpletable.tex}
\input{tmp_\jobname_simpletable.tex}
\begin{tcbverbatimwrite}{tmp_\jobname_table.tex}
\newcolumntype{R}{>{$}r<{$}} %
\newcolumntype{V}[1]{>{[\;}*{#1}{R@{\;\;}}R<{\;]}} %
\begin{table}[tbhp]
{\footnotesize
\captionsetup{position=top}
\caption{Example table adapted from Kolda and Mayo \rm{\cite{KoMa14}}.}\label{tab:KoMa14}
\begin{center}
\subfloat[$\beta=1$]{
\begin{tabular}{|r|R|V{3}|c|r@{\,$\pm$\,}l|} \hline
occ. & \multicolumn{1}{c|}{$\lambda$} & \multicolumn{4}{c|}{$\mathbf{x}$} &
fevals & \multicolumn{2}{c|}{time (sec.)}\\ \hline
718 & 11.3476 & 0.5544 & 0.3155 & 1.2018 & 0.0977 & 45 & 0.17 & 0.06 \\ \hline
134 & 3.7394 & 0.2642 & -1.1056 & 0.2657 & -0.3160 & 31 & 0.12 & 0.05 \\ \hline
4 & \multicolumn{6}{c|}{\emph{--- Failed to converge ---}} & 0.21 & 0.10 \\ \hline
\end{tabular}}
\subfloat[$\beta=-1$]{
\begin{tabular}{|r|R|V{3}|c|r@{\,$\pm$\,}l|} \hline
occ. & \multicolumn{1}{c|}{$\lambda$} & \multicolumn{4}{c|}{$\mathbf{x}$} &
fevals & \multicolumn{2}{c|}{time (sec.)}\\ \hline
72 & -1.1507 & 0.2291 & 0.6444 & 0.3540 & -0.8990 & 34 & 0.14 & 0.06 \\ \hline
624 & -6.3985 & 0.1003 & 0.1840 & 0.5305 & 1.2438 & 48 & 0.19 & 0.08 \\ \hline
2 & \multicolumn{6}{c|}{\emph{--- Failed to converge ---}} & 0.23 & 0.02 \\ \hline
\end{tabular}}
\end{center}
}
\end{table}
\end{tcbverbatimwrite}
\examplefile[label={ex:table},%
before upper={\preamble[\scriptsize]{\bs usepackage\{array\}}\\[-0.4em]
\noindent\preamble[\scriptsize]{\bs usepackage[caption=false]\{subfig\}}},%
listing only, listing options={%
style=siamlatex,basicstyle=\ttfamily\scriptsize}]%
{Example table with subtables.}{tmp_\jobname_table.tex}
\input{tmp_\jobname_table.tex}
\section{Figures}
\label{sec:fig}
It is recommended that all figures be generated in high resolution.
In the past, SIAM has required encapsulated
postscript (EPS) format for final production. This is still an
acceptable format, but SIAM also now allows high-resolution
PDF, JPEG, and PNG figures.
If working with EPS images and using \texttt{pdflatex}, we
recommend the package \texttt{epstopdf} to automatically convert EPS
images to PDF for inclusion in PDF documents created by
\texttt{pdflatex}.
\Cref{ex:fig} shows the code to generate \cref{fig:testfig}. This
example uses the \texttt{graphicx} package for the
\code{\includegraphics} command.
\begin{tcbverbatimwrite}{tmp_\jobname_fig.tex}
\begin{figure}[tbhp]
\centering
\subfloat[$\epsilon_{\max}=5$]{\label{fig:a}\includegraphics{lexample_fig1}}
\subfloat[$\epsilon_{\max}=0.5$]{\label{fig:b}\includegraphics{lexample_fig2}}
\caption{Example figure using external image files.}
\label{fig:testfig}
\end{figure}
\end{tcbverbatimwrite}
\examplefile[label={ex:fig},%
before upper={\preamble[\scriptsize]{\bs usepackage\{graphicx,epstopdf\}}\\[-0.4em]
\noindent\preamble[\scriptsize]{\bs usepackage[caption=false]\{subfig\}}},%
listing only, listing options={%
style=siamlatex,basicstyle=\ttfamily\scriptsize}]%
{Example figure with subfigures and external files}{tmp_\jobname_fig.tex}
\input{tmp_\jobname_fig.tex}
Another option for figures is a graphics-generator that is platform- and
format-independent.
PGF is a TeX macro package for generating such graphics and works together with the most important TeX backend drivers, including pdftex and dvips.
The user-friedly syntax layer called TikZ.
Here we show an example using \texttt{PGFPLOTS}, useful for drawing high-quality
plots directly in \LaTeX.
\Cref{ex:data} and \cref{ex:pgfplots} shows the data and code, respectively, to generate \cref{fig:pgfplots}, adapted from \cite{pgfplots}.
\examplefile[label={ex:data},listing only,
listing options={style=siamlatex,basicstyle=\ttfamily\scriptsize}]%
{Example data file (data.dat)}{data.dat}
\begin{tcbverbatimwrite}{tmp_\jobname_tikz.tex}
\begin{figure}[tbhp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[height=2.75in, grid=major,
xlabel={Degrees of Freedom}, ylabel={$L_2$ Error},
legend entries={$d=2$,$d=3$}]
\addplot table [x=d2_dof,y=d2_l2_err] {data.dat};
\addplot table [x=d3_dof,y=d3_l2_err] {data.dat};
\end{loglogaxis}
\end{tikzpicture}
\caption{Example \texttt{PGFPLOTS} figure.}
\label{fig:pgfplots}
\end{figure}
\end{tcbverbatimwrite}
\examplefile[label={ex:pgfplots},%
before upper={\preamble[\scriptsize]{\bs usepackage\{pgfplots\}}},%
listing only, listing options={%
style=siamlatex}]%
{Example TikZ/PGF for platform-independent graphics.}{tmp_\jobname_tikz.tex}
\input{tmp_\jobname_tikz.tex}
\section{Algorithms}
\label{sec:algs}
SIAM automatically includes the \texttt{algorithm} package in the
class definition. This provides the float environment.
Users have the choice of \texttt{algpseudocode},
\texttt{algorithmic}, and other packages for actually formatting the algorithm.
For example, \cref{alg:buildtree} is produced by the code in \cref{ex:alg}.
In order to reference lines within the algorithm, we need to tell the
\texttt{cleveref} package how to do the referencing, which is the
second line of \cref{ex:alg}. Then we can use the code
\code{\cref{line3}} to produce \cref{line3}.
\begin{tcbverbatimwrite}{tmp_\jobname_alg.tex}
\begin{algorithm}
\caption{Build tree}
\label{alg:buildtree}
\begin{algorithmic}[1]
\STATE{Define $P:=T:=\{ \{1\},\ldots,\{d\}$\}}
\WHILE{$\#P > 1$}
\STATE\label{line3}{Choose $C^\prime\in\mathcal{C}_p(P)$ with $C^\prime := \operatorname{argmin}_{C\in\mathcal{C}_p(P)} \varrho(C)$}
\STATE{Find an optimal partition tree $T_{C^\prime}$ }
\STATE{Update $P := (P{\setminus} C^\prime) \cup \{ \bigcup_{t\in C^\prime} t \}$}
\STATE{Update $T := T \cup \{ \bigcup_{t\in\tau} t : \tau\in T_{C^\prime}{\setminus} \mathcal{L}(T_{C^\prime})\}$}
\ENDWHILE
\RETURN $T$
\end{algorithmic}
\end{algorithm}
\end{tcbverbatimwrite}
\examplefile[float=htpb,label={ex:alg},%
before upper={\preamble[\scriptsize]{\bs usepackage\{algorithmic\}}\\[-0.4em]
\preamble[\scriptsize]{\bs Crefname\{ALC@unique\}\{Line\}\{Lines\}}},%
listing only, listing options={%
style=siamlatex,basicstyle=\ttfamily\scriptsize}]%
{Example algorithm}{tmp_\jobname_alg.tex}
\input{tmp_\jobname_alg.tex}
\section{Sections}
\label{sec:sec}
Sections are denoted using standard \LaTeX\ section commands, i.e.,
\code{\section}, \code{\subsection}, etc.
If you wish to end the section title with something other that a period (the default), you have to add the command \code{\nopunct} at the end of the title.
Appendices are created with the normal sectioning commands, following
the command \code{
\subsection{Autocorrelation Functions}
We also analyze the convergence of our MCMC chains to steady state through the use of autocorrelations, trace plots and acceptance ratios.
Trace plots show the state of the chain over each iteration.
Theoretically, if the Markov chain has converged to the posterior density, the states of the chain should not be correlated. {Therefore, the chain is said to be {\em well-mixed} if the trace plot exhibits behavior indicating that the states are independently and identically distributed (iid). On the other hand,} any portion of the Markov chain having high auto-correlation between states would imply that the Markov chain has not {(yet)} converged to the stationary distribution. {As another indicator of mixing}, we use the Autocorrelation Function (ACF) package in Matlab, \cite{acf_matlab}, to compute the correlogram of each Markov Chain recovered through the proposed MCMC sampling techniques
The acceptance ratio {determined from line 9} in Algorithm \ref{alg:MHAlgorithm} is the number of times $x^{cand}$ is accepted as the next sample relative to the total number of iterations. Therefore, a large acceptance ratio {(typically $\alpha > .8$)} may indicate the Markov chain is stuck in a certain region of the posterior probability density and has not explored the entire support of the density. In contrast, a small acceptance ratio {(e.g.~$\alpha < .2$)} may indicate that the chain may be moving too quickly within the domain of the {proposal} density. Further, with a low acceptance ratio, it remains possible that the chain does not explore certain isolated modes of the posterior probability density \cite{MCMCR}. While the ideal acceptance ratio remains an open question for the Metropolis-Hastings algorithm, as a general rule of thumb, it should be close to $.5$ for a one-dimensional problem \cite{MCMCR}. This metric was primarily derived for the Gaussian proposal distribution, which is also used in this paper.
\subsection{Sparse Signal Recovery in One Dimension}
\input{sparsesignal}
\subsection{Signal Recovery in One Dimension}
We now consider reconstructing the signal defined as
\begin{example}\label{ex:true_1D}
\begin{equation*}
x(s) = \begin{cases}
40, &\quad 0.1 \leq s \leq 0.25 \\
10, &\quad 0.35 \leq s \leq 0.325 \\
20\sqrt{2 \pi}e^{-\left(\frac{s-0.75}{0.05}\right)^2}, &\quad s > 0.5 \\
0, & \quad \mathrm{otherwise,}
\end{cases}
\end{equation*}
\end{example}
where $s \in [0,1]$ is the domain of interest. Similar to {the example used in} \cite{BardsleyUQ}, the domain is uniformly discretized so that $s_i = \frac{i}{n}$, $i = 1,\dots,n$ with $n = 80$. For all of our experiments in one dimension, we simulate $J=20$ measurements. {We choose $20$ since it was shown to be a sufficient number of measurements for reconstructing a signal using the VBJS method in \cite{VBJS}.} We also choose $K=20$ and $M=10$ for Algorithm \ref{alg:Cross-Validation}.
Because we are {ultimately} interested in imaging applications such as MRI, SAR and ultrasound, in our experiments we assume that the observable data {$y$} consists of multiple, noisy Fourier measurements of $x$. That is, each measurement is collected according to (\ref{eq:ForwardModel}), with $A$ defined as the discrete Fourier transform matrix and $\sigma >0$. {We point out that} the discrete Fourier transform matrix is better conditioned than the blurring model used in \cite{BardsleyUQ}. Here we allow for the measurements to be collected with significantly higher noise levels, i.e., $\sigma >>0$ {or SNR $<<1$}, however. {In what follows we show how results for $SNR = 5$ dB and $SNR \approx -.2$ dB.}
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{recons_sparse_edge_final.pdf}
\caption{Approximate reconstructions (\ref{eq:final_approx}) of Example \ref{ex:true_1D} displayed along with credibility intervals $C_{0.05}$ when using (left) the Laplace prior (\ref{eq:PosteriorDensitylaplace}) (middle) the proposed support informed sparsity prior (\ref{eq:PosteriorDensitymask}) and (right) the proposed support informed prior described in (\ref{eq:PosteriorDensitymaskl2}). {(top) $\sigma \approx 5.5$ ($SNR = 5$ dB), (bottom) $\sigma = 10$ ($SNR \approx -.2$ dB).}}
\label{fig:recons}
\end{figure}
Figure \ref{fig:recons} displays the mean of the Markov chain (\ref{eq:final_approx}) recovered on each grid point in the domain when sampling the posteriors described in (\ref{eq:final_posterior}) with (left) the Laplace prior (\ref{eq:laplace_prior}) and (right) the proposed support informed sparsity prior (\ref{eq:SISP_prior}) for {(top) $\sigma \approx 5.5$ ($SNR = 5$ dB) and (bottom) $\sigma = 10$ ($SNR \approx -.2$ dB)}. Along with the mean, in Figure \ref{fig:recons} we also display $C_{0.05}$ for each reconstruction. Notice that the proposed support informed sparsity prior results in tighter credibility intervals and more accurate reconstructions, {especially in the $\sigma = 10$ case}.
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{distributions.pdf}
\caption{{Estimated posterior distributions at the points in the domain indicated by the red crosses in Figure \ref{fig:mask_demo}. Table \ref{tab:distros} reports each distribution and its corresponding parameters. (top) $\sigma \approx 5.5$ ($SNR = 5$ dB); (bottom) $\sigma = 10$ ($SNR \approx -.2$ dB).}}
\label{fig:distros}
\end{figure}
Figure \ref{fig:distros} shows the resulting distributions at the points in the domain indicated by the red crosses in Figure \ref{fig:mask_demo}. Each distribution is estimated given all samples of the Markov chain at the desired point in the domain after discarding the first $B$ states. Distributions are optimized based on the Pearson Chi-Squared goodness of fit metric and the mean-squared error when compared to the histogram. The resulting distributions and their corresponding parameters are displayed in Table \ref{tab:distros}. Consistent with the remainder of our results, we see that the distributions of the support informed $\ell_1$ prior at edge locations is more closely centered around the true function values. In smooth regions, we see compelling evidence to use the support informed $\ell_2$ prior. This motivates future work into exploring the use of different priors in different regions of the domain.
\begin{table}[hbt]
\begin{center}
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Posterior & Noise & Location 1 & Location 2 & Location 3 & Location 4 \\ \hline
\multirow{2}{*}{(\ref{eq:PosteriorDensitylaplace})} & \multirow{2}{*}{ 5 dB} & Beta & $\chi^2$ & Normal & Log-Normal \\
& & (12.45,11.94) & (247.60) & (0.46, 1.50) & (0.01) \\ \hline
\multirow{2}{*}{(\ref{eq:PosteriorDensitymask})} & \multirow{2}{*}{ 5 dB} & $\chi^2$ & Pearson type III & Normal & Log-Normal \\
& & (95.34) & (-0.07) & (-1,1.29) & (0.06) \\ \hline
\multirow{2}{*}{(\ref{eq:PosteriorDensitymaskl2})} & \multirow{2}{*}{ 5 dB} & Beta & Beta & Normal & Normal \\
& & (22.15,18.08) & ($5.13\times 10^6$, 196.99) & (0.15, 1.61) & (48.48, 1.53) \\ \hline
\multirow{2}{*}{(\ref{eq:PosteriorDensitylaplace})} & \multirow{2}{*}{ -.2 dB} & Beta &Weibull (min) & Beta & Pearson type III \\
& & (23.53,34.94) & (3.61) & (399.66,$1.15\times 10^7$) & (-0.11) \\ \hline
\multirow{2}{*}{(\ref{eq:PosteriorDensitymask})} & \multirow{2}{*}{ -.2 dB} & Log-Normal & Weibull (max) & Log-Normal & Weibull (min) \\
& & (0.03) & (3.32) & (0.01) & (3.2) \\ \hline
\multirow{2}{*}{(\ref{eq:PosteriorDensitymaskl2})} & \multirow{2}{*}{ -.2 dB} & Beta & Beta & Weibull (max) & Beta \\
& & (325.45,$3.32\times 10^3$) & (16.56,19.12) & (3.9) & (48.37,44.73) \\ \hline
\end{tabular}}
\end{center}
\caption{{The distributions and corresponding parameters resulting from a best-fit analysis of the last $M-B$ Markov chain samples (\ref{eq:MarkovMatrix}) at the locations in the domain indicated by the red crosses seen in Figure \ref{fig:mask_demo}. The distributions along with the histogram of samples are visualized in Figure \ref{fig:distros}.} }
\label{tab:distros}
\end{table}
\begin{remark}
We do not report on results using the posterior given by \eqref{eq:PosteriorDensityhierarchical} for Example \ref{ex:true_1D} because the MH algorithm {as implemented in Algorithm \ref{alg:MHAlgorithm}} did not converge in this case. We speculate that the (unmasked) $\ell_2$ norm in the prior is prohibitively large for numerical computation. {It may be possible to calculate \eqref{eq:acceptance} so as to avoid this issue, for example by using logarithms.} Alternatively, one could use a Gibbs sampler with an appropriate SPD operator to replace ${\mathcal L}$ in \eqref{eq:PosteriorDensityhierarchical}, see for example the discussion in \cite{BardsleyUQ}. {Since our previous observations in the sparse signal case suggest that the resulting confidence interval would still be larger than those displayed in Figure \ref{fig:recons}, and as comparing different sampling algorithms is not the focus of this investigation, we give no further consideration to \eqref{eq:PosteriorDensityhierarchical}.}
\end{remark}
\begin{figure}[htb]
\centering
\includegraphics[width=.9\textwidth]{ar_final.pdf}
\caption{Acceptance ratios resulting from sampling the posterior defined using the Laplace prior (\ref{eq:PosteriorDensitylaplace}), support informed sparsity prior (\ref{eq:PosteriorDensitymask}) and support informed $\ell_2$ prior (\ref{eq:PosteriorDensitymaskl2}). (left) $\sigma \approx 5.5$ ($SNR = 5$ dB); (right) $\sigma = 10$ ($SNR \approx -.2$ dB).}
\label{fig:ar}
\end{figure}
We now evaluate the mixing and convergence of the MCMC chains. Figure \ref{fig:ar} displays the acceptance ratio across all MH iterations. With our strategy to adjust the variance of the proposal distribution, we see that each chain exhibits an acceptable acceptance ratio, with the acceptance ratio for the support informed sparsity prior tending closer to the desirable $.5$ value. We note that we did not try to optimize $\sigma_q$ in Algorithm \ref{alg:MHAlgorithm}, and that its initial tuning as described in the paragraph below Algorithm \ref{alg:MHAlgorithm} was done in the same way for each choice of posterior.} Figure \ref{fig:acf_lines} displays the correlograms associated with sampling the posterior distributions at each grid point marked with a red cross in Figure \ref{fig:mask_demo}. {The results affirm that prior information about the support locations in the sparse domain, which is achievable $SNR = 5$ dB, does indeed reduce the uncertainty.} When comparing the correlograms resulting from {the posteriors in \eqref{eq:PosteriorDensitylaplace}, \eqref{eq:PosteriorDensitymask}, and \eqref{eq:PosteriorDensitymaskl2} for lower SNR, it is evident that the proposed support informed priors yield less or similarly correlated samples near discontinuities than the more standard Laplace prior does.} Consistent with the results in the sparse signal case, the Laplace prior demonstrates earlier mixing in smooth regions than the support informed sparse prior, but {the mixing is comparable to that achieved using} the support informed $\ell_2$ prior. {These results, confirmed in Figure \ref{fig:abs_error},} suggest that as the SNR decreases a better result overall may be obtained when the choice of posterior is itself domain dependent; that is, when the support informed $\ell_2$ prior and the support informed sparse prior are adopted in different regions once the support in the sparse domain has been approximated. This will be discussed in future investigations, as such a ``splitting'' may also yield more efficient computation.
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{acf_lines.pdf}
\caption{The autocorrelation function (ACF) calculated for various lag lengths at the points in the domain indicated by the red crosses in Figure \ref{fig:mask_demo}. Samples are taken from the posterior distributions defined using the Laplace prior (\ref{eq:PosteriorDensitylaplace}), support informed sparsity prior (\ref{eq:PosteriorDensitymask}) and support informed $\ell_2$ prior (\ref{eq:PosteriorDensitymaskl2}). {(top) $\sigma \approx 5.5$ ($SNR = 5$ dB), (bottom) $\sigma = 10$ ($SNR \approx -.2$ dB).}}
\label{fig:acf_lines}
\end{figure}
Figure \ref{fig:trace} plots the value of each sample drawn from the posterior distribution {determined via Algorithm \ref{alg:MHAlgorithm}} described by (top) (\ref{eq:PosteriorDensitylaplace}), (middle) (\ref{eq:PosteriorDensitymask}), and (bottom) \eqref{eq:PosteriorDensitymaskl2} at the values in the domain indicated by the red crosses in Figure \ref{fig:mask_demo} {for $SNR = 5$ dB ($\sigma \approx 5.5$).} {While each trace plot yields similar behavior, it is evident that at all points considered, using the posterior defined with the support informed sparsity prior (\ref{eq:PosteriorDensitymask}) results in samples that are more uniformly distributed about the true function value, indicated by the horizontal red line.} {We note that comparable results were obtained for $\sigma = 10$ ($SNR \approx -.2$ dB) and are not displayed here.}
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{trace.pdf}
\caption{Trace plots calculated at the points described by the red crosses in Figure \ref{fig:mask_demo}. Samples are taken from the posterior distributions defined using the (top) Laplace prior (\ref{eq:PosteriorDensitylaplace}); (second row) support informed sparsity prior (\ref{eq:PosteriorDensitymask}); and (bottom) support informed $\ell_2$ prior (\ref{eq:PosteriorDensitymaskl2}). Here $\sigma = 10$ ($SNR \approx -.2$ dB). {The trace plots display similar behavior when $\sigma \approx 5.5$ ($SNR = 5$ dB).}}
\label{fig:trace}
\end{figure}
{As was done for the sparse signal recovery case,} we also evaluate the accuracy of the reconstructions (\ref{eq:final_approx}) for Example \ref{ex:true_1D} given {discrete Fourier} measurements acquired with increasing levels of complex additive Gaussian white noise in (\ref{eq:ForwardModel}). Figure \ref{fig:rel_error} displays the relative error, \textcolor{red}{\eqref{eq:rel_error}},
and Figure \ref{fig:abs_error} displays the absolute error, $|x_i-x^{true}_i|$, at the points in the spatial domain described by the red crosses in Figure \ref{fig:mask_demo}. {Each point $x_i$ across the domain is computed by averaging the results over 20 independent MCMC trials.}
{Once again we see that} the support informed priors consistently yield either better or comparable results for all levels of SNR. {This is true in smooth regions as well as near discontinuities, in which case it is clear that using the Laplace prior in regions where the underlying signal is {\em not} sparse in the presumably sparse domain yields dramatically worse results.}
\begin{figure}[htb]
\centering
\includegraphics[width=.5\textwidth]{rel_error_runs_20.png}
\caption{The relative reconstruction error (\ref{eq:rel_error}) calculated across the entire spatial domain for decreasing SNR (\ref{eq:SNR}).}
\label{fig:rel_error}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=\textwidth]{error_1D_SNR.pdf}
\caption{Absolute error comparison for decreasing SNR (\ref{eq:SNR}) at each of the red crosses in Figure \ref{fig:mask_demo}}
\label{fig:abs_error}
\end{figure}
\subsection{Point Estimation}
\label{sec:PointEstimation}
\label{sec:map}
{While the main goal of this investigation is to draw samples of the full posterior, it is instructive to first demonstrate how the MAP point estimate is used to recover a mode of the unknown distribution, \cite{murphy2012machine}.}
For any realization of \eqref{eq:ForwardModel}, $y_j \sim \mathbf{Y}$, $j = 1,\dots,J$, the MAP estimate for either posterior in \eqref{eq:final_posterior} or \eqref{eq:final_posteriorl2} is computed as
\begin{equation}
\label{eq:MAP1}
x_{MAP}^j \equiv \argmax{x \in \mathbb{R}^n} \left\{\log \left[ \hat{f}_{\mathbf{X}, \bm{\lambda}|\mathbf{Y}}(x, \hat{\lambda}|y_j)\right]\right\}, \quad j = 1,\dots,J.
\end{equation}
A corresponding convex optimization problem is then easily derived for the posteriors in \eqref{eq:final_posterior} or \eqref{eq:final_posteriorl2} as
\begin{align}
x_{MAP}^j &=
\argmax{x \in \mathbb{R}^n} \Big\{ \log(C) - \Big( \frac{\hat{\lambda}}{\nu} ||\mathcal{T}x||_\nu^\nu + \frac{1}{2 \sigma^2}||y_j-Ax||_2^2 \Big) \Big\} \nonumber \\
&= \argmin{x \in \mathbb{R}^n} \Big\{ \frac{\hat{\lambda}}{\nu} ||\mathcal{T}x||_\nu^\nu + \frac{1}{2 \sigma^2} ||y_j-Ax||_2^2 \Big\}, \quad j = 1,\dots,J, \label{eq:MAPEstimate}
\end{align}
{where $\nu = 1$ or $2$, $\mathcal{T} = \mathcal{L}$ or $\mathcal{ML}$, and $C = C_L$ or $C_M$.}
Since the MAP estimation can be reduced to an optimization problem, \eqref{eq:MAPEstimate} is often understood from a non-Bayesian perspective. {In particular, note that the MAP estimate used in Step 7 of Algorithm \ref{alg:Cross-Validation} is similarly constructed, with $\hat{\lambda}$ given by $\tilde{\lambda}$ in Step 6, $\mathcal{T} = \mathcal{L}$, and $\nu = 1$. We will also use the average MAP estimate given by (\ref{eq:MAPEstimate}) over the $J$ measurement vectors to set the initial state $x^0$ for the Metropolis-Hastings Algorithm given in Algorithm \ref{alg:MHAlgorithm}. Finally, although not discussed in this current investigation, we point out that there may be situations for which there are regions in a signal domain where efficiency is more important than quantifying uncertainty. In such cases using \eqref{eq:MAPEstimate} to construct an approximation within these regions may be more appropriate.}
\subsection{{Drawing} samples of the posterior distribution}
\label{sec:IntervalEstimation}
We seek to sample the posterior probability distribution of $\mathbf{X}$ described by (\ref{eq:final_posterior}) or \eqref{eq:final_posteriorl2} for the model given by \eqref{eq:ForwardModel}. This allows us both to quantify the uncertainty in our signal recovery as well as to discern modes of the posterior distribution that are unavailable in a MAP estimate. By Bayes' rule, \eqref{eq:full_posterior}, it is clear that {sampling} $\mathbf{X}$ in \eqref{eq:ForwardModel} from observations in \eqref{eq:PointEstimate} requires an approximation to the probability density $f_{\mathbf{Y}}(y)$, which typically does not have a closed form solution. Moreover, numerical quadrature rules are impractical whenever the dimension of the sample space $\mathbb{R}^n$ is high \cite{kaipio2006statistical}. We therefore choose to use the Markov Chain Monte Carlo (MCMC) method which is designed to construct an ergodic Markov chain that converges to the stationary probability density ${f}_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x,\lambda|y)$ in \eqref{eq:final_posterior}, which is proportional to the posterior we seek. Overviews of MCMC may be found in \cite{MCMC,MCMCR, kaipio2006statistical}. Because our {approach} does not assume conjugate priors, the posterior density (\ref{eq:full_posterior}) we derive does not represent a known posterior distribution. Hence we employ the particular MCMC method known as the Metropolis-Hastings (MH) algorithm \cite{hastings, metropolis}, {which we briefly describe below}
\subsubsection{The Metropolis Hastings (MH) Algorithm}
\label{sec:MH}
In the MH algorithm, a Markov chain in ${\Omega}$ is constructed so that it converges to the conditional posterior density ${f}_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x,\lambda|y)$ in \eqref{eq:full_posterior}.
The MH algorithm first defines a \emph{proposal distribution} $q: \Omega \times \Omega \rightarrow [0, 1]$ such that when the Markov chain is at state $x^{k-1}$, it proposes a \emph{candidate} state $x^{cand} \sim q$. Its acceptance of the next state, $x^k$, is determined by an \emph{acceptance probability} $\alpha: \Omega \times \Omega \rightarrow [0,1]$ given by
\begin{align}
\alpha(x^{cand}|x^{k-1})
&= \min \Bigg\{ 1, \frac{q(x^{k-1}|x^{cand}){f}_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x^{cand},\lambda|y)}{q(x^{cand}|x^{k-1}){f}_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x^{k-1},\lambda|y)} \Bigg\}\nonumber\\
&{\approx} \min \Bigg\{ 1, \frac{q(x^{k-1}|x^{cand})\hat{f}_{\mathbf{X},\bm{{{\lambda}}}|\mathbf{Y}}(x^{cand},{\hat{\lambda}}|y)}{q(x^{cand}|x^{k-1})\hat{f}_{\mathbf{X},\bm{{{\lambda}}}|\mathbf{Y}}(x^{k-1},{\hat{\lambda}}|y)} \Bigg\},
\label{eq:acceptance}
\end{align}
where {$\hat{f}_{\mathbf{X},{\bm{\lambda}}|\mathbf{Y}}$ is defined in \eqref{eq:final_posterior}}. In this way the MH algorithm is able to evaluate the full posterior (\ref{eq:full_posterior}) without needing to calculate the evidence $f_{\mathbf{Y}}(y)$ in (\ref{eq:MarginalDensity}). The general MH algorithm is summarized in Algorithm \ref{alg:MHAlgorithm}.
\begin{algorithm}[H]
\caption{The Metropolis-Hastings Algorithm}
\label{alg:MHAlgorithm}
\begin{algorithmic}[1]
\STATE{\textbf{Input:} The initial state $x^0$, {the choice of proposal distribution $q$, and its corresponding variance $\sigma_q$.}}
\STATE{\textbf{Output:} Markov chain $\{x^1, x^2, x^3, \ldots, x^M\}$.}
\STATE{Initialize $x^0$ and $k = 1$.}
\WHILE{$k\leq M$}
\STATE{Propose $x^{cand} \sim q(x^k|x^{k-1})$.}
\STATE{Compute $\alpha(x^{cand}|x^{k-1})$ using (\ref{eq:acceptance}) or (\ref{symmetricproposal})}.
\STATE{Sample $u \sim \mathcal{U}[0, 1]$.}
\IF{$u < \alpha$}
\STATE{Accept the candidate: set $x^k = x^{cand}$.}
\ELSE
\STATE{Reject the candidate: set $x^k = x^{k-1}$.}
\ENDIF
\ENDWHILE
\end{algorithmic}
\end{algorithm}
Algorithm \ref{alg:MHAlgorithm} requires user inputs of the initial state $x^0$ and the proposal distribution $q$. For our experiments we set the initial state $x^0$ as the average MAP estimate in (\ref{eq:MAPEstimate}) over all measurement vectors $y_j$, $j = 1,\dots, J$, and $q$ to be Gaussian, such that $x^{cand} \sim \mathcal{N}(x^{k-1}, \sigma_q)$. {As noted previously, we also use $y = \frac{1}{J} \sum_{j = 1}^J y_j$, the mean of the observable data.}
Since the Gaussian distribution is symmetric, i.e.~$q(x^{k-1}|x^{cand}) = q(x^{cand}|x^{k-1})$, the acceptance probability (\ref{eq:acceptance}) is simplified to
\begin{equation}\label{symmetricproposal}
\alpha(x^{cand}|x^{k-1}) = \min \Bigg\{ 1, \frac{\hat{f}_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x^{cand},\hat{\lambda}|y)}{\hat{f}_{\mathbf{X},\bm{\lambda}|\mathbf{Y}}(x^{k-1},\hat{\lambda}|y)} \Bigg\}.
\end{equation}
While the ideal acceptance ratio is an open question for the Metropolis-Hastings algorithm, a typical range for an acceptance ratio is $[.2,.8]$. This enables the method to ``reasonably'' explore the space, preventing the chain from getting stuck in one place or moving too quickly from one component to another, \cite{MCMCR}. Because the acceptance ratio explicitly depends on $\sigma_q$, for the purposes of this investigation, we simply chose $\sigma_q$ so that it fell within this interval. Specifically, for all posteriors given in (\ref{eq:final_posterior}) and \eqref{eq:final_posteriorl2}, we ran a few iterations of Algorithm \ref{alg:MHAlgorithm} and observed the behavior of the acceptance ratio. If it fell outside the interval, we adjusted the input parameter $\sigma_q$ accordingly.
It is important to note that $x^0$ may not provide a good initial guess. This is a standard problem in all MCMC algorithms and is addressed by implementing what is known a {\em burn-in} period. That is, if $M$ is the total number of iterations computed by Algorithm \ref{alg:MHAlgorithm}, the first $B< M$ (called the burn-in rate) iterations are discarded, so that the total length of the chain is $M-B$. {We chose $M = 50,000$ and $B = 25,000$ in all of our experiments, the same values used in studies in which convergence is consistently observed, \cite{ConvergenceMCMC}.}
In this regard we point out that our strategy for choosing $x^0$, and more importantly our use of the support informed sparse prior, allowed us to pick smaller burn-in rate and chain length while still reducing the uncertainty in all of our experiments. More studies are needed to confirm this is the case in general.
The iterations of the Markov chain obtained using Algorithm \ref{alg:MHAlgorithm} may be represented by a matrix $\mathcal{M} \in \mathbb{R}^{n \times M}$ {given by}
\begin{equation}
\label{eq:MarkovMatrix}
\mathcal{M}=
\begin{bmatrix}
x_{1}^0 & x_{1}^1 & x_{1}^2 & \dots & x_{1}^M \\
x_{2}^0 & x_{2}^1 & \dots & x_{2}^3 \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
x_{n}^0 & x_{n}^1 & x_{n}^2 & \dots & x_{n}^M
\end{bmatrix},
\end{equation}
where each column of the matrix $\mathcal{M}$ represents one state in the Markov chain. In particular, an entry $x^t_i$ is the $i^{th}$ {component} of the $t^{th}$ state in the Markov chain. As in \cite{BardsleyUQ}, we then calculate a final estimate for the unknown $x$ {at each component} as the mean of the samples after discarding the first $B$ states in the Markov chain:
\begin{equation}\label{eq:final_approx}
x_i= \frac{1}{M -B}\sum_{t = B+1}^M x_i^t.
\end{equation}
A summary of our proposed empirical Bayesian inference approach to solving the statistical inverse problem defined in (\ref{eq:ForwardModel}) is provided in Algorithm \ref{alg:final_alg}. {The mean in \eqref{eq:final_approx} is obtained after discarding the first $B < M$ burn-in iterations.} {The Algorithm is written for the support informed sparse $\ell_1$ prior, but the $\ell_2$ prior case given in \eqref{eq:PosteriorDensitymaskl2} follows analogously.}
\begin{algorithm}[H]
\caption{{Support Informed Sparse Prior} Empirical Bayesian Inference Algorithm}
\label{alg:final_alg}
\begin{algorithmic}[1]
\STATE{\textbf{Input:} Multiple measurement vectors, $y_j$, for $j = 1,...,J$.}
\STATE{\textbf{Output:} Markov chain $\mathcal{M}$ in \eqref{eq:MarkovMatrix} describing samples from $f_{\textbf{X},\bm{\lambda}|\mathbf{Y}}(x,\lambda|y)$.}
\STATE{Estimate the hyper-prior $\hat{\lambda}$ according to Algorithm \ref{alg:Cross-Validation}.}
\STATE{Calculate the support informed sparsity prior using Algorithm \ref{alg:SISP}.}
\STATE{Sample the posterior (\ref{eq:PosteriorDensitymask}) using Algorithm \ref{alg:MHAlgorithm}.}
\end{algorithmic}
\end{algorithm}
|
\section{Introduction}
Softly broken supersymmetry alleviates the electroweak hierarchy
problem \cite{Witten:1981nf,Sakai:1981gr}. In the limit of exact
supersymmetry there are no quadratically divergent radiative
corrections to the masses of Higgs bosons. Increasing lower bounds of
the masses of superparticles from searches at the LHC \cite{pdg} imply
sizable loop corrections to the masses of Higgs bosons in realistic
supersymmetric models. This leads to finetuning if soft breaking
masses are treated as uncorrelated \cite{Barbieri:1987fn,
Ellis:1986yg}. However, this source of finetuning might be greatly
reduced in models with a small number of independent soft breaking
parameters, typically defined at a very large renormalization scale,
which introduce correlations between weak--scale parameters
\cite{Baer:2012cf}.
On the other hand, the minimal supersymmetric extension of the
standard model (MSSM) \cite{Haber:1984rc,Nilles:1983ge} also requires
a supersymmetric contributions to Higgs and higgsino masses; this mass
parameter is usually called $\mu$. Searches for charginos imply
\cite{pdg} $|\mu| \:\raisebox{-0.5ex}{$\stackrel{\textstyle>}{\sim}$}\: 100$ GeV. Since $\mu$ enters the Higgs
potential already at tree--level, a value of $|\mu|$ much above $M_Z$
inevitably leads to finetuning. There is thus general agreement in the
discussion of finetuning issues in the MSSM that -- given the
experimental lower bound -- smaller values of $|\mu|$ are preferred,
with the necessary degree of finetuning increasing like $\mu^2$.
The same parameter $\mu$ also sets the mass scale for the higgsinos in
the MSSM;\footnote{This can be avoided only if one introduces additional
non--holomorphic soft--breaking higgsino mass terms \cite{Ross:2017kjc};
however, most supersymmetry breaking mechanisms do not generate such
terms. We will therefore not consider this possibility here.} as already
noted, this is the origin of the lower bound on $|\mu|$. This establishes
a connection between finetuning and the phenomenology of the neutralinos
and charginos in the MSSM.
One of the attractive features of the MSSM with exact $R-$parity is
that it automatically contains a candidate particle to form the
cosmological dark matter whose existence can be inferred from a host
of observations, assuming only that Einsteinian (or indeed Newtonian)
gravity is applicable at the length scales of (clusters of) galaxies
\cite{Bertone:2004pz}. This candidate is the lightest neutralino
$\tilde \chi_1^0$ \cite{Jungman:1995df}, whose mass is bounded from
above by $|\mu|$. Given that naturalness arguments prefer a small
value of $|\mu|$, one might assume that the most natural dark matter
candidate in the MSSM is a light higgsino--like neutralino. However,
in minimal cosmology a higgsino--like neutralino has the correct relic
density only for $|\mu| \simeq 1.2$ TeV \cite{Edsjo:1997bg}, which
would lead to permille--level electroweak finetuning. A wino--like LSP
would have to be even heavier. One can appeal to non--standard
cosmologies, e.g. including non--thermal production mechanisms, in
order to give a lighter higgsino--light neutralino the correct relic
density; however, such scenarios are already quite strongly constrained
by indirect dark matter searches \cite{Baer:2018hpb}.
In this article we therefore assume that the bino mass parameter
$M_1 < |\mu|$, so that the LSP eigenstate is dominated by the bino
component; $M_1$ can be taken positive without loss of
generality. Also in this case finetuning would prefer $|\mu|$ to be
not far above $M_1$. On the other hand, if $|\mu| \simeq M_1$ the LSP
has sizable higgsino and bino components, and hence generically
sizable couplings to the neutral Higgs bosons of the MSSM. Such mixed
neutralinos tend to have rather large scattering cross sections on
matter, in potential conflict with strong lower bounds on this
quantity from direct dark matter searches \cite{pdg}. This is true in
particular for the so--called ``well tempered neutralino''
\cite{ArkaniHamed:2006mb}, a bino--higgsino mixture with the correct
relic density in minimal cosmology.
In this article we explore this connection quantitatively, by deriving a
lower bound on the difference $|\mu| - M_1$ from the upper bound on
the neutralino--nucleon scattering cross section found by the Xenon
collaboration \cite{Aprile:2018dbl}. We do this both numerically, and
using a simple approximation for the bino--like neutralino eigenstate
which is very accurate in the relevant region of parameter space. The
resulting lower bound on $|\mu|$ is much stronger than the trivial
constraint $|\mu| > M_1$ which follows from the requirement of a
bino--like LSP; this is true in particular if $M_1$ and $\mu$ have the
same sign. An upper bound on $|\mu|$ from finetuning considerations
therefore leads to a considerably stronger upper bound on $M_1$ from
direct dark matter searches.
A bino--like neutralino will often have too large a relic density in
minimal cosmology. This can be cured either by assuming non--standard
cosmology (e.g. a period of late entropy production)
\cite{Kamionkowski:1990ty, Gelmini:2006pw}, or -- for not too small
$M_1$ -- by arranging for co--annihilation with a charged
superparticle, e.g. a $\tilde \tau$ slepton \cite{Ellis:1999mm}, which
can still have escaped detection by collider experiments if it is
close in mass to the lightest neutralino. Neither of these
modifications changes the cross section for neutralino--proton
scattering significantly. By not imposing any relic density constraint
our result thus becomes less model--dependent. This, as well as the
use of the more recent, and considerably stronger, Xenon--1T
constraint and the approximate analytical derivation of the constraint
on $|\mu|$, distinguishes our analysis from that of
ref.\cite{Profumo:2017ntc}.
The rest of this article is organized as follows. In the following section
we review neutralino mixing, both exact and using a simple approximation.
We also give the relevant expressions for the neutralino--nucleon scattering
cross section. In section~3 we present the resulting lower bound on the
difference $|\mu| - M_1$ as a function of $M_1$, before concluding in
section~4.
\section{Formalism}
In this section we briefly review the neutralino masses and mixings
in the MSSM, as well as the spin--independent contribution to
neutralino--nucleon scattering.
\subsection{The Neutralinos in the MSSM}
The neutralinos are mixtures of the two neutral gauginos (the bino
$\tilde B$ and the neutral wino $\widetilde W_3$) and two neutral
higgsinos $\tilde h_d^0, \, \tilde h_u^0$ associated with the two
$SU(2)$ Higgs doublets required in the MSSM. The resulting mass
matrix in the
$\tilde B,\, \widetilde W_3, \, \tilde h_d^0,\, \tilde h_u^0$ basis is
given by \cite{Nilles:1983ge}:
\begin{equation} \label{eq:matrix}
{\cal M}_0 = \mbox{$ \left( \begin{array}{cccc}
M_1 & 0 & - M_Z \cos \! \beta \sin \! \theta_W & M_Z \sin \! \beta \sin \!
\theta_W \\
0 & M_2 & M_Z \cos \! \beta \cos \! \theta_W & -M_Z \sin \! \beta \cos \!
\theta_W \\
- M_Z \cos \! \beta \sin \! \theta_W & M_Z \cos \! \beta \cos \! \theta_W &
0 & -\mu \\
M_Z \sin \! \beta \sin \! \theta_W & -M_Z \sin \! \beta \cos \! \theta_W &
-\mu & 0
\end{array} \right)\,. $}
\end{equation}
Here $M_1$ and $M_2$ are soft breaking masses for the bino and wino,
respectively, $\mu$ is the higgsino mass parameter, $\theta_W$ is the
weak mixing angle, $M_Z \simeq 91$ GeV is the mass of the $Z$ boson,
and $\tan\beta = \langle H_u^0 \rangle / \langle H_d^0 \rangle$ is the
ratio of the vacuum expectation values (VEVs) of the two neutral Higgs
fields. The mass matrix is diagonalized by the $4 \times 4$ matrix
$N$, such that the $i-$th neutralino eigenstate is given by
\begin{equation} \label{eq:es}
\tilde \chi_i^0 = N_{i1} \tilde B + N_{i2} \widetilde W_3 +
N_{i3} \tilde h_d^0 + N_{14} \tilde h_u^0\,.
\end{equation}
Here we are interested in scenarios where the lightest neutralino
$\tilde \chi_1^0$, which is our dark matter candidate, is dominated by
the bino component. This requires $|M_1| < |M_2|, \, |\mu|$. We will
assume that these mass parameters are real; nontrivial complex phases
would contribute to CP violation, which is strongly constrained by
upper bounds on the electric dipole moments of the electron and
neutron \cite{Arbey:2019pdb}. Without loss of generality we take $M_1$
to be positive, but allow both signs for $\mu$. $|M_2|$ is constrained
significantly by searches for charginos and neutralinos at the LHC
\cite{pdg}; as long as $|M_2| > M_1$ the sign of $M_2$ is essentially
irrelevant for our analysis.
Evidently the mixing between the gaugino and higgsino states is
controlled by the mass of the $Z$ boson. If the difference between the
gaugino masses and $|\mu|$ is larger than $M_Z$, all mixing angles
will therefore be quite small, allowing for an approximate
perturbative diagonalization of the mass matrix (\ref{eq:matrix}). In
particular, the components of a bino--like $\tilde \chi_1^0$ can then be
approximated by \cite{Drees:1992am, Choi:2001ww}
\begin{equation} \label{eq:es-approx}
\begin{split}
N_{12} &\simeq -M_Z^2 \cos{\theta_W} \sin{\theta_W}
\frac {M_1 +\mu \sin{2\beta}} {(M_1-M_2) (M_1^2-\mu^2)} \,; \\
N_{13} &\simeq -M_Z \sin{\theta_W}
\frac {M_1 \cos{\beta} + \mu \sin{\beta}} {(M_1^2-\mu^2)} \,; \\
N_{14} &\simeq M_Z \sin{\theta_W}
\frac{M_1 \sin{\beta} +\mu \cos{\beta}} {(M_1^2-\mu^2)} \,; \\
N_{11} &= \sqrt{1-N_{12}^2-N_{13}^2-N_{14}^2} \,.
\end{split}
\end{equation}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{Nij.png}
\caption {The bino and higgsino components of the lightest
neutralino eigenstate as a function of $\mu$, for fixed $M_1 = 300$
GeV, $M_2 = 3$ TeV and $\tan\beta=10$. The simple approximation of
eqs.(\ref{eq:es-approx}), shown in black, describes the exact (blue)
results very well once $\mu - M_1 \geq 2 M_Z$.}
\label{fig:Nij}
\end{figure}
In Fig.~\ref{fig:Nij} we compare these approximate expression with
exact results. Evidently the approximation works very well for
$\mu - M_1 \geq 2 M_Z$ or so. Since we took a very large value of
$M_2$ the wino component essentially vanishes in this example;
however, the first eq.(\ref{eq:es-approx}) shows that it only appears
at second order in $M_Z$, and is therefore always much smaller than
the higgsino components in the region of interest. In this figure we
have chosen $\mu$ to be positive. The second and third
eq.(\ref{eq:es-approx}) shows that this increases the higgsino
components. As a result, for $\mu < 0$ the approximation
(\ref{eq:es-approx}) becomes very accurate already for
$|\mu| - M_1 \geq 1.5 M_Z$.
\subsection{Neutralino--Nucleon Scattering in the MSSM}
In the limit of vanishing neutralino velocity only two kinds of
interactions contribute to neutralino--nucleon scattering. One of them
depends on the spin of the target nucleus; this contribution is
usually sub--dominant for heavy target nuclei like Xenon or
Germanium, which currently yield the tightest constraints for
neutralino masses above 10 GeV or so. The spin independent contributions
dominate because their contribution to the scattering cross section of
heavy nuclei scales like the square of the nucleon number. They originate
from the effective Lagrangian \cite{Jungman:1995df}
\begin{equation} \label{eq:Leff}
L_{\rm SI}^{\rm eff} = f_q \overline{\tilde\chi_1^0} \tilde\chi_1^0 \bar{q}q
\end{equation}
which describes the interactions of neutralinos with quarks. Here we
have limited ourselves to the leading, dimension--6, operator; strong
lower bounds on squark masses \cite{pdg} imply that dimension--8
operators due to squark exchange \cite{Drees:1993bu} can safely be
ignored.
Squark exchange also contributes at dimension $6$. However, this
contribution is proportional to the mass of the quark
\cite{Jungman:1995df}. That is of course also true for Higgs
exchange contributions. However, at least the lighter neutral Higgs
boson, whose mass we now know to be $125$ GeV, lies about an order of
magnitude below the current lower bound on first generation squark
masses. Squark exchange diagrams therefore contribute only at the
$1\%$ level at best, which is well below the uncertainty of the Higgs
exchange contribution. We therefore ignore them in our analysis.
However, we do allow for the contribution of the heavier neutral Higgs
boson. The effective neutralino--quark coupling $f_q$ is thus given
by
\begin{equation} \label{eq:fq}
f_q = \sum_{\phi=h,H} m_q \frac{g_{\phi \bar \chi \chi} g_{\phi \bar qq}}{m_{\phi}^2}\,.
\end{equation}
Note that we factored the quark mass out of the Higgs couplings to quarks,
making the latter independent of the quark mass; the couplings to
up-- and down--type quarks still differ, however. They are given by
\cite{Gunion:1984yn}:
\begin{equation} \label{eq:yuk}
\begin{split}
g_{h \bar uu} &= \frac{-g \cos\alpha}{2M_W \sin\beta} \simeq\frac{-g}{2M_W}\,;\\
g_{h\bar dd} &= \frac{g \sin\alpha}{2M_W \cos\beta}\simeq\frac{-g}{2M_W}\,;\\
g_{H\bar uu} &= \frac{-g \sin\alpha}{2M_W \sin\beta}
\simeq\frac{g}{2M_W\tan\beta}\,;\\
g_{H \bar dd} &= \frac{-g \cos\alpha}{2M_W \cos\beta}
\simeq \frac{-g\tan\beta}{2M_W}\,.
\end{split}
\end{equation}
Here $\alpha$ is the mixing angle between the two neutral Higgs
bosons, $g$ is the $SU(2)$ gauge coupling and $M_W \simeq 80$ GeV is
the mass of the $W$ boson. The couplings of the 125 GeV Higgs boson
are known to be quite close to those of the SM Higgs boson
\cite{pdg}. Moreover, none of the heavier Higgs bosons of the MSSM
have yet been found. Both observations can easily be satisfied in the
so--called decoupling limit, where the mass of the neutral CP--odd
Higgs boson satisfies $m_A^2 \gg M_Z^2$. In that limit the other heavy
MSSM Higgs bosons also have masses very close to $m_A$, and the mixing
angle $\alpha$ satisfies
$\cos\alpha \simeq \sin\beta,\ \sin\alpha \simeq -\cos\beta$. This
leads to the simplifications in the Higgs couplings to quarks given
after the $\simeq$ signs in eqs.(\ref{eq:yuk}). In particular, the
couplings of the lighter Higgs boson $h$ then approach those of the SM
Higgs, in agreement with observation. The couplings of the heavier
neutral Higgs boson $H$ to up--type quarks is suppressed by
$1/\tan\beta$, while its couplings to down--type quarks are enhanced
by $\tan\beta$.
The Higgs bosons couple to one gaugino and one higgsino current state.
As a result, their couplings to neutralino current eigenstates are
proportional to the product of gaugino and higgsino components
\cite{Gunion:1984yn}:
\begin{equation} \label{eq:ghcc}
\begin{split}
g_{h\bar\chi \chi} =& \frac{1}{2}(g N_{12} - g'N_{11})
(N_{13}\sin\alpha + N_{14} \cos {\alpha}) \,; \\
g_{H\bar\chi \chi} =& \frac{1}{2}(g N_{12} - g'N_{11})
(N_{14} \sin\alpha - N_{13}\cos{\alpha}) \,,
\end{split}
\end{equation}
where $g'$ is the $U(1)_Y$ coupling.
The total spin--independent neutralino--proton scattering cross
section can be written as \cite{Jungman:1995df}
\begin{equation} \label{eq:cs}
\sigma_{SI}^{\chi p} = \frac{4\mu_{\chi p}^2} {\pi} |G_s^p|^2 \,,
\end{equation}
where
$\mu_{\chi p} = m_p m_{\tilde \chi_1^0}/ (m_p + m_{\tilde\chi_1^0})$
is the reduced mass of the neutralino--proton system, and the
effective neutralino--proton coupling is given by
\begin{equation} \label{eq:G}
G_s^p = -\sum_{q=u,d,..} \langle p| m_q\bar{q} q| p \rangle
\sum_{\phi=h,H} \frac{g_{\phi\bar\chi \chi} g_{\phi \bar qq}}{m_\phi^2} \,.
\end{equation}
For the light $u,d,s$ quarks, the hadronic matrix elements have to be
computed using non--perturbative methods. Once these are known, the
contribution from heavy $c,b,t$ quarks can be computed perturbatively
through a triangle diagram coupling to two gluons
\cite{Shifman:1978zn}. One usually parameterizes
$\langle N| m_q\bar{q} q| N \rangle = f_{T_q} m_p$. We use the
numerical values from DarkSUSY \cite{Bringmann:2018lay}:
\begin{equation} \label{eq:fT}
f_{TU} \equiv \sum_{q = u,c,t} f_{Tq} = 0.14\,; \ \
f_{TD} \equiv \sum_{q = d,s,b} f_{Tq} = 0.23\,.
\end{equation}
We note that these numbers are somewhat uncertain, but our values are
rather conservative \cite{Ellis:2018dmb}.
Putting everything together, using the approximate expressions
(\ref{eq:es-approx}) for the lightest neutralino eigenstate and
assuming the decoupling limit, we find:
\begin{equation} \label{eq:G-approx}
\begin{split}
\left. G_S^p \right|_h &\simeq -A \ m_p \left( f_{TU} + f_{TD} \right)
\left( \frac{M_1+\mu \sin{2\beta}} {m_h^2 (\mu^2-M_1^2)} \right)\,; \\
\left. G_S^p \right|_H &\simeq -A \ m_p \left( \frac{ f_{TU}} {\tan\beta}
- f_{TD} \tan\beta\right) \times \left(
\frac{\mu \cos{2\beta}}{m_H^2(\mu^2-M_1^2)} \right)\,.
\end{split}
\end{equation}
Here we have introduced the constant
\begin{equation} \label{eq:A}
A = \frac{g g' M_Z \sin\theta_W}{4 M_W} = 0.032\,.
\end{equation}
It should be noted that $\tan\beta > 1$ implies $\cos(2\beta) < 0$.
Eqs.(\ref{eq:fT}) imply that the term $\propto f_{TD}$ always
dominates $H$ exchange. Hence $h$ and $H$ exchange contribute with the
same sign if $\mu > 0$ or $\mu \sin2\beta < -M_1$; they interfere
destructively for $0 > \mu > -M_1/\sin2\beta$.
\section{Results}
We are now ready to present our numerical results. We wish to determine
the lower bound on $|\mu|$ that follows from the non--observation of
neutralinos, which we assume to form all of (galactic) dark matter.
We will not be concerned with very light neutralinos, where current bounds
from direct dark matter search are still quite poor \cite{pdg}. For
masses above $20$ GeV the most stringent current bound comes from the
Xenon--1T experiment \cite{Aprile:2018dbl}. In this range the bound
is well parameterized by
\begin{equation} \label{sig-max}
\begin{split}
\sigma^{\rm max}(m_{\tilde \chi_1^0})_{\rm XENON} &=
\left( \frac {m_{\tilde\chi_1^0}} {10 \ {\rm GeV}}
+ \frac {2.7 \cdot 10^4 \ {\rm GeV}^3 }{m_{\tilde \chi_1^0}^3} \right)
\cdot 10^{-47} \ {\rm cm}^2 \\
&= \left( \frac {m_{\tilde\chi_1^0}} {3.9 \ {\rm GeV}}
+ \frac {7 \cdot 10^4 \ {\rm GeV}^3 }{m_{\tilde \chi_1^0}^3} \right)
\cdot 10^{-20} \ {\rm GeV}^{-2}\,.
\end{split}
\end{equation}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{CROSS.png}
\caption{The predicted neutralino--proton scattering cross section
for $M_1 = 150$ GeV, $M_2 = 3$ TeV, $m_A = 1.8$ TeV and $\tan\beta = 10$
as function of $|\mu|$, for positive (green) and negative (blue) $\mu$. The
bound from the Xenon--1T collaboration is shown as black dot--dashed line.}
\label{fig:fig3}
\end{figure}
Fig.~3 shows that this bound constrains the MSSM parameter space quite
severely, if we assume that the lightest neutralino $\tilde \chi_1^0$
forms all of dark matter. Here we have chosen $M_1 = 150$ GeV,
$m_A = 1.8$ TeV, $M_2 = 3$ TeV and $\tan\beta = 10$; the exact value
of $M_2$ is basically irrelevant as long as it is significantly larger
than $M_1$. As expected the predicted cross section is largest for
$|\mu| \simeq M_1$, which leads to strong bino--higgsino
mixing. However, the Xenon--1T bound requires $|\mu|$ well above
$M_1$, i.e. the lightest neutralino has to be bino--like; note that
values of $|\mu|$ below 100 GeV have not been considered here since
they are excluded by chargino searches at LEP for the given (large)
value of $M_2$. In the allowed range of $|\mu|$ the approximate
diagonalization of eqs.(\ref{eq:es-approx}) works quite
well. Eq.(\ref{eq:G-approx}) then explains why the lower bound on $|\mu|$
is considerably weaker for $\mu < 0$: evidently the two terms in the
numerator of $\left. G_S^p \right|_h$ tend to cancel (add up) for
positive (negative) $\mu$. As a result, for $\mu < 0$ the contribution
from $H$ exchange, which contributes with opposite sign than the
dominant $h$ exchange term, is relatively more important and further
reduces the cross section. In fact, the $h$ exchange contribution
vanishes at $\mu = -M_1/\sin(2\beta) \simeq 760$ GeV. Due to $H$
exchange the actual zero of the cross section \cite{Drees:1993bu} --
the so--called ``blind spot'' \cite{cheung2013} -- already occurs at
$\mu \simeq -660$ GeV. In contrast, for $\mu > 0$ the (small) $H$
exchange contribution slightly strengthens the lower bound on $\mu$,
which saturates at $\sim 590$ GeV for $m_H \rightarrow \infty$. This
value is already uncomfortably large in view of finetuning
considerations.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{bound-pos.png}
\caption{The lower bound on $|\mu| - M_1$, in units of $M_Z$, that
follows from the upper bound on the neutralino--proton scattering
cross section derived by the Xenon--1T collaboration, for $\mu >
0$. We have again chosen $M_2 = 3$ TeV and $m_A = 1.8$ TeV. The
solid lines show numerical results obtained using DarkSUSY for
different values of $\tan\beta$, while the dashed curves are based
on the approximate analytical diagonalization of the neutralino mass
matrix, see eq.(\ref{eq:mubound}).}
\label{fig:fig4}
\end{figure}
This conclusion is reinforced by Fig.~\ref{fig:fig4}, which shows the
lower bound on $\mu - M_1$ in units of $M_Z$ as a function of $M_1$
for four different values of $\tan\beta$; the values of $M_2$ and
$m_A$ are as in Fig.~\ref{fig:fig3}. The solid colored lines have been
derived numerically using DarkSUSY, whereas the black dashed lines are
based on the approximate diagonalization of the neutralino mass
matrix. Recall that we ignore squark exchange, so that only $h$ and $H$
exchange contribute to the spin--independent scattering cross section.
Using eqs.(\ref{eq:G-approx}) the extremal values of $\mu$ that
saturate the experimental upper bound on the cross section can be
computed analytically. To this end we introduce the quantities
\begin{equation} \label{eq:pars}
\begin{split}
\kappa &= \frac {\sqrt{\pi \sigma^{\rm max}}} {2 A m_p^2}\,; \\
c_\mu &= \frac { (f_{TD} + f_{TU}) \sin 2\beta} {m_h^2}
- \frac{f_{TD} \tan\beta - f_{TU} \cot\beta } {m_H^2}\,; \\
c_1 &= \frac { f_{TD} + f_{TU}} {m_h^2}\,.
\end{split}
\end{equation}
The quantity $A$ has been defined in eq.(\ref{eq:A}), and the
normalized hadronic matrix elements $f_{TD}$ and $f_{TU}$ in
eqs.(\ref{eq:fT}). As noted above, the contribution $\propto f_{TU}$
to the $H$ exchange contribution is essentially negligible. $c_\mu$
collects terms in the Higgs exchange amplitude that are proportional
to $\mu$; only $h$ exchange contributes to $c_1$, which gets
multiplied with $M_1$ in this amplitude. The extremal values of $\mu$
are then given by
\begin{equation} \label{eq:mubound}
\mu_\pm = \frac {c_\mu} {2 \kappa} \pm \sqrt { \frac {c_\mu^2} {4 \kappa^2}
+ M_1^2 + \frac{c_1 M_1}{\kappa} }\,.
\end{equation}
The dashed lines in Fig.~\ref{fig:fig4} correspond to the positive
solution $\mu_+$.
We see that this approximation describes the numerical results very
well even for the smallest value of $M_1$ we consider. In particular,
the rather strong dependence on $\tan\beta$ originates from the
$\sin2\beta$ factor in the $h$ exchange contribution to $c_\mu$, see
the second eq.(\ref{eq:pars}); $H$ exchange is always subdominant for
our choice $m_H = 1.8$ TeV. The $\tan\beta$ dependence becomes
somewhat weaker for larger values of $M_1$, where the $c_1$ term
becomes more important which is independent of $\tan\beta$.
Evidently the Xenon--1T bound is quite constraining for $\mu > 0$. For
example, if we interpret electroweak finetuning considerations as
requiring $|\mu| < 500$ GeV, one finds $\tan\beta > 10$ for
$M_1 \geq 20$ GeV, and $M_1 < 115 \ (165)$ GeV for
$\tan\beta = 20 \ (50)$.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{bound-neg.png}
\caption{The lower bound on $|\mu| - M_1$, in units of $M_Z$, that
follows from the upper bound on the neutralino--proton scattering
cross section derived by the Xenon--1T collaboration, for $\mu < 0$;
the regions to the left of the upper red and green curves are also
excluded. Parameter values and notation are as in
Fig.~\ref{fig:fig4}.}
\label{fig:fig5}
\end{figure}
In eq.(\ref{eq:mubound}) we have assumed that
$c_\mu \mu + c_1 M_1 > 0$, which is always true for $\mu > 0$. It
remains true for values of $|\mu|$ below the ``blind spot''; if $H$
exchange is negligible this corresponds to $|\mu| \sin2\beta <
M_1$. The negative solution in eq.(\ref{eq:mubound}) then gives the
value of $\mu$ where the cross section decreases below the lower bound
when coming from $\mu = 0$. Fig.~\ref{fig:fig5} shows that this again
describes the exact numerically derived bound quite well: although the
bound on $|\mu| - M_1$ is considerably weaker than for positive $\mu$,
we saw in eqs.(\ref{eq:es-approx}) that there are cancellations in the
small entries of the $\tilde \chi_1^0$ eigenstate if $\mu < 0$; since
corrections to this approximation are of order of the squares of these
small entries, for given $|\mu|$ the approximation works better for
$\mu < 0$. In this figure we only show results for $M_1 \geq 40$ GeV,
since for smaller values of $M_1$ the bound on $|\mu|$ often is below
chargino search limit of about $100$ GeV. Note that now the lower
bound on $|\mu|$ {\em in}creases with increasing $\tan\beta$. This is
because the ``blind spot'' $\mu = - M_1/\sin2\beta$ moves to larger
values of $|\mu|$ for larger $\tan\beta$. In the limit
$\tan\beta \rightarrow \infty$ the bound on $|\mu|$ becomes
independent of the sign of $\mu$.
Beyond the blind spot the sign of $c_\mu \mu + c_1 M_1$ flips. This
region of parameter space can still be described by
eq.(\ref{eq:mubound}), by simply setting $\kappa \rightarrow - \kappa$
everywhere. If $H$ exchange is negligible, this can easily make the
argument of the root in eq.(\ref{eq:mubound}) negative, signaling that
no solution exists. In this case the cross section remains below the
experimental bound for all values of $\mu$ below the $\mu_-$ solution
in the original eq.(\ref{eq:mubound}). For the parameters of
Fig.~\ref{fig:fig5} we find that this is true for $\tan\beta > 8$. For
smaller values of $\tan\beta$ a sizable region of parameter space (to
the left and below the red and green solid lines) beyond the blind
spot is again excluded.\footnote{These curves are reproduced very
accurately by the approximate diagonalization of the neutralino mass
matrix; we do not show these results in order not to clutter up the
figure too much.}
So far we have assumed that the heavy Higgs boson is very heavy, so
that its contribution to neutralino--proton scattering is
subdominant. In fact, there are several constraints on the masses of
the heavy Higgs bosons in the MSSM, which can be characterized by the
mass of the CP--odd Higgs boson, $m_A$. The most robust bounds come
from direct searches for the heavy neutral Higgs bosons; the most
sensitive ones exploit their decay into $\tau^+\tau^-$ pairs. In
particular, a recent ATLAS analysis \cite{Aad:2020zxo} is sensitive to
$m_A$ up to about 2 TeV, for very large $\tan\beta$. For
$\tau^+\tau^-$ invariant masses around $400$ GeV there seems to be
some excess of events. While not statistically significant, it leads
to a bound which is somewhat worse than the expected
sensitivity.\footnote{CMS did not yet publish the corresponding
analysis for the full Run--2 data set.} We therefore chose a
parameter set just on the exclusion line, $m_A = 400$ GeV and
$\tan\beta = 8$, in order to illustrate the maximal possible effect
from heavy Higgs exchange. It should be noted that this choice leads
to a sizable contribution from charged Higgs boson loops to radiative
$b \rightarrow s \gamma$ decays \cite{Bertolini:1990if,
Borzumati:1994kk}; however, these can be compensated by postulating
some amount of squark flavor mixing \cite{Okumura:2003hy}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.475\textwidth]{ma400-pos.png}
\includegraphics[width=0.475\textwidth]{ma400-neg.png}
\caption{The lower bound on $|\mu| - M_1$, in units of $M_Z$, that
follows from the upper bound on the neutralino--proton scattering
cross section derived by the Xenon--1T collaboration, for $\mu > 0$
(left) and $\mu < 0$ (right). We have chosen $M_2 = 3$ TeV and
$\tan\beta = 8$; the black (green) lines are for $m_A = 0.4 \ (5)$
TeV. In the right frame the enclosed region as well as the region
above the first line are excluded.}
\label{fig:fig6}
\end{figure}
The bound on $|\mu|$ that results from the Xenon--1T constraint for
this choice of parameters is shown by the black lines in
Figs.~\ref{fig:fig6}; for comparison we also show results for
negligible $H$ exchange (green curves). As noted earlier for $\mu > 0$
both Higgs bosons always contribute with equal sign, so maximizing the
$H$ exchange contribution greatly strengthens the lower bound on
$\mu$. The effect is especially strong at smaller $M_1$ since $H$
exchange only contributes to $c_\mu$, not to $c_1$. The resulting
lower bound on $\mu$ is always above 950 GeV, i.e. in this region of
parameter space a bino--like lightest neutralino would yield little
benefit regarding electroweak finetuning compared with the canonical
thermal higgsino--like neutralino with $\mu \simeq 1.2$ TeV. Of
course, the black line shows the {\em maximal} effect from $H$
exchange. For somewhat larger $m_A$, away from the ATLAS lower bound,
the bound on $\mu$ will fall somewhere between the black and green
lines.
In sharp contrast, for $\mu < 0$ the $H$ exchange contribution reduces
the lower bound on $|\mu|$ even further, by moving the blind spot to
smaller values of $|\mu|$. However, for $M_1 \leq 170$ GeV the cross
section beyond the blind spot again increases above the Xenon--1T
bound, leading to a second excluded region. In fact, the allowed
region around the blind spot is very narrow for $M_1 \leq 150$ GeV.
The right frame of Fig.~\ref{fig:fig6} therefore shows that for
$m_A = 400$ GeV values of $M_1$ below about 150 GeV require
significant finetuning, either to hone in on the blind spot, or in the
electroweak sector due to the large values of $|\mu|$ required by the
Xenon--1T constraint away from the blind spot.
\section{Summary and Conclusions}
In this paper we have shown that the upper bound on the
neutralino--proton cross section from the Xenon--1T experiment leads
to strong lower bounds on $|\mu|$ if the bino mass parameter $M_1$
exceeds 20 GeV. We have assumed that the lightest neutralino forms all
of dark matter, but did not require the correct thermal relic density
in minimal cosmology. This constraint causes tension with electroweak
finetuning arguments, since in the MSSM with holomorphic soft breaking
terms the higgsino mass parameter $\mu$ also contributes to the masses
of the Higgs bosons. The bound is usually significantly stronger for
$\mu > 0$; however, also for $\mu < 0$ we found very strong
constraints if the heavy Higgs bosons are close in mass to current
experimental constraints and $M_1 \leq 150$ GeV. This argument can be
turned around to derive an upper bound on $M_1$ from an upper bound on
$|\mu|$ from electroweak finetuning; the latter is, however, not easy
to quantify unambiguously \cite{Drees:1995hj}. Our analytical
expressions will help to easily update these constraints when future
direct dark matter searches are published.
The Xenon--1T bound becomes considerably weaker for neutralino masses
below 20 GeV, which we did not consider in this paper. While even very
small values of $M_1$ remain experimentally allowed as long as all
sfermions are sufficiently heavy \cite{Dreiner:2009ic}, they do not
appear particularly plausible given the ever strengthening lower
bounds on the masses of the other gauginos (electroweak winos as well
as gluinos), chiefly from searches at the LHC \cite{pdg}. In fact,
many models of supersymmetry breaking predict fixed ratios between
these masses \cite{Choi:2007ka}, leading to strong lower bounds on
$M_1$. Our analysis would then lead to even stronger lower bounds on
$|\mu|$.
\subsubsection*{Acknowledgments}
GG thanks the Deutsche Akademische Auslandsdienst (DAAD) for partial financial
support and the Bethe Center for Theoretical Physics (BCTP) for its
hospitality. We thank Hessamaddin Arfaei for discussions.
|
\section{Introduction}
The Capsule network (CapsNet)~\cite{sabour2017dynamic} was proposed to address the intrinsic limitations of convolutional networks (ConvNet)~\cite{lecun1998gradient}, such as the exponential inefficiency and the lack of robustness to affine transformations. In recent years, It has been sugested that CapsNets have the potential to surpass the dominant convolutional networks in these aspects~\cite{sabour2017dynamic,hinton2018matrix,rajasegaran2019deepcaps,gu2020improving,de2020introducing,mazzia2021efficient}. However, there lack comprehensive comparisons to support this assumption, and even for some reported improvements, there are no solid ablation studies to figure out which ones of the components in CapsNets are, in fact, effective.
In this paper, we first carefully examine the major differences in design between the capsule networks and the common convolutional networks adopted for image classification. A common convolutional network follows a simple algorithm flow, using a backbone convolutional network to extract image features, a global average pooling layer plus a linear layer to produce the classification logits (or optionally several fully connected layers~\cite{krizhevsky2017imagenet}), and an $N$-way SoftMax loss to drive the learning. To be better aligned with the capsule (vector) representations, the capsule networks introduce several special components. These components involve (see Fig.~\ref{fig:overview} for detailed architectures):
\begin{itemize}
\item a non-shared transformation module, in which the primary capsules are transformed to execute votes by non-shared transformation matrices;
\item a dynamic routing layer to automatically group input capsules to produce output capsules with high agreements in each output capsule;
\item a squashing function, which is applied to \textit{squash} the capsule vectors such that their lengths distribute in the range of $[0, 1)$;
\item a marginal classification loss to work together with the \emph{squashed} capsule representations;
\item a class-conditional reconstruction sub-network with a reconstruction loss, targeting at recovering the original image from the capsule representations. This sub-network acts as a regularization force, in complementary to the classification loss.
\end{itemize}
Unlike previous studies~\cite{sabour2017dynamic,hinton2018matrix} which usually takes CapsNet as a whole to test its robustness, we instead try to study the effects of each of the above components in their effectiveness on robustness. We consider the three different aspects shown in~\cite{sabour2017dynamic}:
\begin{itemize}
\item the robustness to affine transformations,
\item the ability to recognizing overlapping digits,
\item the semantic representation compactness.
\end{itemize}
Our investigations reveal that some widely believed benefits of Capsule networks could be wrong:
\begin{enumerate}
\item The ConvNets baseline adopted in comparison with CapsNets is weak~\cite{sabour2017dynamic}. Concretely, there is no global average pooling layer before the classification head in this baseline, which sacrifices the ability of spatial invariance to some extent and is harmful for generalization to novel views. In fact, a ConvNet with an additional global average pooling layer can outperform CapsNet by a large margin in the robustness to affine transformation;
\item The dynamic routing actually may harm the robustness to input affine transformation, in contrast to the common belief;
\item The high performance of CapsNets to recognize overlapping digits can be mainly attributed to the extra modeling capacity brought by the transformation matrices.
\item Some components of CapsNets are indeed beneficial for learning semantic representations, e.g., the conditional reconstruction and the squashing function, but they are mainly auxiliary components and can be applied beyond CapsNets.
\end{enumerate}
In addition to these findings, we also enhance common ConvNets by the useful components of CapsNet, and achieve greater robustness. The paper is organized as follows: Sec.~\ref{sec:related_work} introduces the CapsNet and related work. In Sec.~\ref{sec:study}, we examine the behavior of CapsNets and ConvNets on three kinds of robustness, one by one, and component by component. The last section concludes our work and discusses future work.
\section{Background and Related Works}
\label{sec:related_work}
\paragraph{Capsule Network with Dynamic Routing~\cite{sabour2017dynamic}:} The CapsNet architecture is shown in Fig.~\ref{fig:overview}. CapsNet first extracts feature maps of shape $(C, H, W)$ from pixel intensities with two standard convolutional layers where $C$, $H$, $W$ are the number of channels, the height, and the width of the feature maps, respectively. The extracted feature maps are reformulated as primary capsules $(C/D_{in}, H, W, D_{in})$ where $D_{in}$ is the dimensions of the primary capsules. There are $M = C/D_{in} * H * W$ primary capsules in total. Each capsule $\pmb{u}_i$, a $D_{in}$-dimensional vector, consists of $D_{in}$ units across $D_{in}$ feature maps at the same location. Each primary capsule is transformed to make a vote with a transformation matrix $\pmb{W}_{ij} \in \mathbb{R}^{(D_{in} \times N*D_{out})}$, where $N$ is the number of output classes and $D_{out}$ is the dimensions of output capsules. The vote is
\begin{equation}
\pmb{\hat{u}}_{j|i} = \pmb{u}_i\pmb{W}_{ij}.
\label{equ:vote}
\end{equation}
The routing mechanism takes all votes into consideration and identify a weight $c_{ij}$ for each vote $\pmb{\hat{u}}_{j|i}$. Concretely, the routing process iterates over the following three steps
\begin{equation}
\small
\begin{split}
\pmb{s}^{(t)}_j &= \sum^N_i c^{(t)}_{ij} \pmb{\hat{u}}_{j|i}, \\
\pmb{v}^{(t)}_j &= g(\pmb{s}^{(t)}_j), \\
c^{(t+1)}_{ij} &= \frac{\exp(b_{ij} + \sum_{r=1}^t \pmb{v}^{(r)}_j \pmb{\hat{u}}_{j|i}) }{\sum_k \exp(b_{ik} + \sum_{r=1}^t \pmb{v}^{(r)}_k \pmb{\hat{u}}_{k|i} )},
\end{split}
\label{eq:rounting_process}
\end{equation}
where the superscript $t$ is the index of an iteration starting from 1 and $g(\cdot)$ is a squashing function that maps the length of the vector $\pmb{s}_j$ into the range of $[0, 1)$. The $b_{ik}$ is the log prior probability. The squashing function is
\begin{equation}
\pmb{v}_j= g(\pmb{s}_j) = \frac{\norm{\pmb{s}_j}^2}{1+\norm{\pmb{s}_j}^2} \frac{\pmb{s}_j}{\norm{\pmb{s}_j}}.
\label{equ:squa}
\end{equation}
The length of the final output capsule $\pmb{v}_j$ corresponds to the output probability of the $j$-th class. The margin loss function is applied to compute the classification loss
\begin{equation}
\begin{split}
L_k = & T_k \max(0, m^+ - \norm{\mathbf{v}_k})^2 \\
& + \lambda (1 -T_k) \max(0, \norm{\mathbf{v}_k} - m^-)^2
\end{split}
\label{equ:marginloss}
\end{equation}
where $T_k = 1$ if the object of the $k$-th class is present in the input. As in~\cite{sabour2017dynamic}, the hyper-parameters are often empirically set as $m^+ = 0.9$, $m^- = 0.1$ and $\lambda=0.5$.
\begin{figure*}[t]
\centering
\includegraphics[width=1.01\textwidth]{figures/Caps2CNNs}
\caption{The overview of ConvNet and CapsNet architectures: The ConvNet-FC is a naive ConvNet architecture, while ConvNet-Avg is the one commonly used in image classifications. The CapsNet consists of primary capsule extraction, a transformation process, a routing process, and a class-conditional reconstruction, which is far more complex than ConvNets.}
\label{fig:overview}
\end{figure*}
A reconstruction sub-network reconstructs the input image from all $N$ output capsules with a masking mechanism. The ones corresponding to the non-ground-truth classes are masked with zeros before being transfered to the reconstruction sub-network. Due to the masking mechanism, only the capsule of the ground-truth class is visible for the reconstruction. Hence, the reconstruction process is called class-conditional reconstruction. The reconstruction loss is computed as a regularization term in the loss function.
\paragraph{Capsule Network Follow-Ups:} Many routing mechanisms have been proposed to improve the performance of CapsNet, such as Expectation-Maximization Routing~\cite{hinton2018matrix}, Self-Routing~\cite{hahn2019self}, Variational Bayes Routing~\cite{ribeiro2020capsule}, Straight-Through Attentive Routing~\cite{ahmed2019star}, and Inverted Dot-Product Attention routing~\cite{Tsai2020Capsules}. To reduce the parameters of CapsNet, a matrix or a tensor has been used to represent an entity instead of a vector~\cite{hinton2018matrix,rajasegaran2019deepcaps}. The size of the learnable transformation matrix can be reduced by the matrix/tensor representations. Another way to improve CapsNets is to integrate advanced modules of ConvNets into CapsNets, e.g., by skip connections~\cite{he2016deep,rajasegaran2019deepcaps} and dense connections~\cite{huang2017densely,phaye2018multi}.
Besides, the robustness of CapsNet has also been intensively investigated. Both new routing mechanisms~\cite{hinton2018matrix} and new architectures~\cite{kosiorek2019stacked} can improve the affine transformation robustness. The work~\cite{gu2020improving} achieves the best performance on the transformation robustness benchmark by simply removing the dynamic routing and by sharing the transformation matrix. The work also revealed that the high transformation robustness of CapsNets could not be attributed to the dynamic routing mechanism. The work~\cite{gu2020interpretable} replaces the dynamic routing with a multi-head attention-based graph pooling approach to achieve better interpretability. The replacement of the routing does not harm the robustness of CapsNet, even though it is the fundamental part of CapsNets. These claims further motivate us to investigate the individual components of CapsNet.
Additionally, CapsNet with new routing mechanisms can achieve high adversarial robustness~\cite{hahn2019self}. However, the work~\cite{michels2019vulnerability} shows CapsNet can be fooled as easily as CovNet. Recent work shows that the class-conditional reconstruction sub-network of CapsNet is useful to detect adversarial examples~\cite{qin2019detecting,qin2020deflecting}. The work~\cite{gu2021effective} designs the first attack method specific for CapsNet, which reduces the robust accuracy and increases the rate to pass the adversarial detection. Due to the attack-defense arms race, it is difficult to draw a solid conclusion on the adversarial robustness of CapsNet. Hence, in this work, we mainly focus on the advantage of CapsNet demonstrated in~\cite{sabour2017dynamic}.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.31\textwidth}
\includegraphics[width=\textwidth]{figures/data_vis_mnist}
\caption{MNIST dataset}
\label{fig:data_vis_mnist}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}[b]{0.31\textwidth}
\includegraphics[width=\textwidth]{figures/data_vis_affNIST}
\caption{AffNIST dataset}
\label{fig:data_vis_affNIST}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}[b]{0.31\textwidth}
\includegraphics[width=\textwidth]{figures/data_vis_MultiMNIST}
\caption{MultiMNIST dataset}
\label{fig:data_vis_MultiMNIST}
\end{subfigure}
\caption{Visualization of datasets: While MNIST dataset corresponds to standard hand-written digits, AffNIST dataset consists of affine-transformed MNIST images. MultiMNIST dataset consists of images with two overlapping digits. In the figure, the two overlapping digits are marked with two different colors, i.e., yellow and magenta.}
\label{fig:data_vis}
\end{figure*}
\section {Empirical Studies on Capsule Network}
\label{sec:study}
In this section, we conduct empirical studies on the robustness of CapsNets. Before we dive into the studies, we first introduce the architectures of CapsNets and ConvNets. The CapsNet we focus on in this work is Capsule Networks with dynamic routing~\cite{sabour2017dynamic}. Since the research on CapsNets is still at a primary stage, the work~\cite{sabour2017dynamic} compares their CapsNet with a LeNet-type ConvNet~\cite{lecun1998gradient}, called ConvNet-FC. The ConvNet-FC and CapsNet are illustrated in Fig.~\ref{fig:overview} on 28$\times$28 MNIST images. The notation Conv(C, K, S) stands for a convolutional layer where C, K, S are the number of channels, the kernel size, and the stride size, respectively. FC(N) is a fully connected layer where $N$ is the number of output units. All Conv and FC are followed by a ReLU activation function.
\textbf{ConvNet-FC}: The simple ConvNet baseline used in~\cite{sabour2017dynamic} is Conv(256, 5, 1) + Conv(256, 5, 1) + Conv(128, 5, 1) + FC(328) + FC(192) + Softmax(10). The three standard convolutional layers and two fully connected layers are applied to extract features from input images. An $N$-way Softmax is applied to obtain the output distribution. During training, cross-entropy loss is typically applied.
\textbf{CapsNet}: The CapsNet with Dynamic Routing in~\cite{sabour2017dynamic} is Conv(256, 9, 1) + Conv(256, 9, 2) + Dynamic Routing, followed by a reconstruction sub-network, FC(512) + FC(1024) + FC(28$\times$28). The feature maps are computed with the two standard convolutional layers. The extracted feature maps (256, H, W) is reshaped into primary capsules (32*H*W, 8) where H and W are the height and width of feature maps. The primary capsules are squashed by the squashing function in Equation (\ref{equ:squa}) and then transformed to make votes with the learned transformation matrices (32*H*W, 8, 160). The vote of each primary capsule is 160-dimensional. The dynamic routing in Equation (\ref{eq:rounting_process}) is applied to the votes to identify their weights. The output of the dynamic routing is 160-dimensional, i.e. representing 10 16-dimensional output capsules. The squashing function in Equation (\ref{equ:squa}) is applied to output capsules to map their lengths into [0, 1). The length of an output capsule is interpreted as a class output probability. In the training process, the margin loss in Equation (\ref{equ:marginloss}) is applied as the classification loss. In the class-conditional reconstruction process, a masking mechanism is applied to output capsules, where the capsules, corresponding to non-ground-truth classes, are masked with zeros. The input image is reconstructed from the masked output capsules. The reconstruction loss is used to regularize the training process.
By comparing the two networks, we can summarize 5 major differences between ConvNets and CapsNets, namely, a transformation process, a dynamic routing layer, a squashing function, the use of a marginal loss instead of a cross-entropy loss, and a class-conditional reconstruction regularization. With these differences, CapsNet outperforms ConvNet-FC in terms of robustness to affine transformation and overlapping digits recognition as well as in learning compact semantic representations. In this section, we will investigate these advantages one by one. In each of our studies, we attempt to answer the following questions:
\vspace{-0.1cm}
\begin{enumerate}
\setlength\itemsep{0.em}
\item Do ConvNet-FC and CapsNets perform differently?
\item Which components of CapsNets make the difference?
\item How bridge the gap between the two networks?
\end{enumerate}
\subsection{Robustness to Input Affine Transformation}
\label{sec:abl_aff}
\textbf{Settings}: To examine the transformation robustness of both models, we use the popular benchmark~\cite{sabour2017dynamic,gu2020improving} where models are trained on MNIST and tested on AffNIST. In AffNIST~\cite{sabour2017dynamic}, the original 28$\times$28 MNIST images are first padded with 6 pixels to 40$\times$40 image and then affine transformed, namely, rotation within 20 degrees, shearing within 45 degrees, scaling from 0.8 to 1.2 in both vertical and horizontal directions, and translation within 8 pixels in each direction. In the training dataset, the 28$\times$28 MNIST images are placed randomly on a black background of 40$\times$40 pixels without further transformation. The image examples are visualized in Fig.~\ref{fig:data_vis}. The performance on both MNIST and AffNIST test datasets is reported. All scores are averaged over 5 runs across this paper.
Besides ConvNet-FC and CapsNet, we include the state-of-the-art model on the benchmark in this experiment, namely, Aff-CapsNet. It simplifies CapsNet by removing dynamic routing and sharing the transformation matrix in the transformation process.
Following~\cite{sabour2017dynamic,gu2020improving}, the Adam optimizer is used to train the models with an initial learning rate of 0.001 and a batch size of 128. In CapsNet, the reconstruction loss is scaled down by 0.0005 so that it does not dominate the margin loss during training. It is hard to decide which model is more robust to affine transformations when they achieved different accuracy on untransformed examples. To eliminate this confounding factor, we stopped training the models when they achieve similar performance (i.e. about 99.22\%), following~\cite{sabour2017dynamic}.
\begin{table}[h]
\begin{center}
\footnotesize
\setlength\tabcolsep{0.14cm}
\begin{tabular}{P{2.8cm}P{1.4cm}P{1.4cm}P{1.4cm}}
\toprule
Models & \#Para. & MNIST & AffNIST \\
\midrule
GE-CapsNet~\cite{lenssen2018group}& - & 98.42 & 89.10 \\
SPARSECAPS~\cite{rawlinson2018sparse} & - & 99 & 90.12 \\
SCAE~\cite{kosiorek2019stacked} & - & 98.5 & 92.21 \\
EM-CapsNet~\cite{hinton2018matrix} & - & 99.2 & 93.1 \\
\midrule
ConvNet-FC~\cite{sabour2017dynamic} & 35.4M & 99.22 & 66 \\
CapsNet~\cite{sabour2017dynamic} & 13.5M & 99.23 & 79 \\
CapsNet-NoR~\cite{gu2020improving} & 13.5M & 99.22 & 81.81 \\
Aff-CapsNet-DR~\cite{gu2020improving} & 7.5M & 99.22 & 89.03 \\
Aff-CapsNet~\cite{gu2020improving} & 7.5M & 99.23 & 93.21 \\
\midrule
\textbf{ConvNet-Avg} & 5.3M & 99.22 & \textbf{94.11} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Comparison on the transformation robustness benchmark: The generalization performance to AffNIST is reported when models achieve similar performance on MNIST test dataset. Our simple ConvNet-Avg is more robust than CapsNet to input affine transformations.}
\label{tab:affNoRSOTA}
\vspace{-0.1cm}
\end{table}
\begin{table*}[t]
\begin{center}
\footnotesize
\begin{tabular}{c | c | c | c | c | c | c | c | c}
\hline
Factors & Routing & Shared TransM & Squash-fn & Reconstion & Loss & Train-MNIST & Test-MNIST & Test-AffNIST \\
\hline
\hline
\multirow{2}{*}{Routing} & \textcolor{colormark}{NoR} & \checkmark & \checkmark & \checkmark & MarginLoss & 100 & 99.29{\tiny ($\pm$ 0.13)} & 93.55{\tiny ($\pm$ 1.47)} \\
\cline{2-9}
& \textcolor{colormark}{DR} & - & - & - & - & 100 & 99.21{\tiny ($\pm$ 0.31)} & \textbf{90.07}{\tiny ($\pm$ 0.98)} \\
\hline
\hline
\multirow{2}{*}{Shared TransM} & NoR & \textcolor{colormark}{$\pmb{\checkmark}$} & \checkmark & \checkmark & MarginLoss & 100 & 99.29{\tiny ($\pm$ 0.13)} & 93.55{\tiny ($\pm$ 1.47)} \\
\cline{2-9}
& - & \textcolor{colormark}{$\pmb{\times}$} & - & - & - & 100 & 98.98{\tiny ($\pm$ 0.04)} & \textbf{80.49}{\tiny ($\pm$ 0.34)} \\
\hline
\hline
\multirow{2}{*}{Squash-fn} & NoR & \checkmark & \textcolor{colormark}{$\pmb{\checkmark}$} & \checkmark & MarginLoss & 100 & 99.29{\tiny ($\pm$ 0.13)} & 93.55{\tiny ($\pm$ 1.47)} \\
\cline{2-9}
& - & - & \textcolor{colormark}{$\pmb{\times}$} & - & - & 99.75 & 97.93{\tiny ($\pm$ 0.13)} & \textbf{80.42}{\tiny ($\pm$ 0.39)} \\
\hline
\hline
\multirow{3}{*}{Reconstruction} & NoR & \checkmark & \checkmark & \textcolor{colormark}{conditional} & MarginLoss & 100 & 99.29{\tiny ($\pm$ 0.13)} & 93.55{\tiny ($\pm$ 1.47)} \\
\cline{2-9}
& - & - & - & \textcolor{colormark}{normal} & - & 100 & 99.43{\tiny ($\pm$ 0.28)} & 95.09{\tiny ($\pm$ 0.56)} \\
\cline{2-9}
& - & - & - & \textcolor{colormark}{$\pmb{\times}$} & - & 100 & 99.39{\tiny ($\pm$ 0.26)} & 93.49{\tiny ($\pm$ 0.46)} \\
\hline
\hline
\multirow{2}{*}{Loss} & NoR & \checkmark & \checkmark & \checkmark & \textcolor{colormark}{MarginLoss} & 100 & 99.29{\tiny ($\pm$ 0.13)} & 93.55{\tiny ($\pm$ 1.47)} \\
\cline{2-9}
& - & - & - & - & \textcolor{colormark}{CE Loss} & 100 & 99.27{\tiny ($\pm$ 0.05)} & 94.67{\tiny ($\pm$ 0.43)} \\
\hline
\end{tabular}
\end{center}
\caption{The performance on MNIST training dataset, MNIST test dataset, and AffMNIST test dataset are reported, respectively (in percentage \%). Dynamic Routing (DR) and Margin loss are even harmful to the transformation robustness, while the squashing function (Squash-fn) and the shared transformation matrix (Shared TransM) are beneficial.}
\label{tab:perf_affcaps}
\end{table*}
\textbf{Results and Analysis}: The performance is reported in Tab.~\ref{tab:affNoRSOTA}. We can observe that there is a gap between ConvNet-FC and CapsNet. As reported in~\cite{sabour2017dynamic,gu2020improving}, the CapsNet outperforms ConvNet-FC, and Aff-CapsNet outperforms CapsNet. We take Aff-CapsNet (a simplified CapsNet) as a baseline and conduct further ablation studies on the components of CapsNet in Tab.~\ref{tab:perf_affcaps}. We report the model test performance on both un-transformed MNIST test images and novel affine-transformed ones. No early stopping is applied in the ablation studies.
The transformation process can be seen as a fully connected (FC) layer since the transformation matrices therein are equivalent to the parameters of an FC layer. \textit{Why is Aff-CapsNet more robust than CapsNet?} The transformation robustness of CapsNet can be improved by sharing the transformation matrix. When the transformation matrix is shared and no routing is applied in Aff-CapsNet, the transformation process is essential to conduct group $1\times1$ convolutional operations, global average pooling operations, and an average operation on the pooling results of different groups. A further study shows that the number of groups has no effect on the robustness (see Supplement A). Hence, we attribute the superior performance of the sharing transformation matrix to the global average pooling operation. \textit{Why is CapsNet more robust than ConvNet-FC?} The ConvNet-FC has two fully connected layers, while CapsNet has a functionally similar one. Another difference between them is the kernel size. Our study shows that large kernels are also beneficial to achieve transformation robustness (see Tab.~\ref{tab:perf_affcaps_kernel}). This argument also echoes our claim above. Namely, both global average pooling and large kernels improve the robustness by increasing receptive fields.
In Tab.~\ref{tab:perf_affcaps}, the dynamic routing is even harmful to the transformation robustness, which is also supported by the Tab.~\ref{tab:affNoRSOTA}. In addition, when no squashing function is applied, CapsNet has to regress the capsule length to extreme values (e.g., 0 or 1), which is a hard task and leads to unsatisfying performance (even on the training dataset). The margin loss can slightly weaken the transformation robustness of CapsNet, while reconstruction makes no difference to it. The non-conditional reconstruction slightly improves the performance since it updates all capsules in each training iteration.
Based on our findings, we propose a new simple ConvNet baseline, called \textbf{ConvNet-Avg}. It starts with the two convolutional layers and terminates with a global average pooling and an output layer, which is also a common architecture used in image classification. The cross-entropy loss is applied to train the model. To make a fair comparison, we use the same convolutional layers as in CapsNet and Aff-CapsNet, namely, Conv(256, 9, 1) + Conv(256, 9, 2) + Global AvgPool + FC(10) (see Fig.~\ref{fig:overview}). It is hard to decide which model is better at generalizing to affine transformations when they achieved different accuracy on untransformed examples. We follow previous work and stop training the models when they achieve similar test performance (99.22\%). As shown in Tab.~\ref{tab:affNoRSOTA}, our simple ConvNet-Avg achieves slightly better performance with fewer parameters.
\textbf{Conclusions}: 1) Compared to ConvNet-FC, CapsNet achieves better test performance with fewer parameters on AffNIST. We attribute the gap to the kernel size. 2) Dynamic routing can harm the transformation robustness of CapsNet. When the routing is removed, the uniform average of votes (i.e., NoR) aggregates the global information better. 3) Our baseline ConvNet-Avg outperforms CapsNets significantly. It consists of only convolutional layers and a global average pooling layer, and no advanced component from SOTA ConvNets. The simplicity of ConvNet-Avg indicates that CapsNets are even less robust to affine transformation than ConvNets in a fair comparison.
\begin{table*}[t]
\begin{center}
\footnotesize
\begin{tabular}{c | P{0.6cm}P{0.6cm}P{0.6cm} | P{0.6cm}P{0.6cm}P{0.6cm} | P{0.6cm}P{0.6cm}P{0.6cm} | P{0.6cm}P{0.6cm}P{0.6cm} | P{0.6cm}P{0.6cm}P{0.6cm}}
\hline
Kernels & \multicolumn{3}{c|}{\textcolor{colormark}{K(3, 3)}} & \multicolumn{3}{c|}{\textcolor{colormark}{K(5, 5)}} & \multicolumn{3}{c|}{\textcolor{colormark}{K(7, 7)}} & \multicolumn{3}{c|}{\textcolor{colormark}{K(9, 9)}} & \multicolumn{3}{c}{\textcolor{colormark}{K(11, 11)}} \\
\hline
Models & \#Para. & A$_{std}$ & A$_{aff}$ & \#Para. & A$_{std}$ & A$_{aff}$ & \#Para. & A$_{std}$ & A$_{aff}$ & \#Para. & A$_{std}$ & A$_{aff}$ & \#Para. & A$_{std}$ & A$_{aff}$ \\
\hline
CapsNet & 16.1M & 96.31 & 61.36 & 14.4M & 98.18 & 70.34 & 13.5M & 98.74 & 75.82 &13.5M & 99.26 & 79.12 & 14.3M & 99.1 & 86.79 \\
\hline
ConvNet-FC & 49.5M & 96.54 & 64.57 & 35.4M & 99.23 & 66.08 & 25.2M & 99.03 & 66.76 & 18.8M & - & - & 16.19M & - & - \\
\hline
ConvNet-Avg & 0.59M & 97.14 & \textbf{86.58} & 1.70M & 98.58 & \textbf{90.95} & 3.23M & 99.1 & \textbf{92.31} & 5.30M & 99.22 & \textbf{94.11} & 7.96M & 99.34 & \textbf{90.58} \\
\hline
\end{tabular}
\end{center}
\caption{The effect of the kernel sizes on the transformation robustness of different models: Both standard accuracy (A$_{std}$) and the generalization accuracy (A$_{aff}$) on transformed data are reported. The large kernels make positive contributions to the transformation robustness. When the same kernel size is applied, ConvNet-Avg outperforms both ConvNet-FC and CapsNet.}
\label{tab:perf_affcaps_kernel}
\end{table*}
\subsection{Recognizing overlappping digits}
\label{sec:over_digits}
\textbf{Settings}: The work~\cite{sabour2017dynamic} shows that the CapsNet is able to recognize overlapping digits by segmenting them. To check this property, we use the MultiMNIST dataset, which is generated by overlaying a digit on top of another digit but from a different class. Specifically, a 28$\times$28 MNIST image with a digit is first shifted up to 4 pixels in each direction resulting in a 36$\times$36 image. The resulting image is overlaid to another image from different classes but the same set (training dataset or test dataset). For each image in MNIST, we can create $N$ (from 1 to 1K) images. See Fig.~\ref{fig:data_vis_MultiMNIST} for some examples from data.
The classification of an image with overlapping digits is correct if both digits are correctly classified (the top 2 output classes match the ground truth). The margin loss can be applied to compute the classification loss. In the ConvNet baselines, the sigmoid function is applied to logits instead of softmax to obtain output probabilities since this is a multi-target classification task, and the binary cross-entropy loss is applied to compute the classification loss.
In the training process, the CapsNet is first applied to the overlapping digits to obtain output capsules. During reconstruction, a ground-truth class is picked at a time, and the capsule corresponding to the class is kept for the reconstruction while others are masked with zeros. In other words, we run the reconstruction sub-network twice, each for one digit. The reconstruction loss can be computed similarly since the images of individual digits are available.
\begin{table*}[t]
\begin{center}
\footnotesize
\begin{tabular}{c | c | c | c | c | c | c | c }
\hline
Factors & Routing & Shared TransM & Squash-fn & Reconstion & Loss & Train-MultiMNIST & Test-MultiMNIST \\
\hline
\hline
\multirow{2}{*}{Routing} & \textcolor{colormark}{DR} & $\pmb{\times}$ & \checkmark & \checkmark & MarginLoss & 94.03 & 93.26{\tiny ($\pm$ 0.24)} \\
\cline{2-8}
& \textcolor{colormark}{NoR} & - & - & - & - & 90.28 & 90.07{\tiny ($\pm$ 0.29)} \\
\hline
\hline
\multirow{2}{*}{Shared TransM} & DR & \textcolor{colormark}{$\pmb{\times}$} & \checkmark & \checkmark & MarginLoss & 94.03 & 93.26{\tiny ($\pm$ 0.24)} \\
\cline{2-8}
& - & \textcolor{colormark}{$\pmb{\checkmark}$} & - & - & - & 86.92 & \textbf{86.44}{\tiny ($\pm$ 0.37)} \\
\hline
\hline
\multirow{2}{*}{Squash-fn} & DR & $\pmb{\times}$ & \textcolor{colormark}{$\pmb{\checkmark}$} & \checkmark & MarginLoss & 94.03 & 93.26{\tiny ($\pm$ 0.24)} \\
\cline{2-8}
& - & - & \textcolor{colormark}{$\pmb{\times}$} & - & - & 87.71 & \textbf{87.24}{\tiny ($\pm$ 0.53)} \\
\hline
\hline
\multirow{3}{*}{Reconstruction} & DR & $\pmb{\times}$ & \checkmark & \textcolor{colormark}{conditional} & MarginLoss & 94.03 & 93.26{\tiny ($\pm$ 0.24)} \\
\cline{2-8}
& - & - & - & \textcolor{colormark}{normal} & - & 93.83 & 93.19{\tiny ($\pm$ 0.30)} \\
\cline{2-8}
& - & - & - & \textcolor{colormark}{$\pmb{\times}$} & - & 90.28 & 90.17{\tiny ($\pm$ 0.26)} \\
\hline
\hline
\multirow{2}{*}{Loss} & DR & $\pmb{\times}$ & \checkmark & \checkmark & \textcolor{colormark}{MarginLoss} & 94.03 & 93.26{\tiny ($\pm$ 0.24)} \\
\cline{2-8}
& - & - & - & - & \textcolor{colormark}{BCE Loss} & 91.64 & 91.19{\tiny ($\pm$ 0.35)} \\
\hline
\end{tabular}
\end{center}
\caption{The ablation study on components of CapsNet: The performance of models trained on 6M overlapping digits. All individual components make positive contributions to the ability to recognize overlapping digits. The transformation matrices contribute the most; the performance drops dramatically if a shared transformation matrix is applied.}
\label{tab:perf_multi_mnist}
\end{table*}
\textbf{Results and Analysis}: The overlapping digit recognition performance is reported in Tab.~\ref{tab:perf_multi_mnist} where the individual components of CapsNets are ablated. The reconstruction sub-network helps to improve the recognition performance. However, it does not have to be class-conditional. The reconstruction loss regularizes the training process so that the information about both digits is encoded in features and high-level capsules. The margin loss can be directly applied to a multi-target classification task, which outperforms the standard binary cross-entropy loss. Both the reconstruction and the margin loss can be applied to enhance a ConvNet.
When a vector representation is applied, the squashing function plays an important role. When applying the squashing function to the primary capsules, the feature maps are group-wise normalized. The information is communicated across different channels, which can help to better disentangle overlapping digits. Additionally, CapsNet has to regress the non-squashed capsule length to certain values. Since the regression task is hard, CapsNets achieve unsatisfying performance on both the training and test dataset. The analysis echoes the one in Sec.~\ref{sec:abl_aff}.
The dynamic routing process identifies the weights for votes, which results in a higher modeling capacity than the uniform averaging operation on votes. Other components that support the CapsNet's modeling capacity are the transformation matrices. When a shared transformation matrix is applied, the model performance drops dramatically. We check the ConvNet-Avg on this task and observe that CapsNet outperforms ConvNet-Avg significantly. The reason behind this is that the global pooling operation can be harmful for recognizing overlapping digits since it aggregates a feature map into a single unit. The convolutional layer itself is not able to disentangle the overlapping digits into different feature maps. In CapsNet, the transformation process acts as a fully connected layer, which avoids the global average pooling. Hence, we argue that the high modeling capacity is the essential reason why CapsNet performs well on the overlapping digits recognition task.
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{figures/multimnist_accu}
\caption{ConvNet-FC(LK) outperforms CapsNet on MultiMNIST dataset with different data sizes.}
\label{fig:datasize_MultiMNIST}
\end{figure}
FC layers in ConvNet-FC can maintain richer information (features at all locations) for distinguishing overlapping digits. Note that the baseline ConvNet-FC in~\cite{sabour2017dynamic} has a smaller kernel size than in CapsNet. Hence, we propose to apply ConvNet-FC with large kernels (ConvNet-FC(LK)) to this overlapping digits recognition task. In ConvNet-FC(LK), we also reduce the units of fully connected layers to save parameters so that it can be compared to CapsNets. When the same large kernel is applied, ConvNet-FC(LK) outperforms the CapsNet and sets a new SOTA on this benchmark (97.11\% vs. 95.18\%). When different training data sizes and different kernel sizes are applied in the experiments, the simple ConvNet-FC(LK) outperforms the CapsNet consistently (See Fig.~\ref{fig:datasize_MultiMNIST} and Supplement B).
\textbf{Conclusions}: 1) All the components contribute to the ability of CapsNet to recognize overlapping digits. 2) The transformation process with a non-shared transformation matrix and a dynamic routing to weight votes bring high modeling capacity, which essentially supports the high performance of CapsNet in this task. 3) The simple ConvNet-FC(LK) with similar parameters performs better than CapsNet on this benchmark, which indicates that CapsNet is not more robust than ConvNet to recognize overlapping digits.
\subsection{Semantic Capsule Representations}
\textbf{Settings:} In CapsNets, when a single element in a capsule is perturbated, the reconstructed images are visually changed correspondingly~\cite{sabour2017dynamic}, see Fig.~\ref{fig:ConvNet_Caps_perb}. The visual changes often correspond to human-understandable semantic object variations. In this experiment, we investigate which components support the semantic representations. Since this property is mainly demonstrated by a reconstruction sub-network, we introduce three models below:
\begin{figure*}
\centering
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/ConvNet_r_vis}
\caption{on ConvNet-R}
\label{fig:ConvNet_R_perb}
\end{subfigure}\hspace{0.2cm}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/ConvNet_cr_vis}
\caption{on ConvNet-CR}
\label{fig:ConvNet_CR_perb}
\end{subfigure}\hspace{0.2cm}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/capsnet_no_rout_vis}
\caption{on ConvNet-CR-SF}
\label{fig:ConvNet_CR_SF_perb}
\end{subfigure}\hspace{0.2cm}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/capsnet_agreement_vis}
\caption{on CapsNet}
\label{fig:ConvNet_Caps_perb}
\end{subfigure}
\caption{The reconstructed images are shown when a single unit is perturbed. The reconstruction only helps when the class-conditional masking mechanism is applied. The squashing function improves the visual response further.}
\label{fig:comapct_rep_vis}
\end{figure*}
\begin{figure*}
\centering
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/ConvNet_R}
\caption{Reconstruction in ConvNet-R}
\label{fig:ConvNet_R_recons}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/ConvNet_CR}
\caption{Reconstruction in ConvNet-CR}
\label{fig:ConvNet_CR_recons}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/ConvNet_CR_SF}
\caption{Reconstruction in ConvNet-CR-SF}
\label{fig:ConvNet_CR_SF_recons}
\end{subfigure}
\caption{The reconstruction from feature space to the input space: In ConvNet-R, the capsule representations of different classes are entangled in feature space; the ones in ConvNet-CR are clearly separated due to the class-conditional masking mechanism. When a squashing function is applied to squash the vector, the representations live within a manifold. The representation constraints improve the network's ability to extrapolate object variations.}
\label{fig:recon_analysis}
\end{figure*}
\textbf{ConvNet-CR}: This ConvNet baseline has the same number of parameters as in CapsNet and the same reconstruction sub-network. Its architecture is Conv(256, 9, 1) + Conv(256, 9, 2) + FC(160), where 160 corresponds to the dimensions of output capsules and the parameters in FC(160) corresponds to the non-shared transformation matrices of CapsNet. The 160 activations are grouped into 10 groups where each group corresponds to an output capsule. The sum of 16 activations in each vector corresponds to a logit. The sigmoid function is applied to each logit to obtain the output probability. The reconstruction sub-network reconstructs the input from (the 160 activations) with a masking mechanism, similar to that in CapsNet.
\textbf{ConvNet-R}: In this baseline, an output layer FC(10) is built on the 160 activations of ConvNet-CR instead of grouping them. The reconstruction sub-network of ConvNet-CR reconstructs the input from the 160 activations directly, without the masking mechanism.
\textbf{ConvNet-CR-SF}: This baseline equips ConvNet-CR with the squashing function in Equation (\ref{equ:squa}). The feature maps from Conv(256, 9, 2) are mapped into vectors with the same shape of primary capsules, and the vectors are squashed. Each element of the vectors is fully connected to 160 units of the next layer. The 160 activations are grouped to obtain the 10 output vectors. The vectors are similarly squashed so that their lengths stand for the output probability of the corresponding class. This baseline is equivalent to CapsNet without a routing mechanism (CapsNet-NoR).
In CapsNet, several units can correspond to a similar semantic concept. An interesting question to investigate is, what percentage of neurons strongly react to changes of a given latent factor. We propose a metric to evaluate such compactness. Given a latent factor $z$ (e.g. rotation) and an image X, we compute the semantic compactness score with the following steps:
\begin{enumerate}
\setlength\itemsep{0.em}
\item \hspace{-0.15cm} Creating a list of images with different rotation degrees;
\item \hspace{-0.15cm} Obtaining their representation vectors via forward inferences (the vectors of ground-truth classes are kept);
\item \hspace{-0.15cm} Computing the variance of the vectors in each dimension $\pmb{Var}$ and normalize them by their sum $\pmb{Var}_n$;
\item \hspace{-0.15cm} Computing the KL divergence between the normalized variance values $\pmb{Var}_n$ and a uniform prior.
\end{enumerate}
\vspace{-0.1cm}
The compactness score is averaged over the whole dataset. The higher the score is, the more compact the semantic representation becomes. The intuition behind the score is that, if only one unit changes when images are rotated, the normalized variance will be one-hot, and the relative entropy to uniform prior is the maximum.
\textbf{Results and Analysis}: After training, we perform the capsule perturbation experiments on the 160 activations, as in~\cite{sabour2017dynamic}. In CapsNet, we tweak one dimension of capsule representations by intervals of 0.05 in the range [-0.2, 0.2]. The reconstructed images are visualized in Fig.~\ref{fig:ConvNet_Caps_perb}. The semantic changes of images can be observed, e.g., the rotation and the stroke thickness. We find that the reconstructed images in ConvNets stay almost unchanged visually when perturbing the corresponding activation with the same range. The observation can be caused by the too-small perturbation range for the unit activations. Hence, we increase the range gradually until the reconstructed image cannot be recognized where we reach the range of [-8, 8]. The reconstructed images are shown in Fig.~\ref{fig:comapct_rep_vis}. In ConvNet-R, the semantics of reconstructed images is not sensitive to all individual dimensions in Fig.~\ref{fig:ConvNet_R_perb}. In ConvNet-CR, where the class-conditional reconstruction is applied, the changes of representation unit also cause the semantic changes of reconstructed images in Fig.~\ref{fig:ConvNet_CR_perb}. When the squashing function is applied, the representations in ConvNet-CR-CF strongly react to the perturbations in Fig.~\ref{fig:ConvNet_CR_SF_perb}.
\begin{table}[t]
\begin{center}
\footnotesize
\setlength\tabcolsep{0.06cm}
\begin{tabular}{P{1.9cm}P{0.cm}P{1.cm}P{1.cm}P{0.8cm}P{1.cm}P{1.cm}}
\toprule
Datasets & \multicolumn{6}{|c}{MNIST} \\
\midrule
Factors & \multicolumn{1}{|c}{Rotation} & Trans-X & Trans-Y & Scale & Shear-X & Shear-Y \\
\midrule
ConvNet-R & \multicolumn{1}{|c}{0.0003} & 0.0016 & 0.0009 & 0.0004 & 0.0003 & 0.0007 \\
\midrule
ConvNet-CR & \multicolumn{1}{|c}{0.0028} & 0.0038 & 0.0032 & 0.0052 & 0.0058 & 0.0022 \\
\midrule
ConvNet-CR-SF & \multicolumn{1}{|c}{\textbf{0.0325}} & \textbf{0.2010} & \textbf{0.3192} & \textbf{0.0146} & \textbf{0.0476} & \textbf{0.0506} \\
\midrule
CapsNet & \multicolumn{1}{|c}{0.0031} & 0.0107 & 0.0464 & 0.0026 & 0.0098 & 0.0021 \\
\bottomrule
\end{tabular}
\end{center}
\caption{The representation compactness: The class-conditional reconstruction and the squashing function improve the compactness, while dynamic routing reduces it.}
\label{tab:comapct_rep_eval}
\end{table}
Both the class-conditional reconstruction mechanism and the squashing function can help ConvNets to learn meaningful semantic representations. The two components characterize the function learned by the reconstruction sub-network, which maps representations from feature space back to input space. We illustrate the characteristics of these functions in Fig.~\ref{fig:recon_analysis}, using an example with a 2D input space and 3 output classes. The ConvNet-R reconstructs inputs from the features that are entangled to some degree. In ConvNet-CR, the features of different classes are perfectly separated since the features are class-conditional. The ConvNet-CR-CF constrains the feature space further by squashing the vectors so that they live inside a manifold. We also report the compactness score of each model in Tab.~\ref{tab:comapct_rep_eval}. We speculate that it is these constraints that improve the representation's compactness. More experiments on the FMNIST dataset can be found in Supplement C.
\textbf{Conclusions}: Both the class-conditional reconstruction and the squashing function help CapsNet learn meaningful semantic representations, while dynamic routing is even harmful. The two components can be integrated into ConvNets, where ConvNet-CR-SF learns better semantic compact representations than CapsNets.
\section{Conclusion}
We reveal 5 major differences between CapsNets and ConvNets and study 3 properties of CapsNets. We show that dynamic routing is harmful to CapsNets in terms of transformation robustness and semantic representations. In each presented task, a simple ConvNet can be built to outperform the CapsNet significantly. We find that there is no single ConvNet that can outperform CapsNet in all cases. Hence, we conclude that \textit{CapsNets with dynamic routing are not more robust than ConvNets}. We leave further explorations for future work, e.g., concerning different datasets, and other properties of CapsNets, and other CapsNets.
The dynamic routing aggregates information from low-level entities into high-level ones. The aggregation can be also be done by a graph pooling operation~\cite{gu2020interpretable}. In ConvNets, the relationship between low-level entities is also explored in aggregation~\cite{hu2018relation,hu2019local}. More aggregation approaches wil be explored in future work.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
{\em Frozen planet orbits} are periodic orbits in the helium atom
which play an important role in its semiclassical
treatment~\cite{tanner-richter-rost, wintgen-richter-tanner}. In such an
orbit both electrons lie on a line on the same side of the
nucleus. The inner electron undergoes consecutive collisions with the
nucleus while the outer electron (the actual ``frozen planet'') stays
almost stationary at some distance. See the following figure.
\bigskip
\setlength{\unitlength}{1mm}
\begin{picture}(1,1)
\put(0,0){\circle*{4}}
\put(2,0){\line(1,0){9}}
\put(22,0){\vector(-1,0){11}}
\put(23,0){\circle*{2}}
\put(24,0){\vector(1,0){11}}
\put(35,0){\line(1,0){9}}
\put(68,0){\line(1,0){2}}
\put(74,0){\vector(-1,0){4}}
\put(75,0){\circle*{2}}
\put(76,0){\vector(1,0){4}}
\put(80,0){\line(1,0){2}}
\put(22,-3){$q_2$}
\put(74,-3){$q_1$}
\end{picture}
\bigskip
\bigskip
An interesting aspect of frozen planet orbits is that they cannot be
obtained using perturbative
methods. Indeed, if the interaction between the two electrons is
switched off both electrons just fall into the nucleus. In order to
deal with this problem, the second author replaced in~\cite{frauenfelder}
the instantaneous interaction between the two electrons by a mean
interaction and showed analytically that in this case there exists a
unique nondegenerate frozen planet orbit.
\smallskip
{\bf Variational setup. }
In this paper we present a variational approach to frozen planet
orbits with instantaneous or mean interaction. One difficulty
lies in the collisions of the inner electron with the nucleus, which
need to be regularized in order to obtain a good fuctional analytic
setup. A traditional way to regularize two-body collisions is the
Levi-Civita regularization \cite{levi-civita}. In case of mean
interactions our problem has delay and the application of the
traditional Levi-Civita regularization becomes
problematic. Fortunately, in a recent interesting paper by Barutello,
Ortega and Verzini \cite{barutello-ortega-verzini} a new nonlocal
regularization was discovered. This regularization is motivated by the
traditional Levi-Civita regularization, but in sharp contrast to the
latter it is defined on the loop space and therefore fits well with
our problem. Even for loops without collisions this transformation is
quite intriguing. It is not smooth in the usual sense but scale smooth
in the sense of Hofer, Wysocki and Zehnder~\cite{hofer-wysocki-zehnder}.
\smallskip
We study two functionals $\mathcal{B}_{av}$ and $\mathcal{B}_{in}$
arising from regularizing frozen planet configurations for the mean
and instantaneous interaction, respectively, as well as their
linear interpolation $\mathcal{B}_r=r\mathcal{B}_{in}+(1-r)\mathcal{B}_{av}$, $r\in[0,1]$.
In general, it is not
clear that critical points of a regularized action functional
correspond precisely to the regularized solutions. It might happen
that new exotic solutions appear as critical points,
see~\cite{barutello-ortega-verzini} for examples of this
phenomenon. Excluding such exotic critical points requires a careful
analysis and this analysis occupies the main part of this paper. In
particular, we prove (see Theorem~\ref{thm:interpol})
\smallskip
\textbf{Theorem\,A: }
For each $r\in[0,1]$, critical points of the regularized action
functional $\mathcal{B}_r$ correspond precisely to frozen planet orbits for
the interpolated interaction.
\smallskip
{\bf Symmetries. }
There is a special case of frozen planet orbits referred to as
\emph{symmetric frozen planet orbits}. These are frozen planet orbits
in which the outer electron has vanishing velocity whenever the inner
electron collides with the nucleus or is at maximal distance from the
nucleus, see~\cite{frauenfelder2}. We have (see
Theorem~\ref{thm:interpol-symm})
\smallskip
\textbf{Theorem\,B: }
The regularized action functional $\mathcal{B}_r$ is invariant under an
involution such that the critical points fixed by the involution are
precisely the symmetric frozen planet orbits.
\smallskip
In view of Theorem\,B one can think of symmetric frozen planet orbits as a
nonlocal generalization of brake orbits.
\smallskip
{\bf Regularity. }
In order to study critical points of $\mathcal{B}_r$, we introduce its
$L^2$-gradient as a map $\nabla\mathcal{B}_r:X\to Y$ from a suitable Hilbert
manifold $X$ to a Hilbert space $Y$. It satisfies (see
Theorem~\ref{thm:Fredholm})
\smallskip
\textbf{Theorem\,C: }
For each $r\in[0,1]$, the $L^2$-gradient $\nabla\mathcal{B}_r:X\to Y$ is a
$C^1$-Fredholm map of index $0$.
\smallskip
This result is nontrivial because $\nabla\mathcal{B}_r$ involves time
reparametrizations depending on points of $X$ as well as singular terms.
Inspection of its proof shows that $\nabla\mathcal{B}_r$ is {\em not} of class $C^2$.
Theorem\,C makes the functionals $\mathcal{B}_r$ amenable to classical
variational methods such as index theory or Morse theory. In this
paper we content ourselves with computing the {\em mod $2$ Euler
number} $\chi(\nabla\mathcal{B}_r)$, i.e., the count of zeroes modulo $2$
(after suitable perturbation, see Appendix~\ref{sec:Euler}). On
symmetric loops we find
$$
\chi(\nabla\mathcal{B}_{in}) = \chi(\nabla\mathcal{B}_{av})=1,
$$
where the first equality follows from homotopy invariance of the mod
$2$ Euler number and the second one from a further deformation and
explicit computation (see Appendix~\ref{sec:Euler-mean-interaction}).
As a consequence, we obtain (see Theorem~\ref{thm:existence}):
\textbf{Corollary\,C: }
For every $E<0$ there exists a symmetric frozen planet orbit of energy $E$.
\smallskip
{\bf Hamiltonian formulation. }
The regularized action functional $\mathcal{B}_r$ has an intriguing
structure. It consists of two terms. The first term can be interpreted
as a kinetic energy, but for a nonlocal metric which depends on the
whole loop. The second term can be interpreted as the negative of a
nonlocal potential which is defined on the loop space of the
configuration space. We explain how in this situation a \emph{nonlocal
Legendre transform} can be carried out which produces a nonlocal
Hamiltonian (see Section~\ref{sec:Ham}):
\smallskip
\textbf{Theorem\,D: }
After applying the nonlocal Legendre transform to the regularized
action functional $\mathcal{B}_r$, the corresponding Hamiltonian delay
equation reproduces frozen planet orbits.
\smallskip
Thus there are two nonlocal approaches to frozen planet orbits, one
Lagrangian and one Hamiltonian. This produces food for thought for
many interesting research projects. For instance, the Lagrangian
action functional has a Morse index at each frozen planet orbit. On
the other hand, the Hessian of the Hamiltonian action functional is a
Fredholm operator which gives rise to a nonlocal Conley-Zehnder index
at each frozen planet orbit. Since in the local case the Morse index
corresponds to the Conley-Zehnder index after Legendre transform we
may ask
\smallskip
\textbf{Question\,1: }
How are the Morse index and the Conley-Zehnder index for frozen planet
orbits related?
\smallskip
The correspondence between these indices is an important ingredient in
the celebrated adiabatic limit argument by Salamon and Weber
\cite{salamon-weber} relating Floer homology with the heat flow on
chain level.
\smallskip
\textbf{Question\,2: }
Is there an analogon of the adiabatic limit argument of Salamon and
Weber for frozen planet orbits?
\section{The Kepler problem}\label{sec:Kepler}
When the interaction between the electrons is ignored the system
decouples into two one-electron systems, each of which is equivalent
to the Kepler problem in celestial mechanics.
In this section we recall the regularization
procedure of Barutello, Ortega and Verzini~\cite{barutello-ortega-verzini}
for the Kepler problem in the plane. It is based on the Levi-Civita
regularization map $\mathcal{C}\to\mathcal{C}$, $z\mapsto q=z^2$. Since we are interested
in the case that the position $q$ of the electron remains on the
positive real line, we consider the restriction $\mathbb{R}\to\mathbb{R}_{\geq 0}$,
$z\mapsto q=z^2$ of the Levi-Civita map to the real line.
\subsection{Levi-Civita transformation}\label{ss:LC}
In this subsection we describe the Levi-Civita transformation
without worrying about the regularity of the involved maps; precise
statements will be given in the following subsections.
We abbreviate by $S^1=\mathbb{R}/\mathbb{Z}$ the circle. We denote the
$L^2$-inner product of $z_1,z_2\in L^2(S^1,\mathbb{R})$ we by
$$
\langle z_1,z_2\rangle := \int_0^1 z_1(\tau) z_2(\tau) d\tau,
$$
and the $L^2$-norm of $z \in L^2(S^1, \mathbb{R})$ by
$$
\|z\| := \sqrt{\langle z, z\rangle}.
$$
In the sequel we will work with Sobolev spaces $H^k=W^{k,2}$, but
the only relevant norms and inner products will be the ones from $L^2$.
Consider two maps
$$
q:S^1\to\mathbb{R}_{\geq 0},\qquad z:S^1\to\mathbb{R}
$$
related by the {\em Levi-Civita transformation}
\begin{equation}\label{eq:LC}
q(t) = z(\tau)^2
\end{equation}
for a time change $t\longleftrightarrow \tau$ satisfying
$0\longleftrightarrow 0$ and
\begin{equation}\label{eq:t-tau}
\frac{dt}{q(t)} = \frac{d\tau}{\|z\|^2}.
\end{equation}
This implies that the mean values of $q$ and $1/q$ are given by
\begin{equation}\label{bov1}
\overline{q} :=
\int_0^1 q(t)dt
= \int_0^1 \frac{z(\tau)^4}{\|z\|^2} d\tau
= \frac{\|z^2\|^2}{\|z\|^2}
\end{equation}
and
\begin{equation}\label{bov2}
\int_0^1\frac{dt}{q(t)} = \frac{1}{\|z\|^2}.
\end{equation}
We will denote derivatives with respect to $t$ by a dot and
derivatives with respect to $\tau$ by a prime. Then the first and
second derivatives of $q$ and $z$ (where they are defined) are related by
\begin{equation}\label{bov3}
\dot q(t)
= 2z(\tau)z'(\tau)\frac{d\tau}{dt} = \frac{2\|z\|^2z'(\tau)}{z(\tau)}
\end{equation}
and
\begin{equation}\label{eq:ddot-q-z}
\ddot q(t)
= 2\|z\|^2\frac{z''(\tau)z(\tau)-z'(\tau)^2}{z(\tau)^2}\frac{d\tau}{dt}
= \frac{2\|z\|^4}{z(\tau)^4}\bigl(z''(\tau)z(\tau)-z'(\tau)^2\bigr).
\end{equation}
Substituting $z^2$ and $z'^2$ by~\eqref{eq:LC} and~\eqref{bov3} this becomes
\begin{equation}\label{eq:ddot-q}
\ddot q(t)
= \frac{1}{q(t)}\Bigl(2\|z\|^4\frac{z''(\tau)}{z(\tau)} - \frac{\dot q(t)^2}{2}\Bigr).
\end{equation}
The $L^2$-norm of the derivative of $q$ is given by
\begin{eqnarray}\label{bov4}
\|\dot{q}\|^2
= \int_0^1 \dot{q}(t)^2 dt
= \int_0^1
\frac{4\|z\|^4z'(\tau)^2}{z(\tau)^2}\,\frac{z(\tau)^2}{\|z\|^2}d\tau
= 4\|z\|^2\|z'\|^2.
\end{eqnarray}
\subsection{Inverting the Levi-Civita transformation}\label{ss:LC-inv}
In this subsection we prove that, under suitable technical hypotheses,
the Levi-Civita transformation defines a 2-to-1 covering.
We begin with a precise definition of the Levi-Civita transformation.
Let $z\in C^0(S^1,\mathbb{R})$ be a continuous function with finite zero set
$$
Z_z:=z^{-1}(0).
$$
We associate to $z$ a $C^1$-map $t_z:S^1\to S^1$ by
\begin{equation}\label{eq:tz}
t_z(\tau) := \frac{1}{\|z\|^2}\int_0^\tau z(\sigma)^2d\sigma.
\end{equation}
Note that $t_z(0)=0$ and
\begin{equation}\label{eq:dert}
t_z'(\tau) = \frac{z(\tau)^2}{\|z\|^2}.
\end{equation}
Since $z$ has only finitely many zeroes, this shows that $t_z$ is
strictly increasing and we conclude
\begin{lemma}\label{lem:tz-homeo}
If $z\in C^0(S^1,\mathbb{R})$ has only finitely many zeroes, then
the map $t_z:S^1\to S^1$ defined by~\eqref{eq:tz} is a homeomorphism.
\hfill$\square$
\end{lemma}
It follows that $t_z \colon S^1 \to S^1$ has a continuous inverse
$$
\tau_z:=t_z^{-1} \colon S^1 \to S^1.
$$
Since $t_z$ is of class $C^1$, the function $\tau_z$ is also of class $C^1$ on the
complement of the finite set $t_z(Z)$ with derivative
\begin{equation}\label{eq:taudot}
\dot\tau_z(t) = \frac{\|z\|^2}{z(\tau_z(t))^2}.
\end{equation}
We define a continuous map $q:S^1\to \mathbb{R}_{\ge 0}$ by
\begin{equation}\label{eq:LC-def}
q(t) := z(\tau_z(t))^2.
\end{equation}
Then the two maps $z,q$ are related by the Levi-Civita
transformation~\eqref{eq:LC} with $\tau=\tau_z$. Their zero sets
$$
Z_z=z^{-1}(0)\quad \text{and}\quad Z_q:=q^{-1}(0)=t_z(Z_z)
$$
are in bijective correspondence via $t_z$ (or equivalently $\tau_z$).
Moreover, by~\eqref{bov2} we have
$$
\int_0^1\frac{ds}{q(s)} = \frac{1}{\|z\|^2} <\infty.
$$
Conversely, suppose we are given a map $q\in C^0(S^1,\mathbb{R}_{\geq 0})$
with finite zero set $Z_q$ satisfying $\int\frac{ds}{q(s)}<\infty$.
We associate to $q$ the time reparametrization $\tau_q:S^1\to S^1$,
\begin{equation}\label{eq:tauq}
\tau_q(t) :=
\Bigl(\int_0^1\frac{ds}{q(s)}\Bigr)^{-1} \int_0^t\frac{1}{q(s)}ds.
\end{equation}
Then $\tau_q$ is of class $C^1$ outside the zero set $Z_q=q^{-1}(0)$
with derivative
\begin{equation}\label{eq:tauqdot}
\tau_q'(t) =
\Bigl(\int_0^1\frac{ds}{q(s)}\Bigr)^{-1}\frac{1}{q(t)}, \qquad t\in
S^1\setminus Z_q.
\end{equation}
By~\cite[Lemma 2.1]{barutello-ortega-verzini} the map $\tau_q:S^1\to S^1$
is a homeomorphism whose inverse $t_q:=\tau_q^{-1}$ is of class $C^1$
and satisfies
\begin{equation}\label{eq:tqdot}
t_q'(\tau) = \Bigl(\int_0^1\frac{ds}{q(s)}\Bigr)\, q(t_q(\tau)),\qquad
\tau\in S^1.
\end{equation}
Suppose that $z:S^1\to\mathbb{R}$ is a continuous function satisfying
\begin{equation}\label{eq:LC-inv}
z(\tau)^2 = q(t_q(\tau)).
\end{equation}
Then $z$ has finite zero set $Z_z=\tau_q(Z_q)$, so we can associate to
$z$ the homeomorphism $t_z:S^1\to S^1$ defined by~\eqref{eq:tz} and
its inverse $\tau_z$. We claim that
\begin{equation}\label{eq:homeoqz}
\tau_q=\tau_z\quad \text{and}\quad t_q=t_z
\end{equation}
It is enough to check the second equality. For this we compute
$$
\int_0^\tau z(\sigma)^2d\sigma
=\int_0^\tau q(t_q(\sigma))d\sigma
\stackrel{(*)}{=}\Bigl(\int_0^1\frac{ds}{q(s)}\Bigr)^{-1}
\int_0^{t_q(\tau)}ds
=\Bigl(\int_0^1\frac{ds}{q(s)}\Bigr)^{-1}t_q(\tau),
$$
where $(*)$ follows from the coordinate change $\sigma=\tau_q(s)$
and~\eqref{eq:tauqdot}. Evaluating at $\tau=1$ gives us
$$
\frac{1}{\|z\|^2}=\int_0^1\frac{ds}{q(s)}.
$$
Therefore,
$$
t_z(\tau) = \frac{1}{||z||^2}\int_0^\tau z(\sigma)^2d\sigma = t_q(\tau)
$$
and~\eqref{eq:homeoqz} is established. Hence $q$ is the Levi-Civita
transform of $z$ defined by~\eqref{eq:LC-def}.
Equation~\eqref{eq:LC-inv} does not uniquely determine $z$ for given
$q$ because the sign of $z$ can be arbitrarily chosen on each connected
component of $S^1\setminus Z_z$. If $Z_z$ consists of an {\em even}
number of points, then we can determine $z$ up to a global sign by the
requirement that $z$ changes its sign at each zero. Therefore, the
preceding discussion shows
\begin{lemma}\label{lem:LC}
The Levi-Civita transformation $z\mapsto q$ given by~\eqref{eq:LC-def}
defines for each even integer $m\in2\mathbb{N}$ a surjective 2-to-1 map
\begin{align*}
\mathcal{L}\colon &\{z\in C^0(S^1,\mathbb{R})\mid \text{$z$ has precisely $m$ zeroes and
switches sign at each zero}\} \\
&\longrightarrow \{q\in C^0(S^1,\mathbb{R}_{\geq 0})\mid \text{$z$ has precisely $m$
zeroes and }\int_0^1\frac{ds}{q(s)}<\infty\}.
\end{align*}
\hfill$\square$
\end{lemma}
The following lemma shows how additional regularity properties
translate between $z$ and $q$. Near each zero $t_*$ of $q$ we define
the local sign function
$$
s_*(t) := \begin{cases}
-1 & t<t_*, \\+1 & t>t_* \,.\end{cases}
$$
If $q$ is of class $C^1$ outside its zero set, we denote by
\begin{equation}\label{eq:energy-q}
E_q(t) := \frac{\dot q^2(t)}{2}-\frac{N}{q(t)},\qquad t\in S^1\setminus Z_q
\end{equation}
its {\em Kepler energy} at time $t$, for some fixed
$N>0$. By~\eqref{bov4} it corresponds under the Levi-Civita
transformation to
\begin{equation}\label{eq:energy-z}
E_z(\tau) := \frac{2\|z\|^4z'(\tau)^2-N}{z(\tau)^2},\qquad \tau\in S^1\setminus Z_z.
\end{equation}
\begin{lemma}\label{lem:LC-reg}
Let $z,q$ be as in Lemma~\ref{lem:LC} related by the Levi-Civita
transform~\eqref{eq:LC-def}. Then the following hold.
(a) $z\in H^1(S^1,\mathbb{R})$ if and only if $q\in H^1(S^1,\mathbb{R}_{\geq 0})$;
(b) $z$ is of class $C^k$ outside its zeroes if and only if $q$ is of
class $C^k$ outside its zeroes;
(c) $z$ is of class $C^1$ on all of $S^1$ if and only if $q$ is of class
$C^1$ outside $Z_q$ and for each $t_*\in
Z_q$ the following limit exists:
$$
\lim_{t_*\neq t\to t_*}s_*(t)\sqrt{q(t)}\dot q(t)\,;
$$
(d) $z$ is of class $C^1$ with transverse zeroes if and only if $q$ is
of class $C^1$ outside $Z_q$ and for each $t_*\in Z_q$ the limit in
(c) exists and is positive.
(e) the energy $E_z:S^1\setminus Z_z\to\mathbb{R}$ is defined and extends to a
continuous function $S^1\to\mathbb{R}$ if and only if the energy $E_q:S^1\setminus
Z_q\to\mathbb{R}$ is defined and extends to a continuous function $S^1\to\mathbb{R}$;
(f) the conditions in (e) imply those in (d).
\end{lemma}
{\bf Proof: }
Part (a) follows immediately from formula~\eqref{bov4}. For (b)
just note that if $z$ is of class $C^k$ outside $Z_z$ then $t_z$ is of
class $C^{k+1}$ outside $Z_z$, so $\tau_z$ and therefore also $q$ is of class
$C^{k+1}$ outside $t_z(Z_z)=Z_q$, and the same in the reverse direction.
For (c) and (d) suppose that $z,q$ are of class $C^1$ outside their zero sets.
In the following we will always denote by $\tau,t$ times related by
the time transformation $t=t_z(\tau)=t_q(\tau)$.
Consider a zero $t_*\in Z_q$ with corresponding $\tau_*\in Z_z$.
Since $z$ switches sign at $\tau_*$ we can write
$$
z(\tau) = \varepsilon s_*(t)\sqrt{q(t)}
$$
for $\tau\neq\tau_*$ near $\tau_*$, with some sign $\varepsilon\in\{-1,1\}$.
Inserting this into~\eqref{bov3} and solving for $z'(\tau)$ yields
\begin{equation}
z'(\tau) = \frac{\varepsilon s_*(t)\sqrt{q(t)}\dot q(t)}{2\|z\|^2},
\end{equation}
from which (c) and (d) follow.
Part (e) follows immediately from $E_z(\tau)=E_q(t_z(\tau))$. To see
that (e) implies (d), note first that the existence and continuity of
$E_q:S^1\setminus Z_q\to\mathbb{R}$ implies that $q$ is of class $C^1$ on
$S^1\setminus Z_q$. Suppose now that $E_q$ extends to a continuous
function $S^1\to\mathbb{R}$, so for each $t_*\in Z_q$ the limit
$$
\lim_{t_8\neq t\to t_*}\Bigl(\frac{\dot q(t)^2}{2}-\frac{N}{q(t)}\Bigr)
$$
exists. This implies that $\dot q(t)^2\to\infty$ as $t\to t^*$, in
particular $\dot q(t)\neq 0$ for all $t\neq t_*$ close to $t_*$.
Since $q(t)>0$ for $t\neq t_*$ and $q(t_*)=0$, this implies that
$$
\dot q(t) = s_*(t)\sqrt{\dot q(t)^2} = s_*(t)\sqrt{2E_q(t)+\frac{2N}{q(t)}}
$$
for all $t\neq t_*$ close to $t_*$. It follows that
$$
\lim_{t_*\neq t\to t_*}s_*(t)\sqrt{q(t)}\dot q(t)
= \lim_{t_*\neq t\to t_*}\sqrt{2E(t)q(t)+2N} = \sqrt{2N} > 0,
$$
which is the condition in (d). This proves the lemma.
\hfill$\square$
Note that $z\in H^1$ and the extension of $E_q$ to a continuous
function $S^1\to\mathbb{R}$ implies the existence of the integral
$$
\int_0^1E_q(t)dt = \frac{\|\dot q\|^2}{2}-\int_0^1\frac{dt}{q(t)},
$$
and therefore $\int_0^1\frac{ds}{q(s)}<\infty$. Hence
Lemma~\ref{lem:LC-reg} implies
\begin{cor}\label{cor:LC}
For each $m\in 2\mathbb{N}$ the Levi-Civita map $\mathcal{L}$ of Lemma~\ref{lem:LC}
restricts to a surjective 2-to-1 map
$$
\mathcal{L}\colon \mathcal{C}^1_{ce}(S^1,\mathbb{R})\to \mathcal{H}^1_{ce}(S^1,\mathbb{R}_{\geq 0}),\quad\text{where}
$$
\begin{itemize}
\item $\mathcal{C}^1_{ce}(S^1,\mathbb{R})$ denotes the set of $z\in C^1(S^1,\mathbb{R})$ with
precisely $m$ zeroes such that all zeroes are transverse and the energy $E_z$ extends to a
continuous function $S^1\to\mathbb{R}$, and
\item $\mathcal{H}^1_{ce}(S^1,\mathbb{R}_{\geq 0})$ denotes the set of $q\in H^1(S^1,\mathbb{R}_{\geq 0})$ with
precisely $m$ zeroes such that $q$ is of class $C^1$ outside its
zeroes and the energy $E_q$ extends to a continuous function $S^1\to\mathbb{R}$.
\end{itemize}
\hfill$\square$
\end{cor}
\subsection{Variational characterization of generalized solutions}
An electron moving in the electric field of a fixed nucleus with
charge $N>0$ is described by Newton's equation
\begin{equation}\label{eq:Kepler}
\ddot q(t) = -\frac{N}{q(t)^2}.
\end{equation}
Alternatively, it describes the Kepler problem a body of mass $1$
moving in the gravitational field of a body of mass $N$. Its solutions
are the critical points of the Lagrangian action functional
$$
\mathcal{S}(q) := \frac12\int_0^1\dot q(t)^2dt + \int_0^1\frac{N}{q(t)}dt.
$$
Let $q$ and $z$ be related by the Levi-Civita transformation~\eqref{eq:LC}.
Using the relations of the preceding subsection, we rewrite the
Lagrangian action of $q$ in terms of $z$:
\begin{align*}
\mathcal{S}(q)
&= \frac12\int_0^1\frac{4\|z\|^4z'(\tau)^2}{z(\tau)^2}\frac{z(\tau)^2}{\|z\|^2}d\tau
+ \frac{N}{\|z\|^2} \cr
&= 2\|z\|^2\|z'\|^2 + \frac{N}{\|z\|^2}.
\end{align*}
We denote the resulting action functional of $z$ by
\begin{equation}\label{eq:Q}
\mathcal{Q}: H^1(S^1,\mathbb{R})\setminus\{0\}\to\mathbb{R},\qquad
\mathcal{Q}(z) := 2\|z\|^2\|z'\|^2 + \frac{N}{\|z\|^2}.
\end{equation}
Following~\cite{barutello-ortega-verzini} we call $q\in
H^1(S^1,\mathbb{R}_{\geq 0})$ a {\em generalized solution of~\eqref{eq:Kepler}} if
\begin{enumerate}
\item the zero set $Z=q^{-1}(0)\subset S^1$ is finite and has an even
number of elements;
\item on $S^1\setminus Z$ the map $q$ is smooth and satisfies~\eqref{eq:Kepler};
\item the energy
$$
E(t) := \frac{\dot q(t)^2}{2}-\frac{N}{q(t)},\qquad t\in
S^1\setminus Z
$$
extends to a continuous function $E:S^1\to\mathbb{R}$.
\end{enumerate}
Note that the energy $E$ is then constant (by conservation of energy)
and negative (for $q$ to be bounded).
\begin{thm}[Barutello, Ortega and Verzini~\cite{barutello-ortega-verzini}]\label{thm:BOV}
Under the Levi-Civita tranformation~\eqref{eq:LC} with time
change~\eqref{eq:t-tau}, critical points $z:S^1\to\mathbb{R}$ of the action
functional $\mathcal{Q}$ defined in~\eqref{eq:Q} are in 2-to-1 correspondence
with generalized solutions $q:S^1\to\mathbb{R}_{\geq 0}$ of~\eqref{eq:Kepler}.
\end{thm}
In the remainder of this section we will spell out the proof of this
theorem because it uses some ingredients that will also be needed in
later sections.
\subsection{From critical points to generalized solutions}\label{ss:z-to-q-Kepler}
The differential of $\mathcal{Q}$ at $z\in H^1(S^1,\mathbb{R})\setminus\{0\}$ in
direction $v\in H^1(S^1,\mathbb{R})$ is given by
\begin{equation}\label{eq:DQ}
D\mathcal{Q}(z)v
= 4\|z\|^2 \langle z',v'\rangle
+ 4\|z'\|^2\langle z,v\rangle
- \frac{2N}{\|z\|^4}\langle z,v\rangle
\end{equation}
This shows that a critical point $z$ has a weak second derivative and
satisfies the second order ODE with constant coefficient
\begin{equation}\label{eq:z-Kepler}
z''(\tau) = a\,z(\tau),\qquad a = \frac{\|z'\|^2}{\|z\|^2} - \frac{N}{2\|z\|^6}.
\end{equation}
It follows that $z$ is smooth. Moreover, $z\in H^1(S^1,\mathbb{R})$ forces $a<0$,
so $z$ is a shifted sine function. In particular, $z$ has {\em transverse
zeroes} in the sense that $z'(\tau)\neq 0$ whenever $z(\tau)=0$.
In particular, its zero set
$$
Z:=\{\tau\in S^1\mid z(\tau)=0\}
$$
is finite. We associate to $z$ the smooth map $t_z:S^1\to S^1$ defined
by~\eqref{eq:tz}. By Lemma~\ref{lem:tz-homeo}, the map $t_z$ is a
homeomorphism with continuous inverse $\tau_z=t_z^{-1} \colon S^1 \to
S^1$. Since $t_z$ is smooth, the function $\tau_z$ is also smooth on the
complement of the finite set $t_z(Z)$ with derivative given by
equation~\eqref{eq:taudot}.
We define a continuous map $q:S^1\to S^1$ by
$$
q(t) := z(\tau_z(t))^2.
$$
Then the two maps $z,q:S^1\to\mathbb{R}$ are smooth except at finitely many
points and related by the Levi-Civita transformation~\eqref{eq:LC}
with $\tau=\tau_z$.
Subsituting $z''$ by~\eqref{eq:z-Kepler} in equation~\eqref{eq:ddot-q}
we get the following ODE for $q$ at points $t\in S^1\setminus t_z(Z)$:
\begin{equation}\label{eq:ddot-q-Kepler}
\ddot q(t)
= \frac{1}{q(t)}\Bigl(2\|z\|^4a-\frac{\dot q(t)^2}{2}\Bigr)
= \frac{1}{q(t)}\Bigl(c-\frac{\dot q(t)^2}{2}\Bigr)
\end{equation}
with the constant (using~\eqref{bov2} and~\eqref{bov4})
$$
c := 2\|z\|^4a = 2\|z'\|^2\|z\|^2 - \frac{N}{\|z\|^2}
= \frac{\|\dot q\|^2}{2} - \int_0^1\frac{N}{q(s)}ds.
$$
Since at a local maximum $t$ of $q$ we must have $\dot q(t)=0$ and
$\ddot q(t)<0$, it follows that $c<0$, hence $\ddot q<0$ outside its zeroes.
Consider now two consecutive zeroes $t_-<t_+$ of $q$ and the smooth map
$$
\beta := \frac{\ddot q}{q}: (t_-,t_+)\to\mathbb{R}_-\,.
$$
Then (omitting the $t$) we have $\beta q^2 = q\ddot q = c-\dot q^2/2$
and taking a time derivative yields
$$
\dot\beta q^2+2\beta q\dot q = -\dot q\dot q = -\beta q\dot q,
$$
hence
\begin{equation}\label{eq:betaqu}
\dot\beta q = -3\beta\dot q.
\end{equation}
\begin{lemma}\label{lem:betamu}
Equation~\eqref{eq:betaqu} for functions $q>0$ and $\beta<0$ on
$(t_-,t_+)$ implies that
\begin{equation}\label{eq:betamu}
\beta = -\frac{\mu}{q^3}
\end{equation}
on $(t_-,t_+)$ for some constant $\mu>0$.
\end{lemma}
\textbf{Proof: }
Dividing both sides of equation~\eqref{eq:betamu} by $q\beta$ yields
$$
\frac{d}{dt} \log(-\beta)=-3\frac{d}{dt} \log(q),
$$
which by integration implies the lemma.
\hfill $\square$
\\
The lemma implies that
\begin{equation}\label{eq:Kepler-mu}
\ddot q = -\mu/q^2
\end{equation}
on $(t_-,t_+)$. Combining this with~\eqref{eq:ddot-q-Kepler} yields
\begin{equation}\label{eq:mu-q}
-\frac{\mu}{q} = q\ddot q
= c-\frac{\dot q^2}{2}
= \frac{\|\dot q\|^2}{2} - \int_0^1\frac{N}{q(s)}ds -\frac{\dot q^2}{2}
\end{equation}
on $(t_-,t_+)$. Multiplying this equation by $-q$, smoothness of $q$ gives
$$
\mu = \lim_{t\to t_\pm}-q(t)\Bigl(\frac{\|\dot q\|^2}{2} -
\int_0^1\frac{N}{q(s)}ds -\frac{\dot q(t)^2}{2}\Bigr).
$$
This shows that the constant $\mu$ is the same for each interval
between consecutive zeroes of $q$, so equation~\eqref{eq:mu-q} holds
on all of $S^1\setminus t_z(Z)$. Now integrating~\eqref{eq:mu-q} over
$S^1$ yields
$$
\mu = N,
$$
so~\eqref{eq:Kepler-mu} becomes Newton's equation~\eqref{eq:Kepler}.
Inserting $\mu=N$ in~\eqref{eq:mu-q} shows that the energy
$$
E = \frac{\dot q^2}{2} - -\frac{N}{q}
= \frac{\|\dot q\|^2}{2} - \int_0^1\frac{N}{q(s)}ds
$$
is constant, so $q$ is a generalized solution of~\eqref{eq:Kepler}.
\subsection{From generalized solutions to critical points}\label{ss:q-to-z-Kepler}
Let now $q\in H^1(S^1,\mathbb{R}_{\geq 0})$ be a generalized solution of
equation~\eqref{eq:Kepler}. Integrating the constant energy yields
$$
E = \int_0^1\frac{\dot q(t)^2}{2}dt - \int_0^1\frac{N}{q(t)}dt.
$$
Since $q\in H^1$, the first term on the right hand side is finite and
it follows that
$$
\int_0^1\frac{1}{q(t)}dt < \infty.
$$
As in Section~\ref{ss:LC-inv}, we associate to $q$ the time
reparametrization $\tau_q:S^1\to S^1$ defined by~\eqref{eq:tauq} and
its inverse $t_q=\tau_q^{-1}$. Recall that $\tau_q$
is smooth outside the zero set $Z_q=q^{-1}(0)$ and $t_q$ is of class $C^1$.
We define a continuous function $z:S^1\to\mathbb{R}$ by the condition
$$
z(\tau)^2 = q(t_q(\tau))
$$
and the requirement that $z$ changes its sign at each zero. This is
possible because $q$ has an even number of zeroes, and it determines
$z$ up to a global sign. Using the change of variable $\tau=\tau_q(t)$
we find
$$
\|z\|^2 = \int_0^1z(\tau)^2d\tau
= \Bigl(\int_0^1\frac{1}{q(s)}ds\Bigr)^{-1}\int_0^1z(\tau_q(t))^2 \frac{1}{q(t)}dt
= \Bigl(\int_0^1\frac{1}{q(s)}ds\Bigr)^{-1}.
$$
It follows that $z$ and $q$ are related by the Levi-Civita
transformation~\eqref{eq:LC} with time change
$t\longrightarrow\tau=\tau_q$ satisfying~\eqref{eq:t-tau}.
In the sequel we will drop the arguments $t$ and $\tau$.
Combining equations~\eqref{eq:ddot-q-z} and~\eqref{eq:Kepler}
outside $Z_q$ we obtain
$$
-\frac{N}{z^4} = -\frac{N}{q^2} = \ddot q = 2\frac{\|z\|^4}{z^4}(z''z-z'^2),
$$
hence
\begin{equation}\label{eq:Kepler-z}
z''(\tau)z(\tau)-z'(\tau)^2 = -\frac{N}{2\|z\|^4},\qquad \tau\in S^1\setminus\tau_q(Z_q).
\end{equation}
Consider the function
$$
\beta := \frac{z''}{z}: S^1\setminus\tau(Z_q)\to\mathbb{R}.
$$
Inserting this into equation~\eqref{eq:Kepler-z} and using
equations~\eqref{eq:LC} and~\eqref{bov3} we find
$$
-\frac{N}{2\|z\|^4} = \beta z^2 - z'^2
= q\Bigl(\beta - \frac{\dot q^2}{4\|z\|^4}\Bigr),
$$
and solving for $\beta$ yields
$$
\beta(\tau_q(t))
= \frac{1}{2\|z\|^4}\Bigl(\frac{\dot q(t)^2}{2} - \frac{N}{q(t)}\Bigr)
= \frac{1}{2\|z\|^4}E(t),\qquad t\in S^1\setminus Z_q.
$$
Since $q$ is a generalized solution, the energy $E$ is constant and
negative, thus $\beta(\tau) \equiv E/2\|z\|^4 < 0$ and the definition
of $\beta$ implies
$$
z''(\tau) = \frac{E}{2\|z\|^4}z(\tau),\qquad \tau\in S^1\setminus\tau_q(Z_q).
$$
The solutions of this ODE are shifted sine functions. So the condition
that $z$ switches sign at each zero implies that it defines a smooth
function $z:S^1\to\mathbb{R}$ solving the ODE on all of $S^1$.
Rewriting the energy via~\eqref{bov2} and~\eqref{bov4} as
$$
E = \frac{\|\dot q\|^2}{2} - \int_0^1\frac{N}{q(t)}dt
= 2\|z\|^2\|z'\|^2 - \frac{N}{\|z\|^2}
$$
and inserting this into the ODE, we see that $z$ satisfies the
ODE~\eqref{eq:z-Kepler} and is therefore a critical point of $\mathcal{Q}$.
Together with the previous subsection this concludes the proof of
Theorem~\ref{thm:BOV}.
\section{Mean interaction}\label{sec:mean-interaction}
In this section we consider a ``helium atom'' in which the two electrons interact by
the {\em mean values} $\overline{q}_i = \int_0^1q_i(t)dt$ according to
\begin{equation}\label{eq:mean-interaction}
\left\{\;
\begin{aligned}
\ddot{q}_1(t) &= -\frac{2}{q_1(t)^2}+\frac{1}{(\overline{q}_1-\overline{q}_2)^2}, \cr
\ddot{q}_2(t) &= -\frac{2}{q_2(t)^2}-\frac{1}{(\overline{q}_1-\overline{q}_2)^2}
\end{aligned}
\right.
\end{equation}
where we impose the condition
\begin{equation}\label{eq:mean-ineq}
\overline q_1 > \overline q_2.
\end{equation}
\subsection{Variational characterization of generalized solutions}\label{ss:var-mean}
Solutions of~\eqref{eq:mean-interaction} are the critical points of the action functional
$$
\mathcal{S}_{av}(q_1,q_2)
:= \sum_{i=1}^2\Bigl(\frac12\int_0^1\dot q_i(t)^2dt + \int_0^1\frac{2}{q_i(t)}dt\Bigr)
- \frac{1}{\overline q_1-\overline q_2}.
$$
For $i=1,2$ let $q_i$ and $z_i$ be related by Levi-Civita transformations
\begin{equation}\label{eq:LC-i}
q_i(t) = z_i(\tau_i(t))^2
\end{equation}
for time changes $\tau_i(t)$ satisfying $\tau_i(0)=0$ and
\begin{equation}\label{eq:t-tau-i}
\frac{dt}{q_i(t)} = \frac{d\tau_i(t)}{\|z_i\|^2}.
\end{equation}
Note that we perform different times changes for the two electrons.
Then all the relations in Section~\eqref{ss:LC} hold with $(q,z,\tau)=(q_i,z_i,\tau_i)$.
In particular, we can use equation~\eqref{bov1} to rewrite the
interaction term in terms of the $z_i$:
\begin{align*}
- \frac{1}{\overline q_1-\overline q_2}
= - \frac{1}{\frac{\|z_1^2\|^2}{\|z_1\|^2} - \frac{\|z_2^2\|^2}{\|z_2\|^2}}
= - \frac{\|z_1\|^2\|z_2\|^2}{\|z_1^2\|^2\|z_2\|^2 - \|z_2^2\|^2\|z_1\|^2}.
\end{align*}
We denote the resulting {\em mean interaction} of $(z_1,z_2)$ by
\begin{equation}\label{eq:A}
\mathcal{A}(z_1,z_2) := - \frac{\|z_1\|^2\|z_2\|^2}{\|z_1^2\|^2\|z_2\|^2 - \|z_2^2\|^2\|z_1\|^2}.
\end{equation}
This quantity is naturally defined on the space
\begin{equation}\label{eq:Hav}
\mathcal{H}_{av}^1 := \Bigg\{z=(z_1,z_2) \in
H^1(S^1,\mathbb{R}^2)\;\Bigl|\; ||z_1||>0,\,\,||z_2||>0,\,\,
\frac{||z_1^2||^2}{||z_1||^2}>\frac{||z_2^2||^2}{||z_2||^2}\Bigg\}.
\end{equation}
Note that $\mathcal{H}_{av}^1$ is an open subset of the Hilbert space
$H^1(S^1,\mathbb{R}^2)$ and the last condition in its definition
corresponds to condition~\eqref{eq:mean-ineq}. On $\mathcal{H}_{av}^1$
we consider the functional
\begin{equation}\label{eq:Bav}
\mathcal{B}_{av} \colon \mathcal{H}_{av}^1 \to \mathbb{R}, \qquad
\mathcal{B}_{av}(z_1,z_2) := \mathcal{Q}(z_1) + \mathcal{Q}(z_2) + \mathcal{A}(z_1,z_2),
\end{equation}
with the functionals
\begin{equation*
\mathcal{Q}(z_i) = 2\|z_i\|^2\|z_i'\|^2 + \frac{2}{\|z_i\|^2}
\end{equation*}
from equation~\eqref{eq:Q} with charge $N=2$.
We call $(q_1,q_2)\in H^1(S^1,\mathbb{R}_{\geq 0}\times\mathbb{R}_{\geq 0})$ a {\em
generalized solution of~\eqref{eq:mean-interaction}} if for $i=1,2$
the following holds:
\begin{enumerate}
\item $\overline q_1>\overline q_2$;
\item the zero sets $Z_i=q_i^{-1}(0)\subset S^1$ are finite and each have an even
number of elements;
\item the restricted maps $q_i:S^1\setminus Z_i\to\mathbb{R}_{\geq 0}$ are
smooth and satisfy~\eqref{eq:mean-interaction};
\item the energies
$$
E_i(t) := \frac{\dot q_i(t)^2}{2}-\frac{2}{q_i(t)},\qquad t\in
S^1\setminus Z_i
$$
extend to continuous functions $E_i:S^1\to\mathbb{R}$.
\end{enumerate}
Note that the individual energies $E_i$ need not be constant, but
their sum is constant and negative.
\begin{thm}[Generalized solutions with mean interaction]\label{thm:mean-interaction}
Under the Levi-Civita tranformations~\eqref{eq:LC-i} with time
changes~\eqref{eq:t-tau-i}, critical points $(z_1,z_2)$ of the action
functional $\mathcal{B}_{av}$ defined in~\eqref{eq:Bav} are in 4-to-1 correspondence
with generalized solutions $(q_1,q_2)$ of~\eqref{eq:mean-interaction}.
\end{thm}
The proof of this theorem will take up the remainder of this section.
\subsection{The differential of $\mathcal{B}_{av}$}\label{ss:diffav}
The differential of the mean interaction $\mathcal{A}$ at
$(z_1,z_2)\in\mathcal{H}_{av}^1$ in direction $(v_1,v_2)\in
H^1(S^1,\mathbb{R}^2)$ is given by
\begin{eqnarray}\label{eq:DA}
D\mathcal{A}[z_1,z_2](v_1,v_2) \nonumber
&=& -2\frac{||z_2||^2 \cdot \langle z_1,v_1\rangle+||z_1||^2 \cdot
\langle z_2,v_2 \rangle}{||z_1^2||^2\cdot ||z_2||^2-||z_2^2||^2
\cdot ||z_1||^2}\\ \nonumber
& &+2\frac{||z_1||^2 \cdot ||z_2||^2\Big(2||z_2||^2 \cdot \langle
z_1^3, v_1 \rangle+||z_1^2||^2 \cdot \langle z_2,v_2\rangle\Big) }
{\big(||z_1^2||^2\cdot ||z_2||^2-||z_2^2||^2 \cdot
||z_1||^2\big)^2 }\\ \nonumber
& &-2\frac{||z_1||^2 \cdot ||z_2||^2\Big(2||z_1||^2 \cdot \langle
z_2^3,v_2 \rangle+||z_2^2||^2 \cdot \langle z_1,v_1\rangle
\Big)}{\big(||z_1^2||^2\cdot ||z_2||^2-||z_2^2||^2 \cdot
||z_1||^2\big)^2 }\\ \nonumber
&=& -2\frac{||z_2||^4 \cdot ||z_1^2||^2}{\big(||z_1^2||^2\cdot ||z_2||^2
-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\langle z_1,v_1 \rangle \\ \nonumber
& &+2\frac{||z_1||^4 \cdot ||z_2^2||^2}{\big(||z_1^2||^2\cdot ||z_2||^2
-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\langle z_2,v_2 \rangle\\ \nonumber
& &+4\frac{||z_1||^2\cdot ||z_2||^4}{\big(||z_1^2||^2\cdot
||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2\big)^2} \langle
z_1^3,v_1\rangle\\ \nonumber
& &-4\frac{||z_1||^4\cdot ||z_2||^2}{\big(||z_1^2||^2\cdot
||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2\big)^2} \langle z_2^3,v_2\rangle
\end{eqnarray}
Combined with equation~\eqref{eq:DQ} with $z=z_i$, $v=v_i$ and $N=2$
for the differentials $D\mathcal{Q}(z_i)v_i$, this yields the differential of $\mathcal{B}_{av}$:
\begin{eqnarray}\label{eq:DBav}
D\mathcal{B}_{av}[z_1,z_2](v_1,v_2) \nonumber
&=& 4 \sum_{i=1}^2 \bigg(||z_i||^2 \langle z_i',v_i'\rangle
+ ||z_i'||^2 \cdot \langle z_i,v_i\rangle
- \frac{\langle z_i,v_i\rangle}{||z_i||^4}\bigg) \\ \nonumber
& & -2\frac{||z_2||^4 \cdot ||z_1^2||^2}{\big(||z_1^2||^2\cdot ||z_2||^2
-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\langle z_1,v_1 \rangle \\
& & +2\frac{||z_1||^4 \cdot ||z_2^2||^2}{\big(||z_1^2||^2\cdot ||z_2||^2
-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\langle z_2,v_2 \rangle\\ \nonumber
& &+4\frac{||z_1||^2\cdot ||z_2||^4}{\big(||z_1^2||^2\cdot
||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2\big)^2} \langle z_1^3,v_1\rangle\\ \nonumber
& &-4\frac{||z_1||^4\cdot ||z_2||^2}{\big(||z_1^2||^2\cdot
||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2\big)^2} \langle z_2^3,v_2\rangle
\end{eqnarray}
\subsection{Critical points of $\mathcal{B}_{av}$}
Equation~\eqref{eq:DBav} leads to the characterization of critical
points of $\mathcal{B}_{av}$:
\begin{prop}\label{prop:critav}
A point $(z_1,z_2)\in \mathcal{H}_{av}^1$ is a critical point of
$\mathcal{B}_{av}$ if and only if $(z_1,z_2)$ is smooth and solves
the system of (uncoupled!) ODEs
\begin{equation}\label{eq:z-av}
\left\{
\begin{aligned}
z_1'' &= a_1 z_1 + b_1z_1^3 \\
z_2'' &= a_2 z_2 + b_2z_2^3
\end{aligned}
\right.
\end{equation}
with the constants
\begin{eqnarray*}
a_1&=&\frac{||z_1'||^2}{||z_1||^2}-\frac{1}{||z_1||^6}-\frac{||z_2||^4 \cdot ||z_1^2||^2}{2||z_1||^2 \cdot\big(||z_1^2||^2\cdot ||z_2||^2
-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\\
b_1&=&+\frac{||z_2||^4}{\big(||z_1^2||^2\cdot ||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\\
a_2&=&\frac{||z_2'||^2}{||z_2||^2}-\frac{1}{||z_2||^6}+\frac{||z_1||^4 \cdot ||z_2^2||^2}{2||z_2||^2 \cdot\big(||z_1^2||^2\cdot ||z_2||^2
-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\\
b_2&=&-\frac{||z_1||^4}{\big(||z_1^2||^2\cdot ||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2\big)^2}.
\end{eqnarray*}
\end{prop}
\textbf{Proof: }
From equation~\eqref{eq:DBav} we see that $(z_1,z_2)$ is a critical
point of $\mathcal{B}_{av}$ if and only if $z_1$ and $z_2$ have
weak second derivatives and satisfy the system of
ODEs~\eqref{eq:z-av}. Bootstrapping these equations
we conclude that $z_1$ and $z_2$ are smooth and the proposition follows.
\hfill $\square$
\begin{cor}\label{cor:mean-transverse-zeroes}
Suppose that $(z_1,z_2)$ is a critical point of
$\mathcal{B}_{av}$. Then $z_1$ and $z_2$ have transverse zeros. In
particular, their zero sets
$$
Z_i := \big\{\tau \in S^1 \mid z_i(\tau)=0\big\}, \qquad i=1,2
$$
are finite.
\end{cor}
\textbf{Proof: }
Arguing by contradiction, suppose that there exists a point
$\tau_0\in S^1$ such that $z_1(\tau_0)=z_1'(\tau_0)=0$.
Then the function $z_1$ and the zero function both solve the first
equation in~\eqref{eq:z-av} with the same initial conditions at
$\tau_0$. By uniqueness of solutions of ODEs we conclude $z_1\equiv
0$, contradicting the condition $\|z_1\|>0$ in the definition of
$\mathcal{H}_{av}$. An analogous argument applies to $z_2$.
\hfill $\square$
\subsection{From critical points to generalized solutions}\label{ss:z-to-q-mean}
Let $(z_1,z_2)\in \mathcal{H}_{av}^1$ be a critical point of $\mathcal{B}_{av}$,
so by Proposition~\ref{prop:critav} the maps $z_1,z_2:S^1\to\mathbb{R}$ are
smooth and satisfy~\eqref{eq:z-av}.
For $i=1,2$ we define the smooth maps $t_{z_i}:S^1\to S^1$ by
\begin{equation}\label{eq:ti}
t_{z_i}(\tau) := \frac{1}{\|z_i\|^2}\int_0^\tau z_i(\sigma)^2d\sigma.
\end{equation}
Since by Corollary~\ref{cor:mean-transverse-zeroes} the map $z_i$ has
only finitely many zeroes, it follows from Lemma~\ref{lem:tz-homeo}
that $t_{z_i}:S^1\to S^1$ is a homeomorphism with continuous inverse
$\tau_{z_i} \colon S^1 \to S^1$. We define continuous maps $q_i:S^1\to S^1$ by
$$
q_i(t) := z_i(\tau_{z_i}(t))^2.
$$
Then for $i=1,2$ the maps $z_i,q_i:S^1\to\mathbb{R}$ are smooth except at finitely many
points and related by the Levi-Civita transformation~\eqref{eq:LC}
with $\tau=\tau_{z_i}$. as explained in Section~\ref{ss:var-mean}, the
last condition in the definition of $\mathcal{H}_{av}^1$ implies
$$
\overline q_1>\overline q_2.
$$
Let us now focus on $i=1$. Subsituting $z_1''$ by~\eqref{eq:z-av} in
equation~\eqref{eq:ddot-q} with $q=q_1$ and $z=z_1$
we compute at points $t\in S^1\setminus t_{z_1}(Z_1)$:
\begin{eqnarray}
\ddot{q}_1 \nonumber
&=&\bigg(2||z_1'||^2 \cdot
||z_1||^2-\frac{2}{||z_1||^2}-\frac{||z_1||^2 \cdot ||z_2||^4 \cdot
||z_1^2||^2}{\big(||z_1^2||^2\cdot ||z_2||^2
-||z_2^2||^2 \cdot ||z_1||^2\big)^2}\bigg)\frac{1}{q_1}\\ \nonumber
& &+\frac{2||z_1||^4 \cdot ||z_2||^4}{\big(||z_1^2||^2\cdot ||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2\big)^2}
-\frac{\dot{q}_1^2}{2q_1}\\ \nonumber
&=&\bigg(2||z_1'||^2 \cdot ||z_1||^2-\frac{2}{||z_1||^2}-\frac{||z_1||^4 \cdot ||z_2||^4 \cdot \frac{||z_1^2||^2}{
||z_1||^2}}{||z_1||^4 \cdot ||z_2||^4 \cdot\Big(\frac{||z_1^2||^2}{||z_1||^2}
-\frac{||z_2^2||^2}{||z_2||^2}\Big)^2}-\frac{\dot{q}_1^2}{2}\bigg)\frac{1}{q_1}\\ \nonumber
& &+\frac{2||z_1||^4 \cdot ||z_2||^4}{||z_1||^4 \cdot ||z_2||^4 \cdot\Big(\frac{||z_1^2||^2}{||z_1||^2}
-\frac{||z_2^2||^2}{||z_2||^2}\Big)^2}\\ \nonumber
&=&\bigg(2||z_1'||^2 \cdot ||z_1||^2-\frac{2}{||z_1||^2}-\frac{\frac{||z_1^2||^2}{
||z_1||^2}}{\Big(\frac{||z_1^2||^2}{||z_1||^2}
-\frac{||z_2^2||^2}{||z_2||^2}\Big)^2}-\frac{\dot{q}_1^2}{2}\bigg)\frac{1}{q_1}\\ \nonumber
& &+\frac{2}{\Big(\frac{||z_1^2||^2}{||z_1||^2}
-\frac{||z_2^2||^2}{||z_2||^2}\Big)^2}\\ \nonumber
&=&\bigg(\frac{||\dot{q}_1||^2}{2}-\int_0^1 \frac{2}{q_1(s)}ds-\frac{\overline{q}_1}{(\overline{q}_1-\overline{q}_2)^2}
-\frac{\dot{q}_1^2}{2}\bigg)\frac{1}{q_1}+\frac{2}{(\overline{q}_1-\overline{q}_2)^2}.
\end{eqnarray}
Thus $q_1$ satisfies the ODE
\begin{equation}\label{bov6}
\ddot{q}_1 = \bigg(c_1
-\frac{\dot{q}_1^2}{2}\bigg)\frac{1}{q_1}+\frac{2}{(\overline{q}_1-\overline{q}_2)^2}
\end{equation}
with the constant
\begin{equation}\label{eq:c1-mean}
c_1 = \frac{||\dot{q}_1||^2}{2}-\int_0^1
\frac{2}{q_1(s)}ds-\frac{\overline{q}_1}{(\overline{q}_1-\overline{q}_2)^2}.
\end{equation}
At the global maximum $t_{\rm max}$ of $q_1$ equation~\eqref{bov6} becomes
$$
\frac{c_1}{q_1(t_{\rm max})} + \frac{2}{(\overline{q}_1-\overline{q}_2)^2} =
\ddot q_1(t_{\rm max})\leq 0,
$$
hence
\begin{equation}\label{eq:c1-ineq}
c_1 \leq -\frac{2q_1(t_{\rm max})}{(\overline{q}_1-\overline{q}_2)^2}.
\end{equation}
Let now $t_-<t_+$ be adjacent zeroes of $q_1$ and consider the smooth map
$$
\beta_1:=\frac{\ddot{q}_1-\frac{1}{(\overline{q}_1-\overline{q}_2)^2}}{q_1}
\colon (t_-,t_+) \to \mathbb{R}.
$$
From (\ref{bov6}) we obtain
$$
\beta_1 q_1^2
= c_1 -\frac{\dot{q}_1^2}{2} + \frac{q_1}{(\overline{q}_1-\overline{q}_2)^2}.
$$
With $q_1\leq q_1(t_{\rm max})$ and inequality~\eqref{eq:c1-ineq} this implies
$$
\beta_1 q_1^2
\leq -\frac{\dot{q}_1^2}{2} - \frac{q_1(t_{\rm
max})}{(\overline{q}_1-\overline{q}_2)^2} < 0,
$$
hence $\beta_1<0$ on $(t_-,t_+)$. Differentiating both sides of the
equation for $\beta q_1^2$ we get
$$\dot{\beta}_1 q_1^2+2 \beta_1 q_1 \dot{q}_1=-\ddot{q}_1 \dot{q}_1+\frac{\dot{q}_1}{(\overline{q}_1-\overline{q}_2)^2}=-\beta_1 q_1 \dot{q}_1$$
and therefore
\begin{equation*
\dot{\beta}_1 q_1 = -3\beta_1 \dot{q}_1
\end{equation*}
By Lemma~\ref{lem:betamu} this implies that
\begin{equation*
\beta_1 = -\frac{\mu}{q_1^3}
\end{equation*}
on $(t_-,t_+)$ for some constant $\mu>0$. By definition of $\beta_1$
this yields
\begin{equation}\label{bov7}
\ddot{q}_1(t) =
-\frac{\mu}{q_1(t)^2}+\frac{1}{(\overline{q}_1-\overline{q}_2)^2}
\end{equation}
for $t\in (t_-,t_+)$. Plugging this into~\eqref{bov6} we infer
\begin{equation}\label{bov7a}
\mu = -\bigg(\frac{\|\dot{q}_1\|^2}{2}-\int_0^1 \frac{2}{q_1(s)}ds
-\frac{\overline{q}_1}{(\overline{q}_1-\overline{q}_2)^2}
-\frac{\dot{q}_1(t)^2}{2}\bigg)q_1(t)-\frac{q_1(t)^2}{(\overline{q}_1-\overline{q}_2)^2}
\end{equation}
for $t\in(t_-,t_+)$. In particular,
\begin{equation}\label{bov8}
\mu = \lim_{t \to t_\pm} \frac{\dot{q}_1(t)^2 q_1(t)}{2}
= 2||z_1||^4 z_1'\big(\tau_{z_1}(t_\pm)\big)^2.
\end{equation}
We deduce from this that equation~\eqref{bov7} holds on $S^1 \setminus
t_{z_1}(Z_1)$ with a fixed $\mu$ independent of
the connected component in $S^1 \setminus t_{z_1}(Z_1)$. We divide (\ref{bov7a}) by $q_1$ to get
\begin{equation}\label{bov7b}
\frac{\mu}{q_1(t)}
= -\frac{\|\dot{q}_1\|^2}{2}+\int_0^1
\frac{2}{q_1(s)}ds+\frac{\overline{q}_1}{(\overline{q}_1-\overline{q}_2)^2}
+\frac{\dot{q}_1(t)^2}{2}-\frac{q_1(t)}{(\overline{q}_1-\overline{q}_2)^2}.
\end{equation}
Integrating this equation yields
$$
\mu\int \frac{1}{q_1(t)}dt=2\int_0^1 \frac{1}{q_1(s)}ds,
$$
and therefore
$$\mu=2.$$
Thus equation~\eqref{bov7} becomes the first equation in~\eqref{eq:mean-interaction}.
Similarly, one obtains for $q_2$ the equation
$$\ddot{q}_2=\bigg(\frac{||\dot{q}_2||^2}{2}-\int_0^1 \frac{2}{q_2}dt+\frac{\overline{q}_2}{(\overline{q}_1-\overline{q}_2)^2}
-\frac{\dot{q}_2^2}{2}\bigg)\frac{1}{q_2}-\frac{2}{(\overline{q}_1-\overline{q}_2)^2}.$$
Setting
$$\beta_2:=\frac{\ddot{q}_2+\frac{1}{(\overline{q}_1-\overline{q}_2)^2}}{q_1}$$
one gets
$$\dot{\beta}_2 q_2=-3\beta_2 \dot{q}_2$$
implying that there exists $\mu \in \mathbb{R}$ such that
$$\beta_2=-\frac{\mu}{q_2^3}.$$
Arguing as above one deduces from this that $\mu=2$ and thus $q_2$
satisfies the second equation in~\eqref{eq:mean-interaction}.
To see the continuity of $E_1$, we solve equation~\eqref{bov7b} (with
$\mu=2$) for
$$
E_1(t) = \frac{\dot q_1(t)^2}{2} - \frac{2}{q_1(t)}
= \frac{\|\dot{q}_1\|^2}{2} - \int_0^1 \frac{2}{q_1(s)}ds
- \frac{\overline{q}_1}{(\overline{q}_1-\overline{q}_2)^2}
-\frac{q_1(t)}{(\overline{q}_1-\overline{q}_2)^2}
$$
and note that the right hand side is continuous as a function of $t\in [0,1]$.
Continuity of $E_2$ follows similarly, and we have shown that
$(q_1,q_2)$ is a generalized solution of equation~\eqref{eq:mean-interaction}.
\subsection{From generalized solutions to critical points}\label{ss:q-to-z-mean}
Let now $(q_1,q_2)\in H^1(S^1,\mathbb{R}_{\geq 0}\times\mathbb{R}_{\geq 0})$ be a
generalized solution of equation~\eqref{eq:mean-interaction}.
The definition of a generalized solution implies that
$q_1,q_2\in \mathcal{H}_{ce}^1(S^1,\mathbb{R}_{\geq 0})$. Corollary~\ref{cor:LC}
implies that the set $\mathcal{L}^{-1}(q_1)\times \mathcal{L}^{-1}(q_2)$ consists of
$4$ elements. The goal of this section is to show that each
$(z_1,z_2)\in \mathcal{L}^{-1}(q_1)\times \mathcal{L}^{-1}(q_2)$ is a critical point
of $\mathcal{B}_{av}$. To see this, observe that smoothness of $q_i$ on the
complement of its zero set $Z_{q_i}$ implies smoothness of $z_i$ on the
complement of its zero set $Z_{z_i}$, $i\in \{1,2\}$. In particular,
second derivatives of $z_i$ make sense there and we can
make the following statement.
\begin{lemma}\label{lem:eqzmean}
Any $(z_1,z_2)\in \mathcal{L}^{-1}(q_1)\times \mathcal{L}^{-1}(q_2)$ satisfies the critical point
equation~\eqref{eq:z-av} on the complement of the set $Z_{z_1}\cup Z_{z_2}$.
\end{lemma}
Assuming this lemma for the moment, recall that $z_1$ and $z_2$ are of
class $C^1$ and \eqref{eq:z-av} expresses $z_i''$ through
$z_i$. Thus bootstrapping \eqref{eq:z-av} implies that $z_1$ and
$z_2$ are smooth and \eqref{eq:z-av} holds on the whole
$S^1$. Therefore, it remains to prove Lemma~\ref{lem:eqzmean}.
{\bf Proof of Lemma~\ref{lem:eqzmean}: }
We will show the desired equation for $z_1$.
A similar argument will do the job for $z_2$. Recall the
equation satisfied by $q_1$,
\begin{equation}\label{eq:q1mean}
\ddot q_1=-\frac{2}{q_1^2}+\frac{1}{(\overline{q}_1-\overline{q}_2)^2}.
\end{equation}
Set
\begin{equation}\label{eq:beta-mean}
\beta_1:=\frac{\ddot{q}_1-\frac{1}{(\overline{q}_1-\overline{q}_2)^2}}{q_1}
\end{equation}
on $S^1\setminus Z_{q_1}$. Then by~\eqref{eq:q1mean} we have
$$
\beta_1=-\frac{2}{q_1^3},
$$
and taking time derivative we obtain
$$
\dot\beta_1=\frac{3\cdot 2}{q_1^4}\dot q_1=-\frac{3\beta_1 \dot q_1}{q_1}.
$$
We multiply both sides with $q_1^2$ to get
$$
\dot\beta_1q_1^2=-3\beta_1\dot q_1q_1\,.
$$
We bring $-2\beta_1\dot q_1q_1$ to the other side to continue
$$
\dot\beta_1q_1^2+2\beta_1\dot q_1q_1=-\beta_1\dot q_1q_1\,.
$$
We substitute the original definition~\eqref{eq:beta-mean}
of $\beta_1$ in the right hand side to get
$$
\dot\beta_1q_1^2+2\beta_1\dot q_1q_1=-\dot q_1\ddot q_1+
\frac{\dot q_1}{(\overline{q}_1-\overline{q}_2)^2}.
$$
Integrating both sides from $0$ to $t$ we get
$$
\beta_1 q_1^2
= C -\frac{\dot{q}_1^2}{2} + \frac{q_1}{(\overline{q}_1-\overline{q}_2)^2}
$$
with some constant $C\in \mathbb{R}$.
We substitute the original definition~\eqref{eq:beta-mean}
of $\beta_1$ in the left hand side to get
$$
\ddot{q}_1q_1 = C
-\frac{\dot{q}_1^2}{2}+\frac{2q_1}{(\overline{q}_1-\overline{q}_2)^2}.
$$
Observe that modulo the exact value of $C$ this is exactly equation~\eqref{bov6},
which is equivalent the first equation of~\eqref{eq:z-av}.
Therefore, we are left with computing the constant $C$.
For this we solve the last equation for $C$ and use~\eqref{eq:q1mean} to obtain
$$
C=\frac{\dot{q}_1^2}{2}-\frac{2}{q_1}-\frac{q_1}{(\overline{q}_1-\overline{q}_2)^2},
$$
and integrating from $0$ to $1$ gives us
$$
C = \frac{||\dot{q}_1||^2}{2}-\int_0^1
\frac{2}{q_1(s)}ds-\frac{\overline{q}_1}{(\overline{q}_1-\overline{q}_2)^2},
$$
which matches the constant $c_1$ in~\eqref{eq:c1-mean}. This concludes
the proof of Lemma~\ref{lem:eqzmean}, and therefore of
Theorem~\ref{thm:mean-interaction}.
\hfill$\square$
\section{Instantaneous interaction}\label{sec:inst-interaction}
In this section we consider the real helium atom in which the two electrons interact by
their Coulomb repulsion according to
\begin{equation}\label{eq:inst-interaction}
\left\{\;
\begin{aligned}
\ddot{q}_1(t) &= -\frac{2}{q_1(t)^2}+\frac{1}{(q_1(t)-q_2(t))^2}, \cr
\ddot{q}_2(t) &= -\frac{2}{q_2(t)^2}-\frac{1}{(q_1(t)-q_2(t))^2}
\end{aligned}
\right.
\end{equation}
where we impose the condition
\begin{equation}\label{eq:inst-ineq}
q_1(t) > q_2(t)\geq 0\qquad\text{for all }t\in S^1.
\end{equation}
\subsection{Variational characterization of generalized
solutions}\label{ss:gen-solutions-inst}
Solutions of~\eqref{eq:inst-interaction} are the critical points of the action functional
$$
\mathcal{S}_{in}(q_1,q_2)
:= \sum_{i=1}^2\Bigl(\frac12\int_0^1\dot q_i(t)^2dt + \int_0^1\frac{2}{q_i(t)}dt\Bigr)
- \int_0^1\frac{1}{q_1(t)-q_2(t)}dt.
$$
For $i=1,2$ let $q_i$ and $z_i$ be related by Levi-Civita transformations
\begin{equation}\label{eq:LC-zi}
q_i(t) = z_i(\tau_{z_i}(t))^2,
\end{equation}
where $\tau_{z_i}:S^1\to S^1$ is the inverse of $t_{z_i}:S^1\to S^1$
defined in equation~\eqref{eq:tz} with $z=z_i$. As in Section it
follows that
\begin{equation}\label{eq:t-tau-zi}
\frac{dt}{q_i(t)} = \frac{d\tau_{z_i}(t)}{\|z_i\|^2},
\end{equation}
so all the relations in Section~\eqref{ss:LC} hold with $(q,z,\tau)=(q_i,z_i,\tau_{z_i})$.
In particular, we can rewrite the last integral in terms of the $z_i$
as the {\em instantaneous interaction}
\begin{align}\label{eq:I}
\mathcal{I}(z_1,z_2)
&:= - \frac{1}{\|z_1\|^2}\int_0^1\frac{z_1(\tau)^2}{z_1^2(\tau)-z_2^2
(\tau_{z_2}(t_{z_1}(\tau)))}d\tau \\
&= - \int_0^1\frac{1}{z_1^2(\tau_{z_1}(t)) - z_2^2(\tau_{z_2}(t))}dt \cr
&= \frac{1}{\|z_2\|^2}\int_0^1\frac{z_2(\tau)^2}{z_2^2(\tau)-z_1^2(\tau_{z_1}(t_{z_2}(\tau)))}d\tau,
\end{align}
where in the last two equalities we have changed the integration
variable to $\tau=\tau_{z_1}(t)$ resp.~$\tau=\tau_{z_2}(t)$ using
equation~\eqref{eq:t-tau-zi}. Note that the functional $\mathcal{I}$ is {\em
nonlocal} due to the appearance of the time changes, which we have
written as $\tau_{z_i}$ rather than $\tau_i$ to indicate their
dependence on $z_i$.
The instantaneous interaction $\mathcal{I}$ is naturally defined on the space
\begin{equation}\label{eq:Hin}
\begin{aligned}
\mathcal{H}_{in}^1 := \Bigl\{
&z=(z_1,z_2) \in H^1(S^1,\mathbb{R}^2)\;\Bigl|\;
\|z_1\|>0,\,\,\|z_2\|>0,\,\, \cr
&\ \ z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))>0 \,\,\text{for
all}\,\,\tau\in S^1\Bigr\}.
\end{aligned}
\end{equation}
Note that $\mathcal{H}_{in}^1$ is an open subset of the Hilbert space
$H^1(S^1,\mathbb{R}^2)$ and the last condition in its definition
corresponds to condition~\eqref{eq:inst-ineq}. Since integrating
condition~\eqref{eq:inst-ineq} leads to the averaged
condition~\eqref{eq:mean-ineq}, which is in turn equivalent to the
last condition in the definition of $\mathcal{H}_{av}^1$ in
Section~\ref{ss:var-mean}, we have
$$
\mathcal{H}_{in}^1\subset \mathcal{H}_{av}^1.
$$
On $\mathcal{H}_{in}^1$ we consider the functional
\begin{equation}\label{eq:Bin}
\mathcal{B}_{in} \colon \mathcal{H}_{in}^1 \to \mathbb{R}, \qquad
\mathcal{B}_{in}(z_1,z_2) := \mathcal{Q}(z_1) + \mathcal{Q}(z_2) + \mathcal{I}(z_1,z_2),
\end{equation}
with the functionals
\begin{equation*
\mathcal{Q}(z_i) = 2\|z_i\|^2\|z_i'\|^2 + \frac{2}{\|z_i\|^2}
\end{equation*}
from equation~\eqref{eq:Q} with charge $N=2$.
We call $(q_1,q_2)\in H^1(S^1,\mathbb{R}_{\geq 0}\times\mathbb{R}_{\geq 0})$ a {\em
generalized solution of~\eqref{eq:inst-interaction}} if for $i=1,2$
the following holds:
\begin{enumerate}
\item $q_1(t)>q_2(t)$ for all $t\in S^1$;
\item the zero sets $Z_i=q_i^{-1}(0)\subset S^1$ are finite and each have an even
number of elements;
\item the restricted maps $q_i:S^1\setminus Z_i\to\mathbb{R}_{\geq 0}$ are
smooth and satisfy~\eqref{eq:inst-interaction};
\item the energies
$$
E_i(t) := \frac{\dot q_i(t)^2}{2}-\frac{2}{q_i(t)},\qquad t\in
S^1\setminus Z_i
$$
extend to continuous functions $E_i:S^1\to\mathbb{R}$.
\end{enumerate}
Note that the individual energies $E_i$ need not be constant, but
the total energy
$$
E = E_1(t) + E_2(t) + \frac{1}{q_1(t)-q_2(t)}
$$
is constant and negative.
\begin{thm}[Generalized solutions with instantaneous
interaction]\label{thm:inst-interaction}
Under the Levi-Civita tranformations~\eqref{eq:LC-zi} with time
changes~\eqref{eq:t-tau-zi}, critical points $(z_1,z_2)$ of the action
functional $\mathcal{B}_{in}$ defined in~\eqref{eq:Bin} are in 4-to-1 correspondence
with generalized solutions $(q_1,q_2)$ of~\eqref{eq:inst-interaction}.
\end{thm}
The proof of this theorem will take up the remainder of this section.
\subsection{The differential of $\mathcal{B}_{in}$}\label{ss:diffin}
In this subsection we compute the differential of $\mathcal{B}_{in}$
at $(z_1,z_2)\in\mathcal{H}_{in}^1$ in direction $(v_1,v_2)\in H^1(S^1,\mathbb{R}^2)$.
For this we will need for $i=1,2$ the derivative of the time change
$t_{z_i}$ with respect to $z_i$ in direction
$v_i$. By~\eqref{eq:dt} it is given by
\begin{equation}\label{eq:dt-i}
Dt_{z_i}(v_i)(\tau) = \frac{2}{\|z_i\|^2}\int_0^\tau z_i(\sigma) v_i(\sigma) d\sigma
- \frac{2\langle z_i,v_i\rangle}{\|z_i\|^4}\int_0^\tau z_i(\sigma)^2 d\sigma\,.
\end{equation}
Using this we now compute the differential of the instantaneous interaction
$\mathcal{I}$ at $(z_1,z_2)\in\mathcal{H}_{in}^1$ with respect to
$z_1$ in direction $v_1\in H^1(S^1,\mathbb{R})$. Using the first expression
for $\mathcal{I}$ in~\eqref{eq:I} we obtain
\begin{eqnarray*}
& & D_1\mathcal{I}[z_1,z_2](v_1)\\
&=& \frac{2\langle z_1,v_1\rangle}{||z_1||^4}
\int_0^1\frac{z_1^2(\tau)}{z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))}d\tau\\
& & +\frac{2}{||z_1||^2}\int_0^1 \frac{z_1(\tau)
z_2^2(\tau_{z_2}(t_{z_1}(\tau)))v_1(\tau)}{\big(z_1^2(\tau) -
z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau\\
& & -\frac{2}{||z_1||^2}\int_0^1 \frac{z_1^2(\tau)
z_2(\tau_{z_2}(t_{z_1}(\tau)))z_2'(\tau_{z_2}(t_{z_1}(\tau)))\dot{\tau}_{z_2}(t_{z_1}(\tau))
Dt_{z_1}(v_1)(\tau)}{\big(z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau
\end{eqnarray*}
We rewrite the third term on the right hand side as
\begin{eqnarray*}
& & -\frac{2}{||z_1||^2}\int_0^1 \frac{z_1^2(\tau)
z_2(\tau_{z_2}(t_{z_1}(\tau)))z_2'(\tau_{z_2}(t_{z_1}(\tau)))\dot{\tau}_{z_2}(t_{z_1}(\tau))
Dt_{z_1}(v_1)(\tau)}{\big(z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau\\
&=& -\frac{2||z_2||^2}{||z_1||^2}\int_0^1 \frac{z_1^2(\tau)
z_2'(\tau_{z_2}(t_{z_1}(\tau)))Dt_{z_1}(v_1)(\tau)}
{z_2(\tau_{z_2}(t_{z_1}(\tau)))\big(z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau\\
&=& -\frac{4||z_2||^2}{||z_1||^4}\int_0^1 \frac{z_1^2(\tau)
z_2'(\tau_{z_2}(t_{z_1}(\tau)))}{z_2(\tau_{z_2}(t_{z_1}(\tau)))\big(z_1^2(\tau)
-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}
\bigg(\int_0^\tau z_1(\sigma)v_1(\sigma) d\sigma\bigg) d\tau\\
& & +\frac{4||z_2||^2\langle z_1,v_1\rangle}{||z_1||^6}\int_0^1
\frac{z_1^2(\tau) z_2'(\tau_{z_2}(t_{z_1}(\tau)))\int_0^\tau z_1(\sigma)^2d\sigma
}{z_2(\tau_{z_2}(t_{z_1}(\tau)))\big(z_1^2(\tau)
-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2} d\tau\\
&=& -\frac{4||z_2||^2}{||z_1||^4}\int_0^1\bigg(\int_\sigma^1
\frac{z_1^2(\tau)z_2'(\tau_{z_2}(t_{z_1}(\tau)))}{z_2(\tau_{z_2}(t_{z_1}(\tau)))
\big(z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau\bigg)
z_1(\sigma) v_1(\sigma) d\sigma\\
& & +\frac{4||z_2||^2\langle z_1,v_1\rangle}{||z_1||^4}\int_0^1
\frac{z_1^2(\tau) z_2'(\tau_{z_2}(t_{z_1}(\tau)))
t_{z_1}(\tau)}{z_2(\tau_{z_2}(t_{z_1}(\tau)))\big(z_1^2(\tau)
-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau.\\
\end{eqnarray*}
Here for the first equality we haved used formula~\eqref{eq:taudot} for
$\dot\tau_{z_i}$, and for the second one we have used equation~\eqref{eq:dt-i}.
For the third equality we have switched the order of integration in the
first integral, and we have used equation~\eqref{eq:ti} to replace
$\int_0^\tau z_1(\sigma)^2d\sigma$ by $\|z_1\|^2t_{z_1}(\tau)$.
By symmetry, the expression for $D_2\mathcal{I}[z_1,z_2](v_2)$ is the same with
a global minus sign and the roles of $(z_1,v_1)$ and $(z_2,v_2)$ reversed.
Putting everything together,
we obtain the differential of $\mathcal{B}_{in}$:
\begin{eqnarray}\label{eq:DBin} \nonumber
& &d\mathcal{B}_{in}[z_1,z_2](v_1,v_2)\\ \nonumber
&=&4 \sum_{i=1}^2 \Bigg(||z_i'||^2 \cdot \langle z_i,v_i\rangle
+||z_i||^2 \langle z_i',v_i'\rangle-\frac{\langle z_i,v_i\rangle}{||z_i||^4}\Bigg)\\ \nonumber
& &+\frac{2\langle z_1,v_1\rangle}{||z_1||^4}\int_0^1\frac{z_1^2(\tau)}{z_1^2(\tau)
-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))}d\tau\\ \nonumber
& &+\frac{2}{||z_1||^2}\int_0^1 \frac{z_1(\tau)
z_2^2(\tau_{z_2}(t_{z_1}(\tau)))v_1(\tau)}{\big(z_1^2(\tau)
-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau\\ \nonumber
& &-\frac{4||z_2||^2}{||z_1||^4}\int_0^1\Bigg(\int_\sigma^1 \frac{z_1^2(\tau) z_2'(\tau_{z_2}(t_{z_1}(\tau)))}{z_2(\tau_{z_2}(t_{z_1}(\tau)))\big(z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau
\Bigg) z_1(\sigma) v_1(\sigma) d\sigma\\
& &+\frac{4||z_2||^2\langle z_1,v_1\rangle}{||z_1||^4}\int_0^1
\frac{z_1^2(\tau) z_2'(\tau_{z_2}(t_{z_1}(\tau)))
t_{z_1}(\tau)}{z_2(\tau_{z_2}(t_{z_1}(\tau)))\big(z_1^2(\tau)
-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}d\tau\\ \nonumber
& &-\frac{2\langle
z_2,v_2\rangle}{||z_2||^4}\int_0^1\frac{z_2^2(\tau)}
{z_2^2(\tau)-z_1^2(\tau_{z_1}(t_{z_2}(\tau)))}d\tau\\ \nonumber
& &-\frac{2}{||z_2||^2}\int_0^1 \frac{z_2(\tau)
z_1^2(\tau_{z_1}(t_{z_2}(\tau)))v_2(\tau)}
{\big(z_2^2(\tau)-z_1^2(\tau_{z_1}(t_{z_2}(\tau)))\big)^2}d\tau\\ \nonumber
& &+\frac{4||z_1||^2}{||z_2||^4}\int_0^1\Bigg(\int_\sigma^1 \frac{z_2^2(\tau) z_1'(\tau_{z_1}(t_{z_2}(\tau)))}{z_1(\tau_{z_1}(t_{z_2}(\tau)))\big(z_2^2(\tau)-z_1^2(\tau_{z_1}(t_{z_2}(\tau)))\big)^2}d\tau\Bigg)
z_2(\sigma) v_2(\sigma)d\sigma\\ \nonumber
& &-\frac{4||z_1||^2\langle z_2,v_2\rangle}{||z_2||^4}\int_0^1
\frac{z_2^2(\tau) z_1'(\tau_{z_1}(t_{z_2}(\tau)))
t_{z_2}(\tau)}{z_1(\tau_{z_1}(t_{z_2}(\tau)))\big(z_2^2(\tau)
-z_1^2(\tau_{z_1}(t_{z_2}(\tau)))\big)^2}d\tau\\ \nonumber
\end{eqnarray}
\subsection{Critical points of $\mathcal{B}_{in}$}
Equation~\eqref{eq:DBin} leads to the characterization of critical
points of $\mathcal{B}_{in}$:
\begin{prop}\label{prop:critin}
A point $(z_1,z_2)\in \mathcal{H}_{in}^1$ is a critical point of $\mathcal{B}_{in}$ if and only if
$(z_1,z_2)$ is smooth and solves the following system of coupled
nonlocal integral--differential equations:
\begin{eqnarray}\label{eq:critinst}
z_1''(\tau)&=&\frac{||z_1'||^2z_1(\tau)}{||z_1||^2}-\frac{z_1(\tau)}{||z_1||^6}\\ \nonumber
& &+\frac{z_1(\tau)}{2||z_1||^6}\int_0^1\frac{z_1^2(\sigma)}{z_1^2(\sigma)-z_2^2(\tau_{z_2}(t_{z_1}(\sigma)))}d\sigma\\ \nonumber
& &+\frac{z_1(\tau)z_2^2(\tau_{z_2}(t_{z_1}(\tau)))}{2||z_1||^4\big(z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}\\ \nonumber
& &-\frac{||z_2||^2z_1(\tau)}{||z_1||^6}
\int_\tau^1 \frac{z_1^2(\sigma) z_2'(\tau_{z_2}(t_{z_1}(\sigma)))}{z_2(\tau_{z_2}(t_{z_1}(\sigma)))\big(z_1^2(\sigma)-z_2^2(\tau_{z_2}(t_{z_1}(\sigma)))\big)^2}d\sigma\\ \nonumber
& &+\frac{||z_2||^2 z_1(\tau)}{||z_1||^6}
\int_0^1 \frac{z_1^2(\sigma) z_2'(\tau_{z_2}(t_{z_1}(\sigma))) t_{z_1}(\sigma)}
{z_2(\tau_{z_2}(t_{z_1}(\sigma)))\big(z_1^2(\sigma)-z_2^2(\tau_{z_2}(t_{z_1}(\sigma)))\big)^2}d\sigma\\
\nonumber
z_2''(\tau)&=&\frac{||z_2'||^2z_2(\tau)}{||z_2||^2}-\frac{z_2(\tau)}{||z_2||^6}\\ \nonumber
& &-\frac{z_2(\tau)}{2||z_2||^6}\int_0^1\frac{z_2^2(\sigma)}{z_2^2(\sigma)-z_1^2(\tau_{z_1}(t_{z_2}(\sigma)))}d\sigma\\ \nonumber
& &-\frac{z_2(\tau)z_1^2(\tau_{z_1}(t_{z_2}(\tau)))}{2||z_2||^4\big(z_2^2(\tau)-z_1^2(\tau_{z_1}(t_{z_2}(\tau)))\big)^2}\\ \nonumber
& &+\frac{||z_1||^2z_2(\tau)}{||z_2||^6}
\int_\tau^1 \frac{z_2^2(\sigma) z_1'(\tau_{z_1}(t_{z_2}(\sigma)))}{z_1(\tau_{z_1}(t_{z_2}(\sigma)))\big(z_2^2(\sigma)-z_1^2(\tau_{z_1}(t_{z_2}(\sigma)))\big)^2}d\sigma\\ \nonumber
& &-\frac{||z_1||^2 z_2(\tau)}{||z_2||^6}
\int_0^1 \frac{z_2^2(\sigma) z_1'(\tau_{z_1}(t_{z_2}(\sigma))) t_{z_2}(\sigma)}{z_1(\tau_{z_1}(t_{z_2}(\sigma)))\big(z_2^2(\sigma)-z_1^2(\tau_{z_1}(t_{z_2}(\sigma)))\big)^2}d\sigma
\end{eqnarray}
\end{prop}
\textbf{Proof: }
From equation~\eqref{eq:DBin} we see that $(z_1,z_2)$ is a critical
point of $\mathcal{B}_{in}$ if and only if $z_1$ and $z_2$ have
weak second derivatives and satisfy the system of
equations~\eqref{eq:critinst}. Bootstrapping these equations
we conclude that $z_1$ and $z_2$ are smooth and the proposition follows.
\hfill $\square$
\begin{cor}\label{cor:inst-transverse-zeroes}
Suppose that $(z_1,z_2)$ is a critical point of
$\mathcal{B}_{in}$. Then $z_1$ has no zeroes and $z_2$ has transverse
zeros. In particular, their zero sets
$$
Z_i := \big\{\tau \in S^1 \mid z_i(\tau)=0\big\}, \qquad i=1,2
$$
are finite.
\end{cor}
\textbf{Proof: }
Let $(z_1,z_2)\in\mathcal{H}_{in}^1$ be a critical point of
$\mathcal{B}_{in}$. It follows directly from the definition of
$\mathcal{H}_{in}^1$ that $z_1(t)>0$ for all $t\in S^1$.
Next note that at $(z_1,z_2)$ solves the system of
coupled nonlinear integral--differential
equations~\eqref{eq:critinst} which has the form
\begin{equation}\label{eq:critinst1}
\left\{
\begin{aligned}
z_1''(\tau) &= f_1(\tau)z_1(\tau)\\
z_2''(\tau) &= f_2(\tau)z_2(\tau),
\end{aligned}
\right.
\end{equation}
where $f_i:S^1\to \mathbb{R}$, $i=1,2$ are smooth functions depending on
$(z_1,z_2)$. Disregarding the dependence of the $f_i$ on $(z_1,z_2)$,
we can view $(z_1,z_2)$ as a solution to the system of {\em uncoupled
linear ODEs}~\eqref{eq:critinst1}.
Arguing by contradiction, suppose now that there exists a point
$\tau_0\in S^1$ such that $z_2(\tau_0)=z_2'(\tau_0)=0$.
Then the function $z_2$ and the zero function both solve the second
equation in~\eqref{eq:critinst1} with the same initial conditions at
$\tau_0$. By uniqueness of solutions of ODEs we conclude $z_2\equiv
0$, contradicting the condition $\|z_2\|>0$ in the definition of
$\mathcal{H}_{in}$.
\hfill $\square$
\subsection{From critical points to generalized solutions}\label{ss:z-to-q-inst}
Let $(z_1,z_2)\in \mathcal{H}_{in}^1$ be a critical point of $\mathcal{B}_{in}$,
so by Proposition~\ref{prop:critin} the maps $z_1,z_2:S^1\to\mathbb{R}$ are
smooth and satisfy~\eqref{eq:critinst}.
As in trhe previous section, for $i=1,2$ we define the smooth maps
$t_{z_i}:S^1\to S^1$ by~\eqref{eq:ti}.
Since by Corollary~\ref{cor:inst-transverse-zeroes} the map $z_i$ has
only finitely many zeroes, it follows from Lemma~\ref{lem:tz-homeo}
that $t_{z_i}:S^1\to S^1$ is a homeomorphism with continuous inverse
$\tau_{z_i} \colon S^1 \to S^1$. We define continuous maps $q_i:S^1\to S^1$ by
\begin{equation}\label{eq:q-i-z-i}
q_i(t) := z_i(\tau_{z_i}(t))^2.
\end{equation}
Then for $i=1,2$ the maps $z_i,q_i:S^1\to\mathbb{R}$ are smooth except at finitely many
points and related by the Levi-Civita transformation~\eqref{eq:LC}
with $\tau=\tau_{z_i}$. The
last condition in the definition of $\mathcal{H}_{in}^1$ implies
$$
q_1(t) > q_2(t) \quad\text{for all }t\in S^1.
$$
Let us now focus on $i=1$. Although by Corollary~\ref{cor:inst-transverse-zeroes}
the function $z_1$ has no zeroes, in the following argument we
will allow $z_1$ to have a finite set $Z_1$ of zeroes; this will
ensure that the same argument carries over to $z_2$ (which may have zeroes).
At points $t\in S^1\setminus t_{z_1}(Z_1)$ we compute:
\begin{eqnarray}\label{qinst}
\ddot{q}_1(t) q_1(t) \nonumber
&=& 2\|z_1\|^4 \frac{z_1''(\tau_{z_1}(t))}{z_1(\tau_{z_1}(t))}
-\frac{\dot{q}_1(t)^2}{2}\\ \nonumber
&=& 2||z_1||^2\cdot ||z_1'||^2
-\frac{2}{||z_1||^2}
-\frac{\dot{q}_1^2(t)}{2}\\ \nonumber
& & +\frac{1}{||z_1||^2}\int_0^1\frac{z_1(\tau_{z_1}(s))^2}
{z_1^2(\tau_{z_1}(s))-z_2^2(\tau_{z_2}(s))}\dot{\tau}_{z_1}(s)ds\\ \nonumber
& &+\frac{z_2^2(\tau_{z_2}(t))}{\big(z_1^2(\tau_{z_1}(t))-z_2^2(\tau_{z_2}(t))\big)^2}\\ \nonumber
& &-\frac{2||z_2||^2}{||z_1||^2}\int_t^1 \frac{z_1^2(\tau_{z_1}(s))z_2'(\tau_{z_2}(s))}{z_2(\tau_{z_2}(s))
\big(z_1^2(\tau_{z_1}(s))-z_2^2(\tau_{z_2}(s))\big)^2}\dot{\tau}_{z_1}(s)ds\\ \nonumber
& &+\frac{2||z_2||^2}{||z_1||^2}\int_0^1
\frac{z_1^2(\tau_{z_1}(s))z_2'(\tau_{z_2}(s))s}{z_2(\tau_{z_2}(s))
\big(z_1^2(\tau_{z_1}(s))-z_2^2(\tau_{z_2}(s))\big)^2}\dot{\tau}_{z_1}(s)ds\\ \nonumber
&=& \frac{||\dot{q}_1||^2}{2}
-\int_0^1 \frac{2}{q_1(s)}ds
-\frac{\dot{q}_1^2(t)}{2}
+\int_0^1\frac{1}{q_1(s)-q_2(s)}ds\\ \nonumber
& &+\frac{q_2(t)}{\big(q_1(t)-q_2(t)\big)^2}
-2||z_2||^2 \int_t^1 \frac{\frac{d}{ds}z_2(\tau_{z_2}(s))}{\dot{\tau}_{z_2}(s)
z_2(\tau_{z_2}(s))\big(q_1(s)-q_2(s)\big)^2}ds\\ \nonumber
& &+2||z_2||^2 \int_0^1 \frac{s\frac{d}{ds}z_2(\tau_{z_2}(s))}{\dot{\tau}_{z_2}(s)
z_2(\tau_{z_2}(s))\big(q_1(s)-q_2(s)\big)^2}ds\\ \nonumber
&=& \frac{||\dot{q}_1||^2}{2}
-\int_0^1 \frac{2}{q_1(s)}ds
-\frac{\dot{q}_1^2(t)}{2}
+\int_0^1\frac{1}{q_1(s)-q_2(s)}ds\\ \nonumber
& &+\frac{q_2(t)}{\big(q_1(t)-q_2(t)\big)^2}
-\int_t^1 \frac{2z_2(\tau_{z_2}(s))\frac{d}{ds}z_2(\tau_{z_2}(s))}{\big(q_1(s)-q_2(s)\big)^2}ds\\ \nonumber
& &+\int_0^1 \frac{2sz_2(\tau_{z_2}(s))\frac{d}{ds}z_2(\tau_{z_2}(s))}{\big(q_1(s)-q_2(s)\big)^2}ds\\ \nonumber
&=& \frac{||\dot{q}_1||^2}{2}
-\int_0^1 \frac{2}{q_1(s)}ds
-\frac{\dot{q}_1^2(t)}{2}
+\int_0^1\frac{1}{q_1(s)-q_2(s)}ds\\ \nonumber
& &+\frac{q_2(t)}{\big(q_1(t)-q_2(t)\big)^2}
-\int_t^1 \frac{\frac{d}{ds}z_2^2(\tau_{z_2}(s))}
{\big(q_1(s)-q_2(s)\big)^2}ds\\ \nonumber
& &+\int_0^1 \frac{s\frac{d}{ds}z_2^2(\tau_{z_2}(s))}{\big(q_1(s)-q_2(s)\big)^2}ds\\ \nonumber
&=& \frac{||\dot{q}_1||^2}{2}
-\int_0^1 \frac{2}{q_1(s)}ds
-\frac{\dot{q}_1^2(t)}{2}
+\int_0^1\frac{1}{q_1(s)-q_2(s)}ds\\ \nonumber
& &+\frac{q_2(t)}{\big(q_1(t)-q_2(t)\big)^2}
-\int_t^1 \frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds\\
& &+\int_0^1 \frac{s\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds\,.
\end{eqnarray}
Here the first equality comes from equation~\eqref{eq:ddot-q} with $q=q_1$ and $z=z_1$.
In the second one we substitute $z_1''$ by~\eqref{eq:critinst} and change
variables $\sigma=\tau_{z_1}(s)$ in the integrals.
In the third one we use the following replacements from~\eqref{eq:q-i-z-i},
\eqref{bov2}, \eqref{bov4} and~\eqref{eq:taudot}:
\begin{gather*}
z_i(\tau_{z_i}(s))^2 = q_i(s),\qquad
\frac{1}{||z_1||^2} = \int_0^1 \frac{1}{q_1(s)}ds,\cr
4||z_1||^2 \cdot ||z_1'||^2 = ||\dot{q}_1||^2,\qquad
\dot\tau_{z_1}(s) = \frac{\|z_1\|^2}{z_1(\tau_{z_1}(s))^2}
\end{gather*}
and the chain rule for $\frac{d}{ds} z_2(\tau_{z_2}(s))$.
In the fourth one we use~\eqref{eq:taudot} to replace $\dot\tau_{z_2}(s)$,
in the fifth one the chain rule for $\frac{d}{ds}{z_2^2(\tau_{z_2}(s))}$,
and in the sixth one~\eqref{eq:q-i-z-i} to insert $q_2(s)$.
Thus $q_1$ satisfies the integral--differential equation
\begin{equation}\label{qinst2}
\ddot{q}_1 = \bigg(c_1
-\frac{\dot{q}_1^2}{2}
+\frac{q_2}{\big(q_1-q_2\big)^2}
-\int_t^1 \frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds
\bigg)\frac{1}{q_1}
\end{equation}
with the constant
$$
c_1 = \frac{||\dot{q}_1||^2}{2}
-\int_0^1 \frac{2}{q_1(s)}ds
+\int_0^1\frac{1}{q_1(s)-q_2(s)}ds
+\int_0^1 \frac{s\,\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds\,.
$$
Let now $t_-<t_+$ be adjacent zeroes of $q_1$ and consider the smooth map
$$
\beta_1:=\frac{\ddot{q}_1-\frac{1}{(q_1-q_2)^2}}{q_1}
\colon (t_-,t_+) \to \mathbb{R}.
$$
From~\eqref{qinst2} we obtain
\begin{eqnarray*}
\beta_1 q_1^2
&=& \ddot{q}_1q_1 - \frac{q_1}{(q_1-q_2)^2}\\
&=& c_1 -\frac{\dot{q}_1^2}{2}
+\frac{1}{q_2-q_1}
-\int_t^1 \frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds\,.
\end{eqnarray*}
Taking the time derivative of this expression for $\beta_1 q_1^2$ we obtain
\begin{eqnarray*}
\dot{\beta}_1 q_1^2+2\beta_1 q_1\dot{q}_1
&=&-\ddot{q}_1 \dot{q}_1 -\frac{\dot{q}_2-\dot{q}_1}{(q_2-q_1)^2}
+\frac{\dot{q}_2}{(q_1-q_2)^2}\\
&=&-\ddot{q}_1 \dot{q}_1+\frac{\dot{q}_1}{(q_1-q_2)^2}\\
&=&-\beta_1 q_1 \dot{q}_1\,.
\end{eqnarray*}
Therefore, dividing both sides by $q_1$, we get
$$
\dot\beta_1q_1=-3\beta_1\dot q_1\,.
$$
This is exactly equation~\eqref{eq:betaqu}. We apply Lemma~\ref{lem:betamu} to get
$$
\beta_1 = -\frac{\mu}{q_1^3}
$$
on $(t_-,t_+)$ for some constant $\mu>0$. Here positivity of $\mu$ follows again
because by definition $\beta_1<0$ at the maximum of $q_1$ on $(t_-,t_+)$.
We substitute the definition of $\beta_1$ in the last displayed equation and solve for $\ddot{q}_1$
to get
\begin{equation}\label{eq:ddotqmu}
\ddot{q}_1(t)=-\frac{\mu}{q_1(t)^2}+\frac{1}{(q_1(t)-q_2(t))^2}, \qquad t \in (t_-,t_+).
\end{equation}
We solve the last equation for $\mu/q_1$ and substitute $\ddot q_1 q_1$
by~\eqref{qinst} to get
\begin{eqnarray}\label{eq:muq}
\frac{\mu}{q_1(t)} \nonumber
&=& -\ddot q_1(t)q_1(t) + \frac{q_1(t)}{(q_1(t)-q_2(t))^2}\\ \nonumber
&=& -\frac{||\dot q_1 ||^2}{2}+\int_0^1\frac{2ds}{q_1(s)}+\frac{\dot q_1(t)^2}{2}
-\int_0^1\frac{ds}{q_1(s)-q_2(s)}+\frac{1}{q_1(t)-q_2(t)}\\
& & +\int_t^1\frac{\dot q_2(s)ds}{(q_1(s)-q_2(s))^2}
-\int_0^1\frac{s\,\dot q_2(s)ds}{(q_1(s)-q_2(s))^2}\,.
\end{eqnarray}
We multiply both sides of the last equation by $q_1(t)$
take the limit $t\to t_{\pm}$ and recall that $q_1(t_\pm)=0$ to get
\begin{equation
\mu = \lim_{t \to t_\pm} \frac{\dot{q}_1(t)^2 q_1(t)}{2}=2||z_1||^4 z_1'\big(\tau_{z_1}(t_\pm)\big)^2.
\end{equation}
Therefore, $\mu$ is a global constant independent of the interval $(t_-,t_+)$.
(This argument is actually only needed for $q_2$ in place of $q_1$;
for $q_1$ itself, since it has no zeroes, we can replace $(t_-,t_+)$
by all of $S^1$ from the outset).
Now note that by Fubini's theorem for every integrable function
$f:[0,1]\to\mathbb{R}$ we have
$$
\int_0^1dt\int_t^1f(s)ds = \int_0^1ds\,f(s)\int_0^s dt = \int_0^1sf(s)ds\,.
$$
Integrating both sides of qquation~\eqref{eq:muq} from $0$ to $1$ and
applying this identity to
$$
f(s):=\frac{\dot q_2}{(q_1(s)-q_2(s))^2}
$$
we obtain
$$
\mu=2\,.
$$
Thus equation~\eqref{eq:ddotqmu} becomes the first equation in~\eqref{eq:inst-interaction}.
Similarly, we deduce that $q_2$ satisfies the second equation
in~\eqref{eq:inst-interaction} outside its zero set.
To see the continuity of the energy $E_1$, we solve equation~\eqref{eq:muq} (with
$\mu=2$) for
\begin{eqnarray*}\nonumber
E_1(t)
&=& \frac{\dot q_1(t)^2}{2} - \frac{2}{q_1(t)}\\
&=& \frac{||\dot q_1 ||^2}{2} - \int_0^1\frac{2\,ds}{q_1(s)}
+ \int_0^1\frac{ds}{q_1(s)-q_2(s)} - \frac{1}{q_1(t)-q_2(t)}\\ \nonumber
& & - \int_t^1\frac{\dot q_2(s)ds}{(q_1(s)-q_2(s))^2}
+ \int_0^1\frac{s\,\dot q_2(s)ds}{(q_1(s)-q_2(s))^2}
\end{eqnarray*}
and note that the right hand side is continuous as a function of $t\in [0,1]$.
Continuity of $E_2$ follows similarly, and we have shown that
$(q_1,q_2)$ is a generalized solution of equation~\eqref{eq:inst-interaction}.
\subsection{From generalized solutions to critical points}\label{ss:q-to-z-inst}
Let now $(q_1,q_2)\in H^1(S^1,\mathbb{R}_{\geq 0}\times\mathbb{R}_{\geq 0})$ be a
generalized solution of equation~\eqref{eq:inst-interaction}. The
definition of a generalized solution implies that $q_1,q_2\in
\mathcal{H}_{ce}^1(S^1,\mathbb{R}_{\geq 0})$. Corollary~\ref{cor:LC} implies that the
set $\mathcal{L}^{-1}(q_1)\times \mathcal{L}^{-1}(q_2)$
consists of $4$ elements. The goal of this section is to show that each
$(z_1,z_2)\in \mathcal{L}^{-1}(q_1)\times \mathcal{L}^{-1}(q_2)$ is a critical point
of $\mathcal{B}_{in}$. To see this, recall that $q_1$ has no zeroes, so
smoothness of $q_1$ implies smoothness of $z_1$. Smoothness of $q_2$ on the
complement of its zero set $Z_{q_2}$ implies smoothness of $z_2$ on
the complement of its zero set $Z_{z_2}$. In
particular, second derivatives of $z_2$ make sense there and we can
make the following statement.
\begin{lemma}\label{lem:eqzinst}
Any $(z_1,z_2)\in Q^{-1}(q_1)\times Q^{-1}(q_2)$ satisfies the critical point
equation \eqref{eq:critinst} on the complement of the set $Z_{z_2}$.
\end{lemma}
Assuming this lemma for the moment, recall that $z_2$ is of class $C^1$ and
\eqref{eq:critinst} expresses $z_2''$ through $z_1,\, z_1',\, z_2,\,z_2'$.
Thus bootstrapping \eqref{eq:critinst} implies that $z_2$
is smooth and \eqref{eq:critinst} holds on the whole $S^1$. Therefore,
it remains to prove Lemma~\eqref{lem:eqzinst}.
{\bf Proof of Lemma~\ref{lem:eqzinst}: }
We will show the desired equation for $z_1$ pretending that it has
a possibly nonempty zero set $Z_{q_1}$. A similar argument will do the
job for $z_2$. Recall the equation satisfied by $q_1$,
\begin{equation}\label{eq:q1}
\ddot q_1=-\frac{2}{q_1^2}+\frac{1}{(q_1-q_2)^2}.
\end{equation}
Set
\begin{equation}\label{eq:beta-inst}
\beta_1:=\frac{\ddot{q}_1-\frac{1}{(q_1-q_2)^2}}{q_1}
\end{equation}
on $S^1\setminus Z_{q_1}$. Then by~\eqref{eq:q1} we have
$$
\beta_1=-\frac{2}{q_1^3},
$$
and taking a time derivative we obtain
$$
\dot\beta_1=\frac{3\cdot 2}{q_1^4}\dot q_1=-\frac{3\beta_1 \dot q_1}{q_1}.
$$
We multiply both sides with $q_1^2$ to get
$$
\dot\beta_1q_1^2=-3\beta_1\dot q_1q_1\,.
$$
We bring $-2\beta_1\dot q_1q_1$ to the other side to continue
$$
\dot\beta_1q_1^2+2\beta_1\dot q_1q_1=-\beta_1\dot q_1q_1\,.
$$
We substitute the original definition \eqref{eq:beta-inst}
of $\beta_1$ in the right hand side to get
$$
\dot\beta_1q_1^2+2\beta_1\dot q_1q_1=-\dot q_1\ddot q_1+\frac{\dot q_2}{(q_1-q_2)^2}
-\frac{\dot q_2-\dot q_1}{(q_2-q_1)^2}.
$$
Integrating both sides from $t$ to $1$ we get
\begin{equation}\label{eq:aux1}
\beta_1q_1^2=-\frac{\dot{q}_1^2}{2}-\int_t^1 \frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds
+\frac{q_2-q_1}{(q_2-q_1)^2}+C
\end{equation}
for some constant $C\in \mathbb{R}$. We bring $-\frac{q_1}{(q_2-q_1)^2}$ to the left hand side and
use \eqref{eq:beta-inst} to simplify it to
$$
\beta_1q_1^2+\frac{q_1}{(q_2-q_1)^2}=\ddot q_1 q_1\,.
$$
Therefore \eqref{eq:aux1} transforms to
\begin{equation}\label{eq:aux2}
\ddot q_1 q_1=-\frac{\dot{q}_1^2}{2}-\int_t^1 \frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds
+\frac{q_2}{(q_2-q_1)^2}+C
\end{equation}
Observe that modulo the exact value of $C$ this is exactly equation~\eqref{qinst2},
which is equivalent the first equation of \eqref{eq:critinst}
An analogous discussion applied to $q_2$
will lead to the second equation of \eqref{eq:critinst}. Therefore, we are left with
computing the constant $C$.
To compute $C$ we use \eqref{eq:q1} to get rid of $\ddot q$ on the left hand side of
\eqref{eq:aux2} and obtain
$$
-\frac{2}{q_1}+\frac{q_1}{(q_1-q_2)^2}
=-\frac{\dot{q}_1^2}{2}-\int_t^1
\frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds
+\frac{q_2}{(q_2-q_1)^2}+C\,.
$$
We solve the last equation for $C$,
$$
C=-\frac{2}{q_1}+\frac{1}{q_1-q_2}
+\frac{\dot{q}_1^2}{2}+\int_t^1 \frac{\dot{q}_2}{\big(q_1-q_2\big)^2}ds\,,
$$
and integrate both sides from $0$ to $1$. Noting that
$$
\int_0^1dt\int_t^1 \frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds=
\int_0^1ds\int_0^s dt\frac{\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}=
\int_0^1\frac{s\dot{q}_2(s)ds}{\big(q_1(s)-q_2(s)\big)^2},
$$
this gives us
$$
C=-\int_0^1\frac{2ds}{q_1(s)}+\int_0^1\frac{ds}{q_1(s)-q_2(s)}
+\frac{||\dot{q}_1||^2}{2}+\int_0^1\frac{s\dot{q}_2ds}{\big(q_1-q_2\big)^2},
$$
which matches the constant in equation~\eqref{qinst}.
This concludes the proof of Lemma~\ref{lem:eqzinst}, and therefore of
Theorem~\ref{thm:inst-interaction}.
\hfill$\square$
\section{Interpolation}\label{sec:interpol}
We now interpolate linearly between the instantaneous and mean
interactions. That is, for $r\in[0,1]$ we consider the system of coupled ODEs
\begin{equation}\label{eq:interpol}
\left\{\;
\begin{aligned}
\ddot{q}_1(t) &= -\frac{2}{q_1(t)^2} + \frac{r}{(q_1(t)-q_2(t))^2}
+ \frac{1-r}{(\overline{q}_1-\overline{q}_2)^2}, \cr
\ddot{q}_2(t) &= -\frac{2}{q_2(t)^2} - \frac{r}{(q_1(t)-q_2(t))^2}
- \frac{1-r}{(\overline{q}_1-\overline{q}_2)^2}.
\end{aligned}
\right.
\end{equation}
For $r=0$ this agrees with the system~\eqref{eq:mean-interaction} for
mean interaction, and for $r=1$ with the
system~\eqref{eq:inst-interaction} for instantaneous interaction.
Solutions of~\eqref{eq:interpol} are crtitcal points of the functional
$r\mathcal{S}_{in}+(1-r)\mathcal{S}_{av}$, which under Levi-Civita
transformation corresponds to the functional
\begin{eqnarray}\label{eq:Br}
\mathcal{B}_r(z_1,z_2)
&:=& r\mathcal{B}_{in}(z_1,z_2) + (1-r)\mathcal{B}_{av}(z_1,z_2) \\ \nonumber
&=&2\sum_{i=1}^2\Bigg(||z_i||^2 \cdot ||z_i'||^2
+\frac{1}{||z_i||^2}\Bigg)
-\frac{r||z_1||^2 \cdot ||z_2||^2}
{||z_1^2||^2\cdot ||z_2||^2-||z_2^2||^2 \cdot ||z_1||^2}\\ \nonumber
& &-\frac{1-r}{||z_1||^2}\int_0^1\frac{z_1(\tau)^2}{z_1^2(\tau)-z_2^2
(\tau_{z_2}(t_{z_1}(\tau)))}d\tau
\end{eqnarray}
defined on $\mathcal{H}_{in}$.
We define a {\em generalized solution $(q_1,q_2)$ of~\eqref{eq:interpol}} as in
Section~\ref{ss:gen-solutions-inst}, only that instead
of~\eqref{eq:inst-interaction} it now solves~\eqref{eq:interpol}.
Then we have the following generalization of Theorems~\ref{thm:mean-interaction}
and~\ref{thm:inst-interaction}.
\begin{thm}[Generalized solutions for interpolated interaction]\label{thm:interpol}
Under the Levi-Civita tranformations~\eqref{eq:LC-zi} with time
changes~\eqref{eq:t-tau-zi}, critical points $(z_1,z_2)$ of the action
functional $\mathcal{B}_r$, $r\in[0,1$, are in 4-to-1 correspondence
with generalized solutions $(q_1,q_2)$ of~\eqref{eq:interpol}.
\end{thm}
{\bf Proof: }
The proof is very similar to the proofs of Theorems~\ref{thm:mean-interaction}
and~\ref{thm:inst-interaction}. Critical points of $\mathcal{B}_r$ are solutions of the problem
\begin{eqnarray}\label{eq:z-interpol}
z_1''(\tau) \nonumber
&=&\frac{||z_1'||^2z_1(\tau)}{||z_1||^2}-\frac{z_1(\tau)}{||z_1||^6}+\frac{(1-r)z_1(\tau)}{2||z_1||^6}\int_0^1\frac{z_1(\sigma)^2}{z_1^2(\sigma)-z_2^2(\tau_{z_2}(t_{z_1}(\sigma)))}d\sigma\\ \nonumber
& &+\frac{(1-r)z_1(\tau)z_2^2(\tau_{z_2}(t_{z_1}(\tau)))}{2||z_1||^4\big(z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))\big)^2}\\ \nonumber
& &-\frac{(1-r)||z_2||^2z_1(\tau)}{||z_1||^6}
\int_\tau^1 \frac{z_1^2(\sigma) z_2'(\tau_{z_2}(t_{z_1}(\sigma)))}{z_2(\tau_{z_2}(t_{z_1}(\sigma)))\big(z_1^2(\sigma)-z_2^2(\tau_{z_2}(t_{z_1}(\sigma)))\big)^2}d\sigma\\ \nonumber
& &+\frac{(1-r)||z_2||^2 z_1(\tau)}{||z_1||^6}
\int_0^1 \frac{z_1^2(\sigma) z_2'(\tau_{z_2}(t_{z_1}(\sigma))) t_{z_1}(\sigma)}
{z_2(\tau_{z_2}(t_{z_1}(\sigma)))\big(z_1^2(\sigma)-z_2^2(\tau_{z_2}(t_{z_1}(\sigma)))\big)^2}d\sigma\\ \nonumber
& &-\frac{r||z_2||^4 ||z_1^2||^2 z_1(\tau)}{2||z_1||^2 \big(||z_1^2||^2 ||z_2||^2-||z_2^2||^2
||z_1||^2\big)^2}\\ \nonumber
& &+\frac{r||z_2||^4 z_1^3(\tau)}{\big(||z_1^2||^2 ||z_2||^2-||z_2^2||^2||z_1||^2\big)^2}\\ \nonumber
z_2''(\tau)&=&\frac{||z_2'||^2z_2(\tau)}{||z_2||^2}-\frac{z_2(\tau)}{||z_2||^6}-\frac{(1-r)z_2(\tau)}{2||z_2||^6}\int_0^1\frac{z_2(\sigma)^2}{z_2^2(\sigma)-z_1^2(\tau_{z_1}(t_{z_2}(\sigma)))}d\sigma\\ \nonumber
& &-\frac{(1-r)z_2(\tau)z_1^2(\tau_{z_1}(t_{z_2}(\tau)))}{2||z_2||^4\big(z_2^2(\tau)-z_1^2(\tau_{z_1}(t_{z_2}(\tau)))\big)^2}\\ \nonumber
& &+\frac{(1-r)||z_1||^2z_2(\tau)}{||z_2||^6}
\int_\tau^1 \frac{z_2^2(\sigma) z_1'(\tau_{z_1}(t_{z_2}(\sigma)))}{z_1(\tau_{z_1}(t_{z_2}(\sigma)))\big(z_2^2(\sigma)-z_1^2(\tau_{z_1}(t_{z_2}(\sigma)))\big)^2}d\sigma\\ \nonumber
& &-\frac{(1-r)||z_1||^2 z_2(\tau)}{||z_2||^6}
\int_0^1 \frac{z_2^2(\sigma) z_1'(\tau_{z_1}(t_{z_2}(\sigma))) t_{z_2}(\sigma)}{z_1(\tau_{z_1}(t_{z_2}(\sigma)))\big(z_2^2(\sigma)-z_1^2(\tau_{z_1}(t_{z_2}(\sigma)))\big)^2}d\sigma\\ \nonumber
& &+\frac{r||z_1||^4||z_2^2||^2 z_2(\tau)}{2||z_2||^2\big(||z_1^2||^2||z_2||^2-||z_2^2||^2||z_1||^2\big)^2}\\
& &-\frac{r||z_1||^4 z_2^3(\tau)}{\big(||z_1^2||^2||z_2||^2-||z_2^2||^2||z_1||^2\big)^2}
\end{eqnarray}
which interpolates between problems~\eqref{eq:z-av}
and~\eqref{eq:critinst}.
Suppose now that $(z_1,z_2)$ is a critical point of $\mathcal{B}_r$ and define
$$q_i(t) := z_i^2(\tau_{z_i}(t)), \qquad i \in \{1,2\}.$$
From~\eqref{eq:z-interpol} we obtain for $q_1$ the equation
\begin{eqnarray}\label{qint}
2\ddot{q}_1(t) q_1(t)&=&||\dot{q}_1||^2-\int_0^1 \frac{4}{q_1(s)}ds-\dot{q}_1^2(t)+\int_0^1\frac{2(1-r)}{q_1(s)-
q_2(s)}ds\\ \nonumber
& &+\frac{2(1-r)q_2(t)}{\big(q_1(t)-q_2(t)\big)^2}-\int_t^1 \frac{2(1-r)\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds\\ \nonumber
& &+\int_0^1 \frac{2(1-r)s\dot{q}_2(s)}{\big(q_1(s)-q_2(s)\big)^2}ds
-\frac{2r\overline{q}_1}{(\overline{q}_1-\overline{q}_2)^2}
+\frac{4r q_1(t)}{(\overline{q}_1-\overline{q_2})^2}
\end{eqnarray}
which interpolates between (\ref{bov6}) and (\ref{qinst}). As before,
outside collisions we define
$$\beta_1:=\frac{\ddot{q}_1-\frac{r}{(q_1-q_2)^2}-\frac{1-r}{(\overline{q}_1-\overline{q}_2)^2}}{q_1}.$$
Again $\beta_1$ solves the ODE
$$\dot{\beta}_1 q_1=-3\beta_1 \dot{q}_1$$
and it follows that
$$\beta_1=-\frac{\mu}{q_1^3},$$
where $\mu$ is locally constant on the complement of collisions on the
circle. Using the continuity of $z_1'$, it follows again that $\mu$ is
constant and from~\eqref{qint} we conclude that $\mu=2$.
Therefore, outside of collisions $q_1$ solves the first equation
in~\eqref{eq:interpol}. Similarly, it follows that outside collisions
$q_2$ solves the second equation in~\eqref{eq:interpol}.
The converse direction is proved similarly as in the previous cases.
\hfill$\square$
\section{Existence of symmetric frozen planet orbits}
A {\em simple frozen planet orbit of period $T>0$} is a map
$(q_1,q_2)\in H^1(\mathbb{R}/T\mathbb{Z},\mathbb{R}^2)$ with the following properties:
\begin{enumerate}
\item $q_1(t)>q_2(t)\geq 0$ for all $t\in\mathbb{R}/T\mathbb{Z}$;
\item $q_2$ has a unique zero at $t=0$;
\item $(q_1,q_2):(0,T)\to\mathbb{R}^2$ is
smooth and satisfies~\eqref{eq:inst-interaction};
\item the energies
$$
E_i(t) := \frac{\dot q_i(t)^2}{2}-\frac{2}{q_i(t)},\qquad t\in
(0,T)
$$
extend to continuous functions $E_i:\mathbb{R}/T\mathbb{Z}\to\mathbb{R}$.
\end{enumerate}
Here simplicity corresponds to the second condition, and every frozen
planet orbit is a multiple cover of a simple one.
Recall that the individual energies $E_i$ need not be constant, but
the total energy
$$
E = E_1(t) + E_2(t) + \frac{1}{q_1(t)-q_2(t)}
$$
is constant and negative. A simple frozen planet orbit of period $T>0$
is called {\em symmetric} if its satisfies in addition
$$
q(t) = q(T-t)\quad\text{for all }t\in\mathbb{R}/T\mathbb{Z}.
$$
In this section we prove the following result which corresponds to
Theorem\,C in the Introduction.
\begin{thm}\label{thm:existence}
For every $E<0$ there exists a symmetric simple frozen planet orbit of energy $E$.
\end{thm}
{\bf Rescaling. }
Let $q=(q_1,q_2):\mathbb{R}\to(\mathbb{R}_{\geq0})^2$ be a generalized solution
of~\eqref{eq:inst-interaction} of period $T$ and energy $E$.
Direct computation shows that for each $c>0$,
$$
q_c(t):=c^2q(c^{-3}t)
$$
is again a generalized solution
of~\eqref{eq:inst-interaction} of period $c^3T$ and energy $c^{-2}E$.
Therefore, given a generalized solution $q$ of~\eqref{eq:inst-interaction}
of period $1$ and negative energy, rescaling yields similar
solutions with any prescribed period, or alternatively with any
prescribed negative energy.
As a consequence, we will from now on restrict our discussion to
generalized solutions of period $1$.
\subsection{Symmetries}\label{ss:symmetries}
In this subsection we describe the symmetries of the variational
problems in Sections~\ref{sec:Kepler}, \ref{sec:mean-interaction}
and~\ref{sec:inst-interaction}.
{\bf The Kepler problem. }
The functional $\mathcal{Q}$ in~\eqref{eq:Q} is obviously invariant under the
following transformations:
\begin{itemize}
\item time shift $T_sz(\tau):=z(s+\tau)$, $s\in S^1$;
\item time reversal $Rz(\tau):=z(-\tau)$;
\item sign reversal $Sz(\tau):= -z(\tau)$.
\end{itemize}
\begin{lemma}\label{lem:time-rev}
The homeomorphism $t_z:S^1\to S^1$ defined in~\eqref{eq:tz} and its
inverse $\tau_z$ transform under time reversal as
$$
t_{Rz}(\tau) = -t_z(-\tau),\qquad \tau_{Rz}(t) = -\tau_z(-t).
$$
\end{lemma}
{\bf Proof: }
The first equation follows from
\begin{eqnarray*}
t_{Rz}(\tau)
&=& \frac{\int_0^\tau (Rz)^2(\sigma)d\sigma}{\|Rz\|^2}
= \frac{\int_0^\tau z^2(-\sigma)d\sigma}{\|z\|^2}
= -\frac{\int_0^{-\tau} z^2(\sigma)d\sigma}{\|z\|^2}
= -t_{z}(-\tau),
\end{eqnarray*}
and the second equation follows from the first one by checking that the
two maps are inverse to one another:
\begin{align*}
t_{Rz}(\tau_{Rz}(t))
= t_{Rz}(-\tau_{z}(-t))
= -t_{z}(\tau_{z}(-t))
= -(-t) =t.
\end{align*}
\hfill $\square$
It follows that under the Levi-Civita transformation
$q(t)=z(\tau_z(t))^2$ the above symmetries correspond to time shift
$T_sq(t)=q(s+t)$, time reversal $Rq(t)=q(-t)$, and the identity $Sq(t)=q(t)$.
Thus the first two symmetries of $z$ correspond to actual symmetries
of the Kepler problem, while the sign change $Sz=-z$ just expresses
the fact the the Levi-Civita transformation defines a 2-to-1 correspondence.
\smallskip
{\bf Mean and instantaneous interactions. }
The problems with interaction have the following symmetries.
\begin{lemma}\label{lem:inv}
The functionals $\mathcal{B}_{av}$ and $\mathcal{B}_{in}$ in~\eqref{eq:Bav}
and~\eqref{eq:Bin} are invariant under the
following transformations of $z=(z_1,z_2)$:
\begin{itemize}
\item joint time shift $T_sz(\tau):=z(s+\tau)$, $s\in S^1$;
\item joint time reversal $Rz(\tau):=z(-\tau)$;
\item separate sign reversals $S_1(z_1,z_2):=(-z_1,z_2)$ and $S_2(z_1,z_2):=(z_1,-z_2)$.
\end{itemize}
\end{lemma}
\textbf{Proof: }
Invariance of $\mathcal{B}_{av}$ and $\mathcal{B}_{in}$ under joint time shift $T_s$ and
separate sign changes $S_i$ is obvious, and so is invariance of $\mathcal{B}_{av}$ under
joint time reversal $R$ (in fact, $\mathcal{B}_{av}$ is even invariant under
separate time reversals). For invariance of $\mathcal{B}_{in}$ under $R$ we
write the instantaneous interaction from~\eqref{eq:I} in the form
$$
\mathcal{I}(z_1,z_2)=\int_0^1\frac{dt}{z_2^2(\tau_{z_2}(t))-z_1^2(\tau_{z_1}(t))}.
$$
Using Lemma~\ref{lem:time-rev} we compute
$$
(Rz_i)(\tau_{Rz_i}(t)) = z_i(-\tau_{Rz_i}(t)) = z_i(\tau_{z_i}(-t))
$$
for $i=1,2$, and therefore
\begin{eqnarray*}
\mathcal{I}(R(z_1,z_2))
&=& \int_0^1\frac{dt}{z_2^2(\tau_{z_2}(-t))-z_1^2(\tau_{z_1}(-t))}\\
&=& \int_0^1\frac{dt}{z_2^2(\tau_{z_2}(t))-z_1^2(\tau_{z_1}(t))}
= \mathcal{I}(z_1,z_2).
\end{eqnarray*}
\hfill $\square$
\subsection{Twisted loops}\label{ss:twisted}
Theorems~\ref{thm:BOV}, ~\ref{thm:mean-interaction} and~\ref{thm:mean-interaction}
establish a correspondence between critical points of the functionals
$\mathcal{Q}$, $\mathcal{B}_{av}$ and $\mathcal{B}_{in}$ and generalized solutions with an
even number of zeroes. In this subsection we explain how to deal with
generalized solutions with an {\em odd} number of zeroes.
{\bf The Kepler problem. }
Consider $q\in H^1(S^1,\mathbb{R}_{\geq 0})$ satisfying all conditions on a
generalized solution of~\eqref{eq:Kepler} except that it has an odd
number of zeroes. By slight abuse of notation we will still refer to
such $q$ as a ``generalized solution''.
Recall that $S^1=\mathbb{R}/\mathbb{Z}$. Then we can view $q$ as a map
$\widetilde q\in H^1(\mathbb{R}/2\mathbb{Z},\mathbb{R}_{\geq 0})$ with an even number of zeroes and
associate to it as in Section~\ref{ss:q-to-z-Kepler} a
homeomorphism $\tau_q:\mathbb{R}/2\mathbb{Z}\to \mathbb{R}/2\mathbb{Z}$ and a map $z\in H^1(\mathbb{R}/2\mathbb{Z},\mathbb{R})$
such that
$$
z(\tau_q(t))^2=\widetilde q(t),\qquad t\in\mathbb{R}/2\mathbb{Z}.
$$
By construction these maps satisfy
$$
\widetilde q(t+1) = \widetilde q(t),\qquad
\tau_q(t+1) = \tau_q(t),\qquad
z(t+1) = -z(t),\qquad t\in\mathbb{R}/2\mathbb{Z}.
$$
This leads us to introduce for each $k\in\mathbb{N}_0$ the Hilbert space of
{\em twisted loops}
$$
\widetilde H^k(S^1,\mathbb{R}) := \{z\in H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R})\mid z(\tau+1)=-z(\tau)
\text{ for all }\tau\}
$$
with the inner product
$$
\langle z,v\rangle := \int_0^1z(\tau)v(\tau)d\tau = \frac12\int_0^2z(\tau)v(\tau)d\tau.
$$
Note that $\widetilde H^k(S^1,\mathbb{R})$ is the fixed point set of the linear involution
$$
\sigma=S\circ T_1:H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R}) \to H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R}),\qquad \sigma z(\tau)=-z(\tau+1),
$$
where $T_1$ and $S$ are the time shift by $1$ and sign reversal from
Section\ref{ss:symmetries}. Thus $\widetilde H^k(S^1,\mathbb{R})$ is a closed linear
subspace of $H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R})$, with inner product the one induced from
$H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R})$ divided by $2$.
We define
\begin{equation*
\mathcal{Q}: \widetilde H^1(S^1,\mathbb{R})\setminus\{0\}\to\mathbb{R},\qquad
\mathcal{Q}(z) := 2\|z\|^2\|z'\|^2 + \frac{N}{\|z\|^2}
\end{equation*}
by the same formula as in~\eqref{eq:Q}. As in
Section~\ref{ss:z-to-q-Kepler} it follows that $z\in \widetilde H^1(S^1,\mathbb{R})\setminus\{0\}$
is a critical point of $\mathcal{Q}$ if and only if
\begin{equation}\label{eq:DQ-twisted}
\langle z',v'\rangle + \langle az,v\rangle = 0\qquad \text{for all }v\in \widetilde H^1(S^1,\mathbb{R}),
\end{equation}
with the constant
\begin{equation*
a = \frac{\|z'\|^2}{\|z\|^2} - \frac{N}{2\|z\|^6}.
\end{equation*}
To proceed we need the following lemma.
\begin{lemma}\label{lem:twisted-var}
Suppose that $f,g\in\widetilde H^0(S^1,\mathbb{R})$ satisfy
$$
\langle f,v'\rangle + \langle g,v\rangle = 0 \qquad\text{for all }v\in\widetilde H^1(S^1,\mathbb{R}).
$$
Then $f\in\widetilde H^0(S^1,\mathbb{R})$ with weak derivative $f'=g$.
\end{lemma}
{\bf Proof: }
By definition of $\widetilde H^0(S^1,\mathbb{R})$ we have $f,g\in H^0(\mathbb{R}/2\mathbb{Z},\mathbb{R})$
with $f(\tau+1)=-f(\tau)$ and $g(\tau+1)=-g(\tau)$, so the hypothesis of
the lemma reads
\begin{equation}\label{eq:weak-der}
\int_0^2 f(\tau)v'(\tau)d\tau + \int_0^2g(\tau)v(\tau)d\tau = 0
\end{equation}
for all $v\in H^1(\mathbb{R}/2\mathbb{Z},\mathbb{R})$ with $v(\tau+1)=-v(\tau)$.
By definition of the weak derivative, we need to show that~\eqref{eq:weak-der}
holds for all $v\in H^1(\mathbb{R}/2\mathbb{Z},\mathbb{R})$ (not necessarily satisfying $v(\tau+1)=-v(\tau)$).
Multiplying by bump functions and using linearity, it suffices to show
this for $v\in H^1(\mathbb{R}/2\mathbb{Z},\mathbb{R})$ with support in an interval
$I\subset\mathbb{R}/2\mathbb{Z}$ of length less than $1$. Given such $v$ we define
$\widetilde v\in H^1(\mathbb{R}/2\mathbb{Z},\mathbb{R})$ by
$$
\widetilde v(\tau) := \begin{cases}
v(\tau) & \tau \in I, \\
-v(\tau-1) & \tau-1 \in I, \\
0 & \text{otherwise}.
\end{cases}
$$
Using $f(\tau+1)=-f(\tau)$ and $\widetilde v'(\tau+1)=\widetilde v'(\tau)$ we compute
\begin{align*}
\int_0^2 f(\tau)\widetilde v'(\tau)d\tau
&= \int_I f(\tau)\widetilde v'(\tau)d\tau + \int_I f(\tau+1)\widetilde v'(\tau+1)d\tau\cr
&= \int_I f(\tau)\widetilde v'(\tau)d\tau + \int_I f(\tau)\widetilde v'(\tau)d\tau\cr
&= 2\int_0^2 f(\tau)v'(\tau)d\tau,
\end{align*}
and similarly
$$
\int_0^2 g(\tau)\widetilde v(\tau)d\tau
= 2\int_0^2 g(\tau)v(\tau)d\tau.
$$
Since~\eqref{eq:weak-der} holds with $\widetilde v$ in place of $v$, we conclude
\begin{align*}
0
&= \int_0^2 f(\tau)\widetilde v'(\tau)d\tau + \int_0^2g(\tau)\widetilde v(\tau)d\tau
= 2\bigg(\int_0^2 f(\tau)v'(\tau)d\tau + \int_0^2g(\tau)v(\tau)d\tau\bigg)
\end{align*}
and the lemma is proved.
\hfill $\square$
Let us now return to the critical point $z\in \widetilde
H^1(S^1,\mathbb{R})\setminus\{0\}$ of $\mathcal{Q}$ satisfying~\eqref{eq:DQ-twisted} above.
Applying Lemma~\ref{lem:twisted-var} with $f=z'$ and $g(z)=az$, we
conclude that $z\in\widetilde H^2(S^1,\mathbb{R})$ and its second weak derivative satisfies
$$
z'' = az.
$$
This is the same ODE~\eqref{eq:z-Kepler} as for critical points in
the untwisted case. Now the arguments in Section~\ref{sec:Kepler}
carry over without further changes to show the following twisted version of
Theorem~\ref{thm:BOV}.
\begin{thm}\label{thm:BOV-twisted}
Under the Levi-Civita tranformation~\eqref{eq:LC} with time
change~\eqref{eq:t-tau}, critical points $z$ of the action
functional $\mathcal{Q}: \widetilde H^1(S^1,\mathbb{R})\setminus\{0\}\to\mathbb{R}$ on twisted
loops are in 2-to-1 correspondence
with generalized solutions $q:S^1\to\mathbb{R}_{\geq 0}$ of~\eqref{eq:Kepler}
having an odd number of zeroes.
\hfill $\square$
\end{thm}
{\bf Mean and instantaneous interactions. }
Consider now the helium atom with. Recall that in a frozen planet
configuration the inner electron $q_2$ should undergo repeated
collisions with the nucleus while the outer electron $q_1$ experiences no
collisions. If in period $1$ the inner electron undergoes an odd number of
collisions (for example a single one), the Levi-Civita transformed
maps $z_1,z_2$ will be $2$-periodic and satisfy
$$
z_1(\tau+1) = z_1(\tau),\qquad z_2(\tau+1) = -z_2(\tau).
$$
This leads us to introduce for each $k\in\mathbb{N}_0$ the Hilbert space of
{\em twisted loops}
\begin{align*}
\widetilde H^k(S^1,\mathbb{R}^2)
:= \bigl\{&z=(z_1,z_2)\in H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R})\;\bigl| \\
&z_1(\tau+1) = z_1(\tau),\ z_2(\tau+1)=-z_2(\tau)\text{ for all }\tau\bigr\}
\end{align*}
with the inner product
$$
\langle z,v\rangle := \sum_{i=1}^2\int_0^1z(\tau)v(\tau)d\tau
= \frac12\sum_{i=1}^2\int_0^2z(\tau)v(\tau)d\tau.
$$
Note that $\widetilde H^k(S^1,\mathbb{R}^2)$ is the fixed point set of the linear involution
$$
\sigma=S_2\circ T_1:H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R}^2) \to H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R}^2),\quad \sigma
z(\tau)=\bigl(z_1(\tau+1),-z_2(\tau+1)\bigr),
$$
where $T_1$ and $S_2$ are the joint time shift by $1$ and sign reversal in
the second component from
Section~\ref{ss:symmetries}. Thus $\widetilde H^k(S^1,\mathbb{R}^2)$ is a closed linear
subspace of $H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R}^2)$, with inner product the one induced from
$H^k(\mathbb{R}/2\mathbb{Z},\mathbb{R}^2)$ divided by $2$.
We define open subsets
\begin{equation*
\widetilde{\mathcal{H}}_{av}^1 := \Bigg\{z=(z_1,z_2) \in
\widetilde H^1(S^1,\mathbb{R}^2)\;\Bigl|\; ||z_1||>0,\,\,||z_2||>0,\,\,
\frac{||z_1^2||^2}{||z_1||^2}>\frac{||z_2^2||^2}{||z_2||^2}\Bigg\}
\end{equation*}
and
\begin{equation}\label{eq:Hin-twisted}
\begin{aligned}
\widetilde{\mathcal{H}}_{in}^1 := \Bigl\{
&z=(z_1,z_2) \in \widetilde H^1(S^1,\mathbb{R}^2)\;\Bigl|\;
\|z_1\|>0,\,\,\|z_2\|>0,\,\, \cr
&\ \ z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))>0 \,\,\text{for
all}\,\,\tau\in S^1\Bigr\}
\end{aligned}
\end{equation}
as in Sections~\ref{sec:mean-interaction} and~\ref{sec:inst-interaction},
and we define the mean and instantaneous interaction functionals and
their interpolation
\begin{equation*}
\mathcal{B}_{av}:\widetilde{\mathcal{H}}_{av}^1\to\mathbb{R},\qquad
\mathcal{B}_{in},\mathcal{B}_r:\widetilde{\mathcal{H}}_{in}^1\to\mathbb{R}
\end{equation*}
by the same formulas as in~\eqref{eq:Bav}, ~\eqref{eq:Bin} and~\eqref{eq:Br}.
Note that in formulas~\eqref{eq:DBav} and~\eqref{eq:DBin} for the
derivatives of $\mathcal{B}_{av}$ and $\mathcal{B}_{in}$ the leading order terms $\langle
z_i',v_i'\rangle$ are the same as in the Kepler problem. Arguing as above
using Lemma~\ref{lem:twisted-var}, we thus conclude that critical
points of $\mathcal{B}_{av}$ and $\mathcal{B}_{r}$ on twisted loops
still satisfy the same equations~\eqref{eq:z-av}
and~\eqref{eq:z-interpol} as in the untwisted case. Therefore, in
analogy with Theorem~\ref{thm:BOV-twisted}, we obtain the following
twisted version of Theorems~\ref{thm:mean-interaction}
and~\ref{thm:interpol}.
\begin{thm}\label{thm:interpol-twisted}
Under the Levi-Civita tranformations~\eqref{eq:LC-i} with time
changes~\eqref{eq:t-tau-i}, critical points $(z_1,z_2)$ of the action
functional $\mathcal{B}_r:\widetilde{\mathcal{H}}_{av}^1\to\mathbb{R}$ on twisted
loops are in 4-to-1 correspondence
with generalized solutions $(q_1,q_2)$ of~\eqref{eq:mean-interaction}
with $q_1$ having an even and $q_2$ an odd number of zeroes.
Similarly, for each $r\in[0,1]$ critical points $(z_1,z_2)$ of the action
functional $\mathcal{B}_{r}:\widetilde{\mathcal{H}}_{in}^1\to\mathbb{R}$ on twisted
loops are in 4-to-1 correspondence
with generalized solutions $(q_1,q_2)$ of~\eqref{eq:interpol}
with $q_1$ having an even and $q_2$ an odd number of zeroes.
\hfill $\square$
\end{thm}
\subsection{Symmetric loops}\label{ss:symmetric}
In order to study symmetric frozen planet orbits, we introduce
for each $k\in\mathbb{N}_0$ the Hilbert space of {\em symmetric loops}
\begin{align*}
\widehat H^k(S^1,\mathbb{R}^2)
:= \bigl\{&z=(z_1,z_2)\in \widetilde H^k(S^1,\mathbb{R}^2)\;\bigl| \\
&z(\tau) = z(1-\tau)\text{ for all }\tau\bigr\}
\end{align*}
with the inner product induced from $\widehat H^k(S^1,\mathbb{R}^2)$.
Thus $\widehat H^k(S^1,\mathbb{R}^2)$ is the fixed point set of the linear involution
$$
\rho=R\circ T_1:\widetilde H^k(S^1,\mathbb{R}^2) \to \widetilde H^k(S^1,\mathbb{R}^2),\quad \rho
z(\tau)=z(1-\tau),
$$
where $T_1$ and $R$ are the joint time shift by $1$ and joint time
reversal in from Section~\ref{ss:symmetries}.
Elements $(z_1,z_2)\in \widehat H^k(S^1,\mathbb{R}^2)$ are $2$-periodic loops
satisfying for all $\tau\in\mathbb{R}/2\mathbb{Z}$ the conditions
$$
z_1(\tau) = z_1(\tau+1) = z_1(1-\tau),\qquad
z_2(\tau) = -z_2(\tau+1) = z_2(1-\tau).
$$
Taking derivatives they imply
$$
z_1'(\tau) = z_1'(\tau+1) = -z_1'(1-\tau),\qquad
z_2'(\tau) = -z_2'(\tau+1) = -z_2'(1-\tau).
$$
In particular, at $\tau=0$ and $\tau=1/2$ they imply
$$
z_1'(0) = z_2(0) = 0,\qquad z_1'(1/2) = z_2'(1/2) = 0.
$$
Thus $z_1$ is $1$-periodic with critical points at $\tau=0$ and
$\tau=1/2$, while $z_2$ is $2$-periodic with zeroes at $\tau=0$ and
$\tau=1$ and with critical points at $\tau=1/2$ and $\tau=3/2$.
The images $q_i(t) = z_i(\tau_{z_i}(t))$ under the Levi-Civita transformation
are $1$-periodic and {\em symmetric}, i.e.
$$
q_i(t)=q_i(1-t),\qquad t\in S^1.
$$
In particular, if $(q_1,q_2)$ satisfies in addition the
ODE~\eqref{eq:inst-interaction},
then $(q_1,q_2)$ is a symmetric frozen planet orbit.
Again, we define open subsets
\begin{equation*
\widehat{\mathcal{H}}_{av}^k := \Bigg\{z=(z_1,z_2) \in
\widehat H^k(S^1,\mathbb{R}^2)\;\Bigl|\; ||z_1||>0,\,\,||z_2||>0,\,\,
\frac{||z_1^2||^2}{||z_1||^2}>\frac{||z_2^2||^2}{||z_2||^2}\Bigg\}
\end{equation*}
and
\begin{equation}\label{eq:Hin-symm}
\begin{aligned}
\widehat{\mathcal{H}}_{in}^k := \Bigl\{
&z=(z_1,z_2) \in \widehat H^k(S^1,\mathbb{R}^2)\;\Bigl|\;
\|z_1\|>0,\,\,\|z_2\|>0,\,\, \cr
&\ \ z_1^2(\tau)-z_2^2(\tau_{z_2}(t_{z_1}(\tau)))>0 \,\,\text{for
all}\,\,\tau\in S^1\Bigr\},
\end{aligned}
\end{equation}
and we define the mean and instantaneous interaction functionals and
their interpolation
\begin{equation*}
\mathcal{B}_{av}:\widehat{\mathcal{H}}_{av}^1\to\mathbb{R},\qquad
\mathcal{B}_{in},\mathcal{B}_r:\widehat{\mathcal{H}}_{in}^1\to\mathbb{R}
\end{equation*}
by the same formulas as in~\eqref{eq:Bav}, ~\eqref{eq:Bin} and~\eqref{eq:Br}.
Arguing as in the previous subsection, using a variant of
Lemma~\ref{lem:twisted-var}, we conclude that critical
points of $\mathcal{B}_{av}$ and $\mathcal{B}_{r}$ on symmetric loops
still satisfy the same equations~\eqref{eq:z-av}
and~\eqref{eq:z-interpol} and we obtain the following
symmetric version of Theorems~\ref{thm:mean-interaction}
and~\ref{thm:interpol}.
\begin{thm}\label{thm:interpol-symm}
Under the Levi-Civita tranformations~\eqref{eq:LC-i} with time
changes~\eqref{eq:t-tau-i}, critical points $(z_1,z_2)$ of the action
functional $\mathcal{B}_{av}:\widehat{\mathcal{H}}_{av}^1\to\mathbb{R}$ on symmetric
loops are in 4-to-1 correspondence with symmetric
generalized solutions $(q_1,q_2)$ of~\eqref{eq:mean-interaction}
with $q_1$ having an even and $q_2$ an odd number of zeroes.
Similarly, for each $r\in[0,1]$ critical points $(z_1,z_2)$ of the action
functional $\mathcal{B}_{r}:\widehat{\mathcal{H}}_{in}^1\to\mathbb{R}$ on symmetric
loops are in 4-to-1 correspondence with symmetric
generalized solutions $(q_1,q_2)$ of~\eqref{eq:interpol}
with $q_1$ having an even and $q_2$ an odd number of zeroes
(i.e.~with symmetric frozen planet orbits).
\hfill $\square$
\end{thm}
\subsection{Proof of the Existence Theorem~\ref{thm:existence}}\label{ss:existence-proof}
Now we are ready to prove Theorem~\ref{thm:existence}. By the
rescaling discussion following the theorem, it suffices to consider the case of
period $1$ without prescribing the energy.
Using the notation of the previous subsection we define
\begin{equation}\label{def:X}
X := \{z=(z_1,z_2)\in\widehat\mathcal{H}^2_{in}\mid
z_i(\tau)>0\text{ for all }\tau\in(0,1)\text{ and }i=1,2\}.
\end{equation}
This is an open subset of the Hilbert space $\widehat H^2(S^1,\mathbb{R}^2)$ and
thus a Hilbert manifold. Note that the twisting conditions
$z_1(\tau+1)=z_1(\tau)$ and $z_2(\tau+1)=-z_2(\tau)$ imply that
$z_1(\tau)>0$ for all $\tau\in\mathbb{R}$ and $z_2(\tau)<0$ for $\tau\in(1,2)$. This
ensures that $z_2$ is simple of minimal period $2$ and it
removes the symmetries $z_i\mapsto\pm z_i$.
We consider the Hilbert space
$$
Y := \widehat H^0(S^1,\mathbb{R}^2)
$$
and the $L^2$-gradient of the interpolation functional $\mathcal{B}_r=(1-r)\mathcal{B}_{in}+r\mathcal{B}_{av}$,
$$
\nabla\mathcal{B}_r=(1-r)\nabla\mathcal{B}_{in}+r\nabla\mathcal{B}_{av}:X\to Y,\qquad r\in[0,1].
$$
According to Theorem~\ref{thm:Fredholm}, for each $r\in[0,1]$ this is
a $C^1$-Fredholm map of index zero. Thus
$$
F:[0,1]\times X\to Y,\qquad (r,z)\mapsto \nabla\mathcal{B}_r(z)
$$
is a $C^1$-Fredholm operator of index $1$.
According to Theorem~\ref{thm:interpol-symm}, for each $r\in[0,1]$
zeroes $z\in X$ of $\nabla\mathcal{B}_r$ correspond under the Levi-Civita
transformation to symmetric generalized solutions
$q=(q_1,q_2):S^1\to\mathbb{R}^2$ of~\eqref{qint}. The condition $z_2(\tau)>0$
for all $\tau\in(0,1)$ in the definition of $X$ implies that $q_2$ has
a unique zero at $t=0$.
By the main result in~\cite{frauenfelder2} there exists a constant
$\kappa$ such that
$$
\max_{t\in S^1}\bigg\{q_1(t),\frac{1}{q_1(t)-q_2(t)}\bigg\}\leq\kappa
$$
for all such solutions $q$ of~\eqref{qint} and all $r\in[0,1]$. Thus
on $(r,z)\in F^{-1}(0)$ the function $z_1^2(t)$ is uniformly bounded
from above and the difference
$z_1^2(\tau_{z_1}(t))-z_2^2(\tau_{z_2}(t))$ is uniformly bounded away
from zero. In view of the ODE~\eqref{eq:z-interpol}, this implies that
the zero set $F^{-1}(0)\subset[0,1]\times X$ is compact. Hence
$F:[0,1]\times X\to Y$ is a homotopy as in Theorem~\ref{thm:zero-Fredholm}
between $f_0=\nabla\mathcal{B}_{in}$ and $f_1=\nabla\mathcal{B}_{av}$, and it follows that
$f_0,f_1$ have well-defined mod $2$ Euler numbers satisfying
$$
\chi(\nabla\mathcal{B}_{in}) = \chi(\nabla\mathcal{B}_{av}).
$$
By Theorem~\ref{thm:Euler-mean-intersection}, the mod $2$ Euler number
of $\nabla\mathcal{B}_{av}$ satisfies
$$
\chi(\nabla\mathcal{B}_{av}) = 1.
$$
Together with the previous displayed equation this shows that
$\chi(\nabla\mathcal{B}_{in})=1$, so $\nabla\mathcal{B}_{in}$ possesses a zero
whose Levi-Civita transform is the desired symmetric frozen planet orbit.
This concludes the proof of Theorem~\ref{thm:existence}.
\section{Hamiltonian formulation}\label{sec:Ham}
In this section we present the Hamiltonian formulations of the
problems described in the previous sections. They will all be derived
from a general result proved in the first subsection.
\subsection{Legendre transform}\label{ss:Legendre-transform}
In this subsection we describe an abstract Legendre transform which
will be applied to the helium atom in the following subsections.
For concreteness we restrict to the case that the configuration
space is $\mathbb{R}^n$, but everything could be easily extended to more general
configuration manifolds.
For $k\in\mathbb{N}_0$ we abbreviate $H^k:=H^k(S^1,\mathbb{R}^n)$. We denote the
derivative of $q\in H^1$ by $\dot q$. Suppose we are given
an open subset $\mathcal{U}^1\subset H^1$ and a Lagrange function
$$
\mathcal{L}: \mathcal{U}^1\times H^0\to\mathbb{R},\qquad (q,v)\mapsto \mathcal{L}(q,v).
$$
We say that $\mathcal{L}$ possesses a {\em continuous $L^2$-gradient} if $\mathcal{L}$
is of class $C^1$ there exists a continuous map
$$
\nabla\mathcal{L} = (\nabla_1\mathcal{L},\nabla_2\mathcal{L}):\mathcal{U}^1\times H^0\to H^0\times H^0
$$
uniquely defined by the conditions
$$
\langle\nabla_i\mathcal{L}(q,v),w\rangle = D_i\mathcal{L}(q,v)w \quad \text{for all }w\in H^1,
$$
where $\langle\ ,\ \rangle$ is the $L^2$-inner product and $D_i\mathcal{L}$ denotes
the derivative with respect to the $i$-th variable.
We associate to such $\mathcal{L}$ its {\em Lagrangian action}
$$
\mathcal{S}_\mathcal{L}: \mathcal{U}^1\to\mathbb{R},\qquad q\mapsto \mathcal{L}(q,\dot q).
$$
This is a $C^1$-function whose Fr\'echet derivative at $q\in\mathcal{U}^1$ in direction $w\in H^1$ is
\begin{align*}
D\mathcal{S}_\mathcal{L}(q)w
&= D_1\mathcal{L}(q,\dot q)w + D_2\mathcal{L}(q,\dot q)\dot w
= \langle\nabla_1\mathcal{L}(q,\dot q),w\rangle + \langle\nabla_2\mathcal{L}(q,\dot q),\dot w\rangle,
\end{align*}
It follows that $q\in\mathcal{U}^1$ is a critical point of $\mathcal{S}$ if
and only if $\nabla_2\mathcal{L}(q,\dot q)\in H^1$ and the following
{\em Euler-Lagrange equation} holds:
\begin{equation}\label{eq:EL}
\frac{d}{dt}\nabla_2\mathcal{L}(q,\dot q) = \nabla_1\mathcal{L}(q,\dot q).
\end{equation}
Let us impose the following condition on $\mathcal{L}$:
\begin{description}
\item[(L)] {\em There exists a differentiable map
$$
F:\mathcal{U}^1\times H^0\to H^0,\qquad (q,p)\mapsto F(q,p)
$$
such that for each $q\in\mathcal{U}^1$ the map $v\mapsto \nabla_2\mathcal{L}(q,v)$ is
a homeomorphism with inverse $p\mapsto F(q,p)$.}
\end{description}
In particular, the map $F$ then satisfies
\begin{equation}\label{eq:F}
\nabla_2\mathcal{L}\bigl(q,F(q,p)\bigr) = p.
\end{equation}
Then we associate to $\mathcal{L}$ its {\em fibrewise Legendre transform}
$$
\mathcal{H}:\mathcal{U}^1\times H^0\to\mathbb{R},\qquad \mathcal{H}(q,p):=\langle p,F(q,p)\rangle - \mathcal{L}(q,F(q,p)).
$$
Using equation~\eqref{eq:F} we compute for $w\in H^1$:
\begin{eqnarray*}
&& D_1\mathcal{H}(q,p)w \\
&=& \langle p,D_1F(q,p)w\rangle - D_1\mathcal{L}(q,F(q,p))w -
D_2\mathcal{L}(q,F(q,p))D_1F(q,p)w \\
&=& \langle p-\nabla_2\mathcal{L}(q,F(q,p)),D_1F(q,p)w\rangle
- \langle\nabla_1\mathcal{L}(q,F(q,p)),w\rangle \\
&=& - \langle\nabla_1\mathcal{L}(q,F(q,p)),w\rangle\,, \\
&& D_2\mathcal{H}(q,p)w \\
&=& \langle F(q,p),w\rangle + \langle p,D_2F(q,p)w\rangle -
D_2\mathcal{L}(q,F(q,p))D_2F(q,p)w \\
&=& \langle F(q,p),w\rangle
+ \langle p - \nabla_2\mathcal{L}(q,F(q,p)),D_2F(q,p)w\rangle \\
&=& \langle F(q,p),w\rangle\,.
\end{eqnarray*}
This shows that $\mathcal{H}$ has a continuous $L^2$-gradient which is related
to that of $\mathcal{L}$ by
\begin{equation}\label{eq:H-L}
\begin{aligned}
\nabla_1\mathcal{H}(q,p) &= -\nabla_1\mathcal{L}(q,F(q,p)),\cr
\nabla_2\mathcal{H}(q,p) &= F(q,p).
\end{aligned}
\end{equation}
On the other hand, to any Hamilton function $\mathcal{H}:\mathcal{U}^1\times
H^0\to\mathbb{R}$ with continuous $L^2$-gradient we can associate its {\em Hamiltonian action}
$$
\mathcal{A}_\mathcal{H}:\mathcal{U}^1\times H^0\to\mathbb{R},\qquad
\mathcal{A}_\mathcal{H}(q,p) := \langle p,\dot q\rangle - \mathcal{H}(q,p).
$$
Its derivatives in direction $w\in H^1$ are given by
\begin{align*}
D_1\mathcal{A}_\mathcal{H}(q,p)w
&= \langle p,\dot w\rangle - \langle\nabla_1\mathcal{H}(q,p),w\rangle,\cr
D_2\mathcal{A}_\mathcal{H}(q,p)w
&= \langle \dot q,w\rangle - \langle\nabla_2\mathcal{H}(q,p),w\rangle.
\end{align*}
It follows that $(q,p)\in\mathcal{U}^1\times H^0$ is a critical point of $\mathcal{A}_\mathcal{H}$ if
and only if $p\in H^1$ and the following {\em Hamilton equations} hold:
\begin{equation}\label{eq:Ham}
\begin{aligned}
\dot p &= -\nabla_1\mathcal{H}(q,p),\cr
\dot q &= \nabla_2\mathcal{H}(q,p).
\end{aligned}
\end{equation}
\begin{prop}\label{prop:Legendre-transform}
Let $\mathcal{L}:\mathcal{U}^1\times H^0\to\mathbb{R}$ be a Lagrange function with continuous gradient
satisfying condition (L) and $\mathcal{H}:\mathcal{U}^1\times H^0\to\mathbb{R}$ its fibrewise
Legendre transform. Then the assignments $(q,p)\mapsto q$ and
$q\mapsto (q,p=\nabla_2\mathcal{L}(q,\dot q))$ define a one-to-one
correspondence between critical points $(q,p)$ of $\mathcal{A}_\mathcal{H}$ and
critical points $q$ of $\mathcal{S}_\mathcal{L}$.
\end{prop}
{\bf Proof: }
Let $q\in\mathcal{U}^1$ be a critical point of $\mathcal{S}_\mathcal{L}$, so
$p:=\nabla_2\mathcal{L}(q,\dot q)\in H^1$ and $q$ solves~\eqref{eq:EL}.
Then condition (L) and the second equation in~\eqref{eq:H-L} give
$$
\dot q = F(q,p) = \nabla_2\mathcal{H}(q,p),
$$
and~\eqref{eq:EL} and the first equation in~\eqref{eq:H-L} give
$$
\dot p = \frac{d}{dt}\nabla_2\mathcal{L}(q,\dot q) = \nabla_1\mathcal{L}(q,\dot q) =
-\nabla_1\mathcal{L}(q,F(q,p)) = \nabla_1\mathcal{H}(q,p).
$$
So $(q,p)$ solves~\eqref{eq:Ham} and is therefore a critical point of $\mathcal{A}_\mathcal{H}$.
Conversely, let $(q,p)\in\mathcal{U}^1\times H^0$ be a critical point of $\mathcal{A}_\mathcal{H}$,
so $p\in H^1$ and $(q,p)$ solves~\eqref{eq:Ham}.
Then the second equation in~\eqref{eq:Ham} and the second equation in~\eqref{eq:H-L} give
$$
\dot q = \nabla_2\mathcal{H}(q,p) = F(q,p),
$$
which by condition (L) implies $\nabla_2\mathcal{L}(q,\dot q)=p\in H^1$.
Now the first equation in~\eqref{eq:Ham} and the first equation
in~\eqref{eq:H-L} give
$$
\frac{d}{dt}\nabla_2\mathcal{L}(q,\dot q) = \dot p = \nabla_1\mathcal{H}(q,p)
= -\nabla_1\mathcal{L}(q,F(q,p)) = \nabla_1\mathcal{L}(q,\dot q).
$$
So $q$ solves~\eqref{eq:EL} and is therefore a critical point of $\mathcal{S}_\mathcal{L}$.
\hfill $\square$
\begin{rem}
Inspection of the preceding proof shows that formulae for the
derivatives of $\mathcal{H}$ do not involve derivatives of $F$. This suggests
that Proposition~\ref{prop:Legendre-transform} should still hold if
in condition (L) we only assume continuity of $F$ rather than
differentiability.
\end{rem}
\begin{ex}\label{ex:density}
Classically, the Lagrangian $\mathcal{L}:\mathcal{U}^1\times H^0\to\mathbb{R}$ has the form
$$
\mathcal{L}(q,v) = \int_0^1 L(q(t),v(t))dt
$$
for a smooth Lagrangian density $L:U\times\mathbb{R}^n\to\mathbb{R}$, where $U\subset\mathbb{R}^n$
is an open subset and $\mathcal{U}^1=H^1(S^1,U)$. Then
$$
\mathcal{H}(q,p) = \int_0^1 H(q(t),p(t))dt
$$
with $H:U\times\mathbb{R}^n\to\mathbb{R}$ the fibrewise Legendre transform of $L$, and
the Euler-Lagrange and Hamilton's equations take the traditional form
for $i=1,\dots,n$:
$$
\frac{d}{dt}\frac{\partial L}{\partial\dot q_i} = \frac{\partial L}{\partial q_i}
$$
and
$$
\dot p_i = -\frac{\partial H}{\partial q_i},\qquad
\dot q_i = \frac{\partial H}{\partial p_i}.
$$
\end{ex}
This example covers the instantaneous interaction Lagrangian
$\mathcal{L}_{in}$ for helium in the original coordinates $q=(q_1,q_2)\in\mathbb{R}_+^2$
away from collisions. The more general setting in
Proposition~\ref{prop:Legendre-transform} will be needed in the
following subsections to deal with the Lagrangians $\mathcal{B}_{av}$ and $\mathcal{B}_{in}$
in the new coordinate $z=(z_1,z_2)$, which do not have the form in
Example~\ref{ex:density}.
\subsection{The Kepler problem}
For the Kepler problem, the function $\mathcal{Q}$ defined
in~\eqref{eq:Q} is the Lagrangian action $\mathcal{S}_\mathcal{L}$ associated
to the Lagrange function
\begin{equation}\label{eq:L-Kepler}
\mathcal{L}: \mathcal{U}^1\times H^0(S^1,\mathbb{R})\to\mathbb{R},\qquad
\mathcal{L}(z,w) = 2\|z\|^2\|w\|^2 + \frac{N}{\|z\|^2}
\end{equation}
with $\mathcal{U}^1=H^1(S^1,\mathbb{R})\setminus\{0\}$.
The computation of the differential in Section~\ref{ss:z-to-q-Kepler}
shows that $\mathcal{L}$ has a continuous $L^2$-gradient.
The associated momentum $\eta$ is given by
$$
\eta = \nabla_2\mathcal{L}(z,w) = 4\|z\|^2w,
$$
which can be solved for $w$ as
$$
w = \frac{\eta}{4\|z\|^2} = F(z,\eta).
$$
Note that the map $F$ is smooth in $(z,\eta)$.
It follows that $\|\eta\|^2=16\|z\|^4\|w\|^2$ and the associated Hamilton
function becomes
\begin{equation}\label{eq:H-Kepler}
\mathcal{H}(z,\eta) = \langle\eta,F(z,\eta)\rangle - \mathcal{L}(z,F(z,\eta))
= \frac{\|\eta\|^2}{8\|z\|^2} - \frac{N}{\|z\|^2},
\end{equation}
with Hamiltonian action
\begin{equation}\label{eq:AH-Kepler}
\mathcal{A}_\mathcal{H}(z,\eta) = \langle\eta,z'\rangle
- \frac{\|\eta\|^2}{8\|z\|^2} + \frac{N}{\|z\|^2},
\end{equation}
By Proposition~\ref{prop:Legendre-transform}, critical points of
$\mathcal{A}_\mathcal{H}$ are in one-to-one correspondence to critical points of $\mathcal{Q}$.
\subsection{Mean interaction}
For the helium atom with mean interaction, the function $\mathcal{B}_{av}$ defined
in~\eqref{eq:Bav} is the Lagrangian action $\mathcal{S}_{\mathcal{L}_{av}}$ associated
to the Lagrange function
\begin{equation*
\mathcal{L}_{av}: \mathcal{H}^1_{av}\times H^0(S^1,\mathbb{R}^2)\to\mathbb{R},\qquad
\mathcal{L}_{av}(z,w) = \mathcal{L}(z_1,w_1) + \mathcal{L}(z_2,w_2) + \mathcal{A}(z_1,z_2)
\end{equation*}
where $\mathcal{H}^1_{av}$ and $\mathcal{A}$ are defined in~\eqref{eq:Hav}
and~\eqref{eq:A}, and $\mathcal{L}$ is the Kepler Lagrangian
from~\eqref{eq:L-Kepler} with charge $N=2$.
The computation of the differential in Section~\ref{ss:diffav}
shows that $\mathcal{L}_{av}$ has a continuous $L^2$-gradient.
Since the interaction term $\mathcal{A}$ does not depend on the $w_i$,
the associated momenta $\eta_i$ are given as in the Kepler case by
$$
\eta_i = 4\|z_i\|^2w_i,\qquad
w_i = \frac{\eta_i}{4\|z_i\|^2} = F(z_i,\eta_i)
$$
and the associated Hamilton function becomes
\begin{eqnarray*
\mathcal{H}_{av}(z,\eta)
&=& \mathcal{H}(z_1,\eta_1) + \mathcal{H}(z_2,\eta_2) - \mathcal{A}(z_1,z_2) \\
&=& \sum_{i=1}^2\bigg(\frac{\|\eta_i\|^2}{8\|z_i\|^2} -
\frac{2}{\|z_i\|^2}\bigg)
+ \frac{\|z_1\|^2\|z_2\|^2}{\|z_1^2\|^2\|z_2\|^2 - \|z_2^2\|^2\|z_1\|^2},
\end{eqnarray*}
with Hamiltonian action
\begin{eqnarray*
\mathcal{A}_{\mathcal{H}_{av}}(z,\eta)
&=& \langle\eta_1,z_1'\rangle + \langle\eta_2,z_2'\rangle - \mathcal{H}_{av}(z,\eta) \\
&=& \sum_{i=1}^2\bigg(\langle\eta_1,z_1'\rangle - \frac{\|\eta_i\|^2}{8\|z_i\|^2} +
\frac{2}{\|z_i\|^2}\bigg)
- \frac{\|z_1\|^2\|z_2\|^2}{\|z_1^2\|^2\|z_2\|^2 - \|z_2^2\|^2\|z_1\|^2}.
\end{eqnarray*}
By Proposition~\ref{prop:Legendre-transform}, critical points of
$\mathcal{A}_{\mathcal{H}_{av}}$ are in one-to-one correspondence to critical points of $\mathcal{B}_{av}$.
\subsection{Instanteneous interaction}
For the helium atom with instantaneous interaction, the function $\mathcal{B}_{in}$ defined
in~\eqref{eq:Bin} is the Lagrangian action $\mathcal{S}_{\mathcal{L}_{in}}$ associated
to the Lagrange function
\begin{equation*
\mathcal{L}_{in}: \mathcal{H}^1_{in}\times H^0(S^1,\mathbb{R}^2)\to\mathbb{R},\qquad
\mathcal{L}_{in}(z,w) = \mathcal{L}(z_1,w_1) + \mathcal{L}(z_2,w_2) + \mathcal{I}(z_1,z_2)
\end{equation*}
where $\mathcal{H}^1_{in}$ and $\mathcal{I}$ are defined in~\eqref{eq:Hin}
and~\eqref{eq:I}, and $\mathcal{L}$ is the Kepler Lagrangian
from~\eqref{eq:L-Kepler} with charge $N=2$.
The computation of the differential in Section~\ref{ss:diffin}
shows that $\mathcal{L}_{in}$ has a continuous $L^2$-gradient.
Since the interaction term $\mathcal{I}$ does not depend on the $w_i$,
the associated momenta $\eta_i$ are given as in the Kepler case by
$$
\eta_i = 4\|z_i\|^2w_i,\qquad
w_i = \frac{\eta_i}{4\|z_i\|^2} = F(z_i,\eta_i)
$$
and the associated Hamilton function becomes
\begin{eqnarray*
\mathcal{H}_{in}(z,\eta)
&=& \mathcal{H}(z_1,\eta_1) + \mathcal{H}(z_2,\eta_2) - \mathcal{I}(z_1,z_2) \\
&=& \sum_{i=1}^2\bigg(\frac{\|\eta_i\|^2}{8\|z_i\|^2} -
\frac{2}{\|z_i\|^2}\bigg)
+ \int_0^1\frac{1}{z_1^2(\tau_{z_1}(t)) - z_2^2(\tau_{z_2}(t))}dt,
\end{eqnarray*}
with Hamiltonian action
\begin{eqnarray*
\mathcal{A}_{\mathcal{H}_{in}}(z,\eta)
&=& \langle\eta_1,z_1'\rangle + \langle\eta_2,z_2'\rangle - \mathcal{H}_{in}(z,\eta) \\
&=& \sum_{i=1}^2\bigg(\langle\eta_1,z_1'\rangle - \frac{\|\eta_i\|^2}{8\|z_i\|^2} +
\frac{2}{\|z_i\|^2}\bigg)
- \int_0^1\frac{1}{z_1^2(\tau_{z_1}(t)) - z_2^2(\tau_{z_2}(t))}dt\,.
\end{eqnarray*}
By Proposition~\ref{prop:Legendre-transform}, critical points of
$\mathcal{A}_{\mathcal{H}_{in}}$ are in one-to-one correspondence to critical points of
$\mathcal{B}_{in}$.
|
\section*{Results}
We strive to characterize the patterns by which post-mortem memory evolves during the year immediately following the death of public figures.
When considering this time frame, prior work has primarily taken a qualitative stance, asking \textit{how,} linguistically, the mainstream and social media speak about small sets of deceased people~\cite{starck_death_2008,hanusch_valuing_2008,gach_control_2017,stone_trauma_2002,sanderson_tweeting_2010,goh_analysis_2011,radford_grief_2012}.
In contrast, enabled by a comprehensive corpus of news and social media posts, we take a quantitative stance, asking \textit{about whom} the media speak \textit{how much} after death.
At the core of our analysis are time series of mention frequency.
A person $i$'s \textit{raw mention time series} specifies, for each day $t$ relative to $i$'s day of death ($t=0$), the base-10 logarithm of the fraction
$S_i(t)$
of documents in which person $i$ was mentioned, out of all documents published on day~$t$.
To reduce noise, we also generated \textit{smoothed mention time series} using a variable span smoother based on local linear fits \cite{friedman_variable_1984}.
For each person, separate time series were computed for the news and for Twitter (examples in \Figref{fig:gallery}; additional details about mention time series construction in \textit{Materials and Methods}.)
\begin{figure
\centering
\subfigure[News]{
\label{fig:gallery_NEWS}
\includegraphics[width=0.25\linewidth]{mention_curve_Benoit_Mandelbrot_NEWS.pdf}
\hspace{-4mm}
\includegraphics[width=0.25\linewidth]{mention_curve_Kashiram_Rana_NEWS.pdf}
\hspace{-4mm}
\includegraphics[width=0.25\linewidth]{mention_curve_George_McGovern_NEWS.pdf}
\hspace{-4mm}
\includegraphics[width=0.25\linewidth]{mention_curve_Amy_Winehouse_NEWS.pdf}
}
\subfigure[Twitter]{
\label{fig:gallery_TWITTER}
\includegraphics[width=0.25\linewidth]{mention_curve_Benoit_Mandelbrot_TWITTER.pdf}
\hspace{-4mm}
\includegraphics[width=0.25\linewidth]{mention_curve_Kashiram_Rana_TWITTER.pdf}
\hspace{-4mm}
\includegraphics[width=0.25\linewidth]{mention_curve_George_McGovern_TWITTER.pdf}
\hspace{-4mm}
\includegraphics[width=0.25\linewidth]{mention_curve_Amy_Winehouse_TWITTER.pdf}
}
\caption{
\textbf{Examples of mention time series for four deceased public figures,}
as observed \textbf{(a)}~in the news and \textbf{(b)}~on Twitter.
In mention time series, the $x$-axis specifies the number of days since death, and the $y$-axis, the base-10 logarithm of the fraction of documents in which the person was mentioned that day, out of all documents published that day.
Light circles correspond to raw mention time series, dark curves, to their smoothed versions.
}
\label{fig:gallery}
\end{figure}
\subsubsection*{The shape of post-mortem memory: communicative and cultural memory}
Averaging the 2\,362\xspace raw mention time series (\Figref{fig:avg_mention_curve}) exposes a sharp spike in the interest in public figures in the immediate wake of their death (on days 0 and 1),
followed by a steep drop up until around day 30,
where the curves elbow into a long, much flatter phase, which is only slightly disrupted by a small secondary spike on day 365 after death.
The main spike is so strong that, without logarithmically transforming the fractions $S_i(t)$ of mentioning documents, no interesting information besides the dominant main spike would be visually discernible.
\begin{figure*
\centering
\subfigure[News]{
\label{fig:avg_mention_curve_NEWS}
\includegraphics[width=0.47\textwidth]{avg_mention_curve_NEWS.pdf}
}
\subfigure[Twitter]{
\label{fig:avg_mention_curve_TWITER}
\includegraphics[width=0.47\textwidth]{avg_mention_curve_TWITTER.pdf}
}
\caption{
\textbf{Average mention time series,}
obtained via the arithmetic mean of the individual raw mention time series of the 2\,362\xspace people included in the study,
\textbf{(a)}~in the news and
\textbf{(b)}~on Twitter
(see \Figref{fig:gallery} for examples of individual mention time series).
On average, the mention frequency of deceased public figures spikes by about 9\,400\% in the news, and by about 28\,000\% on Twitter, when they die, and fades quickly thereafter, with a minor secondary spike on the death anniversary.
\trackedChange{
We also plot the best fit of the shifted power law model (\textit{black}),
which decomposes the total collective memory $S(t)=u(t)+v(t)$ on day $t$
into a sum of
communicative memory $u(t)=at^{-b}$ (\textit{green})
and cultural memory $v(t)=c$ (\textit{red}).
}
Insets show the same data and fits on logarithmic $x$-axes.
}
\label{fig:avg_mention_curve}
\end{figure*}
\trackedChangeNew{
In a model that is conceptually similar to Candia \abbrevStyle{et al.}\xspace's biexponential model \cite{candia_universal_2019}, we decompose the post-mortem collective memory $S(t)$ into a sum of two components, $S(t)=u(t)+v(t)$,
where $u(t)$ captures \textit{communicative memory,} and $v(t)$, \textit{cultural memory} (see introduction).
Communicative memory is modeled as a decaying power law, \abbrevStyle{i.e.}\xspace, $u(t) = a t^{-b}$,%
\footnote{With $u(t)=a t^{-b}$, we have $\d u(t) / \d t = -(b/t)\ u(t)$, so the forgetting rate $b/t$ (fraction of memory lost at time $t$) decreases over time, in line with what is suggested by the psychological literature~\cite{wixted_form_1991}.}
whereas cultural memory is modeled as constant during the time frame considered here (400 days after death), \abbrevStyle{i.e.}\xspace, $v(t)=c$.
We refer to this model as a \textit{shifted power law}.
It fits the empirical average mention time series ($R^2=0.99$; details on model fitting in \textit{Materials and Methods}) significantly better than any of eight alternative models from the literature~\cite{candia_universal_2019,rubin_one_1996,wixted_form_1991} (details in \SupFig{2}{}), including the biexponential model \cite{candia_universal_2019}.
The best shifted power law fit is shown as a black line in \Figref{fig:avg_mention_curve}; the communicative and cultural memory components are plotted separately in green and red, respectively.
The fitted decay parameter of communicative memory is similar
for the news ($b=1.34$) and
for Twitter ($b=1.54$).
Communicative memory starts high on day $t=1$, but drops below cultural memory quickly, after 14 and 18 days in the news and on Twitter, respectively,
and accounts for only 25\%
of total collective memory after 31 and 36 days, respectively, which constitutes an inflection point where communicative memory levels off strongly (\SupFig{3}).
Moreover, even though no pre-mortem data was used in fitting the model, the constant cultural memory $c$ closely approximates the average pre-mortem fraction of mentioning documents in both media (cf.\ \Figref{fig:avg_mention_curve}).
This suggests that, on average, public figures build up a certain baseline amount of (cultural) memory during their lifetime, on top of which a burst of quickly fading communicative memory is layered in the wake of death.
Note that, although collective memory rapidly reverts to the pre-mortem level when averaging over all people, it need not be so for individual people, as we shall see below (\Figref{fig:megafigure}(b)).}
Further evidence in support of two distinct memory modes comes from the fact that the average length of documents that mention a public figure dropped sharply with death (possibly due to brief death notes and obituaries) and reached the pre-mortem level again after about 30 days (\SupFig{4}{}), \abbrevStyle{i.e.}\xspace, around the inflection point where communicative memory levels off according to the fitted model.
We hence divide the post-mortem period into two phases: short-term (days 0 through 29) and long-term (days 30 through~360).
Based on this distinction, in order to reason about the shape of mention time series, we summarize each time series by four characteristic numbers (depicted graphically in \Figref{fig:megafigure}(a)):
\begin{figure*
\centering
\includegraphics[width=\linewidth]{megafigure_INTEGER-COUNTS_CROPPED.pdf}
(a) \hspace{85mm} (b)
\caption{
\textbf{Cluster analysis of mention time series.}
\textbf{(a)}~From each mention time series, we extract four characteristic numbers:
pre-mortem mean, short-term boost, long-term boost, and halving time.
In the resulting four\hyp dimensional space, time series are clustered using the $k$-means algorithm.
According to the average silhouette criterion, the optimal number of clusters is $k=4$ both in the news and on Twitter.
\textbf{(b)}~Nearly identical clusters (C1 through C4) emerge independently in the news and on Twitter, in terms of both cluster centroids and cluster sizes.
Cluster centroids are depicted as bar charts in the right (news) and bottom (Twitter) margins; average mention time series for each cluster, in the left (news) and top (Twitter) margins.
As captured by the confusion matrix, whose diagonal entries are much larger than under a null model that assumes the two media to be independent, a given person tends to fall into the corresponding clusters in the two media.
}
\label{fig:megafigure}
\end{figure*}
\begin{sidewaystable}
\begin{center}
\caption{
Linear regression models of short- and long-term boosts in news and Twitter. Standard errors of coefficients in parentheses.
}
\label{table:coefficients}
\begin{tabular}{l D{)}{)}{11)3} D{)}{)}{11)3} D{)}{)}{11)3} D{)}{)}{11)3} }
\toprule
& \multicolumn{1}{c}{\shortstack{Short-term boost\\(News)}} & \multicolumn{1}{c}{\shortstack{Short-term boost\\(Twitter)}} & \multicolumn{1}{c}{\shortstack{Long-term boost\\(News)}} & \multicolumn{1}{c}{\shortstack{Long-term boost\\(Twitter)}} \\
\midrule
(Intercept) & 2.322 \; (.063)^{***} & 2.670 \; (.067)^{***} & .088 \; (.014)^{***} & .095 \; (.015)^{***} \\
Pre-mortem mean (relative rank) & .804 \; (.093)^{***} & .948 \; (.100)^{***} & .031 \; (.020) & .086 \; (.022)^{***} \\
Manner of death: unnatural & .618 \; (.095)^{***} & .282 \; (.100)^{**} & .097 \; (.021)^{***} & .106 \; (.022)^{***} \\
Language: non-anglophone & -.316 \; (.074)^{***} & -.116 \; (.078) & -.061 \; (.016)^{***} & -.037 \; (.017)^{*} \\
Language: unknown & -.446 \; (.086)^{***} & -.325 \; (.091)^{***} & -.079 \; (.019)^{***} & -.081 \; (.020)^{***} \\
Gender: female & .083 \; (.072) & -.034 \; (.076) & .034 \; (.016)^{*} & .006 \; (.017) \\
Notability type: academia/engineering & .181 \; (.197) & .340 \; (.208) & -.032 \; (.043) & .023 \; (.046) \\
Notability type: general fame & .070 \; (.124) & .132 \; (.131) & -.010 \; (.027) & -.008 \; (.029) \\
Notability type: known for death & -.107 \; (.099) & -.088 \; (.106) & -.021 \; (.022) & .008 \; (.023) \\
Notability type: leadership & .152 \; (.083) & .113 \; (.087) & -.058 \; (.018)^{**} & -.040 \; (.019)^{*} \\
Notability type: sports & .049 \; (.083) & .072 \; (.088) & -.034 \; (.018) & -.034 \; (.020) \\
Age: 20--29 & .162 \; (.170) & .718 \; (.180)^{***} & .060 \; (.037) & .192 \; (.040)^{***} \\
Age: 30--39 & .400 \; (.167)^{*} & .649 \; (.177)^{***} & .028 \; (.037) & .118 \; (.039)^{**} \\
Age: 40--49 & -.046 \; (.126) & .351 \; (.133)^{**} & -.001 \; (.028) & .100 \; (.030)^{***} \\
Age: 50--59 & -.075 \; (.099) & .181 \; (.104) & -.058 \; (.022)^{**} & -.024 \; (.023) \\
Age: 60--69 & -.109 \; (.082) & .130 \; (.086) & -.050 \; (.018)^{**} & -.025 \; (.019) \\
Age: 80--89 & .022 \; (.078) & .021 \; (.082) & -.018 \; (.017) & -.013 \; (.018) \\
Age: 90--99 & .174 \; (.098) & .034 \; (.103) & -.011 \; (.021) & -.024 \; (.023) \\
\midrule
R$^2$ & .213 & .192 & .123 & .178 \\
Adj. R$^2$ & .197 & .176 & .106 & .161 \\
Num. obs. & 870 & 870 & 870 & 870 \\
RMSE & .772 & .815 & .169 & .181 \\
\bottomrule
\multicolumn{5}{l}{\scriptsize{$^{***}p<0.001$, $^{**}p<0.01$, $^*p<0.05$}}
\end{tabular}
\end{center}
\end{sidewaystable}
\begin{table}
\begin{center}
\caption{
Linear regression models of news-minus-Twitter difference in short- and long-term boosts.
Standard errors of coefficients in parentheses.
}
\label{table:coefficients News vs. Twitter}
\begin{tabular}{l D{)}{)}{11)3} D{)}{)}{11)3} }
\toprule
& \multicolumn{1}{c}{Short-term boost} & \multicolumn{1}{c}{Long-term boost} \\
\midrule
(Intercept) & -.427 \; (.047)^{***} & -.015 \; (.014) \\
Pre-mortem mean (relative-rank diff.) & -.212 \; (.083)^{*} & -.034 \; (.025) \\
Manner of death: unnatural & .348 \; (.070)^{***} & -.008 \; (.021) \\
Language: non-anglophone & -.219 \; (.054)^{***} & -.023 \; (.017) \\
Language: unknown & -.052 \; (.063) & .012 \; (.019) \\
Gender: female & .091 \; (.053) & .029 \; (.016) \\
Notability type: academia/engineering & -.048 \; (.146) & -.046 \; (.044) \\
Notability type: general fame & -.016 \; (.092) & .002 \; (.028) \\
Notability type: known for death & .105 \; (.074) & -.015 \; (.022) \\
Notability type: leadership & .200 \; (.062)^{**} & -.006 \; (.019) \\
Notability type: sports & .059 \; (.062) & .009 \; (.019) \\
Age: 20--29 & -.577 \; (.126)^{***} & -.135 \; (.038)^{***} \\
Age: 30--39 & -.235 \; (.124) & -.089 \; (.038)^{*} \\
Age: 40--49 & -.374 \; (.093)^{***} & -.101 \; (.028)^{***} \\
Age: 50--59 & -.204 \; (.073)^{**} & -.029 \; (.022) \\
Age: 60--69 & -.175 \; (.061)^{**} & -.021 \; (.018) \\
Age: 80--89 & .014 \; (.058) & -.006 \; (.018) \\
Age: 90--99 & .164 \; (.072)^{*} & .015 \; (.022) \\
\midrule
R$^2$ & .101 & .052 \\
Adj. R$^2$ & .083 & .034 \\
Num. obs. & 870 & 870 \\
RMSE & .571 & .174 \\
\bottomrule
\multicolumn{3}{l}{\scriptsize{$^{***}p<0.001$, $^{**}p<0.01$, $^*p<0.05$}}
\end{tabular}
\end{center}
\end{table}
\begin{enumerate}
\item \textbf{Pre-mortem mean:} arithmetic mean of days 360 through 30 before death.
\item \textbf{Short-term boost:} maximum of days 0 through 29 after death, minus the pre-mortem mean.
\item \textbf{Long-term boost:} arithmetic mean of days 30 through 360 after death, minus the pre-mortem mean.
\item \textbf{Halving time:} number of days required to accumulate half of the total area between the post-mortem curve (including the day of death) and the minimum post-mortem value.
\end{enumerate}
All characteristics were computed on the smoothed time series, with the exception of the maximum used in the short-term boost, which was computed on the raw time series.
The 29 days immediately before death were excluded from the pre-mortem mean in an effort to exclude a potential rise in interest in people whose impending death might have been anticipated, \abbrevStyle{e.g.}\xspace, due to illness.
Since the time series capture logarithmic mention frequencies, the (arithmetic) pre-mortem mean corresponds to the logarithm of the geometric mean mention frequency, and the short- and long-term boosts, to the logarithm of the multiplicative increase over the pre-mortem geometric mean mention frequency.
\subsubsection*{Magnitude of short- and long-term boosts}
Aggregating the short-term boost over all public figures allows us to quantify the sharp spike immediately after death observed in \Figref{fig:avg_mention_curve}.
The median short-term boost was 1.98 (95\% CI $[1.90,2.03]$) in the news,
and 2.45 (95\% CI $[2.37,2.50]$) on Twitter.
(All curve characteristics are summarized in \SupTab{2}{} and \SupFig{5}.)
The boost was significantly stronger on Twitter
(Wilcoxon's signed-rank test: $W=477\,893$, two-sided $p < 10^{-15}$),
where it approximately corresponded to a 28\,000\% increase on the linear scale
($10^{2.45} \approx 281$),
compared to a 9\,400\% increase in the news
($10^{1.98} \approx 95$).
After the immediate spike, media interest tended to fade quickly.
In the news, no important long-term boost was observed (median 0.000545, 95\% CI $[-0.000908, 0.00171]$),
whereas on Twitter, we measured a significantly larger
(Wilcoxon's signed-rank test: $W=881\,590$, two-sided $p < 10^{-15}$)
long-term boost of 0.0160 in the median (95\% CI $[0.0133, 0.0175]$),
translating to a 3.8\% increase on the linear scale
($10^{0.016} \approx 1.038$).
\subsubsection*{Cluster analysis of mention time series}
Mention time series expose a great variety of curve shapes, a glimpse of which is given by the examples of \Figref{fig:gallery}.
We hypothesized that, despite their diversity, mention time series could be grouped into distinct classes, a hypothesis that we explored in a cluster analysis.
Time series were represented by their four characteristic numbers
(pre-mortem mean,
short-term boost,
long-term boost,
halving time)
in $z$-score\hyp standardized form and clustered using the $k$-means algorithm.
A separate clustering was performed for the news and for Twitter.
Evaluating all numbers of clusters $k \in \{2, \dots, 30\}$ via the average silhouette criterion \cite{rousseeuw_silhouettes_1987} revealed a clear optimum for $k=4$ clusters for both the news and Twitter (\SupFig{7}).
The cluster centroids are visualized in the right and bottom margins of \Figref{fig:megafigure}(b);
the right margin shows the centroids for the news,
the bottom margin, for Twitter.
Moreover, we plot the average smoothed mention time series for each cluster in the left (news) and top (Twitter) margins.
(An overlay of all time series per cluster is plotted in \SupFig{8}.)
Strikingly, although the clustering was performed independently for the news and for Twitter, respectively, the centroids that emerged---as well as the number of data points in each cluster---are nearly identical.
The resulting clusters, which we name C1 through C4 in order of decreasing size, can be described as follows:
\begin{enumerate}
\item[\textbf{C1}] \textbf{(``blip''):} Average mention frequency pre- as well as post-mortem, with a short-term boost of average magnitude in between (62\% of people in the news; 59\% on Twitter).
\item[\textbf{C2}] \textbf{(``silence''):} Average mention frequency pre- as well as post-mortem, with a faint short-term boost of below-average magnitude in between (28\% in the news; 26\% on Twitter).
\item[\textbf{C3}] \textbf{(``rise''):} High pre-mortem mention frequency, large short-term boost, followed by an extreme long-term boost (7\% in the news; 11\% on Twitter).
\item[\textbf{C4}] \textbf{(``decline''):} Extreme pre-mortem mention frequency, above-average short-term boost, followed by a below-average long-term boost (3\% in the news; 4\% on Twitter).
\end{enumerate}
In both media, over half of the people (59--62\%) fall into cluster C1; their time series resemble the overall average (see \Figref{fig:avg_mention_curve}), with a brief spike after death and a quick drop to the---usually low---pre-mortem level.
About half of the remaining people (26--28\%) fall into cluster C2; their time series are similar to those of C1, with the difference that the death of people in C2 went largely unnoticed.
About half of the people outside of C1 and C2 (7--11\%) fall into C3, which mostly contains people who were popular already before death and experienced a large boost in attention in both the short and the long term.
The final cluster, C4, is composed of a tiny elite (3--4\%) of people of an extreme pre-mortem popularity that tended to fade post-mortem. The long-term decrease was considerably stronger in the news than on Twitter in this cluster.
Not only do nearly identical clusters of nearly identical size emerge in the news as on Twitter;
a given person also tends to fall into the corresponding clusters in the two media,
as captured by the cluster confusion matrix (\Figref{fig:megafigure}(b)), which counts, for all $i,j \in \{1,2,3,4\}$, the number of people falling into news cluster $i$ and Twitter cluster $j$.
Using Pearson's $\chi^2$ test, we reject the null hypothesis under which cluster membership is assumed to be independent in the news \abbrevStyle{vs.}\xspace\ Twitter, given the empirical marginal cluster sizes
($\chi^2 = 1\,739, p < 10^{-5}$).
In particular, all diagonal entries of the confusion matrix are strongly over\hyp represented,%
\footnote{
The empirical trace (the sum of diagonal elements) is 1\,727.
Under the null hypothesis, the expected trace would be 1\,059.
Given the marginal constraints, the minimum and maximum traces that could possibly be attained are 505 and 2\,236, respectively (determined via linear programming).
That is, the empirical trace attains 71\% of the range between the minimum and the maximum, whereas the null model attains only 32\%.
}
whereas all but two off-diagonal entries are under\hyp represented, and ``C3 in news, C4 on Twitter'' is the only off-diagonal entry to be significantly over\hyp represented
(one-sample proportions test with continuity correction: $\chi^2 = 135, p < 10^{-15}$).
\subsubsection*{Biographic correlates of post-mortem memory}
Next, we aim to understand what pre-mortem properties of a person are associated with their post-mortem mention frequency.
A na\"ive correlational analysis would not suffice for this purpose, as personal properties are correlated with one another; \abbrevStyle{e.g.}\xspace, leaders (politicians, CEOs, etc.)\ in the dataset are more likely to have died old and of a natural death, and are more likely to be men, compared to artists.
In order to disentangle such correlations, we performed a regression analysis.
We fitted linear regression models for two outcomes:
\begin{enumerate}
\itemsep -2pt\topsep-10pt\partopsep-10pt
\item short-term boost,
\item long-term boost,
\end{enumerate}
and with six predictors in either case:
\begin{enumerate}
\itemsep -2pt\topsep-10pt\partopsep-10pt
\item pre-mortem mean mention frequency,
\item age at death (factor with eight levels: 20--29, 30--39, \dots, 90--99),
\item manner of death (factor with two levels: natural, unnatural),
\item notability type (factor with six levels, specifying a profession or role for which the person was most known:
arts,
sports,
\trackedChange{leadership [including politicians, business\slash organization leaders, religious leaders, military, etc.],}
\trackedChange{known for death [including disaster victims],}
general fame,
academia\slash engineering),
\item \trackedChange{language} (factor with three levels: anglophone, non-anglophone, unknown),
\item gender (factor with two levels: female, male).
\end{enumerate}
\begin{figure*
\centering
\includegraphics[width=0.32\linewidth]{age_coefs_NEWS_peak_mean_boost.pdf}
\includegraphics[width=0.32\linewidth]{age_coefs_TWITTER_peak_mean_boost.pdf}
\includegraphics[width=0.32\linewidth]{age_coefs_NEWS-VS_TWITTER_peak_mean_boost_diff.pdf}
\subfigure[News]{
\label{fig:age_coefs_NEWS}
\includegraphics[width=0.32\linewidth]{age_coefs_NEWS_perm_boost.pdf}
}
\hspace{-3.3mm}
\subfigure[Twitter]{
\label{fig:age_coefs_TWITTER}
\includegraphics[width=0.32\linewidth]{age_coefs_TWITTER_perm_boost.pdf}
}
\hspace{-3.3mm}
\subfigure[News \abbrevStyle{vs.}\xspace\ Twitter]{
\label{fig:age_coefs_NEWS_VS_TWITTER}
\includegraphics[width=0.32\linewidth]{age_coefs_NEWS-VS_TWITTER_perm_boost_diff.pdf}
}
\caption{
\textbf{Dependence of post-mortem mention frequency on age at death}
for
\textbf{(a)}~the news and
\textbf{(b)}~Twitter,
in terms of short-term (\textit{top}) and long-term (\textit{bottom}) boost,
\trackedChange{defined as the base-10 logarithm of the post-to-pre-mortem ratio of fractions of mentioning documents.}
Results were obtained via linear regression models that controlled for pre-mortem mean, notability type, \trackedChange{language}, and gender.
\textbf{(c)}~Per-person news-minus-Twitter difference in short-term (\textit{top}) and long-term (\textit{bottom}) boosts.
Error bars capture 95\% confidence intervals (approximated as $\pm$2 standard errors).
These plots show
that unnatural deaths lead to larger attention boosts across age brackets, both short- and long-term;
that the news increases attention most for those who die very young or very old, whereas Twitter increases attention more the younger the deceased person;
and that the difference between attention boosts in news \abbrevStyle{vs.}\xspace\ Twitter is larger for those who die older (short- and long-term) and for those who die an unnatural death, across age brackets (short-term).
}
\label{fig:age_coefs}
\end{figure*}
Out of all 2\,362\xspace people, the regression analysis only included the 870 people for whom all factors took one of the above\hyp defined levels (details on definition of biographic features in \textit{Materials and Methods;} distribution summarized in in \SupTab{1}{}).
All factor variables were dummy-coded as binary indicators.
The pre-mortem mean was first rank-transformed and then linearly scaled and shifted to the interval $[-0.5,0.5]$.
Additionally, 70--79 years (which contains the mean and median age at death) was chosen as the default age level,
and the most frequent level was chosen as the default for all other factors, such that the regression intercept captures the average outcome for a ``baseline persona'' representing
male anglophone artists of median pre-mortem popularity who died a natural death at age 70--79.
\trackedChange{With the above, a coefficient $\beta$ for a binary predictor corresponds to an additive boost increase of $\beta$ with respect to the baseline persona,
or---since boosts are base-10 logarithms (of post-to-pre-mortem ratios of mention frequencies)---to a multiplicative post-to-pre-mortem ratio increase of $10^\beta$.}
A separate regression model was fitted for each combination of medium (news or Twitter) and outcome (short- or long-term boost), for a total of four models.
The model coefficients (summarized in \Tabref{table:coefficients}) paint a largely consistent picture for the news \abbrevStyle{vs.}\xspace\ Twitter.
We observe that, in both media, \textit{ceteris paribus,}
both the short- and the long-term boost was larger
for people who died an unnatural death,
for people with an anglophone background,
and for people who were already popular pre-mortem.
No significant gender variation was detected,
with the exception of the long-term boost in the news, which was slightly larger for women.
The only significant notability type was leadership,
\trackedChange{whose long-term boost was smaller than that of the baseline (arts) in both media.}
The dependence of short- and long-term boosts on the age at death is displayed visually in \Figref{fig:age_coefs_NEWS} and \ref{fig:age_coefs_TWITTER}.
In order to determine whether the above finding that attention increased more for people who died an unnatural death holds across age brackets, the plots are based on a slightly modified model with an additional ``age by manner of death'' interaction term.
This allows us to estimate the average post-mortem attention boost attained by each age bracket separately for people who died natural \abbrevStyle{vs.}\xspace\ unnatural deaths (as before, the estimates are for male anglophone artists of median pre-mortem popularity).
Inspecting the curves of \Figref{fig:age_coefs_NEWS} and \ref{fig:age_coefs_TWITTER}, we make two observations.
First, across age brackets, people who died an unnatural death received larger boosts, both short- and long-term, and both in the news and on Twitter.
Second, the curves have a non\hyp monotonic U-shape for the news (\Figref{fig:age_coefs_NEWS}), but are monotonically decreasing for Twitter (\Figref{fig:age_coefs_TWITTER});
\abbrevStyle{i.e.}\xspace, the news increased attention most for those who died either very young or very old, whereas Twitter increased attention more the younger the deceased.
\subsubsection*{News \abbrevStyle{vs.}\xspace\ Twitter}
The above analyses were done separately for the news and Twitter.
In order to understand how post-mortem memory of the same person differed between the two media, we conducted a pairwise analysis.
We again fitted linear regression models with the same predictors as above, but this time with outcomes defined by the news-minus-Twitter difference in short- and long-term boosts.
Accordingly, the rank-transformed and scaled pre-mortem mean predictor was replaced with the news-minus-Twitter difference in rank-transformed and scaled pre-mortem means.
Given this setup, large positive coefficients mark groups of people who received particularly strong boosts in the news compared to Twitter, and large negative coefficients mark groups of people who received particularly strong boosts on Twitter compared to the news.
The model coefficients (summarized in \Tabref{table:coefficients News vs. Twitter}) reveal that those who died an unnatural death, as well as leaders, received
particularly large short-term attention boosts in the news compared to Twitter.
Conversely, pre-mortem popular people and those with a non\hyp anglophone background received
particularly large short-term attention boosts on Twitter compared to the news,
the latter possibly because English is the most globally connected language~\cite{ronen_links_2014}, such that Twitter posts, even though all written in English, stemmed from a more geographically and culturally diverse set of writers than news articles, which originated exclusively from anglophone outlets by design.
Other than leaders, no further notability type was significantly associated with either outcome, and no significant gender variation was observed.
Finally, the age dependence is visualized in \Figref{fig:age_coefs_NEWS_VS_TWITTER}, which shows that, the older a person, the larger the news-minus-Twitter difference in boosts, confirming that news media favored older people more than Twitter did, both short- and long-term.
\section*{Discussion}
Our analysis of mention frequencies over time revealed that, for the majority of public figures, a sharp pulse of media attention immediately followed death, whereby mention frequency increased by 9\,400\% in the news, and by 28\,000\% on Twitter, in the median.
The average mention frequency then declined sharply, with an inflection point around one month after death, from where on it decayed more slowly, eventually converging toward the pre-mortem level.
These two stages are consistent with a model that posits two components of collective memory:
\trackedChangeNew{
a constant baseline level of cultural memory built up during life,
and an added layer of communicative memory that is sparked by death and usually decays in a matter of days according to a power law.}
A cluster analysis of the mention time series revealed a set of four prototypical memory patterns (``blip'', ``silence'', ``rise'', and ``decline'').
The same set of patterns emerged independently in the news and on Twitter, and the same person tended to fall into the same cluster across the two media.
In our regression analysis of biographic correlates of post-mortem memory, out of all notability types (arts, sports, leadership, known for death, general fame, academia\slash engineering), only leadership (politicians, business leaders, \abbrevStyle{etc.}\xspace)\ stood out significantly, being associated with a particularly low boost in long-term memory (\Tabref{table:coefficients}).
One might wonder if this fact could simply be explained by a regression to the mean, since leaders had the highest pre-mortem mention frequencies in the news (\SupFig{6}(a)). Note, however, that on Twitter, too, leaders saw the lowest long-term boosts,
despite the fact that, on Twitter, it is artists---not leaders---who had the highest pre-mortem mention frequencies (\SupFig{6}(b)). We thus consider a regression to the mean an unlikely explanation for the low long-term boosts of leaders, as it could not simultaneously explain the situation in both media.
Rather, we speculate that the lives and legacies of people of different notability types might differ systematically.
Considering that nearly all (eight out of 10) long-term boost coefficients for notability types are negative in \Tabref{table:coefficients} (and none are significantly positive), the distinction to be made is in fact not that between leadership and the rest, but rather that between arts (the default notability type) and the rest.
Based on this observation, we hypothesize that artists remain more present in the collective memory because they not only tend to be active performers during their lifetime, but also frequently leave a legacy of artwork that can long survive them,
whereas leaders, athletes, \abbrevStyle{etc.}\xspace, are noteworthy primarily for the actions they take during their lifetime, and are of much decreased interest to the media once they cannot take action anymore---an effect that seems to be most pronounced for leaders.%
\footnote{In terms of interesting exceptions to this rule, certain leaders saw large long-term boosts in the news (\abbrevStyle{e.g.}\xspace, Zimbabwean military Solomon Mujuru, who died in a fire in circumstances considered suspicious by some), on Twitter (\abbrevStyle{e.g.}\xspace, World War II veterans Richard Winters and Maurice Brown), or in both media (\abbrevStyle{e.g.}\xspace, entrepreneur and activist Aaron Swartz, suicide at age~26; Governor of Punjab Salman Taseer, assassinated at age~66).}
The low coefficients of determination (adjusted $R^2$) of the linear regression models, ranging from
0.106 to 0.197
(\Tabref{table:coefficients}),
serve as a testimony of the richness of human lives and legacies, which cannot be captured by statistical models relying on just a few biographic variables.
Given the inherent unpredictability of social systems~\cite{martin_exploring_2016,salganik_experimental_2006}, this would be unlikely to change even if more biographic variables and more data points became available, and if more complex statistical models were to be used.
We emphasize, however, that despite the inherent limits of predictability all model fits were highly significant
($p<10^{-15}$ for the $F$-statistics of all models of \Tabref{table:coefficients}, cf.\ \textit{Regression modeling} in Supplementary Information).
Also, and most importantly, the effects were not only significant, but also large.
\trackedChange{
As mentioned, a coefficient $\beta$ for a binary predictor corresponds to a multiplicative increase of $10^\beta$ in the post-to-pre-mortem ratio of mention frequencies, compared to the baseline persona, a male anglophone artist of median pre-mortem popularity who died a natural death at age 70--79.
For example, \textit{ceteris paribus,} an unnatural death quadrupled ($10^{0.618} \approx 4.15$) the short-term post-to-pre-mortem mention\hyp frequency ratio in the news, and nearly doubled it ($10^{0.282} \approx 1.92$) on Twitter.
The effect of age at death was also large.
For instance, on Twitter, \textit{ceteris paribus,} the short-term post-to-pre-mortem mention\hyp frequency ratio for the 30--39 age bracket was twice that of the neighboring, 40--49 age bracket ($10^{0.649-0.351} \approx 1.99$);
and that of the youngest age bracket was nearly five times that of the oldest age bracket ($10^{0.718-0.034} \approx 4.83$).
}
One of the key contributions of this study is the comparison between mainstream news and Twitter---a prominent social media platform---on a fixed set of attention subjects, thus extending a rich literature on the interplay between the two media \cite{grinberg_fake_2019,bovet_influence_2019,zhao_comparing_2011,kwak_what_2010}.
Despite the striking similarity of prototypical mention time series emerging from the cluster analysis (\Figref{fig:megafigure}), the regression analysis revealed several noteworthy differences between post-mortem memory in the news \abbrevStyle{vs.}\xspace\ Twitter.
First, whereas on Twitter the post-mortem boost was monotonically and negatively associated with age at death (\Figref{fig:age_coefs_TWITTER}),
we observed a non\hyp monotonic U-shaped relationship in the news (\Figref{fig:age_coefs_NEWS}), which provided the largest post-mortem boost to both those who died very young and to those who died very old, an effect that even held for a fixed person (\Figref{fig:age_coefs_NEWS_VS_TWITTER}).
Second, the increased short-term boost associated with unnatural deaths was even more pronounced in the news than on Twitter (\Tabref{table:coefficients News vs. Twitter}), across age groups (\Figref{fig:age_coefs_NEWS_VS_TWITTER}).
And third, leaders were boosted more by the news than by Twitter, both short- and long-term (\Tabref{table:coefficients News vs. Twitter}).
Taken together, these findings could be interpreted as the result of two simultaneous roles played by mainstream news media:
on the one hand, as heralds catering to the public curiosity stirred by a young or unnatural death;
on the other hand, as stewards of collective memory when an old person or an accomplished leader dies after a life of achievement.
On the contrary, the extent to which Twitter plays both roles is weaker:
on the one hand, unnatural deaths were followed by a less pronounced short-term boost on Twitter than in the news;
on the other hand, Twitter users paid less attention when an old public figure or a leader died.
The present study showed that even the simple counting of mentions yields nuanced insights into \textit{who} is remembered after death.
Future studies may go further by also asking \textit{how} deceased public figures are remembered, by studying how the language, tone, and attitude toward them change in the wake of death.
By considering a diverse set of thousands of public figures such as ours, future work will be able to quantify, \abbrevStyle{e.g.}\xspace, to what extent news and social media abide by the old Latin adage \textit{``De mortuis nihil nisi bonum''} (``Of the dead, speak no evil'').
\trackedChange{The analysis could be further enriched by going beyond the coarse biographic categories considered here and leveraging manually curated repositories of more fine-grained information about public figures~\cite{yu_pantheon_2016}.}
\trackedChange{
We also emphasize that media attention cannot capture all aspects of collective memory, so we encourage researchers to apply our methodology to further measures of popularity, in particular those capturing the consumption, rather than production, of content, including songs, movies, books, Wikipedia articles, \abbrevStyle{etc.}\xspace
}
Finally, this study started from an elite of people considered worthy of being included in the Freebase knowledge base (which roughly equals the set of people with a Wikipedia article).
This notability bias was further increased by discarding people whose pre-mortem mention frequency was too low in the news or on Twitter (see \textit{Materials and Methods}), a restriction necessary in order to compare the coverage of a fixed person across the two media.
Since the bar for being mentioned in the news~\cite{shor_large-scale_2019} as well as for being included in Freebase and Wikipedia~\cite{wagner_women_2016} is higher for women than for men, the women included in the study are likely to be more inherently\hyp noteworthy than the men included.
This might in turn affect our estimate of the association of gender with post-mortem memory:
although we identified only small and mostly insignificant effects, it is conceivable that different effects might appear if the inherent noteworthiness was equalized across genders in the dataset by lowering the bar for inclusion for women or raising it for men---an important methodological challenge.
Going forward, researchers should also strive to remove the bar for inclusion in a study of post-mortem memory altogether, by moving from a noteworthy elite of public figures to a representative set of regular people.
With the widespread adoption of social media, we may, for the first time in history, not only ask, but also answer, who is remembered after they die.
\section*{Methods}
\subsubsection*{News and Twitter corpora}
We compiled a large corpus of media coverage via the online media aggregation service Spinn3r, which provides ``social media, weblogs, news, video, and live web content''~\cite{spinn3r_doc}.
We had full access to the Spinn3r data stream and collected a complete copy over the course of more than five years (June 2009 to September 2014) via the Spinn3r API, for a total of around 40~terabytes of data.
Besides the main text content, documents consist of a title, a URL, and a publication date.
Twitter posts (\textit{tweets}) were easy to recognize automatically in the Spinn3r data, whereas news articles were not explicitly labeled as such.
In order to identify news articles, we started from a comprehensive list of all 151K online news articles about Osama bin Laden's killing (2~May 2011) indexed by Google News \cite{bharat2011google}.
Assuming that every relevant news outlet had reported on bin Laden's death, we labeled as news articles all documents in the Spinn3r crawl that were published on one of the 6\,608 Web domains that also published an English news article about bin Laden's death according to the Google News list.
We included in our analysis all English-language news articles and tweets collected between 11~June 2009 and 30~September 2014. The resulting corpus comprises, for each day, hundreds of thousands of news articles and tens of millions of tweets (\SupFig{1}).
Although Spinn3r does not publicly disclose its data collection strategy, we assess the corpus as highly comprehensive (\SupTab{3}{}).
\subsubsection*{Detecting people mentions}
In order to construct mention time series (\Figref{fig:gallery}), we had to identify documents that contain the names of dead public figures.
This is not a trivial task, since names may be ambiguous. Entity disambiguation is a well-studied task, but unfortunately natural language processing--based methods were too resource\hyp intensive for our 40-terabyte corpus, so we resorted to a simpler method:
in addition to fully\hyp unambiguous names and aliases (henceforth simply ``names'') belonging to a single entity,
we identified a set of mostly\hyp unambiguous names, which refer to the same entity at least 90\% of the time in English Wikipedia, and we mapped each mention of such a name in the Spinn3r corpus to the entity it most frequently referred to in Wikipedia (people without any highly unambiguous name were excluded).
For Twitter, we considered a tweet to mention a given person if the tweet mentioned at least one full name of the person.
For the news, we additionally required at least one additional mention of the person (via a full name or a token-based prefix or suffix of a full name), in order to reduce spurious mentions (\abbrevStyle{e.g.}\xspace, when the person was merely mentioned in a link to another article, rather than in the core article content).
\subsubsection*{Inclusion criteria}
In order to compile a set of dead public figures, we started from the 33\,340 people listed in the 28~September 2014 version of the Freebase knowledge base as having died during the period spanned by our media corpus (11~June 2009 to 30~September 2014).
On 86 out of these 1\,936 days, Spinn3r failed to provide data due to technical problems.
We excluded people for whom the 100 days immediately following death included at least one of the 86 missing days (in order to obtain better estimates of short- and long-term boosts), who died within less than 360 days of the corpus boundary dates (in order to compute pre-mortem means and long-term boosts in the same way for everyone), who were mentioned on fewer than five days in either of the news or Twitter during the 360 days before death (in order to avoid extremely noisy pre-mortem means), or whose names on English Wikipedia contained parentheses, \abbrevStyle{e.g.}\xspace, ``John Spence (Trinidad politician)'' (as such names are unlikely to be used in prose).
These criteria reduced the set of people from 33\,340 to 2\,362\xspace.
\subsubsection*{Biographic features}
Each public figure was described by the following biographic features, extracted or computed from Freebase:
age at death,
gender,
manner of death (natural or unnatural, inferred from the more detailed cause of death, cf.\ \textit{Taxonomy of causes of death} in Supplementary Information),
language (``anglophone'' for citizens of the U.S., Canada, the U.K., Ireland, Australia, New Zealand, or South Africa; ``non-anglophone'' for citizens of other countries; ``unknown'' for people with no nationality listed in Freebase),
and notability type (a profession or role for which the person was most known, \abbrevStyle{e.g.}\xspace, ``singer'' for Whitney Houston; manually grouped into six classes:
arts, sports, leadership, known for death, general fame, academia\slash engineering;
cf.\ \textit{Taxonomy of notability types} in Supplementary Information).
The distribution of these features is summarized in \SupTab{1}{} for all public figures, for the people included in the study, and for the subset retained for the regression analysis.
\subsubsection*{Mention time series}
To avoid taking logarithms of zero when constructing mention time series, a constant value of $\epsilon$ was added to each individual time series before taking logarithms, where $\epsilon$ was the minimum non-zero value across all individual time series (but note that the raw time series of \Figref{fig:gallery} were drawn without adding $\epsilon$).
Separate values of $\epsilon$ were computed for the news and for Twitter.
When smoothing time series via the variable span smoother~\cite{friedman_variable_1984}, we considered the pre- and post-mortem periods separately, in order to not smooth out the spike that usually immediately followed death.
Missing days were interpolated linearly.
\trackedChangeNew{
\subsubsection*{Model fitting}
In order to fit the shifted power law model to the data, we operate on the logarithmic scale, by finding the nonlinear least-squares estimates
\begin{equation}
\argmin_{a,b,c} \,
\sum_{t=1}^{400}
\left(
\langle \log S_i(t) \rangle -
\log \left( a t^{-b} + c \right)
\right)^2,
\end{equation}
where $\langle \log S_i(t) \rangle$ is the arithmetic mean of the empirically measured $\log S_i(t)$ over all persons $i$.
The following optimal parameters were obtained:
\begin{eqnarray}
\text{News:} & a=5.58 \times 10^{-5}, \;\;\; b=1.34, \;\;\; c=1.75 \times 10^{-6}\\
\text{Twitter:} & a=1.90 \times 10^{-6}, \;\;\; b=1.54, \;\;\; c=2.35 \times 10^{-8}
\end{eqnarray}
}
\subsubsection*{Data and code availability}
All analysis code, as well as mention frequency data and supplementary data, have been deposited in GitHub (\url{https://github.com/epfl-dlab/post-mortem-memory}).
\section*{Acknowledgments}
R.W. was partly supported by Swiss National Science Foundation grant 200021\_185043, Collaborative Research on Science and Society (CROSS), and gifts from Google, Facebook, and Microsoft.
J.L. is a Chan Zuckerberg Biohub investigator and was partly supported by NSF grants OAC-1835598 (CINES), OAC-1934578 (HDR), CCF-1918940 (Expeditions), and IIS-2030477 (RAPID); Stanford Data Science Initiative; and Chan Zuckerberg Biohub.
We thank Janice Lan for help with taxonomy construction,
Spinn3r for data access,
and Ahmad Abu-Akel, Micha\l{} Kosi\'nski, and Michele Catasta for helpful discussions.
\renewcommand\refname{References}
\bibliographystyle{plain}
|
\section{ResponSeNotes}
\label{sec:51}
\textcolor{red}{Hiyerarşik yapı çalışırken,kullanılan kod imgeleri tek tek alıp tahmin etmek üzerine kurulu. İmgeyi diskten okunduktan sonra ilk sınıflandırıcıya gidiyor. Burada birinci seviye sınıflara ayrıldıktan sonra if ve switch case ifadeleri kullanılarak ayrıntı sınıfları için tekrar sınıflandırılıyor.
Ölçüm yaparken; konvolüsyonel modellerin ve I/O işleminin süreyi domine etmesi dikkate alınmıştır. Bu sebeple Keras ImageGenerator sınıfı kullanılarak batch hazırlanmıştır. Modellerin tümü özdeş olduğu için, ImageGenerator Eval fonksiyonu giriş ve çıkışı arasındaki zaman ölçülmüştür. Eval fonksiyonu çıkışında ImageGenerator ile oluşturulmuş olan batch doğrudan ikinci modele gönderilmiş ve burada Eval fonksiyonu tekrar çalıştırılmıştır. Eval fonksiyomu çıkışında alınan zaman iki modelin peşpeşe çalıştığı toplam zamanı vermektedir. İlk modelin tek başına çalışması ile de Basix Classifier süreleri hesaplanmıştır.}
\section{Introduction}
Thanks to the available satellite and aerial imaging technologies, the earth's surface can be observed and analyzed with images in various spatial, spectral, and temporal resolutions. The research in remote sensing scene image analysis aims to classify scene images that are obtained from available instruments in various forms to some predefined, discrete set of semantic categories, e.g. residential area, forest, bridge, etc. In this context, a scene image refers to a manually created local image patch that ideally contains a scene belonging to a particular semantic category obtained from large-scale remote sensing images \cite{Ref-NW45}. Automating the content labeling of local patches in large scale remote sensing images is useful for many application areas such as environment monitoring \cite{phinn2012multi}, urban planning \cite{
li2013object,mishra2014mapping,kim2009forest}, analysis of agricultural regions \cite{janssen1992knowledge}, managing natural disasters \cite{stumpf2011object,cheng2013automatic}, land use/cover determination \cite{chen2014pyramid,chen2016land} and so on \cite{Cheng_2020}.
Early works in this domain focused more on the identification of pixel labels in aerial scenes since the resolution of such images are very low and sometimes, a sub-pixel analysis may also be necessary \cite{janssen1992knowledge}. However, the emergence of high-resolution images made interpretation of pixels harder and meaningless, since the semantic context of a scene can be interpreted by considering the relationships of multiple pixels, belonging to different objects and their surroundings \cite{blaschke2001s}. The next direction was the identification of objects, such as buildings, cars, trees, etc., from remote sensing images \cite{blaschke2014geographic,druaguct2006automated,eisank2011generic}. However, in higher resolution images, semantic content is identified correctly with the interaction of multiple objects, or some of the content is ambiguous without the surrounding context if only object appearances are taken as references; most object-classification based approaches struggle among similar scenes in the scene classification task. For instance, trees, roads, and houses could be visible in a dense residential area. The content of the high-resolution remote sense images are usually heterogeneous; hence, closing the semantic gap between the low-level semantics of the image pixels/objects and high-level (i.e. scene level) semantics become crucial \cite{bratasanu2010bridging}. Therefore, recent works focus more on a semantic-level remote sensing scene classification \cite{zhao2016feature, xia2015accurate, cheng2015effective,zhang2013semi, zheng2012automatic,Ref-UCM,zhong2015scene,zhang2016semantic,hu2015unsupervised,Ref-NW45}.
Since the publication of the UC Merced dataset in 2010 \cite{Ref-UCM}, various remote sensing scene image datasets have been made publicly available to support the research in the remote sensing scene classification (RSSC) domain. Most of the related publicly available datasets provide images in red, green and blue (RGB) spectral bands, in various spatial and image resolutions \cite{Ref-UCM, Ref-AID, Ref-WHU, Ref-RSICB, Ref-NW45}. Among these, the most challenging dataset, which is NWPU-RESIS45 (NW45 for short), provides more samples per 45 distinct semantic scene categories with high image variations and diversity, in a range of spatial resolution from 0.2 to 30 meters. In this dataset, scenes contain a broad category of images including land use and land cover classes (e.g., farmland, forest, residential area), man-made object classes (e.g., airplane, ship, bridge, church), and natural object classes (e.g., beach, river, cloud, etc.).
Remote sensing scene classification (RSSC) is a challenging problem, mainly due to high variations in pose and appearance, viewpoint, illumination, and backgrounds of the scene images. The images have high intra-class variation and inter-class similarities. There are various computational approaches for the automatic classification of remote sensing scene images in the literature (Section \ref{sec:2}). Recent advancements in the computer vision and machine learning domains, particularly deep learning approaches, accelerated the success of the computational approaches considerably in this field \cite{Cheng_2020}. The most successful performances are achieved using the convolutional neural network (CNN) based models in this domain. Promising classification accuracies are obtained by applying transfer learning to some pre-trained deep CNN models that are trained on the ImageNet dataset \cite{deng2009imagenet}. Adaptation of the models to the RSSC domain improves classification performances, which is otherwise low due to the relative scarcity of the number of images that are required for training deep models in the available RSSC benchmarks \cite{Ref-UCM, Ref-AID, Ref-WHU, Ref-RSICB, Ref-NW45}.
Most of the existing solutions to the RSSC problem use deep CNN-based models to directly predict the fine-grained class labels from a given scene image. In this work, we investigate the RSSC problem from a very basic, yet unexplored perspective; using a two-layer hierarchical structure, where we utilize the semantic relationships between the classes to define the hierarchy. Given the hierarchy tree, we construct a hierarchical model to solve the problem by using high capacity deep CNN models in each node of the hierarchy tree. Although some hierarchic solutions exist in the literature in the RSSC domain, the hierarchy is usually defined based on clustering similarities of the collected samples' features.
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{Figure1.pdf}
\caption{Flow diagram of the proposed solution.}
\label{fig:Fig_Flow}
\end{figure}
Our hierarchic solution has two primary advantages: (1) More general distinguishing features of a category that are ignored by a fine-level category label can be captured in a coarse-level classifier, (2) a dedicated classifier for hard-to-distinguish classes can learn more discriminative features utilizing a larger dedicated learning capacity. With this in mind, we designed a two-level hierarchical framework utilizing the defined semantic categories of the scene classes, \textit{i.e. buildings, transportation, water areas, natural lands, and constructed lands}, then each node in this representation is sub-divided into related fine-grained classes, \textit{(e.g. beach, lake, river, etc. under the water areas category)}. We have a backbone CNN (pre-trained DenseNet \cite{huang2018densely}) that serves as the building block in our solution; we then fine-tune the coarse-level independent layers in the first layer of our hierarchy and the fine-level independent layers in the second layer of the hierarchy. The models are trained at each node independently, sharing the low-level features of the building block CNN. The flow diagram of our model is depicted in Fig. \ref{fig:Fig_Flow}.
Our model is similar more to the hierarchical model proposed in HD-CNN architecture \cite{ICCV2015_HDCNN}, which is designed for object classification tasks for CIFAR100 and ImageNet benchmark datasets; yet differently from them, we conditionally run the fine category model components considering the coarse-level maximum prediction probabilities. There are two primary reasons behind our design: (1) We want to be able to evaluate the independent performances of the fine-level classifiers, in case it is preferred in particular domains, (2) to optimize the computation time of the hierarchic model, by truncating alternative paths in the hierarchy tree considering the likelihoods. We reduce the computational time considerably by the conditional execution of the fine-level CNN models without sacrificing the model performance; only one branch, with the maximum prediction probability, is executed for each image.
In this research, we want to inquire about the effects of the increased capacity utilization on model performance in the RSSC domain, particularly for addressing the ambiguities arising from high intra-class variances and inter-class similarities in scenes. Even though existing deep learning models have high capacities, i.e. capable of learning and using millions of parameters for effective non-linear transformations, the nature of the data in the RSSC domain is still challenging due to the ambiguities in the class boundaries \cite{sen2019scene}. In this context, we experimented with a hierarchical scheme that we designed in a two-layer tree structure, where each node of the tree is itself a deep model dedicated to classifying a subset of semantically grouped fine-grained categories. We worked with (NW45) \cite{Ref-NW45} by re-arranging the categories in this dataset in two layers. The contributions of this paper can be summarized as follows:
\begin{itemize}
\item We compared the performances of a hierarchical (two-level) and a non-hierarchical (one-level) models using the class hierarchies that we formed from the NW45 dataset. Both schemes are designed using the same backbone CNN model architectures, i.e. DenseNet-121. To the best of our knowledge, this is the first time in this domain that the semantic hierarchy is created as we presented in this work, using independent deep CNN models at the nodes of the hierarchy tree.
\item We computed the approximate overall prediction probabilities of the hierarchical model branches statistically by taking account of the independent prediction performances of the individual models in the structure. The empirical results are validated with these estimates.
\item Even though we applied pruning in the execution of hierarchical tree branches, by following only the most likely branch in the second level for computational efficiency, the performance of the hierarchical model is comparable to the state-of-the-art non-hierarchic model performances.
\item We empirically observed and statistically validated that due to the accumulation of errors in the cascading layers of the hierarchical structure, the overall performance of the hierarchical model can not exceed that of the similarly designed non-hierarchical model. Yet, the independent fine-grained classifier accuracies in the second level of our hierarchical structure are higher as we expected and can be utilized in related contexts as separate models.
\end{itemize}
The remainder of this paper is organized as follows: In Section \ref{sec:2}, we provide related works particularly focusing on deep learning-based approaches in the RSSC domain. In Section \ref{sec:3}, we explain our dataset and CNN models. Following that, in Section \ref{sec:32}, we provide our proposed method in detail. In Section \ref{sec:4}, we provide our experiment results, and finally, in Section \ref{sec:5}, we provide the conclusion part.
\section{Related Works} \label{sec:2}
From a broad perspective, three main classification approaches exist in the RSSC literature \cite{Cheng_2020}: pixel-level, object-level and scene-level approaches. Pixel-level methods are based on semantic segmentation, where each pixel in an image is assigned to one of the pre-defined semantic categories. It is an active research topic, particularly for the multispectral and hyperspectral analysis of the scene images \cite{tuia2011survey, he2017recent}. After the improvements in the spatial resolution of the RS images, detecting objects in the images became more feasible, so researchers focused more on object-level segmentation and classification \cite{blaschke2001s, yan2006comparison}. Object-level classification aims to detect the objects and classify images depending on the relationships between these objects. This approach provides better results, however, it usually suffers from the diversity of object appearances in remote sensing scene images. More recently, especially after the increase in the availability of high spatial resolution aerial and satellite images, and public datasets with scene-based class labels, the research efforts are concentrated more on scene-level classification (RSSC), rather than pixel or object-level classification.
RSSC is an image classification problem and computational solutions to this problem require the extraction of useful features from scene images. Xia et al. \cite{Xia_2017} categorized the feature extraction approaches in scene classification problem into three levels: low-level, mid-level and high level. Low-level and mid-level features are hand-crafted features that are coded manually using various computer vision algorithms \cite{inproceedingsSantos2010}, \cite{yang2008comparing}, \cite{chen2011evaluation}, \cite{luo2013indexing}.
Despite the promising results presented in the literature, classifying a remote sensing image using hand-crafted features has its limitations in the RSSC domain. Representational power of features computed using traditional methods falls short against complex variations in remote sensing images. Instead, high-level features, which are also named deep features, are extracted using deep neural networks. Starting with the seminal work of Krizhevsky et al. \cite{krizhevsky2012imagenet}, deep CNN models have been utilized successfully in various application domains and became one of the crucial methods, especially for image classification and semantic segmentation tasks from various domains \cite{SurveyClassification, SurveySegmentation}. Convolutional neural networks are deep learning structures where feature extraction and the classification phases are combined and learned end-to-end during training.
Despite their end-to-end structures and high success rates in classification, CNN architectures have two major drawbacks: the complexity of training and the need for large amounts of labeled data. To avoid these two drawbacks, transfer learning methods are used to adapt pre-trained models, which are trained using ImageNet \cite{deng2009imagenet} dataset, to different domains. ImageNet is a large natural image database that includes thousands of labeled examples on each class. Most of the well-known architectures, such as AlexNet\cite{krizhevsky2012imagenet}, DenseNet \cite{huang2018densely}, ResNet \cite{he2015deep}, are trained using this dataset. Retraining of these pre-trained networks with various approaches provides higher accuracies in other domains where labeled samples are scarce.
Recently, research in the RSSC domain has also been dominated by deep learning-based methods. Although there are large amounts of aerial or satellite images, there are only a few public benchmark datasets with labels in the RSSC domain \cite{marbhal2020evaluation}. To prevent the overfitting problem, which is a serious issue with deep model training using small datasets, custom-designed shallow architectures are utilized frequently, such as the one presented in Luus et al.'s work \cite{luus2015multiview}. They designed a small CNN with four convolutional layers. Nogueira et al, \cite{nogueira2017towards} compared the efficiency of training deep models from scratch and fine-tuning of pre-trained networks in the RSSC domain using different layers of CNN network as feature extractors.
The lack of a large number of labeled images in the RSSC domain and the advantages of transfer learning has led to the utilization of transfer learning-based approaches. Hu et al.\cite{hu2015transferring} and Cheng et al. \cite{cheng2016scene} analyzed how to use pre-trained CNNs in the RSSC domain. Many follow-up works used different pre-trained CNN models and evaluated the performances of these models applying transfer learning and data augmentation techniques; in \cite{castelluccio2015land} using GoogLeNet \cite{szegedy2014going} and CaffeNet \cite{jia2014caffe}, in \cite{scott2017training} using CaffeNet, ResNet and GoogLeNet models. In \cite{sen2019scene}, we evaluated the generalization performances of CNN models in the RSSC domain with a different perspective and reported that transfer learning of many moderate level CNN models performs better than their measured accuracies.
Recently, different methods have employed new strategies to increase the classification performances of CNN-based models on the available datasets. \cite{Cheng2018WhenDL} employed metric learning to address the inter-class similarity and intra-class diversity problems in RSSC domain; instead of the optimization based on cross-entropy loss, they optimized a discriminative loss function during CNN trainings. In \cite{HierarchicalAtt&Fusion2020}, the authors used channel attention modules to improve the features that are obtained from different layers of the ResNet50 model.
In \cite{MetaLearning_2020_CVPR_Workshops}, model agnostic meta learning is applied in few shot classification setting for land use classification to handle high diversity of the scene images in RSSC domains. Some works utilized siamese and triplet networks \cite{liu2017scene, liu2019siamese}, some used model ensembles to increase the model performances \cite{zhang2015scene, scott2017fusion, scott2018enhanced}.
\input{Table1.tex}
Land use/cover changes are tracked to observe urban and rural development. Many countries have constructed schemes which are arranged in multiple levels \cite{LAGRO2005321}, \cite{CORINNE20190510}. Although classifying land use/cover classes in a multilevel hierarchically structured scheme is a common method, research using hierarchical, multi-layered, structures are fewer.
Most of the works that use machine learning techniques create the hierarchical feature representation using clustering-based methods and then that hierarchical representation is used to improve the fine-grained predictions. In \cite{HierarchicalMetricLern_2017}, the authors organized the classes as binary trees using iterative max-margin clustering strategy and then trained each non-leaf node of this hierarchy to learn different metric transformations to increase the discrimination between the classes in that layer. In NW45, they reported $84.6\%$ classification accuracy with $80\%$ training data. Without metric learning training, the accuracy they report is $77.4\%$. (They use VGGNet-16 features in their work). In \cite{HierarhicalWassersteinCNN_2019}, the authors build a binary hierarchy three using the fusion of the deep features from AlexNet, GoogleNet, and VGGNet-16. They extracted features from these deep models and used clustering algorithms to determine categories in each branch of the hierarchy at that level. They obtained the best result in NW45, i.e. $96.37\%$, by fusing all the features from these three deep models while creating the binary hierarchy trees.
The hierarchical scheme that we apply in this work is different from the above-mentioned works. In the computer vision domain, there is a vast literature on hierarchic structures \cite{Survey2012}. One of the pre-deep learning era works that use object hierarchy trees in the classification of objects is \cite{ObjectHierarchy_ICCV_2007}. Relying on the fact that human cognition exploits hierarchic representation of world objects \cite{JOHNSON1998515}, they constructed sample object hierarchies as a tree structure and trained a classifier at each node of this tree using the natural images obtained from a set of objects from real life, such as car-side, school-bus under \textit{the closed frame vehicles category}; duck, ostrich under \textit{the winged animals' category}, etc. This is conceptually the most similar design to our approach in this domain. When identifying an object/scene, human beings follow a coarse to fine identification scheme. To exploit this; here, we create a two-layer hierarchy, where at each node we fine-tune a deep CNN model to the classes defined at that semantic layer using the ground truth labels for supervised training. Our work is also related to the hierarchical scheme proposed in \cite{ICCV2015_HDCNN} for object classification in ImageNet 1000-class dataset. The components in their model is also independently pretrained similar to our model. Based on our prior observations that the coarse-level classification is performed with high prediction probabilities, we avoided computationally expensive probabilistic averaging in the final prediction stage that is applied in \cite{ICCV2015_HDCNN}, which requires evaluation of all model predictions in the second level. Instead, we applied pruning to select the branch with the maximum prediction probability in the second level of our solution.
Note that, the recent model performance improvement techniques, such as training model ensembles, metric learning, various forms of attention mechanisms, multimodel feature fusion, etc. are orthogonal to the research that we presented here. Since any performance improvement technique in the deep learning domain can be applied to the primary CNN models in our hierarchy, we are providing the performances of the vanilla models in our hierarchic scheme as a baseline hierarchic model. We obtained comparable classification performance, i.e. $94.10\pm0.65\%$ (see Table \ref{tab:table4}), with the existing state-of-the-art methods in this dataset, without applying any performance improvement techniques.
\section{Materials}\label{sec:3
We train our models using the NWPU-RESIS45 (NW45) dataset and two instances of the DenseNet architecture as the backbone CNN model. In this part, we will first briefly explain the NW45 dataset and provide the hierarchical scheme that we designed using the annotations in this dataset. Then, we provide the details of our hierarchical solution.
\subsection{NWPU-RESIS45 Dataset} \label{sec:24}
NW45 is one of the largest datasets that is recently released by Cheng et al. \cite{Ref-NW45} for RSSC studies. It has 45 classes and each class includes 700 RGB images obtained from Google Earth images. The pixel size of all images is 256x256, and the spatial resolution varies from 20cm to 30m per pixel.
This dataset contains a high degree of similarity between images from different classes and high diversity in images within the same class. These are the two important challenges that are faced frequently and need to be addressed in the RSSC domain. Some sub-categories have remarkably high semantic overlaps, such as commercial area and industrial area, tennis court, and basketball court, which makes this data set more challenging compared to other smaller-scale datasets in this domain.
\subsection{Hierarchical Interpretation of the NW45 Dataset}\label{sec:311}
NW45 dataset is not originally designed and labeled in a multilayered hierarchical scheme. We re-organized the classes in a two-layer hierarchy and used that scheme to evaluate our model performances. To the best of our knowledge, this is the first work that tries this scheme on this dataset. In this context, we reinterpret the scenes in the original dataset in a two-layered scheme. In this interpretation, there are 5 main classes on the first layer: \textit{Buildings, Transportation, Natural Lands, Constructed Lands, and Water Areas}. All the 45 classes, which are originally defined in the NW45 dataset, are grouped semantically under these top-layer classes. The proposed two-layered hierarchical structure is summarized in Table \ref{tab:table1}.
\subsection{CNN Architectures }\label{sec:312}
DenseNet architecture has different versions, depending on the number of convolutional layers, i.e. DenseNet-121, DenseNet-160, DenseNet-201 etc. Considering the performances, computational complexities, and the number of parameters in these models, we opted for an instance of DenseNet-121 that is pre-trained with the Imagenet \cite{deng2009imagenet} dataset. We experimented with two different model architectures as ablation studies in all the models that we developed in this research while forming the proposing hierarchical and non-hierarchical model structures.
\input{Table2.tex}
\textit{D121-Full}: This is the full form of the DenseNet-121 architecture. We changed only the number of classes in the last layer, i.e. softmax, of the model. We fine-tuned the last 7 layers of \textit{Conv4block[1...16]}. Other layers of the network are kept unchanged during training to protect the weights of the pre-trained model. Hence, in the hierarchic schemes, all the models, at the coarse and fine-levels, in the hierarchy tree share the corresponding fixed low-level layers of this model (Fig. \ref{fig:Fig_Flow}).
\textit{D121-Half}: This architecture is designed using only the first half of the DenseNet-121 model. We kept the layers only up to the \textit{pool3} layer and fine-tuned the last 4 layers of \textit{Conv3block[1..12]}. Hence, all the models in the hierarchy tree, at the coarse and fine-levels, share the corresponding fixed low-level layers of this model (Fig. \ref{fig:Fig_Flow}).
We used the compact network especially to observe the performance of using small models for high granularity categorical classification, (in the second level), in the proposed hierarchy. The comparison of these two CNN architectures according to their parameters are shown in Table-\ref{tab:table2}.
\begin{figure*}[ht!]
\centering
\includegraphics{Figure2.pdf}
\caption{Structure of the proposed hierarchical model.}
\label{fig:Fig_ProposedModel}
\end{figure*}
\section{Proposed Model} \label{sec:32}
The semantic class relations in the proposed hierarchical model architecture are depicted in Fig.\ref{fig:Fig_ProposedModel}. (The flow diagram of the same model is depicted in Fig. \ref{fig:Fig_Flow}.) This architecture makes the classification process in two cascaded layers: in the first layer (the coarse-level), we have a \textit{First-level Classifier}. This model takes the input images and classifies them as one of the coarse-level categorical classes (Table-\ref{tab:table1}). According to the classification in the first layer, the input image is then sent to the related Second-level Classifier model, where the maximum predictive probability (MPP) is obtained as the coarse predictions, to get the finer-level class labels in that category. For example, if the input image is classified as \textit{Water Areas} using the First-level Classifier, with the MPP, then it is sent to WaterAreas Classifier to identify the fine-grained sub-categories of the related class, like river, beach, etc. A detailed explanation of the classifiers in our architectures is explained below:
\begin{itemize}
\item First-level Classifier: This is used for the classification of the image in a broader context, i.e. coarse-layer of the hierarchy. Output categories of the First-level Classifier are: \textit{Buildings, Transportation, Natural Lands, Constructed Lands, Water Areas} classes.
\item Second-level Classifiers: The classifiers after the First-level Classifier are referred to as the second-level classifiers, which are used to determine the fine-level categorical labels of given images. There are 5 second-level classifiers in our hierarchy; \textit{Buildings-Classifier, Transportation-Classifier, NaturalLands-Classifier, ConstructedLands-Classifier, WaterAreas-Classifier}. Each class in this level is trained using the related sub-level labels in the original NW45 dataset (Table \ref{tab:table1}).
\item Basic (Non-Hierarchic) Classifier: This classifier takes an input image and classifies it into one of the 45 categories that is defined in the NW45 dataset. This model is trained end-to-end using the ground truth labels of the NW45 dataset and detects the category of a given image directly.
\end{itemize}
\subsection{Model Training}\label{sec:training}
We trained all models using the 5-fold cross-validation method. We shuffled the data and created five-folds, each containing 448 sample images for training, 112 sample images for validation, and 140 sample images for testing. We applied data augmentation during training, by rotating the images 90, 180, and 270 degrees clockwise; and by taking the horizontal and vertical reflections of the images. To compare the performances of our complex and simple models, we trained both D121-Full and D121-Half model architectures as specified, for all the models in this work.
\subsection{Experiment Settings}\label{sec:experiments}
We performed three different experiments using our 7 different models (Section \ref{sec:32}) which are trained using the same procedures explained in Section~\ref{sec:training}. For the hierarchical structure, we have a First-level Classifier, i.e. 5Class-Classifier, which has 5 categorical outputs, that are designed according to the proposed hierarchical scheme (Fig.~\ref{fig:Fig_ProposedModel}). For each output category, we have a Second-level Classifier, as we explained before. In the non-hierarchical scheme, we only have a Basic (non-hierarchic) Classifier, which classifies the images into 45-class categories that are defined in the NW45 dataset. We also refer to that model as a 45Class-Classifier. The details of the three experiments are explained below.
\textbf{\textit{Experiment-1: Hierarchical top-down}}
After training all sub-models of the First-level and Second-level Classifier models, we constructed our proposed hierarchical structure (Fig.~\ref{fig:Fig_Flow}). In this experiment, First-level Classifier, which we also refer to as the 5Class-Classifier, takes an input image and classifies it as one of the 5 categorical outputs belonging to second-level categories. Then the input is classified with the related Second-level Classifier for which the MPP is obtained.
\textbf{\textit{Experiment-2: Hierarchical top-down}}
In this setting, we skip the First-level classifier and directly send the input images to the related Second-level Classifier, which is determined using the ground-truth labels of the images. We want to see the test performance of the proposed hierarchical scheme when the First-level Classifier works with 100\% accuracy.
\textbf{\textit{Experiment-3: Hierarchical bottom-up}}
In this experiment, we perform a hierarchical classification from the bottom-up direction. We first classify a given image using our Basic (45Class) Classifier. According to the output of this classifier, the second-level category of the image is determined according to our hierarchical scheme (Table \ref{tab:table1}). Then, the input image is re-evaluated using the corresponding Second-level Classifier. For example, for a given \textit{river} image, assume that it is classified as a \textit{beach} by the Basic (45Class) Classifier. Since in our hierarchical scheme beach class is under the WaterAreas category, we send the same image to WaterAreas Classifier for re-evaluation. Then the category of the image is determined considering the sub-categories of the WaterAreas result, i.e. may be corrected as a river, in the second evaluation. The order is, therefore, bottom-up, where we follow the direction from a fine-category (bottom) to a coarse category (top).
In addition to these experiments, we also evaluated the performance of the Basic (45Class) Classifier in a non-hierarchical setting.
\section{Results and Discussions} \label{sec:4}
\input{Table3.tex}
In this research, we want to analyze the performance of the deep models that are configured to identify coarser semantic categories of images first, and then identify the finer category in the context of the coarse category in a hierarchical setting, in the challenging RSSC domain. We first start by evaluating the performances of the 7 basic models that we defined in detail in Section \ref{sec:experiments}. The experiment results are provided in Table~\ref{tab:table3}. As can be seen from the performances in this table, Second-level Classifiers perform very well on the related sub-category classification tasks. Except for the Building-Classifier, all the models perform higher than 96\% accuracy with the D121 Full model. Considering the high performance of the First-level Classifier, i.e. more than 97\% accuracy, and the performances of the Second-level Classifiers, at a first glance, it seems as though the hierarchical structure will work better than the Basic (45Class) Classifier model. However, when we compute the overall accuracies, we observe that it's not actually the case (Table~\ref{tab:table4}).
Overall accuracies of the three experiments are provided in Table~\ref{tab:table4}, together with the performance of the non-hierarchic Basic Classifier. Our intuition behind modeling a hierarchical architecture was that the classification task in this model will be less challenging for the deep models in each layer. From the top-to-bottom, the model in the first level will only need to predict 5 broad categories, instead of learning all finer details of 45 categories; in the second level, each deep model will only need to learn the finer specifications of the less number of classes, which could potentially increase the overall accuracy rates.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\textwidth]{Figure3.pdf}
\caption{tSNE visualizations of feature space. Top row: non-hierarchic \textit{45Class Classifier}, middle, and bottom rows: hierarchic fine-level classifiers; from left to right \textit{Lands Constructed Classifier}, \textit{Transportation Classifier}, \textit{Buildings Classifier}, \textit{Lands Natural Classifier}.}
\label{fig:Fig_Projections}
\end{figure}
To understand the distribution of data in the trained models' embedding space better, we analyzed the feature distribution of all the models using the t-Distributed Stochastic Neighbor Embedding algorithm (tSNE) \cite{tSNEVis2008}. tSNE performs a nonlinear projection from a high dimensional data space to lower dimensions by optimizing a function to preserve neighborhood relation in the data as much as possible; hence very useful for visualizations of complex data distributions in higher dimensions. We took the feature vectors from the D121-Full model from the last layer, (right before the softmax layer, as an 1152 dimensional vector for each sample image) and project these vectors to 2-dimensional planes with tSNE. The results from 4 distinct models are displayed in Fig. \ref{fig:Fig_Projections}. As can be seen, 45 class non-hierarchic model's feature space is very crowded; although we can identify clusters of individual classes (each one is shown with a different color), some of the classes (in the central part of the distribution) have quite close proximities and many samples violating class boundaries between different clusters are visible. On the other hand, we can easily see that the hierarchic model's feature distributions are better-separated thanks to the dedicated capacity for a fewer number of classes.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\textwidth]{Figure4.pdf}
\caption{Predicted class probabilities. Correct class probabilities are depicted with red boxes. Left column: test image, middle column: non hierarchic (45Class) classifier outputs, right column: hierarchic (second-level) classifier outputs.}
\label{fig:Fig_OutputScores}
\end{figure}
As supported also by tSNE projections, independent evaluations from hierarchic levels show that our intuition is correct, however, the combined hierarchic model experiments - as we implemented here by pruning - showed us that it is not the case in the combined model. As can be seen, the performance of the primary Hierarchic model is less than all the other models, i.e. 92.61\% with the D121-Half and 94.10\% with the D121-Full model. In Experiment-2, the overall accuracy of the Second-level Classifiers, which is 96.73\% with the D121-Full model, is higher than that of the Basic Classifier performance, i.e. 95.40\%. We also analyzed the output prediction probabilities of Basic Classifier and Second-level classifiers for some samples that are misclassified by the Basic classifier. As shown in Fig. \ref{fig:Fig_OutputScores}, the confidence of the Second-level Classifiers are considerably higher for the true categories, while they are misclassified with a similar category with the Basic Classifier.
Our experiments show that the performance of the First-level Classifier in Experiment-1 affects the accuracy results considerably. Note that, in Experiment-2, we used the ground truth labels to determine the high-level category of each test sample. Even with the half-size model, Experiment-2 results are better than Experiment-1 accuracies. Experiment-3 performances show that re-evaluating the predictions of the Basic classifier does not improve the performance of the model as we thought at the beginning. The motivation was primarily to correct misclassifications in some upper categories that are caused by the inter-class similarities in the finer-level classes. The improvement is too small, hence in practice, it makes more sense to trust the predictions of Basic Classifier in one pass, rather than waiting for re-evaluation.
\input{Table4.tex}
\input{Table5.tex}
\begin{figure*}[ht!]
\centering
\includegraphics[scale=0.55]{Figure5.pdf}
\caption{A statistical estimation of the classification accuracies. The accuracies are obtained by multiplication of the prediction accuracies in two layers. Accuracies in top (blue) boxes (on the left side of branches) show the categorical prediction accuracies for the Full First-level model; the bottom (red) boxes show that of the Half First-level model (for the category on the right). Overall prediction accuracies are computed using the prior test data distribution (depicted with the black boxes).}
\label{fig:Fig_OverAllResults}
\end{figure*}
We also created the average accuracies of the Non-hierarchic Basic Classifier and Hierarchic top-down classifier (the model in Experiment-1) based on the 5 high-level categories that we defined in our hierarchical scheme (Table~\ref{tab:tabel5}). In both models, the fine level predictions are obtained first, i.e. river, bridge, etc. Then, the accuracies are computed on the related high-level category basis, considering the ground truth labels. For instance, for the N number of test samples in the Buildings category, we get the accuracy of the predictions of the models for the corresponding N samples. The empirical results provided in Table~\ref{tab:tabel5} show that the increase in the capacity of the model, i.e. using the D121-Full model, closes the performance gap between the Hierarchical model and the Basic model. The overall performance difference is 1.45\% with the D121 Half models, and 1.30\% with the D121-Full models. On the categorical basis, the hierarchic model makes most of its mistakes in the \textit{Water Areas} upper category.
All the results in our experiments, including the ones in Table~\ref{tab:tabel5}, are obtained using 5-fold cross-validation. In addition to the empirical analysis, we also conducted statistical analysis to validate the hierarchical model performance using the performances of the individual models in the hierarchy. The reported \textit{overall accuracies} that we computed in our experiments, in Table~\ref{tab:table3}, can be considered as the correct prediction probabilities of the models for each category. In our proposed hierarchical architecture, since each model evaluates the category of a given input independently from each other, the prediction probabilities of the hierarchical branches can be calculated using the multiplication of the probabilities of the cascaded layers. As a result, the overall performance of the model decreases in a two-layer scheme. The statistical analysis of our hierarchical model to estimate an approximate prediction accuracy for each upper category is depicted in Fig.~\ref{fig:Fig_OverAllResults}. As stated, each model's independent classification probabilities are obtained from Table~\ref{tab:table3} for this estimation. We also computed the overall prediction accuracies of the First-level model for each category; hence these accuracies are used as class conditional probabilities. Prior data distribution for the test set is used to weight the computed accuracies in each branch in the overall accuracy computation. As can be seen from the joint probabilities, although sub-category performances are slightly different, the overall accuracies of the statistical estimates are very close to the empirical results. This analysis validates our empirical results and better depicts the reason behind the performance reduction in our hierarchical models; although the individual model performances are superior in both levels, the small errors in two-cascaded layers are accumulated. For instance, the performance of the Buildings-Classifier in the hierarchical model (Second-level classifier), i.e. 93.44\%, is superior to that of the Basic Classifier, i.e. 91.52\%; yet, the joint probability with the First-level model takes it down, i.e. 90.65\%. As can also be seen from the statistical analysis, the Building class score, i.e. 90.49\% (Full model), is close to our empirical result and below that of the single-layer deep model, i.e. Basic Classifier.
The observation of the performance of Experiment-2 is also crucial (Table~\ref{tab:table4}). As also supported with the results in Table~\ref{tab:table3} and with the selected challenging samples in Fig. \ref{fig:Fig_OutputScores}, Second-level Classifiers perform very well, even with a small architecture like D121-Half. They can produce accurate results when we want to classify a special region among only particular sub-categories. For example, if we want to find out forests in a particular region, we can use a Natural-Lands Classifier or we can consider training a more specific dichotomy for special queries when needed.
\input{MyTables}
\input{MyTables_CPU}
\textbf{Computational Times of Models:} We estimated the computational times of the models in our hierarchical framework and the basic classifier and presented the estimated times in Tables \ref{tab:time6} and \ref{tab:time7}. The estimations are obtained both on GPU and CPU for a batch of 135 images multiple times (10 times) and the average of these estimations is reported. As can be seen, the running time of the hierarchical model is approximately 2 times more than the basic model (on both GPU and CPU runs) as we expected; since the hierarchic model contains the same CNN model with the Basic Classifier in a cascaded two-layer scheme. Note that, I/O times are not included in these estimations to be able to compare only the computational times of the deep models.
\section{Conclusion} \label{sec:5}
In this work, we proposed a hierarchical framework to assess the performances of deep models in this scheme. All the models are trained using two instances of DenseNet 121 models, one with high capacity and one with low capacity. We obtained empirical performances for each model using a 5-fold cross-validation with the NW45 dataset. Our empirical analysis depicted that the hierarchical model performs comparable to the non-hierarchical model, yet it can not exceed the performance of our non-hierarchic model. We supported the estimated empirical results with the statistical estimates of the model prediction probabilities.
Image patches provided in the datasets in this domain do not always contain scenes related to one fine category; we have shown some examples in our previous study \cite{sen2019scene} depicting such ambiguities on NW45 and five other related datasets. Hence, there are some dataset biases imposed in the evaluations of the models in this domain. There is a need for standardization of the datasets in this domain and reducing the ambiguities in scene labeling. As of now, there is little agreement even on the category names in related datasets. In this work, we tried generating a two-layer hierarchical interpretation of one of the largest datasets in this field to contribute more to the semantic interpretation of the classes in a hierarchy and encourage future datasets to define similar schemes for further assessments.
\section*{Acknowledgements}
The numerical calculations reported in this paper were fully/partially performed at TUBITAK ULAKBIM, High Performance and Grid Computing Center (TRUBA resources).
\bibliographystyle{ieeetr}
\footnotesize
|
\section{Introduction}
Symmetry protected topological (SPT) phases are symmetric gapped phases whose ground states are unique on any closed manifold.
In the low energy limit, SPT phases are expected to be described by invertible quantum field theories, which are classified by using bordism groups \cite{Kap2014, KTTW2015, FH2016, Yon2019}.
In particular, in 1+1 dimensions, the low energy limit of SPT phases is believed to become topological quantum field theories (TQFTs).
Therefore, the classification of 1+1d SPT phases with symmetry $G$ reduces to the classification of invertible $G$-equivariant TQFTs.
When $G$ is an internal symmetry, the low energy TQFTs are defined on oriented manifolds.
Such TQFTs are called oriented TQFTs.
The algebraic descriptions of oriented $G$-equivariant TQFTs enable us to show that 1+1d bosonic SPT phases with finite internal symmetry $G$ are classified by group cohomology $H^2(G, \mathrm{U}(1))$ \cite{Tur1999, MS2006, KT2017, CGLW2013}.
On the other hand, when $G$ involves time-reversal symmetry, the low energy limit is described by unoriented TQFTs, which can be defined on unoriented manifolds.
The algebraic descriptions of unoriented $G$-equivariant TQFTs for finite groups $G$ are given in \cite{KT2017, Swe2013}.
In particular, it is shown that invertible unoriented $G$-equivariant TQFTs are classified by twisted group cohomology $H^2(G, \mathrm{U}(1)_{\rho})$ where the homomorphism $\rho: G \rightarrow \mathbb{Z}_2$ determines which elements reverse the orientation of time \cite{KT2017, CGLW2013}.
We can extend the classification of SPT phases along with the generalizations of the notion of symmetry.
In general, symmetries are characterized by the algebraic relations of topological defects.
For example, ordinary group symmetries are associated with invertible topological defects with codimension 1.
One possible generalization of the ordinary group symmetries is symmetries generated by topological defects with higher codimensions.
Such generalized symmetries are called higher form symmetries \cite{GKSW2015}.
Another generalization is to consider non-invertible topological defects with codimension 1, which do not form a group.
In particular, in 1+1 dimensions, the algebraic relations of finite numbers of topological defect lines including non-invertible ones are described by unitary fusion categories \cite{BT2018, CLSWY2019}.
The corresponding symmetries are called fusion category symmetries.
As well as ordinary group symmetries, we can discuss anomalies and gauging of them \cite{FRS2002, FFRS2010, BCP2014a, BCP2014b, BCP2015, CR2016, BT2018, TW2019, CLSWY2019}.
Fusion category symmetries have a long history in the study of two-dimensional rational conformal field theories \cite{Ver1988, PZ2001, FFRS2004, FFRS2007, CLSWY2019}.
We can also find fusion category symmetries in lattice models such as anyonic chains \cite{BG2017, FTLTKWF2007, TTWL2008, TAFHLT2008, GATLTW2009, GATHLTW2013, PBTLTV2012}.
The thermodynamic limit of anyonic chains is often described by conformal field theories, whose energy spectra cannot be gapped without breaking fusion category symmetries.
There is also a way to construct two-dimensional statistical mechanical models with general fusion category symmetries \cite{AMF2016, AFM2020}.
These models are naturally described in terms of three-dimensional topological field theories on a manifold with boundaries.
Similarly, fusion category symmetries on the boundary of 2+1d topologically ordered states are investigated in \cite{LTLSB2020}.
Recently, fusion category symmetries are also studied in the context of two-dimensional topological field theories \cite{BT2018, TW2019, KORS2020, HL2021}.
Remarkably, it is shown in \cite{TW2019, KORS2020} that oriented bosonic TQFTs with fusion category symmetry are classified by the module categories of the fusion category.
This includes the classification of bosonic fusion category SPT phases, which are gapped phases with fusion category symmetry whose ground states are unique on a circle.
In this paper, we generalize the classification of bosonic fusion category SPT phases to the case with time-reversal symmetry.
Our approach is to axiomatize two-dimensional unoriented TQFTs with fusion category symmetry and solve them under the condition that the Hilbert space on a circle without topological defect is one-dimensional.
Along the way, we reproduce the classification of bosonic fusion category SPT phases without time-reversal symmetry.
The rest of the paper is organized as follows.
In section \ref{sec: Preliminary: fusion category symmetries}, we review fusion category symmetries in two-dimensional quantum field theories.
The contents of this section are not restricted to TQFTs.
In section \ref{sec: Oriented TQFTs with fusion category symmetry}, we classify bosonic fusion category SPT phases without time-reversal symmetry.
Section \ref{sec: Consistency conditions of oriented TQFTs} is devoted to the review of oriented bosonic TQFTs with fusion category symmetry.
By solving the consistency conditions of oriented TQFTs explicitly in section \ref{sec: Bosonic fusion category SPT phases without time-reversal symmetry}, we show that fusion category SPT phases without time-reversal symmetry are classified by isomorphism classes of fiber functors.\footnote{A fiber functor is a tensor functor from a fusion category to the category of vector spaces.}
This agrees with the result in previous papers \cite{TW2019, KORS2020}, although the approach is slightly different.
Our derivation clarifies the physical interpretation of fiber functors.
In section \ref{sec: Unoriented TQFTs with fusion category symmetry}, we classify bosonic fusion category SPT phases with time-reversal symmetry.
We first formulate unoriented TQFTs with fusion category symmetry in section \ref{sec: Consistency conditions of unoriented TQFTs} and then classify SPT phases by solving the consistency conditions of unoriented TQFTs in section \ref{sec: Bosonic fusion category SPT phases with time-reversal symmetry}.
We will see that bosonic fusion category SPT phases with time-reversal symmetry are classified by the algebraic data $(Z, M, i, s, \phi)$ where $(Z, M, i)$ represents bosonic fusion category SPT phases without time-reversal symmetry, $s$ represents bosonic SPT phases only with time-reversal symmetry, and $\phi$ represents the action of orientation-reversing symmetry.
Finally, in section \ref{sec: Examples (unoriented)}, we discuss examples including the classification of SPT phases with duality symmetry.
In some cases, duality symmetries do not admit time-reversal invariant SPT phases, although they admit SPT phases without time-reversal symmetry.
This may be thought of as mixed anomalies between time-reversal symmetry and the duality symmetries.
Throughout the paper, we assume that the total symmetry splits into time-reversal symmetry and finite internal symmetry.
\section{Preliminary: fusion category symmetries}
\label{sec: Preliminary: fusion category symmetries}
In this section, we briefly review fusion category symmetries to fix the notation by following \cite{BT2018}.
For details about fusion categories, see e.g. \cite{EGNO2015}.
The basic ingredients of fusion category symmetry are topological defect lines and topological point operators.
A topological defect line of a theory with fusion category symmetry $\mathcal{C}$ is labeled by an object $x$ of a unitary fusion category $\mathcal{C}$.
In particular, the trivial defect line corresponds to the unit object $1$, which is simple.
A topological point operator that changes a topological defect $x$ to another topological defect $y$ is labeled by a morphism $f \in \mathop{\mathrm{Hom}}(x, y)$, where $\mathop{\mathrm{Hom}}(x, y)$ is a $\mathbb{C}$-vector space of morphisms from $x$ to $y$.\footnote{We note that $\mathop{\mathrm{Hom}}(x, y)$ is a subspace of topological point operators between $x$ and $y$. For example, topological point operators on which a simple topological line ends are excluded because $\mathop{\mathrm{Hom}(1, x)}$ is empty when $x$ is simple.}
Every topological point operator $f \in \mathop{\mathrm{Hom}}(x, y)$ has the adjoint $f^{\dagger} \in \mathop{\mathrm{Hom}}(y, x)$.
When we have two topological defects $x$ and $y$ that run parallel with each other, we can regard them as a single topological defect, which is labeled by the tensor product $x \otimes y$.
We can also think of the tensor product $x \otimes y$ as the fusion of $x$ and $y$.
The trivial defect $1$ acts as the unit of the fusion:
\begin{equation}
1 \otimes x \cong x \otimes 1 \cong x.
\end{equation}
The isomorphisms $l_x: 1 \otimes x \rightarrow x$ and $r_x: x \otimes 1 \rightarrow x$ are called the left and right units respectively, which are taken to be the identity morphism $\mathrm{id}_x$ by identifying $1 \otimes x$ and $x \otimes 1$ with $x$.
The order of the fusion of three topological defects $x$, $y$, and $z$ is changed by using the isomorphism $\alpha_{xyz}: (x \otimes y) \otimes z \rightarrow x \otimes (y \otimes z)$ that is called the associator.
The associators satisfy the following commutative diagram called the pentagon equation:
\begin{equation}
\begin{tikzcd}
& (x \otimes y) \otimes (z \otimes w) \arrow[dr, "\alpha_{x, y, z \otimes w}"] & \\
((x \otimes y) \otimes z) \otimes w \arrow[ru, "\alpha_{x \otimes y, z, w}"] \arrow[d, "\alpha_{xyz} \otimes \mathrm{id}_w"'] & & x \otimes (y \otimes (z \otimes w))\\
(x \otimes (y \otimes z)) \otimes w \arrow[rr, "\alpha_{x, y \otimes z, w}"'] & & x \otimes ((y \otimes z) \otimes w) \arrow[u, "\mathrm{id}_x \otimes \alpha_{yzw}"']
\end{tikzcd}
\label{eq: pentagon}
\end{equation}
The orientation reversal of a topological defect $x$ is labeled by the dual object $x^{*}$.
In particular, the topological defect $x^{**}$ whose orientation is reversed twice is equal to the original topological defect $x$ up to natural isomorphism $a_x: x \rightarrow x^{**}$ that is called the pivotal structure.
Folding $x$ to the left is described by morphisms $\mathrm{ev}_x^L: x^* \otimes x \rightarrow 1$ and $\mathrm{coev}_x^L: 1 \rightarrow x \otimes x^*$, which are called the left evaluation morphism and the left coevaluation morphism respectively.
Similarly, folding $x$ to the right is described by the right evaluation morphism $\mathrm{ev}_x^R: x \otimes x^* \rightarrow 1$ and the right coevaluation morphism $\mathrm{coev}_x^R: 1 \rightarrow x^* \otimes x$, which are given by the adjoints of the left evaluation and coevaluation morphisms
\begin{equation}
\mathrm{ev}_x^R = (\mathrm{coev}_x^L)^{\dagger}, \quad \mathrm{coev}_x^R = (\mathrm{ev}_x^L)^{\dagger}.
\label{eq: adjoint}
\end{equation}
For a unitary fusion category, there is a canonical pivotal structure that is given by $a_x = (\mathrm{id}_{x^{**}} \otimes \mathrm{ev}_x^L) \circ ((\mathrm{ev}_{x^*}^L)^{\dagger} \otimes \mathrm{id}_x)$, see e.g. \cite{HP2017}.\footnote{The associator is omitted in this expression. In the subsequent sections, we will sometimes omit associators to simplify the notation. One can restore them wherever needed.}
We note that the left and right (co)evaluation morphisms are related by the canonical pivotal structure as follows:
\begin{equation}
\mathrm{ev}_x^L = \mathrm{ev}_{x^*}^R \circ (\mathrm{id}_{x^*} \otimes a_x), \quad (a_x \otimes \mathrm{id}_{x^*}) \circ \mathrm{coev}_x^L = \mathrm{coev}_{x^*}^R.
\label{eq: canonical}
\end{equation}
These evaluation and coevaluation morphisms are represented diagrammatically as shown in figure \ref{fig: folding}.
\begin{figure}
\begin{minipage}{0.25 \hsize}
\begin{center}
\includegraphics[width = 3.5cm]{left_ev.pdf}
\end{center}
\end{minipage}%
\begin{minipage}{0.25 \hsize}
\begin{center}
\includegraphics[width = 3.5cm]{left_coev.pdf}
\end{center}
\end{minipage}%
\begin{minipage}{0.25 \hsize}
\begin{center}
\includegraphics[width = 3.5cm]{right_ev.pdf}
\end{center}
\end{minipage}%
\begin{minipage}{0.25 \hsize}
\begin{center}
\includegraphics[width = 3.5cm]{right_coev.pdf}
\end{center}
\end{minipage}
\caption{The left and right (co)evaluation morphisms describe the operation of folding topological defect lines to the left and right respectively. In the folded diagrams, topological point operators corresponding to (co)evaluation morphisms are often implicit.}
\label{fig: folding}
\end{figure}
In particular, the zigzag-shaped defect constructed from these folding operations cannot be distinguished from the straight defect:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.6cm]{zigzag1.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 1.6cm]{zigzag2.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 0.2cm]{zigzag3.pdf}.
\label{eq: zigzag axiom}
\end{equation}
The loop diagram evaluates the quantum dimension
\begin{equation}
\mathop{\mathrm{dim}}x = \adjincludegraphics[valign = c, width = 1.8cm]{qdim.pdf},
\end{equation}
which does not depend on whether the loop is clockwise or counterclockwise, i.e. $\mathop{\mathrm{dim}} x = \mathop{\mathrm{dim}} x^*$.
We can also consider the sum $x \oplus y$ of topological defects $x$ and $y$.
When the topological defect $x \oplus y$ is put on a path $P$, the correlation function is calculated as the sum of the correlation functions in the presence of topological defects $x$ and $y$ on $P$:
\begin{equation}
\braket{\cdots (x \oplus y)(P) \cdots} = \braket{\cdots x(P) \cdots} + \braket{\cdots y(P) \cdots}.
\end{equation}
\section{Oriented TQFTs with fusion category symmetry}
\label{sec: Oriented TQFTs with fusion category symmetry}
\subsection{Consistency conditions of oriented TQFTs}
\label{sec: Consistency conditions of oriented TQFTs}
In this section, we review two-dimensional bosonic oriented TQFTs with fusion category symmetry $\mathcal{C}$.
We will enumerate the consistency conditions on transition amplitudes by following \cite{BT2018}.
We first assign a vector space $V_x$ to a circle with a topological defect $x$.
When a topological defect is a direct sum $x \oplus y$, the vector space $V_{x \oplus y}$ is given by the direct sum of two vector spaces $V_x$ and $V_y$.
To define the vector space on a circle with multiple topological defects, we need to specify the base point on the circle.
Once we fix the base point, the vector space on a circle with topological defects $x, y, z, \cdots$ is defined as the vector space on a circle with a single topological defect labeled by the tensor product $((x \otimes y) \otimes z) \otimes \cdots$, see figure \ref{fig: vector space on a circle}.
\begin{figure}
\begin{center}
\includegraphics[width = 6cm]{circle.pdf}
\caption{The vector space on a circle with multiple topological defects is the same as the vector space on a circle with a single topological defect labeled by the tensor product of the topological defects. To determine the order of the tensor product, we need to specify the base point on a circle, which is represented by the cross mark in the above figure.}
\label{fig: vector space on a circle}
\end{center}
\end{figure}
The vector space on a disjoint union of circles is given by the tensor product of the vector spaces on each circle.
Next, we consider transition amplitudes between the vector spaces.
Let us begin with the transition amplitude for a cylinder.
There are two types of cylinder amplitudes as shown in figure \ref{fig: cylinder amplitude}: one is the transition amplitude corresponding to a topological point operator $f \in \mathop{\mathrm{Hom}}(x, x^{\prime})$, and the other is the transition amplitude corresponding to a change of the base point.
\begin{figure}
\begin{center}
\includegraphics[width = 6cm]{cylinder_amplitude.pdf}
\caption{There are two types of transition amplitudes for a cylinder. The left figure represents the transition amplitude $Z(f): V_x \rightarrow V_{x^{\prime}}$ corresponding to a topological point operator $f \in \mathop{\mathrm{Hom}}(x, x^{\prime})$. The right figure represents the transition amplitude $X_{xy}: V_{x \otimes y} \rightarrow V_{y \otimes x}$ corresponding to a change of the base point.}
\label{fig: cylinder amplitude}
\end{center}
\end{figure}
The former is denoted by $Z(f): V_x \rightarrow V_{x^{\prime}}$.
In particular, the identity morphism $\mathrm{id}_x$ corresponds to the identity map $\mathrm{id}_{V_x}$.
The latter is denoted by $X_{xy}: V_{x \otimes y} \rightarrow V_{y \otimes x}$ when the base point on a circle with two topological defects $x$ and $y$ is moved from the left of $x$ to the right of $x$.
We call the trajectory of the base point the auxiliary line.
To make the cylinder amplitude well-defined, we require that the composition of morphisms $f$ and $g$ gives rise to the composition of the transition amplitudes $Z(f)$ and $Z(g)$, see figure \ref{fig: composition on a cylinder}:
\begin{equation}
Z(g \circ f) = Z(g) \circ Z(f).
\tag{O1}
\label{eq: composition on a cylinder}
\end{equation}
This makes $Z$ a functor from $\mathcal{C}$ to $\mathrm{Vec}$ where $Z(x)$ is defined as the vector space $V_x$ for every object $x$ of $\mathcal{C}$.
\begin{figure}
\begin{center}
\includegraphics[width = 6cm]{cylinder_composition.pdf}
\caption{The transition amplitude $Z(g \circ f)$ corresponding to the composition of topological point operators agrees with the composition $Z(g) \circ Z(f)$ of the transition amplitudes corresponding to each topological point operator. This is because putting a topological point operator $g \circ f$ can be regarded as putting two topological point operators $f$ and $g$ successively.}
\label{fig: composition on a cylinder}
\end{center}
\end{figure}
We demand that this functor is $\mathbb{C}$-linear in morphisms because a linear combination of topological point operators $f, f^{\prime} \in \mathop{\mathrm{Hom}}(x, x^{\prime})$ results in a linear combination of transition amplitudes, see figure \ref{fig: C-linear}:
\begin{equation}
Z(\alpha f + \alpha^{\prime} f^{\prime}) = \alpha Z(f) + \alpha^{\prime} Z(f^{\prime}), \quad \forall \alpha, \alpha^{\prime} \in \mathbb{C}.
\tag{O2}
\label{eq: C-linear}
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[width = 9cm]{C-linear.pdf}
\caption{The functor $Z: \mathcal{C} \rightarrow \mathrm{Vec}$ is $\mathbb{C}$-linear in morphisms.}
\label{fig: C-linear}
\end{center}
\end{figure}
Furthermore, we require that the transition amplitudes do not depend on the shape of the auxiliary line if we fix the base points on the initial and final circles.
In particular, the transition amplitudes are invariant under winding the auxiliary line around a cylinder, see figure \ref{fig: isomorphism X}.
\begin{figure}[t]
\begin{center}
\includegraphics[width = 8cm]{isomorphism_X.pdf}
\caption{Winding the auxiliary line around a circle is equivalent to moving the auxiliary line back and forth, both of which are equivalent to doing nothing. This implies that $X_{yx}$ is the inverse of $X_{xy}$. In particular, $X_{xy}: V_{x \otimes y} \rightarrow V_{y \otimes x}$ is an isomorphism.}
\label{fig: isomorphism X}
\end{center}
\end{figure}
This implies that the linear map $X_{xy}: V_{x \otimes y} \rightarrow V_{y \otimes x}$ is an isomorphism, whose inverse is given by $X_{yx}$:
\begin{equation}
X_{xy} = X_{yx}^{-1}.
\tag{O3}
\label{eq: C2}
\end{equation}
We also require that a change of the base point commutes with morphisms $f \in \mathop{\mathrm{Hom}}(x, x^{\prime})$ and $g \in \mathop{\mathrm{Hom}}(y, y^{\prime})$ as shown in figure \ref{fig: Z(f)X = XZ(f)}:
\begin{equation}
\begin{aligned}
Z(\mathrm{id}_y \otimes f) \circ X_{xy} & = X_{x^{\prime}y} \circ Z(f \otimes \mathrm{id}_y),\\
Z(g \otimes \mathrm{id}_x) \circ X_{xy} & = X_{x y^{\prime}} \circ Z(\mathrm{id}_x \otimes g).
\end{aligned}
\tag{O4}
\label{eq: C4}
\end{equation}
These equations indicate that the auxiliary line is transparent to topological point operators.
\begin{figure}
\begin{center}
\includegraphics[width = 8cm]{commutativity_of_X_and_Zf.pdf}
\caption{The change of the base point $X_{xy}$ commutes with a topological point operator on $x$ because we can prevent the auxiliary line from intersecting with the topological defect $x$ by changing the path of the auxiliary line as shown in the middle figure. This means that the auxiliary line is transparent to a topological point operator on $x$. The auxiliary line is also transparent to a topological point operator on $y$ for the same reason.}
\label{fig: Z(f)X = XZ(f)}
\end{center}
\end{figure}
The last constraint on the cylinder amplitude is that moving a base point across the topological defects $x$ and $y$ at the same time is equivalent to moving a base point across $x$ and $y$ successively.
This leads to the following consistency condition:
\begin{equation}
X_{y, z \otimes x} \circ Z(\alpha_{yzx}) \circ X_{x, y \otimes z} \circ Z(\alpha_{xyz}) = Z(\alpha_{zxy}^{-1}) \circ X_{x \otimes y, z}.
\tag{O5}
\label{eq: coherence}
\end{equation}
In particular, if we choose $x = y = 1$, we obtain $X_{1, x} = \mathrm{id}_{V_x}$, which means that moving the base point across a trivial defect does not affect the transition amplitude.\footnote{We recall that the left and right units are chosen to be the identity morphism.}
This completes the consistency conditions on the cylinder amplitude.
For later convenience, we define the generalized associator $\mathcal{A}_{p \rightarrow p^{\prime}}: V_{p} \rightarrow V_{p^{\prime}}$ as a composition of isomorphisms $X$ and $Z(\alpha)$, where $p$ is a tensor product of any number of objects and $p^{\prime}$ is a cyclic permutation of $p$ with arbitrary parentheses.
For example, the generalized associator $\mathcal{A}_{(x \otimes y) \otimes z \rightarrow (z \otimes x) \otimes y}$ is given by both sides of eq. (\ref{eq: coherence}).
In general, there are many ways to construct an isomorphism from $V_{p}$ to $V_{p^{\prime}}$ only from $X$ and $Z(\alpha)$.
However, they give rise to the same isomorphism due to eq. (\ref{eq: coherence}).
This means that the generalized associator $\mathcal{A}_{p \rightarrow p^{\prime}}$ is unique, although it has many distinct-looking expressions.
The transition amplitude for a general surface can be constructed from the cylinder amplitude and the following four basic elements: the unit $i: \mathbb{C} \rightarrow V_1$, the counit $\epsilon: V_1 \rightarrow \mathbb{C}$, the multiplication $M_{xy}: V_x \otimes V_y \rightarrow V_{x \otimes y}$, and the comultiplication $\Delta_{xy}: V_{x \otimes y} \rightarrow V_x \otimes V_y$, see also figure \ref{fig: basic elements} for diagrammatic representations.
\begin{figure}
\begin{center}
\includegraphics[width = 11cm]{basic_elements.pdf}
\caption{The above figures represent the basic elements of the transition amplitude for a general surface: from left to right, the unit, the counit, the multiplication, and the comultiplication.}
\label{fig: basic elements}
\end{center}
\end{figure}
For unitary quantum field theories, the counit $\epsilon$ and the comultiplication $\Delta_{xy}$ are the adjoints of the unit $i$ and the multiplication $M_{xy}$ respectively (with respect to the non-degenerate pairing (\ref{eq: non-degenerate})):
\begin{equation}
\epsilon = i^{\dagger}, \quad \Delta_{xy} = M_{xy}^{\dagger}.
\end{equation}
More generally, a transition amplitude changes to its adjoint when the corresponding surface is turned upside down, or equivalently when the time direction is reversed.
Any two-dimensional surface $\Sigma$ can be obtained by successively gluing cylinders and the basic elements shown in figure \ref{fig: basic elements}.
The transition amplitude for a surface $\Sigma$ is defined as the composition of the linear maps corresponding to cylinders and the basic elements that appear in the decomposition of $\Sigma$.
However, the way to decompose $\Sigma$ is not unique in general.
Therefore, we need to impose consistency conditions so that the transition amplitudes do not depend on a decomposition.
Such consistency conditions can be listed as follows \cite{BT2018}:
\begin{description}
\item [Non-degenerate pairing] (figure \ref{fig: non-degenerate})\\
\begin{figure}
\begin{center}
\includegraphics[width = 3.6cm]{non-degenerate.pdf}
\caption{The composition of the multiplication and the counit defines a non-degenerate pairing, which can be used to regard $V_{x^*}$ as the dual vector space of $V_x$.}
\label{fig: non-degenerate}
\end{center}
\end{figure}
\begin{equation}
\text{The pairing }\epsilon \circ Z(\mathrm{ev}_x^L) \circ M_{x^*, x}: V_{x^*} \otimes V_x \rightarrow \mathbb{C}\text{ is non-degenerate}.
\tag{O6}
\label{eq: non-degenerate}
\end{equation}
This non-degenerate pairing is related to the inner product of the Hilbert space via $CPT$ conjugation \cite{KT2017, KTY2017}.
\item [Unit constraint] (figure \ref{fig: unit constraint})\\
\begin{figure}
\begin{center}
\includegraphics[width = 8cm]{unit_constraint.pdf}
\caption{The multiplication $M$ has the unit $i(1) \in V_1$.}
\label{fig: unit constraint}
\end{center}
\end{figure}
The unit $i$ acts as the unit of the multiplication $M$:
\begin{equation}
M_{1x}(i(1) \otimes v_x) = M_{x1}(v_x \otimes i(1)) = v_x, \quad \forall v_x \in V_x.
\tag{O7}
\label{eq: unit constraint}
\end{equation}
\item [Associativity] (figure \ref{fig: associativity constraint})\\
\begin{figure}
\begin{center}
\includegraphics[width = 8cm]{associativity_constraint.pdf}
\caption{The multiplication $M$ is associative up to associator. The topological point operators corresponding to associators are omitted in the above figure.}
\label{fig: associativity constraint}
\end{center}
\end{figure}
The multiplication $M$ is associative up to associator:
\begin{equation}
Z(\alpha_{xyz}) \circ M_{x \otimes y, z} \circ (M_{xy} \otimes \mathrm{id}_{V_z}) = M_{x, y \otimes z} \circ (\mathrm{id}_{V_x} \otimes M_{yz}) \circ \alpha_{V_x V_y V_z}^{\mathrm{Vec}},
\tag{O8}
\label{eq: associativity constraint}
\end{equation}
where $\alpha_{V_x V_y V_z}^{\mathrm{Vec}}$ is the associator of the category of vector spaces $\mathrm{Vec}$, which is chosen to be the identity map $\alpha_{V_x V_y V_z}^{\mathrm{Vec}} = \mathrm{id}_{V_x \otimes V_y \otimes V_z}$ by identifying $(V_x \otimes V_y) \otimes V_z$ with $V_x \otimes (V_y \otimes V_z)$.
\item [Twisted commutativity] (figure \ref{fig: twisted commutativity})\\
\begin{figure}[t]
\begin{center}
\includegraphics[width = 6cm]{twisted_commutativity.pdf}
\caption{The multiplication $M$ is commutative up to isomorphism $X_{xy}$.}
\label{fig: twisted commutativity}
\end{center}
\end{figure}
The multiplication $M$ is commutative up to the change of the base point:
\begin{equation}
M_{yx} \circ c_{V_x V_y} = X_{xy} \circ M_{xy},
\tag{O9}
\label{eq: twisted commutativity}
\end{equation}
where $c_{V_x V_y}: V_x \otimes V_y \rightarrow V_y \otimes V_x$ is the symmetric braiding of $\mathrm{Vec}$, which is defined by $c_{V_x V_y}(v_x \otimes v_y) := v_y \otimes v_x$ for any $v_x \in V_x$ and $v_y \in V_y$.
\item [Commutativity of topological point operators and the multiplication] (figure \ref{fig: commutativity of a morphism and the multiplication})\\
\begin{figure}
\begin{center}
\includegraphics[width = 6cm]{commutativity_of_f_and_M.pdf}
\caption{The transition amplitude does not depend on whether a topological point operator $f \in \mathop{\mathrm{Hom}}(x, x^{\prime})$ is inserted before or after the multiplication $M_{xy}$. The same holds for $g \in \mathop{\mathrm{Hom}}(y, y^{\prime})$.}
\label{fig: commutativity of a morphism and the multiplication}
\end{center}
\end{figure}
Topological point operators commute with the multiplication $M$:
\begin{equation}
\begin{aligned}
M_{x^{\prime}y} \circ (Z(f) \otimes \mathrm{id}_{V_y}) & = Z(f \otimes \mathrm{id}_y) \circ M_{xy}, \quad \forall f \in \mathop{\mathrm{Hom}}(x, x^{\prime}),\\
M_{xy^{\prime}} \circ (\mathrm{id}_{V_x} \otimes Z(g)) & = Z(\mathrm{id}_x \otimes g) \circ M_{xy}, \quad \forall g \in \mathop{\mathrm{Hom}}(y, y^{\prime}).
\end{aligned}
\tag{O10}
\label{eq: commutativity of a morphism and the multiplication}
\end{equation}
\item [The uniqueness of the multiplication] (figure \ref{fig: uniqueness of the multiplication})\\
\begin{figure}[t]
\begin{center}
\includegraphics[width = 6cm]{uniqueness_of_M.pdf}
\caption
The auxiliary line does not intersect with topological defects on the left-hand side, while it intersects with topological defect $z$ three times on the right-hand side.
They give the same transition amplitude.}
\label{fig: uniqueness of the multiplication}
\end{center}
\end{figure}
The multiplication $M$ uniquely determines a map for topologically equivalent configurations of topological defects on the pants diagram:
\begin{equation}
\begin{aligned}
& Z((\mathrm{id}_x \otimes \mathrm{ev}_z^L) \otimes \mathrm{id}_{y^*}) \circ \mathcal{A}_{(x \otimes z^*) \otimes (z \otimes y^*) \rightarrow (x \otimes (z^* \otimes z)) \otimes y^*} \circ M_{x \otimes z^*, z \otimes y^*}\\
= ~ & Z(\mathrm{id}_{x \otimes y^*} \otimes \mathrm{ev}_z^R) \circ \mathcal{A}_{(z^* \otimes x) \otimes (y^* \otimes z) \rightarrow (x \otimes y^*) \otimes (z \otimes z^*)} \circ M_{z^* \otimes x, y^* \otimes z} \circ (X_{x, z^*} \otimes X_{z, y^*}).
\end{aligned}
\tag{O11}
\label{eq: uniqueness of the multiplication}
\end{equation}
\item [Consistency on the torus] (figure \ref{fig: consistency on the torus})\\
\begin{figure}
\begin{center}
\includegraphics[width = 7cm]{consistency_on_the_torus.pdf}
\caption{The transition amplitude for a twice-punctured torus does not depend on a time function on it. The blue lines represent the initial time-slice and the red lines represent the final time-slice. A time-slice at an intermediate time, which consists of two disjoint circles, is represented by the green lines. Two different ways to split the initial circle into the intermediate circles give rise to the same transition amplitude. The auxiliary lines are omitted in the above figure.}
\label{fig: consistency on the torus}
\end{center}
\end{figure}
Different ways to decompose a twice-punctured torus give the same transition amplitude:
\begin{equation}
\begin{aligned}
& \mathcal{A}_{(y \otimes x) \otimes (w \otimes z) \rightarrow (x \otimes w) \otimes (z \otimes y)} \circ M_{y \otimes x, w \otimes z} \circ (X_{xy} \otimes X_{zw}) \circ \Delta_{x \otimes y, z \otimes w}\\
= ~ & M_{x \otimes w, z \otimes y} \circ (X_{wx} \otimes X_{yz}) \circ \Delta_{w \otimes x, y \otimes z} \circ \mathcal{A}_{(x \otimes y) \otimes (z \otimes w) \rightarrow (w \otimes x) \otimes (y \otimes z)}.
\end{aligned}
\tag{O12}
\label{eq: consistency on the torus}
\end{equation}
\end{description}
In summary, a two-dimensional bosonic oriented TQFT with fusion category symmetry is given by a quadruple $(Z, X, M, i)$ that satisfies the consistency conditions (\ref{eq: composition on a cylinder})--(\ref{eq: consistency on the torus}).\footnote{In \cite{BT2018}, there is an additional consistency condition called the cyclic symmetry of the multiplication. However, this condition turns out to be satisfied as a consequence of (\ref{eq: non-degenerate}) and (\ref{eq: associativity constraint}).}
\subsection{Bosonic fusion category SPT phases without time-reversal symmetry}
\label{sec: Bosonic fusion category SPT phases without time-reversal symmetry}
In this section, we classify bosonic SPT phases with fusion category symmetry $\mathcal{C}$ by solving the consistency conditions (\ref{eq: composition on a cylinder})--(\ref{eq: consistency on the torus}).
To focus on the solutions that describe SPT phases, we further impose the condition that the ground state is unique on a circle without topological defects, i.e. $\mathop{\mathrm{dim}} V_1 = 1$.
More generally, the dimension of the vector space $V_x$ for an SPT phase is equal to the quantum dimension of the object $x$
\begin{equation}
\mathop{\mathrm{dim}} V_x = \mathop{\mathrm{dim}} x,
\label{eq: dim = qdim}
\end{equation}
which can be seen from modular invariance of the partition function on a torus \cite{CLSWY2019}.
\begin{comment}
If the topological defect $x$ is put on a temporal cycle of a torus, the partition function is given by the trace of the identity operator over the defect vector space, which means that the partition function is equal to $\mathop{\mathrm{dim}} V_x$.
On the other hand, if the topological defect $x$ is put on a spatial cycle, the partition function is given by the trace of the topological line operator $U_x$ over the vector space $V_1$.
In particular, for SPT phases, the action of $U_x$ on $V_1$ agrees as the quantum dimension of $x$.
Hence, the partition function is equal to $\mathop{\mathrm{dim}} x$.
Since the temporal cycle and the spacial cycle are exchanged by the modular $S$ transformation, modular invariance implies that the above two partition functions are the same, see figure \ref{fig: modular invariance}.
\begin{figure}
\begin{center}
\includegraphics[width = 6cm]{modular_invariance.pdf}
\caption{For SPT phases, the dimension of the defect vector space $V_x$ agrees with the quantum dimension of the topological defect $x$ due to modular invariance.}
\label{fig: modular invariance}
\end{center}
\end{figure}
This shows eq. (\ref{eq: dim = qdim}).
\end{comment}
For SPT phases, the unit $i: \mathbb{C} \rightarrow V_1$ is an isomorphism because it is a non-zero linear map between one-dimensional vector spaces.
In particular, $i$ is unitary if the partition function on a sphere is unity:
\begin{equation}
i^{\dagger} \circ i = 1.
\label{eq: unitary unit}
\end{equation}
In the following, we assume that this condition is satisfied.
Furthermore, we assume that the multiplication $M_{xy}: V_{x} \otimes V_y \rightarrow V_{x \otimes y}$ is also unitary:
\begin{equation}
M_{xy}^{\dagger} = M_{xy}^{-1} \Leftrightarrow \adjincludegraphics[valign = c, width = 1.4cm]{Delta_M.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.4cm]{identity.pdf}.
\label{eq: unitary M}
\end{equation}
This equation reduces to eq. (\ref{eq: dim = qdim}) if we form a loop of $x$ by choosing $y = x^{*}$ and close the punctures with a cap and its adjoint.
Now, we solve the consistency conditions (\ref{eq: composition on a cylinder})--(\ref{eq: consistency on the torus}) under the constraints (\ref{eq: unitary unit}) and (\ref{eq: unitary M}).
We first notice that $Z: \mathcal{C} \rightarrow \mathrm{Vec}$ is a functor by (\ref{eq: composition on a cylinder}), where an object $x$ is mapped to a vector space $V_x$ and a morphism $f \in \mathop{\mathrm{Hom}}(x, y)$ is mapped to a linear map $Z(f): V_x \rightarrow V_y$.
We recall that the functor $Z$ is additive in objects in the sense that the direct sum of objects $x$ and $y$ is mapped to the direct sum of vector spaces $V_x$ and $V_y$.
This functor is also $\mathbb{C}$-linear in morphisms due to (\ref{eq: C-linear}).
Furthermore, (\ref{eq: unit constraint}) and (\ref{eq: associativity constraint}) indicate that the triple $(Z, M, i)$ is a monoidal functor where the naturality of the isomorphism $M_{xy}$ follows from (\ref{eq: commutativity of a morphism and the multiplication}).
Thus, $(Z, M, i)$ is a tensor functor from $\mathcal{C}$ to $\mathrm{Vec}$, namely a fiber functor.
In this way, we can extract the data of a fiber functor from an SPT phase with fusion category symmetry.
Conversely, we can show that the data of a fiber functor $(Z, M, i)$ is sufficient to construct a solution of (\ref{eq: composition on a cylinder})--(\ref{eq: consistency on the torus}) as we will see below.
In other words, the other consistency conditions are automatically satisfied when $(Z, M, i)$ is a fiber functor.
Since $M_{xy}$ is unitary (\ref{eq: unitary M}), the isomorphism $X_{xy}$ is determined by (\ref{eq: twisted commutativity}) as
\begin{equation}
X_{xy} = M_{yx} \circ c_{V_x V_y} \circ M_{xy}^{-1}.
\label{eq: X = McM^{-1}}
\end{equation}
This satisfies (\ref{eq: C2}) because $c_{V_x V_y}$ is the symmetric braiding of $\mathrm{Vec}$.
Furthermore, the naturality of $M_{xy}$ and $c_{V_x V_y}$ implies that $X_{xy}$ is also natural (\ref{eq: C4}).
The non-degeneracy of the pairing (\ref{eq: non-degenerate}) is guaranteed by the equality
\begin{equation}
\adjincludegraphics[valign = c, width = 1.25cm]{non-degenerate1.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 2.1cm]{non-degenerate2.pdf},
\end{equation}
which follows from the Frobenius relation represented by the following commutative diagram:
\begin{equation}
\begin{tikzcd}[row sep = 12pt]
V_1 \otimes V_x \arrow[r, "M_{1x}"] \arrow[dd, "Z(\mathrm{coev}_x^L)"'] & V_{1 \otimes x} \arrow[r, "Z(l_x)"] \arrow[rd, "Z(\mathrm{coev}_x^L \otimes \mathrm{id}_x)"'] & V_x \arrow[r, "Z(r_x^{-1})"] & V_{x \otimes 1} \arrow[r, "M_{x1}^{-1}"] & V_x \otimes V_1\\
& & V_{x \otimes x^* \otimes x} \arrow[ru, "Z(\mathrm{id}_x \otimes \mathrm{ev}_x^L)"'] & & \\
V_{x \otimes x^*} \otimes V_x \arrow[rru, "M_{x \otimes x^*, x}"] \arrow[rr, "M_{x, x^*}^{-1}"'] & & V_x \otimes V_{x^*} \otimes V_x \arrow[rr, "M_{x^*, x}"'] & & V_x \otimes V_{x^* \otimes x} \arrow[llu, "M_{x, x^* \otimes x}"'] \arrow[uu, "Z(\mathrm{ev}_x^L)"']
\end{tikzcd}
\end{equation}
To proceed further, we notice that a linear map from $V_{x_1 \otimes y_1 \otimes \cdots} \otimes V_{x_2 \otimes y_2 \otimes \cdots} \otimes \cdots$ to $V_{x^{\prime}_1 \otimes y^{\prime}_1 \otimes \cdots} \otimes V_{x^{\prime}_2 \otimes y^{\prime}_2 \otimes \cdots} \otimes \cdots$ that consists only of $M$ and $c$ is uniquely determined by the permutation of the topological defect $(x_1, y_1, \cdots, x_2, y_2, \cdots) \rightarrow (x^{\prime}_1, y^{\prime}_1, \cdots, x^{\prime}_2, y^{\prime}_2, \cdots)$ and the change of the parentheses.
The same holds for a linear map that consists only of $M$, $c$, $Z(\alpha)$, and $X$ because $Z(\alpha)$ and $X$ can be expressed in terms of $M$ and $c$ due to (\ref{eq: associativity constraint}) and (\ref{eq: twisted commutativity}).
This shows (\ref{eq: coherence}) and (\ref{eq: consistency on the torus}).
This also reduces the consistency condition (\ref{eq: uniqueness of the multiplication}) to the following commutative diagram:
\begin{equation}
\begin{tikzcd}
& V_1 &\\
V_{z^* \otimes z} \arrow[ru, "Z(\mathrm{ev}_z^L)"] & \mathbb{C} \arrow[u, "i"] & V_{z \otimes z^*} \arrow[lu, "Z(\mathrm{ev}_z^R)"'] \\
V_{z^*} \otimes V_z \arrow[u, "M_{z^*, z}"] \arrow[ru, "\mathrm{ev}_{V_z}^L"] \arrow[rr, "c_{V_{z^*} V_z}"'] & & V_z \otimes V_{z^*} \arrow[lu, "\mathrm{ev}_{V_z}^R"'] \arrow[u, "M_{z, z^*}"']
\end{tikzcd}
\label{eq: C10 commutative diagram}
\end{equation}
The commutativity of the left square in the above diagram defines the evaluation morphism $\mathrm{ev}_{V_z}^L: V_{z^{*}} \otimes V_z \rightarrow \mathbb{C}$, which is the non-degenerate pairing (\ref{eq: non-degenerate}).
Accordingly, we can consider $V_{z^*}$ as the dual vector space of $V_z$.
If we denote a basis of $V_z$ as $\{ e_i \mid i = 1, 2, \cdots, \mathop{\mathrm{dim}} V_z \}$, the dual basis $\{ e^i \mid i = 1, 2, \cdots, \mathop{\mathrm{dim}} V_{z^*} = \mathop{\mathrm{dim}} V_z \}$ of $V_{z^*}$ is defined by
\begin{equation}
\mathrm{ev}_{V_z}^L(e^i \otimes e_j) = \delta^i_j,
\label{eq: ev dual basis}
\end{equation}
where $\delta^i_j$ is the Kronecker delta.
Equation (\ref{eq: zigzag axiom}) implies that the coevaluation morphism $\mathrm{coev}_{V_z}^L: \mathbb{C} \rightarrow V_z \otimes V_{z^*}$ is given by the embedding
\begin{equation}
\mathrm{coev}_{V_z}^L (1) = \sum_{i = 1}^{\mathop{\mathrm{dim}} V_z} e_i \otimes e^i,
\label{eq: coev dual basis}
\end{equation}
whose adjoint agrees with the right evaluation morphism $\mathrm{ev}_{V_z}^R$ defined by the commutativity of the right square in the diagram (\ref{eq: C10 commutative diagram}):
\begin{equation}
\mathrm{ev}_{V_z}^R (e_i \otimes e^j) = \delta_i^j.
\label{eq: ev R dual basis}
\end{equation}
From eqs. (\ref{eq: ev dual basis}) and (\ref{eq: ev R dual basis}), we find that the bottom triangle in (\ref{eq: C10 commutative diagram}) is commutative.
In summary, the solutions of the consistency conditions (\ref{eq: composition on a cylinder})--(\ref{eq: consistency on the torus}) for fusion category SPT phases are in one-to-one correspondence with fiber functors of the fusion category $\mathcal{C}$.
To classify fusion category SPT phases, we regard naturally isomorphic fiber functors as the same SPT phase because a natural isomorphism between fiber functors corresponds to a change of the bases of the vector spaces.
Therefore, bosonic SPT phases with fusion category symmetry $\mathcal{C}$ are classified by isomorphism classes of fiber functors of $\mathcal{C}$.
\section{Unoriented TQFTs with fusion category symmetry}
\label{sec: Unoriented TQFTs with fusion category symmetry}
\subsection{Consistency conditions of unoriented TQFTs}
\label{sec: Consistency conditions of unoriented TQFTs}
A two-dimensional unoriented TQFT with finite group symmetry is formulated in \cite{KT2017, Swe2013} as a set of consistency conditions on algebraic data.\footnote{A two-dimensional unoriented TQFT without internal symmetry is formulated in \cite{TT2006}. Quantum field theories on unoriented manifolds are of interest in recent studies on topological phases of matter \cite{HSCRL2014, CHMR2015, HCR2016, CTR2016, BBCJW2019, Wit2016c, TY2017a, TY2017b, Bha2017, SR2017, SSR2017, SSR2017b, SSGR2018, Kob2019, Tur2020}.}
In this section, we reformulate 2d unoriented TQFTs to incorporate fusion category symmetries.
We assume that the total symmetry splits into time-reversal symmetry and finite internal symmetry.
In this case, we can treat orientation-reversing defects and symmetry defects separately.
The algebraic data of an unoriented TQFT consist of the following linear maps in addition to the algebraic data of an oriented TQFT: the orientation-reversing isomorphism $\phi_x: V_x \rightarrow V_{x^{*}}$ and the cross-cap amplitude $\theta_{x, y}: V_{x \otimes y} \rightarrow V_{x^{*} \otimes y}$, see also figure \ref{fig: unoriented amplitudes}.
\begin{figure}
\begin{center}
\includegraphics[width = 5cm]{unoriented_amplitudes.pdf}
\caption{Every unoriented surface can be decomposed into oriented surfaces and the above two unoriented surfaces. The left diagram represents the orientation-reversing diffeomorphism of a circle, which induces an isomorphism between the vector spaces $V_x$ and $V_{x^{*}}$. The right diagram represents a cylinder with a cross-cap, which induces a linear map from $V_{x \otimes y}$ to $V_{x^{*} \otimes y}$.}
\label{fig: unoriented amplitudes}
\end{center}
\end{figure}
In particular, $\phi$ represents the linear action of $CP$ symmetry, which we assume to be unitary, rather than the anti-linear action of time-reversal symmetry.
We can avoid anti-linear maps by using $CPT$ symmetry \cite{KT2017}.
A variant of the cross-cap amplitude as shown in figure \ref{fig: variant of theta} can be computed by composing the cross-cap amplitude $\theta$ and the generalized associators $\mathcal{A}$.
\begin{figure}
\begin{center}
\includegraphics[width = 1.8cm]{variant_of_theta.pdf}
\caption{The transition amplitude for the above cylinder with a cross-cap is a linear map from $V_{(x \otimes y) \otimes z}$ to $V_{(x \otimes y^*) \otimes z}$, which is given by $\mathcal{A}_{y^* \otimes (z \otimes x) \rightarrow (x \otimes y^*) \otimes z} \circ \theta_{y, z \otimes x} \circ \mathcal{A}_{(x \otimes y) \otimes z \rightarrow y \otimes (z \otimes x)}$.}
\label{fig: variant of theta}
\end{center}
\end{figure}
A general unoriented surface can be decomposed into the following building blocks and their adjoints: a cylinder, a cap, a pair of pants, and a cylinder with a cross-cap.
Correspondingly, the transition amplitude for a general unoriented surface is given by the composition of the transition amplitudes for these elements that appear in the decomposition.
However, the decomposition is not unique in general.
Therefore, we need to impose consistency conditions so that the transition amplitude does not depend on a decomposition.
In the following, we list the consistency conditions that involve $\phi$ and $\theta$.
\begin{description}
\item[Invariance of the unit]~\\
The unit $i: \mathbb{C} \rightarrow V_1$ is invariant under the action of the orientation-reversing isomorphism
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U0-LHS.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 1.5cm]{U0-RHS.pdf}.
\tag{U1}
\label{eq: invariant unit}
\end{equation}
\item[Involution]~\\
The orientation-reversing isomorphism is involutive up to pivotal structure $a_x: x \rightarrow x^{**}$
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U1-LHS.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{U1-RHS.pdf}.
\tag{U2}
\label{eq: involution}
\end{equation}
\item[Commutativity of the orientation reversal and the change of the base point] ~\\
The orientation-reversing isomorphism commutes with the change of the base point:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U2-LHS.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{U2-RHS.pdf}.
\tag{U3}
\label{eq: commutativity of X and phi}
\end{equation}
\item[Orientation reversal of a topological point operator] ~\\
The orientation-reversing symmetry acts on a topological point operator $f \in \mathop{\mathrm{Hom}}(x, y)$ and turn it into another topological point operator $\overline{f} \in \mathop{\mathrm{Hom}}(x^*, y^*)$ such that
\begin{equation}
\adjincludegraphics[valign = c, width = 1.35cm]{U3-LHS1.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.35cm]{U3-RHS1.pdf}, \quad
\adjincludegraphics[valign = c, width = 1.5cm]{U3-LHS2.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{U3-RHS2.pdf}.
\tag{U4}
\label{eq: orientation reversal of f}
\end{equation}
\item[Cross-cap on a folded topological defect] ~\\
We can move the position of a cross-cap along a folded topological defect line:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U5-LHS.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{U5-RHS.pdf}.
\tag{U5}
\label{eq: compatibility of left-theta and right-theta}
\end{equation}
\item[Deformation of a topological defect through a cross-cap] ~\\
A topological defect that passes a cross-cap twice can avoid the cross-cap by a deformation:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U6-LHS0.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{U6-RHS0.pdf}, \quad \quad
\adjincludegraphics[valign = c, width = 1.7cm]{U6-LHS1.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{U6-RHS1.pdf}.
\tag{U6}
\label{eq: reconnection}
\end{equation}
We also have similar equations for the diagrams flipped horizontally.
\item[Commutativity of the cross-cap and the orientation reversal] ~\\
The orientation-reversing isomorphism commutes with the cross-cap amplitude:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.35cm]{U7-LHS.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.35cm]{U7-RHS.pdf}.
\tag{U7}
\label{eq: compatibility of theta and phi}
\end{equation}
\item[M\"{o}bius identity] ~\\
Two different ways to compute the transition amplitude for the punctured M\"{o}bius strip lead to the same result:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U8-LHS.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.35cm]{U8-RHS.pdf}.
\tag{U8}
\label{eq: Mobius}
\end{equation}
\item[Equivariance of the multiplication] ~\\
The multiplication is equivariant with respect to the orientation-reversing isomorphism:
\begin{equation}
\adjincludegraphics[valign = c, width = 2.25cm]{U10-LHS.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 2.25cm]{U10-RHS.pdf}.
\tag{U9}
\label{eq: equivariance}
\end{equation}
\item[Compatibility of the cross-cap amplitude and the multiplication] ~\\
We can move the position of a cross-cap through the pants diagram:
\begin{equation}
\adjincludegraphics[valign = c, width = 2.25cm]{U11-LHS1.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 2.25cm]{U11-RHS1.pdf}, \quad
\adjincludegraphics[valign = c, width = 2.25cm]{U11-LHS2.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 2.25cm]{U11-RHS2.pdf}.
\tag{U10}
\label{eq: compatibility of theta and M}
\end{equation}
\item[Klein identity] ~\\
Two different decompositions of the twice-punctured Klein bottle give rise to the same transition amplitude:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U12-LHS.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 2cm]{U12-RHS.pdf}.
\tag{U11}
\label{eq: Klein}
\end{equation}
\end{description}
We need to check the consistency of the assignment of the transition amplitudes.
There are three types of consistency as follows:
\begin{enumerate}
\item consistency under changes of the time function,
\item consistency under deformations of topological defects on each basic element, and
\item consistency under the composition of the orientation-reversing isomorphism.
\end{enumerate}
Let us begin with the consistency of type 1.
The transition amplitudes must be invariant under changes of the time function.
As discussed in the appendix of \cite{KT2017}, we have two consistency conditions of this type for unoriented surfaces: one is the consistency on a pair of pants with a cross-cap and the other is the consistency on a twice-punctured Klein bottle.
These consistency conditions are given by (\ref{eq: compatibility of theta and M}) and (\ref{eq: Klein}) respectively.
Next, we consider the consistency of type 2.
Topologically equivalent configurations of topological defects on a basic element give rise to the same transition amplitude.
For this type of consistency conditions, it suffices to consider the consistency on a cylinder with a cross-cap because the consistency on the other basic elements is already considered in oriented TQFTs.
The consistency on a cylinder with a cross-cap is given by (\ref{eq: compatibility of left-theta and right-theta}) and (\ref{eq: reconnection}).
Finally, we have the consistency conditions of type 3 for each basic element as follows: (\ref{eq: involution}) and (\ref{eq: commutativity of X and phi}) for a cylinder, (\ref{eq: invariant unit}) for a cap, (\ref{eq: equivariance}) for a pair of pants, and (\ref{eq: compatibility of theta and phi}) and (\ref{eq: Mobius}) for a cylinder with a cross-cap.
The remaining equation (\ref{eq: orientation reversal of f}) is the consistency of topological point operators.
This completes the consistency conditions on transition amplitudes for unoriented surfaces.
Therefore, a two-dimensional bosonic unoriented TQFT with fusion category symmetry $\mathcal{C}$ is described by a sextuple $(Z, X, M, i, \phi, \theta)$ that satisfies (\ref{eq: composition on a cylinder})--(\ref{eq: consistency on the torus}) and (\ref{eq: invariant unit})--(\ref{eq: Klein}).
As we will see in appendix \ref{sec: Unoriented equivariant TQFTs}, this reduces to an unoriented equivariant TQFT when the symmetry is given by a finite group $\mathbb{Z}_2^T \times G$.
\subsection{Bosonic fusion category SPT phases with time-reversal symmetry}
\label{sec: Bosonic fusion category SPT phases with time-reversal symmetry}
In this section, we classify fusion category SPT phases with time-reversal symmetry by solving the consistency conditions (\ref{eq: invariant unit})--(\ref{eq: Klein}) under the constraints (\ref{eq: unitary unit}) and (\ref{eq: unitary M}).
We first recall that the triple $(Z, M, i): \mathcal{C} \rightarrow \mathrm{Vec}$ is a fiber functor and $X$ is given by eq. (\ref{eq: X = McM^{-1}}).
In particular, the multiplication $M_{x, y}$ is unitary.
Therefore, the cross-cap amplitude $\theta_{x, y}$ is determined by $\theta_{x, 1}$ via eq. (\ref{eq: compatibility of theta and M}) as follows:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{theta-L2.pdf} ~ = ~ \adjincludegraphics[valign = c, width =2cm]{M-theta-M.pdf}.
\label{eq: M-theta-M}
\end{equation}
Conversely, the consistency condition (\ref{eq: compatibility of theta and M}) follows from the above equation.
Furthermore, the Klein identity (\ref{eq: Klein}) reduces to
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{phi2.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{theta-theta.pdf}.
\label{eq: reduced Klein}
\end{equation}
This implies that the cross-cap amplitude $\theta_{x, y}$ is an isomorphism.
By using eqs. (\ref{eq: M-theta-M}) and (\ref{eq: reduced Klein}), the equivariance (\ref{eq: equivariance}) can be simplified as follows:
\begin{equation}
\adjincludegraphics[valign = c, width = 2.2cm]{equivariance1.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 2.2cm]{equivariance2.pdf}
\Leftrightarrow
\adjincludegraphics[valign = c, width = 1.5cm]{equivariance5.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{equivariance6.pdf}
\Leftrightarrow
\adjincludegraphics[valign = c, width = 1.5cm]{equivariance7.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{equivariance8.pdf}.
\label{eq: involutive cross-cap}
\end{equation}
More generally, we obtain
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{involution1.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 1.5cm]{involution2.pdf},
\label{eq: involutive cross-cap2}
\end{equation}
from (\ref{eq: involution}).
The above equations indicate that the orientation-reversing isomorphism acts trivially on the vector space $V_1$, which shows (\ref{eq: invariant unit}).
As we will see below, it turns out that the remaining consistency conditions (\ref{eq: commutativity of X and phi})--(\ref{eq: Mobius}) follow from eqs. (\ref{eq: M-theta-M}), (\ref{eq: reduced Klein}), (\ref{eq: involutive cross-cap2}) and the consistency conditions of oriented TQFTs.
We first notice that (\ref{eq: commutativity of X and phi}), (\ref{eq: orientation reversal of f}), (\ref{eq: compatibility of theta and phi}), and (\ref{eq: Mobius}) immediately follow from eqs. (\ref{eq: M-theta-M}), (\ref{eq: reduced Klein}), and (\ref{eq: involutive cross-cap2}).
To show the consistency condition (\ref{eq: compatibility of left-theta and right-theta}), it is convenient to compose the cross-cap amplitude to both sides:
\begin{equation}
\adjincludegraphics[valign = c, width = 1.5cm]{U5-1.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 1.5cm]{U5-2.pdf}
\label{eq: U5 prime}
\end{equation}
This equation is equivalent to (\ref{eq: compatibility of left-theta and right-theta}) because the cross-cap amplitude is an isomorphism.
The left-hand side can be written as
\begin{equation}
\mathrm{(LHS)} = \adjincludegraphics[valign = c, width = 1.5cm]{U5-3.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 1.5cm]{U5-4.pdf},
\label{eq: U5 LHS}
\end{equation}
where the second equality follows from the relation (\ref{eq: canonical}) between the left and right coevaluation morphisms.
On the other hand, the right-hand side can be computed as
\begin{equation}
\mathrm{(RHS)} = \adjincludegraphics[valign = c, width = 1.5cm]{U5-5.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 1.5cm]{U5-6.pdf} ~ = ~
\adjincludegraphics[valign = c, width = 1.5cm]{U5-4.pdf} = \mathrm{(LHS)}.
\label{eq: U5 RHS}
\end{equation}
The second equality follows from the fact that the orientation reversal of the right coevaluation morphism $\mathrm{coev}^R_{x}$ is given by $\overline{\mathrm{coev}^R_{x}} = \mathrm{coev}^L_{x^*}$, which can be seen from the unitarity of $\phi_x$ and the first equation of (\ref{eq: orientation reversal of f}).
Indeed, when $\phi_x$ is unitary, the orientation reversal of the coevaluation morphism defined by the first equation of (\ref{eq: orientation reversal of f}) satisfies
\begin{equation}
\adjincludegraphics[valign = c, width = 1.35cm]{coev_bar1.pdf} ~ = ~ \adjincludegraphics[valign = c, width = 2.25cm]{coev_bar2.pdf},
\end{equation}
which implies $\overline{\mathrm{coev}^R_{x}} = \mathrm{coev}^L_{x^*}$ and $\overline{\mathrm{coev}^L_{x}} = \mathrm{coev}^R_{x^*}$.
We note that this shows the first equation of (\ref{eq: reconnection}).
Finally, we consider the second equation of (\ref{eq: reconnection}).
The left-hand side and the right-hand side can be written as
\begin{equation}
\mathrm{(LHS)} = \adjincludegraphics[valign = c, width = 2.25cm]{U6-LHS3.pdf}, \quad \quad
\mathrm{(RHS)} = \adjincludegraphics[valign = c, width = 1.35cm]{U6-RHS3.pdf}.
\end{equation}
We notice that both the left-hand side and the right-hand side have the orientation-reversing isomorphism and the cross-cap amplitude in common, which can be eliminated from the equality $(\mathrm{LHS}) = (\mathrm{RHS})$ because they are isomorphisms.
Hence, this equation reduces to the equality of the transition amplitudes for oriented surfaces.
By considering the commutativity of diagrams that consist only of $M$, $c$, $Z(\alpha)$, and $X$ as we discussed in section \ref{sec: Bosonic fusion category SPT phases without time-reversal symmetry}, we find that the second equation of (\ref{eq: reconnection}) is equivalent to the following commutative diagram:
\begin{equation}
\begin{tikzcd}[column sep = large, row sep = 12pt]
V_{x \otimes y} \arrow[rr, "X_{x, y}"] \arrow[dd, "Z(\mathrm{coev}_y^R)"'] \arrow[dr, "Z(\mathrm{coev}_y^R)"'] & & V_{y \otimes x}\\
& V_{x \otimes y \otimes y^{*} \otimes y} \arrow[dr, "X_{x, y \otimes y^* \otimes y}"] & \\
V_{y^* \otimes y \otimes x \otimes y} \arrow[rr, "X_{y^* \otimes y \otimes x, y}"'] \arrow[ur, "X_{y^* \otimes y, x \otimes y}"] & & V_{y \otimes y^* \otimes y \otimes x} \arrow[uu, "Z(\mathrm{ev}_y^R)"']
\end{tikzcd}
\end{equation}
The transition amplitudes for the surfaces flipped horizontally satisfy a similar equation.
Thus, the solution of the consistency conditions (\ref{eq: composition on a cylinder})--(\ref{eq: consistency on the torus}) and (\ref{eq: invariant unit})--(\ref{eq: Klein}) for an SPT phase with symmetry $\mathbb{Z}_2^T \times \mathcal{C}$ is given by a quintuple $(Z, M, i, \phi, \theta)$ such that the triple $(Z, M, i)$ is a fiber functor of $\mathcal{C}$, the isomorphism $\theta_{x, y}: V_{x \otimes y} \rightarrow V_{x^* \otimes y}$ is given by $\theta_{x, y} = M_{x^*, y} \circ (\theta_{x, 1} \otimes \mathrm{id}_{V_y}) \circ M_{x, y}^{-1}$ where $\theta_{x, 1}$ satisfies $\theta_{x^*, 1} \circ \theta_{x, 1} = Z(a_x)$, and the isomorphism $\phi_{x}: V_{x} \rightarrow V_{x^*}$ is given by $\phi_{x \otimes y} = \theta_{y, x^*} \circ X_{x^*, y} \circ \theta_{x, y}$ where $X_{x, y}$ is defined by eq. (\ref{eq: X = McM^{-1}}).
The quintuple $(Z, M, i, \phi, \theta)$ that satisfies the above conditions is equivalent to the following algebraic data $(Z, M, i, s, \phi)$:
\begin{itemize}
\item a fiber functor $(Z, M, i): \mathcal{C} \rightarrow \mathrm{Vec}$,
\item a sign $s \in \mathbb{Z}_2 = \{ \pm 1\}$,
\item an isomorphism $\phi_x: V_x \rightarrow V_{x^*}$ that is involutive up to pivotal structure $a_x$ and commutes with the multiplication $M$ up to isomorphism $X$ defined by eq. (\ref{eq: X = McM^{-1}}):
\begin{equation}
\phi_{x^*} \circ \phi_x = Z(a_x), \quad
X_{x^*, y^*} \circ M_{x^*, y^*} \circ (\phi_x \otimes \phi_y) = \phi_{x \otimes y} \circ M_{x, y}.
\label{eq: involutive equivariance}
\end{equation}
\end{itemize}
When we have a quintuple $(Z, M, i, \phi, \theta)$, the sign $s \in \mathbb{Z}_2$ and the isomorphism that satisfies eq. (\ref{eq: involutive equivariance}) are given by $\theta_{1, 1}$ and $\phi$ respectively.
Conversely, when we have a quintuple $(Z, M, i, s, \phi)$, we can define the cross-cap amplitude as $\theta_{x, y} := M_{x^*, y} \circ (\theta_{x, 1} \otimes \mathrm{id}_{V_y}) \circ M^{-1}_{x, y}$ where $\theta_{x, 1} := s \phi_x$.
The quintuples $(Z, M, i, s, \phi)$ and $(Z^{\prime}, M^{\prime}, i^{\prime}, s^{\prime}, \phi^{\prime})$ correspond to the same SPT phase if they are related by a change of the bases of the vector spaces. The change of the bases is described by a tensor natural isomorphism $\eta: Z \rightarrow Z^{\prime}$, which guarantees that the transition amplitudes for oriented surfaces transform appropriately under the change of the bases.
For the classification of SPT phases with time-reversal symmetry, we need to impose an additional condition on the tensor natural isomorphism $\eta$ so that the transition amplitudes for unoriented surfaces also transform appropriately.
Specifically, we require that the following diagram commutes:
\begin{equation}
\begin{tikzcd}
Z(x) \arrow[r, "\phi_x"] \arrow[d, "\eta_x"'] & Z(x^{*}) \arrow[d, "\eta_{x^*}"]\\
Z^{\prime}(x) \arrow[r, "\phi^{\prime}_x"'] & Z^{\prime}(x^*)
\end{tikzcd}
\end{equation}
The quintuples $(Z, M, i, s, \phi)$ and $(Z^{\prime}, M^{\prime}, i^{\prime}, s^{\prime}, \phi^{\prime})$ are said to be equivalent when they are related by a tensor natural isomorphism $\eta$ that satisfies the above commutative diagram.
Therefore, bosonic fusion category SPT phases with time-reversal symmetry are classified by equivalence classes of quintuples $(Z, M, i, s, \phi)$.
\subsection{Examples}
\label{sec: Examples (unoriented)}
\subsubsection{Finite group symmetry}
1+1d bosonic SPT phases with finite group symmetry $\mathbb{Z}_2^T \times G$ are classified by the group cohomology $H^2(\mathbb{Z}_2^T \times G, \mathrm{U}_T(1))$ where $\mathrm{U}_T(1)$ is $\mathrm{U}(1)$ on which time-reversal symmetry $\mathbb{Z}_2^T$ acts as the complex conjugation \cite{CGLW2013, KT2017}.
By using a K\"{u}nneth formula, we can factorize the group cohomology $H^2(\mathbb{Z}_2^T \times G, \mathrm{U}_T(1))$ as
\begin{equation}
H^2(\mathbb{Z}_2^T \times G, \mathrm{U}_T(1)) = \mathbb{Z}_2 \times H^1(G, \mathrm{U}(1))/2H^1(G, \mathrm{U}(1)) \times {}_2H^2(G, \mathrm{U}(1)),
\label{eq: Kunneth}
\end{equation}
where $2H^1(G, \mathrm{U}(1))$ is the group of even elements of $H^1(G, \mathrm{U}(1))$ and ${}_2H^2(G, \mathrm{U}(1))$ is the 2-torsion subgroup of $H^2(G, \mathrm{U}(1))$, which consists of elements $x \in H^2(G, \mathrm{U}(1))$ such that $2x = 0$.
We note that the first factor $\mathbb{Z}_2$ captures the classification of 1+1d bosonic SPT phases only with time-reversal symmetry.
In the following, we will reproduce this classification as a special case of the classification of fusion category SPT phases.
We begin with noticing that the sign $s \in \mathbb{Z}_2$ gives the first term on the right-hand side of eq. (\ref{eq: Kunneth}).
Furthermore, for a finite group symmetry $G$,\footnote{As a fusion category symmetry, a finite group symmetry $G$ is described by the category of $G$-graded vector spaces $\mathrm{Vec}_G$.} fiber functors $(Z, M, i)$ are classified by the group cohomology $H^2(G, \mathrm{U}(1))$.
However, some elements of $H^2(G, \mathrm{U}(1))$ are not allowed when the time-reversal symmetry is taken into account.
We will see shortly that $H^2(G, \mathrm{U}(1))$ reduces to its 2-torsion subgroup ${}_2H^2(G, \mathrm{U}(1))$ in the presence of time-reversal symmetry.
The remaining term $H^1(G, \mathrm{U}(1)) / 2H^1(G, \mathrm{U}(1))$ comes from the classification of isomorphisms $\phi$ that satisfy eq. (\ref{eq: involutive equivariance}).
Once we fix a basis $b_g$ of each vector space $V_g$, the isomorphism $\phi_g: V_g \rightarrow V_{g^{-1}}$ can be represented by a phase factor $\phi(g)$ as $\phi_g(b_g) = \phi(g) b_{g^{-1}}$.
For this choice of the basis, equation (\ref{eq: involutive equivariance}) can be written as
\begin{equation}
\phi(g^{-1}) \phi(g) = 1, \quad \phi(g) \phi(h) = \frac{\nu(g, h)}{\nu(h^{-1}, g^{-1})} \phi(gh),
\label{eq: phi equivariance}
\end{equation}
where a 2-cocycle $\nu \in Z^2(G, \mathrm{U}(1))$ represents the multiplication $M_{g, h} (b_g \otimes b_h) = \nu(g, h) b_{gh}$ with respect to the basis $\{b_g\}$.
If we choose the basis so that $\nu(g, h) = \nu(h^{-1}, g^{-1})^{-1}$, the second equation indicates that the cohomology class of $\nu^2$ vanishes.
This means that $\nu$ takes values in ${}_2H^2(G, \mathrm{U}(1))$, which gives the last term on the right-hand side of eq. (\ref{eq: Kunneth}).
Given a 2-cocycle $\nu$, the solutions of eq. (\ref{eq: phi equivariance}) are in one-to-one correspondence with the first group cohomology $H^1(G, \mathrm{U}(1))$.
To see this, we suppose that $\phi_1$ and $\phi_2$ are solutions of eq. (\ref{eq: phi equivariance}).
The ratio $\Phi := \phi_1 / \phi_2$ satisfies
\begin{equation}
\Phi(g) \Phi(h) = \Phi(gh),
\end{equation}
which indicates that $\Phi$ is a homomorphism from $G$ to $\mathrm{U}(1)$.
Therefore, if eq. (\ref{eq: phi equivariance}) has a solution, the solutions of eq. (\ref{eq: phi equivariance}) are in one-to-one correspondence with the set of homomorphisms from $G$ to $\mathrm{U}(1)$, which is the first group cohomology $H^1(G, \mathrm{U}(1))$.
The existence of a solution can be seen by explicitly constructing a solution in the following way \cite{KT2017}.
Let $\widetilde{\nu} \in Z^2(\mathbb{Z}_2^T \times G, \mathrm{U}_T(1))$ be a group 2-cocycle that agrees with $\nu \in Z^2(G, \mathrm{U}(1))$ on the orientation-preserving subgroup $G$, i.e. $\widetilde{\nu}(g, h) = \nu(g, h)$ for all $g, h \in G$.
If we define
\begin{equation}
\phi_0(g) := \frac{\widetilde{\nu}(T, g^{-1}) \widetilde{\nu}(Tg^{-1}, T) \widetilde{\nu}(g, g^{-1})}{\widetilde{\nu}(T, T)}
\end{equation}
where $T$ is the generator of time-reversal symmetry $\mathbb{Z}_2^T$, we find that $\phi_0$ satisfies (\ref{eq: phi equivariance}).
Thus, the set of the orientation-reversing isomorphisms are in one-to-one correspondence with $H^1(G, \mathrm{U}(1))$.
The pairs $(\nu, \phi)$ and $(\nu^{\prime}, \phi^{\prime})$ are equivalent if they are related by a change of the basis $b_g \rightarrow b^{\prime}_g := \mu(g) b_g$ where $\mu$ is a $\mathrm{U}(1)$-valued function on $G$.
Under this change of the basis, $\nu$ and $\phi$ transform as
\begin{equation}
\nu(g, h) \rightarrow \nu^{\prime}(g, h) = \frac{\mu(g)\mu(h)}{\mu(gh)} \nu(g, h), \quad
\phi(g) \rightarrow \phi^{\prime}(g) = \frac{\mu(g)}{\mu(g^{-1})} \phi(g).
\end{equation}
To classify the solutions $\phi$ of eq. (\ref{eq: phi equivariance}) for a fixed $\nu$, we demand that the multiplication $\nu$ is invariant under the change of the basis.
This means that $\mu$ is a group homomorphism from $G$ to $\mathrm{U}(1)$.
Thus, equivalent solutions $\phi$ and $\phi^{\prime}$ are related by the multiplication of a squared group homomorphism $\mu(g)/\mu(g^{-1}) = \mu(g)^2$, which forms an abelian group $2H^1(G, \mathrm{U}(1))$.
\begin{comment}
Two solutions $\phi$ and $\phi^{\prime}$ are equivalent if and only if they are related by the change of the basis $b_g \rightarrow b^{\prime}_g := \mu(g) b_g$ where $\mu$ is a $\mathrm{U}(1)$-valued function on $G$.
Under this change of the basis, $\phi$ transforms as
\begin{equation}
\phi(g) \rightarrow \phi^{\prime}(g) = \frac{\mu(g)}{\mu(g^{-1})} \phi(g).
\end{equation}
As we discussed in the previous paragraph, the ratio $\phi^{\prime}(g) / \phi(g) = \mu(g) / \mu(g^{-1})$ is a homomorphism from $G$ to $\mathrm{U}(1)$.
However, not every homomorphism can be realized as $\mu(g)/\mu(g^{-1})$.
It turns out that for every $\mathrm{U}(1)$-valued function $\mu$, there exists a homomorphism $f: G \rightarrow \mathrm{U}(1)$ such that
\begin{equation}
\frac{\mu(g)}{\mu(g^{-1})} = f(g)^2.
\end{equation}
Conversely, for every homomorphism $f: G \rightarrow \mathrm{U(1)}$, there exists a $\mathrm{U}(1)$-valued function $\mu$ such that the above equation is satisfied.
Thus, homomorphisms $\mu(g)/\mu(g^{-1})$ form a group $2H^1(G, \mathrm{U}(1))$, which is the group of even elements of $H^1(G, \mathrm{U}(1))$.
To obtain this result, we first decompose $\mu$ into the symmetric part $\mu_S$ and the anti-symmetric part $\mu_A$ as
\begin{equation}
\mu(g) = \mu_S(g) \mu_A(g),
\end{equation}
where $\mu_S$ and $\mu_A$ are given by $\mu_S(g) := \left[ \mu(g) \mu(g^{-1}) \right]^{1/2}$ and $\mu_A(g) := \left[ \mu(g)/\mu(g^{-1}) \right]^{1/2}$ respectively.
The ratio $\mu(g)/\mu(g^{-1})$ depends only on the anti-symmetric part $\mu_A$ because the symmetric part $\mu_S$ cancels out in this expression:
\begin{equation}
\mu(g) / \mu(g^{-1}) = \mu_A(g) / \mu_A(g^{-1}) = \mu_A(g)^2.
\label{eq: squared muA}
\end{equation}
Since the left-hand side $\mu(g)/\mu(g^{-1})$ is a group homomorphism, we find $\delta \mu_A^2(g, h) = 1$, which implies that $\delta \mu_A(g, h)$ is a $\mathbb{Z}_2$-valued coboundary.
In particular, there exists a $\mathbb{Z}_2$-valued function $\epsilon: G \rightarrow \mathbb{Z}_2$ such that $\delta \mu_A(g, h) = \delta \epsilon (g, h)$.
Therefore, if we define a $\mathrm{U}(1)$-valued function $f$ by $f(g) := \mu_A(g) \epsilon(g)$, $f$ becomes a group homomorphism and satisfy
\begin{equation}
\mu(g) / \mu(g^{-1}) = f(g)^2.
\end{equation}
This shows the existence of a group homomorphism $f$ such that $\mu(g)/\mu(g^{-1}) = f(g)^2$.
Conversely, a squared group homomorphism $f(g)^2$ can always be written as $\mu(g)/ \mu(g^{-1})$ because $f(g)^2 = f(g)/ f(g^{-1})$, i.e. we can choose $\mu(g) = f(g)$.
Thus, group homomorphisms in the form of $\mu(g)/\mu(g^{-1})$ are in one-to-one correspondence with $2H^1(G, \mathrm{U}(1))$.
\end{comment}
Therefore, orientation-reversing isomorphisms are classified up to change of the basis by the quotient group $H^1(G, \mathrm{U}(1)) / 2H^1(G, \mathrm{U}(1))$, which gives the second term on the right-hand side of (\ref{eq: Kunneth}).
This completes the classification of 1+1d bosonic SPT phases with symmetry $\mathbb{Z}_2^T \times G$.
\subsubsection{Duality symmetry}
\label{sec: Duality symmetry (unoriented)}
The invariance of a theory under gauging a finite abelian group symmetry $A$ is called duality symmetry.
A theory with duality symmetry is said to be self-dual.
The duality symmetry is described by a Tambara-Yamagami category $\mathrm{TY}(A, \chi, \epsilon)$ where $\chi: A \times A \rightarrow \mathrm{U}(1)$ is a non-degenerate symmetric bicharacter of $A$ and $\epsilon \in \{ \pm 1\}$ is a sign \cite{TY1998}.
The set of simple objects of $\mathrm{TY}(A, \chi, \epsilon)$ consists of group-like objects $a \in A$ and the duality object $m$, whose fusion rules are given by
\begin{equation}
a \otimes b = ab, \quad m \otimes a = a \otimes m = m, \quad m \otimes m = \bigoplus_{a \in A} a.
\label{eq: TY fusion rule}
\end{equation}
From the above fusion rules, we find that the quantum dimensions of a group-like object $a$ and the duality object $m$ are given by $\mathop{\mathop{\mathrm{dim}}} a = 1$ and $\mathop{\mathop{\mathrm{dim}}} m = \sqrt{|A|}$ respectively.
The nontrivial associators of $\mathrm{TY}(A, \chi, \epsilon)$ are summarized as follows:
\begin{equation}
\alpha_{amb} = \chi(a, b) \mathrm{id}_m,\quad
\alpha_{mam} = \bigoplus_{b \in A} \chi(a, b) \mathrm{id}_b,\quad
(\alpha_{mmm})_{ab} = \frac{\epsilon}{\sqrt{|A|}} \chi(a, b)^{-1} \mathrm{id}_m.
\label{eq: TY associators}
\end{equation}
The last equation means that the associator $\alpha_{mmm}$ consists of the morphisms $\epsilon \chi(a, b)^{-1} \mathrm{id}_m / \sqrt{|A|}$ from the $a$th component of the source object $\bigoplus m$ to the $b$th component of the target object $\bigoplus m$.
In the following, we classify bosonic self-dual SPT phases with time-reversal symmetry by group-theoretical data.
We first review the classification of fiber functors of the Tambara-Yamagami category $\mathrm{TY}(A, \chi, \epsilon)$ \cite{Tam2000}.
To simplify the notation, we denote the isomorphisms $M_{xy}: V_x \otimes V_y \rightarrow V_{x \otimes y}$ for simple objects $x$ and $y$ as
\begin{align}
v_a v_b & := M_{ab}(v_a \otimes v_b), \label{eq: TY multiplication1}\\
v_a \cdot w & := M_{am}(v_a \otimes w), \label{eq: TY multiplication2}\\
w \cdot v_a & := M_{ma}(w \otimes v_a), \label{eq: TY multiplication3}\\
[w, w^{\prime}] & := M_{mm}(w \otimes w^{\prime}), \label{eq: TY multiplication4}
\end{align}
where $v_a \in V_a$, $v_b \in V_b$, and $w, w^{\prime} \in V_m$.
The first equation (\ref{eq: TY multiplication1}) gives an associative multiplication on $V = \bigoplus V_a$.
In general, this multiplication is twisted by a group 2-cocycle $\xi \in Z^2(A, \mathrm{U}(1))$
\begin{equation}
u_a u_b = \xi(a, b) u_{ab},
\end{equation}
where $u_a$ is a basis of $V_a$.
The second equation (\ref{eq: TY multiplication2}) and the third equation (\ref{eq: TY multiplication3}) give the left and right actions of $V$ on $V_m$ respectively.
The left and right actions are related by an involutive anti-automorphism $f: V \rightarrow V$ as follows:
\begin{equation}
w \cdot v_a = f(v_a) \cdot w.
\end{equation}
The involutive anti-automorphism $f$ on $V$ induces an involutive automorphism $\sigma$ on $A$ as
\begin{equation}
f(u_a) = \nu(a) u_{\sigma(a)},
\label{eq: def of nu}
\end{equation}
where $\nu(a)$ is a complex number with absolute value $1$.
We can show that a fiber functor of the Tambara-Yamagami category $\mathrm{TY}(A, \chi, \epsilon)$ is characterized by a triple $(\sigma, \xi, \nu)$ such that
\begin{align}
\chi(a, b) & = \xi(a, \sigma(b)) / \xi(\sigma(b), a),\label{eq: triple1}\\
\nu(a) \nu(b) / \nu(ab) & = \xi(a, b) / \xi(\sigma(b), \sigma(a)), \label{eq: triple2}\\
\nu(a) \nu(\sigma(a)) & = 1, \label{eq: triple3}\\
\sum_{a \in A \text{ s.t. } \sigma(a) = a} \nu(a) & = \epsilon \sqrt{|A|}. \label{eq: triple4}
\end{align}
We note that the cohomology class of $\xi$ must be nontrivial due to the non-degeneracy of $\chi$.
If we change the basis of $V_a$ from $u_a$ to $u^{\prime}_a := \psi(a) u_a$, the triple $(\sigma, \xi, \nu)$ changes to
\begin{equation}
\sigma^{\prime} = \sigma, \quad \xi^{\prime}(a, b) = \frac{\psi(a) \psi(b)}{\psi(ab)} \xi(a, b), \quad \nu^{\prime}(a) = \frac{\psi(a)}{\psi(\sigma(a))} \nu(a).
\label{eq: change of basis}
\end{equation}
Thus, the fiber functors characterized by the triples $(\sigma, \xi, \nu)$ and $(\sigma^{\prime}, \xi^{\prime}, \nu^{\prime})$ are naturally isomorphic if they are related by eq. (\ref{eq: change of basis}).
Such triples are said to be equivalent to each other.
Therefore, 1+1d bosonic SPT phases with duality symmetry are classified by equivalence classes of triples $(\sigma, \xi, \nu)$ that satisfy eqs. (\ref{eq: triple1})--(\ref{eq: triple4}).
To classify self-dual SPT phases with time-reversal symmetry, we also need to consider the orientation-reversing isomorphisms $\phi_x: V_x \rightarrow V_{x^*}$ that satisfy eq. (\ref{eq: involutive equivariance}).
Considering that the pivotal structure of the duality object $m$ is given by $a_m = \epsilon \mathop{\mathrm{id}}_m$ (see e.g. \cite{Shi2011}), we find that eq. (\ref{eq: involutive equivariance}) for the Tambara-Yamagami category $\mathrm{TY}(A, \chi, \epsilon)$ reduces to
\begin{align}
\phi_{a^{-1}} (\phi_a(v_a)) & = v_a, \label{eq: phi-a involution}\\
\phi_m(\phi_m(w)) & = \epsilon w, \label{eq: phi-m involution}\\
\phi_b(v_b) \phi_a(v_a) & = \phi_{ab}(v_a v_b), \label{eq: phi-b phi-a}\\
\phi_m(w) \cdot \phi_a(v_a) & = \phi_m(v_a \cdot w), \label{eq: phi-m phi-a}\\
\phi_a(v_a) \cdot \phi_m(w) & = \phi_m(w \cdot v_a), \label{eq: phi-a phi-m}\\
[\phi_m(w^{\prime}), \phi_m(w)]_{a} & = \phi_{a^{-1}} ([w, w^{\prime}]_{a^{-1}}). \label{eq: phi-m phi-m}
\end{align}
The subscript $a$ on the left-hand side of the last equation indicates the $a$-twisted sector $V_a$ in $V_m \otimes V_m = \bigoplus V_a$.
By choosing a basis $u_a \in V_a$, we can represent $\phi_a$ by a phase factor $\phi(a)$ as $\phi_a(u_a) = \phi(a) u_{a^{-1}}$.
Similarly, by choosing a basis of $V_m$, we can represent $\phi_m$ by a $\sqrt{|A|} \times \sqrt{|A|}$ unitary matrix $\Phi$ as $\phi_m(w) = \Phi w$.
In terms of $\phi(a)$ and $\Phi$, eqs. (\ref{eq: phi-a involution}) and (\ref{eq: phi-m involution}) can be written as
\begin{align}
\phi(a) \phi(a^{-1}) & = 1, \label{eq: phi(a) involution}\\
\Phi^2 & = \epsilon. \label{eq: Phi involution}
\end{align}
If we recall that the multiplication on $V = \bigoplus V_a$ is twisted by a 2-cocycle $\xi$, we can write eq. (\ref{eq: phi-b phi-a}) as
\begin{equation}
\phi(a) \phi(b) \xi(b^{-1}, a^{-1}) = \phi(ab) \xi(a, b),
\label{eq: quintuple1}
\end{equation}
which includes eq. (\ref{eq: phi(a) involution}) as a special case.
Equation (\ref{eq: phi-m phi-a}) is equivalent to $f(\phi_a(u_a)) \cdot \phi_m(w) = \phi_m(u_a \cdot w)$, whose matrix representation is given by
\begin{equation}
\phi(a) \nu(a^{-1}) U_{\sigma(a^{-1})} \Phi = \Phi U_a
\label{eq: phi-m phi-a matrix rep}
\end{equation}
where $U_a$ is the matrix representation of the left action of $u_a$.
We note that $U_a$ obeys the same algebra as $u_a$, i.e. $U_a U_b = \xi(a, b) U_{ab}$.
Equation (\ref{eq: phi-m phi-a matrix rep}) means that the unitary transformation $U_a \mapsto \Phi U_a \Phi^{-1}$ induced by the orientation-reversing isomorphism $\Phi$ on $V_m$ agrees with $f \circ \phi: U_a \mapsto \phi(a) \nu(a^{-1}) U_{\sigma(a^{-1})}$.
Put differently, the linear map $f \circ \phi$ is represented by the unitary matrix $\Phi$, which gives the orientation-reversing isomorphism on $V_m$.
Similarly, eq. (\ref{eq: phi-a phi-m}) gives rise to the same equation (\ref{eq: phi-m phi-a matrix rep}).
Furthermore, as we will see in appendix \ref{sec: Matrix representation of eq. (phi-m phi-m)}, eq. (\ref{eq: phi-m phi-m}) reduces to
\begin{equation}
c(\sigma(a)) \nu(\sigma(a)) = c(a)
\label{eq: symmetry of Gamma Phi}
\end{equation}
where $c(a)$ is the $U_a$-component of $\Phi$, i.e. $\Phi = \sum_{a \in A} c(a) U_a$.
This expansion is always possible because $\{U_a\}$ spans the vector space of all $\sqrt{|A|} \times \sqrt{|A|}$ matrices, which will also be shown in appendix \ref{sec: Matrix representation of eq. (phi-m phi-m)}.
Therefore, a self-dual SPT phase with time-reversal symmetry is given by the algebraic data $(\sigma, \xi, \nu, s, \phi, \Phi)$ where $s \in \mathbb{Z}_2$ is a sign and $(\sigma, \xi, \nu, \phi, \Phi)$ satisfies (\ref{eq: triple1})--(\ref{eq: triple4}) and (\ref{eq: Phi involution})--(\ref{eq: symmetry of Gamma Phi}).
If we change the bases of $V_a$ and $V_m$ by a phase factor $\psi(a)$ and a unitary matrix $U$ respectively, $(\sigma, \xi, \nu)$ transforms as eq. (\ref{eq: change of basis}) and $(s, \phi, \Phi)$ transforms as
\begin{equation}
s^{\prime} = s, \quad \phi^{\prime}(a) = \frac{\psi(a)}{\psi(a^{-1})} \phi(a), \quad \Phi^{\prime} = U \Phi U^{-1}.
\label{eq: change of basis (unoriented)}
\end{equation}
The sextuples $(\sigma, \xi, \nu, s, \phi, \Phi)$ and $(\sigma^{\prime}, \xi^{\prime}, \nu^{\prime}, s^{\prime}, \phi^{\prime}, \Phi^{\prime})$ are equivalent if they are related by the change of the bases (\ref{eq: change of basis}) and (\ref{eq: change of basis (unoriented)}).
Thus, time-reversal invariant bosonic SPT phases with duality symmetry $\mathrm{TY}(A, \chi, \epsilon)$ are classified by equivalence classes of $(\sigma, \xi, \nu, s, \phi, \Phi)$.
As a concrete example, we explicitly compute the classification of self-dual SPT phases when $A = \mathbb{Z}_2 \times \mathbb{Z}_2 = \{(0, 0), (1, 0), (0, 1), (1, 1)\}$.
In this case, we have four different Tambara-Yamagami categories $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi_{\pm}, \pm 1)$ where the non-degenerate symmetric bicharacters $\chi_{\pm}$ are given by \cite{TY1998}
\begin{align}
\chi_+((1, 0), (1, 0)) & = \chi_+((0, 1), (0, 1)) = +1, \quad \chi_+((1, 0), (0, 1)) = -1,\\
\chi_-((1, 0), (1, 0)) & = \chi_-((0, 1), (0, 1)) = -1, \quad \chi_-((1, 0), (0, 1)) = +1.
\end{align}
The involutive automorphism $\sigma$ is uniquely determined by eq. (\ref{eq: triple1}) where $\xi$ is a representative of the nontrivial cohomology class in $H^2(\mathbb{Z}_2 \times \mathbb{Z}_2, \mathrm{U}(1)) = \mathbb{Z}_2$.
Specifically, $\sigma$ is equal to the identity map when $\chi = \chi_+$, while $\sigma$ is a nontrivial involution that maps $(1, 0)$ to $(0, 1)$ when $\chi = \chi_-$.
Here, we choose a group 2-cocycle $\xi$ as $\xi(a, b) = (-1)^{a_1 b_2}$ where $a = (a_1, a_2)$ and $b = (b_1, b_2)$.
For this choice of $\xi$, the two-dimensional projective representation of $\mathbb{Z}_2 \times \mathbb{Z}_2$ is given by using the Pauli matrices as follows:
\begin{equation}
U_{(0, 0)} = 1, \quad U_{(1, 0)} = \sigma_x, \quad U_{(0, 1)} = \sigma_z, \quad U_{(1, 1)} = i \sigma_y.
\end{equation}
Furthermore, the solutions $\phi$ of eq. (\ref{eq: quintuple1}) are classified by $H^1(\mathbb{Z}_2 \times \mathbb{Z}_2, \mathrm{U}(1)) / 2H^1(\mathbb{Z}_2 \times \mathbb{Z}_2, \mathrm{U}(1)) = \mathbb{Z}_2 \times \mathbb{Z}_2$ as we discussed in the previous subsection.
If we denote $\phi$ as $\phi = (\phi(0, 0), \phi(1, 0), \phi(0, 1), \phi(1, 1))$, the solutions of (\ref{eq: quintuple1}) are explicitly written as
\begin{equation}
\phi_1 = (1, -1, 1, 1), \quad \phi_2 = (1, 1, -1, 1), \quad \phi_3 = (1, 1, 1, -1), \quad \phi_4 = (1, -1, -1, -1).
\end{equation}
These solutions do not necessarily give rise to time-reversal invariant self-dual SPT phases due to eqs. (\ref{eq: phi-m phi-a matrix rep}) and (\ref{eq: symmetry of Gamma Phi}).
In the following, we classify SPT phases $(\sigma, \xi, \nu, s, \phi, \Phi)$ for each $\mathbb{Z}_2 \times \mathbb{Z}_2$ Tambara-Yamagami category.
Since the classification of $(\sigma, \xi, \nu)$ is already known \cite{Tam2000, TW2019}, it suffices to consider the classification of $(s, \phi, \Phi)$.
In particular, $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi_-, -1)$ does not admit SPT phases because eqs. (\ref{eq: triple1})--(\ref{eq: triple4}) do not have any solution.
For $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi_+, +1)$, we have three inequivalent $\nu$, which we denote by $\nu_1$, $\nu_2$, and $\nu_3$:
\begin{align}
\nu_1(1, 0) & = -1, \quad \nu_1(0, 0) = \nu_1(0, 1) = \nu_1(1, 1) = 1, \\
\nu_2(0, 1) & = -1, \quad \nu_2(0, 0) = \nu_2(1, 0) = \nu_2(1, 1) = 1, \\
\nu_3(1, 1) & = -1, \quad \nu_3(0, 0) = \nu_3(1, 0) = \nu_3(0, 1) = 1.
\end{align}
We find that there are four equivalence classes of the solutions of (\ref{eq: Phi involution})--(\ref{eq: symmetry of Gamma Phi}) for each $\nu$:\footnote{Any solution of eqs. (\ref{eq: Phi involution})--(\ref{eq: symmetry of Gamma Phi}) is equivalent to one of the solutions listed in eqs. (\ref{eq: solution1})--(\ref{eq: solution3}). For example, we have a solution $(\nu, \phi, \Phi) = (\nu_1, \phi_2, -\sigma_y)$, which is equivalent to $(\nu_1, \phi_2, \sigma_y)$.}
\begin{align}
\nu = \nu_1 \Rightarrow (\phi, \Phi) & = (\phi_1, +1), (\phi_1, -1), (\phi_2, \sigma_y), (\phi_3, \sigma_z), \label{eq: solution1}\\
\nu = \nu_2 \Rightarrow (\phi, \Phi) & = (\phi_1, \sigma_y), (\phi_2, +1), (\phi_2, -1), (\phi_3, \sigma_x), \label{eq: solution2}\\
\nu = \nu_3 \Rightarrow (\phi, \Phi) & = (\phi_1, \sigma_z), (\phi_2, \sigma_x), (\phi_3, +1), (\phi_3, -1). \label{eq: solution3}
\end{align}
If we take into account the choice of a sign $s \in \mathbb{Z}_2$, we obtain 24 different SPT phases in total.
For $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi_+, -1)$, we have a single $\nu$: $\nu(0, 0) = 1, \nu(1, 0) = \nu(0, 1) = \nu(1, 1) = -1$.
The solutions of (\ref{eq: Phi involution})--(\ref{eq: symmetry of Gamma Phi}) are given by $(\phi, \Phi) = (\phi_4, +i), (\phi_4, -i)$, which are not equivalent to each other.
Since the classification of SPT phases also depends on the choice of a sign $s \in \mathbb{Z}_2$, the bosonic SPT phase with duality symmetry $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi_+, -1)$ splits into four different SPT phases in the presence of time-reversal symmetry.
For $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi_-, +1)$, we have one equivalence class of $\nu$, whose representative is given by $\nu(a) = 1$ for all $a \in \mathbb{Z}_2 \times \mathbb{Z}_2$.
The solutions of (\ref{eq: Phi involution})--(\ref{eq: symmetry of Gamma Phi}) are in the equivalence class of $(\phi, \Phi) = (\phi_3, (\sigma_x + \sigma_z)/\sqrt{2})$.
Therefore, the bosonic SPT phase with duality symmetry $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi_-, +1)$ splits into two different time-reversal invariant SPT phases, which are distinguished by the choice of a sign $s \in \mathbb{Z}_2$.
The number of bosonic SPT phases for each duality symmetry $\mathrm{TY}(\mathbb{Z}_2 \times \mathbb{Z}_2, \chi, \epsilon)$ is summarized in tables \ref{tab: duality SPT phases1} and \ref{tab: duality SPT phases2}.
\begin{table}
\begin{center}
\begin{minipage}{0.475 \hsize}
\begin{center}
\begin{tabular}{c|cc}
& $\epsilon = +1$ & $\epsilon = -1$\\ \hline
$\chi = \chi_+$ & 3 & 1\\
$\chi = \chi_-$ & 1 & 0
\end{tabular}
\caption{The number of $\mathbb{Z}_2 \times \mathbb{Z}_2$ self-dual SPT phases without time-reversal symmetry.}
\label{tab: duality SPT phases1}
\end{center}
\end{minipage}%
\begin{minipage}{0.05 \hsize}
\begin{center}
\end{center}
\end{minipage}%
\begin{minipage}{0.475 \hsize}
\begin{center}
\begin{tabular}{c|cc}
& $\epsilon = +1$ & $\epsilon = -1$\\ \hline
$\chi = \chi_+$ & 24 & 4\\
$\chi = \chi_-$ & 2 & 0
\end{tabular}
\caption{The number of $\mathbb{Z}_2 \times \mathbb{Z}_2$ self-dual SPT phases with time-reversal symmetry.}
\label{tab: duality SPT phases2}
\end{center}
\end{minipage}
\end{center}
\end{table}
Finally, we notice that some duality symmetries do not admit time-reversal invariant SPT phases even if they admit SPT phases without time-reversal symmetry.
For example, a $\mathbb{Z}_{2n+1} \times \mathbb{Z}_{2n+1}$ self-duality $\mathrm{TY}(\mathbb{Z}_{2n+1} \times \mathbb{Z}_{2n+1}, \chi, +1)$ admits an SPT phase in the absence of time-reversal symmetry due to Proposition 4.2 of \cite{Tam2000}.
Here, the non-degenerate symmetric bicharacter $\chi$ is given by
\begin{equation}
\chi(a, a) = \chi(b, b) = 1, \quad \chi(a, b) = e^{2 \pi i/ (2n+1)},
\end{equation}
where $a$ and $b$ are the generators of $\mathbb{Z}_{2n+1} \times \mathbb{Z}_{2n+1}$.
However, this self-duality does not admit time-reversal invariant SPT phases because the 2-torsion subgroup of $H^2(\mathbb{Z}_{2n+1} \times \mathbb{Z}_{2n+1}, \mathrm{U}(1)) = \mathbb{Z}_{2n+1}$ is trivial and hence there is no nontrivial cocycle $\xi$ that satisfies eqs. (\ref{eq: triple1}) and (\ref{eq: quintuple1}) simultaneously.
This may be regarded as a mixed anomaly between time-reversal symmetry and the duality symmetry $\mathrm{TY}(\mathbb{Z}_{2n + 1} \times \mathbb{Z}_{2n + 1}, \chi, +1)$.
\section{Conclusion}
In this paper, we discussed the classification of 1+1d bosonic fusion category SPT phases with and without time-reversal symmetry.
We showed that bosonic fusion category SPT phases without time-reversal symmetry are classified by isomorphism classes of fiber functors, which agrees with the previous result \cite{TW2019, KORS2020}.
We obtained this result by explicitly solving the consistency conditions of oriented TQFTs with fusion category symmetry formulated in \cite{BT2018}.
Our derivation revealed that the data of a fiber functor naturally appear in the low energy limit of SPT phases.
We also classified bosonic fusion category SPT phases with time-reversal symmetry when the total symmetry splits into time-reversal symmetry and fusion category symmetry.
To accomplish the classification, we axiomatized unoriented TQFTs with fusion category symmetry by generalizing the oriented case.
We found that unoriented TQFTs for SPT phases are in one-to-one correspondence with equivalence classes of quintuples $(Z, M, i, s, \phi)$ where $(Z, M, i)$ is a fiber functor, $s$ is a sign, and $\phi$ is a collection of isomorphisms that satisfy eq. (\ref{eq: involutive equivariance}).
As an application, we specified the group-theoretical data that classify bosonic self-dual SPT phases with time-reversal symmetry and found that some duality symmetries do not have SPT phases only when time-reversal symmetry is imposed.
There are several future directions.
One is to generalize our result to the case where the total symmetry does not split into time-reversal symmetry and fusion category symmetry.
For a finite group symmetry that is not necessarily a direct product of time-reversal symmetry and internal symmetry, the axioms of unoriented equivariant TQFTs are given in \cite{KT2017}.
In particular, invertible unoriented equivariant TQFTs are shown to be classified by twisted group cohomology.
It would be interesting to extend this result to more general fusion category symmetries.
Another future direction is the classification of fermionic fusion category SPT phases.
For fermionic theories, fusion category symmetry can further be generalized to superfusion categories, which incorporate the information of Majorana fermions that reside on topological defect lines \cite{KORS2020, NR2020, GK2016, BGK2017, BE2017, Ush2018, BGHNPRW2017, ALW2019, LSCH2020}.
Moreover, fermionic topological field theories, which describe the low energy limit of fermionic SPT phases, depend on a choice of a variant of spin structure.
It would be interesting to formulate fermionic topological field theories with superfusion category symmetry and solve them to classify fermionic fusion category SPT phases.
\section*{Acknowledgments}
The author thanks Ryohei Kobayashi and Masaki Oshikawa for comments on the manuscript.
The author is supported by FoPM, WINGS Program, the University of Tokyo.
|
\chapter*{Introduction}
Since the seminal work of Dedekind and Weber \cite{MR1579901}, the similarity between number fields and fields of algebraic functions of one variable has been known and has deeply influenced researches in algebraic geometry and number theory. Inspired by the discovery of Hensel and Hasse on embeddings of a number field into diverse local fields, Weil \cite{Weil39} considered all places, finite or infinite, of a number field, which made a decisive step toward the unification of number theory and algebraic geometry. Many works have then been done along this direction. On the one hand, the analogue of Diophantine problems (notably Mordell's conjecture) in the function field setting has been studied by Manin \cite{MR0154868}, Grauert \cite{MR222087} and Samuel \cite{MR204430}; on the other hand, through Weil's height machine \cite{MR42169} and the theory of N\'{e}ron-Tate's height \cite{MR179173}, methods of algebraic geometry have been systematically applied to the research of Diophantine problems, and it has been realized that the understanding of the arithmetic of algebraic varieties over a number field, which should be analogous to algebraic geometry over a smooth projective curve, is indispensable in the geometrical approach of Diophantine problems. Under such a circonstance Arakelov \cite{MR0466150,MR0472815} has developed the arithmetic intersection theory for arithmetic surfaces (namely relative curves over $\operatorname{Spec}\mathbb Z$). Note that the transcription of the intersection theory to the arithmetic setting is by no means automatic. The key idea of Arakelov is to introduce transcendental objects, notably Hermitian metrics or Green functions, over the infinite place, in order to ``compactify'' arithmetic surfaces. To each pair of compactified arithmetic divisor, he attached a family of local intersection numbers parametrized by the set of places of the base number field. The global intersection number is obtained by taking the sum of local intersection numbers. Arakelov's idea has soon led to spectacular advancements in Diophantine geometry, especially Faltings' proof \cite{MR740897} of Mordell's conjecture.
The fundament of Arakelov geometry for higher dimensional arithmetic varieties has been established by Gillet and Soul\'{e}. The have introduced an arithmetic intersection theory \cite{MR770447,MR1087394} for general arithmetic varieties and proved an ``arithmetic Riemann-Roch theorem'' \cite{MR1055224}. They have introduced the arithmetic Chow group, which is an hybride construction of the classic Chow group in algebraic geometry and currents in complex analytic geometry. Applications of arithmetic intersection theory in Diophantine geometry have then been developed, notably to build up an intrinsic height theory for arithmetic projective varieties (see for example \cite{MR1109353,MR1260106}). Arakelov's height theory becomes now an important tool in arithmetic geometry. Upon the need of including several constructions of local heights (such as the canonical local height for subvarieties in an Abelian variety) in the setting of Arakelov geometry, Zhang \cite{MR1311351} has introduced the notion of adelic metrics for ample line bundles on a projective variety over a number field, which could be considered as uniform limit of Hermitian line bundles (with possibly different integral models).
Inspired by the similarity between Diophantine analysis and Nevanlinna theory, Gubler \cite{MR1472498} has proposed a vast generalization of height theory in the framework of $M$-fields. Recall that a $M$-field is a field $K$ equipped with a measure space $M$ and a map from $K\times M$ to $\mathbb R_{\geqslant 0}$ which behaves almost everywhere like absolute values on $K$. Combining the intersection product of Green currents in the Archimedean case and the local height of Chow forms, he has introduced local heights (parametrized by the measure space $M$) for a projective variety over an $M$-field. Assuming the integrability of the function of local heights on the measure space $M$, he has defined the global height of the variety as the integral of local heights. Interesting examples have been discussed in the article, which show that in many cases the function of local heights is indeed integrable.
In \cite{MR3783789}, we have developed an Arakelov geometry over adelic curves. Our framework is similar to $M$-field of Gubler, with a slightly different point of view: an adelic curve is a field equipped with a family of absolute values parametrized by a measure space (in particular, we require the absolute values to be defined everywhere). These absolute values play the role of places in algebraic number theory. Hence we can view an adelic curve as a measure space of ``places'' of a given field, except that we allow possibly equivalent absolute values in the family, or even copies of the same absolute values. Natural examples of adelic curves contain global fields, countably generated fields over global fields (as we will show in the second chapter of the current article), field equipped with copies of the trivial absolute value, and also the amalgamation of different adelic structures of the same field. Our motivation was to establish a theory of adelic vector bundles (generalizing previous works of Stuhler \cite{MR424707}, Grayson \cite{MR780079}, Bost \cite{MR1423622} and Gaudron \cite{MR2431505}), which is analogous to geometry of numbers and hence provides tools to consider Diophantine analysis in a general and flexible setting. By using the theory of adelic vector bundles, the arithmetic birational invariants are discussed in a systematic way.
The first contribution of the current article is to discuss transcendental coverings of adelic curves. Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ be an adelic curve, where $K$ is a countable field, $(\Omega,\mathcal A,\nu)$ is a measure space, and $\phi:\omega\mapsto|\raisebox{.4ex}{.}|_\omega$ is a map from $\Omega$ to the set of all absolute values of $K$, such that, for any $a\in K^{\times}$, the function $(\omega\in\Omega)\mapsto\ln|a|_\omega$ is measurable. In \cite[Chapter 3]{MR3783789}, for any algebraic extension $L/K$, we have constructed a measure space $(\Omega_{L},\mathcal A_L,\nu_L)$, which is fibered over $(\Omega,\mathcal A,\nu)$ and admits a family of disintegration probability measures. To each $\omega\in\Omega$, we correspond the fiber $\Omega_{L,\omega}$ to the family of all absolute values of $L$ extending $|\raisebox{.4ex}{.}|_\omega$. Thus we obtain a structure of adelic curve on $L$ which is called an \emph{algebraic covering} of $S$.
In \cite[\S3.2.5]{MR3783789}, we have illustrated the construction of an adelic curve structure on $\mathbb Q(T)$, which takes into account the arithmetic of $\mathbb Q$ and the geometry of $\mathbb P^1$. In the current article, we generalizes and systemize such a construction on a purely transcendental and countably generated extension of the underlying field $K$ of the adelic curve $S$. For simplicity, we explain here the case of rational function of finitely many variables. Let $n$ be an integer such that $n\geqslant 1$ and $\boldsymbol{T}=(T_1,\ldots,T_n)$ be variables. Let $L$ be the rational function field $K(\boldsymbol{T})=K(T_1,\ldots,T_n)$, which is by definition the field of fractions of the polynomial ring $K[\boldsymbol{T}]=K[T_1,\ldots,T_n]$.
To each $\omega\in\Omega$ such that the absolute value $|\raisebox{.4ex}{.}|_{\omega}$ is non-Archimedean, by Gauss's lemma, we extends $|\raisebox{.4ex}{.}|_{\omega}$ to be an absolute value on $L$ such that
\[\forall\,f=\sum_{\boldsymbol{d}\in\mathbb N^n}a_{\boldsymbol{d}}(f)\boldsymbol{T}^{\boldsymbol{d}}\in K[\boldsymbol{T}],\quad |f|_{\omega}=\max_{\boldsymbol{d}\in\mathbb N^n}|a_{\boldsymbol{d}}|_\omega.\]
We then take $\Omega_{L,\omega}$ to be the one point set $\{\omega\}$, which is equipped with the trivial probability measure.
In the case where the absolute value $|\raisebox{.4ex}{.}|_\omega$ is Archimedean, we fix an embedding $\iota_{\omega}:K\rightarrow\mathbb C$ such that $|\raisebox{.4ex}{.}|_{\omega}$ is the composition of the usual absolute value $|\raisebox{.4ex}{.}|$ on $\mathbb C$ with $\iota_\omega$ (by a measurable selection argument, we can arrange that the family of $\iota_{\omega}$ parametrized by Archimedean places is $\mathcal A$-measurable). We let
\[ \Omega_{L,\omega}:=\left\{ (t_1,\ldots,t_n) \in [0,1]^n \left|
\begin{array}{l}
\text{\small $( e(t_1),\ldots,e(t_n))$ is algebraically} \\
\text{\small independent over $\iota_{\omega}(K)$}
\end{array}
\right\}\right.,
\]
where for each $t\in[0,1]$, $e(t)$ denotes $\mathrm{e}^{2\pi it}$. Note that, if we equip $[0,1]^n$ with the Borel $\sigma$-algebra and the uniform probability measure, then $\Omega_{L,\omega}$ is a Borel set of measure $1$. Moreover, each element $\boldsymbol{t}=(t_1,\ldots,t_n)\in\Omega_{L,\omega}$ gives rise to an absolute value $|\raisebox{.4ex}{.}|_{\boldsymbol{t}}$ on $L$ such that
\[\forall\,f=\sum_{\boldsymbol{d}\in\mathbb N^n}a_{\boldsymbol{d}}(f)\boldsymbol{T}^{\boldsymbol{d}}\in K[\boldsymbol{T}],\quad |f|_{\boldsymbol{t}}=\bigg|\sum_{\boldsymbol{d}\in\mathbb N^n}\iota_\omega(a_{\boldsymbol{d}}(f))e(t_1)^{d_1}\cdots e(t_n)^{d_n}\bigg|.\]
It turns out that the disjoint union $\Omega_{L}$ of $(\Omega_{L,\omega})_{\omega\in\Omega}$ forms a structure of adelic curve on the field $L$, which is fibered over that of $S$, and admits a family of disintegration probability measures. We denote by $S_{L}=(L,(\Omega_{L},\mathcal A_{\Omega_{L}},\nu_{L}),\phi_{L})$ the corresponding adelic curve.
In the case where the adelic curve $S$ is proper, namely the following equality holds for any $a\in K^{\times}$
\[\int_{\Omega}\ln|a|_\omega\,\nu(\mathrm{d}\omega)=0,\]
it is not true in general that the adelic curve $S_{L}$ is also proper. In the article, we propose several natural ``compactifications'' of the adelic curve. Here we explain one of them which has an ``arithmetic nature''. We say that two irreducible polynomials $P$ and $Q$ in $K[T_1,\ldots,T_n]$ are equivalent if they differ by a factor of non-zero element of $K$. This is an equivalence relation on the set of all irreducible polynomials In each equivalence class we pick a representative to form a family $\mathscr P$ of irreducible polynomials. Then every non-zero element $f$ of $K$ can be written in a unique way as
\[f=a(f)\prod_{F\in\mathscr P}F^{\operatorname{ord}_F(f)},\]
where $a(f)\in K^{\times}$, and $\operatorname{ord}_F(\raisebox{.4ex}{.}):L\rightarrow\mathbb Z\cup\{+\infty\}$ is the discrete valuation associated with $F$, we denote by $|\raisebox{.4ex}{.}|_F=\mathrm{e}^{-\operatorname{ord}_F(\raisebox{.4ex}{.})}$ the corresponding absolute value on $L$. Moreover, the degree function on $K[\boldsymbol{T}]$ extends naturally to $L$ so that $-\deg(\raisebox{.4ex}{.})$ is a discrete valuation on $L$. Moreover, the following equality holds (see Proposition \ref{prop:admissible:fibration:K:T})
\[\forall\,f\in K(\boldsymbol{T}),\quad \sum_{F\in\mathscr P}\deg(F)\operatorname{ord}_F(f)=\deg(f).\]
We let $|\raisebox{.4ex}{.}|_\infty$ be the absolute value on $L$ such that $|\raisebox{.4ex}{.}|_\infty=\mathrm{e}^{\deg(\raisebox{.4ex}{.})}$. Note that, for any $F\in\mathscr P$, one has
\[h_{S_L}(F):=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\geqslant 0.\]
We fix a positive real number $\lambda$. Let $(\Omega_L^\lambda,\mathcal A_{L}^{\lambda},\nu_L^\lambda)$ be the disjoint union of $(\Omega_L,\mathcal A_L,\nu_L)$ and $\mathscr P\cup\{\infty\}$, which is equipped with the measure $\nu_L^\lambda$ extending $\nu_L$ and such that $\nu_L^\lambda(\{\infty\})=\lambda$ and
\[\forall\, F\in\mathscr P,\quad \nu_L^\lambda(\{F\})=h_{S_L}(F)+\lambda\deg(F).\]
Let $\phi_L^\lambda$ be the map from $\Omega_L^\lambda$ to the set of absolute values on $L$, sending $x\in\Omega_L^\lambda$ to $|\raisebox{.4ex}{.}|_x$.
Then we establish the following result (see \S\ref{sec:Arithmetic adelic structure}, notably Propositions \ref{prop:proper:L:lambda} and \ref{prop:Northcott:property}, see also Proposition \ref{Pro: compactification 2} for the general construction).
\begin{enonce*}{Theorem A} Assume that the adelic curve $S$ is proper.
\begin{enumerate}[label=\rm(\arabic*)]
\item For any $\lambda>0$, the adelic curve $S_L^\lambda=(L,(\Omega_L^\lambda,\mathcal A_L^\lambda,\nu_L^\lambda),\phi_L^\lambda)$ is proper.
\item If the adelic curve $S$ satisfies the Northcott property, namely, for any $C\geqslant 0$, the set
\[\bigg\{a\in K\,\bigg|\,\int_{\Omega}\max\{\ln|a|_\omega,0\}\,\nu(\mathrm{d}\omega)\leqslant C\bigg\}\]
is finite, then, for any $\lambda>0$, the adelic curve $S_L^\lambda$ satisfies the Northcott property.
\end{enumerate}
\end{enonce*}
Together with the algebraic covering of adelic curves mentioned above. This construction provides a large family of adelic structures for finitely generated extensions of $\mathbb Q$, which behaves well from the view of geometry of numbers. Note however that the compactification $S_L^\lambda$ is not fibered over $S$, but rather fibered over the amalgamation of $S$ with copies of the trivial absolute value on $K$. This phenomenon suggest that it is a need of dealing with the trivial absolute value in the consideration of the relative geometry of adelic curves.
To build up a more complete picture of Arakelov geometry over an adelic curve, it is important to develop an arithmetic intersection theory and relate it to the heights of projective varieties over an adelic curve. Although the local intersection theory is now well understood, thanks to works such as \cite{MR1472498,MR1629925,MR2244803,MR3498148}, it remains a challenging problem to show that the local intersection numbers form an integrable function over the parametrizing measure space. In this article, we resolve this integrability problem and thus establish a global intersection theory in the framework of Arakelov geometry over adelic curves. Recall that the function of local heights for an adelic line bundle is only well defined up to the function of absolute values of a non-zero scalar. One way to make explicit the local height function is to fix a family of global sections of the line bundle which intersect properly. Note that each global sections determine a Cartier divisor on the projective variety, and the adelic metrics of the adelic line bundle determine a family of Green functions of the Cartier divisor parametrized by the measure space of ``places''. For this reason, we choose to work in the framework of adelic Cartier divisors.
Let $S$ be an adelic curve, which consists of a field $K$, a measure space $(\Omega,\mathcal A,\nu)$ and a family $(|\raisebox{.4ex}{.}|_{\omega})_{\omega\in\Omega}$ of absolute values on $K$ parametrized by $\Omega$. Let $X$ be a projective scheme over $\operatorname{Spec} K$ and $d$ be the Krull dimension of $X$. By adelic Cartier divisor on $X$, we mean the datum $\overline D$ consisting of a Cartier divisor $D$ on $X$ together with a family $g=(g_\omega)_{\omega\in\Omega}$ parametrized by $\Omega$, where $g_\omega$ is a Green function of $D_\omega$, the pull back of $D$ on $X_\omega=X\otimes_K K_\omega$, with $K_\omega$ being the completion of $K$ with respect to $|\raisebox{.4ex}{.}|_\omega$. Conditions of measurability and dominancy (with respect to $\omega\in\Omega$) for the family $g$ are also required (see \S\S\ref{Sec: Reminder on adelic vector bundles}--\ref{Sec: Integrability of local intersection numbers} for more details). We first introduce the local intersection product for adelic Cartier divisors. More precisely, if $\overline D_i=(D_i,g_i)$, $i\in\{0,\ldots,d\}$, form a family of integrable metrized Cartier divisors on $X$ (namely a Cartier divisor equipped with a Green function) such that $D_0,\ldots,D_d$ intersect properly, we define, for any $\omega\in\Omega$ a local intersection number
\[(\overline D_0,\ldots,\overline D_d)_\omega\in\mathbb R\]
in a recursive way using Bedford-Taylor theory \cite{MR445006} and its non-Archimedean analogue \cite{MR2244803}.
In the case where $|\raisebox{.4ex}{.}|_{\omega}$ is a trivial absolute value, we need a careful definition of
the local intersection number
(see Definition~\ref{Def: intersection in trivial valuation case}, for details).
Note the local intersection number is a multi-linear function on the set of $(d+1)$-uplets $(\overline D_0,\ldots,\overline D_d)$ such that $D_0,\ldots,D_d$ intersect properly.
To establish a global intersection theory, we need to show that the function of local intersection numbers
\[(\omega\in\Omega)\longmapsto (\overline D_0,\ldots,\overline D_d)_\omega \]
is measurable and integrable with respect to $\nu$, where the measurability part is more subtle. Although the Green function families of $\overline D_0,\ldots,\overline D_d$ are supposed to be measurable, the corresponding products of Chern currents (or their non-Archimedean analogue) depend on the local analytic geometry relatively to the absolute values $|\raisebox{.4ex}{.}|_\omega$. It seems to be a difficult (but interesting) problem to precisely describe the measurability of the local geometry of the analytic spaces $X_\omega^{\mathrm{an}}$. For places $\omega$ which are Archimedean, as we can embed all local completions $K_{\omega}$ in the same field $\mathbb C$, by a measurable selection theorem one can show that the family of Monge-Amp\`{e}re measures is measurable with respect to $\omega$ (see Theorem \ref{thm:measurable:archimedean}). However, for non-Archimedean places, such embeddings in a common valued field do not exist in general, and the classic approach of taking a common integral model for all non-Archimedean places is not adequate in the setting of adelic curves, either.
To overcome this difficulty, our approach consist in relating the local intersection number to the local length of the mixed resultant and hence reduce the problem to the measurability of the function of local length of the mixed resultant, which is known by the theory of adelic vector bundles developed in \cite{MR3783789}. This approach is inspired by previous results of Philippon \cite{MR876159} on height of algebraic cycles via the theory of Chow forms and the comparison \cite{MR1092175,MR1306549,MR1144338,MR1260106} between Philippon's height and Faltings height (defined by the arithmetic intersection theory). Note that the similar idea has also been used in \cite{MR1472498} to construct the local height in the setting of $M$-fields.
Let us briefly recall the theory of mixed resultant. It is a multi-homogeneous generalization of Chow forms, which allows to describe the interactions of several embeddings of a variety in projective spaces by a multi-homogeneous polynomial. One of its original forms is the discriminant of a quadratic polynomial, or more generally the resultant of $n+1$ polynomials $P_0,\ldots,P_n$ in $n$ variables over an algebraically closed field, which is an irreducible polynomial in the coefficients of $P_0,\ldots,P_n$, which vanishes precisely when these polynomials have a common root. The modern algebraic approach of resultants goes back to the elimination theory of Cayley \cite{Cayley48}, where he related resultant to the determinant of Koszul complex. We use here a geometric reformulation as in the book \cite{MR1264417} of Gel'fand, Kapranov and Zelevinsky. In Diophantine geometry, mixed resultant has been used by R\'emond \cite{MR1837827} to study multi-projective heights.
We assume that the Cartier divisors $D_i$ are very ample and thus determine closed immersions $f_i$ from $X$ to the projective space of the linear system $E_i$ of the divisor $D_i$. By \emph{incidence variety} of $(f_0,\ldots,f_d)$, we mean the closed subscheme $I_X$ of $X\times_K\mathbb P(E_0^\vee)\times_K\cdots\times_K\mathbb P(E_{d}^\vee)$ parametrizing points $(x,\alpha_0,\ldots,\alpha_d)$ such that
\[\alpha_0(x)=\cdots=\alpha_d(x)=0.\]
One can also consider $I_X$ as a multi-projective bundle over $X$ (of $E_i^\vee$ quotient by the tautological line subbundle).
Therefore, the projection of $I_X$ in $\mathbb P(E_0^\vee)\times_K\cdots\times_K\mathbb P(E_d^\vee)$ consists of a family of hyperplanes in $\mathbb P(E_0),\ldots,\mathbb P(E_d)$ respectively, which contain at least one common point of $X$. It turns out that this projection is actually a multi-homogeneous hypersurface of $\mathbb P(E_0^\vee)\times_K\cdots\times_K\mathbb P(E_d^\vee)$, which is defined by a multi-homogeneous polynomial $R_{f_0,\ldots,f_d}^X$, called a \emph{resultant} of $X$ with respect to the embeddings of $f_0,\ldots,f_d$. We refer the readers to \cite[\S3.3]{MR1264417} for more details, see also \cite{MR3098424} for applications in arithmetic Nullstellensatz. When $K$ is a number field, the height of the polynomial $R_{f_0,\ldots,f_d}^X$ can be viewed as a height of the arithmetic variety $X$, and, in the particular case where the image of $D_i$ in the Picard group are colinear, an explicit comparison between the height of resultant and the Faltings height of $X$ has been discussed in \cite[Theorem 4.3.2]{MR1260106} (see also \S4.3.4 of \emph{loc. cit.}).
Usually the resultant is well defined up to a factor in $K^{\times}$. In the classic setting of number field, this is anodyne for the study of the global height, thanks to the product formula. However, in our setting, this dependence on the choice of a non-zero scalar could be annoying, especially when the adelic curve does not satisfy a product formula. In order to obtain a local height equality, we introduce, for each vector
\[(s_0,\ldots,s_d)\in E_0\times\cdots \times E_d\]
such that $\operatorname{div}(s_0),\ldots,\operatorname{div}(s_d)$ intersect properly on $X$, a specific resultant $R^{X,s_0,\ldots,s_d}_{f_0,\ldots,f_d}$ of $X$ with respect to the embeddings, which is the only resultant such that
\[R^{X,s_0,\ldots,s_d}_{f_0,\ldots,f_d}(s_0,\ldots,s_d)=1.\] We then show that the local height for this resultant coincides with the local height of $X$ defined by the local intersection theory. By using this comparison of local height and properties of adelic vector bundles over an adelic curve (see \cite[\S4.1.4]{CMArakelovAdelic}), we prove the integrability of the local height function on non-Archimedean places. Moreover, the integral of the local height equalities leads to a equality between the global height of the resultant and the arithmetic intersection number (see Remark \ref{Rem: global heights equality}), which generalizes the height comparison results in \cite{MR1092175,MR1260106}. In resume, we obtain the following result (see Theorems \ref{Thm: equality of local height} and \ref{Thm: integrability of local intersection}).
\begin{enonce*}{Theorem B}
Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ be an adelic curve, $X$ be a projective scheme over $S$, $d$ be the dimension of $X$, $D_0,\ldots, D_d$ be Cartier divisors on $X$, which are equipped with Green function families $g_0,\ldots,g_d$, respectively, such that $(D_{i,\omega},g_{i,\omega})$ is integrable for any $\omega\in\Omega$ and $i\in\{0,\ldots,d\}$.
\begin{enumerate}[label=\rm(\arabic*)]
\item\label{Item: local height formula} Assume that the Cartier divisors $D_0,\ldots,D_d$ are very ample. For any $i\in\{0,\ldots,d\}$, let $E_i=H^0(X,\mathcal O_X(D_i))$, $f_i:X\rightarrow\mathbb P(E_i)$ be the closed embedding and $s_i\in E_i$ be the regular meromorphic section of $\mathcal O_X(D_i)$ corresponding to $D_i$. Assume that the continuous metric family $\varphi_{g_i}$ corresponding to the Green function family $g_i$ is consisting of the orthogonal quotient metrics induces by a Hermitian norm family $\xi_i=(\|\raisebox{.4ex}{.}\|_{i,\omega})_{\omega\in\Omega}$ on $E_i$. Then, for any $\omega\in\Omega$, then following equalities hold.
\begin{enumerate}[label=\rm(1.\alph*)]
\item In the case where $|\raisebox{.4ex}{.}|_\omega$ is non-Archimedean, one has
\[(\overline D_0\cdots\overline D_d)_\omega=\ln\Big\|R_{f_0,\ldots,f_d}^{X,s_0,\ldots,s_d}\Big\|_{\omega,\varepsilon},\]
where the norm $\|\raisebox{.4ex}{.}\|_{\omega,\varepsilon}$ on the space of multi-homogeneous polynomials is the $\varepsilon$-tensor product of $\varepsilon$-symmetric power of $\|\raisebox{.4ex}{.}\|_{i,\omega,*}$.
\item In the case where $|\raisebox{.4ex}{.}|_\omega$ is Archimedean, one has
\[\begin{split}(\overline D_0\cdots\overline D_d)_\omega=\int_{\mathbb S(E_{0,\omega})\times\cdots\times\mathbb S(E_{d,\omega})}&\ln\Big|R_{f_0,\ldots,f_d}^{X,s_0,\ldots s_d}(z_0,\ldots,z_d)\Big|_\omega\,\mathrm{d}z_0\cdots\mathrm{d}z_d\\
&+\frac 12\sum_{i=0}^d\delta_i\sum_{\ell=1}^{r_i}\frac{1}{\ell},
\end{split}\]
where $\mathbb{S}(E_{i,\omega})$ denotes the unit sphere of $(E_{i,\omega},\|\raisebox{.4ex}{.}\|_{i,\omega})$, $\mathrm{d}z_i$ is the Borel probability measure on $\mathbb{S}(E_{i,\omega})$ invariant by the unitary group, $r_i$ is the dimension of $E_i$, and $\delta_i$ is the intersection number
\[(D_0\cdots D_{i-1}D_{i+1}\cdots D_d).\]
\end{enumerate}
\item Assume that, either the $\sigma$-algebra $\mathcal A$ is discrete, or the field $K$ admits a countable subfield, which is dense in each $K_\omega$. If all couples $\overline D_i=(D_i,g_i)$ are integrable adelic Cartier divisors on $X$, the the function
\[(\omega\in\Omega)\longrightarrow (\overline D_0\cdots\overline D_d)_\omega\]
is $\nu$-integrable.
\end{enumerate}
\end{enonce*}
As an application, we can define the multi-height of the projective scheme $X$ with respect to $\overline D_0,\ldots,\overline D_d$ as
\[h_{\overline D_0\cdots\overline D_d}(X)=\int_{\Omega}(\overline D_0\cdots\overline D_d)_\omega\,\nu(\mathrm{d}\omega),\]
and, under the assumptions of the point \ref{Item: local height formula} in the above theorem, we can relate the multi-height with the height of the resultant, by taking the integral of the local height equality.
From the methodological point of view, the approach of \cite{MR1092175} works within $\mathbb P^N(\mathbb C)$ and uses elimination theory and complex analysis of the Fubini-Study metric; that of \cite{MR1260106} relies on a choice of integral model and computations in the arithmetic Chow groups. In our setting, we need to deal with general non-Archimedean metrics. Hence these approaches do not fit well with the framework of adelic curves. Our method consists in computing the local height of
\[X\times_K\mathbb P(E_0^\vee)\times_K\cdots\times_K\mathbb P(E_{d}^\vee)\]
in two ways (see Lemma \ref{Lem: h s in h IX}
for details). We first consider this scheme as a fibration of multi-projective space over $X$ and relate this local height to that of $X$ by taking the local intersection along the fibers. We then relate the height of this product scheme to that of the incidence subscheme $I_X$ and then use the identification of $I_X$ with a multi-projective bundle over $X$ to compute recursively the height of $I_X$. Our method allows to obtain a local height equality in considering the Archimedean case and the non-Archimedean case in a uniform way.
It is worth mentioning that an intersection theory of arithmetic cycles and a Riemann-Roch theory could be expected for the setting of adelic curves. However, new ideas are needed to establish a good formulation of the measurability for various arithmetic objects arising in such a theory.
The rest of the article is organized as follows. In the first chapter, we remind several basic constructions used in the article, including multi-linear subset and multi-linear functions, Cartier divisors on general scheme, proper intersection of Cartier divisors on a projective scheme, multi-homogeneous polynomials, incidence subscheme and resultants, and linear projections of closed subschemes in a projective space. The second chapter is devoted to the construction of adelic structures. After a brief reminder on the definition of adelic curves and their algebraic covers, we introduce transcendental fibrations of adelic curves and their compactifications. These constructions provide a large family of examples of adelic curves. In the third chapter, we consider the local intersection theory in the setting of projective schemes over a complete valued field. We first remind the notions of continuous metrics on an invertible sheaf and its semi-positivity. Then we explain the notion of Green functions of Cartier divisors and their relation with continuous metrics. The construction of Monge-Amp\`{e}re mesures and local intersection numbers is then discussed. The last sections are devoted to establish the link between the local intersection number and the length (in the non-Archimedean case) or Mahler measure (in the Archimedean case) of the corresponding resultant, respectively. In the fourth and last chapter, we prove the integrability of the local height function and construct the global multi-height.
\chapter{Reminder and preliminaries}
\section{Symmetric and multi-linear subsets}
In this section, we fix a commutative and unitary ring $k$.
\begin{defi}\label{def:domain}
Let $d$ be a non-negative integer and $V$ be a $k$-module.
We say that a subset $S$ of $V^{d+1}$ is \emph{multi-linear} if, for any $j\in\{0,\ldots,d\}$ and for any $(x_0,\ldots,x_{j-1},x_{j+1},\ldots,x_d)\in V^d$, the subset
\[\{x_j\in V\,|\,(x_0,\ldots,x_{j-1},x_j,x_{j+1},\ldots,x_d)\in S\}\]
of $V$ is either empty or a sub-$k$-module. If in addition
\[(x_0,\ldots,x_d)\in S\;\Longrightarrow\;(x_{\sigma(0)},\ldots,x_{\sigma(d)})\in S\]
for any bijection $\sigma:\{0,\ldots,d\}\rightarrow\{0,\ldots,d\}$, we say that the multi-linear subset $S$ is \emph{symmetric}.
\end{defi}
\begin{prop}
Let $d\in{\mathbb Z}_{\geqslant 0}$, $V$ be a $k$-module and $S$ be a multi-linear subset of $V^{d+1}$.
For any $j\in\{0,\ldots,d\}$, let $I_j$ be a non-empty finite set, $(x_{j,i})_{i\in I_j}$ be a family of elements of $V$, $(\lambda_{j,i})_{i\in I_j}$ be a family of elements of $k$, and $y_j=\sum_{i\in I_j}\lambda_{j,i}x_{j,i}$. Assume that, for any $(i_0,\ldots,i_d)\in I_0\times\cdots\times I_d$, one has $(x_{0,i_0},\ldots,x_{d,i_d})\in S$. Then $(y_0,\ldots,y_d)\in S$.
\end{prop}
\begin{proof}
We reason by induction on $d$. In the case where $d=0$, $S$ is a sub-$k$-module of $V$ when it is not empty. Since $y_0$ is a $k$-linear combination of elements of $S$, we obtain that $y_0\in S$.
We now assume that $d\geqslant 1$ and that the statement holds for multi-linear subsets of $V^d$. Let \[S'=\{(z_0,\ldots,z_{d-1})\in V^d\,|\,(z_0,\ldots,z_{d-1},y_d)\in S\}.\]
Since $S$ is a multi-linear subset of $V^{d+1}$, for any $(i_0,\ldots,i_{d-1})\in I_0\times\cdots\times I_{d-1}$, one has $(x_{i_0},\ldots,x_{i_{d-1}},y_d)\in S$ and hence $(x_{i_0},\ldots,x_{i_{d-1}})\in S'$. Moreover, $S'$ is a multi-linear subset of $V^d$. Hence the induction hypothesis leads to $(y_0,\ldots,y_{d-1})\in S'$ and thus $(y_0,\ldots,y_d)\in S$.
\end{proof}
\begin{defi}Let $d\in {\mathbb Z}_{\geqslant 0}$, $V$ and $W$ be two $k$-modules, and $S$ be a multi-linear subset of $V^{d+1}$. We say that a map $f:S\rightarrow W$ is \emph{multi-linear} if, for any $j\in\{0,\ldots,d\}$ and for any $(x_0,\ldots,x_{j-1},x_{j+1},\ldots,x_d)\in V^d$, the map
\[\{x_j\in V\,|\,(x_0,\ldots,x_{j-1},x_j,x_{j+1},\ldots,x_d)\in S\}\longrightarrow W,\quad x_j\mapsto f(x_0,\ldots,x_d),\]
is $k$-linear once \[\{x_j\in V\,|\,(x_0,\ldots,x_{j-1},x_j,x_{j+1},\ldots,x_d)\in S\}\] is not empty. If in addition $S$ is symmetric and $f(x_0,\ldots,x_d)=f(x_{\sigma(0)},\ldots,x_{\sigma(d)})$ for any $(x_0,\ldots,x_d)\in S$ and any bijection $\sigma:\{0,\ldots,d\}\rightarrow\{0,\ldots,d\}$, we say that $f$ is a \emph{symmetric multi-linear map}.
\end{defi}
\begin{prop}\label{pro:domain}
Let $d\in{\mathbb Z}_{\geqslant 0}$, $V$ and $W$ be two $k$-modules, $S$ be a multi-linear subset of $V^{d+1}$, and $f:S\rightarrow W$ be a multi-linear map.
Let $( x_{j,i} )_{(j,i)\in\{0,\ldots,d\}^2}$ be a matrix consisting of elements of $V$
such that $(x_{0,i_0}, \ldots, x_{d,i_d}) \in S$ for any $(i_0, \ldots, i_d) \in \{ 0 , \ldots, d\}^{d+1}$.
Then
\begin{equation}\label{Equ:somme alternee}\begin{split}
\sum_{\sigma \in \mathfrak{S}(\{0, \ldots, d\})} f&(x_{0,\sigma(0)}, \ldots, x_{d,\sigma(d)}) \\
&= \sum_{\emptyset \not= I \subseteq \{ 0, \ldots, d \}} (-1)^{d+1 - \#I} f\Big(\sum\nolimits_{i_0 \in I} x_{0,i_0} , \ldots,\sum\nolimits_{i_d \in I} x_{d,i_d} \Big),\end{split}
\end{equation}
where $\mathfrak{S}(\{0, \ldots, d\})$ is the permutation group of $\{ 0, \ldots, d\}$.
\end{prop}
\begin{proof} By the multi-linearity of $f$, we can rewrite the right-hand side of the equality \eqref{Equ:somme alternee} as
\[\begin{split}&\quad\;\sum_{\varnothing\neq I\subseteq\{0,\ldots,d\}}(-1)^{d+1-\#I}\sum_{(i_0,\ldots,i_d)\in I^{d+1}}f(x_{0,i_0},\ldots,x_{d,i_d})\\
&=\sum_{(i_0,\ldots,i_d)\in\{0,\ldots,d\}^{d+1}}\bigg(\sum_{\{i_0,\ldots,i_d\}\subseteq I\subseteq\{0,\ldots,d\}}(-1)^{d+1-\#I}\bigg)f(x_{0,i_0},\ldots,x_{d,i_d}).
\end{split}\]
Note that, for $(i_0,\ldots,i_d)\in\{0,\ldots,d\}^{d+1}$ such that $\{i_0,\ldots,i_d\}\subsetneq \{0,\ldots,d\}$, one has
\[\sum_{\{i_0,\ldots,i_d\}\subseteq I\subseteq\{0,\ldots,d\}}(-1)^{d+1-\#I}=(-1)^{d+1-\#\{i_0,\ldots,i_d\}}\sum_{J\subseteq\{0,\ldots,d\}\setminus\{i_0,\ldots,i_d\}}(-1)^{-\#J}=0\]
since
\[\sum_{J\subseteq\{0,\ldots,d\}\setminus\{i_0,\ldots,i_d\}}(-1)^{-\#J}=(1+(-1))^{d+1-\#\{i_0,\ldots,i_d\}}=0.\]
Therefore the equality \eqref{Equ:somme alternee} holds.
\end{proof}
\section{Cartier divisors}
In this section, let us recall the notion of Cartier divisor on a general scheme. The main references are \cite[$\text{IV}_4$, \S\S20-21]{EGA} and \cite{MR570309}.
\begin{defi}Let $X$ be a locally ringed space. We denote by $\mathcal O_X$ the structural sheaf of $X$. Let $\mathscr M_X$ be the sheaf of \emph{meromorphic functions on $X$}. Recall that $\mathscr M_X$ is the sheaf of commutative and unitary rings associated with the presheaf
\[U\longmapsto\mathcal O_X(U)[S_X(U)^{-1}],\]
where $S_X(U)$ denotes the multiplicative sub-monoid of $\mathcal O_X(U)$ consisting of local non-zero-divisors of $\mathcal O_X(U)$, that is,
$s\in\mathcal O_X(U)$ such that the homothety
\[\mathcal O_{X,x}\longrightarrow\mathcal O_{X,x},\quad a\longmapsto as_x\]
is injective for any $x\in U$
(here $s_x$ denotes the canonical image of $s$ in the local ring $\mathcal O_{X,x}$). We refer the readers to \cite{MR570309} for a clarification on the construction of the sheaf of meromorphic functions comparing to \cite[$\text{IV}_4$.(20.1.3)]{EGA}.
\end{defi}
\begin{rema}Note that, for any $x\in X$, $\mathscr M_{X,x}$ identifies with $\mathcal O_{X,x}(S_{X,x}^{-1})$, where $S_{X,x}$ denotes the direct limit of $S_X(U)$ with $U$ running over the set of open neighbourhoods of $x$, viewed as a multiplicative submonoid of $\mathcal O_{X,x}$, which is contained in the sub-monoid of non-zero-divisors. Therefore, $\mathscr M_{X,x}$ could be considered as a sub-ring of the total fraction ring of $\mathcal O_{X,x}$, namely the localization of $\mathcal O_{X,x}$ with respect to the set of non-zero-divisors. In general the local ring $\mathscr M_{X,x}$ is different from the ring of total fractions of $\mathcal O_{X,x}$ even if $X$ is an affine scheme. The equality holds notably when $X$ is a locally Noetherian scheme or a reduced scheme whose set of irreducible component is locally finite. We refer the readers to \cite{MR570309} for counter-examples and more details.
\end{rema}
\begin{defi}Let $X$ be a locally ringed space. We denote by $\mathscr M_X^{\times}$ the subsheaf of multiplicative monoids of $\mathscr M_X$ consisting of invertible elements. In other words, for any open subset $U$ of $X$, $\mathscr M_X^{\times}(U)$ is consisting of sections $s\in\mathscr M_X^{\times}(U)$ such that, for any $x\in U$, the homothety
\[\mathscr M_{X,x}\longrightarrow\mathscr M_{X,x},\quad a\longmapsto as_x\]
is an isomorphism of $\mathscr M_{X,x}$-modules. An element of $\mathscr M_X^{\times}(U)$ is called a \emph{regular meromorphic function} on $X$. Similarly, let $\mathcal O_X^{\times}$ be the subsheaf of multiplicative monoids of $\mathcal O_X$ consisting of invertible elements~: for any open subset $U$ of $X$, $\mathcal O_X^{\times}(U)$ consists of sections $s\in\mathcal O_X(U)$ such that, for any $x\in U$, the homothety
\[\mathcal O_{X,x}\longrightarrow\mathcal O_{X,x},\quad a\longmapsto as_x\]
is an isomorphism of $\mathcal O_{X,x}$-modules. Note that, for each $s\in\mathcal O_X(U)$, the homothety $s_x:\mathcal O_{X,x}\rightarrow\mathcal O_{X,x}$ induces by passing to localisation an homothety $\mathscr M_{X,x}\rightarrow\mathscr M_{X,x}$, which is an isomorphism of $\mathscr M_{X,x}$-modules if $s_x:\mathcal O_{X,x}\rightarrow\mathcal O_{X,x}$ is an isomorphism. Therefore, the canonical morphism $\mathcal O_X\rightarrow\mathscr M_X$ induces a morphism of sheaves of abelian groups $\mathcal O_X^{\times}\rightarrow\mathscr M_X^{\times}$.
\end{defi}
\begin{defi}
We call \emph{Cartier divisor} on $X$ any global section of the sheaf $\mathscr M_X^{\times}/\mathcal O_X^{\times}$.
By definition, a Cartier divisor $D$ is represented by the following data: (i) an open covering $X = \bigcup_i U_i$ of $X$ and (ii) $f_i \in \mathscr M_X^{\times}(U_i)$ for each $i$ such that $f_i/f_j \in \mathcal O^{\times}_X$ on $U_i \cap U_j$ for all $i, j$.
The regular meromorphic function $f_i$ is called a \emph{local equation} of $D$ over $U_i$.
The group of Cartier divisors is denoted by $\operatorname{Div}(X)$ and the group law of $\operatorname{Div}(X)$ is written additively. Note that the exact sequence
\[\xymatrix{1\ar[r]&\mathcal O_X^{\times}\ar[r]&\mathscr M_X^{\times}\ar[r]&\mathscr M_X^{\times}/\mathcal O_X^{\times}\ar[r]&0}\]
induces an exact sequence of cohomological groups
\begin{equation}\label{Equ: cohomological exact sequence}\xymatrix@C=1.17pc{1\ar[r]&\Gamma(X,\mathcal O_{X}^{\times})\ar[r]&\Gamma(X,\mathscr M_X^{\times})\ar[r]&\operatorname{Div}(X)\ar[r]&H^1(X,\mathcal O_X^{\times})\ar[r]&H^1(X,\mathscr M_X^{\times})}.\end{equation}
We denote by $\operatorname{div}(\raisebox{.4ex}{.})$ the morphism $\Gamma(X,\mathscr M_X^{\times})\rightarrow\operatorname{Div}(X)$ in this exact sequence. Since the group law of $\operatorname{Div}(X)$ is written additively, one has
\[\operatorname{div}(fg)=\operatorname{div}(f)+\operatorname{div}(g)\]
for any couple of regular meromorphic functions $f$ and $g$ on $X$. A Cartier divisor belonging to the image of $\operatorname{div}(\raisebox{.4ex}{.})$ is said to be \emph{principal}. If $D_1$ and $D_2$ are two Cartier divisors such that $D_1-D_2$ is principal, we say that $D_1$ and $D_2$ are \emph{linearly equivalent}, denoted by $D_1\sim D_2$.
\end{defi}
\begin{rema}Recall that $H^1(X,\mathcal O_X^{\times})$ identifies with the Picard group $\operatorname{Pic}(X)$ of $X$, namely the group of isomorphism classes of invertible $\mathcal O_X$-modules (see \cite[{\bf0}.(5.6.3)]{MR3075000}). Similarly, $H^1(X,\mathscr M_X^{\times})$ identifies with the group of isomorphism classes of invertible $\mathscr M_X$-modules. If $L$ is an invertible $\mathcal O_X$-module, then $\mathscr M_X\otimes_{\mathcal O_X}L$ is an invertible $\mathscr M_X$-module. The homomorphism $H^1(X,\mathcal O_X^{\times})\rightarrow H^1(X,\mathscr M_X^{\times})$ sends the isomorphism class of an invertible $\mathcal O_X$-module $L$ to that of the invertible $\mathscr M_X$-module $\mathscr M_X\otimes_{\mathcal O_X}L$.
\end{rema}
\begin{defi}
Let $L$ be an invertible $\mathcal O_X$-module and $U$ be a non-empty open subset of $X$. We call \emph{regular meromorphic section} of $L$ on $U$ any element of $\Gamma(U,\mathscr M_X\otimes_{\mathcal O_X}L)$ which defines an isomorphism from $\mathscr M_U$ to $\mathscr M_U\otimes_{\mathcal O_U}L|_U$. Therefore, $\mathscr M_X\otimes_{\mathcal O_X}L$ is isomorphic as $\mathscr M_X$-module to $\mathscr M_X$ if and only if $L$ admits a regular meromorphic section on $X$.
\end{defi}
\begin{rema}\label{Rem: regular meromorphic section}
Let $X$ be a locally Noetherian scheme or a reduced scheme whose set of irreducible component is locally finite. For any $x\in X$, the local ring $\mathscr M_{X,x}$ identifies with the ring of total fractions of $\mathcal O_{X,x}$. Therefore, if $L$ is an invertible $\mathcal O_X$-module and if $U$ is an open subset of $X$, an element $s\in \Gamma(U,\mathscr M_X\otimes_{\mathcal O_X}L)$ is a regular meromorphic section of $L$ on $U$ if and only if it defines an injective homomorphism from $\mathcal O_U$ to $\mathscr M_U\otimes_{\mathcal O_U}L$. In particular, an element $s\in\Gamma(U,L)$ defines a regular meromorphic section of $L$ on $U$ if and only if, for any $x\in U$, $s_x\in\mathcal O_{X,x}\otimes_{\mathcal O_X}L$ is of the form $f_xs_{0,x}$, where $f_x$ is a non-zero-divisor of $\mathcal O_{X,x}$, and $s_{0,x}$ is a local trivialization of $L$ at $x$. This condition is also equivalent to $s(y)\neq 0$ for any associate point $y\in U$. Recall that a point $y\in X$ is called an \emph{associated point} if there exists $a\in\mathcal O_{X,y}$ such that the maximal ideal of $\mathcal O_{X,y}$ identifies with \[\operatorname{ann}(a):=\{f\in\mathcal O_{X,y}\,|\,af=0\}.\]
Let $x$ be a point of $X$. Assume that $s_x=f_xs_{0,x}$ where $f_x$ is a zero-divisor in $\mathcal O_{X,x}$, then $f_x$ belongs to an associated prime ideal of $\mathcal O_{X,x}$, which corresponds to an associated point $y\in X$ such that $x\in\overline{\{y\}}$ and $s(y)=0$.
By \cite[$\text{IV}_4.(21.3.5)$]{EGA}, if $X$ is a Noetherian scheme, which admits an ample invertible $\mathcal O_X$-module, then the set of all associated points of $X$ is contained in an affine open subset of $X$, and any invertible $\mathcal O_X$-module admits a regular meromorphic section.
\end{rema}
\begin{defi}\label{Def: divisor of a section}
Let $D$ be a Cartier divisor on $X$. The homomorphism $\operatorname{Div}(X)\rightarrow H^1(X,\mathcal O_X^{\times})$ in the exact sequence \eqref{Equ: cohomological exact sequence} sends $D$ to an isomorphism class of invertible $\mathcal O_X$-modules. One can actually construct explicitly an invertible $\mathcal O_X$-module $\mathcal O_X(D)$ in this class as follows. Let $(U_i)_{i\in I}$ be an open cover of the topological space such that $D$ is represented on each $U_i$ by a regular meromorphic function $f_i\in\Gamma(U_i,\mathscr M_{U_i}^{\times})$. For any couple $(i,j)\in I^2$, $f_i|_{U_i\cap U_j}f_j|_{U_i\cap U_j}^{-1}$ defines an isomorphism\[(f_i^{-1}\mathcal O_{U_i})|_{U_i\cap U_j}\longrightarrow(f_j^{-1}\mathcal O_{U_j})|_{U_i\cap U_j}.\]Moreover, these isomorphisms clearly satisfy the cocycle condition. Thus the gluing of the sheaves $f_i^{-1}\mathcal O_{U_i}$ leads to an invertible sub-$\mathcal O_X$-module of $\mathscr M_X$ which we denote by $\mathcal O_X(D)$. Note that the gluing of meromorphic sections \[f_i\otimes f_i^{-1}\in\Gamma(U_i,\mathscr M_{U_i}\otimes\mathcal O_X(D))\] leads to a global regular meromorphic section of $\mathcal O_X(D)$, which we denote by $s_D$ and call \emph{canonical regular meromorphic section} of $\mathcal O_X(D)$. Hence $\mathscr M_X\otimes_{\mathcal O_X}\mathcal O_X(D)$ is canonically isomorphic to $\mathscr M_X$. Note that two Cartier divisors $D_1$ and $D_2$ are linearly equivalent if and only if the invertible $\mathcal O_X$-modules $\mathcal O_X(D_1)$ and $\mathcal O_X(D_2)$ are isomorphic.
Conversely, the exactness of the diagram \eqref{Equ: cohomological exact sequence} shows that, an invertible $\mathcal O_X$-module $L$ is isomorphic to an invertible $\mathcal O_X$-module of the form $\mathcal O_X(D)$ if and only if it admits a regular meromorphic section on $X$. One can also construct explicitly a Cartier divisor from a regular meromorphic section $s$ of $L$. In fact, let $(U_i)_{i\in I}$ be an open cover of $X$ such that each $L|_{U_i}$ is trivialized by a section $s_i\in L(U_i)$. For any $i\in I$, let $f_i$ be the unique regular meromorphic function on $U_i$ such that $s=f_is_i$. Then the family $(f_i)_{i\in I}$ of regular meromorphic functions defines a Cartier divisor on $X$ which we denote by $\operatorname{div}(L;s)$, or by $\operatorname{div}(s)$ for simplicity.
\end{defi}
\begin{rema}
In the case where $X$ is a quasi-projective scheme over a field, any invertible $\mathcal O_X$-module admits a global regular meromorphic section and therefore is isomorphic to an invertible $\mathcal O_X$-module of the form $\mathcal O_X(D)$, where $D$ is a Cartier divisor. Hence one has an exact sequence
\[\xymatrix{1\ar[r]&\Gamma(X,\mathcal O_{X}^{\times})\ar[r]&\Gamma(X,\mathscr M_X^{\times})\ar[r]&\operatorname{Div}(X)\ar[r]&H^1(X,\mathcal O_X^{\times})\ar[r]&1}.\]
\end{rema}
\begin{rema}\label{remark:Cartier:div:on:0:dim}
Let $X$ be a $0$-dimensional projective scheme over a field $k$.
Then there is a $k$-algebra $A$ which is finite-dimensional as a vector space over $k$, and such that
$X = \operatorname{Spec}(A)$.
Note that the canonical homomorphism
$A \to \bigoplus_{x \in X} A_x$ is an isomorphism.
Let $f_x$ be a regular element of $A_x$. As the homotethy map $A_x \to A_x$, $a\mapsto f_x a$, is injective and $A_x$ is a finite-dimensional vector space over $k$, this homothety map is actually an isomorphism, that is,
$f_x \in A^{\times}_x$. Thus $\mathscr M_X^{\times} = \mathcal O_{X}^{\times}$.
Therefore, every Cartier divisor on $X$ can be represented by $1 \in A$.
\end{rema}
\begin{rema}\label{remark:expansion:Cartier:div:as:cycle}
Let $X$ be a Noetherian scheme. We denote by $X^{(1)}$ the set of all height 1 points of $X$, that is,
$x \in X$ with $\dim (\mathcal O_{X, x}) = 1$. For $x \in X^{(1)}$ and a regular element $f$ of $\mathcal O_{X, x}$,
we set \[\operatorname{ord}_x(f) := \operatorname{length}_{\mathcal O_{X, x}}(\mathcal O_{X, x}/f\mathcal O_{X, x}).\]
Then $\operatorname{ord}_x(fg) = \operatorname{ord}_x(f) + \operatorname{ord}_x(g)$ for
all regular elements $f, g$ of $\mathcal O_{X, x}$ (cf \cite[the last paragraph of Section~1.3]{MArakelov}), so that
$\operatorname{ord}_x$ extends to a homomorphism $\mathscr M^{\times}_{X,x} \to \mathbb Z$.
Let $D$ be a Cartier divisor on $X$ and $f$ a local equation of $D$ at $x$.
Then it is easy to see that $\operatorname{ord}_{x}(f)$ does not depend on the choice of $f$, so that $\operatorname{ord}_x(f)$ is denoted by $\operatorname{ord}_x(D)$.
We call the cycle
\[\sum_{x \in X^{(1)}} \operatorname{ord}_x(D) \overline{\{ x \}}\]
\emph{the cycle associated with $D$}, which is denoted by $z(D)$.
Let $X_1, \ldots, X_{\ell}$ be the irreducible components of $X$ and $\eta_1, \ldots, \eta_{\ell}$
be the generic points of $X_1, \ldots, X_{\ell}$, respectively.
Then
\begin{equation}\label{eqn:remark:expansion:Cartier:div:as:cycle:01}
z(D) = \sum_{j=1}^{\ell} \operatorname{length}_{{\mathscr{O}}_{X, \eta_j}}({\mathscr{O}}_{X, \eta_j}) z(\rest{D}{X_j}).
\end{equation}
Indeed, by \cite[(6) of Lemma~1.7]{MArakelov}, $\operatorname{ord}_x(D) = \sum_{j \in J_x} b_j \operatorname{ord}_x(\rest{D}{X_j})$, where $b_j =
\operatorname{length}_{{\mathscr{O}}_{X, \eta_j}}({\mathscr{O}}_{X, \eta_j})$ and $J_x = \{ j \mid x \in X_j \}$. Thus if we set
\[
a_{x, j} = \begin{cases} \operatorname{ord}_x(\rest{D}{X_j}) & \text{if $x \in X_j$}, \\
0 & \text{if $x\not\in X_j$},
\end{cases}
\]
then $\operatorname{ord}_x(D) = \sum_{j=1}^{\ell} a_{x, j} b_j$. Thus
\begin{align*}
z(D) & = \sum_{x \in X^{(1)}} \operatorname{ord}_x(D) \overline{\{ x \}} = \sum_{x \in X^{(1)}} \Big( \sum_{j=1}^{\ell} a_{x, j} b_j \Big)\overline{\{ x \}} \\
& = \sum_{j=1}^{\ell} b_j \sum_{x \in X^{(1)}} a_{x, j} \overline{\{ x \}} = \sum_{j=1}^{\ell} b_j \sum_{x \in X_j^{(1)}} \operatorname{ord}_x(\rest{D}{X_j}) \overline{\{ x \}} =
\sum_{j=1}^{\ell} b_j z(\rest{D}{Z_j}),
\end{align*}
as required.
Let $L$ be an invertible ${\mathscr{O}}_X$-module and $s$ a regular meromorphic section of $L$ over $X$.
For $x \in X^{(1)}$, $\operatorname{ord}_x(s)$ is defined by $\operatorname{ord}_x(f)$, where $f$ is given by $s = f \omega$ for some local basis $\omega$ of $L$ around $x$.
Note that $\operatorname{ord}_x(s)$ does not depend on the choice of the local basis $\omega$ around $x$.
Then the cycle $z(L; s)$ associated with $\operatorname{div}(L; s)$ is defined by
\[ z(L;s) := \sum_{x \in X^{(1)}} \operatorname{ord}_x(s) \overline{\{ x \}}.\]
\end{rema}
{
\begin{defi}
Let $\varphi:X\rightarrow Y$ be a morphism of locally ringed space. If $U$ is an open subset of $Y$, we denote by $S_\varphi(U)$ the preimage of $S_X(\varphi^{-1}(U))$ by the structural ring homomorphism
\[\mathcal O_Y(U)\longrightarrow\mathcal O_X(\varphi^{-1}(U)).\]
We denote by $\mathscr M_\varphi$ the sheaf of commutative and unitary rings associated with the presheaf
\[U\longmapsto \mathcal O_Y(U)[S_\varphi(U)^{-1}].\]
It is a subsheaf of $\mathscr M_Y$. Moreover, the structural morphism of sheaves $\mathcal O_Y\rightarrow \varphi_*(\mathcal O_X)$ induces by localization a morphism $\mathscr M_\varphi\rightarrow \varphi_*(\mathscr M_X)$, which defines a morphism of locally ringed spaces $(X,\mathscr M_X)\rightarrow (Y,\mathscr M_\varphi)$.
\end{defi}
\begin{rema}\label{Rem: pull back}
There are several situations in which $\mathscr M_\varphi$ identifies with $\mathscr M_Y$, notably when one of the following conditions is satisfied (see \cite[$\text{IV}_4.(21.4.5)$]{EGA}):
\begin{enumerate}[label=\rm(\arabic*)]
\item $\varphi$ is flat, namely for any $x\in X$, the morphism of rings $\varphi_x:\mathcal O_{Y,\varphi(x)}\rightarrow\mathcal O_{X,x}$ defines a structure of flat $\mathcal O_{Y,\varphi(x)}$-algebra on $\mathcal O_{X,x}$,
\item $X$ and $Y$ are locally Noetherian schemes, and $f$ sends any associated point of $X$ to an associated point of $Y$,
\item $X$ and $Y$ are schemes, the set of irreducible components of $Y$ is locally finite, $X$ is reduced, and any irreducible component of $X$ dominates an irreducible component of $Y$.
\end{enumerate}
\end{rema}
\begin{defi}\label{Def: pull back}
Let $\varphi:X\rightarrow Y$ be a morphism of locally ringed spaces, and $D$ be a Cartier divisor on $Y$. Assume that both $D$ and $-D$ are global sections of $(\mathscr M_Y^{\times}\cap\mathscr M_\varphi)/\mathcal O_X^{\times}$, or equivalently, for any local equation $f$ of $D$ over an open subset $U$ of $Y$, one has $\{f,f^{-1}\}\subset\mathscr M_\varphi(U)$. Then the canonical regular meromorphic section $s_D$ of $\mathcal O_Y(D)$ actually defines an isomorphism
\[\mathscr M_\varphi\longrightarrow \mathscr M_\varphi\otimes_{\mathcal O_Y}\mathcal O_Y(D).\]
which induces an isomorphisme
\[\varphi^*(s_D):\mathscr M_X\longrightarrow\mathscr M_{X}\otimes_{\mathcal O_X}\varphi^*(\mathcal O_Y(D)).\]
We denote by $\varphi^*(D)$ the Cartier divisor $\operatorname{div}(\varphi^*(\mathcal O_Y(D));\varphi^*(s_D))$ corresponding to this regular meromorphic section, and call it the \emph{pull-back} of $D$ by $\varphi$. In the case where $\varphi$ is an immersion, the Cartier divisor $\varphi^*(D)$ is also denoted by $D|_X$.
\end{defi}
}
Finally let us consider the following lemma.
\begin{lemm}\label{lem:eq:regular:mero:generic:total}
Let $\mathfrak o$ be an integral domain, $A$ be an $\mathfrak o$-algebra and $S := \mathfrak o \setminus \{ 0 \}$.
If $A$ is flat over $\mathfrak o$, then we have the following:
\begin{enumerate}[label=\rm(\arabic*)]
\item For $s \in S$, the homomorphism $s \cdot : A \to A$ given by $a \mapsto s \cdot a$ is injective.
In particular, the structure homomorphism $\mathfrak o \to A$ is injective, so that in the following,
$\mathfrak o$ is considered as a subring of $A$.
\item The natural homomorphism $A \to A_S$ is injective.
\item For $a \in A$, $a$ is a non-zero-divisor in $A$ if and only if $a/1$ is a non-zero-divisor in $A_S$.
In particular, a non-zero-divisor of $A_S$ can be written by a form $a/s$ where
$a$ is a non-zero-divisor of $A$ and $s \in S$.
\item Let $Q(A)$ and $Q(A_S)$ be the total quotient rings of $A$ and $A_S$, respectively.
The homomorphism $Q(A) \to Q(A_S)$ induced by $A \to A_S$ is well-defined and bijective.
In particular, $Q(A)^{\times} = Q(A_S)^{\times}$.
\end{enumerate}
\end{lemm}
\begin{proof}
(1) is obvious because $\mathfrak o$ is an integral domain and $A$ is flat over $\mathfrak o$.
(2) follows from (1).
(3) The assertion follows from (1) and the following commutative diagram:
\[
\begin{CD}
A @>>> A_S \\
@V{a\cdot}VV @VV{a\cdot}V \\
A @>>> A_S
\end{CD}
\]
(4) By (3), if $a \in A$ is a non-zero-divisor, then $a/1$ is a non-zero-divisor in $A_S$, so that
$Q(A) \to Q(A_S)$ is well-defined. The injectivity of $Q(A) \to Q(A_S)$ follows from (2).
For its surjectivity, observe the following:
\[
\frac{b/t}{a/s} = \frac{(st/1)(b/t)}{(st/1)(a/s)} = \frac{sb/1}{ta/1}.
\]
\end{proof}
\section{Proper intersection}
Let $d$ be a non-negative integer and $X$ be a $d$-dimensional scheme of finite type over a field $k$.
Let $D$ be a Cartier divisor on $X$. We define the support of $D$ to be
\[
\operatorname{Supp}(D) := \{ x \in X \mid f_x \not\in \mathcal O_{X, x}^{\times} \},
\]
where $f_x$ is a local equation of $D$ at $x$. Note that the above definition does not depend on the choice of $f_x$ since two local equations of $D$ at $x$ differ by a factor in $\mathcal O_{X,x}^{\times}$.
\begin{prop}\label{prop:Supp:closed}
\begin{enumerate}[label=\rm(\arabic*)]
\item $\operatorname{Supp}(D)$ is a Zariski closed subset of $X$.
\item $\operatorname{Supp}(D + D') \subseteq \operatorname{Supp}(D) \cup \operatorname{Supp}(D')$.
\end{enumerate}
\end{prop}
\begin{proof}
(1) Clearly we may assume that $X$ is affine and $D$ is principal, that is,
$X = \operatorname{Spec}(A)$ and $D$ is defined by a regular meromorphic function $f$ on $X$, which could be considered as an element of the total fraction ring of $A$ (that is, the localization of $A$ with respect to the subset of non-zero-divisors). By \cite{MR570309}, for any prime ideal $\mathfrak p$ of $A$, there is a canonical ring homomorphism from the total fraction ring of $A$ to that of $A_{\mathfrak p}$.
We set $\mathfrak a = \{a \in A \mid af \in A \}$ and $\mathfrak b = \mathfrak af$. Then $\mathfrak a$ and $\mathfrak b$ are ideals of $A$. Note that, for $\mathfrak p \in \operatorname{Spec}(A)$,
\[\mathfrak a_{\mathfrak p} = \{ u \in A_{\mathfrak p} \mid u f \in A_{\mathfrak p} \}.\] In fact, clearly one has $\mathfrak a_{\mathfrak p}\subseteq\{u\in A_{\mathfrak p}\,|\,uf\in A_{\mathfrak p}\}$. Conversely, if $u=a/s$ (with $a\in A$ and $s\in A\setminus\mathfrak p$) is an element of $A_{\mathfrak p}$ such that $uf\in A_{\mathfrak p}$, then there exists $t\in A\setminus\mathfrak p$ such that $at\in\mathfrak a$ and hence $u=at/st\in\mathfrak a_{\mathfrak p}$.
Thus
\[
\mathfrak p \not\in \operatorname{Supp}(D) \Longleftrightarrow f \in A_\mathfrak p^{\times} \Longleftrightarrow \text{$\mathfrak a_{\mathfrak p} = A_{\mathfrak p}$ and $\mathfrak b_{\mathfrak p} = A_{\mathfrak p}$}
\Longleftrightarrow \mathfrak p \not\in V(\mathfrak a) \cup V(\mathfrak b),
\]
that is, $\operatorname{Supp}(D) = V(\mathfrak a) \cup V(\mathfrak b)$, as desired.
\medskip
(2) Let $f_x$ and $f'_x$ be local equations of $D$ and $D'$ at $x$, respectively.
Then
\begin{align*}
x \not\in \operatorname{Supp}(D) \cup \operatorname{Supp}(D') & \Longrightarrow f_x, f'_x \in \mathcal O^{\times}_{X, x} \Longrightarrow
f_xf'_x \in \mathcal O^{\times}_{X, x} \\
& \Longrightarrow x \not\in \operatorname{Supp}(D + D'),
\end{align*}
as required.
\end{proof}
\begin{defi}\label{def:meet:properly}
Let $n$ be an integer such that $0 \leqslant n \leqslant d$.
Let $D_0, \ldots, D_n$ be Cartier divisors on $X$.
We say that \emph{$D_0, \ldots, D_n$ intersect properly} if, for any non-empty subset $J$ of $\{ 0, \ldots, n\}$,
\[\dim \Big( \bigcap\nolimits_{j \in J} \operatorname{Supp}(D_j) \Big) \leqslant d - \operatorname{card}(J).\] By convention, $\dim(\emptyset)$ is defined to be $-1$.
We set
\[
\mathcal{IP}_X^{(n)} := \{ (D_0, \ldots, D_n) \in \operatorname{Div}(X)^{n+1} \mid \text{$D_0, \ldots, D_n$ intersect properly} \}.
\]
In the case where $n=d$, we often denote $\mathcal{IP}_X^{(n)}$ by $\mathcal{IP}_X$.
\end{defi}
\begin{lemm}\label{lemm:equality:supp:extension:field}
Let $k'/k$ be an extension of fields.
Let $A$ be a $k$-algebra and $A' := A \otimes_k k'$.
Let $\pi : \operatorname{Spec}(A') \to \operatorname{Spec}(A)$ be the morphism induced by the natural homomorphism $A \to A'$.
Let $Q(A)$ (resp. $Q(A')$) be the total quotient ring of $A$ (resp. $A'$).
Let $\alpha \in Q(A)^{\times}$ and $\alpha' := \alpha \otimes_k 1 \in Q(A) \otimes_k k'$.
If we set
\[
\begin{cases}
\operatorname{Supp}(\alpha) := \{ P \in \operatorname{Spec}(A) \mid \alpha \not\in A_P^{\times} \}, \\
\operatorname{Supp}(\alpha') := \{ P' \in \operatorname{Spec}(A') \mid \alpha' \not\in {A'}_{P'}^{\times} \},
\end{cases}
\]
then $\operatorname{Supp}(\alpha') = \pi^{-1}(\operatorname{Supp}(\alpha))$.
\end{lemm}
\begin{proof}
First of all, note that $Q(A) \otimes_k k' \subseteq Q(A')$ and $\alpha' \in (Q(A) \otimes_k k')^{\times} \subseteq Q(A')^{\times}$ because $\pi$ is flat.
Let
$I := \{ a \in A \mid a \alpha \in A \}$, $J:= I \alpha$,
$I':= \{ a' \in A' \mid a' \alpha' \in A' \}$ and $J':= I' \alpha'$.
Then one has the following.
\begin{enonce}{Claim}\label{Claim:lemm:equality:supp:extension:field:01}
\begin{enumerate}[label=\rm(\arabic*)]
\item $\operatorname{Supp}(\alpha) = \operatorname{Spec}(A/I) \cup \operatorname{Spec}(A/J)$ and $\operatorname{Supp}(\alpha') = \operatorname{Spec}(A'/I') \cup \operatorname{Spec}(A'/J')$.
\item $I' = I \otimes_k k'$ and $J'=J \otimes_k k'$.
\item $\operatorname{Spec}(A'/I') = \pi^{-1}(\operatorname{Spec}(A/I))$ and $\operatorname{Spec}(A'/J') = \pi^{-1}(\operatorname{Spec}(A/J))$.
\end{enumerate}
\end{enonce}
\begin{proof}
Let $\{ x_\lambda \}_{\lambda\in\Lambda}$ be a basis of $k'$ over $k$.
Note that $V \otimes_k k' = \bigoplus_{\lambda \in \Lambda} V \otimes_k k x_{\lambda}$ for a $k$-module $V$.
(1) It is sufficient to prove the first equality. The second is similar to the first.
Note that $I_P = \{ a \in A_P \mid a \alpha \in A_P^{\times} \}$.
Thus, if $\alpha \in A_P^{\times}$, then $I_P = J_P = A_P$, so that $P \not\in \operatorname{Spec}(A/I) \cup \operatorname{Spec}(A/J)$.
Conversely, we assume that $P \not\in \operatorname{Spec}(A/I) \cup \operatorname{Spec}(A/J)$, that is, $I \not\subseteq P$ and
$J \not\subseteq P$. Thus $I_P = J_P = A_P$, and hence $\alpha \in A_P^{\times}$.
(2)
Obviously $I \otimes_k k' \subseteq I'$. We assume $a' \in I'$.
Then there are $a_{\lambda}$'s such that $a_{\lambda} \in A$ and $a' = \sum_{\lambda} a_{\lambda} \otimes x_{\lambda}$. By our assumption,
we can find $b_{\lambda}$'s such that $b_{\lambda} \in A$ and
\[
\sum\nolimits_{\lambda} a_{\lambda} \alpha \otimes x_{\lambda} = a' \alpha' = \sum\nolimits_{\lambda} b_{\lambda} \otimes x_{\lambda},
\]
so that $a_{\lambda}\alpha = b_{\lambda} \in A$ for all $\lambda$. Thus $a_{\lambda} \in I$.
Therefore the first assertion follows. The second is a consequence of the first.
(3) follows from (2).
\end{proof}
By using (1) and (3) of the above claim,
\begin{align*}
\pi^{-1}(\operatorname{Supp}(\alpha)) & = \pi^{-1}(\operatorname{Spec}(A/I) \cup \operatorname{Spec}(A/J)) \\
& = \pi^{-1}(\operatorname{Spec}(A/I)) \cup \pi^{-1}(\operatorname{Spec}(A/J)) \\
& = \operatorname{Supp}(A'/I') \cup \operatorname{Supp}(A'/J') = \operatorname{Supp}(\alpha'),
\end{align*}
as required.
\end{proof}
\begin{rema}\label{Rem: extension of scalars}
Let $k'/k$ be an extension of fields, $X_{k'}=X\times_{\operatorname{Spec} k}\operatorname{Spec} k'$ and $\pi:X_{k'}\rightarrow X$ be the morphism of projection. Since the canonical morphism $\operatorname{Spec} k'\rightarrow\operatorname{Spec} k$ is flat, also is the morphism of projection $\pi$ (see \cite[$\text{IV}_1.(2.1.4)$]{EGA}). Therefore, for any Cartier divisor $D$ on $X$, the pull-back $\pi^*(D)$ is well defined as a Cartier divisor on $X_{k'}$, which we denote by $D_{k'}$.
By Lemma~\ref{lemm:equality:supp:extension:field}, one has
\[\operatorname{Supp}(D_{k'})=\pi^{-1}(\operatorname{Supp}(D)).\]
In particular, if $D_0,\ldots,D_n$ are Cartier divisors on $X$, which intersect properly, then, for any subset $J$ of $\{0,\ldots,n\}$, one has (see for example \cite[Proposition 5.38]{MR2675155} for the equality in the middle)
\[\begin{split}\dim \Big( \bigcap\nolimits_{j \in J} \operatorname{Supp}(D_{j,k'}) \Big)&= \dim \Big( \pi^{-1}\Big(\bigcap\nolimits_{j \in J} \operatorname{Supp}(D_j) \Big)\Big)\\&=\dim \Big( \bigcap\nolimits_{j \in J} \operatorname{Supp}(D_j) \Big)\leqslant d-\operatorname{card}(J).
\end{split}\]
Therefore, the Cartier divisors $D_{0,k'},\ldots,D_{n,k'}$ on $X_{k'}$ intersect properly.
\end{rema}
\begin{lemm}\label{lem:meet:properly}
The set $\mathcal{IP}_X^{(n)}$ forms a symmetric and multi-linear subset of $\operatorname{Div}(X)^{n+1}$ in the sense of Definition~\ref{def:domain}.
\end{lemm}
\begin{proof}
It is sufficient to show that if $(D_0, D_1, \ldots, D_n), (D'_0, D_1, \ldots, D_n) \in \mathcal{IP}_X^{(n)}$, then
$(D_0 + D'_0, D_1, \ldots, D_n) \in \mathcal{IP}_X^{(n)}$. We set
\[
D''_i = \begin{cases}
D_0 + D'_0, & \text{if $i = 0$}, \\
D_i, & \text{if $i \geqslant 1$}.
\end{cases}
\]
If $(D''_0, D''_1, \ldots, D''_n) \not\in \mathcal{IP}_X^{(n)}$, then there is a non-empty subset $J$ of $\{ 0, \ldots, n \}$ such that
\[
\dim \Big( \bigcap\nolimits_{j \in J} \operatorname{Supp}\big(D''_{j} \big) \Big) > d - \#(J).
\]
Clearly $0 \in J$.
We can find a schematic point $P \in X$ such that $\dim \overline{\{ P \}} > d - \#(J)$ and
$P \in \operatorname{Supp}\big(D''_{j} \big)$ for all $j \in J$, so that $P \in \operatorname{Supp}(D_0 + D'_0)$ and $P \in \operatorname{Supp}(D_j)$ for $j \in J \setminus \{ 0 \}$.
Thus, by Proposition~\ref{prop:Supp:closed}, $P \in \operatorname{Supp}(D_0)$ or $P \in \operatorname{Supp}(D'_0)$, which is a contradiction.
\end{proof}
\begin{lemm}\label{lem:IP:projective:ample}
We assume that $X$ is projective. Let $n$ be an integer such that $0 \leqslant n \leqslant d$.
\begin{enumerate}[label=\rm(\arabic*)]
\item
Let $L_0, \ldots, L_n$ be invertible $\mathcal O_X$-modules.
Then there are regular meromorphic sections $s_0, \ldots, s_n$ of $L_0, \ldots, L_n$,
respectively, such that, if we set $D_i = \operatorname{div}(L_i; s_i)$ for $i \in \{ 0, \ldots, n\}$,
then $D_0, \ldots, D_n$ intersect properly.
\item
If $(D_0, D_1, \ldots, D_n), (D'_0, D'_1, \ldots, D'_n)
\in \mathcal{IP}^{(n)}_X$ and $D_0 \sim D'_0$, then there is $D''_0$ such that
$D''_0 \sim D_0 \,(\sim D'_0)$ and $(D''_0, D_1, \ldots, D_n), (D''_0, D'_1, \ldots, D'_n)
\in \mathcal{IP}_X^{(n)}$.
\end{enumerate}
\end{lemm}
\begin{proof}
(1) We prove it by induction on $n$ in incorporating the proof of the initial case in the induction procedure.
By the hypothesis of induction (when $n\geqslant 1$),
there are regular meromorphic sections $s_0, \ldots, s_{n-1}$ of $L_0, \ldots, L_{n-1}$,
respectively, such that if we set $D_i = \operatorname{div}(L_i; s_i)$ for $i \in \{ 0, \ldots, n-1\}$,
then $D_0, \ldots, D_{n-1}$ intersect properly.
We now introduce the following claim, which (in the case where $n=0$) also prove the initial case of induction.
\begin{enonce}{Claim}\label{Claim: decomposition}
There exist very ample invertible $\mathcal O_X$-modules $L'_n$ and $L''_n$, and global sections $s'_n$ and $s''_n$ of $L'_n$ and $L''_n$, which satisfy the following conditions~:
\begin{enumerate}[label=\rm(\roman*)]
\item $L_n = L'_n \otimes {L''_n}^{-1}$,
\item $s'_n$ and $s''_n$ define regular meromorphic sections of $L_n'$ and $L_n''$, respectively,
\item\label{Item intersection properly} if we set $D'_n = \operatorname{div}(L'_n; s'_n)$ and $D''_n = \operatorname{div}(L''_n; s''_n)$,
then both families of Cartier divisors $D_0, \ldots, D_{n-1}, D'_n$ and $D_0, \ldots, D_{n-1}, D''_n$
intersect properly.
\end{enumerate}
\end{enonce}
\begin{proof}[Proof of Claim \ref{Claim: decomposition}]
Since $X$ is projective, there exists a very ample $\mathcal O_X$-module $L$. By \cite[II.(4.5.5)]{EGA}, there exists an integer $\ell_0\in\mathbb N_{\geqslant 1}$ such that both invertible $\mathcal O_X$-modules $L^{\otimes \ell_0}$ and $L^{\otimes\ell_0}\otimes L_n^{-1}$ are generated by global sections. Let $\Sigma$ be the set of generic points of
\[\bigcap_{i=0}^{n-1}\operatorname{Supp}(D_i).\]
We equip the set $\Sigma\cup\operatorname{Ass}(X)$ with the order $\succ$ of generalization, namely $x\succ y$ if and only if $y$ belongs to the Zariski closure of $\{x\}$. We denote by $\{y_1,\ldots,y_b\}$ the set of all minimal elements of the set $\Sigma\cup\operatorname{Ass}(X)$ .
For any $i\in\{1,\ldots,b\}$, one has
\[y_i\in X\setminus\bigcup_{\begin{subarray}{c}j\in\{1,\ldots,b\}\\
j\neq i\end{subarray}}\overline{\{y_j\}}.\]
By \cite[II.(4.5.4)]{EGA}, for any $i\in\{1,\ldots,b\}$, there exists $\ell_i\in\mathbb N_{\geqslant 1}$ and a section $t_i\in H^0(X,L^{\otimes n})$ such that $t_i(y_i)\neq 0$ and that $t_i(y_j)=0$ for any $j\in\{1,\ldots,b\}\setminus\{i\}$. Moreover, by replacing the global sections $t_1,\ldots,t_b$ by suitable powers, we may assume, without loss of generality, that all $\ell_1,\ldots,\ell_b$ are equal to a positive integer $\ell$. For any $i\in\{1,\ldots,b\}$, let $u_i\in H^0(X,L^{\otimes{\ell_0}})$ and $v_i\in H^0(X,L^{\otimes\ell_0}\otimes L_n^{-1})$ be such that $u_i(y_i)\neq 0$ and $v_i(y_i)\neq 0$. These sections exist since the invertible $\mathcal O_X$-modules $L^{\otimes\ell_0}$ and $L^{\otimes\ell_0}\otimes L_n^{-1}$ are generated by global sections. Now we take
\[L_n'=L^{\otimes(\ell_0+\ell)},\qquad L_n''=L_n'\otimes L_n^{-1}=(L^{\ell_0}\otimes L_n^{-1})\otimes L^{\otimes\ell}, \]
and
\[s_n'=\sum_{i=1}^bu_it_i,\qquad s_n''=\sum_{i=1}^bv_it_i.\]
Then, for any $i\in\{1,\ldots,b\}$, one has $s_n'(y_i)\neq 0$ and $s_n''(y_i)\neq 0$. In particular, $s_n'$ and $s_n''$ do not vanish on any of the associated points of $X$ and hence are regular meromorphic sections (see Remark \ref{Rem: regular meromorphic section}). Moreover, since these sections do not vanish on any point of $\Sigma$, we obtain the condition \ref{Item intersection properly} above.
\end{proof}
Thus, by Lemma \ref{lem:meet:properly},
we can see that $D_1, \ldots, D_{n-1}, D_n$ intersect properly,
where $D_n = D'_n - D''_n = \operatorname{div}(L_n; s_n \otimes {s_n}^{-1})$,
as required.
\medskip
(2) We can find very ample Cartier divisors $A$ and $B$ on $X$ such that $D_0 = A - B$.
Then, by the same argument as the induction procedure in the proof of (1), we obtain that there are $A'$ and $B'$ such that $A' \sim A$, $B' \sim B$ and
\[(A', D_1, \ldots, D_n), (A', D'_1, \ldots, D'_n),(B', D_1, \ldots, D_n), (B', D'_1, \ldots, D'_n)
\in \mathcal{IP}_X^{(n)}.\]
Thus if we set $D''_0 = A' - B'$, then, by Lemma~\ref{lem:meet:properly}, one has the conclusion.
\end{proof}
\begin{rema}\label{Rem: further explanation to Claim}
Claim \ref{Claim: decomposition} has its own interest and will be used in further chapters in the following way. Let $X$ be a $d$-dimensional projective scheme over $\operatorname{Spec} k$ and $D_0,\ldots,D_d$ be Cartier divisors on $X$. We suppose that $D_0,\ldots, D_d$ intersect properly. Let $D_0=A_0-A_0'$ be a decomposition of $D_0$ into the difference of two very ample Cartier divisors. A priori $A_0,D_1,\ldots,D_d$ do not intersect properly. However, by Claim \ref{Claim: decomposition}, one can find a very ample invertible $\mathcal O_X$-module $L$ and a global section $s$ of $L\otimes\mathcal O_X(A_0)$ defining a regular meromorphic section, such that $\operatorname{div}(L;s),D_1,\ldots,D_d$ intersect properly. Let $B=\operatorname{div}(L;s)-A_0$. This is a very ample Cartier divisor since $\mathcal O_X(B)$ is isomorphic to $L$. Moreover, both $(A_0+B,D_1,\ldots,D_d)$ and $(A_0'+B,D_1,\ldots,D_d)$ belong to $\mathcal {IP}^{(d)}_X$ since the former one and their difference do.
\end{rema}
\section{Multi-homogeneous polynomials}\label{Sec: multi-homogeneous polynomial}
Let $k$ be a field and $(E_i)_{i=0}^d$ be a family of finite-dimensional vector spaces over $k$. Let $(\delta_0,\ldots,\delta_{d})$ be a multi-index in $\mathbb N^{d+1}$.
\begin{defi}We call \emph{multi-homogeneous polynomial of multi-degree $(\delta_0,\ldots,\delta_d)$} on $E_0\times\cdots\times E_d$ any element of
\[S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee),\]
where $S^{\delta_i}(E_i^\vee)$ denotes the $\delta_i$-th symmetric power of the vector space $E_i^\vee$.
\end{defi}
Recall that one dual vector space of $S^{\delta_i}(E_i^\vee)$ is given by \[\Gamma^{\delta_i}(E_i):=(E_i^{\otimes \delta_i})^{\mathfrak S_{\delta_i}},\]
where $\mathfrak S_{\delta_i}$ is the symmetric group on $\{1,\ldots,\delta_i\}$, which acts on $E_i^{\otimes \delta_i}$ by permuting tensor factors (see \cite[Chapitre IV, \S5, no. 11, proposition 20]{MR0274237}). Therefore, one dual vector space of $S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee)$ is given by
\[\Gamma^{\delta_0}(E_0)\otimes_k\cdots\otimes_k\Gamma^{\delta_d}(E_d).\]
If $R\in S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee)$ is a multi-homogeneous polynomial of multi-degree $(\delta_0,\ldots,\delta_d)$, for any $(s_0,\ldots,s_d)\in E_0\times\cdots\times E_d$, we denote by $R(s_0,\ldots,s_d)$ the value
\[R(s_0^{\otimes \delta_0}\otimes\cdots\otimes s_d^{
\otimes \delta_d})\]
in $k$. Thus $R$ determines a function on $E_0\times\cdots\times E_d$ valued in $K$ (which we still denote by $R$ by abuse of notation). Note that, as an element of $S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee)$, $R$ is uniquely determined by the corresponding function on $E_0\times\cdots\times E_d$ since each vector space $\Gamma^{\delta_i}(E_i)$ is spanned over $k$ by elements of the form $s_i^{\otimes \delta_i}$, $s_i\in E_i$ (see \cite[Chapitre IV, \S5, no. 5, proposition 5]{MR0274237}). This observation allows us to consider, for any $i\in\{0,\ldots,d\}$ and $s_i\in E_i$, the specification
\[R(\cdots,\underset{\mathclap{\shortstack{\scriptsize $\uparrow$\\[-.2ex]$i$-th coordinate}}
}{s_i},\cdots) \] of $R$ at $s_i$ as an element of
\[S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_k S^{\delta_{i-1}}(E_{i-1}^\vee)\otimes_k S^{\delta_{i+1}}(E_{i+1}^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee)\]
or as a multi-homogeneous polynomial function on
\[E_0\times\cdots\times E_{i-1}\times E_{i+1}\times\cdots\times E_d.\]
\begin{rema}
Note that an element of $S^{\delta_0}(E_0^{\vee}) \otimes_k \cdots \otimes_k S^{\delta_d}(E_d^{\vee})$ yields
a multi-homogeneous polynomial function on $\mathbb A(E_0^{\vee}) \times_k \cdots \times_k \mathbb A(E_d^{\vee})$ and
the set of $k$-rational points of $\mathbb A(E_0^{\vee}) \times_k \cdots \times_k \mathbb A(E_d^{\vee})$ is naturally isomorphic to
$E_0 \times \cdots \times E_d$, where $\mathbb A(E_i^{\vee}) = \operatorname{Spec}(\bigoplus_{\delta=0}^{\infty} S^{\delta}(E_i^{\vee}))$
for each $i$.
\end{rema}
\section{Incidence subscheme}
Let $k$ be a field and $E$ be a finite-dimensional vector space over $k$.
We denote $\operatorname{Proj}(\bigoplus_{\delta=0}^{\infty} S^{\delta}(E))$ by $\mathbb P(E)$.
Recall that the projective space $\mathbb P(E)$ represents the contravariant functor from the category of $k$-schemes to that of sets, which sends a $k$-scheme $\varphi:S\rightarrow\operatorname{Spec} k$ to the set of all invertible quotient $\mathcal O_S$-modules of $\varphi^*(E)$. In particular, if we denote by $\pi_E:\mathbb P(E)\rightarrow\operatorname{Spec} k$ the structural scheme morphism, then the universal object of the representation of this functor by $\mathbb P(E)$ is a quotient $\mathcal O_{\mathbb P(E)}$-module of $\pi_E^*(E)$, which we denote by $\mathcal O_E(1)$ and which we call \emph{universal invertible sheaf} on $\mathbb P(E)$. For any positive integer $n$, we let $\mathcal O_E(n):=\mathcal O_E(1)^{\otimes n}$ and $\mathcal O_E(-n):=(\mathcal O_E(1)^\vee)^{\otimes n}$. Note that the quotient homomorphism $\pi_E^*(E)\rightarrow\mathcal O_E(1)$ induces by passing to dual modules an injective homomorphism
\[\mathcal O_E(-1)\longrightarrow\pi_E^*(E^\vee).\]
We now consider the fibre product of projective spaces $\mathbb P(E)\times_k\mathbb P(E^\vee)$. Let \[p_1:\mathbb P(E)\times_k\mathbb P(E^\vee)\longrightarrow \mathbb P(E)\quad\text{ and }\quad p_2:\mathbb P(E)\times_k\mathbb P(E^\vee)\longrightarrow \mathbb P(E^\vee)\]
be morphisms of projection. Note that the following diagram of scheme morphisms is cartesian
\[\xymatrix{\mathbb P(E)\times_k\mathbb P(E^\vee)\ar[r]^-{p_2}\ar[d]_-{p_1}&\mathbb P(E^\vee)\ar[d]^-{\pi_{E^\vee}}\\
\mathbb P(E)\ar[r]_-{\pi_E}&\operatorname{Spec} k}\]
The composition
of the homomorphisms
\begin{equation}\label{Equ: incident subscheme}p_1^*(\mathcal O_E(-1))\longrightarrow p_1^*(\pi_E^*(E^\vee))\cong p_2^*(\pi_{E^\vee}^*(E^\vee))\longrightarrow p_2^*(\mathcal O_{E^\vee}(1))\end{equation}
determines a global section of the invertible sheaf \[\mathcal O_E(1)\boxtimes\mathcal O_{E^\vee}(1):= p_1^*(\mathcal O_E(1))\otimes p_2^*(\mathcal O_{E^\vee}(1)).\]
\begin{defi}
We call \emph{incidence subscheme} of $\mathbb P(E)\times_k\mathbb P(E^\vee)$ and we denote by $I_E$ the closed subscheme of $\mathbb P(E)\times_k\mathbb P(E^\vee)$ defined by the vanishing of the global section of $\mathcal O_E(1)\boxtimes\mathcal O_{E^\vee}(1)$ determined by \eqref{Equ: incident subscheme}. In particular, the cycle class of $I_E$ modulo the linear equivalence is
\[c_1(\mathcal O_E(1)\boxtimes\mathcal O_{E^\vee}(1))\cap[\mathbb P(E)\times_k\mathbb P(E^\vee)].\]
\end{defi}
The following proposition shows that the incidence subscheme can be realized as a projective bundle over $\mathbb P(E)$.
\begin{prop}\label{Pro: indcidence variety as bundle}
Let $Q_{E^\vee}$ be the quotient sheaf of $\pi_E^*(E^\vee)$ by the canonical image of $\mathcal O_E(-1)$. Then the incidence subscheme $I_E$ is isomorphic as a $\mathbb P(E)$-scheme to the projective bundle $\mathbb P(Q_{E^\vee}\otimes\mathcal O_E(1))$. Moreover, under this isomorphism, the restriction of $\mathcal O_E(1)\boxtimes\mathcal O_{E^\vee}(1)$ to $I_E$ is isomorphic to the universal invertible sheaf of the projective bundle $\mathbb P(Q_{E^\vee}\otimes\mathcal O_E(1))$.
\end{prop}
\begin{proof}
It suffices to identify $p_1:\mathbb P(E)\times_k\mathbb P(E^\vee)\rightarrow\mathbb P(E)$ with the projective bundle \[\mathbb P(\pi_E^*(E^\vee)\otimes\mathcal O_E(1))\longrightarrow\mathbb P(E).\]
Note that the universal invertible sheaf of this projective bundle is isomorphic to $\mathcal O_E(1)\boxtimes\mathcal O_{E^\vee}(1)$. Under this identification, the vanishing locus of \eqref{Equ: incident subscheme} coincides with the projective bundle $\mathbb P(Q_{E^\vee}\otimes\mathcal O_E(1))$.
\end{proof}
\begin{rema}\label{Rem: fibre incidence variety}
As a scheme over $\mathbb P(E)$, the incident subscheme $I_E$ also identifies with the projective bundle $\mathbb P(Q_{E^\vee})$. However, the universal invertible sheaf of this projective bundle is the restriction of $p_2^*(\mathcal O_{E^\vee}(1))$. Moreover, we can also consider the morphism of projection from the incidence subscheme to $\mathbb P(E^\vee)$. By the duality between $E$ and $E^\vee$, the incidence subscheme $I_E$ also identifies with the projective bundle of $Q_{E}:=\pi_{E^\vee}^*(E)/\mathcal O_{E^\vee}(-1)$ over $\mathbb P(E^\vee)$. In particular, if $x$ is a point of $\mathbb P(E^\vee)$, then the fibre of the incidence subscheme $I_{E}$ over $x$ identifies with \[\mathbb P((E\otimes_k\kappa(x))/x^*\mathcal O_E(-1)),\]
which is a hyperplane in $\mathbb P(E\otimes_k\kappa(x))$ defined by the vanishing locus of any non-zero element of the one-dimensional $\kappa(x)$-vector subspace of $E\otimes_k\kappa(x)$ defining the point $x$.
\end{rema}
\section{Resultants}
Let $k$ be a field and $X$ be an integral projective $k$-scheme, and $d$ be the Krull dimension of $X$. For any $i\in\{0,\ldots,d\}$, we fix a finite-dimensional vector space $E_i$ over $k$ and a closed embedding $f_i:X\rightarrow\mathbb P(E_i)$, and we denote by $L_i$ the pull-back of $\mathcal O_{E_i}(1)$ by $f_i$. For each $i\in\{0,\ldots,d\}$, we let $r_i$ be the Krull dimension of $\mathbb P(E_i)$, which identifies with $\operatorname{dim}_k(E_i)-1$. For each $i\in\{0,\ldots,d\}$, we let $\delta_i$ be the intersection number
\[\deg\big(c_1(L_0)\cdots c_1(L_{i-1})c_1(L_{i+1})\cdots c_1(L_d)\cap[X]\big)\]
Let $\mathbb P=\mathbb P(E_0)\times_k\cdots\times_k\mathbb P(E_d)$ be the product of $k$-schemes $(\mathbb P(E_i))_{i=0}^d$. The family $(f_i)_{i=0}^d$ induces a closed embedding $f:X\rightarrow \mathbb P$. Let \[\check{\mathbb P}:=\mathbb P(E_0^\vee)\times_k\cdots\times_k\mathbb P(E_d^\vee)\]
be the product of dual projective spaces. We identify $\mathbb P\times_k\check{\mathbb P}$ with
\[(\mathbb P(E_0)\times_k\mathbb P(E_0^\vee))\times_k\cdots\times_k(\mathbb P(E_d)\times_k\mathbb P(E_d^\vee))\]
and we denote by \[I_{\mathbb P}:=I_{E_0}\times_k\cdots\times_kI_{E_d}\] the fibre product of incidence subschemes,
so that the class of $I_{\mathbb P}$ modulo the linear equivalence coincides with the intersection product
\[
c_1(r_0^*(\mathcal O_{E_0}(1)\boxtimes\mathcal O_{E_0^\vee}(1))) \cdots c_1(r_n^*(\mathcal O_{E_n}(1)\boxtimes\mathcal O_{E_n^\vee}(1)))\cap[\mathbb P\times_k\check{\mathbb P}],
\] where $r_i : \mathbb P\times_k\check{\mathbb P} \to \mathbb P(E_i)\times_k\mathbb P(E_i^\vee)$ is the $i$-th projection.
By Proposition \ref{Pro: indcidence variety as bundle} (see also Remark \ref{Rem: fibre incidence variety}), $I_{\mathbb P}$ is isomorphic to a fiber product of projective bundles
\[\mathbb P(Q_{E_0})\times_k\cdots\times_k\mathbb P(Q_{E_d}).\]
\begin{defi}
We denote by $I_X$ the fibre product
$X\times_{\mathbb P}I_{\mathbb P}$, called the \emph{incidence subscheme} of $X\times_{k}\check{\mathbb P}$. As an $X$-scheme, it identifies with
\[\mathbb P(Q_{E_0}|_X)\times_{X} \cdots\times_{X}\mathbb P(Q_{E_d}|_X).\]
and hence is an integral closed subscheme of dimension
\[d+(r_0-1)+\cdots+(r_d-1)=r_0+\cdots+r_d-1\]
of $\mathbb P\times_k\check{\mathbb P}$.
In particular, for any extension $K$ of $k$ and any element
\[(x,\alpha_0,\ldots,\alpha_d)\in X(K)\times\mathbb P(E_0^\vee)(K)\times\cdots\times\mathbb P(E_d^\vee)(K),\]
if we denote by $H_i$ the hyperplane in $\mathbb P(E_{i,K})$ defined by the vanishing of $\alpha_i$, then $(x,\alpha_0,\ldots,\alpha_d)$ belongs to $I_X(K)$ if and only if $f_{i,K}(x)\in H_i$ for any $i\in\{1,\ldots,d\}$.
In addition,
the cycle class of $I_X$ modulo the linear equivalence is
the intersection product
\begin{equation} \label{eqn:class:incidence:subscheme}
c_1\left( p^*(L_0) \otimes q^*q_0^*(\mathcal O_{E_0^{\vee}}(1))\right) \cdots c_1\left( p^*(L_d) \otimes q^*q_d^*(\mathcal O_{E_d^{\vee}}(1))\right)\cap[X\times_k\check{\mathbb P}], \end{equation}
where $p : X \times_k \check{\mathbb P} \to X$, $q : X \times_k \check{\mathbb P} \to \check{\mathbb P}$ and $q_i : \check{\mathbb P} \to \mathbb P(E_i^{\vee})$ are the projections.
\end{defi}
\if00
\begin{prop}
The direct image by the projection $q : X \times_k\check{\mathbb P}\rightarrow\check{\mathbb P}$ of $I_X$ is a multi-homogeneous hypersurface of multi-degree $(\delta_0,\ldots,\delta_d)$.
\end{prop}
\begin{proof}
It is sufficient to see that $q_*(I_X)$ belongs to the cycle class
\[
c_1(\mathcal O_{E_0^{\vee}}(\delta_0) \boxtimes \cdots \boxtimes \mathcal O_{E_d^{\vee}}(\delta_d))\cap[\check{\mathbb P}].
\]
Note that, for any $(i_1,\ldots,i_n)\in\{0,\ldots,d\}^n$ such that $i_1,\ldots, i_n$ are distinct,
\[
q_*(c_1(p^*(L_{i_1})) \cdots c_1(p^*(L_{i_n}))\cap[X\times_k\check{\mathbb P}])\]
is equal to
\[c_1(\mathcal O_{E_{i_1}^\vee}(1)\boxtimes\cdots\boxtimes\mathcal O_{E_{i_n}^\vee}(1))\cap [\check{\mathbb P}]
\]
if $n=d$, and is equal to the zero cycle class otherwise.
Therefore, the assertion follows from \eqref{eqn:class:incidence:subscheme}.
\end{proof}
\else
\begin{prop}
The direct image by the morphism of projection $\mathbb P\times_k\check{\mathbb P}\rightarrow\check{\mathbb P}$ of the effective algebraic cycle in $\mathbb P\times_k\check{\mathbb P}$ associated with $I_X$ is a multi-homogeneous hypersurface of multi-degree $(\delta_0,\ldots,\delta_d)$.
\end{prop}
\begin{proof}
We reason by induction on $d$. First we assume that $d=0$. Without loss of generality, we may assume that $X=\operatorname{Spec} k'$, where $k'$ is a finite extension of $k$. Let $x$ be the image of $f_0:X\rightarrow\mathbb P(E_0)$, which corresponds to a one-dimensional $k'$-vector subspace of $E_0\otimes_kk'$
\[x^*\mathcal O_{E_0}(-1)\longrightarrow E_0^\vee\otimes_kk'. \]
The scheme $I_X=X\times_{\mathbb P}I_{\mathbb P}$ identifies with $\mathbb P(Q_{E_0})$, where \[Q_{E_0}=(E_0^\vee\otimes_kk')/x^{*}\mathcal O_{E_0}(-1).\]
This is a hyperplane of \[\mathbb P(E_0^\vee\otimes_kk')\cong\mathbb P(E_0^\vee)\times_{\operatorname{Spec} k}\operatorname{Spec} k'.\]
Note that the field of definition of the quotient vector space $Q_{E_0}$ identifies with the residue field $\kappa(x)$ of the point $x$. Hence the projection of $\mathbb P(Q_{E_0})$ in $\mathbb P(E_0^\vee)$ is a hypersurface of degree $[\kappa(x):k]$.
We now proceed with the case $d\geqslant 1$ in supposing that the proposition is true for equidimensional projective $K$-scheme of dimension $d-1$. Without loss of generality, we may assume that $X$ is an integral projective $k$-scheme. Let $\eta$ be the generic point of $\mathbb P(E_d^\vee)$ and $k_d$ be the field of rational functions on $\mathbb P(E_d^\vee)$. Note that the point $\eta$ corresponds to a one-dimension vector subspace of \[E_d\otimes_kk_d=H^0(\mathbb P(E_d\otimes_kk_d),\mathcal O_{E_d\otimes_kk_d}(1)).\] We pick a non-zero element $s$ in this one-dimensional vector subspace and let $Z$ be the effective cycle of $X\times_{\operatorname{Spec} k}\operatorname{Spec} k_d$ defined by the vanishing of $f_{d,k_d}^*(s)$. Note that
$I_{X}\times_{\mathbb P(E_d^\vee),\eta}\operatorname{Spec} k_d$
is isomorphic to $I_{Z_s}$ (see Remark \ref{Rem: fibre incidence variety}).
We apply the induction hypothesis to $Z_s$ and obtain that the projection of the effective cycle associated with $I_{Z_s}$ is a multi-homogeneous hypersurface of multi-degree $(\delta_0,\ldots,\delta_{d-1})$. Therefore the projection of $I_X$ in $\check{\mathbb P}$ is also a hypersurface, which is multi-homogeneous of multi-degree $(\delta_0,\ldots,\delta_{d-1})$ in the first $d-1$ coordinates. The same argument applied to fibres of $I_X$ over the generic points of $\mathbb P(E_0^\vee),\ldots,\mathbb P(E_{d-1}^\vee)$ shows that the hypersurface is actually multi-homogeneous in all coordinates and is of multi-degree $(\delta_0,\ldots,\delta_d)$.
\end{proof}
\fi
\begin{defi}
Let $X$ be an integral projective $k$-scheme of dimension $d$. We call \emph{resultant of $X$ with respect to $(f_i)_{i=0}^d$} any multi-homogeneous polynomial of multi-degree $(\delta_0,\ldots,\delta_d)$ on $E_0\times\cdots\times E_d$, whose vanishing cycle in \[\mathbb P(E_0^\vee)\times_k\cdots\times_k\mathbb P(E_d^\vee)\] identifies with the projection of the cycle associated with the incidence subscheme $I_X$. Note that the resultant of $X$ with respect to $(f_i)_{i=0}^d$ is unique up to a factor of scalar in $k\setminus\{0\}$ as an element of $S^{\delta_0}(E_0^{\vee}) \otimes_k \cdots \otimes_k S^{\delta_d}(E_d^{\vee})$.
In general, if $X$ is a projective $k$-scheme of dimension $d$ and if
\[\sum_{i=1}^nm_i X_i\]
is the $d$-dimensional part of the fundamental cycle of $X$, where $X_1,\ldots, X_n$ are $d$-dimensional irreducible components of $X$, and $m_i$ is the local multiplicity of $X$ at the generic point of $X_i$, we define the \emph{resultant} of $X$ with respect to $(f_i)_{i=0}^d$ as
any multi-homogeneous polynomial of the form
\[(R_{f_0|_{X_1},\ldots,f_d|_{X_1}}^{X_1})^{m_1}\cdots (R_{f_0|_{X_n},\ldots, f_d|_{X_n}}^{X_n})^{m_n},\]
where each $R_{f_0|_{X_i},\ldots,f_d|_{X_i}}^{X_i}$ is a resultant of $X_i$ with respect to $(f_i|_{X_i})_{i=0}^d$.
\end{defi}
\begin{exem}\label{Exe: explicit construction of resultant}
We consider the particular case where $d=0$. Let $f_0:X\rightarrow\mathbb P(E_0)$ be a close embedding. We first assume that $X$ is integral. In this case $f_0$ sends $X$ to a closed point $x$ of $\mathbb P(E_0)$. Let $\kappa(x)$ be the residue field of $x$ and $\delta_0=[\kappa(x):k]$ be the degree of $x$. Let $s_0$ be an element of $E_0$. We assume that, if we view $s_0$ as a global section of $\mathcal O_{E_0}(1)$, one has $s_0(x)\neq 0$. We construct an element $R^{X,s_0}_{f_0}\in S^{\delta_0}(E_0^\vee)$ as follows. Let \[\varphi_0:E_0\otimes_K\kappa(x)\longrightarrow\mathcal O_{E_0}(1)(x)\]
be the surjective $\kappa(x)$-linear map corresponding to the closed point $x$, and
\[\varphi_0^\vee:\mathcal O_{E_0}(-1)(x)\longrightarrow E_0^\vee\otimes_K\kappa(x)\]
be the dual $\kappa(x)$-linear map of $\varphi_0$, which is an injective linear map. Let $s_0(x)^\vee$ be the unique $\kappa(x)$-linear form on $\mathcal O_{E_0}(1)(x)$ taking the value $1$ at $s_0(x)$. We let \[R_{f_0}^{X,s_0}:=N_{\kappa(x)/K}(\varphi_0^\vee(s_0(x)^\vee))\in S^{\delta_0}(E_0^\vee),\]
which is defined as the determinant of the following homothety endomorphism of the free module $\operatorname{Sym}(E_0^{\vee})\otimes_K\kappa(x)$ of rank $\delta_0$ over the symmetric algebra $\operatorname{Sym}(E_0^\vee)$
\[\xymatrix@C+2pc{\operatorname{Sym}(E_0^{\vee})\otimes_K\kappa(x)\ar[r]^-{\varphi_0^\vee(s_0(x)^\vee)}&\operatorname{Sym}(E_0^{\vee})\otimes_K\kappa(x).}\]
Note that \[\varphi_0^\vee(s_0(x)^\vee)(s_0\otimes 1)=s_0(x)^\vee(s_0(x))=1.\]
Therefore the following equality holds
\[R_{f_0}^{X,s_0}(s_0)=1.\]
Assume that $X$ is not irreducible.
We let $X_1,\ldots,X_n$ be irreducible components of $X$ (namely points of $X$). For each $i\in\{1,\ldots,n\}$, let $x_i=f_0(X_i)$ and $a_i$ be the local multiplicity of $X$ at $X_i$. Then\[a_1x_1+\cdots+a_nx_n\] is the decomposition of $f(X)$ as a zero-dimensional cycle in $\mathbb P(E_0)$, where $x_1,\ldots,x_n$ are closed points of $\mathbb P(E_0)$ and $a_1,\ldots,a_n$ are positive integers. If $s_0$ is a global section of $\mathcal O_{E_0}(1)$, which does not vanish on any of the points $x_1,\ldots,x_n$, we define
\[R_{f_0}^{X,s_0}:=\prod_{i=1}^n(R_{f_0|_{X_i}}^{X_i,s_0})^{a_i}.\]
Then $R_{f_0}^{X,s_0}$ is a resultant of $X$ with respect to the closed embedding $f_0$, which satisfies $R_{f_0}^{X,s_0}(s_0)=1$.
\end{exem}
\begin{exem}
Let $n$ and $m$ be positive integers, and let \[f : \mathbb P^1 \longrightarrow \mathbb P^n, \qquad (x,y) \longmapsto (x^iy^{n-i})_{i=0}^n\] and \[g : \mathbb P^1 \longrightarrow \mathbb P^m,\qquad (x,y) \longmapsto (x^jy^{m-j})_{j=0}^m\] be the Veronese embeddings of degree $n$ and $m$, respectively. Note that the resultant $R$ of $\mathbb P^1$ with respect to $f$ and $g$ is the usual resultant, that is,
{\small\[
R(a_0, \ldots, a_n, b_0, \ldots, b_m) =
\det \begin{pmatrix}
a_0 & a_{1} & \cdots & & a_n & & & \\
& a_0 & a_{1} & \cdots & & a_n & & \\
& & \ddots & & & & \ddots & \\
& & & a_0 &a_{1}&\cdots & & a_n \\
b_0 & b_{1} & \cdots & & b_m & & & \\
& b_0 & b_{1} & \cdots & & b_m & & \\
& & \ddots & & & & \ddots & \\
& & & b_0 &b_{1}&\cdots & & b_m \\
\end{pmatrix}
\begin{array}{l}
\\ [-10ex]
\rdelim\}{4.7}{5ex}[$m$ rows] \\ \\ \\[5ex] \rdelim\}{4.7}{5ex}[$n$ rows] \\ \\
\end{array}
\]}
\end{exem}
\begin{rema}\label{Rem: resultant induction formula}
Let $R_{f_0,\ldots,f_d}^X$ be a resultant of $X$ with respect to $(f_i)_{i=0}^d$. If $K/k$ is an extension and if $s$ is an element of $E_d\otimes_kK$, defining a global section of $\mathcal O_{\mathbb P(E_d\otimes_kK)}(1)$, which intersects properly with all irreducible components $X\times_{\operatorname{Spec} k}\operatorname{Spec} K$, then, viewed as a multi-homogeneous polynomial on
\[(E_0\otimes_kK)\times\cdots\times(E_d\otimes_kK)\]
by extension of scalars, the resultant $R^X_{f_0,\ldots,f_d}$ specified on the last coordinate at $s$, is a resultant of $\operatorname{div}(s)\cap X_{K}$ with respect to $(f_{i,K})_{i=0}^{d-1}$. This observation motivates the following explicit construction of the resultant polynomial by induction.
\end{rema}
\begin{defi}\label{Def:resultant precise}
Let $(s_0,\ldots,s_d)\in E_0\times\cdots\times E_d$. We assume that, for any irreducible component $Z$ of $X$, the divisors $\operatorname{div}(s_0),\ldots,\operatorname{div}(s_d)$ intersect properly on $Z$. We denote by $R_{f_0,\ldots,f_d}^{X,s_0,\ldots,s_d}$ the unique resultant of $X$ with respect to $f_0,\ldots,f_d$ such that
\[R_{f_0,\ldots,f_d}^{X,s_0,\ldots,s_d}(s_0,\ldots,s_d)=1.\]
\end{defi}
\begin{rema}\label{Rem: resultant extension}
Let $k'/k$ be an extension of fields. For any $i\in\{0,\ldots,d\}$, the morphism $f_i:X\rightarrow\mathbb P(E_i)$ induces by base change a closed embedding $f_i'$ from $X':=X\times_{\operatorname{Spec} k}\operatorname{Spec} k'$ to $\mathbb P(E_i')$, where $E_i':=E_i\otimes_kk'$.
Note that the incidence subscheme of \[X'\times_{k'}\mathbb P(E_{0}'^\vee)\times_{k'}\cdots\times_{k'}\mathbb P(E_{d}'^\vee)\]
identifies with $I_X\times_{\operatorname{Spec} k}\operatorname{Spec} k'$. Therefore, if $R_{f_0,\ldots,f_d}^X$ is a resultant of $X$ with respect to $(f_i)_{i=0}^d$, then
\[R_{f_0,\ldots,f_d}^{X}\otimes 1\in (S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee))\otimes_kk'\cong S^{\delta_0}(E_{0}'^\vee)\otimes_{k'}\cdots\otimes_{k'}S^{\delta_d}(E_{d}'^\vee)\]
is a resultant of $X'$ with respect to $(f_{i}')_{i=0}^d$. Similarly, if $(s_0,\ldots,s_d)$ is an element of $ E_0\times\cdots\times E_d$ such that the divisors $\operatorname{div}(s_0),\ldots,\operatorname{div}(s_d)$ intersect properly on each irreducible component of $X$, then the following equality holds
\[R_{f_{0}',\ldots,f_{d}'}^{X',s_0',\ldots,s_d'}=R_{f_0,\ldots,f_d}^{X,s_0,\ldots,s_d}\otimes 1,\]
where for each $i\in\{0,\ldots,d\}$, $s_i'$ denotes the element $s_i\otimes 1$ in $E_i'=E_i\otimes_kk'$.
\end{rema}
\section{Projection to a projective space}
Let $k$ be an infinite field, $n$ be an integer such that $n\geqslant 1$, and $V$ be a vector space of dimension $n+1$ over $k$. Let $\mathbb P(V)$ be the projective space associated with the $k$-vector space $V$ and $\mathcal O_V(1)$ be the universal invertible sheaf on $\mathbb P(V)$. Recall that for any $k$-algebra $A$, any $k$-point of $\mathbb P(V)$ valued in $A$ corresponds to a quotient invertible $A$-module of $V\otimes_kA$. In particular, if $x$ is a scheme point of $\mathbb P(V)$ and $\kappa(x)$ is the residue field of $x$, then the scheme point $x$ corresponds to a non-zero $\kappa(x)$-linear map $p_x:V\otimes_k\kappa(x)\rightarrow\kappa(x)$, which is unique up to a unique homothety $\kappa(x)\rightarrow\kappa(x)$ by an element of $\kappa(x)^{\times}$.
\begin{defi}
We call \emph{rational linear subspace} of $\mathbb P(V)$ any Zariski closed subset of $\mathbb P(V)$ defined by the vanishing of all sections in a $k$-linear subspace of $V=H^0(\mathbb P(V),\mathcal O_{V}(1))$. If $Y$ is a rational linear subspace of $\mathbb P(V)$ which is of codimension $1$, we say that $Y$ is a \emph{rational hyperplane} in $\mathbb P(V)$.
\end{defi}
\begin{exem}
\begin{enumerate}[label=\rm(\arabic*)]
\item The scheme $\mathbb P(V)$ is a rational linear subspace of $\mathbb P(V)$. It is defined by the vanishing of the zero vector in $V$.
\item Let $x$ be a rational point of $\mathbb P(V)$, which corresponds to a non-zero $k$-linear map $\pi_x:V\rightarrow k$. Then $\{x\}$ is the vanishing locus of sections in $\operatorname{Ker}(\pi_x)$ and hence is a rational linear subspace of $\mathbb P(V)$.
\item The empty subset of $\mathbb P(V)$ is a rational linear subspace, which identifies with the vanishing locus of all sections in $V$. By convention, the dimension of the empty subset of $\mathbb P(V)$ is defined as $-1$.
\end{enumerate}
\end{exem}
\begin{rema}\label{Rem: linear proj}
If $Y$ is a rational linear subspace of $\mathbb P(V)$ which is the vanishing locus of a $k$-vector subspace $W$ of $V$, then the $k$-scheme $Y$ is isomorphic to $\mathbb P(V/W)$. We call \emph{linear projection with center $Y$} the $k$-morphisme $\pi_Y:\mathbb P(V)\setminus Y \rightarrow \mathbb P(W)$ which sends, for any commutative $k$-algebra $A$, any quotient invertible $A$-module $p_L:V\otimes_kA\rightarrow L$ in $(\mathbb P(V)\setminus Y)(A)$ to the composition \[W\otimes_kA\lhook\joinrel\longrightarrow V\otimes_kA\stackrel{p_L}{\longrightarrow}L,\]
which is an element of $\mathbb P(W)(A)$.
We assume that $Y=\{y\}$ is the set of one rational point of $\mathbb P(V)$, which corresponds to a non-zero $k$-linear map $p_y:V\rightarrow k$ whose kernel is $W$. Let $z$ be a scheme point of $\mathbb P(V)$, $\kappa(z)$ be the residue field of $z$, and $p_z:V\otimes_k\kappa(z)\rightarrow\kappa(z)$ be the non-zero $\kappa(z)$-linear map corresponding to the scheme point $z$. Note that $\kappa(z)$ is generated by elements of the form $p_z(f\otimes 1)/p_z(g\otimes 1)$, where $f$ and $g$ are elements of $V$ such that $p_z(g\otimes 1)\neq 0$. Assume that $y$ does not belong the Zariski closure of $\{z\}$. Then there exists at least an element $s\in V\setminus W$ such that $p_z(s\otimes 1)= 0$. Let $z'$ be the image of $z$ by the linear projection $\pi_Y$. The residue field of $z'$ identifies with the sub-extension of $\kappa(x)/k$ generated by elements of the form $p_z(f'\otimes 1)/p_z(g'\otimes 1)$, where $f'$ and $g'$ are elements of $W$ such that $p_z(g'\otimes 1)\neq 0$. As $W$ is of codimension $1$ in $V$ and $s$ is an element of $V\setminus W$ such that $p_z(s\otimes 1)=0$, we obtain that, for any $f\in V$, there exists $f'\in W$ such that $p_z(f\otimes 1)=p_z(f'\otimes 1)$. Therefore we obtain that $\kappa(z)=\kappa(z')$. In particular, if $X$ is a closed subset of $\mathbb P(V)$ which does not contain $y$, then $\pi_Y(X)$ has the same dimension as $X$.
\end{rema}
\begin{prop}\label{prop:projection:to:projective:space}
Let $d\in\{0,\ldots,n\}$.
Let $X$ be a Zariski closed set of $\mathbb P(V)$ such that $\dim(X) \leq d$.
Then we have the following:
\begin{enumerate}[label=\rm(\arabic*)]
\item
There is a rational linear subspace $M$ of $\mathbb P(V)$ such that $\dim(M) = n-1-d$ and $X \cap M = \emptyset$.
\item
Let $T$ be a rational linear subspace of $\mathbb P(V)$ such that $\dim(T) > n-d-1$, and that $X$ and $T$ meets properly.
Then
there is a rational linear subspace $M$ of $\mathbb P(V)$ such that $M \subseteq T$, $\dim(M) = n-1-d$ and $X \cap M = \emptyset$.
\item We assume that $X$ is irreducible and $\dim(X) = d$.
Let $M$ be a rational linear subspace of $\mathbb P(V)$ such that $\dim(M) = n-1-d$ and $M \cap X = \emptyset$, which is the vanishing locus of a vector space $W$ of $V$.
Let $\pi_M : \mathbb P(V) \setminus M \to \mathbb P(W)$ be the projection with the center $M$. Then $\pi := \rest{\pi_M}{X} : X \to \mathbb P^d_k$
is finite and surjective and $\pi^*(\mathcal O_{\mathbb P^d_k}(1)) = \rest{\mathcal O_{\mathbb P^n_k}(1)}{X}$.
\end{enumerate}
\end{prop}
\begin{proof}
(1) We prove the assertion by induction on $n-d$.
If $n=d$, then the assertion is obvious by choosing $M$ as the empty set, so that we assume that $n > d$.
Since $X \not= \mathbb P(V)$ and $k$ is an infinite field, there is a rational point $x \in \mathbb P(V)$ which does not belong to $X$. Let $W$ be the set of sections $s\in V=H^0(\mathbb P(V),\mathcal O_V(1))$ which vanish at $x$. This is a vector subspace of $V$.
Let $\pi : \mathbb P(V) \setminus \{ x \} \to \mathbb P(W)$
be the projection with center $\{x$\}. Since $x \not\in X$, by Remark \ref{Rem: linear proj} we obtain that $X$ and $X'$ have the same dimension.
In particular, $\dim(X') \leq d$. As $(n-1) -d < n-d$, by the hypothesis of induction, there is a linear
subspace $M'$ in $\mathbb P(W)$ such that $\dim (M') = n-2-d$ and $X' \cap M' = \emptyset$
Thus if we set $M = \pi^{-1}(M') \cup \{ x \}$, then one has the desired subspace.
(2) Assume that $T$ is defined by the vanishing of sections in a $k$-vector subspace $W$ of $V$. If we set $X' = X \cap T$ and $t = \dim T$, then $\dim X' \leq d - (n-t)$ and $T \simeq \mathbb P(V/W)$.
As
\[
t - (d-(n-t)) = n -d \geq 0,
\]
by (1), there is linear subspace $M$ in $T$ such that $\dim M = t - 1 - (d-(n-t))$ and $M \cap X' = \emptyset$.
Thus one has (2).
(3) Let $T$ be a linear subspace of $\mathbb P(V)$ such that $M \subseteq T$ and $\dim(T) = n -d$.
It is sufficient to show that $\dim (T \cap X) = 0$. Note that
$M$ is a rational hyperplane in $T$,
so that if $\dim (T \cap X) \geq 1$, then $M \cap X \not= \emptyset$. Therefore $\dim (T \cap X) = 0$.
\end{proof}
\chapter{Adelic curves and their constructions}
\section{Adelic structures}
In this section, we recall the notion of adelic curves. Let $K$ be a field.
An \emph{adelic structure} of $K$ consists of data
$((\Omega, \mathcal A, \nu), \phi)$ satisfying the following properties:
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item $(\Omega, \mathcal A,\nu)$ is a measure space, that is,
$\mathcal A$ is a $\sigma$-algebra of $\Omega$ and $\nu$ is a measure on $(\Omega, \mathcal A)$.
\item
The last $\phi$ is a map
from $\Omega$ to $M_K$, where $M_K$ is the set of all absolute values of $K$. For any $\omega\in\Omega$, we denote the absolute value $\phi(\omega)$ by $|\raisebox{.4ex}{.}|_{\omega}$.
\item For any $\omega\in\Omega$ and any $a \in K^{\times}$, the function $(\omega \in \Omega) \mapsto \ln |a|_{\omega}$ is $\nu$-integrable.
\end{enumerate}
The field $K$ equipped with an adelic structure is called an \emph{adelic curve}. Moreover, the adelic structure $((\Omega,\mathcal A,\nu),\phi)$ is said to be \emph{proper} if
\begin{equation}\label{eqn:product:formula}
\int_{\Omega} \ln |a|_{\omega}\, \nu(\mathrm{d}\omega) = 0
\end{equation}
holds for all $a \in K^{\times}$. If the adelic structure $((\Omega,\mathcal A,\nu),\phi)$ is proper, we also say that the adelic curve $(K,(\Omega,\mathcal A,\nu),\phi)$ is \emph{proper}. The equation \eqref{eqn:product:formula} is
called \emph{product formula}. For details, see \cite[Chapter 3]{CMArakelovAdelic}.
We denote the set of all $\omega \in \Omega$ with $|\raisebox{.4ex}{.}|_{\omega}$ Archimedean
(resp. non-Archimedean)
by $\Omega_{\infty}$ (resp. $\Omega_{\operatorname{fin}}$). The restriction of $\mathcal A$ to $\Omega_\infty$ (resp. $\Omega_{\operatorname{fin}}$) is denoted by $\mathcal A_\infty$ (resp. $\mathcal A_{\operatorname{fin}}$).
Note that $\Omega_{\infty}$ and $\Omega_{\operatorname{fin}}$ belong to $\mathcal{A}$ (see \cite[Proposition 3.1.1]{CMArakelovAdelic}).
For each $\omega \in \Omega_{\infty}$, there exist
an embedding $\iota_{\omega} : K \to {\mathbb{C}}$ and $\kappa_{\omega} \in (0, 1]$
such that $|a|_{\omega} = |\iota_{\omega}(a)|^{\kappa_{\omega}}$ for all $a \in K$,
where $|\raisebox{.4ex}{.}|$ is the usual absolute value of ${\mathbb{C}}$.
Note that the invariant $\kappa_{\omega}$ does not depend on the choice of the embedding
$\iota_{\omega} : K \to {\mathbb{C}}$.
From now on, we always assume that $\kappa_{\omega} = 1$ for all $\omega \in \Omega_{\infty}$.
For $(a_1, \ldots, a_n) \in K^n \setminus \{ (0, \ldots, 0) \}$,
the height $h_S(a_1, \ldots, a_n)$ of $(a_1, \ldots, a_n)$ with respect to the adelic curve
$S = (K,(\Omega, \mathcal A, \nu), \phi)$ is defined to be
\begin{equation}\label{eqn:height:function:wrt:S}
h_S(a_1, \ldots, a_n) := \int_{\Omega} \ln( \max \{ |a_1|_{\omega}, \ldots, |a_n|_{\omega} \})
\nu(\mathrm{d}\omega).
\end{equation}
Note that if $S$ is proper, then $h_S(a) = 0$ for all $a \in K^{\times}$.
\begin{rema}
Many classic constructions in algebraic geometry and arithmetic geometry, such as algebraic curves, rings of algebraic integers, polarized projective varieties and arithmetic varieties, can be interpreted as adelic curves. For example, on the filed $\mathbb Q$ of rational numbers there is an adelic structure consisting of all places of $\mathbb Q$ (namely the set $\Omega_{\mathbb Q}$ of all prime numbers and $\infty$) equipped with the discrete $\sigma$-algebra and the measure $\nu$ such that $\nu(\{\omega\})=1$ for any $\omega\in\Omega_{\mathbb Q}$, where $|\raisebox{.4ex}{.}|_{\infty}$ is the usual absolute value on $\mathbb Q$ and $|\raisebox{.4ex}{.}|_p$ is the $p$-adic absolute value for any prime number $p$. The product formula for this adelic curve is just the logarithmic version of the usual product formula for rational numbers
\[\forall\,a\in\mathbb Q^{\times},\quad |a|_\infty\cdot\prod_{p}|a|_p=1.\]
We call this adelic structure \emph{the standard adelic structure on $\mathbb Q$}. We refer the readers to \cite[\S 3.2]{CMArakelovAdelic} for more examples.
\end{rema}
\begin{defi}\label{Def: covering of adelic curves}
Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ and $S'=(K',(\Omega',\mathcal A',\nu'),\phi')$ be two adelic curves. We call \emph{morphism} from $S'$ to $S$ any triplet $\alpha=(\alpha^{\#},\alpha_{\#}, I_\alpha)$, where
\begin{enumerate}[label=\rm(\arabic*)]
\item $\alpha^{\#}:K\rightarrow K'$ is a field homomorphism,
\item $\alpha_{\#}:(\Omega',\mathcal A')\rightarrow(\Omega,\mathcal A)$ is a measurable map, such that, for any $\omega'\in\Omega'$, \[|\raisebox{.4ex}{.}|_{\omega'}\circ\alpha^{\#}=|\raisebox{.4ex}{.}|_{\alpha_{\#}(\omega')},\]
and that the direct image of $\nu'$ by $\alpha_{\#}$ coincides with $\nu$, namely for any $\mathcal A$-measurable function $f:\Omega\rightarrow\mathbb R$ which is either non-negative or integrable, one has
\[\int_{\Omega'}f(\alpha_{\#}(\omega'))\,\nu'(\mathrm{d}\omega')=\int_\Omega f(\omega)\,\nu(\mathrm{d}\omega),\]
\item \[I_\alpha:\mathscr L^1(\Omega',\mathcal A',\nu')\longrightarrow \mathscr L^1(\Omega,\mathcal A,\nu)\] is a linear map sending positive integrable functions on $(\Omega',\mathcal A',\nu')$ to positive integrable functions on $(\Omega,\mathcal A,\nu)$ such that, for any $f\in\mathscr L^1(\Omega',\mathcal A',\nu')$,
\[\int_{\Omega}I_\alpha(f)(\omega)\,\nu(\mathrm{d}\omega)=\int_{\Omega'}f(\omega')\,\nu'(\mathrm{d}\omega').\]
\end{enumerate}
If in addition for any $g\in\mathscr L^1(\Omega,\mathcal A,\nu)$, one has \[g\circ\alpha_{\#}\in\mathscr L^1(\Omega',\mathcal A',\nu')\quad\text{ and }\quad I_\alpha(g\circ\alpha_{\#})=g,\] we say that $\alpha$ is a \emph{covering} of adelic curves.
\end{defi}
\section{Algebraic coverings of adelic curves}\label{Sec: algebraic coverings}
Adelic curves are very flexible constructions. On a field there exist many adelic structures. It is also possible to construct new adelic structures from given ones. Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ be an adelic curve. In \cite[\S 3.2]{CMArakelovAdelic} it has been explained how to construct, for any algebraic extension $L/K$, a natural adelic curve \[S\otimes_KL=(L,(\Omega_{L},\mathcal A_{L},\nu_{L}),\phi_{L})\] on $L$ such that $\Omega_{L}=\Omega\times_{M_K,\phi}M_{L}$. The projection map $\pi_{L/K}:\Omega_{L}\rightarrow\Omega$ satisfies the relation
\[\nu=(\pi_{L/K})_*(\nu_L).\]
Moreover, for any $\omega\in\Omega$, the fibre $\pi_{L/K}^{-1}(\{\omega\})$ is equipped with a natural $\sigma$-algebra and a probability measure $\nu_{L,\omega}$, such that, for any positive $\mathcal A_{L}$-measurable function $f$ on $\Omega_{L}$, one has
\[\int_{\Omega_{L}}g(x)\,\nu_{L}(\mathrm{d}x)=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\pi_{L/K}^{-1}(\omega)}g(x)\,\nu_{L,\omega}(\mathrm{d}x).\]
In other words, the family of measures $(\nu_{L,\omega})_{\omega\in\Omega}$ form an disintegration of $\nu_{L}$ over $\nu$.
If the adelic curve $S$ is proper, then also is $S\otimes_KL$, see \cite[Proposition 3.4.10]{CMArakelovAdelic}. If we denote by $i_{K,L}:K\rightarrow L$ the inclusion map, and \[I_{L/K}:\mathscr L^1(\Omega_L,\mathcal A_L,\nu_L)\longrightarrow\mathscr L^1(\Omega,\mathcal A,\nu)\]
the linear map of fiber integrals, which sends $g\in L^1(\Omega_L,\mathcal A_L,\nu_L)$ to the function
\[(\omega\in\Omega)\longmapsto\int_{\pi_{L/K}^{-1}(\omega)}g(x)\,\nu_{L,\omega}(\mathrm{d}x),\]
then the triplet
$(i_{K,L},\pi_{L/K},I_{L/K})$ forms a covering of adelic curves in the sense of Definition \ref{Def: covering of adelic curves}.
\begin{lemm}\label{lemma:discrete:algebraic}
Let $K'$ be an algebraic extension of $K$ and
$S \otimes K' := (K', (\Omega', \mathcal{A}', \nu'), \phi')$.
Suppose that $K$ and $\Omega_{\operatorname{fin}}$ are countable sets. If $(\Omega_{\operatorname{fin}},\mathcal{A}_{\operatorname{fin}})$ is discrete, also is $(\Omega'_{\operatorname{fin}},\mathcal{A}'_{\operatorname{fin}})$.
\end{lemm}
\begin{proof}Since $K'/K$ is an algebraic extension, and $K$ and $\Omega_{\operatorname{fin}}$ are countable sets, we obtain that the sets $K'$ and $\Omega'_{\operatorname{fin}}$ are countable, so that it is sufficient to see that $\{ \omega' \}\in\mathcal{A}_{\operatorname{fin}}'$
for all $\omega' \in \Omega'_{\operatorname{fin}}$.
First we consider the case
where $K'$ is finite over $K$.
Let $\omega' \in \Omega'$ and $\omega = \pi(\omega')$,
where $\pi : \Omega' \to \Omega$ is the canonical map.
Then as $\{ \omega \}$ is $\mathcal{A}_{\operatorname{fin}}$-measurable and
$\pi$ is measurable, $\pi^{-1}(\{\omega\})\in\mathcal{A}'_{\operatorname{fin}}$.
If $|\raisebox{.4ex}{.}|_{\omega}$ is trivial, then $\pi^{-1}(\{\omega\}) = \{ \omega' \}$,
so that the assertion is obvious. Next we assume that $|\raisebox{.4ex}{.}|_{\omega}$
is non-trivial. Let us see that, for any $(x, x') \in \pi^{-1}(\{\omega\})^2$
with $x \not= x'$, $|\raisebox{.4ex}{.}|_x$ is not equivalent to $|\raisebox{.4ex}{.}|_{x'}$.
Otherwise, there is $\kappa \in {\mathbb{R}}_{>0}$ such that $|\raisebox{.4ex}{.}|_{x'} =
|\raisebox{.4ex}{.}|_x^{\kappa}$. As $|\raisebox{.4ex}{.}|_{\omega}$ is non-trivial,
there is $a \in K$ such that $|a|_{\omega} < 1$. Then
\[
|a|_{\omega} = |a|_{x'} = |a|_{x}^{\kappa} = |a|_{\omega}^{\kappa},
\]
and hence $\kappa = 1$, which is a contradiction. Therefore,
there is $a' \in K'$ such that $|a'|_{\omega'} < 1$ and
$|a'|_{x} > 1$ for all $x \in \pi^{-1}(\{\omega\}) \setminus \{ \omega' \}$
(cf. \cite[the proof of Theorem~3.4]{Neukirch}).
Note that \[\Delta := \{ \chi \in \Omega' \,:\, |a'|_\chi < 1 \}\] is $\mathcal{A}'_{\operatorname{fin}}$-measurable,
so that $\{ \omega' \} = \pi^{-1}(\{\omega\}) \cap \Delta$ is $\mathcal{A}'_{\operatorname{fin}}$-measurable.
In general, for $a \in K'$, let \[(K(a),(\Omega_{K(a)},\mathcal A_{K(a)},\nu_{K(a)}),\phi_{K(a)})=S\otimes K(a)\] and let
$\pi_{K'/K(a)} : \Omega' \to \Omega_{K(a)}$ be the canonical map.
By the previous case, $\{\pi_{K'/K(a)}(\omega')\}\in\mathcal A_{K(a)}$, so that
$\pi_{K'/K(a)}^{-1}(\{\pi_{K'/K(a)}(\omega')\})\in\mathcal A'$.
Therefore, as $K'$ is countable,
\[
\bigcap_{a \in K'} \pi_{K'/K(a)}^{-1}(\pi_{K'/K(a)}(\omega')).
\]
belongs to $\mathcal A'$. Thus it suffices to prove
\begin{equation}\label{eqn:prop:discrete:algebraic:01}
\{\omega'\} = \bigcap_{a \in K'} \pi_{K'/K(a)}^{-1}(\pi_{K'/K(a)}(\omega')).
\end{equation}
Indeed, if $x \in \bigcap_{a \in K'} \pi_{K'/K(a)}^{-1}(\{\pi_{K'/K(a)}(\omega')\})$,
then, for any $a \in K'$, $\pi_{K'/K(a)}(x) = \pi_{K'/K(a)}(\omega')$,
so that $|a|_x = |a|_{\omega'}$,
which means that $x = \omega'$.
\end{proof}
\section{Transcendental fibrations of adelic curves}
\label{Sec: Transcendental fibration}
The purpose of this section is to discuss the extension of an adelic structure to a transcendental extension of the field. We fix an adelic curve $S=(K,(\Omega,\mathcal A,\nu),\phi)$. For any $\omega\in\Omega$, let $K_\omega$ be the completion of $K$ with respect to the absolute value $\phi(\omega)$. Let $B$ be a $K$-algebra. Note that $B$ is not necessarily of finite type over $K$.
We assume that $B$ is a unique factorization domain and the set $B^{\times}$ of units in $B$ coincides with $K^{\times}$. We say that two irreducible elements of $B$ are \emph{equivalent} if they differ by a unit as a factor. This defines an equivalence relation on the set of all irreducible elements of $B$. We pick a representative in each of the equivalence classes to form a subset $\mathscr P_B$ of $B$ consisting of non-equivalent irreducible elements. Let $L$ be the field of fractions of $B$.
Recall that any non-zero element $g\in L$ can be written in a unique way as
\[c(g)\prod_{F\in\mathscr P_B}F^{\operatorname{ord}_F(g)},\]
where $c(g)$ is an element of $K^{\times}=B^{\times}$, and for each $F\in\mathscr P_B$, $\operatorname{ord}_F(g)$ is an integer. Note that $\operatorname{ord}_F(\raisebox{.4ex}{.})$ is a discrete valuation on the field $L$, and $\operatorname{ord}_F(a)=0$ for any $a\in K^{\times}=B^{\times}$.
\begin{defi} \label{Def: admissible family}
For any $\omega\in\Omega$, let $S_{L,\omega}=(L,(\Omega_{L,\omega},\mathcal A_{L,\omega},\nu_{L,\omega}),\phi_{L,\omega})$ be an adelic curve such that $\nu_{L,\omega}$ is a probability measure. We say that the family $(S_{L,\omega})_{\omega\in\Omega}$ is an \emph{admissible fibration with respect to $(B,\mathscr P_B)$} over the adelic curve $S$ if the following conditions are satisfied:
\begin{enumerate}[label=
\rm(\alph*)]
\item for any $\omega\in\Omega$ and any $x\in \Omega_{L,\omega}$, the absolute value $\phi_{L,\omega}(x)$ on $L$ is an extension of $\phi(\omega)$ on $K$,
\item\label{Cond: mesurabilit} for any element $g\in B\setminus\{0\}$, any finite family $(F_j)_{j=1}^n$ of elements of $\mathscr P_B$ containing $\{F\in\mathscr P_B\,|\,\operatorname{ord}_{F}(g)\neq 0\}$ and any $(C_j)_{j=1}^n\in\mathbb R_{\geqslant 0}^n$, the function
\[(\omega\in\Omega)\longmapsto\int_{\Omega_{L,\omega}}|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\,\nu_{L,\omega}(\mathrm{d}x)\]
is $\mathcal A$-measurable,
\item\label{Cond: integrability} for any $\omega\in\Omega$ and any element $F$ of $\mathscr P_B$, the function
\[(\omega\in\Omega)\longmapsto\int_{\Omega_{L,\omega}}\ln|F|_x\, \nu_{L,\omega}(\mathrm{d}x)\]
is integrable with respect to $\nu$.
\end{enumerate}
Let $(S_{L,\omega})_{\omega\in\Omega}$ be an admissible fibration over the adelic curve $S$. We define $\Omega_L$ as the disjoint union of $(\Omega_{L,\omega})_{\omega\in\Omega}$ and let $\phi_L$ be the map from $\Omega_L$ to the set of all absolute values on $L$, whose restriction on each $\Omega_{L,\omega}$ is equal to $\phi_{L,\omega}$. Let $\pi_{L/K}:\Omega_L\rightarrow\Omega$ be the projection map, sending the elements of $\Omega_{L,\omega}$ to $\omega$. We equip $\Omega_L$ with the $\sigma$-algebra $\mathcal A_L$ generated by the projection map $\pi_{L/K}$ and all functions of the form $(x\in\Omega_{L})\mapsto |g|_x$, where $g$ runs over the set $L$.
\end{defi}
\begin{prop}\label{Pro: disintegration}
Let $f$ be a non-negative $\mathcal A_L$-measurable function on $\Omega_L$. For any $\omega\in\Omega$, the function $f$ is $\mathcal A_{L,\omega}$-measurable on $\Omega_{L,\omega}$. Moreover, the function
\[(\omega\in\Omega)\longmapsto\int_{\Omega_{L,\omega}}f(x)\,\nu_{L,\omega}(\mathrm{d}x)\in[0,+\infty]\]
is $\mathcal A$-measurable.
\end{prop}
\begin{proof}
Let $\mathcal H$ be the set of all bounded non-negative $\mathcal A_{L}$-measurable functions $g$ on $\Omega_{L}$ which is $\mathcal A_{L,\omega}$-measurable on $\Omega_{L,\omega}$ for any $\omega\in\Omega$ and such that the function
\[(\omega\in\Omega)\longmapsto\int_{\Omega_{L,\omega}}f(x)\,\nu_{L,\omega}(\mathrm{d}x)\]
is $\mathcal A$-measurable. Note that, for any non-negative bounded $\mathcal A$-measurable function $\varphi$ on $\Omega$, one has $\varphi\circ\pi\in\mathcal H$ since it is constant on each fiber $\Omega_{L,\omega}$ and
\[\int_{\Omega_{L,\omega}}\varphi(\pi(x))\,\nu_{L,\omega}(\mathrm{d}x)=\int_{\Omega_{L,\omega}}\varphi(\omega)\,\nu_{L,\omega}(\mathrm{d}x)=\varphi(\omega).\] In particular, all non-negative constant functions belong to $\mathcal H$. Clearly, for any $(g_1,g_2)\in\mathcal H\times\mathcal H$ and any $(a_1,a_2)\in\mathbb R_{\geqslant 0}\times\mathbb R_{\geqslant 0}$, one has $a_1g_1+a_2g_2\in\mathcal H$. For any increasing sequence of functions $(g_n)_{n\in\mathbb N}$ in $\mathcal H$, the pointwise limit of $(g_n)_{n\in\mathbb N}$ belongs to $\mathcal H$. Moreover, for functions $g_1$ and $g_2$ in $\mathcal H$ such that $g_2\geqslant g_1$, then one $g_2-g_1\in\mathcal H$.
Let $\mathcal S$ be the set of functions of the form \[(x\in\Omega_{L})\longmapsto|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\varphi(\pi(x)),\] where $g$ is an element of $B\setminus\{0\}$, $(F_j)_{j=1}^n$ is a finite family of elements of $\mathscr P_B$ containing $\{F\in\mathscr P_B\,|\,\operatorname{ord}_{F}(g)\neq 0\}$, $(C_j)_{j=1}^n$ is a family of positive constant and $\varphi$ is a non-negative and bounded $\mathcal A$-measurable function on $\Omega$. Clearly the set $\mathcal S$ is stable by multiplication. Note that the function sending $\omega\in\Omega$ to
\[\begin{split}&\quad\;\int_{\Omega_{L,\omega}}|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\varphi(\pi(x))\,\nu_{L,\omega}(\mathrm{d}x)\\
&=\varphi(\omega)\int_{\Omega_{L,\omega}}|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\,\nu_{L,\omega}(\mathrm{d}x)
\end{split}\]
takes real values and is $\mathcal A$-measurable by the condition \ref{Cond: mesurabilit} above. Therefore, $\mathcal S$ is a subset of $\mathcal H$. Since the $\sigma$-algebra $\mathcal A_L$ is generated by $\mathcal S$, by monotone class theorem (see \cite[\S2.2]{Yan}, see also \cite[\S A.1]{CMArakelovAdelic}), $\mathcal H$ contains all bounded non-negative $\mathcal A_{L}$-measurable functions. Finally, since any non-negative $\mathcal A_{L}$-measurable function $f$ can be written as the limit of an increasing sequence of bounded non-negative $\mathcal A_{L}$-measurable functions, the assertion of the proposition is true.
\end{proof}
\begin{defi}\label{Def: adelic structure coming from a fibration}
Let $(S_{L,\omega})_{\omega\in\Omega}$ be an admissible fibration over $S$ (see Definition \ref{Def: admissible family}), where $S_{L,\omega}=(L,(\Omega_{L,\omega},\mathcal A_{L,\omega},\nu_{L,\omega}),\phi_{L,\omega})$. By Proposition \ref{Pro: disintegration}, there is a measure $\nu_{L}$ on the measurable space $(\Omega_{L},\mathcal A_{L})$ such that, for any non-negative $\mathcal A_{L}$-measurable function $f$ on $\Omega_{L}$, one has
\[\int_{\Omega_{L}}f(x)\,\nu_{L}(\mathrm{d}x)=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}f(x)\,\nu_{L,\omega}(\mathrm{d}x).\]
Therefore $S_L := (L,(\Omega_{L},\mathcal A_{L},\nu_{L}),\phi_{L})$ is an adelic curve, called the \emph{adelic curve associated with the admissible fibration $(S_{L,\omega})_{\omega\in\Omega}$.} Since $\nu_{L,\omega}$ are probability measures, if we denote by $i_{K,L}:K\rightarrow L$ the inclusion map, by $\pi_{L/K}:\Omega_L\rightarrow \Omega$ the map sending the elements of $\Omega_{L,\omega}$ to $\omega$, and by \[I_{L/K}:\mathscr L^1(\Omega_L,\mathcal A_L,\nu_L)\longrightarrow\mathscr L^1(\Omega,\mathcal A,\nu)\]
the linear map of fiber integrals, then the triplet $(i_{K,L},\pi_{L/K},I_{L/K})$ forms a covering of adelic curves in the sense of Definition \ref{Def: covering of adelic curves}.
\end{defi}
\section{Intrinsic compactification of admissible fibrations}
Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ be a proper adelic curve, $B$ be a $K$-algebra which is a unique factorization domain, and $\mathscr P_B$ be a representative family of irreducible elements as in the previous section. Let $L$ be the field of fractions of $B$ and \[\big(S_{L,\omega}=(L,(\Omega_{L,\omega},\mathcal A_{L,\omega},\nu_{L,\omega}),\phi_{L,\omega})\big)_{\omega\in\Omega}\] be an admissible fibration with respect to $(B,\mathscr P_B)$. In the previous section, we have constructed an adelic curve $S_L := (L,(\Omega_{L},\mathcal A_{L},\nu_{L}),\phi_{L})$ which fibers over $S$ and such that the measure $\nu_{L}$ disintegrates over $\nu$ by the family of measures $(\nu_{L,\omega})_{\omega\in\Omega}$ on the fibers. This construction looks similar to algebraic coverings of adelic curves. However, even in the case where the adelic structure $((\Omega,\mathcal A,\nu),\phi)$ is proper, the adelic structure $((\Omega_L,\mathcal A,\nu_L),\phi_L)$ is not necessarily proper. In this section, we show that, under a mild condition on the admissible fibration $(S_\omega)_{\omega\in\Omega}$ over $S$, we can naturally ``compactify'' the adelic structure $((\Omega_{L},\mathcal A_{L},\nu_{L}),\phi_{L})$. For any element $F\in\mathscr P_B$, we denote by $|\raisebox{.4ex}{.}|_F$ the absolute value on $K(T)$ such that
\[\forall\,g\in L^{\times},\quad |g|_F:=\operatorname{e}^{-\operatorname{ord}_F(g)}.\]
Thus we obtain a map $\phi_{L}'$ from $\mathscr P_B$ to $M_{L}$ sending $F$ to $|\raisebox{.4ex}{.}|_F$. Let $(\Omega_{L}^*,\mathcal A_{L}^*)$ be the disjoint union of the measurable spaces $(\Omega_{L},\mathcal A_{L})$ and $\mathscr P_B$ equipped with the discrete $\sigma$-algebra. Let $\phi_L^*:\Omega_L^*\rightarrow M_L$ be the map extending $\phi_L$ on $\Omega_L$ and $\phi_L'$ on $\mathscr P_B$.
\begin{prop}\label{Pro: compactification}
Let $(S_\omega)_{\omega\in\Omega}$ be an admissible fibration over $S$. We assume that, for any element $F\in\mathscr P_B$,
\begin{equation}\label{Equ: positivity of height}h_{S_L}(F):=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\geqslant 0.\end{equation}
Let $\nu_{L}^*$ be the measure on $(\Omega_{L}^*,\mathcal A_{L}^*)$ which coincides with $\nu_{L}$ on $(\Omega_{L},\mathcal A_{L})$ and such that
\begin{gather*}\forall\,F\in\mathscr P_B,\quad\nu_{L}^*(\{F\})=h_{S_L}(F).\end{gather*}
Then $S^*_L:= (L,(\Omega_{L}^*,\mathcal A_{L}^*,\nu_{L}^*),\phi_L^*)$ is a proper adelic curve.
\end{prop}
\begin{proof}
For any $g\in L^{\times}$, one has
\begin{equation}\label{Equ: computation ln g}\begin{split}&\quad\;\int_{\Omega_{L}}\ln|g|_x\,\nu_{L}(\mathrm{d}x)=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}\ln|g|_x\,\nu_{L,\omega}(\mathrm{d}x)\\
&=\sum_{F\in\mathscr P_B}\operatorname{ord}_F(g)\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)=\sum_{F\in\mathscr P_B}\operatorname{ord}_F(g)h_{S_L}(F).
\end{split}\end{equation}
Thus
\[\int_{\Omega_{L}^*}\ln|g|_x\,\nu_{L}^*(\mathrm{d}x)=\int_{\Omega_L}\ln|g|_x\,\nu_L(\mathrm{d}x)+\sum_{F\in\mathscr P_B}h_{S_L}(F)\ln|g|_F=0.
\]
\end{proof}
\begin{defi}\label{def:canonical:compactification}
Under the assumption \eqref{Equ: positivity of height}, the adelic curve $S^*_L$ is called the \emph{canonical
compactification} of $S_L$.
\end{defi}
\begin{rema}
Let $\mathcal A_B$ be the discrete $\sigma$-algebra on $\mathscr P_B$, $\nu_B$ be the measure on $(\mathscr P_B,\mathcal A_B)$ such that
\[\nu_B(\{F\})=h_{S_L}(F)\]
for any $F\in\mathscr P_B$, and $\phi_B:\mathscr P_B\rightarrow M_K$ be the map sending any element of $\mathscr P_B$ to the trivial absolute value on $K$. Then $S_B:=(K,(\mathscr P_B,\mathcal A_B,\nu_B),\phi_B)$ forms an adelic curve having $K$ as the underlying field. Let $S^*$ be the amalgamation of $S$ and $S_B$. Then, the inclusion map $K\rightarrow L$, the projection \[\pi_{L/K}\amalg \operatorname{Id}_{\mathscr P_B}:\Omega_L^*=\Omega_L\amalg\mathscr P_B\longrightarrow\Omega^*=\Omega\amalg\mathscr P_B\]
and the integral along fibers form a covering of adelic curves.
\end{rema}
\section{Non-intrinsic compactification of admissible fibrations}
We keep the notation of the previous section. In this section, we assume that the family of absolute values $(|\raisebox{.4ex}{.}|_{F})_{F\in\mathscr P_B}$ can be included in a proper adelic structure. We will show that a weaker positivity condition than \eqref{Equ: positivity of height} would be enough to ensure the existence of (non-intrinsic) compactifications of the adelic structure $((\Omega_{L},\mathcal A_{L},\nu_{L}),\phi_{L})$. In the rest of the subsection, we assume that there exists a \emph{proper} adelic structure $((\Omega_{L}',\mathcal A_{L}',\nu_L'),\phi_L')$ on $L$ which satisfies the following conditions:
\begin{enumerate}[label=\rm(\arabic*)]
\item $\Omega_L'$ contains $\mathscr P_B$ as a discrete measurable sub-space and $\nu_L'(\{F\})>0$ for any $F\in\mathscr P_B$,
\item for any $F\in\mathscr P_B$, one has $\phi_L'(F)=|\raisebox{.4ex}{.}|_F$.
\end{enumerate}
Note that the existence of such an adelic structure is is true when $K$ is of characteristic $0$ and $\operatorname{Spec} B$ is a smooth $K$-scheme of finite type. In this case there exists a projective $K$-scheme $X$ and an open immersion from $B$ into $X$. Then one can construct an adelic consisting of prime divisors of $X$, by choosing a polarization on $X$. We refer the readers to \cite[\S3.2.4]{CMArakelovAdelic} for more details.
\begin{prop}\label{Pro: compactification 2}
Let $(S_\omega)_{\omega\in\Omega}$ be an admissible fibration over $S$. For any element $F\in\mathscr P_B$, let
\begin{equation}\label{Equ: positivity of height :02}h_{S_L}(F):=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x).\end{equation}
Let $\delta$ be a positive constant. We assume that
\[\forall\,F\in\mathscr P_B,\quad h_{S_L}(F)+\delta\nu_L'(\{F\})\geqslant 0.\]
Let $(\Omega_L'',\mathcal A_L'')$ be the disjoint union of $(\Omega_L,\mathcal A_L)$ and $(\Omega_L',\mathcal A_L')$, $\phi_L'':\Omega_L''\rightarrow M_L$ be the map extending $\phi_L$ and $\phi_L'$, and $\nu_L^\delta$ be the measure on
$(\Omega_L'',\mathcal A_L'')$ which coincides with $\nu_L$ on $(\Omega_L,\mathcal A_L)$ and coincides with
\[\delta\nu_L'+\sum_{F\in\mathscr P_B}h_{S_L}(F)\operatorname{Dirac}_F\]
on $(\Omega_L',\nu_L')$, where $\operatorname{Dirac}_F$ denotes the Dirac measure at $F$. Then $((\Omega_{L}'',\mathcal A_{L}'',\nu_{L}^\delta),\phi_L'')$ is a proper adelic structure on $L$.
\end{prop}
\begin{proof}
For any $g\in L^{\times}$, one has
\[\int_{\Omega_L^*}\ln|g|_x\,\nu_L^\delta(\mathrm{d}x)=\int_{\Omega_L}\ln|g|_x\,\nu_L(\mathrm{d}x)+\delta\int_{\Omega_L'}\ln|g|_x\,\nu_L'(\mathrm{d}x)+\sum_{F\in\mathscr P_B}h_{S_L}(F)\ln|g|_F.
\]
By \eqref{Equ: computation ln g}, one has
\[\int_{\Omega_L}\ln|g|_x\,\nu_L(\mathrm{d}x)+\sum_{F\in\mathscr P_B}h_{S_L}(F)\ln|g|_F=0.\]
Moreover, since $((\Omega_L',\mathcal A_L',\nu_L'),\phi_L')$ is a proper adelic structure, one has
\[\int_{\Omega_L'}\ln|g|_x\,\nu_L'(\mathrm{d}x)=0.\]
Therefore we obtain
\[\int_{\Omega_L''}\ln|g|_x\,\nu_L^\delta(\mathrm{d}x)=0.\]
\end{proof}
\section{Purely transcendental fibration of adelic curves}
\label{subsec:adelic:structure:of:purely:transcendental:fibration}
In this section, we apply the results obtained in previous sections to the study of adelic structures on a purely transcendental extension of the underlying field of an adelic curve. Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ be an adelic curve and $I$ be a non-empty set. We consider the polynomial ring $K[\boldsymbol{T}_I]$ spanned by $I$, where $\boldsymbol{T}_I=(T_i)_{i\in I}$ denotes the variables. Let $\mathbb N^{\oplus I}$ be the set of vectors $\boldsymbol{d}=(d_i)_{i\in I}\in\mathbb N^I$ such that $d_i=0$ for all but a finite number of $i\in I$. For any vector $\boldsymbol{d}=(d_i)_{i\in I}\in\mathbb N^{\oplus I}$, we denote by $\boldsymbol{T}^{
\boldsymbol{d}}$ the monomial
\[\prod_{i\in I,\, d_i>0}T_i^{d_i}.\]
If $g$ is an element of $K[\boldsymbol{T}_I]$, for any $\boldsymbol{d}\in\mathbb N^{\oplus I}$ we denote by $a_{\boldsymbol{d}}(g)$ the coefficient of $\boldsymbol{T}^{\boldsymbol{d}}$ in the writing of $g$ as a $K$-linear combination of monomials.
For convenience, $K[\boldsymbol{T}_I]$ means $K$ in the case where $I = \emptyset$.
\begin{lemm}\label{lem:UFD:K:Lambda}
\begin{enumerate}[label=\rm(\arabic*)]
\item
Let $J$ be a subset of $I$. If $f$ and $g$ are two elements of $K[\boldsymbol{T}_I]$ such that $fg$ belongs to $K[\boldsymbol{T}_J]$, then both polynomials $f$ and $g$ belong to $K[\boldsymbol{T}_J]$.
\item The ring $K[\boldsymbol{T}_I]$ is a unique factorization domain and $K[\boldsymbol{T}_I]^{\times}=K^{\times}$.
\end{enumerate}
\end{lemm}
\begin{proof}
(1) For $i \in I$ and $\varphi\in K[\boldsymbol{T}_I]$, the degree of $\varphi$ with
respect to $x_i$ is denoted by $\deg_i(\varphi)$. Note that the function $\deg_i(\raisebox{.4ex}{.})$ satisfies the equality $\deg_i(fg) = \deg_i(f) + \deg_i(g)$, so that $\deg_i(f) = \deg_i(g) = 0$ once $i \in I \setminus J$,
which means that $g$ and $h$ belong to $K[\boldsymbol{T}_J]$.
\medskip
(2)
For any finite subset $J$ of $I$, it is well-known that $K[\boldsymbol{T}_J]$ is a unique factorization domain.
Moreover, for $f \in K[\boldsymbol{T}_I] \setminus \{ 0 \}$,
there is a finite subset $J$ of $I$ such that $f \in K[\boldsymbol{T}_J]$.
Thus the first assertion follows from (1). The second assertion is a direct consequence of (1) in the particular case where $J=\varnothing$.
\end{proof}
Let $L=K(\boldsymbol{T}_I)$ be the field of fractions of $K[\boldsymbol{T}_I]$. As in \S\ref{Sec: Transcendental fibration}, we pick in each equivalent class of irreducible polynomials in $K[\boldsymbol{T}_I]$, a representative to form a subset $\mathscr P_{K[\boldsymbol{T}_I]}$. For each element $F\in\mathscr P_{K[\boldsymbol{T}_I]}$, we let $\mathrm{ord}_F(\raisebox{.4ex}{.})$ be the discrete valuation on $L$ defined by $F$ and let $|\raisebox{.4ex}{.}|_F:=\mathrm{e}^{-\operatorname{ord}_F(\raisebox{.4ex}{.})}$ be the corresponding absolute value. Let $\deg(\raisebox{.4ex}{.})$ be the degree function on $K[\boldsymbol{T}_I]$. Note that for any $(f,g)\in K[\boldsymbol{T}_I]^2$ one has
\[\deg(f+g)\leqslant\max\{\deg(f),\deg(g)\},\quad
\deg(fg)=\deg(f)+\deg(g).\]
Therefore the function $-\deg(\raisebox{.4ex}{.})$ extends to a discrete valuation on $L$. Denote by $|\raisebox{.4ex}{.}|_\infty$ the corresponding absolute value, defined as
\[|\raisebox{.4ex}{.}|_{\infty}=\mathrm{e}^{\deg(\raisebox{.4ex}{.})}.\]
Note that the following product formula holds
\[\forall\, g\in L\setminus\{0\},\quad \ln|g|_{\infty}+\sum_{F\in\mathscr P_{K[\boldsymbol{T}_I]}}\deg(F)\ln|g|_F=0.\]
In other words, if we equip $\Omega_L':=\mathscr P_{K[\boldsymbol{T}_I]}\amalg\{\infty\}$ with the discrete $\sigma$-algebra $\mathcal A_L'$ and the measure $\nu_L'$ such that
\[\nu_L'(\{\infty\})=1\text{ and }\nu_L'(\{F\})=\deg(F)\] for any $F\in\mathscr P_{K[\boldsymbol{T}_I]}$, then $(L,(\Omega_L',\mathcal A_L',\nu_L'),\phi_L')$ forms a proper adelic curve, where \[\phi_L':\mathscr P_{K[\boldsymbol{T}_I]}\amalg\{\infty\}\rightarrow M_L\] sends $x$ to $|\raisebox{.4ex}{.}|_x$.\medskip
\begin{rema}
Let $\boldsymbol{X}_{I \cup \{ \infty \}} = \{ X_i \}_{i \in I} \cup \{ X_{\infty} \}$ be the variables indexed by $I \cup \{ \infty \}$. Let $\varphi : K[\boldsymbol{X}_{I \cup \{ \infty \}}] \to K[\boldsymbol{T}_I]$ be the homomorphism given by
$\varphi(f) = f((T_i)_{i \in I}, 1)$. If $f$ is an irreducible homogeneous polynomial in $K[\boldsymbol{X}_{I \cup \{ \infty \}}]$ and $f \not= X_{\infty}$, then $\varphi(f)$ is an irreducible polynomial in $K[\boldsymbol{T}_I]$. Moreover, for any irreducible polynomial $g$ in $K[\boldsymbol{T}_I]$,
there is an irreducible homogeneous polynomial $f$ in $K[\boldsymbol{X}_{I \cup \{ \infty \}}]$ such that $\varphi(f) = g$.
Note that the above $|\raisebox{.4ex}{.}|_{\infty}$ comes from the irreducible polynomial $X_{\infty}$, so that
the corresponding element is $1 = \varphi(X_{\infty})$.
\end{rema}
\begin{lemm}[Gauss's Lemma]
\label{lem:Gauss:lemma}
Let $|\raisebox{.4ex}{.}|$ be a non-Archimedean absolute value on $K$.
We fix $\pmb{e} = (e_i)_{i \in I} \in {\mathbb{R}}_{>0}^{I}$. For $\boldsymbol{d} = (d_i)_{i\in I} \in\mathbb N^{\oplus I}$, we set $\pmb{e}^{\boldsymbol{d}} := \prod_{i \in I} e_i^{d_i}$.
We denote by $|\raisebox{.4ex}{.}|_{\pmb{e},L}$ the function on $K[\boldsymbol{T}_I]$ sending $f\in K[\boldsymbol{T}_I]$ to \[\max_{\boldsymbol{d}\in\mathbb N^{\oplus I}}|a_{\boldsymbol{d}}(f)|\pmb{e}^{\boldsymbol{d}}.\]
Then, for any $(f,g)\in K[\boldsymbol{T}_I]^2$ one has
\[|fg|_{\pmb{e},L}=|f|_{\pmb{e},L}\cdot|g|_{\pmb{e},L}\quad\text{and}\quad|f+g|_{\pmb{e},L}\leqslant\max\{|f|_{\pmb{e},L},|g|_{\pmb{e},L}\}. \] In particular,
$|\raisebox{.4ex}{.}|_{\pmb{e},L}$ extends to an absolute value on $L=K(\boldsymbol{T}_I)$.
\end{lemm}
\begin{proof}
If we set $f = \sum_{\boldsymbol{d}'\in\mathbb N^{\oplus I}} a_{\boldsymbol{d}'} \boldsymbol{T}^{\boldsymbol{d}'}$ and
$g = \sum_{\boldsymbol{d}''\in\mathbb N^{\oplus I}} b_{\boldsymbol{d}''} \boldsymbol{T}^{\boldsymbol{d}''}$, then
\[
fg = \sum_{\boldsymbol{d}\in\mathbb N^{\oplus I}} \left( \sum_{\substack{\boldsymbol{d}', \boldsymbol{d}'' \in\mathbb N^{\oplus I}, \\ \boldsymbol{d}' + \boldsymbol{d}'' = \boldsymbol{d}}}
a_{\boldsymbol{d}'} b_{\boldsymbol{d}''} \right) \boldsymbol{T}^{\boldsymbol{d}}
\quad\text{and}\quad
f + g = \sum_{\boldsymbol{d}\in\mathbb N^{\oplus I}} (a_{\boldsymbol{d}} + b_{\boldsymbol{d}}) \boldsymbol{T}^{\boldsymbol{d}}
\]
Thus it is easy to see
\begin{equation}\label{eqn:lem:Gauss:lemma:01}
\begin{cases}
|fg|_{\pmb{e},L}\leqslant|f|_{\pmb{e},L}\cdot|g|_{\pmb{e},L}, \\
|f+g|_{\pmb{e},L}\leqslant\max\{|f|_{\pmb{e},L},|g|_{\pmb{e},L}\}.
\end{cases}
\end{equation}
Let $\Sigma_f = \{ \boldsymbol{d}'\in\mathbb N^{\oplus I} \mid |a_{\boldsymbol{d}'}|\pmb{e}^{\boldsymbol{d}'} =
|f|_{\pmb{e},L} \}$ and
$\Sigma_g = \{ \boldsymbol{d}''\in\mathbb N^{\oplus I} \mid |b_{\boldsymbol{d}''}|\pmb{e}^{\boldsymbol{d}''} =
|g|_{\pmb{e},L} \}$. Let $\leqslant_{\mathrm{lex}}$ be the lexicographic order on $\mathbb N^{\oplus I}$.
We choose $\pmb{\delta}(f) \in \Sigma_f$ and $\pmb{\delta}(g) \in \Sigma_g$ such that
$\boldsymbol{d}' \leqslant_{\mathrm{lex}} \pmb{\delta}(f)$ and
$\boldsymbol{d}'' \leqslant_{\mathrm{lex}} \pmb{\delta}(g)$
for all $\boldsymbol{d}' \in \Sigma_f$ and $\boldsymbol{d}'' \in \Sigma_g$.
\begin{enonce}{Claim}\label{claim:lem:Gauss:lemma:01}
One has $|a_{\boldsymbol{d}'}||b_{\boldsymbol{d}''}| \leqslant |a_{\pmb{\delta}(f)}||b_{\pmb{\delta}(g)}|$
for all $\boldsymbol{d}', \boldsymbol{d}'' \in \mathbb N^{\oplus I}$ with
$\boldsymbol{d}' + \boldsymbol{d}'' = \pmb{\delta}(f) + \pmb{\delta}(g)$.
Moreover, the equality holds if and only if $\boldsymbol{d}' = \pmb{\delta}(f)$
and $\boldsymbol{d}'' = \pmb{\delta}(g)$.
\end{enonce}
\begin{proof}
As $|a_{\boldsymbol{d}'}| \pmb{e}^{\boldsymbol{d}'} \leqslant |f|_{\pmb{e}, L}$ and
$|b_{\boldsymbol{d}''}| \pmb{e}^{\boldsymbol{d}''} \leqslant |g|_{\pmb{e}, L}$, one has
\[
|a_{\boldsymbol{d}'}||b_{\boldsymbol{d}''}| \leqslant \frac{|f|_{\pmb{e}, L} |g|_{\pmb{e}, L}}{\pmb{e}^{\boldsymbol{d}'+ \boldsymbol{d}''}} = \frac{|a_{\pmb{\delta}(f)}|\pmb{e}^{\pmb{\delta}(f)} |b_{\pmb{\delta}(g)}|\pmb{e}^{\pmb{\delta}(g)}}{\pmb{e}^{\boldsymbol{d}'+ \boldsymbol{d}''}} = |a_{\pmb{\delta}(f)}||b_{\pmb{\delta}(g)}|.
\]
We assume that the equality holds. Then $\boldsymbol{d}' \in \Sigma_f$ and $\boldsymbol{d}'' \in \Sigma_g$, so that
$\boldsymbol{d}' \leqslant_{\mathrm{lex}} \pmb{\delta}(f)$ and $\boldsymbol{d}'' \leqslant_{\mathrm{lex}} \pmb{\delta}(g)$.
Therefore, one has the assertion because $\boldsymbol{d}' + \boldsymbol{d}'' = \pmb{\delta}(f) + \pmb{\delta}(g)$.
\end{proof}
The above claim implies that
\[
\left| \sum_{\substack{\boldsymbol{d}', \boldsymbol{d}'' \in\mathbb N^{\oplus I}, \\ \boldsymbol{d}' + \boldsymbol{d}'' = \pmb{\delta}(f) + \pmb{\delta}(g)}}
a_{\boldsymbol{d}'} b_{\boldsymbol{d}''} \right| \pmb{e}^{\pmb{\delta}(f) + \pmb{\delta}(g)}= |a_{\pmb{\delta}(f)}|\pmb{e}^{\pmb{\delta}(f)} |b_{\pmb{\delta}(g)}|\pmb{e}^{\pmb{\delta}(g)} = |f|_{\pmb{e}, L} |g|_{\pmb{e}, L},
\]
which means that $|fg|_{\pmb{e}, L} \geqslant |f|_{\pmb{e}, L}|g|_{\pmb{e}, L}$, as required.
\end{proof}
For any $\omega\in\Omega\setminus \Omega_{\infty}$, let $|\raisebox{.4ex}{.}|_{\omega,L}$ be the absolute value on $L$ such that
\[\forall\,g=\sum_{\boldsymbol{d}\in\mathbb N^{\oplus I}} a_{\boldsymbol{d}}(g)\boldsymbol{T}_I^{\boldsymbol{d}}\in K[\boldsymbol{T}_I],\quad
|g|_{\omega,L}:=\sup_{\boldsymbol{d}\in\mathbb N^{\oplus I}}|a_{\boldsymbol{d}}(g)|_\omega.\]
By Lemma~\ref{lem:Gauss:lemma}, this absolute value is an extension of $|\raisebox{.4ex}{.}|_{\omega}$ on $K$. Let \[((\Omega_{L,\omega},\mathcal A_{L,\omega},\nu_{L,\omega}),\phi_{L,\omega})\] be the adelic structure on $L$ which consists of a single copy of the absolute value $|\raisebox{.4ex}{.}|_{\omega,L}$, equipped with the unique probability measure. We denote by $S_{L,\omega}$ the adelic curve $(L,(\Omega_{L,\omega},\mathcal A_{L,\omega},\nu_{L,\omega}),\phi_{L,\omega})$.
\begin{prop}\label{Pro: S is an admissible family}
If $\Omega_{\infty} = \emptyset$, then
family $(S_{L,\omega})_{\omega\in\Omega}$ is an admissible fibration over $S$.
\end{prop}
\begin{proof}
Let
$g$ be a non-zero element of $K[\boldsymbol{T}_I]$, $(F_j)_{j=1}^n$ be a finite family of elements of $\mathscr P_{K[\boldsymbol{T}_I]}$ containing $\{F\in \mathscr P_{K[\boldsymbol{T}_I]}\,|\,\operatorname{ord}_F(g)\neq 0\}$ and $(C_j)_{j=1}^n$ be a family of non-negative constants. One has
\[\int_{\Omega_{L,\omega}}|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\,\nu_{L,\omega}(\mathrm{d}x)=\max_{\boldsymbol{d}\in\mathbb N^{\oplus I}}|a_{\boldsymbol{d}}(g)|_\omega\cdot\prod_{j=1}^n\prod_{\boldsymbol{d}\in\mathbb N^{\oplus I}}1\hspace{-0.25em}\mathrm{l}_{|a_{\boldsymbol{d}}(F_j)|_\omega\leqslant C_j}.\]
Therefore the function
\[(\omega\in\Omega)\longmapsto\int_{\Omega_{L,\omega}}|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\,\nu_{L,\omega}(\mathrm{d}x).\]
is $\mathcal A$-measurable. Moreover, for any element $F$ of $\mathscr P_{K[\boldsymbol{T}_I]}$, one has
\begin{equation}\label{Equ: local integral for finite places}\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)=\max_{\boldsymbol{d}\in\mathbb N^{\oplus I},\,a_{\boldsymbol{d}}(F)\neq 0}\ln|a_{\boldsymbol{d}}(F)|_\omega.\end{equation}
Therefore the function
\[(\omega\in\Omega)\longmapsto\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\]
is $\nu$-integrable.
\end{proof}
\begin{rema}
In the case where $\Omega_{\infty} = \emptyset$ and
the adelic curve $S$ is proper, for any $\boldsymbol{d}$ such that $a_{\boldsymbol{d}}(F)\neq 0$, one has
\[\int_{\omega\in\Omega}\ln|a_{\boldsymbol{d}}(F)|_\omega\,\nu(\mathrm{d}\omega)=0,\]
and hence
\[h_{S_L}(F)=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\geqslant 0.\]
\end{rema}
\section{Arithmetic adelic structure}
\label{sec:Arithmetic adelic structure}
In this section, we provides a ``standard'' construction of an adelic structure for a countable field of characteristic zero. More precisely, for any countable field $E$ of characteristic zero, we will construct an adelic curve $S_E=(E,(\Omega_E,\mathcal A_E,\nu_E),\phi_E)$, which satisfies the following properties:
\begin{enumerate}[label=\rm(\arabic*)]
\item\label{Item: propre} $S_E$ is proper.
\item For any $\omega \in \Omega_E$, the absolute value $\phi_E(\omega)$ is not trivial.
\item The set $\Omega_{E,\operatorname{fin}}$ of $\omega\in\Omega$ such that $\phi_E(\omega)$ is non-Archimedean is infinite but countable.
\item\label{Item: Northcott} Let $E^{\mathrm{ac}}$ be an algebraic closure of $E$.
If $E_0$ is a subfield of $E^{\mathrm{ac}}$ such that $E_0$ is finitely generated over ${\mathbb{Q}}$,
then
\[\{ a \in E^{\mathrm{ac}} \mid \text{$h_{S_E \otimes_E E^{\mathrm{ac}}} (1,a) \leqslant C$ and $[E_0(a) : E_0] \leqslant \delta$} \}\] is finite for all $C \in {\mathbb{R}}_{\geqslant 0}$ and
$\delta \in {\mathbb{Z}}_{\geqslant 1}$.
\end{enumerate}
\begin{defi}\label{Def: arithmetic adelic structure}
Let $K$ be a countable field of characteristic $0$. An adelic structure of $K$ which satisfies the above conditions \ref{Item: propre}--\ref{Item: Northcott} is said to be \emph{arithmetic}.
\end{defi}
\begin{rema}
Note that the condition \ref{Item: Northcott} is analogous to Northcott's property in Diophantine geometry. In Arakelov geometry of adelic curve, we say that an adelic curve $S=(K,(\Omega,\mathcal A,\nu),\phi)$ \emph{has Northcott property} if the set
\[\{a\in K\,|\,h_S(1,a)\leqslant C\}\]
is finite for any $C\geqslant 0$ (see \cite[Definition 3.5.2]{CMArakelovAdelic}). In the case where the adelic curve $S$ is proper and has Northcott property, an analogue of Northcott's theorem holds (see \cite[Definition 3.5.3]{CMArakelovAdelic})
\end{rema}
In the remaining of the section, we fix a countable field $K$ of characteristic $0$ and a countable non-empty set $I$. We equip $K$ with an adelic structure $((\Omega,\mathcal A,\nu),\phi)$ to form an adelic curve, which we denote by $S$.
We also fix a family $(\iota_\omega)_{\omega\in\Omega_\infty}$ of embeddings from $K$ to $\mathbb C$ such that $|\raisebox{.4ex}{.}|_\omega=|\iota_\omega(\raisebox{.4ex}{.})|$ for any $\omega\in\Omega_\infty$ and that the map $(\omega\in\Omega_\infty)\mapsto\iota_\omega(a)$ is measurable for each $a\in K$ (see Theorem \ref{thm:measurable:family:embeddings}). For any element $f\in K[\boldsymbol{T}_I]$, we denote by $\iota_\omega(f)$ the polynomial in $\mathbb C[\boldsymbol{T}_I]$ defined as
\[\iota_\omega(f):=\sum_{\boldsymbol{d}\in\mathbb N^{\oplus I}}\iota_\omega(a_{\boldsymbol{d}}(f))\boldsymbol{T}_I^{\boldsymbol{d}}.\]
This defines a ring homomorphism from $K[\boldsymbol{T}_I]$ to $\mathbb C[\boldsymbol{T}_I]$, which extends to a homomorphism of fields form $K(\boldsymbol{T}_I)$ to $\mathbb C(\boldsymbol{T}_I)$, which we still denote by $\iota_\omega(\raisebox{.4ex}{.})$.
\begin{enonce}[remark]{Notation}
For convenience, for any $f\in K[\boldsymbol{T}_I]$, the complex polynomial $\iota_{\omega}(f)\in\mathbb C[\boldsymbol{T}_I]$ is often denoted by $f_{\omega}$.
\end{enonce}
For any $t\in[0,1]$, we denote by $e(t)$ the complex number $\mathrm{e}^{2\pi t\sqrt{-1}}$. For any $\omega\in\Omega_\infty$, we denote by $\Omega_{L,\omega}$ the set
\[
\Omega_{L,\omega}:=\left\{ (t_i)_{i\in I} \in [0,1]^I \left|
\begin{array}{l}
\text{\small $( e(t_i))_{i\in I}$ is algebraically} \\
\text{\small independent over $\iota_{\omega}(K)$}
\end{array}
\right\}\right..
\]
Note that by definition one has
\begin{equation}\label{Equ: transcendental points}[0,1]^I\setminus\Omega_{L,\omega}=\bigcup_{f\in K[\boldsymbol{T}_I]\setminus\{0\}}\{(t_i)_{i\in I}\in[0,1]^I\,:\,f_\omega\big((e(t_i))_{i\in I}\big)=0\}.\end{equation}
We equip $[0,1]^I$ with the product $\sigma$-algebra (namely the smallest $\sigma$-algebra making measurable the projection maps to the coordinates) and the product of the uniform probability measure on $[0,1]$, denoted by $\eta_I$ (see \cite[\S4.2]{MR0651017} for the product of an arbitrary family of probability spaces).
\begin{lemm}\label{Lem: outside Omega negligible}
For any $\omega\in\Omega_\infty$, the subset $\Omega_{L,\omega}$ of $[0,1]^I$ is measurable, and $[0,1]^I\setminus\Omega_{L,\omega}$ is $\eta_I$-negligible.
\end{lemm}
\begin{proof} The measurability of $\Omega_{L,\omega}$ follows from \eqref{Equ: transcendental points}.
For any non-zero element of $K[\boldsymbol{T}_I]$, let
\[V_I(f)=\{(t_i)_{i\in I}\in[0,1]^I\,:\,f_\omega\big((e(t_i))_{i\in I}\big)=0\}.\]
Since $K$ and $I$ are countable, $K[\boldsymbol{T}_I]$ is a countable set. Therefore, by \eqref{Equ: transcendental points}, to prove the second statement it suffices to show that $\eta_I(V_I(f))=0$. We first treat the case where $I$ is a finite set. Without loss of generality, we assume that $I=\{1,\ldots,n\}$, where $n\in\mathbb N$. The case where $n=0$ (namely $I=\varnothing$) is trivial since in this case $V_I(f)$ is empty. Assume that $n\geqslant 1$. For $t\in[0,1]$, let $f_t$ be the polynomial
\[\iota_{\omega}(f)(T_1,\ldots,T_{n-1},e(t))\in \iota_\omega(K)(e(t))[T_1,\ldots,T_{n-1}].\]
Then by Fubini's theorem, one has
\[\eta_{\{1,\ldots,n\}}(V_{\{1,\ldots,n\}}(f))=\int_{[0,1]}\eta_{\{1,\ldots,n-1\}}(V_{\{1,\ldots,n-1\}}(f_t))\,\mathrm{d}t=0,\]
where the second equality comes from the induction hypothesis.
We now consider the general case. Let $J$ be a finite subset of $I$ such that $f\in K[(T_i)_{i\in J}]$. By the definition of the product measure, one has
\[\eta_I(V_I(f))=\eta_J(V_J(f))=0.\]
\end{proof}
For any $\omega\in\Omega_\infty$, we equip $\Omega_{L,\omega}$ with the restriction of the product $\sigma$-algebra on $[0,1]^I$ and the restriction of the product probability measure $\eta_I$ to obtain a probability space denoted by $(\Omega_{L,\omega},\mathcal A_{L,\omega},\nu_{L,\omega})$.
Let $\phi_{L,\omega}:\Omega_{L,\omega}\rightarrow M_L$ be the map sending $x=(t_i)_{i\in I}\in\Omega_{L,\omega}$ to the absolute value
\[(f\in L)\longmapsto |f|_x:=\Big|f_\omega\big((e(t_i))_{i\in I}\big)\Big|.\]
Thus we obtain an adelic curve $S_{L,\omega}:=(L,(\Omega_{L,\omega},\mathcal A_{L,\omega},\nu_{L,\omega}),\phi_{L,\omega})$.
We recall Jensen's formula for Mahler measure of polynomials (see \cite{MR1554908} for a proof).
\begin{lemm}[Jensen's formula]
Let \[P(T)=a_d(T-\alpha_1)\cdots(T-\alpha_d)\in\mathbb C[T]\] be a complex polynomial of one variable $T$, with $a_d\in\mathbb C\setminus\{0\}$ and $(\alpha_1,\ldots,\alpha_d)\in\mathbb C^d$. One has
\[\int_0^1\ln|P(e(t))|\,\mathrm{d}t=\ln|a_d|+\sum_{j=1}^d\ln(\max\{1,|\alpha_j|\})\geqslant\ln|a_d|.\]
\end{lemm}
\begin{prop}\label{prop:admissible:fibration:K:T}
The family of adelic curves $(S_{L,\omega})_{\omega\in\Omega}$ is an admissible fibration over the adelic curve $S$. Moreover, in the case where the adelic curve $S$ is proper, for any $F\in\mathscr P_{K[T]}$, one has
\[h_{S_L}(F):=\int_{\Omega}\nu(\mathrm{d}\omega)\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\geqslant 0.\]
\end{prop}
\begin{proof}
{\bf Step 1.} By construction, for any $\omega\in\Omega$ and any $x\in\Omega_{L,\omega}$, the absolute value $\phi_{L,\omega}(x)$ on $L$ extends the absolute value $\phi(\omega)$ on $K$.
{\bf Step 2.} Let $g$ be a non-zero element of $K[\boldsymbol{T}_I]$, $(F_j)_{j=1}^n$ be elements of $\mathscr P_{K[\boldsymbol{T}_I]}$ containing \[\{F\in\mathscr P_{K[\boldsymbol{T}_I]}\,:\,\operatorname{ord}_F(g)\neq 0\},\] and $(C_j)_{j=1}^n\in\mathbb R_{\geqslant 0}^n$. We show that the function
\begin{equation}\label{Equ: interal sur Omega}(\omega\in\Omega)\longmapsto\int_{\Omega_{L,\omega}}|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\,\nu_{L,\omega}(\mathrm{d}x)\end{equation}
is $\mathcal A$-measurable. We choose a finite subset $J$ of $I$ such that $g$, $F_1,\ldots,F_n$ belong to $K[(T_i)_{i\in J}]$. By Lemma \ref{Lem: outside Omega negligible}, one has
\[\begin{split}
&\quad\int_{\Omega_{L,\omega}}|g|_x1\hspace{-0.25em}\mathrm{l}_{|F_1|_x\leqslant C_1,\ldots,|F_n|_x\leqslant C_n}\,\nu_{L,\omega}(\mathrm{d}x)\\&=\int_{[0,1]^I}\Big|g_\omega\big((e(t_i))_{i\in I}\big)\Big|\prod_{j=1}^n1\hspace{-0.25em}\mathrm{l}_{|F_{j,\omega}((e(t_i))_{i\in I})|\leqslant C_j}\,\eta_I(\mathrm{d}(t_i)_{i\in I})\\
&=\int_{[0,1]^J}\Big|g_\omega\big((e(t_i))_{i\in J}\big)\Big|\prod_{j=1}^n1\hspace{-0.25em}\mathrm{l}_{|F_{j,\omega}((e(t_i))_{i\in I})|\leqslant C_j}\,\eta_J(\mathrm{d}(t_i)_{i\in J})
\end{split}\]
Note that $[0,1]^J$ is a separable
compact metric space. By the criterion of measurability for functions on product measurable space proved in \cite[Lemma 9.2]{MR44536} and the measurability of integrals with parameter (see \cite[Lemma 1.26]{MR1464694}), we obtain the measurability of the function \eqref{Equ: interal sur Omega} on $\Omega_\infty$. The measurability of this function on $\Omega\setminus\Omega_\infty$ follows from Proposition \ref{Pro: S is an admissible family}.
{\bf Step 3.} It remains to show that the function
\begin{equation}\label{Equ: int ln F}(\omega\in\Omega)\longmapsto \int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\end{equation}
is well defined and is integrable for any $F\in\mathscr P_{K[T]}$. By Proposition \ref{Pro: S is an admissible family} again, it suffices to show its integrability on $\Omega_\infty$. Let
\[\Theta:=\{\boldsymbol{d}\in\mathbb N^{\oplus I}\,:\,a_{\boldsymbol{d}}(F)\neq 0\}.\]
One has
\[\ln|F|_x\leqslant\max_{\boldsymbol{d}\in\Theta}\ln|a_{\boldsymbol{d}}(F)|_\omega+\ln(\operatorname{card}(\Theta)).\]
Therefore, for $\omega\in\Omega_\infty$, the integral
\[\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\]
is well defined and the following inequality holds:
\begin{equation}\label{Equ: upper bound of the integral}\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\leqslant\max_{\boldsymbol{d}\in\Theta}\ln|a_{\boldsymbol{d}}(F)|_\omega+\ln(\operatorname{card}(\Theta)).\end{equation} Moreover, by an argument similar to that in Step 2, it can be shown that the function
\[(\omega\in\Omega_\infty)\longrightarrow \int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\]
is measurable. Finally, by writing
\[\int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\]
as successive integrals, and then by applying Jensen's formula in a recursive way, we obtain that there exists $\boldsymbol{d}_0\in\Theta$ such that
\begin{equation}\label{Equ: lower bound for int ln}\forall\,\omega\in\Omega_\infty,\quad \int_{\Omega_{L,\omega}}\ln|F|_x\,\nu_{L,\omega}(\mathrm{d}x)\geqslant\ln|a_{\boldsymbol{d}_0}(F)|_\omega.\end{equation}
Combining this inequality with \eqref{Equ: upper bound of the integral} and the fact that $\nu(\Omega_\infty)<+\infty$ (see \cite[Proposition 3.1.2]{CMArakelovAdelic}), we obtain the integrability of the function \eqref{Equ: int ln F} on $\Omega_\infty$. Finally, applying \eqref{Equ: local integral for finite places} to $\omega\in\Omega\setminus\Omega_\infty$, the inequality \eqref{Equ: lower bound for int ln} leads to
\[h_{S_L}(F)\geqslant\int_{\omega\in\Omega}\ln|a_{\boldsymbol{d_0}}(F)|_\omega
\,\nu(\mathrm{d}\omega)=0\]
provided that the adelic curve $S$ is proper. The proposition is thus proved.
\end{proof}
\begin{rema}
Note that, for $f \in L$,
\[
h_{S_L}(f) = \int_{\Omega_{\infty}} \nu(\mathrm{d}\omega) \int_{\Omega_{L,\omega}} \ln \Big|f_\omega\big(( e(t_i) )_{i \in I}\big)\Big|\, \eta_I (\mathrm{d} (t_i)_{i \in I}) + \int_{\Omega_{\rm{fin}}}
\ln |f|_{\omega} \,\nu(\mathrm{d}\omega).
\]
Thus $h_{S_L}(1) = 0$ and $h_{S_L}(T_i) = 0$ for all $i \in I$.
\end{rema}
\begin{defi}\label{def:S:L:lambda}
As a corollary, to the admissible fibration $(S_{L,\omega})_{\omega\in\Omega}$ one can associate an adelic structure $((\Omega_L,\mathcal A_L,\nu_L),\phi_L)$ on $L$ as in Definition \ref{Def: adelic structure coming from a fibration}. We fix $\lambda \in {\mathbb{R}}_{\geqslant 0}$.
Let $S_L^\lambda:=(L,(\Omega_L^\lambda,\mathcal A_L^\lambda,\nu_L^\lambda),\phi_L^\lambda)$ be an adelic curve with underlying field $L$ such that
\begin{enumerate}[label=\rm(\arabic*)]
\item $(\Omega_L^\lambda,\mathcal A_L^\lambda,\nu_L^\lambda)$ is the disjoint union of $(\Omega_L,\mathcal A_L,\nu_L)$ and $\mathscr P_{K[\boldsymbol{T}_I]} \cup \{ \infty \}$ equipped with the discrete $\sigma$-algebra and the measure satisfying
\[
\nu_L^\lambda(\{ F \}) = h_{S_L}(F) + \lambda \deg(F) \quad\text{and}\quad
\nu_L^\lambda(\{ \infty \}) = \lambda
\]
for any $F\in\mathscr P_{K[\boldsymbol{T}_I]}$.
\item the map $\phi_L^\lambda:\Omega_L^\lambda\rightarrow M_L$ extends $\phi_L$ and the map \[(x\in\mathscr P_{K[\boldsymbol{T}_I]} \cup \{\infty\} )\longmapsto|\raisebox{.4ex}{.}|_x.\]
\end{enumerate}
The adelic curve $S^\lambda_L$ is called the \emph{$\lambda$-twisted compactification} of $S_L$.
\end{defi}
\begin{rema}\label{Rem:S:L:lambda}
Note that if $\lambda = 0$, then $S_L^{\lambda} = S_L^*$.
Moreover, if $K$ and $\Omega_{\rm{fin}}$ are countable and
$\mathcal A_{\Omega_{\rm{fin}}}$ is discrete, then
$L$ and $\Omega_{L,\rm{fin}}^*$ are countable and
$\mathcal A_{\Omega_{L,\rm{fin}}^*}$ is discrete.
\end{rema}
\begin{prop}\label{prop:proper:L:lambda}
The adelic curve $S_L^\lambda = (L,(\Omega_L^\lambda,\mathcal A_L^\lambda,\nu_L^\lambda),\phi_L^\lambda)$ is proper.
\end{prop}
\begin{proof}
If $\lambda = 0$, then the assertion follows from Proposition~\ref{Pro: compactification} and Proposition~\ref{prop:admissible:fibration:K:T}.
Note that
\begin{equation}\label{Equ: deg equality}
\deg(g) = \sum_{F \in \mathscr P_{K[\boldsymbol{T}_I]}} \deg(F) \operatorname{ord}_F(g)
\end{equation}
for $g \in L^{\times}$,
so that
\[
\sum_{F \in \mathscr P_{K[\boldsymbol{T}_I]}} (h_{S_L}(F) + \lambda \deg(F)) (-\operatorname{ord}_F(g)) + \lambda \deg(g)
= \sum_{F \in \mathscr P_{K[\boldsymbol{T}_I]}} h_{S_L}(F) (-\operatorname{ord}_F(g)),
\]
as required.
\end{proof}
\begin{rema}
The above result can be considered as a particular case of Proposition \ref{Pro: compactification 2}. In fact, if we equip $\mathscr P_{K[\boldsymbol{T}_I]}\cup\{\infty\}$ with the discrete $\sigma$-algebra $\mathcal A'$ and the measure $\nu'$ such that $\nu'(\{\infty\})=1$ and $\nu'(\{F\})=\deg(F)$, then \[(L,(\mathscr P_{K[\boldsymbol{T}_I]}\cup\{\infty\},\mathcal A',\nu'),\phi')\]
forms an adelic curve, where $\phi'$ sends $x\in\mathscr P_{K[\boldsymbol{T}_I]}\cup\{\infty\}$ to the absolute value $|\raisebox{.4ex}{.}|_x$. Then the equality \eqref{Equ: deg equality} shows that this adelic curve is proper. Note that the restriction of $\nu^\lambda_L$ on $\mathscr P_{K[\boldsymbol{T}_I]}\cup\{\infty\}$ coincides with
\[\lambda\nu_L'+\sum_{F\in\mathscr P_{K[\boldsymbol{T}_I]}}h_{S_L}(F)\operatorname{Dirac}_F.\]
Therefore the statement of Proposition \ref{prop:proper:L:lambda} follows from Proposition \ref{Pro: compactification 2}.
\end{rema}
\begin{lemm}\label{lem:deg:mu:formula}
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item
If $F_0, \ldots, F_r \in K[\boldsymbol{T}_I]$ with $(F_0, \ldots, F_r) \not= (0, \ldots, 0)$, then
{\allowdisplaybreaks
\begin{multline*}
\kern3em h_{S_L^\lambda}(F_0, \ldots, F_r) \leqslant
\int_{\Omega_{L,\infty}} \ln \max \{ |F_0|_{x}, \ldots, |F_r|_{x} \}
\nu_{L,\infty}(\mathrm{d} x) \\
\kern10em + \int_{\Omega_{\operatorname{fin}}} \ln \max \{ |F_0|_{\omega}, \ldots, |F_r|_{\omega} \}
\nu_{\operatorname{fin}}(\mathrm{d} \omega) \\
+ \lambda \max \{ \deg(F_0), \ldots, \deg(F_r) \}.
\end{multline*}}
Moreover, if $\mathrm{G.C.D}(F_0, \ldots, F_r) = 1$, then the equality holds.
\item
Fix $n \in I$ and let $I' = I \setminus \{ n \}$ and $L' = K(\boldsymbol{T}_{I'})$.
For $F \in K[\boldsymbol{T}_I] \setminus \{ 0 \}$, if we set
$F = a_0 T_n^d + a_1 T_n^{d-1} + \cdots + a_d$ such that $a_0, a_1, \ldots, a_d \in K[\boldsymbol{T}_{I'}]$ and $a_0 \not= 0$,
then \[h_{S_{L'}^\lambda}(a_0, \ldots, a_d) \leqslant h_{S_L}(F) + \deg(F) (\lambda + \ln(2)
\nu(\Omega_{\infty})).\]
\end{enumerate}
\end{lemm}
\begin{proof}
(1) Note that
\[
\max \{ |F_0|_{\xi}, \ldots, |F_r|_{\xi} \} \begin{cases}
\leqslant 1 & \text{in general}, \\
= 1 & \text{if $\mathrm{G.C.D}(F_0, \ldots, F_r) = 1$},
\end{cases}
\]
for $\xi \in\mathscr P_{K[\boldsymbol{T}_I]}$,
so that the assertion follows.
\medskip
(2)
Note that $d \leqslant \deg(F)$. We set $f = F/a_0$.
For $y \in \Omega_{L', \infty}$,
let \[f_{y} = T_n^d + \iota_{y}(a_1/a_0)T_n^{d-1} + \cdots + \iota_{y}(a_d/a_0) = (T_n - \alpha_1) \cdots (T_n - \alpha_d)\] be the irreducible decomposition in ${\mathbb{C}}[T_n]$.
Then,
\[
\iota_{y}(a_k/a_0) = (-1)^k \sum_{1 \leqslant i_1 < \cdots < i_k \leqslant d} \alpha_{i_1} \cdots \alpha_{i_k},
\]
so that {\allowdisplaybreaks
\begin{align*}
|a_k/a_0|_{y} & \leqslant \sum_{1 \leqslant i_1 < \cdots < i_k \leqslant d} |\alpha_{i_1}| \cdots |\alpha_{i_k}| \leqslant \sum_{1 \leqslant i_1 < \cdots < i_k \leqslant d} \max \{1, |\alpha_{i_1}|\} \cdots\max \{1, |\alpha_{i_k}|\} \\
& \leqslant \sum_{1 \leqslant i_1 < \cdots < i_k \leqslant d} \max \{1, |\alpha_{1}|\} \cdots\max \{1, |\alpha_{d}|\} \\
& \leqslant 2^{\deg(F)} \max \{1, |\alpha_{1}|\} \cdots\max \{ 1, |\alpha_{d}|\}
\end{align*}}
because $ \binom{d}{k} \leqslant 2^d \leqslant 2^{\deg(F)}$,
and hence one has
\[
\max \{ 1, |a_k/a_0|_{y}\} \leqslant 2^{\deg(F)} \max \{1, |\alpha_{1}|\} \cdots\max \{1, |\alpha_{d}|\}.
\]
On the other hand, by Jensen's formula,
\[
\int_{0}^1 \ln |f_{y}(e(t_n))| \,\mathrm{d}t_n = \sum_{i=1}^d \ln \max \{ 1, |\alpha_i| \}.
\]
Therefore, one obtains
\[
\ln \max\{ 1, |a_k/a_0|_{y} \} \leqslant \int_{0}^1 \ln |f_{y}(e(t_n))| \,\mathrm{d}t_n +\deg(F)\ln(2)
\]
for all $k=1, \ldots, d$,
so that
{\allowdisplaybreaks
\begin{align*}
\ln \max\{ |a_0|_{y}, |a_1|_{y}, \ldots, |a_d|_{y} \} \\
& \kern-10em =
\ln |a_0|_{y} +
\ln \max\{ 1, |a_1/a_0|_{y}, \ldots, |a_d/a_0|_{y} \} \\
& \kern-10em \leqslant \ln |a_0|_{y} + \int_{0}^1 \ln |f_{y}(e(t_n))|\, \mathrm{d}t_n + \deg(F)\ln(2) \\
& \kern-10em = \int_{0}^1 \ln |F_{y}(e(t_n))|\, \mathrm{d}t_n + \deg(F)\ln(2).
\end{align*}}
Thus, by
Fubini's theorem,
{\allowdisplaybreaks
\begin{align*}
\int_{\Omega_{L,\infty}} \ln |F|_{x} \nu_{L, \infty}(\mathrm{d} x) & =
\int_{\Omega_{L', \infty} \times [0,1]} \ln |F_{y}(e(t_n))|\,
\nu_{L'}(\mathrm{d}y)\, \mathrm{d} t_n \\
& \kern-8em= \int_{\Omega_{L', \infty}} \left(
\int_{0}^1 \ln |F_{y}(e(t_n))|\, \mathrm{d}t_n\right)\, \nu_{L'}(\mathrm{d} y) \\
& \kern-8em\geqslant \int_{\Omega_{L', \infty}} \left( \ln \max\{|a_0|_{y}, \ldots, |a_d|_{y} \} - \deg(F) \ln(2) \right) \, \nu_{L'}(\mathrm{d} y) \\
& \kern-8em = \int_{\Omega_{L', \infty}} \ln \max\{ |a_0|_{y}, \ldots, |a_d|_{y}\}\, \nu_{L'}(\mathrm{d} y) - \deg(F) \ln(2)\nu_{L'}(\Omega_{L', \infty}) \\
& \kern-8em = \int_{\Omega_{L', \infty}} \ln \max\{ |a_0|_{y}, \ldots, |a_d|_{y}\} \, \nu_{L'}(\mathrm{d} y) - \deg(F) \ln(2)\nu(\Omega_{\infty}).
\end{align*}}
On the other hand, note that \[
|F|_{\omega} = \max \{ |a_0|_{\omega}, \ldots, |a_d|_{\omega}\}
\]
for $\omega \in \Omega_{\rm{fin}}$, so that
{\allowdisplaybreaks
\begin{multline*}
\int_{\Omega_{L', \infty}} \ln \max\{ |a_0|_{y}, \ldots, |a_d|_{y}\} \, \nu_{L'}(\mathrm{d} y) \\
+
\int_{\Omega_{\operatorname{fin}}} \ln \max\{ |a_0|_{\omega}, \ldots, |a_d|_{\omega}\}\, \nu(\mathrm{d} \omega) \\
\leqslant h_{S_L}(F) + \deg(F) \ln(2)\,\nu(\Omega_{\infty}).
\end{multline*}}{\allowdisplaybreaks
\begin{align*}
h_{S_{L'}^{\lambda}}(a_0,\ldots, a_d) & \leqslant \int_{\Omega_{L', \infty}} \ln \max\{ |a_0|_{y}, \ldots, |a_d|_{y}\} \,\nu_{L'}(\mathrm{d} y) \\
& \kern5em + \int_{\Omega_{\mathrm{fin}}}
\ln \max \{ |a_0|_{\omega}, \ldots, |a_d|_{\omega} \}\,\nu (\mathrm{d} \omega) \\
& \kern10em + \lambda\max \{\deg(a_0), \ldots, \deg(a_d) \} \\
& \leqslant h_{S_L}(F) + \deg(F) (\lambda + \ln(2)\nu(\Omega_{\infty})),
\end{align*}}
as required.
\end{proof}
Fix $n \in I$ and let $I' = I \setminus \{ n \}$ and $L' = K(\boldsymbol{T}_{I'})$.
For $F \in K[\boldsymbol{T}_I] \setminus \{ 0 \}$, we set
$F = a_0 T_n^{d} + \cdots + a_d$ such that $a_0, \ldots, a_d \in K[\boldsymbol{T}_{I'}]$ and $a_0 \not= 0$. We define $\bbnu(F)$ to be
\[\bbnu(F) := F/a_0 = T_n^{d} + (a_1/a_0)T_n^{d-1} + \cdots + (a_d/a_0).\] Note that $\bbnu(F)$ is a monic polynomial over $L'$.
\begin{prop}\label{prop:finiteness:deg:mu:bounded}
If $S_{L'}^{\lambda}$ has Northcott's property, then,
for $C \in {\mathbb{R}}$ and $\delta \in {\mathbb{Z}}_{\geqslant 1}$, then the set
\[\{ \bbnu(F) \mid \text{$F \in K[\boldsymbol{T}_I] \setminus \{ 0 \}$, $h_{S_L}(F) \leqslant C$ and
$\deg(F) \leqslant \delta$} \}\] is finite.
\end{prop}
\begin{proof}
Let $\Theta := \{ F \in K[\boldsymbol{T}_I] \setminus \{ 0 \} \mid \text{$h_{S_L}(F) \leqslant C$ and
$\deg(F) \leqslant \delta$} \}$ and $\vartheta : \Theta \to {\mathbb{P}}^{\delta}(L')$
be a map given by the following way:
for \[F = a_0 T_n^{d} + \cdots + a_d \in \Theta\quad (\text{$a_0, \ldots, a_d \in K[\boldsymbol{T}_I]$ and $a_0 \not= 0$}),\] \[\vartheta(F) :=
\overbrace{(a_0 : \cdots : a_d : 0 : \cdots : 0)}^{\delta+1}
\in {\mathbb{P}}^{\delta}(L').\]
By Lemma~\ref{lem:deg:mu:formula},
{\allowdisplaybreaks
\begin{multline*}
h_{S_{L'}^{\lambda}}(\vartheta(F)) \leqslant h_{S_L}(F) + \deg(F) (\lambda + \ln(2) \nu(\Omega_{\infty})) \leqslant C + \delta(\lambda + \ln(2) \nu(\Omega_{\infty})).
\end{multline*}}
Thus the assertion of the proposition is a consequence of Northcott's property of $S_{L'}^{\lambda}$.
\end{proof}
\begin{prop}\label{prop:Northcott:property}
If $S$ has Northcott's property, $\operatorname{card}(I) < \infty$ and $\lambda > 0$, then $S_L^{\lambda}$ has also Northcott's property.
\end{prop}
\begin{proof}
We prove it by induction on $\operatorname{card}(I)$. If $\operatorname{card}(I)=0$, then the assertion is obvious because
$S_{L}^{\lambda} = S$.
Fix $n \in I$ and let $I' = I \setminus \{ n \}$ and $L' = K(\boldsymbol{T}_{I'})$.
It is sufficient to see that $\{ f \in L^{\times} \mid h_{S_L^{\lambda}} (f, 1 ) \leqslant C \}$ is finite for any $C$.
For $f \in L^{\times}$, let us choose $F_1, F_2 \in K[\boldsymbol{T}_{I}] \setminus \{ 0 \}$
such that $f = F_1/F_2$, and $F_1$ and $F_2$ are relatively prime.
We set
\[
\begin{cases}
F_1 = a_{10}T_n^{d_1} + \cdots + a_{1d_1}\quad(\text{$a_{10}, \ldots, a_{1d_1} \in K[\boldsymbol{T}_{I'}]$ and $a_{10} \not= 0$}),\\
F_2 = a_{20}T_n^{d_2} + \cdots + a_{2d_2}\quad(\text{$a_{20}, \ldots, a_{2d_2} \in K[\boldsymbol{T}_{I'}]$ and $a_{20} \not= 0$}).
\end{cases}
\]
\begin{enonce}{Claim}\label{claim:thm:Northcott:property:01}
If $h_{S_L^{\lambda}}(f , 1) \leqslant C$, then one has the following:
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item $\max \{ \deg(F_1), \deg(F_2) \} \leqslant C/\lambda$ and $\max \{ h_{S_L}(F_1), h_{S_L}(F_2) \} \leqslant C$.
\item $h_{S_{L'}^{\lambda}}(a_{10} , a_{20}) \leqslant C$.
\end{enumerate}
\end{enonce}
\begin{proof}
(1) As $C \geqslant h_{S_L^{\lambda}}(f, 1) = h_{S_L^{\lambda}}(F_1, F_2)$ and
$F_1$ and $F_2$ are relatively prime, by (1) in Lemma~\ref{lem:deg:mu:formula}, one has
{\allowdisplaybreaks
\begin{multline}\label{eqn:thm:Northcott:property:01}
C \geqslant \lambda \max \{ \deg(F_1), \deg(F_2) \} + \int_{\Omega_{L, \infty}} \ln \max \{ |F_1|_{x}, |F_2|_{x} \}\,
\nu_{L}(\mathrm{d} x) \\
+ \int_{\Omega_{\mathrm{fin}}} \ln \max \{ |F_1|_{\omega}, |F_2|_{\omega} \}\, \nu(\mathrm{d} \omega).
\end{multline}}
Thus,
{\allowdisplaybreaks
\begin{align*}
C & \geqslant \lambda \max \{ \deg(F_1), \deg(F_2) \} + \max \{ h_{S_L}(F_1), h_{S_L}(F_2) \}
\end{align*}}
Therefore, (1) follows because $h_{S_L}(F_1), h_{S_L}(F_2) \geqslant 0$.
\medskip
(2) By (1) in Lemma~\ref{lem:deg:mu:formula},
{\allowdisplaybreaks
\begin{align*}
h_{S_{L'}^{\lambda}}(a_{10}, a_{20}) & \leqslant \lambda \max \{ \deg(a_{10}), \deg(a_{20}) \} \\
& \kern5em + \int_{\Omega_{L', \infty}} \ln \max \{ |a_{10}|_{y}, |a_{20}|_{y} \}\,\nu_{L'}(\mathrm{d} y) \\
& \kern8em + \int_{\Omega_{\mathrm{fin}}} \ln \max \{ |a_{10}|_{\omega}, |a_{20}|_{\omega} \} \,\nu(\mathrm{d} \omega).
\end{align*}}
Therefore, by \eqref{eqn:thm:Northcott:property:01},
it is sufficient to see the following:
{\allowdisplaybreaks
\begin{equation}\label{eqn:thm:Northcott:property:02}
\int_{\Omega_{L, \infty}} \ln \max \{ |F_1|_{x}, |F_2|_{x} \}
\nu_{L,\infty}(\mathrm{d} x)
\geqslant
\int_{\Omega_{L', \infty}} \ln \max \left\{ |a_{10}|_{y}, |a_{20}|_{y}
\right\} \,
\nu_{L'}(\mathrm{d} y)
\end{equation}}
and
{\allowdisplaybreaks
\begin{equation}\label{eqn:thm:Northcott:property:03}
\int_{\Omega_{\mathrm{fin}}} \ln \max \{ |F_1|_{\omega}, |F_2|_{\omega} \} \,\nu(\mathrm{d} \omega) \geqslant
\int_{\Omega_{\mathrm{fin}}} \ln \max \{ |a_{10}|_{\omega}, |a_{20}|_{\omega} \} \,\nu(\mathrm{d} \omega).
\end{equation}}
Indeed, by Jensen's formula together with Fubini's theorem,
{\allowdisplaybreaks
\begin{align*}
\int_{\Omega_{L, \infty}} \ln \max \{ |F_1|_{x}, |F_2|_{x} \}\,
\nu_{L}(\mathrm{d} x) \\
&\kern-8em= \int_{\Omega_{L', \infty} \times [0,1]} \ln \max_{i=1,2} \{ |F_{iy}(e(t_n))| \}
\,\nu_{L'}(\mathrm{d}y) \mathrm{d}t_n \\
&\kern-8em= \int_{\Omega_{L', \infty}} \left(\int_{0}^1 \ln \max_{i=1,2} \{ |F_{iy}(e(t_n))| \}\mathrm{d}t_{n} \right) \,\nu_{L'}(\mathrm{d}y) \\
& \kern-8em\geqslant \int_{\Omega_{L', \infty}} \max_{i=1,2} \left\{ \int_{0}^1 \ln |F_{iy}(e(t_n))|\mathrm{d}t_{n} \right\}\,\nu_{L'}(\mathrm{d}y) \\
& \kern-8em\geqslant \int_{\Omega_{L', \infty}} \max_{i=1,2} \left\{ \ln |a_{i0y}| \right\} \,\nu_{L'}(\mathrm{d}y) \\
&\kern-8em= \int_{\Omega_{L', \infty}} \ln \max \left\{ |a_{10}|_{y}, |a_{20}|_{y}
\right\}
\,\nu_{L'}(\mathrm{d} y),
\end{align*}}
as required for \eqref{eqn:thm:Northcott:property:02}. Further, since $|F_1|_{\omega} \geqslant |a_{10}|_{\omega}$ and $|F_2|_{\omega} \geqslant |a_{20}|_{\omega}$, one has \eqref{eqn:thm:Northcott:property:03}.
\end{proof}
If we set
\[
\begin{cases}
\Delta= \{ \bbnu(F) \mid \text{$F \in K[\boldsymbol{T}_I] \setminus \{ 0 \}$, $h_{S_L}(F) \leqslant C$ and $\deg(F) \leqslant C/\lambda$ }\}, \\
\Delta' = \{ a \in K(\boldsymbol{T}_{I'}) \mid \text{$h_{S_{L'}^{\lambda}}(a, 1) \leqslant C$} \},
\end{cases}
\]
then, by Proposition~\ref{prop:finiteness:deg:mu:bounded} together with the hypothesis of
induction, $\Delta$ and $\Delta'$ are finite.
Moreover, by Claim~\ref{claim:thm:Northcott:property:01},
if $h_{S_L^{\lambda}}(f, 1) \leqslant C$, then
\[\bbnu(F_1), \bbnu(F_2) \in \Delta\quad\text{and}\quad
a_{10}/a_{20} \in \Delta'.\]
Thus the assertion follows because $f = (a_{10}/a_{20})(\bbnu(F_1)/\bbnu(F_2))$.
\end{proof}
\begin{rema}
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item Note that $h_{S_L^{\lambda}}(1, T_n) = \lambda$ for all $n \in I$,
so that Northcott's property does not hold for $S_L^{\lambda}$ if $\Lambda$ is
infinite.
\item Let $S_{{\mathbb{Q}}}$ be the standard adelic structure of ${\mathbb{Q}}$. Then, it is easy to see that
\[
h_{(S_{{\mathbb{Q}}})_{{\mathbb{Q}}(T)}^*}(1, T^n-1) = \int_{0}^1 \ln \max \{1, |e(nt) - 1|\}\, \mathrm{d}t \leqslant \ln 2
\]
for all $n \geqslant 0$, so that the Northcott's property does not hold for $S_{{\mathbb{Q}}(T)}^*$.
\end{enumerate}
\end{rema}
\begin{theo}\label{thm:Northcott:S:Lambda:L}
We use the same notation as in Section~\ref{subsec:adelic:structure:of:purely:transcendental:fibration}.
We assume that $S$ has Northcott's propery and $\lambda > 0$.
Let $E$ be an algebraic closure of $L = K(\boldsymbol{T}_{I})$.
If $E_0$ is a subfield of $E$ such that $E_0$ is finitely generated over $K$,
then $S_L^{\lambda} \otimes_L E$ has Northcott's property over $E_0$, that is, \[\Big\{ a \in E \mid \text{$h_{S_L^{\lambda} \otimes E}(1 , a) \leqslant C$ and $[E_0(a) : E_0] \leqslant \delta$} \Big\}\] is finite
for any $C \in {\mathbb{R}}_{\geqslant 0}$ and $\delta\in{\mathbb{Z}}_{\geqslant 1}$.
\end{theo}
\begin{proof}
Since $E_0$ is finitely generated over $K$ and $E$ is algebraic over $L$,
we can choose a finite subset $I'$ of $I$ such that
$E_0(\boldsymbol{T}_{I'})$ is finite over $K(\boldsymbol{T}_{I'})$.
It is sufficient to see that the set
\begin{equation}\label{eqn:lemma:Northcott:S:Lambda:L:01}
\Big\{ \alpha \in E \mid
\text{$h_{S_L^{\lambda} \otimes E}(1, \alpha) \leqslant C$ and $[ K(\boldsymbol{T}_{I'})(\alpha):K(\boldsymbol{T}_{I'})] \leqslant \delta$}\Big\}
\end{equation}
is finite for any $C \in {\mathbb{R}}_{\geqslant 0}$ and $\delta \in {\mathbb{Z}}_{\geqslant 1}$. Indeed, note that
\begin{align*}
[K(\boldsymbol{T}_{I'})(\alpha):K(\boldsymbol{T}_{I'})] & \leqslant [E_0(\boldsymbol{T}_{I'})(\alpha):K(\boldsymbol{T}_{I'})] \\
& = [E_0(\boldsymbol{T}_{I'})(\alpha):E_0(\boldsymbol{T}_{I'})][E_0(\boldsymbol{T}_{I'}):K(\boldsymbol{T}_{I'})] \\
& \leqslant [E_0(\alpha) : E_0][E_0(\boldsymbol{T}_{I'}):K(\boldsymbol{T}_{I'})],
\end{align*}
so that
\begin{multline*}
\Big\{ a \in E \mid
\text{$h_{S_L^{\lambda} \otimes E}(1, a) \leqslant C$ and $[ E_0(a):E_0] \leqslant \delta$}\Big\} \\
\kern-5em \subseteq
\Big\{ \alpha \in E \mid
\text{$h_{S_L^{\lambda} \otimes E}(1, \alpha) \leqslant C$ and }\\
\kern10em \text{$[K(\boldsymbol{T}_{I'})(\alpha):K(\boldsymbol{T}_{I'})] \leqslant \delta[ E_0(\boldsymbol{T}_{I'}):K(\boldsymbol{T}_{I'})]$}\Big\}.
\end{multline*}
Let $\alpha$ be an element of the set \eqref{eqn:lemma:Northcott:S:Lambda:L:01}.
Let $f(t)$ be the minimal polynomial of $\alpha$ over $K(\boldsymbol{T}_{I'})$.
As $K(\boldsymbol{T}_{I})$ is a regular extension over $K(\boldsymbol{T}_{I'})$,
\[
K(\boldsymbol{T}_{I})[t]/f(t)K(\boldsymbol{T}_{I})[t] \simeq \left(K(\boldsymbol{T}_{I'})[t]/f(t)K(\boldsymbol{T}_{I'})[t]\right) \otimes_{K(\boldsymbol{T}_{I'})} K(\boldsymbol{T}_{I})
\]
is an integral domain, so that $f(t)$ is irreducible over $K(\boldsymbol{T}_{I})$, and hence $f(t)$ is also the minimal polynomial of $\alpha$ over $K(\boldsymbol{T}_{I})$.
We set
\[f = t^d + a_1t^{d-1} + \cdots + a_d\quad(a_1, \ldots, a_d \in K(\boldsymbol{T}_{I'})).\]
Then, in the same arguments as \cite[Theorem~3.5.3]{CMArakelovAdelic},
one has
\[
h_{S_L^{\lambda}}(1, a_1, \ldots, a_d) \leqslant \delta C + (\delta - 1) \ln(2) \nu(\Omega_{\infty}),
\]
so that
$h_{S_{K(\boldsymbol{T}_{I'})}^{\lambda}}(1, a_1, \ldots, a_d) \leqslant \delta C + (\delta - 1) \ln(2) \nu(\Omega_{\infty})$.
Therefore, the assertion is a consequence of Proposition~\ref{prop:Northcott:property}.
\end{proof}
\begin{theo}\label{thm:adelic:structure:countable}
If $E$ is a countable field of characteristic zero, then
$E$ has an arithmetic adelic structure (see Definition \ref{Def: arithmetic adelic structure}).
\end{theo}
\begin{proof}
We denote by $S$ the standard adelic curve with $\mathbb Q$ as underlying field. Recall that the measure space of $S$ is given by the set of all places of $\mathbb Q$ equipped with the discrete $\sigma$-algebra and the counting measure.
Let $\{ x_n \}_{n=1}^N$ be a transcendental basis of $E$ over ${\mathbb{Q}}$. Note that $N$ might be $+\infty$. Moreover, $E$ is algebraic over $L:={\mathbb{Q}}((x_n)_{n=1}^N)$. Let $\lambda$ be a positive number.
Starting from the adelic curve $S$, by the way in Subsetion~\ref{subsec:adelic:structure:of:purely:transcendental:fibration},
let $S_{L}^{\lambda}$ be the $\lambda$-twisted compactification of $S_L$.
We claim that the adelic curve $S_L^\lambda\otimes_L E$ satisfies the properties (1) -- (4) characterizing an arithmetic adelic curve.
The property (1) follows from Proposition~\ref{prop:proper:L:lambda} and \cite[Proposition~3.4.10]{CMArakelovAdelic}.
The property (2) is obvious.
For (3), see Lemma~\ref{lemma:discrete:algebraic} and Remark~\ref{Rem:S:L:lambda}.
Finally the property (4) follows from Theorem~\ref{thm:Northcott:S:Lambda:L}.
\end{proof}
\subsection{Density of Fermat property over arithmetic function fields}
In this subsection, let us consider a simple application of Theorem~\ref{thm:adelic:structure:countable}
together with Faltings' theorem \cite{MR1307396}.
Let $K$ be a field. We denote by $\mu(K)$ the subgroup of $K^{\times}$
consisting of roots of unity in $K$, that is, \[\mu(K) := \{ a \in K \mid \text{$a^n = 1$ for some $n \in {\mathbb{Z}}_{>0}$} \}.\]
Let $N$ be a positive integer and let $F_N := \operatorname{Spec}({\mathbb{Z}}[X, Y]/(X^N + Y^N - 1))$.
We say that \emph{$F_N$ has Fermat's property over $K$} if $x, y \in \mu(K) \cup \{ 0 \}$ for all $(x, y) \in F_N(K)$.
Then one has the following theorem.
\begin{theo}\label{thm:density:Fermat:conj}
If $K$ is an arithmetic function field, then
\[
\lim_{m\to\infty} \frac{\#\{ N \in {\mathbb{Z}} \mid \text{$1 \leq N \leq m$ and $F_N$ has Fermat's property over $K$} \}}{m}
= 1.
\]
\end{theo}
\begin{proof}
Let $S$ be a proper adelic structure of $K$ with Northcott's property (cf. Theorem~\ref{thm:adelic:structure:countable}).
Let us begin with the following claim:
\begin{enonce}{Claim}
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item
For $x, y \in K$, $h_S(x,y,1) = 0$ if and only if $x, y \in \mu(K) \cup \{ 0 \}$.
\item If $N \geq 4$, then there is a positive integer $m_0$ such that
$F_{Nm}$ has Fermat's property of every integer $m \geq m_0$.
\end{enumerate}
\end{enonce}
\begin{proof}
(1) We assume that $h_S(x,y,1) = 0$ for $x, y \in K$. Then $h_S(x^n, y^n, 1) = n h_S(x, y, 1) = 0$ for all $n \in {\mathbb{Z}}_{>0}$, so that,
by Northcott's property, \[\{ (x^n, y^n) \mid n \in {\mathbb{Z}}_{>0} \}\] is finite. Therefore, there are $n, n' \in {\mathbb{Z}}_{>0}$ such that
$n > n'$ and $(x^n, y^n) = (x^{n'}, y^{n'})$, and hence $x, y \in \mu(K) \cup \{ 0 \}$.
The converse is obvious.
\smallskip
(2) First of all,
note that $F_N(K)$ is finite by Faltings' theorem \cite{MR1307396}.
We set
\[
\begin{cases}
H := \max \{h_S(x, y, 1) \mid (x, y) \in F_N(K) \}, \\
a := \inf \{ h_S(x, y, 1) \mid \text{$x, y \in K$ and $h_S(x, y, 1) > 0$} \}.
\end{cases}
\]
Note that $a > 0$ by Northcott's property. For a positive integer $m$ with $m \geq \exp(H/a)$, we assume that $h_S(x, y, 1) > 0$ for some $(x, y) \in F_{Nm}(K)$.
Then, as $(x^m, y^m) \in F_N(K)$,
\[
H \geq h_S(x^m, y^m, 1) = m h_S(x, y, 1) \geq m a,
\]
so that $\exp(H/a) \geq \exp(m)$, and hence $m \geq \exp(m)$. This is a contradiction.
Therefore, $h_S(x, y, 1) = 0$ for all $(x, y) \in F_{Nm}(K)$. Thus, by (1), $F_{Nm}$ has Fermat's property.
\end{proof}
By (2) together with \cite[Lemma~5.16]{IKMFaltings},
one can conclude the assertion of the theorem.
\end{proof}
In the case where $K = {\mathbb{Q}}$, it was proved by \cite{MR734070, MR777765, MR766439} (cf. \cite{MR1719329}). A general number field case is treated in \cite{IKMFaltings}. The above theorem gives an evidence of the following conjecture:
\begin{enonce}{Conjecture}[Fermat's conjecture over an arithmetic function field] Let K be an arithmetic function field.
Then is there a positive integer $N_0$ depending on $K$ such that $F_N$ has Fermat's property over K for all $N \geq N_0$?
\end{enonce}
\section{Polarized adelic structure}\label{sec:Polarized adelic structure}
In this subsection, we recall an adelic structure induced by a polarization of a field.
Let $K$ be a finitely generated field over $\mathbb Q$ and $n$ be the transcendental degree of $K$ over ${\mathbb{Q}}$.
Let $\mathscr B\rightarrow\operatorname{Spec}\mathbb Z$ be a normal projective arithmetic variety such that the function field of $\mathscr B$ is $K$.
Note that $\dim \mathscr B = n+1$.
Let \[\big(\mathscr B; \overline{\mathscr H}_1 = (\mathscr H_1, h_1), \ldots, \overline{\mathscr H}_n = (\mathscr H_n, h_n)\big)\] be data with the following properties:
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item $\mathscr H_1, \ldots, \mathscr H_n$ are invertible ${\mathscr{O}}_{\mathscr B}$-modules that
are nef along all fibers of $\mathscr B \to \operatorname{Spec}({\mathbb{Z}})$.
\item The second entries $h_1, \ldots, h_n$ are semipositive metrics of $\mathscr H_1, \ldots, \mathscr H_n$ on
$\mathscr B({\mathbb{C}})$, respectively.
\item For each $i=1,\ldots,n$, the associated height function with $\overline{\mathscr H}_i$ is non-negative
\end{enumerate}
According to \cite{MR1779799}, the data $(\mathscr B; \overline{\mathscr H}_1, \ldots, \overline{\mathscr H}_n)$
is called a \emph{polarization of $K$}.
Let $x$ be a ${\mathbb{C}}$-valued point of $\mathscr B$, that is, there are a unique scheme point $p_x \in \mathscr B$ and
a unique homomorphism $\phi_x :{\mathscr{O}}_{\mathscr B, p_x} \to {\mathbb{C}}$ such that
$x$ is given by $\phi_x$.
We say $x$ is \emph{generic} if $p_x$ is the generic point of $\mathscr B$.
We denote the set of all generic ${\mathbb{C}}$-valued points by $\mathscr B({\mathbb{C}})_{\mathrm{gen}}$.
Note that the measure of $\mathscr B({\mathbb{C}}) \setminus \mathscr B({\mathbb{C}})_{\mathrm{gen}}$ is zero.
The polarization $(\mathscr B; \overline{\mathscr H}_1, \ldots, \overline{\mathscr H}_n)$ yields a proper
adelic structure of $K$ in the following way.
First of all, we set
\[
\begin{cases}
\Omega_{\infty} := \mathscr B({\mathbb{C}})_{\mathrm{gen}},\\
\Omega \setminus \Omega_{\infty} :=
\text{the set of all prime divisors on $\mathscr B$}.
\end{cases}
\]
For each element of $\omega \in \Omega$, $|\raisebox{.4ex}{.}|_{\omega}$ is give by
\[
\begin{cases}
| f |_x := |\phi_x(f)| & \text{if $x \in \Omega_{\infty}$},\\
|f|_{\Gamma} := \exp\left(-\operatorname{ord}_{\Gamma}(f)
\right) &
\text{if $\Gamma \in \Omega \setminus \Omega_{\infty}$}
\end{cases}
\]
for $f \in K$. Note that $\Omega_{\infty}$ is a measurable subset of a projective space, so that
one can give the standard measurable space structure and
a measure on $\Omega_{\infty}$ is given by $c_1(\overline{\mathscr H}_1)\wedge \cdots \wedge c_1(\overline{\mathscr H}_n)$.
The measurable space structure on $\Omega \setminus \Omega_{\infty}$ is discrete and a measure $\nu$ on $\Omega \setminus \Omega_{\infty}$ is given by $\nu(\{ \Gamma \}) = (\overline{\mathscr H}_1 \cdots \overline{\mathscr H}_n \cdot (\Gamma, 0))$.
This adelic structure is called the \emph{polarized adelic structure by the polarization
$(\mathscr B; \overline{\mathscr H}_1, \ldots, \overline{\mathscr H}_n)$}.
\begin{enonce}[remark]{Example}
Let $h$ be the metric of ${\mathscr{O}}_{{\mathbb{P}}^1_{{\mathbb{C}}}}(1)$ on ${\mathbb{P}}^1_{{\mathbb{C}}} = \operatorname{Proj}({\mathbb{C}}[T_0, T_1])$ given by
\[
|aT_0 + bT_1|_{h}(\zeta_0, \zeta_1) := \frac{|a\zeta_0 + b \zeta_1|}{\max \{ |\zeta_0|, |\zeta_1| \}}.
\]
Then $({\mathscr{O}}_{{\mathbb{P}}^1_{{\mathbb{Z}}}}(1), h)$ gives rise to a semipositive metrized invertible ${\mathscr{O}}_{{\mathbb{P}}^1_{{\mathbb{Z}}}}$-module, so
that
\[
\Big(({\mathbb{P}}^1_{{\mathbb{Z}}})^n; p_1^*({\mathscr{O}}_{{\mathbb{P}}^1_{{\mathbb{Z}}}}(1), h), \ldots, p_n^*({\mathscr{O}}_{{\mathbb{P}}^1_{{\mathbb{Z}}}}(1), h)\Big)
\]
yields to an adelic structure of the purely transcendental extension ${\mathbb{Q}}(x_1, \ldots, x_n)$
over ${\mathbb{Q}}$, where $p_i : ({\mathbb{P}}^1_{{\mathbb{Z}}})^n \to {\mathbb{P}}^1_{{\mathbb{Z}}}$ is the projection to the $i$-th factor. Note that it is nothing more than the adelic structure described in Section~\ref{subsec:adelic:structure:of:purely:transcendental:fibration} and
Section~\ref{sec:Arithmetic adelic structure}.
\end{enonce}
\chapter{Local intersection number and local height}
In this chapter, we fix a field $k$
equipped with an absolute value $|\raisebox{.4ex}{.}|$, such that $k$ is complete under the topology induced by the absolute value $|\raisebox{.4ex}{.}|$.
In the case where $|\raisebox{.4ex}{.}|$ is Archimedean, $k$ is equal to $\mathbb R$ or $\mathbb C$. In this case we always assume that $|\raisebox{.4ex}{.}|$ is the usual absolute value on $\mathbb R$ or $\mathbb C$. Note that the absolute value $|\raisebox{.4ex}{.}|$ extends in a unique way to any algebraic extension of $k$ (see \cite{Neukirch} Chapter II, Theorem 6.2). In particular, we fix an algebraic closure $k^{\operatorname{ac}}$, on which the absolute value $|\raisebox{.4ex}{.}|$ extends in a unique way.
Throughout this chapter, we denote the pair $(k, |\raisebox{.4ex}{.}|)$ by $v$. In the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean, we denote by $\mathfrak o_v$ the valuation ring of $v=(k,|\raisebox{.4ex}{.}|)$, and by $\mathfrak m_v$ the maximal ideal of $\mathfrak o_v$.
\section{Reminder on completion of an algebraic closure}We denote by $\mathbb C_k$ the completion of an algebraic closure $k^{\operatorname{ac}}$ of $k$, on which the absolute value $|\raisebox{.4ex}{.}|$ extends by continuity. Recall that $\mathbb C_k$ is algebraically closed. A proof for the case where $k=\mathbb Q_p$ can for example be found in \cite[(10.3.2)]{MR2392026}, by using Krasner's lemma. The positive characteristic case is quite similar, but a supplementary argument is needed to show that there is no inseparable algebraic extension of $\mathbb C_k$. For the convenience of the readers, we include the proof here (see also \cite[Theorem~17.1]{MR2444734} for another proof).
\begin{lemm}\label{Lem: completion is perfect}
Let $K$ be a field equipped with an absolute value $|\raisebox{.4ex}{.}|$ and $\widehat{K}$ be the completion of $K$. If the field $K$ is perfect, then also is $\widehat{K}$.
\end{lemm}
\begin{proof}
Clearly it suffices to treat the case where the characteristic of $K$ is $p>0$. To prove that the completed field $\widehat{K}$ is perfect, we need to show that any element $a$ of $\widehat{K}$ has a $p$-th root in $\widehat{K}$. We choose a sequence $(a_n)_{n\in\mathbb N}$ of elements of $K$ which converges to $a$. Since $K$ is supposed to be perfect, for each $n\in\mathbb N$ we can choose $b_n\in K$ such that $b_n^p=a_n$. For any $(n,m)\in\mathbb N^2$ one has
\[|b_n-b_m|^p=|(b_n-b_m)^p|=|b_n^p-b_m^p|=|a_n-a_m|.\]
Hence $(b_n)_{n\in\mathbb N}$ is a Cauchy sequence in $K$, which converges to an element $b\in\widehat{K}$. Therefore
\[b^p=\lim_{n\rightarrow+\infty}b_n^p=\lim_{n\rightarrow+\infty}a_n=a,\]
as required.
\end{proof}
\begin{prop}
The field $\mathbb C_k$ is algebraically closed.
\end{prop}
\begin{proof}
It suffices to treat the case where the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean.
We begin with proving that the field $\mathbb C_k$ is separably closed. Let $\mathbb C_k^{s}$ be a separable closure of $\mathbb C_k$, on which $|\raisebox{.4ex}{.}|$ extends in a unique way. Let $\alpha$ be a non-zero element of $\mathbb C_k^s$ and
\[f(T)=T^r+a_1T^{r-1}+\cdots+a_r\in\mathbb C_k[T]\]
be the minimal polynomial of $\alpha$. Assume that $r\geqslant 2$. Let $\alpha_2,\ldots,\alpha_r$ be conjugates of $\alpha$ in $\mathbb C_k^s$ which are different from $\alpha$, and let
\[\varepsilon=\min_{j\in\{2,\ldots,r\}}|\alpha-\alpha_j|.\] Since $k^{\operatorname{ac}}$ is dense in $\mathbb C_k$, there exists a polynomial
\[g(T)=T^r+b_1T^{r-1}+\cdots+b_r\in k^{\operatorname{ac}}[T]\]
such that
\[\max_{i\in\{1,\ldots,r\}}|\alpha|^{r-i}|b_i-a_i|<\varepsilon^r.\]
Since $k^{\operatorname{ac}}$ is algebraically closed, there exist elements $\beta_1,\ldots,\beta_r$ such that
\[g(T)=(T-\beta_1)\cdots(T-\beta_r).\]
One has
\[\prod_{i=1}^r|\alpha-\beta_i|=|g(\alpha)|=|g(\alpha)-f(\alpha)|\leqslant\max_{i\in\{1,\ldots,r\}}|\alpha|^{r-i}|b_i-a_i|<\varepsilon^n.\]
Hence there exists $\beta\in\{\beta_1,\ldots,\beta_r\}$ such that $|\alpha-\beta|<\varepsilon$. However, for any $\sigma\in\operatorname{Gal}(\mathbb C_k^s/\mathbb C_k)$, one has
\[|\alpha-\beta|=|\sigma(\alpha-\beta)|=|\sigma(\alpha)-\beta|.\]
This implies $|\alpha-\sigma(\alpha)|<\varepsilon$, which leads to a contradiction. Therefore one has $r=1$, or equivalently, $\alpha\in\mathbb C_k$.
To show that $\mathbb C_k$ is algebraic closed, it suffices to check that $\mathbb C_k$ does not admit any algebraic inseparable extension, or equivalently, $\mathbb C_k$ is a perfect field. Note that any algebraic closed field is perfect (see \cite[Chapitre V, \S1, no.5, Proposition 5]{MR643362}). Hence the result follows from Lemma \ref{Lem: completion is perfect}.
\end{proof}
\section{Continuous metrics}
If $X$ is a projective $k$-scheme, we denote by $X^{\mathrm{an}}$ the analytification of $X$. If $k=\mathbb C$ and $|\raisebox{.4ex}{.}|$ is the usual absolute value, then $X^{\mathrm{an}}$ is a complex analytic space; if $|\raisebox{.4ex}{.}|$ is non-Archimedean, then the analytification $X^{\mathrm{an}}$ is defined in the sense of Berkovich (see \cite[\S4.3]{MR1070709}). Recall that any element $x$ of $X^{\mathrm{an}}$ consists of a scheme point of $X$ and an absolute value $|\raisebox{.4ex}{.}|_x$ on the residue field of the scheme point, which extends the absolute value $|\raisebox{.4ex}{.}|$ on $k$. We denote by $\widehat{\kappa}(x)$ the completion of the residue field of the scheme point with respect to the absolute value $|\raisebox{.4ex}{.}|_x$, on which the absolute value extends by continuity. In the remaining of the section, we fix a projective $k$-scheme $X$.
\begin{defi}\label{Defi: continuous metric}
Let $L$ be an invertible $\mathcal O_X$-module. We call \emph{continuous metric} on $L$ any family $\varphi=(|\raisebox{.4ex}{.}|_{\varphi}(x))_{x\in X^{\mathrm{an}}}$, where for each $x\in X^{\mathrm{an}}$, $|\raisebox{.4ex}{.}|_\varphi(x)$ is a norm on $L(x):=L\otimes_{\mathcal O_X}\widehat{\kappa}(x)$, such that, for any section $s$ of $L$ on a Zariski open subset $U$ of $X$, the map $|s|_\varphi$ from $U^{\mathrm{an}}$ to $\mathbb R_{\geqslant 0}$ sending $(x\in U^{\mathrm{an}})$ to $|s(x)|_{\varphi}(x)$ is a continuous function on $U^{\mathrm{an}}$. If $\varphi$ and $\psi$ are continuous metrics on $L$, we define
\[d(\varphi,\psi):=\sup_{x\in X^{\mathrm{an}}}\bigg|\ln\frac{|\raisebox{.4ex}{.}|_{\varphi}(x)}{|\raisebox{.4ex}{.}|_\psi(x)}\bigg|,\]
where
\[\frac{|\raisebox{.4ex}{.}|_{\varphi}(x)}{|\raisebox{.4ex}{.}|_\psi(x)}:=\frac{|\ell|_{\varphi}(x)}{|\ell|_\psi(x)}
\quad\text{for any $\ell\in L(x)\setminus\{0\}$.}\]
\end{defi}
\begin{exem}\phantomsection\label{Exa: Fubini-Study}
\begin{enumerate}[label=\rm(\arabic*)]
\item\label{Item: quotient metrics} Let $L$ be an invertible $\mathcal O_X$-module and $n$ be a positive integer. Let $(E,\|\raisebox{.4ex}{.}\|)$ be a finite-dimensional normed vector space over $k$. We assume that $p:E\otimes_{k}\mathcal O_X\rightarrow L^{\otimes n}$ is a surjective homomorphism of $\mathcal O_X$-modules, which induces a $k$-morphism $f:X\rightarrow\mathbb P(E)$ such that $L^{\otimes n}$ is isomorphic to $f^*(\mathcal O_E(1))$, where $\mathcal O_E(1)$ denotes the universal invertible sheaf on the projective space $\mathbb P(E)$ (see \cite[II.(4.2.3)]{EGA}). For each point $x\in X^{\mathrm{an}}$ the norm $\mathopen{\|}{\raisebox{.4ex}{.}}\mathclose{\|}$ induces a quotient norm $|\raisebox{.4ex}{.}|(x)$ on $L(x)$ such that, for any $\ell\in L(x)\setminus\{0\}$,
\[|\ell|(x)=\inf_{\begin{subarray}{c}s\in E,\,\lambda\in\widehat{\kappa}(x)^{\times}\\
p(s)(x)=\lambda\ell^{\otimes n}
\end{subarray}}\big(|\lambda|_x^{-1}\mathopen{\|}s\mathclose{\|}\big)^{1/n}.\]
The quotient norms $(|\raisebox{.4ex}{.}|(x))_{x\in X^{\mathrm{an}}}$ define a continuous metric on $L$, called the \emph{quotient metric induced by $\mathopen{\|}{\raisebox{.4ex}{.}}\mathclose{\|}$}. By definition, if $\|\raisebox{.4ex}{.}\|_1$ and $\|\raisebox{.4ex}{.}\|_2$ are two norms on $E$, and if $\varphi_1$ and $\varphi_2$ are quotient metrics induced by $\|\raisebox{.4ex}{.}\|_1$ and $\|\raisebox{.4ex}{.}\|_2$, respectively, then one has
\begin{equation}\label{Equ: distance of quotient metrics}d(\varphi_1,\varphi_2)\leqslant d(\|\raisebox{.4ex}{.}\|_1,\|\raisebox{.4ex}{.}\|_2):=\sup_{s\in E\setminus\{0\}}\Big|\ln\|s\|_1-\ln\|s\|_2\Big|.\end{equation}
\item Let $L$ be an invertible $\mathcal O_X$-module and $\varphi=(|\raisebox{.4ex}{.}|_{\varphi}(x))_{x\in X^{\mathrm{an}}}$ be a continuous metric on $L$. The dual norms of $|\raisebox{.4ex}{.}|_{\varphi}(x)$ on $L(x)^\vee$ form a continuous metric on $L^\vee$, which we denote by $\varphi^{\vee}$. Recall that for any $\ell\in L(x)\setminus\{0\}$, one has
\[|\ell^\vee|_{\varphi^\vee}=|\ell|_{\varphi}^{-1},\]
where $\ell^\vee$ denotes the linear form on $L(x)$ such that $\ell^\vee(\lambda\ell)=\lambda$ for any $\lambda\in\widehat{\kappa}(x)$.
\item Let $L_1$ and $L_2$ be invertible $\mathcal O_X$-modules, and $\varphi_1$ and $\varphi_2$ be continuous metrics on $L_1$ and $L_2$ respectively. Then the tensor product norms of $|\raisebox{.4ex}{.}|_{\varphi_1}(x)$ and $|\raisebox{.4ex}{.}|_{\varphi_2}(x)$ form a continuous metric on $L_1\otimes L_2$, which we denote by $\varphi_1\otimes\varphi_2$. Note that, for any $\ell_1\in L_1(x)$ and $\ell_2\in L_2(x)$, one has
\[|\ell_1\otimes\ell_2|_{\varphi_1\otimes\varphi_2}(x)=|\ell_1|_{\varphi_1}(x)\cdot|\ell_2|_{\varphi_2}(x).\]
\item Let $f:Y\rightarrow X$ be a $k$-morphism of projective $k$-schemes. Let $L$ be an invertible $\mathcal O_X$-module, equipped with a continuous metric $\varphi$. Then the metric $\varphi$ induces by pull-back a continuous metric $f^*(\varphi)$ on $f^*(L)$ such that, for any $y\in Y^{\operatorname{an}}$ and any $\ell\in L(f^{\operatorname{an}}(y))$, one has
\[|f^*(\ell)|_{f^*(\varphi)}(y)=|\ell|_{\varphi}(f^{\operatorname{an}}(y)).\]
The metric $f^*(\varphi)$ is called the \emph{pull-back} of $\varphi$ by $f$.
\item\label{Item: extension of scalars metric} Let $k'/k$ be an extension of fields. We assume that the absolute value $|\raisebox{.4ex}{.}|$ extends to $k'$ and that the field $k'$ is complete with respect to the topology induced by the extended absolute value. Let $X_{k'}$ be the fiber product $X\times_{\operatorname{Spec} k}\operatorname{Spec} k'$. We denote by $\pi:X_{k'}\rightarrow X$ the morphism of projection. Then the map \begin{equation}\label{Equ: pi natural}\pi^{\natural}:X_{k'}^{\operatorname{an}}\longrightarrow X^{\mathrm{an}},\end{equation}
sending any point $x'=(j(x'),|\raisebox{.4ex}{.}|_{x'})\in X_{k'}^{\mathrm{an}}$ to the pair consisting of the scheme point $\pi(j(x'))$ of $X$ and the restriction of $|\raisebox{.4ex}{.}|_{x'}$ to the residue field of $\pi(j(x'))$, is continuous (see \cite[Proposition 2.1.17]{CMArakelovAdelic}), where $j:X_{k'}^{\mathrm{an}}\rightarrow X_{k'}$ denotes the map sending a point in the analytic space to its underlying scheme point.
Let $L$ be an invertible $\mathcal O_X$-module, equipped with a continuous metric $\varphi$. Let $L_{k'}$ be the pull-back of $L$ by the morphism of projection $\pi$. The continuous metric $\varphi$ induces a continuous metric $\varphi_{k'}$ on $L_{k'}$ such that, for any $x'\in X_{k'}^{\operatorname{an}}$ and any $\ell\in L(\pi^{\natural}(x'))$, one has
\[\forall\,a\in\widehat{\kappa}(x'),\quad |a\otimes\ell|_{\varphi_{k'}}(x')=|a|_{x'}\cdot|\ell|_{\varphi}(\pi^{\natural}(x')). \]
In particular, if $\psi$ is another continuous metric on $L$, then one has
\begin{equation}\label{Equ: distance extension of scalars}
d(\varphi_{k'},\psi_{k'})\leqslant d(\varphi,\psi).
\end{equation}
\end{enumerate}
\end{exem}
\begin{defi}\label{Def: extension of sclars}
Let $(E,\|\raisebox{.4ex}{.}\|)$ be a finite-dimensional normed vector space over $k$. We assume that the norm $\|\raisebox{.4ex}{.}\|$ is either ultrametric or induced by an inner product. Let $k'/k$ be an extension of fields, on which the absolute value $|\raisebox{.4ex}{.}|$ extends. We assume that field $k'$ is complete with respect to the extended absolute value. We denote by $\|\raisebox{.4ex}{.}\|_{k'}$ the following norm on $E_{k'}:=E\otimes_kk'$.
\begin{enumerate}[label=\rm(\arabic*)]
\item In the case where the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean and the norm $\|\raisebox{.4ex}{.}\|$ is ultrametric, $\|\raisebox{.4ex}{.}\|_{k'}$ is the $\varepsilon$-extension of scalars of the norm $\|\raisebox{.4ex}{.}\|$. Namely, for any $t=s_1\otimes \lambda_1+\cdots+s_m\otimes \lambda_m\in E\otimes_kk'$
\[ \|t\|_{k'}:=\sup_{f\in E^\vee\setminus\{0\}}\frac{|\lambda_1f(s_1)+\cdots+\lambda_mf(s_m)|}{\|f\|_*},\]
where $\|\raisebox{.4ex}{.}\|_*$ denotes the dual norm of $\|\raisebox{.4ex}{.}\|$, which is defined as
\[\|f\|_*:=\sup_{x\in E\setminus\{0\}}\frac{|f(s)|}{\|s\|}.\]
This is an ultrametric norm on $E_{k'}$ such that $\|s\otimes a\|_{k'}=\|s\|\cdot|a|$ (see \cite[Proposition 1.3.1]{CMArakelovAdelic}). Moreover, if $(e_i)_{i=1}^r$ is an orthonormal basis of $(E,\|\raisebox{.4ex}{.}\|)$, then $(e_i\otimes 1)_{i=1}^r$ is an orthonormal basis of $(E_{k'},\|\raisebox{.4ex}{.}\|_{k'})$ (see \cite[Proposition 1.3.13]{CMArakelovAdelic}).
\item In the case where the absolute value $|\raisebox{.4ex}{.}|$ is Archimedean, $k=\mathbb R$, $k'=\mathbb C$, and $\|\raisebox{.4ex}{.}\|$ is induced by an inner product $\langle\kern.15em,\kern-.02em\rangle$, $\|\raisebox{.4ex}{.}\|_{\mathbb C}$ is the orthogonal extension of scalars of $\|\raisebox{.4ex}{.}\|$. Namely, for any $(s,t)\in E\times E$,
\[\|s\otimes 1+t\otimes\sqrt{-1}\|_{\mathbb C}:=(\|s\|^2+\|t\|^2)^{1/2}.\]
Clearly, for any $s\in E$ one has $\|s\otimes 1\|_{\mathbb C}=\|s\|$.
Note that the norm $\|\raisebox{.4ex}{.}\|_{\mathbb C}$ is induced by an inner product $\langle\kern.15em,\kern-.02em\rangle_{\mathbb C}$ on $E_{\mathbb C}$ such that, for any $u=s\otimes 1+t\otimes\sqrt{-1}$ and $u'=s'\otimes 1+t'\otimes\sqrt{-1}$ in $E_{\mathbb C}$,
\[\mathopen{\langle}u,u'\mathclose{\rangle}=\langle s,s'\rangle+\langle t,t'\rangle+\sqrt{-1}(\langle s,t'\rangle-\langle t,s'\rangle).\]
Moreover, if $(e_i)_{i=1}^r$ is an orthonormal basis of $(E,\|\raisebox{.4ex}{.}\|)$, then $(e_i\otimes 1)_{i=1}^r$ is an orthonormal basis of $(E_{\mathbb C},\|\raisebox{.4ex}{.}\|_{\mathbb C})$.
\end{enumerate}
\end{defi}
\begin{rema}
\label{Rem: extension of normss}
Let $n$ be a positive integer. Assume that $p:E\otimes_{k}\mathcal O_X\rightarrow L^{\otimes n}$ is a surjective homomorphism of $\mathcal O_X$-modules, which induces a $k$-morphism $f:X\rightarrow\mathbb P(E)$ such that $L^{\otimes n}=f^*(\mathcal O_E(1))$. We equip $L$ with the quotient metric $\varphi$ induced by $\|\raisebox{.4ex}{.}\|$. In the case where the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean, for any point $x\in X^{\mathrm{an}}$, the norm $|\raisebox{.4ex}{.}|_{\varphi^{\otimes n}}(x)$ on $L^{\otimes n}(x)$ coincides with the quotient norm on $L^{\otimes n}(x)$ induced by the norm $\|\raisebox{.4ex}{.}\|_{\widehat{\kappa}(x)}$ on $E\otimes_{k}{\widehat{\kappa}(x)}$ and the quotient map $p_x:E\otimes_{k}{\widehat{\kappa}(x)}\rightarrow L^{\otimes n}$. We refer the readers to \cite[Proposition 1.3.26 (i)]{CMArakelovAdelic} for a proof. As for the Archimedean case with $k=\mathbb R$ and $\widehat{\kappa}(x)=\mathbb C$, note that, if $s$ and $t$ are elements of $E$ and $a$ and $b$ are complex numbers such that
\[p_x(s)=a\ell^{\otimes n},\quad p_x(t)=b\ell^{\otimes n},\]
where $\ell$ is a fixed non-zero element of $L(x)$. Then one has
\[p_x(s\otimes 1+t\otimes\sqrt{-1})=(a+b\sqrt{-1})\ell^{\otimes n}\]
and hence
\[\frac{(\|s\|^2+\|t\|^2)^{\frac 12}}{|a+b\sqrt{-1}|}\geqslant\frac{(\|s\|^2+\|t\|^2)^{\frac 12}}{|a|+|b|}\geqslant \frac{1}{\sqrt{2}}\frac{\|s\|+\|t\|}{|a|+|b|}\geqslant \frac{1}{\sqrt{2}}|\ell|_{\varphi^{\otimes n}}(x).\]
Therefore, the quotient norm on $L^{\otimes n}$ induced by $\|\raisebox{.4ex}{.}\|_{\widehat{\kappa}(x)}$ and the quotient map \[p_x:E\otimes_k\widehat{\kappa}(x)\longrightarrow L^{\otimes n}(x),\]
which is bounded from above by $|\raisebox{.4ex}{.}|_{\varphi^{\otimes n}}(x)$ by definition, is actually bounded from below by $\frac{1}{\sqrt{2}}|\raisebox{.4ex}{.}|_{\varphi^{\otimes n}}(x)$.
Let $k'/k$ be a valued extension of $(k,|\raisebox{.4ex}{.}|)$ which is complete. By extension of scalars, we obtain a surjective homomorphism of $\mathcal O_{X_{k'}}$-modules \[p_{k'}:E_{k'}\otimes_{k'}\mathcal O_{X_{k'}}\longrightarrow L_{k'}^{\otimes n},\] which corresponds to the $k'$-morphism $f_{k'}:X_{k'}\rightarrow\mathbb P(E_{k'})$. Let $\varphi$ be the quotient metric on $L$ induced by $\|\raisebox{.4ex}{.}\|$. In the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean, it turns out that the quotient metric on $L_{k'}$ induced by $\|\raisebox{.4ex}{.}\|_{k'}$ coincides with $\varphi_{k'}$. This fact follows from \cite[Proposition 1.3.15 (i)]{CMArakelovAdelic} and the above identification of the quotient metric to a family of quotient norms. In the Archimedean case with $k=\mathbb R$ and $k'=\mathbb C$, by the above estimate, in general the quotient metric $\varphi'$ on $L_{\mathbb C}$ induced by $\|\raisebox{.4ex}{.}\|_{\mathbb C}$ is different from $\varphi_{\mathbb C}$. The above estimate actually shows that, for any $x\in X_{\mathbb C}^{\operatorname{an}}$ one has
\[2^{-\frac{1}{2n}}|\raisebox{.4ex}{.}|_{\varphi_{\mathbb C}}(x)\leqslant |\raisebox{.4ex}{.}|_{\varphi'}(x)\leqslant|\raisebox{.4ex}{.}|_{\varphi_{\mathbb C}}(x).\]
Note that the metric $\varphi_{\mathbb C}$ is still a quotient metric. In fact, if we consider the $\pi$-extension of scalars $\|\raisebox{.4ex}{.}\|_{\mathbb C,\pi}$ on $E_{\mathbb C}$ defined as
\[\forall\,t\in E_{\mathbb C},\quad\|t\|_{\mathbb C,\pi}:=\inf_{t=s_1\otimes\lambda_1+\cdots+s_m\otimes\lambda_m}\sum_{i=1}^m|\lambda_i|\cdot\|s_i\|.\]
Then the metric $\varphi_{\mathbb C}$ identifies with the quotient metric induced by $\|\raisebox{.4ex}{.}\|_{\mathbb C,\pi}$.
\end{rema}
\begin{defi}\label{Def: Fubini-Study metric}
Let $L$ be an invertible $\mathcal O_X$-module and $n$ be a positive integer.
Let $(E,\|\raisebox{.4ex}{.}\|)$ be a finite-dimensional normed vector space over $k$. We assume that the norm $\|\raisebox{.4ex}{.}\|$ is either ultrametric or induced by an inner product. Let $p:E\otimes_k\mathcal O_X\rightarrow L^{\otimes n}$ be a surjective homomorphism of $\mathcal O_X$-module, which induces a $k$-morphism $f:X\rightarrow\mathbb P(E)$ such that $L^{\otimes n}$ is isomorphic to $f^*(\mathcal O_E(1))$. For each point $x\in X^{\mathrm{an}}$, the norm $\|\raisebox{.4ex}{.}\|_{\widehat{\kappa}(x)}$ on $E\otimes_k\widehat{\kappa}(x)$ induces by quotient a norm $|\raisebox{.4ex}{.}|(x)$ on $L^{\otimes n}(x)$. There then exists a unique continuous metric $\varphi$ on $L$ such that $|\raisebox{.4ex}{.}|_{\varphi^{\otimes n}}(x)=|\raisebox{.4ex}{.}|(x)$ for any $x\in X^{\mathrm{an}}$. The metric $\varphi$ is called the \emph{orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|$}.
Note that, in the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean or $(k,|\raisebox{.4ex}{.}|)$ is $\mathbb C$ equipped with the usual absolute value, the orthogonal quotient metric identifies with the quotient metric induced by $\|\raisebox{.4ex}{.}\|$ introduced in Example \ref{Exa: Fubini-Study} \ref{Item: quotient metrics}. Moreover, for any complete valued extension $k'/k$, the metric $\varphi_{k'}$ identifies with the orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|_{k'}$ (see Remark \ref{Rem: extension of normss} above).
\end{defi}
\begin{defi}
Let $L$ be a semi-ample invertible $\mathcal O_X$-module and $\varphi$ be a continuous metric on $L$. If there exists a sequence of quotient metrics $\varphi_n$ on $L$ such that \[\lim_{n\rightarrow+\infty}d(\varphi_n,\varphi)=0,\] we say that the metric $\varphi$ is \emph{semi-positive} (see \cite[\S2.2]{MR3783789}). In the case where $|\raisebox{.4ex}{.}|$ is Archimedean and $k=\mathbb C$, this definition is equivalent to the plurisubharmonicity of the metric $\varphi$ (see for example \cite[Theorem 3.5]{MR1254133}).
\end{defi}
\begin{rema}\label{Rem: positivity extension scalar}
Let $L$ be an invertible $\mathcal O_X$-module. Let $k'/k$ be a complete valued extension of $k$, $X_{k'}$ be the fiber product $X\times_{\operatorname{Spec} k}\operatorname{Spec} k'$ and $\pi^{\natural}:X_{k'}^{\mathrm{an}}\rightarrow X^{\mathrm{an}}$ be the map defined in \eqref{Equ: pi natural}. If $\varphi$ and $\psi$ are two continuous metrics on $L$, then the metrics $\varphi_{k'}$ and $\psi_{k'}$ satisfy the relation (see \eqref{Equ: distance extension of scalars})
\[d(\varphi_{k'},\psi_{k'})\leqslant d(\varphi,\psi).\]
Therefore, if $\varphi$ is a semi-positive metric on $L$, then $\varphi_{k'}$ is also a semi-positive metric.
\end{rema}
\begin{defi}\label{Def: Fubini-Study}
Let $L$ be a very ample invertible $\mathcal O_X$-module and $\varphi$ be a continuous metric on $L$. For any positive integer $m$, the continuous metric $\varphi$ induces a seminorm $\|\raisebox{.4ex}{.}\|_{\varphi^{\otimes m}}$ on $H^0(X,L^{\otimes m})$ as follows:
\[\forall\,s\in H^0(X,L^{\otimes m}),\quad \|s\|_{\varphi^{\otimes m}}=\sup_{x\in X^{\operatorname{an}}}|s|_{\varphi^{\otimes m}}(x).\]
This seminorm is a norm notably when the scheme $X$ is reduced.
For each point $x\in X^{\mathrm{an}}$, the seminorm $\|\raisebox{.4ex}{.}\|_{\varphi^{\otimes m}}$ induces a quotient seminorm $|\raisebox{.4ex}{.}|_{\varphi^{(m)}}(x)$ on $L(x)$ such that, for any $\ell\in L(x)\setminus\{0\}$
\[|\ell|_{\varphi^{(m)}}(x)=\inf_{\begin{subarray}{c}
s\in H^0(X,L^{\otimes m}),\,\lambda\in\widehat{\kappa}(x)^{\times}\\
s(x)=\lambda\ell^{\otimes m}
\end{subarray}}(|\lambda|^{-1}\|s\|_{\varphi^{\otimes m}})^{1/m}.\]
This seminorm is actually a norm and is bounded from below by $|\raisebox{.4ex}{.}|_\varphi(x)$. The norms $(|\raisebox{.4ex}{.}|_{\varphi^{(m)}}(x))_{x\in X^{\operatorname{an}}}$ form a continuous metric on $L$, which we denote by $\varphi^{(m)}$.
\end{defi}
\begin{prop}\label{Pro: distance varphi 1mvarphi2m}
Let $L$ be a very ample invertible $\mathcal O_X$-module. If $\varphi_1$ and $\varphi_2$ are two continuous metrics on $L$, then the following inequalities hold:
\[\forall\,m\in\mathbb N_{\geqslant 1},\quad d(\varphi_1^{(m)},\varphi_2^{(m)})\leqslant d(\varphi_1,\varphi_2).\]
\end{prop}
\begin{proof}
By definition, one has
\[\sup_{\begin{subarray}{c}
s\in H^0(X,L^{\otimes m})\\
\|s\|_{\varphi_1^{\otimes m}}\neq 0
\end{subarray}}\bigg|\ln\frac{\|s\|_{\varphi_1^{\otimes m}}}{\|s\|_{\varphi_2^{\otimes m}}}\bigg|\leqslant d(\varphi_1^{\otimes m},\varphi_2^{\otimes m})=m\,d(\varphi_1,\varphi_2).\]
Therefore,
\[d(\varphi_1^{(m)},\varphi_2^{(m)})\leqslant \frac 1md(\varphi_1^{\otimes m},\varphi_2^{\otimes m})\leqslant d(\varphi_1,\varphi_2).\]
\end{proof}
\begin{rema}\label{Rem: Fubini-Study of quotient metric}Let $(E,\|\raisebox{.4ex}{.}\|)$ be a finite-dimensional vector space over $k$, $m$ be a positive integer and $p:\pi^*(E)\rightarrow L^{\otimes m}$ be a surjective homomorphism of $\mathcal O_X$-modules, where $\pi:X\rightarrow\operatorname{Spec} k$ denotes the structural morphism of schemes. Let $\varphi$ be the quotient metric induced by $\|\raisebox{.4ex}{.}\|$. Note that $p$ induces by adjunction between $\pi^*$ and $\pi_*$ a $k$-linear map $\alpha:E\rightarrow H^0(X,L^{\otimes m})$. Let $s$ be an element of $H^0(X,L^{\otimes m})$. For any $x\in X^{\mathrm{an}}$, one has
\[|s|_{\varphi^{\otimes m}}(x)=\inf_{\begin{subarray}{c}t\in E,\,\lambda\in\widehat{\kappa}(x)^{\times}\\
\alpha(t)(x)=\lambda s(x)
\end{subarray}}\frac{\|t\|}{|\lambda|_x}.\]
In particular, for any $s$ in the image of the linear map $\alpha$, one has
\[\|s\|_{\varphi^{\otimes m}}\leqslant\inf_{t\in E,\,\alpha(t)=s}\|t\|.\]
Therefore, for $x\in X^{\mathrm{an}}$ and $\ell\in L(x)\setminus\{0\}$, one has
\[|\ell|_{\varphi^{(m)}}(x)=\inf_{\begin{subarray}{c}s\in H^0(X,L^{\otimes m}),\,\lambda\in\widehat{\kappa}(x)^{\times}\\
s(x)=\lambda\ell^{\otimes m}
\end{subarray}}\Big(\frac{\|s\|_{\varphi^{\otimes m}}}{|\lambda|_x}\Big)^{1/m}\leqslant\inf_{\begin{subarray}{c}t\in E,\,\lambda\in\widehat{\kappa}(x)^{\times}\\
\alpha(t)(x)=b\ell^{\otimes m}
\end{subarray}}\Big(\frac{\|t\|}{|\lambda|_x}\Big)^{1/m}=|\ell|_{\varphi}(x).\]
Combining with the inequality $|\ell|_{\varphi^{(m)}}(x)\geqslant|\ell|_{\varphi}(x)$, we obtain the equality $\varphi^{(m)}=\varphi$.
\end{rema}
\begin{prop}\label{Pro: varphi 1 as a continuous metric}
Let $L$ be a very ample invertible $\mathcal O_X$-module, equipped with a continuous metric $\varphi$. Let $\|\raisebox{.4ex}{.}\|$ be a norm on the vector space $H^0(X,L^{\otimes n})$. For any $a>0$, let $\|\raisebox{.4ex}{.}\|_{a}$ be the norm on $H^0(X,L^{\otimes n})$ defined by
\[\forall\,s\in H^0(X,L^{\otimes n}),\quad\|s\|_{a}=\max\{\|s\|_{\varphi},a\|s\|\}=\max\Big\{\sup_{x\in X^{\operatorname{an}}}|s|_{\varphi}(x),a\|s\|\Big\},\]
and let $\varphi_{a}$ be the quotient metric on $L$ induced by $\|\raisebox{.4ex}{.}\|_a$. Then, for any $x\in X^{\mathrm{an}}$
\begin{equation}\label{Equ: metric varphi1 bounded by varphi eps}|\raisebox{.4ex}{.}|_{\varphi^{(1)}}(x)\leqslant|\raisebox{.4ex}{.}|_{\varphi_{a}}(x),\end{equation}
and there exists $a_0>0$ such that $\varphi_{a}=\varphi^{(1)}$ when $0<a\leqslant a_0$.
\end{prop}
\begin{proof}
By definition, one has $\|\raisebox{.4ex}{.}\|_{a}\geqslant\|\raisebox{.4ex}{.}\|_{\varphi}$. Hence the inequality \eqref{Equ: metric varphi1 bounded by varphi eps} holds.
Let $N_{\|\raisebox{.4ex}{.}\|_{\varphi}}$ be the null space of the norm $\|\raisebox{.4ex}{.}\|_{\varphi}$, which is defined as
\[N_{\|\raisebox{.4ex}{.}\|_{\varphi}}=\{s\in H^0(X,L)\,|\,\|s\|_{\varphi}=0\}.\] Let $E$ be the quotient vector space $H^0(X,L)/N_{\|\raisebox{.4ex}{.}\|_{\varphi}}$ and $\pi:H^0(X,L)\rightarrow E$ be the projection map. We denote by $\|\raisebox{.4ex}{.}\|_E$ the quotient norm of $\|\raisebox{.4ex}{.}\|$ on $E$ and $\|\raisebox{.4ex}{.}\|_{\varphi,E}$ be the quotient seminorm of $\|\raisebox{.4ex}{.}\|_{\varphi}$ on $E$, which is actually a norm satisfying the relation
\begin{equation}\label{Equ: quotient seminorm}\forall\,s\in H^0(X,L),\quad \|\pi(s)\|_{\varphi,E}=\|s\|_{\varphi}.\end{equation}
Since all norms on $E$ are equivalent, there exists $C>0$ such that $\|\raisebox{.4ex}{.}\|_{E}\leqslant C\|\raisebox{.4ex}{.}\|_{\varphi,E}$. Therefore, for any $x\in X^{\mathrm{an}}$, and any $\ell\in L(x)\setminus\{0\}$ one has
\[\begin{split}|\ell|_{\varphi_a}(x)&=\inf_{\begin{subarray}{c}
s\in H^0(X,L^{\otimes n}),\,\lambda\in\widehat{\kappa}(x)^{\times}\\
s(x)=\lambda\ell^{\otimes n}
\end{subarray}}\Big(\frac{\max\{\|s\|_{\varphi},a\|s\|\}}{|\lambda|}\Big)^{\frac 1n}\\
&=\inf_{\begin{subarray}{c}
s\in H^0(X,L^{\otimes n}),\,\lambda\in\widehat{\kappa}(x)^{\times}\\
s(x)=\lambda\ell^{\otimes n}
\end{subarray}}\Big(\frac{\max\{\|\pi(s)\|_{\varphi,E},a\|\pi(s)\|_E\}}{|\lambda|}\Big)=|\ell|_{\varphi^{(1)}}(x)
\end{split}\]
once $a<C^{-1}$, where the second equality comes from the fact that $s(x)=0$ when $s\in N_{\|\raisebox{.4ex}{.}\|_{\varphi}}$.
\end{proof}
\begin{prop}\label{Pro: convergence of varphin}
Let $L$ be a very ample invertible $\mathcal O_X$-module, equipped with a semi-positive continuous metric $\varphi$. Then one has
\[\lim_{m\rightarrow+\infty}d(\varphi^{(m)},\varphi)=0.\]
\end{prop}
\begin{proof}First of all, for positive integers $m$ and $m'$, one has
\[\forall\,x\in X^{\mathrm{an}},\;\forall\,\ell\in L(x)\setminus\{0\},\quad |\ell|_{\varphi^{(m+m')}}^{m+m'}(x)\leqslant |\ell|_{\varphi^{(m)}}^m\cdot|\ell|_{\varphi^{(m')}}^{m'}. \]
Therefore \[(m+m')d(\varphi^{(m+m')},\varphi)\leqslant md(\varphi^{(m)},\varphi)+m'd(\varphi^{(m')}).\]
By Fekete's lemma we obtain that the sequence
\[d(\varphi^{(m)},\varphi),\quad m\in\mathbb N,\;m\geqslant 1\]
converges to a non-negative real number, which is also equal to
\[\inf_{m\in\mathbb N,\;m\geqslant 1}d(\varphi^{(m)},\varphi).\]
Moreover,
since the metric $\varphi$ is semi-positive, there exist a sequence of positive integers $(m_n)_{n\in\mathbb N}$, a sequence of finite-dimensional normed vector spaces $((E_n,\|\raisebox{.4ex}{.}\|_n))_{n\in\mathbb N}$ and surjective homomorphisms of $\mathcal O_X$-modules $p_n:E_n\otimes_k\mathcal O_X\rightarrow L^{\otimes m_n}$ such that, if we denote by $\varphi_n$ the quotient metric on $L$ induced by $\|\raisebox{.4ex}{.}\|_n$, then one has
\[\lim_{n\rightarrow+\infty}d(\varphi_n,\varphi)=0.\]
By Remark \ref{Rem: Fubini-Study of quotient metric}, one has $\varphi_n^{(m_n)}=\varphi_n$ and hence
\[d(\varphi^{(m_n)},\varphi)\leqslant d(\varphi^{(m_n)},\varphi_n)+d(\varphi_n,\varphi)=d(\varphi^{(m_n)},\varphi_n^{(m_n)})+d(\varphi_n,\varphi)\leqslant 2d(\varphi_n,\varphi),\]
where the last inequality comes from Proposition \ref{Pro: distance varphi 1mvarphi2m}. By taking the limite when $n\rightarrow +\infty$, we obtain that
\[\inf_{m\in\mathbb N,\,m\geqslant 1}d(\varphi^{(m)},\varphi)=0.\]
\end{proof}
\begin{defi}
Let $(L,\varphi)$ be a metrized invertible $\mathcal O_X$-module. We say that $(L,\varphi)$ is \emph{integrable} if there exist ample invertible $\mathcal O_X$-modules $L_1$ and $L_2$ equipped with semi-positive metrics $\varphi_1$ and $\varphi_2$ respectively, such that $L=L_1\otimes L_2^\vee$ and $\varphi=\varphi_1\otimes\varphi_2^\vee$.
\end{defi}
\begin{defi}
We assume that $v$ is non-Archimedean.
Let $(L,\varphi)$ be a metrized invertible $\mathcal O_X$-module.
We say $\varphi$ is a \emph{model metric} if there are a positive integer $n$ and
a model $(\mathscr X, \mathscr L)$ of $(X, L^{\otimes n})$
such that $\varphi^{\otimes n}$ coincides with the metric arising from
the model $(\mathscr X, \mathscr L)$ (cf. \cite[Subsection~2.3.2]{CMArakelovAdelic}).
In the above definition, we may assume that $\mathscr X$ is flat over $\mathfrak o_v$ (for details, see \cite[Subsection~2.3.2]{CMArakelovAdelic}).
In the case where $L$ is nef, if $\mathscr L$ is nef along the special fiber of
$\mathscr X \to \operatorname{Spec}(\mathfrak o_v)$, then
the model $(\mathscr X, \mathscr L)$ is said to be \emph{nef} and
$\varphi$ is called a \emph{nef model metric}.
\end{defi}
\begin{rema}\label{rem:model:nef:metric}
Let $(\mathscr X, \mathscr L)$ be a model of $(X, L)$, $\mathscr X_{\mathrm{red}}$ be the reduced scheme associated with $\mathscr X$ and $\mathscr L_{\mathrm{red}} := \rest{\mathscr L}{X_{\mathrm{red}}}$.
For $x \in X^{\operatorname{an}}$, the morphism $\operatorname{Spec} (\mathfrak o_x) \to \mathscr X$ factors through
$\operatorname{Spec} (\mathfrak o_x) \to \mathscr X_{\mathrm{red}} \to \mathscr X$, and hence
$\varphi_{\mathscr L}$ coincides with $\varphi_{\mathscr L_{\mathrm{red}}}$.
Moreover, $\mathscr L$ is nef with respect to $\mathscr{X} \to \operatorname{Spec}(\mathfrak o_v)$ if and only if
$\mathscr L_{\mathrm{red}}$ is nef with respect to $\mathscr{X}_{\mathrm{red}} \to \operatorname{Spec}(\mathfrak o_v)$.
\end{rema}
\begin{defi}
Let $(L,\varphi)$ be a metrized invertible $\mathcal O_X$-module.
We say $(L,\varphi)$ is \emph{smooth} if one of the following conditions is satisfied:
\begin{enumerate}[label=\rm(\roman*)]
\item if $v$ is Archimedean, $\varphi$ is a $C^{\infty}$-metric;
\item if $v$ is non-Archimedean, $\varphi$ is a model metric.
\end{enumerate}
If $L$ is nef
and $v$ is non-Archimedean,
then $\varphi$ is said to be \emph{$M$-semi-positive}
if there is a sequence $\{ \varphi_m\}_{m=1}^{\infty}$ of nef model metrics of $L$
such that $\lim\limits_{m\to\infty} d(\varphi, \varphi_m) = 0$.
\end{defi}
\begin{lemm}\label{lemma:finitely:presented:model}
We assume that $v$ is non-Archimedean.
Let $L$ be an invertible ${\mathscr{O}}_X$-module and $(\mathscr X, \mathscr L)$ be a model of $(X, L)$.
Then there is a model $(\mathscr X', \mathscr L')$ of $(X, L)$ with the following properties:
\begin{enumerate}[label=\rm(\arabic*)]
\item $\mathscr X' \to \operatorname{Spec}(\mathfrak o_v)$ is finitely presented, that is, $(\mathscr X', \mathscr L')$ is a coherent model of $(X,L)$ (cf. \cite[Subsection~2.3.2]{CMArakelovAdelic}).
\item $\mathscr X$ is a closed subscheme of $\mathscr X'$.
\item The special fiber of $\mathscr X' \to \operatorname{Spec}(\mathfrak o_v)$ coincides with
the special fiber of $\mathscr X \to \operatorname{Spec}(\mathfrak o_v)$.
\item $\rest{\mathscr L'}{\mathscr X} = \mathscr L$.
\end{enumerate}
\end{lemm}
\begin{proof}
By \cite[Corollary~5.16 in Chapter~II]{Hart77}, there are a polynomial ring $A := \mathfrak o_v[T_0, \ldots, T_N]$ over $\mathfrak o_v$ and
a homogeneous ideal $I$ of $A$ such that $\mathscr X = \operatorname{Proj} (A/I)$. We set $R := A/I$. Let $p : A \to R$ and $\pi : A \to A \otimes_{\mathfrak o_v} \mathfrak o_v/\mathfrak m_v = (\mathfrak o_v/\mathfrak m_v)[T_0, \ldots, T_N]$ be the natural homomorphisms.
There are homogeneous elements $h_1, \ldots, h_e$ of $R$ and $g_{ij} \in R_{(h_ih_j)}$ ($i, j \in \{ 1, \ldots, e \}$) such that $\mathscr X = \bigcup_{i=1}^e D_+(h_i)$ and
$\{ g_{ij} \}_{i, j \in \{ 1, \ldots, e \}}$ gives transition functions of $\mathscr L$,
where $R_{(h)}$ (for a homogenous element $h$) is the homogeneous localization with respect to $h$.
We choose a homogeneous element $H_i$ of $A$ such that $p(H_i) = h_i$.
Since \[\emptyset = \bigcap_{i=1}^e V_+(h_i) = V_+(h_1 R + \cdots + h_e R),\] we have
$R_+ \subseteq \operatorname{rad}(h_1 R + \cdots + h_e R)$
by \cite[Lemma~3.35 in Section~2.3]{MR1917232}, that is,
there is a positive integer $a$ such that $p(T_0)^a, \ldots, p(T_N)^a \in h_1 R + \cdots + h_e R$, so that
\begin{equation}\label{eqn:thm:equiv:semi-positive:M:semi-positive:01}
T_0^a, \ldots, T_N^a \in H_1 A + \cdots + H_e A + I.
\end{equation}
We also choose $G_{ij} \in A_{(H_iH_j)}$ such that $p(G_{ij}) = g_{ij}$ and $G_{ii} = 1$. As $g_{ij}g_{jl} = g_{il}$ on $R_{(h_ih_jh_l)}$, one can see
\begin{equation}\label{eqn:thm:equiv:semi-positive:M:semi-positive:02}
G_{ij}G_{jl} - G_{il} \in I_{(H_iH_jH_l)}
\end{equation}
for all $i,j, l\in \{ 1, \ldots, e \}$. Let $S = \mathfrak o_v \setminus \{ 0 \}$.
Since $I_S$ and $\pi(I)$ are homogeneous ideals of $k[T_0, \ldots, T_N]$ and $(\mathfrak o_v/\mathfrak m_v)[T_0, \ldots, T_N]$, respectively, $I_S$ and $\pi(I)$ are finitely generated ideals.
Therefore, by using \eqref{eqn:thm:equiv:semi-positive:M:semi-positive:01} and \eqref{eqn:thm:equiv:semi-positive:M:semi-positive:02}, one can find a finitely generated
homogeneous ideal $I'$ of $A$ such that
\[
\begin{cases}
I' \subseteq I,\quad I'_S = I_S,\quad \pi(I') = \pi(I), \\
T_0^a, \ldots, T_N^a \in H_1 A + \cdots + H_e A + I', \\
G_{ij}G_{jl} - G_{il} \in I'_{(H_iH_jH_l)}\quad(\forall\ i, j, l \in \{ 1, \ldots, e \}).
\end{cases}
\]
Let $R' := A/I'$, $\mathscr X' := \operatorname{Proj}(R')$ and $p' : A \to R'$ be the natural homomorphism.
Obviously $\mathscr X$ is a closed subscheme of $\mathscr X'$.
We set $h'_i = p'(H_i)$ and $g'_{ij} = p'(G_{ij})$.
Then $p'(T_0)^a, \ldots, p'(T_N)^a \in h'_1 R' + \cdots + h'_eR'$, which means that
$\mathscr X' = \bigcup_{i=1}^e D_+(h'_i)$ by \cite[Lemma~3.35 in Section~2.3]{MR1917232}.
Moreover, $g'_{ij}g'_{il} = g'_{il}$.
In particular, $g'_{ij}g'_{ji} = g'_{ii} = 1$, so that $g'_{ij} \in {R'}_{(h'_ih'_j)}^{\times}$.
This means that $\{ g'_{ij} \}_{i,j \in \{ 1, \ldots, e \}}$ gives rise to an invertible ${\mathscr{O}}_{\mathscr X'}$-module
$\mathscr L'$ such that $\rest{\mathscr L'}{\mathscr X} = \mathscr L$.
Moreover, $(\mathscr X', \mathscr L')$ is a model of $(X, L)$ and
the special fiber of $\mathscr X' \to \operatorname{Spec}(\mathfrak o_v)$ is same as
the special fiber of $\mathscr X' \to \operatorname{Spec}(\mathfrak o_v)$, as required.
\end{proof}
\begin{prop}\label{prop:ample:every:fiber:ample}
Let $\mathscr X \to \operatorname{Spec}(\mathfrak o_v)$ be a model of $X$ and $\mathscr L$ be an invertible ${\mathscr{O}}_{\mathscr X}$-module.
If $\mathscr L$ is ample on every fiber of $\mathscr X \to \operatorname{Spec}(\mathfrak o_v)$, then $\mathscr L$ is ample.
\end{prop}
\begin{proof}
By Lemma~\ref{lemma:finitely:presented:model}, there are a coherent model of $\mathscr X'$ of $X$ and an invertible ${\mathscr{O}}_{\mathscr X'}$-module $\mathscr L'$
such that
$\mathscr X$ is a closed subscheme of $\mathscr X'$, $\rest{\mathscr L'}{\mathscr X} = \mathscr L$ and
the special fiber of $\mathscr X' \to \operatorname{Spec}(\mathfrak o_v)$ coincides with the special fiber of $\mathscr X \to \operatorname{Spec}(\mathfrak o_v)$.
Note that $\mathscr L'$ is ample on every fiber of $\mathscr X' \to \operatorname{Spec}(\mathfrak o_v)$, and hence
$\mathscr L'$ is ample by \cite[IV-3, Corollaire~(9.6.4)]{EGA} because
$\mathscr X' \to \operatorname{Spec}(\mathfrak o_v)$ is finitely presented. Therefore $\mathscr L$ is ample.
\end{proof}
\begin{theo}
\label{thm:equiv:semi-positive:M:semi-positive}
We assume that $v$ is non-Archimedean and $|\raisebox{.4ex}{.}|$ is not trivial.
Let $L$ be a semi-ample invertible ${\mathscr{O}}_X$-module and $\varphi$ be a continuous metric of $L$.
Then $\varphi$ is semi-positive if and only if $\varphi$ is $M$-semi-positive.
\end{theo}
\begin{proof}
First we assume that $\varphi$ is semi-positive.
By Remark~\ref{rem:model:nef:metric}, we may assume that $X$ is reduced.
As $L$ is semi-positive, there is a positive integer $n_0$ such that $L^{\otimes n_0}$ is generated by global sections,
so we may assume that $L$ is generated by global sections, and hence
$L^{\otimes n}$ is generated by global sections for all $n \geqslant 1$.
Fix $\lambda \in \mathopen{]}0,1\mathclose{[}$ such that $\lambda < \sup \{ |a| : a \in k^{\times}, |a| < 1 \}$.
By \cite[Proposition~1.2.22]{CMArakelovAdelic},
there is a finitely generated lattice $\mathcal E_n$ of $H^0(X, L^{\otimes n})$ such that
$d(\|\raisebox{.4ex}{.}\|_{\mathcal E_n}, \|\raisebox{.4ex}{.}\|_{n\varphi}) \leqslant \log(\lambda^{-1})$.
Note that there is a morphism $f_n : X \to {\mathbb{P}}(H^0(X, L^{\otimes n}))$ with $f_n^*({\mathscr{O}}_{{\mathbb{P}}(H^0(X, L^{\otimes n}))}(1)) = L^{\otimes n}$, so we can find a morphism
$\mathcal F_n : \mathscr X_n \to {\mathbb{P}}(\mathcal E_n)$ over $\mathfrak o_v$
such that $\mathscr X_n$ is flat and projective over $\mathfrak o_v$ and
$\mathcal F_n$ is an extension of $f_n$ over $\mathfrak o_v$. If we set $\mathscr L_n = \mathcal F_n^*({\mathscr{O}}_{{\mathbb{P}}(\mathcal E_n)}(1))$,
then $(\mathscr X_n, \mathscr L_n)$ is a flat model of $(X, L^{\otimes n})$.
As $\mathcal E_n \otimes_{\mathfrak o_v} {\mathscr{O}}_{{\mathbb{P}}(\mathcal E_n)} \to {\mathscr{O}}_{{\mathbb{P}}(\mathcal E_n)}(1)$ is surjective,
one also has the sujectivity of $\mathcal E_n \otimes_{\mathfrak o_v} {\mathscr{O}}_{\mathscr X_n} \to \mathscr L_n$.
Therefore, by \cite[Proposition~2.3.12]{CMArakelovAdelic}, the model metric $\varphi_{\mathscr L_n}$ coincides
with the quotient metric induced by $\|\raisebox{.4ex}{.}\|_{\mathcal E_n}$.
Therefore, if we denote by $\varphi_n$ the quotient metic induced by $\|\raisebox{.4ex}{.}\|_{n\varphi}$, then, by \cite[Proposition~2.2.20]{CMArakelovAdelic},
\[
d(\varphi_{\mathscr L_n}, \varphi_n) \leqslant d(\|\raisebox{.4ex}{.}\|_{\mathcal E_n}, \|\raisebox{.4ex}{.}\|_{n\varphi})
\leqslant \log(\lambda^{-1}),
\]
which implies
\begin{multline*}
d((1/n)\varphi_{\mathscr L_n}, \varphi) \leqslant d((1/n)\varphi_{\mathscr L_n}, (1/n)\varphi_n) + d( (1/n)\varphi_n, \varphi) \\
\leqslant (1/n) \log(\lambda^{-1}) + d( (1/n)\varphi_n, \varphi),
\end{multline*}
and hence $\lim\limits_{n\to\infty} d((1/n)\varphi_{\mathscr L_n}, \varphi) = 0$.
Thus $\varphi$ is $M$-semi-positive because $\mathscr L_n$ is nef.
\medskip
Let us see the converse.
Let $(\mathscr X, \mathscr L)$ be a model of $(X, L)$ such that $\mathscr L$ is nef along
the special fiber of $\mathscr X \to \operatorname{Spec}(\mathfrak o_v)$.
Let $\varphi_{\mathscr L}$ be the metric arising from the model $(\mathscr X, \mathscr L)$.
It is sufficient to see that $\varphi_{\mathscr L}$ is semi-positive.
Let $\mathscr A$ be an ample invertible ${\mathscr{O}}_{\mathscr X}$-module.
Then, for $n \geqslant 1$, $\mathscr A \otimes \mathscr L^{\otimes n}$ is ample
on every fiber of $\mathscr X \to \operatorname{Spec}(\mathfrak o_v)$, and hence,
by Proposition~\ref{prop:ample:every:fiber:ample}, $\mathscr A \otimes \mathscr L^{\otimes n}$ is ample
on $\mathscr X$ for all $n \geqslant 1$.
Therefore, by \cite[Proposition~2.3.17]{CMArakelovAdelic}, $\varphi_{\mathscr L}$ is semi-positive.
\end{proof}
\section{Green functions}
In this section, we fix a projective $k$-scheme $X$.
\begin{defi}
Let $D$ be a Cartier divisor on $X$. We call \emph{Green function} of $D$ any real-valued continuous function on $(X\setminus\operatorname{Supp}(D))^{\mathrm{an}}$ such that, for any regular meromorphic function $f\in\Gamma(U,\mathscr M_X^{\times})$ which defines the Cartier divisor locally on a Zariski open subset $U$, the function $g+\log|f|$ on $(U\setminus\operatorname{Supp}(D))^{\mathrm{an}}$ extends to a continuous function on $U^{\mathrm{an}}$. A pair $(D,g)$ consisting of a Cartier divisor $D$ on $X$ and a Green function $g$ of $D$ is called a \emph{metrized Cartier divisor}. We denote by $\widehat{\operatorname{Div}}(X)$ the set of all metrized Cartier divisors on $X$.
Further $g$ is said to be \emph{smooth} if $({\mathscr{O}}_X(D), |\raisebox{.4ex}{.}|_g)$ is smooth.
A smooth Green function of $D=0$ is called a \emph{smooth function} on $X^{\operatorname{an}}$.
\end{defi}
\begin{exem}\label{Rem: Green function of several points}
In the case where $D$ is the zero Cartier divisor, Green functions of $D$ are continuous functions on $X^{\mathrm{an}}$. In particular, if the Krull dimension of $X$ is zero, then $X^{\mathrm{an}}$ consists of isolated points. In this case any Cartier divisor $D$ on $X$ is trivial (see Remark \ref{remark:Cartier:div:on:0:dim}) and hence Green functions identify with elements in the real vector space spanned by $X^{\operatorname{an}}$.
In the case where $D$ is a principal Cartier divisor, namely a Cartier divisor of the form $\operatorname{div}(f)$, where $f$ is a regular meromorphic function, then by definition $-\ln|f|$ is a Green function of $\operatorname{div}(f)$. We denote by $\widehat{\operatorname{div}}(f)$ the pair $(\operatorname{div}(f),-\ln|f|)$. Such a metrized Cartier divisor is said to be \emph{principal}.
\end{exem}
\begin{rema}\label{Rem: metric induced by a green function}
Metrized Cartier divisors are closely related to metrized invertible sheafs. Let $D$ be a Cartier divisor on $X$. We denote by $\mathcal O_X(D)$ the sub-$\mathcal O_X$-module of $\mathscr M_X$ generated by $-D$. Let $(U_i)_{i\in I}$ be an open covering of $X$ such that, on each $U_i$ the Cartier divisor is defined by a regular meromorphic function $s_i$. Then the restriction of $\mathcal O_X(D)$ at $U_i$ is given by $\mathcal O_{U_i}s_i^{-1}$. If $g$ is a Green function of $D$, then it induces a continuous metric $\varphi_g=(|\raisebox{.4ex}{.}|_g(x))_{x\in X^{\mathrm{an}}}$ on $\mathcal O_X(D)$ such that \[|s_i^{-1}|_{g}:=\exp(-g-\ln|s_i|)\text{ on $U_i^{\mathrm{an}}$}.\]
Note that the metric of the canonical regular meromorphic section (see Definition \ref{Def: divisor of a section}) is given by
\[|s_D|_g=|s_i\otimes s_i^{-1}|_g=\exp(-g)\text{ on $U_i$}.\]
Conversely, given an invertible $\mathcal O_X$-module $L$, any non-zero rational section $s$ of $L$ defines a Cartier divisor $\operatorname{div}(L;s)$. Moreover, if $\varphi$ is a continuous metric on $L$, then $-\ln|s|_{\varphi}$ is a Green function of $\operatorname{div}(L;s)$.
We denote by $\operatorname{\widehat{div}}(\overline L;s)$ (or by $\operatorname{\widehat{div}}(s)$ for simplicity) the metrized Cartier divisor $(\operatorname{div}(L;s),-\ln|s|_\varphi)$.
The above relation between metrized Cartier divisors and metrized invertible sheaves is important to define the following composition law on the set of metrized Cartier divisors. Let $(D_1,g_1)$ and $(D_2,g_2)$ be metrized Cartier divisors. Note that $\mathcal O_X(D_1+D_2)$ is canonically isomorphic to $\mathcal O_X(D_1)\otimes_{\mathcal O_X}\mathcal O_{X}(D_2)$. Moreover, under the canonical isomorphism
\[\mathcal O_{X}(D_1+D_2)\stackrel{\sim}{\longrightarrow}\mathcal O_X(D_1)\otimes_{\mathcal O_X}\mathcal O_{X}(D_2),\]
the regular meromorphic section $s_{D_1+D_2}$ corresponds to $s_{D_1}\otimes s_{D_2}$. We equip the invertible sheaf $\mathcal O_{X}(D_1)$ and $\mathcal O_X(D_2)$ with the metrics $\varphi_{g_1}=(|\raisebox{.4ex}{.}|_{g_1}(x))_{x\in X^{\operatorname{an}}}$ and $\varphi_{g_2}=(|\raisebox{.4ex}{.}|_{g_2}(x))_{x\in X^{\operatorname{an}}}$ respectively, and $\mathcal O_X(D_1+D_2)$ with the tensor product metric $\varphi_{g_1}\otimes\varphi_{g_2}$. We then denote by $g_1+g_2$ the Green function in the metrized Cartier divisor $\widehat{\operatorname{div}}(s_{D_1+D_2})$. Clearly, for any $x\in \big(X\setminus(\operatorname{Supp}(D_1)\cup\operatorname{Supp}(D_2))\big)^{\operatorname{an}}$, one has
\[(g_1+g_2)(x)=g_1(x)+g_2(x).\]
Note that the set $\widehat{\operatorname{Div}}(X)$ of metrized Cartier divisors equipped with this composition law forms a commutative group.
\end{rema}
\begin{defi}
Let $(A,g)$ be a metrized Cartier divisor such that $\mathcal O_X(A)$ is an ample invertible $\mathcal O_X$-module (namely the Cartier divisor $A$ is ample). We say that the Green function $g$ is \emph{plurisubharmonic} if the metric $|\raisebox{.4ex}{.}|_g$ on $\mathcal O_X(A)$ is semi-positive. We refer to \cite[\S6.8]{ChamDucros} and \cite[\S6]{MR3975640} for a local version of positivity conditions.
We say that a metrized Cartier divisor $(D, g)$ is \emph{integrable}
if there are ample Cartier divisors $A_1$ and $A_2$ together with
plurisubharmonic Green functions $g_1$ and $g_2$ of $A_1$ and $A_2$, respectively, such that
$(D, g) = (A_1, g_1) - (A_2, g_2)$. We denote by $\widehat{\operatorname{Int}}(X)$ the set of all integrable metrized Cartier divisors. This is a subgroup of the group $\widehat{\operatorname{Div}}(X)$ of metrized Cartier divisors.
\end{defi}
\begin{rema}\label{Rem: extension scalar Green function}
Let $k'/k$ be a valued extension which is complete. Let $X_{k'}$ be the fiber product $X\times_{\operatorname{Spec} k}\operatorname{Spec} k'$, and $\pi:X_{k'}\rightarrow X$ be the morphism of projection. Let $(D,g)$ be a metrized Cartier divisor on $X$. Then the pull-back $D_{k'}$ of $D$ by the morphisme $\pi$ is well defined (see Definition \ref{Def: pull back} and Remark \ref{Rem: extension of scalars}). Note that $\mathcal O_{X_{k'}}(D_{k'})$ is isomorphic with the pull-back of $\mathcal O_X(D)$ by $\pi$, and the canonical meromorphic section $s_{D_{k'}}$ of $D_{k'}$ identifies with the pull-back of $s_D$ by $\pi$. Let $\varphi_g$ be the continuous metric on $\mathcal O_{X}(D)$ induced by the Green function $g$. We denote by $g_{k'}$ the Green function of $D_{k'}$ defined as
\[g_{k'}=-\ln|s_{D_{k'}}|_{\varphi_{g,k'}},\]
where $\varphi_{g,k'}$ is the continuous metric on $\pi^*(\mathcal O_X(D))\cong\mathcal O_{X_{k'}}(D_{k'})$ induced by $\varphi_g$ (see Example \ref{Exa: Fubini-Study} \ref{Item: extension of scalars metric}). Note that, for any element $x'\in X_{k'}^{\mathrm{an}}$ such that $\pi^{\natural}(x')\in (X\setminus\operatorname{Supp}(D))^{\mathrm{an}}$, one has
\[g_{k'}(x')=g(\pi^\natural(x')).\]
Moreover, the composition of $g$ with the restriction of $\pi^{\natural}$ to $(X_{k'}\setminus\operatorname{Supp}(D_{k'}))^{\mathrm{an}}$ forms a Green function of $D_{k'}$. We denote by $g_{k'}$ this Green function. By Remark \ref{Rem: positivity extension scalar}, if $\mathcal O_X(D)$ is semi-ample and $g$ is plurisubharmonic, then $g_{k'}$ is also plurisubharmonic. If $(D,g)$ is integrable, then $(D_{k'},g_{k'})$ is also integrable. Therefore the correspondance
$(D,g)\mapsto (D_{k'},g_{k'})$
defines a group homomorphism from $\widehat{\operatorname{Div}}(X)\rightarrow\widehat{\operatorname{Div}}(X_{k'})$, whose restriction to $\widehat{\operatorname{Int}}(X)$ defines a group homomorphism $\widehat{\operatorname{Int}}(X)\rightarrow\widehat{\operatorname{Int}}(X_{k'})$.
\end{rema}
\begin{theo}
\label{thm:approximation:smooth:funcs}
Let $X$ be a $d$-dimensional projective and integral scheme over $k$. Let
$D$ be a nef and effective Cartier divisor and $g$ be a Green function of $D$ such that either
\begin{enumerate}[label=\rm(\alph*)]
\item if $v$ is Archimedean, the metric of $|\raisebox{.4ex}{.}|_g$ of ${\mathscr{O}}_X(D)$ is $C^{\infty}$ and semi-positive, or
\item if $v$ is non-Archimedean, the metric of $|\raisebox{.4ex}{.}|_g$ of ${\mathscr{O}}_X(D)$ is a nef model metric.
\end{enumerate}
Then there is a sequence $( \psi_n )_{n\in\mathbb N}$ of smooth functions on $X^{\operatorname{an}}$ with the following properties:
\begin{enumerate}[label=\rm(\arabic*)]
\item for all $n \in {\mathbb{Z}_{\geqslant 1}}$, $\psi_n \leqslant g$, $\psi_n \leqslant \psi_{n+1}$.
\item for each point $x \in X^{\operatorname{an}}$, $\sup \{ \psi_n(x) ; n \in {\mathbb{Z}_{\geqslant 1}} \} = g(x)$.
\item for all $n \in {\mathbb{Z}_{\geqslant 1}}$, $g - \psi_n$ is a Green function of $D$ such that either
\begin{enumerate}[label=\rm(\arabic{enumi}.\alph*)]
\item if $v$ is Archimedean, the metric of $|\raisebox{.4ex}{.}|_{g-\psi_n}$ of ${\mathscr{O}}_X(D)$ is $C^{\infty}$ and semi-positive, or
\item if $v$ is non-Archimedean, the metric of $|\raisebox{.4ex}{.}|_{g-\psi_n}$ of ${\mathscr{O}}_X(D)$ is a nef model metric.
\end{enumerate}
\end{enumerate}
\end{theo}
\begin{proof}
This theorem is nothing more than \cite[Th\'eor\`eme~3.1]{MR2244803}. In the case where $v$ is non-Archimedean, it is proved under
the additional assumption that $v$ is discrete. However, their proof works well by slight modifications.
For reader's convenience, we reprove it here.
\medskip
We may assume that $v$ is non-Archimedean.
If the theorem holds for $(mD, mg)$ for some positive number $m$, then
it also holds for $(D, g)$, so that we may assume that
there is a flat model $(\mathscr X, \mathscr L)$ of $(X, {\mathscr{O}}_X(D))$ such that
$|\raisebox{.4ex}{.}|_g = |\raisebox{.4ex}{.}|_{\varphi_{\mathscr L}}$ and $\mathscr L$ is nef along the special fiber of $\mathscr X \to \operatorname{Spec}({\mathfrak{o}}_v)$.
By Lemma~\ref{lem:eq:regular:mero:generic:total}, there is a Cartier divisor $\mathscr D$ on $\mathscr X$
such that ${\mathscr{O}}_{\mathscr X}(\mathscr D) = \mathscr L$, $\rest{\mathscr D}{X} = D$ and
$g$ is the Green function arising from $(\mathscr X, \mathscr D)$.
Let $\mathscr X = \bigcup_{i=1}^N \operatorname{Spec} (\mathscr A_i)$ be an affine open covering of $\mathscr X$ such that
$\mathscr D$ is given by a local equation $f_i$ on $\operatorname{Spec} (\mathscr A_i)$.
Since $D$ is effective, one has $f_i \in (\mathscr A_i)_S$, that is, $s_i f_i \in \mathscr A_i$ for some $s_i \in S$,
where $S := \mathfrak o_v \setminus \{ 0 \}$, so that
if we set $s = s_1 \cdots s_N$, then $s f_i \in \mathscr A_i$ for all $i = 1, \ldots, N$.
Let \[g' := g - \log |s|,\quad \mathscr L' := \mathscr L \otimes {\mathscr{O}}_{\mathscr X} s^{-1}\quad\text{and}\quad
\mathscr D' := \mathscr D + \operatorname{div}(s).\]
Then $\mathscr D'$ is effective, ${\mathscr{O}}_{\mathscr X}(\mathscr D') = \mathscr L'$ and $|\raisebox{.4ex}{.}|_{g'} = |\raisebox{.4ex}{.}|_{\mathscr L'}$. Thus, if the theorem holds for $g'$, then one has the assertion for $g$, and hence we may further assume that $\mathscr D$ is
effective.
Fix $a \in S$ such that $|a| < 1$, and set
\[ \psi_n = \min \{ g, -n \log |a|\}\quad (\forall\ n \in {\mathbb{Z}_{\geqslant 1}}).\]
The properties (1) and (2) are obvious, so
we need to see (3).
Let $\mathscr I_n$ be the ideal sheaf of ${\mathscr{O}}_{\mathscr X}$ generated by a local equation of $\mathscr D$ and $a^n$.
Let $p_n : \mathscr Y_n \to \mathscr X$ be the blowing-up in terms of the ideal sheaf $\mathscr I_n$.
Note that $\mathscr I_n {\mathscr{O}}_{\mathscr Y}$ is a locally principal ideal sheaf of ${\mathscr{O}}_{\mathscr Y_n}$
whose support is contained in the special fiber of $\mathscr Y_n \to \operatorname{Spec}(\mathfrak o_v)$, that is,
there is an effective Cartier divisor $\mathscr E_n$ on $\mathscr Y_n$ such that
${\mathscr{O}}_{\mathscr Y_n}(-\mathscr E_n) = \mathscr I_n {\mathscr{O}}_{\mathscr Y_n}$ and $\rest{\mathscr E_n}{X} = 0$.
Obviously $\psi_n$ is a smooth function arising from the model $(\mathscr Y_n, \mathscr E_n)$.
Therefore, it is sufficient to show that $p_n^*(\mathscr D) - \mathscr E_n$ is nef along the special fiber
$\mathscr Y_n \to \operatorname{Spec}(\mathfrak o_v)$.
Let $\mathscr X = \bigcup_{i=1}^N \operatorname{Spec} (\mathscr A_i)$ be an affine open covering of $\mathscr X$ as before.
Note that
$\mathscr D$ is given by $f_i \in \mathscr A_i$ on $\operatorname{Spec} \mathscr A_i$ for each $i$.
Then \[p_n^{-1}(\operatorname{Spec} \mathscr A_i) = \operatorname{Proj} (\mathscr A_i[T_0, T_1]/(f_i T_0 - a^n T_1)).\]
If we set $p_n^{-1}(\operatorname{Spec} \mathscr A_i)_\alpha = \{ T_\alpha \not = 0 \}$ for $\alpha \in \{ 0, 1 \}$,
then $f_i = a^n (T_1/T_0)$ on $p_n^{-1}(\operatorname{Spec} \mathscr A_i)_0$ and
$a^n = f_i (T_0/T_1)$ on $p_n^{-1}(\operatorname{Spec} \mathscr A_i)_1$, so that
\begin{equation}
\label{eqn:thm:approximation:smooth:funcs:01}
\begin{cases}
\rest{{\mathscr{O}}_{\mathscr Y_n}(-\mathscr E_n)}{p_n^{-1}(\operatorname{Spec} \mathscr A_i)_0} = a^n {\mathscr{O}}_{p_n^{-1}(\operatorname{Spec} \mathscr A_i)_0}, \\
\rest{{\mathscr{O}}_{\mathscr Y_n}(-\mathscr E_n)}{p_n^{-1}(\operatorname{Spec} \mathscr A_i)_1} = f_i {\mathscr{O}}_{p_n^{-1}(\operatorname{Spec} \mathscr A_i)_1}.
\end{cases}
\end{equation}
Therefore, one can see that
$p_n^*(\mathscr D) - \mathscr E_n$ and $\mathrm{div}(a^n) - \mathscr E_n$ are effective.
Let us see $(p_n^*(\mathscr D) - \mathscr E_n \cdot C) \geqslant 0$
for any irreducible curve $C$ on the special fiber of $\mathscr Y_n \to \operatorname{Spec}(\mathfrak o_v)$.
Let $\xi$ be the generic point of $C$. We choose $i$ such that $\xi \in p_n^{-1}(\operatorname{Spec} \mathscr A_i)$.
If $\xi \not\in \operatorname{Supp}(p_n^*(\mathscr D) - \mathscr E_n)$, then the assertion is obvious because $p_n^*(\mathscr D) - \mathscr E_n$ is effective.
Otherwise, by \eqref{eqn:thm:approximation:smooth:funcs:01}, $\xi \in p_n^{-1}(\operatorname{Spec} \mathscr A_i)_0$.
Then, by \eqref{eqn:thm:approximation:smooth:funcs:01} again, $\xi \not\in \operatorname{Supp}(\mathrm{div}(a^n) - \mathscr E_n)$,
so that $((\mathrm{div}(a^n) - \mathscr E_n) \cdot C) \geqslant 0$ by the reason of
the effectivity of $\mathrm{div}(a^n) - \mathscr E_n$.
Note that $p_n^*(\mathscr D) - \mathscr E_n$ is linearly equivalent to $p_n^*(\mathscr D) + (\mathrm{div}(a^n) - \mathscr E_n)$. Thus it is sufficient to show that $(p_n^*(\mathscr D) \cdot C) \geqslant 0$,
which is obvious because of the projection formula and the nefness of $\mathscr D$.
\end{proof}
\section{Local measures}\label{sec:local:measures}
In this section, we assume that $k$ is \emph{algebraically closed}. Let $X$ be a
projective $k$-scheme and let $d$ be the dimension of $X$. Assume given a family $(L_i)_{i=1}^d$ of semi-ample invertible $\mathcal O_X$-modules. For any $i\in\{1,\ldots,d\}$, let $\varphi_i$ be a semi-positive continuous metric on $L_i$.
First we assume that $X$ is integral.
In the case where $|\raisebox{.4ex}{.}|$ is Archimedean (and hence $k=\mathbb C$), by Bedford-Taylor theory \cite{MR445006} one can construct a Borel measure
\[c_1(L_1,\varphi_1)\cdots c_1(L_{d},\varphi_d)\]
having
\[\deg(c_1(L_1)\cdots c_1(L_d)\cap[X])\]
as its total mass. In the non-Archimedean case, an analoguous measure has been proposed by Chambert-Loir \cite{MR2244803}, assuming that the field $k$ admits a dense countable subfield (see also \cite[\S5]{ChamDucros} for a general non-Archimedean analogue of Bedford-Taylor theory).
In any case, the measure $c_1(L_1,\varphi_1)\cdots c_1(L_{d},\varphi_d)$ is often denoted by
$\mu_{(L_1,\varphi_1)\cdots (L_{d},\varphi_d)}$.
Note that the measure $\mu_{(L_1,\varphi_1)\cdots (L_d,\varphi)}$ is additive with respect to each $(L_i,\varphi_i)$. More precisely, if $i\in\{1,\ldots,d\}$ and if $(M_i,\psi_i)$ is another semi-positively metrized invertible $\mathcal O_X$-module, then the measure
\[\mu_{(L_1,\varphi_1)\cdots(L_{i-1},\varphi_{i-1})(L_i\otimes M_i,\varphi_i\otimes\psi_i)(L_{i+1},\varphi_{i+1})\cdots(L_d,\varphi_d)}\]
is equal to
\[\mu_{(L_1,\varphi_1)\cdots(L_d,\varphi_d)}+\mu_{(L_1,\varphi_1)\cdots(L_{i-1},\varphi_{i-1})(M_i,\psi_i)(L_{i+1},\varphi_{i+1})\cdots(L_d,\varphi_d)}\]
Moreover, for any permutation $\sigma:\{1,\ldots,d\}\rightarrow\{1,\ldots,d\}$, one has
\[\mu_{(L_{\sigma(1)},\varphi_{\sigma(1)})\cdots(L_{\sigma(d)},\varphi_{\sigma(d)})}=\mu_{(L_1,\varphi_1)\cdots(L_d,\varphi_d)}.\]
In general, let $X_1, \ldots, X_n$ be irreducible components of $X$ which are of dimension $d$, and $\eta_1, \ldots, \eta_n$ the generic points of $X_1, \ldots, X_n$, respectively.
Let $\xi_i : X_i \hookrightarrow X$ be the canonical closed embedding for each $i$.
Then a measure $\mu_{(L_1,\varphi_1)\cdots (L_{d},\varphi_d)}$ on $X^{\mathrm{an}}$ is defined to be
\begin{multline}\label{def:measure:assocated:invertible:sheaves}
\mu_{(L_1,\varphi_1)\cdots (L_{d},\varphi_d)} := \\
\sum_{j=1}^n \operatorname{length}_{{\mathscr{O}}_{X,\eta_j}}({\mathscr{O}}_{X,\eta_j})(\xi^{\mathrm{an}}_j)_*\Big(c_1\big(\xi_j^*(L_1,\varphi_1)\big)\cdots c_1\big(\xi_j^*(L_{d},\varphi_d)\big)\Big).
\end{multline}
\begin{defi}
Let $(L_1,\varphi_1),\ldots,(L_d,\varphi_d)$ be a family of integrable metrized invertible $\mathcal O_X$-modules. For each $i\in\{1,\ldots,d\}$, we let $(L_i',\varphi_i')$ and $(L_i'',\varphi_i'')$ be ample invertible $\mathcal O_X$-modules equipped with semi-positive metrics, such that $L_i=L_i'\otimes (L_i'')^\vee$ and $\varphi_i=\varphi_i'\otimes(\varphi_i'')^\vee$. We define a signed Radon measure $\mu_{(L_1,\varphi_1)\cdots(L_d,\varphi_d)}$ on $X^{\mathrm{an}}$ as follows:
\[\mu_{(L_1,\varphi_1)\cdots(L_d,\varphi_d)}:=\sum_{I\subseteq\{1,\ldots,d\}}(-1)^{\operatorname{card}(I)}\mu_{(L_{1,I},\varphi_{1,I})\cdots(L_{d,I},\varphi_{d,I})},\]
where $(L_{j,I},\varphi_{j,I})=(L_j'',\varphi_j'')$ if $j\in I$, and $(L_{j,I},\varphi_{j,I})=(L_j',\varphi_j')$ if $j\in\{1,\ldots,d\}\setminus I$.
\end{defi}
\begin{exem}\label{Exe:chambert-loir measure}We recall the explicit construction of Chambert-Loir's measure in a particular case as explained in \cite[\S2.3]{MR2244803}.
Assume that the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean and that the $k$-scheme $X$ is integral and normal. Let $k^\circ$ be the valuation ring of $(k,|\raisebox{.4ex}{.}|)$ and $\mathfrak m$ be the maximal ideal of $k^\circ$. Suppose given an integral model of $X$, namely, a flat and normal projective $k^\circ$-scheme $\mathscr X$ such that \[\mathscr X\times_{\operatorname{Spec} k^\circ}\operatorname{Spec} k\cong X.\]
Let $\mathscr X_{\mathfrak m}$ be the fibre of $\mathscr X$ over the closed point of $\operatorname{Spec} k^\circ$. It turns out that the reduction map from $X^{\mathrm{an}}$ to $\mathscr X_{\mathfrak m}$ is surjective. Let $Z_1,\ldots,Z_n$ be irreducible components of $\mathscr X_{\mathfrak m}$. For any $j\in\{1,\ldots,n\}$, there exists a unique point $z_j\in X^{\mathrm{an}}$ whose reduction identifies with the generic point of $Z_j$.
Assume that each metric $\varphi_j$ is induced by an integral model $\mathscr L_i$, which is an invertible sheaf on $\mathscr X$ such that $\mathscr L_i|_X\cong L_i$. Then the measure
\[c_1(L_1,\varphi_1)\cdots c_1(L_{d},\varphi_d)\]
is given by
\[\sum_{j=1}^d \operatorname{mult}_{Z_j}(\mathscr X_{\mathfrak m})\deg(c_1(\mathscr L_1|_{\mathscr X_{\mathfrak m}})\cdots c_1(\mathscr L_{d}|_{\mathscr X_{\mathfrak m}})\cap[Z_j])\operatorname{Dirac}_{z_j},\]
where $\operatorname{mult}_{Z_j}(\mathscr X_{\mathfrak m})$ is the multiplicity of $Z_j$ in $\mathscr X_{\mathfrak m}$, and $\operatorname{Dirac}_{z_j}$ denotes the Dirac measure at $z_j$.
\end{exem}
\begin{rema}\label{rem:approximation:CL:measure}
We assume that $X$ is integral. Let
$\{ \varphi_{1, n} \}_{n=1}^{\infty}, \ldots, \{ \varphi_{d, n} \}_{n=1}^{\infty}$ be sequences of semi-positive
metrics of $L_1, \ldots, L_d$, respectively such that \[\lim\limits_{n\to\infty} d(\varphi_{i, n}, \varphi_i) = 0\]
for all $i=1, \ldots, d$. Then, by using \cite[Corollary~(3.6)]{Demagbook} and \cite[Corollaire~(5.6.5)]{ChamDucros}, one can see
\[
\lim_{n\to\infty} \int_{X^{\operatorname{an}}} f \mu_{(L_1, \varphi_{1, n}) \cdots (L_d, \varphi_{d, n})} =
\int_{X^{\operatorname{an}}} f \mu_{(L_1, \varphi_{1}) \cdots (L_d, \varphi_{d})}
\]
for any smooth function $f$ on $X^{\operatorname{an}}$.
\end{rema}
\begin{defi}
Let $\overline D_1=(D_1,g_1),\ldots,\overline D_d=(D_d,g_d)$ be a family of integrable metrized Cartier divisors on $X$. For any $i\in\{1,\ldots,d\}$, we write $(D_i,g_i)$ as the difference of two metrized Cartier divisors $(D_i',g_i')-(D_i'',g_i'')$, where $D_i'$ and $D_i''$ are ample, and $g_i'$ and $g_i''$ are plurisubharmonic. We define a signed Radon measure $\mu_{\overline D_1\cdots \overline D_d}$ on $X^{\mathrm{an}}$
to be
\begin{equation*}\mu_{\overline D_1\cdots\overline D_d}
:=\sum_{I\subseteq\{1,\ldots,d\}}(-1)^{\operatorname{card}(I)}\mu_{\overline{D}_{1,I}\cdots\overline D_{d,I}},
\end{equation*}
where $\overline D_{j,I}=(D_i'',g_{i}'')$ if $j\in I$, and $\overline D_{j,I}=(D_i',g_i')$ if $j\in\{1,\ldots,d\}\setminus I$.
Note that this signed measure does not depend on the choice of the decompositions.
Let $X_1, \ldots, X_n$ be irreducible components of $X$ and $\eta_1, \ldots, \eta_n$ be the generic points
of $X_1, \ldots, X_n$, respectively.
Let $\xi_j : X_j \hookrightarrow X$ be the canonical closed embedding.
Then it is easy to see
\begin{equation}\label{eqn:measure:non:reduced:scheme}
\mu_{(D_1, g_1) \cdots (D_d, g_d)} = \sum_{j=1}^n \operatorname{length}_{{\mathscr{O}}_{X, \eta_j}}({\mathscr{O}}_{X, \eta_j})
(\xi_j^{\mathrm{an}})_*\Big(\mu_{\xi_j^*(D_1, g_1) \cdots \xi_j^*(D_d, g_d)}\Big).
\end{equation}
\end{defi}
\begin{prop}\label{prop:CH:measure:projection:formula}
Let $\pi : Y \to X$ be a surjective morphism between integral projective schemes over $k$.
We set $e = \dim X$ and $d = \dim Y$.
Let $(L_1, \varphi_1), \ldots, (L_d, \varphi_d)$ be integrable metrized invertible ${\mathscr{O}}_X$-modules.
Then one has the following:
\begin{enumerate}[label=\rm(\arabic*)]
\item If $d > e$, then $\pi_*(\mu_{\pi^*(L_1, \varphi_1) \cdots \pi^*(L_d, \varphi_d)}) = 0$.
\item If $d = e$, then
$\pi_*(\mu_{\pi^*(L_1, \varphi_1) \cdots \pi^*(L_d, \varphi_d)}) = (\deg \pi)\mu_{(L_0, \varphi_0) \cdots (L_d, \varphi_d)}$.
\end{enumerate}
\end{prop}
\begin{proof}
We may assume that $L_1, \ldots, L_d$ are ample and $\varphi_1, \ldots, \varphi_d$ are semi-positive.
If $\varphi_1, \ldots, \varphi_d$ are smooth, then the assertion is well-known (cf. \cite[Proposition~10.4]{GKTropical}).
Let $\{ \varphi_{1,n} \}_{n=1}^{\infty}, \ldots, \{ \varphi_{d,n} \}_{n=1}^{\infty}$ be regularizations of
$\varphi_1, \ldots, \varphi_d$, that is, $\varphi_{1,n}, \ldots, \varphi_{d,n}$ are smooth and
semi-positive for $i=1, \ldots, d$ and $n \geqslant 1$, and
$\lim\limits_{n\to\infty} d(\varphi_{i}, \varphi_{i, n}) = 0$ for $i=1, \ldots, d$
(for example, see \cite{MR3028754} for the Archimedean case and Theorem~\ref{thm:equiv:semi-positive:M:semi-positive} for the non-Archimedean case).
Let $f$ be a smooth function on $X^{\operatorname{an}}$. Then, by using \cite[Corollary~(3.6)]{Demagbook} and \cite[Corollaire~(5.6.5)]{ChamDucros}, one can see that
\begin{align*}
& \lim_{n\to\infty} \int_{X^{\operatorname{an}}} \pi^*(f) \mu_{\pi^*(L_{1}, \varphi_{1,n}) \cdots \pi^*(L_d, \varphi_{d,n})}
= \int_{X^{\operatorname{an}}} \pi^*(f) \mu_{\pi^*(L_{1}, \varphi_{1}) \cdots \pi^*(L_d, \varphi_{d})} \\
\intertext{and if $d = e$, then}
& \lim_{n\to\infty} \int_{Y^{\operatorname{an}}} f \mu_{(L_{1}, \varphi_{1,n}) \cdots (L_d, \varphi_{d,n})}
= \int_{X^{\operatorname{an}}} f \mu_{(L_{1}, \varphi_{1}) \cdots (L_d, \varphi_{d})}.
\end{align*}
Thus the assertions follow.
\end{proof}
\begin{rema}Let $X$ and $Y$ be two projective schemes over $\operatorname{Spec} k$, of Krull dimension $d$ and $n$, respectively. Let $\overline L_1,\ldots,\overline L_d$ be integrable metrized invertible $\mathcal O_X$-modules, $\overline M_1,\ldots,\overline M_d$ be integrable $\mathcal O_Y$-modules. We consider the fiber product $X\times_kY$ and let $\pi_1:X\times_kY\rightarrow X$ and $\pi_2:X\times_kY\rightarrow Y$ be the two morphisms of projection. In the case where $k$ is Archimedean, the analytic space $(X\times_k Y)^{\mathrm{an}}$ is homeomorphic to $X^{\mathrm{an}}\times Y^{\mathrm{an}}$ and the measure \[\mu_{\pi_1^*(\overline L_1)\cdots\pi_1^*(\overline L_d)\pi_2^*(\overline M_1)\cdots\pi_2^*(\overline M_n)}\]
on $(X\times_kY)^{\operatorname{an}}$ identifies with
\[\mu_{\overline L_1\cdots\overline L_d}\otimes\mu_{\overline M_1\cdots\overline M_n}.\]
In the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean, in general the topological space $(X\times_kY)^{\mathrm{an}}$ is not homeomorphic to $X^{\operatorname{an}}\times Y^{\operatorname{an}}$. However, there is a natural continuous map \[\alpha:(X\times_kY)^{\operatorname{an}}\longrightarrow X^{\mathrm{an}}\times Y^{\mathrm{an}}.\] Then the following equality holds (see \cite[\S2.8]{MR2244803})
\begin{equation*}
\alpha_*\Big(\mu_{\pi_1^*(\overline L_1)\cdots\pi_1^*(\overline L_d)\pi_2^*(\overline M_1)\cdots\pi_2^*(\overline M_n)}\Big)=\mu_{\overline L_1\cdots\overline L_d}\otimes\mu_{\overline M_1\cdots\overline M_n}.
\end{equation*}
In particular, if $g$ is a measurable function on $Y^{\mathrm{an}}$ which is integrable with respect to $\mu_{\overline M_1\cdots\overline M_n}$, one has
\begin{equation}\label{Equ:Fubini}\int_{(X\times_kY)^{\mathrm{an}}}(g\circ\pi_2^{\operatorname{an}})\,\mathrm{d}\mu_{\pi_1^*(\overline L_1)\cdots\pi_1^*(\overline L_d)\pi_2^*(\overline M_1)\cdots\pi_2^*(\overline M_n)}=\int_{Y^{\mathrm{an}}}g\,\mathrm{d}\mu_{\overline M_1\cdots\overline M_n}.\end{equation}
\end{rema}
\begin{defi}
Let $E$ be a finite-dimensional vector space over $k$. We say that a norm $\|\raisebox{.4ex}{.}\|$ on $E$ is \emph{orthonormally decomposable} if
\begin{enumerate}[label=\rm(\arabic*)]
\item in the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean, the norm $\|\raisebox{.4ex}{.}\|$ is ultrametric, and $(E,\|\raisebox{.4ex}{.}\|)$ admits an orthonormal basis $(e_j)_{j=0}^{r}$, namely, \[\forall\,(\lambda_j)_{j=0}^{r}\in k^{r+1},\quad \|\lambda_0e_0+\cdots+\lambda_{r}e_{r}\|=\max_{j\in\{0,\ldots,r\}}|\lambda_j|;\]
\item in the case where $|\raisebox{.4ex}{.}|$ is Archimedean, the norm $\|\raisebox{.4ex}{.}\|$ is induced by an inner product $\mathopen{\langle}\cdot,\cdot\mathclose{\rangle}$.
\end{enumerate}
Note that for each valued extension $(k',|\raisebox{.4ex}{.}|')$ of $(k,|\raisebox{.4ex}{.}|)$, there is a unique norm $\|\raisebox{.4ex}{.}\|_{k'}$ on $E\otimes_kk'$, which is either ultrametric or induced by an inner product, such that any orthonormal basis of $(E,\|\raisebox{.4ex}{.}\|)$ is also an orthonormal basis of the extended normed vector space $(E\otimes_kk',\|\raisebox{.4ex}{.}\|_{k'})$ (see Definition \ref{Def: extension of sclars}).
\end{defi}
\begin{rema}\label{Rem: approximation by good norms}
Let $E$ be a finite-dimensional vector space over $k$, and $\|\raisebox{.4ex}{.}\|$ be an orthonormally decomposable norm on $E$. For any $s\in E$, the real number $\|s\|$ belongs to the image of the absolute value $|\raisebox{.4ex}{.}|$. In particular, if $s$ is non-zero, then there exists $\lambda\in k$ such that $\|\lambda s\|=1$.
In the case where the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean, it is \emph{not} true that any ultrametrically normed vector space admits an orthonormal basis (see \cite[Example 2.3.26]{MR2598517}). However, if $(E,\|\raisebox{.4ex}{.}\|)$ is a finite-dimensional ultrametrically normed vector space over $k$, for any $\alpha\in\mathbb R$ such that $0<\alpha<1$, there exists an $\alpha$-orthogonal basis of $E$ (cf. \cite[\S 2.3]{MR2598517}, see also \cite[\S1.2.6]{CMArakelovAdelic} for details), namely a basis $(e_i)_{i=1}^r$ such that, for any $(\lambda_i)_{i=1}^r\in k^r$,
\[ \alpha\max_{i\in\{1,\ldots,r\}}|\lambda_i|\cdot\|e_i\|\leqslant\|\lambda_1e_1+\cdots+\lambda_re_r\|\leqslant\max_{i\in\{1,\ldots,r\}}|a_i|\cdot\|e_i\|.\]
Moreover, since $k$ is assumed to be algebraically closed, in the case where absolute value $|\raisebox{.4ex}{.}|$ is non-trivial, the image of $|\raisebox{.4ex}{.}|$ is dense in $\mathbb R$. In fact, if $a$ is an element of $k$ such that $|a|\neq 1$, for any non-zero rational number $p/q$ with $p\in\mathbb Z$ and $q\in\mathbb Z_{>0}$, any element $x\in k$ satisfying the polynomial equation
\[x^q=a^p\]
has $|a|^{p/q}$ as absolute value. Therefore, by possibly delating the vectors $(e_i)_{i=1}^r$ we may assume that
\[\alpha\leqslant\|e_i\|\leqslant 1\]
for any $i\in\{1,\ldots,r\}$. Therefore, if we denote by $\|\raisebox{.4ex}{.}\|_\alpha$ the norm on $E$ under which $(e_i)_{i=1}^r$ is an orthonormal basis of $E$, then for any $x=\lambda_1e_1+\cdots+\lambda_re_r$ in $E$, one has
\[\|x\|_\alpha=\max_{i\in\{1,\ldots,r\}}|\lambda_i|\leqslant\alpha^{-1}\max_{i\in\{1,\ldots,r\}}|\lambda_i|\cdot\|e_i\|\leqslant\alpha^{-2}\|x\|,\]
and
\[\|x\|\leqslant\max_{i\in\{1,\ldots,r\}}|\lambda|\cdot\|e_i\|\leqslant\max_{i\in\{1,\ldots,r\}}|\lambda_i|=\|x\|_\alpha.\]
Therefore, one has
\[d(\|\raisebox{.4ex}{.}\|_\alpha,\|\raisebox{.4ex}{.}\|):=\sup_{x\in E\setminus\{0\}}\Big|\ln\|x\|_\alpha-\ln\|x\|\Big|\leqslant -2\ln(\alpha).\]
Thus we can approximate the ultrametric norm $\|\raisebox{.4ex}{.}\|$ by a sequence of ultrametric norms which are orthonormally decomposable.
\end{rema}
\begin{prop}\label{Pro: decomposable quotient norm}Let $(E,\|\raisebox{.4ex}{.}\|)$ be a finite-dimensional vector space over $k$, equipped with an orthonormally decomposable norm. Then any element $s_0\in E$ such that $\|s_0\|=1$ belongs to an orthonormal basis. Moreover, for any quotient vector space $G$ of $E$, the quotient norm on $F$ is orthonormally decomposable.
\end{prop}
\begin{proof}
The statement is classic when $|\raisebox{.4ex}{.}|$ is Archimedean, which follows from the Gram-Schmidt process. In the following, we assume that $|\raisebox{.4ex}{.}|$ is non-Archimedean. Let $k^\circ$ be the valuation ring of $(k,|\raisebox{.4ex}{.}|)$.
Let $(e_j)_{j=0}^r$ be an orthonormal basis of $(E,\|\raisebox{.4ex}{.}\|)$. Without loss of generality, we may assume that $s_0=\lambda_0e_0+\cdots+\lambda_re_r$ with $(\lambda_0,\ldots,\lambda_r)\in (k^\circ)^{r+1}$ and $|\lambda_0|=1$. We then construct an upper triangular matrix $A$ of size $(r+1)\times(r+1)$, such that the first row $A$ is $(\lambda_0,\ldots,\lambda_r)$ and the diagonal coordinates of $A$ are elements of absolute value $1$ in $k$. Then the matrix $A$ belongs to $\operatorname{GL}_{r+1}(k^\circ)$. Let $(s_j)_{j=0}^r$ be the basis of $E$ such that
\[(s_0,\ldots,s_r)^{T}=A(e_0,\ldots,e_r)^T.\]
For any $j\in\{0,\ldots,r\}$, one has $\|s_j\|=1$. Moreover, for any $(b_0,\ldots,b_r)\in k^r$, one has
\[b_0s_0+\cdots+b_rs_r=(b_0,\ldots, b_r)A(e_0,\ldots,e_r)^T.\]
Let $(c_0,\ldots,c_r)=(b_0,\ldots,b_r)A$. Since $(e_0,\ldots,e_r)$ is an orthonormal basis, one has
\[\|b_0s_0+\cdots+b_rs_r\|=\max_{j\in\{0,\ldots,r\}}|c_j|.\]
Note that $(b_0,\ldots,b_r)=(c_0,\ldots,c_r)A^{-1}$. Since $A^{-1}$ belongs to $\operatorname{GL}_{r+1}(k^\circ)$, one has
\[\forall\,i\in\{0,\ldots,r\},\quad |b_i|\leqslant\max_{j\in\{0,\ldots,r\}}|c_j|.\]
Therefore one obtains
\[\|b_0s_0+\cdots+b_rs_r\|\geqslant\max_{i\in\{0,\ldots,r\}}|b_i|.\]
Combined with the strong triangle inequality, we obtain
\[\|b_0s_0+\cdots+b_rs_r\|=\max_{i\in\{0,\ldots,r\}}|b_i|.\]
Therefore $(s_j)_{j=0}^r$ is an orthonormal basis of $(E,\|\raisebox{.4ex}{.}\|)$. In particular, the image of $(s_1,\ldots,s_{r})$ in $E/ks_0$ forms an orthonormal basis of $E/ks_0$ with respect to $\|\raisebox{.4ex}{.}\|$. Therefore the quotient norm on $E/ks_0$ is orthonormally decomposable. By induction we can show that all quotient norms of $\|\raisebox{.4ex}{.}\|$ are orthonormally decomposable.
\end{proof}
In the remaining of this section, we fix a finite-dimensional vector space $E$ equipped with an orthonormally decomposable norm $\|\raisebox{.4ex}{.}\|$. We also choose an orthonormal basis $(e_j)_{j=1}^r$ of $(E,\|\raisebox{.4ex}{.}\|)$. Let $\mathbb P(E)$ be the projective space of $E$ and $\mathcal O_E(1)$ be the universal invertible sheaf on $\mathbb P(E)$. We equip $\mathcal O_E(1)$ with the orthogonal quotient metric $(|\raisebox{.4ex}{.}|(x))_{x\in\mathbb P(E)^{\mathrm{an}}}$ (see Definition \ref{Def: Fubini-Study metric}) and denote by $\overline{\mathcal O_E(1)}$ the corresponding metrized invertible sheaf. Recall that each point $x\in\mathbb P(E)^{\mathrm{an}}$ corresponds to a one-dimensional quotient vector space
\[E\otimes_K\widehat{\kappa}(x)\longrightarrow\mathcal O_E(1)(x),\]
where $\widehat{\kappa}(x)$ denotes the completed residue field of $x$. Then the norm $|\raisebox{.4ex}{.}|(x)$ on $\mathcal O_E(1)(x)$ is by definition the quotient norm of $\|\raisebox{.4ex}{.}\|_{\widehat{\kappa}(x)}$.
\begin{defi}Assume that $|\raisebox{.4ex}{.}|$ is non-Archimedean. We denote by $\xi$ the point in $\mathbb P(E)^{\mathrm{an}}$ which is the generic point of $\mathbb P(E)^{\mathrm{an}}$ equipped with the absolute value \[\textstyle|\raisebox{.4ex}{.}|_\xi:k\big(\frac{e_0}{e_r},\ldots,\frac{e_{r-1}}{e_r}\big)\longrightarrow\mathbb R_{\geqslant 0}\]
such that, for any
\[ P=\sum_{\boldsymbol{a}=(a_0,\ldots,a_{r-1})\in\mathbb N^d}\lambda_{\boldsymbol{a}}\Big(\frac{e_0}{e_r}\Big)^{a_0}\cdots\Big(\frac{e_{r-1}}{e_r}\Big)^{a_{r-1}}\in \textstyle{k\big[\frac{e_0}{e_r},\ldots,\frac{e_{r-1}}{e_r}\big]},\]
one has
\[|P|_\xi=\max_{\boldsymbol{a}\in\mathbb N^d}|\lambda_{\boldsymbol{a}}|.\]
Note that the point $\xi$ does not depend on the choice of the orthonormal basis $(e_j)_{j=0}^r$. In fact, the norm $\|\raisebox{.4ex}{.}\|$ induces a symmetric algebra norm on $k[E]$ (which is often called a \emph{Gauss norm}) and hence defines an absolute value on the fraction field of $k[E]$. The restriction of this absolute value to the field of rational functions on $\mathbb P(E)$ identifies with $|\raisebox{.4ex}{.}|_\xi$. Hence $\xi$ is called the \emph{Gauss point} of $\mathbb P(E)^{\mathrm{an}}$.
\end{defi}
\begin{prop}\label{Pro: non Archimedean projective space} Assume that the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean.
The following equality holds
\[c_1(\overline{\mathcal O_E(1)})^r=\operatorname{Dirac}_\xi,\]
where $\operatorname{Dirac}_\xi$ denotes the Dirac measure at $\xi$.
\end{prop}
\begin{proof} Let $k^\circ$ be the valuation ring of $(k,|\raisebox{.4ex}{.}|)$, $\mathfrak m$ be the maximal ideal of $k^\circ$, and $\kappa=k^\circ/\mathfrak m$ be the residue field of $k^\circ$.
Let $\mathcal E$ be the free $k^\circ$-module generated by $\{e_0,\ldots,e_r\}$. Then $\mathbb P(\mathcal E)$ is a projective flat $k^\circ$-scheme such that
\[\mathbb P(\mathcal E)\times_{\operatorname{Spec} k^\circ}\operatorname{Spec} k\cong\mathbb P(E).\]
Note that the fibre product
\[\mathbb P(\mathcal E)\times_{\operatorname{Spec} k^\circ}\operatorname{Spec}\kappa\]
is isomorphic to $\mathbb P(\mathcal E\otimes_{k^\circ}\kappa)$, which is an integral $\kappa$-scheme. Therefore, one has (see Example \ref{Exe:chambert-loir measure})
\[c_1(\overline{\mathcal O_E(1)})^r=\deg(c_1(\mathcal O_{\mathcal E_{\kappa}}(1))^r\cap\mathbb P(\mathcal E_{\kappa}))\operatorname{Dirac}_\xi=\operatorname{Dirac}_\xi.\]
\end{proof}
\begin{rema} \label{Rem: mesure complexe} Assume that $k=\mathbb C$ and $|\raisebox{.4ex}{.}|$ is the usual absolute value. Let $(E,\|\raisebox{.4ex}{.}\|)$ be a Hermitian space and \[\mathbb{S}(E^\vee,\|\raisebox{.4ex}{.}\|_*)=\{\alpha\in E^\vee\,|\,\|\alpha\|_*=1\}\] be the unit sphere in $E$, where $\|\raisebox{.4ex}{.}\|_*$ denotes the dual norm of $\|\raisebox{.4ex}{.}\|$, which is also a Hermitian norm. Note that $\mathbb P(E)^{\mathrm{an}}$ identifies with the quotient of $\mathbb S(E^\vee,\|\raisebox{.4ex}{.}\|_*)$ by the action of the unit sphere $\mathbb S(\mathbb C)=\{z\in\mathbb C\,|\,|z|=1\}$ in $\mathbb C$. We equip the universal invertible sheaf $\mathcal O_E(1)$ with the orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|$ and equip $\mathbb S(E^\vee,\|\raisebox{.4ex}{.}\|_*)$ with the unique $U(E^\vee,\|\raisebox{.4ex}{.}\|_*)$-invariant Borel probability measure
$\eta_{\mathbb S(E^\vee,\|\raisebox{.4ex}{.}\|_*)}$ which is locally equivalent to Lebesgue measure. Then the measure \[c_1(\overline{\mathcal O_E(1)})^{\dim_{\mathbb C}(E)-1}\] identifies with the direct image of $\eta_{\mathbb S(E^\vee,\|\raisebox{.4ex}{.}\|_*)}$ by the projection map from $\mathbb S(E^\vee,\|\raisebox{.4ex}{.}\|_*)$ to $\mathbb P(E)^{\mathrm{an}}$ (see for example \cite[(1.4.7)]{MR1260106} for more details).
\end{rema}
\begin{theo}\label{thm:integrability:Green:func}
Let $\overline{L} = (L, \varphi),\overline{L}_1 = (L_1, \varphi_1), \ldots, \overline{L}_d = (L_d, \varphi_d)$
be integrable metrized invertible ${\mathscr{O}}_X$-modules.
Let $s$ be a regular meromorphic section of $L$.
Then $g =-\log |s|_\varphi$ is integrable with respect to $\mu_{\overline{L}_1 \cdots \overline{L}_d}$.
\end{theo}
\begin{proof}
The proof of this theorem is same as \cite[Th\'eor\`eme~4.1]{MR2244803}.
We prove it without using the local intersection numbers.
Clearly we may assume that $X$ is integral, $L, L_1, \ldots, L_d$ are ample and
$\overline{L}, \overline{L}_1, \ldots, \overline{L}_d$ are semi-positive.
Let $\mathcal I$ be the ideal sheaf of ${\mathscr{O}}_X$ given by
\[\mathcal I_x = \{ a \in {\mathscr{O}}_{X,x} \mid a s_x \in L_x \}.\]
Choose a positive number $m$ and a non-zero section $t_1 \in H^0(X, \mathcal{I} L^{\otimes m}) \setminus \{ 0 \}$.
If we set $t_2 = t_1 \otimes s$,
then $s = t_2 \otimes t_1^{-1}$ and $t_2 \in H^0(X, L^{\otimes m+1}) \setminus \{ 0 \}$ and
$g = -\log |t_2|_{(m+1)\varphi} + \log |t_1|_{m \varphi}$, so that we may assume that $s \in H^0(X, L)\setminus \{ 0 \}$.
Let $\varphi'$ be a metric of $L$ such that either
(a) if $v$ is Archimedean, $\varphi'$ is $C^{\infty}$ and semi-positive, or
(b) if $v$ is non-Archimedean, $\varphi'$ is a nef model metric.
Then $-\log |s|_{\varphi} + \log |s|_{\varphi'}$ is a continuous function, so that
we may assume that $\varphi = \varphi'$.
By Theorem~\ref{thm:approximation:smooth:funcs},
there is a sequence $\{ \psi_n \}_{n\in\mathbb N}$ of smooth functions on $X^{\operatorname{an}}$ with the following properties:
\begin{enumerate}[label=\rm(\arabic*)]
\item for all $n \in {\mathbb{Z}_{\geqslant 1}}$, $\psi_n \leqslant g$, $\psi_n \leqslant \psi_{n+1}$.
\item for each point $x \in X^{\operatorname{an}}$, $\sup \{ \psi_n(x) ; n \in {\mathbb{Z}_{\geqslant 1}} \} = g(x)$.
\item for all $n \in {\mathbb{Z}_{\geqslant 1}}$, $g - \psi_n$ is a Green function of $D$ such that either
\begin{enumerate}[label=\rm(\arabic{enumi}.\alph*)]
\item if $v$ is Archimedean, the metric of $|\raisebox{.4ex}{.}|_{g-\psi_n}$ of $L$ is $C^{\infty}$ and semi-positive, or
\item if $v$ is non-Archimedean, the metric of $|\raisebox{.4ex}{.}|_{g-\psi_n}$ of $L$ is a nef model metric.
\end{enumerate}
\end{enumerate}
We prove the assertion by induction on the number
\[
e := \operatorname{Card}\{ i \in \{ 1, \ldots, d \} \mid \text{$\varphi_i$ is not smooth} \}.
\]
If $e = 0$, that is, $\varphi_i$ is smooth for all $i$, then the assertion is obvious.
We assume that $e > 0$. Obviously we may assume that $\varphi_1$ is not smooth.
Let $\varphi'_1$ be a semi-positive and smooth metric of $L_1$.
If we choose a continuous function $\vartheta$ such that $|\raisebox{.4ex}{.}|_{\varphi_1} = \exp(-\vartheta)|\raisebox{.4ex}{.}|_{\varphi'_1}$, then $c_1(\overline{L}_1) = c_1(\overline{L}'_1) + dd^c(\vartheta)$, where
$\overline{L}'_1 = (L_1,\varphi'_1)$.
Let us consider the following integral:
\[
I_n := \int_{X^{\operatorname{an}}} \psi_n c_1(\overline{L}_1) \cdots c_1(\overline{L}_d).
\]
Note that $\psi_n$ and $\vartheta$ are locally written by differences of plurisubharmonic functions, so that,
by \cite[Proposition~2.3]{MR2244803},
\begin{align*}
I_n & = \int_{X^{\operatorname{an}}} \psi_n c_1(\overline{L}'_1) \cdots c_1(\overline{L}_d) + \int_{X^{\operatorname{an}}} \psi_n dd^c(\vartheta) c_1(\overline{L}_2)\cdots c_1(\overline{L}_d) \\
& = \int_{X^{\operatorname{an}}} \psi_n c_1(\overline{L}'_1) \cdots c_1(\overline{L}_d) + \int_{X^{\operatorname{an}}} \vartheta dd^c(\psi_n) c_1(\overline{L}_2) \cdots c_1(\overline{L}_d).
\end{align*}
By the hypothesis of induction,
\[
\lim_{n\to\infty} \int_{X^{\operatorname{an}}} \psi_n c_1(\overline{L}'_1) \cdots c_1(\overline{L}_d)
\]
exists. Moreover, by the same arguments as the last part of \cite[Th\'eor\`em~4.1]{MR2244803},
one can see
\begin{multline*}
\lim_{n\to\infty}\int_{X^{\operatorname{an}}} \vartheta dd^c(\psi_n) c_1(\overline{L}_2) \cdots c_1(\overline{L}_d) \\
= \int_{X^{\operatorname{an}}} \vartheta c_1(\overline{L}) c_1(\overline{L}_2) \cdots c_1(\overline{L}_d) -
\int_{\operatorname{div}(s)^{\operatorname{an}}} \vartheta c_1(\overline{L}_2) \cdots c_1(\overline{L}_d).
\end{multline*}
Therefore $\lim_{n\to\infty} I_n$ exists, as required.
\end{proof}
\section{Local intersection number over an algebraically closed field}
Let $k$ be an algebraically closed field equipped with a non-trivial absolute value $|\raisebox{.4ex}{.}|$ such that $k$ is complete with respect to the topology defined by $|\raisebox{.4ex}{.}|$.
The pair $(k, |\raisebox{.4ex}{.}|)$ is denoted by $v$.
Let $X$ be a projective scheme over $k$ and $d$ be its dimension.
Recall that any element $x$ of $X^{\mathrm{an}}$ consists of a scheme point of $X$ and an absolute value $|\raisebox{.4ex}{.}|_x$ of the residue field of the scheme point. We denote by $\widehat{\kappa}(x)$ the completion of the residue field of the scheme point with respect to the absolute value $|\raisebox{.4ex}{.}|_x$, on which the absolute value extends by continuity.
\begin{defi}\label{def:local:intersection}
Let $(D_0, g_0), \ldots, (D_d, g_d)$ be integrable metrized Cartier divisors on $X$.
We assume that $D_0, \ldots, D_d$ intersect properly, that is, $(D_0, \ldots, D_d) \in {\mathcal{IP}}_X$ (see Definition~\ref{def:meet:properly}).
According to \cite{MR2244803}, we define \emph{the local intersection number $\big((D_0, g_0) \cdots (D_d, g_d)\big)_v$ at $v$} as follows.
In the case where $d = 0$, one has $X=\operatorname{Spec}(A)$ for some $k$-algebra with $\dim_k(A) < \infty$.
By Remark~\ref{remark:Cartier:div:on:0:dim} and Example~\ref{Rem: Green function of several points},
\[A = \bigoplus_{x \in \operatorname{Spec}(A)} A_x\;\text{ and }\;(D_0, g_0) = \sum_{x \in \operatorname{Spec}(A)} (0, a_x),\]
where $a_x \in \mathbb{R}$ for all $x \in \operatorname{Spec}(A)$.
Then
\begin{equation}\label{eqn:def:local:intersection:dim:0}
\big((D_0, g_0)\big)_v := \sum_{x \in \operatorname{Spec}(A)} \operatorname{length}_{A_x}(A_x)\,a_x.
\end{equation}
Note that $\operatorname{length}_{A_x}(A_x) = \dim_k (A_x)$ because $k$ is algebraically closed.
If $d > 0$ and $\sum_{i=1}^n a_i Z_i$ is the cycle associated with $D_d$ (cf. Remark~\ref{remark:expansion:Cartier:div:as:cycle}), then the local intersection number $\big((D_0, g_0) \cdots (D_d, g_d)\big)_v$ is defined in a recursive way with respect to $d=\dim(X)$ as
\begin{equation}\label{eqn:def:local:intersection}
\sum_{i=1}^n a_i \Big(\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\Big)_v + \int_{X^{\operatorname{an}}} g_{d}(x)\, \mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}(\mathrm{d}x).
\end{equation}
For the integrability of $g_d$ with respect to the measure $\mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}$, see Theorem~\ref{thm:integrability:Green:func}.
\end{defi}
\begin{prop}\label{prop:formula:local:intersection:irreducible:components}
Let $X_1, \ldots, X_\ell$ be irreducible components of $X$ and $\eta_1, \ldots, \eta_\ell$ be the generic points of $X_1, \ldots, X_\ell$, respectively.
Then
\[
\big((D_0, g_0) \cdots (D_d, g_d)\big)_v = \sum_{j=1}^{\ell} \operatorname{length}_{\mathcal O_{X, \eta_j}}(\mathcal O_{X, \eta_j}) \big(\rest{(D_0, g_0)}{X_j} \cdots \rest{(D_d, g_d)}{X_j}\big)_v.
\]
\end{prop}
\begin{proof}
In the case where $d = 0$, the assertion is obvious.
We assume that $d > 0$. By the definition of $\mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}$ (cf. Section~\ref{sec:local:measures}), if we set
\[b_j = \operatorname{length}_{\mathcal O_{X, \eta_j}}(\mathcal O_{X, \eta_j}),\] then one has
\[
\int_{X^{\mathrm{an}}} g_d(x)\, \mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}(\mathrm{d}x) \\
= \sum_{j=1}^{\ell} b_j \int_{X_j^{\mathrm{an}}}
g_d(x)\, \mu_{\rest{(D_0, g_0)}{X_j} \cdots \rest{(D_{d-1}, g_{d-1})}{X_j}}(\mathrm{d}x).
\]
If $\sum_{i=1}^n a_i Z_i$ and $\sum_{i=1}^n a_{ji} Z_i$ are the cycles associated with $D_d$ and $\rest{D_d}{X_j}$, respectively,
then, by \eqref{eqn:remark:expansion:Cartier:div:as:cycle:01},
$a_i = \sum_{j=1}^{\ell} b_j a_{ji}$, so that
\begin{multline*}
\sum_{i=1}^n a_i \Big(\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\Big)_v \\
= \sum_{i=1}^n \sum_{j=1}^{\ell} b_j a_{ji} \Big(\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\Big)_v \\
= \sum_{j=1}^{\ell} b_j \sum_{i=1}^n a_{ji}\Big(\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\Big)_v.
\end{multline*}
Therefore, since
\begin{multline*}
\big(\rest{(D_0, g_0)}{X_j} \cdots \rest{(D_d, g_d)}{X_j}\big)_v = \sum_{i=1}^n a_{ji}\Big(\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\Big)_v \\
+ \int_{X_j^{\mathrm{an}}}
g_d(x)\,\mu_{\rest{(D_0, g_0)}{X_j} \cdots \rest{(D_{d-1}, g_{d-1})}{X_j}}(\mathrm{d}x),
\end{multline*}
one has the desired formula.
\end{proof}
\begin{prop}\label{prop:multilinear:symmetric:semiample:case}
Let $(D_0, g_0) \ldots, (D_i, g_i), (D'_i, g'_i), \ldots, (D_d, g_d)$ be integrable metrized Cartier divisors on $X$ such that
$(D_0, \ldots, D_i, \ldots, D_d)$ and $(D_0, \ldots, D'_i, \ldots, D_d)$ belong to ${\mathcal{IP}}_X$.
Then one has the following:
\begin{enumerate}[label=\rm(\arabic*)]
\item
The local intersection pairing is multi-linear, that is,
\[
\kern2em\begin{cases}
\big( (D_0, g_0) \cdots (D_i + D'_i, g_i + g'_i) \cdots (D_d, g_d) \big)_v \\
\kern0.7em = \big( (D_0, g_0) \cdots (D_i, g_i) \cdots (D_d, g_d) \big)_v +
\big( (D_0, g_0) \cdots (D'_i, g'_i) \cdots (D_d, g_d) \big)_v. \\[2ex]
\big( (D_0, g_0) \cdots (-D_i, -g_i) \cdots (D_d, g_d) \big)_v = - \big( (D_0, g_0) \cdots (D_i, g_i) \cdots (D_d, g_d) \big)_v.
\end{cases}
\]
\item
We assume that $D_0, \ldots, D_d$ are ample and $g_0, \ldots, g_d$ are plurisubharmonic.
For each $i$, let $\{ g_{i,n} \}_{n=1}^{\infty}$ be a sequence of plurisubharmonic Green functions of $D_i$
such that $\lim_{n\to\infty} \| g_i - g_{i,n}\|_{\sup} = 0$. Then
\[
\lim_{n\to\infty} \big( (D_0, g_{0,n}) \cdots (D_{d}, g_{d,n}) \big)_v = \big( (D_0, g_0) \cdots (D_d, g_d) \big)_v
\]
\item
The local intersection pairing is symmetric, that is, for any bijection $\sigma:\{0,\ldots,d\}\rightarrow\{0,\ldots,d\}$ one has
\[
\big( (D_{\sigma(0)}, g_{\sigma(0)}) \cdots \cdots (D_{\sigma(d)}, g_{\sigma(d)}) \big)_v
= \big( (D_0, g_0) \cdots (D_d, g_d) \big)_v.\]\end{enumerate}
\end{prop}
\begin{proof}
Clearly we may assume that $X$ is integral.
We prove (1), (2) and (3) by induction on $d$. In the case $d=0$, the assertion is obvious, so that we assume $d > 0$.
\medskip
(1) If $0 \leqslant i < d$, the assertions follow from the hypothesis of induction and
the multi-linearity of the measure $\mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}$ with respect to $(D_0, g_0)$, $\ldots$, $(D_{d-1}, g_{d-1})$, so that we may assume that $i=d$.
Let $D_d = a_1 Z_1 + \cdots + a_n Z_n$ and $D'_d = a'_1 Z_1 + \cdots + a'_n Z_n$ be
the decompositions of $D_d$ and $D'_{d}$ as cycles.
Then $D_d + D'_d = (a_1 + a'_1) Z_1 + \cdots + (a_n + a'_n) Z_n$ and
$-D_d = (-a_1)Z_1 + \cdots + (-a_n)Z_n$, so that the assertions are obvious.
\medskip
(2) By \eqref{eqn:def:local:intersection} and the hypothesis of induction,
it is sufficient to see
\[
\lim_{n\to\infty} \int_{X^{\operatorname{an}}} g_{d,n} \mu_{(D_0, g_{0,n}) \cdots (D_{d-1}, g_{d-1, n})} =
\int_{X^{\operatorname{an}}} g_{d} \mu_{(D_0, g_{0}) \cdots (D_{d-1}, g_{d-1})},
\]
which follows from \cite[Corollary~(3.6)]{Demagbook} and \cite[Corollaire~(5.6.5)]{ChamDucros}.
\medskip
(3) We may assume that $D_0, \ldots, D_d$ are ample and $g_0, \ldots, g_d$ are plurisubharmonic.
By (2) together with regularizations of metrics, we may further assume that metrics $|\raisebox{.4ex}{.}|_{g_0}, \ldots, |\raisebox{.4ex}{.}|_{g_d}$ are smooth. It suffices to prove the assertion in the particular case where $\sigma$ is a transposition exchanging two indices $i$ and $j$ with $i<j$. If $j < d$, then the assertion follows from the hypothesis of induction,
so that we may assume that $j = d$. If $i < d-1$, then
\begin{multline*}
\big( (D_0, g_0) \cdots (D_i, g_i) \cdots (D_{d-1}, g_{d-1}) \cdot (D_d, g_d) \big)_v \\
= \big( (D_0, g_0) \cdots (D_{d-1}, g_{d-1}) \cdots (D_i, g_i) \cdot (D_d, g_d) \big)_v.\end{multline*}
by the hypothesis of induction. Therefore we may assume that $i = d-1$.
Let $D_d = a_1 Z_1 + \cdots + a_n Z_n$ and $\rest{D_{d-1}}{Z_i} = a_{i1} Z_{i1} + \cdots + a_{in} Z_{in}$
be the decomposition as cycles. Then
\begin{multline*}
\big( (D_0, g_0) \cdots (D_{d-1}, g_{d-1}) \cdot (D_d, g_d) \big)_v \\
= \sum_{i,j} a_i a_{ij} \big(\kern-.3em\rest{(D_0, g_0)}{Z_{ij}} \cdots \rest{(D_{d-2}, g_{d-2})}{Z_{ij}}\big)_v \\
\kern7em+ \sum_{i} a_i \int_{Z_i^{\operatorname{an}}} g_{d-1}(x)\,\mu_{\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-2}, g_{d-2})}{Z_i}}(\mathrm{d}x) \\
+ \int_{X^{\operatorname{an}}} g_d(x)\, \mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}(\mathrm{d}x).
\end{multline*}
In the same way, if $D_{d-1} = a'_1 Z'_1 + \cdots + a'_n Z'_n$ and $\rest{D_{d}}{Z'_i} = a'_{i1} Z'_{i1} + \cdots + a'_{in} Z'_{in}$
be the decomposition as cycles, then
\begin{multline*}
\big( (D_0, g_0) \cdots (D_{d}, g_{d}) \cdot (D_{d-1}, g_{d-1}) \big)_v \\
= \sum_{i,j} a'_i a'_{ij} \big(\kern-.3em\rest{(D_0, g_0)}{Z'_{ij}} \cdots \rest{(D_{d-2}, g_{d-2})}{Z'_{ij}} \big)_v \\
\kern7em + \sum_{i} a'_i \int_{(Z'_i)^{\operatorname{an}}} g_{d}(x)\, \mu_{\rest{(D_0, g_0)}{Z'_i} \cdots \rest{(D_{d-2}, g_{d-2})}{Z'_i}} (\mathrm{d}x)\\
+ \int_{X^{\operatorname{an}}} g_{d-1}(x)\, \mu_{(D_0, g_0) \cdots (D_{d-2}, g_{d-2}) \cdot (D_{d}, g_{d})}(\mathrm{d}x).
\end{multline*}
By \cite[Proposition~5.2 (2)]{MArakelov}, one has $\sum_{ij} a_i a_{ij}Z_{ij} = \sum_{ij} a'_i a'_{ij}Z'_{ij}$ as cycles, so that it is sufficient to show that
\begin{multline*}
\sum_{i} a_i \int_{Z_i^{\operatorname{an}}} g_{d-1}(x)\,\mu_{\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-2}, g_{d-2})}{Z_i}}(\mathrm{d}x)
+ \int_{X^{\operatorname{an}}} g_d(x)\, \mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}(\mathrm{d}x) \\
= \sum_{i} a'_i \int_{(Z'_i)^{\operatorname{an}}} g_{d}(x)\,\mu_{\rest{(D_0, g_0)}{Z'_i} \cdots \rest{(D_{d-2}, g_{d-2})}{Z'_i}}(\mathrm{d}x) \\
+ \int_{X^{\operatorname{an}}} g_{d-1}(x)\, \mu_{(D_0, g_0) \cdots (D_{d-2}, g_{d-2}) \cdot (D_{d}, g_{d})}(\mathrm{d}x),
\end{multline*}
which is nothing more than \cite[Theorem~5.6]{MArakelov} for the Archimedean case and \cite[Proposition~11.5]{GKTropical} for the non-Archimedean case.
\end{proof}
\begin{prop}\label{prop:intersection:finite:morphism}
Let $\pi : Y \to X$ be a surjective morphism of integral projective schemes over $k$.
We set $e = \dim X$ and $d = \dim Y$.
Let $(D_0, g_0), \ldots, (D_d, g_d)$ be integrable metrized Cartier divisors on $X$ such that
$(\pi^*(D_0), \ldots, \pi^*(D_d)) \in {\mathcal{IP}}_Y$.
Then one has the following:
\begin{enumerate}[label=\rm(\arabic*)]
\item If $d > e$, then $(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d))_v = 0$.
\item If $d = e$ and $(D_0,\ldots,D_d) \in {\mathcal{IP}}_X$, then
\[(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d))_v = (\deg \pi)((D_0, g_0) \cdots (D_d, g_d))_v.\]
\end{enumerate}
\end{prop}
\begin{proof}
We prove (1) and (2) by induction on $e$.
If $e = 0$, then (2) is obvious. For (1), as $\pi^*(D_0, g_0) = (0, a_0), \ldots, \pi^*(D_d, g_d) = (0, a_d)$
for some $a_0, \ldots, a_d \in {\mathbb{R}}$, then
\[
(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d))_v = \int_{X^{\operatorname{an}}} a_d \mu_{(0, a_0)\cdots(0, a_d)} = 0,
\]
as desired.
We assume $e > 0$.
Let $D_d = a_1 Z_1 + \cdots + a_n Z_n$ and $\pi^*(D_d) = b_{1} Z'_{1} + \cdots + b_{N}Z'_{N}$
be the decompositions as cycles.
By \eqref{eqn:def:local:intersection},
\begin{multline*}
\big(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d)\big)_v = \sum_{j=1}^N b_j \big(\kern-.3em\rest{\pi^*(D_0, g_0)}{Z'_j} \cdots \rest{\pi^*(D_{d-1}, g_{d-1})}{Z'_j}\big)_v \\
+ \int_{Y^{\operatorname{an}}} g_{d}(\pi^{\mathrm{an}}(y)) \,\mu_{\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d)}(\mathrm{d}y),
\end{multline*}
Note that if $e < d$, then $\dim \pi(Z'_j) < \dim Z'_j$
and $\pi_*(\mu_{\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d)}) = 0$ by Proposition~\ref{prop:CH:measure:projection:formula}, so that one has (1).
Next we assume that $e = d$.
For each $i$, we set $J_i = \{ j \in \{ 1, \ldots, N \} \mid \pi(Z'_j) = Z_i \}$.
We set $J_0 = \{ 1, \ldots, N \} \setminus (J_1 \cup \cdots \cup J_n)$. By the hypothesis of induction for (1),
$\big(\kern-.3em\rest{\pi^*(D_0, g_0)}{Z'_j} \cdots \rest{\pi^*(D_{d-1}, g_{d-1})}{Z'_j}\big)_v = 0$ for all
$j \in J_0$, so that, by the hypothesis of induction for (2) and Proposition~\ref{prop:CH:measure:projection:formula}, the above equation implies
\begin{align*}
\big(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d)\big)_v \\
& \kern-5em = \sum_{i=1}^n \sum_{j \in J_i} b_j \big(\kern-.3em\rest{\pi^*(D_0, g_0)}{Z'_j} \cdots \rest{\pi^*(D_{d-1}, g_{d-1})}{Z'_j}\big)_v \\
& + \int_{Y^{\operatorname{an}}} g_{d}(\pi^{\mathrm{an}}(y)) \,\mu_{\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d)}(\mathrm{d}y) \\
& \kern-5em = \sum_{i=1}^n \big(\kern-.3em\rest{(D_0, g_0)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\big)_v \sum_{j \in J_i} b_j \deg(\rest{\pi}{Z'_j}) \\
& + \deg(\pi) \int_{X^{\operatorname{an}}} g_{d}(x)\, \mu_{(D_0, g_0) \cdots (D_d, g_d)}(\mathrm{d}x).
\end{align*}
Therefore, the assertion follows because $\sum_{j \in J_i} b_j \deg(\rest{\pi}{Z'_j}) = \deg(\pi)a_i$ (cf. \cite[Lemma~1.12]{MArakelov}).
\end{proof}
\begin{prop}\label{prop:intersection:principal:div}
Let $f$ be a regular meromorphic function on $X$ and $(D_1, g_1), \ldots, (D_d, g_d)$
be integrable metrized Cartier divisors on $X$ such that $(\operatorname{div}(f), D_1, \ldots, D_d) \in {\mathcal{IP}}_X$.
If we set $D_1 \cdots D_{d} = \sum_{x \in X_{(0)}} a_x x$ as cycle, then
\begin{equation} \label{Equ: intersection principal divisor}\big(\widehat{\operatorname{div}}(f) \cdot (D_1, g_1) \cdots (D_{d}, g_{d})\big)_v = \sum_{x \in X_{(0)}} a_x \big(-\log |f|(x^{\operatorname{an}})\big),\end{equation}
where $X_{(0)}$ is the set of all closed point of $X$ and $x^{\operatorname{an}}$ is the associated absolute value at $x$.
Note that in the case where $\dim(X) = 0$, the above formula means that
\[\big(\widehat{\operatorname{div}}(f) \big)_v = 0.\]
\end{prop}
\begin{proof}
Let $X = a_1 X_1 + \cdots + a_n X_n$ be the decomposition as cycles.
Then
\begin{equation*} \big(\widehat{\operatorname{div}}(f) \cdot (D_1, g_1) \cdots (D_{d}, g_{d})\big)_v
=
\sum_{i=1}^n a_i \big({\widehat{\operatorname{div}}(f)}|_{X_i} \cdot \rest{(D_1, g_1)}{X_i} \cdots \rest{(D_{d}, g_{d})}{X_i}\big)_v
\end{equation*}
and
\[
D_1 \cdots D_{d} = \sum_{i=1}^n a_i \big(\rest{D_1}{X_i} \cdots\rest{D_d}{X_i}\big),
\]
so that we may assume that $X$ is integral.
We prove the equality \eqref{Equ: intersection principal divisor} by induction on $d=\dim(X)$. In the case where $\dim(X) = 0$, the assertion is obvious
because $f$ is a unit.
We assume that $\dim(X) \geqslant 1$. Let $D_d = a_1 Z_1 + \cdots + a_n Z_n$
be the decomposition as cycles. Let $\sum_{x \in X_{(0)}} b_{ix} x$ be the decomposition of
$\rest{D_1}{Z_i} \cdots \rest{D_{d-1}}{Z_i} = D_1 \cdots D_{d-1} \cdot Z_i$ as cycles.
Then
\[
\sum_{i=1}^n a_i \sum_{x \in X_{(0)}} b_{ix} x = \sum_{x \in X_{(0)}} a_x x,
\]
so that $a_x = \sum_{i=1}^n a_i b_{ix}$. On the other hand, by hypothesis of induction,
\[
\big(\widehat{\operatorname{div}}(f)|_{Z_i} \cdot \rest{(D_1, g_1)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\big)_v = \sum_{x \in X_{(0)}} b_{ix} (-\log |f|(x^{\operatorname{an}})).
\]
Therefore,
\begin{multline*}
\sum_{x \in X_{(0)}} a_x (-\log |f|(x^{\operatorname{an}})) \\
= \sum_{x \in X_{(0)}} \Big( \sum_{i=1}^n a_i b_{ix} \Big) (-\log |f|(x^{\operatorname{an}}))
=\sum_{i=1}^n a_i \sum_{x \in X} b_{ix} (-\log |f|(x^{\operatorname{an}})) \\
= \sum_{i=1}^n a_i \big(\widehat{\operatorname{div}}(f)|_{Z_i} \cdot \rest{(D_1, g_1)}{Z_i} \cdots \rest{(D_{d-1}, g_{d-1})}{Z_i}\big)_v.
\end{multline*}
Note that $\mu_{(\widehat{\operatorname{div}}(f) \cdot (D_1, g_1) \cdots (D_d, g_d))} = 0$, and hence the assertion follows by \eqref{eqn:def:local:intersection}.
\end{proof}
\begin{prop}\label{prop:intersection:trivial:divisor}
Let $(D_0, g_0), \ldots, (D_{d-1}, g_{d-1}), (0,g)$ be integrable metrized Cartier divisors on $X$ with $(D_0, \ldots, D_{d-1}, 0) \in {\mathcal{IP}}_X$.
We assume that $D_0, \ldots, D_{d-1}$ are semiample and $g_0, \ldots, g_{d-1}$ are plurisubharmonic.
Then
\[
\big((D_0, g_0) \cdots (D_{d-1}, g_{d-1}) \cdot (0, g)\big)_v = \int_{X^{\operatorname{an}}} g(x) \mu_{(D_0, g_0) \cdots (D_{d-1}, g_{d-1})}(\mathrm{d}x).
\]
In particular,
\begin{multline*}
\min \{ g(x) \mid x \in X^{\operatorname{an}} \} (D_0 \cdots D_{d-1}) \\\
\leqslant
\big((D_0, g_0) \cdots (D_{d-1}, g_{d-1}) \cdot (0, g)\big)_v \\
\leqslant \max \{ g(x) \mid x \in X^{\operatorname{an}} \} (D_0 \cdots D_{d-1}).
\end{multline*}
\end{prop}
\begin{proof}
This is trivial by the definition.
\end{proof}
\begin{coro}\label{cor:intersection:trivial:divisor}
Let $(D_0, g_0), \ldots, (D_d, g_d)$ be integrable arithmetic Cartier divisors on $X$ with $(D_0, \ldots, D_d) \in {\mathcal{IP}}_X$.
We assume that $D_0, \ldots, D_d$ are semiample and $g_0, \ldots, g_d$ are plurisubmarmonic.
Let $g'_0, \ldots, g'_d$ be another plurisubharmonic Green functions of $D_0, \ldots, D_d$, respectively.
Then one has
\begin{multline*}
\left| \big((D_0, g'_0) \cdots (D_d, g'_d)\big)_{v} - \big((D_0, g_0) \cdots (D_d, g_d)\big)_{v} \right| \\
\leqslant
\sum_{i=0}^d \max \{|g'_i - g_i|(x) \mid x \in X^{\operatorname{an}} \} (D_0 \cdots D_{i-1} \cdot D_{i+1} \cdots D_d).
\end{multline*}
\end{coro}
\begin{proof}
By using Proposition~\ref{prop:multilinear:symmetric:semiample:case},
\begin{multline*}
\big((D_0, g'_0) \cdots (D_d, g'_d)\big) - \big((D_0, g_0) \cdots (D_d, g_d)\big) \\
=
\sum_{i=0}^d ((D_0, g_0) \cdots (D_{i-1}, g_{i-1}) \cdot (0, g'_i - g_i) \cdot (D_{i+1}, g'_{i+1}) \cdots (D_d, g'_d) \big),
\end{multline*}
so that the assertion follows from Proposition~\ref{prop:intersection:trivial:divisor}.
\end{proof}
\begin{prop}\label{prop:local:intersection:PP}
We assume that $X = {\mathbb{P}}^d_k$ and $L = {\mathscr{O}}_{{\mathbb{P}}^d}(1)$. Let $\{ T_0, \ldots, T_d \}$ be a basis of $H^0({\mathbb{P}}^d_k, {\mathscr{O}}_{{\mathbb{P}}^d}(1))$ over $k$. We view $(T_0 : \cdots : T_d)$ as a homogeneous coordinate of
${\mathbb{P}}^d_k$.
Let $\|\raisebox{.4ex}{.}\|$ be a norm of $H^0({\mathbb{P}}^d_k, {\mathscr{O}}_{{\mathbb{P}}^d}(1))$ given by
\[
\| a_0 T_0 + \cdots + a_d T_d \| = \begin{cases}
\sqrt{|a_0|^2 + \cdots + |a_d|^2} & \text{if $v$ is Archimedean},\\[2ex]
\max \{ |a_0|, \ldots, |a_d| \} & \text{if $v$ is non-Archimedean}.
\end{cases}
\]
Let $\varphi$
be the orthogonal quotient metric of ${\mathscr{O}}_{{\mathbb{P}}^d}(1)$ given by the surjective homomorphism
$H^0({\mathbb{P}}^d_k, {\mathscr{O}}_{{\mathbb{P}}^d}(1)) \otimes {\mathscr{O}}_{{\mathbb{P}}^d} \to {\mathscr{O}}_{{\mathbb{P}}^d}(1)$ and the above norm $\|\raisebox{.4ex}{.}\|$.
We set $H_i = \{ T_i = 0 \}$ and $h_i = -\log |T_i|_{\varphi}$. Then
\[
((H_0, h_0) \cdots (H_d, h_d))_v = \begin{cases}
\widehat{\operatorname{deg}}\Big(\widehat{c}_1({\mathscr{O}}_{{\mathbb{P}}^d_{{\mathbb{Z}}}}(1), \varphi)^{d+1}\Big) & \text{if $v$ is Archimedean},\\[2ex]
0 & \text{if $v$ is non-Archimedean},
\end{cases}
\]
where $\widehat{\operatorname{deg}}\Big(\widehat{c}_1({\mathscr{O}}_{{\mathbb{P}}^d_{{\mathbb{Z}}}}(1), \varphi)^{d+1}\Big)$ is the self-intersection number of
the arithmetic first Chern class $\widehat{c}_1({\mathscr{O}}_{{\mathbb{P}}^d_{{\mathbb{Z}}}}(1), \varphi)$ on
the $d$-dimensional projective space ${\mathbb{P}}^d_{{\mathbb{Z}}}$ over ${\mathbb{Z}}$.
\end{prop}
\begin{proof}
If we set
\[
a_m := \int_{{\mathbb{P}}^m_k} -\log |T_m|_{\varphi}(x)\,\mu_{({\mathscr{O}}_{{\mathbb{P}}^m}(1), \varphi_{\rm{FS}})^m}(\mathrm{d}x)
\]
for a positive integer $m$, then
\[
((H_0, h_0) \cdots (H_d, h_d))_v = \sum_{m=1}^d a_m.
\]
In the following, we set $x_i = T_i/T_0$.
\medskip
{\bf $\bullet$ Archimedean case} :\quad
The algorithms of the calculation is exactly same as one on ${\mathbb{P}}^d_{{\mathbb{Z}}}$, so that we have the assertion.
\bigskip
{\bf $\bullet$ non-Archimedean case} :\quad
If we set
$|f|_* = \max_{i_1, \ldots, i_m} \{ |c_{i_1, \ldots, i_m}| \}$ for \[f = \sum_{i_1,\ldots,i_m} c_{i_1, \ldots, i_m} x_1^{i_1} \cdots x_m^{i_m} \in k[x_1, \ldots, x_m],\] then $|\raisebox{.4ex}{.}|_*$ extends to an absolute value of $k(x_1, \ldots, x_m)$ (cf. Lemma~\ref{lem:Gauss:lemma}). We set $U = \{ T_m \not= 0\}$. Note that if $\xi \in U^{\operatorname{an}}$, then
\[
|T_m|_{\varphi}(\xi) = \frac{|x_m|_\xi}{\max\{1, |x_1|_\xi, \ldots, |x_m|_\xi\}}.
\]
Let ${\mathfrak{o}}_v$ be the valuation ring of $v$. Note that $\varphi$ coincides with
the metric of the model $({\mathbb{P}}^d_{{\mathfrak{o}}_v}, {\mathscr{O}}_{{\mathbb{P}}^d_{{\mathfrak{o}}_v}}(1))$ by \cite[Proposition~2.3.12]{CMArakelovAdelic},
so that $\mu_{({\mathscr{O}}_{{\mathbb{P}}^m}(1), \varphi)^m} = \delta_{|\raisebox{.4ex}{.}|_*}$. Thus
\[
a_m = -\log \frac{|x_m|_*}{\max\{1, |x_1|_*, \ldots, |x_m|_*\}} = 0,
\]
and hence the assertion follows.
\end{proof}
\section{Local intersection number over a general field}\label{Sec: Local intersection number over a general field}
In this section, we consider the local intersection product and local height formula in the non-necessarily algebraically closed case. We fix in this section a complete valued field $v=(k,|\raisebox{.4ex}{.}|)$ such that $|\raisebox{.4ex}{.}|$ is not trivial. Let $\mathbb C_k$ be the completion of an algebraic closure of $k$. Note that the absolute value $|\raisebox{.4ex}{.}|$ extends naturally to $\mathbb C_k$ and the valued field $(\mathbb C_k,|\raisebox{.4ex}{.}|)$ is both algebraically closed and complete. We denote by $v^{\operatorname{ac}}$ the couple $(\mathbb C_k,|\raisebox{.4ex}{.}|)$. We also fix a projective morphism $\pi:X\rightarrow\operatorname{Spec} k$ and we denote by $X_{\mathbb C_k}$ the fiber product $X\times_{\operatorname{Spec} k}\operatorname{Spec} \mathbb C_k$. Let $d$ be the Krull dimension of $X$, which is also equal to the Krull dimension of $X_{\mathbb C_k}$.
\begin{defi}\label{Def: intersection pairing general case}
Let $(D_0,g_0),\ldots,(D_d,g_d)$ be a family of metrized Cartier divisor on $X$ such that $D_0,\ldots,D_d$ intersect properly and that $g_0,\ldots,g_d$ are integrable Green functions. By Remark \ref{Rem: extension of scalars}, the Cartier divisors $D_{0,\mathbb C_k},\ldots,D_{d,\mathbb C_k}$ intersect properly. Moreover, by Remark \ref{Rem: extension scalar Green function}, the Green functions $g_{0,\mathbb C_k},\ldots,g_{d,\mathbb C_k}$ are integrable. We then define the local intersection number of $(D_0,g_0),\ldots,(D_d,g_d)$ as
\[\big( (D_0, g_0) \cdots (D_d, g_d) \big)_v:=\big( (D_{0,\mathbb C_k}, g_{0,\mathbb C_k}) \cdots (D_{d,\mathbb C_k}, g_{d,\mathbb C_k}) \big)_{v^{\operatorname{ac}}}\]
\end{defi}
Several properties of the local intersection number follow directly from the results of the previous section. We gather them below.
\begin{rema}\label{rem:properties:local:intersection:general:field}
Recall that $\operatorname{\widehat{Int}}(X)$ denotes the group of integrable metrized Cartier divisors on $X$. Let $\widehat{{\mathcal{IP}}}_X$ be the subset of $\operatorname{\widehat{Int}}(X)^{d+1}$ consisting of elements
\[\big((D_0,g_0),\ldots,(D_d,g_d)\big)\]
such that the Cartier divisors $D_0,\ldots,D_d$ intersect properly.
\begin{enumerate}[label=\rm(\arabic*)]
\item The set $\widehat{{\mathcal{IP}}}_X$ forms a symmetric multi-linear subset of the group $\operatorname{\widehat{Int}}(X)^{d+1}$. Moreover, the function of local intersection number
\[\big( (D_0, g_0) \cdots (D_d, g_d) \big)\longmapsto\big( (D_0, g_0) \cdots (D_d, g_d) \big)_v\]
form a symmetric multi-linear map from $\widehat{{\mathcal{IP}}}_X$ to $\mathbb R$. These statements follow from Proposition \ref{prop:multilinear:symmetric:semiample:case}.
\item
Let $\pi : Y \to X$ be a surjective morphism of geometrically integral projective schemes over $k$.
We set $e = \dim X$ and $d = \dim Y$.
Let $(D_0, g_0), \ldots, (D_d, g_d)$ be integrable metrized Cartier divisors on $X$ such that
$(\pi^*(D_0), \ldots, \pi^*(D_d)) \in {\mathcal{IP}}_Y$.
Then one has the following:
\begin{enumerate}[label=\rm(\roman*)]
\item If $d > e$, then $(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d))_v = 0$.
\item If $d = e$ and $(D_0,\ldots,D_d) \in {\mathcal{IP}}_X$, then
\[(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d))_v = (\deg \pi)((D_0, g_0) \cdots (D_d, g_d))_v.\]
\end{enumerate}
We refer to Proposition \ref{prop:intersection:finite:morphism} for a proof.
\item Let $f$ be a regular meromorphic function on $X$ and $(D_1,g_1),\ldots,(D_d,g_d)$ be integrable metrized Cartier divisors on $X$ such that $(\operatorname{div}(f),D_1,\ldots,D_d)\in{\mathcal{IP}}_X$. Suppose that
\[D_1\cdots D_d=\sum_{x\in X_{(0)}}a_xx\]
as a cycle, then
\[\big(\widehat{\operatorname{div}}(f) \cdot (D_1, g_1) \cdots (D_{d}, g_{d})\big)_v = \sum_{x \in X_{(0)}} a_x[\kappa(x):k]_s \big(-\log |f|(x^{\operatorname{an}})\big),\]
where $[\kappa(x):k]_s$ denotes the separable degree of the residue field $\kappa(x)$ over $k$. We refer to Proposition \ref{prop:intersection:principal:div} for more details.
\item Let $\big((D_0,g_0),\ldots,(D_d,g_d)\big)$ be an element of $\widehat{{\mathcal{IP}}}_X$. We assume that $D_0,\ldots,D_{d-1}$ are semi-ample, $g_0,\ldots,g_{d-1}$ are plurisubharmonic, and $D_{\textcolor{mred}{d}}=0$. Then one has
\[\delta\min_{x\in X^{\mathrm{an}}}g_d(x)\leqslant \big((D_0,g_0),\ldots,(D_d,g_d)\big)_v\leqslant\delta\max_{x\in X^{\mathrm{an}}}g_d(x),\]
where $\delta=(D_0,\ldots,D_{d-1})$. See Proposition \ref{prop:intersection:trivial:divisor} for more details.
\item Let $\big((D_0,g_0),\ldots,(D_d,g_d)\big)$ and $\big((D_0,g_0'),\ldots,(D_d,g_d')\big)$ be two elements of $\widehat{{\mathcal{IP}}}_X$ having the same family of underlying Cartier divisors. One has
\[\begin{split}\Big|\big((D_0,g_0),\ldots,(D_d,g_d)\big)_v&-\big((D_0,g_0'),\ldots,(D_d,g_d')\big)_v\Big|\\&\quad\leqslant\sum_{i=0}^d\max_{x\in X^{\operatorname{an}}}|g_i'-g_i|(x)(D_0\cdots D_{i-1}\cdot D_{i+1}\cdots D_d).\end{split}\] See Corollary \ref{cor:intersection:trivial:divisor} for more details.
\end{enumerate}
\end{rema}
\section{Local height}
In this section, we fix a complete valued field $v=(k,|\raisebox{.4ex}{.}|)$ and a projective scheme $X$ over $\operatorname{Spec} k$. Let $d$ be the dimension of $X$.
\begin{defi}\label{Def: intersection number}
Let $\overline L_i=(L_i,\varphi_i)$, $i\in\{0,\ldots,d\}$ be a family of metrized invertible $\mathcal O_X$-modules, where each $L_i$ is an invertible $\mathcal O_X$-module, and $\varphi_i$ is a continuous and integrable metric on $L_i$. For any $i\in\{0,\ldots,d\}$, we let $s_i$ be a regular meromorphic section of $L_i$ on $X$. Assume that the Cartier divisors $\operatorname{div}(s_0),\ldots,\operatorname{div}(s_d)$ intersect properly. We define the \emph{local height} of $X$ with respect to the family of metrized invertible $\mathcal O_X$-modules $(\overline L_i)_{i=0}^d$ and the family of regular meromorphic sections $(s_i)_{i=0}^d$ as the local intersection number (see Definition \ref{Def: intersection pairing general case})
\[h_{\overline L_0,\ldots,\overline L_d}^{s_0,\ldots,s_d}(X):=\big(\operatorname{\widehat{div}}(s_0)\cdots\operatorname{\widehat{div}}(s_d)\big)_v.\]
\end{defi}
\begin{enonce}[remark]{Notation}\label{Not: simplified notation}
We often encounter the situation where each $\overline L_i$ is the pull-back by a projective morphism $f_i:X\rightarrow Y_i$ of a metrized invertible $\mathcal O_{Y_i}$-module $\overline{M_i}$ and $s_i$ is the pull-back of a regular meromorphic section $t_i$. In such a situation, for simplicity of notation,
we often use the expressions $h_{\overline M_0,\ldots,\overline M_{d}}^{t_0,\ldots,t_{d}}(X)$ or $h_{\overline L_0,\ldots,\overline L_{d}}^{t_0,\ldots,t_{d}}(X)$ to denote $h_{\overline L_0,\ldots,\overline L_d}^{s_0,\ldots,s_{d}}(X)$.
\end{enonce}
\begin{rema} We keep the notation of Definition \ref{Def: intersection number} in assuming that the field $k$ is algebraically closed.
Let $X_1,\ldots,X_n$ be irreducible components of $X$, considered as reduced closed subscheme of $X$. For any $j\in\{1,\ldots,n\}$, let $\operatorname{mult}_{X_j}(X)$ be the multiplicity of the component $X_j$, which is by definition the length of the Artinian local ring of $\mathcal O_X$ at the generic point of $X_j$. Then, for any $j\in\{1,\ldots,n\}$, the divisors on $X_j$ associated with the restricted sections $(s_i|_{X_j})_{i=0}^d$ intersect properly on $X_j$.
Assume firstly that $d=0$. In this case, each $X_j$ consists of a closed point $x_j$ of $X$, which is actually a rational point since $k$ is supposed to be algebraically closed. Hence $X_j^{\operatorname{an}}$ only contains one point, which we denote by $x_j^{\operatorname{an}}$. Note that $s_0$ does not vanish at any of the closed points $X_j$. By definition, $h_{\overline L_0}^{s_0}(X)$ is equal to
\begin{equation}\label{Equ: height of one point}-\sum_{j=1}^n\operatorname{mult}_{X_j}(X)\ln|s_0|_{\varphi_0}(x_j^{\operatorname{an}}).\end{equation}
In the case where $d\geqslant 1$, the induction formula in Definition \ref{def:local:intersection}
for local intersection number leads to the following formula for the local height.
\begin{equation}\label{Equ: height recursive formula}\begin{split}h_{\overline L_0,\ldots,\overline L_d}^{s_0,\ldots,s_d}(X)=
\sum_{i=1}^n & a_i h_{\overline L_0,\ldots,\overline L_{d-1}}^{s_0,\ldots,s_{d-1}}(Z_i)\\&-\int_{X^{\mathrm{an}}}\ln|s_d|_{\varphi_d}(x) \,\mu_{(L_0,\varphi_0)\cdots (L_{d-1},\varphi_{d-1})}(\mathrm{d}x),\end{split}\end{equation}
where $\sum_{i=1}^n a_i Z_i$ is the cycle associated with $\operatorname{div}(L_d; s_d)$.
\end{rema}
\begin{defi}
Let $(E,\|\raisebox{.4ex}{.}\|)$ be a finite-dimensional normed vector space over $k$, and $r$ be the rank of $E$. We denote by $\|\raisebox{.4ex}{.}\|_{\det}$ the norm on the one-dimensional vector space $\det(E):=\Lambda^r(E)$ such that,
\[\forall\,\eta\in\det(E),\quad
\|\eta\|_{\det}:=\inf_{\eta=t_1\wedge\cdots\wedge t_r}\|t_1\|\cdots\|t_r\|.\]
Note that, if the norm $\|\raisebox{.4ex}{.}\|$ is ultrametric or induced by an inner product, for any complete valued extension $k'$ of $k$, one has (see Definition \ref{Def: extension of sclars})
\begin{equation}\label{Equ: determinant norm}\|\raisebox{.4ex}{.}\|_{k',\det}=\|\raisebox{.4ex}{.}\|_{\det,k'},\end{equation}
if we identify $\det(E)\otimes_kk'$ with $\det(E\otimes_kk')$.
We refer the readers to \cite[Proposition 1.3.19]{CMArakelovAdelic} for a proof.
\end{defi}
\begin{prop}\label{Pro: height projective space}
Let $E$ be a finite-dimensional vector space over $k$, equipped with a norm $\|\raisebox{.4ex}{.}\|$ which is either ultrametric or induced by an inner product, $r=\operatorname{dim}_k(E)$, and $L=\mathcal O_E(1)$ be the universal invertible sheaf on $\mathbb P(E)$. We equip $L$ with the orthogonal quotient metric $\varphi$ induced by $\|\raisebox{.4ex}{.}\|$. Let $(s_j)_{j=0}^r$ be a basis of $E$ over $k$. If $|\raisebox{.4ex}{.}|$ is non-Archimedean, then
\[h_{\overline{L},\ldots,\overline{L}}^{s_0,\ldots,s_r}(\mathbb P(E))=-\ln\|s_0\wedge\cdots\wedge s_r\|_{\det};\] if $|\raisebox{.4ex}{.}|$ is Archimedean, then
\[h_{\overline{L},\ldots,\overline{L}}^{s_0,\ldots,s_r}(\mathbb P(E))=-\ln\|s_0\wedge\cdots\wedge s_r\|_{\det}+\sigma_r,\]
where
\[\sigma_r=\frac 12\sum_{m=1}^r\sum_{\ell=1}^m\frac{1}{\ell}\]
is the $r$-th Stoll number.
\end{prop}
\begin{proof}
First, the metric $\varphi_{\mathbb C_k}$ identifies with the orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|_{\mathbb C_k}$. Therefore, by \eqref{Equ: determinant norm} we may assume without loss of generality that $k$ is algebraically closed.
By Remark \ref{Rem: approximation by good norms}, one can find a sequence $(\|\raisebox{.4ex}{.}\|_n)_{n\in\mathbb N}$ of orthonormally decomposable norms such that
\[\lim_{n\rightarrow+\infty}d(\|\raisebox{.4ex}{.}\|_n,\|\raisebox{.4ex}{.}\|)=0.\]
By \eqref{Equ: distance of quotient metrics}, if we denote by $\varphi_n$ the orthogonal quotient metric on $L$ induced by $\|\raisebox{.4ex}{.}\|_n$, then one has
\[\lim_{n\rightarrow+\infty}d(\varphi_n,\varphi)=0.\]
By Corollary \ref{cor:intersection:trivial:divisor}, one has
\[\lim_{n\rightarrow+\infty}h_{(L,\varphi_n),\ldots,(L,\varphi_n)}^{s_0,\ldots,s_r}(\mathbb P(E))=h_{\overline{L},\ldots,\overline{L}}^{s_0,\ldots,s_r}(\mathbb P(E)).\]
Moreover, by \cite[Proposition 1.1.64]{CMArakelovAdelic} one has
\[0\leqslant d(\|\raisebox{.4ex}{.}\|_{n,\det},\|\raisebox{.4ex}{.}\|_{\det})\leqslant r d(\|\raisebox{.4ex}{.}\|_n,\|\raisebox{.4ex}{.}\|) \]
and hence
\[\lim_{n\rightarrow+\infty}d(\|\raisebox{.4ex}{.}\|_{n,\det},\|\raisebox{.4ex}{.}\|_{\det})=0.\]
Therefore, without loss of generality, we may assume that the norm $\|\raisebox{.4ex}{.}\|$ is orthonormally decomposable.
We reason by induction on $r$. In the case where $r=0$, the vector space $E$ is one-dimensional, and $s_0$ is a non-zero element of $E$. One has
\[h_{\overline L}^{s_0}(\mathbb P(E))=-\ln\|s_0\|.\]
We now assume that $r\geqslant 1$. Let $G$ be the quotient vector space of $E$ by $ks_r$. Note that the quotient norm $\|\raisebox{.4ex}{.}\|_{\operatorname{quot}}$ on $G$ is orthonormally decomposable (see Proposition \ref{Pro: decomposable quotient norm}). For $j\in\{0,\ldots,r-1\}$, let $\overline{s}_j$ be the class of $s_j$ in $G$. We can also view $\overline{s}_j$ as the restriction of $s_j$ to the closed subscheme $\mathbb P(G)$ of $\mathbb P(E)$. We apply the induction hypothesis to $(G,\|\raisebox{.4ex}{.}\|_{\operatorname{quot}})$ and obtain (see Notation \ref{Not: simplified notation})
\[h_{\overline L,\ldots,\overline L}^{s_0,\ldots,s_{r-1}}(\mathbb P(G))=-\ln\|\overline{s}_0\wedge\cdots\wedge\overline{s}_{r-1}\|_{\operatorname{quot},\det}\]
when $|\raisebox{.4ex}{.}|$ is non-Archimedean and
\[h_{\overline L,\ldots,\overline L}^{s_0,\ldots,s_{r-1}}(\mathbb P(G))=-\ln\|\overline{s}_0\wedge\cdots\wedge\overline{s}_{r-1}\|_{\operatorname{quot},\det}+\sigma_{r-1}\]
We now compute the integral
\[-\int_{\mathbb P(E)^{\mathrm{an}}}\ln|s_r|_{\varphi}\,\mathrm{d}\mu_{\overline{L}^r}.\]
We first consider the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean. By Proposition \ref{Pro: non Archimedean projective space} one has
\[\int_{\mathbb P(E)^{\mathrm{an}}}\ln|s_r|_{\varphi}\,\mathrm{d}\mu_{\overline{L}^r}=-\ln|s_r|_{\varphi}(\xi)=-\ln\|s_r\|,\]
where $\xi$ denotes the Gauss point of $\mathbb P(E)^{\mathrm{an}}$.
Therefore, by \cite[Proposition 1.2.51]{CMArakelovAdelic} we obtain
\[h_{\overline{L},\ldots,\overline{L}}^{s_0,\ldots,s_r}(\mathbb P(E))=-\ln\|\overline s_0\wedge\cdots\wedge\overline s_{r-1}\|_{\operatorname{quot},\det}-\ln\|s_r\|=-\ln\|s_0\wedge\cdots\wedge s_r\|_{\det}.\]
In the case where $|\raisebox{.4ex}{.}|$ is Archimedean, by \cite[\S1.4.3]{MR1260106} Remark (iii), one has
\[-\int_{\mathbb P(E)^{\mathrm{an}}}\ln|s_r|_{\varphi_r}\,\mathrm{d}\mu_{\overline{L}^r}=-\ln\|s\|+\frac 12\sum_{\ell=1}^r\frac{1}{\ell}.\]
Therefore \[\begin{split}h_{\overline{L},\ldots,\overline{L}}^{s_0,\ldots,s_r}(\mathbb P(E))&=-\ln\|\overline{s}_0\wedge\cdots\wedge\overline{s}_{r-1}\|_{\operatorname{quot},\det}-\ln\|s_r\|+\frac 12\sum_{m=1}^r\sum_{\ell=1}^m\frac{1}{\ell}\\
&=-\ln\|s_0\wedge\cdots\wedge s_r\|_{\det}+\frac 12\sum_{m=1}^r\sum_{\ell=1}^m\frac{1}{\ell}.
\end{split}\]
\end{proof}
In the remaining of the section, we consider a family \[(E_i,\|\raisebox{.4ex}{.}\|_i),\quad i\in\{0,\ldots,d\}\] of finite-dimensional vector spaces over $k$ equipped with norms which are either ultrametric or induced by inner products. For each $i\in\{0,\ldots,d\}$, we let $(E_i^\vee,\|\raisebox{.4ex}{.}\|_{i,*})$ be the dual normed vector space of $(E_i,\|\raisebox{.4ex}{.}\|_i)$, $r_i:=\operatorname{dim}_k(E_i)-1$, $(s_{i,j})_{j=0}^{r_i}$ be a basis of $E_i$ over $k$, and $(\alpha_{i,j})_{j=0}^{r_i}$ be the dual basis of $(s_{i,j})_{j=0}^{r_i}$, namely
\[\alpha_{i,j}(s_{i,j})=1\quad \text{and}\quad \alpha_{i,j}(s_{i,\ell})=0\quad \text{if $j\neq\ell$}.\]
Let $\check{\mathbb P}$ be the product projective space
\[\mathbb P(E_0^\vee)\times_k\cdots\times_k\mathbb P(E_d^\vee).\]
For any $i\in\{0,\ldots,d\}$, let $\pi_i:\check{\mathbb P}\rightarrow\mathbb P(E_i^\vee)$ be the morphism of projection to the $i^{\text{th}}$ coordinate, and $L_i=\pi_i^*(\mathcal O_{E_i^\vee}(1))$. We equip $L_i$ with the orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|_{i,*}$, which we denote by $\varphi_i$. Let $(\delta_0,\ldots,\delta_d)$ be an element of $\mathbb N^{d+1}$,
\[L=\pi_0^*(\mathcal O_{E_0^\vee}(\delta_0))\otimes\cdots\otimes\pi_d^*(\mathcal O_{E_d^\vee}(\delta_d))=L_0^{\otimes\delta_0}\otimes\cdots\otimes L_d^{\otimes\delta_d}.\]
We equip $L$ with the metric
\[\varphi:=\varphi_0^{\otimes\delta_0}\otimes\cdots\otimes\varphi_d^{\otimes\delta_d}.\]
Let $R$ be a non-zero element of
\[S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_k S^{\delta_d}(E_d^\vee),\]
which is considered as a global section of $L$, and also as a multi-homogenous polynomial of multi-degree $(\delta_0,\ldots,\delta_d)$ on $E_0\times\cdots\times E_d$. For any $i\in\{0,\ldots,d\}$, let
\[{\overline{\boldsymbol{L}}_i}=(\underbrace{\overline L_i,\ldots,\overline L_i}_{r_i\text{ copies}}),\quad\boldsymbol{\alpha}_i:=(\alpha_{i,j})_{j=1}^{r_i}.\]
The purpose of this section is to compute the local height $h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P})$.
\begin{prop}\label{Pro: hight of multi-projective} Assume that the sections $R$ and
\[\alpha_{i,j},\quad i\in\{0,\ldots,d\},\quad j\in\{1,\ldots,r_i\}\]
intersect property on $\check{\mathbb P}$. If the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean, then
\[h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P})=-\ln|R(s_{0,0},\ldots,s_{d,0})|-\sum_{i=0}^d\delta_i\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det};\]
if the absolute value $|\raisebox{.4ex}{.}|$ is Archimedean, then
\[h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P})=-\ln|R(s_{0,0},\ldots,s_{d,0})|-\sum_{i=0}^d\delta_i\big(\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}-\sigma_{r_i}\big).\]
\end{prop}
\begin{proof}
By the same argument as in the beginning of the proof of Proposition \ref{Pro: height projective space}, we may assume without loss of generality that $k$ is algebraically closed and that all norms $\|\raisebox{.4ex}{.}\|_i$ are orthonormally decomposable.
We reason by induction on $r_0+\cdots+r_d$. Consider first the case where $r_0=\cdots=r_d=0$. One has
\[h_{\overline L}^R(\check{\mathbb P})=-\ln|R(s_{0,0},\ldots,s_{d,0})|.\]
In the following, we assume that $r_0+\cdots+r_d>0$. Let $i$ be an element of $\{0,\ldots,d\}$ such that $r_i>0$. We consider the quotient vector space $G_i^\vee=E_i^\vee/k\alpha_{i,r_i}$. For $j\in\{0,\ldots,r_i-1\}$, let $\overline \alpha_{i,j}$ be the class of $\alpha_{i,j}$ in $G_i$. Let $\overline{\boldsymbol{\alpha}}_i:=(\overline \alpha_{i,j})_{j=1}^{r_i-1}$ and
\[\check{\mathbb P}'=\mathbb P(E_0)\times_k\cdots\times\mathbb P(E_{i-1})\times_k\mathbb P(G_i)\times_k\mathbb P(E_{i+1})\times_k\cdots\times_k\mathbb P(E_d).\]
By the same argument as in Proposition \ref{Pro: non Archimedean projective space}, we obtain that, in the case where the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean, one has
\[\mu_{\overline L\,\overline L_0^{r_0}\cdots\overline L_{i-1}^{r_{i-1}}\overline L_i^{r_i-1}\overline{L}_{i+1}^{r_{i+1}}\cdots\overline L_d^{r_d}}=\delta_i\operatorname{Dirac}_\xi,\]
where $\operatorname{Dirac}_\xi$ denotes the Dirac measure at the Gauss point $\xi$ of $\check{\mathbb P}^{\operatorname{an}}$. Hence,
by \eqref{Equ: height recursive formula}, one has
\[\begin{split}h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P})&=h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_{i-1},\overline{\boldsymbol{L}}_i',\overline{\boldsymbol{L}}_{i+1},\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_{i-1},\overline{\boldsymbol{\alpha}}_i,\boldsymbol{\alpha}_{i+1},\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P}')-\delta_i\ln|\alpha_{i,r_i}|_{\varphi_i}(\xi)\\&=h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_{i-1},\overline{\boldsymbol{L}}_i',\overline{\boldsymbol{L}}_{i+1},\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_{i-1},\overline{\boldsymbol{\alpha}}_i,\boldsymbol{\alpha}_{i+1},\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P}')-\delta_i\ln\|\alpha_{i,r_i}\|_{i,*},
\end{split}\]
where
\[\overline{\boldsymbol{L}}_i':=(\underbrace{\overline L_i,\ldots,\overline L_i}_{r_i-1\text{ copies}}).\]
By the induction hypothesis, we obtain
\[\begin{split}h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P})&=-\ln|R(s_{0,0},\ldots,s_{d,0})|-\sum_{j\in\{0,\ldots,d\}\setminus\{i\}}\delta_j\ln\|\alpha_{j,0}\wedge\cdots\wedge\alpha_{j,r_j}\|_{j,*,\det}\\
&\qquad\quad -\delta_i\ln\|\overline{\alpha}_{i,0}\wedge\cdots\wedge\overline{\alpha}_{i,r_i-1}\|_{i,*,\operatorname{quot},\det}-\delta_i\ln\|\alpha_{i,r_i}\|_{i,*}\\
&=-\ln|R(s_{0,0},\ldots,s_{d,0})|-\sum_{j=0}^d\delta_j\ln\|\alpha_{j,0}\wedge\cdots\wedge{\alpha_{j,r_j}}\|_{j,*,\det},
\end{split}\]
where the last equality comes from \cite[Proposition 1.2.51]{CMArakelovAdelic}.
In the case where $|\raisebox{.4ex}{.}|$ is Archimedean, by \cite[\S1.4.3]{MR1260106} Remark (iii) one has
\[h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P})=h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_{i-1},\overline{\boldsymbol{L}}_i',\overline{\boldsymbol{L}}_{i+1},\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_{i-1},\overline{\boldsymbol{\alpha}}_i,\boldsymbol{\alpha}_{i+1},\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P}')-\delta_i\Big(\ln\|\alpha_{i,r_i}\|_{i,*}-\frac 12\sum_{\ell=1}^{r_i}\frac{1}{\ell}\Big).\]
Thus the induction hypothesis leads to
\[\begin{split}h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\check{\mathbb P})&=-\ln|R(s_{0,0},\ldots,s_{d,0})|-\hspace{-1.5ex}
\sum_{\begin{subarray}{c}j\in\{0,\ldots,d\}\\j\neq i\end{subarray}}\hspace{-1.5ex}
\delta_j\big(\ln\|\alpha_{j,0}\wedge\cdots\wedge\alpha_{j,r_j}\|_{j,*,\det}-\sigma_{r_j}\big)\\
&\hspace{-8ex} -\delta_i\big(\ln\|\overline{\alpha}_{i,0}\wedge\cdots\wedge\overline{\alpha}_{i,r_i-1}\|_{i,*,\operatorname{quot},\det}-\sigma_{r_i-1}\big) -
\delta_i\Big(\ln\|\alpha_{i,r_i}\|_{i,*}-\frac 12\sum_{\ell=1}^{r_i}\frac{1}{\ell}\Big)\\&=-\ln|R(s_{0,0},\ldots,s_{d,0})|-\sum_{j=0}^d\delta_j\big(\ln\|\alpha_{j,0}\wedge\cdots\wedge\alpha_{j,r_j}\|-\sigma_{r_j}\big),
\end{split}\]
as required.
\end{proof}
\section{Local height of the resultant}\label{Sec: local height of resultant}
The purpose of this subsection is to relate local heights of a projective variety and its resultant. As in the previous section, $v=(k,|\raisebox{.4ex}{.}|)$ denotes a complete valued field such that $|\raisebox{.4ex}{.}|$ is not trivial. We fix a projective $k$-scheme $X$ and we let $d$ be the Krull dimension of $X$. Let $(E_i)_{i=0}^d$ be a family of finite-dimensional vector spaces over $k$. For each $i\in\{0,\ldots,d\}$, we denote by $r_i:=\dim_k(E_i)-1$ and let $\|\raisebox{.4ex}{.}\|_{i}$ be a norm on $E_i$, which is supposed to be either ultrametric or induced by an inner product. Let $f_i:X\rightarrow\mathbb P(E_i)$ be a closed immersion. We pick elements $s
_0,\ldots,s_d$ of $E_0,\ldots,E_d$ respectively, such that \[\operatorname{div}(s_0|_X),\ldots,\operatorname{div}(s_d|_X)\] intersect properly on $X$. For simplicity of notation, we denote by
\[\boldsymbol{s}:=(s_0,\ldots,s_d).\]
Let $\check{\mathbb P}:=\mathbb P(E_0^\vee)\times_k\cdots\times_k\mathbb P(E_d^\vee)$, and let \[p:X\times_k\check{\mathbb P}\longrightarrow X\quad\text{and}\quad q:X\times_k\check{\mathbb P}\longrightarrow\check{\mathbb P}\]
be morphisme of projections.
For any $i\in\{0,\ldots,d\}$, let $\pi_i:\check{\mathbb P}\rightarrow\mathbb P(E_i^\vee)$ be the projection to the $i$-th coordinate and let $q_i=\pi_i\circ q$.
For $i\in\{0,\ldots,d\}$, let $\overline L_i$ be $L_i=\pi_i^*({\mathcal O_{E_i^\vee}(1)})$ equipped with the pull-back of the orthogonal quotient metric on $\mathcal O_{E_i^\vee}(1)$ associated with $\|\raisebox{.4ex}{.}\|_{i,*}$, and let \[\overline{\boldsymbol{L}}_i:=(\underbrace{\overline{L}_i,\ldots,\overline{L}_i}_{r_i\text{ copies}})\]
and \[(\alpha_{i,0},\boldsymbol{\alpha}_i=(\alpha_{i,1},\ldots,\alpha_{i,r_i}))\] be a basis of $E_i^\vee$ such that \[\alpha_{i,0}(s_i)=1\quad\text{and}\quad \alpha_{i,j}(s_i)=0\text{ for $j\in\{1,\ldots,r_i\}$.}\] For simplicity, we denote by $R$ the resultant
\[R_{f_0,\ldots,f_d}^{X,s_0,\ldots,s_d}\]
as in Definition \ref{Def:resultant precise}, considered as a global section of
\[L=\pi_0^*(L_0)^{\otimes\delta_0}\otimes\cdots\otimes\pi_d^*(L_d)^{\otimes\delta_d},\]
where
\[\delta_i:=\deg\big(c_1(L_0)\cdots c_1(L_{i-1})c_1(L_{i+1})\cdots c_1(L_d)\cap[X]\big).\] Note that one has $R(s_0,\ldots,s_d)=1$. Moreover, the Cartier divisors \[\operatorname{div}(R), \operatorname{div}(\pi_0^*(\alpha_{0,1})),\ldots, \operatorname{div}(\pi_0^*(\alpha_{0,r_0})),\ldots,\operatorname{div}(\pi_d^*(\alpha_{d,1})),\ldots, \operatorname{div}(\pi_d^*(\alpha_{d,r_d}))\] intersect properly.
\begin{lemm}\label{Lem: h div R h IX}
Assume that the field $k$ is algebraically closed and $X$ is integral. One has
\[h_{\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}\big(\operatorname{div}(R)\big)=h_{q^*(\boldsymbol{\overline L}_0),\ldots,q^*(\boldsymbol{\overline L}_d)}^{q_0^*(\boldsymbol{\alpha}_0),\ldots,q_d^*(\boldsymbol{\alpha}_d)}(I_X).\]
\end{lemm}
\begin{proof}
The projection $q:I_X\rightarrow\operatorname{div}(R)$ is a birational morphism (see the proof of \cite[Proposition 3.1]{MR1264417}). Hence the equality follows from the induction formula \eqref{Equ: height recursive formula} and \cite[Proposition 2.4.11 (4)]{MR3498148}.
\end{proof}
\begin{defi}\label{Def: epsilon tensor product}
Assume that the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean. We equip each symmetric power $S^{\delta_i}(E_i^\vee)$ with the $\varepsilon$-symmetric power norm of $\|\raisebox{.4ex}{.}\|_{i,*}$, namely the quotient norm of the $\varepsilon$-tensor power of $\|\raisebox{.4ex}{.}\|_{i,*}$ (see Remark \ref{Rem: extension of normss} for the definition of the dual norm $\|\raisebox{.4ex}{.}\|_{i,*}$).
Recall that the $\varepsilon$-tensor power of the norm $\|\raisebox{.4ex}{.}\|_{i,*}$ is the norm $\|\raisebox{.4ex}{.}\|_{i,*,\varepsilon}$ on $(E_i^{\vee})^{\otimes_k\delta_i}$ defined as (see \cite[Definition 1.1.52]{CMArakelovAdelic})
\[\|T\|_{i,*,\varepsilon}=\sup_{\begin{subarray}{c}(t_1,\ldots,t_{\delta_i})\in E_i^{\delta_i}\\
\forall\,j\in\{1,\ldots,\delta_i\},\;t_j\neq 0
\end{subarray}}\frac{|T(t_1,\ldots,t_{\delta_i})|}{\|t_1\|_i\cdots\|t_{\delta_i}\|_i}.\]
We then equip the vector space $S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_k S^{\delta_d}(E_d^\vee)$ with the $\varepsilon$-tensor product of the $\varepsilon$-symmetric power norms, which we denote simply by $\|\raisebox{.4ex}{.}\|$.
\end{defi}
\begin{rema}\label{Rem: extensions scalars}
Note that, by \cite[Definition 1.1.58]{CMArakelovAdelic}, the norm $\|\raisebox{.4ex}{.}\|$ also identifies with the quotient norm by the canonical quotient map
\[(E_0^{\vee})^{\otimes_k\delta_0}\otimes_k\cdots\otimes_k(E_d^{\vee})^{\otimes_k\delta_d}\longrightarrow S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_k S^{\delta_d}(E_d^\vee)\]
of the $\varepsilon$-tensor product of $\delta_i$ copies of $\|\raisebox{.4ex}{.}\|_{i,*}$, $i\in\{0,\ldots,d\}$. By Propositions 1.3.20 and 1.3.21 of \cite{CMArakelovAdelic}, we obtain that, for any complete valued extension $k'$ of $k$, the norm $\|\raisebox{.4ex}{.}\|_{k'}$ on
\[\big(S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_k S^{\delta_d}(E_d^\vee)\big)\otimes_{k}k'\cong S^{\delta_0}(E_{0,k'}^\vee)\otimes_{k'}\cdots\otimes_{k'}S^{\delta_d}(E_{d,k'}^\vee)\]
identifies with the $\varepsilon$-tensor product of $\delta_i$ copies of $\|\raisebox{.4ex}{.}\|_{i,k',*}$, $i\in\{0,\ldots,d\}$.
\end{rema}
\begin{lemm}\label{Lem: h div r et h p}
In the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean and $k$ is algebraically closed, one has
\begin{equation}\label{Equ: passage a hauteur de P}h_{\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}(\operatorname{div}(R))=h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}(\check{\mathbb P})+\ln\|R\|.\end{equation}
\end{lemm}
\begin{proof}
Let $\xi$ be the Gauss point of $\check{\mathbb P}^{\mathrm{an}}$. It suffices to observe that
\[|R|_{\varphi}(\xi)=\|R\|,\]
where $\varphi$ is tensor product of orthogonal quotient metrics. In fact, if we consider the Veronese-Segre embedding
\[\check{\mathbb P}\longrightarrow\mathbb P(S^{\delta_0}(E_0^\vee))\times_k\cdots\times_k\mathbb P(S^{\delta_d}(E_d^\vee))\longrightarrow\mathbb P(S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee)),\]
then the metric $\varphi$ identifies with the quotient metric induced by $\|\raisebox{.4ex}{.}\|$ (see \cite[Proposition 1.1.58]{CMArakelovAdelic}). Moreover, one has
\[\mu_{\overline L_0^{r_0}\cdots \overline L_d^{r_d}}=\operatorname{Dirac}_\xi.\]
Therefore the equality \eqref{Equ: passage a hauteur de P} follows from the induction formula \eqref{Equ: height recursive formula}.
\end{proof}
\begin{lemm}\label{Lem: h div R etc}
In the case where $|\raisebox{.4ex}{.}|$ is Archimedean and $k=\mathbb C$, one has
\[\begin{split}h_{\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}&(\operatorname{div}(R))=h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}(\check{\mathbb P})\\&+\int_{\mathbb S_0\times\cdots\times \mathbb S_d}\ln|R(z_0,\ldots,z_d)|\,\eta_{\mathbb S_0}(\mathrm{d}z_0)\otimes\cdots\otimes\eta_{\mathbb S_d}(\mathrm{d}z_d),
\end{split}\]
where $\mathbb S_i$ is the unit sphere of $(E_{i,\mathbb C},\|\raisebox{.4ex}{.}\|_{i,\mathbb C})$, and $\mu_{\mathbb S_i}$ is the $U(E_{i,\mathbb C},\|\raisebox{.4ex}{.}\|_{i,\mathbb C})$-invariant Borel probability measure on $\mathbb S_i$.
\end{lemm}
\begin{proof}
This is a direct consequence of the induction formula \eqref{Equ: height recursive formula} and Remark \ref{Rem: mesure complexe}.
\end{proof}
\begin{lemm}\label{Lem: h s in h IX}
Assume that the field $k$ is algebraically closed and $X$ is integral. For any $i\in \{0,\ldots,d\}$, we equip $\mathcal O_{E_i}(1)$ with the orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|_i$, and denote by $M_i'$ the restriction of $\mathcal O_{E_i}(1)$ to $X$ and equip it with the restricted metric. If $|\raisebox{.4ex}{.}|$ is non-Archimedean, then one has
\[h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)=h_{q^*(\boldsymbol{\overline L}_0),\ldots,q^*(\boldsymbol{\overline L}_d)}^{q_0^*(\boldsymbol{\alpha}_0),\ldots,q_d^*(\boldsymbol{\alpha}_d)}(I_X) +\sum_{i=0}^d\delta_i\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det};\]
if $|\raisebox{.4ex}{.}|$ is Archimedean, then one has
\begin{multline*}h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)=h_{q^*(\boldsymbol{\overline L}_0),\ldots,q^*(\boldsymbol{\overline L}_d)}^{q_0^*(\boldsymbol{\alpha}_0),\ldots,q_d^*(\boldsymbol{\alpha}_d)}(I_X)+\sum_{i=0}^d\delta_i\big(\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}-\sigma_{r_i-1}\big),
\end{multline*}
where
\[\sigma_{r_i-1}=\frac 12\sum_{m=1}^{r_i-1}\sum_{\ell=1}^m\frac{1}{\ell}.\]
\end{lemm}
\begin{proof}
For $i\in\{0,\ldots,d\}$, let $t_i$ be the global section of $\mathcal O_{E_i}(1)\boxtimes\mathcal O_{E_i^\vee}(1)$ on $\mathbb P(E_i)\times_k\mathbb P(E_i^\vee)$ defining the incidence subscheme. Then $t_i$ corresponds to the restriction of the trace element of $E_i\otimes_kE_i^\vee$ via the Segre embedding \[\mathbb P(E_i)\times_k\mathbb P(E_i^\vee)\longrightarrow\mathbb P(E_i\otimes_kE_i^\vee).\]
Let $\boldsymbol{t}=(t_0,\ldots,t_{d})$. For any $i\in\{0,\ldots,d\}$, let
\[(s_i,s_{i,1},\ldots,s_{i,r_i})\]
be the dual basis of $(\alpha_{i,j})_{j=0}^{r_i}$. By definition one has
\[t_i=s_i\otimes\alpha_{i,0}+s_{i,1}\otimes\alpha_{i,1}+\cdots+s_{i,r_i}\otimes\alpha_{i,r_i}.\]
For $i\in\{0,\ldots,d\}$, let $L_i:=q_i^*(\mathcal O_{E_i^\vee}(1))$, $M_i=p^*(\mathcal O_{E_i}(1)|_X)$ and $N_i= L_i\otimes M_i$. We use two methods to compute the following local height of $X\times\check{\mathbb P}$ (see Notation \ref{Not: simplified notation})
\[h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(X\times_k\check{\mathbb P}),\]
where $\overline{\boldsymbol{N}}=(\overline N_0,\ldots,\overline{N}_d)$.
We will show by induction that
\begin{equation}\label{Equ: height of product}h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(X\times_k\check{\mathbb P})=h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)-\sum_{i=0}^d\delta_i\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}\end{equation}
if $|\raisebox{.4ex}{.}|$ is non-Archimedean, and
\begin{equation}\label{Equ: height of product Archimedean}h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(X\times_k\check{\mathbb P})=h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)-\sum_{i=0}^d\delta_i\big(\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}-\sigma_{r_i}\big)\end{equation}
if $|\raisebox{.4ex}{.}|$ is Archimedean.
Let $i\in\{0,\ldots,d\}$ be such that $r_i>0$. Let $G_i^\vee=E_i^\vee/k\alpha_{i,r_i}$, $\overline{\boldsymbol{\alpha}}_i=(\overline \alpha_{i,j})_{j=1}^{r_i-1}$, and
\[\check{\mathbb P}'=\mathbb P(E_0)\times_k\cdots\times\mathbb P(E_{i-1})\times_k\mathbb P(G_i)\times_k\mathbb P(E_{i+1})\times_k\cdots\times_k\mathbb P(E_d).\]
Then, with the notation \[\overline{\boldsymbol{L}}_i':=(\underbrace{\overline L_i,\ldots,\overline L_i}_{r_i-1\text{ copies}}),\]
by \eqref{Equ: height recursive formula} one can write $h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(X\times_k\check{\mathbb P})$ as
\[h_{\overline{\boldsymbol{N}},\overline {\boldsymbol L}_0,\ldots,\overline {\boldsymbol L}_{i-1},\overline {\boldsymbol L}_i',\overline {\boldsymbol L}_{i+1},\ldots,\overline {\boldsymbol L}_{d}}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_{i-1},\overline{\boldsymbol{\alpha}}_i,\boldsymbol{\alpha}_{i+1},\ldots,\boldsymbol{\alpha}_d}(X\times\check{\mathbb P})-\int_{(X\times\check{\mathbb P})^{\mathrm{an}}}\hspace{-2mm}\ln|\alpha_{i,r_i}|\,\mathrm{d}\mu_{\overline N_0\cdots\overline N_d\overline L_0^{r_0}\cdots\overline L_{i-1}^{r_{i-1}}\overline L_i^{r_i-1}\overline L_{i+1}^{r_{i+1}}\cdots\overline L_d^{r_d}},\]
which is equal to
\[h_{\overline{\boldsymbol{N}},\overline {\boldsymbol L}_0,\ldots,\overline {\boldsymbol L}_{i-1},\overline {\boldsymbol L}_i',\overline {\boldsymbol L}_{i+1},\ldots,\overline {\boldsymbol L}_{d}}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_{i-1},\overline{\boldsymbol{\alpha}}_i,\boldsymbol{\alpha}_{i+1},\ldots,\boldsymbol{\alpha}_d}(X\times\check{\mathbb P})-\int_{(X\times\check{\mathbb P})^{\mathrm{an}}}\ln|\alpha_{i,r_i}|\,\mathrm{d}\mu_{\overline M_0\cdots\overline{M}_{i-1}\overline M_{i+1}\cdots\overline M_d\overline L_0^{r_0}\cdots\overline L_d^{r_d}}.\]
If $|\raisebox{.4ex}{.}|$ is non-Archimedean, it identifies with
\[h_{\overline{\boldsymbol{N}},\overline {\boldsymbol L}_0,\ldots,\overline {\boldsymbol L}_{i-1},\overline {\boldsymbol L}_i',\overline {\boldsymbol L}_{i+1},\ldots,\overline {\boldsymbol L}_{d}}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_{i-1},\overline{\boldsymbol{\alpha}}_i,\boldsymbol{\alpha}_{i+1},\ldots,\boldsymbol{\alpha}_d}(X\times\check{\mathbb P})-\delta_i\ln\|\alpha_{i,r_i}\|_{i,*}\]
In the case where $|\raisebox{.4ex}{.}|$ is Archimedean, it equals
\[h_{\overline N,\overline {\boldsymbol L}_0,\ldots,\overline {\boldsymbol L}_{i-1},\overline {\boldsymbol L}_i',\overline {\boldsymbol L}_{i+1},\ldots,\overline {\boldsymbol L}_{d}}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_{i-1},\overline{\boldsymbol{\alpha}}_i,\boldsymbol{\alpha}_{i+1},\ldots,\boldsymbol{\alpha}_d}(X\times\check{\mathbb P})-\delta_i\Big(\ln\|\alpha_{i,r_i}\|_{i,*}-\frac 12\sum_{\ell=1}^{r_i}\frac{1}{\ell}\Big).\]
Hence by induction we obtain \eqref{Equ: height of product} and \eqref{Equ: height of product Archimedean} according to the nature of $|\raisebox{.4ex}{.}|$.
Now let $\boldsymbol{t}'=(t_0,\ldots,t_{d-1})$ and $\overline{\boldsymbol{N}}'=(\overline N_0,\ldots,\overline N_{d-1})$, still by \eqref{Equ: height recursive formula} one can write $h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(X\times_k\check{\mathbb P})$ as
\[\begin{split}&\quad\; h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t}',\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\operatorname{div}(t_d)) -\int_{(X\times_k\check{\mathbb P})^{\mathrm{an}}}\ln|t_d|\,\mathrm{d}\mu_{\overline N_0\cdots\overline{N}_{d-1}\overline L_0^{r_0}\cdots\overline L_d^{r_d}}\\
&=h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t}',\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(\operatorname{div}(t_d)) -\int_{(X\times_k\check{\mathbb P})^{\mathrm{an}}}\ln|t_d|\,\mathrm{d}\mu_{\overline M_0\cdots\overline M_{d-1}\overline L_0^{r_0}\cdots\overline L_d^{r_d}}
\end{split}\]
Note that for any element $z\in(X\times_k\check{\mathbb P})^{\mathrm{an}}$ represented by
\[(\beta,x_0,\ldots,x_d)\in E_{d,\widehat{\kappa}(z)}^\vee\times E_{0,\widehat{\kappa}(z)}\cdots\times E_{d,\widehat{\kappa}(z)}\]
one has
\begin{equation}\label{Equ: ln td}\ln|t_d|(z)=\ln\frac{|\beta(x_d)|_{z}}{\|\beta\|_{d,\widehat{\kappa}(z)}\cdot\|x_d\|_{d,\widehat{\kappa}(z)}}.\end{equation}
In the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean, this leads to \[\int_{(X\times_k\check{\mathbb P})^{\mathrm{an}}}\ln|t_d|\,\mathrm{d}\mu_{\overline M_0\cdots\overline M_{d-1}\overline L_0^{r_0}\cdots\overline L_d^{r_d}}=0\]
by using \eqref{Equ:Fubini} and
\[\int_{\mathbb P(E_{d,\widehat{\kappa}(z)}^\vee)^{\mathrm{an}}}\ln|\beta|\,\mathrm{d}\mu_{\overline{\mathcal O_{E_d}(d)}^{r_d}}=\ln\|\beta\|_{d,*,\widehat{\kappa}(x)}.\]
In the case where $|\raisebox{.4ex}{.}|$ is Archimedean, by \cite[\S1.4.3]{MR1260106} Remark (iii), \eqref{Equ: ln td} leads to
\[-\int_{(X\times_k\check{\mathbb P})^{\mathrm{an}}}\ln|t_d|\,\mathrm{d}\mu_{\overline M_0\cdots\overline{M}_{d-1}\overline L_0^{r_0}\cdots\overline L_d^{r_d}}=\frac {\delta_d}2\sum_{\ell=1}^{r_i}\frac{1}{\ell}.\]
Then by induction we obtain
\begin{equation}\label{Equ: h X times P}h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(X\times_k\check{\mathbb P})=h_{\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(I_X)\end{equation}
when $|\raisebox{.4ex}{.}|$ is non-Archimedean and
\begin{equation}\label{Equ: h X times P2}h_{\overline{\boldsymbol{N}},\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{t},\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(X\times_k\check{\mathbb P})=h_{\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d}(I_X)+\frac 12\delta_i\sum_{i=0}^d\sum_{\ell=1}^{r_i}\frac{1}{\ell}\end{equation}
when $|\raisebox{.4ex}{.}|$ is Archimedean. Combining \eqref{Equ: h X times P} with \eqref{Equ: height of product}, and \eqref{Equ: h X times P2} with \eqref{Equ: height of product Archimedean}, we obtain the result.
\end{proof}
\begin{theo}\label{Thm: equality of local height} For any $i\in \{0,\ldots,d\}$, we equip $\mathcal O_{E_i}(1)$ with the orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|_i$, and denote by $M_i'$ the restriction of $\mathcal O_{E_i}(1)$ to $X$ and equip it with the restricted metric. In the case where $|\raisebox{.4ex}{.}|$ is non-Archimedean, one has
\[h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)=\ln\|R\|,\]
where the norm $\|\raisebox{.4ex}{.}\|$ was introduced in Definition \ref{Def: epsilon tensor product}. In the case where $|\raisebox{.4ex}{.}|$ is Archimedean, one has
\begin{multline*}h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)=\int_{\mathbb S_0\times\cdots\times \mathbb S_d}\hspace{-1mm}\ln|R(z_0,\ldots,z_d)|\,\eta_{\mathbb S_0}(\mathrm{d}z_0)\otimes\cdots\otimes\eta_{\mathbb S_d}(\mathrm{d}z_d)+\frac 12\sum_{i=0}^d\delta_i\sum_{\ell=1}^{r_i}\frac{1}{\ell},\end{multline*}
where $\mathbb S_i$ is the unit sphere of $(E_{i,\mathbb C},\|\raisebox{.4ex}{.}\|_{i,\mathbb C})$, and $\eta_{\mathbb S_i}$ is the $U(E_{i,\mathbb C},\|\raisebox{.4ex}{.}\|_{i,\mathbb C})$-invariant Borel probability measure on $\mathbb S_{i,\sigma}$.
\end{theo}
\begin{proof} By Remark \ref{Rem: resultant extension}, \[R\otimes 1\in (S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee))\otimes_k\mathbb C_k\]
is the resultant of $X_{\mathbb C_k}$ with respect to $f_{0,\mathbb C_k},\ldots,f_{d,\mathbb C_k}$, which takes value $1$ at $(s_0,\ldots,s_d)$. Therefore, by extension of scalars, we may assume without loss of generality that $k$ is algebraically closed and $X$ is integral.
We treat firstly the non-Archimedean case.
By Lemma \ref{Lem: h s in h IX}, one has
\[h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)=h_{q^*(\boldsymbol{\overline L}_0),\ldots,q^*(\boldsymbol{\overline L}_d)}^{q_0^*(\boldsymbol{\alpha}_0),\ldots,q_d^*(\boldsymbol{\alpha}_d)}(I_X)+\sum_{i=0}^d\delta_i\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}.\]
By Lemma \ref{Lem: h div R h IX}, this is also equal to
\[h_{\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}(\operatorname{div}(R))+\sum_{i=0}^d\delta_i\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}.\]
By Lemma \ref{Lem: h div r et h p}, it is equal to
\[h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}(\check{\mathbb P})+\ln\|R\|+\sum_{i=0}^d\delta_i\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}.\]
By Proposition \ref{Pro: hight of multi-projective} and the relation (see Definition \ref{Def:resultant precise})
\[R(s_0,\ldots,s_d)=1,\]
we obtain
\[h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)=\ln\|R\|.\]
The case where $|\raisebox{.4ex}{.}|$ is Archimedean is quite similar. We have
\[\begin{split}
&\quad\; h_{\overline{M_0'},\ldots,\overline{M_d'}}^{s_0,\ldots,s_d}(X)=h_{q^*(\boldsymbol{\overline L}_0),\ldots,q^*(\boldsymbol{\overline L}_d)}^{q_0^*(\boldsymbol{\alpha}_0),\ldots,q_d^*(\boldsymbol{\alpha}_d)}(I_X)+\sum_{i=0}^d\delta_i\big(\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}-\sigma_{r_i-1}\big)
\\&=h_{\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}(\operatorname{div}(R))+\sum_{i=0}^d\delta_i\big(\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}-\sigma_{r_i-1}\big)\\
&=h_{\overline L,\overline{\boldsymbol{L}}_0,\ldots,\overline{\boldsymbol{L}}_d}^{R,\pi_0^*(\boldsymbol{\alpha}_0),\ldots,\pi_d^*(\boldsymbol{\alpha}_d)}(\check{\mathbb P})\\
& \qquad\quad+\int_{\mathbb S_0\times\cdots\times \mathbb S_d}\ln|R(z_0,\ldots,z_d)|\,\eta_{\mathbb S_0}(\mathrm{d}z_0)\otimes\cdots\otimes\eta_{\mathbb S_d}(\mathrm{d}z_d)\\
&\qquad\quad+\sum_{i=0}^d\delta_i\big(\ln\|\alpha_{i,0}\wedge\cdots\wedge\alpha_{i,r_i}\|_{i,*,\det}-\sigma_{r_i}\big)+\frac 12\sum_{i=0}^d\delta_i\sum_{\ell=1}^{r_i}\frac{1}{\ell}\\
&=\int_{\mathbb S_0\times\cdots\times \mathbb S_d}\ln|R(z_0,\ldots,z_d)|\,\eta_{\mathbb S_0}(\mathrm{d}z_0)\otimes\cdots\otimes\eta_{\mathbb S_d}(\mathrm{d}z_d)+\frac 12\sum_{i=0}^d\delta_i\sum_{\ell=1}^{r_i}\frac{1}{\ell},
\end{split}\]
where the first equality comes from Lemma \ref{Lem: h s in h IX}, the second one from Lemma \ref{Lem: h div R h IX}, the third one from Lemma \ref{Lem: h div R etc}, and the last one from Proposition \ref{Pro: hight of multi-projective}.
\end{proof}
\begin{rema} Note that the result of Theorem \ref{Thm: equality of local height} does not depend on the choice of the vectors $\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d$.
If we are only interested in the equalities in the theorem, we could choose $\boldsymbol{\alpha}_0,\ldots,\boldsymbol{\alpha}_d$ carefully to make the computation simpler. However, the formulae in the lemmas proving the theorem are of their proper interest, especially in the computations of height of homogeneous hypersurfaces in multi-projective spaces, and hence are worth to be detailed.
\end{rema}
\begin{prop}\label{prop:invariance:local:intersection:field:extension}
Assume that the absolute value $|\raisebox{.4ex}{.}|$ is non-Archimedean. Let $K$ be an extension of $k$, on which the absolute value extends. We assume that $K$ is complete with respect to the extended absolute value. Let $X$ be a projective scheme over $\operatorname{Spec} k$, $d$ be the dimensional of $X$, and $\overline D_i=(D_i,g_i)$ be a family of integrable metrised Cartier divisors, where $i\in\{0,\ldots,d\}$, such that $D_0,\ldots,D_d$ intersect properly. For each $i\in\{0,\ldots,d\}$, let $\overline D_{i,K}:=(D_{i,k},g_{i,k})$. Then the following equality holds:
\begin{equation}\label{Equ: intersection invariant by extenison of scalars}
(\overline D_0\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}=(\overline D_{0,K}\cdots\overline D_{d,K})_{(K,|\raisebox{.4ex}{.}|)}.
\end{equation}
\end{prop}
\begin{proof}
{\bf Step 1}: In this step, we assume that $D_0,\ldots,D_d$ are very ample, and, for each $i\in\{0,\ldots,d\}$, there exist a positive integer $m_i$ and an ultrametric norm $\|\raisebox{.4ex}{.}\|_i$ on $E_i=H^0(X,\mathcal O_X(m_iD_i))$, such that $\varphi_{g_i}$ identifies with the quotient metric induced by $\|\raisebox{.4ex}{.}\|_i$.
For each $i\in\{0,\ldots,d\}$, let $f_i:X\rightarrow\mathbb P(E_i)$ be the canonical closed embedding. Note that $\mathcal O_X(m_iD_i)\cong f_i^*(\mathcal O_{E_i}(1))$. In order to simplify the notation, we let $L_i$ be the line bundle $\mathcal O_X(m_iD_i)$ and $s_i$ be the canonical regular meromorphic section of $L_i$. Let $R$ be the resultant
\[R_{f_0,\ldots,f_d}^{X,s_0,\ldots,s_d},\]
which is considered as an element of
\[S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee),\]
and \[\delta_i=(D_0\cdots D_{i-1}D_{i+1}\cdots D_d).\]
Then, by Theorem \ref{Thm: equality of local height}, the equality
\[(\overline D_0\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}=\ln\|R\|\]
holds, where $\|{\raisebox{.4ex}{.}}\|$ denotes the $\varepsilon$-tensor product of $\varepsilon$-tensor powers of $\|\raisebox{.4ex}{.}\|_{i,*}$.
Similarly, by Remarks \ref{Rem: resultant extension} and \ref{Rem: extensions scalars}, one has
\[(\overline{D}_{0,K},\ldots,\overline D_{d,K})_{(K,|\raisebox{.4ex}{.}|)}=\ln\|R\otimes 1\|_{K}.\]
By \cite[Proposition 1.3.1 (1)]{CMArakelovAdelic}, one has $\|R\otimes 1\|_K=\|R\|$. Hence the equality \eqref{Equ: intersection invariant by extenison of scalars} follows.
{\bf Step 2}: In this step, we still assume that $D_0,\ldots,D_d$ are very ample. However, the Green functions $g_0,\ldots,g_d$ are only supposed to be plurisubharmonic.
For any $i\in\{0,\ldots,d\}$ and any positive integer $m$, let $g_i^{(m)}$ be the Green function associated with the quotient metric $\varphi_{g_i}^{(m)}$ as in Definition \ref{Def: Fubini-Study}, and let $\overline D_i^{(m)}=(D_i,g_i^{(m)})$. By Proposition \ref{Pro: convergence of varphin}, we obtain that, for any $i\in\{0,\ldots,d\}$,
\begin{equation}\label{Equ: convergence of Green functions}\lim_{m\rightarrow+\infty}\sup_{x\in X^{\mathrm{an}}}|g_i^{(m)}-g_i|(x)=0,\end{equation}
Therefore, by Corollary \ref{cor:intersection:trivial:divisor} (see also \S\ref{Sec: Local intersection number over a general field}), we obtain
\begin{equation}\label{Equ: intersection m}\lim_{m\rightarrow+\infty}(\overline D_0^{(m)}\cdots\overline D_d^{(m)})_{(k,|\raisebox{.4ex}{.}|)}=(\overline D_0\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}.\end{equation}
Moreover, \eqref{Equ: convergence of Green functions} leads to
\[\lim_{m\rightarrow+\infty}\sup_{x\in X_K^{\mathrm{an}}}|g_{i,K}^{(m)}-g_{i,K}|(x)=0.\]
Hence, similarly to \eqref{Equ: intersection m}, we have
\[\lim_{m\rightarrow+\infty}(\overline D_{0,K}^{(m)}\cdots\overline D_{d,K}^{(m)})_{(K,|\raisebox{.4ex}{.}|)}=(\overline D_{0,K}\cdots\overline D_{d,K})_{(K,|\raisebox{.4ex}{.}|)}\]
Note that, by \cite[Proposition 1.3.16]{CMArakelovAdelic}, $g_{i,K}^{(m)}$ is also the Green function associated with a quotient metric. Therefore, by the result in Step 1, we obtain that
\[(\overline D_0^{(m)}\cdots\overline D_d^{(m)})_{(k,|\raisebox{.4ex}{.}|)}=(\overline D_{0,K}^{(m)}\cdots\overline D_{d,K}^{(m)})_{(K,|\raisebox{.4ex}{.}|)}\]
for any $m$, so that, by passing to limit when $m\rightarrow+\infty$, we obtain \eqref{Equ: intersection invariant by extenison of scalars}.
{\bf Step 3:} We now treat the general case. For each $i\in\{-1,0,\ldots,d\}$, we consider the following condition $(C_r)$:
\begin{quote}
\it For any $i\in\{0,\ldots,d\}$ such that $1\leqslant i\leqslant r$, the Cartier divisor $D_i$ is very ample and the Green function $g_i$ is plurisubharmonic.
\end{quote}
We will show by inverted induction on $r$ that, under the condition $(C_r)$, the equality \eqref{Equ: intersection invariant by extenison of scalars} holds. Note that the initial case where $r=d$ is proved in Step 2. We suppose that the equality \eqref{Equ: intersection invariant by extenison of scalars} is true under the condition $(C_r)$ and will prove it under the condition $(C_{r-1})$. Since $\overline D_r$ is integrable, there exists very ample Cartier divisors $A_r'$ and $A_r''$, and plurisubharmonic Green functions $h_r'$ and $h_r''$ of $A_r'$ and $A_r''$, respectively, such that
\[(D_r,g_r)=(A_r',h_r')-(A_r'',h_r'').\]
By Claim \ref{Claim: decomposition} (see also Remark \ref{Rem: further explanation to Claim}), there exists a very ample Cartier divisor $B_r$ such that \[(D_0,\ldots,D_{r-1},B_r+A_r',D_{r+1},\ldots,D_d) \in\mathcal {IP}_X^{(d)}.\] Since $\mathcal {IP}_X^{(d)}$ is a multilinear subset of $\operatorname{Div}(X)^{n+1}$, we obtain that \[(D_0,\ldots,D_{r-1},B_r+A_r'',D_{r+1},\ldots,D_d) \in\mathcal {IP}_X^{(d)}.\] We pick arbitrarily a plurisubharmonic Green function $l_r$ on $B_r$. Let
\[\overline D_r'=(B_r+A_r',l_r+h_r'),\quad \overline D_r''=(B_r+A_r'',l_r+h_r'')\]Then the induction hypothesis shows that
\begin{gather*}(\overline D_0\cdots\overline D_{r-1}\overline D_r'\overline D_{r+1}\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}=(\overline D_{0,K}\cdots\overline D_{r-1,K}\overline D_{r,K}'\overline D_{r+1,K}\cdots\overline D_{d,K})_{(K,|\raisebox{.4ex}{.}|)},\\
(\overline D_0\cdots\overline D_{r-1}\overline D_r''\overline D_{r+1}\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}=(\overline D_{0,K}\cdots\overline D_{r-1,K}\overline D_{r,K}''\overline D_{r+1,K}\cdots\overline D_{d,K})_{(K,|\raisebox{.4ex}{.}|)}.\end{gather*}
Taking the difference, we obtain \eqref{Equ: intersection invariant by extenison of scalars}
\end{proof}
\begin{rema}\label{Rem: extension of scalars archimedean}
If $K$ is a subfield of $\mathbb C_k$, the assertion of Proposition~\ref{prop:invariance:local:intersection:field:extension}
is obvious by its definition (cf. Definition~\ref{Def: intersection pairing general case}). In particular, the statement of Proposition \ref{prop:invariance:local:intersection:field:extension} is also true when $|\raisebox{.4ex}{.}|$ is Archimedean.
Proposition~\ref{prop:invariance:local:intersection:field:extension} guarantees the invariance of intersection number under any field extension.
\end{rema}
\section{Trivial valuation case}
In this section, we fix a field $k$ and equip it with the trivial absolute value $|\raisebox{.4ex}{.}|$, namely $|a|=1$ for any $a\in k^{\times}$. Let $K=k(T)$ be the field of rational functions over $k$, and $u$ be a positive constant such that $u\not=1$.
By Lemma \ref{lem:Gauss:lemma}, there exists a non-Archimedean absolute value $|\raisebox{.4ex}{.}|_u$ on $K$ which extends the above absolute value $|\raisebox{.4ex}{.}|$ on $k$, such that,
\[\forall\,f=a_0+a_1T+\cdots+a_nT^n\in k[T],\quad |f|_u=\max_{i\in\{0,\ldots,n\}}|a_i|u^i.\]
Note that $|\raisebox{.4ex}{.}|_u$ is not trivial.
\begin{defi}\label{Def: intersection in trivial valuation case}
Let $X$ be a projective scheme of dimension $d$ over $\operatorname{Spec} k$. If $\overline D_i=(D_i,g_i)$, $i\in\{0,\ldots,d\}$, is a family of integrable metrized Cartier divisors, such that $D_0,\ldots,D_d$ intersect properly. We denote by $(\overline D_0\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}$ the intersection number
\[((D_{0,K},g_{0,K})\cdots(D_{d,K},g_{d,K}))_{(K,|\raisebox{.4ex}{.}|_u)}\]
\end{defi}
\begin{enonce}{Notation and assumptions}\label{Notation and assumptions}\rm
Let $((E_i,\|\raisebox{.4ex}{.}\|_i))_{i=0}^d$ be a family of finite-dimensional ultrametrically normed vector space over $k$. For any $i\in\{0,\ldots,d\}$, let $r_i=\dim_{k}(E_i)-1$, $f_i:X\rightarrow\mathbb P(E_i)$ be a closed immersion, and $s_i$ an element of $E_i$, viewed as a global section of $\mathcal O_{E_i}(1)$. We assume that the restriction of $s_i$ to $X$ defines a regular meromorphic section of $L_i:=\mathcal O_{E_i}(1)|_X$ and that the Cartier divisors
\[D_i=\operatorname{div}(s_i|_X), \quad i\in\{0,\ldots, d\}\]
intersect properly. We equip each $D_i$ with the Green function associated with the quotient metric induced by $\|\raisebox{.4ex}{.}\|_i$. Let $R$ be the resultant
\[R=R_{f_0,\ldots,f_d}^{X,s_0,\ldots,s_d}\in S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee),\]
where
\[\delta_i=(D_0\cdots D_{i-1}D_{i+1}\cdots D_d).\]
\end{enonce}
\begin{prop} Under Notation and assumptions \ref{Notation and assumptions},
the following equality holds
\begin{equation}\label{Equ: local intersection number trivial valuation}(\overline D_0\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}=\ln\|R\|,\end{equation}
where $\|\raisebox{.4ex}{.}\|$ denotes the $\varepsilon$-tensor product of $\varepsilon$-tensor power norms of $\|\raisebox{.4ex}{.}\|_{i,*}$.
\end{prop}
\begin{proof}
Under the isomorphism of $K$-vector spaces
\[(S^{\delta_0}(E_0^\vee)\otimes_k\cdots\otimes_kS^{\delta_d}(E_d^\vee))\otimes_kK\cong S^{\delta_0}(E_{0,K}^\vee)\otimes_K\cdots\otimes_KS^{\delta_d}(E_{d,K}^\vee),\]
the element $R\otimes 1$ coincides with the resultant (see Remark \ref{Rem: resultant extension})
\[R_{f_{0,K},\ldots,f_{d,K}}^{X_K,s_{0}\otimes 1,\ldots,s_{d}\otimes 1}.\]
By Theorem \ref{Thm: equality of local height} and Remark \ref{Rem: extensions scalars}, one has
\[(\overline D_0\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}=\ln\|R\otimes 1\|_K.\]
By \cite[Proposition 1.3.1 (1)]{CMArakelovAdelic}, one has $\|R\otimes 1\|_K=\|R\|$. Hence we obtain the equality \eqref{Equ: local intersection number trivial valuation}.
\end{proof}
\begin{coro}
Let $X$ be a projective scheme of dimension $d$ over $\operatorname{Spec} k$. If $\overline D_i=(D_i,g_i)$, $i\in\{0,\ldots,d\}$, is a family of integrable metrized Cartier divisors, such that $D_0,\ldots,D_d$ intersect properly. Then the intersection number $(\overline D_0\cdots\overline D_d)_{(k,|\raisebox{.4ex}{.}|)}$ does not depend on
the choice of $u$.
\end{coro}
\begin{proof}
By the multi-linearity of the intersection number, it suffices to treat the case where all Cartier divisors $D_i$ are very ample and all $g_i$ are plurisubharmonic. Moreover, by Proposition \ref{Pro: convergence of varphin} and Corollary \ref{cor:intersection:trivial:divisor} we can further reduce the problem to the case of Notation and assumptions \ref{Notation and assumptions}. In that case the assertion follows from \eqref{Equ: local intersection number trivial valuation}.
\end{proof}
\begin{rema}
By using Remark~\ref{rem:properties:local:intersection:general:field}, one has the following properties.
\begin{enumerate}[label=\rm(\arabic*)]
\item The set $\widehat{{\mathcal{IP}}}_X$ forms a symmetric multi-linear subset of the group $\operatorname{\widehat{Int}}(X)^{d+1}$. Moreover, the function of local intersection number
\[\big( (D_0, g_0) \cdots (D_d, g_d) \big)\longmapsto\big( (D_0, g_0) \cdots (D_d, g_d) \big)_v\]
form a symmetric multi-linear map from $\widehat{{\mathcal{IP}}}_X$ to $\mathbb R$.
\item
Let $\pi : Y \to X$ be a surjective morphism of geometrically integral projective schemes over $k$.
We set $e = \dim X$ and $d = \dim Y$.
Let $(D_0, g_0), \ldots, (D_d, g_d)$ be integrable metrized Cartier divisors on $X$ such that
$(\pi^*(D_0), \ldots, \pi^*(D_d)) \in {\mathcal{IP}}_Y$.
Then one has the following:
\begin{enumerate}[label=\rm(\roman*)]
\item If $d > e$, then $(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d))_v = 0$.
\item If $d = e$ and $(D_0,\ldots,D_d) \in {\mathcal{IP}}_X$, then
\[(\pi^*(D_0, g_0) \cdots \pi^*(D_d, g_d))_v = (\deg \pi)((D_0, g_0) \cdots (D_d, g_d))_v.\]
\end{enumerate}
\item Let $f$ be a regular meromorphic function on $X$ and $(D_1,g_1),\ldots,(D_d,g_d)$ be integrable metrized Cartier divisors on $X$ such that $(\operatorname{div}(f),D_1,\ldots,D_d)\in{\mathcal{IP}}_X$. Then
\[\big(\widehat{\operatorname{div}}(f) \cdot (D_1, g_1) \cdots (D_{d}, g_{d})\big)_v = 0. \]
Note that $-\log|f|(x^{\mathrm{an}}) = 0$ for any $x \in X_{(0)}$ in Remark~\ref{rem:properties:local:intersection:general:field}
because $|\raisebox{.4ex}{.}|$ is trivial.
\end{enumerate}
Let $(L_0, \varphi_0), \ldots, (L_d, \varphi_d)$ be a family of integrable metrized invertible ${\mathscr{O}}_X$-modules.
By the property (3), the local intersection number $\big( (L_0, \varphi_0) \cdots (L_d, \varphi_d) \big)_v$ is well-defined.
\end{rema}
\begin{rema}
In \cite{Chen_Moriwaki2020}, an intersection product of metrized divisors has been introduced in the setting of curves over a trivially valued field $(k,|\raisebox{.4ex}{.}|)$. Let $X$ be a regular projective curve over $\operatorname{Spec} k$. Recall that the Berkovich space $X^{\operatorname{an}}$ is an infinite tree
\vspace{3mm}
\begin{center}
\begin{tikzpicture}
\filldraw(0,1) circle (2pt) node[align=center, above]{$\eta_0$};
\filldraw(-3,0) circle (2pt) ;
\draw (-1,0) node{$\cdots$};
\filldraw(-2,0) circle (2pt) ;
\filldraw(-0,0) circle (2pt) node[align=center, below]{$x_0$} ;
\filldraw(1,0) circle (2pt) ;
\draw (2,0) node{$\cdots$};
\filldraw(3,0) circle (2pt) ;
\draw (0,1) -- (0,0);
\draw (0,1) -- (-3,0);
\draw (0,1) -- (1,0);
\draw (0,1) -- (-2,0);
\draw (0,1) -- (3,0);
\end{tikzpicture}
\end{center}
\vspace{3mm}
where the root point $\eta_0$ corresponds to the generic point of $X$ together with the trivial absolute value on $\kappa(\eta)$, and each leaf $x_0$ corresponds to the closed point $x$ together with the trivial absolute value on $\kappa(x)$. Moreover, each branch $\mathopen{]}\eta_0,x_0\mathclose{[}$ is parametrized by $\mathrm{]}0,+\infty\mathrm{[}$, where $t\in\mathrm{]}0,+\infty\mathrm{[}$ corresponds to the generic point $\eta$ together with the absolute value \[|\raisebox{.4ex}{.}|_{x,t}=\exp(-t\operatorname{ord}_x(\raisebox{.4ex}{.})).\]
We denote by $t(\raisebox{.4ex}{.}):X^{\mathrm{an}}\rightarrow [0,+\infty]$ the parametrization map, where $t(\eta_0)=0$ and $t(x_0)=+\infty$.
Let $D$ be a Cartier divisor on $X$. Recall that a Green function $g$ of $D$ is of the form
\[g=g_D+\varphi_g,\]
where $g_D$ is the canonical Green function of $D$, which is defined as
\[g_D(\xi)=\operatorname{ord}_x(D)t(\xi),\]
and $\varphi_g$ is a continuous real-valued function on $X^{\mathrm{an}}$ (which is hence bounded since $X^{\mathrm{an}}$ is compact). Then, the intersection number of two integrable metrized Cartier divisor $\overline D_0=(D_0,g_0)$ and $\overline D_1=(D_1,g_1)$ has been defined as
\begin{equation}\label{Equ: alternative intersection product}g_1(\eta_0)\deg(D_0)+g_0(\eta_0)\deg(D_1)-\sum_{x\in X^{(1)}}[\kappa(x):k]\int_0^{+\infty}\varphi'_{g_0\circ\xi_x}(t)\varphi'_{g_1\circ\xi_x}(t)\,\mathrm{d}t,
\end{equation}
where $X^{(1)}$ is the set of closed points of $X$, $\xi_x:[0,+\infty]\rightarrow[\eta_0,x_0]$ is the map sending $t\in[0,+\infty]$ to the point in $[\eta_0,x_0]$ of parameter $t$, and the function $\varphi_{g_1\circ\xi_x}'(\raisebox{.4ex}{.})$ should be considered as right-continuous version of the Radon-Nikodym density of the function $\varphi_{g_1\circ\xi_x}(\raisebox{.4ex}{.})$ with respect to the Lebesgue measure.
Let $(L, \varphi_0)$ and $(L_1, \varphi_1)$ be integrable metrized invertible ${\mathscr{O}}_X$-modules.
By \cite[Remark~7.3]{Chen_Moriwaki2020}, the above local intersection number with respect to $(L, \varphi_0)$ and $(L_1, \varphi_1)$ is well-defined.
To destinguish this intersection number with the intersection number defined in Definition~\ref{Def: intersection in trivial valuation case},
it is denoted by $((L_0, \varphi_0) \cdot (L_1, \varphi_2))_v'$.
Let $(E,\|\raisebox{.4ex}{.}\|)$ be a ultrametrically normed vector space and $f:X\rightarrow\mathbb P(E)$ be a closed embedding. Let $L$ be the pull-back of the universal invertible sheaf $\mathcal O_E(1)$ by $f$. We equip $L$ with the quotient metric induced by the norm $\|\raisebox{.4ex}{.}\|$. Let $E\otimes_{k}\kappa(\eta)\rightarrow L(\eta)$ be the universal quotient $\kappa(\eta)$-linear map and
\[L(\eta)^\vee\longrightarrow E_{\kappa(\eta)}^\vee\]
be the dual linear map. By this injective linear map we identify $L(\eta)^\vee$ with a one-dimensional vector subspace of $E_{\kappa(\eta)}^\vee$. Consider an element $s\in E\setminus\{0\}$ which is viewed as a global section of $\mathcal O_E(1)$. We suppose that $s$ does not vanish at $f(\eta)$. Let $x$ be a closed point of $X$, $t\in\mathopen{[}0,+\infty\mathclose{[}$, and $\xi$ the element of $\mathopen{[}\eta_0,x_0\mathclose{[}$ having $t$ as its parameter, one has
\[|s|(\xi)=\frac{|\beta(s)|_{x,t}}{\|\beta\|_{x,t,*}},\]
where $\beta$ is an arbitrary non-zero element of $L(\eta)^\vee$, and $\|\raisebox{.4ex}{.}\|_{x,t}$ is the norm on $E_{\kappa(\eta)}=E\otimes_k\kappa(\eta)$ constructed from $\|\raisebox{.4ex}{.}\|$ by extension of scalars to $(\kappa(\eta),|\raisebox{.4ex}{.}|_{x,t})$. In particular, if we pick an orthogonal basis of $(E,\|\raisebox{.4ex}{.}\|)$ of the form $\{s_0,s_1,\ldots,s_r\}$ with $s_0=s$, and pick a vector $\beta\in L(\eta)^\vee$ of the form
\[\beta=\lambda_0s_0^\vee+\lambda_1s_1^\vee+\cdots+\lambda_rs_r^\vee,\quad (\lambda_0,\ldots,\lambda_r)\in\kappa(\eta)^r,\]
then one has
\[|s|(\xi)=\frac{|\lambda_0|_{x,t}}{
\displaystyle\max_{i\in\{0,\ldots,r\}}|\lambda_i|_{x,t}\cdot\|s_i\|^{-1}}=
\min_{i\in\{0,\ldots,r\}}
\|s_i\|\exp(\operatorname{ord}_x(\lambda_i)-\operatorname{ord}_x(\lambda_0))^t.\]
Therefore, the Green function $g$ of $\operatorname{div}(s)$ corresponding to the quotient metric is given by
\[g(\xi)=-\min_{i\in\{0,\ldots,r\}}\Big(t\big(\operatorname{ord}_x(\lambda_i)-\operatorname{ord}_x(\lambda_0)\big)+\ln\|s_i\|\Big)\text{ on $\mathopen{]}\eta_0,x_0\mathclose{[}$}.\]
Moreover, one has
\[g(\eta_0)=-\min_{\begin{subarray}{c}i\in\{0,\ldots,r\}\\
\lambda_i\neq 0
\end{subarray}}\ln\|s_i\|\]
and
\begin{equation}\label{Equ: computation of varphi g}\varphi_g(\xi)=-\min_{i\in\{0,\ldots,r\}}
\bigg(t\Big(\operatorname{ord}_x(\lambda_i)-\min_{j\in\{0,\ldots,r\}}\operatorname{ord}_x(\lambda_j)\Big)+\ln\|s_i\|\bigg).\end{equation}
\end{rema}
We now illustrate the comparison of \eqref{Equ: alternative intersection product} and the local intersection product introduced in Definition \ref{Def: intersection in trivial valuation case} in the particular case.
\begin{prop}
Let $(E,\|\raisebox{.4ex}{.}\|)$ be a finite-dimensional
ultrametrically normed vector space over $k$.
Let $X$ be a regular projective curve over $k$ and $L$ be an invertible ${\mathscr{O}}_X$-module.
We assume that there is a surjective homomorphism $E \otimes_k {\mathscr{O}}_X \to L$.
Let $\varphi$ be the Fubini-Study metric of $L$ induced by the above homomorphism and $(E,\|\raisebox{.4ex}{.}\|)$.
If either (1) $X = {\mathbb{P}}^1_k$ and $L={\mathscr{O}}_{{\mathbb{P}}^1}(1)$, or (2) $\dim E = 2$, then
\[
((L, \varphi) \cdot (L, \varphi))_v = ((L, \varphi) \cdot (L, \varphi))_v'.
\]
\end{prop}
\begin{proof}
(1) (the case where $X = {\mathbb{P}}^1$ and $L = {\mathscr{O}}_{{\mathbb{P}}^1}(1)$)
Let $E \to H^0(X, L)$ be the natural homomorphism, which is surjective because
$E \otimes_k {\mathscr{O}}_X \to L$ is surjective.
Let $\|\raisebox{.4ex}{.}\|'$ be the quotient norm of $H^0(X, L)$ by $\|\raisebox{.4ex}{.}\|$ of $E$.
Then the Fubini-Study metric of $L$ induced by $(H^0(X, L), \|\raisebox{.4ex}{.}\|')$ coincides with $\varphi$,
so that we may assume that $E = H^0(X, L)$ and $X = {\mathbb{P}}(E)$.
Let $\{s_0,s_1\}$ be an orthogonal basis of $E$.
Let $\overline D_0=(D_0,g_0)$ and $\overline D_1=(D_1,g_1)$ be the metrized Cartier divisors $\operatorname{\widehat{\mathrm{div}}}(s_0)$ and $\operatorname{\widehat{\mathrm{div}}}(s_1)$, respectively.
We need to prove
\begin{equation}\label{Equ: equality of two definitions}\begin{split}(\overline D_0\cdot\overline D_1)_{(k,|\raisebox{.4ex}{.}|)}=g_1(\eta_0)&\deg(D_0)+g_0(\eta_0)\deg(D_1)\\
&-\sum_{x\in\mathbb P(E)^{(1)}}[\kappa(x):k]\int_0^{+\infty}\varphi_{g_0\circ\xi_x}'(t)\varphi_{g_1\circ\xi_x}'(t)\,\mathrm{d}t.
\end{split}\end{equation}
Note that $\{s_0,s_1\}$ forms an orthogonal basis of $E\otimes_kK$ (see \cite[Proposition 1.3.13]{CMArakelovAdelic}). Therefore, Proposition \ref{Pro: height projective space} leads to
\[(\overline D_0\cdot\overline D_1)_{(k,|\raisebox{.4ex}{.}|)}=-\ln\|s_0\wedge s_1\|_{K,\det}=-\ln\|s_0\|_K-\ln\|s_1\|_K=-\ln\|s_0\|-\ln\|s_1\|,\]
where the last equality comes from \cite[Proposition 1.3.1]{CMArakelovAdelic}.
Note that $\varphi = \varphi_{g_0}=\varphi_{g_1}$, which we denote by $\varphi_g$.
Note that the field of rational functions of $\mathbb P(E)$ is given by the subfield of $k(E)$ generated by $\tau=s_1/s_0$. Moreover, the universal one-dimensional $k(\tau)$-linear subspace
\[L^\vee\otimes_kk(\tau)\longrightarrow E^\vee\otimes_kk(\tau)\]
is spanned by the vector
\[\beta=s_0^\vee+\tau s_1^\vee,\]
where $\{s_0^\vee,s_1^\vee\}$ is the dual basis of $\{s_0,s_1\}$. Let $x_0$ and $x_1$ be the vanishing point of $s_0$ and $s_1$, respectively. These are rational points of $\mathbb P(E)$. By \eqref{Equ: computation of varphi g}, we obtain that, for $x\in\mathbb P(E)^{(1)}\setminus\{x_0,x_1\}$, one has
\begin{equation}\label{Equ: varphi g xi x}\varphi_{g\circ\xi_x}(t)=-\min\{\ln\|s_0\|,\ln\|s_1\|\},\end{equation}
and
\begin{gather*}
\varphi_{g\circ\xi_{x_0}}(t)=-\min\{t+\ln\|s_0\|,\ln\|s_1\|\},\\
\varphi_{g\circ\xi_{x_1}}(t)=-\min\{\ln\|s_0\|,t+\ln\|s_1\|\}.
\end{gather*}
Therefore, one obtains that
\begin{gather*}
\varphi_{g\circ\xi_{x_0}}'=-1\hspace{-0.25em}\mathrm{l}_{[0,\max\{\ln\frac{\|s_1\|}{\|s_0\|},0\})},\\
\varphi_{g\circ\xi_{x_1}}'=-1\hspace{-0.25em}\mathrm{l}_{[0,\max\{\ln\frac{\|s_0\|}{\|s_1\|},0\})},
\end{gather*}
which leads to
\[-\sum_{x\in\mathbb P(E)^{(1)}}[\kappa(x):k]\int_0^{+\infty}\varphi_{g_0\circ\xi_x}'(t)^2\,\mathrm{d}t=-\big|\ln\|s_0\|-\ln\|s_1\|\big|.\]
Moreover, \eqref{Equ: varphi g xi x} also implies that
\[g_0(\eta_0)=g_1(\eta_0)=-\min\{\ln\|s_0\|,\ln\|s_1\|\}. \]
Hence the right hand side of \eqref{Equ: equality of two definitions} is equal to
\[-2\min\{\ln\|s_0\|,\ln\|s_1\|\}-\big|\ln\|s_0\|-\ln\|s_1\|\big|=-\ln\|s_0\|-\ln\|s_1\|,\]
as desired.
\bigskip
(2) (the case where $\dim E = 2$)
Then one has a finite surjective morphism $f : X \to {\mathbb{P}}(E)$ such that $f^*({\mathscr{O}}_{{\mathbb{P}}(E)}(1)) = L$.
Let $\psi$ be the Fubini-Study metric induced by $E \otimes {\mathscr{O}}_{{\mathbb{P}}(E)} \to {\mathscr{O}}_{{\mathbb{P}}(E)}(1)$ and $\|\raisebox{.4ex}{.}\|$.
Then $f^*(\psi) = \varphi$, that is, $(L, \varphi) = f^*({\mathscr{O}}_{{\mathbb{P}}(E)}(1), \psi)$. Therefore one can see
\[
\begin{cases}
((L, \varphi) \cdot (L,\varphi))_v = \deg(f) (({\mathscr{O}}_{{\mathbb{P}}(E)}(1), \psi) \cdot ({\mathscr{O}}_{{\mathbb{P}}(E)}(1), \psi))_v, \\
((L, \varphi) \cdot (L,\varphi))_v' = \deg(f) (({\mathscr{O}}_{{\mathbb{P}}(E)}(1), \psi) \cdot ({\mathscr{O}}_{{\mathbb{P}}(E)}(1), \psi))_v'.
\end{cases}
\]
Thus the assertion follows from (1).
\end{proof}
\begin{rema}
The above proposition suggests that \eqref{Equ: alternative intersection product} should be equal to the intersection number introduced in Definition \ref{Def: intersection in trivial valuation case}. We expect that an explicit computation of the resultant in the projective curve case would establish such an equality by using Theorem \ref{Thm: equality of local height}.
\end{rema}
\chapter{Global intersection number}
Let $K$ be a field and $S = (K, (\Omega, {\mathcal A}, \nu), \phi)$ be an adelic curve the underlying field of which is $K$. For any $\omega\in\Omega$, we denote by $K_\omega$ the completion of $K$ with respect to $|\raisebox{.4ex}{.}|_\omega$.
We assume that, either the $\sigma$-algebra $\mathcal A$ is discrete, or there exists a countable subfield $K_0$ of $K$ which is dense in each $K_\omega$, $\omega\in\Omega$.
Let $X$ be a $d$-dimensional projective scheme over $K$. For any $\omega\in\Omega$, let $X_\omega$ be the fiber product $X\times_{\operatorname{Spec} K}\operatorname{Spec} K_\omega$. Note that the morphism $\operatorname{Spec} K_\omega\rightarrow\operatorname{Spec} K$ is flat. Hence the morphism of projection $X_\omega\rightarrow X$ is also flat (see \cite[$\text{IV}_1.(2.1.4)$]{EGA}).
\section{Reminder on adelic vector bundles}
\label{Sec: Reminder on adelic vector bundles}
\begin{defi}
Let $E$ be a finite-dimensional vector space over $K$. We call \emph{norm family} of $E$ any family $\xi=(\|\raisebox{.4ex}{.}\|_\omega)_{\omega\in\Omega}$, where each $\|\raisebox{.4ex}{.}\|_\omega$ is a norm on $E_{K_\omega}:=E\otimes_KK_\omega$. If for any $\omega\in\Omega$, the norm $\|\raisebox{.4ex}{.}\|_\omega$ is either ultrametric (when $\omega$ is non-Archimedean) or induced by an inner product (when $\omega$ is Archimedean), we say that the norm family $\xi$ is \emph{Hermitian}.
If $\xi=(\|\raisebox{.4ex}{.}\|_\omega)_{\omega\in\Omega}$ and $\xi'=(\|\raisebox{.4ex}{.}\|_\omega)_{\omega\in\Omega}$ are two norm families of $E$, we define the \emph{local distance function} of $\xi$ and $\xi'$ as the function
\[(\omega\in\Omega)\longmapsto d_\omega(\xi,\xi'):=\sup_{s\in E_{K_\omega}\setminus\{0\}}\Big|\ln\|s\|_\omega-\ln\|s\|_\omega'\Big|.\]
\end{defi}
\begin{exem}\label{Exe: norm family of a basis}
Let $\boldsymbol{e}=(e_i)_{i=1}^r$ be a basis of $E$ over $K$. For any $\omega\in\Omega$, we let $\|\raisebox{.4ex}{.}\|_{\boldsymbol{e},\omega}$ be the norm on $E_{K_\omega}$ such that, for $(\lambda_1,\ldots,\lambda_r)\in K_\omega^r$,
\[ \|\lambda_1e_1+\cdots+\lambda_re_r\|_{\boldsymbol{e},\omega}=\begin{cases}
\max\{|\lambda_1|_\omega,\ldots,|\lambda_r|_\omega\},&\text{$\omega$ is non-Archimedean},\\
(|\lambda_1|_\omega^2+\cdots+|\lambda_r|_\omega^2)^{1/2}, &\text{$\omega$ is Archimedean}.
\end{cases}\]
Then $(\|\raisebox{.4ex}{.}\|_{\boldsymbol{e},\omega})_{\omega\in \Omega} $ forms a Hermitian norm family of $E$, which we denote by $\xi_{\boldsymbol{e}}$.
\end{exem}
\begin{defi}
Let $E$ be a finite-dimensional vector space over $K$ and $\xi$ be a norm family on $E$. We say that $\xi$ is \emph{measurable} if for any $s\in E$ the function
\[(\omega\in\Omega)\longrightarrow\|s\|_\omega\]
is $\mathcal A$-measurable. We say that the family $\xi$ is \emph{strongly dominated} if there exists a basis $\boldsymbol{e}$ of $E$ such that the local distance function
\[(\omega\in\Omega)\longmapsto d_\omega(\xi,\xi_{\boldsymbol{e}})\]
is bounded from above by an integrable function. If $\xi$ is measurable and strongly dominated, we say that $(E,\xi)$ is a \emph{strongly adelic vector bundle}.
We refer the readers to \cite[\S4.1.4]{CMArakelovAdelic} for more details about this definition, and also to the Proposition 4.1.24 (1.b) for the measurability of the dual norm family of $\xi$ under our assumption on the adelic curve.
If $(E,\xi)$ is a strongly adelic vector bundle, for any non-zero element $s$ of $E$, the function
\[(\omega\in\Omega)\longmapsto \ln\|s\|_{\omega}\]
is integrable. We denote by $\widehat{\deg}_\xi(s)$, or simply by $\widehat{\deg}(s)$ if there is no ambiguity on the norm family, the integral
\[-\int_{\Omega}\ln\|s\|_\omega\,\nu(\mathrm{d}\omega),\]
called \emph{Arakelov degree} of $s$ (with respect to $\xi$).
\end{defi}
\begin{defi}
Let $X$ be a projective $K$-scheme and $L$ be an invertible $\mathcal O_X$-module. For any $\omega\in\Omega$, we denote by $L_\omega$ the pull-back of $L$ by the morphism of projection $X_\omega\rightarrow X$. We call \emph{metric family} of $L$ and family $\varphi=(\varphi_\omega)_{\omega\in\Omega}$, where each $\varphi_\omega$ is a continuous metric on $L_\omega$ (see Definition \ref{Defi: continuous metric}). Note that the dual metrics $(\varphi_\omega^\vee)_{\omega\in\Omega}$ form a metric family on the dual invertible $\mathcal O_X$-module $L^\vee$, which we denote by $\varphi^\vee$. If $L_1$ and $L_2$ are invertible $\mathcal O_X$-modules, and $\varphi_1$ and $\varphi_2$ are metric families on $L_1$ and $L_2$, respectively, then the metrics $(\varphi_{1,\omega}\otimes\varphi_{2,\omega})_{\omega\in\Omega}$ form a metric family of $L_1\otimes L_2$, which we denote by $\varphi_1\otimes\varphi_2$.
If $\varphi$ and $\varphi'$ are two metric metrics of the same invertible $\mathcal O_X$-module $L$, we define the \emph{local distance function} between $\varphi$ and $\varphi'$ as the function
\[(\omega\in\Omega) \longmapsto d_{\omega}(\varphi, \varphi') := \sup_{x\in X_{\omega}^{\operatorname{an}}}\bigg|\ln\frac{|\raisebox{.4ex}{.}|_{\varphi_\omega}(x)}{|\raisebox{.4ex}{.}|_{\varphi_\omega'}(x)}\bigg|\]
\end{defi}
\begin{rema}
In the case where $X$ is the spectrum of a finite extension $K'$ of $K$, an invertible $\mathcal O_X$-module $L$ can be considered as a one-dimensional vector space over $K'$, and a metric family on $L$ identifies with a norm family of $L$ if we consider the adelic curve $S\otimes_KK'$.
\end{rema}
\begin{defi}
Let $f:Y\rightarrow X$ be a projective $K$-morphism of projective $K$-schemes. Let $L$ be an invertible $\mathcal O_X$-module, equipped with a metric family $\varphi=(\varphi_\omega)_{\omega\in\Omega}$. For any $\omega\in\Omega$, let $f_\omega:Y_\omega\rightarrow X_\omega$ be the $K_\omega$-morphism induce by $f$ by extension of scalars. Then, for any $\omega\in\Omega$, the metric $\varphi_\omega$ induces by pull-back a continuous metric $f_\omega^*(\varphi_\omega)$ on $f_\omega^*(L_\omega)$ such that, for any $y\in Y_\omega^{\operatorname{an}}$ and any $\ell\in L_\omega(f^{\operatorname{an}}(y))$, one has
\[|f_\omega^*(\ell)|_{f_\omega^*(\varphi_\omega)}(y)=|\ell|_{\varphi_\omega}(f^{\operatorname{an}}(y)).\]
We denote by $f^*(\varphi)$ the metric family $(f_\omega^*(\varphi_\omega))_{\omega\in\Omega}$ and call it the \emph{pull-back of $\varphi$ by $f$}. In the case where $f$ is an immersion, $f^*(\varphi)$ is also called \emph{restriction of $\varphi$}.
\end{defi}
\begin{exem}\label{Exe: quotient metric family}
A natural example of metric family is the \emph{quotient metric family} induced by a norm family. Denote by $\pi:X\rightarrow\operatorname{Spec} K$ the structural morphism. Let $E$ be a finite-dimensional vector space over $K$ and $f:\pi^*(E)\rightarrow L^{\otimes n}$ be a surjective homomorphism of $\mathcal O_X$-modules, where $n$ is a positive integer. For any $\omega\in\Omega$, the homomorphism $f$ induces by pull-back a surjective homomorphism of $\mathcal O_{X_\omega}$-modules $f_\omega:\pi_{K_\omega}^*(E)\rightarrow L_\omega$.
Assume given a norm family $\xi=(\|\raisebox{.4ex}{.}\|_\omega)_{\omega\in\Omega}$ of $E$. We denote by $\varphi_{\xi}$ the metric family of $L$ consisting of quotient metrics associated with $\|\raisebox{.4ex}{.}\|_{\omega}$ (see Example \ref{Exa: Fubini-Study} \ref{Item: quotient metrics}), and call it the \emph{quotient metric family} induced by $\xi$.
Assume that the norm family $\xi$ is Hermitian. For each $\omega\in\Omega$, let $\varphi_{\xi,\omega}^{\operatorname{ort}}$ be the orthogonal quotient metric induced by $\|\raisebox{.4ex}{.}\|_\omega$ (see Definition \ref{Def: Fubini-Study metric}). Note that this metric coincides with $\varphi_{\xi,\omega}$ when $|\raisebox{.4ex}{.}|_\omega$ is non-Archimedean or $K_\omega$ is complex. The metric family $\varphi_{\xi}^{\operatorname{ort}}$ is called \emph{orthogonal quotient metric family} induced by $\xi$.
\end{exem}
\begin{exem}
Let $X$ be a projective $K$-scheme, $L$ be an invertible $\mathcal O_X$-module, and $\varphi=(\varphi_\omega)_{\omega\in\Omega}$ be a metric family on $L$. Let $K'/K$ be an algebraic extension of the field $K$, and \[S\otimes K'=(K',(\Omega',\mathcal A',\nu'),\phi')\] be the corresponding algebraic covering of the adelic curve $S$ (see \S\ref{Sec: algebraic coverings}). Recall that $\Omega'$ is defined as $\Omega\times_{M_K,\phi}M_{K'}$, where $M_K$ and $M_{K'}$ are the sets of all absolute values of $K$ and of $K'$, respectively.
Let $X'$ be the fiber product $X\times_{\operatorname{Spec} K}\operatorname{Spec} K'$ and $L'$ be the pull-back of $L$ on $X'$. If $\omega'$ is an element of $\Omega'$ and $\omega$ is the image of $\omega'$ in $\Omega$ by the projection map
\[\Omega'=\Omega\times_{M_{K},\phi}M_{K'}\longrightarrow\Omega,\]
then one has
\[X'_{\omega'}:=X'\times_{\operatorname{Spec} K'}\operatorname{Spec}{K'_{\omega'}}\cong (X\times_KK_\omega)\times_{K_\omega}K'_{\omega'}.\]
Moreover, the pull-back of $L_\omega$ on $X'_{\omega'}$ identifies with $L'_{\omega'}$. We denote by $p_{\omega'}$ the morphism of projection from $X'_{\omega'}$ to $X_\omega$. Then the map \[p_{\omega'}^{\natural}:(X'_{\omega'})^{\operatorname{an}}\longrightarrow X_\omega^{\mathrm{an}},\]
sending any point $x'=(j(x'),|\raisebox{.4ex}{.}|_{x'})$ to the pair consisting of the scheme point $p_{\omega'}(j(x'))$ of $X_\omega$ and the restriction of $|\raisebox{.4ex}{.}|_{x'}$ on the residue field of $p_{\omega'}(j(x'))$, is continuous (see \cite[Proposition 2.1.17]{CMArakelovAdelic}), where $j:(X_{\omega'}')^{\mathrm{an}}\rightarrow X'_{\omega'}$ denotes the map sending a point in the analytic space to its underlying scheme point. Therefore, the continuous metric $\varphi_\omega$ induces by composition with $p^{\natural}$ a continuous metric $\varphi_{\omega'}$ such that, for any $x'\in (X'_{\omega'})^{\operatorname{an}}$ and any $\ell\in L_\omega(p^{\natural}(x'))$, one has
\[\forall\,a\in\widehat{\kappa}(x'),\quad |a\otimes\ell|_{\varphi_{\omega'}}(x')=|a|_{x'}\cdot|\ell|_{\varphi_\omega}(x). \]
Therefore, $(\varphi_{\omega'})_{\omega'\in\Omega'}$ forms a metric family of $L'$ which we denote by $\varphi_{K'}$.
\end{exem}
\begin{defi}\label{Def: adelic line bundle}
Let $L$ be an invertible $\mathcal O_X$-module and $\varphi=(\varphi_\omega)_{\omega\in\Omega}$ be a metric family of $L$.
\begin{enumerate}[label=\rm(\arabic*)]
\item We say that $\varphi$ is \emph{dominated} if there exist invertible $\mathcal O_X$-modules $L_1$ and $L_2$, respectively equipped with metric families $\varphi_1$ and $\varphi_2$, which are quotient metric families associated with dominated norms families, such that $L\cong L_1\otimes L_2^\vee$ and that the local distance function \[(\omega\in\Omega)\longmapsto d_\omega(\varphi,\varphi_1\otimes\varphi_2^\vee)\]
is bounded from above by a $\nu$-integrable function (see \cite[\S6.1.1]{CMArakelovAdelic});
\item We say that $\varphi$ is \emph{measurable} if the following conditions are satisfied (see \cite[\S6.1.4]{CMArakelovAdelic}):
\begin{enumerate}[label=\rm(2.\roman*)]
\item for any closed point $P$ of $X$, the norm family $P^*(\varphi)$ of $P^*(L)$ is measurable,
\item for any $\xi\in X^{\mathrm{an}}$ (where we consider the trivial absolute value on $K$ in the construction of $X^{\mathrm{an}}$) whose associated scheme is of dimension $1$ and such that the exponent\footnote{Since the schematic point associated with $\xi$ is of dimension $1$, the absolute value $|\raisebox{.4ex}{.}|_\xi$ is discrete and hence is of the form $|\raisebox{.4ex}{.}|_\xi=\exp(- t\operatorname{ord}_\xi(\raisebox{.4ex}{.}))$, where the (surjective) map $\operatorname{ord}_\xi(\raisebox{.4ex}{.}):\widehat{\kappa}(\xi)\rightarrow\mathbb Z\cup\{+\infty\}$ is the discrete valuation corresponding to the absolute value $|\raisebox{.4ex}{.}|_\xi$. The non-negative real number $t$ is called the \emph{exponent} of the absolute value $|\raisebox{.4ex}{.}|_\xi$.} of the absolute value $|\raisebox{.4ex}{.}|_\xi$ is rational, and for any $\ell\in L\otimes_{\mathcal O_X}\widehat{\kappa}(\xi)$, the function
\[(\omega\in\Omega_0)\longmapsto |\ell|_{\varphi_\omega}(\xi)\]
is measurable, where $\Omega_0$ is the subset of $\omega\in\Omega$ such that $|\raisebox{.4ex}{.}|_\omega$ is trivial, and we consider the restriction of the $\sigma$-algebra $\mathcal A$ to $\Omega_0$.
\end{enumerate}
\end{enumerate}
If $\varphi$ is both dominated and measurable, we say that the pair $\overline L=(L,\varphi)$ is an \emph{adelic line bundle}.
\end{defi}
\begin{prop}\label{Pro: measurable direct image}
Let $\pi:X\rightarrow\operatorname{Spec} K$ be a projective scheme over $\operatorname{Spec} K$, $L$ be an invertible $\mathcal O_X$-module, $\varphi$ be a metric family of $L$, and $E=H^0(X,L)$. We equip $E$ with a norm family $\xi=(\|\raisebox{.4ex}{.}\|_{\omega})_{\omega\in\Omega}$. Consider the following norm family $\xi'=(\|\raisebox{.4ex}{.}\|_{\omega}')_{\omega\in\Omega}$ defined as
\[\forall\,s\in H^0(X_\omega,L_\omega),\quad\|s\|_{\omega}':=\max\bigg\{\sup_{x\in X_\omega^{\operatorname{an}}}|s|_{\varphi_\omega}(x),\|s\|_\omega\bigg\}.\]
Then one has the following:
\begin{enumerate}[label=\rm(\arabic*)]
\item\label{Item: xi prime measurable} If $\varphi$ and $\xi$ are both measurable, then $\xi'$ is also measurable.
\item\label{Item: xi prime dominated} If $\varphi$ is dominated and $\xi$ is strongly dominated, then $\xi'$ is strongly dominated.
\end{enumerate}
\end{prop}
\begin{proof}
\ref{Item: xi prime measurable} For any $\omega\in\Omega$, we let $\|\raisebox{.4ex}{.}\|_{\varphi_\omega}$ be the seminorm on $E\otimes_KK_\omega=H^0(X_\omega,L_\omega)$ defined as
\[\forall\,s\in H^0(X_\omega,L_\omega),\quad\|s\|_{\varphi_\omega}:=\sup_{x\in X_\omega^{\operatorname{an}}}|s|_{\varphi_\omega}(x).\]
By \cite[Propositions 6.1.20 and 6.1.26]{CMArakelovAdelic}, for any $s\in H^0(X,L)$, the function
\[(\omega\in\Omega)\longmapsto\|s\|_{\varphi_\omega}\]
is measurable. Therefore the function
\[(\omega\in\Omega)\longmapsto\|s\|_{\omega}'=\max\{\|s\|_{\varphi_\omega},\|s\|_\omega\}\]
is also measurable once the norm family $\xi$ is measurable.
\ref{Item: xi prime dominated} We may assume without loss of generality that there exists a basis $\boldsymbol{e}=(e_i)_{i=1}^r$ of $E$ such that, for any $\omega\in\Omega$
\[\forall\,(\lambda_1,\ldots,\lambda_r)\in K_\omega^r,\quad \|\lambda_1e_1+\cdots+\lambda_re_r\|_{\omega}=\max_{i\in\{1,\ldots,r\}}|\lambda_i|_\omega. \]
By \cite[Remark 6.1.17]{CMArakelovAdelic}, for any $s\in H^0(X,L)$, the function
\[(\omega\in\Omega)\longmapsto \ln\|s\|_{\varphi_\omega}\]
is bounded from above by an integrable function. Let $A:\Omega\rightarrow\mathbb R_{\geqslant 0}$ be a positive integrable function on $\Omega$ such that
\[\forall\,\omega\in\Omega,\quad \max_{i\in\{1,\ldots,r\}}\ln\|e_i\|_{\varphi_\omega}\leqslant A(\omega).\]
For any $\omega\in\Omega\setminus\Omega_\infty$ and any $(\lambda_1,\ldots,\lambda_r)\in K_\omega^r$, one has
\[\ln\|\lambda_1e_1+\cdots+\lambda_re_r\|_\omega\leqslant \ln\|\lambda_1e_1+\cdots+\lambda_re_r\|_\omega'\leqslant\max_{i\in\{1,\ldots,r\}}(\ln|\lambda_i|_\omega+\ln\|e_i\|_\omega').
\]
Note that $\|e_i\|=1$ and hence
\[\ln\|e_i\|_\omega'=\max\{\ln\|e_i\|_{\varphi_\omega},\ln(1)\}\leqslant A(\omega).\]
Therefore one has
\[d(\|\raisebox{.4ex}{.}\|_\omega,\|\raisebox{.4ex}{.}\|_{\omega}')\leqslant A(\omega).\]
In the case where $\omega\in\Omega_\infty$, for any $(\lambda_1,\ldots,\lambda_r)\in K_\omega^r$ one has
\[\ln\|\lambda_1e_1+\cdots+\lambda_re_r\|_\omega\leqslant \ln\|\lambda_1e_1+\cdots+\lambda_re_r\|_\omega'\leqslant\max_{i\in\{1,\ldots,r\}}\ln|\lambda_i|_\omega+A(\omega)+\ln(r).\]
Finally we obtain that
\[\forall\,\omega\in\Omega,\quad d_{\omega}(\xi,\xi')\leqslant A(\omega)+\ln(r)1\hspace{-0.25em}\mathrm{l}_{\Omega_\infty}(\omega).\]
Hence the norm family $\xi'$ is strongly dominated (see \cite[Proposition 3.1.2]{CMArakelovAdelic} for the fact that $\nu(\Omega_\infty)$ is finite).
\end{proof}
\begin{lemm}\label{lem:measurable:integrable:adelic:algebraic:cover}
Let $S = (K, (\Omega, \mathcal A, \nu), \phi)$ be an adelic curve, $K'$ be an algebraic extension of $K$ and
$S_{K'} = S \otimes_K K' = (K', (\Omega', \mathcal A', \nu'), \phi')$. Let $f$ be a function on $\Omega$.
Then one has the following:
\begin{enumerate}[label=\rm(\arabic*)]
\item $f$ is measurable if and only if $f \circ \pi_{K'/K}$ is measurable.
\item $f$ is integrable if and only if $f \circ \pi_{K'/K}$ is integrable.
\end{enumerate}
\end{lemm}
\begin{proof}
Clearly we may assume that $f$ is non-negative, so that
it is a consequence of \cite[Proposition~3.4.8 and Proposition~3.4.9]{CMArakelovAdelic}.
\end{proof}
\section{Integrability of local intersection numbers}
\label{Sec: Integrability of local intersection numbers}
In this section, we fix a projective $K$-scheme $X$. Let $d$ be the dimension of $X$.
\begin{defi}\label{def:adelic:Cartier:divisor}
Let $D$ be a Cartier divisor on $X$. For any $\omega\in\Omega$, let $D_\omega$ be the pull-back of $D$ by the morphism of projection $X_\omega\rightarrow X$, which is well defined since the morphism of projection $X_\omega\rightarrow X$ is flat (see Remark \ref{Rem: pull back} and Definition \ref{Def: pull back}). We call \emph{Green function family} of $D$ any family $(g_\omega)_{\omega\in\Omega}$ parametrized by $\Omega$, where each $g_\omega$ is a Green function of $D_\omega$. We denote by $\varphi_g$ the metric family $(|\raisebox{.4ex}{.}|_{g_\omega})_{\omega\in\Omega}$ of $\mathcal O_X(D)$, where $|\raisebox{.4ex}{.}|_{g_\omega}$ is the continuous metric on $\mathcal O_{X_\omega}(D_\omega)$ induced by the Green function $g_\omega$ (see Remark \ref{Rem: metric induced by a green function}). If the metric family $\varphi_g$ is measurable, we say that the Green function family $g$ is \emph{measurable}. If the metric $\varphi_g$ is dominated, we say that the Green function family $g$ is \emph{dominated}. We refer to Definition \ref{Def: adelic line bundle} for the dominancy and measurability of metrics. If $g$ is both dominated and measurable, we say that $(D,g)$ is an \emph{adelic Cartier divisor}.
Let $D$ be an invertible $\mathcal O_X$-module and $g$ be a Green function family of $D$. If $D$ is ample and all metrics in the family $\varphi_g$ are semi-positive, we say that the Green function family $g$ is \emph{semi-positive}. We say that $(D, g)$ is \emph{integrable} if there exist ample Cartier divisors $D_1$ and $D_2$, together with semi-positive Green function families $g_1$ and $g_2$ of $D_1$ and $D_2$ respectively, such that $D=D_1-D_2$ and $g=g_1-g_2$. Similarly, we say that an adelic line bundle $(L,\varphi)$ is \emph{integrable} if there exists ample invertible $\mathcal O_X$-modules $L_1$ and $L_2$, and metric families consisting of semi-positive metrics $\varphi_1$ and $\varphi_2$ on $L_1$ and $L_2$, respectively, such that $L=L_2\otimes L_1^\vee$ and $\varphi=\varphi_2\otimes\varphi_1^\vee$.
Let $D_0,\ldots,D_d$ be a family of Cartier divisors, which intersect properly. For any $i\in\{0,\ldots,d\}$, let $g_i$ be a Green function family of $D_i$ such that $(D_i,g_i)$ is integrable. Then, for any $\omega\in\Omega$, a local intersection number
\[((D_{0,\omega},g_{0,\omega}),\ldots,(D_{d,\omega},g_{d,\omega}))_{(K_\omega,|\cdot|_\omega)}\]
has been introduced in Definition \ref{def:local:intersection}, which we denote by
\[(\overline D_0\cdots\overline D_d)_\omega\]
for simplicity. Thus the local intersection numbers define a function
\[(\omega\in\Omega)\longrightarrow (\overline D_0\cdots\overline D_d)_\omega.\]
\end{defi}
\begin{defi}
Let $D_1$ and $D_2$ be Cartier divisors on $X$, and $g_1$ and $g_2$ be Green function families of $D_1$ and $D_2$, respectively. We say that $(D_1,g_1)$ and $(D_2,g_2)$ are \emph{linearly equivalent} and we note
\[(D_1,g_1)\sim(D_2,g_2)\] if $\mathcal O_X(D_1)$ is isomorphic to $\mathcal O_X(D_2)$ and if there exists an isomorphism of $\mathcal O_X$-modules $\mathcal O_X(D_1)\rightarrow\mathcal O_X(D)$ which identifies the metric $\varphi_{g_1}$ to $\varphi_{g_2}$.
\end{defi}
\begin{prop}\label{prop:reduction:same:class} Assume that, for all Cartier divisors $E_0,\ldots,E_d$ which intersect properly, and measurable (resp. dominated) Green function families $h_0,\ldots,h_d$ of $E_0,\ldots,E_d$ respectively, such that all $(E_i,h_i)$ are integrable and linearly equivalent, the function of local intersection number
\[(\omega\in\Omega)\longmapsto(\overline E_0\cdots\overline E_d)_\omega\]
is measurable (resp. dominated). Then, for all Cartier divisors $D_0,\ldots,D_d$ which intersect properly and measurable (resp. dominated) Green function families $g_0,\ldots,g_d$ of $D_0,\ldots,D_d$ respectively, such that all $(D_i,g_i)$ are integrable (but not necessarily linearly equivalent), the function of local intersection number
\[(\omega\in\Omega)\longmapsto (\overline D_0\cdots\overline D_d)_\omega\]
is measurable (resp. dominated).
\end{prop}
\begin{proof}
First of all, by Lemma~\ref{lem:measurable:integrable:adelic:algebraic:cover}, we may assume that
$K$ is algebraically closed.
By Lemma~\ref{lem:IP:projective:ample},
we can choose a matrix \[( D_{i,j} )_{(i,j)\in\{0,\ldots,d\}^2}\]
consisting of Cartier divisors on $X$ such that $(D_{i_0,0}, \ldots, D_{i_d,d}) \in {\mathcal{IP}}_X$ for any
$(i_0, \ldots, i_d) \in \{ 0, \ldots, d \}^{d+1}$, and that $D_{i,j} \sim D_i$ for all $(i, j) \in \{0, \ldots, d \}^2$.
Let $g_{i,j}$ be a family of integrable Green functions of $D_{i,j}$ such that $(D_{i,j}, g_{i,j}) \sim (D_i, g_i)$.
By Proposition~\ref{pro:domain},
\begin{multline*}
\sum_{\sigma \in \mathfrak{S}(\{0, \ldots, d\})} \big(\overline{D}_{0,\sigma(0)} \cdots \overline{D}_{d,\sigma(d)}\big)_\omega =
\sum_{\sigma \in \mathfrak{S}(\{0, \ldots, d\})} \big(\overline{D}_{\sigma(0),0} \cdots \overline{D}_{\sigma(d),d}\big)_\omega\\
= \sum_{\emptyset \not= I \subseteq \{ 0, \ldots, d \}} (-1)^{(d+1) - \operatorname{card}(I)} \Bigg(\Big(\sum\nolimits_{i \in I} \overline{D}_{i,0} \Big) \cdots \Big(\sum\nolimits_{i \in I} \overline{D}_{i,d}\Big) \Bigg)_\omega,
\end{multline*}
where $\overline{D}_{i,j} = (D_{i,j}, g_{i,j})$.
Note that $\sum_{i \in I} \overline{D}_{i,a} \sim \sum_{i \in I} \overline{D}_{i,b}$, so that
\[
(\omega \in \Omega) \mapsto \sum_{\emptyset \not= I \subseteq \{ 0, \ldots, d \}} (-1)^{(d+1) - \operatorname{card}(I)} \Bigg(\Big(\sum\nolimits_{i \in I} \overline{D}_{i,0} \Big) \cdots \Big(\sum\nolimits_{i \in I} \overline{D}_{i,d}\Big) \Bigg)_\omega
\]
is measurable (resp. dominant) by our assumption. Moreover, by Proposition~\ref{prop:intersection:principal:div},
for each $\sigma \in \mathfrak{S}(\{0, \ldots, d\})$, there is an integrable function $A_{\sigma}$ on $\Omega$ such that
\[
(\overline{D}_{0,\sigma(0)} \cdots \overline{D}_{d,\sigma(d)})_\omega =
(\overline{D}_0 \cdots \overline{D}_d)_\omega + A_{\sigma}(\omega).
\]
Thus the assertion follows. Note that $\int_{\Omega} A_{\sigma}(\omega)\, \nu(d\omega) = 0$
if $S$ is proper.
\end{proof}
\begin{theo}\label{Thm: measurability non archimedean}Assume that $\Omega_\infty=\emptyset$.
Let $(L_i)_{i=0}^d$ be a family of invertible $\mathcal O_X$-modules. For each $i\in\{0,\ldots,d\}$, let $s_i$ be a regular meromorphic section of $L_i$ and $D_i=\operatorname{div}(s_i)$. We suppose that $D_0,\ldots,D_d$ intersect properly. For any $i\in\{0,\ldots,d\}$, let $\varphi_i=(\varphi_{i,\omega})_{\omega\in\Omega}$ be a measurable metric family on $L_i$ such that $(L_{i,\omega},\varphi_{i,\omega})$ is integrable, and let $g_i=(g_{i,\omega})_{\omega\in\Omega}$ be the family of Green functions of $D_i$ corresponding to $\varphi_i$.
Then the
function of local intersection numbers
\begin{equation}\label{Equ: local intersection measurable}(\omega\in\Omega)\longrightarrow (\overline D_0\cdots\overline D_d)_\omega\end{equation}
is $\mathcal A$-measurable.
\end{theo}
\begin{proof}
By Lemma~\ref{lem:measurable:integrable:adelic:algebraic:cover}, we may assume that
$K$ is algebraically closed.
By using Proposition~\ref{prop:multilinear:symmetric:semiample:case}, we may further assume that $L_0,\ldots,L_d$ are very ample.
For any $i\in\{0,\ldots,d\}$, we denote by $\delta_i$ the intersection number
\[\deg(c_1(L_0)\cdots c_1(L_{i-1})c_1(L_{i+1})\cdots c_1(L_d)\cap[X]).\]
We introduce, for each $r\in\{-1,\ldots,d\}$, then following condition $(C_r)$:
\begin{quote}
\it\hskip\parindent For each $i\in\{0,\ldots,d\}$ such that $0\leqslant i\leqslant r$, there exist a positive integer $m_i$ and a measurable Hermitian norm family $\xi_i=(\|\raisebox{.4ex}{.}\|_{i,\omega})_{\omega\in\Omega}$ on $H^0(X,L_i^{\otimes m_i})$, such that $\varphi_i$ identifies with the quotient metric family induced by $\xi_i$.
\end{quote}
We will prove by inverted induction on $r$ that, under the condition $(C_r)$, the function \eqref{Equ: local intersection measurable} is $\mathcal A$-measurable. Note that the condition $(C_{-1})$ is always true and hence the measurability of \eqref{Equ: local intersection measurable} under $(C_{-1})$ is just the statement of the theorem. We begin with the case where $r=d$. For any $i\in\{1,\ldots,d\}$, let $E_i=H^0(X,L_i^{\otimes m_i})$ and $f_i:X\rightarrow\mathbb P(E_i)$ be the canonical closed embedding. Note that $L_i^{\otimes m_i}$ is isomorphic to $f_i^*(\mathcal O_{E_i}(1))$. We denote by $R$ the resultant
\[R_{f_0,\ldots,f_d}^{X,s_0^{\otimes m_0},\ldots,s_d^{\otimes m_d}},\]
which is an element of
\[S^{\delta_0 N_0}(E_0^\vee)\otimes_K\cdots\otimes_KS^{\delta_dN_d}(E_d^\vee),\]
where
\[N_i=\frac{m_0\cdots m_d}{m_i}.\]
We equip this vector space with the family of $\varepsilon$-tensor product of $\varepsilon$-symmetric power norms of $\|\raisebox{.4ex}{.}\|_{i,\omega,*}$ (see Definition \ref{Def: epsilon tensor product}), which we denote by $\xi=(\|\raisebox{.4ex}{.}\|_\omega)_{\omega\in\Omega}$. By \cite[Proposition 4.1.24]{CMArakelovAdelic}, the norm family $\xi$ is measurable. By Theorem \ref{Thm: equality of local height}, one has
\[m_0\cdots m_d(\overline D_0\cdots\overline D_d)_{\omega}=(m_0\overline D_0\cdots m_d\overline D_d)_\omega=\ln\|R\|_\omega.\]
Hence the function
\[(\omega\in\Omega)\longmapsto (\overline D_0\cdots\overline D_d)_\omega\] is measurable.
We prove the measurability of \eqref{Equ: local intersection measurable} under $(C_{r-1})$ in assuming that the measurability of \eqref{Equ: local intersection measurable} is true under $(C_{r})$, where $r\in\{0,\ldots,d\}$. For any positive integer $m$, we let $g_r^{(m)}$ be the Green function family of $D_r$ corresponding to the metric family $\varphi_r^{(m)}=(\varphi_{r,\omega}^{(m)})_{\omega\in\Omega}$ (see Definition \ref{Def: Fubini-Study}). We first show that the function
\[(\omega\in\Omega)\longmapsto (\overline D_0\cdots\overline D_{r-1}(D_r,g_r^{(m)})\overline D_{r+1}\cdots\overline D_d)_\omega \]
is measurable. For this purpose, we choose arbitrarily a measurable norm family $\xi_r=(\|\raisebox{.4ex}{.}\|_{\omega})_{\omega\in\Omega}$ on the vector space $H^0(X,L^{\otimes m})$ (one can choose $\xi_r=\xi_{\boldsymbol{e}}$, where $\boldsymbol{e}$ is a basis of $H^0(X,L^{\otimes m})$, see Example \ref{Exe: norm family of a basis}). For any $a>0$ and any $\omega\in\Omega$, we let $\varphi_{r,a,\omega}^{(m)}$ be the quotient metric on $L_r$ induced by the norm
\[\|\raisebox{.4ex}{.}\|_{a,\omega}:=\max\{\|\raisebox{.4ex}{.}\|_{\varphi_r^{\otimes m}},a\|\raisebox{.4ex}{.}\|_{\omega}\}\]
on $H^0(X_\omega,L_\omega^{\otimes m})$, and let $g_{r,a}^{(m)}$ be the Green function of $D_r$ corresponding to the metric $\varphi_{r,a,\omega}^{(m)}$.
By Proposition \ref{Pro: measurable direct image}, the norm family $\xi_{r,a}:=(\|\raisebox{.4ex}{.}\|_{a,\omega})_{\omega\in\Omega}$ is measurable. Therefore $\overline D_0,\ldots,\overline D_{r-1},(D_r,g_{r,a}^{(m)}),\overline D_{r+1}\cdots\overline D_d$ satisfy the condition $(C_r)$. By the induction hypothesis, we obtain that
the function
\[(\omega\in\Omega)\longmapsto (\overline D_0\cdots\overline D_{r-1}(D_r,g_{r,a}^{(m)})\overline D_{r+1}\cdots\overline D_d)_\omega \]
is measurable. Moreover, by Proposition \ref{Pro: varphi 1 as a continuous metric}, we obtain that, for any $\omega\in\Omega$, there exists $a_{\omega}>0$ such that $g_{r,a}^{(m)}=g_r^{(m)}$ when $0<a<a_\omega$. Therefore one has
\[(\overline D_0\cdots\overline D_{r-1}(D_r,g_r^{(m)})\overline D_{r+1}\cdots\overline D_d)_\omega=\lim_{a\in\mathbb Q,\,a\rightarrow 0+}(\overline D_0\cdots\overline D_{r-1}(D_r,g_{r,a}^{(m)})\overline D_{r+1}\cdots\overline D_d)_\omega\]
and hence the function
\[(\omega\in\Omega)\longmapsto (\overline D_0\cdots\overline D_{r-1}(D_r,g_r^{(m)})\overline D_{r+1}\cdots\overline D_d)_\omega \]
is measurable. Finally, by Proposition \ref{Pro: convergence of varphin} and Corollary \ref{cor:intersection:trivial:divisor}, one has
\[(\overline D_0\cdots\overline D_d)_\omega=\lim_{m\rightarrow+\infty}(\overline D_0\cdots\overline D_{r-1}(D_r,g_r^{(m)})\overline D_{r+1}\cdots\overline D_d)_\omega\]
and therefore the function
\[(\omega\in\Omega)\longmapsto (\overline D_0\cdots\overline D_d)_\omega\]
is measurable.
\end{proof}
In the following, we study the measurability of the function of local intersection number over Archimedean places. We assume that $\Omega_{\infty}=\Omega$.
If $K$ contains a square root $\sqrt{-1}$ of $-1$, then, by Lemma~\ref{lemma:measurable:family:embeddings},
for each $\omega \in \Omega$, there is an embedding
$\sigma_{\omega} : K \hookrightarrow {\mathbb{C}}$
with the following properties:
\begin{enumerate}[label=\rm(\arabic*)]
\item $|\raisebox{.4ex}{.}|_{\omega} = |\sigma_{\omega}(\raisebox{.4ex}{.})|$ for all $\omega \in \Omega$.
\item $\sigma_{\omega}(\sqrt{-1}) = i$, so that $\sigma_{\omega}(a + \sqrt{-1}b) = a + ib$ for all $a, b \in {\mathbb{Q}}$, where $i$ is the usual imaginary unit in $\mathbb C$.
\item For $a \in K$, $(\omega \in \Omega) \mapsto \sigma_{\omega}(a)$ is measurable.
\end{enumerate}
\begin{prop}\label{prop:measurable:partial:X}
We assume that $\Omega = \Omega_{\infty}$ and $\sqrt{-1} \in K$.
Let $n$ and $d$ be non-negative integers with $n \geq d$ and
$\pi: {\mathbb{A}}^n_K \to {\mathbb{A}}^d_K$ be the projection given by $(x_1, \ldots, x_n) \mapsto (x_1, \ldots, x_d)$.
Let $U$ be a non-empty Zariski open set of ${\mathbb{A}}^d_K$ and $X$ be a reduced closed subscheme of $\pi^{-1}(U)$ such that $\rest{\pi}{X} : X \to U$ is finite, surjective and \'etale.
We assume that either (i) $n=d$
and $X = \pi^{-1}(U)$, or (ii) $K$ is algebraically closed field.
Let $f= \{ f_{\omega} \}_{\omega \in \Omega}$ be a family of functions indexed by $\Omega$ such that
$f_{\omega}$ is a $C^{\infty}$-function on $\pi_{\omega}^{-1}(U_{\omega})$ and that, for any $K$-rational point $P \in \pi^{-1}(U)(K)$,
the function given by $(\omega \in \Omega) \mapsto f_{\omega}(P_{\omega})$ is measurable.
If we set $g_{\omega} = \rest{f_{\omega}}{X_{\omega}}$ for $\omega \in \Omega$, then, for any $P \in X(K)$ and $l \in \{ 1, \ldots, d\}$,
\[
(\omega \in \Omega) \mapsto \frac{\partial g_{\omega}}{\partial z_{l\omega}}(P_{\omega})\quad\text{and}\quad
(\omega \in \Omega) \mapsto \frac{\partial g_{\omega}}{\partial \overline{z}_{l\omega}}(P_{\omega})
\]
are measurable, where $(z_{1\omega}, \ldots, z_{d\omega})$ be the coordinate of ${\mathbb{A}}^d \times_{\sigma_{\omega}} {\mathbb{C}}$.
\end{prop}
\begin{proof}
{\bf Case (i)}: $n=d$ (so that $\pi = \operatorname{id}$)
and $X = \pi^{-1}(U)$.
Let $x_{l\omega}$ (resp. $y_{l\omega}$) be the real part (resp. the imaginary part) of $z_{l\omega}$.
It is sufficient to show that
\[
(\omega \in \Omega) \mapsto \frac{\partial f_{\omega}}{\partial x_{l\omega}}(P_{\omega})\quad\text{and}\quad
(\omega \in \Omega) \mapsto \frac{\partial f_{\omega}}{\partial y_{l\omega}}(P_{\omega})
\]
are measurable.
We set $P_{\omega} = \sigma_{\omega}(P) = (a_{1\omega} + i b_{1\omega}, \ldots, a_{n\omega} + ib_{n\omega})$.
Then, for $\varepsilon \in {\mathbb{Q}}^{\times}$,
\[
\begin{cases}
(P + \varepsilon e_l)_{\omega} = \sigma_{\omega}(P + \varepsilon e_l) = (a_{1\omega} + i b_{1\omega}, \ldots, (a_{l\omega} + \varepsilon) + ib_{l\omega}, \ldots, a_{n\omega} + ib_{n\omega}),\\
(P + \varepsilon ie_l)_{\omega} = \sigma_{\omega}(P + \varepsilon i e_l) = (a_{1\omega} + i b_{1\omega}, \ldots, a_{l\omega} + i(b_{l\omega}+\varepsilon), \ldots, a_{n\omega} + ib_{n\omega}),
\end{cases}
\]
where $\{ e_1, \ldots, e_n \}$ is the standard basis of $K^n$, so that
\[
\begin{cases}
{\displaystyle \lim_{\substack{\varepsilon \in {\mathbb{Q}}^{\times}\\ \varepsilon \to 0}} \frac{f_{\omega}((P + \varepsilon e_l)_{\omega}) - f_{\omega}(P_{\omega})}{\varepsilon} = \frac{\partial f_{\omega}}{\partial x_{l\omega}}(P_{\omega}),}\\
{\displaystyle \lim_{\substack{\varepsilon \in {\mathbb{Q}}^{\times}\\ \varepsilon \to 0}} \frac{f_{\omega}((P + \varepsilon ie_l)_{\omega}) - f_{\omega}(P_{\omega})}{\varepsilon} = \frac{\partial f_{\omega}}{\partial y_{l\omega}}(P_{\omega}).}
\end{cases}
\]
Note that
\[
\begin{cases}
{\displaystyle (\omega \in \Omega) \mapsto \frac{f_{\omega}((P + \varepsilon e_l)_{\omega}) - f_{\omega}(P_{\omega})}{\varepsilon},}\\[1ex]
{\displaystyle (\omega \in \Omega) \mapsto \frac{f_{\omega}((P + \varepsilon ie_l)_{\omega}) - f_{\omega}(P_{\omega})}{\varepsilon}.}
\end{cases}
\]
are measurable. Thus the assertion follows.
\bigskip
{\bf Case (ii)}: $K$ is algebraically closed field.
By replacing $U$ and $X$ by $U \setminus \pi(P)$ and $X \setminus P$, we may assume that $P=(0, \ldots, 0)$.
If we set $Q = \pi(P)$, then $(\rest{\pi}{X})^* : {\mathscr{O}}_{U, Q}^h \overset{\sim}{\longrightarrow} {\mathscr{O}}_{X, P}^h$, where
${\mathscr{O}}_{U, Q}^h$ and ${\mathscr{O}}_{X, P}^h$ are the Henselizations of ${\mathscr{O}}_{U, Q}$ and ${\mathscr{O}}_{X, P}$, respectively.
Thus there are $\varphi_{d+1}, \ldots, \varphi_{n} \in {\mathscr{O}}_{U, Q}^h$ such that $(\rest{\pi}{X})^*(\varphi_{j}) = \rest{x_j}{X}$ for
$j \in \{ d+1, \ldots, n \}$.
We set
\[
\varphi_j = \sum_{e_1\cdots e_d \in {\mathbb{Z}}_{\geq 0} } a_{j, e_1\cdots e_d} X_1^{e_1} \cdots X_d^{e_d}
\]
as an element of $K[\![ X_1, \ldots, X_{d} ]\!]$. Note that if we set
\[
\varphi_{j\omega} = \sum_{e_1\cdots e_d \in {\mathbb{Z}}_{\geq 0} } \sigma_{\omega}(a_{j, e_1\cdots e_d}) X_1^{e_1} \cdots X_d^{e_d},
\]
then \[g_{\omega} = f_{\omega}(z_{1\omega},\ldots,z_{d\omega},\varphi_{d+1 \omega}(z_{1\omega},\ldots,z_{d\omega}), \ldots, \varphi_{n \omega}(z_{1\omega},\ldots,z_{d\omega}))
\]
as a function on $U$ around $Q$. Then, for $l \in \{ 1, \ldots, d \}$,
\[
\begin{cases}
{\displaystyle \frac{\partial g_{\omega}}{\partial z_{l\omega}}(P_{\omega}) = \frac{\partial f_{\omega}}{\partial z_{l\omega}}(0, \ldots, 0) + \sum_{j=d+1}^n \frac{\partial f_{\omega}}{\partial z_{j \omega}}(0, \ldots, 0)\frac{\partial \varphi_{j\omega}}{\partial z_{l\omega}}(0, \ldots, 0)}, \\[2ex]
{\displaystyle \frac{\partial g_{\omega}}{\partial \overline{z}_{l\omega}}(P_{\omega}) = \frac{\partial f_{\omega}}{\partial \overline{z}_{l\omega}}(0, \ldots, 0)}.
\end{cases}
\]
If we denote $a_{j, e_1, \ldots, e_d}$ by $a_{j, l}$ in the case where $e_1 = 0, \ldots, e_l = 1, \ldots, e_d = 0$, then
\[\begin{cases}
{\displaystyle \frac{\partial g_{\omega}}{\partial z_{l\omega}}(P_{\omega}) = \frac{\partial f_{\omega}}{\partial z_{l\omega}}(0, \ldots, 0) + \sum_{j=d+1}^n \frac{\partial f_{\omega}}{\partial z_{j \omega}}(0, \ldots, 0)\sigma_{\omega}(a_{j, l})}, \\[2ex]
{\displaystyle \frac{\partial g_{\omega}}{\partial \overline{z}_{l\omega}}(P_{\omega}) = \frac{\partial f_{\omega}}{\partial \overline{z}_{l\omega}}(0, \ldots, 0)},
\end{cases}
\]
so that the assertions follow from the case (i).
\end{proof}
\begin{prop}\label{prop:measurable:vol:int:A}
We assume that $\Omega = \Omega_{\infty}$ and $\sqrt{-1} \in K$.
Let $U$ be a non-empty Zariski open set of ${\mathbb{A}}^n_K$.
Let $h= \{ h_{\omega} \}_{\omega \in \Omega}$ be a family of functions indexed by $\Omega$ such that
$h_{\omega}$ is a $C^{\infty}$-function on $U_{\omega}$ and that, for any $K$-rational point $P \in U(K)$,
the function given by $(\omega \in \Omega) \mapsto h_{\omega}(P_{\omega})$ is measurable.
For each $\omega \in \Omega$, let $(z_{1\omega}, \ldots, z_{n\omega})$ is the coordinate of ${\mathbb{A}}^n \otimes_{\sigma_{\omega}} {\mathbb{C}}$. If
\[\int_{U_{\omega}} \Big(\frac i2\Big)^nh_{\omega} (z_{1\omega},\ldots,z_{n\omega})\, \mathrm{d}z_{1\omega} \wedge \mathrm{d}\bar{z}_{1\omega} \wedge \cdots \wedge \mathrm{d}z_{n\omega} \wedge \mathrm{d}\bar{z}_{n\omega} \] exists for any $\omega \in \Omega$, then
\[
(\omega \in \Omega) \mapsto \int_{U_{\omega}} \Big(\frac i2\Big)^nh_{\omega} (z_{1\omega},\ldots,z_{n\omega}) \,\mathrm{d}z_{1\omega} \wedge \mathrm{d}\bar{z}_{1\omega} \wedge \cdots \wedge \mathrm{d}z_{n\omega} \wedge \mathrm{d}\bar{z}_{n\omega}
\]
is measurable.
\end{prop}
\begin{proof}
Shrinking $U$ if necessarily, we may assume that ${\mathbb{A}}^n_K \setminus U$ is defined by $\{ F = 0 \}$ for some $F \in K[X_1, \ldots, X_n] \setminus \{ 0 \}$.
We set
\[
U_{\omega, N} = \Big\{ (z_{1\omega},\ldots,z_{n\omega}) \in {\mathbb{C}}^n \,\Big|\,
\text{$\max_{j\in\{1, \ldots, n\}} |z_{j\omega}| \leq N$ and $|F(z_{1\omega}, \ldots, z_{n\omega})| \geqslant 1/N$} \Big\}.
\]
Let $x_{i\omega}$ (resp. $y_{i\omega}$)
be the real part (resp. imaginary part) of $z_{i\omega}$.
Then
\begin{multline*}
\Big(\frac i2\Big)^n h_{\omega} \,\mathrm{d}z_{1\omega} \wedge \mathrm{d}\bar{z}_{1\omega} \wedge \cdots \wedge \mathrm{d}z_{n\omega} \wedge \mathrm{d}\bar{z}_{n\omega} \\
=
h_{\omega}\, \mathrm{d}x_{1\omega} \wedge \mathrm{d}y_{1\omega} \wedge \cdots
\wedge \mathrm{d}x_{n\omega} \wedge \mathrm{d}y_{n\omega}.
\end{multline*}
Moreover,
\begin{multline}\label{eqn:prop:measurable:projective:space:01}
\int_{U_{\omega, N}} h_{\omega}\, \mathrm{d}x_{1\omega} \wedge \mathrm{d}y_{1\omega} \wedge \cdots
\wedge \mathrm{d}x_{n\omega} \wedge \mathrm{d}y_{n\omega} \\
\kern-12em = \lim_{m\to\infty} \sum_{\substack{a_1, b_1, \ldots, a_n, b_n \in {\mathbb{Z}} \\
\big(\frac{a_1 + i b_1}{m}, \ldots, \frac{a_n + i b_n}{m}\big) \in U_{\omega, N}}} \frac{1}{m^{2n}} h_{\omega}\Big(\frac{a_1 + i b_1}{m}, \ldots, \frac{a_n + i b_n}{m}\Big).
\end{multline}
Note that
\[
(\omega \in \Omega_{\infty}) \longmapsto h_{\omega}\Big(\frac{a_1 + i b_1}{m}, \ldots, \frac{a_n + i b_n}{m}\Big)
\]
is measurable, so that \eqref{eqn:prop:measurable:projective:space:01} means that
\[
(\omega \in \Omega_{\infty}) \longmapsto \int_{U_{\omega, N}} h_{\omega}\, \mathrm{d}x_{1\omega} \wedge \mathrm{d}y_{1\omega} \wedge \cdots
\wedge \mathrm{d}x_{n\omega} \wedge \mathrm{d}y_{n\omega}
\]
is measurable. Therefore, one has the assertion because
\begin{multline*}
\lim_{N\to\infty} \int_{U_{\omega, N}} h_{\omega}\, \mathrm{d}x_{1\omega} \wedge \mathrm{d}y_{1\omega} \wedge \cdots
\wedge \mathrm{d}x_{n\omega} \wedge \mathrm{d}y_{n\omega}
\\ = \int_{U_{\omega}} h_{\omega}\, \mathrm{d}x_{1\omega} \wedge \mathrm{d}y_{1\omega} \wedge \cdots
\wedge \mathrm{d}x_{n\omega} \wedge \mathrm{d}y_{n\omega}. \end{multline*}
\end{proof}
\begin{theo}\label{thm:L2:norm:very:ample}
We assume that $\Omega = \Omega_{\infty}$ and $K$ is algebraically closed.
Let $X$ be a $d$-dimensional
projective and integral variety over $K$ and $L$ be a very ample invertible ${\mathscr{O}}_X$-module.
Let $\{ \|\raisebox{.4ex}{.}\|_{\omega} \}_{\omega \in \Omega}$ be a measurable family of hermitian norms on $H^0(X, L)$.
Let $\varphi = \{ \varphi_{\omega} \}_{\omega \in \Omega}$ be a family of metrics on $L$ induced by the surjective homomorphism $H^0(X, L) \otimes {\mathscr{O}}_X \to L$ and
$\{ \|\raisebox{.4ex}{.}\|_{\omega} \}_{\omega \in \Omega}$. For $s \in H^0(X, L) \setminus \{ 0 \}$,
\[(\omega \in \Omega) \mapsto \int_{X_{\omega}} \log |s|_{\varphi_{\omega}} c_1(L_{\omega}, \varphi_{\omega})^{\wedge d}\]
is measurable.
\end{theo}
\begin{proof}
Let $n = \dim_K H^0(X, L) -1$ and $X \hookrightarrow {\mathbb{P}}^n_K$ be the embedding by $L$. Note that $L = \rest{{\mathscr{O}}_{{\mathbb{P}}^n_K}(1)}{X}$.
Since $H^0({\mathbb{P}}^n_K, {\mathscr{O}}_{{\mathbb{P}}^n_K}(1)) \simeq H^0(X, L)$, one has $t \in H^0({\mathbb{P}}^n_K, {\mathscr{O}}_{{\mathbb{P}}^n_K}(1))$ with $\rest{t}{X} = s$.
Let $\psi = \{ \psi_{\omega} \}_{\omega \in \Omega}$ be a family of metics of ${\mathscr{O}}_{{\mathbb{P}}^n_K}(1)$ induced by the surjective homomorphism
$H^0({\mathbb{P}}^n_K, {\mathscr{O}}_{{\mathbb{P}}^n_K}(1)) \otimes {\mathscr{O}}_{{\mathbb{P}}^n_K} \to {\mathscr{O}}_{{\mathbb{P}}^n_K}(1)$ and
$\{ \|\raisebox{.4ex}{.}\|_{\omega} \}_{\omega \in \Omega}$. Note that $\rest{\psi}{X} = \varphi$.
By Proposition~\ref{prop:projection:to:projective:space},
we can choose a linear subspace $M$ in ${\mathbb{P}}^n_K$ such that $\operatorname{codim} M = d+1$, $M \cap X = \emptyset$ and
$M \subseteq \{ t = 0 \}$, so that, by Proposition~\ref{prop:projection:to:projective:space} again,
the morphism $\pi : X \to {\mathbb{P}}^d_K$ induced by
the projection $\pi_M : {\mathbb{P}}^n_K \setminus M \to {\mathbb{P}}^d_K$ with the center $M$ is finite and surjective.
We choose a homogenous coordinate $(T_0:\ldots:T_n)$ on ${\mathbb{P}}^n_K$ such that \[t = T_0\quad\text{and}\quad
M = \{ T_0 = \cdots = T_d = 0 \}.\] Then
$\pi_M$ is given by $(T_0 : \cdots : T_n) \mapsto (T_0 : \cdots : T_d)$.
Let $U$ be a non-empty open of ${\mathbb{P}}^d_K$ such that $\pi : X \to {\mathbb{P}}^d_K$ is \'{e}tale over $U$.
We may assume that $U \subseteq \{ T_0 \not = 0 \}$. We set $X_j = T_j/T_0$ ($j=1, \ldots, n$).
Then \[
\begin{cases} {\mathbb{P}}^n_K \setminus \{ T_0 = 0 \} = \operatorname{Spec}(K[X_1, \ldots, X_n]) = {\mathbb{A}}^n_K, \\
{\mathbb{P}}^d_K \setminus \{ T_0 = 0 \} = \operatorname{Spec}(K[X_1, \ldots, X_d]) = {\mathbb{A}}^d_K
\end{cases}
\] and $\pi_M$ on ${\mathbb{P}}^n_K \setminus \{ T_0 = 0 \}$
is given by $(X_1, \ldots, X_n) \mapsto (X_1, \ldots, X_d)$.
Let \[(z_{1\omega}, \ldots, z_{n\omega})\quad\text{and}\quad (z_{1\omega}, \ldots, z_{d\omega})\] be
the coordinates of ${\mathbb{A}}^n_K \otimes_{\sigma_{\omega}} {\mathbb{C}}$ and ${\mathbb{A}}^d_K \otimes_{\sigma_{\omega}} {\mathbb{C}}$,
respectively. Note that $f_{\omega} := \log |t|_{\psi_{\omega}}$ is $C^{\infty}$ on ${\mathbb{A}}^n_K \otimes_{\sigma_{\omega}} {\mathbb{C}}$.
Then, by Proposition~\ref{prop:measurable:partial:X},
if we set
\[
\rest{f_{\omega}}{X_{\omega}} c_1(L_{\omega}, \varphi_{\omega})^{\wedge d} = i^d h_{\omega} (dz_{1\omega} \wedge d\bar{z}_{1\omega}) \wedge
\cdots \wedge (dz_{d\omega} \wedge d\bar{z}_{d\omega})
\]
on $\pi_{\omega}^{-1}(U_{\omega})$, then, for $P \in \pi^{-1}(U)$,
$(\omega \in \Omega) \mapsto h(P_{\omega})$ is measurable.
Note that
\begin{align*}
\int_{X_{\omega}} \log |s|_{\varphi_{\omega}} c_1(L_{\omega}, \varphi_{\omega})^{\wedge d} & = \int_{\pi_{\omega}^{-1}(U_{\omega})} \rest{f_{\omega}}{X_{\omega}} c_1(L_{\omega}, \varphi_{\omega})^{\wedge d} \\
& = \int_{\pi_{\omega}^{-1}(U_{\omega})} i^d h_{\omega} (dz_{1\omega} \wedge d\bar{z}_{1\omega}) \wedge
\cdots \wedge (dz_{d\omega} \wedge d\bar{z}_{d\omega}) \\
& = \int_{U_{\omega}} i^d (\pi_{\omega})_*(h_{\omega}) (z_{1\omega} \wedge d\bar{z}_{1\omega}) \wedge
\cdots \wedge (dz_{d\omega} \wedge d\bar{z}_{d\omega}).
\end{align*}
Moreover, $(\pi_{\omega})_*(h_{\omega})$ is $C^{\infty}$ over $U_{\omega}$. Further,
for $P \in U(K)$, if we set $\pi^{-1}(P) = \{ Q_1, \ldots, Q_r \}$, then
\[
(\pi_{\omega})_*(h_{\omega})(P_{\omega}) = \sum_{i=1}^r h_{\omega}(Q_{i\omega}),
\]
so that $(\omega \in \Omega) \mapsto (\pi_{\omega})_*(h_{\omega})(P_{\omega})$ is measurable.
Therefore, by Proposition~\ref{prop:measurable:vol:int:A},
\[
(\omega \in \Omega) \mapsto \int_{U_{\omega}} i^d (\pi_{\omega})_*(h_{\omega}) (dz_{1\omega} \wedge d\bar{z}_{1\omega}) \wedge
\cdots \wedge (dz_{d\omega} \wedge d\bar{z}_{d\omega})
\]
is measurable. Thus the assertion follows.
\end{proof}
\begin{theo}\label{thm:measurable:archimedean}
We assume that $\Omega = \Omega_{\infty}$.
Let $X$ be a projective scheme over $K$ and $L$ be an ample invertible ${\mathscr{O}}_X$-module.
Let $\varphi = \{ \varphi_{\omega} \}_{\omega \in \Omega}$ be a measurable family of semipositive metrics.
Then, for $s \in H^0(X, L) \setminus \{ 0 \}$,
\[ (\omega \in \Omega) \longmapsto \int_{X_{\omega}} \log |s|_{\varphi_{\omega}} c_1(L_{\omega},\varphi_{\omega})^d \]
is measurable.
\end{theo}
\begin{proof}
By Lemma~\ref{lem:measurable:integrable:adelic:algebraic:cover},
we may assume that $K$ is algebraically closed.
We choose a positive integer $N$ such that $L^{\otimes n}$ is very ample for
for all $n \geqslant N$.
Let $\varphi_n = \{ \varphi_{n,\omega} \}_{\omega \in \Omega}$ be the Fubini-Study metric of $L^{\otimes n}$ induced by $H^0(X, L^{\otimes n}) \otimes {\mathscr{O}}_X \to L^{\otimes n}$ and $\|\raisebox{.4ex}{.}\|_{n\varphi} = \{ \|\raisebox{.4ex}{.}\|_{n\varphi_{\omega}} \}_{\omega \in \Omega}$.
Moreover, by \cite[Theorem~4.1.26]{CMArakelovAdelic}, there is a measurable Hermitian norm family
$\{ \|\raisebox{.4ex}{.}\|_{n, \omega}^H \}_{\omega \in \Omega}$ on $H^0(X, L^{\otimes n})$ such that
\[
\|\raisebox{.4ex}{.}\|_{n\varphi_\omega} \leqslant \|\raisebox{.4ex}{.}\|_{n, \omega}^H \leqslant (h^0(L^{\otimes n}) + 1)^{1/2} \|\raisebox{.4ex}{.}\|_{n\varphi_\omega}
\]
for $\omega \in \Omega$.
Let $\varphi^H_{n,\omega}$ be the Fubini-Study metric of $L^{\otimes n}$ induced by $H^0(X, L^{\otimes n}) \otimes {\mathscr{O}}_X \to L^{\otimes n}$ and $\|\raisebox{.4ex}{.}\|^H_{n, \omega}$.
Note that
\[
d_{\omega}\left(\frac{1}{n}\varphi_{n}, \frac{1}{n}\varphi^H_{n}\right) \leqslant \frac{ d_{\omega}(\|\raisebox{.4ex}{.}\|_{n\varphi},
\|\raisebox{.4ex}{.}\|_{n}^H)}{n} \leqslant \frac{\ln (h^0(L^{\otimes n}) + 1)}{2n}.
\]
Therefore, if we set
$\psi_{n,\omega} = (1/n)\varphi^H_{n,\omega}$,
then $\lim_{n\to\infty} d_{\omega}(\varphi, \psi_n) = 0$ for all $\omega \in \Omega$
because $\lim_{n\to\infty} d_{\omega}(\varphi, (1/n)\varphi_n) = 0$.
By Theorem~\ref{thm:L2:norm:very:ample},
\[
(\omega \in \Omega_{\infty}) \mapsto \int_{X_{\omega}} \log |s|_{\psi_{n,\omega}} c_1(L_{\omega}, \psi_{n, \omega})^d
= \frac{1}{n^{d+1}} \int_{X_{\omega}} \log |s^n|_{\varphi^H_{n,\omega}} c_1(nL_{\omega}, \varphi^H_{n, \omega})^d
\]
is measurable. Further, by \cite[Corollary~3.6]{Demagbook},
\[
\lim_{n\to\infty} \int_{X_{\omega}} \log |s|_{\psi_{n,\omega}} c_1(L_{\omega}, \psi_{n, \omega})^d = \int_{X_{\omega}} \log |s|_{\varphi_{n}} c_1(L_{\omega}, \varphi_{\omega})^d.
\]
Therefore, the assertion follows.
\end{proof}
Combining Theorems \ref{Thm: measurability non archimedean} and \ref{thm:measurable:archimedean}, we obtain the following result.
\begin{theo}\label{Thm: measurability}
Let $X\rightarrow\operatorname{Spec} K$ be a projective scheme over $K$ and $d$ be the dimension of $X$. Let $D_0,\ldots,D_d$ be Cartier divisors on $X$, which intersect properly. We equip each $D_i$ with a measurable Green function $g_i$ such that $(D_{i},g_{i})$ is integrable. Then the local intersection function
\[(\omega\in\Omega)\longmapsto ((D_0,g_0)\cdots(D_d,g_d))_{\omega}\]
is $\mathcal A$-measurable.
\end{theo}
\begin{proof}The measurability over $\Omega\setminus\Omega_\infty$ follows directly from Theorem \ref{Thm: measurability non archimedean}. Moreover, in view of Theorem \ref{thm:measurable:archimedean}, the measurability over $\Omega_\infty$ follows from Proposition~\ref{prop:intersection:trivial:divisor} and the multi-linearity of
the local intersection measure.
\end{proof}
\begin{theo}\label{thm:integrability:local:intersection}
Let $X\rightarrow\operatorname{Spec} K$ be a projective scheme over $K$ and $d$ be the dimension of $X$. Let $D_0,\ldots,D_d$ be Cartier divisors on $X$, which intersect properly. We equip each $D_i$ with a dominated Green function $g_i$ such that $(D_{i},g_{i})$ is integrable. Then the local intersection function
\begin{equation}\label{eqn:thm:integrability:local:intersection:01}(\omega\in\Omega)\longmapsto ((D_0,g_0)\cdots(D_d,g_d))_{\omega}\end{equation}
is dominated.
\end{theo}
\begin{proof}
By Lemma~\ref{lem:measurable:integrable:adelic:algebraic:cover}, we may assume that
$K$ is algebraically closed.
By using Proposition~\ref{prop:multilinear:symmetric:semiample:case},
we may further assume that $D_0, \ldots, D_d$ are very ample.
Moreover, by Proposition~\ref{prop:reduction:same:class}, we may assume without loss of generality that
there are an integrable adelic line bundle $(L, \varphi)$ and non-zero rational sections $s_0, \ldots, s_d$ of $L$
such that ${\mathscr{O}}_X(D_i) = L$ and $g_i = -\log |s_i|_{\varphi}$ for $i \in \{0,\ldots, d\}$.
Note that $L$ is very ample. Thus, by Proposition~\ref{prop:projection:to:projective:space},
there is a finite and surjective morphism $\pi : X \to {\mathbb{P}}^d_K$ such that
$L = \pi^*({\mathscr{O}}_{{\mathbb{P}}^d}(1))$. Let $(T_0 : \cdots : T_d)$ be a homogeneous coordinate of ${\mathbb{P}}^d_K$.
Let $\varphi_{\rm{FS}}$ be a Fubini-Study metric of ${\mathbb{P}}^d_K$ and $H_i = \{ T_i = 0 \}$ as in Proposition~\ref{prop:local:intersection:PP}. Moreover, we set $h_i = -\log |T_i|_{\varphi_{\rm{FS}}}$.
First we assume that $\varphi = \pi^*(\varphi_{\rm{FS}})$.
If $D_i = \pi^*(H_i)$ for $i \in \{ 0, \ldots, d \}$, then the dominancy of \eqref{eqn:thm:integrability:local:intersection:01} follows from
Proposition~\ref{prop:intersection:finite:morphism} and Proposition~\ref{prop:local:intersection:PP}.
In general, there are non-zero rational functions $f_0, \ldots, f_d$ on $X$
such that $D_i = \pi^*(H_i) + (f_i)$ for $i \in \{ 0, \ldots, d \}$. Then, by Proposition~\ref{prop:intersection:principal:div},
there is an integrable function $\theta$ on $\Omega$ such that
\[
((D_0, g_0) \cdots (D_d, g_d))_{\omega} = ((\pi^*(H_0), \pi^*(h_0)) \cdots (\pi^*(H_d), \pi^*(h_d)))_{\omega} + \theta(\omega).
\]
Thus one has the dominancy of \eqref{eqn:thm:integrability:local:intersection:01}.
In general, there is a family $g$ of integrable continuous functions such that $\varphi = \exp(g) \pi^*(\varphi_{\rm{FS}})$.
In this case, the dominancy of \eqref{eqn:thm:integrability:local:intersection:01} follows from
Corollary~\ref{cor:intersection:trivial:divisor}. \end{proof}
Finally, we obtain the following integrability theorem.
\begin{theo}\label{Thm: integrability of local intersection}
Let $X$ be a projective $K$-scheme of dimension $d$, and $\overline D_0,\ldots,\overline D_d$ be a family of integrable adelic Cartier divisors. Assume that the underlying Cartier divisors $D_0,\ldots,D_d$ intersect properly. Then the function of local intersection numbers
\begin{equation}\label{Equ: local intersection function}(\omega\in\Omega)\longmapsto (\overline D_0\cdots\overline D_d)_\omega\end{equation}
is integrable on the measure space $(\Omega,\mathcal A,\nu)$.
\end{theo}
\begin{defi}
Let $X$ be a projective $K$-scheme of dimension $d$, and $\overline D_0,\ldots,\overline D_d$ be a family of integrable adelic Cartier divisors, such that $D_0,\ldots,D_d$ intersect properly. We define the global intersection number of $\overline D_0,\ldots,\overline D_d$ as
\[(\overline D_0\cdots\overline D_d)_{S}:=\int_{\omega\in\Omega}(\overline D_0\cdots\overline D_d)_\omega\,\nu(\mathrm{d}\omega).\]
\end{defi}
\begin{rema}\label{Rem: global heights equality}
Let $X$ be a projective $K$-scheme of dimension $d$. For any $i\in\{0,\ldots,d\}$, let \[(E_i,\xi_i=(\|\raisebox{.4ex}{.}\|_{i,\omega})_{\omega\in\Omega})\] be a Hermitian adelic vector bundle on $S$, and $f_i:X\rightarrow\mathbb P(E_i)$ be a closed embedding. Let $L_i$ be the restriction of $\mathcal O_{E_i}(1)$ to $X$, which is equipped with the orthogonal quotient metric family $\varphi_i$ induced by $\xi_i$. We choose a global section $s_i$ of $L_i$ such that $s_0,\ldots,s_d$ intersect properly. For each $i\in\{0,\ldots,d\}$, let $D_i$ be the Cartier divisor $\operatorname{div}(s_i)$ and $g_i$ be the Green function family of $D_i$ corresponding to $\varphi_i$. By Theorem \ref{Thm: equality of local height}, if we denote by $R$ the resultant
\[R^{X,s_0,\ldots,s_d}_{f_0,\ldots,f_d}\in S^{\delta_0}(E_0^\vee)\otimes_K\cdots\otimes_KS^{\delta_d}(E_d^\vee),\]
where $\delta_i=(D_0\cdots D_{i-1}D_{i+1}\cdots D_d)$,
then the following equality holds
\[\begin{split}&(\overline D_0\cdots\overline D_d)=\int_{\omega\in\Omega\setminus\Omega_\infty}\ln\|R\|_\omega\,\nu(\mathrm{d}\omega)\\&\quad+
\int_{\sigma\in\Omega_\infty}\nu(\mathrm{d}\sigma)\int_{\mathbb S_{0,\sigma}\times\cdots\times \mathbb S_{d,\sigma}}\ln|R_\sigma(z_0,\ldots,z_d)|\,\eta_{\mathbb S_{0,\sigma}}(\mathrm{d}z_0)\otimes\cdots\otimes\eta_{\mathbb S_{d,\sigma}}(d\mathrm{z}_d)\\
&\qquad+\nu(\Omega_\infty)\frac 12\sum_{i=0}^{d}\delta_i\sum_{\ell=1}^{r_i}\frac{1}{\ell},
\end{split}\]
where
\begin{enumerate}[label=\rm(\arabic*)]
\item $\|\raisebox{.4ex}{.}\|_\omega$ is the $\varepsilon$-tensor product of $\delta_i$-th $\varepsilon$-symmetric tensor power of $\|\raisebox{.4ex}{.}\|_{i,\omega,*}$,
\item $R_\sigma$ is the element of
\[S^{\delta_0}(E_{0,\mathbb C_\sigma}^\vee)\otimes_{\mathbb C_\sigma}\cdots\otimes_{\mathbb C_\sigma}S^{\delta_{d,\mathbb C_\sigma}}(E_{d,\mathbb C_\sigma}^\vee)\]
indued by $R$,
\item $\mathbb S_{i,\sigma}$ is the unique sphere of $(E_{i,\mathbb C_\sigma},\|\raisebox{.4ex}{.}\|_{i,\sigma,\mathbb C_\sigma})$,
\item $\eta_{\mathbb S_{i,\sigma}}$ is the $U(E_{i,\mathbb C_\sigma},\|\raisebox{.4ex}{.}\|_{i,\mathbb C_\sigma})$-invariant Borel probaility measure on $\mathbb S_{i,\sigma}$.
\end{enumerate}
\end{rema}
\section{Invariance of intersection number by coverings}
Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ be an adelic curve. Consider a covering \[\alpha=(\alpha^{\#},\alpha_{\#},I_\alpha)\] from another adelic curve $S'=(K',(\Omega',\mathcal A',\nu'),\phi')$ to $S$ (see Definition \ref{Def: covering of adelic curves}). We assume that, either both $\sigma$-algebra $\mathcal A$ and $\mathcal A'$ is discrete, or there exist countable subfields $K_0$ and $K_0'$ of $K$ and $K'$ respectively, such that $K_0$ is dense in each $K_\omega$ with $\omega\in\Omega$, and $K_0'$ is dense in each $K'_{\omega'}$ with $\omega'\in\Omega'$. Recall that $\alpha^{\#}:K\longrightarrow K'$ is a field homomorphism, \[\alpha_{\#}:(\Omega',\mathcal A')\rightarrow (\Omega,\mathcal A)\] is a measurable map, and \[I_\alpha:\mathscr L^1(\Omega',\mathcal A',\nu')\longrightarrow\mathscr L^1(\Omega,\mathcal A,\nu)\] is a disintegration kernel of $\nu'$ over $\nu$ such that, for any $g\in\mathscr L^1(\Omega,\mathcal A,\nu)$, one has $g\circ\alpha_{\#}\in\mathscr L^1(\Omega',\mathcal A',\nu')$ and
$I_\alpha(g\circ\alpha_{\#})=g$.
In this section, we consider a projective scheme $X$ of dimension $d$ over $\operatorname{Spec} K$ and a family \[\overline D_0=(D_0,g_0),\ldots,\overline{D}_d=(D_d,g_d)\] of adelic Cartier divisors, such that $D_0,\ldots,D_d$ intersect properly. The purpose of this section is to define the extension of scalars $\overline D_{i,\alpha}$ of each adelic Cartier divisor $\overline D_i$ by $\alpha$ and show the following equality
\[(\overline D_{0,\alpha}\cdots\overline D_{d,\alpha})_{S'}=(\overline D_0\cdots\overline D_d)_S.\]
\begin{defi}
Let $D$ be a Cartier divisor on $X$ and $g=(g_\omega)_{\omega\in\Omega}$ be a Green function family of $D$ (see Definition \ref{def:adelic:Cartier:divisor}). Let $X_{\alpha}$ be the fiber product \[X\times_{\operatorname{Spec} K,\alpha^{\#}}\operatorname{Spec} K'\] and $D_\alpha$ be the pull-back of $D$ by the morphism of projection $X_{\alpha}\rightarrow X$. If $\omega'$ is an element of $\Omega'$ and $\omega=\alpha_{\#}(\omega')$, then the Cartier divisor $D_{\alpha,\omega'}$ identifies with the pull-back of $D_\omega$ by the morphism of projection
\[X_{\alpha,\omega'}=X_{\alpha}\times_{\operatorname{Spec} K'}\operatorname{Spec} K'_{\omega'}\cong X_\omega\times_{\operatorname{Spec} K_\omega}\operatorname{Spec} K'_{\omega'}\longrightarrow X_\omega.\]
We denote by $g_{\alpha,\omega'}$ the Green function $g_{\omega,K'_{\omega'}}$ (see Remark \ref{Rem: extension scalar Green function}). Then the family $g_\alpha:=(g_{\alpha,\omega'})_{\omega'\in\Omega'}$ forms a Green function family of the Cartier divisor $D_\alpha$.
Let $L$ be an invertible $\mathcal O_X$-module and $\varphi=(\varphi_\omega)_{\omega\in\Omega}$ be a metric family on $L$. We denote by $L_\alpha$ the pull-back of $L$ by the morphism of projection $X_{\alpha}\rightarrow X$. If $\omega'$ is an element of $\Omega'$ and $\omega=\alpha_{\#}(\omega')$, then the invertible sheaf $L_{\alpha,\omega'}$ identifies with the pull-back of $L_\omega$ by the morphism of projection $X_{K',\omega'}\rightarrow X_\omega$. We denote by $\varphi_{\alpha,\omega'}$ the continuous metric $\varphi_{\omega,K'_{\omega'}}$ (see Example \ref{Exa: Fubini-Study} \ref{Item: extension of scalars metric}) on $L_{\alpha,\omega'}$. Then the family $\varphi_\alpha:=(\varphi_{\alpha,\omega'})_{\omega'\in\Omega'}$ forms a metric family of $L_\alpha$. Note that, if $s$ is a regular meromorphic section of $L$, $D=\operatorname{div}(s)$ and $g=(g_\omega)_{\omega\in\Omega}$ is the Green function family of $D$ corresponding to the metric family $\varphi$, then $g_\alpha$ is the Green function family of $D_\alpha$ corresponding to $\varphi_{\alpha}$.
\end{defi}
\begin{prop}Let $\pi:X\rightarrow\operatorname{Spec} K$ be a projective $K$-scheme.
\begin{enumerate}[label=\rm(\arabic*)]
\item\label{Item: pull-back adelic line bundle} Let $L$ be an invertible $\mathcal O_X$-module and $\varphi$ be a metric family on $L$. If $\varphi$ is dominated, then $\varphi_\alpha$ is also dominated.
\item\label{Item: pull-back dominated} Let $D$ be a Cartier divisor on $X$ and $g$ is a Green function family of $g$. If $g$ is dominated, then $g_\alpha$ is also dominated.
\end{enumerate}
\end{prop}
\begin{proof}
It suffices to prove the first statement. Assume that $\psi$ is another metric family on $L$. If $\omega'$ is an element of $\Omega'$ and if $\omega=\alpha_{\#}(\omega')$, then by \eqref{Equ: distance extension of scalars} one has
\[d_{\omega'}(\varphi_\alpha,\psi_\alpha)=d_\omega(\varphi,\psi).\]
Therefore, if the function $(\omega\in\Omega)\mapsto d_\omega(\varphi,\psi)$ is dominated, then also is the function $(\omega'\in\Omega')\mapsto d_{\omega'}(\varphi_\alpha,\psi_\alpha)$. To prove that the metric family $\varphi$ is dominated, we can assume without loss of generality that there exist a finite-dimensional vector space over $K$, a strongly dominated norm family $\xi=(\|\raisebox{.4ex}{.}\|_{\omega})_{\omega\in\Omega}$ on $E$, a positive integer $n$ and a surjective homomorphism $f:\pi^*(E)\rightarrow L^{\otimes n}$ such that $\varphi$ identifies with the orthogonal quotient metric family induced by $\xi$ (see Definition \ref{Def: Fubini-Study metric}). We may assume further that $\xi$ is Hermitian and $E$ admits a basis $\boldsymbol{e}$ which is orthonormal with respect to all norms $\|\raisebox{.4ex}{.}\|_\omega$.
For any $\omega'\in\Omega'$, let $\|\raisebox{.4ex}{.}\|_{\omega'}$ be the norm $\|\raisebox{.4ex}{.}\|_{\omega,K'_{\omega'}}$, where $\omega=\alpha_{\#}(\omega')$. Then $\xi_\alpha^{H}=(\|\raisebox{.4ex}{.}\|_{\omega'})_{\omega'\in\Omega'}$ is a norm family on $E_{K'}$. Moreover, if we view $\boldsymbol{e}$ as a basis of $E_{K'}$ over $K'$, then it is orthonormal with respect to all norms $\|\raisebox{.4ex}{.}\|_{\omega'}$. In particular, the norm family $\xi_\alpha^{H}$ is strongly dominated. Since $\varphi_\alpha$ coincides with the orthogonal quotient metric family induced by $\xi_\alpha^{H}$, we deduce that the metric family $\varphi_\alpha$ is also dominated.
\end{proof}
\begin{defi}
Let $E$ be a finite-dimensional vector space over $K$ and $\xi=(\|\raisebox{.4ex}{.}\|_\omega)_{\omega\in\Omega}$ be a norm family on $E$. We denote by $\xi^\vee=(\|\raisebox{.4ex}{.}\|_{\omega,*})_{\omega\in\Omega}$ the \emph{dual norm family} on $E^\vee$, which is defined as
\[\forall\,f\in E^\vee_{K_\omega},\quad \|f\|_{\omega,*}:=\sup_{s\in E_{K_\omega}\setminus\{0\}}\frac{|f(s)|_\omega}{\|s\|_\omega}.\]
By \cite[Proposition 4.1.24 (1.b)]{CMArakelovAdelic}, the norm family $\xi^\vee$ is measurable.
We define $\xi_{\alpha}=(\|\raisebox{.4ex}{.}\|_{\omega'})_{\omega'\in\Omega'}$ the following norm family on $E_\alpha:=E\otimes_{K,\alpha^{\#}}K'$. In the case where $\omega'$ is non-Archimedean, the norm $\|\raisebox{.4ex}{.}\|_{\omega'}$ is the $\varepsilon$-extension of scalars of $\|\raisebox{.4ex}{.}\|_\omega$, where $\omega=\alpha_{\#}(\omega')$; in other words, one has
\[\forall\,s\in E_{\alpha,K'_{\omega'}},\quad \|s\|_{\omega'}=\sup_{f\in E_{K_\omega}^\vee\setminus\{0\}}\frac{|f(s)|_{\omega'}}{\|f\|_{\omega,*}}.\]
In the case where $\omega'$ is Archimedean, the norm $\|\raisebox{.4ex}{.}\|_{\omega'}$ is the $\pi$-extension of scalars of $\|\raisebox{.4ex}{.}\|_\omega$, in other words, one has
\[\forall\,s\in E_{\alpha,K_{\omega'}'},\quad \|s\|_{\omega'}=\inf\left\{|\lambda_1|_{\omega'}\cdot\|s_1\|_\omega+\cdots+|\lambda_N|_{\omega'}\cdot\|s_N\|_\omega\,\left|\,\begin{subarray}{l}
N\in\mathbb N,\,N\geqslant 1\\
(\lambda_1,\ldots,\lambda_N)\in (K'_{\omega'})^N\\
(s_1,\ldots,s_N)\in E_\omega^N\\
s=\lambda_1s_1+\cdots+\lambda_Ns_N
\end{subarray}\right.\right\}.\]
Similarly, we define $\xi_{\alpha,\varepsilon}$ the norm family on $E_\alpha$ consisting of $\varepsilon$-extension of scalars (for both non-Archimedean and Archimedean absolute values).
\end{defi}
\begin{lemm}\label{Lem: norm family extension of scalars}
Let $E$ be a finite-dimension vector space over $K$ and $\xi=(\|\raisebox{.4ex}{.}\|_{\omega})_{\omega\in\Omega}$ be a measurable norm family on $E$. Then the norm families $\xi_{\alpha,\varepsilon}$ and $\xi_{\alpha}$ defined above are also measurable.
\end{lemm}
\begin{proof}
The proof is very similar to that of \cite[Proposition 4.1.24 (1.c)]{CMArakelovAdelic}. The case where $\mathcal A$ and $\mathcal A'$ are discrete is trivial. In the following, we will treat the case where $K$ and $K'$ admit countable subfields $K_0$ and $K_0'$ such that $K_0$ is dense in each $K_\omega$ with $\omega\in\Omega$, and $K_0'$ is dense in each $K'_{\omega'}$ with $\omega'\in\Omega'$, respectively. We first check the measurability of $\xi_{\alpha,\varepsilon}$. For any $\omega'\in\Omega'$, let $\|\raisebox{.4ex}{.}\|_{\omega',\varepsilon}$ be the norm indexed by $\omega'$ in the family $\xi_{\alpha,\varepsilon}$. Let $H_0$ be a finite-dimensional $K_0$-vector subspace of $E^\vee$ which generates $E^\vee$ as a vector space over $K$. Then $H_0\setminus\{0\}$ is dense in $E_{K_\omega}^\vee\setminus\{0\}$ for any $\omega\in\Omega$. If $s$ is an element of $E_\alpha$, then for any $\omega'\in\Omega'$,
\[\|s\|_{\omega',\varepsilon}=\sup_{f\in H_0\setminus\{0\}}\frac{|f(s)|_{\omega'}}{\|f\|_{\omega,*}}.\]
Hence it is the supremum of a countable family of $\mathcal A'$-measurable function in $\omega'$. As for the second statement, it suffices to apply the first statement to $\xi^\vee$ to obtain the measurability of $(\xi^{\vee})_{\alpha,\varepsilon}$. Since $\xi_{\alpha}$ is the dual norm family of $(\xi^{\vee})_{\alpha,\varepsilon}$ (see \cite[Proposition 1.3.20]{CMArakelovAdelic}), by \cite[Proposition 4.1.24 (1.c)]{CMArakelovAdelic} we obtain the measurability of $\xi_\alpha$.
\end{proof}
\begin{prop}
Let $X$ be a projective scheme over $\operatorname{Spec} K$. \begin{enumerate}[label=\rm(\arabic*)]
\item\label{Item: pull-back adelic line bundle} Let $L$ be an invertible $\mathcal O_X$-module and $\varphi$ be a metric family on $L$. We assume that $L$ is ample and all metrics in the family $\varphi$ is semi-positive If $\varphi$ is measurable, then $\varphi_\alpha$ is also measurable.
\item\label{Item: pull-back dominated} Let $D$ be a Cartier divisor on $X$ and $g$ be a Green function family of $g$. Assume that $D$ is ample and $g$ is semi-positive. If $g$ is measurable, then $g_\alpha$ is also measurable.
\end{enumerate}
\end{prop}
\begin{proof}
It suffices to prove the first statement. Similarly to the proof of Theorem \ref{Thm: measurability non archimedean}, for any $m\in\mathbb N_{\geqslant 1}$ such that $L^{\otimes m}$ is very ample we choose a norm family $\xi_m=(\|\raisebox{.4ex}{.}\|_{\omega}^{(m)})_{\omega\in\Omega}$ on $H^0(X,L^{\otimes m})$ such that $H^0(X,L^{\otimes m})$ admet a basis which is orthonormal with respect to each norm $\|\raisebox{.4ex}{.}\|_{\omega}^{(m)}$. This norm family is clearly measurable. For any $b>0$ and any $\omega\in\Omega$, let $\varphi_{b,\omega}^{(m)}$ the quotient metric on $L$ induced by the norm
\[\|\raisebox{.4ex}{.}\|^{(m)}_{b,\omega}=\max\{\|\raisebox{.4ex}{.}\|_{\varphi^{\otimes m}},b\|\raisebox{.4ex}{.}\|^{(m)}_\omega\}\]
on $H^0(X_\omega,L_\omega^{\otimes m})$. By Proposition \ref{Pro: measurable direct image}, the norm family $\xi_{b}^{(m)}:=(\|\raisebox{.4ex}{.}\|_{b,\omega}^{(m)})_{\omega\in\Omega}$ is measurable. By Lemma \ref{Lem: norm family extension of scalars}, we deduce that the norm family $\xi_{b,\alpha}^{(m)}$ of $H^0(X_\alpha,L_\alpha^{\otimes m})$ is $\mathcal A'$-measurable.
Let $\varphi_{b}^{(m)}$ be the quotient metric family on $L$ induced by $\xi_{b}^{(m)}$. By \cite[Remark 2.2.14]{CMArakelovAdelic}, the metric $\varphi_{b,\alpha}^{(m)}$ identifies with the quotient metric family on $L_\alpha$ induced by $\xi_{b,\alpha}^{(m)}$. Since the norm family $\xi_{b,\alpha}^{(m)}$ is measurable, by \cite[Proposition 6.1.30]{CMArakelovAdelic}, the metric family $\varphi_{b,\alpha}^{(m)}$ is measurable. By Proposition \ref{Pro: varphi 1 as a continuous metric}, for any fixed $\omega'\in\Omega'$ and $\omega=\alpha_{\#}(\omega')$, for sufficiently small $b$ one has
$\varphi_{b,\omega}^{(m)}=\varphi^{(m)}_{\omega}$ and hence $\varphi_{b,\alpha,\omega'}^{(m)}=\varphi^{(m)}_{\alpha,\omega'}$. Therefore, by \cite[Proposition 6.1.29]{CMArakelovAdelic} we obtain that $\varphi^{(m)}_\alpha$ is measurable.
By \eqref{Equ: distance extension of scalars}, for any $\omega'\in\Omega'$ and $\omega=\alpha_{\#}(\omega)$, one has
\[d_{\omega'}(\varphi^{(m)}_\alpha,\varphi_\alpha)\leqslant d_{\omega}(\varphi^{(m)},\varphi).\]
Since the metric family $\varphi$ is semi-positive, by Proposition \ref{Pro: convergence of varphin}, we deduce that, for any $\omega'\in\Omega'$, one has
\[\lim_{m\rightarrow+\infty}d_{\omega'}(\varphi_\alpha^{(m)},\varphi_\alpha)=0.\]
Still by \cite[Proposition 6.1.29]{CMArakelovAdelic}, we obtain that the metric family $\varphi$ is measurable.
\end{proof}
\begin{theo}
Let $X$ be a projective scheme over $\operatorname{Spec} K$ and $d$ be the dimension of $X$. Let $D_0,\ldots,D_d$ be Cartier divisors on $X$ which intersects properly. We assume that each Cartier divisor $D_i$ is equipped with an integrable Green function family $g_i$. The the following equality holds
\[((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{\omega'}=((D_0,g_0)\cdots(D_d,g_d))_{\alpha_{\#}(\omega')}.\]
In particular, if all Green function family $g_i$ are dominated (resp. measurable), then the function \[(\omega'\in\Omega')\longmapsto ((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{\omega'}\]is dominated (resp. measurable). If all $(D_i,g_i)$ are adelic Cartier divisors, then the following equality holds
\[((D_0,g_0)\cdots(D_d,g_d))_S=((D_{0,\alpha},g_{d,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{S'}.\]
\end{theo}
\begin{proof}
For any $\omega'\in\Omega'$ and $\omega=\alpha_{\#}(\omega')$, the equality
\[((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{\omega'}=((D_0,g_0)\cdots(D_d,g_d))_\omega\]
follows from \ref{prop:invariance:local:intersection:field:extension} (see also Remark \ref{Rem: extension of scalars archimedean}).
If $g_0,\ldots,g_d$ are measurable, by Theorem \ref{Thm: measurability}, the function
\[(\omega\in\Omega)\longmapsto ((D_0,g_0)\cdots(D_d,g_d))_\omega\]
is $\mathcal A$-measurable. Since $\alpha_{\#}$ is a measurable map, we deduce that the function
\[(\omega'\in\Omega')\longmapsto((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha})_{\omega'}\]
is $\mathcal A'$-measurable.
Assume that the Green function families $g_0,\ldots,g_d$ are dominated. By Theorem \ref{thm:integrability:local:intersection}, there exists an integrable function $F$ on the measure space $(\Omega,\mathcal A,\nu)$ such that
\[\forall\,\omega\in\Omega,\quad |((D_0,g_0)\cdots(D_d,g_d))_{\omega}|\leqslant F(\omega).\]
Hence
\[\forall\,\omega'\in\Omega,\quad |((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{\omega'}|\leqslant F(\alpha_{\#}(\omega')).\]
Hence the function
\[(\omega'\in\Omega')\longmapsto ((D_{0,\alpha},g_{0,\alpha}),\cdots,(D_{d,\alpha},g_{d,\alpha}))_{\omega'}\]
is dominated. Finally, if the function
\[(\omega\in\Omega)\longmapsto ((D_0,g_0)\cdots(D_d,g_d))_\omega\]
is integrable, then also is the function
\[(\omega'\in\Omega')\longmapsto ((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{\omega'}=((D_0,g_0)\cdots(D_d,g_d))_{\alpha_{\#}(\omega')}\]
is also integrable, and one has
\[\begin{split}&\quad\;((D_{0,\alpha},g_{d,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{S'}=\int_{\Omega'}((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{\omega'}\,\nu'(\mathrm{d}\omega')\\
&=\int_{\Omega}I_\alpha(\omega'\longmapsto((D_{0,\alpha},g_{0,\alpha})\cdots(D_{d,\alpha},g_{d,\alpha}))_{\omega'})\,\nu(\mathrm{d}\omega)\\
&=\int_{\Omega}((D_0,g_0)\cdots(D_d,g_d))_\omega\,\nu(\mathrm{d}\omega)=((D_0,g_0)\cdots(D_d,g_d))_S.
\end{split}\]
\end{proof}
\section{Multi-heights}
From now on, we assume that the adelic curve $S$ is proper.
\begin{defi}
Let $X$ be a projective scheme over $\operatorname{Spec} K$. If $f$ is a regular meromorphic function on $X$, we denote by $\widehat{\operatorname{div}}(f)$ the following adelic Cartier divisor
\[(\operatorname{div}(f),(-\ln|f|_\omega)_{\omega\in\Omega}).\]
If $\overline L=(L,\varphi)$ is an adelic line bundle on $X$ and if $s$ is a regular meromorphic section of $L$ on $X$, we denote by $\widehat{\operatorname{div}}(s)$ the following adelic Cartier divisor
\[(\operatorname{div}(s),(-\ln|s|_{\varphi_\omega})_{\omega\in\Omega}).\]
\end{defi}
\begin{prop}\label{prop:global:intersection:proper:adelic}
Let $X$ be a projective $K$-scheme of dimension $d$, and $\overline D_0,\ldots,\overline D_d$ and
$\overline D'_0,\ldots,\overline D'_d$ be families of integrable adelic Cartier divisors, such that $D_0,\ldots,D_d$ and $D'_0,\ldots,D'_d$ intersect properly.
If there is a family of regular meromorphic functions $f_0, \ldots, f_d$ on $X$
such that $\overline D_i = \overline D'_i + \widehat{\operatorname{div}}(f_i)$ for $i \in \{ 0, \ldots, d \}$.
Then \[(\overline D_0\cdots\overline D_d)_S =(\overline D'_0\cdots\overline D'_d)_S.\]
\end{prop}
\begin{proof}
It is sufficient to prove that if $f$ is a regular meromorphic function on $X$ and
$\overline D_1,\ldots,\overline D_d$ are integrable adelic Cartier divisors such that
$\operatorname{div}(f), D_1,\ldots,D_d$ intersect properly,
then $(\widehat{\operatorname{div}}(f) \cdot \overline D_1\cdots\overline D_d)_S = 0$.
Clearly we may assume that $K$ is algebraically closed, so that the assertion follows from
Proposition~\ref{prop:intersection:principal:div} and the product formula.
\end{proof}
\begin{defi}
Let $\overline{L}_0=(L_0,\varphi_0), \ldots, \overline{L}_d=(L_d,\varphi_d)$ be a family of integrable adelic line bundles.
Let $s_0, \ldots, s_d$ be regular meromorphic sections of $L_0, \ldots, L_d$, respectively such that
$\operatorname{div}(s_0), \ldots, \operatorname{div}(s_d)$ intersect properly.
Then, by Proposition~\ref{prop:global:intersection:proper:adelic}, the global intersection number
\[
(\widehat{\operatorname{div}}(s_0) \cdots \widehat{\operatorname{div}}(s_d))_S
\]
does not depend on the choice of $s_0, \ldots, s_d$. The \emph{global intersection number}
\[
(\overline{L}_0 \cdots \overline{L}_d )_S
\]
of $\overline{L}_0 \cdots \overline{L}_d$ over $S$ is then defined as \[(\widehat{\operatorname{div}}(s_0) \cdots \widehat{\operatorname{div}}(s_d))_S.\]
This number is also called the \emph{multi-height} of $X$ with respect to $\overline L_0,\ldots,\overline L_d$ and is denoted by
\[h_{\overline L_0,\ldots\overline L_d}(X).\]
In the particular case where $\overline L_0,\ldots\overline L_d$ are all equal to the same integrable adelic line bundle $\overline L$, the number $h_{\overline L,\ldots,\overline L}(X)$ is denoted by $h_{\overline L}(X)$ in abbreviation, and is called the \emph{height} of $X$ with respect to $\overline L$.
\end{defi}
\begin{prop}
\begin{enumerate}[label=\rm(\arabic*)]
\item The global intersection pairing is a symmetric bilinear form on the group consisting of integrable adelic line bundle.
\item
Let $X_1, \ldots, X_\ell$ be irreducible components of $X$ and $\eta_1, \ldots, \eta_\ell$ be the generic points of $X_1, \ldots, X_\ell$, respectively.
Then
\[
\big(\overline{L}_0 \cdots \overline{L}_d\big)_S = \sum_{j=1}^{\ell} \operatorname{length}_{\mathcal O_{X, \eta_j}}(\mathcal O_{X, \eta_j}) \big(\rest{\overline{L}_0}{X_j} \cdots \rest{\overline{L}_d}{X_j}\big)_S.
\]
\item Let $s_d$ be a regular meromorphic section of $L_d$ and $\operatorname{div}(s_d) = a_1 Z_1 + \cdots + a_n Z_n$ be the decomposition
as cycles. Then
\[\begin{split}
(\overline{L}_0 \cdots \overline{L}_{d})_S =\int_{\Omega} & {\left(\int_{X_{\omega}^{\operatorname{an}}} -\log |s_d|_{\varphi_\omega}(x)\, \mu_{(L_{0,\omega},\varphi_{0,\omega}), \cdots (L_{d-1,\omega},\varphi_{d-1,\omega})}(\mathrm{d}x)\right)}\nu(\mathrm{d}\omega)\\
&+
\sum_{i=1}^n a_i (\rest{\overline{L}_0}{Z_i} \cdots \rest{\overline{L}_{d-1}}{Z_i})_S.
\end{split}
\]
\end{enumerate}
\end{prop}
\begin{proof}
They follows from \eqref{eqn:def:local:intersection} and Proposition~\ref{prop:multilinear:symmetric:semiample:case}.
\end{proof}
Finally let us consider the projection formula for our intersection theory.
For this purpose, we need three lemmas.
\begin{lemm}\label{lem:Artin:local:algebra:formula}
Let $(A,\mathfrak m)$ be a local Artinian ring and $B$ be an $A$-algebra such that $B$ is finitely generated as an $A$-module.
Let $M$ be a finitely generated $B$-module. Then
\[
\operatorname{length}_A(M) = \sum_{\mathfrak n \in \operatorname{Spec}(B)} [B/\mathfrak n : A/\mathfrak m] \operatorname{length}_{B_{\mathfrak n}}(M_{\mathfrak n}).
\]
In particular, if $B$ is flat over $A$, then
\[\operatorname{rk}_A(B) \operatorname{length}_A(A) = \sum_{\mathfrak n \in \operatorname{Spec}(B)} [B/\mathfrak n : A/\mathfrak m] \operatorname{length}_{B_{\mathfrak n}}(B_{\mathfrak n}).\]
\end{lemm}
\begin{proof}
Let $0 \to M' \to M \to M'' \to 0$ be an exact sequence of finitely generated $B$-modules. Then,
both sides of the above first equation are additive with respect to the exact sequence.
Therefore, we may assume that $M = B/\mathfrak{n}$ for some $\mathfrak n \in \operatorname{Spec}(B)$. In this case, it is obvious.
\end{proof}
\begin{lemm}\label{lem:associated:prime:maps:zero:ideal}
Let $A$ be an integral domain and $B$ be a flat $A$-algebra.
If we denote the structure homomorphism $A \to B$ by $\phi$, then
$\phi^{-1}(P) = \{ 0 \}$ for any $P \in \mathrm{Ass}_B(B)$.
\end{lemm}
\begin{proof}
We set $P = \mathrm{ann}(b)$ for some $b \in B \setminus \{ 0 \}$.
If there is $a \in \phi^{-1}(P) \setminus \{ 0 \}$,
then $\phi(a) b = 0$.
Since $B$ is flat over $A$, $\phi(a)$ is regular, so that $b = 0$.
This is a contradiction.
\end{proof}
\begin{lemm}\label{lem:cycle:push:forward:base:change}
Let $f : Y \to X$ be a proper and surjective morphism of integral scheme of finite type over a field $k$ such that
$\dim X = \dim Y$.
For an extension filed $k'$ of $k$, if $X' := X \times_{\operatorname{Spec}(k)}\operatorname{Spec}(k')$, $Y' := Y \times_{\operatorname{Spec}(k)}\operatorname{Spec}(k')$ and
$f' : X' \to Y'$ is the induced morphism, then
\[f'_*([X']) = [k(Y) : k(X)][Y'].\]
\end{lemm}
\begin{proof}
By Lemma~\ref{lem:associated:prime:maps:zero:ideal}, any irreducible component of $X'$ (resp. $Y'$) maps surjectively to $X$ (resp. $Y$) by $X' \to X$ (resp. $Y' \to Y$).
Moreover, we can find a non-empty Zariski open set $U$ of $X$ such that $f^{-1}(U) \to U$ is finite and flat.
Note that if we set $U' := U \times_{\operatorname{Spec}(k)}\operatorname{Spec}(k')$, then
${f'}^{-1}(U') = f^{-1}(U) \times_{\operatorname{Spec}(k)}\operatorname{Spec}(k')$ and ${f'}^{-1}(U') \to U'$ is finite and flat.
Therefore, we may assume that $f$ is finite and flat, so that
the assertion
is a consequence of the second formula in Lemma~\ref{lem:Artin:local:algebra:formula}.
\end{proof}
\begin{defi}
Let $Z = a_1 Z_1 + \cdots + a_r Z_r$ be an $l$-dimensional cycle on $X$ and
$\overline{L}_0, \ldots, \overline{L}_l$ be integrable adelic line bundles.
Then $(\overline{L}_0 \cdots \overline{L}_l \mid Z)_S$ is defined to be
\[
(\overline{L}_0 \cdots \overline{L}_l \mid Z)_S := \sum_{j=1}^r a_j \Big(\rest{\overline{L}_0}{Z_j} \cdots \rest{\overline{L}_l}{Z_j}\Big)_S.
\]
In the case where $\overline L_0,\ldots,\overline{L}_l$ are all equal to the same adelic line bundle $\overline L$, we call it the \emph{height} of the cycle $Z$ with respect to $\overline L$, and denote it by $h_{\overline L}(Z)$.
\end{defi}
\begin{theo}[Projection formula]\label{prop:projection:formula:intersection}
Let $f : Y \to X$ be a morphism of projective schemes over $K$ and $\overline{L}_0, \ldots, \overline{L}_l$
be integrable adelic line bundles on $X$.
For an $l$-cycle $Z$ on $Y$,
\[(f^*(\overline{L}_0) \cdots f^*(\overline{L}_l) \mid Z)_S = (\overline{L}_0 \cdots \overline{L}_l \mid f_*(Z))_S.\]
\end{theo}
\begin{proof}
First let us see the following:
\begin{enonce}{Claim}\label{claim:prop:projection:formula:intersection:01}
If $f$ is a surjective morphism of projective integral schemes over $K$, then
\[
(f^*(\overline{L}_0) \cdots f^*(\overline{L}_l))_S = \begin{cases}
\deg(f) (\overline{L}_0 \cdots \overline{L}_l)_S & \text{if $\dim X = \dim Y$},\\
0 & \text{if $\dim X < \dim Y$}.
\end{cases}
\]
In other words,
\[ (f^*(\overline{L}_0) \cdots f^*(\overline{L}_l) \mid Y)_S = (\overline{L}_0 \cdots \overline{L}_l \mid f_*(Y))_S.
\]
\end{enonce}
\begin{proof}
We choose rational sections $s_0, \ldots, s_d$ of $L_0,\ldots, L_d$, respectively such that
$\operatorname{div}(s_0), \ldots, \operatorname{div}(s_d)$ intersect properly on $X$ and
$f^*(\operatorname{div}(s_0)), \ldots, f^*(\operatorname{div}(s_d))$ intersect properly on $Y$.
Let $K_{\omega}$ be the completion of $K$ with respect to $\omega \in \Omega$,
$X_{\omega} := X \times_{\operatorname{Spec}(K)} \operatorname{Spec}(K_{\omega})$, $Y_{\omega} := Y \times_{\operatorname{Spec}(K)} \operatorname{Spec}(K_{\omega})$
and $f_{\omega} : Y_{\omega} \to X_{\omega}$ be the induced morphism.
Further let $\pi_{X,\omega} : X_{\omega} \to X$ and $\pi_{Y,\omega} : Y_{\omega} \to Y$ be the projections.
Then the following diagram is commutative.
\[
\xymatrix{
Y_{\omega} \ar[r]^-{f_{\omega}}\ar[d]_-{\pi_{Y,\omega}}& X_{\omega}\ar[d]^-{\pi_{X,\omega}} \\
Y \ar[r]_-{f}& X
}
\]
Since $X$ and $Y$ are integral, $f^*(\operatorname{div}(s_i))$ is well defined as a Cartier divisor.
Moreover, $\pi_{Y,\omega}^*(f^*(\operatorname{div}(s_i)))$ and $\operatorname{div}(s_i)_{\omega} := \pi_{X,\omega}^*(\operatorname{div}(s_i))$ are defined because $\pi_{Y,\omega}$ and $\pi_{X,\omega}$ are flat.
Therefore, $f_{\omega}^*(\operatorname{div}(s_i)_{\omega})$ is defined as a Cartier divisor on $Y_{\omega}$ for each
$i = 0, \ldots, d$.
Let $Y_{\omega, 1}, \ldots, Y_{\omega, m_{\omega}}$ (resp. $X_{\omega, 1}, \ldots, X_{\omega, n_{\omega}}$)
be irreducible components of $Y_{\omega}$ (resp. $X_{\omega}$).
First we assume that $\dim X < \dim Y$. Then, by Proposition~\ref{prop:intersection:finite:morphism},
\[
\Big(\rest{f_{\omega}^*(\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega})}{Y_{\omega,j}} \cdots
\rest{f_{\omega}^*(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})}{Y_{\omega,j}}\Big)_{\omega} =
0\]
for all $j = 1, \ldots, m_{\omega}$.
Therefore, \[
\Big(f_{\omega}^*(\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega}) \cdots
f_{\omega}^*(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})\Big)_{\omega} = 0,
\]
and hence
the assertion follows.
Next we assume that $\dim X = \dim Y$.
For each $i \in\{ 1, \ldots, n_{\omega}\}$, let \[J_{\omega, i} := \{ j \in \{ 1,\ldots, m_{\omega} \} \mid f_{\omega}(Y_{\omega, j}) = X_{\omega, i} \}\] and \[J_{\omega, 0} := \{ 1, \ldots, n_{\omega} \} \setminus (J_{\omega, 1} \cup \cdots \cup J_{\omega, n_{\omega}}).\] By Proposition~\ref{prop:intersection:finite:morphism}, if $j \in J_{\omega, i}$ ($i\in\{1, \ldots, n_{\omega}\}$), then
\begin{multline*}
\Big(\rest{f_{\omega}^*(\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega})}{Y_{\omega,j}} \cdots
\rest{f_{\omega}^*(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})}{Y_{\omega,j}}\Big)_{\omega} \\
=
\deg(\rest{f_{\omega}}{Y_{\omega, j}})\Big(\rest{(\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega})}{X_{\omega,i}} \cdots
\rest{(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})}{X_{\omega,i}}\Big)_{\omega}.
\end{multline*}
Moreover, if $j \in J_{\omega, 0}$, then
\[
\Big(\rest{f_{\omega}^*(\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega})}{Y_{\omega,j}} \cdots
\rest{f_{\omega}^*(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})}{Y_{\omega,j}}\Big)_{\omega} \\
= 0.
\]
Thus, by Lemma~\ref{lem:cycle:push:forward:base:change}, one has
\begin{multline*}
\Big(f_{\omega}^*(\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega}) \cdots
f_{\omega}^*(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})\Big)_{\omega} \\
= \deg(f)\Big((\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega}) \cdots
(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})\Big)_{\omega}.
\end{multline*}
Therefore,
\begin{align*}
(f^*(\overline{L}_0) \cdots f^*(\overline{L}_l))_S & \\
& \kern-5em = \int_{\Omega} \Big(f_{\omega}^*(\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega}) \cdots
f_{\omega}^*(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})\Big)_{\omega} \nu(\mathrm{d}\omega) \\
& \kern-5em = \deg(f) \int_{\Omega} \Big((\operatorname{div}(s_0)_{\omega}, -\log |s_0|_{\varphi_\omega}) \cdots
(\operatorname{div}(s_d)_{\omega}, -\log |s_d|_{\varphi_\omega})\Big)_{\omega} \nu(\mathrm{d}\omega) \\
& \kern-5em = \deg(f) (\overline{L}_0 \cdots \overline{L}_l)_S.
\end{align*}
as required.
\end{proof}
In general, if we set $Z = a_1 Z_1 + \cdots + a_r Z_r$, then, by Claim~\ref{claim:prop:projection:formula:intersection:01},
\begin{multline*}
(f^*(\overline{L}_0) \cdots f^*(\overline{L}_l) \mid Z)_S =
\sum_{j=1}^r a_j (f^*(\overline{L}_0) \cdots f^*(\overline{L}_l) \mid Z_j)_S \\
=
\sum_{j=1}^r a_j (\overline{L}_0 \cdots \overline{L}_l \mid f_*(Z_j))_S =
(\overline{L}_0 \cdots \overline{L}_l \mid f_*(Z))_S.
\end{multline*}
\end{proof}
\section{Polarized adelic structure case}
Let $K$ be a finitely generated field over $\mathbb Q$ and $n$ be the transcendental degree of $K$ over ${\mathbb{Q}}$.
Let $(\mathscr B; \overline{\mathscr H}_1, \ldots, \overline{\mathscr H}_n)$ be a polarization of $K$ and
$S = (K, (\Omega, \mathcal A, \nu), \phi)$ be the polarized adelic structure by
$(\mathscr B; \overline{\mathscr H}_1, \ldots, \overline{\mathscr H}_n)$ (for details, see Section~\ref{sec:Polarized adelic structure}).
Let $X$ be a $d$-dimensional projective and integral scheme over $K$.
We choose a projective arithmetic variety $\mathscr X$ and a morphism $\pi : \mathscr X \to \mathscr B$
such that the generic fiber of $\mathscr X \to \mathscr B$ is $X$.
Let $L_0, \ldots, L_d$ be invertible ${\mathscr{O}}_X$-modules.
We assume that there are $C^{\infty}$-metrized invertible ${\mathscr{O}}_{\mathscr X}$-modules
$\overline{\mathscr L}_0 = (\mathscr L, h_0), \ldots, \overline{\mathscr L}_d = (\mathscr L_d, h_d)$ in the usual sense on arithmetic varieties such that
$\mathscr L_0, \ldots, \mathscr L_d$ coincides with $L_0, \ldots, L_d$ on $X$.
Note that, for each $\omega \in \Omega$, $\overline{\mathscr L}_i$ yields a smooth metric $\varphi_{i,\omega}$
of $L_{i,\omega}$, that is,
if $\omega \in \Omega_{\infty}$, then $\varphi_{i,\omega} = \rest{h_i}{\pi^{-1}(\omega)}$;
if $\omega \in \Omega \setminus \Omega_{\infty}$, then $\varphi_{i,\omega}$ is the model metric induced by
the model $(\mathscr X, \mathscr L_i)$.
We denote $\{(L_{i,\omega}, \varphi_{i,\omega})\}_{\omega \in \Omega}$ by $\overline{L}_i$.
\begin{prop}
$(\overline{L}_0\cdots \overline{L}_d)_S = (\overline{\mathscr L}_0 \cdots \overline{\mathscr L}_d \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_n))$
\end{prop}
\begin{proof}
We prove the assertion by induction on $d$.
Clearly we may assume that $\mathscr X$ is normal.
If $d=0$, that is, $\dim \mathscr X = n+1$, then it is an easy consequence of \cite[Lemma~1.12, Lemma~1.15, Proposition~5.3, Lemma~5.15 and Theorem~5.20]{MArakelov}.
We assume that $d > 0$. Let us choose a non-zero rational section $s_0$ of $\mathscr L_0$.
Let $\operatorname{div}(s_0) = a_1 \mathcal Z_1 + \cdots + a_r \mathcal Z_r$ be the decomposition as a cycle.
Then one has
\begin{multline*}
(\overline{\mathscr L}_0 \cdots \overline{\mathscr L}_d \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_n)) \\
\kern-10em = \sum_{i=1}^r a_i (\overline{\mathscr L}_1 \cdots \overline{\mathscr L}_d \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_n) \cdot (\mathcal Z_i, 0)) \\
+ \int_{\mathscr X({\mathbb{C}})} -\log |s_0|_{h_0} c_1(\overline{\mathscr L_1}) \wedge \cdots \wedge c_1(\overline{\mathscr L_d})
\wedge c_1(\pi^*\overline{\mathscr H_1}) \wedge \cdots \wedge c_1(\pi^*\overline{\mathscr H_n}).
\end{multline*}
Note that
\begin{multline*}
\int_{\mathscr X({\mathbb{C}})} -\log |s_0|_{h_0} c_1(\overline{\mathscr L_1}) \wedge \cdots \wedge c_1(\overline{\mathscr L_d})
\wedge c_1(\pi^*\overline{\mathscr H_1}) \wedge \cdots \wedge c_1(\pi^*\overline{\mathscr H_n}) \\
= \int_{\mathscr B({\mathbb{C}})} \left(\int_{\mathscr X({\mathbb{C}})/\mathscr B({\mathbb{C}})} -\log |s_0|_{h_0} c_1(\overline{\mathscr L_1}) \wedge \cdots \wedge c_1(\overline{\mathscr L_d}) \right)
c_1(\overline{\mathscr H_1}) \wedge \cdots \wedge c_1(\overline{\mathscr H_n}).
\end{multline*}
Here we consider the following claim:
\begin{enonce}{Claim}
Let $\psi : \mathscr Y \to \mathscr C$ be a surjective morphism of projective arithmetic varieties.
Let $\overline{\mathscr M}_1, \ldots, \overline{\mathscr M}_d$
(resp. $\overline{\mathscr D}_1, \ldots, \overline{\mathscr D}_n$) be metrized integrable
invertible ${\mathscr{O}}_{\mathscr Y}$-modules
(resp. ${\mathscr{O}}_{\mathscr C}$-modules) such that $d + n = \dim \mathscr Y$.
Let $\mathscr Y_{\eta}$ be the generic fiber of $\psi : \mathscr Y \to \mathscr C$.
Then
\begin{multline*}
(\overline{\mathscr M}_1 \cdots \overline{\mathscr M}_d \cdot \pi^*\overline{\mathscr D}_1 \cdots
\pi^*\overline{\mathscr D}_n) \\
= \begin{cases}
\big(\rest{\mathscr M_1}{\mathscr Y_{\eta}} \cdots \rest{\mathscr M_d}{\mathscr Y_{\eta}}\big) (\overline{\mathscr D}_1 \cdots \overline{\mathscr D}_n), & \text{if $d = \dim \mathscr Y_{\eta}$}, \\
0, & \text{if $d < \dim \mathscr Y_{\eta}$}.
\end{cases}
\end{multline*}
\end{enonce}
\begin{proof}
This is a consequence of the projection formula (cf. \cite[Theorem~5.20]{MArakelov}).
\end{proof}
By the above claim, if $\mathcal Z$ is a prime divisor on $\mathscr X$
with $\pi(\mathcal Z)\not=\mathscr B$, then
\begin{multline*}
\big(\overline{\mathscr L}_1 \cdots \overline{\mathscr L}_n \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_d) \cdot (\mathcal Z, 0)\big) \\
= \begin{cases}
\big(\rest{\mathscr L_1}{\mathcal Z_{\eta}} \cdots \rest{\mathscr L_n}{\mathcal Z_{\eta}}\big) (\overline{\mathscr H}_1 \cdots \overline{\mathscr H}_d \cdot (\pi(\mathcal Z), 0)), & \text{if $\operatorname{codim}(\pi(\mathcal Z); \mathscr B) = 1$},\\
0, & \text{if $\operatorname{codim}(\pi(\mathcal Z); \mathscr B) \geqslant 2$},
\end{cases}
\end{multline*}
where $\mathcal Z_{\eta}$ is the generic fiber of $\mathcal Z \to \pi(\mathcal Z)$. Therefore, if we set
\[
\begin{cases}
I_{h} := \{ i \in \{ 1, \ldots, r \} \mid \pi(\mathcal Z_i) = \mathscr B \}, \\
I_{\Gamma} := \{ i \in \{ 1, \ldots, r \} \mid \pi(\mathcal Z_i) = \Gamma \}
\end{cases}
\]
for $\Gamma \in \Omega \setminus \Omega_{\infty}$, and denote
\[
\sum_{i=1}^r a_i (\overline{\mathscr L}_1 \cdots \overline{\mathscr L}_d \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_n) \cdot (\mathcal Z_i, 0))
\]
by $T$, then,
by Example~\ref{Exe:chambert-loir measure} and hypothesis of induction on $d$, one has
{\allowdisplaybreaks\begin{align*}
T & = \sum_{i \in I_{h}} a_i (\overline{\mathscr L}_1 \cdots \overline{\mathscr L}_d \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_n) \cdot (\mathcal Z_i, 0)) \\
& \kern4em +
\sum_{\Gamma \in \Omega \setminus \Omega_{\infty}}
\sum_{i \in I_{\Gamma}} a_i (\overline{\mathscr L}_1 \cdots \overline{\mathscr L}_d \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_n) \cdot (\mathcal Z_i, 0)) \\
& = \sum_{i \in I_{h}} a_i \big(\rest{\overline{L}_1}{Z_i} \cdots \rest{\overline{L}_d}{Z_i}\big)_S \\
& \kern2em + \sum_{\Gamma \in \Omega \setminus \Omega_{\infty}} (\overline{\mathscr H}_1 \cdots \overline{\mathscr H}_d \cdot
(\Gamma, 0))
\int_{X_{\Gamma}^{\operatorname{an}}} -\log | s_0|_{\varphi_{0,\Gamma}} c_1(L_1, \varphi_{1,\Gamma}) \cdots c_1(L_d, \varphi_{d,\Gamma}),
\end{align*}}
where $Z_i$ is the generic fiber of $\mathcal Z_i \to \mathscr B$ for $i \in I_h$.
Thus, by \eqref{eqn:def:local:intersection},
\begin{multline*}
(\overline{\mathscr L}_0 \cdots \overline{\mathscr L}_d \cdot
\pi^*(\overline{\mathscr H}_1) \cdots \pi^*(\overline{\mathscr H}_n)) = \sum_{i \in I_{h}} a_i \big(\rest{\overline{L}_1}{Z_i} \cdots \rest{\overline{L}_d}{Z_i}\big)_S \\
\kern3em + \int_{\Omega}
\left(\int_{X_{\omega}^{\operatorname{an}}} -\log | s_0|_{\varphi_{0,\omega}} c_1(L_1, \varphi_{1,\omega}) \cdots c_1(L_d, \varphi_{d,\omega})\right) \nu(\mathrm{d}\omega) = (\overline{L}_1 \cdots \overline{L}_d)_S,
\end{multline*}
as required.
\end{proof}
\chapter{}
\section{Measurable family of embeddings}
In this appendix, we consider the following theorem, which was proved in \cite[Step~1 in Theorem~4.1.26]{CMArakelovAdelic}. We present here an alternative proof.
\begin{theo}\label{thm:measurable:family:embeddings}
Let $S=(K,(\Omega,\mathcal A,\nu),\phi)$ be an adelic curve such that $K$ is countable and $\Omega_\infty$ is not empty. There exists
a family $( \iota_{\omega} )_{\omega \in \Omega_{\infty}}$
of embeddings $K \to {\mathbb{C}}$ such that $|\raisebox{.4ex}{.}|_{\omega} = |\iota_{\omega}(\raisebox{.4ex}{.})|$ for all
$\omega \in \Omega_{\infty}$ and that the map $(\omega \in \Omega_{\infty}) \mapsto
\iota_{\omega}(a)$ is measurable for each $a \in K$.
\end{theo}
Let
${\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$ be the set of all sequences $(x_n)_{n=0}^{\infty}$
consisting of complex numbers. For $x = (x_n)_{n=0}^{\infty} \in {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$,
the $n$-th entry of $x_n$ of $x$ is often denoted by $x(n)$.
One can easily check the following proposition (see \cite[\S3.5]{Ovch} for (1), and (2) is straightforward).
\begin{prop}\label{prop:measurability:C:N}
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item
If we define $d : {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}} \times {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}} \to {\mathbb{R}}_{\geqslant 0}$ to be
\[
d(x, y) = \sum_{n=0}^{\infty} 2^{-n}\frac{|x(n) - y(n)|}{1 + |x(n) - y(n)|}
\quad (x, y \in {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}),
\]
then $d$ yields a distance function on ${\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$. Moreover,
the topology determined by $d$ coincides with the product topology, and
$({\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}, d)$ is a complete and second-countable space.
\item
Let $\mathcal{B}_{{\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}}$ be the Borel $\sigma$-algebra by the product topology
on ${\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$ and $\mathcal{B}_{{\mathbb{C}}}$ be the Borel $\sigma$-algebra by
the standard topology on ${\mathbb{C}}$.
Let $(\Omega, \mathcal{A})$ be a measurable space and
$( f_n )_{n=0}^{\infty}$ be a family of maps $\Omega \to {\mathbb{C}}$. Then
$f : (\Omega, \mathcal{A}) \to ({\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}, \mathcal{B}_{{\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}})$ given by $f(\omega) = (f_n(\omega))_{n=0}^{\infty}$ is measurable
if and only if $f_n : (\Omega, \mathcal{A}) \to ({\mathbb{C}}, \mathcal{B}_{{\mathbb{C}}})$ is measurable for all $n \in {\mathbb{Z}_{\geqslant 1}}$.
\end{enumerate}
\end{prop}
\begin{lemm}\label{lemma:measurable:family:embeddings}
Let $S = (K, (\Omega, \mathcal{A}, \nu), \phi)$ be an adelic curve such that $\Omega_\infty$ is not empty. Suppose that $-1\in K$ admits a square root $\xi$ in $K$. Then there is a family $( \iota_{\omega} )_{\omega \in \Omega_{\infty}}$ of embeddings $K \to {\mathbb{C}}$ which satisfy the following conditions:
\begin{enumerate}[label=\rm(\arabic*)]
\item for any $\omega\in\Omega_\infty$,
$\iota_{\omega}(\xi) = \sqrt{-1}$,
\item for any $\omega \in \Omega_{\infty}$, $|\raisebox{.4ex}{.}|_{\omega} = |\iota_{\omega}(\raisebox{.4ex}{.})|$,
\item for any $a \in K$, the function $(\omega \in \Omega_{\infty}) \mapsto
\iota_{\omega}(a)$ is measurable.
\end{enumerate}
\end{lemm}
\begin{proof}
Fix a family $( \sigma_{\omega} )_{\omega \in \Omega_{\infty}}$ of embeddings $K \to {\mathbb{C}}$ such that $|\raisebox{.4ex}{.}|_{\omega} = |\sigma_{\omega}(\raisebox{.4ex}{.})|$ for all $\omega \in \Omega_{\infty}$.
Note that $\sigma_{\omega}(\xi) = \pm \sqrt{-1}$ because
\[(\sigma_{\omega}(\xi))^2 = \sigma_{\omega}(\xi^2) = \sigma_{\omega}(-1) = -1.\]
We define a family $( \iota_{\omega} )_{\omega \in \Omega_{\infty}}$ of embeddings
by
\[
\iota_{\omega} = \begin{cases}
\sigma_{\omega} & \text{if $\sigma_{\omega}(\xi) = \sqrt{-1}$}, \\[1ex]
\overline{\sigma_{\omega}} & \text{if $\sigma_{\omega}(\xi) = -\sqrt{-1}$},
\end{cases}
\]
where $\overline{\phantom{x}}$ means the complex conjugation.
Then $\iota_{\omega}(\xi) = \sqrt{-1}$ for all $\omega \in \Omega_{\infty}$.
Thus one can see
\[
\iota_{\omega}(a) = \left(|a + (1/2)|^2_{\omega} - |a|^2_{\omega} - |1/2|_{\omega}^2\right)
+ \sqrt{-1} \left(|a + (\xi/2)|^2_{\omega} - |a|^2_{\omega} - |\xi/2|_{\omega}^2\right),
\]
as required.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:measurable:family:embeddings}]
By Lemma~\ref{lemma:measurable:family:embeddings}, we may assume that $-1\in K$ does not have any square root in $K$.
We set \[K = (a_n)_{n=1}^{\infty}, \quad L := K(\sqrt{-1})\quad \text{and}\quad
(L, (\Omega_L, \mathcal{A}_L, \nu_L), \phi_L) = S \otimes_K L.\]
Then, by Lemma~\ref{lemma:measurable:family:embeddings} again, there is a family $( \iota_{\chi} )_{\chi \in \Omega_{L,\infty}}$
of embeddings $L \to {\mathbb{C}}$ such that $|\raisebox{.4ex}{.}|_{\chi} = |\iota_{\chi}(\raisebox{.4ex}{.})|$ for all
$\chi \in \Omega_{L,\infty}$ and $(\chi \in \Omega_{L,\infty}) \mapsto \iota_{\chi}(b)$ is
measurable for all $b \in L$. Thus, if we define $h : \Omega_{L,\infty} \to {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$
by $h(\chi) = (\iota_{\chi}(a_n))_{n=1}^{\infty}$, then $h$ is also measurable by (2) in Proposition~\ref{prop:measurability:C:N}, so that,
for an open set $U$ in ${\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$, $h^{-1}(U)$ is a measurable subset in $\Omega_{L,\infty}$.
Let $\pi_{L/K} : \Omega_L \to \Omega$ be the canonical map. We consider a map
$F : \Omega_{\infty} \to \mathcal{P}({\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}})$ given by
$F(\omega) = h(\pi_{L/K}^{-1}(\omega))$.
Then one can see that
\begin{align*}
\{ \omega \in \Omega_{\infty} \mid F(\omega) \cap U \not= \emptyset \} & =
\{ \omega \in \Omega_{\infty} \mid \pi^{-1}_{L/K}(\omega) \cap h^{-1}(U) \not= \emptyset \} \\
& = \{ \omega \in \Omega_{\infty} \mid
I_{L/K}(1\hspace{-0.25em}\mathrm{l}_{h^{-1}(U)})(\omega) > 0 \}.
\end{align*}
By \cite[Theorem~3.3.4]{CMArakelovAdelic}, $I_{L/K}(1\hspace{-0.25em}\mathrm{l}_{h^{-1}(U)})$ is measurable, so that
\[\{ \omega \in \Omega_{\infty} \mid F(\omega) \cap U \not= \emptyset \} \in \mathcal A.\]
Thus, by Kuratowski and Ryll-Nardzewski measurable selection theorem (cf. \cite[Theorem~A.2.1]{CMArakelovAdelic}) together with (1) in Proposition~\ref{prop:measurability:C:N}, there is a measurable map
$f : \Omega_{\infty} \to {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$ such that $f(\omega) \in F(\omega)$ for all $\omega \in \Omega_{\infty}$.
For each $\omega \in \Omega_{\infty}$, we choose $\chi_{\omega} \in \pi_{L/K}^{-1}(\omega)$ such that $f(\omega) = h(\chi_{\omega})$.
If we set $\iota_{\omega} = \rest{\iota_{\chi_{\omega}}}{K}$, then $\iota_{\omega}$ yields an embedding $K \to {\mathbb{C}}$
such that $|a|_{\omega} = |\iota_{\omega}(a)|$ for all $a \in K$.
Moreover, for all $n\in {\mathbb{Z}_{\geqslant 1}}$, $(\omega \in \Omega_{\infty}) \mapsto \iota_{\omega}(a_n)$
is measurable by (2) in Proposition~\ref{prop:measurability:C:N}. Thus the assertion follows.
\end{proof}
\if01
\begin{theo}\label{thm:measurable:family:embeddings}
If $\#(K) = \aleph_0$, then there is a family $\{ \iota_{\omega} \}_{\omega \in \Omega_{\infty}}$
of embeddings $K \to {\mathbb{C}}$ such that $|\raisebox{.4ex}{.}|_{\omega} = |\iota_{\omega}(\raisebox{.4ex}{.})|$ for all
$\omega \in \Omega_{\infty}$ and a map $(\omega \in \Omega_{\infty}) \mapsto
\iota_{\omega}(a)$ is measurable for each $a \in K$.
\end{theo}
\subsubsection{Topology of product space}
Let $\{ (X_n, d_n) \}_{n=1}^{\infty}$ be a family of metric spaces.
Let \[{\mathscr{X}} := \prod_{n=1}^{\infty} X_n\] and
$\mathscr{T}$ be the product topology of ${\mathscr{X}}$, that is,
the coarsest topology such that the projection $p_n : {\mathscr{X}} \to X_n$ to the $n$-th factor is continuous for all $n \in {\mathbb{Z}_{\geqslant 1}}$, where ${\mathbb{Z}_{\geqslant 1}}$ is the set of all positive integers.
For $x = (x_n)_{n=1}^{\infty} \in {\mathscr{X}}$, the $n$-th entry $x_n$ of $x$ is often denoted by $x(n)$.
Let ${\mathscr{O}}_n$ be the set of all open sets in $X_n$ and
\begin{multline*}
\left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)_* := \\
\left\{ (U_n)_{n=1}^{\infty} \in \prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n \mid \text{$U_n = X_n$ except finitely many $n \in {\mathbb{Z}_{\geqslant 1}}$} \right\}.
\end{multline*}
For $(U_n)_{n=1}^{\infty} \in \left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)_*$,
we set
\[
{\mathscr{U}}((U_n)_{n=1}^{\infty}) := \{ x \in X \mid \text{$x(n) \in U_n$ for all $n \in {\mathbb{Z}_{\geqslant 1}}$} \}
= \bigcap_{n=1}^{\infty} p_n^{-1}(U_n).
\]
Note that $\{ {\mathscr{U}}((U_n)_{n=1}^{\infty}) \}_{(U_n)_{n=1}^{\infty} \in \left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)_*}$ forms an open basis of $\mathscr{T}$.
We define $d : {\mathscr{X}} \times {\mathscr{X}} \to {\mathbb{R}}_{\geqslant 0}$ to be
\[
d(x, y) := \sum_{n=1}^{\infty} 2^{-n} \frac{d_n(x(n),y(n))}{1 + d_n(x(n),y(n))}\quad(x, y \in {\mathscr{X}}).
\]
Note that the function $t/(1+t)$ on ${\mathbb{R}}_{\geqslant 0}$
is strictly increasing and subadditive, so that one can see that
the above $d$ yields a distance on ${\mathscr{X}}$.
For $r \in {\mathbb{R}}_{> 0}$, $x \in {\mathscr{X}}$ and $x_n \in X_n$, we set
\[
{\mathscr{U}}(x;r) := \{ y \in X \mid d(x, y) < r\}\quad\text{and}\quad
U_n(x_n; r) := \{ y_n \in X_n \mid d_n(x_n, y_n) < r \}.
\]
\begin{prop}\label{prop:topology:product}
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item
The topology determined by the distance $d$ coincides with the product topology $\mathscr{T}$.
\item
If $(X_n, d_n)$ is complete for all $n \in {\mathbb{Z}_{\geqslant 1}}$, then
$({\mathscr{X}}, d)$ is also complete.
\item We assume that $(X_n, d_n)$ is a second-countable space for all $n \in {\mathbb{Z}_{\geqslant 1}}$, that is,
there is ${\mathscr{O}}'_n \subseteq {\mathscr{O}}_n$ such that
${\mathscr{O}}'_n$ is a basis and $\#({\mathscr{O}}'_n) = \aleph_0$.
If we set
\[
\kern2.2em \left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)'_* = \left\{
(U_n)_{n=1}^{\infty} \in \left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)_* \mid \text{$U_n \in {\mathscr{O}}'_n$ or $U_n = X_n$} \right\},
\]
then $\#(\left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)'_*) = \aleph_0$ and
$\{ {\mathscr{U}}((U_n)_{n=1}^{\infty}) \}_{(U_n)_{n=1}^{\infty} \in \left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)'_*}$ forms a basis of $\mathscr{T}$.
In particular, ${\mathscr{X}}$ is also a second-countable space.
\end{enumerate}
\end{prop}
\begin{proof}
(1) Let $\mathscr{T}'$ be the topology determined by $d$.
It is sufficient to see the following (i) and (ii):
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\roman{enumi})}}
\item The projection $p_n : {\mathscr{X}} \to X_n$ is continuous with respect to $\mathscr{T}'$ for
all $n \in {\mathbb{Z}_{\geqslant 1}}$.
In particular, $\mathscr{T} \subseteq \mathscr{T}'$.
\item For ${\mathscr{U}}(x;r)$ and $y \in {\mathscr{U}}(x;r)$,
there is $(U_n)_{n=1}^{\infty} \in \left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)_*$
such that $y \in {\mathscr{U}}((U_n)_{n=1}^{\infty}) \subseteq {\mathscr{U}}(x;r)$. In particular,
$\mathscr{T}' \subseteq \mathscr{T}$.
\end{enumerate}
(i) follows from
\begin{equation}\label{eqn:prop:topology:product:01}
\frac{d_n(x(n), y(n))}{1 + d_n(x(n), y(n))} \leqslant 2^n d(x, y)
\quad (\forall x, y \in X).
\end{equation}
(ii) As ${\mathscr{U}}(y; r - d(x,y)) \subseteq {\mathscr{U}}(x;r)$, we may assume that $y=x$.
We choose $N \in {\mathbb{Z}_{\geqslant 1}}$ and $r' \in {\mathbb{R}}_{>0}$ such that $\sum_{n=N+1}^{\infty} 2^{-n} < r/2$ and $\sum_{n=1}^{N} 2^{-n} r'/(1 + r') < r/2$.
If we set
\[
U_n := \begin{cases} U_n(x(n);r') & \text{if $n \in \{ 1,\ldots,N\}$},\\
X_n & \text{if $n > N$},
\end{cases}
\]
then $x \in {\mathscr{U}}((U_n)_{n=1}^{\infty})\subseteq {\mathscr{U}}(x;r)$. Indeed, if $y \in {\mathscr{U}}((U_n)_{n=1}^{\infty})$, then
\[
d(y, x) = \sum_{n=1}^{\infty} 2^{-n}\frac{d_n(y(n),x(n))}{1+d_n(y(n),x(n))} <
\sum_{n=1}^N 2^{-n} \frac{r'}{1+ r'} + \sum_{n=N+1}^{\infty} 2^{-n} < r.
\]
\medskip
(2) Let $\{ x_m \}_{m=1}^{\infty}$ be a Cauchy sequence in $({\mathscr{X}}, d)$.
Then, by \eqref{eqn:prop:topology:product:01},
$\{ x_m(n) \}_{m=1}^{\infty}$ is also a Cauchy sequence in $X_n$,
so that there is $x_n \in X_n$ such that \[\lim_{m\to\infty} d_n(x_m(n), x_n) = 0.\]
Let $x \in {\mathscr{X}}$ such that $x(n) = x_n$ for all $n \in {\mathbb{Z}_{\geqslant 1}}$. For any $\varepsilon > 0$,
by the above (ii), there is $(U_n)_{n=1}^{\infty} \in \left(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n\right)_*$
such that $x \in {\mathscr{U}}((U_n)_{n=1}^{\infty}) \subseteq {\mathscr{U}}(x;\varepsilon)$.
If we set $I = \{ n \in {\mathbb{Z}_{\geqslant 1}} \mid U_n \not= X_n \}$,
then there is $m_0 > 0$ such that
$x_m(n) \in U_n$ for all $n \in I$ and $m \geqslant m_0$.
Thus $d(x_m, x) < \varepsilon$,
so that $\lim_{m\to\infty} d(x_m, x) = 0$, as required.
\medskip
(3) is obvious.
\end{proof}
Let $\mathcal{B}_{X_n}$ and $\mathcal{B}_{{\mathscr{X}}}$ be the Borel $\sigma$-algebras
of $X_n$ and ${\mathscr{X}}$, respectively.
\begin{prop}\label{prop:measurability:product}
We assume that $(X_n, d_n)$ is a second-countable space for all $n \in {\mathbb{Z}_{\geqslant 1}}$.
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item
$\mathcal{B}_{{\mathscr{X}}}$ is the smallest $\sigma$-algebra such that $p_n : ({\mathscr{X}}, \mathcal{B}_{{\mathscr{X}}}) \to (X_n, \mathcal{B}_{X_n})$
is measurable for all $n \in {\mathbb{Z}_{\geqslant 1}}$.
\item
Let $(\Omega, \mathcal{A})$ be a measurable space and
$f_n : \Omega \to X_n$ be a map for each $n \in {\mathbb{Z}_{\geqslant 1}}$. Then
$f : (\Omega, \mathcal{A}) \to ({\mathscr{X}}, \mathcal{B}_{{\mathscr{X}}})$ given by $f(\omega) =
(f_n(\omega))_{n=1}^{\infty}$ is measurable
if and only if $f_n : (\Omega, \mathcal{A}) \to (X_n, \mathcal{B}_{X_n})$ is measurable for all $n \in {\mathbb{Z}_{\geqslant 1}}$.
\end{enumerate}
\end{prop}
\begin{proof}
(1)
Let $\mathcal{B}'_{{\mathscr{X}}}$ be the smallest $\sigma$-algebra such that $p_n : ({\mathscr{X}}, \mathcal{B}'_{{\mathscr{X}}}) \to (X_n, \mathcal{B}_{X_n})$
is measurable for all $n \in {\mathbb{Z}_{\geqslant 1}}$. Obviously $\mathcal{B}'_{{\mathscr{X}}} \subseteq \mathcal{B}_{{\mathscr{X}}}$ because
$p_n : {\mathscr{X}} \to X_n$
is continuous for all $n \in {\mathbb{Z}_{\geqslant 1}}$. For its converse, we need to show that ${\mathscr{U}} \in \mathcal{B}'_{{\mathscr{X}}}$ for all open sets ${\mathscr{U}}$ of ${\mathscr{X}}$.
By (3) in Proposition~\ref{prop:topology:product}, we may assume that ${\mathscr{U}} = {\mathscr{U}}((U_n)_{n=1}^{\infty})$ for some $(U_n)_{n=1}^{\infty} \in
(\prod\nolimits_{n=1}^{\infty} {\mathscr{O}}_n)'_*$.
Then \[ {\mathscr{U}} = \bigcap_{n=1}^{\infty} p_n^{-1}(U_n) \in \mathcal{B}'_{{\mathscr{X}}},\] as required.
\medskip
(2) is a consequence of (1).
\end{proof}
As a corollary of Proposition~\ref{prop:topology:product} and Proposition~\ref{prop:measurability:product}. one has the following.
\begin{Corollary}\label{cor:measurability:C:N}
\begin{enumerate}
\renewcommand{\labelenumi}{\textup{(\arabic{enumi})}}
\item
If we define $d : {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}} \times {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}} \to {\mathbb{R}}_{\geqslant 0}$ to be
\[
d(x, y) = \sum_{n=1}^{\infty} 2^{-n}\frac{|x(n) - y(n)|}{1 + |x(n) - y(n)|}
\quad (x, y \in {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}),
\]
then the topology determined by $d$ coincides with the product topology, and
$({\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}, d)$ is a complete and second-countable space.
In particular, $({\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}, d)$ is separable.
\item
Let $\mathcal{B}_{{\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}}$ be the Borel $\sigma$-algebra by the product topology
on ${\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$ and $\mathcal{B}_{{\mathbb{C}}}$ be the Borel $\sigma$-algebra by
the standard topology on ${\mathbb{C}}$.
Let $(\Omega, \mathcal{A})$ be a measurable space and
$\{ f_n \}_{n \in {\mathbb{Z}}_{\geqslant 1}}$ be a family of maps $\Omega \to {\mathbb{C}}$. Then
$f : (\Omega, \mathcal{A}) \to ({\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}, \mathcal{B}_{{\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}})$ given by $f(\omega) = (f_n(\omega))_{n=1}^{\infty}$ is measurable
if and only if $f_n : (\Omega, \mathcal{A}) \to ({\mathbb{C}}, \mathcal{B}_{{\mathbb{C}}})$ is measurable for all $n \in {\mathbb{Z}_{\geqslant 1}}$.
\end{enumerate}
\end{Corollary}
\subsubsection{Proof of Theorem~\ref{thm:measurable:family:embeddings}}
Let us begin with the following lemma.
\beginlemm\label{lemma:measurable:family:embeddings}
Let $S = (K, (\Omega, \mathcal{A}, \nu), \phi)$ be an adelic structure of a field $K$.
If $\sqrt{-1} \in K$, then there is a family $\{ \iota_{\omega} \}_{\omega \in \Omega_{\infty}}$ of embeddings $K \to {\mathbb{C}}$ such that
$\iota_{\omega}(\sqrt{-1}) = \sqrt{-1}$ and $|\raisebox{.4ex}{.}|_{\omega} = |\iota_{\omega}(\raisebox{.4ex}{.})|$
for all $\omega \in \Omega_{\infty}$, and that
a map $(\omega \in \Omega_{\infty}) \mapsto
\iota_{\omega}(a)$ is measurable for all $a \in K$.
\endlemm
\begin{proof}
Fix a family $\{ \sigma_{\omega} \}_{\omega \in \Omega_{\infty}}$ of embeddings $K \to {\mathbb{C}}$ such that $|\raisebox{.4ex}{.}|_{\omega} = |\sigma_{\omega}(\raisebox{.4ex}{.})|$ for all $\omega \in \Omega_{\infty}$.
Note that $\sigma_{\omega}(\sqrt{-1}) = \pm \sqrt{-1}$ because
$(\sigma_{\omega}(\sqrt{-1}))^2 = \sigma_{\omega}(\sqrt{-1}^2) = \sigma_{\omega}(-1) = -1$.
We define a family $\{ \iota_{\omega} \}_{\omega \in \Omega_{\infty}}$ of embeddings
by
\[
\iota_{\omega} = \begin{cases}
\sigma_{\omega} & \text{if $\sigma_{\omega}(\sqrt{-1}) = \sqrt{-1}$}, \\[1ex]
\overline{\sigma_{\omega}} & \text{if $\sigma_{\omega}(\sqrt{-1}) = -\sqrt{-1}$},
\end{cases}
\]
where $\overline{\phantom{x}}$ means the complex conjugation.
Then $\iota_{\omega}(\sqrt{-1}) = \sqrt{-1}$ for all $\omega \in \Omega_{\infty}$.
Thus one can see
\begin{multline*}
\iota_{\omega}(a) = \left(|a + (1/2)|^2_{\omega} - |a|^2_{\omega} - |1/2|_{\omega}^2\right) \\
+ \sqrt{-1} \left(|a + (\sqrt{-1}/2)|^2_{\omega} - |a|^2_{\omega} - |\sqrt{-1}/2|_{\omega}^2\right),
\end{multline*}
as required.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:measurable:family:embeddings}]
By Lemma~\ref{lemma:measurable:family:embeddings}, we may assume that $\sqrt{-1} \not\in K$.
We set \[K = \{ a_n \}_{n=1}^{\infty}, \quad L := K(\sqrt{-1})\quad \text{and}\quad
(L, (\Omega_L, \mathcal{A}_L, \nu_L), \phi_L) = S \otimes_K L.\]
Then, by Lemma~\ref{lemma:measurable:family:embeddings} again, there is a family $\{ \iota_{\chi} \}_{\chi \in \Omega_{L,\infty}}$
of embeddings $L \to {\mathbb{C}}$ such that $|\raisebox{.4ex}{.}|_{\chi} = |\iota_{\chi}(\raisebox{.4ex}{.})|$ for all
$\chi \in \Omega_{L,\infty}$ and $(\chi \in \Omega_{L,\infty}) \mapsto \iota_{\chi}(b)$ is
measurable for all $b \in L$. Thus, if we define $h : \Omega_{L,\infty} \to {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$
by $h(\chi) = (\iota_{\chi}(a_n))_{n=1}^{\infty}$, then $h$ is also measurable by (2) in Corollary~\ref{cor:measurability:C:N}, so that,
for an open set $U$ in ${\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$, $h^{-1}(U)$ is a measurable subset in $\Omega_{L,\infty}$.
Let $\pi_{L/K} : \Omega_L \to \Omega$ be the canonical map. We consider a map
$F : \Omega_{\infty} \to \mathcal{P}({\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}})$ given by
\[
F(\omega) = h(\pi_{L/K}^{-1}(\omega)).
\]
Then one can see that
\begin{align*}
\{ \omega \in \Omega_{\infty} \mid F(\omega) \cap U \not= \emptyset \} & =
\{ \omega \in \Omega_{\infty} \mid \pi^{-1}_{L/K}(\omega) \cap h^{-1}(U) \not= \emptyset \} \\
& = \{ \omega \in \Omega_{\infty} \mid
I_{L/K}(1\hspace{-0.25em}\mathrm{l}_{h^{-1}(U)})(\omega) > 0 \}.
\end{align*}
By \cite[Theorem~3.3.4]{CMArakelovAdelic}, $I_{L/K}(1\hspace{-0.25em}\mathrm{l}_{h^{-1}(U)})$ is measurable, so that
\[\{ \omega \in \Omega_{\infty} \mid F(\omega) \cap U \not= \emptyset \} \in \mathcal A.\]
Thus, by Kuratowski and Ryll-Nardzewski measurable selection theorem (cf. \cite[Theorem~A.2.1]{CMArakelovAdelic}) together with (1) in Corollary~\ref{cor:measurability:C:N}, there is a measurable map
$f : \Omega_{\infty} \to {\mathbb{C}}^{{\mathbb{Z}_{\geqslant 1}}}$ such that $f(\omega) \in F(\omega)$ for all $\omega \in \Omega_{\infty}$.
For each $\omega \in \Omega_{\infty}$, we choose $\chi_{\omega} \in \pi_{L/K}^{-1}(\omega)$ such that $f(\omega) = h(\chi_{\omega})$.
If we set $\iota_{\omega} = \rest{\iota_{\chi_{\omega}}}{K}$, then $\iota_{\omega}$ yields an embedding $K \to {\mathbb{C}}$
such that $|a|_{\omega} = |\iota_{\omega}(a)|$ for all $a \in K$.
Moreover, for all $n\in {\mathbb{Z}_{\geqslant 1}}$, $(\omega \in \Omega_{\infty}) \mapsto \iota_{\omega}(a_n)$
is measurable by (2) in Corollary~\ref{cor:measurability:C:N}. Thus the assertion follows.
\end{proof}
\fi
|
\section{Introduction}
\label{sec:intro}
An atomic nucleus is a compact system composed of nucleon fermions
-- protons and neutrons -- under the strong nuclear force.
The Fermi statistic can be used to describe the single particle motion
inside the nucleus, and the nuclear Fermi momentum is used to describe
the independent nucleon energy level near Fermi surface.
In a nucleus, the nucleons move with an average momentum
which is closely related to the nuclear Fermi momentum.
The motions of the nucleons affect the scattering processes
under the high momentum transfer, such as the quasielastic scattering
\cite{Moniz:1969sr,Moniz:1971mt,Whitney:1974hr,Donnelly:1975ze,Benhar:2006wy},
the transverse momentum spectrum \cite{Fredriksson:1975tp,Fredriksson:1975cm,Yong:2015gma}
and the deep inelastic scattering \cite{Saito:1985ct,Arneodo:1992wf,Geesaman:1995yd,Malace:2014uea}.
What is more? The better understanding of the nuclear Fermi momentum
helps us better classifying the high-momentum tail of the nucleon
momentum distribution and the short-range correlations among the nucleons
\cite{Arrington:2011xs,Hen:2016kwk,Hen:2014nza,Schmidt:2020kcl,Wang:2020uhj},
since the quasielastic scattering is sensitive to both
the single nucleon motion and the nucleon-nucleon correlations.
For a better understanding of the high energy scattering experiments
on the nuclear targets, the nuclear Fermi motion effect is one of the corrections
that needs to be made.
The nuclear Fermi momentum can be deduced via the quasielastic scattering
of a high energy electron on the nuclear target \cite{Moniz:1969sr,Moniz:1971mt,Whitney:1974hr},
due to the weakly electromagnetic interaction which does not disturb much the structure of the target.
The quasielastic scattering can simply be viewed as the electron scattering
from an individual and moving nucleon in the Fermi sea,
with the recoiling nucleon going outside of the Fermi sphere
(escaping from the nucleus for most of the cases).
A quasielastic peak dominates in the spectrum of the electron energy loss
below the continuum region caused by deep inelastic scattering.
The width of the quasielastic peak is broadened by the Fermi motion
of the nucleon, and the position of the peak is influenced by
the separation energy between the struck nucleon and remanent nucleus \cite{Moniz:1969sr,Moniz:1971mt,Whitney:1974hr}.
Therefore measuring the width of the quasielastic peak is the key
of extracting the nuclear Fermi momentum.
Usually the differential cross section of quasielastic scattering
is described with the spectral function $S(\vec{k},E)$, where $\vec{k}$ and $E$
are the initial momentum and energy of the struck nucleon respectively \cite{Benhar:2006wy}.
In plane-wave-impulse approximation, the famous $y$-scaling
for quasielastic scattering implies the nucleonic degrees of freedom
and the initial nucleon momentum distribution \cite{Benhar:2006wy,West:1974ua,Sick:1980ey,Day:1987az,Arrington:1995hs}.
Extraction of the nuclear Fermi momentum gives us some guidance
in understanding the corresponding nuclear effect in high energy experiments.
To better understand the complex underlying neutrino-nucleus interaction,
physicists have measured the quasielastic scattering and resonance production
on the Titanium and Argon targets for the first time, at Hall A, JLab \cite{JeffersonLabHallA:2018zyx,Dai:2018gch}.
Recent high precision data of quasielastic scattering on the deuteron and some heavy nuclei
are provided by CLAS collaboration with the facilities at JLab \cite{Schmookler:2019nvf}.
The experiment was performed with 5 GeV electron beam hitting a dual target system,
with the scattered electrons measured with CLAS spectrometer.
Hence it is interesting to look at the CLAS data for quasielastic scattering peak
at $x_{\rm B}=\frac{Q^2}{2m\omega}\sim 1$,
where $Q^2$ and $\omega$ are the minus four-momentum square
and the energy of the exchanged virtual photon respectively.
The $e$-$A$ quasielastic cross sections are carefully given after
acceptance corrections, radiative corrections, Coulomb corrections,
and bin centering corrections. In the recent publication \cite{Schmookler:2019nvf},
the cross-section ratios are presented to quantify the nuclear effects,
which is also beneficial in reducing the systematic errors.
For the cross-section ratio of a heavy nucleus to deuteron,
it is interesting to note that there is a dip at $x_{\rm B}\sim 1$.
In this work, we demonstrate that the dip can be explained with the quasielastic peaks
of the heavy nucleus and the deuteron, which have quite different peak widths.
We use a simple Fermi smearing model in explaining the experimental data.
Both gaussian distribution and Cauchy distribution are applied
to describe the nuclear quasielastic scattering peak.
We extract the nuclear Fermi momenta of some unmeasured nuclei, especially the lightest
nucleon composite -- the deuteron.
In this work, the Fermi momentum of light nucleus is defined as the width of the peak
of nuclear cross section corresponding to quasi-elastic single-nucleon knock out.
This definition is regarded as an extension of the traditional picture
of Fermi momentum of heavy nucleus.
The method is discussed in Sec. \ref{sec:method}.
The determination of the deuteron Fermi momentum is presented in Sec. \ref{sec:deuteron}.
The determinations of the nuclear Fermi momenta
of aluminium and iron are presented in Sec. \ref{sec:Al-and-Fe}.
Some discussions and a concise summary is given in Sec. \ref{sec:summary}.
\section{Gaussian distribution and Cauchy distribution for nuclear quasielastic scattering peak}
\label{sec:method}
The quasielastic scattering on a nucleus is
viewed as the elastic scattering between the electron
probe and one nucleon inside the nucleus.
Compared to the $e$-$p$ elastic scattering, the energy transfer of quasielastic scattering
at a certain angle is smeared by the nucleon Fermi motion.
To depict the smearing from the randomly moving nucleon,
we apply the naive gaussian distribution and the long-tail Cauchy distribution for simplicity.
Cauchy distribution turns out to be a fine model description
in the narrow range around the quasielastic scattering peak.
As shown in the references \cite{Benhar:2015wva,Benhar:2015ula,Arrington:1998ps},
the inelastic scattering and nucleon-nucleon correlation
also contribute to the cross section around the quasi-elastic scattering peak,
but just a few percent in case of low-energy electron beams.
The differential cross section of nuclear quasielastic scattering over the energy transfer $\omega$
modeled with the gaussian distribution is given by,
\begin{equation}\label{eq:QECrosssection_gaussian}
\frac{1}{\mathrm{A}}\frac{d \sigma_\mathrm{A}(\omega)}{d\omega} =
\sigma_{\mathrm{eN}}\frac{1}{\sqrt{2 \pi}\zeta k_{\mathrm{F},\mathrm{A}}}
e^{-\frac{\left(\omega-\omega_{0}\right)^{2}}{2 \zeta^2 k_{\mathrm{F},\mathrm{A}}^{2}}},
\end{equation}
where $A$ is the mass number of the nucleus,
$\omega=E-E'$ is the energy transfer from the electron to the nucleon,
and $\sigma_{\mathrm{eN}}$ represents the total cross section of $e$-$N$ elastic scattering.
The main assumption of our model is that the width of the differential cross section peak
is proportional to the Fermi momentum $k_{\mathrm{F},\mathrm{A}}$
of the nucleon inside nucleus A \cite{Moniz:1969sr,Moniz:1971mt,Whitney:1974hr}.
Therefore, in Eq. (\ref{eq:QECrosssection_gaussian}), we introduce a pure coefficient $\zeta$,
which is a universal constant for all the nuclear targets in the same experiment.
In experiments, we usually present the cross section as a function of the Bjorken scaling variable $x_\mathrm{B}$.
In the nuclear target rest frame, the energy exchange $\omega$ is connected to $x_\mathrm{B}$,
via the definition of the Bjorken scaling variable.
$\omega$ is related to $x_\mathrm{B}$ as,
\begin{equation}\label{eq:omega_and-xB}
\omega=\frac{Q^{2}}{2 m x_{\mathrm{B}}},
\end{equation}
where $Q^2$ is the minus of the square of four-momentum transfer and $m$ is the nucleon mass.
Note that we see the quasielastic scattering peak of
the differential cross section with a slight shift away from $x_\mathrm{B}=1$,
and this shift of the peak comes from the average nucleon separation energy \cite{Moniz:1969sr,Moniz:1971mt,Whitney:1974hr}.
Again the gaussian smearing model does not work far from $x_\mathrm{B}=1$.
Using Eq. (\ref{eq:omega_and-xB}) and under the gaussian smearing model,
we can rewrite the per-nucleon quasielastic cross section
as a function of $x_\mathrm{B}$, which is written as,
\begin{equation}\label{eq:QECrosssection_dxB_gaussian}
\begin{split}
&\frac{1}{\mathrm{A}} \frac{d \sigma_\mathrm{A}(x_\mathrm{B})}{d x_\mathrm{B}} =
\left|\frac{1}{\mathrm{A}} \frac{d \sigma_\mathrm{A}(\omega)}{d \omega} \frac{d \omega}{d x_\mathrm{B}} \right| \\
&=\frac{\sigma_{\mathrm{e}\mathrm{N}}}{\sqrt{2 \pi}\zeta k_{\mathrm{F},\mathrm{A}}}
\exp\left[-\frac{(\frac{1}{x_\mathrm{B}}-X_0^\mathrm{A})^2}{2W_\mathrm{A}^2}\right]\frac{Q^2}{2 m x_\mathrm{B}^2}, \\
\end{split}
\end{equation}
where we have the parameter $X_0^\mathrm{A}=\frac{2 m \omega_{0}}{Q^2}$ for the central value
and the parameter $W_\mathrm{A}=\frac{2 m \zeta k_{\mathrm{F},\mathrm{A}}}{Q^2}$ for the width.
With $\mathrm{A}=2$ in Eq. (\ref{eq:QECrosssection_dxB_gaussian}), we get the quasielastic cross section for deuteron.
Dividing the quasielastic scattering cross section of a heavy nucleus by that of deuteron,
we then get a master formula for the cross-section ratio near $x_B\sim 1$,
\begin{equation}\label{eq:CrosssectionRatio_gaussian}
\left| \frac{d \sigma_\mathrm{A}/\mathrm{A}}{d \sigma_\mathrm{D}/2}\right| =
\frac{W_\mathrm{D}}{W_\mathrm{A}}\exp\left[\frac{(\frac{1}{x_\mathrm{B}}-X_0^\mathrm{D})^2}{2 W_\mathrm{D}^2}
-\frac{(\frac{1}{x_\mathrm{B}}-X_0^\mathrm{A})^2}{2 W_\mathrm{A}^2}\right].
\end{equation}
Here the subscript $\mathrm{D}$ denotes the deuterium.
Note that $W_\mathrm{A}=\frac{2 m \zeta k_{\mathrm{F},\mathrm{A}}}{Q^2}$ is linearly proportional
to the nuclear Fermi momentum of the nucleus A.
Once the width ratio $W_\mathrm{A_1}/W_\mathrm{A_2}$ is extracted from the data of the cross-section ratio,
the Fermi momentum ratio of the two nuclei is also obtained.
In Eq. (\ref{eq:CrosssectionRatio_gaussian}), the free parameter $X_0$ should be 1 for elastic scattering in theory.
Due to the mass deficit of the nucleon inside the nucleus,
the calculation of $x_{\mathrm{B}}$ should be different for the bound nucleon.
Hence the parameter $X_0$ is slightly away from 1.
In addition to the nuclear Fermi motion, the nucleon-nucleon short-range correlation
results in a small number of high momentum nucleons.
In order to model the long-tail momentum distribution from short-range correlations,
we apply a model using Cauchy distribution for the quasielastic scattering peak.
Similar to the derivations of the gaussian model, the quasielastic cross section
from Cauchy-distribution smearing is given by,
\begin{equation}\label{eq:QECrosssection_dxB_Cauchy}
\begin{split}
&\frac{1}{\mathrm{A}} \frac{d \sigma_\mathrm{A}(x_\mathrm{B})}{d x_\mathrm{B}} =
\left|\frac{1}{\mathrm{A}} \frac{d \sigma_\mathrm{A}(\omega)}{d \omega} \frac{d \omega}{d x_\mathrm{B}} \right| \\
&=\frac{\sigma_{\mathrm{e}\mathrm{N}}}{\pi\Gamma_\mathrm{A}}
\frac{1}{\left[ 1+\left(\frac{(1/x_\mathrm{B})-X_0^\mathrm{A}}{\Gamma_\mathrm{A}}\right)^2 \right]} \frac{Q^2}{2 m x_\mathrm{B}^2}, \\
\end{split}
\end{equation}
in which the full width $\Gamma_{\rm A}$ is proportional to the Fermi momentum of nucleus $A$.
Then the cross-section ratio of a heavy nucleus $A$ to the deuteron ($D$) is written as,
\begin{equation}\label{eq:CrosssectionRatio_Cauchy}
\left| \frac{d \sigma_\mathrm{A}/\mathrm{A}}{d \sigma_\mathrm{D}/2}\right| =
\frac{\Gamma_\mathrm{D}}{\Gamma_\mathrm{A}}
\frac{\left[ 1+\left(\frac{(1/x_\mathrm{B})-X_0^\mathrm{D}}{\Gamma_\mathrm{D}}\right)^2 \right]}
{\left[ 1+\left(\frac{(1/x_\mathrm{B})-X_0^\mathrm{A}}{\Gamma_\mathrm{A}}\right)^2 \right]},
\end{equation}
which is the master equation in the Cauchy-distribution model.
The Fermi momenta of several nuclei were measured decades ago \cite{Moniz:1971mt}, such as $^{12}$C and $^{208}$Pb.
Recently the quasielastic cross-section ratios of some heavy nuclei ($^{12}$C, $^{27}$Al, $^{56}$Fe, and $^{208}$Pb)
to deuteron were analyzed by CLAS Collaboration in the data-mining project of 6 GeV data \cite{Schmookler:2019nvf}.
Therefore this provides us an opportunity to determine the Fermi momentum of deuteron,
with the models expressed in Eq. (\ref{eq:CrosssectionRatio_gaussian}) and Eq. (\ref{eq:CrosssectionRatio_Cauchy}),
combining the measurements of $^{12}$C and $^{208}$Pb.
The main objective of the paper is to provide the Fermi momenta of some unmeasured nuclei
based on the Fermi motion smearing model.
\section{Fermi momentum of deuteron}
\label{sec:deuteron}
Fig. \ref{fig:XsectionRatioCD} and Fig. \ref{fig:XsectionRatioPbD} show the per-nucleon
quasielastic cross-section ratios as a function of $x_\mathrm{B}$ \cite{Schmookler:2019nvf},
for $^{12}$C over deuteron
and for $^{208}$Pb over deuteron, respectively. The least-square fits based on the models
described in Eq. (\ref{eq:CrosssectionRatio_gaussian}) and Eq. (\ref{eq:CrosssectionRatio_Cauchy}) are performed to the data,
which are also shown in the figures.
The fitting results based on gaussian-distribution model and Cauchy-distribution model
are listed in Table \ref{tab:GaussModelFitResults} and Table \ref{tab:CauchyModelFitResults} respectively.
Judged by the $\chi^2/ndf$ value, the model based on Cauchy distribution is much better than
the model based on gaussian distribution.
This implies that there is the high-momentum tail for the nucleon momentum distribution,
which is unveiled directly in the short-range correlation experiments in the past two decades.
We find that the peak position is consistent with $X_0=1$ for the quasielastic scattering on the deuteron
while the peak position $X_0$ is slightly lower than 1 for the heavy nuclei.
This indicates that the separation energy of the nucleon inside the deuteron is negligible.
For the heavy nuclei, the nucleon separation energy is much larger.
The small reduced $\chi^2$ ($<1$) of the fits suggest that the quasielastic scattering model
based on Cauchy distribution is in good consistency with the experimental data (see Table \ref{tab:CauchyModelFitResults}).
Therefore we use this model to extract the nuclear Fermi momentum of the unmeasured deuteron.
In the fits, we let the full width $\Gamma_\mathrm{A}$, the full width ratio $\Gamma_\mathrm{D}/\Gamma_\mathrm{A}$,
the median $X_0^\mathrm{D}$ and the median $X_0^\mathrm{A}$ be the free parameters.
$\Gamma_\mathrm{D}/\Gamma_\mathrm{C}$ is determined to be 0.51 $\pm$ 0.03,
and $\Gamma_\mathrm{D}/\Gamma_\mathrm{Pb}$ is determined to be 0.46 $\pm$ 0.04.
For a fit of four free parameters, the $1\sigma$ uncertainties for
the determined parameters are defined by $\chi^2=\chi^2_{\rm min}+4.88$ \cite{James:1994vla}.
The ratio of the quasielastic peak width actually can be regarded as the ratio of the nuclear Fermi momentum.
As the nuclear Fermi momenta of carbon and lead
are already measured to be 221 $\pm$ 5 MeV/c and 265 $\pm$ 5 MeV/c respectively
from the previous experiments \cite{Moniz:1971mt},
we then deduce the nuclear Fermi momentum of deuteron to be 113 $\pm$ 8 MeV/c
and 122 $\pm$ 11 MeV/c from the above obtained ratios of quasielastic peak widths.
It is delighting to find that the values of the Fermi momentum of deuteron obtained from
the cross-section ratios of different nuclear targets
are consistent with each other within the uncertainty.
To get an average, the Fermi momentum of deuteron is
combined to be 116 $\pm$ 7 MeV/c.
Note that the designed resolution of CLAS drift chamber
for 1 GeV/c charged particle is $\delta p/p \leq 0.5\%$,
which is of quite high precision.
The measured elastic peak resolution is no greater than 16 MeV at beam energies of 2.4 and 4 GeV \cite{Mestayer:2000we}.
Therefore the detector resolution for $x_{\rm B}$ is about $\delta x_{\rm B} \sim (16{\rm MeV}) / (938{\rm MeV}) \sim 0.017$,
which is much smaller than the measured full width of nuclear quasielastic peak.
Moreover, the detector resolution is much smaller than the bin size of $x_{\rm B}$ as well.
Owing to the scattered electrons of high energy were measured precisely with
the drift chamber system of CLAS, we neglect the tiny detector resolution in the analysis.
In how large kinematic region that our models are valid and effective?
To answer this question, we have performed the fits to Carbon/Deutron data
in a wider kinematic range from $x_{\rm B}=0.84$ to $x_{\rm B}=1.14$.
The gaussian-distribution model gives the fitting quality $\chi^2/ndf=106/3=35$
and the Cauchy-distribution model gives the fitting quality $\chi^2/ndf=34/3=11$.
We see that our models do not fit well in a broader kinematic region of $x_{\rm B}$.
Nevertheless for the Cauchy-distribution model,
the determined free parameters in the wider fitting range are consistent
with the ones in the narrow fitting range we applied.
The obtained parameters are $\Gamma_{\rm C}=0.30\pm 0.01$, $\Gamma_{\rm D}/\Gamma_{\rm C}=0.49\pm 0.02$,
$X_0^{\rm D}=1.01\pm 0.01$ and $X_0^{\rm C}=0.99\pm 0.01$
from the fit in the wider kinematic region of $0.84<x_{\rm B}<1.14$.
\begin{figure}
\begin{center}
\includegraphics[width=0.85\linewidth]{Ratio_Xsection_Carbon_Deuterium.pdf}
\caption{The quasielastic cross-section ratio $\left| \frac{d \sigma_\mathrm{A}/\mathrm{A}}{d \sigma_\mathrm{D}/2}\right|$ of carbon to deuteron
as a function of the Bjorken variable $x_\mathrm{B}$. The experimental data are taken from CLAS Collaboration ~\cite{Schmookler:2019nvf}.
The dashed blue curve shows a fit to a model of the assumption that the quasielastic peak is gaussian.
The solid red curve shows a fit to a model of the assumption that the quasielastic peak is Cauchy distribution. }
\label{fig:XsectionRatioCD}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.85\linewidth]{Ratio_Xsection_Lead_Deuterium.pdf}
\caption{The quasielastic cross-section ratio $\left| \frac{d \sigma_\mathrm{A}/\mathrm{A}}{d \sigma_\mathrm{D}/2}\right|$ of lead to deuteron
as a function of the Bjorken variable $x_\mathrm{B}$. The experimental data are taken from CLAS Collaboration ~\cite{Schmookler:2019nvf}.
The dashed blue curve shows a fit to a model of the assumption that the quasielastic peak is gaussian.
The solid red curve shows a fit to a model of the assumption that the quasielastic peak is Cauchy distribution. }
\label{fig:XsectionRatioPbD}
\end{center}
\end{figure}
\begin{table}[h]
\caption{\label{tab:GaussModelFitResults}
The least-square fitting results of the cross-section ratio data,
within the model of gaussian distribution for the quasielastic scattering peak.
The fit quality, the width, the width ratio and the central value of the peak are listed. }
\begin{tabular}{cccccc}
\hline\hline
Fit & $\frac{\chi^2}{ndf}$ & $W_\mathrm{A}$ & $W_\mathrm{D}/W_\mathrm{A}$ & $X_0^\mathrm{D}$ & $X_0^\mathrm{A}$ \\
\hline
C/D & $\frac{11}{1}$ & $0.14\pm0.03$ & $0.54\pm0.07$ & $1.01\pm0.02$ & $0.97\pm0.06$ \\
Pb/D& $\frac{9.6}{1}$ & $0.14\pm0.23$ & $0.53\pm0.64$ & $0.99\pm0.12$ & $0.92\pm0.20$ \\
\hline\hline
Fit & $\frac{\chi^2}{ndf}$ & $W_\mathrm{D}$ & $W_\mathrm{A}/W_\mathrm{D}$ & $X_0^\mathrm{D}$ & $X_0^\mathrm{A}$ \\
\hline
Al/D& $\frac{7.0}{1}$ & $0.073\pm0.005$ & $1.9\pm0.3$ & $1.01\pm0.02$ & $0.97\pm0.06$ \\
Fe/D& $\frac{6.6}{1}$ & $0.074\pm0.005$ & $2.0\pm0.3$ & $1.01\pm0.02$ & $0.96\pm0.06$ \\
\hline\hline
\end{tabular}
\end{table}
\begin{table}[h]
\caption{\label{tab:CauchyModelFitResults}
The least-square fitting results of the cross-section ratio data,
within the model of Cauchy distribution for the quasielastic scattering peak.
The fit quality, the full width, the full-width ratio and the central value of the peak are listed. }
\begin{tabular}{cccccc}
\hline\hline
Fit & $\frac{\chi^2}{ndf}$ & $\Gamma_\mathrm{A}$ & $\Gamma_\mathrm{D}/\Gamma_\mathrm{A}$ & $X_0^\mathrm{D}$ & $X_0^\mathrm{A}$ \\
\hline
C/D & $\frac{0.37}{1}$ & $0.31\pm0.02$ & $0.51\pm0.03$ & $1.01\pm0.01$ & $0.99\pm0.03$ \\
Pb/D& $\frac{0.33}{1}$ & $0.37\pm0.03$ & $0.46\pm0.04$ & $1.01\pm0.01$ & $0.96\pm0.03$ \\
\hline\hline
Fit & $\frac{\chi^2}{ndf}$ & $\Gamma_\mathrm{D}$ & $\Gamma_\mathrm{A}/\Gamma_\mathrm{D}$ & $X_0^\mathrm{D}$ & $X_0^\mathrm{A}$ \\
\hline
Al/D& $\frac{1.7}{1}$ & $0.159\pm0.008$ & $2.0\pm0.2$ & $1.01\pm0.01$ & $0.98\pm0.03$ \\
Fe/D& $\frac{1.2}{1}$ & $0.165\pm0.008$ & $2.1\pm0.2$ & $1.01\pm0.01$ & $0.98\pm0.03$ \\
\hline\hline
\end{tabular}
\end{table}
One of the main motivations for this study is to understand the dip presented
in the cross-section ratio around $x_{\rm B} = 1$.
In addition to the ratio between the cross sections of heavy nucleus and deuteron,
does the smearing model interpret the inclusive nuclear cross section as well?
Fig. \ref{fig:NuclearXsections} shows the electron inclusive scattering
cross sections of $^2$H and $^{12}$C as a function of $x_{\rm B}$.
The experimental data are taken from the references \cite{Arrington:1998ps,Benhar:2006er},
which were measured under the electron beam energy of 4.045 GeV.
Note that the $d\sigma/d\omega$ data were transformed into the $d\sigma/dx_{\rm B}$ data
according to Eq. (\ref{eq:QECrosssection_dxB_gaussian}).
The dashed curves show the fits of the Cauchy distribution for quasi-elastic
scattering peak plus the background for inelastic scattering and nucleon-nucleon correlations.
In performing the fits, the central values and the full widths of the Cauchy distributions
for deuteron and carbon are fixed as the fitted values listed in Table \ref{tab:CauchyModelFitResults}.
We find that the simple Cauchy-distribution model describes well the inclusive cross sections.
Unfixing the central values for Cauchy-distribution peaks,
the fits reproduce the experimental data even better.
We also find that there is significant contribution from inelastic scattering
for the carbon data at high energy around 4 GeV.
Therefore, to extract the nuclear Fermi momentum,
the low-energy data are preferred.
\begin{figure}
\begin{center}
\includegraphics[width=0.85\linewidth]{Electron_Scattering_Xsection.pdf}
\caption{The electron inclusive cross sections on deuteron and carbon targets as a function of $x_{\rm B}$,
at the electron beam energy of 4.045 GeV.
The dashed curves show the fits based on the model of a Cauchy distribution
for quasi-elastic peak plus a continuum background for inelastic scattering.
For the fits, the full widths are fixed at 0.16 and 0.31 for deuteron and carbon respectively.
And the central values of the quasi-elastic peaks are fixed at 1.01 and 0.99 for deuteron and carbon respectively. }
\label{fig:NuclearXsections}
\end{center}
\end{figure}
\section{Fermi momenta of aluminium and iron}
\label{sec:Al-and-Fe}
The quasielastic scattering cross-section ratios between some other heavy nuclei
and the deuteron are also measured by CLAS Collaboration.
Fig. \ref{fig:XsectionRatioAlD} shows the cross-section ratio between $^{27}$Al and deuteron as a function of $x_{\rm B}$;
And Fig. \ref{fig:XsectionRatioFeD} shows the cross-section ratio between $^{56}$Fe and deuteron as a function of $x_{\rm B}$.
By performing the fits to the CLAS data within the model
of Cauchy-distribution peak (Eq. (\ref{eq:CrosssectionRatio_Cauchy})),
we have obtained the ratios of quasielastic peak widths:
$\Gamma_\mathrm{Al}/\Gamma_\mathrm{D}$ = 2.0 $\pm$ 0.2 and $\Gamma_\mathrm{Fe}/\Gamma_\mathrm{D}$ = 2.1 $\pm$ 0.2.
In Figs. \ref{fig:XsectionRatioAlD} and \ref{fig:XsectionRatioFeD},
we see that the Cauchy-distribution model describes the experimental data amazingly well,
with the $\chi^2/ndf$ around 1 (see Table \ref{tab:CauchyModelFitResults}).
The full-width ratio is directly equal to the nuclear-Fermi-momentum ratio.
As we have determined the Fermi momentum of the deuteron in the above section,
we can calculate the Fermi momentum of the heavy nucleus with
the obtained ratio of the Fermi momenta between the heavy nucleus and the deuteron.
Based on the results of the above fits with the Cauchy-distribution model,
we have determined the Fermi momenta
of $^{27}$Al and $^{56}$Fe, which are summarized in Table \ref{tab:FermiMomentumResults}
and comparable with the simple Fermi gas model calculations.
In a recent work \cite{Murphy:2019wed}, the Fermi momentum of $^{27}$Al
is estimated from the super scaling phenomenon around $\Psi\sim 0$.
Our obtained Fermi momentum of aluminum nucleus is obvious smaller than
the value given by the super scaling analysis \cite{Murphy:2019wed},
though the discrepancy between the two values is within the uncertainty of our analysis.
One possible source for the discrepancy is that the electron beam energies
and $Q^2$ are different for the data used in the two analyses,
which result in the different magnitudes of inelastic scattering contributions.
Ignoring the influence of inelastic scattering may introduce
a sizeable systematic uncertainties in our analysis.
The other possible source for the discrepancy is
the violation of super scaling phenomenon.
\begin{figure}
\begin{center}
\includegraphics[width=0.85\linewidth]{Ratio_Xsection_Aluminium_Deuterium.pdf}
\caption{The quasielastic cross-section ratio $\left| \frac{d \sigma_\mathrm{A}/\mathrm{A}}{d \sigma_\mathrm{D}/2}\right|$ of aluminium to deuteron
as a function of the Bjorken variable $x_\mathrm{B}$. The experimental data are taken from CLAS Collaboration ~\cite{Schmookler:2019nvf}.
The dashed blue curve shows a fit to a model of the assumption that the quasielastic peak is gaussian.
The solid red curve shows a fit to a model of the assumption that the quasielastic peak is Cauchy distribution. }
\label{fig:XsectionRatioAlD}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.85\linewidth]{Ratio_Xsection_Iron_Deuterium.pdf}
\caption{The quasielastic cross-section ratio $\left| \frac{d \sigma_\mathrm{A}/\mathrm{A}}{d \sigma_\mathrm{D}/2}\right|$ of iron to deuteron
as a function of the Bjorken variable $x_\mathrm{B}$. The experimental data are taken from CLAS Collaboration ~\cite{Schmookler:2019nvf}.
The dashed blue curve shows a fit to a model of the assumption that the quasielastic peak is gaussian.
The solid red curve shows a fit to a model of the assumption that the quasielastic peak is Cauchy distribution. }
\label{fig:XsectionRatioFeD}
\end{center}
\end{figure}
\begin{table}[h]
\caption{\label{tab:FermiMomentumResults}
Fermi Momenta of some nuclei determined in this work.
$k_{\mathrm{F,~exp.}}$ denotes the Fermi momentum given by our analysis of the CLAS data.
The errors are the statistical errors only.
$k_{\mathrm{F,~theo.}}$ denotes the Fermi momentum given by the calculation
from the Fermi gas model for the nucleus (see Eq. (\ref{eq:FerimiMomentum_density})). }
\begin{tabular}{ccccc}
\hline\hline
&Nucleus &$k_{\mathrm{F,~exp.}}$ (MeV/c)& $k_{\mathrm{F,theo.}}$ (MeV/c)&\\
\hline
& $^{2}$H & $116\pm7$ & 140 &\\
& $^{27}$Al & $232\pm27$ & 226 &\\
& $^{56}$Fe & $244\pm28$ & 231 &\\
\hline\hline
\end{tabular}
\end{table}
\section{Discussions and summary}
\label{sec:summary}
From our analysis based on the simple model of nucleon momentum smearing, we find that
the nuclear Fermi momentum of deuteron is about half of that of a heavy nucleus,
while the Fermi momenta of aluminium and iron are close to the Fermi momentum of heavy nucleus
around 250 MeV/c. An interesting question is that whether or not
the Fermi motion of the nucleon in deuteron satisfies the Fermi gas model description.
In Fermi gas model, the Fermi momentum is directly connected to the nuclear density.
Based on Pauli exclusion principle,
the nucleon fermions fully occupy the quantum states in the nucleus.
Assuming a sphere nucleus, the number of nucleon fermions is simply counted as \cite{d2019introduction},
\begin{equation}\label{eq:nucleon_number}
\tilde{n} = 2 \frac{\left(\frac{4}{3}\pi R^3\right) \left(\frac{4}{3}\pi k_{\rm F}^3\right) }{(2\pi\hbar)^3 },
\end{equation}
where $\tilde{n}$ is the proton number or the neutron number,
$R$ is the radius of a nucleus and $k_{\rm F}$ is the Fermi momentum.
Approximately the nuclear radius is proportional to $A^{1/3}$.
With $R=r_0A^{1/3}$, the proton Fermi momentum $k_{\rm F}^{\rm p}$
in Fermi gas model is calculated as,
\begin{equation}\label{eq:FerimiMomentum_radius}
k_{\mathrm{F}}^{\mathrm{p}}=\frac{1}{r_0}\left(\frac{9}{4}\pi \frac{Z}{A}\right)^{\frac{1}{3}},
\end{equation}
in which $\hbar$ equals to 1 with the natural unit used.
Using the formula $\rho=A/(\frac{4}{3}\pi R^3)$, we can rewrite the proton Fermi momentum
as a function of the nuclear density $\rho$, which is written as,
\begin{equation}\label{eq:FerimiMomentum_density}
k_{\mathrm{F}}^{\mathrm{p}}=\left(3 \pi^2 \frac{Z}{A} \rho \right)^{\frac{1}{3}}.
\end{equation}
Here we assume a spherical nucleus and we set $\hbar$ to 1
with the natural unit used.
We adopt the nuclear density data from Ref. \cite{Gomez:1993ri} (see Table~\ref{tab:nuclear_density}
for the nuclei studied in the reference). These average nuclear densities are calculated
with $\rho(A) = 3A/(4\pi R_e^3)$ and $R_e^2 = 5\left<r^2\right>/3$.
$\sqrt{ \left<r^2\right> }$ is the root-mean-square radius
from the elastic electron scattering on the nucleus \cite{Gomez:1993ri}.
\begin{table}[h]
\caption{\label{tab:nuclear_density} Nucleon densities of some nuclei \cite{Gomez:1993ri}. }
\begin{tabular}{ccccccc}
\hline\hline
&A & $^{2}$H & $^{12}$C & $^{27}$Al & $^{56}$Fe &\\
\hline
&$\rho$ (fm$^{-3}$)& 0.024 & 0.089 & 0.106 & 0.117 &\\
\hline\hline
\end{tabular}
\end{table}
With the nucleon density of $^{12}$C, the Fermi momentum of $^{12}$C is 216 MeV/c based on the Fermi gas model
described above, which is consistent with the previously measured value $221\pm 5$ MeV/c \cite{Moniz:1971mt}.
The Fermi gas model predictions for some other nuclei are also listed in Table \ref{tab:FermiMomentumResults}.
We find that the Fermi momenta of $^{27}$Al and $^{56}$Fe extracted in this work agree well with
the Fermi gas model calculations, while the Fermi momentum of deuteron is slightly smaller than the model prediction.
This implies that the Fermi gas model may not work well for the very light nucleus.
It is of no surprise because the deuteron is too light to fit in the Fermi gas picture.
Moreover, the determined Fermi momentum of deuteron is larger than the prediction 87 MeV/c
from a phenomenological parametrization \cite{Wang:2016mzo}, with
$k_{\mathrm{F}}(Z,N,A)=K_{\mathrm{F}}^{\mathrm{p}}(1-A^{-t_{\mathrm{p}}})\frac{Z}{A} + K_{\mathrm{F}}^{\mathrm{n}}(1-A^{-t_{\mathrm{n}}})\frac{N}{A}$.
The Fermi momentum of deuteron is extracted for the first time
from the quasielastic scattering data from CLAS \cite{Schmookler:2019nvf}.
The Fermi momentum of deuteron is $116\pm 7$ MeV/c, which is much smaller than that of the nucleus of high density.
The good fits of our model to the data indicate that the Cauchy distribution is valid
to fit the quasielastic cross section in a narrow range,
and the width of the quasielastic peak is proportional to the nuclear Fermi momentum.
The cross-section ratio for quasielastic scattering can be described with the formula in Eq. (\ref{eq:CrosssectionRatio_Cauchy}).
Within this formulism and the CLAS data, we have extracted the Fermi momenta of two
unmeasured heavy nuclei, $^{27}$Al and $^{56}$Fe.
The quasielastic scattering by the high energy electron beam on the nuclear target is
not only an important tool to see the short-range correlations among nucleons \cite{Schmookler:2019nvf},
but also a powerful method to acquire the information of nucleon Fermi motion inside the nucleus.
And these nucleon Fermi motion information would be very helpful
to the analyses in the experiments with the nuclear targets to study the properties of neutrons or protons
(such as the deuterium target).
Last but not least, to understand the Fermi momenta of the very light nuclei we need
more theoretical and experimental studies.
\begin{acknowledgments}
This work is supported by the National Natural Science Foundation of China under the Grant NO. 12005266
and the Strategic Priority Research Program of Chinese Academy of Sciences under the Grant NO. XDB34030301.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}\label{sec1}
Let $\varOmega$ be a bounded domain in $\mathbb{R}^n$, $n\ge 2$, with a Lipschitz boundary $\partial\varOmega$.
In the space-time cylinder $\varOmega_T=\varOmega\times (0,T)$, $T\in(0,\infty)$, we consider the following differential equation:
\begin{equation}\label{1.1}
\partial_t u -\Delta u +\varphi\Big(\int_0^T \alpha(\cdot,s)\, u(\cdot,s)\,ds\Big)\, u =f,
\end{equation}
where $u=u(x,t)$ is an unknown scalar function, $x=(x_1,\ldots,x_n)$ the vector of the spatial variables
in $\mathbb{R}^n$, $t$ the time variable on $[0,T]$, $\varphi: \mathbb{R}\to \mathbb{R}$ a
function that will be specified below, $f:\varOmega_T\to \mathbb{R}$ a prescribed function. We suppose that
the following boundary and initial conditions are satisfied:
\begin{equation}\label{1.2}
u(x,t)=0\quad \text{for}\quad x\in\partial\varOmega,\quad t\in [0,T],
\end{equation}
\begin{equation}\label{1.3}
u(x,0)=u_0(x)\quad \text{for}\quad x\in\varOmega,
\end{equation}
where the function $u_0:\varOmega\to \mathbb{R}$ is prescribed.
Equation \eqref{1.1} contains a non-local in time $t$ term with the integral over the
whole interval $(0, T)$, where the problem is considered. For this reason, we call this equation global-in-time.
A problem of this type arose when modeling the chaotic dynamics of a polymer molecule (polymer chain) in an aqueous solution
\cite{StSt}. The function $u(x,t)$ corresponds to the density of probability that the
$t$-th segment of the chain is in a certain region of the space. The role of time in the equation
is played by the arc length parameter along the chain whose total length is $T$.
The equation contains a term responsible for the interaction of chain's segments and the function
$\varphi$ is called the \emph{interaction potential}. Since each segment interacts with all others
through the surrounding fluid, $\varphi$ depends on the integral of the density of probability over the entire chain,
i.e., over the entire time interval $(0,T)$.
Equations with a global-in-time term appear also in other fields, for example, in the filtration theory \cite{L},
in the population dynamics (see \cite{Webb,W1} and the references therein).
Notice that the equation in the population dynamics is ultra-parabolic
and the age of individuals plays the role of the second time. In the steady case, the equation becomes parabolic.
The global terms contain the integral with respect to the age and can be in the data of the problem and in the equation as well.
There are a lot of works, where a global-in-time term stands in the data of the problem.
These are problems with non-local boundary or initial conditions. The latter means that the initial data depend on
the time integral of the solution. This topic is widely represented in the literature, and there is no problem to find
relevant publications. The problems with global-in-time data are very different from ours.
The problem we study looks like parabolic one, however it has some unusual features.
First of all, since the potential $\varphi$ in equation \eqref{1.1} is global-in-time, the state of the system
depends not only on the past but also on the future. That is, the causality principle is violated. Besides, from a mathematical point
of view, the solution of a nonlinear parabolic problem is commonly being constructed locally in time and is extended
afterwards. In our case, this procedure is impossible. Finally, as a rule, the local-in-time uniqueness of the solution implies
the global one. We cannot prove the uniqueness without restrictions on $T$.
There are several papers devoted to problems close to \eqref{1.1}--\eqref{1.3}. In \cite{St1,St2},
the weak solvability of problem \eqref{1.1}--\eqref{1.3} with $\alpha\equiv 1$ and $f=0$ is proven. Notice that
the smoothness of the obtained in this papers weak solution can be easily improved for better initial data.
The strong solvability of the same problem was proven in \cite{W2}, where the semigroup approach was employed.
In \cite{D}, the authors consider the problem with the Laplace operator replaced by a more complex non-local
operator that can be thought as the fractional Laplacian. The uniqueness of the solution is proven for sufficiently
small $T$ in all these works. The smallness condition means that there is a restriction for $T$ which depends on the
data of the problem.
In this paper, we consider a more general problem compared with the previous works. At first, the global-in-time term
contains a weighted integral and the smoothness of the weight $\alpha$ is assumed to be not sufficient to apply the
technique from \cite{St1,St2}. Secondly, equation \eqref{1.1} contains the right-hand side $f$.
The point is that the proof of the main results of the paper employs only the energy estimate and does not use any
smoothness properties of the solution with respect to time. The right-hand side $f$ prevents or makes it difficult to
obtain good time-estimates. Finally, the potential $\varphi$ is assumed to be just a continuous bounded from below function.
There are no conditions on the growth of this function at infinity, no convexity, monotonicity, or differentiability conditions.
In \cite{St2}, the assumption on the potential $\varphi$ also does not contain the growth conditions, but $\varphi$
satisfies other requirements that made it possible to obtain a higher degree of integrability of $\varphi$ than in the presented paper.
The absence of the growth conditions at infinity leads to the fact that the interaction term is not a lower order term
in equation \eqref{1.1}. Notice that in \cite{W2} this term has a lower order.
Although the continuous function $\varphi$ is assumed to be just bounded from below, by standard arguments
the problem can be reduced to the problem with a positive potential. Really,
suppose that $\varphi(\xi)\ge -K$ for some positive number $K$ and all $\xi\in \mathbb{R}$.
If $u$ is a solution of problem~\eqref{1.1}--\eqref{1.3}, then $\bar{u}(x,t)=e^{-Kt}\,u(x,t)$ is a solution
of the following problem
$$
\partial_t \bar{u} -\Delta \bar{u} +\bar{\varphi}(\bar{v})\, \bar{u} =\bar{f},
\quad \bar{u}|_{\partial\varOmega}=0,\quad \bar{u}|_{t=0}=u_0,
$$
where
$$
\bar{\varphi}=\varphi +K,\quad
\bar{v}(x)=\int_0^T \bar{\alpha}(x,t)\, \bar{u}(x,t)\,dt,
\quad \bar{\alpha}(x,t)=e^{Kt}\,\alpha(x,t), \quad \bar{f}(x,t)=e^{-Kt}\,f(x,t).
$$
Thus, we have the same problem but with the non-negative potential $\bar{\varphi}$.
In what follows, we will assume that $\varphi$ is a non-negative function.
We will use the standard Lebesgue and Sobolev spaces $L^p(\varOmega)$, $H^1_0(\varOmega)$,
$L^q(0,T;L^p(\varOmega))$, and $L^q(0,T;H^1_0(\varOmega))$, where $p,q\in [1,\infty]$.
As usual, $H^{-1}(\varOmega)$ is the dual space of $H^1_0(\varOmega)$
with respect to the pivot space $L^2(\varOmega)$. The norm in $L^2(\varOmega)$ will be denoted by $\|\cdot\|$.
\begin{definition}\label{t1.1}
Let $\varphi: \mathbb{R}\to \mathbb{R}$ be a non-negative continuous function, $T\in (0,\infty)$,
$\alpha\in L^1(0,T; L^2(\varOmega))$, $f\in L^1(0,T; L^2(\varOmega))$, and $u_0\in L^2(\varOmega)$.
A function $u:\varOmega_T\to \mathbb{R}$ is said to be a \emph{weak solution of problem \eqref{1.1}--\eqref{1.3}}, if
\begin{enumerate}
\item
$u\in L^2(0,T;H^1_0(\varOmega))$ and $\varphi(\zeta)\,u\in L^1(\varOmega_T)$, where
$\zeta=\int_0^T \alpha (\cdot,t)\, u(\cdot,t)\,dt$;
\item
the following integral identity
$$
\int_0^T \int_\varOmega \big(u\,\partial_t h - \nabla u\cdot\nabla h - \varphi(v)\,u\, h +f\,h\big)\,dx\,dt +
\int_\varOmega u_0 h_0\, dx=0
$$
holds for an arbitrary smooth in the closure of $\varOmega_T$ function $h$ such that $h(x,t)=0$
for $x\in\partial\varOmega$ and for $t=T$. Here, $h_0=h|_{t=0}$.
\hfill\textbullet
\end{enumerate}
\end{definition}
The main result of the paper is Theorem~\ref{t3.1} that states the existence of the weak solution of
problem~\eqref{1.1}--\eqref{1.3}. In the next section, we consider the case where the potential is
assumed to be bounded and prove Theorem~\ref{t2.1} that is an auxiliary result for Theorem~\ref{t3.1}.
We emphasize once again that in the proofs of these theorems we do not use any continuity properties of the solution in time.
This leads to certain difficulties, for example, in proving the energy estimate \eqref{3.1} (see Step~4 of the proof of
Theorem~\ref{t3.1}) which is usually obtained automatically from
the corresponding estimate for the approximate solutions by using the weak semicontinuity of norms in Banach spaces.
\section{Problem with a bounded potential}\label{sec2}
In this section, we consider the case of the problem where the potential $\varphi:\mathbb{R}\to \mathbb{R}$
is a continuous bounded function.
\begin{theorem}\label{t2.1}
Assume that $T\in (0,\infty)$, $\alpha\in L^1(0,T; L^2(\varOmega))$,
$f\in L^1(0,T; L^2(\varOmega))$, $u_0\in L^2(\varOmega)$, and $\varphi:\mathbb{R}\to \mathbb{R}$
is a continuous function such that
$0\le \varphi(\xi)\le K$ for all $\xi\in \mathbb{R}$, where $K$ is a positive number.
Then there exists a weak solution $u\in L^\infty(0,T;L^2(\varOmega))\cap L^2(0,T;H^1_0(\varOmega))$
of problem~\eqref{1.1}--\eqref{1.3}.
\end{theorem}
\noindent
\textsc{Proof.} To prove the solvability of the problem, we employ the Schauder fixed point theorem. Let $B_R$ be the
ball of the radius $R$ centered at the zero in the function space $L^1(\varOmega)$. The number $R$ will be
determined later. Let us define the mapping $\varPsi:B_R\to B_R$ that is referred to in the Schauder theorem.
For every $w\in B_R$, we define $u_w$ as the solution of the following problem:
\begin{equation}\label{2.1}
\partial_t u_w -\Delta u_w +\varphi(w)\, u_w =f,
\quad u_w|_{\partial\varOmega}=0,\quad u_w|_{t=0}=u_0.
\end{equation}
As it follows from the classical theory of parabolic equations (see, e.g., \cite[Ch.\,7]{Evans}, this problem has a unique weak solution.
Now, we set
$$
\varPsi(w)=\int_0^T \alpha(\cdot,t)\, u_w(\cdot,t)\,dt.
$$
It necessary to prove that there exists $R>0$ such that $\varPsi(B_R)\subset B_R$ and $\varPsi:B_R\to L^1(\varOmega)$
is a compact continuous map.
The usual energy estimate for problem \eqref{2.1} looks as follows:
$$
\|u_w(\cdot,t)\|^2+\int_0^t\|\nabla u_w(\cdot,s)\|^2\,ds \le
\Big(\|u_0\| +\int_0^t \|f(\cdot,s)\|\,ds\Big)^2+\frac{1}{2}\,\Big(\int_0^t \|f(\cdot,s)\|\,ds\Big)^2
$$
for almost all $t\in [0,T]$. We have used the positiveness of $\varphi$ although a similar estimate
holds for $\varphi$ which is just bounded. Thus,
\begin{equation}\label{2.2}
\esup_{t\in[0,T]}\|u_w(\cdot,t)\|^2+\int_0^T\|\nabla u_w(\cdot,s)\|^2\,ds \le C_1,
\end{equation}
where $C_1=\big(\|u_0\| +\int_0^T \|f(\cdot,s)\|\,ds\big)^2+\frac{1}{2}\,\big(\int_0^T \|f(\cdot,s)\|\,ds\big)^2$.
As a consequence of this inequality, we find that
$$
\|\varPsi(w)\|_1\le C_2\sqrt{C_1},
$$
where $C_2=\int_0^T \|\alpha(\cdot,s)\|\,ds$. Therefore, $\varPsi(B_R)\subset B_R$ with $R=C_2\sqrt{C_1}$.
In order to prove the continuity of the mapping $\varPsi$, let us take an arbitrary sequence $\{w_k\}$ in $B_R$ that converges
in $L^1(\varOmega)$ to some $w\in B_R$. Then $w_k\to w$ in measure on $\varOmega$ and, due to the Lebesgue
dominated convergence theorem, $\varphi(w_k)\to \varphi(w)$ in $L^p(\varOmega)$ for all $p\in [1,\infty)$.
The difference $u_{w_k}-u_w$ is a weak solution of the following problem:
\begin{gather*}
\partial_t (u_{w_k}-u_w) -\Delta (u_{w_k}-u_w) +\varphi(w)\, (u_{w_k}-u_w) +
\big(\varphi(w_k)-\varphi(w)\big)\, u_{w_k}=0,
\\
\quad (u_{w_k}-u_w)|_{\partial\varOmega}=0,\quad (u_{w_k}-u_w)|_{t=0}=0.
\end{gather*}
Therefore, it satisfies estimate \eqref{2.2} with $f=\big(\varphi(w_k)-\varphi(w)\big)\, u_{w_k}$ and $u_0=0$. Thus,
$$
\esup_{t\in[0,T]}\|(u_{w_k}-u_w)(\cdot,t)\|^2+\int_0^T\|\nabla (u_{w_k}-u_w)(\cdot,s)\|^2\,ds \le c_k,
$$
where $c_k=\frac{3}{2}\,\big(\int_0^T \|\big(\varphi(w_k)-\varphi(w)\big)\, u_{w_k}(\cdot,s)\|\,ds\big)^2$.
Due to the Sobolev embedding theorems, estimate \eqref{2.2} implies that
$$
\int_0^T \|u_{w_k}(\cdot,s)\|_{L^q(\varOmega)}^2\,ds\le C_3,
$$
for all $k\in \mathbb{N}$ and some independent of $k$ constant $C_3$, where $q=2n/(n-2)$,
if $n>2$, and $q=\infty$ otherwise. As a consequence of the H\"older inequality, we have
$$
\int_0^T \|\big(\varphi(w_k)-\varphi(w)\big)\, u_{w_k}(\cdot,s)\|\,ds
\le \|(\varphi(w_k)-\varphi(w)\|_{L^p(\varOmega)}\, \int_0^T \|u_{w_k}(\cdot,s)\|_{L^q(\varOmega)}\,ds,
$$
where $p=2q/(q-2)$. Thus, $c_k\to 0$ and $\esup_{t\in[0,T]}\|(u_{w_k}-u_w)(\cdot,t)\|\to 0$ as $k\to \infty$.
This means that $\varPsi(w_k)\to \varPsi(w)$ in $L^1(\varOmega)$ as $k\to\infty$. Really,
\begin{multline*}
\|\varPsi(w_k)- \varPsi(w)\|_{L^1(\varOmega)}\le \int_0^T \|\alpha(\cdot,s)\|\, \|(u_{w_k}-u_w)(\cdot,s)\|\,ds
\\
\le \int_0^T \|\alpha(\cdot,s)\|\,ds\, \esup_{t\in[0,T]}\|(u_{w_k}-u_w)(\cdot,t)\|\to 0
\quad\text{as}\quad k\to\infty.
\end{multline*}
Thus, the mapping $\varPsi$ is continuous on $B_R$.
It remains to prove the compactness of $\varPsi$. It is possible to draw arguments based on the available
higher smoothness of the solution to problem \eqref{2.1}. In fact, $u_w$ satisfies a stronger estimate than
\eqref{2.2}. However, we will show that this estimate is sufficient to prove the compactness of $\varPsi$.
Let $\{w_k\}$ be a bounded sequence in $B_R$. We need to show that there exists a subsequence $\{w_{k'}\}$
such that $\varPsi(w_{k'})$ converges in $L^1(\varOmega)$. According to estimate \eqref{2.2},
the sequence $\{u_{w_k}\}$ is bounded in $L^\infty(0,T;L^2(\varOmega))\cap L^2(0,T;H^1_0(\varOmega))$.
Besides that, $\|\nabla\int_0^T h \,u_{w_{k}}\,dt\|\le C_4$ for
an arbitrary smooth function $h:\varOmega_T\to \mathbb{R}$. The constant $C_4$ depends, of course, on $h$.
Therefore, there exists a function $v\in L^\infty(0,T;L^2(\varOmega))\cap L^2(0,T;H^1_0(\varOmega))$
and a subsequence $\{u_{w_{k'}}\}$ such that $u_{w_{k'}}\to v$ as $k'\to\infty$
weakly in $L^2(0,T;H^1_0(\varOmega))$ and $*$-weakly in $L^\infty(0,T;L^2(\varOmega))$, and
$$
\int_0^T h \,u_{w_{k'}}\,dt\to \int_0^T h \,v\,dt\quad\text{in}\quad L^2(\varOmega).
$$
Let us take an arbitrary $\varepsilon>0$. There exists a smooth function $\alpha_\varepsilon:\varOmega_T\to \mathbb{R}$ such that
$\|\alpha-\alpha_\varepsilon\|_{L^1(0,T;L^2(\varOmega))}<\varepsilon$. If we take $h=\alpha_\varepsilon$ in the previous
relation, then we find that
\begin{multline*}
\limsup_{k'\to\infty}\Big\|\int_0^T\alpha\,(u_{w_{k'}}-v)\,dt\Big\|_{L^1(\varOmega)}\le
\limsup_{k'\to\infty}\Big\|\int_0^T(\alpha-\alpha_\varepsilon)\,(u_{w_{k'}}-v)\,dt\Big\|_{L^1(\varOmega)}
\\
+\limsup_{k'\to\infty}\Big\|\int_0^T\alpha_\varepsilon (u_{w_{k'}}-v)\,dt\Big\|_{L^1(\varOmega)}
\le C_5\,\|\alpha-\alpha_\varepsilon\|_{L^1(0,T;L^2(\varOmega))} < C_5\,\varepsilon,
\end{multline*}
where the constant $C_5$ is an upper bound for the norm of $u_{w_{k'}}-v$ in $L^\infty(0,T;L^2(\varOmega))$.
Since $\varepsilon$ is arbitrary,
$$
\lim_{k'\to\infty}\Big\|\int_0^T\alpha\,(u_{w_{k'}}-v)\,dt\Big\|_{L^1(\varOmega)}=0.
$$
Thus, the sequence $\{\varPsi(w_{k'})\}$ converges in $L^1(\varOmega)$.
So, all the conditions of the Schauder theorem is fulfilled and there exists $w\in L^1(\varOmega)$ such that
$\varPsi(w)=w$. Then $u=u_w$ is a weak solution of problem~\eqref{1.1}--\eqref{1.3} and the theorem is proven.
\hfill$\square$
\begin{remark}\label{t2.2}
In Theorem~\ref{t2.1}, we assumed that $f\in L^1(0,T; L^2(\varOmega))$. Without changing the proof, we could suppose
that $f\in L^2(0,T; H^{-1}(\varOmega))$. Really, the proof is based on the energy estimate \eqref{2.2}.
This estimate is also true with $C_1=\|u_0\|^2 +c^2 \int_0^T \|f(\cdot,s)\|^2_{H^{-1}(\varOmega)}\,ds$,
where the constant $c$ is such that $\|u\|_{H^1_0(\varOmega)}\le c\,\|\nabla u\|$.
\hfill$\bullet$
\end{remark}
\section{Problem with an unbounded potential}\label{sec3}
In this section, we consider a more general case where the potential $\varphi$
is a non-negative continuous function which is not necessarily bounded.
\begin{theorem}\label{t3.1}
Assume that $T\in (0,\infty)$, $\alpha\in L^\infty(\varOmega; L^2(0,T))$,
$f\in L^1(0,T; L^2(\varOmega))$, $u_0\in L^2(\varOmega)$, and $\varphi:\mathbb{R}\to \mathbb{R}$
is a non-negative continuous function. Then there exists a weak solution $u\in
L^\infty(0,T;L^2(\varOmega))\cap L^2(0,T;H^1_0(\varOmega))$ of problem~\eqref{1.1}--\eqref{1.3} such that
\begin{equation}\label{3.1}
\esup_{t\in[0,T]}\|u(\cdot,t)\|^2+\int_0^T\|\nabla u(\cdot,t)\|^2\,dt
+\int_0^T \int_\varOmega \varphi(\zeta(x))\, u^2(x,t)\, dx\, dt
\le C_1,
\end{equation}
\begin{equation}\label{3.2}
\text{$\varphi(\zeta)$, $\varphi(\zeta)\zeta$, and $\varphi(\zeta)\zeta^2$ are in $L^1(\varOmega)$},
\end{equation}
\begin{equation}\label{3.3}
\text{$\varphi(\zeta)u$ and $\varphi(\zeta)u^2$ are in $L^1(\varOmega_T)$},
\end{equation}
where $\zeta=\int_0^T \alpha (\cdot,t)\, u(\cdot,t)\,dt\in L^2(\varOmega)$ and
$C_1=\big(\|u_0\| +\int_0^T \|f(\cdot,s)\|\,ds\big)^2+\frac{1}{2}\,\big(\int_0^T \|f(\cdot,s)\|\,ds\big)^2$.
\end{theorem}
\noindent
\textsc{Proof.}
For convenience, we split the proof into several steps.
\medskip\noindent
\textsc{Step~1.}
For every $k\in \mathbb{N}$, we define the following function:
$$
\varphi_k(\xi)=\begin{cases}
\varphi(\xi), & \varphi(\xi)\le k,
\\
k, & \text{otherwise}.
\end{cases}
$$
Let us denote by $u_k$ the weak solution of problem~\eqref{1.1}--\eqref{1.3} with the potential $\varphi_k$.
Since $\varphi_k$ is bounded and $\alpha\in L^\infty(\varOmega; L^2(0,T))\subset L^1(0,T; L^2(\varOmega))$,
the existence of $u_k$ is proven in the previous section (Theorem~\ref{t2.1}).
For brevity, we introduce the notation:
$$
\zeta_k=\int_0^T \alpha (\cdot,t)\, u_k(\cdot,t)\,dt.
$$
According to Definition~\ref{t1.1}, $u_k$ satisfies the integral identity
\begin{equation}\label{3.4}
\int_0^T \int_\varOmega \big(u_k\,\partial_t h - \nabla u_k\cdot\nabla h - \varphi_k(\zeta_k)\,u_k\, h
+ f\, h\big)\,dx\,dt + \int_\varOmega u_0 \,h_0\, dx=0
\end{equation}
for an arbitrary smooth in the closure of $\varOmega_T$ function $h$ such that $h(x,t)=0$
for $x\in\partial\varOmega$ and for $t=T$. The energy estimate for $u_k$ looks as follows:
\begin{equation}\label{3.5}
\esup_{t\in[0,T]}\|u_k(\cdot,t)\|^2+\int_0^T\|\nabla u_k(\cdot,t)\|^2\,dt
+\int_0^T \int_\varOmega \varphi_k(\zeta_k(x))\, u_k^2(x,t)\, dx\, dt
\le C_1,
\end{equation}
where the constant $C_1$ is the same as in \eqref{2.2} and \eqref{3.1}. This estimate implies that the sequence $\{u_k\}$
has a subsequence (denoted again by $\{u_k\}$) which converges $*$-weakly in $L^\infty(0,T;L^2(\varOmega))$
and weakly in $L^2(0,T;H^1_0(\varOmega))$ to a function $u$. Our goal is to pass to the limit in
\eqref{3.4} as $k\to\infty$. Clearly,
$$
\lim_{k\to\infty}\int_0^T \int_\varOmega \big(u_k\,\partial_t h - \nabla u_k\cdot\nabla h\big)\,dx\,dt
=\int_0^T \int_\varOmega \big(u\,\partial_t h - \nabla u\cdot\nabla h\big)\,dx\,dt.
$$
The only problem is to prove that
\begin{equation}\label{3.6}
\lim_{k\to\infty}\int_0^T \int_\varOmega \varphi_k(\zeta_k)\,u_k\, h\,dx\,dt=
\int_0^T \int_\varOmega \varphi(\zeta)\,u\, h\,dx\,dt.
\end{equation}
Notice that we also need to justify that the right-hand side in the last relation makes sense.
\medskip\noindent
\textsc{Step~2.}
The reasoning at the end of the proof of Theorem~\ref{t2.1} implies that
$$
\zeta_k\to\zeta\quad\text{in $L^1(\varOmega)$ as $k\to\infty$.}
$$
Consequently, there exists a subsequence (denoted again by $\{\zeta_k\}$) such that
\begin{equation}\label{3.7}
\zeta_k\to\zeta\quad\text{almost everywhere in $\varOmega$ as $k\to\infty$}.
\end{equation}
Due to \eqref{3.5}, the following estimate holds for all $k\in \mathbb{N}$:
\begin{multline}\label{3.8}
\int_\varOmega \varphi_k(\zeta_k)\,\zeta_k^2\, dx=
\int_\varOmega \varphi_k(\zeta_k)\,\Big(\int_0^T \alpha (\cdot,t)\, u_k(\cdot,t)\,dt\Big)^2\, dx
\\
\le \int_\varOmega \varphi_k(\zeta_k)\Big(\int_0^T \alpha^2 (\cdot,t)\,dt \int_0^T u_k^2(\cdot,t)\,dt\Big) dx
\le C_6,
\end{multline}
where $C_6=C_1\,\|\alpha\|^2_{L^\infty(\varOmega; L^2(0,T))}$ and $C_1$ is the constant from \eqref{3.5}.
As a consequence of the Fatou lemma, this estimate together with the continuity of $\varphi$ and \eqref{3.7} implies that
\begin{equation}\label{3.9}
\int_\varOmega \varphi(\zeta)\,\zeta^2\, dx\le C_6
\end{equation}
which is a part of \eqref{3.2}. The rest of the assertions in \eqref{3.2} are simple consequence of \eqref{3.9}.
We can also prove now that $\zeta\in L^2(\varOmega)$, and namely that
\begin{equation}\label{3.10}
\int_\varOmega \zeta^2\, dx \le TC_6,
\end{equation}
where $C_6$ is the constant from estimate \eqref{3.8}. Indeed, exactly as in \eqref{3.8}, we obtain that
$$
\int_\varOmega \zeta_k^2\, dx
\le \int_\varOmega \int_0^T \alpha^2 (\cdot,t)\,dt \int_0^T u_k^2(\cdot,t)\,dt\, dx
\le T \,\|\alpha\|^2_{L^\infty(\varOmega; L^2(0,T))} \esup_{t\in[0,T]}\|u_k(\cdot,t)\|^2
\le TC_6,
$$
and \eqref{3.10} follows from the Fatou lemma. Notice that we can put the constant
from the Poin\-care inequality instead of $T$ in \eqref{3.10}, if we estimate
$\int_\varOmega u_k^2\, dx$ by $\int_\varOmega |\nabla u_k|^2\, dx$ and use again \eqref{3.5}.
\medskip\noindent
\textsc{Step~3.}
In this step, we prove \eqref{3.6}, which completes the proof of the weak solvability of the problem.
Let $h:\varOmega_T\to \mathbb{R}$ be an arbitrary smooth bounded function and $R=\max_{(x,t)\in\varOmega_T}|h(x,t)|$.
We will prove that
\begin{equation}\label{3.11}
\varphi_k(\zeta_k)\int_0^T u_k\,h\,dt \to \varphi(\zeta)\int_0^T u\,h\,dt\quad\text{as $k\to\infty$ in $L^1(\varOmega)$},
\end{equation}
which clearly implies \eqref{3.6} and, as a consequence, the weak solvability of problem~\eqref{1.1}--\eqref{1.3}.
To do this, we will apply the Vitali convergence theorem (see, e.g., \cite[Sec.~4.8.7]{MP}).
Exactly as for the sequence $\{\zeta_k\}$ at the end of the proof of Theorem~\ref{t2.1},
we can establish (selecting, if necessary, a subsequence) that $\int_0^T u_k\,h\,dt\to \int_0^Tu\,h\,dt$ in
$L^2(\varOmega)$ and almost everywhere in $\varOmega$ as $k\to\infty$.
Besides that, due to \eqref{3.7},
\begin{equation}\label{3.12}
\varphi_k(\zeta_k)\to \varphi(\zeta)\quad\text{as $k\to\infty$ almost everywhere in $\varOmega$.}
\end{equation}
Really, let $E\subset \varOmega$ be the set such that $\zeta_k(x)\to \zeta(x)$ as $k\to \infty$
for all $x\in E$. As it follows from \eqref{3.7}, $\mu_n(\varOmega\setminus E)=0$,
where $\mu_n$ is the $n$-dimensional Lebesgue measure.
For an arbitrary $x\in E$, there exists $k_{x,1}\in \mathbb{N}$ such that $\varphi(\zeta(x))+1 < k_{x,1}$.
Since the function $\varphi$ is continuous, there exists $k_{x,2}\in \mathbb{N}$ such that
$|\varphi(\zeta_k(x))-\varphi(\zeta(x))|<1$ for all $k> k_{x,2}$. Therefore,
$\varphi(\zeta_k(x))< k$ and, as a consequence, $\varphi_k(\zeta_k(x))=\varphi(\zeta_k(x))$
whenever $k>\max\{k_{x,1},k_{x,2}\}$. Due to the continuity of $\varphi$,
we conclude that $\varphi_k(\zeta_k(x))=\varphi(\zeta_k(x))\to \varphi(\zeta(x))$ as $k\to\infty$ for all $x\in E$.
Thus,
$$
\varphi_k(\zeta_k)\int_0^T u_k\,h\,dt \to \varphi(\zeta)\int_0^T u\,h\,dt\quad\text{
almost everywhere in $\varOmega$ as $k\to\infty$}.
$$
According to the Vitali theorem, in order to prove \eqref{3.11}, it is necessary to establish that the sequence
$\{\varphi_k(\zeta_k)\int_0^T u_k\,h\,dt\}$ is uniformly integrable on $\varOmega$.
Let $E$ be an arbitrary measurable subset of $\varOmega$, $E_T=E\times [0,T]$, and
$G_M^k=\{(x,t)\in E_T\; |\; |u_k(x,t)|\ge M\}$. Then
$$
\int_{G_M^k}\varphi_k(\zeta_k)|u_k\,h|\,dx\,dt\le \frac{R}{M}\int_{G_M^k}\varphi_k(\zeta_k)\,u_k^2\,dx\,dt
\le \frac{R}{M}\int_{\varOmega_T}\varphi_k(\zeta_k)\,u_k^2\,dx\,dt\le \frac{C_1R}{M},
$$
where $C_1$ is the constant from \eqref{3.5}. Besides that,
$$
\int_{E_T\setminus G_M^k}\varphi_k(\zeta_k)|u_k\,h|\,dx\,dt\le M R\int_{E_T\setminus G_M^k}\varphi_k(\zeta_k)\,dx\,dt
\le MR\,T \int_{E}\varphi_k(\zeta_k)\,dx.
$$
Therefore,
\begin{equation}\label{3.13}
\Big|\int_{E}\varphi_k(\zeta_k)\int_0^T u_k\,h\,dt\,dx\Big|
=\Big|\int_{E_T}\varphi_k(\zeta_k)\, u_k\,h\,dx\,dt\Big|
\le\frac{C_1R}{M}+MR\, T\int_{E}\varphi_k(\zeta_k)\,dx.
\end{equation}
Let us estimate $\int_{E}\varphi_k(\zeta_k)\,dx$.
If $E_N^k=\{x\in E\; |\; |\zeta_k(x)|> N\}$, where $N\in (0,\infty)$, then \eqref{3.8} implies that
$$
\int_{E_N^k}\varphi_k(\zeta_k)\,dx\le
\frac{1}{N^2}\int_{E_N^k}\varphi_k(\zeta_k)\zeta_k^2\,dx
\le \frac{C_6}{N^2}.
$$
Since the function $\varphi$ is continuous, there exists a constant $\gamma_N>0$ such that
$\varphi(\xi)\le \gamma_N$ for all $\xi\in [-N,N]$. Therefore,
$$
\int_{E\setminus E_N^k}\varphi_k(\zeta_k)\,dx\le \int_{E\setminus E_N^k}\varphi(\zeta_k)\,dx
\le \gamma_N \mu_n(E).
$$
Thus,
$$
\int_{E}\varphi_k(\zeta_k)\,dx\le \frac{C_6}{N^2}+\gamma_N \mu_n(E)
$$
and, as it follows from \eqref{3.13},
$$
\Big|\int_{E}\varphi_k(\zeta_k)\int_0^T u_k\,h\,dt\,dx\Big|\le
\frac{C_1R}{M}+\frac{MR\, T C_6}{N^2}+\gamma_N MR\, T \mu_n(E)
$$
For an arbitrary $\varepsilon>0$, we take $M$ and $N$ such that $C_1 R/M< \varepsilon/3$ and
$MR\, T C_6/N^2<\varepsilon/3$. If $\delta= \varepsilon/(3\gamma_N MR\, T)$,
then, as a consequence of the last inequality,
$$
\Big|\int_{E}\varphi_k(\zeta_k)\int_0^T u_k\,h\,dt\,dx\Big|<\varepsilon\quad\text{whenever $\mu_n(E)<\delta$},
$$
which implies the required uniform integrability of the sequence $\{\varphi_k(\zeta_k)\int_0^T u_k\,h\,dt\}$.
Thus, \eqref{3.11} is proven.
\medskip\noindent
\textsc{Step~4.}
In this step, we prove \eqref{3.1} and \eqref{3.3}, which completes the proof of the theorem.
The fact that $\varphi(\zeta) u^2\in L^1(\varOmega_T)$ is a consequence of \eqref{3.1}.
In order to establish \eqref{3.1}, we have to pass to the limit as $k\to\infty$ in \eqref{3.5}.
The first two terms on the left-hand side of \eqref{3.1} are obtained in the usual way due to the weak lower
semicontinuity of the norms. Let us obtain the third term.
Using the H\"older inequality, we get:
\begin{multline*}
\int_\varOmega \Big|\nabla\int_0^T u_k^2\,dt\Big|\, dx
\le 2\int_{\varOmega_T} |u_k|\,|\nabla u_k|\,dx\,dt
\\
\le 2 \Big(\int_{\varOmega_T} |u_k|^2\,dx\,dt\Big)^{1/2}\,\Big(\int_{\varOmega_T} |\nabla u_k|^2\,dx\,dt\Big)^{1/2}
\le 2T^{1/2}C_1,
\end{multline*}
where $C_1$ is the constant from \eqref{3.5}.
This estimate implies that there exists a function $v\in L^1(\varOmega)$ such that (up to a subsequence)
$$
\int_0^T u_k^2\, dt\to v\quad\text{as $k\to \infty$ in $L^1(\varOmega)$ and almost everywhere in $\varOmega$.}
$$
Therefore, due to \eqref{3.12}, $\varphi(\zeta_k)\int_0^T u_k^2\,dt\to \varphi(\zeta)\,v$
almost everywhere in $\varOmega$. As it follows from \eqref{3.5} and the Fatou lemma,
$\varphi(\zeta)\,v\in L^1(\varOmega)$ and
$$
\int_\varOmega \varphi(\zeta)\,v\,dx\le \liminf_{k\to\infty}
\int_0^T\int_\varOmega\varphi_k(\zeta_k)\, u_k^2\, dx\,dt.
$$
Thus, in order to establish \eqref{3.1}, it suffices to show that
$$
\int_0^T\int_\varOmega\varphi(\zeta)\, u^2\, dx\,dt\le \int_\varOmega \varphi(\zeta)\,v\,dx.
$$
Since $\varphi(\zeta)\ge 0$, it is enough to prove that
$$
\int_0^T u^2\,dt\le v\quad\text{almost everywhere in $\varOmega$.}
$$
Let $G$ be an arbitrary measurable subset of $\varOmega$ and $G_T=G\times [0,T]$.
Since $u_k$ converges to $u$ weakly in $L_2(G_T)$ and the norm in $L^2(G_T)$ is weakly lower semicontinuous,
$$
\int_G \int_0^T u^2\,dt\,dx\le \liminf_{k\to\infty} \int_G \int_0^T u_k^2\,dt\,dx=
\int_G v\,dx.
$$
Due to the arbitrariness of $G$, we have the required relation.
Finally, the assertion that $\varphi(\zeta) u\in L^1(\varOmega_T)$ follows from the H\"older inequality and the facts
that $\varphi(\zeta)\in L^1(\varOmega)$ and $\varphi(\zeta) u^2\in L^1(\varOmega_T)$.
The theorem is proven.
\hfill$\square$
Notice that Remark~\ref{t2.2} is also true for Theorem~\ref{t3.1}, since we used only the energy estimate
in its proof.
\bigskip\noindent
\textbf{Acknowledgement} \\
This work was supported by the Russian Science Foundation (Grant No. 19-11-00069).
|
\section{Introduction}
\label{sec:1}
Second harmonic generation is the simplest fundamental process of nonlinear wave physics, which is also in the center of numerous practical applications in laser physics and beyond. The dynamics of the process has been studied exhaustively \cite{SG1,SG2}, which cannot be said about statistics, despite the fact that understanding the influence of noise on the energy conversion is of paramount practical importance, recently enhanced by the use of meta-materials \cite{SGH1}. Here we address this problem by studying theoretically a two-mode resonant system driven by a combination of pumping and random noise. Our motivation is two-fold. Apart from the classical conversion problem, we find this system ideally suited for elucidating the fundamental problems of non-equilibrium physics. When one mode is stochastically forced and another is dissipated, that presents a minimal model of turbulence cascade. The freedom to force either mode allows us to elucidate the basic differences between direct and inverse cascades. Apart from energy, we shall be interested in the entropy of such far-from equilibrium state, which is expected to be much lower than in thermal equilibrium with the same energy.
A remarkable property, common for all systems of resonantly interacting waves and shared with hydrodynamic systems \cite{VSF}, is that the canonical thermal equilibrium has exactly Gaussian statistics, and the modes fluctuate independently, regardless of the interaction strength.
Here we describe how deviations from equilibrium diminish entropy and build correlations between the two modes. Far from equilibrium the joint two-mode statistics is never close to Gaussian, even when the marginal distribution of every mode is close to Gaussian. On the one hand, the entropy decrease means that the statistical distribution is getting more non-uniform, which poses the question: Can it lead all the way to singularity like the celebrated Sinai-Ruelle-Bowen measures in dynamical systems \cite{D,FS}? We show that this is indeed so: the measure in the phase space is getting singular in the double limit of strong non-equilibrium and weak interaction. On the other hand, since entropy is missing information, any entropy decrease poses another question: Where all this extra information about non-equilibrium is encoded? First, we find out how the entropies of the three marginal distributions, of each mode amplitude and their phase difference, go down as the system deviates from equilibrium. Second, we find out which part of the entropy decrease is due to inter-mode correlation. This is properly measured by the mutual information (rather than by the pair correlation function, suitable for Gaussian statistics only).
The process of the second harmonic generation is described by the following model Hamiltonian (assuming perfect resonance)
\begin{align} {\cal H}_0 =\omega |a_1|^2 +2\omega |a_2|^2 + V a_1^{*2}a_2+ V^*a_1^{ 2}a_2^* \ .\label{Ham5}\end{align}
Here $a_1$ and $a_2$ are the complex amplitudes of two non-linearly coupled modes having frequencies $\omega$ and $2\omega$, respectively, and $V$ is the interaction constant (considered real positive without loss of generality).
The two coupled complex equations govern dynamics: $\dot a_k=-i {\partial {\cal H}_0}/{\partial a_k^*}$, $k=1,2$.
We eliminate the linear terms in these equations by introducing the envelopes
\begin{equation}\label{env}
b_1=a_1e^{i\omega t}, b_2=a_2e^{2i\omega t}.
\end{equation}
That results in a strongly interacting system with a cubic Hamiltonian ${\cal H} = V b_1^{*2}b_2+ V^*b_1^{ 2}b_2^* $.
Due to the symmetry $b_1\to b_1e^{i\phi}$, $b_2\to b_2e^{2i\phi}$ the system $\dot b_k=-i {\partial {\cal H}}/{\partial b_k^*}$ has an extra integral of motion $N= |b_1|^2 +2 |b_2|^2$ and is completely integrable; the phase portrait is presented in Appendix~\ref{sec:Ham}.
Let us add dissipation and stochastic pumping:
\begin{eqnarray}
\label{gen00}
&&\dot b_1=-2i V^*b_1^*b_2-\gamma_1b_1+ \xi_1(t),\\
&&\dot b_2=-i Vb_1^2-\gamma_2 b_2+ \xi_2(t).
\label{gen0}\end{eqnarray}
Here $\gamma_1$ and $\gamma_2$ are the damping coefficients, and $\xi_1$ and $\xi_2$ are independent Gaussian random forces with zero mean $\langle \xi_i(t)\rangle=0$ and the variance $\langle \xi_i(t_1) \xi_j^*(t_2)\rangle=P_i\delta_{ij}\delta(t_2-t_1) $.
We mainly focus on the properties of the statistically steady solutions of the system (\ref{gen00},\ref{gen0}) in the case when one mode is forced, while the other is damped.
Since the modes enter the Hamiltonian in a non-symmetric way, there are two possibilities: one either can pump the first (lower frequency) mode and dump the second (higher frequency) mode or vice versa.
The former scenarios qualitatively corresponds to the direct energy cascade, while the second is reminiscent to the inverse cascade.
We wish to understand how much information is needed in order to build a turbulent state and how much one learns about one mode by observing another. For that we will use the metrics from information theory: entropy and mutual information. The answer to the first question is given by the decrease in entropies
\begin{eqnarray} &S_{12}=-\int db_1db_1^* db_2db_2^*\,\rho(b_1,b_2)\ln\rho(b_1,b_2)\,,\nonumber \\&S_1=-\int db_1db_1^*\, \rho(b_1 )\ln \rho(b_1 )\,,\nonumber \\&S_{ 2}=-\int db_2db_2^*\,\rho( b_2)\ln\rho( b_2)\,,\nonumber\end{eqnarray}
where $\rho$ is either full or marginal probability distribution. The answer to the second question is given by the mutual information between the modes:
\begin{equation} I_{12}=S_1+S_2-S_{12}\ .\label{MI}\end{equation}
Fig.\ref{tm_mi_ext} demonstrates the growth of the mutual information versus the degree of non-equilibrium (an analog of the Reynolds number defined below, see (\ref{chi})).
As one of the simplest model of energy transfer, the system of two coupled oscillators has received considerable attention in the literature~\cite{Kumar_2008,Bonetto_2004,Ciliberto_2013,Falasco_2015,Chun_2015,
Mura_2018,Horowitz_2019,Courant1,Courant}. In particular, in the mathematical literature, one finds an analysis of a two-mode system with a quadratic Hamiltonian ${\cal H} = T a_1^{*2}a_2^2$ with the purpose to get insight into the energy transfer in wave turbulence \cite{Courant1,Courant}.
What distinguishes our model is that it directly corresponds to physical reality and allows experimental validation. In addition, an asymmetry between the modes allows us to compare direct and inverse cascades, which turn out quite different. Another distinction is that we add entropic and informational consideration to the energetic analysis.
\begin{figure}[h!]
\includegraphics[width=50mm]{tm_icdc_mi_v2.pdf} \centering
\caption{
Mutual information versus Reynolds number for direct and inverse cascades (red and blue lines respectively). Three dimensional
distribution are computed with bin size $\Delta \theta = 2\pi/32$ and $\Delta\rho^2_{1,2}/n_{1,2} = 1$ (circles) and 0.5 (diamonds).
}
\label{tm_mi_ext}
\end{figure}
The paper is organized from the point of view of entropy: we start from the maximal-entropy equilibrium and investigate near-equilibrium states in Section~\ref{eq}. We then move to study the noise-driven direct and inverse cascades in Section \ref{sec:turb}. We define a dimensionless parameter akin to the Reynolds number and show how entropy decreases as the Reynolds number increases. We begin with the limit of small Reynolds number in \ref{smallRe} and continue to large Reynolds number in \ref{sec:largeRe}, where entropy decreases all the way down as the measure becomes singular in the limit of infinite Reynolds number. In Section~\ref{sec:laser} we consider an instability-driven first harmonic and study the influence of noise on the conversion process, this can serve as a simple model for a laser generating second harmonic. Conclusion~\ref{sec:con} briefly lists our main results.
\section{Near Thermal equilibrium}
\label{eq}
Adopting the language of stochastic thermodynamics, one can call the ratios $P_1/\gamma_1\equiv T_1$ and $2P_2/\gamma_2\equiv T_2$ effective temperatures experienced by two modes $b_1$ and $b_2$ which are governed by the Langevin equations (\ref{gen00}) and (\ref{gen0}).
If $\Delta T=T_1-T_2=0$, then it is straightforward to find from the Fokker-Planck equation or entropy maximum the steady-state joint probability distribution:
\begin{equation} {\cal P}_0={1\over Z} \exp\biggl(-\frac{2|b_1|^2+4|b_2|^2}{T}\biggr)={1\over Z} \exp\biggl(-\frac{2N}{T}\biggr).\label{SG2}\end{equation}
Despite strong interaction, this distribution is exactly Gaussian and the modes are statistically independent.
The later means that the mean energy flux and the mutual information between modes are zero. Thermal equilibrium corresponds to the equipartition of the quadratic invariant: $\langle |b_1|^2\rangle\equiv n_1=2n_2\equiv 2\langle |b_2|^2\rangle$.
What can we say about the system's statistics when modes are subject to different effective temperatures?
Let us introduce the dimensionless measure of non-equilibrium
\begin{equation}
\sigma=\frac{\Delta T}{T},\ \ \text{where} \ \ T=\frac{P_1+2P_2}{2(\gamma_1+\gamma_2)}.
\end{equation}
Another dimensionless parameter quantifies interaction strength relative to the dissipation:
\begin{equation}\chi={(\gamma_1+\gamma_2)^3\over (P_1+2P_2)|V|^2}\ .\label{chi}\end{equation}
\begin{figure*}
\includegraphics[width=175mm]{tm_near_eq_v2.pdf}\centering
\caption{
%
Top row: deviation of entropies from equilibrium ($\Delta T = 0$) and mutual information
for $\gamma_1 = \gamma_2$ and different strength of interaction marked by color.
Equilibrium is Gaussian with
$S_{{\rm eq}1} = 1/\ln(2)$,
$S_{{\rm eq}2} = 1/\ln(2) - 1$,
and
$S_{{\rm eq}\theta} = \log_2(2 \pi)$.
Dashed lines show the entropy deviation of a single mode respectively at $T_1$ and $T_2$
for marginal distributions,
$\Delta S_1 = \log_2 (1 + \Delta T/4T) $ and
$\Delta S_2 = \log_2 (1 - \Delta T/4T)$ .
The entropies of individual modes are affected by the change of variance of (nearly Gaussian)
distribution and by deviation from the Gaussian shape. For close-to-equilibrium cases presented here
the first effect is dominant.
%
Bottom row: panels (e) and (f) show deviation of entropies from the entropies of Gaussian distributions with the same variance;
panel (g) illustrates the quadratic dependence of the entropy of phase distribution,
panel (h) illustrates the quadratic dependence for mutual information
(solid lines for $\Delta T > 0$; dashed lines for $\Delta T <0$). The offset 0.005 is the effect of
finite ensemble and bin size; the offset decreases as the size of data set and resolution improve.
%
3D distributions are computed with bin sizes $2\pi/32$ for the phase and
$0.1 T$ for $\rho_{1,2}$.
}
\label{fig_s12}
\label{fig_mi}
\end{figure*}
Denote $\rho_{1,2}=|b_{1,2}|$ and $\theta=\arg(b_1^2b_2^*)$. From (\ref{gen00}) and (\ref{gen0}), the steady-state equations on the second moments read:
\begin{eqnarray}
\label{moments1}
&&-4V\langle \rho_1^2\rho_2\sin\theta\rangle -2\gamma_1\langle\rho_1^2\rangle+P_1=0,\\
\label{moments2}
&&2V\langle \rho_1^2\rho_2\sin\theta\rangle -2\gamma_2\langle\rho_2^2\rangle+P_2=0,
\end{eqnarray}
The time derivative of the real part of the third moment is given by: $d\langle {\cal H}\rangle/dt=-(2\gamma_1+\gamma_2)\langle \cal H\rangle$, since $\langle \xi_1b_1b_2^*\rangle=\langle \xi_2^*b_1^2\rangle=0$. Therefore, in any steady state, either in thermal equilibrium or out of it, one has
\begin{equation}
\label{moments3}\langle {\cal H}\rangle=2V\langle\rho_1^2\rho_2\cos\theta\rangle=0.
\end{equation}
Equations (\ref{moments1})-(\ref{moments3}) are valid for any values of $\sigma$ and $\chi$.
{At $\sigma\ne0$, the probability density ${\cal P}(b_1,b_2)$ is non-Gaussian in non-equilibrium, yet it is close to Gaussian when $|\sigma|\ll1$ for all values of $\chi$.
The simplest to treat is the limit of small interaction, $\chi \gg1$.
In this case, the first correction to Eq. (\ref{SG2}) is determined by the energy flux between modes, which is small and proportional to the temperature difference:}
\begin{eqnarray}
\label{pert4}
&&\ln{\cal P}(b_1,b_2) \approx - { 2|b_1|^2\over T_1}-{ 4 |b_2|^2 \over T_2}-\\
&&-{4\Delta T\over P_1T_2+P_2T_1}\Im [V^*b_1^{*2}b_2]+O(\chi^{-2})
\end{eqnarray}
Smallness of interaction multiplies the parameter of non-equilibrium $\Delta T/T$ in the right hand side of Eq. (\ref{pert4}), so that this result is valid even when $\Delta T/T$ is not small.
That means that, as long as both temperatures remain finite and interaction is weak, even far from equilibrium the relative entropy is small:
$$D({\cal P}|{\cal P}_0)=\int db_1db_1^*db_2db_2^*{\cal P}\ln({\cal P}/{\cal P}_0)\propto \chi^{-1}\ll1\ ,$$
as well as the mutual information.
In the opposite limit $\chi\ll1$ or $V\to\infty$, the non-Gaussian correction is again proportional to the product of the degree of non-equilibrium and the small parameter $\chi$. In terms of $x=|b_1|^2$ and $y=2|b_2|^2$ we obtain:
$$\ln{\cal P} \approx- {x +y \over T}+{\chi^{1/2}\Delta T\over T }f(x,y,\theta)\ ,$$
where the correction satisfies the equation
$${2x {y}\over\sqrt{yT}}\left[\sin\theta\left({\partial\over\partial y}-{\partial\over\partial x}\right)
+{x-2y\over2xy} \cos\theta{\partial\over\partial \theta}\right]f =x-y\,.$$
The correction is odd in phase difference, $f(-\theta)=-f(\theta)$, and scales linearly with amplitudes, so that it is substantial at small amplitudes.
In the limits, $f\to g(\theta)\sqrt{y/2T}$, where $g=\sin\theta$ at $y\ll x$, and $g=\int d\theta/\cos\theta$ at $y\gg x$.
It makes sense to compare entropies at the same mean quadratic energy $N$. To see how entropy goes down on the way to turbulence we shall subtract the total entropy from its maximal equilibrium value, which quantifies the amount of information one needs to create a turbulent state: $\Delta S(N,n_2/n_1)=S_0- S_{12}$.
Numerics support quadratic decrease of $S_{12}(\Delta/T)$ and increase of $I_{12}(\Delta/T)$ up to $\Delta T\simeq 4T$, see Figure~\ref{fig_s12}.
When $\Delta T/T$ exceeds one, the functions are not even which demonstrates the statistical difference between upward and downward energy conversion. We see stronger deviations from Gaussianity for negative $\Delta T<0$, which corresponds to the downward energy flow and to an inverse cascade at $\Delta T/T \to- \infty$. The physical difference is that the first mode pumps the second one as an additive force, while the second mode pumps the first one as a multiplicative instability. Therefore, it seems natural that the entropy is generally lower and the mutual information higher for an inverse transfer. The analysis of the separate distributions of two amplitudes and the relative phase shows that the entropy of the driven mode (say, $S_1(\Delta/T)$ for a direct transfer) grows with $\Delta T$ slower than the entropy of the dissipated mode and $S_\theta(\Delta/T)$ decrease, see panels (g) and (h) in Figure \ref{fig_mi}.
\section{Turbulent cascades}\label{sec:turb}
Now let us have an energy cascade in our model: pumping one mode and dissipating another. When energy flows from lower frequency mode to higher, i.e $0=\xi_2=\gamma_1$ in Eqs. (\ref{gen00}) and (\ref{gen0}), the cascade is called direct, and inverse when $0=\xi_1=\gamma_2$.
In these cases, the only dimensionless parameter is $\chi=\gamma^3/(P|V|^2)$, where $P$ is the intensity of noise acting upon the driving mode, and $\gamma$ denotes the damping coefficient of the dissipating mode. As we shall see below, $\chi$ to some extent plays the role of the Reynolds number of hydrodynamics in a sense that it determines how low is the entropy and how much the occupation numbers deviate from the equipartition $n_1=2n_2$,
even though the system is not close to thermal equilibrium for however small or large $\chi$.
Balance of the quadratic invariant, $N$, means that the dissipating mode keeps the magnitude of order of its equilibrium value: $n_2= P/4\gamma$ for the direct cascade and $n_1=P/\gamma$ for the inverse cascade. How much the mode which is pumped exceeds the equipartition value is determined by the value of $\chi$, as described below. Note that this parameter can be interpreted as the squared ratio of the dissipation rate $\gamma$ and the nonlinear transfer rate $Vn\simeq V\sqrt{P/\gamma}$.
When $\chi$ is small, the interaction between modes is strong and the energy transfer is fast, so that the occupation numbers are expected to be close to equipartition, yet the statistics is not expected to be close to separable Gaussian form given by (\ref{SG2}). Even though the noise is weak, it is white, that is a singular perturbation destroying integrability everywhere in the phase space \cite{Kurchan}; we shall see below how non-trivial the probability distribution is already in this limit.
One may naively expect that in the opposite limit of large $\chi$, when the noise is strong and interaction is weak, the correlation between modes would be weak too. We shall show below that the opposite is true far from equilibrium: the necessity to carry the flux makes the modes strongly correlated precisely because of a strong noise and weak interaction. It is in this limit we find the lowest entropy and the maximal mutual information between modes, as well as appearance of singular measure in phase space.
\begin{figure*}.
\includegraphics[width=175mm]{tm_icdc_prob_v2.pdf}
\caption{ Probabilities of the occupation numbers and the phase for:
(a, b) inverse cascade, small $\chi$;
(c, d) direct cascade, small $\chi$;
(e, f) inverse cascade, large $\chi$;
(g, h) direct cascade, large $\chi$.
Each dataset contains 20M datapoints, at $\Delta t = 0.01$ for inverse cascade and large $\chi$
and $\Delta t = 0.1$ for all other cases. For the inverse cascade, $\chi = \gamma_1^3 / (2 P_2 V^2)$ and $\nu_1 = P_2/\gamma_1$.
We use $\gamma_{1,2} = 0.01$ for small $\chi$ and $\gamma_{1,2} = 1$ for large $\chi$. In all cases, $V=1$. Broken lines in (g) correspond to the approximation (\ref{invweakP2},\ref{invweakP1}).
}
\label{ICDCprob}
\end{figure*}
\subsection{Small Reynolds number: strong-interaction-weak-noise limit}
\label{smallRe}
\subsubsection{Inverse cascade}
In an inverse cascade, energy goes from high to low frequency, so we set $\gamma_2=0$ and $P_1=0$ in (\ref{gen00},\ref{gen0}):
\begin{eqnarray}
\label{inv1}
&&\dot b_1=-2i V^*b_1^*b_2-\gamma b_1,\\
\label{inv1_0}
&&\dot b_2=-i Vb_1^2+\xi(t).
\end{eqnarray}
In the steady state, the energy input rate $P$ must be equal to the dissipation rate $\gamma n_1$ and to the energy flux from the second mode to the first given by the imaginary part of the third cumulant: $2V\langle \rho_1^2\rho_2\sin\theta\rangle = -P$. Then, from the energy balance we obtain $n_1=P/\gamma$ and ${2V}\langle \rho_1^2\rho_2\sin\theta\rangle =-{P}$,
so that $\langle \rho_1^2\rho_2\rangle \ge {P}/{2V}$.
Also, from the condition $\frac{d}{dt}\langle\ln |b_1|^2\rangle=0$ we find
$2V\langle{\rho_2}\sin\theta\rangle=- {\gamma}$
and, therefore $\langle{\rho_2}\rangle\ge{\gamma}/ {2V} $.
From Eqs. (\ref{inv1},\ref{inv1_0}) it is straightforward to see that when $\chi\ll 1$
the steady-state probability distribution ${\cal P}(b_1,b_2)$ cannot be close to the equilibrium Gaussian (\ref{SG2}) with the temperature $T=2P/\gamma$ and the equipartition $P/\gamma=n_{1}=2n_{2}$.
Indeed, the stationarity of $ \langle {\cal H}^2\rangle=4|V|^2\langle\rho_1^4\rho_2^2\cos^2\theta\rangle$ requires $\langle {\cal H}^2\rangle/\langle \rho_1^4\rangle= \frac{|V|^2P}{2\gamma}$, but
this contradicts the Gaussian ratio which gives $\langle {\cal H}^2\rangle/\langle \rho_1^4\rangle=\frac{\left|V\right|^{2}P}{\gamma}$.
Thus, small value of $\chi$ does not mean that the system is near equipartition.
In agreement with this conclusion, results of numerical modeling presented at two lower left panels of Figure \ref{ICDCprob} show that at neither of marginal distributions of the mode amplitudes is close to Gaussian and that the phase distribution does not become uniform as $\chi$ tends to zero.
This is also reflected in nonzero value $I_{12}(+0)$ of mutual information between modes in this limit,
see the right panel of Figure~\ref{tm_mi_ext}.
\subsubsection{Direct cascade}\label{sec:smalldir}
Direct cascade corresponds to the choice $\gamma_1=0$ and $P_2=0$ in (\ref{gen00},\ref{gen0}):
\begin{eqnarray}
\label{bb1_0}
&&\dot b_1=-2i V^*b_1^*b_2+\xi(t),\\
&&\dot b_2=-i Vb_1^2-\gamma b_2.
\label{bb1}\end{eqnarray}
Now energy goes from low to high frequency.
Again, in the steady-state regime, the energy input rate $P/2$ must be equal to the dissipation rate $2\gamma n_2$ and to the energy flux from the first mode to the second given by the imaginary part of the third cumulant:
$2V\langle \rho_1^2\rho_2\sin\theta\rangle = P/2$, so $n_2=P/4\gamma$.
From $\frac{d}{dt}\langle\ln |b_2|^2\rangle=0$ we find
$\langle \rho_1^2 \sin\theta /\rho_2\rangle= {\gamma}/{V}$. Therefore $\langle \rho_1^2\rho_2\rangle \ge{P}/{4V}$ and $\langle{\rho_1^2}/{\rho_2}\rangle\ge{\gamma}/{V}$.
When $\chi\to0$, the dimensionless flux $ \langle \rho_1^2\rho_2\sin\theta\rangle/n_1n_2^{1/2}=\chi^{1/2}$ is small, which may suggest that phase-space distribution is close to the Gaussian equilibrium (\ref{SG2}) with $T=P/\gamma$
and that the phase distribution is close to uniform. Furthermore, as opposed to
the case of inverse cascade discussed above, the equality obtained from the
stationarity of $\langle {\cal H}^2\rangle$:
\begin{equation} \langle {\cal H}^2\rangle=4|V|^2\langle\rho_1^4\rho_2^2\cos^2\theta\rangle=\frac{2|V|^2 P}{\gamma}\langle \rho_1^2\rho_2^2\rangle\ ,\label{K2dir}\end{equation}
is achieved by the Gaussian distribution with $n_1=2n_2=P/2\gamma$.
However, numerical data, as can be seen from two upper left panels of Figure~\ref{ICDCprob}, shows that even though the marginal distributions of amplitudes are close to Gaussian with equipartition, $n_1\approx2n_2$, the phase distribution is far from flat and deviation from equilibrium is substantial.
The mutual information between modes as a function of $\chi$ exhibits a non-zero value of $I_{12}(+0)$ (see Figure~\ref{tm_mi_ext}) which is also a clear footprint of non-equilibrium.
\subsection{Large Reynolds number: weak interaction, strong noise limit}
\label{sec:largeRe}
\subsubsection{Inverse cascade}\label{sec:inv}
The pair of complex equations (\ref{inv1},\ref{inv1_0}) can be rewritten as three real ones since the overall phase drops out:
\begin{eqnarray}
\label{inv001}
&& \dot \rho_1= - 2|V|\rho_1\rho_2\sin\theta-\gamma \rho_1,\\
\label{inv002}
&&\dot \rho_2= |V|\rho_1^2\sin\theta+{P\over4\rho_2}+{\zeta(t)\over\sqrt2},\\
\label{inv003}
&&\dot\theta=|V|\frac{\rho_1^2-4\rho_2^2}{\rho_2}\cos\theta+{\frac{\zeta(t)}{\sqrt{2}\rho_2}},
\label{inv}\end{eqnarray}
where $\zeta(t)$ is the real white noise with zero mean $\langle \zeta(t)\rangle=0$ and the pair correlation function $\langle \zeta(t_1)\zeta(t_2)\rangle=P\delta(t_1-t_2)$.
When $\chi\gg 1$, Eqs. (\ref{inv001})-(\ref{inv003}), can be further simplified by assuming that relative phase is locked on $\theta=-\pi/2$ most the time.
Then, one gets the following closed equations for the amplitudes dynamics
\begin{eqnarray}
\label{inv200}
&&\dot \rho_1= 2|V|\rho_1\rho_2-\gamma \rho_1,\\
&&\dot \rho_2=-|V|\rho_1^2+{P\over4\rho_2}+{\zeta(t)\over\sqrt2}.
\label{inv2}
\end{eqnarray}
A hypothesis that the modes are statistically independent in this limit is shown incorrect in the Appendix~\ref{sec:ap2}. This result is in sharp contrast with the model described in \cite{Courant1}, where
authors found the factorized joint probability density ${\cal P}(\rho_1,\rho_2)$ of mode amplitudes in the limit when their analogue of the parameter $\chi$ is large.
While constructing the probability densities for inverse cascade at $\chi\to\infty$ turns out to be a tricky task, it is straightforward to describe general features of stochastic dynamics dictated by Eqs. (\ref{inv200}) and (\ref{inv2}).
Namely, these pair of nonlinearly coupled equations suggest the following cyclical evolution: $\rho_1$ stays close to zero most of the time while $\rho_2$ undergoes diffusion in a repulsive logarithmic potential; when $\rho_2$ sufficiently outgrows the threshold level $\gamma/2|V|$, $\rho_1$ shoots up and quickly diminishes $\rho_2$; after that $\rho_1$ also resets to the near-zero level and the stochastic dynamics of $\rho_2$ starts from scratch.
The mode dynamics during the intermittent burst events can be described by simplified equations
\begin{eqnarray}
\label{noise_free1}
\dot{\rho}_1=2|V|\rho_1\rho_2-\gamma\rho_1,\\
\label{noise_free2}
\dot{\rho}_2=-|V|\rho_1^2.
\end{eqnarray}
Compared with Eqs. (\ref{inv200}) and (\ref{inv2}), we neglected the terms associated with noise.
\begin{figure}[t]
\centerline{\includegraphics[width=60mm]{tm_burst.pdf}}
\caption{An individual realization
of the modes trajectories during one of the events. The continues lines are obtained from numerical simulations of Eqs. (\ref{inv1}) and (\ref{inv1_0}, and the dashed lines represent theoretical fit based on Eqs. (\ref{burst_rho1}) and (\ref{burst_rho2})). }
\label{pic:burst}
\end{figure}
Equations (\ref{noise_free1}) and (\ref{noise_free2}) are exactly solvable yielding
\begin{equation} {\rho}_1^2(t)= {r_1^2-2\left(\rho_2(t)-\frac{\gamma}{2 V }\right)^2+2\left(r_2-\frac{\gamma}{2|V|}\right)^2}\,,\label{burst_rho1}\end{equation}
\begin{widetext}
\begin{eqnarray}
\label{burst_rho2}
&&{\rho}_2(t)=\frac{\gamma}{2V}+\sqrt{\left(r_2-\frac{\gamma}{2V}\right)^2+\frac{r^2_1}{2}} \tanh\left[\frac12\ln\frac{\sqrt{\bigl(r_2-
\frac{\gamma}{2V}\bigr)^2+\frac{r^2_1}{2}}+r_2-\frac{\gamma}{2V}}{\sqrt{\bigl(r_2-
\frac{\gamma}{2V}\bigr)^2+\frac{r^2_1}{2}}-r_2+\frac{\gamma}{2V}}
-2Vt\sqrt{\Bigl(r_2-\frac{\gamma}{2V}\Bigr)^2+\frac{r^2_1}{2}} \right],
\end{eqnarray}
\end{widetext}
where $r_1=\rho_1(0)$, $r_2=\rho_2(0)$ are the initial conditions.
Estimating $r_2\sim {\gamma/V}$ and $r_1\ll r_2$, we see from Eq. (\ref{burst_rho2}) that the duration of such burst event is $\sim \gamma^{-1}$, which is much smaller than the typical inter-events period $\sim {\gamma^2}/{PV^2}$.
As follows from Eq. (\ref{burst_rho1}), the amplitude of the dissipated mode grows from the initial value $r_1\ll r_2$ to the maximum value $\rho_{1\text{max}}=\sqrt{r_1^2+2(r_2-\frac{\gamma}{2V})^2}\approx \sqrt2(r_2-\frac{\gamma}{2V})$ (attaining it at the moment when $\rho_2(t)=\frac{\gamma}{2V}$) and finally returns to the starting level $\rho_1(\infty)=r_1$. Such bursts are likely responsible for pulses running in shell models, which are chains of interacting triplets.
Numerical simulations confirm the intermittent nature of system dynamics described above.
Namelly, panel (f) of Figure \ref{ICDCprob} reveals that the relative phase is indeed locked at $-\pi/2$.
Figure \ref{pic:burst} illustrates that analytical prediction based on the assumption of phase locking (see Eqs. (\ref{burst_rho1}) and (\ref{burst_rho2})) are in excellent agreement with numerical data extracted from simulations of Eqs. (\ref{inv1}) and (\ref{inv1_0}).
As visible in panel (e) of Figure \ref{ICDCprob}, the tails of the amplitudes probability distributions strongly depend on $\chi$; the fits of ${\cal P}(\rho_1)$ and ${\cal P}(\rho_2)$ by the broken lines in Figure \ref{ICDCprob} are empirical.
As for the mutual information,
from Figure~\ref{tm_mi_ext} we see that $I_{12}\propto \ln\chi$ for inverse cascade with $\chi\gg1$.
\subsubsection{Direct cascade}
We conclude the treatment of our turbulent cascades with the case of a direct cascade in the limit of large Reynolds number, $\chi\rightarrow\infty$. It is in this limit we find a window to the way a singular measure is formed far away from equilibrium. We indeed find that in this limit the full probability distribution is singular and thus corresponds to the lowest entropy state $S_{12}\rightarrow-\infty$.
\label{sec:dir}
In addressing the weak-interaction-strong-noise limit $\gamma\to\infty$ ($\chi\rightarrow\infty$), it is convenient to
express $b_2$ from (\ref{bb1}) as an integral, which
in the leading order shows that the second amplitude is enslaved to the first one:
\begin{equation}
\label{bb3}
b_2(t)=-iV\int_{-\infty}^tb_1^2(t')e^{\gamma(t'-t)}dt'\approx -\frac{iVb_1^2}{\gamma}.
\end{equation}
Substituting this relation into (\ref{bb1_0}), we get a closed equation on the stochastic dynamics of $b_1$
\begin{equation} \dot b_1= -\frac{2V^2}{\gamma}b_1^*b_1^2+\xi(t),\label{SG4}\end{equation}
from which one finds the following expressions for the marginal probability distributions
\begin{align}
{\cal P}(b_1)&=Z_1^{-1}\exp\biggl(-\frac{2 V^2}{\gamma P}|b_1|^4\biggr),
\label{invweakP1_01}\\
{\cal P}(b_2)&=Z_2^{-1}\exp\biggl(-\frac{2\gamma}{P}|b_2|^2\biggr),
\label{invweakP2}
\end{align}
which are valid at $\rho_1,\rho_2\ll {\gamma}/{|V|}$. Thus, the whole probability density in the four-dimensional phase space is singular at $\chi\to\infty$, sitting on a three-dimensional manifold
\begin{equation}\!\!\! {\cal P}(b_1,b_2)=\frac{2}{\sqrt{\pi P\gamma}}\exp\biggl(-\frac{4V^2|b_1|^4}{P\gamma}\biggr)\delta(b_2-\frac{iV}{\gamma}b_1^2)\,,\label{SG5}\end{equation}
so that the total entropy $S_{12}\to-\infty$.
Note also that Eq. (\ref{SG5}) yields large ratio of the typical mode amplitudes: $\rho_1^2/\rho_2^2\simeq \sqrt{\gamma^3/P|V|^2}=\sqrt\chi\gg1$.
Since the distribution over the overall phase is flat, one can integrate it out and conclude that the distribution in the the three-dimensional space of variables $\rho_1,\rho_2,\theta$ concentrates on the
curve $\rho_2\propto\rho_1^2$.
Interestingly, with increasing $\chi$ the joint probability distribution ${\cal P}(b_1,b_2)$ is getting sharper than Gaussian along this curve. This is different from the model discussed in \cite{Courant1} where the driving mode is nearly Gaussian in this limit, the relative amplitudes of both modes fluctuate, and only the relative phase is fixed, that is the joint probability density is only singular with respect to the phase difference $\theta$.
Considering large but finite $\chi$, instead of delta-function in Eq. (\ref{SG5}) one obtains the distribution with a finite width which is the variance of the difference $b_2-Vb_1^2/\gamma$.
To estimate this width we further expand Eq. (\ref{bb3})
\begin{align}
\label{bb3_0}
b_2(t)\approx&-i V\int_{-\infty}^t\left[b_1^2(t)+(t'-t){d b_1^2(t)\over dt}\right]e^{\gamma(t'-t)}dt'\\
=&-{i Vb_1^2(t)\over \gamma}+{2i Vb_1(t) \over \gamma^2}{db_1(t)\over dt}.
\end{align}
From Eq. (\ref{bb3_0}) we get
\begin{equation}\langle |b_2+iVb_1^2/\gamma|^2\rangle={4|V|^2\over \gamma^4}\langle| b_1\dot b_1|^2\rangle= {4|V|^2P^2 \over \gamma^4}\ .\label{width}\end{equation}
Dividing this result by
$\langle |b_2|^2\rangle=P/\gamma$, one obtains that the relative squared width behaves as $1/\chi$
(as expected, it tends to zero when $\chi\to\infty$).
The non-zero width at finite values of $\chi$ entails the finite entropy of the distribution ${\cal P}(b_1,b_2)$: $S_{12}\simeq -\ln \chi$.
This analytical prediction is supported by Figure~\ref{tm_mi_ext}.
For comparison, the naive Gaussian ansatz yields $S_1+S_2\simeq \ln n_1n_2\propto \ln \chi^{-1/2}$, since $n_2=P/4\gamma$ and $n_1\simeq \sqrt{P\gamma}/|V|$.
Now let us plug Eq. (\ref{bb3}) into Eq.(\ref{bb1}), then solving the corresponding steady Fokker-Planck equation (see \ref{sec:apLD}) one arrives at the next order correction for the marginal probability distribution for $\rho_1\ll {\gamma}/{|V|}$:
\begin{equation}
{\cal P}(\rho_1)\propto \rho_1(1-\frac{4|V|^2}{\gamma^2}\rho_1^2)\exp(-\frac{2|V|^2}{\gamma P}\rho_1^4+\frac{16|V|^4}{3\gamma^3P}\rho_1^6)\,.
\label{invweakP1}
\end{equation}
which is more accurate than Eq. (\ref{invweakP1_01}). Unfortunately, extracting similar correction for the probability density ${\cal P}(\rho_2)$ as well as the further high order corrections to ${\cal P}(\rho_1)$ is more challenging.
As can be seen from panel (g) of Figure \ref{ICDCprob}, Eqs. (\ref{invweakP2}) and (\ref{invweakP1}) allow us to fit the numerical data quite accurately in the range $\rho_1,\rho_2\ll {\gamma}/{|V|}$.
Expectedly, the agreement between numeric and analytic results improves with the growth of $\chi$.
\section{Laser generation}\label{sec:laser}
Let us now pump the first harmonic by an instability, for instance, in a laser. Consider first pure dynamics, setting $\xi_1=\xi_2=0$ in (\ref{gen00},\ref{gen0}) and changing sign in front of $\gamma_1$, which now describes gain for an optical signal.
Then the resulting evolution satisfies three closed equations:
\begin{eqnarray}
\label{SG600}
&&{d\rho_1^2 \over dt}=-2F+2\gamma_1\rho_1^2,\\
\label{SG60}
&&{d\rho_2^2\over dt}=F-2\gamma_2\rho_2^2,\\
\label{SG6}
&& {dF\over dt}=(2\gamma_1-\gamma_2)F+2|V|^2\bigl( \rho_1^4-4\rho_1^2\rho_2^2\bigr),
\end{eqnarray}
where $F=2V\rho_1^2\rho_2\sin\theta$ is the instantaneous flux.
Apart from the trivial unstable fixed point $\rho_1=\rho_2=0$, Eqs. (\ref{SG600})-(\ref{SG6}) have the stationary point $\bar\rho_1^2=\gamma_1\gamma_2/2|V|^2$, $\bar\rho_2^2=\gamma_1^2/4|V|^2$, $ \bar \theta=\pi/2$ (and, thus, $\bar F=\gamma_2\gamma_1^2/2|V|^2$).
This means that in the the degenerate case $\gamma_2=2\gamma_1$, the system possesses the steady state $\rho_1=2\rho_2$ existing for any $\theta$. This marginal stability turns into an instability of the steady state at $\gamma_2<2\gamma_1$ and into stability at $\gamma_2>2\gamma_1$.
In what follows, we consider $\gamma_2>2\gamma_1$.
Note that in practice both $\gamma_1$ and $\gamma_2$ often depend on the amplitudes, for instance, due to gain saturation or/and nonlinear damping. However, our main focus here is on the noise impact on the steady state, so we will treat $\gamma_1,\gamma_2$ taken near this state as constants.
{
Let us now add a random pumping and study its influence on the efficiency and statistics of conversion.
The modes amplitudes, $\rho_1$ and $\rho_2$ and the relative phase $\theta$ are governed by the following equations
\begin{eqnarray}
&&\dot \rho_1= - 2V\rho_1\rho_2\sin\theta+\gamma_1\rho_1 +{P\over4\rho_1}+\frac{\zeta_1(t)}{\sqrt{2}},\\
\label{laser2}
&&\dot \rho_2= V\rho_1^2\sin\theta-\gamma_2 \rho_2\,,\\
&& \dot\theta={\rho_1^2-4\rho_2^2\over\rho_2}V\cos\theta+{\sqrt2\zeta_2(t)\over\rho_1}.
\label{laser2'}
\end{eqnarray}
Here $\zeta_1$ and $\zeta_2$ are two independent real white noises with zero mean values $\langle \zeta_i(t)\rangle=0$, and the pair correlator $\langle \zeta_i(t_1)\zeta_j(t_2)\rangle=P\delta_{ij}\delta(t_1-t_2)$.
In the limit of weak noise, $P|V|^2\ll\gamma_1^2(\gamma_2-2\gamma_1)$, one can
apply a linear approximation near the fixed point.
More specifically, we substitute decomposition $\rho_1(t)=\bar\rho_1+u(t),\rho_2(t)=\bar\rho_2+v(t), \theta(t)=\bar \theta+\phi(t)$ into Eqs. (\ref{laser2})-(\ref{laser2'}) and keep only the first order terms with respect to $u$, $v$ and $\phi$.
This procedure yields
\begin{eqnarray}
\label{laser3'}
&&\dot u= - 2V\bar\rho_1v +{P\over4\bar\rho_1}+\frac{\zeta_1(t)}{\sqrt{2}},\\
&&\dot v= 2V\bar\rho_1u-\gamma_2 v,\\
&&\dot\phi={(2\gamma_{1}-\gamma_{2})}\phi+{\sqrt{2}\zeta_2(t)\over\bar\rho_1}.\label{laser3}
\end{eqnarray}
From Eqs. (\ref{laser3'})-(\ref{laser3}) we immediately find the
variances
\begin{eqnarray}
&& \langle u^2\rangle- \langle u\rangle^2=\frac{\left(2\gamma_{1}+\gamma_{2}\right)P}{8\gamma_1\gamma_2}\,,\\
&&
\langle v^2\rangle-\left\langle v\right\rangle ^{2}=\frac{P}{4\gamma_{2}}\,,\\
&&\langle \phi^2\rangle-\langle \phi\rangle^2=\frac{2V^{2}P}{\gamma_{1}\gamma_{2}\left(\gamma_{2}-2\gamma_{1}\right)}.
\end{eqnarray}
We see that level of fluctuations in relative phase grows when one approaches the stability threshold. Note also that far from the threshold, i.e. at $2\gamma_1\ll\gamma_2$, the fluctuations of the second harmonic are suppressed: $\langle v^2\rangle/\langle u^2\rangle\approx \gamma_1/\gamma_2\ll1$. In this case, the noise of the first harmonic only weakly influences the conversion into the second one. However, the conversion is least effective in this limit: $\bar\rho_2^2/\bar\rho_1^2=\gamma_1/2\gamma_2\ll1$.
\section{Conclusion}
\label{sec:con}
Our most important finding is the explicit formula (\ref{SG5}) for the singular measure of a direct cascade in the limit of strong noise and weak interaction. We believe that this is a meaningful advance in non-equilibrium statistics, as it opens a window to the study of the formation of singular measures in systems driven far away from equilibrium. We have described also the approach to this limit and have shown that the total entropy decays and the inter-mode mutual information grows logarithmically with the Reynolds number. In the inverse cascade case in this limit, the phase is locked on $-\pi/2$ and the system exhibits an intermittent dynamics of bursts, which we were able to describe analytically. Such bursts are perhaps responsible for pulses running in shell models, which are chains of interacting triplets used in modeling hydrodynamic incompressible turbulence. It is thus may be interesting to apply the methods developed here to the popular shell model $\dot u_i=u_{i-1}^2-u_iu_{i+1}$ \cite{DN}. After some elementary transformations, this model can be turned into that with the Hamiltonian ${\cal H}=\sum_iV_i(a_i^2a_{i+1}^*+c.c)$, that is the interacting chain built of our pairs.
The opposite limit of weak noise and strong interaction is a singular one: the probability distribution is not close to a Gaussian distribution determined by $N$ for however small $\chi$, despite occupation numbers being close to equipartition and the marginal one mode distributions close to quadratic. It is expressed, in particular, in the nonzero mutual information $I_{12}(\chi)$ at the limit $\chi\to+0$. Figure~\ref{tm_mi_ext} combines the mutual information data for both cascades. We see that $I_{12}(\chi)-I_{12}(+0)\propto \chi^2$ at $\chi\ll1$. We failed to find an analytic solution in this limit either in the direct or inverse cascade, even though it is likely that the probability distribution can be expressed in terms of $N $ and ${\cal H}^2$, which are the conserved quantities of the unforced undamped system.
We conclude with suggesting an interesting application of our model to wave turbulence. In a set of $M+2$ interacting waves, one may consider to model the interaction of a resonant couple with the other $M$ waves by dissipation and random forcing.
When $M\gg1$ we can treat forces from all other modes as a white noise, so that our model (\ref{gen0}) applies. In this case, different limits in $\chi$ correspond to different situations. If we assume an almost continuous distribution of other modes and estimate from the wave kinetic equation $\gamma\simeq V^2Mn/\omega$ and $P\simeq \gamma n$ \cite{ZLF}, then $\chi=\gamma^3/PV^2\simeq V^2Mn/\omega^2\ll1$, which is the original parameter of nonlinearity assumed to be small. In this case, we come to the surprising conclusion that a resonant mode within turbulence, when $\Delta T\simeq T$, has a relative entropy of order unity and independent of $V$. If, however, we have a set of well-isolated resonant interactions, then it makes more sense to assume that the interaction with a given mode is $M$ times smaller so that $\chi$ is large (as $M$ or $\sqrt{M}$), then the relative entropy is small. Note that in most cases the number of resonant interactions,
is much less than the total number of modes in the system.
The work was supported by the Scientific Excellence Center and Ariane de Rothschild Women Doctoral Program at WIS, grant 662962 of the Simons foundation, grant 075-15-2019-1893 by the Russian Ministry of Science, grant 873028 of the EU Horizon 2020 programme, and grants of ISF, BSF and Minerva. NV was in part supported by NSF grant number DMS-1814619. This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by NSF grant number ACI-1548562, allocation DMS-140028.
\nocite{*}
\bibliographystyle{unsrt}
|
\section{Introduction}
Bounded cohomology of topological spaces, introduced by Gromov \cite{Gr}, is a variant of singular cohomology which has deep connections with geometry
and group theory. The bounded cohomology of a topological space $X$ (with coefficients in $\mathbb{R}$) is defined by restricting to those singular $\mathbb{R}$-cochains $\phi \in C^n(X; \mathbb{R})$, $n \geq 0$, whose values $\{\phi(\sigma) \ | \ \Delta^n \xrightarrow{\sigma} X\}$ in $\mathbb{R}$ form a bounded subset.
This seemingly small modification turns out to have significant consequences and leads to a theory which differs substantially from
singular cohomology. On the one hand, the singular $\mathbb{R}$-cochain complex
$$C^{\bullet}(-; \mathbb{R}) \colon \mathrm{Top}^{\mathrm{op}} \to \mathrm{Ch}_{\mathbb{R}}, \ X \mapsto C^{\bullet}(X; \mathbb{R}),$$
preserves weak equivalences and is excisive, i.e., the functor $C^{\bullet}(-;\mathbb{R})$ sends weak homotopy equivalences to quasi-isomorphisms, and homotopy colimits of topological spaces to homotopy limits of cochain complexes. These fundamental properties (which are also true for general coefficients) allow us to determine and compute the singular cohomology of $X$ from the singular cohomologies of the spaces in a homotopy colimit decomposition of $X$. On the other hand, the functor of bounded $\mathbb{R}$-cochains
$$C^{\bullet}_b(-; \mathbb{R}) \colon \mathrm{Top}^{\mathrm{op}} \to \mathrm{Ch}_{\mathbb{R}}, \ X \mapsto C^{\bullet}_b(X; \mathbb{R}),$$
preserves weak equivalences \cite{Iv}, but it is not excisive. The failure to satisfy excision makes bounded cohomology difficult to compute in general.
\smallskip
The inclusion of bounded $\mathbb{R}$-cochains into all singular $\mathbb{R}$-cochains defines a natural comparison map of cochain complexes
$$c_X \colon C^{\bullet}_b(X; \mathbb{R}) \rightarrow C^{\bullet}(X; \mathbb{R})$$
and corresponding comparison maps after passing to cohomology
$$c_X^n \colon H^n_b(X; \mathbb{R}) \rightarrow H^n(X; \mathbb{R}).$$
The study of the properties of these comparison maps to singular cohomology is an important tool in bounded cohomology theory and its applications. Moreover, these comparison maps yield information about the simplicial volume \cite{Gr}: if $X$ is an oriented closed connected $n$-manifold, then the simplicial volume $||X||$ of $X$ vanishes if and only if $c^n_X$ is the zero map.
Gromov \cite{Gr} and Ivanov \cite{Iv, Iv2} established interesting properties of the comparison maps $\{c^n_X\}_{n \geq 0}$ in the case where $X$ admits an open cover with suitable properties (see also \cite{FM, LS}). We recall that an open cover $\mathcal{U} = \{U_i\}_{i \in I}$ of a topological space $X$ is called \emph{amenable} if the image of the homomorphism $\pi_1(U_i, x) \to \pi_1(X, x)$ is amenable for every $i \in I$ and $x \in U_i$; we do not assume here that $U_i$ is path-connected. For any open cover $\mathcal{U} = \{U_i \}_{i \in I}$ of $X$, we denote by $\mathcal{I}_{\mathcal{U}}$ the poset of finite subsets $\sigma \subseteq I$ such that $U_{\sigma} : = \bigcap_{i \in \sigma} U_i \neq \varnothing$, ordered by reverse inclusion of subsets. Let $|\mathcal{I}_{\mathcal{U}}|$ denote the classifying space of $\mathcal{I}_{\mathcal{U}}$ -- this is the geometric realization of the nerve of $\mathcal{I}_{\mathcal{U}}$, that is, the simplicial complex associated to the poset of chains in $\mathcal{I}_{\mathcal{U}}$.
\begin{theorem}[Gromov \cite{Gr}, Ivanov \cite{Iv}] \label{nerve_theorem}
Let $X$ be a path-connected topological space and let $\mathcal{U} = \{U_i\}_{i \in I}$ be an amenable open cover of $X$. Then the following hold:
\begin{itemize}
\item[(1)](Covering theorem) Suppose that $U_{\sigma}$ is path-connected for every $\sigma \in \mathcal{I}_{\mathcal{U}}$. Then there are $\mathbb{R}$-linear maps
$$\phi^n \colon H^n_b(X; \mathbb{R}) \rightarrow H^n(|\mathcal{I}_{\mathcal{U}}|; \mathbb{R}), \ \ n \geq 0,$$
such that the following diagrams commute:
$$
\xymatrix{
H^n_b(X; \mathbb{R}) \ar[r]^{c^n_X} \ar[rd]_{\phi^n} & H^n(X; \mathbb{R}). \\
& H^n(|\mathcal{I}_{\mathcal{U}}|; \mathbb{R}) \ar[u]_{H^n(p_{F_{\mathcal{U}}}; \mathbb{R})}
}
$$
\item[(2)](Vanishing theorem) If the dimension of $|\mathcal{I}_{\mathcal{U}}|$ is $\leq m-1$, then the comparison map $c^n_X$ vanishes for $n \geq m$.
\end{itemize}
\end{theorem}
Theorem \ref{nerve_theorem}(2) is originally due to Gromov \cite{Gr}; several proofs of this result have appeared in the literature (see Ivanov\cite{Iv}, Frigerio--Moraschini \cite{FM} and L\"oh-Sauer \cite{LS}). Theorem \ref{nerve_theorem}(1) is originally due to Ivanov (see \cite{Iv}); alternative proofs of this result can also be found in \cite{FM, LS}. More recently, Ivanov \cite{Iv2} obtained generalizations of Theorem \ref{nerve_theorem}(1) to more general classes of covers.
\smallskip
In this paper we present another approach to Theorem \ref{nerve_theorem} which also leads to stronger versions of the theorem and generalizations. Let us outline the general idea of this approach in the specific context of Theorem \ref{nerve_theorem}. Note that for any open cover $\mathcal{U} = \{U_i \}_{i \in I}$ of $X$, there is an associated diagram of topological spaces
$$F_{\mathcal{U}} \colon \mathcal{I}_{\mathcal{U}} \to \mathrm{Top}, \ \sigma \mapsto U_{\sigma}: = \bigcap_{i \in \sigma} U_i.$$
A fundamental property of $F_{\mathcal{U}}$ is that it is a homotopy colimit diagram, i.e., the canonical map $$\mathrm{hocolim}_{\mathcal{I}_{\mathcal{U}}} F_{\mathcal{U}} \rightarrow X$$
is a weak homotopy equivalence \cite{DI}. In addition, for the constant diagram at the one-point space $\ast$,
$$\delta(\ast) \colon \mathcal{I}_{\mathcal{U}} \to \mathrm{Top}, \ \sigma \mapsto \ast,$$
there is a canonical weak homotopy equivalence
$\mathrm{hocolim}_{\mathcal{I}_{\mathcal{U}}} \delta(\ast) \simeq |\mathcal{I}_{\mathcal{U}}|.$
The natural transformation $F_{\mathcal{U}} \to \delta(\ast)$ yields the canonical map (that appears in Theorem \ref{nerve_theorem})
$$p_{F_{\mathcal{U}}} \colon X \simeq \mathrm{hocolim}_{\mathcal{I}_{\mathcal{U}}} F_{\mathcal{U}} \to \mathrm{hocolim}_{\mathcal{I}_{\mathcal{U}}} \delta(\ast) \simeq |\mathcal{I}_{\mathcal{U}}|.$$
Based on the naturality of $c_X$ and using the fact that singular cohomology is excisive, it is possible to obtain factorizations of $c_X$ from natural factorizations of the maps $U_{\sigma} \subseteq X$. In the case of Theorem \ref{nerve_theorem}, the key assumption that $\mathcal{U}$ is amenable implies that for each inclusion $U_{\sigma} \subseteq X$, $\sigma \in \mathcal{I}_{\mathcal{U}}$, the natural Moore--Postnikov factorization (at $\pi_1$)
$$U_{\sigma} \to V_{\sigma} \to X$$
has the property that the bounded cohomology of $V_{\sigma}$ is concentrated in degree $0$ -- this is a consequence of Gromov's fundamental \emph{Mapping Theorem} \cite{Gr, Iv, FM}. Moreover, if $U_{\sigma}$ is path-connected, then $V_{\sigma}$ has the (bounded) cohomology of the point. These properties yield information about the homotopy limit of the diagram of bounded cochain complexes $\sigma \mapsto C^{\bullet}_b(V_{\sigma}; \mathbb{R})$. In particular, passing to the homotopy limit (for $\sigma \in \mathcal{I}_{\mathcal{U}}$) in the diagrams of cochain complexes
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[d] \ar@/^3pc/[rrdd]^{c_X} & & \\
C^{\bullet}_b(V_{\sigma}; \mathbb{R}) \ar[r]^{c_{V_{\sigma}}} & C^{\bullet}(V_{\sigma};\mathbb{R}) \ar[d] & \\
& C^{\bullet}(U_{\sigma};\mathbb{R}) & C^{\bullet}(X;\mathbb{R}) \ar[l]
}
$$
we obtain a canonical factorization of $c_X$ with the required properties. This approach to Theorem \ref{nerve_theorem}, based on homotopy colimits and the properties of $c_X$, leads to refinements of Theorem \ref{nerve_theorem} in several directions:
\begin{itemize}
\item[(a)] The context of Theorem \ref{nerve_theorem} extends to general homotopy (or $\mathbb{R}$-cohomology) colimit decompositions of $X$. For each such homotopy colimit diagram that is equipped with a factorization (in the sense of Section \ref{sec_factorizations}), there is a corresponding canonical factorization of $c_X$ (Proposition \ref{factorization_general}).
Moreover, for factorizations of the homotopy colimit diagram which satisfy certain vanishing properties, our generalized versions of the \emph{Covering theorem} (Theorem \ref{nerve_theorem_general1}) and of the \emph{Vanishing theorem} (Theorem \ref{nerve_theorem_general2}) show specific properties of the induced factorization of $c_X$ analogous to Theorem \ref{nerve_theorem}. The precise connection with Theorem \ref{nerve_theorem} is discussed in Section \ref{amenable_covers}.
\item[(b)] The method yields canonical factorizations of the comparison map $c_X$ in the homotopy theory of cochain complexes before passing to cohomology.
Working directly with cochain complexes provides the statements with more structure and allows for new variations of the results by direct application of homotopy-theoretic constructions. In Section \ref{pullback_fact}, we discuss some examples of this in the context of \emph{parametrized factorizations} (Subsections \ref{general_parametrized} and \ref{pullback_covers}) as well as for relative bounded cohomology (Subsection \ref{subsec:rel-cov-van}).
\item[(c)] The same method applies similarly to factorizations of homotopy (or $\mathbb{R}$-cohomology) colimit decompositions which satisfy bounded acyclicity only in appropriate ranges of degrees. Specifically, our version of the \emph{Covering theorem} (Theorem \ref{nerve_theorem_general1}) treats also \emph{boundedly $k$-acyclic factorizations} and our version of the \emph{Vanishing theorem} (Theorem \ref{nerve_theorem_general2}) applies also to \emph{boundedly $k$-truncating factorizations}. Amenable covers give rise to $0$-truncating factorizations (see Section \ref{amenable_covers}). In view of the renewed interest and the recent advances in bounded acyclicity, these technical refinements of Theorem \ref{nerve_theorem} could potentially prove useful for future applications.
\end{itemize}
As the general method is essentially based only on the general properties of the comparison map $c_X$, similar arguments can also be used for other such examples of natural transformations. The main ingredient of our approach is to view the comparison map as a natural map of cochain complexes and use its homotopy-theoretic properties in order to identify it as the \emph{coassembly map for bounded cohomology} (see Subsection \ref{sec:comparison_map}). For another related example where the same method can be applied, let us mention the dual context of the comparison map for $\ell^1$-homology
$$H_*(X; \mathbb{R}) \to H^{\ell^1}_*(X;\mathbb{R})$$
which can be treated similarly and the proofs are essentially the same. Further generalizations are discussed briefly in Subsection \ref{generalizations}.
Since several proofs of Theorem \ref{nerve_theorem} (and its variations) are available in the literature, this paper makes no great claim to originality. The main purpose of the paper is to show how a specific homotopy-theoretic viewpoint offers another perspective on Theorem \ref{nerve_theorem} and leads naturally to stronger versions of this result with short proofs; moreover, we hope that the methods might be of further interest in the study of the homotopy-theoretic aspects of bounded cohomology and its applications.
\medskip
\noindent \textbf{Acknowledgements.} I would like to thank Clara L\"oh and Marco Moraschini for many interesting discussions about bounded cohomology and for their useful comments on this work. This work was partially supported by \emph{SFB 1085 -- Higher Invariants} (Universit\"at Regensburg) funded by the DFG.
\section{Preliminaries} \label{sec:prelim}
We briefly review the context of the comparison map for bounded cohomology using the language $\infty$-categories; the theory of $\infty$-categories is
not necessary for obtaining our results, but it will allow us here, as in many other of its applications, to identify and focus on the essential homotopy-theoretic features of the problem. We refer the interested reader to \cite[Ch. 1]{HTT} for an introduction to $\infty$-categories.
\subsection{Spaces and chain complexes} Let $\mathcal{S}$ denote the $\infty$-category of spaces (see, for example, \cite{HTT}). It will be convenient to use
the following model for $\mathcal{S}$. Let $\mathrm{Top}$ denote the (ordinary) category of topological spaces and let $\mathcal{W}_{\mathrm{whe}}$ be the class of weak homotopy equivalences in $\mathrm{Top}$. We define
$$\mathcal{S} : = \mathrm{Top}[\mathcal{W}_{\mathrm{whe}}^{-1}]$$
to be the $\infty$-categorical localization of $\mathrm{Top}$ at $\mathcal{W}_{\mathrm{whe}}$ \cite[Ch. 7]{Ci}, \cite{Hi}. There is no substantive difference between the approach using $(\mathrm{Top}, \mathcal{W}_{\mathrm{whe}})$ (as a category with weak equivalences or a model category) and the associated $\infty$-category $\mathcal{S}$.
Let $\mathrm{Ch}_{\mathbb{R}}$ denote the (ordinary) category of chain complexes of $\mathbb{R}$-vector spaces and let $\mathcal{W}_{\mathrm{qi}}$ be the class of quasi-isomorphisms in $\mathrm{Ch}_ {\mathbb{R}}$. We consider similarly the model for the derived $\infty$-category $\mathcal{D}(\mathbb{R})$ that is given by the $\infty$-categorical localization of $\mathrm{Ch}_{\mathbb{R}}$ at $\mathcal{W}_{\mathrm{qi}}$:
$$\mathcal{D}(\mathbb{R}) : = \mathrm{Ch}_{\mathbb{R}}[\mathcal{W}_{\mathrm{qi}}^{-1}].$$
\subsection{(Co)limit diagrams} Given a simplicial set $\mathcal{I}$ (e.g., the nerve of a small category), we denote by $\mathcal{I}^{\triangleright} := \mathcal{I} \star \Delta^0$ the associated simplicial set (or the nerve of the associated category) that is defined by joining a terminal object $\star$ to $\mathcal{I}$.
A cone on a diagram $F \colon \mathcal{I} \to \mathcal{S}$ is a diagram $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}$ extending $F$. A colimit diagram $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}$ is a cone on $F = \widetilde{F}_{|\mathcal{I}} \colon \mathcal{I} \to \mathcal{S}$ which is initial in the $\infty$-category $\mathcal{S}_{F/}$ of cones on the diagram $F$. Limits in $\mathcal{S}$ (or in any other $\infty$-category) are defined dually. For a colimit diagram $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}$ of $F$, we will sometimes refer to the cone object $\widetilde{F}(\star) = X \in \mathcal{S}$ as the colimit of $F$ and denote this by $\mathrm{colim}_{\mathcal{I}} F$.
Every small diagram in $\mathcal{S}$ (or in $\mathcal{D}(\mathbb{R})$) admits a (co)limit (see \cite{Ci, HTT}).
We say that a cone $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}$ is an \textit{$\mathrm{H}\mathbb{R}$-colimit diagram} on $F = \widetilde{F}_{|\mathcal{I}} \colon \mathcal{I} \to \mathcal{S}$ if the canonical map in $\mathcal{S}$
$$\mathrm{colim}_{\mathcal{I}} F \to \widetilde{F}(\star)$$
induces an isomorphism in singular $\mathbb{R}$-cohomology. A colimit diagram is clearly an $\mathrm{H}\mathbb{R}$-colimit diagram.
Given a diagram $F \colon \mathcal{I} \to \mathcal{S}$, a (colimit, $\mathrm{H}\mathbb{R}$-colimit) diagram $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}$ which extends $F$ and has cone object $X \in \mathcal{S}$ determines (and is determined by) its adjoint diagram in the $\infty$-category $\mathcal{S}_{/X}$ of spaces over $X$, also denoted by
$$\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X).$$
We will call a diagram in $\mathcal{S}_{/X}$ that arises in this way from a colimit diagram in $\mathcal{S}$ (resp. $\mathrm{H}\mathbb{R}$-colimit diagram) again a \emph{colimit diagram} (resp. $\mathrm{H}\mathbb{R}$\emph{-colimit diagram}).
We refer to \cite{Ci, HTT} for a detailed exposition of the properties of (co)limits in higher category theory. There is no essential difference between such colimit diagrams and homotopy colimit diagrams in $\mathrm{Top}$ in the standard sense (see \cite[Ch. 4]{HTT} and \cite[Ch. 7]{Ci} for an analysis of this comparison). We also recommend \cite{Du} and \cite{DHKS} for the theory of homotopy (co)limits in homotopical algebra.
\begin{example} For a simplicial set $\mathcal{I}$, we denote the associated space ($\infty$-groupoid) by $|\mathcal{I}|$; this has the weak homotopy type of (the classifying space of) $\mathcal{I}$ (when $\mathcal{I}$ corresponds to the nerve of an ordinary category). The colimit of the constant diagram at the one-point/contractible space,
$$\delta(\ast) \colon \mathcal{I} \to \mathcal{S}, \ \ i \mapsto \ast,$$ is canonically identified with $|\mathcal{I}|$, i.e., $\mathrm{colim}_{\mathcal{I}} \delta(\ast) \simeq |\mathcal{I}|.$
\noindent For every diagram $F \colon \mathcal{I} \to \mathcal{S}$, there is an obvious natural transformation $F \rightarrow \delta(\ast)$ which induces the canonical map:
$$p_F \colon \mathrm{colim}_{\mathcal{I}} F \to \mathrm{colim}_{\mathcal{I}} \delta(\ast) \simeq |\mathcal{I}|.$$
\end{example}
\begin{example}[Open covers] \label{covering-colim} Let $\mathcal{U} = \{U_i \}_{i \in I}$ be an open cover of a topological space $X$. For a subset $\sigma \subseteq I$, we write $U_{\sigma} := \bigcap_{i \in \sigma} U_i \subseteq X$. Let $\mathcal{I}_{\mathcal{U}}$ denote the (nerve of the) poset of finite subsets $\sigma \subseteq I$ such that $U_{\sigma} \neq \varnothing$, ordered by reverse inclusion of subsets. Passing to the $\infty$-category of spaces, there is an associated diagram:
$$\widetilde{F}_{\mathcal{U}} \colon \mathcal{I}_{\mathcal{U}} \to \mathcal{S}_{/X}, \ \sigma \mapsto (U_{\sigma} \xrightarrow{i_{\sigma}} X).$$
It is a fundamental property of the diagram $\widetilde{F}_{\mathcal{U}}$ that it defines a colimit diagram, that is, the canonical map
\begin{equation*} \label{di}
\mathrm{colim}_{\sigma \in \mathcal{I}_{\mathcal{U}}} U_{\sigma} \rightarrow X
\end{equation*}
is an equivalence in $\mathcal{S}$ (see \cite{DI}).
\end{example}
\subsection{Singular cohomology} Let $C^{\bullet}(-;\mathbb{R}) \colon \mathrm{Top}^{\mathrm{op}} \to \mathrm{Ch}_{\mathbb{R}}$ denote the usual singular $\mathbb{R}$-cochain complex functor. It will be convenient here to regard the singular cochain complex as a chain complex concentrated in nonpositive degrees. Since $C^{\bullet}(-;\mathbb{R})$ sends weak homotopy equivalences to quasi-isomorphisms, we obtain canonically an induced functor
$$C^{\bullet}(-; \mathbb{R}) \colon \mathcal{S}^{\mathrm{op}} \to \mathcal{D}(\mathbb{R}).$$
It follows from classical results that this functor is \emph{excisive}, i.e., $C^{\bullet}(-;\mathbb{R})$ sends colimits in $\mathcal{S}$ to limits in $\mathcal{D}(\mathbb{R})$ -- the case
of coproducts is obvious and the case of pushouts follows from the Mayer-Vietoris sequence for singular cohomology.
\subsection{Bounded cohomology} Let $C^{\bullet}_b(-; \mathbb{R}) \colon \mathrm{Top}^{\mathrm{op}} \to \mathrm{Ch}_{\mathbb{R}}$ denote the functor of the bounded cochain complex \cite{Gr, Iv}, which we regard here as a chain complex concentrated in nonpositive degrees. We recall that a singular cochain
$$\phi \colon C_n(X; \mathbb{R}) \to \mathbb{R}$$
is \emph{bounded} if the subset $\{\phi(\sigma) \in \mathbb{R} \ | \ \sigma \colon \Delta^n \to X\} \subseteq \mathbb{R}$ is bounded. For $n \geq 0$, the bounded cohomology of $X$ is defined as follows:
$$H^{n}_b(X; \mathbb{R}) := H_{-n}(C^{\bullet}_b(X; \mathbb{R})).$$
Since $C^{\bullet}_b(-; \mathbb{R})$ sends weak homotopy equivalences to quasi-isomorphisms \cite{Iv}, we obtain canonically an induced functor
$$C^{\bullet}_b(-; \mathbb{R}) \colon \mathcal{S}^{\mathrm{op}} \to \mathcal{D}(\mathbb{R}).$$
(Note that a weak homotopy equivalence of topological spaces induces a simplicial homotopy equivalence between the corresponding singular sets, since these are Kan complexes. In turn, this induces a chain homotopy equivalence between the singular cochain complexes which also restricts to bounded cochains.)
\subsection{The comparison map} \label{sec:comparison_map} The inclusion of bounded cochains $C^{\bullet}_b(-;\mathbb{R})$ into all singular cochains $C^{\bullet}(-;\mathbb{R})$ induces a natural transformation between the functors $C^{\bullet}_b(-; \mathbb{R}), C^{\bullet}(-;\mathbb{R}) \colon \mathcal{S}^{\mathrm{op}} \to \mathcal{D}(\mathbb{R})$,
$$c \colon C^{\bullet}_b(-;\mathbb{R}) \to C^{\bullet}(-;\mathbb{R}).$$
We will refer to this as the \emph{comparison map} (for bounded cohomology). We also denote the induced comparison maps on cohomology by
$$c^n \colon H^n_b(-; \mathbb{R}) \to H^n(-;\mathbb{R}).$$
A natural transformation $C^{\bullet}_b(-; \mathbb{R}) \to H$ to an excisive functor $H \colon \mathcal{S}^{\mathrm{op}} \to \mathcal{D}(\mathbb{R})$ is determined essentially uniquely by its restriction to $X = \ast$; this is because every (topological) space can be written canonically as a (homotopy) colimit of (weakly) contractible spaces. The comparison map for bounded cohomology is an equivalence for $X = \ast$, that is, the map
$$c_{\{\ast\}} \colon C^{\bullet}_b(\ast ; \mathbb{R}) \to C^{\bullet}(\ast ;\mathbb{R})$$
is an equivalence in $\mathcal{D}(\mathbb{R})$ (it is, in fact, an isomorphism of chain complexes in $\mathrm{Ch}_{\mathbb{R}}$). This map $c_{\{\ast\}}$ together with the fact that $C^{\bullet}(-;\mathbb{R})$ is excisive completely determine the comparison map $c$. Note that the comparison map is not an equivalence in general exactly because $C^{\bullet}_b(-;\mathbb{R})$ fails to be excisive. But $c$ is the initial transformation to an excisive functor. More generally,
for any cocomplete (resp. complete) $\infty$-category $\mathcal{D}$, there is an adjunction of (not necessarily locally small) $\infty$-categories:
$$(-)_{\%} \colon \mathcal{D} \rightleftarrows \mathrm{Fun}(\mathcal{S}, \mathcal{D})\colon \mathrm{ev} \ \ (\text{resp. } \mathrm{ev} \colon \mathrm{Fun}(\mathcal{S}^{\mathrm{op}}, \mathcal{D}) \rightleftarrows \mathcal{D} \colon (-)^{\&})$$
where the right (resp. left) adjoint $\mathrm{ev}$ is given by the evaluation at the contractible space $\ast$ and the associated counit (resp. unit) transformation
$$\smallint \colon (\mathrm{ev}(F))_{\%} \to F \ \ (\text{resp. } \nabla \colon F \to (\mathrm{ev}(F))^{\&})$$
is the \emph{assembly map for $F$}. The functor $(-)^{\&}$ (resp. $(-)_{\%}$) is fully faithful and its essential image consists of the excisive functors. The natural transformation $c$ is an instance of an assembly map in the dual context of contravariant functors
-- $c$ is the \emph{coassembly map for bounded cohomology}. See \cite{WW} for more details about the general construction of assembly maps and their properties, and \cite[6.3-6.4]{Ci} for related general results in the language of $\infty$-categories.
\section{Factorizations} \label{sec_factorizations}
\subsection{General factorizations} Let $X \in \mathcal{S}$ be a space and let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}$ be a diagram of spaces over $X$. A \emph{factorization of $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}$} consists of a diagram $\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G(i) \to X),$ together with a natural transformation $\eta \colon \widetilde{F} \to\widetilde{G}$.
\smallskip
The following proposition shows that factorizations of $\mathrm{H}\mathbb{R}$-colimit diagrams yield factorizations of the comparison map for bounded cohomology.
\begin{proposition} \label{factorization_general}
Let $X \in \mathcal{S}$ be a space and let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$
be an $\mathrm{H}\mathbb{R}$-colimit diagram. For every factorization $\eta \colon \widetilde{F} \to \widetilde{G}$, where
$$\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G(i) \to X),$$
is a diagram of spaces over $X$, there is a canonical factorization of $c_X$ in $\mathcal{D}(\mathbb{R})$:
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[r]^{c_X} \ar[d] & C^{\bullet}(X; \mathbb{R}). \\
\lim_{\mathcal{I}^{\mathrm{op}}}C^{\bullet}_b(G(-); \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G(-); \mathbb{R}) \ar[u]
}
$$
(The maps will be specified in the proof below.)
\end{proposition}
\begin{proof}
There is a canonical diagram in $\mathcal{D}(\mathbb{R})$:
$$\small{
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[d] \ar[rr]^{c_X} && C^{\bullet}(X; \mathbb{R}) \ar[d] \\
C^{\bullet}_b(\mathrm{colim}_{\mathcal{I}} G; \mathbb{R}) \ar[d] \ar[rr] && C^{\bullet}(\mathrm{colim}_{\mathcal{I}} G; \mathbb{R}) \ar[d] \\
C^{\bullet}_b(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}) \ar[rr] && C^{\bullet}(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}). \\
}}
$$
The horizonal maps are induced by the natural comparison map; the vertical maps are given by the functoriality of $C^{\bullet}_b(-;\mathbb{R})$ and $C^{\bullet}(-;\mathbb{R})$. The right vertical composite
$$C^{\bullet}(X; \mathbb{R}) \xrightarrow{\simeq} C^{\bullet}(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R})$$ is an equivalence because the canonical map $\mathrm{colim}_{\mathcal{I}} F \rightarrow X$ is an $\mathbb{R}$-cohomology equivalence by assumption. (Note that the left vertical composite will also be an equivalence and the bounded cochain complex $C^{\bullet}_b(X; \mathbb{R})$ will be a retract of $C^{\bullet}_b(\mathrm{colim}_{\mathcal{I}} G; \mathbb{R})$ if $\widetilde{F}$ is actually a colimit diagram.) Using the fact that $C^{\bullet}(-;\mathbb{R})$ is excisive
and the naturality of the comparison map $c$, the bottom square is identified with the composite diagram:
$$\small{
\xymatrix{
C^{\bullet}_b(\mathrm{colim}_{\mathcal{I}} G; \mathbb{R}) \ar[d] \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G(-); \mathbb{R}) \ar[d] \ar[r]^(0.36){c_{G(-)}} & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G(-); \mathbb{R}) \xleftarrow{\simeq} C^{\bullet}(\mathrm{colim}_{\mathcal{I}} G; \mathbb{R}) \ar@<8ex>[d] \ar@<-8ex>[d] \\
C^{\bullet}_b(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(F(-); \mathbb{R}) \ar[r]^(0.36){c_{F(-)}} & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(F(-); \mathbb{R}) \xleftarrow{\simeq} C^{\bullet}(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}). \\
}}
$$
The left and right horizontal maps are the canonical maps to the respective limits; the middle maps are induced by the natural comparison map.
Then the combination of the two diagrams produces the required factorization.
\end{proof}
\begin{example}[Tautological factorization] \label{ex3}
Let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram and let $\eta \colon \widetilde{F} = \widetilde{F}$ be the identity. Then we obtain a canonical factorization of the comparison map $c_X$ in $\mathcal{D}(\mathbb{R})$:
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[r]^{c_X} \ar[d] & C^{\bullet}(X; \mathbb{R}) \\
\lim_{\mathcal{I}^{\mathrm{op}}}C^{\bullet}_b(F(-); \mathbb{R}) \ar[ur]_q &
}
$$
where the vertical map is the canonical map to the limit and $q$ is the composition
$$\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(F(-); \mathbb{R}) \xrightarrow{c_{F(-)}} \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(F(-); \mathbb{R}) \simeq C^{\bullet}(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}) \simeq C^{\bullet}(X; \mathbb{R}).$$
More generally, note that the factorization of $c_X$ in Proposition \ref{factorization_general} happens essentially on the side of bounded cohomology, i.e., the comparison map factors canonically through the limit of the \emph{bounded} cohomologies of a diagram
of spaces (which depends on $\widetilde{G}$).
\end{example}
\begin{example}[Moore--Postnikov truncation] \label{Moore-Postnikov}
We recall that every map $f \colon Y \to X$ in $\mathcal{S}$ admits canonical \emph{Moore--Postnikov $n$-truncations} \cite[Ch. IX]{Wh}, \cite[IV.3]{GJ}, \cite[6.5]{HTT}; in the case of $\pi_1$, this is a natural factorization of $f$,
$$Y \xrightarrow{p_f} \mathcal{T}(f) \xrightarrow{j_f} X,$$
which is characterized essentially uniquely by the following properties:
\begin{itemize}
\item[(i)] $p_f$ induces a bijection $\pi_0(Y) \xrightarrow{\cong} \pi_0(\mathcal{T}(f))$. Moreover, for every point $y$ in $Y$, the induced homomorphism
$$\pi_1(p_f, y) \colon \pi_1(Y, y) \to \pi_1(\mathcal{T}(f), p_f(y))$$
is surjective.
\item[(ii)] For every point $z$ in $\mathcal{T}(f)$, the homomorphism induced by $j_f$,
$$\pi_1(j_f, z) \colon \pi_1(\mathcal{T}(f), z) \to \pi_1(X, j_f(z)),$$
is injective. Moreover, for every $n > 1$, $j_f$ induces an isomorphism
$$\pi_n(j_f, z) \colon \pi_n(\mathcal{T}(f), z) \xrightarrow{\cong} \pi_n(X, j_f(z)).$$
\end{itemize}
Let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (f(i) \colon F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram. Applying the Moore--Postnikov truncation (at $\pi_1$) to the maps
$(f(i) \colon F(i) \rightarrow X)$ yields a new diagram
$$\mathcal{T}(\widetilde{F}) \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (\mathcal{T}(f(i)) \xrightarrow{j_{f(i)}} X),$$
together with a factorization $p \colon \widetilde{F} \to \mathcal{T}(\widetilde{F}),$ whose components are given by the maps $p_{f(i)}$ in $\mathcal{S}_{/X}$.
Therefore, by Proposition \ref{factorization_general}, we have a canonical factorization in $\mathcal{D}(\mathbb{R})$ as follows:
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[r]^{c_X} \ar[d] & C^{\bullet}(X; \mathbb{R}). \\
\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(\mathcal{T}(f(-)); \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(\mathcal{T}(f(-)); \mathbb{R}) \ar[u]
}
$$
\end{example}
\subsection{(Boundedly) acyclic/truncating factorizations}
The factorization of the comparison map $c_X$ in Proposition \ref{factorization_general} will only be useful in the case of factorizations $\eta \colon \widetilde{F} \to \widetilde{G}$, for which we can determine some interesting properties of the limit of bounded cochain complexes $i \mapsto C^{\bullet}_b(G(i); \mathbb{R})$.
We will consider two special classes of factorizations: \emph{boundedly ($k$-)acyclic} and \emph{boundedly ($k$-)truncating} factorizations.
\begin{definition} \label{def-acyclic-fact}
Let $X \in \mathcal{S}$ be a space, let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}$ be a diagram, and let $k \geq 0$ be an integer.
\begin{itemize}
\item[(1)] We say that $\eta \colon \widetilde{F} \to \widetilde{G}$ is an \emph{acyclic factorization of $\widetilde{F}$} (\emph{with respect to $C^{\bullet}_b(-;\mathbb{R})$}) if the diagram
$$\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G(i) \to X),$$ satisfies the following property: for each $i \in \mathcal{I}$,
the canonical map
$$H^n_b(\ast; \mathbb{R}) \xrightarrow{H^n_b(G(i) \to \ast; \mathbb{R})} H^n_b(G(i); \mathbb{R})$$
is an isomorphism for every $n \geq 0$.
\item[(2)] A factorization $\eta \colon \widetilde{F} \to \widetilde{G}$ is a \emph{k-acyclic factorization of $\widetilde{F}$} (\emph{with respect to $C^{\bullet}_b(-;\mathbb{R})$}) if the diagram
$$\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G(i) \to X),$$ satisfies the following property: for each $i \in \mathcal{I}$,
the canonical map
$$H^n_b(\ast; \mathbb{R}) \xrightarrow{H^n_b(G(i) \to \ast; \mathbb{R})} H^n_b(G(i); \mathbb{R})$$
is an isomorphism for every $n \leq k$ (and a monomorphism in degree $k+1$).
\end{itemize}
\end{definition}
The factorization in Proposition \ref{factorization_general} has interesting implications when we restrict to ($k$-)acyclic factorizations. As we will see in Section \ref{amenable_covers}, the following theorem is a refinement and generalization of the \emph{Covering theorem} (Theorem \ref{nerve_theorem}(1)).
\begin{theorem} \label{nerve_theorem_general1}
Let $X \in \mathcal{S}$ be a space and let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram.
\begin{itemize}
\item[(1)] If $\eta \colon \widetilde{F} \to \widetilde{G}$ is an acyclic factorization of $\widetilde{F}$, then there is a canonical factorization of $c_X$ in $\mathcal{D}(\mathbb{R})$:
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[rr]^{c_X} \ar[dr] && C^{\bullet}(X; \mathbb{R}) \\
& C^{\bullet}(|\mathcal{I}|; \mathbb{R}) \ar[ur]_{\rho_F} &
}
$$
where $\rho_F \colon C^{\bullet}(|\mathcal{I}|; \mathbb{R}) \xrightarrow{C^{\bullet}(p_{F}; \mathbb{R})} C^{\bullet}(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}) \simeq C^{\bullet}(X; \mathbb{R}).$
\item[(2)] If $\eta \colon \widetilde{F} \to \widetilde{G}$ is a $k$-acyclic factorization of $\widetilde{F}$, for some $k \geq 0$, then for every $n \leq k$ there is a canonical factorization of $c^n_X$:
$$
\xymatrix{
H^n_b(X; \mathbb{R}) \ar[rr]^{c^n_X} \ar[dr] && H^n(X; \mathbb{R}) \\
& H^n(|\mathcal{I}|; \mathbb{R}) \ar[ur]_{\rho^n_F} &
}
$$
where $\rho^n_F$ is induced by $\rho_F$ as defined in (1).
\end{itemize}
\end{theorem}
\begin{proof}
(1): Using the acyclicity of $\eta$ and the equivalence $c_{\{\ast\}} \colon C^{\bullet}_b(\ast; \mathbb{R}) \simeq C^{\bullet}(\ast;\mathbb{R})$, we obtain the following canonical identifications in $\mathcal{D}(\mathbb{R})$
$$\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G(-);\mathbb{R}) \simeq \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(\ast; \mathbb{R}) \simeq \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(\ast; \mathbb{R}).$$
Moreover, since $C^{\bullet}(-; \mathbb{R})$ is excisive, we also have canonical identifications:
$$\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(\ast; \mathbb{R}) \simeq C^{\bullet}(\mathrm{colim}_{\mathcal{I}} \delta(\ast); \mathbb{R}) \simeq C^{\bullet}(|\mathcal{I}|; \mathbb{R}).$$
Then the desired factorization follows from Proposition \ref{factorization_general}. Moreover, the identification of $\rho_F$ can be seen from the following diagram in $\mathcal{D}(\mathbb{R})$:
$$
\xymatrix{
\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G(i); \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G(i); \mathbb{R}) \simeq C^{\bullet}(\mathrm{colim}_{\mathcal{I}} G; \mathbb{R}) \rightarrow C^{\bullet}(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}) \\
\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(\ast; \mathbb{R}) \ar[r]^(0.3){\simeq} \ar[u]^{\simeq} & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(\ast ;\mathbb{R}) \simeq C^{\bullet}(|\mathcal{I}|; \mathbb{R}) \xrightarrow{\rho_F} C^{\bullet}(X;\mathbb{R}). \ar@<3ex>[u] \ar@<17ex>[u] \ar@<-17ex>[u]_{\simeq}
}
$$
(Note that the fact that the bottom left map is an equivalence is actually not necessary for the proof.)
\smallskip
\noindent (2): Similarly to the proof of (1), we consider the canonical map in $\mathcal{D}(\mathbb{R})$
$$\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(\ast; \mathbb{R}) \longrightarrow \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G(-); \mathbb{R})$$
which is induced by the natural transformation $G \to \delta(\ast)$. Using that the factorization $\eta$ is $k$-acyclic, it follows that this map also induces an isomorphism in cohomology in degrees $\leq k$ (and a monomorphism in degree $k+1$), since this property is closed under taking limits in $\mathcal{D}(\mathbb{R})$. Then the desired factorization follows from Proposition \ref{factorization_general} and the same diagram in $\mathcal{D}(\mathbb{R})$ as above (after passing to cohomology).
\end{proof}
\begin{example}[Boundedly acyclic $\mathrm{H}\mathbb{R}$-colimit diagrams] \label{bounded-acyclic-colimits}
Let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram and let $\eta \colon \widetilde{F} = \widetilde{F}$ be the tautological factorization (Example \ref{ex3}). Suppose that the values of $F$ are \emph{boundedly acyclic}, that is, for each $i \in \mathcal{I}$, the map
$$H^n_b(\ast; \mathbb{R}) \xrightarrow{H^n_b(F(i) \to \ast; \mathbb{R})} H^n_b(F(i); \mathbb{R})$$
is an isomorphism for $n \geq 0$. In other words, $\eta$ is an acyclic factorization of $\widetilde{F}$, so we obtain by Proposition \ref{factorization_general} and Theorem \ref{nerve_theorem_general1} a canonical factorization of the comparison map $c_X$ in $\mathcal{D}(\mathbb{R})$
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[r]^{c_X} \ar[d] & C^{\bullet}(X; \mathbb{R}). \\
\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(F(-); \mathbb{R}) \ar[r]^(.6){\simeq} & C^{\bullet}(|\mathcal{I}|; \mathbb{R}) \ar[u]_{\rho_F}
}
$$
\end{example}
\begin{remark} \label{acyclic-fact-gromov} Let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram and let $\eta \colon \widetilde{F} \to \widetilde{G}$ be a factorization, where $\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, i \mapsto (G(i) \to X)$. As a consequence of Gromov's \emph{Mapping Theorem} \cite{Gr, Iv}, the factorization $\eta$ is acyclic if and only if the space $G(i)$ is connected and its fundamental group has trivial bounded
cohomology for all $i \in \mathcal{I}$ (see also \cite{MR}). This shows that the factorizations of Example \ref{Moore-Postnikov} have a special importance for the applications of Theorem \ref{nerve_theorem_general1}. We will explore this further in Section \ref{amenable_covers}.
\end{remark}
On the other hand, the notion of a boundedly $k$-truncating factorization requires vanishing properties in the complementary range of degrees.
\begin{definition} \label{def-truncating-fact}
Let $X \in \mathcal{S}$ be a space, let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}$ be a diagram, and let $k \geq 0$ be an integer.
A factorization $\eta \colon \widetilde{F} \to \widetilde{G}$ is a \emph{k-truncating factorization of $\widetilde{F}$} (\emph{with respect to $C^{\bullet}_b(-;\mathbb{R})$}) if the diagram
$$\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G(i) \to X),$$ satisfies the following property: $H^n_b(G(i); \mathbb{R}) \cong 0$ for every $n > k$ and each $i \in \mathcal{I}$.
\end{definition}
The application of the factorization in Proposition \ref{factorization_general} to boundedly truncating factorizations yields the following vanishing result for the comparison map. As we will see in Section \ref{amenable_covers}, this result is a refinement and generalization of the \emph{Vanishing theorem} (Theorem \ref{nerve_theorem}(2)).
\begin{theorem} \label{nerve_theorem_general2}
Let $X \in \mathcal{S}$ be a space and let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram.
\begin{itemize}
\item[(1)] If $\eta \colon \widetilde{F} \to \widetilde{G}$ is a $0$-truncating factorization of $\widetilde{F}$ and $\mathcal{I}$ is (Joyal) equivalent to a simplicial set of dimension $< m$, then the comparison map $c_X$ vanishes in cohomology in degrees $\geq m$.
\item[(2)] More generally, if $\eta \colon \widetilde{F} \to \widetilde{G}$ is a $k$-truncating factorization of $\widetilde{F}$ and $\mathcal{I}$ is (Joyal) equivalent to a simplicial set of dimension $d$, then the comparison map $c_X$ vanishes in cohomology in degrees $ > d + k$.
\end{itemize}
\end{theorem}
\begin{proof} (1): Since $\eta$ is a $0$-truncating factorization, the map $G(i) \to \pi_0(G(i))$, $i \in \mathcal{I}$, induces an equivalence in $\mathcal{D}(\mathbb{R})$
$$C^{\bullet}_b(G(i); \mathbb{R}) \simeq C^{\bullet}_b(\pi_0(G(i)); \mathbb{R}).$$
Then, combining the factorization in Proposition \ref{factorization_general} with the diagram
$$
\xymatrix{
\lim_{\mathcal{I}^{\mathrm{op}}}C^{\bullet}_b(G(-); \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G(-); \mathbb{R}), \\
\lim_{\mathcal{I}^{\mathrm{op}}}C^{\bullet}_b(\pi_0(G(-)); \mathbb{R}) \ar[u]^{\simeq} \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(\pi_0(G(-)); \mathbb{R}) \ar[u]
}
$$
we obtain a factorization of the comparison map $c_X$ through the limit of singular (bounded) cochain complexes
$$D^{\bullet}:= \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(\pi_0(G(-)); \mathbb{R}) \ \ (\text{resp. } \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(\pi_0(G(-)); \mathbb{R})).$$
Both of these are (homotopy) limits of ($\mathbb{Z}$-graded) chain complexes which are concentrated in degree $0$ (up to quasi-isomorphism). Suppose without loss of generality that $\mathcal{I}$ has dimension $< m$. Then the homology of the limit $D^{\bullet}$ is concentrated in nonpositive degrees $k$ with $|k| < m$; this follows by induction on the dimension of $\mathcal{I}$ using \cite[Proposition 4.4.2.2]{HTT} and standard facts from homological algebra. In other words, the cohomology of $D^{\bullet}$ vanishes in cohomological degrees $\geq m$, as required.
\smallskip
\noindent (2): The proof is similar to (1). By Proposition \ref{factorization_general}, the comparison map $c_X$ factors canonically through the limit
$$\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G(-); \mathbb{R}).$$
By assumption, this is a limit of bounded cochain complexes, each of whose cohomology is concentrated in degrees $\leq k$ -- here we are using the cohomological grading convention. We may suppose without loss of generality that $\mathcal{I}$ has dimension $d$. Then, by induction on the dimension of $\mathcal{I}$, using \cite[Proposition 4.4.2.2]{HTT} and standard facts from homological algebra, it follows that the cohomology of the limit is concentrated in degrees $\leq d + k$. (Note that the proof shows the stronger statement that $c_X$ factors through a cochain complex concentrated in degrees $\leq d + k$.)
\end{proof}
\begin{remark} \label{acyclic-fact-gromov2} Let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram and let $\eta \colon \widetilde{F} \to \widetilde{G}$ be a factorization, where $\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, i \mapsto (G(i) \to X)$. Similarly to Remark \ref{acyclic-fact-gromov}, if the factorization $\eta$ is $0$-truncating, then each component of $G(i)$ has trivial bounded cohomology, therefore the fundamental group $\pi_1(G(i), x)$ has trivial bounded cohomology for all $i \in \mathcal{I}$ and $x$ in $G(i)$. The converse also holds if each space $G(i)$, $i \in \mathcal{I}$, has finitely many components; more generally, the converse holds under the assumption that the components of $G(i)$, for each $i \in \mathcal{I}$, satisfy the uniform boundary condition uniformly (see \cite[Appendix A]{LLM}). Again this demonstrates the importance of the factorizations of Example \ref{Moore-Postnikov} for the applications of Theorem \ref{nerve_theorem_general2}. We will explore this further in Section \ref{amenable_covers}.
\end{remark}
\subsection{Generalizations} \label{generalizations} Proposition \ref{factorization_general} and Theorems \ref{nerve_theorem_general1} and \ref{nerve_theorem_general2} apply in much greater generality with the same proofs.
Specifically, Proposition \ref{factorization_general} generalizes to natural transformations $c \colon B \to H$ where $B, H \colon \mathcal{S}^{\mathrm{op}} \to \mathcal{D}$ are functors to a complete $\infty$-category $\mathcal{D}$ and $H$ is excisive. In this context, $\mathrm{H}\mathbb{R}$-colimit diagrams are replaced by diagrams
$$\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, i \mapsto (F(i) \to X),$$ such that the canonical map $\mathrm{colim}_{\mathcal{I}} F \to X$ maps to an equivalence in $\mathcal{D}$ under $H$.
Theorem \ref{nerve_theorem_general1}(1) also holds in this generality (with the same proof). In this context, a factorization $\eta \colon \widetilde{F} \to \widetilde{G}$ of $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}$ is called \emph{acyclic} (with respect to $B$) if $$B(\ast) \xrightarrow{B(G(i) \to \ast)} B(G(i))$$
is an equivalence in $\mathcal{D}$ for every $i \in \mathcal{I}$. Moreover, Theorem \ref{nerve_theorem_general1}(2) and Theorem \ref{nerve_theorem_general2} also apply more generally, e.g., they hold for natural transformations $c \colon B \to H$ as above, where $\mathcal{D} = \mathcal{D}(R)$ is the derived $\infty$-category of a ring $R$ and the homology of $B(\ast) \in \mathcal{D}(R)$ is concentrated in degree $0$. In addition, a close inspection of these proofs indicates that it suffices to know only that $H$ acts excisively for the given diagram in the corresponding range of degrees.
\smallskip
In particular, Proposition \ref{factorization_general} and Theorems \ref{nerve_theorem_general1} and \ref{nerve_theorem_general2} generalize to analogous results about the comparison map for bounded cohomology with respect to general normed rings or with twisted coefficients (coming from $X$).
\smallskip
By passing to the opposite $\infty$-categories, we also obtain versions of these results for natural transformations $c \colon H \to B$ where $B, H \colon \mathcal{S} \to \mathcal{D}$ are functors to a cocomplete $\infty$-category $\mathcal{D}$ and $H$ is excisive (i.e., $H$ preserves small colimits). In particular, this yields analogous results to Proposition \ref{factorization_general} and Theorems \ref{nerve_theorem_general1} and \ref{nerve_theorem_general2} for the comparison map for $\ell^1$-homology. In this case, the comparison map corresponds to the assembly map for $\ell^1$-homology (in the sense of Section \ref{sec:prelim}); this should not be confused with the (other) excisive approximation to $\ell^1$-homology (in the sense of Goodwillie calculus) which is shown to be trivial \cite{Wi}.
\section{Amenable factorizations} \label{amenable_covers}
Amenable covers (or amenable factorizations) are in practice the main source of acyclic or $0$-truncating factorizations (cf. Remarks \ref{acyclic-fact-gromov} and \ref{acyclic-fact-gromov2}).
\begin{definition}
Let $X \in \mathcal{S}$ be a space and let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}$ be a diagram. A factorization $\eta \colon \widetilde{F} \to \widetilde{G}$ is called \emph{amenable} if the diagram
$$\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G(i) \to X),$$ satisfies the following property: the fundamental group $\pi_1(G(i), x)$ is amenable for every $i \in \mathcal{I}$ and $x$ in $G(i)$.
\end{definition}
\begin{corollary} \label{amenable-fact}
Let $X \in \mathcal{S}$ be a space and let $\widetilde{F} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F(i) \to X),$ be an $\mathrm{H}\mathbb{R}$-colimit diagram. Suppose that $\eta \colon \widetilde{F} \to \widetilde{G}$ is an amenable factorization of $\widetilde{F}$, where $\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, i \mapsto (G(i) \to X)$. Then the following hold:
\begin{itemize}
\item[(1)] If the space $G(i)$ is connected for all $i \in \mathcal{I}$, then there is a canonical factorization of $c_X$ in $\mathcal{D}(\mathbb{R})$:
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[rr]^{c_X} \ar[rd] && C^{\bullet}(X; \mathbb{R}) \\
& C^{\bullet}(|\mathcal{I}|; \mathbb{R}) \ar[ru]_{\rho_F} &
}
$$
where $\rho_F \colon C^{\bullet}(|\mathcal{I}|; \mathbb{R}) \xrightarrow{C^{\bullet}(p_{F}; \mathbb{R})} C^{\bullet}(\mathrm{colim}_{\mathcal{I}} F; \mathbb{R}) \simeq C^{\bullet}(X; \mathbb{R}).$
\item[(2)] If $\mathcal{I}$ is (Joyal) equivalent to a simplicial set of dimension $< m$, then the comparison map $c_X$ vanishes in cohomology in degrees $\geq m$.
\end{itemize}
\end{corollary}
\begin{proof} We claim that the amenable factorization $\eta$ is $0$-truncating. It follows from the \emph{Mapping Theorem} \cite{Gr, Iv} that the components of $G(i)$ (or the associated fundamental groups) have trivial bounded cohomology for all $i \in \mathcal{I}$. As $G(i)$ may have infinitely many components, in order to show that the bounded cohomology of $G(i)$ itself is concentrated in degree $0$, it is necessary to look at the general properties of the cohomology of bounded products (see \cite[Appendix A]{LLM}). First, there is a uniform bound for the uniform boundary condition of a -- possibly infinite -- family of amenable groups \cite[Example A.10]{LLM}. By the proof of the \emph{Mapping Theorem} (combined with \cite[Proposition A.3]{LLM}), the same property holds also for families of connected spaces with amenable fundamental groups. It follows that the bounded cohomology of $G(i)$, $i \in \mathcal{I}$, is the bounded product of the bounded cohomologies of its components (see \cite[Theorem A.15]{LLM}), therefore $H^{\bullet}_b(G(i); \mathbb{R})$, $i \in \mathcal{I}$, is concentrated in degree $0$. This proves the claim that amenable factorizations are $0$-truncating. Then (2) is a special case of Theorem \ref{nerve_theorem_general2}(1). Under the additional assumption that $G(i)$ is connected for all $i \in \mathcal{I}$, the $0$-truncating factorization $\eta$ is actually acyclic. Therefore (1) follows from Theorem \ref{nerve_theorem_general1}(1).
\end{proof}
Amenable factorizations often arise in practice from amenable covers of topological spaces (cf. Example \ref{covering-colim}). Let $\mathcal{U} = \{U_i \}_{i \in I}$ be an open cover of a topological space $X$. We recall from Example \ref{covering-colim} that there is an associated colimit diagram
$$\widetilde{F}_{\mathcal{U}} \colon \mathcal{I}_{\mathcal{U}} \to \mathcal{S}_{/X}, \ \sigma \mapsto (U_{\sigma} \xrightarrow{i_{\sigma}} X).$$
As explained in Example \ref{Moore-Postnikov}, the Moore--Postnikov truncation (at $\pi_1$) of the maps $(i_{\sigma} \colon U_{\sigma} \rightarrow X)$ defines a new diagram
$$\mathcal{T}(\widetilde{F}_{\mathcal{U}}) \colon \mathcal{I}_{\mathcal{U}} \to \mathcal{S}_{/X}, \ \sigma \mapsto (V_{\sigma} := \mathcal{T}(i_{\sigma}) \xrightarrow{j_{\sigma}} X)$$
together with a factorization $p \colon \widetilde{F}_{\mathcal{U}} \to \mathcal{T}(\widetilde{F}_{\U})$.
We recover the statements in Theorem \ref{nerve_theorem} as special cases:
\smallskip
\noindent (1) \emph{Covering theorem}. Suppose that $p \colon \widetilde{F}_{\mathcal{U}} \to \mathcal{T}(\widetilde{F}_{\mathcal{U}})$ is an acyclic factorization of $\widetilde{F}_{\mathcal{U}}$, i.e., the space $V_{\sigma}$ have trivial bounded cohomology for every $\sigma \in \mathcal{I}_{\mathcal{U}}$. Then there is a canonical factorization in $\mathcal{D}(\mathbb{R})$ (Theorem \ref{nerve_theorem_general1}):
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[rr]^{c_X} \ar[rd] && C^{\bullet}(X; \mathbb{R}). \\
& C^{\bullet}(|\mathcal{I}_{\mathcal{U}}|; \mathbb{R}) \ar[ur]_{\rho_{F_{\mathcal{U}}}} &
}
$$
\smallskip
\noindent The factorization $p$ is acyclic in the case where $U_{\sigma} \subseteq X$ is path-connected for every $\sigma \in \mathcal{I}_{\mathcal{U}}$ and the image of the homomorphism $(\pi_1(U_i, x) \to \pi_1(X, x))$ is amenable for every $i \in I$. In this case, for every $\sigma \in \mathcal{I}_{\mathcal{U}}$, the space $V_{\sigma}$ is connected in $\mathcal{S}$ with amenable fundamental group -- this uses the fact that subgroups of amenable groups are again amenable. In other words, the factorization $p \colon \widetilde{F}_{\mathcal{U}} \to \mathcal{T}(\widetilde{F}_{\mathcal{U}})$ is amenable and satisfies the assumptions of Corollary \ref{amenable-fact}(1). This case recovers Theorem \ref{nerve_theorem}(1).
\smallskip
\noindent We also refer to the recent work of Ivanov \cite{Iv2} for closely related results.
\smallskip
\noindent (2) \emph{Vanishing theorem}. Suppose that $p \colon \widetilde{F}_{\mathcal{U}} \to \mathcal{T}(\widetilde{F}_{\mathcal{U}})$ is a $0$-truncating factorization of $\widetilde{F}_{\mathcal{U}}$. If the dimension of $\mathcal{I}_{\mathcal{U}}$ is $< m$ (i.e., if $U_{\sigma} = \varnothing$ when $\sigma$ has $m+1$ elements), then the comparison map
$$c^n_X \colon H^n_b(X; \mathbb{R}) \to H^n(X; \mathbb{R})$$
is zero for $n \geq m$ (Theorem \ref{nerve_theorem_general2}).
\smallskip
\noindent The factorization $p \colon \widetilde{F}_{\mathcal{U}} \to \mathcal{T}(\widetilde{F}_{\mathcal{U}})$ is $0$-truncating when $p$ is an amenable factorization, that is, when the image of the homomorphism $(\pi_1(U_i, x) \to \pi_1(X, x))$ is amenable for every $x \in U_i$ and $i \in I$ (see the proof of Corollary \ref{amenable-fact}). We recall that the open cover $\mathcal{U}$ is called \emph{amenable} in this case -- note that $U_i$ are not required to be path-connected. This case recovers Theorem \ref{nerve_theorem}(2).
\smallskip
More generally,
\smallskip
\noindent (3) Suppose that $p \colon \widetilde{F}_{\mathcal{U}} \to \mathcal{T}(\widetilde{F}_{\mathcal{U}})$ is a $k$-truncating factorization of $\widetilde{F}_{\mathcal{U}}$. In this case, we call the open cover $\mathcal{U}$ $k$-\emph{truncating}. If $\mathcal{I}_{\mathcal{U}}$ has dimension $d$, then the comparison map
$$c_X^n \colon H^n_b(X; \mathbb{R}) \to H^n(X; \mathbb{R})$$
is zero for $n > d + k$ (Theorem \ref{nerve_theorem_general2}). On the other hand, if $p \colon \widetilde{F}_{\mathcal{U}} \to \mathcal{T}(\widetilde{F}_{\mathcal{U}})$ is a $k$-acyclic factorization, then the comparison map $c_X^n$ factors through the map $\rho_{F_{\mathcal{U}}}^n \colon H^n(|\mathcal{I}_{\mathcal{U}}|; \mathbb{R}) \to H^n(X; \mathbb{R})$ for $n \leq k$ (Theorem \ref{nerve_theorem_general1}).
\medskip
We note that these results (esp. the vanishing result in (2)) are useful for showing the vanishing of the simplicial volume of $X$, when $X$ is an oriented closed connected manifold; for example, see \cite[3.1]{Gr}.
\begin{application}[Families of amenable covers] This example of application is a slight variation of a recent result of L\"oh--Moraschini \cite[Corollary 1.2]{LM}. Let $p \colon E \to B$ be a fiber bundle between path-connected spaces and let $\mathcal{U} = \{U_i\}_{i \in I}$ be an open cover of $B$ by trivializing neighborhoods for $p$. In addition, suppose that the path-components of $U_{\sigma}$ have amenable fundamental groups for every $\sigma \in \mathcal{I}_{\mathcal{U}}$. Let $d$ denote the dimension of $\mathcal{I}_{\mathcal{U}}$. If $B$ is a finite simplicial complex (or a smooth closed manifold), then there is such an open cover of $B$ with $d = \mathrm{dim}(B)$.
Let $F$ denote the fiber of $p$ (at a basepoint $b \in B$) and suppose that $\mathcal{V} = \{V_j\}_{j \in J}$ is an amenable (or $k$-truncating) open cover of $F$. We denote by $\kappa$ the dimension of $\mathcal{I}_{\mathcal{V}}$. Let $\mathcal{U}' = \{U'_i = p^{-1}(U_i)\}_{i \in I}$ denote the induced open cover of $E$. Using an identification $U'_i \cong U_i \times F$, the amenable open cover $\{U_i \times V_j\}_{j \in J}$ of $U_i \times F$ determines an amenable open cover of $U'_i$, denoted by $\mathcal{V}_i = \{V_{ij}\}_{j \in J}$ -- this is amenable because amenable groups are closed under products.
These choices of open covers for $\{U'_i\}_{i \in I}$ determine an amenable open cover $\mathcal{U} \times \mathcal{V}:= \bigcup_{i \in I} \mathcal{V}_i$ of $E$ -- this is amenable because amenable groups are closed under quotients. An elementary combinatorial argument shows that the dimension of $\mathcal{I}_{\mathcal{U} \times \mathcal{V}}$ is less than $(d+1)(\kappa+1)$.
Therefore the comparison map
$$c^n_E: H^n_b(E; \mathbb{R}) \to H^n(E;\mathbb{R})$$
vanishes for $n \geq (d+1)(\kappa + 1)$. Assuming $E$ is an oriented closed connected manifold, this vanishing result implies the vanishing of the simplicial volume of $E$ when $\mathrm{dim}(E) \geq (d+1)(\kappa + 1)$.
\end{application}
\begin{example}[The plus construction] Let $X$ be a based connected space (for example, $X = BG$ for a discrete group $G$) and
let $X^+_P$ denote the plus construction of $X$ associated to a normal perfect subgroup $P$ of $\pi_1(X, *)$ \cite{HH, Ra}. We recall that every connected space is equivalent to $BG^+_P$ for some discrete group $G$ and a normal perfect subgroup $P \unlhd G$ \cite{KT}.
Then every colimit diagram $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}_{/X}$ gives rise to an $\mathrm{H}\mathbb{R}$-colimit diagram
$\iota_*\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}_{/X^+_P}$ by composition with the canonical acyclic map
$\iota \colon X \to X^+_P$. Moreover, a ($k$-acyclic/truncating or amenable) factorization of $\widetilde{F}$ defines also a ($k$-acyclic/truncating or amenable) factorization of $\iota_*\widetilde{F}$.
Conversely, given a factorization $\eta \colon \iota_*\widetilde{F} \to \widetilde{G}$ of $\iota_*\widetilde{F}$, there is an associated factorization of $\widetilde{F},$
$$\iota^*\eta \colon \widetilde{F} \to \iota^* \widetilde{G},$$
where $\iota^*\widetilde{G} \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G(i) \times_{X^+_P} X \to X),$ is defined by pulling back $\widetilde{G}$ along the map $\iota$; in addition, there are natural acyclic maps $G(i) \times_{X^+_P} X \to G(i)$ for every $i \in \mathcal{I}$. Thus, applying Proposition \ref{factorization_general}, we obtain a factorization of $c_X$ in $\mathcal{D}(\mathbb{R})$:
$$
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[r]^{c_X} \ar[d] & C^{\bullet}(X; \mathbb{R}). \\
\lim_{\mathcal{I}^{\mathrm{op}}}C^{\bullet}_b(G(-) \times_{X^+_P} X; \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G(-) \times_{X^+_P} X; \mathbb{R}) \ar[u] \simeq \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G(-); \mathbb{R})
}
$$
In this way, we may obtain factorizations of the comparison map $c_X$ of $X$ using factorizations of the comparison map $c_{X^+_P}$ of $X^+_P$ whose fundamental group may be simpler in general.
\end{example}
\section{Parametrized factorizations} \label{pullback_fact}
\subsection{General parametrized factorizations} \label{general_parametrized} For any map $p \colon X \to B$ in $\mathcal{S}$, we may carry out a similar analysis of the properties of the comparison map $c_X \colon C^{\bullet}_{b}(X;\mathbb{R}) \to C^{\bullet}(X;\mathbb{R})$ in a parametrized setting. In this section we briefly sketch some of the details and discuss some specific examples and applications.
\smallskip
Let $\mathcal{S}^{\to} := \mathrm{Fun}(\Delta^1, \mathcal{S})$ denote the $\infty$-category of maps in $\mathcal{S}$. Note that a diagram $$\widetilde{F} = (\widetilde{F}_0 \to \widetilde{F}_1) \colon \mathcal{I} \to (\mathcal{S}^{\to})_{/p}$$ corresponds to a cone $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to \mathcal{S}^{\to}$ on $F:=\widetilde{F}_{| \mathcal{I}} \colon \mathcal{I} \to \mathcal{S}^{\to}$, written $i \mapsto F(i) = (F_0(i) \to F_1(i))$, with cone object $p \in \mathcal{S}^{\to}$. Such a diagram in $(\mathcal{S}^{\to})_{/p}$ is a \emph{colimit diagram} (resp. \emph{$\mathrm{H}\mathbb{R}$-colimit diagram}) if the associated cone is a (pointwise) colimit diagram (resp. $\mathrm{H}\mathbb{R}$-colimit diagram) in $\mathcal{S}^{\to}$.
A diagram $\widetilde{G} = (\widetilde{G}_0 \to \widetilde{G}_1) \colon \mathcal{I} \to (\mathcal{S}^{\to})_{/p}$, $i \mapsto \big(G(i) = (G_0(i) \to G_1(i)) \to p\big),$ together with a natural transformation $$\eta \colon \widetilde{F} \to\widetilde{G}$$ is called a \emph{factorization} of $\widetilde{F} \colon \mathcal{I} \to (\mathcal{S}^{\to})_{/p}$. Explicitly, a factorization $\eta \colon \widetilde{F} \to \widetilde{G}$ of $\widetilde{F}$ consists of a factorization $\eta_0 \colon \widetilde{F}_0 \to \widetilde{G}_0$ of $\widetilde{F}_0 \colon \mathcal{I} \to \mathcal{S}_{/X}$ and a factorization $\eta_1 \colon \widetilde{F}_1 \to \widetilde{G}_1$ of $\widetilde{F}_1 \colon \mathcal{I} \to \mathcal{S}_{/B}$ which fit into a natural diagram for all $i \in \mathcal{I}$:
$$
\xymatrix{
F_0(i) \ar[d] \ar[r]^{\eta_{0, i}} & G_0(i) \ar[d] \ar[r] & X \ar[d]^p \\
F_1(i) \ar[r]^{\eta_{1,i}} & G_1(i) \ar[r] & B.
}
$$
\begin{example}[Pullback factorizations] Let $\widetilde{F}_1 \colon \mathcal{I} \to \mathcal{S}_{/B}, \ i \mapsto (F_1(i) \to B),$ be a (colimit) diagram and $\eta_1 \colon \widetilde{F}_1 \to \widetilde{G}_1$ a factorization of $\widetilde{F}_1$, where $\widetilde{G}_1 \colon \mathcal{I} \to \mathcal{S}_{/B}, \ i \mapsto (G_1(i) \to B)$. These diagrams induce by pullback along $p$ a (colimit) diagram (because pulling back commutes with colimits in $\mathcal{S}$)
$$\widetilde{F}_0 \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (F_0(i) = F_1(i) \times_B X \to X),$$
and a factorization $\eta_0 \colon \widetilde{F}_0 \to \widetilde{G}_0$, where
$$\widetilde{G}_0 \colon \mathcal{I} \to \mathcal{S}_{/X}, \ i \mapsto (G_0(i) = G_1(i) \times_B X \to X).$$
Moreover, the canonical maps $F_0(i) \to F_1(i)$ and $G_0(i) \to G_1(i)$ determine diagrams $\widetilde{F} = (\widetilde{F}_0 \to \widetilde{F}_1) \colon \mathcal{I} \to (\mathcal{S}^{\to})_{/p}$ and $\widetilde{G} = (\widetilde{G}_0 \to \widetilde{G}_1) \colon \mathcal{I} \to (\mathcal{S}^{\to})_{/p}$ together with a factorization $\eta = (\eta_0 \to \eta_1)\colon \widetilde{F} \to \widetilde{G}$.
\end{example}
The use of (parametrized) factorizations for the study of the comparison map is based on the diagram below (cf. Proposition \ref{factorization_general}). Let $\widetilde{F} = (\widetilde{F}_0 \to \widetilde{F}_1) \colon \mathcal{I} \to (\mathcal{S}^{\to})_{/p}$ be a diagram such that $\widetilde{F}_0 \colon \mathcal{I} \to \mathcal{S}_{/X}$ is an $\mathrm{H}\mathbb{R}$-colimit diagram. Given another diagram $\widetilde{G} = (\widetilde{G}_0 \to \widetilde{G}_1) \colon \mathcal{I} \to (\mathcal{S}^{\to})_{/p}$ and a factorization $\eta \colon \widetilde{F} \to \widetilde{G}$ of $\widetilde{F}$, then we obtain the following diagram in $\mathcal{D}(\mathbb{R})$:
\begin{equation*} \label{fact_param}
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[r]^{c_X} \ar[d] & C^{\bullet}(X; \mathbb{R}). \ar[d] \ar[dr]^{\simeq} & \\
\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G_0(-); \mathbb{R}) \ar[r]^{[**]} & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G_0(-); \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(F_0(-); \mathbb{R}) \\
\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G_1(-); \mathbb{R}) \ar[u]^{[*]} \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G_1(-); \mathbb{R}) \ar[u] \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(F_1(-); \mathbb{R}) \ar[u] \\
}
\end{equation*}
Moreover, if $\widetilde{F}_1 \colon \mathcal{I} \to \mathcal{S}_{/B}$ is also an $\mathrm{H}\mathbb{R}$-colimit diagram, then we have a canonical equivalence $\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(F_1(-); \mathbb{R}) \simeq C^{\bullet}(B; \mathbb{R})$ such that the right vertical composition in the diagram is identified with the map $C^{\bullet}(p;\mathbb{R}) \colon C^{\bullet}(B;\mathbb{R}) \to C^{\bullet}(X;\mathbb{R})$.
\smallskip
Special properties of $\eta$ or $p$ would generally allow us to conclude useful properties of the maps indicated in the diagram by [$\ast$] and [$\ast\ast$], and therefore obtain information about the comparison map $c_X$. In analogy with Theorems \ref{nerve_theorem_general1} and \ref{nerve_theorem_general2}, the goal of this analysis is either to relate the comparison map $c_X$ with the bounded cohomology of $B$ or show its vanishing in a range of degrees. In order to view Theorems \ref{nerve_theorem_general1} and \ref{nerve_theorem_general2} as special cases of the present context applied to the map $p \colon X \to \ast$, it suffices to note that $\mathcal{S}_{/X}$ can be identified with the full subcategory of $(\mathcal{S}^{\to})_{/p}$, which is spanned by objects of the form
$$
\xymatrix{
A \ar[r] \ar[d] & X \ar[d] \\
\ast \ar@{=}[r] & \ast.}
$$
In this way we may view diagrams and their factorizations in $\mathcal{S}_{/X}$ as diagrams and factorizations in $(\mathcal{S}^{\to})_{/p}$. (The present parametrized setting can also be extended to objects $p \colon K \to \mathcal{S}$ and diagrams and their factorizations in $\mathrm{Fun}(K, \mathcal{S})_{/p}$ for a based simplicial set $K$.)
\begin{example}\label{pullback_fact_1}
In the notation introduced above, suppose that the map $G(i)$ induces isomorphisms in bounded cohomology for all $i \in \mathcal{I}$, that is, the induced map $H^n_b(G_1(i);\mathbb{R}) \to H^n_b(G_0(i);\mathbb{R})$ is an isomorphism for every $n \geq 0$. In this case, the map [$\ast$] in the diagram is an equivalence in $\mathcal{D}(\mathbb{R})$, so we obtain a canonical factorization:
\begin{equation} \label{fact_5.2} \tag{\ref{pullback_fact_1}}
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[rr]^{c_X} \ar[dr] && C^{\bullet}(X; \mathbb{R}). \\
& \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(F_1(-); \mathbb{R}) \ar[ur] &
}
\end{equation}
If $\widetilde{F}_1 \colon \mathcal{I} \to \mathcal{S}_{/B}$ is an $\mathrm{H}\mathbb{R}$-colimit diagram, then this factorization \eqref{fact_5.2} is identified with the factorization:
\begin{equation*}
\xymatrix{
C^{\bullet}_b(X; \mathbb{R}) \ar[rr]^{c_X} \ar[dr] && C^{\bullet}(X; \mathbb{R}). \\
& C^{\bullet}(B; \mathbb{R}) \ar[ru]_{C^{\bullet}(p; \mathbb{R})} \ar[ur] &
}
\end{equation*}
On the other hand, if $B=\ast$ and $\widetilde{F}_1 \colon \mathcal{I} \to \mathcal{S}$ is the constant diagram $\delta(\ast)$, then the factorization \eqref{fact_5.2} is identified with the factorization of Theorem \ref{nerve_theorem_general1}(1). So we may view the factorization \eqref{fact_5.2} as a relative version of Theorem \ref{nerve_theorem_general1}(1).
Under appropriate assumptions on the maps $G(i)$, $i \in \mathcal{I}$, there are similar versions of Theorem \ref{nerve_theorem_general1}(2) and Theorem \ref{nerve_theorem_general2} in this context, too. We leave the details to the interested reader.
\end{example}
\subsection{Pullbacks of open covers} \label{pullback_covers} We briefly discuss the special case of (possibly only $1$-categorical) pullback factorizations arising from open covers (Example \ref{covering-colim}).
\smallskip
Let $p \colon E \to B$ be a surjective map of topological spaces and let $\mathcal{U} = \{U_i\}_{i \in I}$ be an open cover of $B$. Let $\mathcal{U}' = \{U'_i = p^{-1}(U_i)\}_{i \in I}$ denote the induced open cover of $E$ -- note that $\mathcal{I}_{\mathcal{U}'} = \mathcal{I}_{\mathcal{U}}$. An important example is when $p: E \to B$ is a fiber bundle between oriented closed connected manifolds and $\mathcal{U}$ is an open cover of $B$ by trivializing neighborhoods for $p$. In this case, the factorizations of this subsection are often useful for showing the vanishing of the simplicial volume of $E$.
The open covers $\mathcal{U}$ and $\mathcal{U}'$ yield a diagram $\widetilde{F}_{\mathcal{U}/p} \colon \mathcal{I}_{\mathcal{U}} \to (\mathcal{S}^{\to})_{/p}$ which is defined for each $\sigma \in \mathcal{I}_{\mathcal{U}}$ by the square of topological spaces
$$
\xymatrix{
U'_{\sigma} = p^{-1}(U_{\sigma}) \ar[d]_{p_{\sigma}} \ar[r] & E \ar[d]^p \\
U_{\sigma} \ar[r] & B.
}
$$
\smallskip
\noindent \emph{(a) $\mathcal{U}$-local bounded cohomology equivalence} (cf. Example \ref{pullback_fact_1}). Suppose that the following property is satisfied: the map induced by $p_{\sigma} = p_{| U'_{\sigma}} \colon U'_{\sigma} \to U_{\sigma}$,
$$C^{\bullet}_b(U_{\sigma}; \mathbb{R}) \to C^{\bullet}_b(U'_{\sigma}; \mathbb{R}),$$
is an equivalence in $\mathcal{D}(\mathbb{R})$ for every $\sigma \in \mathcal{I}_{\mathcal{U}}$. This happens, for example, when $U_{\sigma}$ is weakly contractible and $U'_{\sigma}$ is path-connected with amenable (or boundedly acyclic) fundamental group for every $\sigma \in \mathcal{I}_{\mathcal{U}}$.
Then, using the tautological factorization (Example \ref{ex3}) of $\widetilde{F}_{\mathcal{U}/p}$ and the equivalence $\lim_{\sigma \in\mathcal{I}^{\mathrm{op}}_{\mathcal{U}}} C^{\bullet}(U_{\sigma}; \mathbb{R}) \simeq C^{\bullet}(B; \mathbb{R})$, we obtain a canonical factorization in $\mathcal{D}(\mathbb{R})$:
\begin{equation*}
\xymatrix{
C^{\bullet}_b(E; \mathbb{R}) \ar[rr]^{c_E} \ar[dr] && C^{\bullet}(E; \mathbb{R}). \\
& C^{\bullet}(B; \mathbb{R}) \ar[ur]_{C^{\bullet}(p; \mathbb{R})}
}
\end{equation*}
\medskip
\noindent \emph{(b) $\mathcal{U}$-local boundedly equivalent factorizations} (cf. Example \ref{pullback_fact_1}).
Let $V_{\sigma}$ (resp. $V'_{\sigma}$) be the Moore--Postnikov truncation of the inclusion map $U_{\sigma} \to B$ (resp. $U'_{\sigma} \to E$) as in Example \ref{Moore-Postnikov} and Section \ref{amenable_covers}. There are natural maps $p'_{\sigma} \colon V'_{\sigma} \to V_{\sigma}$ by the functoriality of the Moore--Postnikov truncation, yielding a parametrized factorization $\eta \colon \widetilde{F}_{\mathcal{U}/p} \to \widetilde{G}$; explicitly, this is given for each $\sigma \in \mathcal{I}_{\mathcal{U}}$ by the diagram
$$
\xymatrix{
U'_{\sigma} \ar[r] \ar[d]_{p_{\sigma}} & V'_{\sigma} \ar[d]_{p'_{\sigma}} \ar[r] & E \ar[d]^p \\
U_{\sigma} \ar[r] & V_{\sigma} \ar[r] & B.
}
$$
Suppose that the following property is satisfied: the map induced by $p'_{\sigma} \colon V'_{\sigma} \to V_{\sigma},$
$$C^{\bullet}_b(V_{\sigma}; \mathbb{R}) \to C^{\bullet}_b(V'_{\sigma}; \mathbb{R}),$$
is an equivalence in $\mathcal{D}(\mathbb{R})$ for every $\sigma \in \mathcal{I}_{\mathcal{U}}$. For example, this property is satisfied when the spaces $U_{\sigma}, U'_{\sigma}$ are path-connected for every $\sigma \in \mathcal{I}_{\mathcal{U}}$ and $\pi_1(p'_{\sigma})$ is surjective with boundedly acyclic kernel (see \cite{MR}). Then, proceeding as in Example \ref{pullback_fact_1}, we obtain a canonical factorization in $\mathcal{D}(\mathbb{R})$:
\begin{equation*}
\xymatrix{
C^{\bullet}_b(E; \mathbb{R}) \ar[rr]^{c_E} \ar[dr] && C^{\bullet}(E; \mathbb{R}). \\
& C^{\bullet}(B; \mathbb{R}) \ar[ur]_{C^{\bullet}(p; \mathbb{R})} &
}
\end{equation*}
\subsection{Relative covering/vanishing theorems} \label{subsec:rel-cov-van} For an inclusion of topological spaces $i \colon A \subseteq X$, the \emph{relative bounded cohomology $C^{\bullet}_b(X, A; \mathbb{R}) \in \mathcal{D}(\mathbb{R})$} of the pair $(X,A)$ is defined by
$$C^{\bullet}_b(X, A;\mathbb{R}) := \mathrm{fib}\big(C^{\bullet}_b(X;\mathbb{R}) \xrightarrow{C^{\bullet}_b(i; \mathbb{R})} C^{\bullet}_b(A;\mathbb{R})\big).$$
Moreover, we may define similarly the relative bounded cohomology of an arbitrary map in $\mathcal{S}$. Similarly to the absolute case, there is also an induced natural comparison map in $\mathcal{D}(\mathbb{R})$ in the relative setting
$$c_{X,A} \colon C^{\bullet}_b(X,A;\mathbb{R}) \to C^{\bullet}(X,A;\mathbb{R}).$$
The analysis of parametrized factorizations in Subsections \ref{general_parametrized} and \ref{pullback_covers} can be adjusted to the relative setting simply by passing to the respective fibers -- this is another instance where working directly in $\mathcal{D}(\mathbb{R})$ can be useful.
More precisely, given an $\mathrm{H}\mathbb{R}$-colimit cone $\widetilde{F} \colon \mathcal{I}^{\triangleright} \to (\mathcal{S}^{\to})_{/i}$ together with a factorization $\eta \colon \widetilde{F} \to \widetilde{G}$ (as in Subsection \ref{general_parametrized}), we obtain a factorization of $c_{X,A}$ in $\mathcal{D}(\mathbb{R})$:
$$\small{
\xymatrix{
C^{\bullet}_b(X, A; \mathbb{R}) \ar[r]^{c_{X,A}} \ar[d] & C^{\bullet}(X, A; \mathbb{R}). \ar[d] \ar@/^6pc/[dd]^{\simeq} \\
\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G_1(-), G_0(-); \mathbb{R}) \ar[r] & \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(G_1(-), G_0(-); \mathbb{R}) \ar[d] \\
& \lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}(F_1(-), F_0(-); \mathbb{R})
}
}
$$
Similarly to the absolute case, we may identify conditions on the factorization $\eta$ such that the relative bounded cohomology groups $H^{\bullet}_b(G_1(i), G_0(i);\mathbb{R})$ vanish in certain degrees, and thus obtain a description of the limit in $\mathcal{D}(\mathbb{R})$
$$\lim_{\mathcal{I}^{\mathrm{op}}} C^{\bullet}_b(G_1(-), G_0(-);\mathbb{R})$$
(as in Theorem \ref{nerve_theorem_general1} and Example \ref{pullback_fact_1}) or conclude the vanishing of $c_{X,A}$ in certain degrees (as in Theorem \ref{nerve_theorem_general2}).
The possible examples are numerous; we will illustrate the application of this general method with the following example, which shows versions of the covering and vanishing theorems in the relative context. Similar results (with different sets of assumptions) were also obtained recently by Li--L\"oh--Moraschini \cite{LLM}.
\begin{application}[Relative covering and vanishing theorems via open covers] \label{relative-open-covers}
Let $\mathcal{U} = \{U_i\}_{i \in I}$ be an amenable (or $0$-truncating) open cover of $X$ and let $\mathcal{U}_A = \{U_i \cap A\}_{i \in I}$ denote the induced open cover of $A$. Together they determine a colimit cone $\widetilde{F}_{\mathcal{U}, \mathcal{U}_A} \colon \mathcal{I}_{\mathcal{U}}^{\triangleright} \to (\mathcal{S}^{\to})_{/i}$, which sends $\sigma \in \mathcal{I}_{\mathcal{U}}$ to the square
$$
\xymatrix{
U_{\sigma} \cap A \ar[r] \ar[d] & A \ar[d]^i \\
U_{\sigma} \ar[r] & X.
}
$$
Note that $U_{\sigma} \cap A$ may be empty for $\sigma \in \mathcal{I}_{\mathcal{U}}$, so $\mathcal{I}_{\mathcal{U}_A}$ may be strictly smaller than $ \mathcal{I}_{\mathcal{U}}$; $\mathcal{I}_{\mathcal{U}_A}$ is upward closed in $\mathcal{I}_{\mathcal{U}}$ (viewed as posets). Let $d$ denote the dimension of (the nerve of) $\mathcal{I}_{\mathcal{U}}$ and $d_A$ the dimension of $\mathcal{I}_{\mathcal{U}_A}$.
Assume that the open cover $\mathcal{U}_A$ of $A$ is again amenable (or $0$-truncating). For example, this holds if the inclusion $i \colon A \subseteq X$ is $\pi_1$-injective (for all basepoints) because amenability is preserved under passing to subgroups.
\smallskip
The Moore--Postnikov truncation (at $\pi_1$), applied seperately to the domain and target of an object in $\mathcal{S}^{\to}$, determines a factorization $\widetilde{F}_{\mathcal{U}, \mathcal{U}_A} \to \widetilde{G}:= \mathcal{T}(\widetilde{F}_{\mathcal{U}, \mathcal{U}_A})$. Using our assumptions, it follows that the cochain complexes
$C^{\bullet}_b(G_1(\sigma);\mathbb{R})$ and $C^{\bullet}_b(G_0(\sigma);\mathbb{R})$ are concentrated in degree $0$. As a consequence, the limit in $\mathcal{D}(\mathbb{R})$
$$\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(G_1(-); \mathbb{R})$$ is concentrated in cohomological degrees $\leq d$. On the other hand, we observe that $C^{\bullet}_b(G_0(\sigma);\mathbb{R}) \simeq 0$ if $\sigma \notin \mathcal{I}_{\mathcal{U}_A}$, so the canonical map
$$\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(G_0(-); \mathbb{R}) \to \lim_{\mathcal{I}_{\mathcal{U}_A}^{\mathrm{op}}} C^{\bullet}_b(G_0(-); \mathbb{R})$$
is an equivalence in $\mathcal{D}(\mathbb{R})$ and the latter cochain complex is concentrated in cohomological degrees $\leq d_A$. Hence, by passing to the fibers, the relative comparison map $c_{X,A}$ factors through the cochain complex
$$\mathrm{fib}\big(\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(G_1(-);\mathbb{R}) \longrightarrow \lim_{\mathcal{I}_{\mathcal{U}_A}^{\mathrm{op}}} C^{\bullet}_b(G_0(-);\mathbb{R})\big)$$
whose cohomology is concentrated in degrees $\leq \mathrm{max}\{d, d_A +1\}$. This implies that the relative comparison map (\emph{relative vanishing theorem})
\begin{equation} \tag{5.3(i)}
c^n_{X,A} \colon H^n_b(X, A; \mathbb{R}) \to H^n(X, A;\mathbb{R})
\end{equation}
vanishes for $n > \mathrm{max}\{d, d_A + 1\}$. Applied to the simplicial volume of an oriented compact connected $n$-manifold $(M, \partial M)$
(see also \cite[Proposition 2.18]{LMR}), this vanishing result recovers the relative vanishing theorem in \cite[Theorem 3.13]{LMR} as a special case.
\smallskip
\noindent In addition, noting that the relative comparison map $c_{X, A}$ factors through the cochain complex
\begin{equation} \tag{5.3(ii)}
\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(G_1(-), G_0(-);\mathbb{R}),
\end{equation}
we observe that for $\sigma \in \mathcal{I}_{\mathcal{U}}$:
\begin{itemize}
\item[(a)] $C^{\bullet}_b(G_1(\sigma), G_0(\sigma);\mathbb{R})$ is concentrated in cohomological degrees $\leq 1$;
\item[(b)] $C^{\bullet}_b(G_1(\sigma), G_0(\sigma);\mathbb{R})$ is concentrated in degree $0$ if and only if the map $\pi_0(U_{\sigma} \cap A) \to \pi_0(U_{\sigma})$ is injective;
\item[(c)] $C^{\bullet}_b(G_1(\sigma), G_0(\sigma);\mathbb{R}) \simeq \mathbb{R}$, if $U_{\sigma}$ is path-connected and $U_{\sigma} \cap A$ is empty;
\item[(d)] $C^{\bullet}_b(G_1(\sigma), G_0(\sigma);\mathbb{R}) \simeq 0$ if and only if the map $\pi_0(U_{\sigma} \cap A) \to \pi_0(U_{\sigma})$ is a bijection.
\end{itemize}
As a consequence, if each $U_{\sigma}$ or $U_{\sigma} \cap A$ is either empty or path-connected for all finite $\sigma \subseteq I$, then the relative comparison map $c_{X,A}$ factors canonically through
\begin{align*}
\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(G_1(-), G_0(-); \mathbb{R}) & \simeq \mathrm{fib}\big(\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(G_1(-); \mathbb{R}) \to \lim_{\mathcal{I}_{\mathcal{U}_{A}}^{\mathrm{op}}}C^{\bullet}_b(G_0(-); \mathbb{R}) \big) \\ & \simeq \mathrm{fib}\big(\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(\ast; \mathbb{R}) \to \lim_{\mathcal{I}_{\mathcal{U}_{A}}^{\mathrm{op}}}C^{\bullet}_b(\ast; \mathbb{R}) \big) \\ & \simeq \mathrm{fib}\big(\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}(\ast; \mathbb{R}) \to \lim_{\mathcal{I}_{\mathcal{U}_{A}}^{\mathrm{op}}}C^{\bullet}(\ast; \mathbb{R}) \big) \\ & \simeq \mathrm{fib}\big(C^{\bullet}(|\mathcal{I}_{\mathcal{U}}|; \mathbb{R}) \to C^{\bullet}(|\mathcal{I}_{\mathcal{U}_{A}}|; \mathbb{R}) \big) \\ & \simeq C^{\bullet}(|\mathcal{I}_{\mathcal{U}}|, |\mathcal{I}_{\mathcal{U}_A}|; \mathbb{R}).
\end{align*}
More precisely, a close inspection of these identifications (cf. the proof of Theorem \ref{nerve_theorem_general1}) shows that $c_{X,A}$ admits a canonical factorization in $\mathcal{D}(\mathbb{R})$ (\emph{relative covering theorem}):
\begin{equation} \tag{5.3(iii)}
\xymatrix{
C^{\bullet}_b(X,A; \mathbb{R}) \ar[rr]^{c_{X,A}} \ar[dr] && C^{\bullet}(X,A; \mathbb{R}) \\
& C^{\bullet}(|\mathcal{I}_{\mathcal{U}}|, |\mathcal{I}_{\mathcal{U}_A}|; \mathbb{R}) \ar[ur]_{\ \ \rho_{F_{\mathcal{U}}, F_{\mathcal{U}_A}}} &
}
\end{equation}
where $\rho_{F_{\mathcal{U}}, F_{\mathcal{U}_A}}$ is induced by the corresponding maps $\rho_{F_{\mathcal{U}}}$ and $\rho_{F_{\mathcal{U}_A}}$ (as defined in Theorem \ref{nerve_theorem_general1} and Section \ref{amenable_covers}).
\smallskip
\noindent On the other hand, if (b) is satisfied for all $\sigma \in \mathcal{I}_{\mathcal{U}}$ and the subposet $$\mathcal{J} = \{\sigma \in \mathcal{I}_{\mathcal{U}} \ | \ \sigma \text{ does not satisfy } (d)\} \subseteq \mathcal{I}_{\mathcal{U}}$$ is upward closed of dimension (of its nerve) denoted by $d_{\mathcal{J}}$, then we have a canonical equivalence in $\mathcal{D}(\mathbb{R})$:
\begin{equation} \tag{5.3(iv)}
\lim_{\mathcal{I}_{\mathcal{U}}^{\mathrm{op}}} C^{\bullet}_b(G_1(-), G_0(-); \mathbb{R}) \xrightarrow{\simeq} \lim_{\mathcal{J}^{\mathrm{op}}} C^{\bullet}_b(G_1(-), G_0(-); \mathbb{R})
\end{equation}
and therefore the comparison map $c_{X,A}$ factors through a cochain complex concentrated in cohomological degrees $\leq d_{\mathcal{J}}$. In this case, the relative comparison map
\begin{equation} \tag{5.3(v)}
c^n_{X,A} \colon H^n_b(X, A; \mathbb{R}) \to H^n(X, A;\mathbb{R})
\end{equation}
vanishes for $n > d_{\mathcal{J}}$.
\end{application}
|
\section{Introduction}
\IEEEPARstart{D}{eep} Convolutional Neural Networks (DCNNs) have achieved state-of-the-art performance on a variety of tasks \cite{lecun2015deep, schmidhuber2015deep} and have revolutionized Computer Vision in both classification \cite{Zoph_2018_CVPR, touvron2019FixRes} and representation \cite{Cao18, deng2019arcface}.
In DCNNs, both representation and classification are typically jointly learned in a single network.
The classification layer placed at the end of such models transforms the $d$-dimension of the network internal feature representation to the $K$-dimension of the output class probabilities. Despite the large number of trainable parameters that this layer adds to the model (i.e. $d \times K$), it has been verified that its removal only causes a slight increase
in error \cite{zeiler2014visualizing}.
Moreover, the most recent architectures tend to avoid the use of fully connected layers \cite{lin2013network} \cite{szegedy2015going} \cite{he2016deep}.
It is also well known that DCNNs can be trained to perform metric learning without the explicit use of a classification layer \cite{chopra2005learning} \cite{schroff2015facenet} \cite{hoffer2015deep}.
In particular, it has been shown that excluding from learning the parameters of the classification layer causes little or no decline in performance while allowing a reduction in the number of trainable parameters \cite{hoffer2018fix}.
Fixed classifiers also have an important role in the theoretical convergence analysis of training models with batch-norm \cite{Ioffe17}. Very recently it has been shown that DCNNs with a fixed classifier and batch-norm in each layer establish a principle of equivalence between different learning rate schedules \cite{li2019exponential}.
All these works seem to suggest that the final fully connected layer used for classification is somewhat redundant and does not have a primary role in learning and generalization. In this paper we show that a special set of fixed classification layers \emph{has a key role in modeling the internal feature representation} of DCNNs, while ensuring little or no loss in classification accuracy and a significant reduction in memory usage.
In DCNNs the internal feature representation for an input sample is the feature vector $\mathbf{f}$ generated by the penultimate layer, while the last layer (i.e. the classifier) outputs score values according to the inner product as:
\begin{equation}
z_i = \mathbf{w}_i^\top \cdot \mathbf{f}
\label{eq_logit}
\end{equation}
for each class $i$, where $\mathbf{w}_i$ is the weight vector of the classifier for the class $i$. To evaluate the loss, the scores are further normalized into probabilities via the softmax function \cite{goodfellow2016deep}.
\begin{figure}[t]
\includegraphics[width=0.99\columnwidth]{Figures/IntroFig_cropped.pdf}
\caption{ Regular Polytope Networks (RePoNet).
The fixed classifiers derived from the three regular polytopes available in $\mathbb{R}^d$ with $d \geq 5$ are shown.
From left: the $d$-Simplex, the $d$-Cube and the $d$-Orthoplex fixed classifier. The trainable parameters $\mathbf{w}_i$ of the classifier are replaced with fixed values taken from the coordinate vertices of a regular polytope (shown in red).
}
\label{fig_IntroFig}
\end{figure}
Since the values of $z_i$ can be also expressed as $z_i = \mathbf{w}^\top_i \cdot \mathbf{f} = ||\mathbf{w}_i|| \: ||\mathbf{f}|| \cos(\theta)$, where $\theta$ is the angle between $\mathbf{w}_i$ and $\mathbf{f}$, the score for the correct label with respect to the other labels is obtained by optimizing the length of the vectors $||\mathbf{w}_i||$, $||\mathbf{f}||$ and the angle $\theta$ they are forming.
This simple formulation of the final classifier provides the intuitive explanation of how feature vector directions and weight vector directions align simultaneously with each other at training time so that their average angle is made as small as possible.
\begin{figure*}[htbp]
\hspace{-0.56cm}
\centering
\subfigure[]{\label{fig:a}\includegraphics[height=0.178\linewidth,valign=t]{Figures/mnist_slow_overlap_feature_polygonal2d_raster.pdf}}
\subfigure[]{\label{fig:b}\includegraphics[height=0.178\linewidth,valign=t]{Figures/mnist_slow_overlap_angles_polygonal2d.pdf}}\hspace{1pt}
\subfigure[]{\label{fig:c}\includegraphics[height=0.178\textwidth,valign=t]{Figures/mnist_slow_overlap_feature_trainable_raster.pdf}}%
\subfigure[]{\label{fig:d}\includegraphics[height=0.178\textwidth,valign=t]{Figures/mnist_slow_overlap_angles_trainable.pdf}}\hspace{1pt}%
\raisebox{-0.6pt}{\includegraphics[height=0.1609\textwidth,width=0.047\textwidth,valign=t]{Figures/mnist_slow_overlap_legend.png}}
\caption{Feature learning on the MNIST dataset in a 2D embedding space. Fig.~(a) and Fig.~(c) show the 2D features learned by RePoNet and by a standard trainable classifier respectively.
Fig.~(b) and Fig.~(d) show the training evolution of the classifier weights (dashed) and their corresponding class feature means (solid) respectively. Both are expressed according to their angles. Although the two methods achieve the same classification accuracy, features in the proposed method are both stationary and maximally separated.
}
\label{fig:intro_fig_mnist}
\end{figure*}
If the parameters $\mathbf{w}_i$ of the classifier in Eq.~\ref{eq_logit} are fixed (i.e. set as non-trainable), \emph{only the feature vector directions} can align toward the classifier weight vector directions and not the opposite. Therefore, weights can be regarded as fixed angular references to which features align.
According to this, we obtain a precise result on the spatio-temporal statistical properties of the generated features during the learning phase. Supported by the empirical evidence in \cite{hoffer2018fix} we show that not only the final classifier of a DCNN can be set as non-trainable with no loss of accuracy and with a significant reduction in memory usage, but that an appropriate set of values assigned to its weights allows learning a maximally separated and strictly stationary embedding while training. That is, the features generated by the Stochastic Gradient Descent (SGD) optimization have constant mean and are angularly centered around their corresponding fixed class weights.
Constant known mean implies that features cannot have non-constant trends while learning.
Maximally separated features and their stationarity are obtained by setting the classifier weights according to values following a highly symmetrical configuration in the embedding space.
DCNN models with trainable classifiers are typically convergent
and therefore, after a sufficient learning time has elapsed, some form of stationarity in the learned features can still be achieved. However, until that time, it is not possible to know where the features will be projected by the learned model in the embedding space.
An advantage of the approach proposed in this paper is that it allows to define (and therefore to know in advance) where the features will be projected before starting the learning process.
Our result can be understood by looking at the basic functionality of the final classifier in a DCNN. The main role of a trainable classifier is to dynamically adjust the decision boundaries to learn class feature representations. When the classifier is set as non-trainable this dynamic adjustment capability is no longer available and it is automatically demanded to all the previous layers. Specifically, the work \cite{hoffer2018fix} reports empirical evidence that the expressive power of DCNN models is large enough to account for the missing dynamic adjustment capability of the classifier.
We provide more systematic empirical evidence confirming and broadening the general validity of DCNNs with fixed classifiers (Sec.~\ref{sec_Exchangeability_Assumption}).
We show that our approach can be theoretically justified and easily implemented by setting the classifier weights to values taken from the coordinate vertices of a regular polytope in the embedding space.
Regular polytopes are the generalization in any number of dimensions of regular polygons and regular polyhedra (i.e. Platonic Solids). Although there are infinite regular polygons in $\mathbb{R}^2$ and 5 regular polyhedra in $\mathbb{R}^3$, there are only three regular polytopes in $\mathbb{R}^d$ with $d \geq 5$, namely the $d$-Simplex, the $d$-Cube and the $d$-Orthoplex. Having different symmetry, geometry and topology, each regular polytope will reflect its properties into the classifier and the embedding space which it defines. Fig.~\ref{fig_IntroFig} illustrates the three basic architectures defined by the proposed approach termed Regular Polytope Networks (RePoNet). Fig.~\ref{fig:intro_fig_mnist} provides a first glance at our main result in a 2D embedding space. Specifically, the main evidence from Fig.~\ref{fig:a} and \ref{fig:b} is that the features learned by RePoNet remain aligned with their corresponding fixed weights and maximally exploit the available representation space directly from the beginning of the training phase.
We apply our method to multiple vision datasets showing that it is possible to generate stationary and maximally separated features without affecting the generalization performance of DCNN models and with a significant reduction in GPU memory usage at training time. A preliminary exploration of this work was presented in \cite{Pernici_2019_CVPR_Workshops, pernici2020class}.
\section{Related Work}
\textbf{Fixed Classifier}.
Empirical evidence shows that convolutional neural networks with a fixed classification layer (i.e. not subject to learning) initialized by random numbers does not worsen the performance on the CIFAR-10 dataset \cite{hardt2016identity}. A recent paper \cite{hoffer2018fix} explores in more detail the idea of excluding from learning the parameters $\mathbf{w}_i$ in Eq.\ref{eq_logit}. The work shows that a fixed classifier causes little or no reduction in classification performance for common datasets while allowing a significant reduction in trainable parameters, especially when the number of classes is large. Setting the last layer as not trainable also reduces the computational complexity for training as well as the communication cost in distributed learning. The paper in question sets the classifier with the coordinate vertices of orthogonal vectors taken from the columns of the Hadamard\footnote{The Hadamard matrix is a square matrix whose entries are either $+1$ or $-1$ and whose rows are mutually orthogonal.} matrix and does not investigate on the internal feature representation.
A major limitation of this method is that, when the number of classes is higher than the dimension of the feature space, it is not possible to have mutually orthogonal columns. As a consequence some of the classes are constrained to lie in a common subspace which causes a reduction in classification performance.
In our solution, we improve and generalize this work by finding a novel set of unique directions overcoming the limitations of the Hadamard matrix.
The work \cite{sablayrolles2018spreading} trains a neural network according to the triplet loss with a set of fixed vertices on a hyper-sphere (i.e. a sphere lattice). The work aims at learning a function that maps real-valued vectors to a uniform distribution over a $d$-dimensional sphere.
As shown in \cite{li2019exponential}, fixed classifiers are also related to BatchNorm \cite{Ioffe17} and learning rate schedules.
BatchNorm parametrizes the weights of a layer to “normalize” its activations (i.e. the features), but typically it is not applied to normalize the outputs of the classifier (i.e. logits).
They show that, with BatchNorm layers and a fixed classifier layer, training with $L2$ regularization is equivalent to training with an exponentially increasing learning rate.
\textbf{Softmax Angular Optimization}.
As originally described in \cite{liu2016large}, under softmax loss\footnote{The combination of cross-entropy loss and the softmax function at the last fully connected layer.} the class prediction is \emph{largely} determined by the angular similarity since softmax loss can be factorized, as shown in Eq.~\ref{eq_logit}, into an amplitude component and an angular component.
Several papers have followed this intuition and proposed to train DCNNs by direct angle optimization \cite{ranjan2017l2,Liu2017CVPR,Liu_2018_CVPR,wang2017normface}. The angle encodes the required discriminative information for class recognition. The wider the angles the better the classes are separated from each other and, accordingly, their representation is more discriminative. The common idea of these works is to constrain the features and/or the classifier to be unit normalized.
The works \cite{liu_2017_coco_v1}, \cite{hasnat2017mises} and \cite{wang2017normface} normalize both features and the classifier weights thus obtaining an exact optimization of the angle in Eq.~\ref{eq_logit}.
With weight normalization only, label prediction is largely determined by the angular similarity \cite{liu2016large} \cite{Liu2017CVPR}. This is not only because Eq.~\ref{eq_logit} can be e factorized into amplitude and angular
component, but also because decision boundaries between adjacent classes are determined by their angular bisectors.
Differently from weight normalization, feature normalization cannot directly perform angle optimization but encourages intra-class compactness of learned features \cite{ranjan2017l2}.
Specifically, \cite{ranjan2017l2} also proposes adding a multiplicative scale parameter after feature normalization based on the property that increasing the norm of samples can decrease the softmax loss \cite{wang2017normface,yuan2017feature}.
Although with a different goal than learning discriminative features, the work \cite{hoffer2018fix}, in addition to fixing the classifier, normalizes both the weights and the features and applies the multiplicative scale parameter.
In agreement with \cite{wang2018additive,hoffer2018fix,ranjan2017l2} and \cite{wang2017normface} we found that applying the feature normalization and the multiplicative scale parameter
makes optimization hard with general datasets, having a significant dependence on image quality. According to this, we follow the work \cite{Liu2017CVPR} that normalizes the classifier weights.
Normalizing classifier weights typically also includes setting the classifier biases to zero. As discussed in \cite{wang2017normface} and in \cite{yuan2017feature} this encourages well-separated features to have bigger magnitudes. This avoids features collapsing into the origin, making angles between the weights and features a reliable metric for classification.
As conjectured in \cite{wang2017normface}, if all classes are well-separated, weight normalization will roughly correspond to computing the mean of features in each class.
The maximal and fixed separation proposed in this paper further strengthens the conjecture, producing features more centered around their fixed weights as the training process progresses.
Another close related work to ours is \cite{virtualclass} in which separability of learned features is improved by injecting a single dynamic virtual negative class into the original softmax. A virtual class is a class that is active in the classifier but has no data available from which to learn.
Injecting the virtual class enlarges the inter-class margin and compresses intra-class distribution by strengthening the decision boundary constraint.
In our case, we can profitably exploit virtual classes when the number of classes of the fixed classifier does not match the number of vertices of a regular polytope.
While all the above works impose large angular distances between the classes, they provide solutions to enforce such constraint in a local manner without considering global inter-class separability and intra-class compactness. For this purpose, very recently the works \cite{HypersphericalEnergy2018}, \cite{Zhao_2019_CVPR} and \cite{Duan_2019_CVPR} add a regularization loss to specifically force the classifier weights to be far from each other in a global manner. These works draw inspiration from a well-known problem in physics -- the Thomson problem \cite{thomson1904xxiv}, where given $K$ charges confined to the surface of a sphere, one seeks to find an arrangement of the charges which minimizes the total electrostatic energy. Electrostatic force repels charges each other inversely proportional to their mutual distance. In \cite{HypersphericalEnergy2018}, \cite{Zhao_2019_CVPR} and \cite{Duan_2019_CVPR} global equiangular features are obtained by adding to the standard categorical cross-entropy loss a further loss inspired by the Thomson problem.
In our research, we follow a similar principle for global separability. We consider that minimal energies are often concomitant with special geometric configurations of charges that recall the geometry of Platonic Solids in high dimensional spaces \cite{batle2016generalized}.
We have reported a few preliminary and qualitative results of using Regular Polytope Networks for compact feature learning in \cite{Pernici_2019_CVPR_Workshops}, where we have demonstrated that the angular parameter of the margin loss of \cite{deng2019arcface} can be analytically determined to maximize feature compactness.
\section{Main Contributions:}
Our technical contributions can be summarized as follows:
\begin{enumerate}
\item We generalize the concept of fixed classifiers and show they can generate stationary and maximally separated features at training time with no loss of performance and in many cases with slightly improved performance.
\item We performed extensive evaluations across a range of datasets and modern CNN architectures reaching state-of-the-art performance. We observed faster speed of convergence and a significant reduction in model parameters.
\item We further provide a formal characterization of the class decision boundaries according to the dual relationship between regular polytopes and statistically verify the validity of our method on random permutations of the labels.
\end{enumerate}
\section{Regular Polytopes and Maximally Separated Stationary Embeddings}
We are basically concerned with the following question: \emph{How should the non-trainable weights of the classifier be distributed in the embedding space such that they generate stationary and maximally separated features?}
Let $\mathbb{X} = \{(x_i, y_i)\}_{i=1}^{N}$ be the training set containing $N$ samples, where $x_i$ is the raw input to the DCNN and $y_{i} \in \{1,2,\cdots,K\}$ is the label of the class that supervises the output of the DCNN. Then, the cross entropy loss can be written as:
\begin{equation}
\mathcal{L}=-\frac{1}{N}\sum_{i=1}^{N} \log\Bigg( \frac {\exp({\mathbf{w}_{y_i}^{\top}\mathbf{f}_{i}+\mathbf{b}_{y_i}})} {\sum_{j=1}^{K}\exp({\mathbf{w}_{j}^{\top}\mathbf{f}_{i} + \mathbf{b}_{j}})} \Bigg),
\label{softmax_loss}
\end{equation}
where $\mathbf{W}=\{ \mathbf{w}_j \}_{j=1}^{K}$ are the classifier weight vectors for the $K$ classes.
Following the discussion in \cite{Liu2017CVPR} we normalize the weights and zero the biases ($\hat{\mathbf{w}}_j = \frac{\mathbf{w}_j}{||\mathbf{w}_j||}$, $\mathbf{b}_j=0$) to directly optimize angles, enabling the network to learn angularly distributed features.
Angles therefore encode the required discriminative information for class recognition and the wider they are, the better the classes are represented. As a consequence, the representation in this case is maximally separated when features are distributed at \emph{equal angles} maximizing the available space.
If we further consider the feature vector parametrized by its unit vector as $\mathbf{f}_i=\kappa_i \, \hat{\mathbf{f}}_i$ where $\kappa_i=||\mathbf{f}_i||$ and $\hat{\mathbf{f}}_i = \frac{\mathbf{f}_i}{||\mathbf{f}_i||}$, then Eq.\ref{softmax_loss} can be rewritten as:
\begin{equation}
\mathcal{L}=-\frac{1}{N}\sum_{i=1}^{N}\log\Bigg( \frac {\exp ( { \kappa_i \hat{\mathbf{w}}_{y_i}^{\top}\hat{\mathbf{f}}_i })} {\sum_{j=1}^{K}\exp({ \kappa_i \hat{\mathbf{w}}_{j}^{\top}\hat{\mathbf{f}}_i })} \Bigg)
\label{softmax_loss_von}
\end{equation}
The equation above can be interpreted as if $N$ realizations from a set of $K$ von Mises-Fisher distributions with different concentration parameters $\kappa_i$ are passed through the softmax function. The probability density function of the von Mises-Fisher distribution for the random $d$-dimensional unit vector $\hat{\mathbf{f}}$ is given by: $ P(\hat{\mathbf{f}} ;\hat{\mathbf{w}},\kappa ) \propto \exp \big ( {\kappa \hat{\mathbf{w}}^{\top} \hat{\mathbf{f}} } \big ) $ where $ \kappa \geq 0$. Under this parameterization $\hat{\mathbf{w}}$ is the mean direction on the hypersphere and $\kappa$ is the concentration parameter. The greater the value of $\kappa$ the higher the concentration of the distribution around the mean direction $\hat{\mathbf{w}}$. The distribution is unimodal for $\kappa>0$ and is uniform on the sphere for $\kappa=0$.
As with this formulation each weight vector is the mean direction of its associated features on the hypersphere, equiangular features maximizing the available space can be obtained by arranging accordingly their corresponding weight vectors around the origin. This problem is equivalent to distributing points uniformly on the sphere and is a well-known geometric problem, called Tammes problem \cite{tammes1930origin} which is a generalization of the physic problem firstly addressed by Thomson \cite{thomson1904xxiv}.
In 2D the problem is that of placing $K$ points on a circle so that they are as far as possible from each other. In this case the optimal solution is that of placing the points at the vertices of a regular $K$-sided polygon.
The 3D analogous of regular polygons are Platonic Solids. However, the five Platonic solids are not always the unique solutions of the Thomson problem. In fact, only the tetrahedron, octahedron and the icosahedron are the unique solutions for $K = 4$, $6$ and $12$ respectively. For $K = 8$: the cube is not optimal in the sense of the Thomson problem. This means that the energy stabilizes at a minimum in configurations that are not symmetric from a geometric point of view. The unique solution in this case is provided by the vertices of an irregular polytope \cite{bagchi1997stay}.
The non geometric symmetry between the locations causes the global charge to be different from zero. Therefore in general, when the number of charges is arbitrary, their position on the sphere cannot reach a configuration for which the global charge vanishes to zero. A similar argument holds in higher dimensions for the so called generalized Thomson problem \cite{batle2016generalized}. According to this, we argue that, \emph{the geometric limit to obtain a zero global charge in the generalized Thomson problem is equivalent to the impossibility to learn maximally separated features for an arbitrary number of classes}.
However, since classification is not constrained in a specific dimension as in the case of charges, our approach addresses this issue by \emph{selecting the appropriate dimension of the embedding space so as to have access to symmetrical fixed classifiers directly from regular polytopes}.
In dimensions five and higher, there are only three ways to do that (See Tab.~\ref{tab_polytopes}) and they involve the symmetry properties of the three well known regular polytopes available in any high dimensional spaces \cite{coxeter1963regular}. These three special classes exist in every dimension and are: the $d$-Simplex, the $d$-Cube and the $d$-Orthoplex.
In the next paragraphs the three fixed classifiers derived from them are presented.
\begin{table}[t]
\centering
\caption{Number of regular Polytopes as dimension $d$ increases.}
\begin{tabular}{@{}l|lllll@{}}
\toprule
\toprule
Dimension $d$ & $1$ & $2$ & $3$ & $4$ & $\geq 5$ \\ \midrule
Number of Regular Polytopes & $1$ & $\infty$ & $5$ & $6$ & $\mathbf{3}$ \\
\bottomrule
\bottomrule
\end{tabular}
\label{tab_polytopes}
\end{table}
\textbf{The $d$-Simplex Fixed Classifier.}
In geometry, a simplex is a generalization of the notion of a triangle or tetrahedron to arbitrary dimensions. Specifically, a $d$-Simplex is a $d$-dimensional polytope which is the convex hull of its $d + 1$ vertices. A regular $d$-Simplex may be constructed from a regular $(d-1)$-Simplex by connecting a new vertex to all original vertices by the common edge length. According to this, the weights for this classifier can be computed as:
\begin{equation}
\mathbf{W}_\mathcal{S}=\Big \{e_1,e_2,\dots,e_{d-1}, \alpha \sum_{i=1}^{d-1} e_i \Big \}
\nonumber
\end{equation}
where $\alpha=\frac{1-\sqrt{d+1}}{d}$ and $e_i$ with $i \in \{1,2, \dots, d-1\}$ denotes the standard basis in $\mathbb{R}^{d-1}$. The final weights will be shifted about the centroid and normalized.
The $d$-Simplex fixed classifier defined in an embedding space of dimension $d$ can accommodate a number of classes equal to its number of vertices:
\begin{equation}
K=d+1.
\label{eq_ksimplex}
\end{equation}
This classifier has the largest number of classes that can be embedded in $\mathbb{R}^d$ such that their corresponding class features are equidistant from each other. It can be shown (see appendix) that the angle subtended between \emph{any} pair of weights is equal to:
\begin{equation}
\theta_{\mathbf{w}_i,\mathbf{w}_j}=\arccos\bigg(-\frac{1}{d}\bigg) \quad \forall i,j \in \{ 1, 2, \dots, K\} : i \neq j.
\label{eq_dsimplex_angle}
\end{equation}
\textbf{The $d$-Orthoplex Fixed Classifier.}
This classifier is derived from the $d$-Ortohoplex (or Cross-Polytope) regular polytope that is defined by the convex hull of points, two on each Cartesian axis of an Euclidean space, that are equidistant from the origin. The weights for this classifier can therefore be defined as:
$$\mathbf{W}_\mathcal{O} = \{ \pm e_1,\pm e_2,\dots,\pm e_d \}.$$
Since it has $2d$ vertices, the derived fixed classifier can accommodate in its embedding space of dimension $d$ a number of distinct classes equal to:
\begin{equation}
K=2d.
\label{eq_kortho}
\end{equation}
Each vertex is adjacent to other $d-1$ vertices and the angle between adjacent vertices is
\begin{equation}
\theta_{\mathbf{w}_i,\mathbf{w}_j}= \frac{\pi}{2} \quad \forall \, i,j \in \{ 1, 2, \dots, K\} : j \in C(i)
\label{eq_dortho_angle}
\end{equation}
Where each $j \in C(i)$ is an adjacent vertex and $C$ is the set of adjacent vertices defined as ${C}(i)=\{ j:(i,j) \in {E}\}$. $E$ is the set of edges of the graph ${G}=(\mathbf{W}_\mathcal{O},{E})$.
The $d$-Orthoplex is the dual polytope of the $d$-Cube and vice versa (i.e. the normals of the $d$-Orthoplex faces correspond to the the directions of the vertices of the $d$-Cube).
\textbf{The $d$-Cube Fixed Classifier.}
The $d$-Cube (or Hypercube) is the regular polytope formed by taking two congruent parallel hypercubes of dimension $(d-1)$ and joining pairs of vertices, so that the distance between them is $1$. A $d$-Cube of dimension 0 is one point.
The fixed classifier derived from the $d$-Cube
is constructed by creating a vertex for each binary number in a string of $d$ bits. Each vertex is a $d$-dimensional boolean vector with binary coordinates $-1$ or $1$. Weights are finally obtained from the normalized vertices:
\begin{equation}
\mathbf{W}_\mathcal{C} = \left \{ \mathbf{w} \in \mathbb {R}^{d}: \left [-\frac{1}{\sqrt{d}},\frac{1}{\sqrt{d}} \right ] ^d \right \}.
\nonumber
\end{equation}
The $d$-Cube can accommodate a number of distinct classes equal to:
\begin{equation}
K=2^d.
\label{eq_kcube}
\end{equation}
The vertices are connected by an edge whenever the Hamming distance of their binary numbers is one therefore forming a $d$-connected graph.
It can be shown (see appendix) that the angle between a vertex with its adjacent (i.e. connected) vertices is:
\begin{equation}
\theta_{\mathbf{w}_i,\mathbf{w}_j} = \arccos\bigg(\frac{d-2}{d}\bigg), \forall \, i,j \in \{ 1, \dots, K\} : j \in C(i)
\label{eq_dcube_angle}
\end{equation}
where $C(i)$ is the set of vertices adjacent to vertex $i$.
Fig.~\ref{fig:anglefcn} shows the angle between a weight and its adjacent weights computed from Eqs.~\ref{eq_dsimplex_angle}, \ref{eq_dortho_angle} and \ref{eq_dcube_angle} as the dimension of the embedding space increases. Having the largest angle between the weights, the $d$-Simplex fixed classifier achieves the best inter-class separability. However, as the embedding space dimension increases, its angle tends towards $\pi/2$. Therefore, the larger the dimension is, the more similar it becomes to the $d$-Orthoplex classifier. The main difference between the two classifiers is in their neighbor connectivity. The different connectivity of the three regular polytope classifiers has a direct influence on the evaluation of the loss. In the case of the $d$-Simplex classifier, all the summed terms in the loss of Eq.~\ref{softmax_loss_von} have always comparable magnitudes in a mini batch.
The $d$-Cube classifier has the most compact feature embedding and the angle between each weight and its $d$ neighbors decreases as the dimension increases. Due to this, it is the hardest to optimize.
\begin{figure}[t]
\centering
\includegraphics[width=0.99\columnwidth]{Figures/AnglesRegularPolytopes.pdf}
\caption{The angular space defined by RePoNet classifiers. Curves represent the angle between a weight and its adjacent weights as the dimension of the embedding space increases. The angle between class features follows the same trend.}
\label{fig:anglefcn}
\end{figure}
\subsection{Implementation}
Given a classification problem with $K$ classes, the three RePoNet fixed classifiers can be simply instantiated by defining a non-trainable fully connected layer of dimension $d$, where $d$ is computed from Eqs.~\ref{eq_ksimplex}, \ref{eq_kortho} and \ref{eq_kcube} as summarized in Tab~\ref{tab_layer_dim}.
\begin{table}[h]
\centering
\caption{Feature dimension $d$ as a function of the number of classes $K$.}
\begin{tabular}{@{}c|ccl@{}}
\toprule
\toprule
RePoNet & $d$-Simplex & $d$-Cube & $d$-Orthoplex \\ \midrule
Layer dim. & $d=K-1$ & $d = \lceil \log_2(K) \rceil$ & $d = \big \lceil \frac{K}{2} \big \rceil$ \\ \bottomrule \bottomrule
\end{tabular}
\label{tab_layer_dim}
\end{table}
\FloatBarrier
In order to accommodate different CNN architectures having different convolutional activation output size (e.g., from the 2048 size of the ResNet50 to the feature size of 10 of the fixed $d$-Cube classifier with the 1000 classes of ImageNet), a middle ``junction'' linear layer (without ReLu) is required.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{Figures/virtual_classes.pdf}
\caption{Learning with unassigned classes in a 2D embedding space.
The features of the first four digits of the MNIST dataset are learned using a 10-sided regular polygon in which six of the classes are virtual. Unassigned classes (colored lines inside the shaded region) force a large angular margin region (shaded region) from which features are pushed out.
}
\label{virtual_classes}
\end{figure}
\subsection{Exceeding Vertices as Virtual Negative Classes}
Except for the $d$-Simplex that allows to assign all its vertices for a given number of classes $K$, for both the $d$-Cube and the $d$-Orthoplex classifiers some of the vertices may be in excess for a given number of classes. As implied by Eq.~\ref{eq_kortho}, in the case of the $d$-Orthoplex one vertex remains unassigned when the number of classes $K$ is odd. In the case of the $d$-Cube classifier, due to the exponential dependency in Eq.~\ref{eq_kcube}, a large number of vertices may remain not assigned. For example, assuming $K=100$ the $d$-Cube fixed classifier has $128$ vertices (see Tab.~\ref{tab_layer_dim}) and $28$ of them are not assigned to any class.
As shown in \cite{virtualclass}, unassigned classes act as virtual negative classes forcing a margin around the unassigned weights without affecting the correctness of softmax based cross entropy optimization.
Virtual negative classes do not change substantially the objective function of Eq.~\ref{softmax_loss_von} that can be rewritten as:
\begin{equation}
\resizebox{0.88\hsize}{!}{$
{\displaystyle
\mathcal{L}=-\frac{1}{N}\sum_{i=1}^{N}\log\Bigg( \frac {\exp ( { \kappa_i \hat{\mathbf{w}}_{y_i}^{\top}\hat{\mathbf{f}}_i })} {\sum_{j=1}^{K}\exp({ \kappa_i \hat{\mathbf{w}}_{j}^{\top}\hat{\mathbf{f}}_i }) + \sum_{j=K+1}^{K_V}\exp({ \kappa_i \hat{\mathbf{w}}_{j}^{\top}\hat{\mathbf{f}}_i })} \Bigg)
}
$}
\label{softmax_loss_virtual}
\end{equation}
where $K_{V}$ is the number of virtual classes (i.e. the exceeding polytope vertices).
Fig.~\ref{virtual_classes} illustrates an example similar to Fig.~\ref{fig:intro_fig_mnist}(a) in which a 10-sided polygon fixed classifier is learned to classify the first four digits of the MNIST dataset ($0,1,2$ and $3$). The remaining six ``empty slots'' of the classifier are not assigned to any class data and therefore the classifier acts as a virtual negative classifier forcing a large margin (the shaded region) around the virtual class weights (colored lines).
This result generalizes the proposed method to any arbitrary number of classes.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{Figures/polygon_duality.pdf}
\caption{ The intuition behind the decision boundaries in RePoNet ($10$-sided polygon).
The bisector directions (dotted lines), represent the class decision boundaries. They have the same direction of the normal of the corresponding polygon side (only one shown for clarity).
The decision boundaries form a regular polygon that is related with the classifier $10$-sided polygon according to duality.
For clarity, only one class region is highlighted (shaded region).}
\label{polygonduality}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[width=0.79\columnwidth]{Figures/poly_2.pdf}
}
\\
\subfigure[]{
\includegraphics[width=0.79\columnwidth]{Figures/poly_3.pdf}
}
\\
\subfigure[]{
\includegraphics[width=0.79\columnwidth]{Figures/poly_1.pdf}
}
\caption{The RePoNet fixed classifiers decision boundaries in a 3D embedding space: (\emph{a}): $d$-Orthoplex classifier; (\emph{b}): $d$-Cube classifier; (\emph{c}): $d$-Simplex classifier. On the \emph{left}: the regular polytope classifier (light blue); its dual polytope (grey); a classifier weight $\mathbf{w}$ (red) and its edge decision boundaries $\mathbf{v}_1, \mathbf{v}_2, \dots$~(black). On the \emph{right}: the same entities on the unit sphere. The yellow region indicates where class features are located (only one class weight and the corresponding edge decision boundaries are shown for clarity). The characterization extends to arbitrary dimensions.
}
\label{Duality}
\end{figure}
\subsection{Fixed Classifier Decision Boundaries}
In binary-classification, the posterior probabilities obtained by softmax in Eq.\ref{softmax_loss_von} are:
\begin{equation}
p_{1}=\frac{\exp ( { \kappa \hat{\mathbf{w}}_{1}^{\top}\hat{\mathbf{f}} })}{\exp ( { \kappa \hat{\mathbf{w}}_{1}^{\top}\hat{\mathbf{f}} })+\exp ( { \kappa \hat{\mathbf{w}}_{2}^{\top}\hat{\mathbf{f}} })}
\end{equation}
\begin{equation}
p_{2}=\frac{\exp ( { \kappa \hat{\mathbf{w}}_{2}^{\top}\hat{\mathbf{f}} })}{\exp ( { \kappa \hat{\mathbf{w}}_{1}^{\top}\hat{\mathbf{f}} })+\exp ( { \kappa \hat{\mathbf{w}}_{2}^{\top}\hat{\mathbf{f}} })}
\end{equation}
where $\mathbf{f}$ is the learned feature vector and $\mathbf{w}_1$ $\mathbf{w}_2$ are the fixed classifier weights. The predicted label will be assigned to the class 1 if $p_{1}>p_{2}$ and to the class 2 if $p_{1}<p_{2}$. By comparing the two probabilities $p_{1}$ and $p_{2},$ $ { \kappa \hat{\mathbf{w}}_{1}^{\top}\hat{\mathbf{f}} } + { \kappa \hat{\mathbf{w}}_{2}^{\top}\hat{\mathbf{f}} }
$
determines the classification result. The decision boundary is therefore
$\kappa \hat{\mathbf{w}}_{1}^{\top}\hat{\mathbf{f}} + \kappa \hat{\mathbf{w}}_{2}^{\top}\hat{\mathbf{f}} = 0$. Due to weight normalization the posterior probabilities result in $p_1 = \kappa ||\hat{\mathbf{f}}|| cos(\theta_1)$ and $p_2 = \kappa ||\hat{\mathbf{f}}|| cos(\theta_2)$ and since $p_1$ and $p_2$ share the same feature $\hat{\mathbf{f}}$ the equation $cos(\theta_1)-cos(\theta_2)=0$ is verified at the \emph{angular bisector} between $\mathbf{w}_{1}$ and $\mathbf{w}_{2}$.
Although the above analysis is built on binary-class case, it can be generalized to the multi-class case \cite{Liu2017CVPR}.
In RePoNet angular bisectors define class decision boundaries that follow a symmetry similar to that of the regular polytope defining the classifier. Specifically, the class decision boundaries and the weights of the classifier are related by the duality relationship that holds between regular polytopes. More practically:
\begin{itemize}
\item the set of decision boundaries of the $d$-Simplex classifier is shaped as a $d$-Simplex;
\item the set of the decision boundaries of the $d$-Cube classifier is shaped as a $d$-Orthoplex;
\item the set of the decision boundaries of the $d$-Orthoplex classifier is shaped as a $d$-Cube.
\end{itemize}
Class decision boundaries are still defined as a regular polytope and the features located within such boundaries are therefore maximally separated. The basic intuition behind this result can be better appreciated in 2D exploiting the well known result that all the regular polygons are self-dual \cite{coxeter1963regular}.
That is, the normal of each side of a regular polygon is parallel to the direction from the origin towards the vertex of its dual polygon. Fig.~\ref{polygonduality} shows the example introduced in Fig.~\ref{virtual_classes} in which decision boundaries are highlighted with dotted lines according to the dual regular polygon. Fig.~\ref{Duality} illustrates the duality relationship between the weights of the three fixed classifiers proposed and their decision boundaries in the 3D embedding space.
\section{Experimental Results}
We evaluate the correctness (Sec.~\ref{sec_Exchangeability_Assumption}) and the no loss of performance of our approach with respect to standard baselines using trainable and fixed classifiers across a range of datasets and architectures (Sec.~\ref{Generalization and Performance Evaluation}).
All the experiments are conducted with the well known MNIST, FashionMNIST \cite{FashionMNIST2017}, EMNIST \cite{EMNIST17}, CIFAR-10, CIFAR-100 \cite{cifar-10} and ImageNet (ILSVRC2012) \cite{deng2009imagenet} datasets.
We chose several common CNN architectures (i.e. LeNet, VGG, ResNet, DenseNet), as well as more recent ones (i.e. SeResNeXt50 \cite{Hu_2018_CVPR}, SkResNeXt50 \cite{Li_2019_CVPR} and EfficientNet \cite{DBLP:conf/icml/TanL19})
that have shown to improve performance while maintaining or, in some cases, reducing computational complexity and model size.
\begin{figure}[t]
\hspace{-0.5cm}
\includegraphics[width=0.5\textwidth]{Figures/shuffle_acc_mnist_cifar.pdf}
\caption{Class permutation verification. Average accuracy curves and confidence interval computed from the MNIST, CIFAR-10 and CIFAR-100 datasets (from top to bottom, respectively) under different random permutations of the ground truth labels position.
}
\label{fig_exchangeability}
\end{figure}
\subsection{Hard Permutations Verification}
\label{sec_Exchangeability_Assumption} Since fixed classifiers cannot rely on an adjustable set of subspaces for class feature representation, we verified if some permutations are harder than others for our proposed method. The presence of such hard permutations would preclude the general applicability of our method.
The standard trainable classifier does not suffer from this problem, when features cannot be well separated trainable classifiers can rearrange the feature subspace directions so that the previous convolutional layers can better disentangle the non-linear interactions between complex data patterns. Instead, fixed classifiers demand this capability to all the previous layers.
According to this, we generate random permutations of the ground truth label positions\footnote{This is equivalent to randomly permuting the classifier weight vectors set $\mathbf{W}=\{ \mathbf{w}_j \}_{j=1}^{K}$. } and a new model is learned for each permuted dataset. Fig.~\ref{fig_exchangeability} shows the mean and the $95\%$ confidence interval computed from the accuracy curves of the learned models. To provide further insight into this analysis, 20 out of 500 accuracy curves computed for each dataset are also shown. Specifically, the evaluation is performed on three different datasets with an increasing level of complexity (i.e MNIST, CIFAR-10 and CIFAR-100).
All the models are trained for 200 epochs to make sure that the models trained with CIFAR-100 achieve convergence.
\begin{figure}[t]
\vspace{-0.2cm}
\centering
\hspace{-0.65cm}
\subfigure[]{\label{fig_even_odd:a}
\includegraphics[height=0.47\linewidth,valign=t]{Figures/mnist_even_odd_feature_reordered_duallegend_raster.pdf}
}
\subfigure[]{\label{fig_CIFAR10:b}
\includegraphics[height=0.47\linewidth,valign=t]{Figures/cifar10_polygonal2d_feature.pdf}
}
\caption{ The distribution of features learned using a 10-sided regular polygon. \emph{(a)}: A special permutation of classes is shown in which the MNIST even and odd digits are placed in the positive and negative half-space of the abscissa respectively. \emph{(b)}: The features learned using the CIFAR-10 dataset.
}
\label{fig_even_odd}
\end{figure}
In order to address the most severe possible outcomes that may happen, for this experiment we used the $d$-Cube fixed classifier. Being the hardest to optimize, this experiment can be regarded as a worst case analysis scenario for our method. As shown in the same figure, the performance is substantially insensitive to both permutations and datasets. The average reduction in performance at the end of the training process is negligible and the confidence intervals reflect the complexity of the datasets. Although the space of permutations cannot be exhaustively evaluated even for a small number of classes, we have achieved proper convergence for the whole set of 1500 learned models.
The experiment took 5 days on a \mbox{Nvidia DGX-1}.
On the basis of this evidence, we can conclude that fixing the classifier (therefore not having access to a set of adjustable subspaces for class feature representation) does not affect the expressive power of neural networks.
This experiment also provides a novel and more systematic empirical evidence of the general applicability and correctness of fixed classifiers
with respect to \cite{hoffer2018fix} where only one permutation was tested.
\input{tables/lenet}
\input{tables/cifar_10}
We finally report qualitative results of a learned permuted dataset. Fig.~\ref{fig_even_odd}(a) shows features learned in a $k$-sided polygon (2d embedding space) on the MNIST dataset. In particular the model is learned with a special permutation of the labels (manually selected) that places even and odd digits features respectively on the positive and negative half space of the abscissa.
Fig.~\ref{fig_even_odd}(b) shows the features of on \textsc{CIFAR-10} learned with a similar $10$-sided-polygon. It can be noticed that features are distributed following the same polygonal pattern shown in Fig.~\ref{fig_even_odd}(a).
\input{tables/cifar_100}
\subsection{Generalization and Performance Evaluation}
\label{Generalization and Performance Evaluation}
Having verified that the order position of the class labels does not adversely affect the proposed method, in this section we evaluate the classification performance of RePoNet on the following datasets: MNIST, EMNIST, FashionMNIST, CIFAR-10, CIFAR-100 and ImageNet.
The RePoNet method is compared with CNN baselines with learned classifiers and the fixed classifier method reported in \cite{hoffer2018fix}, that has been implemented for different architectures and different dimensions of the embedding space. Except for the final fixed classifier all the compared methods have exactly the same architecture and training settings as the one that RePoNet uses.
\\
\subsubsection{\textbf{{MNIST and CIFAR}}}
We trained the so called LeNet++ architecture \cite{wen2016discriminative} on all the MNIST family datasets. The network is a modification of the LeNet \cite{lecun1998gradient} to a deeper and wider network including parametric rectifier linear units (pReLU) \cite{he2015delving}.
For the evaluation on the CIFAR-10 and CIFAR-100 datasets, we further trained VGG \cite{SimonyanZ14a} with depth 13 and 19, ResNet50 \cite{he2016deep},
SeNet \cite{hu2018squeeze} and DenseNet169 \cite{Densenet2017}.
Popular network architectures for ImageNet require modifications to adapt to the CIFAR 32x32 input size. According to this, our experiments follow the publicly available implementations\footnote{
\url{https://github.com/bearpaw/pytorch-classification} and \url{https://github.com/kuangliu/pytorch-cifar} }.
We compared all the variants of our approach for each architecture including trainable classifiers with different dimensions of the feature space.
The mini batch size is 256 for both the MNIST family datasets and the CIFAR-10/100 datasets.
For the CIFAR datasets, we compared both a ``vanilla'' learning setup with no hyperparameters tuning based on the Adam optimizer (learning rate $0.0005$) for the VGG architectures and a learning setup based on SGD with a specific learning rate schedule (starting from 0.1 and decreasing by a factor of 10 after 150 and 250 epochs) for ResNet50, SEnet18 and DenseNet169 architectures.
As hyperparameters tuning is an integral part of Deep Learning we provided two opposite learning setup.
\\
\indent
Test-set accuracy for this experiment is reported in Tab.~\ref{table:lenetpp}, \ref{table:cifar10} and \ref{table:cifar100} for MNISTs, CIFAR-10 and CIFAR-100, respectively.
In addition to the well-known MNIST and FashionMnist, we included EMNIST dataset having $47$ classes including lower/upper case letters and digits. This allows to quantify with a specific dataset and architecture, as in CIFAR-10 and CIFAR-100, the classification accuracy with a higher number of classes.
Each entry in the tables report the test-set accuracy. The subscript indicates the specific feature space dimension $d$ used. The results reveal and confirm that the RePoNet method achieves comparable classification accuracy of other trainable classifier models. This evidence is in agreement on all the combinations of datasets, architectures, number of classes and feature space dimensions considered. All the RePoNet variants exhibit similar behavior even in hard combinations such as the \textsc{CIFAR-100} dataset in a low dimensional feature space. For example, the RePoNet $d$-Cube fixed classifier implemented with the VGG19 architecture achieves an accuracy of $65.32\%$ in a $d=7$ dimensional feature space. A fully trainable classifier in a feature space of dimension $d=512$ (i.e. two orders of magnitude larger), achieves a moderate improvement of about $3\%$ ($68.47\%$). On the other hand, with a significantly lower feature dimension of $d=50$, RePoNet $d$-Orthoplex improves the accuracy to $69.76\%$.
All the RePoNet variants exhibit similar behavior also in the case of more sophisticated architectures trained with SGD scheduled learning rates to match state-of-the-art performance.
RePoNet classifiers are both agnostic to architectures and training setup and are able to improve accuracy similar to trainable classifiers.
Results also show that the Hadamard fixed classifier \cite{hoffer2018fix} does not succeed to learn when the number of classes is larger than the number of unique weight directions in the embedding space (i.e. $d < K$). As expected, this effect is present for simple datasets as the MNIST digits dataset, however as reported in \cite{hoffer2018fix} Section 4.2 (Possible Caveats) as the number of classes $K$ increases the effect is less pronounced.
When $d \! \approx \! K$ or $d \! > \! K$, classification performance is similar. However, as shown in Fig.~\ref{fig_learning_speed}(a)
RePoNet converges faster than \cite{hoffer2018fix}, and with the same speed as the trainable baselines. Our conjecture is that with our symmetrical fixed classifiers, each term in the loss function tends to have the same magnitude centered around the mean of the distribution (i.e. the von Mises-Fisher distribution is similar to the Normal distribution) and therefore the average computed in the loss is a good estimator. Instead, in the Hadamard classifier the terms may have different magnitudes and ``important'' errors in the loss may not be taken into account correctly by simple averaging.
\subsubsection{\textbf{ImageNet}}
Finally, we evaluated our method
on the 1000 object category classification problem defined by the ImageNet dataset. This dataset consists of a 1.2M image training set and a 100k image test set.
We compared all the variants of our approach on different combinations of architectures and their relative trainable classifiers. The comparison also includes the Hadamard classifier.
\\
\indent
Experiments have been conducted in two different configurations of the training hyperparameters.
\emph{First}, we performed experiments using the Adam optimizer and simple augmentation based on random cropping and horizontal flipping on well-established networks such as ResNet50 \cite{he2016deep} and DenseNet169 \cite{Densenet2017}. The learning rate is automatically adjusted when a plateau in model performance is detected.
We trained for $250$ epochs with batch size $64$ with an initial learning rate of $0.0005$.
With this configuration, we aim to evaluate our method without performing any specific hyperparameter optimization or exploiting large computational resources.
\emph{Second}, we evaluated our method with more sophisticated CNN architectures, namely SKresNeXt, SEresNeXt, and EfficientNet with related training hyperparameters. With this configuration, the aim is to evaluate whether our method can reach state-of-the-art performance.
The SKresNeXt and SEresNeXt architectures integrate the SE and SK blocks, \cite{hu2018squeeze} and \cite{li2019selective} respectively, with the ResNeXt architecture \cite{xie2017aggregated}. The benefit of these variants is to maintain computational complexity and model size similar to the SEnet and SKnet architectures while further improving performance. The third architecture, EfficientNet \cite{DBLP:conf/icml/TanL19}, achieves state-of-the-art performance using significantly fewer parameters than other state-of-the-art models.
As these architectures typically require a
large effort to tune the training hyperparameters, we trained our method on top of these models following the settings reported in the original papers.
Specifically, we train EfficentNet-B2 following \cite{DBLP:conf/icml/TanL19}: RMSProp optimizer with decay 0.9 and momentum 0.9; batch norm momentum 0.99; initial learning rate 0.256 that decays by 0.97 every 2.4 epochs; weight decay 1e-5.
Analogously, SKresNeXt50 and SEresNeXt50 are trained following the ResNeXt50 \cite{xie2017aggregated}: SGD optimizer, weight decay 0.0001; momentum 0.9; initial learning rate of 0.1, divided by 10 for three times using a specific schedule reported in the paper.
For all the three models we used automated data augmentation techniques from \cite{cubuk2020randaugment} (RandAugment) with distortion magnitude $7$. SeResNeXt50 and SkResNeXt50 were trained for $250$ epochs with $192$ batch size. EfficientNet-B2 was trained for $450$ epochs with $120$ batch size. Our evaluation is based on the pytorch-image-models\footnote{\url{https://github.com/rwightman/pytorch-image-models}} repository.
\input{tables/imagenet}
Tab.~\ref{table:imagenet} summarizes our results.
As can be clearly noticed, except for the $d$-Cube there is no substantial difference between the performance of our fixed classifiers and the learned classifiers. This holds also in the case of the learned classifiers in their original architecture implementation (shown in the bottom line of the Table). The table also shows that RePoNet accuracy is comparable with the Hadamard fixed classifier \cite{hoffer2018fix}.
As in the cases of CIFAR-10 and CIFAR-100, also with ImageNet the accuracy of the \mbox{$d$-Cube} is lower than the corresponding learned classifiers. We argue this is mainly due to the difficulty of performing optimization in the $d=10$ dimensional space
due to the fact that the angle between each class weight vector and its $d$ adjacent weight vectors approaches to zero as the dimension increases (Fig.~\ref{fig:anglefcn}).
However,
the $d$-Cube classifier shows the largest relative improvement as the representational power of the architecture increases (left to right).
For example, the accuracy of the \mbox{$d$-Cube-\textsc{EfficientNet-B2}} fixed classifier is $12.18$ percentage points larger than the \mbox{$d$-Cube-\textsc{ResNet50}} (i.e. $75.62-63.44 = 12.18$). This relative performance improvement is substantially higher than that of the corresponding learned classifier (i.e. $77.42 - 68.82 = 8.6$). This result is quantitatively consistent with the underlying assumption of this paper and provides further support on the fact that the adjustable capability of the final classifier can be successfully demanded to previous layers.
The other two RePoNet variants substantially achieve the same accuracy of the learned classifiers, irrespective whether they have similar ($d= \{ 10, 500, 999 \}$) or higher feature space dimension ($d= \{ 2048, 1669, 1408 \}$) as in their original architecture implementations. They do not show sensible relative performance improvement with increasing representational power of the network. More importantly, both the $d$-Simplex and the $d$-Orthoplex classifiers reach state-of-the-art accuracy (around 80\%) when combined with competitive architectures. This confirms the validity and the absence of a loss of generalization of our method.
\begin{table}[b]
\caption{The number of parameters of each network and the percentage (\%) of saved parameters on the ImageNet dataset ($d$ indicates the feature dimension).}
\centering
\resizebox{.49\textwidth}{!}{%
\begin{tabular}{lcccc}
\toprule
& & & \textsc{Saved Params ( \% )} & \\
\cmidrule(lr){3-5}
\textsc{Architecture} & \textsc{Param\#} & \textsc{$d$-Cube} & \textsc{$d$-Orthoplex} & \textsc{$d$-Simplex} \\
\midrule
DenseNet169 & 14.15M & $11.65_{d=10}$ & $5.89_{d=500}$ & $0.02_{d=999}$ \\
ResNet50 & 25.56M & $7.94_{d=10}$ & $4.01_{d=500}$ & $0.01_{d=999}$ \\
SeResNeXt50 & 27.56M & $7.36_{d=10}$ & $3.72_{d=500}$ & $0.01_{d=999}$ \\
SkResNeXt50 & 27.50M & $7.38_{d=10}$ & $3.73_{d=500}$ & $0.01_{d=999}$ \\
EfficientNet-B2 & 9.11M & $15.31_{d=10}$ & $7.74_{d=500}$ & $0.03_{d=999}$ \\
\bottomrule
\end{tabular}%
}
\label{params1}
\label{params2}
\end{table}
Finally, Tab.~\ref{params2} shows the total number of parameters for each network in comparison with their original learned classifiers (i.e. bottom line in Tab.~\ref{table:imagenet}). The $d$-Orthoplex-\textsc{EfficinetNet-B2} fixed classifier saves $7.74\%$ of the network parameters while achieving the same accuracy (around 80\%).
It is worth to notice that the $d$-Cube-\textsc{EfficinetNet-B2} with 7.7M of parameters ($15.31\%$ savings) achieves similar accuracy of a vanilla ResNet50 baseline (i.e. around 75\% accuracy) having 25.5M of parameters.
\subsection{Training Time}
The time it takes to train a neural network model to address a classification problem is typically considered as the product of the training time per epoch and the number of epochs which need to be performed to reach the desired level of accuracy \cite{DBLP:conf/bigdataconf/JustusBBM18}.
Although in our case the training time per epoch is lower (the weights of the fixed classifier do not require back-propagation), it has a negligible effect due to the number of epochs required to reach a reasonable desired level of accuracy.
In Fig.~\ref{fig_learning_speed} and Fig.~\ref{fig:accuracy_seresnext} we report the classification accuracy over the epochs for
the two different configurations of the training hyperparameters we evaluated.
\\
\indent
Specifically, Fig.~\ref{fig_learning_speed}(a)(\emph{top}) and Fig.~\ref{fig_learning_speed}(a)(\emph{bottom}) show the training error and the classification accuracy, respectively, over the epochs. The curves are obtained on the CIFAR100 dataset, using the VGG19 architecture and training is performed according to the Adam stochastic optimization.
Fig.~\ref{fig_learning_speed}(b) shows the accuracy curves of the proposed three fixed classifiers and the best performing learned classifier (i.e. $d=999$). The curves are obtained on the ImageNet dataset with the DenseNet169 architecture and learned according to the Adam optimizer. As can be noticed, $d$-Simplex and $d$-Orthoplex classifiers have lower or equal time to reach any desired level of accuracy than the learned and Hadamard fixed classifiers. The $d$-Cube classifier is the slowest and does not reach a comparable final performance. This is due to the different feature dimension ($d=10)$ and topology. However, when compared with a learned classifier with same feature dimension (as discussed in the next paragraph) the training time is similar.
\\
\indent
Fig.~\ref{fig:accuracy_seresnext}(a) and Fig.~\ref{fig:accuracy_seresnext}(b) show the time to reach accuracy using SGD+RandAug on SeResNeXt50 for the $d$-Simplex and $d$-Orthoplex (in red) fixed classifiers and the learned classifiers (in blue), respectively.
As evidenced in the figure, the learned classifiers require about 150 epochs to obtain the accuracy that the $d$-Simplex and $d$-Orthoplex achieve in 50 and 90 epochs, respectively. Although this gain reduces as training progresses towards the end, our method achieves consistently better results.
Fig.~\ref{fig:accuracy_seresnext}(c)
shows that the $d$-Cube classifier requires similar training time with slightly lower final accuracy.
\\
\indent
The general behavior of the curves shown in Fig.~\ref{fig_learning_speed} and Fig.~\ref{fig:accuracy_seresnext}
is consistent across combinations of datasets, architectures, classifiers and training strategies. The training time to reach the same accuracy is shorter or equal for our method and the time reduction follows the complexity of the embeddings defined by each regular polytope fixed classifier.
\indent
\\
\indent
Overall, we have demonstrated that Regular Polytope Networks provide a novel, effective and easy approach to fixed classifiers that achieves comparable state-of-the-art performance with the standard trainable classifiers. They provide faster speed of convergence and a significant reduction in model parameters. To facilitate replication of our experiments, the code will be made publicly available.
\begin{figure}[t]
\centering
\subfigure[]{
\hspace{-0.5cm}
\includegraphics[width=0.99\columnwidth]{Figures/cifar100_speed_loss_accuracy_bigger_n.pdf}
\vspace{.2cm}
}
\subfigure[]{
\hspace{-0.5cm}
\vspace{0.1cm}
\includegraphics[width=0.49\textwidth]{Figures/imagenet_rebuttal_final_bigger_fixed_n.pdf}
}
\vspace{-0.09cm}
\caption{Speed of convergence comparison (\textsc{Adam}). \emph{(a)}: Training error curves (\emph{top}) and test accuracy curves (\emph{bottom}) using the CIFAR-100 dataset with the VGG19 architecture. \emph{(b)}: ImageNet learning speed using DenseNet169. As evidenced from the figures, the proposed method has faster convergence.
}
\label{fig_learning_speed}
\end{figure}
\begin{figure}[t]
\centering
\vspace{.02cm}
\subfigure[]{
\hspace{-0.5cm}
\includegraphics[width=0.49\textwidth]{Figures/accuracy-seresnext50-1.pdf}
\vspace{-.1cm}
}
\subfigure[]{
\hspace{-0.5cm}
\includegraphics[width=0.49\textwidth]{Figures/accuracy-seresnext50-2.pdf}
\vspace{-.1cm}
}
\subfigure[]{
\hspace{-0.5cm}
\includegraphics[width=0.49\textwidth]{Figures/accuracy-seresnext50-3.pdf}
}
\caption{Speed of convergence comparison (\textsc{Sgd+RandAug}). Test accuracy curves over the epochs on the ImageNet test set for the SeResNeXt50 architecture using the proposed fixed classifiers (red) and the standard trainable baselines (blue). \textit{(a):} The $d$-Simplex classifier, \textit{(b):} the $d$-Orthoplex classifier and \textit{(c):} the $d$-Cube classifer. The time to reach the same accuracy is shorter or equal for our method.}
\label{fig:accuracy_seresnext}
\end{figure}
\section{Discussion: Potential and Challenges}
Our finding may have implications in those Deep Neural Network learning contexts in which a classifier must be robust against changes of the feature representation while learning. This is the case of incremental learning settings, especially when features are stored in memory banks while learning \cite{DBLP:journals/corr/GravesWD14, DBLP:journals/corr/WestonCB14,graves2016hybrid}. Despite recent advances, methods inspired by memory-augmented deep neural networks are still limited when it comes to incremental learning. The method \cite{DBLP:conf/iclr/KaiserNRB17} simplifies the original fully differentiable end-to-end idea. Except for the nearest-neighbor query to the memory bank, their approach is fully differentiable, can be trained end-to-end and operates in a incremental manner (i.e. without the need of resetting during training). However, the features stored in the memory bank remain fixed (i.e. they are not undergoing learning) and only the memory bank is learned.
Our approach may have a promising potential for learning both the feature and the memory without considering their joint learning. The intuition is that every time the internal feature representation changes the memory bank must be relearned from scratch. Our method can mitigate the need of feature relearning by keeping the compatibility of features between learning steps thanks to their stationarity. Concurrent to this work, \cite{Shen_2020_CVPR} addresses a similar problem in terms of feature ``back-compatibility'' and exploits a pre-trained fixed classifier to avoid re-indexing a memory bank containing the gallery features of a retrieval system that has been updated.
\\
\indent
This basic idea can be in principle applied to the many computer vision tasks that have benefited from memory based learning. Among them we mention \cite{PERNICI2020102983, Pernici_2018_CVPR, pernici2017unsupervised} for cumulative learning of face appearance models from video stream, \cite{Beery_2020_CVPR, DBLP:conf/iccv/DengHSZXMRG19, DBLP:conf/iccv/ShvetsLB19, DBLP:conf/cvpr/WuF0HKG19, DBLP:conf/iccv/WuCWZ19} for object detection, \cite{DBLP:conf/iccv/OhLXK19} for video object segmentation and \cite{DBLP:conf/eccv/YangC18} for visual object tracking. The works \cite{Beery_2020_CVPR, DBLP:conf/iccv/DengHSZXMRG19, DBLP:conf/iccv/ShvetsLB19, DBLP:conf/cvpr/WuF0HKG19, DBLP:conf/iccv/WuCWZ19} accumulate context from pre-computed feature banks (with fixed pre-trained feature extractors i.e. not undergoing learning). The feature banks extend the time horizon of their network up to 60 second in \cite{DBLP:conf/cvpr/WuF0HKG19} or to one month in \cite{Beery_2020_CVPR} and achieve strong results on spatiotemporal localization. The works \cite{PERNICI2020102983, Pernici_2018_CVPR, pernici2017unsupervised} accumulate extracted face features in a memory bank to preserve all the past knowledge without forgetting and at the same time handle the non-stationarity of the data stream.
At a high level, all these approaches can be framed as a non-parametric estimation method (like nearest neighbors) sitting on top of a high-powered parametric function (Faster R-CNN in the case of object detection \cite{Beery_2020_CVPR}, a face feature extractor in \cite{PERNICI2020102983} and \cite{Pernici_2018_CVPR}, a SiamFC feature extractor \cite{DBLP:conf/eccv/BertinettoVHVT16} for object tracking in \cite{DBLP:conf/eccv/YangC18}).
These methods use a fixed representation that is not incrementally learned as it would require re-encoding all the images in the memory bank.
Avoiding re-encoding images can be advantageous in applications where images cannot be stored for privacy reasons (i.e. face recognition, applications in medical imaging, etc.).
Clearly, also Multi-Object Tracking \cite{CIAPARRONE202061,salvagnini2014information} can benefit from memory based learning.
\section{Conclusion}
We have shown that a special set of fixed classifiers based on regular polytopes generates stationary features by maximally exploiting the available representation space. The proposed method is simple to implement and theoretically correct.
Experimental results confirm both the theoretical analysis and the generalization capability of the approach across a range of datasets, baselines and architectures. Our RePoNet solution improves and generalizes the concept of a fixed classifier, recently proposed in \cite{hoffer2018fix}, to a larger class of fixed classifier models exploiting the inherent symmetry of regular polytopes in the feature space.
Our findings may have implications in all of those Deep Neural Network learning contexts in which a classifier must be robust against changes of the feature representation while learning
as in incremental and continual learning settings.
\section*{Acknowledgment}
\input{ZZ_Acknowledgment}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
Distributed systems consist of several individual components.
Each component has incomplete information about the other components.
Asynchronous distributed systems have no fixed rate at which components progress but rather each component progresses at its individual rate between synchronizations with other components.
Implementing correct algorithms for asynchronous distributed systems is difficult because they have to both work with the incomplete information of the components and for every possible scheduling between the components.
\emph{Petri nets}~\cite{DBLP:books/sp/Reisig85a,DBLP:journals/tcs/NielsenPW81} are a natural model for asynchronous distributed systems.
Tokens represent components and transitions with more than one token correspond to synchronizations between the components.
\emph{Petri nets with transits}~\cite{DBLP:conf/atva/FinkbeinerGHO19} extend Petri nets with a transit relation to model the data flow in asynchronous distributed systems.
\emph{Flow-LTL}~\cite{DBLP:conf/atva/FinkbeinerGHO19} is a specification language for Petri nets with transits and allows us to specify linear properties on both the global and the local view of the system.
In particular, it is possible to globally select desired runs of the system with LTL (e.g., only fair and maximal runs) and check the local data flow of only those runs again with LTL.
A model checker for Petri nets with transits against Flow-LTL is implemented in the tool \textsc{AdamMC}~\cite{DBLP:conf/cav/FinkbeinerGHO20}.
\emph{Petri games}~\cite{DBLP:journals/iandc/FinkbeinerO17} define the synthesis of asynchronous distributed systems based on Petri nets and causal memory.
With causal memory, players exchange their entire causal past only upon synchronization.
Without synchronization, players have no information of each other.
For safety winning conditions, the synthesis algorithm for Petri games with a bounded number of controllable components and one uncontrollable component is implemented in \textsc{AdamSYNT}~\cite{DBLP:conf/cav/FinkbeinerGO15}\footnote{\textsc{AdamSYNT}{} was previously called \textsc{Adam}. From now on, \textsc{AdamMC}{} and \textsc{AdamSYNT}{} are combined in the tool \textsc{Adam}{} (\href{https://github.com/adamtool/adam\#readme}{https://github.com/adamtool/adam}).}.
Both tools are command-line tools lacking visual support to model Petri nets with transits or Petri games and the possibility to simulate or interactively explore implementations, counterexamples, and parts of the created state space.
In this paper, we present a web interface\footnote{The web interface is open source (\href{https://github.com/adamtool/webinterface\#readme}{https://github.com/adamtool/webinterface})
and a corresponding artifact to set it all up locally in a virtual machine is available~\cite{GHY20}.} for model checking asynchronous distributed systems with data flows and for the synthesis of asynchronous distributed systems with causal memory from safety specification.
The web interface offers an input for Petri nets with transits and Petri games where the user interactively creates places, transitions, and their connections with a few inputs.
As a back-end, the algorithms of \textsc{AdamMC}{} are used to model check Petri nets with transits against a given Flow-LTL formula as specification.
Internally, the problem is reduced to the model checking problem of Petri nets against LTL.
Both, the input Petri net with transits and the constructed Petri net can be visualized and simulated in the web interface.
For a positive result, the web interface lets the user follow the control flow of the combined system and the data flow of the components.
For a negative result, the web interface simulates the counterexample with a visual separation of the global and each local behavior.
The algorithms of \textsc{AdamSYNT}{} solve the given Petri game with safety specification.
Internally, the problem is reduced to solving a finite two-player game with complete information.
For a positive result, a winning strategy for the Petri game and the two-player game can be visualized and
the former can be simulated.
For a negative result, the web interface lets the user interactively construct strategies of the two-player game and highlights why they violate the specification.
These new intuitive construction methods, interactive features, and visualizations are of great impact when developing asynchronous distributed systems.
\section{Web Interface for Petri Nets with Transits}
\label{sec:PNwT}
The web interface can model check Petri nets with transits against Flow-LTL.
We use an example from software-defined networks to showcase the workflow.
\subsubsection{Workflow for Petri Nets With Transits}
\label{sec:PNwTworkflow}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{screenshotPNWT150.png}
\caption{Screenshot from the web interface for the model checking workflow.}
\label{fig:pnwt}
\end{figure}
One application domain for Petri nets with transits are \emph{software-defined networks (SDNs)}~\cite{DBLP:journals/ccr/McKeownABPPRST08,DBLP:journals/cacm/CasadoFG14}.
The nodes of the network are \emph{switches} which forward \emph{packets} along the edges of the network according to the \emph{routing configuration}.
Packets enter the network at \emph{ingress switches} and leave it at \emph{egress switches}.
SDNs separate the packet forwarding process, called the \emph{data plane}, from the routing process, called the \emph{control plane}.
\emph{Concurrent updates} to the routing configuration are difficult to get right~\cite{DBLP:conf/networking/ForsterMW16}.
The separation of data and control plane and updates to the routing configuration can be encoded into Petri nets with transits~\cite{DBLP:conf/atva/FinkbeinerGHO19}.
Using this encoding, we demonstrate the workflow of the web interface for model checking an asynchronous distributed system with data flows.
The packets of the SDN are modeled by the data flow in the Petri net with transits.
The data flow relation as an extension from Petri nets to Petri nets with transits is depicted as colored and labeled arcs.
In \refFig{pnwt}, the web interface presents the resulting Petri net with transits \(\ensuremath\mathcal{N}\).
First, we use the tools on the left to create for each switch a place \(si\) with \(i\in\{0,\ldots,5\}\) and add a token (cf.\ outer parts of \(\ensuremath\mathcal{N}\)).
Then, we create transitions for the connections between the switches and for the origin of packets in the SDN (cf.\ transition \(\mathit{ingress}\) in the top-left corner) and link them with flows in both directions.
Additionally, we create local transits between the switches corresponding to the forwarding of packets.
They are displayed in light blue and red and are identified by the letters.
This constitutes the \emph{data plane}.
Next, we define the \emph{control plane}, i.e., which forwarding is activated.
Each transition to forward packets is connected to a place \(ai\) with \(i\in\{0,\ldots,5\}\)
which has a token when the forwarding is configured initially (cf.\ places \(a3\), \(a4\), and \(a5\)) and no token otherwise (cf.\ places \(a0\), \(a1\), and \(a2\)).
For the concurrent update, we create places \(ui\) with \(i\in\{0,\ldots,7\}\) and transitions \(ti\) with \(i\in\{6,\ldots,11\}\) with corresponding flows (cf.\ inner parts of \(\ensuremath\mathcal{N}\)).
Transitions for the forwarding are set as weak fair, i.e., whenever a transition is infinitely long enabled in a run, it also has to fire infinitely often, indicated by the purple color of the outer transitions.
Transitions for the update do not require fairness assumptions.
A satisfied Flow-LTL formula is $A\,F\,s5 $ specifying that all packets eventually reach switch $s5$.
An unsatisfied formula is \((G\,u0\Rightarrow A\,F\,s2)\) requiring for runs, where the update is never executed, that all packets are taking the lower-left route.
The fairness assumptions and a maximality assumption, i.e., whenever some transition can fire in a run some transition fires, are automatically added to the formula.
In the screenshot, a counterexample for the unsatisfied formula is displayed on the right.
The first packet takes the upper-right route via transitions $t3$, $t4$, and $t5$ and the update never starts.
\subsubsection{Features for Petri Nets with Transits.}
\label{sec:PNwTgeneral}
\textsc{AdamMC}~\cite{DBLP:conf/cav/FinkbeinerGHO20} is a command-line model checking tool for Petri nets with transits and Flow-LTL~\cite{DBLP:conf/atva/FinkbeinerGHO19}.
The model checking problem of Petri nets with transits against Flow-LTL is solved by a reduction to Petri nets and LTL.
The web interface allows displaying and arranging the nodes of the Petri net from the reduction and the input Petri net with transits.
Automatic layout techniques
are applied to avoid the overlapping of nodes.
A physics control, which modifies the repulsion, link, and gravity strength of nodes, can be used to minimize the overlapping of edges.
Heuristics generate coordinates for the constructed Petri net by using the coordinates of the input Petri net with transits to obtain a similar layout of corresponding parts.
For a positive result, the web interface allows visualizing the data flow trees for given firing sequences of the nets.
For a negative result, the counterexample can be simulated both in the Petri net with transits and in the Petri net from the reduction.
The witness of the counterexample for each flow subformula
and the run violating the global behavior
can be displayed by the web interface.
This functionality is helpful when developing an encoding of a problem into Petri net with transits
to ensure that a counterexample is not an error in the encoding.
The constructed Petri net can be exported into a standard format for Petri net model checking (PNML)
and the constructed LTL formula can be displayed.
\section{Web Interface for Petri Games}
The web interface can synthesize local controllers from safety specifications.
The workflow is showcased for a distributed alarm system given as a Petri game.
\subsubsection{Workflow for Petri Games}
\label{sec:PGworkflow}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{screenshotPG150.png}
\caption{Screenshot from the web interface for the synthesis workflow.}
\label{fig:pg}
\end{figure}
We demonstrate the workflow of the web interface for the synthesis of asynchronous distributed systems with causal memory from safety specifications.
Petri games separate the places of an underlying Petri net into \emph{system places} and \emph{environment places}.
Tokens on system places are \emph{system players} and tokens on environment places are \emph{environment players}.
Each player has \emph{causal memory}: only upon synchronization with other players, they exchange their entire causal past.
For safety specifications, the system players have to avoid that a bad place is reached for all behaviors of the environment players.
We want to obtain two local controllers of a distributed alarm system that should indicate the location of a burglary at both controllers.
In \refFig{pg}, the web interface presents the resulting Petri game on the left and the winning strategy for the alarm system on the right.
The burglar is modeled by an environment player and each component of the distributed alarm system by a system player.
Environment players are on white places and system players on gray ones.
We create five environment places $e0$, $e1$, $e2$, $\mathit{eL}$, and $\mathit{eR}$.
The place $e0$ has a token, $e1$ and $e2$ serve for the decision to burgle a location, and $\mathit{eL}$ and $\mathit{eR}$ for actually burgling the location.
Each component \(x\in\{p,q\}\) of the alarm system has one system place $x0$ with a token, two system places $x1$ and $x2$ to detect a burglary and inform the other component, and two system places $\mathit{xL}$ and $\mathit{xR}$ to sound an alarm with the position of a burglary.
We create rows of transitions for the environment player deciding where to burgle (first row), for the components detecting a burglary (second row), for the communication between the components (third row), and for sounding the alarm at each location (fourth row).
At last, we use transitions $\mathit{fa}i$ with $i\in\{0,\ldots,3\}$ and $\mathit{fr}j$ with $j\in\{0,\ldots,7\}$ connected to the bad place $\mathit{bad}$ to define that the implementation of the distributed alarm system should avoid false alarms and false reports.
A \emph{false alarm} occurs if the burglar did not burgle any location but an alarm occurred, i.e., in every pair of places $\{e0\}\times\{\mathit{pL}, \mathit{pR}, \mathit{qL}, \mathit{qR}\}$.
A \emph{false report} occurs if a burglary happened at a location but a component of the alarm system indicates a burglary at the other location, i.e., in every pair of places $\{e1, \mathit{eL}\} \times \{\mathit{pR}, \mathit{qR}\}$ and $\{e2, \mathit{eR}\} \times \{\mathit{pL}, \mathit{qL}\}$.
We add transitions and flows to $\mathit{bad}$ for these cases.
The web interface finds a winning strategy (depicted on the right in \refFig{pg}) for the Petri game described above.
Each component locally monitors its location ($t2$, $t3$) and simultaneously waits for information about a burglary at the other location ($t4$, $t5$).
When a burglary is detected at the location of the component then it first informs the other component ($t4$, $t5$) and then outputs an alarm for the current location ($t7$, $t8$).
When a component is informed about a burglary at the other location, it outputs an alarm for the other location ($t6$, $t9$).
\subsubsection{Features for Petri Games}
\label{sec:PGgeneral}
\textsc{AdamSYNT}~\cite{DBLP:conf/cav/FinkbeinerGO15} is a command-line tool for Petri games~\cite{DBLP:journals/iandc/FinkbeinerO17}.
The synthesis problem for Petri games with a bounded number of system players, one environment player, and a safety objective is reduced to the synthesis problem for two-player games.
A winning strategy in the two-player game is translated into a winning strategy for the Petri game.
Both can be visualized in the web interface.
Here, the web interface provides the same features for visualizing, manipulating, and automatically
laying out the elements as for model checking.
It uses the order of nodes of the Petri game to heuristically provide a positioning of the strategy
and allows simulating runs of the strategy.
The winning strategy of the two-player game provides
an additional view on the implementation to check if it is not bogus
due to a forgotten case in the Petri game or specification.
For an unrealizable synthesis problem, the web interface allows analyzing
the underlying two-player game via a stepwise creation of strategies.
This guides the user towards changes to make the problem realizable.
\section{Implementation Details}
\label{sec:impldetails}
The server is implemented using the Sparkjava micro-framework~\cite{sparkjava} for incoming HTTP and WebSocket connections.
The client is a single-page application written in Javascript using Vue.js~\cite{vue}, D3~\cite{d3}, and the Vuetify component library~\cite{vuetify}.
We constructed libraries out of the tools \textsc{AdamMC}{} and \textsc{AdamSYNT}{} and implemented one interface handling both libraries.
Common features like the physics control of nodes share the same implementation.
All components of the libraries and the web interface~\cite{webinterface} are open source and available on GitHub~\cite{adamtool}.
\section{Conclusion}
\label{sec:conclusion}
We presented a web interface for two tools:
\textsc{AdamMC}{}, a model checker for data flows in asynchronous distributed systems represented by Petri nets with transits, and
\textsc{AdamSYNT}{}, a synthesis tool for local controllers from safety specifications in asynchronous distributed systems with causal memory represented by Petri games.
The web interface makes the modeling and debugging of Petri nets with transits and Petri games user-friendly as it presents visual representations of the input, all intermediate steps, and the output of the tools.
The interactive features are a great assistance for correctly modeling distributed systems.
We plan to extend the web interface and tool support to model checking Petri nets with transits against Flow-CTL$^*$~\cite{DBLP:conf/atva/FinkbeinerGHO20}, to other classes of Petri games with a decidable synthesis problem~\cite{DBLP:conf/fsttcs/FinkbeinerG17,DBLP:conf/concur/BeutnerFH19}, to the bounded synthesis approach for Petri games~\cite{DBLP:conf/birthday/Finkbeiner15,DBLP:journals/corr/abs-1711-10637,DBLP:journals/corr/Tentrup16,DBLP:conf/atva/Hecking-Harbusch19}, and to high-level Petri games~\cite{DBLP:journals/acta/GiesekingOW20}.
As our web interface is open source and easy to extend, we also plan to connect it to other tools for Petri nets like APT~\cite{apt}, LoLA~\cite{DBLP:conf/apn/Wolf18a}, or TAPAAL~\cite{DBLP:conf/tacas/DavidJJJMS12}.
\bibliographystyle{splncs04}
|
\section{Introduction}
The safety of autonomous driving functions requires robust detection of surrounding objects such as pedestrians, cars, and bicycles. As implemented in the autonomous driving pipeline, the monitor function needs to report abnormal situations on the resulting detection. The filtering is challenging, as no labeled ground-truth in operation is offered to be compared. One common way of checking abnormality is to use other modalities to compare the result. For example, a monitor cross-checks the 3D bounding boxes generated from multiple cameras and 3D bounding boxes produced by other detection pipelines using LiDARs or radars.
The focus of this paper is to develop another type of monitor that does not exploit diversities in sensor modality. The consideration starts with a practical motivation where one also wants to achieve diversity in monitor design. Meanwhile, utilizing other sensor modalities can be further complicated by faults and component malfunctioning as introduced in the ISO~26262 context. Our key approach is to utilize the concept of \textit{abstraction}, where abstraction soundly aggregates all valuations that are considered legal. In contrast to non-abstraction methods (cf. Section~\ref{se:relatedwork} for related work) the benefit lies in the soundness inherently in the monitor. Moving beyond existing results in abstraction of neuron activation patterns~\cite{Cheng19,HenzingerL020}, we propose two new types of abstraction that are not based on extracting valuations of neurons for each image.
\begin{itemize}
\item \textbf{(Data-Label Abstraction)}
Data-label abstraction builds a summary over labels used in training, where abnormalities refer to the encountered input not contained in the abstraction. We present a general framework that builds the abstraction from $\lambda$-consecutive frames; for each frame, generate information based on the divided physical regions.
\item \textbf{(Post-Algorithm Abstraction)}
Post-algorithm abstraction refers to methods that relax the immediate algorithm in the post-processing pipeline to filter infeasible values. We apply post-algorithm abstraction on the tracking algorithm that takes the input as bounding boxes from the object detection module. Recall that standard tracking algorithms perform association based on objects with the same class due to the assumption of fault-absence in object detection; we demonstrate a simple post-algorithm abstraction technique that blurs the object class in the association step, while the association function is changed from 1-1 (in the standard algorithm) to 1-to-many (in the abstract transformer). Elements being associated together by the abstract transformer are then checked against consistency over their concrete, unabstracted values.
\end{itemize}
We have implemented the concept as a research prototype and evaluated the result using public autonomous driving datasets.
The implemented monitors find undesired cases such as objects in unexpected locations as well as objects with wrong classes. Applying a fine-tuned
Faster R-CNN~\cite{ren2016faster} detector over
the KITTI dataset~\cite{geiger2012we}, both types of monitors serve their purpose in filtering the problematic prediction without the need of human-in-the-loop with a success rate of~$84\%$. Simultaneously, the speed of abnormality filtering can reach the speed around~$20$ FPS, matching our minimum real-time requirement for deployment.
The rest of the paper is structured as follows. After Section~\ref{se:relatedwork}
comparing related works with ours, Section~\ref{se:example} provides a motivating example to hint the underlying principle.
Section~\ref{se:concepts} presents required mathematical formulations such as regions. Sections~\ref{se:construction} and~\ref{se:postalg} provide details on data-label abstraction and post-algorithm abstraction. Finally, Section~\ref{se:experimentation} demonstrates experimental results, and Section~\ref{se:conclusion} concludes the paper with future directions.
\section{Related work}\label{se:relatedwork}
In this work, we focus on monitoring methods without exploiting sensor modalities and diversities, where there exist also monitors equipped on the system level for fault checking and recovery~\cite{NFM20,xiang2021runtime}. Our attention is further restricted to monitoring modules implemented using learning-based approaches. An intuitive method to check abnormalities is to set a threshold on the output of the softmax function~\cite{hendrycks2017a}.
Other techniques such as the temperature scaling technique can be applied to calibrate the confidence on final outputs~\cite{guo2017calibration}. The Monte-Carlo dropout technique~\cite{gal2016dropout,MC19} is an online method that utilizes dropout to create ensembles and to compute the Bayesian measure. The above methods are essentially confidence-based and require proper calibration in order to be used in safety-critical systems\footnote{The need for calibration is mentioned in clause 8.5.2. of the autonomous driving safety standard UL~4600~\cite{UL4600}.}. Currently, proper calibration with guaranteed error bounds is known to be hard for autonomous driving, as the data distribution for real-world driving is unknown and can be highly individualized. Even under the binary classification setup, the sharpness of calibrated confidence is hard to be guaranteed without prior knowledge of the distribution~\cite{gupta2020distribution}. Abstraction-based methods build a data manifold that encloses all values from the training dataset. Answering the problem of whether a newly encountered data point falls inside the manifold serves as a proxy of abnormality detection. Existing results in abstraction-based monitoring~\cite{Cheng19,HenzingerL020} build the manifold by extracting the feature vectors inside the neuron computation and are thus dependent on specific network configurations. Our work focus on two types of abstraction (data-label and post-algorithm) that are not directly related to the DNN being analyzed. For our post-algorithm abstraction techniques for monitoring, it is based on utilizing existing tracking algorithms~\cite{wojke2017simple,chen2018real}. However, tracking algorithms as specified in~\cite{wojke2017simple,chen2018real} assume that results from object detection are perfect in the predicted class, while our purpose is to filter imperfections in object detection. This brings a natural relaxation in our implementation, where our abstraction blurs the class association such that class flips can be efficiently filtered. However, the concept of post-algorithm abstraction is generic, and we expect it to be also applied to filter other types of errors for other modules in autonomous driving.
\section{A motivating example}\label{se:example}
We present two consecutive images at timestamp~$t-1$ and~$t$ in Fig.~\ref{fig:example} to assist in understanding the concept of our monitors. The detected objects are marked with bounding boxes in dashed lines and the associated labels over the boxes.
The monitor framework detailed in this paper will check two types of abnormalities at time~$t$:
\begin{itemize}
\item The monitor based on \textit{data-label abstraction} finds that the identified pedestrian can be problematic, because in the training data, region $r_1$ never has a pedestrian.
\item The monitor based on \textit{post-algorithm abstraction} finds that the motorcycle class can be problematic; by abstracting the information of object class, performing tracking, and then matching the tracked result, there seems to be a class flip over the detected object.
\end{itemize}
\begin{figure}[t]
\centering
\subfigure[Time $t-1$]{
\begin{minipage}[t]{0.43\linewidth}
\includegraphics[width=0.95\linewidth]{fig/motivating1.pdf}\label{fig:exmp1}
\end{minipage}
}%
\subfigure[Time $t$]{
\begin{minipage}[t]{0.43\linewidth}
\includegraphics[width=0.95\linewidth]{fig/motivating2.pdf}\label{fig:exmp2}
\end{minipage}
}
\caption{A motivating example for abnormal checking}
\label{fig:example}\vspace*{-\baselineskip}
\end{figure}
\section{Preliminaries}\label{se:concepts}
We construct necessary notations to define the monitor mathematically. Here for simplicity, the formulation uses 2D object detection over the image plane, but the concept can be easily generalized to a 3D setup.
\begin{defn}[Region]
A region is an area denoted by $r=(pos_r, w,h)$, where $pos_r=(x,y)$ is the center point of the region, $w$ is the width of the region, and $h$ is the height of the region.
\end{defn}
An image is divided into a finite set of non-overlapping regions $R=\{r_i\}_{1\leq i \leq m}$. That is, given any $r_i$ and $r_j$ where $i\neq j$, their intersected area should be zero.
We assume that the dataset contains images. For every image, its ground-truth label of an object $\textbf{x}=(class, pos,size)$ has three attributes: the label of the object, the localization $pos=(x,y)$ being the center point of the bounding box for the object in the image, and the size of the bounding box. Denote~$\mathbb{K}$ as the set of all possible classes.
Given two consecutive images, the physical locations of an object may shift.
Then the movement of an object depicted from a sequence of images can be encoded as a finite trace.
\begin{defn}[Trace of an object]
For an object labeled in a sequence of $n$ consecutive images, let $(x_i,y_i)$ be the center point of the object at $i$-th image, and $size_i$ be its size. The trace of the object is denoted by $\tau=((x_1,y_1),size_1)\rightarrow \ldots\rightarrow((x_i,y_i),size_i)\ldots\rightarrow ((x_n,y_n),size_n)$, and we call $\tau^i=((x_i,y_i),size_i)$ the $i$-th \textit{concrete state}.
\end{defn}
Once an image is divided into regions, the location of an object can be mapped to a region based on the location of its center point. That is, an object $\textbf{x}$ is located in region~$r$, denoted by $\textbf{x}\odot r$, if the center point of the bounding box for object $\textbf{x}$ is in the area covered by region~$r$.
Therefore, a finite trace of an object can be mapped to a region-based trace.
\begin{defn}[Region-based trace]
Given a trace of an object $\tau$ with $\tau^j=((x_j,y_j),size_j)$ and $1\leq j\leq n$, the \textit{region-based trace} is $\hat{\tau}=(r_{\langle 1 \rangle},size_1)\rightarrow \ldots \rightarrow (r_{\langle j \rangle},size_j) \rightarrow \ldots \rightarrow (r_{\langle n \rangle},size_n)$ such that $(x_j,y_j)$ is in the area covered by $r_{\langle j \rangle}$ and
$r_{\langle j\rangle}\in R$. The $j$-th \textit{(abstract) state} of a trace is denoted by $s^j=\hat{\tau}^j$, with $\hat{\tau}^j.region$ being the region of the state, and $\hat{\tau}^j.size$ being the size of the state.
\end{defn}
To simplify the notation, the region-based trace $\hat{\tau}$ of an object is written as a vector $((r_{\langle 1\rangle},size_1),\ldots$,$(r_{\langle n\rangle},size_n))$.
\begin{figure}[tp]
\subfigure[Time $t-1$]{
\begin{minipage}[t]{0.325\linewidth}
\includegraphics[width=0.99\linewidth]{fig/image1.pdf}\label{fig:image1}
\end{minipage}\hspace{-5pt}
}%
\subfigure[Time $t$]{
\begin{minipage}[t]{0.325\linewidth}
\includegraphics[width=0.99\linewidth]{fig/image2.pdf}\label{fig:image2}
\end{minipage}\hspace{-5pt}
}%
\subfigure[Time $t+1$]{
\begin{minipage}[t]{0.325\linewidth}
\includegraphics[width=0.99\linewidth]{fig/image3.pdf}\label{fig:image3}
\end{minipage}
}%
\caption{\vspace{-4pt}The ground-truth on a sequence of images}
\label{fig:groundtruth}\vspace*{-\baselineskip
\end{figure}
Consider the ground-truth on a sequence of consecutive images shown in Fig.~\ref{fig:groundtruth}. Every image is divided into six regions. The set of classes on the objects involved in the sequence is $\{\sig{car}, \sig{truck}\}$. For the hatchback categorized as the
\sig{car} object located in region~$r_3$ in Fig.~\ref{fig:image1}, it is also labeled in Figs.~\ref{fig:image2} and~\ref{fig:image3}. Therefore, its region-based trace is $((r_3,2.7),(r_6,3),(r_6,2.8))$ with the size of the vehicle changed from $2.7$ to $3$ to $2.8$ respectively.
For the \sig{truck} located at $r_2$ in Fig.~\ref{fig:image1}, as it is not shown in Fig.~\ref{fig:image3}, its region-based trace is $((r_2,3),(r_5,3.5), (\bot, -1))$. We use $(\bot, -1)$ as a special region for \textit{undefined}, acting as a proxy of object disappearance in the image frame.
\begin{defn}[$\lambda$-trace]
A $\lambda$-trace is a region-based trace, where the length of the trace is $\lambda$. Let $\mathcal{T}_{\lambda,k}$ be the set of $\lambda$-traces for objects in class $k$.
\end{defn}
Reconsider the ground-truth shown in Fig.~\ref{fig:groundtruth}. The set of $2$-traces for the \sig{truck} class is $\mathcal{T}_{2, \sig{truck}}=\{((r_2,3),(r_5,3.5)), ((r_5,3.5),(\bot, -1))\}$. The set of $2$-traces for the \sig{car} class is
\[\begin{array}{lll}
\{((\bot,-1),(r_3,2.8)),& ((r_3,2.8),(r_6,2.9)), & //\textrm{big car}\\
((r_3,2.7),(r_6,3)), & ((r_6,3),(r_6,2.8)),&//\textrm{hatchback} \\
((\bot,-1),(r_2,2)), &((r_2,2),(r_5,3))\}. &//\textrm{small car}
\end{array}
\]
\section{Monitor construction on data-label abstraction}\label{se:construction}
The monitor on data-label abstraction can be regarded as a dictionary mapping from the set of classes $\mathbb{K}$ to an abstraction over the $\lambda$-traces accumulated from the available datasets.
\subsection{Monitor construction}
Let $D_k=\{((r_{\langle 1 \rangle}, [l_1,u_1]), \ldots, (r_{\langle \lambda \rangle}, [l_{\lambda},u_{\lambda}]))\}$ be the dictionary recording the information related to $\lambda$-traces for class~$k$. The dictionary adopts interval $[l,u]$ to replace the attribute of \textit{size} in every state of a region-based trace, to record the size variation of the object in the region.
A $\lambda$-trace $\hat{\tau}\in \mathcal{T}_{\lambda,k}$ is \textit{strongly included} in $D_k$ if $\exists \hat{\tau}_{\alpha}\in D_k$ such that $\forall i\in [1, \lambda]:$ $(\hat{\tau}^i.region=\hat{\tau}^i_{\alpha}.region) \wedge (\hat{\tau}^i.size \in \hat{\tau}^i_{\alpha}.[l,u])$, denoted by $\sig{strcon}(\hat{\tau},D_k)$=\sig{true}.
Additional to strongly inclusion, a $\lambda$-trace $\hat{\tau}\in \mathcal{T}_{\lambda,k}$ can be \textit{weakly included} in $D_k$, denoted by $\sig{wkcon}(\hat{\tau},D_k)$=\sig{true}, if $\sig{strcon}(\hat{\tau},D_k)$=\sig{false} and $\exists \hat{\tau}_{\alpha}\in D_k$ such that $\forall i\in [1, \lambda]: (\hat{\tau}^i.region=\hat{\tau}^i_{\alpha}.region)$, and $\exists j\in [1,\lambda]: \hat{\tau}^j.size \not\in \hat{\tau}^j_{\alpha}.[l,u])$.
The construction of dictionary $D_k$ is depicted in Alg.~\ref{alg:labelbased}.
The dictionary is updated w.r.t. the newly encountered or weakly included $\lambda$-traces.
Given a newly encountered trace, before adding it to the dictionary, the size of every state in the trace is translated to an interval (e.g., from value~$5$ to interval $[5,5]$) by function \sig{size2inter()} (Lines~4,~5).
Facing a weakly included trace (Line 11), the lower and upper bounds for the size intervals of the trace in the dictionary are expanded accordingly (Lines~14,~15).
\begin{algorithm}[t]
\small
\caption{Construction on data-label abstraction}\label{alg:labelbased}
\begin{algorithmic}[1]
\STATE{Input: a set of $\lambda$-traces $\mathcal{T}_{k,\lambda}$ for objects of class $k$, $D_k$}
\STATE{Output: $D_k$}
\FOR{($\forall \hat{\tau}\in \mathcal{T}_{k,\lambda}$)}
\IF{ $D_k=\emptyset\vee \neg \sig{wkcon}(\hat{\tau}, D_k)$}
\STATE{$D_k=D_k\cup \{\sig{size2inter}(\hat{\tau})\}$}
\STATE{\textbf{continue}}
\ENDIF
\IF{($\sig{strcon}(\hat{\tau},D_k)$)}
\STATE{\textbf{continue}}
\ENDIF
\IF{($\sig{wkcon}(\hat{\tau},D_k))$}
\STATE{Let $ \hat{\tau}_{\alpha}\in D_k$ where $\forall i \in [1, \lambda]: \hat{\tau}^i.region=\hat{\tau}^i_{\alpha}.region$}
\FOR{($j=1;j\leq \lambda; j++$)}
\STATE{$\hat{\tau}^j_{\alpha}.l=\min(\hat{\tau}^j.size,\hat{\tau}^j_{\alpha}.l)$}
\STATE{$\hat{\tau}^j_{\alpha}.u=\max(\hat{\tau}^j.size,\hat{\tau}^j_{\alpha}.u)$}
\ENDFOR
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
For the ground-truth presented in Fig.~\ref{fig:groundtruth}, the dictionary for the 2-traces of class \sig{car} is constructed as follows:
\begin{enumerate
\item Initially, the dictionary is empty. The sizes of states $(\bot,-1)$ and $(r_3,2.8)$ in the first trace are replaced by the intervals $[-1,-1]$ and $[2.8,2.8]$, respectively. Then the trace is added. We have $D_{\sig{car}}=\{((\bot,[-1,-1]),(r_3,[2.8,2.8]))\}$.
\item In the second step, we cannot find the inclusion relation between the trace in the dictionary and the one to be added. Therefore, after adding the second trace, the dictionary is
$\{((\bot,[-1$, $-1])$,$(r_3,[2.8,2.8])),((r_3,[2.8, 2.8]), (r_6,[2.9,2.9]))\}$.
\item In the third step, the trace $((r_3,2.7),(r_6,3))$ is weakly included by the dictionary. Consequently, the intervals in the trace $((r_3,[2.8,2.8]),(r_6,[2.9,2.9]))$ are updated, and the dictionary is $\{((\bot,[-1,-1])$, $(r_3,[2.8,2.8])), ((r_3,[2.7,2.8]),(r_6,[2.9,3]))\}$.
\item When the process terminates, the dictionary is \vspace{-6pt}
\[\begin{array}{l}
\resizebox{.9\hsize}{!}{$\{((\bot,[-1,-1]),(r_3,[2.8,2.8])),((r_3,[2.7,2.8]), (r_6,[2.9,3]))$},\\
\resizebox{.86\hsize}{!}{$((r_6,[3,3]),(r_6,[2.8,2.8])),((\bot,[-1, -1]),(r_2,[2,2]))$},\\
\resizebox{.37\hsize}{!}{$((r_2,[2,2]), (r_5,[3,3]))\}$}.
\end{array}
\]
\end{enumerate}
\subsection{Abnormality checking}
To check abnormalities with the monitor on data-label abstraction, we need to compare the traces of the detected objects with the constructed dictionaries.
Given a $\lambda$-trace $\hat{\tau}$ of a detected object in class $k$, we check its inclusion relation with the dictionary $D_k$.
To simplify the discussion, we assume that the first $\lambda-1$~states in the trace are normal.
If the $\lambda$-trace $\hat{\tau}$ is not strongly included in $D_k$, i.e., $\sig{strcon}(\hat{\tau},D_k)=\sig{false}$, the detected result on the object is abnormal. Precisely, provided that $\sig{strcon}(\hat{\tau},D_k)= \sig{false}$, we may further refine the type of abnormality as follows:
\begin{itemize}
\item (Abnormal size) The abnormality occurs, if $\sig{wkcon}(\hat{\tau},D_k)=\sig{true}$.
\item (Abnormal location) The abnormality occurs, if $\sig{wkcon}(\hat{\tau},D_k)=\sig{false}$.
\item (Abnormal lost object) The abnormality occurs, if $\hat{\tau}^{\lambda}.region=\bot$ and
the condition of ``abnormal location" holds. In other words, if the detected $\lambda$-trace shows that the object is lost, and we cannot find such a trace in the dictionary, the loss on the object is abnormal.
\end{itemize}
Consider the established dictionaries from the example in Fig.~\ref{fig:groundtruth}. Given a region-based trace $((r_3,2.8),(r_6,3.2))$ for the \sig{car} class, it is weakly contained by dictionary $D_{\sig{car}}$, and the monitor reports \textit{abnormal size}.
Given a region-based trace $((r_2,3),(r_4,2.8))$ for the \sig{truck} class, region $r_4$ is not contained in the dictionary, and the location is abnormal.
\subsection{Robustness Considerations}\label{subsec.robustness}
The method proposed in this section is based on partitioning on the region of interest. In implementation, one commonly seen issue is related to the \textit{robustness} of the monitor caused by the region partitioning. An example can be seen in Fig.~\ref{fig:robustness}, where the center of the \sig{car} object at time $t-1$ is close to the top-right corner of region~$r_3$. Here we omit technical details, but under such cases, one can also consider $\delta$-ball perturbation on the center point of the bounding box. By doing so, additional three $\lambda$-traces starting with region $r_2$, $r_5$ and $r_6$ will also be added in the monitor construction process, thereby providing \textit{robustness} guarantees on the monitor.
\section{Post-algorithm abstraction}\label{se:postalg}
\begin{figure}[tp]
\centering
\includegraphics[ width=0.85\linewidth]{fig/robustness.pdf}
\caption{Robustness considerations in data-label abstraction}
\label{fig:robustness}\vspace*{-\baselineskip}
\end{figure}
\begin{figure}[htp]
\centering
\subfigure[Detected results]{
\begin{minipage}[t]{0.99\linewidth}
\includegraphics[width=0.95\linewidth]{fig/tracking1.pdf}\label{fig:tracking1}
\end{minipage}
}\\
\subfigure[Post-algorithm abstraction - class relaxation and correct association]{
\begin{minipage}[t]{0.99\linewidth}
\includegraphics[width=0.95\linewidth]{fig/tracking2.pdf}\label{fig:tracking2}
\end{minipage}
}\\
\subfigure[Post-algorithm abstraction - class relaxation but incorrect association]{
\begin{minipage}[t]{0.99\linewidth}
\includegraphics[width=0.97\linewidth]{fig/tracking3.pdf}\label{fig:tracking3}
\end{minipage}
}\\
\subfigure[Post-algorithm abstraction - class relaxation and weak association]{
\begin{minipage}[t]{0.99\linewidth}
\includegraphics[width=0.95\linewidth]{fig/tracking4.pdf}\label{fig:tracking4}
\end{minipage}
}%
\caption{Understanding class relaxation and weak tracking}
\label{fig:tracking}\vspace*{-\baselineskip}
\end{figure}
\subsection{Understanding the mechanism}
We start this section by providing an intuitive rationale on the process of post-algorithm abstraction using Fig.~\ref{fig:tracking}. Consider Fig.~\ref{fig:tracking1} being the result of object detection between two consecutive frames, where the vehicle identified as a \sig{car} object at time~$t-1$ is identified as a \sig{van} object at time~$t$. Therefore, we expect to have a monitor to filter the issue. To do so, we utilize the post-processing algorithm by generating an \textit{abstract transformer}. The abstract transformer performs an abstraction on the detected object such that the class information is blurred. Then perform object tracking which includes prediction and association. Provided that the tracking algorithm is correct similar to the case in Fig.~\ref{fig:tracking2}, it shall associate objects correctly. Then one can perform a \textit{concretization} step to restore the class label information, and compare against class labels. In Fig.~\ref{fig:tracking1}, the class flip of the \sig{car} object at time~$t$ will thus be detected. This is in contrast to a standard tracking algorithm that assumes correct class information, where standard tracking correctly relates the \sig{truck}, while reporting the disappearing of the \sig{car} and the appearing of the \sig{van}.
However, for the above-mentioned abstraction-based monitor to reduce false alarms, it is crucial to have correct object association between frames. This may be sometimes unrealistic due to objects being close by. Consider the association subroutine in the tracking algorithm producing incorrect results as shown in Fig.~\ref{fig:tracking3}. This leads to additional false alarms for the \sig{truck} object. Our mediation is to also \textit{relax the association} such that the association of an object at time~$t-1$ may be associated to multiple objects at time~$t$. This is to compensate the loss of precision due to blurring the class information. As demonstrated in Fig.~\ref{fig:tracking4}, the \sig{truck} object at time~$t-1$ will be weakly associated with the \sig{truck} object and the \sig{van} object at time~$t$, thereby no alarm will be reported. However, the \sig{car} object at time~$t-1$ will still be detected with class flip. One can view the weak association in the post-algorithm abstraction monitor being analogous to the robustifying step for the data-label abstraction monitor highlighted in Section~\ref{subsec.robustness}.
\subsection{Monitor construction}
The construction of the monitor involves the following steps: 1) abstract or blur the classes of objects to be tracked; 2) perform the relaxed tracking over the objects without considering the classes of the objects; 3) enlarge the association candidates with a given bound; 4) restore the classes for the association candidate objects.
Let $\sig{hideclass(\textbf{x})}$ be the function to abstract the class of object $\textbf{x}$, and the blurred object is $\hat{\textbf{x}}=\sig{hideclass(\textbf{x})}$.
Given an object $\textbf{x}^{t-1}$ at time~$t-1$, the tracked object with blurred class at time~$t$ is $\hat{\textbf{x}}^t=\sig{track}(\sig{hideclass}(\textbf{x}^{t-1}))$, where function \sig{track()} is the relaxed tracking algorithm without considering the class information of objects.
We propose the concept of \textit{support set} as the set of potential tracked results for a given object.
\begin{defn}[Support set]
Given an object $\textbf{x}^{t-1}$ at time~$t-1$, the set of candidates specified by the post-algorithm abstraction is those being covered by a certain area, i.e.,
\begin{equation*}
\resizebox{.99\hsize}{!}{$\sig{support}(\textbf{x}^{t-1})=\{\hat{\textbf{x}}^{t}\,|\, \hat{\textbf{x}}^{t}\odot \sig{radius}(\sig{track}(\sig{hideclass}(\textbf{x}^{t-1})),\delta)\}$},
\end{equation*}
where $\sig{radius}(\sig{track}(\sig{hideclass}(\textbf{x}^{t-1})), \delta)$ is the round area with the center point of the tracked object from $\sig{track}(\sig{hideclass}(\textbf{x}^{t-1}))$ as the center and $\delta$ as the radius.
\end{defn}
For example, for the image at time~$t$ shown in Fig.~\ref{fig:tracking4}, the support sets for the object in class \sig{car} and the object in class \sig{truck} at time~$t-1$ are the same, which are covered by the two cycles in blue.
\subsection{Abnormality checking}
The abnormality checking with post-algorithm abstraction is to check whether the class of an object at time~$t-1$ and the class of the tracked object at time~$t$ are consistent.
To check the consistency, we need to restore the class information for the objects in the support set. Given an object $\textbf{x}^{t-1}$ at time~$t-1$, let $\sig{restoreclass}(\sig{support}(\textbf{x}^{t-1}))$ be the set of class labels for the potentially tracked objects from $\textbf{x}^{t-1}$ at time~$t-1$. If $\textbf{x}^{t-1}.class\not\in \sig{restoreclass}(\sig{support}(\textbf{x}^{t-1}))$, we say that a \textit{label flip} happens for the object $\textbf{x}$ at time~$t$.
As a final remark, the soundness of the post-algorithm abstraction relies on the tracking algorithm having a bounded error. In other words, if the tracking algorithm ensures that the predicted associated object only deviates from the real associated object with a distance of~$\delta$, the monitor can guarantee to detect class flips.
\section{Evaluation}\label{se:experimentation}
The prototype for the run-time monitor framework is implemented with Python, which currently supports abnormality checking on \textit{abnormal location}, \textit{abnormal size}, \textit{abnormal object loss} and \textit{label flip}.
The experimentation runs on a PC equipped with Intel [email protected]. We adopt eleven scenarios in the KITTI dataset\footnote{The evaluation using KITTI dataset in this paper is for knowledge dissemination and scientific publication and is not for commercial use.} for data-label abstraction. The other scenarios are adopted in validating the effectiveness of our method. For data-label abstraction, we set~$\lambda$ to~$1$ to record state-less behavior; for post-algorithm abstraction, we set~$\delta$ to~$0$ to avoid including other objects that are also nearby. Detecting ``abnormal object loss" equals to the case of setting $\lambda$ being~$2$ while the abstraction intuitively covers ``normal object loss" on two sides of an image (i.e., the object can disappear due to falling outside the image frame). The precision of the data-label abstraction monitor is highly dependent on the associated region size. For every image of size $1242 \times 375$ in the KITTI dataset, we build a total of $9\times 6 = 54$ regions. As explained in later subsections, the configuration offers a nice balance between computation speed and detectability. Fig.~\ref{fig:kitti} shows examples where errors are filtered by our implemented monitors.
\begin{figure}[t]
\centering
\subfigure[The traffic light on the left is mis-classified as \textsf{van}. The data-label abstraction monitor detects that it is impossible to have a \sig{van} of such a small size at that region.]{
\begin{minipage}[t]{0.95\linewidth}
\includegraphics[width=\linewidth]{fig/location.pdf}\label{fig:kitti1}
\end{minipage}
}\\
\vspace{-2mm}
\subfigure[At time $t-1$, the bus object is detected as the \textsf{truck} class; in KITTI, buses and trucks share the same class. ]{
\begin{minipage}[t]{0.95\linewidth}
\includegraphics[width=\linewidth]{fig/flip1.pdf}\label{fig:kitti2}
\end{minipage}
}\\%
\vspace{-2mm}
\subfigure[The post-algorithm abstraction monitor detects that the \textsf{truck} object at time $t-1$ is flipped to class \textsf{tram} at time $t$.]{
\begin{minipage}[t]{0.95\linewidth}
\includegraphics[width=\linewidth]{fig/flip2.pdf}\label{fig:kitti3}
\end{minipage}
}
\caption{Qualitative evaluation on the KITTI dataset}
\label{fig:kitti}\vspace{-15pt}
\end{figure}
\subsection{Effectiveness evaluation with ground-truth}
To evaluate the effectiveness of our run-time monitoring method, we randomly modify the labels and bounding boxes of labeled objects in the ground-truth of the dataset, and check whether the framework can find abnormalities. If an injected error is filtered, it is a \texttt{true positive} (TP). In other words, TP implies that there exists an error and monitor raises the warning. \texttt{false positive} (FP) implies that there exists no error and but monitor raises the warning. We also adopt \texttt{recall} and \texttt{precision} as the criteria for the evaluation. Table~\ref{tab:table1} provides a summary on the number of filtered abnormalities in each category, subject to the number of injected errors. From the table, we observe that the method can find all the injected label flips, with a slight cost of raising two additional false alarms. The false alarms are resulted from the setting of~$\delta$, which is not enough to cover close-by objects. For the other injected location errors, the monitor can find most of them. However, the precision for checking abnormal location is lower than those of the others. The reason is that the diversity of adopted scenarios for data-label abstraction is low, and some filtered abnormalities do not exist in those scenarios. The performance should be improved by applying more scenarios in various datasets in data-label abstraction.
\begin{table}[t]
\centering
\caption{Effectiveness evaluation with ground-truth}\label{tab:table1}
\begin{lrbox}{\tablebox}
\begin{tabular}{|l| c| c|c|c|c|c|}\hline
alarm type & $\#$ injected & TP & FP & precision & recall \\\hline
abnormal location
& 34 & 27 & 5 & 0.844 & 0.794\\
abnormal size & 50 & 47 & 6 & 0.887 & 0.940 \\
object loss& 112 & 109 & 10 & 0.916 & 0.973 \\
label flip & 180 & 180 & 2 & 0.989 & 1.000\\
\hline
\end{tabular}
\end{lrbox}
\scalebox{0.95}{\usebox{\tablebox}}
\end{table}
\begin{table}[t]
\centering
\caption{Performance evaluation on run-time detected results}\label{tab:table2}
\begin{lrbox}{\tablebox}
\begin{tabular}{|c|l|c|c|c|c|c|}\hline
scenario & alarm type & TP & FP & max OH & min OH & average\\\hline
\multirow{4}{*} { 1} & abnormal location & 1 & 0 & & &\\
& abnormal size & 7 & 1 & 0.155s & 0.025s & 0.064s\\
& object loss & 72 & 8 & & &\\
& label flip & 78 & 4 & & & \\
\hline
\multirow{4}{*} { 2} & abnormal location & 1 & 0 & & &\\
& abnormal size & 4 & 0 & 0.148s & 0.008s & 0.072s\\
& object loss& 157 & 29 & & &\\
& label flip & 16 & 2 & & &\\
\hline
\multirow{4}{*} { 3} & abnormal location & 2 & 0 & & &\\
& abnormal size & 4 & 1 & 0.126s & 0.001s & 0.058s\\
& object loss& 92 & 15 & & &\\
& label flip & 15 & 2 & & &\\
\hline
\multirow{4}{*} { 4} & abnormal location & 0 & 0 & & &\\
& abnormal size & 2 & 0 & 0.076s & 0.012s & 0.039s\\
& object loss& 25 & 2 & & &\\
& label flip & 7 & 0 & & &\\
\hline
\multirow{4}{*} { 5} & wrong location & 0 & 0 & & &\\
& unusual size & 25 & 6 & 0.130s & 0.012s & 0.047s\\
& object & 140 & 25 & & &\\
& label flip & 37 & 6 & & &\\
\hline
\end{tabular}
\end{lrbox}
\scalebox{0.9}{\usebox{\tablebox}}\vspace*{-\baselineskip}
\end{table}
\subsection{Evaluation on monitoring real-time detection results}
To evaluate its performance in monitoring real-time detection results, we collect the alarms and their categories and manually check their correctness. Meanwhile, we provide the run-time overhead (OH) to check whether the speed of the monitor is sufficient to meet the real-time requirement for self-driving. Table~\ref{tab:table2} summarizes the performance of our method in monitoring the detected results from
fine-tuned Faster R-CNN~\cite{ren2016faster} detection network with Tensorflow Object Detection API~\cite{huang2017speed}.
The selected scenarios involve various road conditions and traffic environments.
Overall the average execution time is below $0.1$ second while some worst case can reach $0.155$ seconds. By careful examination on individual steps, it turns out that the weak-tracking is the overhead, as we relax a tracking algorithm utilizing histogram of gradients~\cite{dalal2005histograms} and kernelized correlation filters~\cite{henriques2014high}.
Apart from further fine-tuning our implementations for faster monitoring speed, towards practical deployment with stricter real-time constraints, simple-to-achieve solutions may include either (1) adapting a simpler tracking algorithm or (2) simply reporting time-out when reaching the assigned budget and continue the next monitoring round.
\section{Concluding Remarks}\label{se:conclusion}
In this paper, we considered the practical problem that the performance of object detection modules implemented with deep neural networks may not be reliable. We developed abstraction-based monitoring as a logical framework for checking abnormalities over detected results. The data-label abstraction extracts characteristics of objects from existing datasets. The post-algorithm abstraction relaxes class labels in tracking algorithms to associate the objects in consecutive images.
Our initial evaluation using publicly available object detection datasets demonstrated promise in integrating the developed technologies into autonomous driving products.
The future work involves improving the performance of the prototype with more efficient tracking algorithms, strengthening the robustness of the framework with provable guarantees. Finally, as the concept of abstraction-based monitoring is a generic framework, we also plan to migrate the technique to filter other types of faults as well as move beyond object detection.
\bibliographystyle{IEEEtran}
|
\section*{APPENDIX}
\parindent0pt
\subsection{Proof of Lemma \ref{lem:inexact_oracle}}\label{app:inexact_oracle}
\begin{proof}
The first statement is proved in Lemma 2.1 of \cite{rogozin2020towards}. For the second statement, we have
\begin{align*}
\mathbb{E} \widetilde{g}_{\delta, L, \mu}(\overline x, {\bf x})
&= \frac{1}{n}\sum_{i=1}^n \frac{1}{r} \sum_{j=1}^r \mathbb{E} \nabla{\bf f}_i(x_i,\xi_i^j)
= \frac{1}{n}\sum_{i=1}^n \frac{1}{r} \sum_{j=1}^r \nabla f_i(x_i)
= \frac{1}{n}\sum_{i=1}^n \nabla f_i(x_i) = g_{\delta, L, \mu}(\overline x, {\bf x}).
\end{align*}
It remains to show \eqref{eq:tilde_g_properties_2}.
\begin{align*}
&\mathbb{E} \norm{\widetilde{g}_{\delta, L, \mu}(\overline x, {\bf x}) - g_{\delta, L, \mu}(\overline x, {\bf x})}^2
\leq \mathbb{E} \norm{ \frac{1}{n}\sum_{i=1}^n \frac{1}{r} \sum_{j=1}^r \nabla{\bf f}_i(x_i,\xi_i^j) - \frac{1}{n}\sum_{i=1}^n \nabla f_i(x_i) }^2 \\
&\qquad \leq \frac{1}{n^2} \sum_{i=1}^n \mathbb{E} \norm{ \frac{1}{r} \sum_{j=1}^r \nabla{\bf f}_i(x_i,\xi_i^j) - \nabla f_i(x_i) }^2
\leq \frac{1}{n^2 r^2} \sum_{i=1}^n \sum_{j=1}^r \mathbb{E} \norm{ \nabla{\bf f}_i(x_i,\xi_i^j) - \nabla f_i(x_i) }^2 \\
&\qquad \leq \frac{\sum_{i=1}^n\sigma^2_i}{n^2r}\pd{=
\frac{\sigma_g^2}{nr}}.
\end{align*}
The last inequality follows directly from (\ref{stoch_assumption_on_variance}).
\end{proof}
\subsection{Proof of Theorem \ref{Th:fast_str_conv_adap}}\label{app:fast_str_conv_adap}
For proving the theorem about complexity bounds, we need the following auxiliary Lemma:
\begin{lemma}
\label{lem:opt_subproblem}
Let $\psi$ be convex function. Then for
\begin{equation}
y = \argmin_{x \in Q} \left( \psi(x) + \beta \norm{x-z}^2 + \gamma \norm{x-u}^2 \right),
\end{equation}
where $\beta > 0$ and $\gamma > 0$, the following is true for any $x \in Q$:
\begin{equation}
\psi(x) + \beta \norm{x-z}^2 + \gamma \norm{x-u}^2 \geq \psi(y) + \beta \norm{y-z}^2 + \gamma \norm{y-u}^2 + (\beta + \gamma) \norm{x-y}^2.
\end{equation}
\end{lemma}
\begin{proof}
As $y$ is minimum, the subgradient of function at point $y$ includes $0$:
\begin{equation*}
\exists g: g + \beta \nabla_x \norm{x-z}^2 \vert_{x=y} + \gamma \nabla_x \norm{x-u}^2 \vert_{x=y} = 0.
\end{equation*}
It holds
\begin{align}
\label{eq:opt_subproblem}
\psi(x) - \psi(y) &\geq \angles{g, x-y} = \angles{\beta \nabla_x \norm{x-z}^2 \vert_{x=y} + \gamma \nabla_x \norm{x-u}^2 \vert_{x=y}, y - x} \\
&= \angles{2 \beta (y - z) + 2 \gamma (y - u), y - x}
\end{align}
and we get that
\begin{align*}
2 \angles{y-z,y-x} &= \norm{y}^2 - \norm{z}^2 - 2 \angles{z, y-z}+\norm{x}^2 - \norm{y}^2 - 2 \angles{y, x-y} - \norm{x}^2 + \norm{z}^2 + 2 \angles{z,x-z} \\
&= \norm{y-z}^2 + \norm{x-y}^2 - \norm{x-z}^2.
\end{align*}
After similar manipulations with the $2 \angles{y-u,y-x}$ term and replacing the right part in \eqref{eq:opt_subproblem}, the lemma statement is obtained.
\end{proof}
Now we pass to the proof of Theorem \ref{Th:fast_str_conv_adap} itself.
\begin{proof}
We begin from the right inequality from \eqref{eq:inexact_oracle_def_devolder}:
$$f(y) - \cbraces{f_{\delta_{k+1},L,\mu}(x) + \angles{g_{\delta_{k+1},L,\mu}(x), y - x}} \le \frac{L}{2} \norm{y - x}^2 + \delta_{k+1}.$$
It can be rewritten as
\begin{align*}
f(x^{k+1}) &- \cbraces{f_{\delta_{k+1},L,\mu}(y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1},L,\mu}(y^{k+1}), x^{k+1} - y^{k+1}}} \\ &\leq \angles{g_{\delta_{k+1},L,\mu}(y^{k+1}) - \widetilde{g}^{\ r_{k+1}}_{\delta_{k+1},L,\mu}(y^{k+1}), x^{k+1} - y^{k+1}} + \frac{L}{2} \norm{y^{k+1} - x^{k+1}}^2 + \delta_{k+1}.
\end{align*}
The first term in the right hand side can be estimated using Young inequality:
\begin{align*}
\angles{g_{\delta_{k+1},L,\mu}(y^{k+1}) - \widetilde{g}^{\ r_{k+1}}_{\delta_{k+1},L,\mu}(y^{k+1}), x^{k+1} - y^{k+1}} \leq \frac{L}{2} \norm{x^{k+1} - y^{k+1}}^2 + \frac{1}{2L} \norm{g_{\delta_{k+1},L,\mu}(y^{k+1}) - \widetilde{g}^{\ r_{k+1}}_{\delta_{k+1},L,\mu}(y^{k+1})}^2
\end{align*}
The combination of the last two expressions yields
\begin{align*}
f(x^{k+1}) \leq f_{\delta_{k+1},L,\mu}(y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1},L,\mu}(y^{k+1}), x^{k+1} - y^{k+1}} + L \norm{y^{k+1} - x^{k+1}}^2 + \lambda + \delta_{k+1},
\end{align*}
where
$$
\lambda := \frac{1}{2L}\norm{g_{\delta_{k+1},L,\mu}(y^{k+1}) - \widetilde{g}^{\ r_{k+1}}_{\delta_{k+1},L,\mu}(y^{k+1})}^2
$$
for brevity. We substitute $x^{k+1}, y^{k+1}$ into several terms by their definitions:
\begin{align*}
&f(x^{k+1}) \leq f_{\delta_{k+1},L,\mu}(y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1},L,\mu}(y^{k+1}), \frac{\alpha^{k+1} u^{k+1} + A^k x^k}{A^{k+1}} - y^{k+1}} \\ &+ L \norm{\frac{\alpha^{k+1}u^k + A^k x^k}{A^{k+1}} - \frac{\alpha^{k+1} u^{k+1} + A^k x^k}{A^{k+1}}}^2 + \lambda + \delta_{k+1}.
\end{align*}
As $A^{k+1} = A^k + \alpha^{k+1}$ by definition and as dot product is convex, we get the following:
\begin{align*}
f(x^{k+1}) &\leq \frac{A^k}{A^{k+1}} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1} }\right) \\
&\quad+ \frac{\alpha^{k+1}}{A^{k+1}} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}), u^{k+1} - y^{k+1}} \right) + \frac{L (\alpha^{k+1})^2}{(A^{k+1})^2} \norm{u^k - u^{k+1}}^2 + \lambda + \delta_{k+1}.
\end{align*}
By definition of $\alpha^{k+1}$ we have:
\begin{align*}
f(x^{k+1}) &\leq \frac{A^k}{A^{k+1}} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1} }\right) \\
&\quad + \frac{\alpha^{k+1}}{A^{k+1}} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}), u^{k+1} - y^{k+1}} \right) + \frac{1 + A^k \mu}{A^{k+1}} \norm{u^k - u^{k+1}}^2 + \lambda + \delta_{k+1}.
\end{align*}
We rewrite that as:
\begin{align*}
f(x^{k+1}) &\leq \frac{A^k}{A^{k+1}} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{g_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1}} + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu}(y^{k+1}) - g_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1}} \right) \\
&\quad+ \frac{\alpha^{k+1}}{A^{k+1}} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}), u^{k+1} - y^{k+1}} \right) + \frac{1 + A^k \mu}{A^{k+1}} \norm{u^k - u^{k+1}}^2 + \lambda + \delta_{k+1}.
\end{align*}
Using left part of (\ref{eq:inexact_oracle_def_devolder}), we get:
\begin{align}
\label{eq:convergency_proof_intermediate}
f(x^{k+1}) &\leq \frac{A^k}{A^{k+1}} \left( f
(x^k) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu}(y^{k+1}) - g_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1}} \right) \\ &+ \frac{\alpha^{k+1}}{A^{k+1}} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}), u^{k+1} - y^{k+1}} \right) + \frac{1 + A^k \mu}{A^{k+1}} \norm{u^k - u^{k+1}}^2 + \lambda + \delta_{k+1}.
\end{align}
From lemma \ref{lem:opt_subproblem} for optimization problem at step 5 in Algorithm \ref{alg:FastAlg2_strong} we have:
\begin{align*}
\alpha^{k+1} &\angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,u^{k+1}-y^{k+1}} + (1 + A^k \mu) \norm{u^{k+1} - u^k}^2 + \alpha^{k+1} \mu \norm{u^{k+1} - y^{k+1}}^2 \\
&+ (1 + A^k \mu + \alpha^{k+1} \mu ) \norm{u^{k+1} - x}^2 \leq \alpha^{k+1} \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,x-y^{k+1}} \\ &+ (1 + A^k \mu) \norm{x - u^k}^2 + \alpha^{k+1} \mu \norm{x - y^{k+1}}^2.
\end{align*}
As squared norm is always non-negative, we obtain
\begin{align}
\label{eq:convergency_proof_lemma}
\alpha^{k+1} &\angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,u^{k+1}-y^{k+1}} + (1 + A^k \mu) \norm{u^{k+1} - u^k}^2
\leq -(1 + A^k \mu + \alpha^{k+1} \mu ) \norm{u^{k+1} - x}^2 \\ &+ \alpha^{k+1} \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,x-y^{k+1}} + (1 + A^k \mu) \norm{x - u^k}^2 + \alpha^{k+1} \mu \norm{x - y^{k+1}}^2
\end{align}
Combining inequalities \eqref{eq:convergency_proof_intermediate} and \eqref{eq:convergency_proof_lemma}, we get:
\begin{align*}
A^{k+1} f(x^{k+1}) \leq~&A^k \left( f (x^k) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu}(y^{k+1}) - g_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1}} \right) \\
&+ \alpha^{k+1} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,x-y^{k+1}} + \mu \norm{x - y^{k+1}}^2 \right) \\
&+ (1 + A^k \mu) \norm{x - u^k}^2 - (1 + A^k \mu + \alpha^{k+1} \mu ) \norm{u^{k+1} - x}^2 + A^{k+1} \lambda + \delta_{k+1} A^{k+1} \\
=~&A^k \left( f (x^k) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu}(y^{k+1}) - g_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1}} \right) \\
&+ \alpha^{k+1} \left( f_{\delta_{k+1}, L, \mu} (y^{k+1}) + \angles{{g}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,x-y^{k+1}} \right) \\
&+ \alpha^{k+1} \left( \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}) - {g}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,x-y^{k+1}} + \mu \norm{x - y^{k+1}}^2 \right) \\ &+ (1 + A^k \mu) \norm{x - u^k}^2 - (1 + A^k \mu + \alpha^{k+1} \mu ) \norm{u^{k+1} - x}^2 + A^{k+1} \lambda + \delta_{k+1} A^{k+1}.
\end{align*}
Using the left part of \eqref{eq:inexact_oracle_def_devolder} again results in
\begin{align*}
A^{k+1} f(x^{k+1}) \leq &{A^k} \left( f (x^k) + \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu}(y^{k+1}) - g_{\delta_{k+1}, L, \mu} (y^{k+1}), x^k - y^{k+1}} \right) \\
&+ \alpha^{k+1} \left( f (x)+ \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu} (y^{k+1}) - {g}_{\delta_{k+1}, L, \mu} (y^{k+1}) ,x-y^{k+1}} + \right) \\ &+ (1 + A^k \mu) \norm{x - u^k}^2 - (1 + A^k \mu + \alpha^{k+1} \mu) \norm{u^{k+1} - x}^2 + A^{k+1} \lambda + \delta_{k+1} A^{k+1}.
\end{align*}
We can take expectation and we may see that angles terms go zero as $\mathbb{E} \widetilde{g}^{\ r}_{\delta, L, \mu}(x) = g_{\delta, L, \mu}(x)$ for any $r$, and $\lambda \leq \frac{\sigma^2}{2Lr_{k+1}}$:
\begin{align*}
A^{k+1} \mathbb{E} f(x^{k+1}) &- {A^k} f (x^k) \leq \alpha^{k+1} f (x) \\ &+ (1 + A^k \mu) \norm{x - u^k}^2 - (1 + A^{k+1} \mu ) \mathbb{E} \norm{u^{k+1} - x}^2 + \frac{\sigma^2 A^{k+1}}{2Lr_{k+1}} + \delta_{k+1} A^{k+1}.
\end{align*}
Now we should pay attention to the fact that the expectation is \textit{conditional} because we consider $x^k$ and other $k$-th variables known before the iteration:
\begin{align*}
A^{k+1} \mathbb{E} \left[ f(x^{k+1}) | x^k, \ldots, x^1 \right] &- {A^k} f (x^k) \leq \alpha^{k+1} f (x) + (1 + A^k \mu) \norm{x - u^k}^2 \\ &- (1 + A^{k+1} \mu ) \mathbb{E} \left[ \norm{u^{k+1} - x}^2 | x^k, \ldots, x^1 \right] + \frac{\sigma^2 A^{k+1}}{2Lr_{k+1}} + \mathbb{E}[\delta_{k+1} | x^k, \ldots, x^1] A^{k+1}.
\end{align*}
If we take $x = x^*$, write these inequalities for all $k$ from $0$ to $N-1$ and sum up all of them, we will get the following:
\begin{align*}
\sum_{i=1}^N A^i &\mathbb{E} \left[ f(x^i) | x^i, \ldots, x^1 \right] \leq \sum_{i=0}^{N-1} A^i f(x^i) + \sum_{i=1}^N \alpha_i f(x^*) + \\
&+ \sum_{i=0}^{N-1} (1 + A^i \mu) \norm{u^i - x^*}^2 - \sum_{i=1}^N (1 + A^i \mu) \mathbb{E} \left[ \norm{u^i - x^*}^2 | x^i, \ldots, x^1 \right] + \sum_{i=1}^N A_i\left( \frac{\sigma^2}{2Lr_{i}} + \mathbb{E}[\delta_i | x^i, \ldots, x^1] \right).
\end{align*}
Next, we use the law of total expectation $N$ times and get rid of conditional expectations, and after that get rid of similar terms:
\begin{gather}
\label{eq:agd_bound}
\mathbb{E} A^N f(x^N) \leq A^N f(x^*) + \norm{u^0 - x^*}^2 - (1 + A^N \mu) \mathbb{E} \norm{u^N - x^*}^2 +
\sum_{i=1}^N A_i \left( \frac{\sigma^2}{2Lr_{i}} + \mathbb{E}\delta_i \right)
\end{gather}
Here we also recall that $A_0 = \alpha_0 = 0, \sum_{i=1}^N \alpha_i = A_N$.
\par Finally, we get
$$ \mathbb{E} f(x^N) - f(x^*) \leq \frac{1}{A^N} \left( \norm{u^0 - x^*}^2 + \sum_{i=1}^N A^i \left( \frac{\sigma^2}{2Lr_{i}} + \mathbb{E}\delta_i \right) \right).
$$
The second inequality is obtained from \eqref{eq:agd_bound} and the fact that $f(x) \geq f(x^*)$.
\end{proof}
\subsection{Proof of Lemma \ref{lem:Ak_properties}}\label{app:Ak_properties}
\begin{proof}
In view of definition of sequence $\alpha^{k+1}$, we have:
\begin{align*}
A^N &\leq A^N(1 + \mu A^{N-1}) = L(A^N - A^{N-1})^2\\
&\leq L(\sqrt{A^N} - \sqrt{A^{N-1}})^2(\sqrt{A^N} + \sqrt{A^{N-1}})^2 \leq 4L^{N} A^N (\sqrt{A^N} - \sqrt{A^{N-1}})^2.
\end{align*}
For the case when $\mu > 0$ we obtain:
\begin{align*}
\mu A^{N-1} A^{N} \leq A^N(1 + \mu A^{N-1}) \leq 4L A^N (\sqrt{A^N} - \sqrt{A^{N-1}})^2.
\end{align*}
From the fact that $A^1 = 1 / L$ and the last inequality we can show that
\begin{align*}
\sqrt{A^N} \geq \left(1 + \frac{1}{2}\sqrt{\frac{\mu}{L}}\right)\sqrt{A^{N-1}} \geq \frac{1}{\sqrt{L}}\left(1 + \frac{1}{2}\sqrt{\frac{\mu}{L}}\right)^{(N-1)}.
\end{align*}
For the second statement, we recall the proof of Lemma A.1 in \cite{rogozin2020towards}.
Update rule for $A^k$ writes as
\begin{align}\label{eq:coef_sequence}
1 + \mu A^k = \frac{L(\alpha^{k+1})^2}{A^{k+1}},~ A^{k} = \sum_{i=0}^k \alpha^{i},~ \alpha^0 = 0.
\end{align}
A sequence $\braces{B^k}_{k=0}^\infty$ with a similar update rule is studied in \cite{devolder2013first}.
\begin{align}\label{eq:nesterov_coef_sequence}
L + \mu B^k = \frac{L (\beta^{k+1})^2}{B^{k+1}},~ B^k = \sum_{i=0}^k \beta^i,~ \beta^0 = 1,
\end{align}
and for sequence $\braces{B^k}_{k=0}^\infty$ it is shown $\frac{\sum_{i=0}^k B^i}{B^k}\le 1 + \sqrt{L / \mu}$.
Dividing \eqref{eq:nesterov_coef_sequence} by L yields
\begin{align*}
1 + \mu(B^k / L) = \frac{L (\beta^{k+1} / L)^2}{(B^{k+1} / L)},
\end{align*}
which means that update rule for $B^k / L$ is equivalent to \eqref{eq:coef_sequence}. Since $A^1 = 1 / L = B^0 / L$, it holds $A^{k+1} = B^k / L,~ k\ge 0$ and
\begin{align*}
\frac{\sum_{i=1}^k A^i}{A^k} = \frac{\sum_{i=0}^{k-1} B^i / L}{B^{k-1} / L} \le 1 + \sqrt\frac{L}{\mu}.
\end{align*}
\end{proof}
\subsection{Proof of Lemma \ref{lem:consensus_iters_strongly_convex}}\label{app:consensus_iters_strongly_convex}
\begin{proof}
The proof follows by revisiting proof of Lemma A.3 in \cite{rogozin2020towards} in stochastic setting. First, note that multiplication by a mixing matrix does not change the average of a vector, i.e. $\frac{1}{n}\mathbf{1}_n\onevec^\top{\bf x} = \frac{1}{n}\mathbf{1}_n\onevec^\top {\bf W}^k {\bf x}~ \forall k\ge 0$. This means $\overline{\bf u}^{k+1} = \overline{\bf v}^{k+1}$.
Second, let us use the contraction property of mixing matrix sequence $\{{\bf W}^k\}_{k=0}^\infty$. We have
\begin{align*}
\mathbb{E}\norm{{\bf u}^{k+1} - \overline{\bf u}^{k+1}}^2
&\le (1 - \lambda)^{2(T/\tau)} \mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf u}^{k+1}}^2
\le e^{-2(T/\tau)\lambda} \mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf u}^{k+1}}^2. \\
\end{align*}
Assuming that $\mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf u}^{k+1}}^2\le D$, we only need $T = \frac{\tau}{2\lambda}\log\frac{D}{\delta'}$ iterations to ensure $\mathbb{E}\norm{{\bf u}^{k+1} - \overline{\bf u}^{k+1}}^2\le\delta'$. In the rest of the proof, we show that $\mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf u}^{k+1}} = \mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf v}^{k+1}} \le \sqrt D$.
According to update rule of Algorithm \ref{alg:decentralized_agd}, it holds
\begin{align*}
\mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf v}^{k+1}}
&\le \frac{\alpha^{k+1}\mu \mathbb{E}\norm{{\bf y}^{k+1} - \overline{\bf y}^{k+1}}}{1 + A^{k+1}\mu} + \frac{(1 + A^k\mu)\mathbb{E}\norm{{\bf u}^k - \overline{\bf u}^k}}{1 + A^{k+1}\mu} + \frac{\alpha^{k+1}}{1 + A^{k+1}\mu} \mathbb{E}\norm{\nabla^r F({\bf y}^{k+1})} \\
&\le \sqrt{\delta'} + \frac{\alpha^{k+1}}{1 + A^{k+1}\mu} \mathbb{E}\norm{\nabla^r F({\bf y}^{k+1})}.
\end{align*}
We estimate $\norm{\nabla^r F({\bf y}^{k+1})}$ using $L_{\xi}$-smoothness of $\nabla F$:
\begin{align*}
\norm{\nabla^r F({\bf y}^{k+1})}
&\le \norm{\nabla^r F({\bf y}^{k+1}) - \nabla^r F({\bf x}^*)} + \norm{\nabla^r F({\bf x}^*)} \\
&\le L_{\xi}\underbrace{\norm{{\bf y}^{k+1} - \overline{\bf y}^{k+1}}}_{\le\sqrt{\delta'}} + L_{\xi}\underbrace{\norm{\overline{\bf y}^{k+1} - {\bf x}^*}}_{= \sqrt n \norm{\overline y^{k+1} - x^*}} + \norm{\nabla^r F({\bf x}^*)} \addtocounter{equation}{1}\tag{\theequation}\label{eq:gradeint_y_upper_bound}
\end{align*}
where $x^* = \argmin_{x\in\mathbb{R}^d} f(x),~ {\bf x}^* = \mathbf{1}_n (x^*)^\top$. It remains to estimate $\norm{\overline y^{k+1} - x^*}$.
\begin{align*}
\norm{\overline y^{k+1} - x^*}
&\le \frac{\alpha^{k+1}}{A^{k+1}} \norm{\overline x^{k+1} - x^*} + \frac{A^{k}}{A^{k+1}}\norm{\overline u^{k+1} - x^*}
\le \max\braces{\norm{\overline x^{k+1} - x^*}, \norm{\overline u^{k+1} - x^*}}
\end{align*}
By Lemma \ref{lem:inexact_agd_convergence} and strong convexity of $f$:
\begin{align*}
\mathbb{E}\norm{\overline x^{k+1} - x^*}^2
&\le \frac{2}{\mu}\cbraces{\mathbb{E} f(\overline x^{k+1}) - f(x^*)}
\le \frac{\norm{\overline u^0 - x^*}^2}{A^{k+1}\mu} + \frac{\sum_{i=1}^{k+1} A^i}{A^{k+1}\mu} \cbraces{\frac{\sigma^2}{2Lr} + \delta}
\end{align*}
and therefore
\begin{align*}
\mathbb{E}\norm{\overline y^{k+1} - x^*}^2
&\le \max\braces{\frac{\norm{\overline u^0 - x^*}^2}{A^{k+1}\mu} + \frac{\sum_{i=1}^{k+1} A^i}{A^{k+1}\mu}\cbraces{\frac{\sigma^2}{2Lr} + \delta},~ \frac{\norm{\overline u^0 - x^*}^2}{1 + A^{k+1}\mu} + \frac{\sum_{i=1}^{k+1} A^i}{1 + A^{k+1}\mu}\cbraces{\frac{\sigma^2}{2Lr} + \delta}} \\
&\le \frac{\norm{\overline u^0 - x^*}^2}{A^{k+1}\mu} + \frac{1}{\mu} \cbraces{1 + \sqrt\frac{L}{\mu}} \cbraces{\frac{\sigma^2}{2Lr} + \delta},
\end{align*}
where the last inequality holds by Lemma \ref{lem:Ak_properties}.
Returning to \eqref{eq:gradeint_y_upper_bound}, we get
\begin{align*}
&\norm{\nabla^r F({\bf y}^{k+1})} \\
&\qquad\le L_{\xi}\sqrt{\delta'} + L_{\xi}\sqrt{n} \cbraces{\frac{\norm{\overline u^0 - x^*}^2}{A^{k+1}\mu} + \frac{1}{\mu} \cbraces{1 + \sqrt\frac{L}{\mu}} \cbraces{\frac{\sigma^2}{2Lr} + \delta}}^{1/2} + \norm{\nabla^r F({\bf x}^*)} \\
&\qquad\le L_{\xi}\sqrt{\delta'} + L_{\xi}\sqrt{n} \cbraces{\frac{L}{\mu} \norm{\overline u^0 - x^*}^2 \cbraces{1 + \sqrt\frac{\mu}{2L}}^{-2k} + \frac{2L^{1/2}}{\mu^{3/2}} \cbraces{\frac{\sigma^2}{2Lr} + \delta}}^{1/2} + \norm{\nabla^r F({\bf x}^*)} \\
&\qquad\le L_{\xi}\sqrt{\delta'} + L_{\xi}\sqrt{n} \cbraces{\frac{L}{\mu} \norm{\overline u^0 - x^*}^2 + \frac{2L^{1/2}}{\mu^{3/2}} \cbraces{\frac{\sigma^2}{2Lr} + \delta}}^{1/2} + \norm{\nabla^r F({\bf x}^*)}.
\end{align*}
For distance to consensus of ${\bf v}^{k+1}$, it holds
\begin{align*}
\mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf v}^{k+1}} \le \sqrt{\delta'} + \frac{\alpha^{k+1}}{1 + A^k\mu + \mu} \mathbb{E}\norm{\nabla F\cbraces{{\bf y}^{k+1}}}
\end{align*}
We estimate coefficient by $\mathbb{E}\norm{\nabla F\cbraces{{\bf y}^{k+1}}}$ using the definition of $\alpha^{k+1}$.
\begin{align*}
&1 + A^k\mu = \frac{L(\alpha^{k+1})^2}{A^k + \alpha^{k+1}} \\
&L(\alpha^{k+1})^2 - (1 + A^k\mu)\alpha^{k+1} - (1 + A^k\mu)A^k = 0 \\
&\alpha^{k+1} = \frac{1+ A^k\mu + \sqrt{(1 + A^k\mu)^2 + 4LA^k(1 + A^k\mu)}}{2L} \\
&\frac{\alpha^{k+1}}{1 + A^{k+1}\mu} \le \frac{\alpha^{k+1}}{1 + A^k\mu} = \frac{1}{2L} \cbraces{1 + \sqrt{1 + 4\frac{LA^k}{1 + A^k\mu}}} \\
&\qquad\le \frac{1}{2L}\cbraces{\sqrt{\frac{L}{\mu}} + \sqrt{\frac{L}{\mu} + 4\frac{L}{\mu}}} \le \frac{2}{\sqrt{L\mu}}
\end{align*}
Returning to ${\bf v}^{k+1}$, we get
\begin{align*}
&\mathbb{E}\norm{{\bf v}^{k+1} - \overline{\bf v}^{k+1}} \\
&\qquad\le \cbraces{\frac{2L_{l}}{\sqrt{L\mu}} + 1}\sqrt{\delta'} + L_{l}\sqrt{\frac{n}{L\mu}} \cbraces{\frac{L}{\mu} \norm{\overline u^0 - x^*}^2 + \frac{2L^{1/2}}{\mu^{3/2}} \cbraces{\frac{\sigma^2}{2Lr} + \delta}}^{1/2} + \frac{2\mathbb{E}\norm{\nabla^r F({\bf x}^*)}}{\sqrt{L\mu}} \\
&\qquad\leq \cbraces{\frac{2L_{l}}{\sqrt{L_{g}\mu_{g}}} + 1}\sqrt{\delta'} + \frac{2L_{l}}{\mu_{g}} \sqrt{n} \cbraces{\norm{\overline u^0 - x^*}^2 + \frac{2}{\sqrt{L_{g}\mu_{g}}}\cbraces{\frac{\sigma_{g}^2}{4nL_{g} r^2} + \delta}}^{1/2} + \frac{2nM_\xi}{\sqrt{L_{g}\mu_{g}}} = \sqrt D,
\end{align*}
where in the last inequality we used $\norm{\nabla^r F({\bf x}^*)} \leq n M_\xi$.
\end{proof}
\subsection{Putting the proof of Theorem \ref{th:total_iterations_strongly_convex} together}\label{app:total_iterations_strongly_convex}
Let us show that choice of number of subroutine iterations $T_k = T$ yields
\begin{align*}
\mathbb{E} f(\overline x^k) - f(x^*) \le \frac{1}{A^k}\cbraces{\norm{\overline u^0 - x^*}^2 + \cbraces{\frac{\sigma^2}{2Lr} + \delta} \sum_{i=1}^k A^i}
\end{align*}
by induction. At $k=0$, we have $\norm{{\bf u}^0 - \overline{\bf u}^0} = 0$ and by Lemma \ref{lem:inexact_agd_convergence} it holds
\begin{align*}
\mathbb{E} f(\overline x^1) - f(x^*) &\le \frac{1}{A^1}\cbraces{\norm{\overline u^0 - x^*}^2 + \cbraces{\frac{\sigma^2}{2Lr} + \delta} A^1}.
\end{align*}
For induction pass, assume that $\mathbb{E}\norm{{\bf u}^j - \overline{\bf u}^j}^2\le \delta'$ for $j = 0,\ldots, k$. By Lemma \ref{lem:consensus_iters_strongly_convex}, if we set $T_k = T$, then $\mathbb{E}\norm{{\bf u}^{k+1} - \overline{\bf u}^{k+1}}^2\le \delta'$. Applying Lemma \ref{lem:inexact_agd_convergence} again, we get
\begin{align*}
\mathbb{E} f(\overline x^k) - f(x^*) \le \frac{1}{A^k}\cbraces{\norm{\overline u^0 - x^*}^2 + \cbraces{\frac{\sigma^2}{2Lr} + \delta} \sum_{i=1}^k A^i}.
\end{align*}
Next, we substitute a bound on $A^k$ from Lemma \ref{lem:Ak_properties} and get
\begin{align*}
\mathbb{E} f(\overline x^N) &- f(x^*) \\
&~\le LR^2 \cbraces{1 + \frac{1}{2}\sqrt\frac{\mu}{L}}^{-2(N-1)} + \cbraces{1 + \sqrt\frac{L}{\mu}} \cbraces{\frac{\sigma^2}{2Lr} + \delta} \\
&~ = 2L_{g} R^2\cbraces{1 + \frac{1}{4}\sqrt\frac{\mu_{g}}{L_{g}}}^{-2(N-1)} \hspace{-0.3cm} + \cbraces{1 + 2\sqrt\frac{L_{g}}{\mu_{g}}} \cbraces{\frac{\sigma^2}{4L_{g} r} + \delta}.
\end{align*}
It remains to estimate the number of iterations required for $\varepsilon$-accuracy. In order to satisfy
\begin{align*}
2L_{g}\norm{\overline u^0 - x^*}^2\cbraces{1 + \frac{1}{4}\sqrt\frac{\mu_{g}}{L_{g}}}^{-2(N-1)} &\le \frac{\varepsilon}{2}, \\
\cbraces{1 + 2\sqrt\frac{L_{g}}{\mu_{g}}} \cbraces{\frac{\sigma^2}{4L_{g} r} + \delta} &\le \frac{\varepsilon}{2},
\end{align*}
it is sufficient to choose $\delta' = \dfrac{n\varepsilon}{32} \dfrac{\mu_{g}^{3/2}}{L_{g}^{1/2} L_{l}^2},~ r = \dfrac{2\sigma^2}{\varepsilon\sqrt{L_{g}\mu_{g}}}$ and
\begin{align*}
N = 3\sqrt{\frac{L_{g}}{\mu_{g}}} \log\cbraces{\frac{4L_{g}\norm{\overline u^0 - x^*}^2}{\varepsilon}}.
\end{align*}
\pd{Finally, the total number of stochastic oracle calls per node equals
\begin{align*}
N_{orcl} = N\cdot r = \dfrac{6\sigma_{g}^2}{n\mu_{g}\varepsilon} \log\cbraces{\frac{4L_{g}\norm{\overline u^0 - x^*}^2}{\varepsilon}}.
\end{align*}
Further, the total number of communications is
\begin{align*}
N_{\text{comm}}
&= N\cdot T = 3\sqrt\frac{L_{g}}{\mu_{g}} \log\cbraces{\frac{4L_{g}\norm{\overline u^0 - x^*}^2}{\varepsilon}}\cdot \kappa \cdot \log\frac{D}{\delta'} \\
&= O\left(\sqrt{\frac{L_{g}}{\mu_{g}}} \kappa \cdot \log\cbraces{\frac{4L_{g}\norm{\overline u^0 - x^*}^2}{\varepsilon}} \log\frac{D}{\delta'}\right),
\end{align*}
where $\kappa = \frac{\tau}{2\lambda}$ if the communication network is time-varying and $\kappa = \sqrt{\chi}$ if the communication network is fixed.}
\section{Introduction}
Distributed algorithms have already about half a century history \cite{bor82,tsi84,deg74} with many applications including robotics, resource allocation, power system control, control of drone or satellite networks, distributed statistical inference and optimal transport, multiagent reinforcement learning \cite{xia06,rab04,ram2009distributed,kra13,ned17e,nedic2017fast,uribe2018distributed,kroshnin2019complexity,ivanova2020composite}.
Recently, development of such algorithms has become one of the main topics in optimization and machine learning motivated by large-scale learning problems with privacy constraints and other challenges such as data being produced or stored distributedly \cite{bot10,boy11,aba16,ned16w,ned15}. An important part of this research studies decentralized distributed optimization algorithms over arbitrary networks. In this setting a network of computing agents, e.g. sensors or computers, is represented by a connected graph in which two agents can communicate with each other if there is an edge between them. This imposes communication constraints and the goal of the whole system~\cite{ned09,ram10,daneshmand2021newton} is to cooperatively minimize a global objective using only local communications between agents, each of which has access only to a local piece of the global objective. Due to random nature of the optimized process or randomness and noise in the used data, a particular important setting is distributed stochastic optimization. Moreover, the topology of the network can vary in time, which may prevent fast convergence of an algorithm.
More precisely, we consider the following optimization problem
\begin{align}\label{eq:initial_problem}
\min_{x\in\mathbb{R}^d} \left[f(x) := \frac{1}{n}\sum_{i=1}^n f_i(x)\right], \quad f_i(x):=\mathbb{E}_{\xi_i\sim{\mathcal{D}}_i} {\bf f}_i(x, \xi_i),
\end{align}
where $\xi_i$'s are random variables with probability distributions ${\mathcal{D}}_i$. For each $i=1,...,n$ we make the following assumptions: $f_i(x)$ is a convex function and that almost sure w.r.t. distribution ${\mathcal{D}}_i$, the function ${\bf f}_i(x, \xi_i)$ has gradient $\nabla {\bf f}_i(x, \xi_i)$, which is $L_i(\rev{\xi_i})$-Lipschitz continuous with respect to the Euclidean norm.
Further, for each $i=1,...,n$, we assume that we know a constant $L_i\geqslant 0$ such that $\sqrt{\mathbb{E}_{\rev{\xi_i}} L_i(\rev{\xi_i})^2 } \leq L_i < +\infty$. Under these assumptions, $\mathbb{E}_{\rev{\xi_i}}\nabla {\bf f}_i(x, \xi_i) = \nabla f_i(x)$ and $f$ is $L_i$-smooth, i.e. has $L_i$-Lipschitz continuous gradient with respect to the Euclidean norm. Also, we assume that, for all $x$, and $i$
\begin{equation}
\label{stoch_assumption_on_variance}
\mathbb{E}_{\rev{\xi_i}}[\norm{\nabla {\bf f}_i(x, \xi_i) - \nabla f_i(x)}^2] \leqslant \sigma_i^2,
\end{equation}
where $\norm{\cdot}$ is the Euclidean norm. Finally, we assume that each $f_i$ is $\mu_i$-strongly convex \rev{($\mu_i > 0$)}. Important characteristics of the objective in \eqref{eq:initial_problem} are local strong convexity parameter $\displaystyle \mu_{l} = \min_i\mu_i$ and local smoothness constant $L_{l} = \max_i L_i$, which define local condition number $\kappa_l=L_{l}/\mu_{l}$, as well as their global counterparts $\mu_{g} = \frac{1}{n}\sum_{i=1}^n\mu_i,~ L_{g} = \frac{1}{n}\sum_{i=1}^n L_i$, $\kappa_g=L_{g} /\mu_{g}$.
The global condition number may be significantly better than local (see e.g. \cite{scaman2017optimal} for details) and it is desired to develop algorithms with complexity depending on the global condition number. Moreover, we introduce a worst-case smoothness constant over stochastic realizations $\displaystyle L_{\xi} = \max_i \max_{\xi} L_i(\xi)$ and a maximum gradient norm at optimum $\displaystyle M_\xi = \max_i \max_{\xi} \norm{\nabla f_i(x^*, \xi)}$ and assume that these constants are well-defined \rev{(finite)}. Similarly to global smoothness and strong convexity constants, we introduce $\sigma_{g}^2 = \frac{1}{n}\sum_{i=1}^n \sigma_i^2$.
To introduce the distributed optimization setup, we assume that \rev{communication constraints in the computational network are represented by an undirected communication graph which may vary with time. Namely, the network is modeled with a sequence of graphs $\braces{\mathcal{G}^k = (V, E^k)}_{k=0}^\infty$. We note that the set of vertices remains the same, while set of edges is allowed to change with time.} Each agent in the network corresponds to a graph vertex and communication \rev{at time slot $k$} is possible only between nodes which are connected by an edge \rev{at this time slot}. Further, each agent $i$ has access only to iid samples from ${\mathcal{D}}_i$ and corresponding stochastic gradients $\nabla {\bf f}_i(x, \xi_i)$. The goal of the whole network is to solve the minimization problem \eqref{eq:initial_problem} by using only communication between neighboring nodes. The performance of decentralized optimization algorithms depend\rev{s} on the characteristic number $\chi$ of the network that quantifies its connectivity and how fast the information is spread over the network. The precise definition will be given later.
\subsection{Related work}
In centralized setting optimal methods exist \cite{ghadimi2013optimal} for the considered setting of smooth strongly convex stochastic optimization, as well as many algorithms for other settings \cite{dvurechensky2018parallel,Zhang2018,dvurechensky2021hyperfast,agafonov2021accelerated}.
Decentralized distributed optimization introduces several challenges, one of them being that one has to care about two complexities: number of oracle calls which are made by each agent and the number of communication steps, which are sufficient to reach a given accuracy $\varepsilon$. In the simple case of all constants $\mu_i,L_i,\sigma_i$ being independent on $i$, the oracle complexity lower bound \cite{scaman2017optimal,dvinskikh2019decentralized}
$\Omega\left(\max\left\{\frac{\sigma^2}{n\mu\varepsilon},\sqrt{\frac{L}{\mu}}\ln\frac{1}{\varepsilon}\right\}\right)$ is a clear counterpart of the \rev{centralized} lower bound \cite{Nemirovskii1983}. The lower bound on communications number $\Omega\left(\sqrt{\frac{L}{\mu}\chi}\ln\frac{1}{\varepsilon} \right)$ corresponds to decentralized deterministic optimization and, compared to standard non-distributed accelerated methods \cite{nesterov1983method,nesterov2020primal-dual,guminov2019accelerated}, has an additional network-dependent factor $\sqrt{\chi}$. Existing distributed algorithms \cite{fallah2019robust,dvinskikh2019primal,dvinskikh2019decentralized,olshevsky2019asymptotic,olshevsky2019non} achieve either the lower oracle complexity bound or the lower communication complexity bound, but not both simultaneously. In this paper we propose an algorithm which closes this gap and achieves both bounds simultaneously.
Deterministic decentralized \rev{optimization} is quite well understood with many centralized algorithms having their decentralized counterparts. For example, there are decentralized subgradient method \cite{Nedic2009}, gradient methods \cite{shi2015extra,rogozin2019projected} and many variants of accelerated gradient methods \cite{Qu2017,ye2020multi,li2018sharp,Jakovetic,dvu2018,dvinskikh2019decentralized,li2020revisiting}, which achieve both communication and \rev{oracle} complexity lower bounds \cite{hendrikx2020optimal,li2020optimal,scaman2017optimal,tang2019practicality}.
The negative side of the majority of the accelerated distributed methods is that their complexity depends on the local condition number $\kappa_l$, which may be larger than the global condition number $\kappa_g$, which corresponds to the centralized optimization. A number of methods \cite{scaman2017optimal,Wu2017,Zhang2017,dvu2018,uribe2020dual} require an assumption that the Fenchel conjugate for each $f_i(x)$ is available, which may be restrictive in practice.
In contrast, our complexity bounds depend on the global condition number and we use the primal approach without additional assumptions on $f_i$'s.
Another important part of our paper is decentralized distributed optimization on time-varying networks. This area is not understood as well as the simpler setting of fixed networks. The first method with provable geometric convergence was proposed in \cite{Nedic2017achieving}. Such primal algorithms as Push-Pull Gradient Method \cite{Pu2018} and DIGing \cite{Nedic2017achieving} are robust to network changes and have theoretical guarantees of convergence over time-varying graphs. Recently, a dual method for time-varying architectures was introduced in \cite{Maros2018}. All these methods do not allow to achieve the above lower bounds. \pd{Moreover, we are not aware of any accelerated algorithms for stochastic optimization on time-varying networks.}
\subsection{Our contributions}
\pd{
The main contribution of this paper is twofold. Firstly, when the communication network is fixed, we propose the first optimal up to logarithmic factors accelerated decentralized distributed algorithm for stochastic convex optimization. This means that our algorithm has
oracle per node complexity $\widetilde{O}\left(\max\left\{\frac{\sigma_g^2}{n\mu_g\varepsilon},\sqrt{\frac{L_g}{\mu_g}}\ln\frac{1}{\varepsilon}\right\}\right)$ and communication complexity $\widetilde{O}\left(\sqrt{\frac{L_g}{\mu_g}\chi}\ln\frac{1}{\varepsilon} \right)$.
Importantly, our communication bound depends on global constants $L_g,\mu_g$ whereas existing algorithms, even for deterministic setting, provide bounds which depend on local constants $L_l$, $\mu_l$ that can be much worse than $L_g$, $\mu_g$.
}
\pd{
Secondly, we propose the first accelerated distributed stochastic optimization algorithm over time-varying graphs. This algorithm has the same oracle per node complexity as the above algorithm and the communication complexity $\widetilde{O}\left(\frac{\tau}{\lambda}\sqrt{\frac{L_g}{\mu_g}}\ln\frac{1}{\varepsilon} \right)$, where $\tau$ and $\lambda$ characterize the dynamics of the communication graph (see the precise definition in Assumption \ref{assum:mixing_matrix}).
}
\section{Preliminaries}
\subsection{Problem reformulation}
In order to solve problem \eqref{eq:initial_problem} in a decentralized manner, we assign a local copy of $x$ to each node in the network, which leads to a linearly constrained problem
\begin{align}\label{eq:problem_inflated}
\min_{{\bf x}\in\mathbb{R}^{n\times d}}~ F({\bf x}) = \sum_{i=1}^m f_i(x_i)~ \text{s.t.}~ x_1 = \ldots = x_n,
\end{align}
where ${\bf x} = (x_1\ldots x_n)^\top\in\mathbb{R}^{n\times d}$. We denote the feasible set ${\mathcal{C}} = \braces{x_1 = \ldots = x_n}$ for brevity. Strong convexity and smoothness parameters of $F$ are related to that of functions $f_i$. Namely, $F$ is $L_{l}$-smooth and $\mu_{l}$-strongly convex on $\mathbb{R}^{n\times d}$ and $L_{g}$-smooth and $\mu_{g}$-strongly convex on the set ${\mathcal{C}}$.
\subsection{Consensus procedure}
In this subsection we discuss, how the agents can interact by exchanging information. Importantly, the communication graph $\mathcal{G}$ can change with time. Thus, we consider a sequence of undirected communication graphs $\braces{\mathcal{G}^k = (V, E^k)}_{k=0}^\infty$ and a sequence of corresponding mixing matrices $\braces{{\bf W}^k}_{k=0}^\infty$ associated with it. We impose the following
\begin{assumption}\label{assum:mixing_matrix}
Mixing matrix sequence $\braces{{\bf W}^k}_{k=0}^\infty$ satisfies the following properties.
\begin{itemize}
\item (Decentralized property) $(i, j)\notin E_k \;\Rightarrow \;[{\bf W}^k]_{ij} = 0$.
\item (Double stochasticity) ${\bf W}^k \mathbf{1}_n = \mathbf{1}_n,~ \mathbf{1}_n^\top{\bf W}^k = \mathbf{1}_n^\top$.
\item (Contraction property) There exist $\tau\in\mathbb{Z}_{++}$ and $\lambda\in(0, 1)$ such that for every $k\ge \tau - 1$ it holds
\begin{align*}
\norm{{\bf W}_{\tau}^k {\bf x} - \overline{\bf x}} \le (1 - \lambda)\norm{{\bf x} - \overline{\bf x}},
\end{align*}
where ${\bf W}_\tau^k = {\bf W}^k \ldots {\bf W}^{k-\tau+1}$.
\end{itemize}
\end{assumption}
The contraction property in Assumption \ref{assum:mixing_matrix} was initially proposed in \cite{koloskova2020unified} in a stochastic form. This property generalizes several assumptions in the literature.
\begin{itemize}
\item Time-static connected graphs. In this scenario we have ${\bf W}^k = {\bf W}$. Therefore, $\lambda = 1 - \lambda_2({\bf W})$, where $\lambda_2({\bf W})$ denotes the second largest eigenvalue of ${\bf W}$.
\item Sequence of connected graphs: every $\mathcal{G}_k$ is connected. In this case $\lambda = 1 - \underset{k\ge 0}{\sup}~\lambda_2({\bf W}^k)$.
\item $\tau$-connected graph sequence: for every $k\ge 0$ graph $\mathcal{G}^k_\tau = (V, E^k\cup E^{k+1}\cup\ldots\cup E^{k+\tau-1})$ is connected \cite{Nedic2017achieving}. For $\tau$-connected graph sequences it holds $1 - \lambda = \underset{k\ge 0}{\sup}~\lambda_{\max}({\bf W}_\tau^k - \frac{1}{n}\mathbf{1}_n\onevec^\top)$.
\end{itemize}
During every (synchronized) communication round, the agents pull information from their neighbors and update their local vectors according to the rule
\begin{align*}
x_i^{k+1} = [{\bf W}^k]_{ii} \rev{x_i^k} + \sum_{(i, j)\in E^k} [{\bf W}^k]_{ij} x_j^k,
\end{align*}
which writes as ${\bf x}^{k+1} = {\bf W}^k {\bf x}^k$ in matrix form. The contraction property in Assumption \ref{assum:mixing_matrix} requires a specific choice of weights in ${\bf W}^k$. Choosing Metropolis weighs is sufficient to ensure the contraction property for $\tau$-connected graph sequences (see \cite{Nedic2017achieving} for details):
\begin{align*}
[{\bf W}^k]_{ij} =
\begin{cases}
1 / (1 + \max\{d^k_i, d^k_j\}) &\text{if }(i, j)\in E^k, \\
0 &\text{if } (i, j)\notin E^k, \\
1 - \displaystyle\sum_{(i, m)\in E^k} [{\bf W}^k]_{im} &\text{if } i = j,
\end{cases}
\end{align*}
where $d^k_i$ denotes the degree of node $i$ in graph $\mathcal{G}^k$.
\pd{When the communication graph $\mathcal{G}$ does not change with time, it is possible to apply accelerated consensus procedures by leveraging Chebyshev acceleration~\cite{auzinger2011iterative,scaman2018optimal}: given the reference matrix ${\bf W}$ as above, set $W_T=P_T({\bf W})$ and $P_T(1)=1$ (the latter is to ensure the double stochasticity of $W_T$), with $T$ being the number of consensus steps and $P_T$ being the Chebyshev polynomial of degree $T$. In this case one can guarantee that
\begin{align*}
\norm{W_T {\bf x} - \overline{\bf x}} \le (1 - \sqrt{1-\rho})^T\norm{{\bf x} - \overline{\bf x}},
\end{align*}
where $\rho := \lambda_2({\bf W}) < 1$. In this case we define $\chi=\frac{1}{1-\rho}$.
}
\section{Algorithm and main result}
In this section we describe the proposed algorithm and give its convergence theorem. Our algorithms is an accelerated mini-batch stochastic gradient method equipped with a consensus procedure. Let $\braces{\xi_i^\ell}_{\ell=1}^r$ be independent random variables with distribution ${\mathcal{D}}_i$. For function ${\bf f}_i$ we define its batched gradient of size $r$ as
\begin{align*}
\nabla^r {\bf f}\cbraces{x, \braces{\xi_i^\ell}_{\ell=1}^r} = \frac{1}{r} \sum_{\ell=1}^r \nabla {\bf f}_i(x, \xi_\ell^r).
\end{align*}
Batched gradient for $F({\bf x})$ is defined analogously. Let $\braces{\xi^\ell}_{\ell=1}^r$ be independent, where $\xi^\ell = (\xi_1^\ell \ldots \xi_n^\ell)^\top$ is a random vector consisting of random variables at all nodes. Then we define $\nabla^r F({\bf x}, \braces{\xi}_{\ell=1}^r)$ as a matrix of $\mathbb{R}^{n\times d}$, the $i$-th row of which stores $\nabla^r {\bf f}_i\cbraces{x_i, \braces{\xi_i^\ell}_{\ell=1}^r}$.
For brevity we use notation $\nabla^r {\bf f}_i(x_i),~ \nabla^r F({\bf x})$ for batched gradients of ${\bf f}_i$ and $F$, respectively.
To describe the algorithm we introduce sequences of extrapolation coefficients $\alpha^k, A^k$ similar to that of \cite{stonyakin2020inexact}, which are defined as follows.
\begin{subequations}
\begin{align}\label{eq:Ak_def}
\alpha^0 &= A^0 = 0 \\
(A^k + \alpha^{k+1})(1 + A^k\mu_{g}/2) &= 2L_{g}(\alpha^{k+1})^2, \\
A^{k+1} &= A^k + \alpha^{k+1}.
\end{align}
\end{subequations}
\begin{algorithm}[H]
\caption{Decentralized Stochastic AGD}
\label{alg:decentralized_agd}
\begin{algorithmic}[1]
\REQUIRE{Initial guess ${\bf x}^0\in \mathcal{C}$, constants $L_{g}, \mu_{g} > 0$, \\${\bf u}^0 = {\bf x}^0$}
\FOR{$k = 0, 1, 2,\ldots$}
\STATE{\hspace{-0.25cm}${\bf y}^{k+1} = \frac{\alpha^{k+1} {\bf u}^k + A^k {\bf x}^k}{A^{k+1}}$}
\STATE{\vspace{0.05cm}\hspace{-0.25cm}\label{alg_step:agd_step}${\bf v}^{k+1} = \frac{(\alpha^{k+1}\mu_{g}/2){\bf y}^{k+1} + (1 + A^k\mu_{g}/2){\bf u}^k}{1 + A^{k + 1}\mu_{g}/2} - \frac{\alpha^{k+1}\nabla^r F({\bf y}^{k+1})}{1 + A^{k + 1}\mu_{g}/2}$}
\STATE{\label{alg_step:consensus_update}
$
\hspace{-0.25cm}
{\bf u}^{k+1} = \text{Consensus}({\bf v}^{k+1}, T^k)
$
}
\STATE{\hspace{-0.25cm}${\bf x}^{k+1} = \frac{\alpha^{k+1} {\bf u}^{k+1} + A^k {\bf x}^k}{A^{k+1}}$}
\ENDFOR
\end{algorithmic}
\end{algorithm}
\vspace{-0.5cm}
\begin{algorithm}[H]
\caption{Consensus}
\label{alg:consensus}
\begin{algorithmic}
\REQUIRE{Initial ${\bf x}^0\in\mathcal{C}$, number of iterations $T$.}
\FOR{$t = 1, \ldots, T$}
\STATE{${\bf x}^{t+1} = {\bf W}^t {\bf x}^t$}
\ENDFOR
\end{algorithmic}
\end{algorithm}
In the next theorem, we provide oracle and communication complexities of Algorithm \ref{alg:decentralized_agd}, i.e. we estimate the number of stochastic oracle calls by each node and the number of communication rounds to solve problem \eqref{eq:initial_problem} with accuracy $\varepsilon$.
\begin{theorem}[Main result]\label{th:total_iterations_strongly_convex}
\pd{Let $\varepsilon > 0$ be the desired accuracy. Set}
\begin{align*}
T_k = T = \frac{\tau}{2\lambda}\log\frac{D}{\delta'},~ \delta' = \frac{n\varepsilon}{32} \frac{\mu_{g}^{3/2}}{L_{g}^{1/2} L_{l}^2},~
r = \frac{2\sigma_{\pd{g}}^2}{\varepsilon\sqrt{L_{g}\mu_{g}}},
\end{align*}
where
\begin{align}\label{eq:def_sqrt_D}
&\sqrt D = \cbraces{\frac{2L_{l}}{\sqrt{L_{g}\mu_{g}}} + 1}\sqrt{\delta'} + \frac{2nM_\xi}{\sqrt{L_{g}\mu_{g}}} \\
&+\frac{2L_{l}}{\mu_{g}} \sqrt{n} \cbraces{\norm{\overline u^0 - x^*}^2 + \frac{2}{\sqrt{L_{g}\mu_{g}}}\cbraces{\frac{\sigma_{g}^2}{4nL_{g} r^2} + \delta}}^{1/2} \notag.
\end{align}
\pd{Then, to yield ${\bf x}^N$ such that
\begin{align*}
&\mathbb{E} f(\overline x^N) - f(x^*)\le \varepsilon,~ \mathbb{E}\norm{{\bf x}^N - \overline{\bf x}^N}^2\le \delta'=O(\varepsilon),
\end{align*}
Algorithm \ref{alg:decentralized_agd} requires \pd{no more than}
\begin{align}\label{eq:agd_computational_complexity}
N_{\pd{orcl}} = N\cdot r = \dfrac{6\sigma_{g}^2}{n\mu_{g}\varepsilon} \log\cbraces{\frac{4L_{g}\norm{\overline u^0 - x^*}^2}{\varepsilon}}
\end{align}
stochastic oracle calls at each node and \pd{no more than}
\begin{align}\label{eq:agd_communication_complexity}
N_{\text{comm}}
&= 3\sqrt{\frac{L_{g}}{\mu_{g}}} \kappa \cdot \log\cbraces{\frac{4L_{g}\norm{\overline u^0 - x^*}^2}{\varepsilon}} \log\frac{D}{\delta'},
\end{align}
\rev{communication rounds}, where $\kappa= \frac{\tau}{2\lambda} $ under Assumption \ref{assum:mixing_matrix} and $\kappa=\sqrt{\chi}$
when the communication graph is static.}
\end{theorem}
We provide the proof of Theorem \ref{th:total_iterations_strongly_convex} in Section \ref{sec:analysis}.
\pd{The number of stochastic oracle calls at each node in \eqref{eq:agd_computational_complexity} coincides with the lower bound for centralized optimization up to a constant factor. When the graph is time-varying, the number of communication steps includes an additional factor $\tau/\lambda$, which characterizes graph connectivity. If the communication graph is fixed, \rev{in addition to the lower oracle complexity bound, our algorithm also achieves lower communication bound up to a polylogarithmic factor.}}
\section{Analysis of the algorithm}\label{sec:analysis}
\pd{Analysis of our algorithm consists of three main parts. Firstly, if an approximate consensus is imposed on local variables at each node, this ensures a stochastic inexact oracle for the global objective $f$. Secondly, we analyze an accelerated stochastic gradient method with stochastic inexact oracle. Thirdly, we analyze, how the consensus procedure allows to obtain an approximate consensus. Finally, we combine the building blocks together and prove the main result.}
\subsection{Stochastic inexact oracle via inexact consensus}
In this subsection we show that if a point ${\bf x}\in\mathbb{R}^{n\times d}$ is close to the set $\mathcal{C}$, i.e. it approximately satisfies consensus constraints, then, the mini-batched and averaged among nodes stochastic gradient provides a stochastic inexact oracle developed in \pd{\cite{devolder2013first,dvurechensky2016stochastic,gasnikov2016stochasticInter}}.
Consider $\overline x, \overline y\in\mathbb{R}^{d}$ and define $\overline{\bf x} = \mathbf{1}_n\overline x^\top = (\overline x\ldots \overline x)^\top,~ \overline{\bf y} = \mathbf{1}_n\overline y^\top = (\overline y\ldots \overline y)^\top\in \mathbb{R}^{n\times d}$. Let ${\bf x}\in\mathbb{R}^{n\times d}$ be such that $\Pi_\mathcal{C} ({\bf x}) = \overline{\bf x}$ and $\norm{\overline{\bf x} - {\bf x}}^2\le \delta'$.
\begin{lemma}\label{lem:inexact_oracle}
Define
\begin{align*}
\delta &= \frac{1}{2n}\cbraces{\frac{L_{l}^2}{L_{g}} + \frac{2L_{l}^2}{\mu_{g}} + L_{l} - \mu_{l}} \delta', \addtocounter{equation}{1}\tag{\theequation}\label{eq:delta_inexact_oracle} \\
f_{\delta, L, \mu}(\overline x, {\bf x}) &= \frac{1}{n} \sbraces{F({\bf x}) + \angles{{\nabla} F({\bf x}), \overline{\bf x} - {\bf x}}} \\
&\quad + \cbraces{\frac{\mu_{l}}{2n} - \frac{2L_{l}^2}{2n\mu_{g}}}\norm{\overline{\bf x} - {\bf x}}^2, \\
g_{\delta, L, \mu}(\overline x, {\bf x}) &= \frac{1}{n}\sum_{i=1}^n \nabla f_i(x_i)\\
\widetilde{g}_{\delta, L, \mu}(\overline x, {\bf x}) &= \frac{1}{n}\sum_{i=1}^n \frac{1}{r}\sum_{j=1}^r \nabla{\bf f}_i(x_i,\xi_i^j).
\end{align*}
Firstly, for any $\overline y\in\mathbb{R}^d$ it holds
\begin{align*
\hspace{-0.25cm}\frac{\mu_{g}}{4}\norm{\overline y - \overline x}^2 &\le f(\overline y) - f_{\delta, L, \mu}(\overline x, {\bf x}) - \angles{ g_{\delta, L, \mu}(\overline x, {\bf x}), \overline y - \overline x} \\
\hspace{-0.25cm}L_{g}\norm{\overline y - \overline x}^2 + \delta &\ge f(\overline y) - f_{\delta, L, \mu}(\overline x, {\bf x}) - \angles{ g_{\delta, L, \mu}(\overline x, {\bf x}), \overline y - \overline x}.
\end{align*}
Secondly, $\widetilde{g}_{\delta, L, \mu}(\overline x, {\bf x})$ satisfies
\begin{subequations}\label{eq:tilde_g_properties}
\begin{align}
\mathbb{E} \widetilde{g}_{\delta, L, \mu}(x) &= g_{\delta, L, \mu}(x) \label{eq:tilde_g_properties_1} \\
\mathbb{E} \|\widetilde{g}_{\delta, L, \mu}(\overline x, {\bf x}) - g_{\delta, L, \mu}(\overline x, {\bf x})\|^2 &\leq \frac{\sum_{i=1}^n\sigma^2_i}{n^2r}\pd{=
\frac{\sigma_g^2}{nr}}.
\label{eq:tilde_g_properties_2}
\end{align}
\end{subequations}
\end{lemma}
\vspace{0.1cm}
The first statement is proved in Lemma 2.1 of \cite{rogozin2020towards}; the proof of the second statement is provided in Appendix \ref{app:inexact_oracle}.
\subsection{Similar Triangles Method with Stochastic Inexact Oracle}
In this subsection we present a general algorithm for minimization problems with stochastic inexact oracle. This subsection is independent from the others and generalizes the algorithm and analysis from \cite{stonyakin2020inexact,stonyakin2019gradient} to the stochastic setting.
Let $f(x)$ be a convex function defined on a convex set $Q\subseteq\mathbb{R}^m$. We assume that $f$ is equipped with stochastic inexact oracle having two components. The first component $(f_{\delta,L,\mu}(x), g_{\delta,L,\mu}(x))$ exists at any point $x\in Q$ and satisfies
\begin{align}\label{eq:inexact_oracle_def_devolder}
\frac{\mu}{2}\norm{y - x}^2
&\le f(y) - \cbraces{f_{\delta,L,\mu}(x) + \angles{g_{\delta,L,\mu}(x), y - x}} \notag\\
&\le \frac{L}{2} \norm{y - x}^2 + \delta
\end{align}
for all $y\in Q$. To allow more flexibility, we assume that $\delta$ may change with the iterations of the algorithm.
The second component $\widetilde{g}_{\delta,L,\mu}(x)$ is stochastic, is available at any point $x\in Q$, and satisfies
\begin{align}
\label{eq:inexact_oracle_moments}
\mathbb{E} \widetilde{g}_{\delta, L, \mu}(x) = g_{\delta, L, \mu}(x), \quad \mathbb{E} \|\widetilde{g}_{\delta, L, \mu}(x) - g_{\delta, L, \mu}(x)\|^2 \leq \sigma^2.
\end{align}
We also denote the \textit{batched} version of the stochastic component as
\begin{align}
\label{eq:inexact_oracle_batched_def}
\widetilde{g}^{\ r}_{\delta, L, \mu}(x) = \frac{1}{r} \sum_{i=1}^r \widetilde{g}_{\delta, L, \mu}(x, \xi_i),
\end{align}
where $\xi_i$'s are iid realizations of the random variable $\xi$. It is straightforward that
\begin{subequations}\label{eq:inexact_oracle_batched_prop}
\begin{align}
&\mathbb{E} \widetilde{g}^{\ r}_{\delta, L, \mu}(x) = g_{\delta, L, \mu}(x), \\ &\mathbb{E} \|\widetilde{g}^{\ r}_{\delta, L, \mu}(x) - g_{\delta, L, \mu}(x)\|^2 \leq \frac{\sigma^2}{r}.
\end{align}
\end{subequations}
Let us consider the following algorithm for minimizing $f$. Note that the error $\delta$ of the oracle and batch size $r$ may depend on the iteration counter $k$. Moreover, we let $\delta$ be stochastic.
\begin{algorithm}[ht]
\caption{AGD with stochastic inexact oracle}
\label{alg:FastAlg2_strong}
\begin{algorithmic}[1]
\REQUIRE{Initial guess $x^0$, constants $L, \mu \geq 0$, sequence of batch sizes $\{r_k\}_{k\geq 0}$.}
\par Set $y^0:=x^0, u^0:=x^0, \alpha^0:=0, A^0:=0$
\FOR{$k \geq 0$}
\STATE{Find $\alpha^{k+1}$ as the greater root of:
$$(A^k + \alpha^{k+1})(1 + A^k \mu) = L \left( \alpha^{k+1} \right)^2$$ }
\STATE{\vspace{-0.5cm} Renew the following variables:
\begin{gather*}
A^{k+1}:=A^k + \alpha^{k+1} \\
y^{k+1}:=\frac{\alpha^{k+1}u^k + A^k x^k}{A^{k+1}}
\end{gather*}
}
\STATE{\vspace{-0.5cm}
Define the function:
\begin{align*}
&\phi^{k+1}(x) := \alpha^{k+1} \angles{\widetilde{g}^{\ r_{k+1}}_{\delta_{k+1}, L, \mu}(y^{k+1}),x - y^{k+1}} \\ &\quad + (1 + A^k \mu) \norm{x - u^k}^2 + \alpha^{k+1} \norm{x - y^{k+1}}^2
\end{align*}
}
\STATE{\vspace{-0.5cm} Solve the optimization problem:
$$u^{k+1} := \argmin_{x \in Q} \phi^{k+1}(x)$$}
\STATE{\vspace{-0.5cm} Update $x$:
$$x^{k+1}:=\frac{\alpha^{k+1} u^{k+1} + A^k x^k}{A^{k+1}}$$
}
\ENDFOR
\end{algorithmic}
\end{algorithm}
We analyze convergence of Algorithm \ref{alg:FastAlg2_strong} by revisiting the proof of Theorem 3.4 in \cite{stonyakin2020inexact} and formulate the result in Theorem \ref{Th:fast_str_conv_adap} below. The complete proof is provided in Appendix \ref{app:fast_str_conv_adap}.
\begin{theorem}
\label{Th:fast_str_conv_adap}
Let Algorithm \ref{alg:FastAlg2_strong} be applied to solve the problem $\min_{x\in Q} f(x)$. \pd{Let also $\norm{u^0 - x^*}\leq R$.} Then, after $N$ iterations we have
{\small
\begin{align}
&\mathbb{E} f(x^N) - f(x^*) \leq \frac{1}{A^N} \left( \pd{R^2} + \sum_{i=1}^N A^i \left( \frac{\sigma^2}{2Lr_{i}} + \mathbb{E}\delta_i \right) \right)\\
&\mathbb{E} \norm{u^N - x^*}^2 \leq \frac{1}{1 + A^N \mu} \left(\pd{R^2} + \sum_{i=1}^N A^i \left( \frac{\sigma^2}{2Lr_{i}} + \mathbb{E}\delta_i \right) \right)
\end{align}
}
\end{theorem}
In order to establish the rate, we recall the results of Lemma 5 in \cite{devolder2013first} and Lemma 3.7 in \cite{stonyakin2020inexact} and estimate the growth of coefficients $A^N$.
\begin{lemma}
\label{lem:Ak_properties}
Coefficient $A^N$ can be lower-bounded as following: \rev{$A^N \geq 1/L\cdot\left( 1 + (1/2)\sqrt{{\mu}/L} \right)^{2(N-1)}$}.
Moreover, we have \rev{$\sum_{i=1}^N A^i / A^N \le {1 + \sqrt{{L}/{\mu}}}$}.
\end{lemma}
\subsection{Proof of the main result}
Thoughout this section, we denote $L = 2L_{g},~ \mu = \mu_{g}/2$ and $\sigma^2 = \sigma_{g}^2/(nr)$.
\subsubsection{Outer loop}
\begin{lemma}\label{lem:inexact_agd_convergence}
Provided that consensus accuracy is $\delta'$, i.e. $\mathbb{E}\norm{{\bf u}^{j} - \overline{\bf u}^j}^2\le \delta' \text{ for } j = 1, \ldots, k$, we have
{\small
\begin{align}
&\mathbb{E} f(\overline x^k) - f(x^*) \le \frac{1}{A^k}\cbraces{\norm{\overline u^0 - x^*}^2 + \cbraces{\frac{\sigma^2}{2Lr} + \delta} \sum_{i=1}^k A^i} \label{eq:inexact_agd_function_residual} \\
&\mathbb{E}\norm{\overline u^k - x^*}^2 \le \frac{1}{1 + A^k\mu} \cbraces{\norm{\overline u^0 - x^*}^2 + \cbraces{\frac{\sigma^2}{2Lr} + \delta} \sum_{i=1}^k A^i} \nonumber
\end{align}
}
where $\delta$ is given in \eqref{eq:delta_inexact_oracle}.
\end{lemma}
\begin{proof}
First, assuming that $\mathbb{E}\norm{{\bf u}^{j} - \overline{\bf u}^j}^2\le \delta'$, we show that ${\bf y}^j, {\bf u}^j, {\bf x}^j$ lie in $\sqrt{\delta'}$-neighborhood of $\mathcal{C}$ by induction. At $j=0$, we have $\norm{{\bf x}^0 - \overline{\bf x}^0} = \norm{{\bf u}^0 - \overline{\bf u}^0} = 0$. Using $A^{j+1} = A^j + \alpha^j$, we get an induction pass $j\to j+1$.
\begin{align*}
&\mathbb{E}\norm{{\bf y}^{j+1} - \overline{\bf y}^{j+1}} \\
&\qquad \le \frac{\alpha^{j+1}}{A^{j+1}}\mathbb{E}\norm{{\bf u}^j - \overline{\bf u}^j} + \frac{A^j}{A^{j+1}} \mathbb{E}\norm{{\bf x}^j - \overline{\bf x}^j}\le \sqrt{\delta'}, \\
&\mathbb{E}\norm{{\bf x}^{j+1} - \overline{\bf x}^{j+1}} \\
&\qquad \le \frac{\alpha^{j+1}}{A^{j+1}}\mathbb{E}\norm{{\bf u}^{j+1} - \overline{\bf u}^{j+1}} + \frac{A^j}{A^{j+1}} \mathbb{E}\norm{{\bf x}^j - \overline{\bf x}^j}\le \sqrt{\delta'}.
\end{align*}
Therefore, $g(\overline y) = \frac{1}{n}\sum_{i=1}^n \nabla f(y_i)$ represents the inexact gradient of $f$, and the desired result directly follows from Theorem \ref{Th:fast_str_conv_adap}.
\end{proof}
\subsubsection{Consensus subroutine iterations}
In order to establish communication complexity of Algorithm \ref{alg:FastAlg2_strong}, we estimate the number of consensus iterations in the following Lemma.
\begin{lemma}\label{lem:consensus_iters_strongly_convex}
Let consensus accuracy be maintained at level $\delta'$, i.e. $\mathbb{E}\norm{{\bf u}^j - \overline{\bf u}^j}^2\le \delta' \text{ for } j = 1, \ldots, k$ and let Assumption \ref{assum:mixing_matrix} hold. Then it is sufficient to make $T_k = T = \frac{\tau}{2\lambda}\log\frac{D}{\delta'}$ consensus iterations, where $D$ is defined in \eqref{eq:def_sqrt_D}, in order to ensure $\delta'$-accuracy on step $k+1$, i.e. $\mathbb{E}\norm{{\bf u}^{k+1} - \overline{\bf u}^{k+1}}^2\le \delta'$.
\end{lemma}
Lemma \ref{lem:consensus_iters_strongly_convex} is analogous to Lemma A.3 in \cite{rogozin2020towards} and is proven in Appendix \ref{app:consensus_iters_strongly_convex}.
\pd{In the same way we can prove that if the communication network is static, we can establish a sufficiently accurate consensus in the next iteration.
\begin{lemma}\label{lem:consensus_iters_strongly_convex_fixed}
Let consensus accuracy be maintained at level $\delta'$, i.e. $\mathbb{E}\norm{{\bf u}^j - \overline{\bf u}^j}^2\le \delta' \text{ for } j = 1, \ldots, k$ and let the communication network be static. Then it is sufficient to make $T_k = T = \sqrt{\chi}\log\frac{D}{\delta'}$ consensus iterations, where $D$ is defined in \eqref{eq:def_sqrt_D}, in order to ensure $\delta'$-accuracy on step $k+1$, i.e. $\mathbb{E}\norm{{\bf u}^{k+1} - \overline{\bf u}^{k+1}}^2\le \delta'$.
\end{lemma}
}
\subsubsection{Putting the proof together}
\rev{We derive the expressions for $r$ and $\delta$ to meet the requirement $\mathbb{E} f(\overline x^k) - f(x^*) \le \varepsilon$ according to \eqref{eq:inexact_agd_function_residual}. This is done by combining the results of Lemmas \ref{lem:Ak_properties}, \ref{lem:consensus_iters_strongly_convex}, \ref{lem:consensus_iters_strongly_convex_fixed}.} The details are given in Appendix \ref{app:total_iterations_strongly_convex}.
\section{Numerical tests}
We run Algorithm \ref{alg:decentralized_agd} on L2-regularized logistic regression problem:
\begin{align*}
f(x) = \frac{1}{m}\sum\limits_{i=1}^m \log\left(1 + \exp(-b_i\angles{a_i, x})\right) + \frac{\theta}{2}\norm{x}^2.
\end{align*}
Here $a_1,\ldots,a_m\in\mathbb{R}^d$ are entries of the dataset, $b_1,\ldots,b_m\in\{-1, 1\}$ denote class labels and $\theta > 0$ is a penalty coefficient. Data points $(a_i, b_i)$ are distributed among the computational nodes in the network.
We use LIBSVM datasets \cite{Chang2011} to run our experiments. Work of Algorithm \ref{alg:decentralized_agd} is simulated on a9a data-set with different settings for batch-size $r$ and number of consensus iterations $T$. The random geometric graph has $20$ nodes. \rev{We compare the performance of Algorithm 1 with DSGD \cite{fallah2019robust,olshevsky2019asymptotic,olshevsky2019non}}.
\begin{figure}[H]
\includegraphics[width=0.5\textwidth]{figures/r=10.png}
\caption{Random geometric graph with 20 nodes; batch size $r = 10$}
\end{figure}
We observe a tradeoff between consensus accuracy and convergence speed in function value. A large number of consensus steps results in more accurate consensus and slower convergence, and vice versa. This tradeoff is present for different batch sizes.
\section{Conclusion}
\pd{We propose an accelerated distributed optimization algorithm for stochastic optimization problems in two settings: time-varying graphs and static graphs. For the latter setting we achieve the full acceleration and our method achieves lower bounds both for the communication and oracle per node complexity. }
\pd{
Our approach is based on accelerated gradient method with stochastic inexact oracle which makes it
generic with many possible extensions. In particular, we focus on a specific case of strongly convex smooth functions, but the possible extensions include non-strongly convex and/or non-smooth functions that can be covered by such inexact oracles \cite{devolder2014first,kamzolov2020universal}. Further, we believe that our results can be extended for composite optimization problems, zeroth-order optimization methods \cite{gorbunov2018accelerated,vorontsova2019accelerated,dvurechensky2021accelerated}, and distributed algorithms for saddle-point problems \cite{gasnikov2021accelerated} and variational inequalities.}
\bibliographystyle{IEEEtran.bst}
|
\section{introduction}
In this paper, we consider closed smooth manifolds admitting Riemannian metrics
as the criticals of a curvature functional.
Let $M$ be a closed $n$-dimensional smooth manifold and ${\mathcal M}_1$ be the set of all smooth Riemannian metrics of a unit volume
on $M$. The total scalar curvature functional ${\mathcal S}$ on ${\mathcal M}_1$ is given by
$$
{\mathcal S}(g) = \int_M s_g\, dv_g.
$$
where $s_g$ is the scalar curvature of the metric $g \in {\mathcal M}_1$ and $dv_g$ denotes the volume form of $g$.
Critical points of ${\mathcal S}$ on ${\mathcal M}_1$ are known to be Einstein (cf. \cite{Be}).
Introducing a subset ${\mathcal C}$ of ${\mathcal M}_1$ consisting of metrics with a constant scalar curvature,
the Euler-Lagrange equation of ${\mathcal S}$ restricted to ${\mathcal C}$ can be written in the following
\begin{eqnarray*}
z_g=s_g'^*(f),
\end{eqnarray*}
which is called the critical point equation (CPE in short).
Here, $z_g$ is the traceless Ricci tensor defined by $z_g:= {\rm Ric}_g - \frac{s_g}{n}g$,
${\rm Ric}_g$ is the Ricci tensor, and $s_g'^*(f)$ is given by
$$
s_g'^*(f)=D_gdf -(\Delta_g f)g -f {\rm Ric}_g,
$$
where $D_gdf$ and $\Delta_g f$ are the Hessian and (negative) Laplacian of $f$, respectively.
From the variational problem for curvature functional, the {\it Besse conjecture} \cite{Be} describes that a non-trivial solution of the CPE
\begin{eqnarray}
z_g = D_gdf -(\Delta_g f)g -f {\rm Ric}_g, \label{cpe}
\end{eqnarray}
should be Einstein. By taking the trace of (\ref{cpe}), we have
\begin{eqnarray}
\Delta _g f=-\frac{s_g}{n-1}f \label{cpet}
\end{eqnarray}
and therefore, using $z_g = {\rm Ric}_g - \frac{s_g}{n}g$, (\ref{cpe}) can be rewritten as
\begin{eqnarray}
(1+f)z_g = Ddf + \frac{s_gf}{n(n-1)}g.\label{eqn2020-5-7-2}
\end{eqnarray}
Since $g$ is clearly Einstein when $f=0$, going forward, we assume that $f$ is not trivial.
Besse conjecture describes that such a nontrivial solution $(g,f)$ of (\ref{cpe}) is Einstein.
Notably, some progress has been made to this conjecture. For example,
if the Riemannian manifold is locally conformally flat,
then the metric is Einstein \cite{laf}. It is also known that
Besse conjecture holds if the manifold has a harmonic curvature \cite{ych}, \cite{erra},
or the metric is Bach-flat \cite{qy}.
We say that $(M, g)$ has a harmonic curvature if the divergence of the Riemannian curvature tensor $R$ vanishes, in other words,
$\d R = 0$, and $(M, g)$ is said to be Bach-flat
when the Bach tensor vanishes. Very recently, it was shown \cite{bal} that the Besse conjecture holds if the complete divergence
of the Weyl curvature tensor $\mathcal W$ is free, ${\rm div}^4 {\mathcal W} = 0$ and the radially Weyl curvature vanishes.
If $\min_M f \ge -1$, it is clear
that $(M, g)$ is Einstein. In fact, if we let $i_{\nabla f}z_g =z_g(\nabla f, \cdot)$,
the divergence of $i_{\nabla f}z_g$ can be computed as ${\rm div}\left(i_{\nabla f}z_g\right) = (1+f)|z_g|^2$.
Hence, by integrating it over $M$, we have $z_g=0$ from the divergence theorem.
As a $1$-form or a vector field, the quantity $i_{\nabla f}z_g$ has a crucial structural meaning in CPE. In \cite{ych}, we show that
$i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z_g$ is parallel to $\nabla} \def\e{\epsilon} \def\la{\langle f$ when $(M, g)$ has a harmonic curvature, and this property plays an important role in proving the main theorem.
First, we prove the Besse conjecture when $z_g$ is vanishing in the direction $\nabla} \def\e{\epsilon} \def\la{\langle f$, which is, in fact, a generalization of the main result in \cite{ych}.
\begin{thm} \label{thm2020-8-22-3}
Let $(g,f)$ be a nontrivial solution of (\ref{cpe}). If $z(\nabla f, X)=0$ for any vector field
$X$ orthogonal to $\nabla f$, then $(M,g)$ is isometric to a standard sphere.
\end{thm}
The second objective of this paper is to prove the Besse conjecture under the condition of positive isotropic curvature on $M$.
Let $(M^n, g)$ be an $n$-dimensional Riemannian manifold with $n \ge 4$.
The Riemannian metric $g = \langle \,\, ,\, \rangle$ can be extended either
to a {\it complex bilinear form} $(\,\, , \, )$ or a {\it Hermitian inner product}
$\llangle\,\,, \, \rrangle$ on each complexified tangent space $T_pM \otimes {\Bbb C}$ for $p \in M$.
A complex $2$-plane $\sigma \subset T_pM\otimes {\Bbb C}$ is {\it totally isotropic}
if $ (Z, Z) = 0$ for any $Z \in \sigma$. For any $2$-plane $\sigma \subset T_pM\otimes {\Bbb C}$, we can define the
complex sectional curvature of $\sigma$ with respect to $\llangle \, , \, \rrangle$ by
\begin{eqnarray}
{\rm K}_{\Bbb C}(\sigma) = \llangle {\mathcal R}(Z \wedge W), Z\wedge W\rrangle,
\label{eqn2019-5-17-1}
\end{eqnarray}
where $\mathcal R : \Lambda^2 T_pM \to \Lambda^2 T_pM$ is the curvature operator and
$\{Z, W\}$ is a unitary basis for $\sigma$ with respect to $\llangle \, , \, \rrangle$.
A Riemannian $n$-manifold $(M^n, g)$ is said to have a
{\it positive isotropic curvature} {\rm(PIC in short)} if
the complex sectional curvature on the isotropic planes is positive, that is,
for any totally isotropic $2$-plane $ \sigma \subset T_pM\otimes {\Bbb C}$,
\begin{eqnarray}
{\rm K}_{\Bbb C}(\sigma) >0. \label{eqn2018-4-22-1}
\end{eqnarray}
If $(M, g)$ has a positive curvature operator, then, it has a PIC \cite{mm88}.
Thus, a standard sphere $({\Bbb S}^n, g_0)$ has a PIC.
Additionally, if the sectional curvature of $(M, g)$ is pointwise strictly quarter-pinched, then
$(M, g)$ has a PIC \cite{mm88}.
It is well-known that the product metric on ${\Bbb S}^{n-1}\times {\Bbb S}^1$ also has
a PIC and the connected sum of manifolds with a PIC admits a PIC metric \cite{m-w}.
The existence of Riemannian metrics with a PIC on compact manifolds which fiber
over the circle is discussed in \cite{lab}. One of the main results on manifolds with a positive isotropic
curvature is that, a simply connected compact $n$-dimensional Riemannian manifold with positive
isotropic curvature is homeomorphic to a sphere \cite{mm88}.
In another major result relating the topology of a positive isotropic
curvature manifold, it is proved \cite{mm88} that the homotopy groups $\pi_i(M) = 0$ for $2 \le i \le [\frac{n}{2}]$ when $\dim(M) = n$, and the fundamental group $\pi_1(M)$ cannot contain any subgroup
isomorphic to the fundamental group of a closed orientable surface \cite{fra}, \cite{f-w}.
For even-dimensional manifolds, it was proved \cite{m-w}, \cite{sea} that a PIC implies the vanishing of the second Betti number.
On the contrary, a PIC implies that $g$ has a positive scalar curvature \cite{m-w}.
More details about the PIC are provided in \cite{c-h} or \cite{ses} and the references are also presented therein.
Our second main result can be stated as follows..
\begin{thm} \label{thm1}
Let $M$ be a compact $n$-dimensional smooth manifold with $n\geq 4$.
If $(g,f)$ is a nontrivial solution of (\ref{cpe}) and $(M, g)$ has positive isotropic curvature,
then $(M,g)$ is isometric to a standard sphere.
\end{thm}
Due to \cite{ych} and \cite{erra}, it suffices to prove that $(M,g)$ has harmonic curvature in Theorem~\ref{thm1}.
Then, $g$ should be Einstein, which implies that $(M, g)$ is isometric to a standard sphere due to Obata \cite{Ob}.
To show that $(M, g)$ has a harmonic curvature, we introduce a $2$-form, $df\wedge i_{\nabla f}z_g$, consisting of
the total differential, $df$, of the potential function $f$ and the traceless Ricci tensor
$z_g$ and prove that it vanishes when $(M, g)$ has positive isotropic curvature.
\vskip .5pc
\noindent
{\bf Notations: } Hereafter, for convenience and simplicity, we denote curvatures
${\rm Ric}_g, z_g, s_g$, and the Hessian and Laplacian of $f$, $D_gdf, \Delta_g$ by
$r, z, s$, and $Ddf, \Delta$, respectively, if there is no ambiguity.
We also use the notation $\langle \,\, ,\,\, \rangle$ for metric $g$ or inner product induced by $g$ on tensor spaces.
\section{Some Preliminaries and tensors}
As a preliminary as well as for subsequent use, we recall the Cotton tensor and briefly describe the properties of
the Cotton tensor. Additionally, we introduce a structural tensor $T$, which plays a key role in
proving our main theorems (For the definition of $T$, refer (\ref{defnt}) ).
This structural tensor $T$ has deep relations to the critical point equation (\ref{cpe})
and the Cotton tensor (cf. \cite{h-y}).
\vskip .5pc
\subsection{Cotton tensor}
Let $(M^n, g)$ be a Riemannian manifold of dimension $n$ with the Levi-Civita
connection $D$. We begin with a differential operator acting on the space
of symmetric $2$-tensors. Let $b$ be a symmetric $2$-tensor on $M$. The
differential $d^Db$ can be defined as follows:
$$
d^Db(X, Y, Z) = D_Xb(Y, Z) - D_Yb(X, Z)
$$
for any vectors $X, Y$ and $Z$. The {\it Cotton tensor}
$C \in\Gamma(\Lambda^2 M \otimes T^*M)$ is defined by
\begin{eqnarray}
C = d^D \left(r - \frac{s}{2(n-1)} g\right)
= d^Dr - \frac{1}{2(n-1)} ds \wedge g. \label{eqn2017-4-1-1}
\end{eqnarray}
It is known that, for $n=3$,
$C=0$ if and only if $(M^3, g)$ is locally conformally flat.
Moreover, for $n \ge 4$, the vanishing of the Weyl
tensor $\mathcal W$ implies the vanishing of the Cotton tensor $C$, while
$C=0$ corresponds to the Weyl tensor being harmonic, i.e, $\d \mathcal W
= 0$ due to the identity \cite{Be}:
\begin{eqnarray}
\d \mathcal W = - \frac{n-3}{n-2}d^D \left(r - \frac{s}{2(n-1)} g\right)
= - \frac{n-3}{n-2} C\label{eqn2016-12-3-16}
\end{eqnarray}
under the following identification
$$
\Gamma(T^*M\otimes \Lambda^2M) \equiv \Gamma(\Lambda^2M \otimes T^*M).
$$
Let $\{E_i\}$ be a local frame with $C(E_i, E_j, E_k) = C_{ijk}$. Then, it follows from (\ref{eqn2017-4-1-1}) that
$$
C_{ijk} = R_{jk;i}-R_{ik;j} - \frac{1}{2(n-1)} S_i \d_{jk} + \frac{1}{2(n-1)} S_j \d_{ik},
$$
where $r(E_i, E_j) = R_{ij}, \, R_{ij;k} = D_{E_k}r(E_i, E_j)$ and $ds(E_i) = S_i.$
The first direct observation on $C$ is that the cyclic summation of $C_{ijk}$ is
vanishing. The second observation on $C$ is that $C_{ijk}$ is
skew symmetric for the first two components, which implies that
$C(X, X, \cdot) = 0$ for any vector $X$, and trace-free in any two indices.
Since $\displaystyle{r = z + \frac{s}{n}g}$, equation (\ref{eqn2017-4-1-1}) can be
rewritten as
\begin{eqnarray*}
C = d^D z + \frac{n-2}{2n(n-1)}ds \wedge g.\label{eqn2017-4-2-1}
\end{eqnarray*}
In the local coordinates, we have
$$
C_{ijk} = z_{jk;i} - z_{ik;j} + \frac{n-2}{2n(n-1)} (S_i \d_{jk} - S_j \d_{ik}).
$$
In the CPE, since the scalar curvature $s$ is constant, we have
$$
C = d^D r = d^D z
$$
and
$$
C_{ijk}z_{jk;i} = |C|^2 + C_{ijk}z_{ik;j} = |C|^2 - C_{jik}z_{ik;j}.
$$
Here, we follow the Einstein convention for indices.
i.e.,
\begin{eqnarray*}
\frac{1}{2}|C|^2= C_{ijk}z_{jk;i} = C(E_i, E_j, E_k)D_{E_i}z(E_j, E_k).\label{eqn2017-4-2-2}
\end{eqnarray*}
Moreover,
\begin{eqnarray}
\frac{1}{2}|C|^2
&=&
C_{ijk}z_{jk;i} = (C_{ijk}z_{jk})_{;i} - C_{ijk;i}z_{jk} \nonumber\\
&=&
(C_{ijk}z_{jk})_{;i} + \langle \d C, z\rangle,\label{eqn2018-1-23-1}
\end{eqnarray}
where $\d C (X, Y) = -{\rm div}C(X, Y) = -D_{E_i} C(E_i, X, Y)$.
\vspace{.13in}
\subsection{ The tensor $T$}
We now define a $3$-tensor $T$ by
\begin{eqnarray}
T= \frac 1{n-2}\, df \wedge z +\frac 1{(n-1)(n-2)}\, i_{\nabla f }z \wedge g. \label{defnt}
\end{eqnarray}
Recall that $i_{\nabla} \def\e{\epsilon} \def\la{\langle f}$ denotes the (usual) interior product to the first factor defined by
$i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z(X) = z(\nabla} \def\e{\epsilon} \def\la{\langle f, X)$ for any vector $X$.
As the Cotton tensor $C$, the cyclic summation of $T_{ijk}$ vanishes.
It is also easy to see that the trace of $T$ in any two summands vanishes. In fact, since $T(X, Y, Z) = - T(Y, X, Z)$, we have,
${\rm tr}_{12} T = 0$, and from the definition of $T$ together with ${\rm tr}(z) = 0$, we can show ${\rm tr}_{13}T = 0$ directly.
Finally, since the cyclic summation of $T$ vanishes, we have
${\rm tr}_{23} T = -{\rm tr}_{12}T - {\rm tr}_{13}T = 0$.
The tensor $T$ looks similar to the tensor in \cite{CC} (cf. \cite{k-o})
where the authors used this in
classifying the complete Bach flat gradient shrinking Ricci solitons.
The first relation of the Cotton tensor $C$ to the tensor $T$ is the following
for the CPE.
\begin{lem}\label{lem2019-5-23-10}
Let $(M^n, g, f)$ be a non-trivial solution of the CPE. Then
\begin{eqnarray*}
(1+f) C = {\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \mathcal W - (n-1)T.\label{eqn2017-6-12-10-1}
\end{eqnarray*}
Here $\tilde{i}_{X}$ is the interior product to the last factor defined by
$\tilde{i}_{X}{\mathcal W} (Y, Z, U)= {\mathcal W}(Y, Z, U, X)$ for any vectors
$ Y, Z$, and $U$.
\end{lem}
\begin{proof}
Taking the differential $d^D$ in (\ref{cpe}), we have
\begin{eqnarray*}
(1+f)d^D z = {\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f}\mathcal W - \frac{n-1}{n-2}df\wedge z -
\frac{1}{n-2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z \wedge g.\label{eqn2016-12-3-10-1}
\end{eqnarray*}
(For more details about this, refer \cite{ych}).
From the definition of $T$ together with $C = d^Dz$, we obtain
\begin{eqnarray*}
(1+f) C = {\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \mathcal W - (n-1)T.
\end{eqnarray*}
\end{proof}
\vspace{.13in}
Now, we derive certain expressions relating $T$ and $C$ to their divergences.
For a symmetric $2$-tensor $b$ on a Riemannian manifold $(M^n, g)$,
we define ${\mathring {\mathcal W}}(b)$ by
$$
{\mathring {\mathcal W}}(b)(X, Y) = \sum_{i=1}^n b(\mathcal W (X, E_i)Y, E_i)
= \sum_{i, j=1}^n \mathcal W (X, E_i, Y, E_j)b(E_i, E_j)
$$
for any local frame $\{E_i\}$. For the Riemannian curvature tensor $R$,
${\mathring R}(b)$ is similarly defined.
\begin{lem}\label{lem2018-2-10-1}
Let $(g, f)$ be a solution of the CPE. Then
\begin{eqnarray*}
\d ({\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \mathcal W) = - \frac{n-3}{n-2} \widehat C
+ (1+f) {\mathring {\mathcal W}}(z),
\label{eqn2018-2-10-1}
\end{eqnarray*}
where ${\widehat C}$ is a $2$-tensor defined as
$$
{\widehat C}(X, Y) = C(Y, \nabla} \def\e{\epsilon} \def\la{\langle f, X)
$$
for any vectors $X, Y$.
\end{lem}
\begin{proof}
Let $\{E_i\}$ be a local frame normal at a point in $M$. At the point, it follows from definition
together with the CPE that
\begin{eqnarray*}
\d ({\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \mathcal W) (X, Y) = \d \mathcal W(X, Y, \nabla} \def\e{\epsilon} \def\la{\langle f) - Ddf(E_i, E_j)
\mathcal W (E_i, X, Y, E_j).
\end{eqnarray*}
Since the trace of $\mathcal W$ in the first and fourth components vanishes, applying
(\ref{eqn2016-12-3-16}), we obtain
$$
\d \mathcal W(X, Y, \nabla} \def\e{\epsilon} \def\la{\langle f) - Ddf(E_i, E_j) \mathcal W (E_i, X, Y, E_j)
= - \frac{n-3}{n-2}C(Y, \nabla} \def\e{\epsilon} \def\la{\langle f, X) + (1+f) {\mathring {\mathcal W}}(z) (X, Y).
$$
\end{proof}
Moreover, the divergence of the tensor ${\widehat C}$ has
the following form.
\begin{lem}\label{lem2018-2-14-10}
For any vector $X$, we have
\begin{eqnarray*}
\d {\widehat C} (X) = -(1+f)\langle i_X C, z\rangle.\label{eqn2018-2-9-2-10}
\end{eqnarray*}
\end{lem}
\begin{proof}
Let $\{E_i\}$ be a local geodesic frame around a point $p\in M$.
Using the fact that $\d C$ is symmetric and
that cyclic summation of $C$ is vanishing, we can show
$$
\d {\widehat C} (E_j) = - (1+f)z(E_i, E_k) C(E_j, E_k, E_i)
= - (1+f)\langle i_{E_j}C, z\rangle.
$$
\end{proof}
We are now ready to prove that $T=0$ implies the CPE conjecture.
\begin{thm}\label{thm2018-1-20-11}
Let $(g, f)$ be a non-trivial solution of the CPE.
If $T = 0$, then the Besse conjecture holds.
\end{thm}
\begin{proof}
It suffices to prove that $C=0$. First, on each level hypersurface $f^{-1}(t)$ for a regular value $t$ of $f$,
it follows from the definition of $T$ that
\begin{eqnarray*}
i_{\nabla f}T &=& \frac {|\nabla f|^2}{n-2}\left( z+\frac {\alpha}{n-1}g\right).
\end{eqnarray*}
Since $T=0$, we have
\begin{eqnarray}
z(E_i, E_j)= -\frac {\alpha}{n-1} \delta_{ij} \quad (2\leq i,j\leq n)\label{eqt08}
\end{eqnarray}
for a local orthonormal frame $\{E_1=N,\cdots, E_{n-1}, E_n\}$ with $N = \frac{\nabla} \def\e{\epsilon} \def\la{\langle f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}$. Also, from
$$
0=(n-2)i_{\nabla f}T(E_j, \nabla f)= \frac {n-2}{n-1}|\nabla f|^2 z(E_j, \nabla f)
$$
for $2\leq j\leq n$, we have
\begin{eqnarray}
z(\nabla f, E_j)=0.\label{eqt09}
\end{eqnarray}
Therefore,
\begin{eqnarray}
\qquad \langle i_{\nabla f}C, z\rangle = -\frac {\alpha}{n-1} \sum_{i=2}^{n}
C(\nabla f, E_i, E_i)=\frac {\alpha}{n-1} C(\nabla f, N, N)=0. \label{eqy2}
\end{eqnarray}
On the other hand, from Lemma~\ref{lem2019-5-23-10}, we have
\begin{eqnarray}
(1+f)C = {\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f} {\mathcal W} \label{eqn2018-2-10-7}
\end{eqnarray}
and hence,
\begin{eqnarray*}
C(X, Y, \nabla} \def\e{\epsilon} \def\la{\langle f) = 0 \label{eqn2018-2-10-5}
\end{eqnarray*}
for any vector fields $X$ and $Y$.
Since the cyclic summation of $C$ vanishes, we have
$$
C(Y, \nabla} \def\e{\epsilon} \def\la{\langle f, X) + C(\nabla} \def\e{\epsilon} \def\la{\langle f, X, Y) = 0.
$$
Taking the divergence $\d$ of (\ref{eqn2018-2-10-7}) and applying
Lemma~\ref{lem2018-2-10-1}, we obtain
\begin{eqnarray*}
-i_{\nabla} \def\e{\epsilon} \def\la{\langle f} C + (1+f)\d C = \frac{n-3}{n-2} i_{\nabla} \def\e{\epsilon} \def\la{\langle f} C + (1+f) {\mathring {\mathcal W}}(z).
\end{eqnarray*}
Thus,
\begin{eqnarray}
(1+f)\d C - (1+f) {\mathring {\mathcal W}}z = \frac{2n-5}{n-2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C.\label{eqn2018-2-10-8}
\end{eqnarray}
Note that, by (\ref{eqn2018-2-10-7}) and definition of $\mathring{\mathcal W}(z)$, the following
$$
\mathring{\mathcal W}(z) (\nabla f, X)= -(1+f) \langle i_XC, z\rangle
$$
holds for any vector field $X$. In particular, by (\ref{eqy2}), we have
$$
\mathring{\mathcal W}(z) (\nabla f, \nabla} \def\e{\epsilon} \def\la{\langle f) =0.
$$
Consequently, by (\ref{eqn2018-2-10-8}), we obtain
\begin{eqnarray}
(1+f)\d C(N, N) = 0.\label{eqn2018-2-10-10}
\end{eqnarray}
Finally, it follows from (\ref{eqt08}) and (\ref{eqt09}) together with
$\displaystyle{\sum_{j=2}^{n}C_{ijj;i} = - C_{i11;i}}$ that
\begin{eqnarray*}
\langle \d C, z\rangle &=& - C_{ijk;i}z_{jk} =
\frac{\alpha}{n-1} \sum_{i=1}^n \sum_{j=2}^{n}C_{ijj;i} - \alpha \sum_{i=1}^n C_{i11;i}\\
&=&
-\frac{n\alpha}{n-1} \sum_{i=1}^n C_{i11:i} = \frac{n\alpha}{n-1} \d C(N, N) =0.
\end{eqnarray*}
Hence, integrating (\ref{eqn2018-1-23-1}) over $M$, we have $C = d^D z = 0$,
which implies that $(M, g)$ has a harmonic curvature. Therefore, it follows from
\cite{ych} and \cite{erra} that $(M, g)$ is Einstein.
\end{proof}
\begin{rem}
{\rm
Introducing the Bach tensor on a Riemannian manifold $(M^n, g)$,
we can directly show that $C=0$ in Theorem~\ref{thm2018-1-20-11} without using
the divergence theorem in a higher dimensional case $n \ge 5$.
The {\it Bach tensor} $B$ is
defined as
\begin{eqnarray*}
B = \frac{1}{n-3}\d^D \d \mathcal W + \frac{1}{n-2}{\mathring {\mathcal W}}(r).\label{eqn2018-2-11-2}
\end{eqnarray*}
When the scalar curvature $s$ is constant,
we have ${\mathring W}(r) = {\mathring W}(z)$. Therefore, from (\ref{eqn2016-12-3-16}), we have
\begin{eqnarray*}
B = \frac{1}{n-2}\left(-\d C + {\mathring {\mathcal W}}(z)\right).\label{bach302}
\end{eqnarray*}
Furthermore, the following property holds in general (cf. \cite{CC}) for $n \ge 4$:
for any vector field $X$,
$$
(n-2)\d B (X) = - \frac{n-4}{n-2} \langle i_X C, z\rangle.
$$
The complete divergence of the Bach tensor has the following form:
\begin{eqnarray*}
\d\d B = \frac{n-4}{(n-2)^2} \left(\frac{1}{2} |C|^2 - \langle \d C, z\rangle \right).\label{eqn2018-2-22-1}
\end{eqnarray*}
Using these identities on the Bach tensor and taking the divergence of $T$, the following identities can be obtained:
\begin{itemize}
\item[(i)] $ (n-2)(1+f)B = -i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C + \frac{n-3}{n-2}{\widehat C} + (n-1) \d T.$
\item[(ii)] For any vector field $X$,
\begin{eqnarray*}
\d\d T (X) = \frac{1}{n-2}(1+f)\langle i_X C, z\rangle + \langle i_X T,
z\rangle.\label{eqn2018-2-9-1}
\end{eqnarray*}
\end{itemize}
Therefore, we obtain the following:
Let $(g, f)$ be a non-trivial solution of the CPE and $T = 0$. Then
$\langle i_X C, z \rangle = 0$ for any vector field $X$ by (ii), and hence,
$$
\frac{1}{2}|C|^2 = \langle \d C, z\rangle =0.
$$
}
\end{rem}
Before closing this section, we would like to mention some of the identities on the square norm
and the divergence of $T$, which will be used later in the proof of our main theorem.
\begin{lem}{\rm (\cite{GH1})}\label{lem2019-5-28-5}
Let $(g, f)$ be a solution of the CPE, Then,
$$
|T|^2= \frac 2{(n-2)^2}|\nabla f|^2 \left( |z|^2-\frac n{n-1}|i_Nz|^2\right).
$$
\end{lem}
\begin{lem} \label{lem2020-5-7-10}
Let $(g, f)$ be a solution of the CPE, Then,
$$
(n-1)\d T = \frac{sf}{n-1}z - D_{\nabla} \def\e{\epsilon} \def\la{\langle f}z + \frac{1}{n-2}\widehat C + \frac{n}{n-2} (1+f)z\circ z - \frac{1}{n-2}(1+f) |z|^2 g.
$$
Here $ z\circ z$ is defined by
$$
z\circ z(X, Y) = \sum_{i=1}^n z(X, E_i) z(Y, E_i)
$$
for a local frame $\{E_i\}_{i=1}^n$ of $M$.
\end{lem}
\begin{proof}
By using (\ref{cpet}) and (\ref{eqn2020-5-7-2}), we can compute
\begin{eqnarray*}
{\d} (df\wedge z) = \frac{sf}{n}z - D_{\nabla} \def\e{\epsilon} \def\la{\langle f}z + (1+f) z\circ z.\label{eqn2018-9-10-1}
\end{eqnarray*}
From the fact that $\delta z = 0$ and using (\ref{eqn2020-5-7-2}) again, we can obtain
\begin{eqnarray*}
{\d} (i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z \wedge g) = D_{\nabla} \def\e{\epsilon} \def\la{\langle f}z + {\widehat C} - (1+f)|z|^2 g + (1+f) z\circ z
- \frac{sf}{n(n-1)} z. \label{eqn2018-9-10-2}
\end{eqnarray*}
By combining these identities, the proof follows from the definition of the tensor $T$.
\end{proof}
\section{CPE with $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X)=0$}
For the potential function $f$ of the CPE, we let
$$
\alpha := z(N, N)
$$
for convenience. Here recall that $N = \frac{\nabla} \def\e{\epsilon} \def\la{\langle f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}$.
Note that the function $\alpha$ is well-defined only on the set $M \setminus {\rm Crit}(f)$,
where ${\rm Crit}(f)$ is the set of
all critical points of $f$. However, $|\alpha| \le |z|$, $\alpha$ can be extended to a $C^0$ function on the whole $M$.
Refer \cite{ych} for more details.
\noindent
First, it is easy to compute that
\begin{eqnarray}
N(|\nabla} \def\e{\epsilon} \def\la{\langle f|) = (1+f)\alpha - \frac{sf}{n(n-1)}\label{eqn2020-5-7-3}
\end{eqnarray}
and thus,
\begin{eqnarray}
N\left(\frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}\right) = - \frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2} \left[(1+f)\alpha - \frac{sf}{n(n-1)}\right].\label{eqn2020-5-7-4}
\end{eqnarray}
In particular, we have the following lemma.
\begin{lem}\label{lem2020-8-12-1}
Assume $z(\nabla f, X)=0$ for any vector field $X$ orthogonal to $\nabla f$.
Then $i_{\nabla} \def\e{\epsilon} \def\la{\langle f} z = \alpha \nabla} \def\e{\epsilon} \def\la{\langle f$ as a vector field, and $D_NN = 0.$
\end{lem}
\begin{proof}
It is obvious that $i_{\nabla} \def\e{\epsilon} \def\la{\langle f} z = \alpha \nabla} \def\e{\epsilon} \def\la{\langle f$. From this, by combining (\ref{eqn2020-5-7-2}) and (\ref{eqn2020-5-7-3}),
it is easy to see that $D_NN = 0$.
\end{proof}
From now, throughout this section, we assume that $(M^n, g, f)$ is a nontrivial solution of the CPE satisfying
\begin{eqnarray}
z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0 \label{eqn2020-5-7-1}
\end{eqnarray}
for any vector field $X$ orthogonal to $\nabla} \def\e{\epsilon} \def\la{\langle f$. In this case, by Lemma~\ref{lem2020-8-12-1}, we can write
\begin{eqnarray}
i_{\nabla} \def\e{\epsilon} \def\la{\langle f} z = \alpha df\label{eqn2020-8-22-1}
\end{eqnarray}
as a $1$-form. As the interior product $\tilde i_{\nabla} \def\e{\epsilon} \def\la{\langle f }{\mathcal W}$ of $4$-tensor,
we define the interior product $\tilde i_{\nabla} \def\e{\epsilon} \def\la{\langle f}T$ and $\tilde i _{\nabla} \def\e{\epsilon} \def\la{\langle f}C$ as
$$
\tilde i_{\nabla} \def\e{\epsilon} \def\la{\langle f}T(X, Y) = T(X, Y, \nabla} \def\e{\epsilon} \def\la{\langle f)\quad \mbox{and}\quad \tilde i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C(X, Y) = C(X, Y, \nabla} \def\e{\epsilon} \def\la{\langle f)
$$
for vector fields $X$ and $Y$.
\begin{lem}\label{Cotz}
Assume $z(\nabla f, X)=0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$. Then
$$
\tilde{i}_{\nabla f}T=0 \quad \mbox{and}\quad \tilde{i}_{\nabla f}C=0$$
Additionally, we have
\begin{eqnarray}
(1+f)|z|^2= \nabla f(\alpha) -\frac {sf}{n-1}\alpha. \label{eqn2}
\end{eqnarray}
\end{lem}
\begin{proof}
By definition of $T$, we have
$$
T(X, Y, \nabla f) =\frac {1}{n-1}\, df\wedge i_{\nabla f} z(X,Y)
$$
for any vectors $X$ and $Y$. Since $i_{\nabla} \def\e{\epsilon} \def\la{\langle f} z = \alpha df$, we obviously have
$df \wedge i_{\nabla} \def\e{\epsilon} \def\la{\langle f} z = 0$, which shows $\tilde{i}_{\nabla f}T=0$.
The second equality follows from Lemma~\ref{lem2019-5-23-10}.
For (\ref{eqn2}), we note from the CPE that
$$
\d (i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z) = -(1+f)|z|^2.
$$
Hence, we can obtain (\ref{eqn2}) by taking the divergence of (\ref{eqn2020-8-22-1}).
\end{proof}
For a nontrivial solution $(g, f)$ of the CPE satisfying $ z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$,
there are two important properties that together play a key role in proving our main theorems.
One is that both the functions, $\alpha = z(N, N)$ and $|z|^2$, are constants on each level set $f^{-1}(c)$, and the other is that there are no critical points of $f$ except its maximum and minimum
points in $M$.
\begin{lem}\label{lem2020-8-23-2}
The functions $|\nabla} \def\e{\epsilon} \def\la{\langle f|, \alpha$ and $|z|^2$ are all constants on each level set $f^{-1}(c)$ of $f$.
\end{lem}
\begin{proof}
We may assume that $c$ is a regular value of $f$ by the Sard's theorem.
Let $X$ be a tangent vector on $f^{-1}(c)$. Then, from (\ref{eqn2020-5-7-2}), we have
$$
\frac{1}{2} X(|\nabla} \def\e{\epsilon} \def\la{\langle f|^2) = Ddf(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = (1+f)z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) - \frac{sf}{n(n-1)} g(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0.
$$
Now, let $X$ be a (local) vector field orthogonal to $\nabla f$ such that $z(X, \nabla} \def\e{\epsilon} \def\la{\langle f ) = 0$.
Since $D_NN=0$ by Lemma~\ref{lem2020-8-12-1}, we have
$g(D_NX, N) = - g(X, D_NN) = 0$ and $D_N\nabla} \def\e{\epsilon} \def\la{\langle f = - |\nabla} \def\e{\epsilon} \def\la{\langle f|N\left(\frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}\right) \nabla} \def\e{\epsilon} \def\la{\langle f$. So,
\begin{eqnarray}
D_N z(X, \nabla} \def\e{\epsilon} \def\la{\langle f) = - z(D_N X, \nabla} \def\e{\epsilon} \def\la{\langle f)-z(X, D_N \nabla} \def\e{\epsilon} \def\la{\langle f) = 0.\label{eqn2020-8-26-1}
\end{eqnarray}
Since $\tilde{i}_{\nabla f}C=0$ by Lemma~\ref{Cotz}, we have
$$
0=C(X, N, \nabla f)=D_X z(N, \nabla f)-D_{N}z(X, \nabla f) = D_X z(N, \nabla f)=|\nabla f|X(\alpha),
$$
implying that $\alpha $ is a constant on $f^{-1}(c)$.
The property, $D_NN=0$, also implies $ [X, N]$ is orthogonal to $\nabla} \def\e{\epsilon} \def\la{\langle f$ and therefore,
$$
X(N(\alpha)) = N(X(\alpha)) - [X, N](\alpha) = 0.
$$
Since $\nabla} \def\e{\epsilon} \def\la{\langle \alpha = N(\alpha)N$, it shows $Dd\alpha(X, \nabla} \def\e{\epsilon} \def\la{\langle f) = 0$, and from (\ref{eqn2}), we have $X(|z|^2) = 0$.
Consequently, we conclude that $|z|^2$ is a constant along each level set $f^{-1}(c)$ of $f$.
\end{proof}
\begin{lem} \label{lem2020-8-26-2}
$\alpha$ is nonpositive on $M$.
\end{lem}
\begin{proof}
First assume that each level set $f^{-1}(t)$ is connected.
Suppose that $\alpha >0$ on a level set $f^{-1}(c)$.
If $c \le -1$, the divergence theorem shows that
$$
0\geq \int_{f \le c} (1+f)|z|^2= \int_{f \le c} {\rm div}(i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z) = \int_{f=c} \alpha |\nabla f|,
$$
which is impossible as $\alpha$ is constant on each level set.
Note that $-1$ is a regular value of $f$ (see Appendix).
If $c>-1$, the divergence theorem states that
$$
0\le \int_{f \ge c} (1+f)|z|^2= - \int_{f=c} \alpha |\nabla f|,
$$
which is also impossible. In any case, we get a contradiction.
For general case, i.e., if level sets may not be connected, refer \cite{ych}
(In the proof of Lemma 5.2 there, we used only the CPE and the condition $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$).
\end{proof}
\begin{lem}\label{cor191}
Let $(g,f)$ be a non-trivial solution of (\ref{cpe}) on an $n$-dimensional compact manifold $M$
with $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$. Then, there are no critical points
of $f$ except at its maximum and minimum points unless $g$ is an Einstein.
\end{lem}
\begin{proof}
Note that $|\nabla f|^2$ is constant on each level set of $f$.
From the Bochner-Weitzenb\"ock formula together with (\ref{eqn2020-5-7-2}),
we have
$$
\frac 12 \Delta |\nabla f|^2-\frac 12 \frac {\nabla f(|\nabla f|^2)}{1+f}
+\frac s{n(n-1)}\frac 1{1+f}|\nabla f|^2=|Ddf|^2.
$$
By maximum principle, the function $|\nabla f|^2$ cannot have its local maximum in
$M_0:=\{x\in M \, |\, f(x) < -1\}$.
Let $p$ be a critical point of $f$ in $M$ other than the minimum or maximum points
of $f$. From the argument above, we can see that
$p$ cannot be in $M_0:=\{x\in M \, |\, f(x) < -1\}$ unless $f$ is a constant.
In fact, if $p \in M_0$ and $p$ is not a minimum point, then there should a local maximum point of $|\nabla f|^2$ in $\{x\in M\,|\, f < f(p)\}$, which is impossible.
We also claim that $p$ cannot be in $M^0:=\{x\in M \, |\, f(x) > -1\}$.
Recall that we have $f(p) \ne -1$ and
$$
\nabla f(|\nabla f|^2)= \frac {2s}{n(n-1)}|\nabla f|^2>0
$$
on the set $f=-1$. This shows that $|\nabla f|^2 $ might have its maximum at a point, say $q$, in $M^0$. However, since
$$
0 =\frac 12 \nabla f(|\nabla f|^2)=(1+f) \alpha |\nabla f|^2 -\frac {sf}{n(n-1)}|\nabla f|^2
$$
at the point $q$ and $(1+f)\alpha |\nabla f|^2 \le 0$, $q$ should lie in
$\Omega:=\{x\in M \, |\, -1<f(x)<0\}$.
This implies that there are no critical points of $f$ in $M^0\setminus \Omega$.
Now, if $|z|^2(p)\neq 0$ with $p \in \Omega$, by Lemma 3.9 of \cite{ych},
the point $p$ has to be a local maximum point of $f$, which is not possible because
$$
\Delta f=-\frac s{n-1}f>0
$$
on $\Omega$. Thus, we have $|z|^2(p) = 0$ with $p \in \Omega$. In particular,
if we let $f(p) = c$ with $-1< c <0$, then we have $\alpha =0$ on the level set $f^{-1}(c)$.
Finally, from the divergence theorem, we have
\begin{eqnarray*}
0 \le \int_{f >c}(1+f)|z|^2 &=& - \int_{f=c} \alpha |\nabla} \def\e{\epsilon} \def\la{\langle f| = 0,
\end{eqnarray*}
which means $z=0$ on the set $f \ge c$. Then, by the analyticity of $g$ and $f$,
$z$ must be vanishing on the set $M^0$ and consequently on the entire $M$.
\end{proof}
Furthermore, we can show that the potential function $f$ has only one maximum point and only one minimum point when $(M, g)$ has positive isotropic curvature.
\begin{thm}\label{thm2019-12-23-1}
Let $(g,f)$ be a non-trivial solution of the CPE on an $n$-dimensional compact manifold $M$
with $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$. Then, there are only two isolated critical points of $f$, in other words, only one maximum point and only one minimum point of $f$ on $M$ exist.
\end{thm}
\begin{proof}
Let ${\min_M f = a}$ and suppose $f^{-1}(a)$ is not discrete. By Lemma~\ref{cor191},
all the connected components of the level hypersurface $f^{-1}(t)$ for any regular value
$t$ of $f$ have the same topological type.
In particular, since $M$ is smooth and $f^{-1}(a)$ is not discrete, $f^{-1}(a)$
must be a hypersurface and it also has the same topological type as any
connected component of $f^{-1}(t)$ for any regular value $t$ of $f$.
Moreover, for a sufficiently small $\epsilon >0$, $f^{-1}(a+\epsilon)$ has two
connected components, say $\Sigma_\epsilon^+, \,\, \Sigma_\epsilon^-$.
Note that $\Sigma_0^+=\Sigma_0^- = f^{-1}(a)$.
Let $\nu$ be a unit normal vector field on $\Sigma:= f^{-1}(a)$. Then,
$\nu$ can be extended smoothly to a vector field $\Xi$ defined on a tubular neighborhood
of $f^{-1}(a)$ such that $\Xi|_{f^{-1}(a)} = \nu$ and
$\Xi|_{\Sigma_\epsilon^+} = N = \frac{\nabla} \def\e{\epsilon} \def\la{\langle f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}$,
$\Xi|_{\Sigma_\epsilon^-} = -N = -\frac{\nabla} \def\e{\epsilon} \def\la{\langle f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}$.
Note that
$$
\lim_{\epsilon\to 0+} N = \lim_{\epsilon \to 0-}(-N) = \nu.
$$
On the hypersurface $f^{-1}(a+\epsilon)$ near $f^{-1}(a)$, the Laplacian of $f$ is given by
$$
\Delta f = Ddf(N, N) + m |\nabla} \def\e{\epsilon} \def\la{\langle f|,
$$
where $m$ denotes the mean curvature of $f^{-1}(a+\epsilon)$.
In particular, by letting $\e \to 0$, we have
$$
\Delta f = Ddf(\nu, \nu)
$$
on $f^{-1}(a)$. Note that the mean curvature $m$ does not blow up on $f^{-1}(a)$.
Therefore, from the CPE\,
$ -\frac{sf}{n-1} = (1+f)z(\nu, \nu) - \frac{sf}{n(n-1)}$, we have
\begin{eqnarray}
z(\nu, \nu) = - \frac{sf}{n(1+f)} = - \frac{sa}{n(1+a)} \label{eqn2019-1210-1}
\end{eqnarray}
on the set $\Sigma = f^{-1}(a)$. In particular, we have $z_p(\nu, \nu) <0$ , since we may assume $a<-1$ as mentioned in Introduction.
Now, since $a$ is the minimum value of $f$, for
each point $p \in f^{-1}(a)$, the index of $Ddf_p$ is zero, i.e., for any vector
$v $ at $p$, we have $Ddf_p(v, v) \ge 0$.
Choosing an orthonormal basis $\{e_1 = \nu(p), e_2, \cdots, e_n\}$ on $T_pM$, we obtain
$$
Ddf_p(\nu, \nu ) = (1+f)z_p(\nu, \nu) - \frac{sf}{n(n-1)}\ge 0
$$
and for all $2\le i \le n$,
$$
Ddf_p(e_i, e_i ) = (1+f)z_p(e_i, e_i) - \frac{sf}{n(n-1)}\ge 0.
$$
In particular, from (\ref{eqn2019-1210-1}),
$$
z_p(e_i, e_i) \le \frac{s}{n(n-1)}\cdot \frac{a}{1+a} = - \frac{1}{n-1} z_p(\nu, \nu).
$$
Summing up $i=2, \cdots, n$, we can see
\begin{eqnarray}
z_p(e_i, e_i) = - \frac{1}{n-1} z_p(\nu, \nu) = \frac{s}{n(n-1)}\cdot\frac{a}{1+a} >0\label{eqn2021-1-25-1}
\end{eqnarray}
for each $i=2, \cdots, n$.
Thus, for each $i =2, \cdots, n$,
\begin{eqnarray}
Ddf_p(e_i, e_i) = (1+f)z_p(e_i, e_i) - \frac{sf}{n(n-1)} = 0.\label{eqn2021-1-24-1}
\end{eqnarray}
\vspace{.2in}
Now, we claim that the minimum set $\Sigma = f^{-1}(a)$ is totally geodesic and in particular, the mean curvature is vanishing, i.e., $m=0$ on $\Sigma$.
In fact, fix $i$ for $i=2,3, \cdots, n$ and let $\gamma : [0, l) \to M$ be a unit speed geodesic such that $\gamma(0) =p\in \Sigma$
and $\gamma'(0) = e_i \in T_p\Sigma$.
Then we have
\begin{eqnarray}
D_{\gamma'} N &=&
\langle \gamma', N\rangle N\left(\frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}\right)\nabla} \def\e{\epsilon} \def\la{\langle f + \frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}\left[(1+ f) z(\gamma', \cdot) - \frac{sf}{n(n-1)}\gamma'\right].
\label{eqn2020-6-17-1}
\end{eqnarray}
Recall that, by (\ref{eqn2020-5-7-4})
$$
N\left(\frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}\right) = - \frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2} \left((1+ f)\alpha - \frac{sf}{n(n-1)}\right).
$$
Let
$$
\gamma' = \langle \gamma', N\rangle N + (\gamma')^{\top},
$$
where $\gamma'(t)^{\top}$ is the tangential component of $\gamma'(t)$ to $f^{-1}(\gamma(t))$,
and substituting these into (\ref{eqn2020-6-17-1}), we obtain
$$
|\nabla} \def\e{\epsilon} \def\la{\langle f| D_{\gamma'} N = (1+ f) z((\gamma')^\top, \cdot) - \frac{sf}{n(n-1)}(\gamma')^\top.
$$
Taking the covariant derivative in the direction $N$, we have
\begin{eqnarray*}
&&Ddf(N, N) D_{\gamma'}N + |\nabla} \def\e{\epsilon} \def\la{\langle f|D_ND_{\gamma'}N \\
&&\qquad
=
|\nabla} \def\e{\epsilon} \def\la{\langle f| z((\gamma')^\top, \cdot) + (1+f) D_N[z((\gamma')^{\top}, \cdot)] - \frac{s}{n(n-1)}|\nabla} \def\e{\epsilon} \def\la{\langle f| (\gamma')^\top - \frac{sf}{n(n-1)}D_N \gamma'^{\top}.
\end{eqnarray*}
Letting $t\to 0+$, we obtain
\begin{eqnarray}
-\frac{sa}{n-1} D_{e_i}\nu = (1+a) D_{\nu} [z((\gamma')^{\top}, \cdot)]\bigg|_p - \frac{sa}{n(n-1)} D_{\nu} (\gamma')^{\top}\bigg|_p\label{eqn2021-1-25-2}
\end{eqnarray}
because the covariant derivative depends only on the point $p$ and initial vector $e_i$.
Now since $z(N, X) = 0$ for $X \perp N$, we may assume that $\{e_i\}_{i=2}^n$ diagonalizes $z$ at the point $p$. Then
\begin{eqnarray}
z(D_{\nu} (\gamma')^{\top}, e_i)\bigg|_p = 0.\label{eqn2021-1-25-3}
\end{eqnarray}
\vspace{.15in}
{\bf Assertion:}\, $\nu(z(\gamma'(t), \gamma'(t))|_{t=0} = 0.$
\vspace{.15in}
\noindent
Defining $\vp(t):= f\circ \gamma(t)$, we have $\vp'(0) =0$ and also $\vp''(0) = Ddf(e_i, e_i) = 0$ by (\ref{eqn2021-1-24-1}).
Since $\Sigma$ is the minimum set of $f$, $\vp'(t)$ is nondecreasing when $\vp(t)$ is sufficiently close to $a = \min f$ and
so $\vp''(t) \ge 0$ for sufficiently small $t$. So
\begin{eqnarray}
\vp''(t) = Ddf(\gamma'(t). \gamma'(t)) = [1+\vp(t) ] z(\gamma'(t), \gamma'(t)) - \frac{s}{n(n-1)}\vp(t) \ge 0\label{eqn2021-1-26-1}
\end{eqnarray}
for sufficiently small $t>0$. However, by (\ref{eqn2021-1-25-1}), we have $z(\gamma'(t), \gamma'(t)) >0$ for sufficiently small $t>0$.
So,
$$
0< z(\gamma'(t), \gamma'(t)) \le \frac{s}{n(n-1)}\cdot \frac{\vp(t)}{1+\vp(t)}
$$
for sufficiently small $t>0$. Defining $\xi(t) = z(\gamma'(t), \gamma'(t)) - \frac{s}{n(n-1)}\cdot \frac{\vp(t)}{1+\vp(t)}$, we have
$\xi(t) \le 0$ and $\xi(0) = 0$ by (\ref{eqn2021-1-24-1}). Thus,
$$
\xi'(0) = \frac{d}{dt}\bigg|_{t=0} z(\gamma'(t), \gamma'(t)) \le 0.
$$
Now considering an extension of $\gamma$ to an interval $(-\e, 0]$,
we can see that $\frac{d}{dt}|_{t=0} z(\gamma'(t), \gamma'(t))$ cannot be negative since $\Sigma$ is the minimum set of $f$.
In other words, we must have
$$
\frac{d}{dt}\bigg|_{t=0} z(\gamma'(t), \gamma'(t)) = 0,
$$
which completes the Assertion.
\vspace{.12in}
Let $\{N, E_2, \cdots, E_n\}$ be a local frame around $p$ such that $E_i(p) = e_i$ for $2 \le i$. Then, for $i \ge 2$,
$$
z(E_i, \cdot) = \sum_{j=2}^n z(E_i, E_j) E_j
$$
and
$$
D_N[z(\gamma'^\top, \cdot)] = D_N \left[z(\gamma'^\top, E_j)E_j\right] = N\left(z(\gamma'^\top, E_j)\right)E_j + z(\gamma'^\top, E_j)D_NE_j.
$$
So,
$$
\langle D_N[z(\gamma'^\top, \cdot)], E_i\rangle = N\left(z(\gamma'^\top, E_i)\right) + z(\gamma'^\top, E_j) \langle D_NE_j, E_i\rangle.
$$
Note that $z(\gamma', \gamma') = \Vert \gamma'^\top\Vert z(\gamma'^{\top}, E_i) + \langle \gamma', N\rangle^2 \alpha$ with $\alpha = z(N, N).$
Since $\Vert \gamma'^\top\Vert$ attains its maximum at $p$ and $\langle \gamma', N\rangle^2$ attains its minimum $0$ at $p$, we have
$$
\nu [z(\gamma', \gamma')]\bigg|_p = \Vert \gamma'^\top\Vert \nu[z(\gamma'^{\top}, E_i)] + \langle \gamma', N\rangle^2 \nu(\alpha) =
\nu[z(\gamma'^{\top}, E_i)]\bigg|_p,
$$
which shows that
\begin{eqnarray}
\nu[z(\gamma'^{\top}, E_i)]\bigg|_p = 0\label{eqn2021-2-4-1}
\end{eqnarray}
by {\bf Assertion}.
Letting $t\to 0$, and applying (\ref{eqn2021-1-25-3}) and (\ref{eqn2021-2-4-1}), we obtain
\begin{eqnarray*}
\langle D_\nu [z(\gamma'^\top, \cdot)], E_i\rangle|_p &=& \nu\left(z(\gamma'^\top, E_i)\right)|_{p} + z(e_i, e_i) \langle D_\nu E_i, E_i\rangle|_p\\
&=&0
\end{eqnarray*}
Thus, by (\ref{eqn2021-1-25-2}) and (\ref{eqn2021-1-25-3}) again, we have
$$
- \frac{sa}{n-1} \langle D_{e_i}\nu, e_i\rangle = (1+a) \langle D_{\nu}[z(\gamma'^\top, \cdot)], E_i\rangle|_p
- \frac{sa}{n(n-1)} \langle D_{\nu} E_i, E_i\rangle|_p =0.
$$
That is,
$$
\langle D_{e_i}\nu, e_i\rangle = 0.
$$
Hence, the square norm of the second fundamental form $A$ is given by
$$
|A|^2(p) = \sum_{i=2}^n \left|\left(D_{e_i}e_i\right)^\perp\right|^2 = \sum_{i=2}^n \langle D_{e_i}\nu, e_i\rangle^2 = 0,
$$
which shows $\Sigma = f^{-1}(a)$ is totally geodesic.
\vspace{.12in}
Finally, from (\ref{eqn2019-1210-1}), we have
$$
{\rm Ric}(\nu, \nu) = \frac{s}{n}\cdot \frac{1}{1+a} <0.
$$
Furthermore, the stability operator for hypersurfaces with a vanishing second fundamental form clearly becomes
$$
\int_{\Sigma} \left[|\nabla} \def\e{\epsilon} \def\la{\langle \vp|^2 - {\rm Ric}(\nu, \nu)\vp^2\right] \ge 0
$$
for any function $\vp$ on $\Sigma$. By Fredholm alternative (cf. \cite{f-s}, Theorem 1),
there exists a positive function $\vp>0$ on $\Sigma$ satisfying
$$
\Delta^\Sigma \vp + {\rm Ric}(\nu, \nu)\vp =0.
$$
However, it follows from the maximum principle that
$\vp$ must be a constant since $\Sigma$ is a compact, which is impossible.
A similar argument shows that $f$ has only one maximum point.
\end{proof}
\section{Proof of Theorem~\ref{thm2020-8-22-3}}
In this section, we will prove Theorem~\ref{thm2020-8-22-3}. Let $(M^n, g, f)$ be a nontrivial solution of the CPE satisfying
$z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for any vector field $X$ which is orthogonal to $\nabla} \def\e{\epsilon} \def\la{\langle f$.
From Theorem~\ref{thm2018-1-20-11}, it suffices to prove that $T=0$ or $(M,g)$ has harmonic curvature.
To this end, we introduce a warped product metric involving
$\frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|}$ as a fiber metric on each level
set $f^{-1}(c)$. Consider a warped product metric $\bar{g}$ on $M$ by
\begin{eqnarray*}
\bar{g}= \frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|}+|\nabla f|^2 g_{\Sigma},
\label{eqn2019-8-27-1}
\end{eqnarray*}
where $g_{\Sigma}$ is the restriction of $g$ to $\Sigma:= f^{-1}(-1)$.
Note that, from Theorem~\ref{thm2019-12-23-1},
the metric $\bar g$ is smooth on $M$ except, possibly at two points, the maximum and minimum points of $f$.
Furthermore, applying Morse theory \cite{mil} together with Theorem~\ref{thm2019-12-23-1}, we can see that
$M$ is homeomorphic to ${\Bbb S}^{n}$, and fiber $f^{-1}(t)$ is topologically ${\Bbb S}^{n-1}$ except the two critical points of $f$.
The following lemma shows that $\nabla f$ is a conformal Killing vector field with respect to the metric $\bar{g}$.
\begin{lem}\label{lemt1}
Let $(g,f)$ be a nontrivial solution of the CPE on
an a $n$-dimensional compact manifold $M$ with $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$.
Then,
$$\frac 12 {\mathcal L}_{\nabla f}\bar{g} =N(|\nabla f|)\bar{g}= \frac 1n
(\bar{\Delta }f)\, \bar{g}.
$$
Here, $\mathcal L$ denotes the Lie derivative.
\end{lem}
\begin{proof}
Note that, by (\ref{eqn2020-5-7-2}) we have,
$$
\frac 12 {\mathcal L}_{\nabla f} g =D_gdf =(1+f)z -\frac {sf}{n(n-1)}g.
$$
Let $X$ and $Y$ be two vector fields with $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$ and $Y \perp \nabla} \def\e{\epsilon} \def\la{\langle f$. By the definition of Lie derivative,
\begin{eqnarray*}
\frac 12{\mathcal L}_{\nabla f}(df \otimes df)(X,Y)&=& Ddf(X, \nabla f)df(Y)+df(X)Ddf(Y, \nabla f)\\
&=& 2\left( (1+f)\alpha -\frac {sf}{n(n-1)}\right) \, df\otimes df(X, Y).
\end{eqnarray*}
Therefore, from (\ref{eqn2020-5-7-3}),
\begin{eqnarray}
\frac 12 {\mathcal L}_{\nabla f}\left( \frac {df}{|\nabla f|} \otimes \frac {df}{|\nabla f|} \right)
= N(|\nabla f|) \frac {df}{|\nabla f|} \otimes \frac {df}{|\nabla f|}.\label{eqnt2}
\end{eqnarray}
Since
$$
\frac 12 {\mathcal L}_{\nabla f} (|\nabla f|^2 g_{\Sigma})
= \frac 12 \nabla f(|\nabla f|^2) g_\Sigma
= Ddf(\nabla f, \nabla f)g_\Sigma= N(|\nabla f|) |\nabla f|^2 g_\Sigma,
$$
we conclude that
$$
\frac 12 {\mathcal L}_{\nabla f}\bar{g}=\bar{D}df= N(|\nabla f|) \bar{g}.
$$
In particular, we have $\bar{\Delta} f =n N(|\nabla f|) $.
\end{proof}
\begin{lem} \label{lem2019-6-22-1}
Let $(g,f)$ be a nontrivial solution of the CPE on
an $n$-dimensional compact manifold $M$ with $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$.
Then $T=0$ on $M$.
\end{lem}
\begin{proof}
Let $p, q \in M$ be two points such that $f(p) = \min_M f$ and $f(q) = \max_M f$, respectively, and let $\bar M = M \setminus \{p, q\}$.
Due to Theorem~\ref{thm2019-12-23-1} and Lemma~\ref{lemt1}, we can apply Tashiro's result \cite{tas} and can see that
$(\bar M, \bar g)$ is conformally equivalent
to ${\Bbb S}^n\setminus \{\bar p, \bar q\}$, where $\bar p$ and $\bar q$ are the points in ${\Bbb S}^n$ corresponding to $p$ and $q$, respectively.
In particular, by Theorem 1 in \cite{bgv},
the fiber space $(\Sigma, g|_{\Sigma})$ is a space of constant curvature.
Thus,
$$
(\Sigma, g|_\Sigma) \equiv ({\Bbb S}^{n-1}, \, r \cdot g_{{\Bbb S}^{n-1}}),
$$
where $r>0$ is a positive constant and $g_{{\Bbb S}^{n-1}}$ is a round metric.
Now, replacing $\Sigma = f^{-1}(-1)$ by $\Sigma_t:= f^{-1}(t)$ in
(\ref{eqn2019-8-27-1}), it can be easily concluded that the warped product metric $\bar g_t$ also
satisfies Lemma~\ref{lemt1}, and hence, the same argument mentioned above
shows that,
for any level hypersurface $\Sigma_t:= f^{-1}(t)$,
$$
(\Sigma_t, g|_{\Sigma_t}) \equiv ({\Bbb S}^{n-1}, r(t) \cdot g_{{\Bbb S}^{n-1}}).
$$
Therefore, the original metric $g$ can also be written as
\begin{eqnarray}
g= \frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|}+ b(f)^2 g_\Sigma,\label{eqn2019-12-23-1}
\end{eqnarray}
where $b (f)>0$ is a positive function depending only on $f$.
From (\ref{eqnt2}) and the following identity
$$
\frac 12 {\mathcal L}_{\nabla f}(b^2 g_\Sigma)=b\langle \nabla f, \nabla b\rangle g_\Sigma = b|\nabla f|^2 \frac {db}{df}g_\Sigma,
$$
we obtain
\begin{eqnarray}
\frac 12 {\mathcal L}_{\nabla f}g=N(|\nabla f|)\frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|}+b|\nabla f|^2 \frac{db}{df}g_\Sigma.\label{eqnt5-1}
\end{eqnarray}
On the contrary, from (\ref{cpe}) together with (\ref{eqn2020-5-7-3}) and (\ref{eqn2019-12-23-1}), we have
\begin{eqnarray*}
\lefteqn{\frac 12 {\mathcal L}_{\nabla f}g = Ddf =(1+f)z -\frac {sf}{n(n-1)}g}\\
&=& N(|\nabla f|) \frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|} +(1+f)z-(1+f)\alpha \frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|} -\frac {sf}{n(n-1)}b^2 g_\Sigma.
\end{eqnarray*}
Comparing this to (\ref{eqnt5-1}), we obtain
\begin{eqnarray}
\left( b|\nabla f|^2 \frac {db}{df} +\frac {sf}{n(n-1)}b^2\right) g_\Sigma=(1+f)\left(z-\alpha \frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|}\right).\label{eqnt7-1}
\end{eqnarray}
Now, let $\{E_1, E_2, \cdots, E_n\}$ be a local frame with $E_1=N$. Then, we have
$$
b|\nabla f|^2 \frac {db}{df}= (1+f)z(E_i, E_i)-\frac {sf}{n(n-1)}b^2
$$
for each $2\leq j\leq n$. Summing up these, we obtain
$$
(n-1)b|\nabla f|^2 \frac {db}{df}= -(1+f)\alpha -\frac {sf}nb^2.
$$
Substituting this into (\ref{eqnt7-1}), we get
\begin{eqnarray}
-\frac {\alpha}{n-1} g_\Sigma= z-\alpha \frac {df}{|\nabla f|}\otimes \frac {df}{|\nabla f|}. \label{eqn2020-8-29-1}
\end{eqnarray}
Replacing $(\Sigma, g_\Sigma)$ by $(\Sigma_t , g_{\Sigma_t})$, we can see that the argument mentioned above is also valid.
Thus, (\ref{eqn2020-8-29-1}) shows that, on each level hypersurface $f^{-1}(t)$, we have
$$
z(E_i, E_j)=-\frac {\alpha}{n-1}
$$
for $2\leq j\leq n$. Hence,
$$
|z|^2=\alpha^2+\frac {\alpha^2}{n-1}=\frac n{n-1}\alpha^2=\frac n{n-1}|i_Nz|^2,
$$
since $z(N, E_i)=0$ for $i\geq 2$. As a result, it follows from
Lemma~\ref{lem2019-5-28-5} that $T=0.$
\end{proof}
\begin{rem}\label{rem2020-9-28-3}
{\rm
Let $(g, f)$ be a non-trivial solution of the CPE with $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$.
In Appendix, we show the following result.
\begin{eqnarray}
\frac{s}{n(n-1)} g = R_N +z +\frac{1+f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C- \left(\alpha - \frac{s}{n(n-1)}\right) \frac{df}{|df|}\otimes \frac{df}{|df|}.\label{eqn2020-9-28-1}
\end{eqnarray}
Here, $R_N$ is defined as follows
$$
R_N(X, Y) = R(X, N, Y, N)
$$
for any vector field $X$ and $Y$. Let $\e:= \frac{s}{n(n-1)}$ and
$$
h:= R_N +\left(z- \alpha \frac{df}{|df|}\otimes \frac{df}{|df|}\right) +\frac{1+f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C.
$$
We can then rewrite the metric $g$ as
\begin{eqnarray}
g = \frac{1}{\e} h + \frac{df}{|df|}\otimes \frac{df}{|df|}.\label{eqn2020-9-28-10}
\end{eqnarray}
Then, the following can be proved on the set $f^{-1}(-1)$:
\begin{eqnarray}
{\mathcal L}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \left(\frac{h}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}\right) = 0\label{eqn2020-9-28-2}
\end{eqnarray}
Therefore, we can conclude that $g$ can be expressed as a warped product metric, and that it is,
in fact, equal to the metric $\bar g$ defined at the beginning of Section 4.
Refer to the Appendix for the detailed proofs of (\ref{eqn2020-9-28-1}) and (\ref{eqn2020-9-28-2}).
}
\end{rem}
Combining Theorem~\ref{thm2018-1-20-11} and Lemma~\ref{lem2019-6-22-1}, we obtain the following theorem.
\begin{thm}
Let $(g,f)$ be a nontrivial solution of the CPE on
an $n$-dimensional compact manifold $M$ with $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$.
Then $M$ is isometric to a standard sphere ${\Bbb S}^n$.
\end{thm}
\section{CPE with positive isotropic curvature}
In this section, we will prove that if $(g, f)$ is a nontrivial solution of the CPE with positive isotropic curvature, then $M$ is isometric to a standard sphere. In the view of Theorem~\ref{thm2020-8-22-3},
it suffices to show that,
$$
z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0
$$
for any $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$. To do this, we introduce a $2$-form $\omega$ on $M$ defined as
$$
\omega := df \wedge i_{\nabla f}z
$$
by considering $i_{\nabla f} z$ as a $1$-form.
In this section, the dimension of the manifold $M$ is assumed to be $n \ge 4$.
First, we have the following.
\begin{lem}\label{lem2018-4-30-20}
We have
\begin{eqnarray}
\omega = (n-1) {\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f} T = -(1+f) {\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f}C.\label{keyeqn}
\end{eqnarray}
\end{lem}
\begin{proof}
As in the proof of Lemma~\ref{Cotz}, it follows from the definition of $T$ that
$$
(n-2) T(X. Y, \nabla f)
=\frac {n-2}{n-1}\, df\wedge i_{\nabla f} z(X,Y)=\frac {n-2}{n-1}\, \omega (X,Y)
$$
for vectors $X$ and $Y$. The second equality follows from
Lemma~\ref{lem2019-5-23-10}.
\end{proof}
\begin{lem}\label{simple}
Let $\{E_1, E_2, \cdots, E_n\}$ be a local frame with $E_1=N = \frac{\nabla} \def\e{\epsilon} \def\la{\langle f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|}$.
Then
$$
\omega =0 \quad \mbox{if and only if}\quad \tilde{i}_{\nabla f}C(N, E_j)=0\quad (j \ge 2).
$$
\end{lem}
\begin{proof}
It follows from the definition of $\omega$ that
\begin{eqnarray}
\omega (E_j, E_k)=0\quad\mbox{for all $j, k \ge 2$},\label{eqn2019-5-26-1}
\end{eqnarray}
which shows, by Lemma~\ref{lem2018-4-30-20},
\begin{eqnarray*}
\tilde{i}_{\nabla f}C(E_j, E_k)=0\label{simple-2}
\end{eqnarray*}
for $2\leq j,k\leq n$. Therefore, it is easy to see that
$$
\omega =0 \quad \mbox{if and only if}\quad \tilde{i}_{\nabla f}C(N, E_j)=0
$$
for $2\leq j\leq n$.
\end{proof}
\vspace{.12in}
Next, we prove that $\omega$ is closed, and when $(M, g)$
has positive isotropic curvature, $\omega$ is vanishing.
\begin{lem} \label{lem191}
As a $2$-form, we have the following
$$
\tilde{i}_{\nabla f}C= di_{\nabla f}z.
$$
\end{lem}
\begin{proof}
Choose a local frame $\{E_i\}$ which is normal at a point $p \in M$,
and let $\{\theta ^i\}$ be its dual coframe so that $d\theta^i\vert_{p}=0$.
Since $i_{\nabla f}z=\sum_{l,k=1}^n f_l z_{lk}\theta ^k$ with $E_l(f) = f_l$ and
$z(E_l, E_k) = z_{lk}$, by (\ref{eqn2020-5-7-2}), we have
\begin{eqnarray*}
di_{\nabla} \def\e{\epsilon} \def\la{\langle f}z &=&
\sum_{j,k} \sum_{l} (f_{lj} z_{lk} + f_l z_{lk;j}) \theta^j \wedge \theta^k \\
&=&
\sum_{j<k} \sum_{l} \left\{(f_{lj} z_{lk} -f_{lk} z_{lj}) + f_l (z_{lk;j} - z_{lj;k})\right\}
\theta^j \wedge \theta^k \\
&=&
\sum_{j<k} \sum_{l} \left[\left\{\left( (1+f)z_{lj} -\frac{sf\,\delta_{lj} }{n(n-1)}\right) z_{lk}
- \left( (1+f)z_{lk} -\frac{sf\,\delta_{lk}}{n(n-1)} \right) z_{lj}\right\} \right] \theta^j \wedge \theta^k \\
& &+ \sum_{j<k} \sum_{l} f_l C_{jkl} \theta^j \wedge \theta^k\\
&=&
\sum_{j<k} \sum_{l} f_l C_{jkl} \theta^j \wedge \theta^k =
{\tilde i}_{\nabla} \def\e{\epsilon} \def\la{\langle f}C.
\end{eqnarray*}
\end{proof}
\begin{lem}\label{closedform}
$\omega}\def\d{\delta}\def\vp{\varphi$ is a closed $2$-form, i.e.,
$d\omega}\def\d{\delta}\def\vp{\varphi = 0$.
\end{lem}
\begin{proof}
Choose a local frame $\{E_i\}$ with $E_1 = N = {\nabla} \def\e{\epsilon} \def\la{\langle f}/{|\nabla} \def\e{\epsilon} \def\la{\langle f|}$, and let $\{\theta^i\}$ be its dual coframe.
Then, by Lemma~\ref{simple} and Lemma~\ref{lem191}
$$
di_{\nabla} \def\e{\epsilon} \def\la{\langle f}z =
\sum_{j<k} \sum_{l} f_l C_{jkl} \theta^j \wedge \theta^k
= \sum_{j<k} |\nabla} \def\e{\epsilon} \def\la{\langle f| C_{jk1} \theta^j \wedge \theta^k
=
|\nabla} \def\e{\epsilon} \def\la{\langle f| \sum_{k=2}^n C_{1k1} \theta^1 \wedge \theta^k.
$$
Thus, by taking the exterior derivative of $\omega$ in (\ref{keyeqn}), we have
$$
d\omega}\def\d{\delta}\def\vp{\varphi = - df \wedge d i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z = - |\nabla} \def\e{\epsilon} \def\la{\langle f| \theta^1 \wedge
\left( |\nabla} \def\e{\epsilon} \def\la{\langle f| \sum_{k=2}^n C_{1k1} \theta^1 \wedge \theta^k\right) =0.
$$
\end{proof}
Now, let $\Omega =\{p\in M \, \vert\, \omega_p \neq 0$ on $T_pM\}$. Then, $\Omega$ is an open subset of $M$.
We start with the following observation.
\begin{lem} \label{compr2}
Suppose that $\omega_p\neq 0$ at $p\in M$. Then,
\begin{eqnarray}
|D\omega |^2(p)\geq |\delta \omega|^2(p).\label{eqn2019-12-25-1}
\end{eqnarray}
\end{lem}
\begin{proof}
First of all, since $\omega_p\neq 0$, we have $df_p\neq 0$ and $f(p)\neq -1$
by definition of $\omega$ and Lemma~\ref{lem2018-4-30-20}.
Define $A:T_pM\to T_pM$ by $ g(Au, v)=\omega (u, v)$ for any $u, v\in T_pM$.
\vskip .5pc
\noindent
{\bf Assertion 1}: $\nabla f(p)\notin \ker A$.
Let $\{e_1, e_2, \cdots, e_n\}$ be an orthonormal basis on $T_pM$ with
$e_1=N(p)$. If $\nabla f(p)\in \ker A$, then,
$$
0=\langle Ae_1, e_j\rangle =\omega(e_1, e_j)=-(1+f)\, \tilde{i}_{\nabla f}C(e_1, e_j)
$$
for $j\geq 2$. This implies that $\omega_p=0$ from Lemma~\ref{simple}, a contradiction.
\vskip .5pc
\noindent
{\bf Assertion 2}: $\ker A\subset (\nabla f)^{\bot}$.
Let $u\in \ker A$ so that $\langle Au, v\rangle =0$ for any $v\in T_pM$.
Let $\{e_1, e_2, \cdots, e_n\}$ be an orthonormal basis on $T_pM$ with $e_1=N(p)$.
Then, by Lemma~\ref{lem2018-4-30-20} together with (\ref{eqn2019-5-26-1}), we have
\begin{eqnarray*}
0&=& \langle Au, e_k\rangle = \omega(u, e_k)= -(1+f)\, \tilde{i}_{\nabla f}C(u, e_k)\\
&=& -(1+f) \, \sum_{j=1}^n \langle u, e_j\rangle \, \tilde{i}_{\nabla f}C(e_j, e_k)= -(1+f) \, \langle u, e_1\rangle \, \tilde{i}_{\nabla f}C(e_1, e_k)
\end{eqnarray*}
for any $1\leq k\leq n$. Since $\omega}\def\d{\delta}\def\vp{\varphi_p \ne 0$, we have
$\tilde{i}_{\nabla f}C(e_1, e_k) \ne 0$ for some $k\ge 2$ by Lemma~\ref{simple}.
So, $\langle u, e_1\rangle =0$, which implies that $\nabla} \def\e{\epsilon} \def\la{\langle f(p) \in (\ker A)^\perp$.
\vskip .5pc
\noindent
{\bf Assertion 3}: Let $\{e_1,... , e_n\}$ be an orthonormal basis on $T_pM$
with $e_1=N(p)$. Then,
$$
||Ae_1||=\sup_{\substack{u\in (\ker A)^{\bot}\\ ||u||=1}} ||Au||.
$$
First, by {\bf Assertion 1} and {\bf Assertion 2},
$Ae_1 \ne 0$ and $e_1 \in (\ker A)^{\bot}$.
Since $ g(Ae_1, e_1) = \omega(e_1, e_1)=0$, $Ae_1$ is orthogonal to $e_1$, we may assume that $e_2= {Ae_1}/{||Ae_1||}$.
In particular, we have $e_2 \in (\ker A)^\perp$ by the skew-symmetry of $A$.
Le $u\in (\ker A)^{\bot}$ with $||u||=1$.
Since $ \langle Ae_i, e_j\rangle =\omega (e_i, e_j)=0$ for $i,j\geq 2$
by (\ref{eqn2019-5-26-1}), we have
\begin{eqnarray*}
Au&=& \sum_{j=1}^n\langle Au, e_j\rangle e_j=\sum_{j=1}^n \sum_{i=1}^n \langle u, e_i\rangle \, \langle Ae_i, e_j\rangle e_j\\
&=& \langle u, e_1\rangle Ae_1 -\sum_{i=1}^n\langle u,e_i\rangle \langle e_i, Ae_1\rangle e_1\\
&=&
\langle u, e_1\rangle Ae_1 - \langle u, Ae_1\rangle e_1\\
&=&
||Ae_1||\langle u, e_1\rangle e_2 - ||Ae_1||\langle u, e_2\rangle e_1.
\end{eqnarray*}
Since $||u||^2= \sum_j \langle u, e_j\rangle^2=1,$ we have
$$
||Au||^2= ||Ae_1||^2 (\langle u, e_1\rangle ^2 +\langle u, e_2\rangle^2 ) \leq ||Ae_1||^2
$$
and hence {\bf Assertion 3} is satisfied.
Note that, for $u=e_j$ with $j\geq 2$, we also have
\begin{eqnarray*}
Ae_j = -\langle e_j, Ae_1\rangle e_1 \quad\mbox{and so}\quad
||Ae_j ||\leq ||Ae_1||.\label{compr}
\end{eqnarray*}
\vspace{.102in}
Now, let us show the inequality (\ref{eqn2019-12-25-1}).
Applying the argument in the proof of Lemma 2.2 from \cite{pz}, we may conclude
that there exists a local frame $\{E_1,\cdots, E_{2m}, \cdots, E_n\}$ around a point $p \in M$ such that
$$
\omega= \sum_{i=1}^m a_i \theta^{2i-1} \wedge \theta^{2i},
$$
where $\{ \theta^1,..,\theta^n\}$ is the dual coframe of $\{E_i\}_{i=1}^n$.
In particular, since $\nabla f(p)\notin \ker A$, we may choose
$E_1=N=\nabla f/||\nabla f||$.
Since $\omega(E_j,E_k)= 0$ for $j,k\geq 2$ by Lemma~\ref{simple}, we have
$$
\omega= u \, \theta^1 \wedge \theta^2
$$
for a local smooth function $u$.
Thus, it is easy to see (cf. \cite{pl1}, p.25) that
$$\delta \omega=E_2(u)\theta^1 -E_1(u)\theta^2$$
and so,
$$|\delta \omega|^2 =(E_1(u))^2+(E_2(u))^2.$$
On the contrary, from $\omega= df \wedge i_{\nabla f}z
= u\theta^1 \wedge \theta^2$, we have
$$
\omega (E_1, E_j)= |\nabla f|z(\nabla f, E_j)=0\quad (j \ge 3).
$$
Thus,
\begin{eqnarray*}
D_{E_1}\omega (E_1,E_2)&=&E_1(\omega(E_1,E_2))-\omega(D_{E_1}E_1, E_2)- \omega(E_1, D_{E_1}E_2)\\
&=&
E_1(u) -\sum_{j=3}^n\langle D_{E_1}E_1, E_j\rangle\, \omega (E_j,E_2)
- \sum_{j=3}^n \langle D_{E_1}E_2, E_j\rangle\, \omega(E_1, E_j)\\
&=& E_1(u).
\end{eqnarray*}
Similarly, since $D_{E_1}\omega (E_1,E_2) = E_2(u)$, we may conclude that
$$
|\delta \omega|^2=(E_1(u))^2+(E_2(u))^2 =|D_{E_1}\omega(E_1,E_2)|^2
+|D_{E_2}\omega (E_1,E_2)|^2\leq |D\omega|^2.
$$
\end{proof}
Using Lemma~\ref{compr2} and Bochner-Weitzenb\"ock formula for $2$-forms, we can prove
the following structural property for the CPE with positive isotropic curvature.
\begin{thm}\label{thm2018-5-16-2}
Let $(g,f)$ be a nontrivial solution of (\ref{cpe}) on a compact manifold $M$ of dimension $n \ge 4$.
If $(M,g)$ has positive isotropic curvature, then the $2$-form $\omega}\def\d{\delta}\def\vp{\varphi = df\wedge i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z$ is vanishing.
\end{thm}
\begin{proof}
It suffices to prove that $\Omega =\emptyset$, where $\Omega =\{p\in M \, \vert\, \omega_p \neq 0$ on $T_pM\}$.
Suppose, on the contrary, $\Omega \neq \emptyset$. For $p\in \Omega$, let $\Omega_0$ be a
connected component of $\Omega$ containing $p$.
Note that $ \Delta \omega = -d \delta \omega$ by Lemma~\ref{closedform}.
It follows from the Bochner-Weitzenb\"ock formula for $2$-forms
(cf. \cite{pl1}, \cite{wu}) that
\begin{eqnarray} \frac 12 \Delta |\omega |^2 =\langle \Delta \omega , \omega \rangle +|D\omega|^2 +\langle E(\omega), \omega\rangle,
\label{bwf}\end{eqnarray}
where $E(\omega)$ is a (local) $2$-form containing isotropic curvature terms as its coefficients.
In particular, if $(M,g)$ has positive isotropic curvature, following the Proposition 2.3 in \cite{pz}
(cf. \cite{m-w}, \cite{sea}) we have
\begin{eqnarray} \langle E(\omega), \omega\rangle >0. \label{piccu}
\end{eqnarray}
Therefore, integrating (\ref{bwf}) over $\Omega_0$, we obtain
$$ \frac 12 \int_{\Omega_0} \Delta |\omega|^2 =\int_{\Omega_0} (\langle \Delta \omega, \omega \rangle + |D\omega |^2 )+\int_{\Omega_0} \langle E(\omega), \omega\rangle.
$$
Since $\omega =0$ on the boundary $\partial \Omega_0$ and $\omega$ is a closed form by Lemma~\ref{closedform}, we have
$$ 0=\int_{\Omega_0} |D\omega |^2 -|\delta \omega|^2 +\langle E(\omega), \omega\rangle.
$$
However, by Lemma~\ref{compr2} and the inequality (\ref{piccu}), the above equation is impossible if $\omega$ is nontrivial. Hence, we may conclude that $\omega =0$, or $\Omega_0=\emptyset$.
\end{proof}
\begin{thm} \label{thm1-1}
Let $M$ be an $n$-dimensional compact smooth manifold with $n\geq 4$.
If $(g,f)$ is a nontrivial solution of (\ref{cpe}) and $(M, g)$ has positive isotropic curvature,
then $(M,g)$ is isometric to a standard sphere.
\end{thm}
\section{Appendix}
In the Appendix, we first claim that $-1$ is a regular value of the potential function $f$
when ${\rm Ric}(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for any vector field $X$ orthogonal to $\nabla} \def\e{\epsilon} \def\la{\langle f$ unless $(M, g)$ is Einstein.
Second, we prove (\ref{eqn2020-9-28-1}) and (\ref{eqn2020-9-28-2}) in Remark~\ref{rem2020-9-28-3}.
\subsection{regularity of $f$ on the set $f^{-1}(-1)$}
Define
$$
\vp:= \frac{1}{2}|\nabla} \def\e{\epsilon} \def\la{\langle f|^2 + \frac{s}{2n(n-1)}f^2.
$$
Then we have,
$$
N(\vp) = (1+f)\alpha |\nabla} \def\e{\epsilon} \def\la{\langle f|
$$
and
$$
\Delta \vp = (1+f)^2 |z|^2 + |\nabla} \def\e{\epsilon} \def\la{\langle f|^2 \alpha.
$$
Therefore,
\begin{eqnarray}
\Delta \vp - \frac{|\nabla} \def\e{\epsilon} \def\la{\langle f|}{1+f} N(\vp) = (1+f)^2 |z|^2 \ge 0. \label{eqn2020-10-8-1}
\end{eqnarray}
By the maximum principle,
$$
\max_{f\le -1-\e} \vp = \max_{f=-1-\e} \vp
$$
and
$$
\max_{f \ge -1+\e}\vp = \max_{f = -1+\e}\vp
$$
for sufficiently small $\e>0$. So, letting $\e \to 0$, we have
\begin{eqnarray}
\max_M \vp = \max_{f=-1} \vp.\label{eqn2020-10-8-2}
\end{eqnarray}
Now, as mentioned in Introduction, if $\min_M f \ge -1$, then $(M, g)$ is Einstein,
and so we may assume that $\min_M f <-1$. Let $\min_{x\in M} f(x) = f(x_0)$. Then
for any point $p \in f^{-1}(-1)$, we have
$$
\vp(p) = \frac{1}{2}|\nabla} \def\e{\epsilon} \def\la{\langle f|^2(p) + \frac{s}{2n(n-1)} \ge \vp(x_0) = \frac{s}{2n(n-1)}[f(x_0)]^2 >
\frac{s}{2n(n-1)},
$$
which proves our claim.
\subsection{warped product metric}
\begin{lem}\label{lem7}
Suppose that $z(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = 0$ for $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$.
Then,
\begin{itemize}
\item[(1)] for vectors $X, Y$ orthogonal to $\nabla} \def\e{\epsilon} \def\la{\langle f$,
$$
i_{\nabla} \def\e{\epsilon} \def\la{\langle f}T(X, Y) = \frac{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}{n-2}\left(z+\frac{\alpha}{n-1}g\right)(X, Y).
$$
\item[(2)] $i_{\nabla} \def\e{\epsilon} \def\la{\langle f}T(\nabla} \def\e{\epsilon} \def\la{\langle f, X) = i_{\nabla} \def\e{\epsilon} \def\la{\langle f}T(X, \nabla} \def\e{\epsilon} \def\la{\langle f) = 0$ for any vector $X$.
\end{itemize}
\end{lem}
\begin{proof}
If $\omega}\def\d{\delta}\def\vp{\varphi= 0$, then $i_{\nabla} \def\e{\epsilon} \def\la{\langle f}z = \alpha df$ and hence
$$
T = \frac{1}{n-2} df\wedge \left(z+\frac{\alpha}{n-1}g\right).
$$
\end{proof}
It follows from Lemma~\ref{lem2019-5-23-10} that
\begin{eqnarray}
-|\nabla} \def\e{\epsilon} \def\la{\langle f|^2 \mathcal W_N = (1+f)i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C + (n-1)i_{\nabla} \def\e{\epsilon} \def\la{\langle f}T.\label{eqn2019-5-28-1}
\end{eqnarray}
For the Weyl curvature tensor $\mathcal W$, $\mathcal W_N$ can be similarly defined as $R_N$.
\begin{lem} \label{lem2018-8-15-4}
Let $(g, f)$ be a non-trivial solution of the CPE with $\omega}\def\d{\delta}\def\vp{\varphi = 0$. Then,
$$
\frac{s}{n(n-1)} g = R_N +z +\frac{1+f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C- \left(\alpha - \frac{s}{n(n-1)}\right) \frac{df}{|df|}\otimes \frac{df}{|df|}.
$$
\end{lem}
\begin{proof}
Let
$$
\Phi:= \frac{s}{n(n-1)}g - z - \frac{1+f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C.
$$
For vector fields $X, Y$ with $X \perp \nabla} \def\e{\epsilon} \def\la{\langle f$ and $ Y \perp \nabla} \def\e{\epsilon} \def\la{\langle f$,
from the curvature decomposition
$$
R = \frac{s}{2n(n-1)}g\owedge g + \frac{1}{n-2}z\owedge g + \mathcal W,
$$
we obtain
\begin{eqnarray*}
R_N(X, Y) = \frac{s}{n(n-1)}g(X, Y) + \frac{1}{n-2} z(X, Y) + \frac{\alpha}{n-2} g(X, Y)
+\mathcal W_N(X, Y)
\end{eqnarray*}
Since, by Lemma~\ref{lem7} together with (\ref{eqn2019-5-28-1}),
\begin{eqnarray*}
\mathcal W_N(X, Y) =
- \frac{1+f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C(X, Y) - \frac{n-1}{n-2}z(X, Y) - \frac{\alpha}{n-2}g(X, Y),
\end{eqnarray*}
we have
\begin{eqnarray}
R_N(X, Y) = \Phi(X, Y).\label{eqn2019-5-28-2}
\end{eqnarray}
Now, let $X$ and $Y$ be arbitrary tangent vector fields. Then, $X$ and $Y$ can be decomposed into
$$
X = X_1 + \langle X, N\rangle N,\quad Y = Y_1 +\langle Y, N\rangle N
$$
with $\langle X_1, N\rangle =0 =\langle Y_1, N\rangle.$
Thus,
\begin{eqnarray*}
R_N(X, Y) &=&
R_N(X_1, Y_1)= \Phi (X_1, Y_1)\\
&=&
\Phi(X, Y) - \langle X, N\rangle \langle Y, N\rangle \Phi(N, N)\\
&=&
\Phi(X, Y) + \left(\alpha -\frac{s}{n(n-1)} \right) \frac{df}{|df|}\otimes \frac{df}{|df|}(X, Y).
\end{eqnarray*}
\end{proof}
\begin{lem}
On the set $f^{-1}(-1)$, we have
\begin{eqnarray*}
{\mathcal L}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \left(\frac{h}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}\right) = 0.
\end{eqnarray*}
Here, recall that $h:= R_N +\left(z- \alpha \frac{df}{|df|}\otimes \frac{df}{|df|}\right)
+\frac{1+f}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}i_{\nabla} \def\e{\epsilon} \def\la{\langle f}C.$
\end{lem}
\begin{proof}
First, from CPE, we have
$$
\frac{1}{2}{\mathcal L}_{\nabla} \def\e{\epsilon} \def\la{\langle f} g = Ddf = (1+f)z - fh - \e f \frac{df}{|df|}\otimes \frac{df}{|df|}
$$
with $\epsilon = \frac{s}{n(n-1)}$. It follows from (\ref{eqn2020-5-7-3}), (\ref{eqnt2}) and (\ref{eqn2020-9-28-10}) that
\begin{eqnarray*}
\frac 1{2\e} {\mathcal L}_{\nabla f} h = (1+f) z - f h - (1+f)\alpha \frac {df}{|\nabla f|} \otimes \frac {df}{|\nabla f|}.
\end{eqnarray*}
Thus,
\begin{eqnarray*}
\frac{1}{2}{\mathcal L}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \left(\frac{h}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}\right)
&=&
\frac{1}{2}\nabla} \def\e{\epsilon} \def\la{\langle f (|\nabla} \def\e{\epsilon} \def\la{\langle f|^{-2}) h + \frac{1}{2|\nabla} \def\e{\epsilon} \def\la{\langle f|^2} {\mathcal L}_{\nabla} \def\e{\epsilon} \def\la{\langle f} h\\
&=&
-N(|\nabla} \def\e{\epsilon} \def\la{\langle f|) \frac{h}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2} + \frac{1}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2} \frac{1}{2}{\mathcal L}_{\nabla} \def\e{\epsilon} \def\la{\langle f} h.
\end{eqnarray*}
In particular, on the set $f^{-1}(-1)$, we have
$$
\frac{1}{2}{\mathcal L}_{\nabla} \def\e{\epsilon} \def\la{\langle f} \left(\frac{h}{|\nabla} \def\e{\epsilon} \def\la{\langle f|^2}\right) = 0.
$$
\end{proof}
|
\section{Introduction}
Intelligent transportation \cite{5959985} has been receiving increasing attention recently, and for the applications, such as assisted driving, violation detection, and congestion forecasting, accurate and efficient cognition and reasoning over the traffic events captured by video cameras is extremely important. As shown by previous works \cite{10.3115/1073012.1073017,35179}, well-designed datasets are often crucial for the development, adaptation and evaluation of different data-driven approaches.
This indicates the significance of creating comprehensive and challenging benchmarks for video causal reasoning
and cognitive development of models,
that explore the underlying causal structures of various traffic events. To this end, we introduce a novel dataset, SUTD-TrafficQA (Traffic Question Answering), to facilitate the research of causal reasoning in complex traffic scenarios.
In our dataset, to help develop models for addressing several major and concerning issues in intelligent transportation, we design 6 challenging reasoning tasks, which require exploring the complex causal structures within the inference process of the traffic events.
As shown in Figure \ref{fig:example},
these tasks correspond to various traffic scenarios involving both road-agents and surroundings, and the models are required to forecast future events, infer past situations, explain accident causes, provide preventive advice, and so on.
To present these reasoning tasks, video question answering \cite{zhu2017uncovering} is a natural and effective choice, and is used for our dataset construction, since to accurately answer the given questions, the models need to acquire strong capabilities of performing various levels of logical reasoning and spatio-temporal cognition for the events.
Besides providing the challenging and useful reasoning tasks,
we adopt a combination scheme of online collection and offline capturing to collect videos, such that the data in our benchmark covers various traffic events, diversified road-agents and surroundings, and different capturing perspectives in the wild. With the provided various tasks and diverse videos, our dataset shall be able to serve as a comprehensive benchmark for video reasoning of traffic events.
In some application scenarios, (e.g., assisted drivin
), the computational resource and energy budget can be constrained.
Thus
both the inference accuracy and the computation efficiency are important for video event reasoning in these scenarios.
Existing video QA methods \cite{Kim_2020_CVPR,Le_2020_CVPR,lei-etal-2020-tvqa} mainly focus on strengthening the reasoning accuracy without emphasizing much efficiency, and most of
these works apply fixed computation pipelines
to answer different questions, while ignoring to conduct adaptive and efficient computation resource allocation based on the logic structure behind reasoning over video events.
In this paper, to achieve reliable and efficient video reasoning, we propose \textbf{Eclipse}, an \textbf{E}ffi\textbf{c}ient g\textbf{li}m\textbf{pse} network.
Specifically, considering there is often large redundancy among video frames, via dynamic inference, our network adaptively determines where to skip and glimpse at each step, and what computation granularity needs to be allocated for the glimpsed frame. Such a dynamic reasoning scheme avoids feature extraction for the irrelevant segments in the video, and hence significantly reduces the overall computation cost towards
reliable and efficient reasoning.
It is noteworthy that both the determination of selecting a glimpse frame and the decision of computation granularity for each glimpse are essentially discrete operations, which are
not trivial to optimize. To handle this issue, an effective joint Gumbel-Softmax mechanism is also introduced in this paper, which makes our Eclipse framework fully differentiable and end-to-end trainable.
To the best of our knowledge, this is the first work that simultaneously performs adaptive frame localization and feature granularity determination in a novel dynamic reasoning process for reliable and efficient causal reasoning and video QA. A joint Gumbel-Softmax operation is also introduced in this work to optimize the two decisions jointly.
\section{Related Works}
\textbf{Intelligent Transportation.}
With the rapid development of deep learning techniques \cite{krizhevsky2012imagenet, ren2015faster, vaswani2017attention}, data-driven intelligent transportation \cite{chandra2020forecasting,lou2019veri,you2020traffic,chandra2020forecasting} has emerged as a prominent research topic.
Lou et al. \cite{lou2019veri} presented a dataset together with an adversarial learning model for vehicle re-identification.
Different from existing intelligent transportation datasets and methods,
in this paper, we investigate the problem of causal reasoning with video QA over various traffic scenarios. A new benchmark, SUTD-TrafficQA, together with a novel model, Eclipse, is proposed for this challenging task.
\begin{table*}[t]
\caption{Comparison among SUTD-TrafficQA and some other video QA datasets. Providing challenging \textbf{traffic-scenario reasoning tasks} with \textbf{real-world videos} and \textbf{human-generated QA pairs}, our dataset shall serve as a comprehensive and challenging benchmark for video reasoning over traffic events.}
\vspace{-0.2cm}
\begin{center}
\scriptsize
\setlength\tabcolsep{0.5pt}
\begin{tabular}{c|c|ccccccc} \hline
\multirow{2}{*}{\textbf{Dataset}} & \textbf{Synthetic Videos} & \multicolumn{7}{c}{\textbf{Real-World Videos}} \\
\cline{2-9}
& CLEVRER \cite{yi2019clevrer}
& MovieQA \cite{tapaswi2016movieqa} & MSRVTT-QA \cite{xu2017video} & TGIF-QA \cite{jang2017tgif} & TVQA \cite{lei2018tvqa} & MarioQA \cite{mun2017marioQA} & Social-IQ \cite{zadeh2019social} & \textbf{SUTD-TrafficQA (Ours)} \\ \hline
Basic Understanding & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & \textbf{$\checkmark$} \\
Attribution & $\checkmark$ & $\checkmark$ & $\times$ & $\times$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & \textbf{$\checkmark$} \\
Event Forecasting & $\checkmark$ & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & \textbf{$\checkmark$} \\
Reverse Reasoning & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & \textbf{$\checkmark$} \\
Counterfactual Inference & $\checkmark$ & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & \textbf{$\checkmark$} \\
Introspection & $\times$ &$\times$ & $\times$& $\times$ & $\times$ & $\times$ & $\times$ & \textbf{$\checkmark$} \\
\hline
QA Generation & Automatic & Human & Automatic & Automatic \& Human & Human & Automatic & Human & Human \\ \hline
Topic & Synthetic Object Collision & Movies & Various Scene & GIFs & TV-Shows & Gameplay & Social Behavior & Traffic Events\\ \hline
\end{tabular}
\end{center}
\label{table:dataset_comparison}
\vspace{-0.5cm}
\end{table*}
\textbf{Video QA Datasets.}
Recently, there emerges a great interest in visual reasoning and question answering in videos \cite{10.1007/s11263-016-0987-1,maharaj2017dataset,zhu2017uncovering,ijcai2017-280}, and several video QA datasets \cite{xu2017video,zeng2016leveraging,tapaswi2016movieqa,zadeh2019social,jang2017tgif,garcia2020knowit,emrvqasongMM18,yu2019activityqa} have been developed. Among them,
MovieQA \cite{tapaswi2016movieqa} and TVQA \cite{lei2018tvqa} present the movie and TV-show videos respectively with human-generated questions. More recently,
CLEVRER \cite{yi2019clevrer} focuses on collision event reasoning among several simple visual objects
in a controlled environment using fully synthetic videos.
Differently, our SUTD-TrafficQA focuses on reasoning over the complex traffic scenarios in the wild,
where 6 challenging tasks for traffic event reasoning are introduced based on the diverse real-world
traffic
videos.
\textbf{Video QA Methods.}
Extensive studies have been conducted for video QA \cite{yu2017end,jang2017tgif,garcia2020knowledge,Jiang_2020_CVPR,Kim_2020_CVPR,lei-etal-2020-tvqa,tsai2019GSTEG,li2019beyond,Kim_2018_ECCV,Yu_2018_ECCV,sukhbaatar2015end,wang2018movie,fan2019heterogeneous,liang2018focal,tensoremnlp17,8654010}.
Yu et al. \cite{yu2017end} employed LSTM to encode videos and QA pairs, and adopted an attention mechanism \cite{you2016image}.
Jang et al. \cite{jang2017tgif} used LSTMs with a different attention scheme to capture spatio-temporal patterns in videos.
Different from existing video QA methods, our Eclipse model investigates the direction of learning an effective glimpse policy for adaptive reasoning to achieve reliable reasoning with computation efficiency.
\textbf{Computation-Efficient Models.}
Recent works \cite{Bhardwaj_2019_CVPR,Mu_2019_CVPR,strubell2019energy,figurnov2017spatially, korbar2019scsampler,wu2019adaframe,schwartz2019green,NEURIPS2019_bd853b47, fan2020adaptive} have pointed out the need of improving the computation-efficiency when designing deep models, and
different strategies,
including filter pruning \cite{li2016pruning}, weight sparsification \cite{sun2016sparsifying}, vector quantization \cite{agustsson2017soft}, and dynamic routing \cite{wang2018skipnet}, etc., have been proposed.
In this paper, we
propose an efficient model, Eclipse, the first model that performs dynamic inference and adaptive computation adjustment for video QA, which leverages an effective glimpse policy with the guidance of text and visual context information for both glimpse frame selection and computation granularity determination.
\section{SUTD-TrafficQA Dataset}\label{section:dataset}
Our dataset contains 62,535 QA pairs and 10,080 videos of traffic scenes. Below we first propose 6 challenging traffic-related reasoning tasks, and then introduce the QA collection process and the dataset statistics.
{\bf Basic understanding.} This task evaluates the ability of the models in perceiving and understanding traffic scenarios at the basic level, which consists of multiple sub-tasks including feature-query (e.g., vehicle type, road situation, and environment description), event-query (e.g., accident existence, pedestrian action analysis, and events temporal relation), event classification (e.g., accident type), and counting (e.g., road-agent number).
{\bf Event forecasting.} This task requires a model to infer future events based on observed videos, and the forecasting questions query about the outcome of the current situation.
{\bf Reverse reasoning.} This task is to ask about the events that have happened before the start of a video segment.
{\bf Counterfactual inference.} This task queries the consequent outcomes of certain hypothesis (e.g., what if the blue sedan had not accelerated?). The hypothetical conditions do not occur in the video, so the model needs to reason about the imagined events under the designated condition.
{\bf Introspection.} This task is to test if models are able to provide preventive advice (e.g., what could the pedestrian have done to avoid the collision with the car?).
The candidate answers list actions that could have been taken to avoid traffic accidents or congestion.
{\bf Attribution.} This task seeks the explanation about the causes of traffic events (e.g., what are the reasons of the rear-end crash?), so as to check if models are able to infer the underlying factors leading to the event.
We define all the above reasoning tasks as multiple-choice questions without limiting the number of candidate answers for each question. The number of candidate answers varies from 2 to 12 for different questions.
We then sample among the candidate answers to balance the dataset
and limit the occurrence of the same correct answers within each task to minimize language biases. As summarized in Table \ref{table:dataset_comparison}, by introducing these challenging tasks, our dataset complements existing datasets, and facilitates the exploration of video QA in complex traffic scenarios.
\subsection{QA Collection}
{\bf Videos.} We collected videos by using a combination of online harvesting and offline capturing, to cover various real-world traffic scenarios.
As for online video collection, a variety of video sharing platforms, based in different countries, are used to increase the diversity, including but not limited to YouTube, LiveLeak, Twitter, and Bilibili. More than nine thousand videos were thus collected from these online sources.
As for offline video capturing, a set of videos were captured via handheld cameras by volunteers, while another set were fetched from car-mounted video recorders. These two sets of offline videos were then examined and trimmed into around one thousand video clips.
After combing the online videos and offline captured ones, we obtain a total of $10,080$ videos containing diversities in various aspects, including: a) different weathers (sunny/rainy/windy/snowy); b) different time (daytime/night); c) diverse road situations (congested/sparse, urban/rural roads); d) various traffic events (accidents, vehicle turning, pedestrian behaviors, traffic lights, etc.); e) different video perspectives (surveillance camera perspective/car-mounted video perspective/hand-held camera perspective); and f) various clip lengths (from $1$ to $70$ seconds).
{\bf QA pairs.} As the first step, the 6 tasks were explained to annotators. To ensure they fully understand the design principle of the tasks, multiple example questions were prepared for them to identify which task the questions belong to. Afterwards, each annotator was presented with batches of video folders, where each folder contains 100 clips that were randomly selected from the full video set. Annotators were asked to create at least 3 questions of the reasoning tasks for each video.
We did not impose constraints on question formats to encourage annotators to keep their QA pairs diversified. Specifically, we encourage them to rephrase similar questions or candidate answers to push the models to learn underlying semantics of QA pairs rather than superficial language correlation.
To ensure the quality of QA pairs,
we cross-checked QA annotations on a weekly basis.
In addition, we kept monitoring the distribution of different tasks in QA collection to maintain the balance and diversity of our dataset.
\subsection{Dataset Statistics}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1\linewidth]{stats2.pdf}
\end{center}
\vspace{-0.4cm}
\caption
Statistics of SUTD-TrafficQA. More in supplementary.
}
\label{fig:Dataset_stats}
\vspace{-0.2cm}
\end{figure}
In this part, we present the statistics of our SUTD-TrafficQA dataset. Figure \ref{fig:Dataset_stats} (a) demonstrates the distribution of question length measured by number of words. The average length of questions is $8.6$ words. Figure \ref{fig:Dataset_stats} (b) shows the various question types categorized by their beginning words, which implies the diversity of questions in our dataset.
Figure \ref{fig:Dataset_stats} (c) presents the split of questions in terms of reasoning tasks.
Our dataset covers a broad range of traffic-related reasoning tasks requiring various levels of spatio-temporal understanding and causal reasoning in videos.
\section{Eclipse Network}
To deal with video reasoning, a common solution is to watch the full video and analyze the whole event information carefully. In this manner, generally, a fixed computation architecture \cite{lei2018tvqa,jang2017tgif,arad2018compositional,Le_2020_CVPR} can be applied over the whole video for tackling each question.
However, using a fixed network architecture for handling the video QA task is often computation-heavy and energy-consuming \cite{strubell2019energy,schwartz2019green}, because the video sequence used for reasoning can be very long and contain plenty of frames for processing.
Recalling that, as humans,
to analyze events in a video, we may not be patient enough to scrutinize the frames in the whole video,
instead, we may
adopt an adaptive information ``foraging'' strategy \cite{fitzsimmons2014skim,duggan2011skim}.
Concretely, we may
use a ``dynamic'' inference manner
to skip forth and back over the sequence to progressively infer and select some useful frames based on the task.
Moreover, for the picked frames, we may examine a few of them very carefully while glancing over others.
Such a dynamic and adaptive perception habit \cite{fitzsimmons2014skim,duggan2011skim}
frees us from watching the whole video thoroughly, and often enables fast yet still very accurate video reasoning at a very small frame usage.
Motivated by this, we aim to explore the direction of efficient and dynamic reasoning in complex traffic scenarios. Thus, we propose an \textbf{E}ffi\textbf{c}ient g\textbf{li}m\textbf{pse} (\textbf{Eclipse}) network for video QA, as illustrated in Figure \ref{fig:Eclipse}.
Instead of using a fixed computation architecture over each video and question, our network learns to dynamically skip to and select a useful video frame at each inference step. Moreover, our network adaptively decides the feature computation granularity (i.e., coarse or fine) of the selected frame. To perform such a process, at each inference step, our network takes advantage of the guidance information including the QA pair, the currently selected frame and the historical cues for dynamic reasoning.
Specifically, as shown in Figure \ref{fig:Eclipse},
in our network, to provide QA information for the Interaction Module, the QA Bank stores the representation of the QA pairs.
At each inference step, to assist the dynamic reasoning of selecting the frame and the corresponding feature granularity, the Interaction Module leverages the QA information, the currently selected frame and the information from historically observed frames to derive an expressive representation,
which then serves as the input of the dynamic reasoning process performed by the downstream modules, including the Prediction Module for outputting the reasoning result, and
Glimpse-Determination Module for dynamically determining which frame to be observed at next step.
Besides, the Exit-Policy Module also uses this representation to adaptively decide whether we can exit the reasoning process at current inference step.
Via such a dynamic and recurring reasoning process, our network can derive reliable answer predictions with notable computation efficiency w.r.t. both the frame usage and feature computation granularity. We elaborate the network modules in detail below.
\begin{figure}[tbp]
\begin{center}
\includegraphics[width=1\linewidth]{Eclipse_1_v1_3.pdf}
\end{center}
\vspace{-0.5cm}
\setlength\abovecaptionskip{-2cm}
\setlength\belowcaptionskip{-2cm}
\caption{Architecture of Eclipse for dynamic causal reasoning.}
\label{fig:Eclipse}
\vspace{-0.4cm}
\end{figure}
{\bf QA Bank.} To provide the QA information as the guidance for dynamic reasoning, our QA Bank encodes the representation of the question via a bi-directional LSTM.
By concatenating hidden states of the BiLSTM, the question representation can be denoted as $H^q\in\mathbb{R}^{n_q\times2d}$, where $n_q$ is the number of words in the question, $d$ denotes the dimension of LSTM hidden state. Similarly, we encode all the candidate answers as
${\{ H^{a_i} \}}^N_{i=1}$,
where $H^{a_i}\in\mathbb{R}^{n_{a_i}\times2d}$,
$N$ is the number of candidate answers,
and $n_{a_i}$ is the number of words in the answer $a_i$.
{\bf Interaction Module.} To assist the subsequent dynamic reasoning with rich guidance information, we design an Interaction Module to fuse different kinds of available inputs. This module, consisting of a context-query sub-module and an interaction LSTM, recurrently interacts with a small number of frames selected from the full video sequence containing $T$ frames. As shown in Figure \ref{fig:Eclipse}, at each inference step $t$, this module fuses the QA information ($H^q$ and $\{H^{a_i}\}$),
the currently selected frame ($I_t$), and the historical cues ($h_{t-1}$)
to produce an expressive representation ($h_t$) that can be used for dynamic reasoning.
More formally, at the $t^{th}$ inference step, to first fuse the textual QA information ($H^q$ and $\{H^{a_i}\}$) with the currently selected visual frame feature ($I_t$),
we use a context-query sub-module to perform such a fusion process. We implement this sub-module by following the context-matching module of \cite{lei2018tvqa}, which can effectively fuse the visual feature sequence with the textual sequence to produce a combined representation. Since the original method in \cite{lei2018tvqa} takes a sequence of visual features as an input, and we have only the selected frame feature $I_t$ here, we treat $I_t$ as a visual feature sequence containing a single element. Hence the fusion process in this sub-module can be formulated as:
\begin{equation}\label{eq.context-query}
v^i_t=[I_t;F^{I_t,q};F^{I_t,a_i};I_t\odot F^{I_t,q};I_t\odot F^{I_t,a_i}]
\end{equation}
where $\odot$ represents element-wise product. $F^{I_t,q}$ and $F^{I_t,a_i}$ are obtained by computing the similarity between the visual frame feature ($I_t$) and the textual features ($H^q$ and $H^{a_i}$). More details of this process are referred to \cite{lei2018tvqa} and also our supplementary.
For simplicity, we use $v_t$ to represent the concatenated ${\{v^i_t \}}^N_{i=1}$, as shown in Figure \ref{fig:Eclipse}. The
output $v_t$, that incorporates the information of currently selected frame and the QA embedding, can be fed into the following interaction LSTM for more information interaction, as introduced below.
Besides fusing the QA information and the currently selected frame ($I_t$), to guide the dynamic reasoning at current inference step, we also incorporate the historical cues from
past inference steps. Thus we design an interaction LSTM that takes the $v_t$ as the new input to interact with the historical cues $h_{t-1}$ as:
\begin{equation}\label{eq.lstm}
c_t, h_t = LSTM(v_t, c_{t-1}, h_{t-1};\theta_{LSTM})
\end{equation}
where $\theta_{LSTM}$ are parameters of LSTM. The generated hidden state $h_t$ encodes rich information of all available inputs and historical cues, and thus it can serve as an expressive representation to be fed into the downstream modules for dynamic reasoning as follows.
{\bf Prediction Module.} This module is used to generate the reasoning result (i.e. the probability distribution over candidate answers) at the current inference step $t$. This module computes the reasoning result as: $p_t=f_{p}(h_{t};\theta_{p})$, where $p_t\in\mathbb{R}^{N}$ represents the probability scores for all candidate answers. $f_{p}$ can be implemented with one fully-connected (FC) layer followed by a $Softmax$ classifier.
{\bf Glimpse-Determination Module.} At each inference step $t$, conditioned on $h_t$, this module performs dynamic reasoning by making two decisions simultaneously. The first decision is to select which frame to be observed at next step, and the second is to decide whether to compute fine-grained features or coarse features for this selected frame.
Corresponding to these two decisions, we design the following two branches within this module.
The skip-policy branch selects the frame that we need to skip to at next inference step via the following process: $s_t=f_{s}(h_{t};\theta_{s})$, where the output $s_t$ indicates the decision of the next frame location. Note that our network can skip forth and back over the entire video sequence, which is conceptually similar to the human reasoning process where we need to not only jump forward to find future informative frames but also go back to examine past information.
Besides determining next frame, this module also has a granularity-policy branch that adaptively decides the feature computation granularity for the next selected frame, formulated as:
$g_t=f_{g}(h_t;\theta_{g})$. The output $g_t$,
denotes the decision of feature granularity. In our implementation, we provide two kinds of feature granularity, namely, coarse features computed by a lightweight CNN; fine-grained features computed by a more representative yet computation-heavier CNN, to be chosen from. In the Glimpse-Determination module, both $f_{s}$ and $f_{g}$ are implemented with a FC layer.
{\bf Exit-Policy Module.} To estimate when we can exit the reasoning process to achieve adaptive inference, we design the Exit-Policy Module. At each inference step $t$, this module decides if we can exit the reasoning process at current step based on the guidance information ($h_t$) as: $e_{t}=f_{e}(h_{t};\theta_e)$, where the output $e_t$ denotes
the confidence score of terminating the reasoning at current step. By training the exit-policy,
our network can achieve adaptive inference, such that only a small and flexible number of frames are selected and computed on a per-video basis to derive reliable reasoning result.
{\bf Optimization.} To optimize the above modules in our network, we introduce several loss functions. Specifically, at each inference step $t$, a cross-entropy loss $\mathcal{L}^t_{pred}$ is used to train the classifier of the Prediction Module:
\begin{equation}\label{eq.cross-entropy loss}
\mathcal{L}^t_{pred}= -\sum_{n=1}^{N} y^n \log (p^{n}_{t})
\end{equation}
where $y$ is the ground-truth one-hot label vector for the candidate answers and $p^{n}_{t}$ is the predicted score for the $n^{th}$ answer.
As for Glimpse-Determination Module, to push the skip-policy branch to select a useful frame at each inference step, a simple yet effective loss is used:
\begin{equation}\label{eq.increment_loss}
\mathcal{L}^t_{incre} = -(m_{t} - m_{t-1})
\end{equation}
where $m_{t}=p_{t}^{gt}-\max \{ p_{t}^{c'}\mid c'\neq gt \}$ is the margin between the predicted probability of the correct answer (indexed by $gt$) and the largest probability of other candidate answers. We can simply infer that a larger $m_t$ indicates a more confident and accurate reasoning. Therefore, we use $m_{t}-m_{t-1}$ to encourage the margin to keep growing over the inference steps, which indicates that at each step, we aim to select a useful frame to benefit our dynamic reasoning, considering that the confidence of our network for the correct answer increases when seeing the selected frames.
Meanwhile, to further save computation cost and prevent the granularity-policy branch from constantly using the computation-heavy fine features, we penalize the feature granularity policy
when the fine feature is computed at each inference step $t$ as follows:
\begin{equation}\label{eq.feature_loss}
\mathcal{L}^t_{feat}= g_{t}
\end{equation}
where $g_t=1$ represents that the policy chooses to extract computation-heavy fine features for next step, while $g_t=0$ means it switches to extract computation-cheap coarse features for next step.
To optimize our whole Glimpse-Determination Module, we incorporate the above two loss functions, $\mathcal{L}^t_{incre}$ and $\mathcal{L}^t_{feat}$, into a combined loss:
\begin{equation}\label{eq.glimpse_loss}
\mathcal{L}^t_{glimpse}= \mathcal{L}^t_{incre} + \mathcal{L}^t_{feat}
\end{equation}
Last but not least, we need to train the Exit-Policy to make a reliable decision if we can exit the reasoning process at current inference step. However, there are no ground-truth labels providing feedback on when our network can exit reasoning.
Therefore, we leverage $m_t$
to generate dynamic labels to train Exit-Policy. Recalling that $m_{t}$ is the probability margin between the correct answer and the largest one of other candidate answers at each step, and $m_{t}$ is optimized to keep increasing
under the constraint of Eqn. \ref{eq.increment_loss}.
Therefore, the gap between $m_t$ at different inference steps can be used to estimate the information gain of seeing more frames in our dynamic reasoning process. Given the pre-defined largest reasoning step $T$, the gap between $m_t$ (at the current step) and $m_T$ (of the final step) can estimate the value of remaining information gain by continuing reasoning till the end. Thus the gap can be used to determine whether our network can stop inference at $t^{th}$ step. When $m_t$ is very close to $m_T$, this means the information gain by observing more frames is small, and thus we can exit reasoning in advance to reduce computation without incurring decrease in prediction accuracy.
In particular, at each step $t$, if $m_{T} - m_{t}<\mu(m_{T}-m_1)$, which means $m_t$ is close to $m_T$, i.e., the estimated remaining information gain is small enough, we set the label, $y^t_{exit}$, as 1, indicating our model can exit reasoning at the $t^{th}$ inference step. Otherwise, the label is set to 0, representing we need to seek more frames. Here $\mu>0$ controls how close $m_t$ should be to $m_T$ when the network exits inference. Conditioned on the estimated labels, $y^t_{exit}$, training Exit-Policy can be seen as a binary classification problem. Thus we train this module by minimizing a binary cross-entropy loss:
\begin{equation}\label{eq.bce_loss}
\mathcal{L}^t_{exit}= -[y^t_{exit}\log(e_{t}) + (1-y^t_{exit})\log(1-e_{t})]
\end{equation}
By combining Eqns (\ref{eq.cross-entropy loss}), (\ref{eq.glimpse_loss}),
and (\ref{eq.bce_loss}), the total loss function for each step $t$ can be formulated as:
\begin{equation}\label{eq.final_loss}
\mathcal{L}^t = \mathcal{L}^t_{pred} + \mathcal{L}^t_{exit} + \lambda *\mathcal{L}^t_{glimpse}
\end{equation}
where $\lambda$ is the weight of the combined loss function for optimizing the Glimpse-Determination Module.
In our experiments, we compute the sum: $\sum_{t=1}^{T}{\mathcal{L}^t}$ from all inference steps as the final optimization objective.
Note that Eqn. (\ref{eq.final_loss}) cannot be optimized directly with gradient descent, since the involved decisions of selecting frames and determining feature granularity in our dynamic reasoning are discrete, and sampling from discrete distribution makes the network non-differentiable. To address this issue, we introduce an effective joint Gumbel-Softmax operation.
{\bf Joint Gumbel Softmax.} The original Gumbel-Softmax Sampling \cite{jang2016categorical} is an effective way to transform the original non-differentiable sample from a discrete distribution, to a differentiable decision from a corresponding Gumbel-Softmax distribution. In our task, to sample from the aforementioned two discrete distributions (namely, selecting frames and determining granularity) simultaneously, we here design an effective joint Gumbel-Softmax operation.
In particular, in the Glimpse-Determination Module, at each step $t$, we first derive the logits $z \in\mathbb{R}^{T*2}$ by feeding the hidden state $h_t$ into a fully-connected layer.
Then we use $Softmax$ to obtain a categorical distribution $\pi_t$ from $z$:
$\pi_t = \left\{ p_{i,j} \mid p_{i,j}=\frac{\exp(z_{i,j})}{\sum_{c=1}^T\sum_{k=1}^2\exp(z_{c,k})} \right\}$.
With the Gumbel-Max trick \cite{jang2016categorical}, the discrete sample from the categorical distribution $\pi_t$ can be defined as follows:
\begin{equation}\label{eq.argmax}
\hat{l_t} = \mathop{\arg\max}_{i\in\{1,...,T\},j\in\{1,2\}} (\log p_{i,j} + g_{i,j})
\end{equation}
where $g_{i,j} = -\log (-\log(u_{i,j}))$ denotes the Gumbel noise, and $u_{i,j}$ is the i.i.d. samples drawn from $Uniform(0,1)$. We can further relax the non-differentiable operation $argmax$ with $softmax$ to facilitate gradient-based optimization:
\begin{equation}\label{eq.argmax_frame_relaxation}
\begin{aligned}
l_t = &\left\{ P_{i,j} \mid P_{i,j}=\frac{\exp{((\log p_{i,j} + g_{i,j})/\tau)}}{\sum_{c=1}^{T}\sum_{k=1}^{2}\exp{((\log p_{c,k} + g_{c,k})/\tau)}}\right\},\\
&for\ i\in\{1,...,T\}, j\in\{1,2\}
\end{aligned}
\end{equation}
where $\tau$ is the temperature parameter, which controls the smoothness of the sampling mechanism. When $\tau\to0$, the sampling approximates the $argmax$ operation in Eqn. \ref{eq.argmax}. The output $l_t$ incorporates the output of two decisions: the first dimension of $l_t$ denotes the decision of selecting the frame (i.e, the output $s_t$ of the skip-policy branch) and the second dimension of $l_t$ denotes the decision of the feature granularity of the selected frame (i.e, the output $g_t$ of the granularity-policy branch).
By using the outputs ($s_t$ and $g_t$) of the joint Gumbel-Softmax operation, our network manages to dynamically select the frame for next inference step and specify the feature granularity for the selected frame at each step.
Therefore, by introducing the joint Gumbel-Softmax, our network can learn the two discrete policy decisions jointly in a fully differentiable way.
{\bf Training and testing.} During training, we optimize our network within a fixed number of steps, which means the exit-policy is trained together with other modules but the exit decisions
are not used. However, at the testing phase, if the exit-policy decides to stop reasoning at the $t^{th}$ inference step, we exit the model and use the current prediction result, $p_t$, as the final reasoning result. In such a manner, our model achieves dynamic causal reasoning.
\section{Experiments} \label{sec:exp}
Given that the number of candidate answers for each question is not fixed in our dataset, we evaluate the performance of our network using binary and multi-choice setups. In binary case (denoted as \textbf{Setting-1/2}), the input to the model is a question with an answer, and the model needs to predict the correctness of this answer. In multi-choice setup (denoted as \textbf{Setting-1/4}), models are expected to select the correct answer from 4 candidate answers (i.e, 3 of them are incorrect). These two experiment setups can be treated as binary and four-class classification problems.
{\bf Implementation Details.}
We compute features from the penultimate layer of a pretrained ResNet-101 model \cite{he2016deep} as the fine-grained frame feature,
and a pretrained MobileNetv2 \cite{sandler2018mobilenetv2} is used as the lightweight CNN to extract coarse features. In the QA Bank, we use Glove \cite{Pennington14glove:global} to embed QA text, and then use a BiLSTM with 150-Dimension hidden states to encode the textual sequence. As for the Interaction LSTM, the dimension of hidden states is 300.
We implement the framework using Pytorch and adopts Adam \cite{kingma2014adam} with a learning rate of 3e-4 and a weight-decay of 1e-5. The $\mu$ in the Exit-Policy is set to 0.1 and $\lambda$ is set to 0.01 in the loss function. We follow \cite{jang2016categorical} and set the initial temperature $\tau$ to 5, and gradually anneal it with an exponential decay factor of -0.045 in every epoch. According to evaluation statistics, our network shows a very fast \textbf{inference speed} of $16ms$ per testing video on a Nvidia RTX 2080Ti GPU.
We compare our network with the following baselines.
{\bf Text-only models.} These models only relying on text information without visual input, are relatively weak baselines used to assess language biases in our SUTD-TrafficQA. {\bf Q-type (random)} randomly selects an answer from the answer space. {\bf QE-LSTM} uses Glove \cite{Pennington14glove:global} to embed the input question and then encode it with LSTM \cite{hochreiter1997long}. The final LSTM hidden state is passed to a MLP for predicting the correct answer. Different from {\bf QE-LSTM} using questions only, {\bf QA-LSTM} uses LSTM to encode both question embedding and answer embedding, and the final hidden states are used for predicting the answer.
{\bf Text+video models.} We evaluate the following models that require both video and text inputs.
{\bf VIS+LSTM} \cite{ren2015exploring} uses LSTM to encode image representation and textual features. Since the original method takes a single image as input, we adapt this method by averaging features of all sampled frames in a video as the visual input.
{\bf Avgpooling} uses each frame with the encoded QA features to compute a prediction for each frame.
We then perform mean pooling over all the frame predictions to obtain the final result.
{\bf CNN+LSTM} uses two LSTMs to encode both the video sequence and the QA text respectively. The two final hidden states are concatenated to predict the correct answer.
{\bf I3D+LSTM} uses I3D network \cite{carreira2017quo} to extract video motion features, and then fuse with QA textual features encoded by LSTM to compute the model prediction.
{\bf TVQA} \cite{lei2018tvqa} is a multi-stream network to fuse input features from different modalities to answer the question.
{\bf HCRN} \cite{Le_2020_CVPR} adopts a Hierarchical Conditional Relation Networks to model sophisticated structure for reasoning over videos.
{\bf BERT-VQA} \cite{Yang_2020_WACV} uses BERT \cite{devlin2018bert}
to encode the visual and language information jointly to predict the answer.
{\bf }
\subsection{Results and Analysis}
\begin{figure*}[t]
\setlength\abovecaptionskip{-5cm}
\setlength\belowcaptionskip{-5cm}
\begin{center}
\includegraphics[width=1\linewidth]{stats_1_compressed.pdf}
\end{center}
\vspace{-0.3cm}
\caption{A qualitative example. The numbers above the selected frames show the order of the sequence selected by our network. It shows that our model selects informative frames dynamically and allocates large computation budget using fine features to most relevant frames for causal reasoning. More examples in supplementary.}
\label{fig:policy_example}
\vspace{-0.3cm}
\end{figure*}
\begin{table}[t]
\vspace{-0.2cm}
\caption{Results on SUTD-TrafficQA dataset.
}
\setlength\abovecaptionskip{-2cm}
\setlength\belowcaptionskip{-2cm}
\setlength{\tabcolsep}{10pt}
\scriptsize
\begin{center}
\begin{tabular}{l|ccc}
\hline
Models & Setting-1/4 & Setting-1/2 & GFLOPs \\\hline
Q-type (random)& 25.00 & 50.00 & - \\
QE-LSTM & 25.21 & 50.45 & - \\
QA-LSTM & 26.65 & 51.02 & - \\ \hline
Avgpooling & 30.45& 57.50 & 252.69 \\
CNN+LSTM & 30.78 & 57.64 & 252.95 \\
I3D+LSTM & 33.21 & 54.67 & 108.72 \\
VIS+LSTM \cite{ren2015exploring} & 29.91 & 54.25 & 252.80\\
BERT-VQA \cite{Yang_2020_WACV} & 33.68 & 63.50 & 266.77 \\
TVQA \cite{lei2018tvqa} & 35.16 & 63.15 & 252.11 \\
HCRN \cite{Le_2020_CVPR} & 36.49 & 63.79 & 2051.04 \\
\textbf{Eclipse} & \textbf{37.05} & \textbf{64.77} & \textbf{28.14} \\ \hline
\textit{Human} & 95.43 & 96.78 & -\\ \hline
\end{tabular}
\end{center}
\label{table:baseline_comparison}
\vspace{-0.5cm}
\end{table}
The results in Table \ref{table:baseline_comparison}
demonstrate the minimal language biases in SUTD-TrafficQA, as the text-only baselines perform almost the same as the random choice. In contrast, the models using video input achieve obviously higher accuracy than text-only baselines. This demonstrates that to solve the reasoning tasks in our dataset, the model needs to associate visual content with linguistic cues to infer correct answers. For a fair computation-efficiency comparison, we pick models requiring video input to compute GFLOPs per video, as the visual feature extraction consumes much computation budget, and the metric of GFLOPs is independent of hardware configurations. The results show our Eclipse achieves state\text{-}of\text{-}the\text{-}art reasoning accuracy
with significantly improved computation efficiency. This verifies that compared to conventional video QA methods, through dynamic causal reasoning, our model effectively exploits the spatio-temporal and logical structure of video events to infer correct answers with much smaller frame usage and efficient feature computation.
In addition, three volunteers who did not see the videos and questions before, were invited to pick correct answers, and we use the average prediction accuracy as \textit{Human} performance. The discrepancy between neural networks and the human performance demonstrates the challenging nature of our dataset and the necessity of further research in video reasoning area.
\begin{table}[tbp]
\caption{Results of removing granularity-policy in Eclipse.
}
\vspace{-0.2cm}
\setlength{\tabcolsep}{0.5pt}
\scriptsize
\begin{center}
\begin{tabular}{c|c|c|c}
\hline
Models & Coarse Features Only & Fine Features Only & ~Eclipse~(dynamic granularity) \\\hline
Accuracy & 34.35 & 37.16 & 37.05 \\ \hline
GFLOPs & 10.61 & 133.75 & 28.14 \\ \hline
\end{tabular}
\end{center}
\label{table:granularity-policy}
\vspace{-0.7cm}
\end{table}
\begin{table}[t]
\caption{Results of removing skip-policy in Eclipse.
Uniform-$n$ means uniformly sampling $n$ frames from the video for reasoning.
}
\vspace{-0.1cm}
\setlength{\tabcolsep}{2pt}
\scriptsize
\begin{center}
\setlength\abovecaptionskip{-2cm}
\setlength\belowcaptionskip{-2cm}
\begin{tabular}{c|ccc|c}
\hline
Models & ~~Uniform-10~~ &~~ Uniform-20~~ & ~~Uniform-40~~ & ~Eclipse~(skip-policy) \\\hline
Accuracy & 34.16 & 35.49 & 36.48 & 37.05 \\ \hline
GFLOPs & 32.17 & 51.19 & 68.41 & 28.14 \\ \hline
\end{tabular}
\end{center}
\label{table:skipping-policy}
\vspace{-0.7cm}
\end{table}
\begin{table}[tbp]
\caption{Results of removing exit-policy in Eclipse.
Final-Step Inference refers to that we remove the exit-policy and infer until the final step.
}
\vspace{-0.1cm}
\setlength{\tabcolsep}{10pt}
\scriptsize
\begin{center}
\begin{tabular}{c|c|c}
\hline
Models & Final-Step Inference & ~Eclipse~(exit-policy)\\\hline
Accuracy & 37.11 & 37.05 \\ \hline
GFLOPs & 36.92 & 28.14 \\ \hline
\end{tabular}
\end{center}
\label{table:exit-policy}
\vspace{-0.9cm}
\end{table}
{\bf Ablation Study.} As shown in Table \ref{table:granularity-policy}, compared with the model variant of using fine features only, by adopting the granularity-policy to adaptively decide feature granularity for the selected frame at each step, our network achieves nearly the same accuracy yet using much lower computation cost.
Our network also achieves obviously higher reasoning accuracy than the method of using coarse features only. These results show the effectiveness of our granularity-policy by choosing fine features for most useful frames and coarse features for less important frames.
Furthermore, we remove the skip-policy and simply uniformly sample frames at each step.
As shown in Table \ref{table:skipping-policy}, our Eclipse performs the best yet at small computation cost.
This shows that our skip-policy effectively reduces the computation cost by selecting useful frames for dynamic reasoning.
Moreover, we present the frame location distributions for the first three inference steps of our network in Figure \ref{fig:Freq}.
As shown, our network selects frames dynamically for different videos as we expect.
We also investigate the exit-policy module by comparing it with the method of reasoning until the final inference step. The result in Table \ref{table:exit-policy} shows that our model achieves best accuracy-to-computation ratio via adaptive inference. In Figure \ref{fig:policy_example}, we further present a qualitative example from our dataset to show how Eclipse performs dynamic and efficient reasoning.
\begin{figure}[htbp]
\vspace{-0.3cm}
\begin{center}
\includegraphics[width=0.8\linewidth]{freq_v2.pdf}
\end{center}
\vspace{-0.4cm}
\caption{Distributions of frame location (Left to right:step1, 2, 3).}
\label{fig:Freq}
\vspace{-0.4cm}
\end{figure}
\section{Conclusion}
We create a new video QA dataset, SUTD-TrafficQA, focusing on video reasoning over traffic events. In our dataset, we introduce 6 reasoning tasks requiring various levels of causal reasoning. Besides, we propose the Eclipse network for video QA. By learning the dynamic glimpse policy and adaptive exit policy, our network achieves superior performance with significant computation efficiency.
{\bf Acknowledgement.} We would like to thank Yutian Lin, Renhang Liu, Yingjie Qiao, Xun Long Ng, Tran Nguyen Bao Long, Koh Kai Ting and Christabel Dorothy for their help in dataset collection and running baseline models. This work is supported by SUTD Projects PIE-SGP-Al2020-02 and SRG-ISTD-2020-153.
{\small
\bibliographystyle{ieee_fullname}
|
\section*{Methods}
\noindent To verify the resonant diffraction governed by the superposition of scattering processes owing to higher Fourier harmonic components, one-dimensional (1D) lattices were investigated through the rigorous finite element method (FEM) simulations. We used the commercial software COMSOL multiphysics 5.3a. Since the 1D periodic structures are invariant in the $y$-direction, simulations were performed in 2D $xz$-plane. Computational cell of size $\Lambda \times 15~\Lambda$ was employed to obtain dispersion relations, $Q$ factors, spatial electric field distributions, and transmission spectra. Bloch periodic boundary condition was used in the $x$-direction and perfectly matched layer absorbing boundary condition was employed in the $z$-direction. For reliable simulations, we used user-controlled mesh with extremely fine element size.
\section*{Data availability}
\noindent Source data are provided with this paper. All other data that support the plots within this paper and other findings of this study are available from the corresponding author upon reasonable request.
\section*{acknowledgments}
\noindent This research was supported by grants from the National Research Foundation of Korea funded by the Ministry of Education (No. 2020R1I1A1A01073945) and Ministry of Science and ICT (No. 2020R1F1A1050227), along with the Gwangju Institute of Science and Technology Research Institute.
\section*{Author contributions}
\noindent The resonance-assisted grating equations were developed by S.-G, and numerical simulations were performed by S.-G. and S.-H. C.-K. analysed the numerical data. All authors contributed to discussions and manuscript writing.
\section*{competing interests}
\noindent The authors declare no competing interests.
\section*{Additional information}
\noindent Correspondence and requests for materials should be addressed to S.-G. and C.-K.
\end{document}
|
\section{Introduction}
The precise history of solar activity and its underlying magnetic field is of interest for a number of reasons. Firstly, records of solar activity and of the magnetic field pose an important constraint on models for the enhancement and the evolution of magnetic flux (mainly dynamo models). Secondly, such records are important for understanding the history of the Sun's influence on the Earth (either through changes in its irradiance, or through space weather effects). Thirdly, a long record of solar activity is needed to understand how the Sun compares with other sun-like stars in its level of activity and variability \citep[e.g.,][]{Radick:2018,Reinhold:2020}.
Solar activity, in all its diverse manifestations, is driven by its magnetic field, so that knowledge of the history of solar activity implies knowledge of its magnetic field.
Two widely used quantities describing the global magnetic field of the Sun are the global open and unsigned total magnetic flux. They are, for example, used in heliospheric physics, for the reconstruction of solar irradiance, or as measures of solar activity when comparing with other stars. These quantities, being global in nature, can be reconstructed from more indirect proxies of solar activity and magnetism, such as the sunspot number and concentrations of the cosmogenic isotopes $^{14}$C or $^{10}$Be in terrestrial archives.
A first model to compute the solar open flux from the sunspot number was developed by \citet{Solanki2000}, based on a simple differential equation describing the evolution of the open flux, $\phi_{\rm open}$. In spite of its simplicity, it successfully reproduced the empirically-reconstructed evolution of the open flux by \cite{Lockwood1999} and the $^{10}$Be concentration in ice cores \citep{Beer+al:1990}. This simple model was extended by \citet{Solanki2002} to cover also the total unsigned magnetic flux, $\phi_{\rm total}$, but now requiring the solution of a set of coupled differential equations to describe the evolution of ephemeral regions (ERs) besides that of active regions (ARs) and of the open flux. All three components of magnetic flux contribute to the evolution of the total magnetic flux.
ARs are the large bipolar structures that harbour sunspots at least part of the time, whereas ERs are smaller bipolar regions without sunspots.
The ability of this model to reproduce concentrations of cosmogenic isotopes turned out to be particularly useful \citep[e.g.,][]{usoskin_JGR_02}. Although far more sophisticated models are in the meantime available to compute not just global magnetic quantities, but also the underlying spatial distribution of the magnetic flux and the detailed input from individual emerging ARs etc., the very simplicity of this set of models allowed them to be inverted \citep[e.g.,][]{Lockwood:2003}, so that, e.g., sunspot number could be reconstructed from measured concentrations of cosmogenic isotopes \citep[e.g.,][]{Usoskin2003,Usoskin2004,Solanki2004,usoskin_AA_16,Wu18_composite}.
The model of \citet{Solanki2002} was further extended and combined with the successful SATIRE model \citep[Spectral And Total Irradiance REconstruction;][]{Fligge2000,Krivova2003,Krivova2011} to compute total solar irradiance over the last 400 years \citep{Krivova2007,Krivova2010}.
\citet[][hereafter VS2010]{VieiraSolanki2010} have further refined the model by distinguishing between the short-lived and long-lived components of the open flux
(\citealp{Ihksanov-Ivanov:1999,Cranmer:2002,Crooker+al:2002}; see \citealp{VieiraSolanki2010} for details), which led to an improved reconstruction of the open flux that displayed a better agreement with observations.
This model, with some tuning, has been the basis for further reconstructions of solar spectral irradiance over the telescope era \citep{Krivova2010}, as well as sunspot number and TSI over the Holocene \citep{Vieira+al:2011,Wu18_composite,Wu18_ssi}.
One shortcoming of earlier versions of the model discussed above is that the open flux during a grand minimum, such as the Maunder minimum, i.e. during a long period essentially without sunspots, invariably drops to zero. This is because in this model, the emergence rate of the magnetic field on the solar surface is linearly linked to the sunspot number, so that by design during a grand minimum no magnetic flux is allowed to emerge.
This leads to a zero open and total flux during the grand minima.
It has been shown, however, that signals of solar activity and variability were also present during the Maunder minimum
\citep{Beer+al:1998,Fligge+al:1999,Usoskin2001,Miyahara:2004,Riley:2015}.
This was also confirmed by modelling \citep{owens12} and points to a need for an improvement of the global total and open magnetic flux model.
Furthermore, more recent solar observations provided new insights into the sources, emergence and evolution of the solar magnetic flux.
Thus, \citet{Thornton-Parnell:2011} have combined observations from various sources and found that the
emergence rate of bipolar magnetic regions with fluxes between $10^{16}$~Mx and $10^{23}$~Mx follows a single power law with a slope of $-2.69$.
Here we present a new, strongly revised version of the VS2010 model that builds on these recent solar observational results, replacing the direct proportionality of ERs and ARs by a more up-to-date approach. It does keep the original differential equations, however, so that it is not a completely independent model. As a natural outcome of the model, ERs keep emerging even during a grand minimum when there are no sunspots for multiple decades. This means that neither the open nor the total magnetic flux drop to zero at any time.
The paper is structured as follows. The data used to constrain and test the new model are briefly introduced in Sect.~\ref{sec:data}.
We describe our model and highlight the changes relative to the older version of the model in Sect.~\ref{sec:model}. The results of the model are presented in Sect.~\ref{sec:resu}, while we summarise and discuss our findings in Sect.~\ref{sec:Summary}, where we also provide an outlook on future applications of the new model.
\section{Data}
\label{sec:data}
The model to be detailed in Sect.~\ref{sec:model} starts from a sunspot number time series and computes the total and open magnetic fluxes of the Sun therefrom.
We therefore require a sunspot series as input to the model.
To constrain the free parameters of the model, we compare its output to observations and independent data-based reconstructions of the total and open magnetic fluxes.
Finally, to test the output of the model, we consider further independent time series of the reconstructed open magnetic flux.
As input to the model we use the following sunspot number data sets: (1) the international sunspot number v2.0, referred to hereafter as ISN2.0 \citep{clette16},
and (2) the group sunspot number, or GSN in short \citep{Hoyt:1998}.
The ISN2.0 data set was extended back to 1643 by adding the sunspot data during the Maunder minimum by \citet{Vaquero15} scaled up by the factor 1.67 to match the ISN2.0 definition.
Also in the GSN record, the values before 1710 were replaced by the data from \citet{Vaquero15}, without any scaling.
To constrain the output of the model we make use of observations of the total magnetic flux \citep[see][]{Arge:2002,Wang:2005,Wenzler:2006}
derived from synoptic charts produced by the three solar observatories with the longest running regular magnetographic measurements:
Wilcox Solar Observatory (WSO),
Mount Wilson solar Observatory (MWO), and National Solar Observatory at Kitt Peak (NSO/KP).
These data sets have already been used to constrain earlier versions of this model and we have used the same versions as employed by \citet{Krivova2007,Krivova2010} and \citet{Wu18_ssi}.
To constrain the free parameters of the model, we consider the average over at least two (MWO and WSO during 2002--2009) or all three (1976--2002) records for each Carrington rotation.
Furthermore, to better constrain the free parameters of the model, we also use the empirical reconstruction of the open magnetic flux from the geomagnetic aa-index covering the period from 1845 to 2010 \citep{Lockwood14_geo}.
Finally, the quality of the computed open magnetic flux is tested by comparing it with two other independent data sets (without changing the free parameters of the model):
(1) a compilation of spacecraft-based in-situ measurements by \citet{Owens:2017} since 1998, and
(2)
a reconstruction by \citet{Wu18_composite}
from decadal INTCAL13 $^{14}$C data covering the Holocene prior to 1900 \citep{reimer13}.
\section{Model}
\label{sec:model}
\subsection{Magnetic flux emergence and evolution}
\label{sec:emerge}
Following the approach by \citet{Solanki2000,Solanki2002} and \citet{VieiraSolanki2010}, we describe the evolution of the solar total and open magnetic flux by a system of ordinary differential equations. However, instead of distinguishing between active regions (AR) and ephemeral regions (ER) as the sources of fresh magnetic flux at the solar surface, we distinguish between ARs and what we call Small-Scale-Emergences (SSEs), i.e. all emergences with fluxes smaller than those in active regions. These therefore combine the flux emerging in the form of ephemeral regions and smaller magnetic bipoles all the way down to internetwork fields. The equations describing the evolution of the different (globally averaged) components of the magnetic flux are:
\begin{equation}
\frac{{\rm d}\phi_{\rm AR}}{{\rm d}t} = \varepsilon_{\rm AR} - \frac{\phi_{\rm AR}}{\tau^{\rm 0}_{\rm AR}} - \frac{\phi_{\rm AR}}{\tau^{\rm s}_{\rm AR}} - \frac{\phi_{\rm AR}}{\tau^{\rm r}_{\rm AR}},
\label{eq:ode1}
\end{equation}
\begin{equation}
\frac{{\rm d}\phi_{\rm SSE}}{{\rm d}t} = \varepsilon_{\rm SSE} - \frac{\phi_{\rm SSE}}{\tau^{\rm 0}_{\rm SSE}} - \frac{\phi_{\rm SSE}}{\tau^{\rm s}_{\rm SSE}},
\label{eq:ode2}
\end{equation}
\begin{equation}
\frac{{\rm d}\phi^{\rm r}_{\rm open}}{{\rm d}t} = \frac{\phi_{\rm AR}}{\tau^{\rm r}_{\rm AR}} - \frac{\phi^{\rm r}_{\rm open}}{\tau^{\rm r}_{\rm open}},
\label{eq:ode3}
\end{equation}
\begin{equation}
\frac{{\rm d}\phi^{\rm s}_{\rm open}}{{\rm d}t} = \frac{\phi_{\rm AR}}{\tau^{\rm s}_{\rm AR}} + \frac{\phi_{\rm SSE}}{\tau^{\rm s}_{\rm SSE}} - \frac{\phi^{\rm s}_{\rm open}}{\tau^{\rm s}_{\rm open}},
\label{eq:ode4}
\end{equation}
\begin{equation}
\phi_{\rm open} = \phi^{\rm r}_{\rm open} + \phi^{\rm s}_{\rm open},
\label{eq:ode5}
\end{equation}
\begin{equation}
\phi_{\rm total} = \phi_{\rm AR} + \phi_{\rm SSE} + \phi_{\rm open}.
\label{eq:ode6}
\end{equation}
Here $\phi_{\rm AR}$, $\phi_{\rm SSE}$, $\phi_{\rm open}$ and $\phi_{\rm total}$ refer to the magnetic flux in ARs and SSEs, as well as the open, and total magnetic flux (all magnetic fluxes are global, i.e. referring to their unsigned sum over the entire solar surface).
The open flux is divided into rapidly ($\phi^{\rm r}_{\rm open}$) and slowly ($\phi^{\rm s}_{\rm open}$) evolving components.
$\tau^{\rm 0}_{\rm AR}$, $\tau^{\rm 0}_{\rm SSE}$, $\tau^{\rm r}_{\rm open}$, and
$\tau^{\rm s}_{\rm open}$ are the decay timescales of ARs, SSEs, rapid open flux, and slow open flux, respectively.
The flux transfer timescales from ARs and SSEs to slow open flux are $\tau^{\rm s}_{\rm AR}$ and $\tau^{\rm s}_{\rm SSE}$, respectively, while the corresponding timescale for the transfer of AR flux to the rapid open flux is $\tau^{\rm r}_{\rm AR}$. See \cite{VieiraSolanki2010} for a discussion on the distinction between rapidly and slowly evolving open magnetic flux.
In the original model, the emergence rate of ARs at a given time $t$, $\varepsilon_\mathrm{AR}(t)$, was linked linearly to the sunspot number, SN, at that time:
\begin{equation}
\varepsilon_\mathrm{AR}(t) = \varepsilon^{\mathrm{max},21}_{\rm AR}\frac{\mathrm{SN}(t)}{\mathrm{SN^{\mathrm{max},21}}} \;,
\label{eq:emerg_act_old}
\end{equation}
where $\varepsilon^{\mathrm{max},21}_{\rm AR}$ and $\mathrm{SN^{\mathrm{max},21}}$ are the three-month averaged emergence rate and SN value observed during the maximum of cycle 21 \citep[taken from][]{Schrijver-Harvey:1994}, respectively.
Because at the time that the model was originally developed, large-scale studies of magnetic flux emergence and evolution could not resolve internetwork fields, the earlier model was restricted to ERs as the only magnetic bipoles smaller than ARs.
The emergence rate of the ERs, $\varepsilon_{\mathrm{ER}, n}$, of the cycle $n$ was not well known and was assumed to be a sinusoidal function, $g^n$, which was stretched such that the length of the ER cycle was longer than the respective sunspot cycle.
The amplitude of the ER cycle was simply taken to be proportional to the maximum value of the emergence rate of the ARs of the cycle, so that the emergence rate of ERs over a cycle had the form:
\begin{equation}
\varepsilon_{\mathrm{ER},n}= \varepsilon^{\mathrm{max},n}_\mathrm{AR}Xg^n \;,
\label{eq:emerg_eph_old}
\end{equation}
where
$X$ is an amplitude factor (a free parameter of the model; same for all cycles).
Importantly, Eq.~\ref{eq:emerg_eph_old} implies a linear relationship between the emergence rate $\varepsilon_{\mathrm{ER},n}$ and the sunspot number. This, along with the linear relationship between $\varepsilon_{\mathrm{AR},n}$ and sunspot number (Eq.~\ref{eq:emerg_act_old}) results in an absence of flux emergence during extended periods of spotless days.
\begin{figure}[htb]
\centering
\resizebox{\hsize}{!}{
\includegraphics{krivova-fig1}
}
\caption{Frequency of emergence vs. the unsigned flux of an emergence event (see Eq.~(\ref{eq:powerlaw}) following \citealp{Thornton-Parnell:2011}). $\phi_0$, $\phi_\mathrm{ER}$, $\phi_\mathrm{AR}$ and $\phi_\mathrm{limit}$ represent the limit below which the local dynamo flux dominates, the minimum ephemeral region flux, the minimum active region flux, and the upper limit of the active region flux, respectively.
The horizontal arrows mark the flux ranges corresponding to the internetwork (IN), ephemeral regions (ERs), active regions (ARs), and Small-Scale Emergences (SSEs). The SSE range includes IN fields and ERs.
The slope $m$ of the distribution was derived by \citet{Thornton-Parnell:2011} by fitting various observations at different activity levels. Slopes $m_1$ and $m_2$ represent the corresponding distributions at maximum and minimum of solar activity levels for cycle 21, respectively (see main text for details).
}
\label{fig:powerlaw}
\end{figure}
To overcome this shortcoming, we incorporate more recent solar observations by \citet{Thornton-Parnell:2011} into the model.
Using high-resolution Hinode, Solar Optical Telescope/Narrow-band Filter Imager (SOT/NFI) observations and combining them with earlier published data, they found that the emergence rate of the magnetic flux on the Sun follows a power-law distribution:
\begin{equation}
\frac{\mathrm{d} N}{\mathrm{d} \phi} = \frac{n_0}{\phi_0} \Big( \frac{\phi}{\phi_0} \Big)^{m} \;,
\label{eq:powerlaw}
\end{equation}
where $\phi_0=10^{16}$\,Mx (the smallest flux per feature that they include in their histograms), $n_0=3.14\times10^{-14}$\,cm$^{-2}$\,day$^{-1}$ and $m=-2.69$
(see the illustration in Fig.~\ref{fig:powerlaw}).
We note that \citet{Thornton-Parnell:2011} have summarised the results from multiple studies with a wide range of solar activity levels and observing conditions.
In earlier studies, \cite{Harvey-Zwaan:1993} and \citet{Harvey_PhD:1993} found that the emergence rate of ARs varied significantly more between solar activity minimum and maximum than that of ERs.
Whereas roughly 8.3 times more ARs emerged during the maximum of cycle 21 than during the minimum (the factor generally grows with the size of the regions, but was on average about 8.3 for all the ARs they studied), this ratio was roughly two for ERs. The number of the smallest magnetic features, forming the internetwork magnetic fields and having fluxes of $10^{16}$-- $10^{17}$\,Mx, appears to be nearly invariable over an activity cycle \citep{Buehler+al:2013,Lites:2014}. These features differ from the larger ones in that they are mainly brought about by a small-scale turbulent dynamo \citep{Voegler-Schuessler:2007,Rempel:2014} that produces the same amount of magnetic flux nearly independently of large-scale activity.
To satisfy these observational constraints,
on the one hand, we keep the number of the smallest magnetic features considered here (with a flux per feature of $10^{16}$\,Mx) fixed at all times. On the other hand, we allow the exponent $m$ to vary
\citep[cf.][]{Parnell_2009,Schrijver-Harvey:1989}
around the empirical value $m = -2.69$ found by \citet{Thornton-Parnell:2011}
within the range
$m_1 \geq m \geq m_2$,
where $m_1 = m + \Delta m$ and $m_2 = m - \Delta m$.
The slopes $m_1$ and $m_2$ describe the distributions of emergence rates during periods when the observed sunspot numbers are $\mathrm{SN}_1$ and $\mathrm{SN}_2$ (with SN$_1 > \mathrm{SN}_2$).
In our model, the slope $m$ follows the SN, $m(\mathrm{SN})$,
according to the non-linear relationship:
\begin{equation}
m\left(\mathrm{SN}\right) = m_1 - \left(\mathrm{SN}_1^{\alpha} - \mathrm{SN}^{\alpha}\right) \frac{m_1-m_2}{\mathrm{SN}_1^{\alpha}-\mathrm{SN}_2^{\alpha}},
\label{eq:m}
\end{equation}
where $\alpha$ is a free parameter, fixed by comparing the output of the model to observations and independent reconstructions (see Sect.~\ref{sec:parameters}).
Now, the emergence rate of magnetic flux in ARs and SSEs at any given time can be calculated as:
\begin{equation}
\varepsilon_\mathrm{AR} =\int_{\phi_\mathrm{AR}}^{\phi_\mathrm{limit}} \frac{n_0}{\phi_0} \Big( \frac{\phi}{\phi_0} \Big)^m\;\phi\;\mathrm{d}\phi= \frac{n_0}{(m+2) \;{\phi_0}^{m+1}} \Big( \phi_\mathrm{limit}^{m+2} - \phi_\mathrm{AR}^{m+2} \Big) \;,
\label{eq:AR_emergence}
\end{equation}
and
\begin{equation}
\varepsilon_\mathrm{SSE} = \int_{\phi_0}^{\phi_\mathrm{AR}} \frac{n_0}{\phi_0} \Big( \frac{\phi}{\phi_0} \Big)^m\;\phi\;\mathrm{d}\phi= \frac{n_0}{(m+2) \;{\phi_0}^{m+1}} \Big( \phi_\mathrm{AR}^{m+2} - \phi_0^{m+2} \Big) \;.
\label{eq:ER_emergence}
\end{equation}
Here $\phi_\mathrm{AR} =4\times10^{20}$\,Mx denotes the magnetic flux of the smallest bipolar regions hosting sunspots, i.e. the smallest active regions \citep[e.g.,][]{Zwaan:1978,Schrijver-Zwaan:2000,vanDriel:2015} and
$\phi_\mathrm{limit}$ is
the flux of the largest considered ARs. Since such regions are extremely rare,
the exact value of $\phi_\mathrm{limit}$ is not important.
Following \citet{Parnell_2009} and \citet{Thornton-Parnell:2011},
we take $\phi_\mathrm{limit}=10^{23}$\,Mx, which is somewhat larger than the maximum flux ($3\times 10^{22}$\,Mx) for ARs listed by \citet{Schrijver-Zwaan:2000} and \citet{vanDriel:2015}.
Tests have shown that also the exact value of $\phi_\mathrm{AR}$ adopted here plays only a minor role for the end result in the sense that although the free parameters may have slightly different values, the computed open and total magnetic flux remain almost unchanged.
To estimate how the slope $m$ of the distribution given by Eq.~(\ref{eq:powerlaw}) changes with activity (Eq.~(\ref{eq:m})) we rely on the observations by
\citet{Harvey-Zwaan:1993} and \citet{Harvey_PhD:1993} for cycle~21.
They found that the number of emerging ARs
in cycle 21
varied between the activity maximum and minimum by a factor of roughly 8.3.
The monthly-smoothed sunspot numbers corresponding to these periods (1979 -- 1982 for the maximum, as well as 1975 -- 1976 and 1985 -- 1986 for the preceding and following minima)
are then
$\mathrm{SN}_1 =217$ and $\mathrm{SN}_2 = 17$, respectively.
These values are obtained for ISN2.0.
For GSN, these numbers correspond to $\mathrm{SN}_1 =130$ and $\mathrm{SN}_2 = 10$.
By using the factor of 8.3 found for the emergence frequency of ARs between activity maximum and minimum by \citet{Harvey_PhD:1993} as a constraint, we obtain $\Delta m = 0.0946$ (and thus $m_1 = -2.5954$ and $m_2 = -2.7846)$.
Note that the values of $m$ reach values higher than $m_1$ or lower than $m_2$ at times when
the sunspot number is higher than SN$_1$ or lower than SN$_2$, respectively.
In particular, when the sunspot number is zero, the corresponding $m$ values are $m(0)=-3.952$ for ISN2.0 and $m(0)=-3.677$ for GSN.
The value of $m(0)$ is different for ISN2.0 and GSN because the value of $\alpha$, the free parameter of the model (see Table \ref{tbl:paras}, and Sect.~\ref{sec:parameters}) is different: $\alpha=0.059$ for ISN2.0 and $\alpha=0.075$ for GSN.
The largest value of $m$
that the direct sunspot records give is obtained for the peak of cycle 19, with $\mathrm{max}(\mathrm{SN})=370$ (ISN2.0) and 247 (GSN), which results in $m(370)=-2.552$ (ISN2.0) and $m(247)=-2.542$ (GSN).
Figure \ref{fig:SN_m_isn}a shows the daily (black) and the monthly-smoothed (red) ISN2.0, while the evolution of $m_\mathrm{SN}$ computed from the monthly-smoothed ISN for the solar cycle~21 is shown in Fig.~ \ref{fig:SN_m_isn}b.
\begin{figure}
\centering
\resizebox{\hsize}{!}{
\includegraphics{krivova-fig2}
}
\caption{
(a) Daily (black) and 30-day smoothed (red) sunspot number (ISN, v2.0) over cycle 21, including the adjacent mimima. The average values during the activity maximum (SN$_1$ = 217) and minimum (SN$_2$ = 17) of cycle 21 are marked by the horizontal dashed lines. (b) Evolution of the monthly-smoothed power-law slope $m$ over the same period, the horizontal dashed lines mark $m_1$ = -2.5954 corresponding to SN$_1$ = 217 and $m_2$ = -2.7846 corresponding to SN$_2$ = 17.}
\label{fig:SN_m_isn}
\end{figure}
\subsection{Parameters of the model}
\label{sec:parameters}
\begin{table}[htb]
\centering
\caption{Parameters of the model.
}
\begin{tabular}{l|cc|c}
\hline
Parameter & ISN & GSN & W18 (ISN)\\
\hline
$n_0$, cm$^{-2}$\,day$^{-1}$ & \multicolumn{2}{c|}{$3.14\times10^{-14}$} & \ldots\\
$\phi_0$, Mx & \multicolumn{2}{c|}{$10^{16}$} & \ldots\\
$\phi_\mathrm{AR}$, Mx & \multicolumn{2}{c|}{$4 \times 10^{20}$} & \ldots\\
$\phi_\mathrm{limit}$, Mx & \multicolumn{2}{c|}{$10^{23}$} & \ldots\\
SN$_1$ & 217 & 130 & \ldots\\
SN$_2$ & 17 & 10 & \ldots\\
$m$ & \multicolumn{2}{c|}{-2.69} & \ldots\\
$\Delta m$ & \multicolumn{2}{c|}{0.0946} & \ldots\\
$\tau^{\rm 0}_{\rm AR}$, yrs& \multicolumn{2}{c|}{$0.027$} & \ldots\\
$\tau^{\rm 0}_{\rm SSE}$, yrs& \multicolumn{2}{c|}{$1.1\times 10^{-5}$} & \ldots\\
\hline
$\alpha$ & 0.059 & 0.075 & \ldots\\
$\tau^{\rm r}_{\rm open}$, yrs& 0.16 & 0.09 & 0.14 \\
$\tau^{\rm s}_{\rm open}$, yrs& 3.98 & 3.90 & 3.75 \\
$\tau^{\rm r}_{\rm AR}$, yrs& 2.79 & 1.57 & 2.6 \\
$\tau^{\rm s}_{\rm AR}$, yrs& 88.15 & 89.81 & 88.3 \\
$\tau^{\rm s}_{\rm SSE}$, yrs& 10.15 & 10.11 & 20.6\tablefootmark{*}\\
\hline
\end{tabular}
\label{tbl:paras}
\tablefoot{The upper part of the table lists parameters that are fixed, whereas the lower part lists free parameters. For comparison, the last column (W18) lists the values of the free parameters of the old model from the most recent version by \citet{Wu18_ssi}.\\
\tablefoottext{*} This value was obtained by \citet{Wu18_ssi} for ERs rather than SSEs.}
\end{table}
\begin{figure*}[!htb]
\sidecaption
\includegraphics[width=12cm]{krivova-fig3}
\caption{Evolution of the total magnetic flux over cycles 21--24 computed with the new (red) and old (blue) models, as well as the observed flux. For observations, symbols show individual Carrington rotations and observatories, as indicated in the legend, and black solid line shows their average.
For the models, means of the daily total magnetic flux over the Carrington rotations are shown, and the contribution of the SSEs is reduced by a factor of 0.4 (see text for details).
Panels (a) and (b) use ISN and GSN as input, respectively.
}
\label{fig:flux_compare_total}
\end{figure*}
Our model, described in Sect.~\ref{sec:emerge}, has a number of parameters, summarised in Table~\ref{tbl:paras}.
The upper part of the table lists quantities taken or deduced from the literature and kept fixed throughout the analysis, whereas the five free parameters that are allowed to vary when fitting the data sets described in Sect.~\ref{sec:data} are given in the lower part.
One positive feature of the new model is that it has less free parameters than the old model (the old model required two free parameters to describe the emergence of ERs; see Table~1 of \citealp{Wu18_ssi}), giving it less `wriggle room' for reproducing observational data.
All parameters relevant to the emergence rates of ARs and SSEs (Eqs.~\ref{eq:powerlaw}--\ref{eq:ER_emergence}) have been described in the previous section. Here we additionally comment on the decay and transfer times of the various components of the magnetic flux used in the ordinary differential equations describing the flux evolution (Eqs.~\ref{eq:ode1}--\ref{eq:ode6}).
The decay times of the ARs and SSEs, $\tau^{\rm 0}_{\rm AR}$ and $\tau^{\rm 0}_{\rm SSE}$, are estimated using the observations by \citet{Parnell_2009} and \citet{Thornton-Parnell:2011}.
Whereas \citet{Thornton-Parnell:2011} have analysed the emergence rate of different features as a function of their flux, \citet{Parnell_2009} analysed the magnetic flux for all (i.e. instantaneously) observed features.
By dividing the total number of the features of a given flux observed at a given instance by their emergence rate, we arrive at their mean lifetime.
The lifetime of the features increases with their sizes or fluxes.
For our purpose we make a simplification and calculate the lifetimes of ARs and SSEs as averages over all regions with fluxes above and below $\phi_\mathrm{AR}$, respectively, thus obtaining $\tau^{\rm 0}_{\rm AR}\approx 10$~days and $\tau^{\rm 0}_{\rm SSE}\approx 6$~minutes.
Since the features with the smallest flux dominate the power law distribution, $\tau^{\rm 0}_{\rm AR}$ is short compared with lifetimes of large ARs and is closer to lifetimes of small ARs \citep[e.g., Table 1 of the review by][]{vanDriel:2015}. For the same reason the SSE lifetime is close to that of internetwork elements, rather than of ERs.
To compare better with observations of ERs, we therefore introduce $\phi_\mathrm{ER}=10^{18}$\,Mx (see Fig.~\ref{fig:powerlaw}), which denotes roughly the lowest magnetic flux contained within ERs.
If we now consider as ERs only the regions with $\phi_\mathrm{ER} < \phi < \phi_\mathrm{AR}$, then we obtain a lifetime $\tau^{\rm 0}_\mathrm{ER}\approx 2$~hours, which is comparable to the lifetimes of $\approx$hours to a day for regions with fluxes between $3\times 10^{18}$\,Mx and $10^{20}$\,Mx listed by \citet{vanDriel:2015}.
The maximum to minimum change of the flux emerging in ERs (i.e. with fluxes $\phi_\mathrm{ER} < \phi < \phi_\mathrm{AR}$) in our model is roughly a factor of 2.5, which is consistent with the results by \citet{Harvey_PhD:1993} and \citet{Harvey-Zwaan:1993}.
We stress that this distinction into ERs and internetwork fields is only used for comparison purposes. Within the model they are not distinguished (see Sect.~\ref{sec:emerge}).
The decay timescales of the rapid, $\tau^{\rm r}_{\rm open}$, and slow,
$\tau^{\rm s}_{\rm open}$, open flux,
as well as timescales for the
flux transfer from ARs and SSEs to slow open flux, $\tau^{\rm s}_{\rm AR}$ and $\tau^{\rm s}_{\rm SSE}$, respectively, and the timescale for the transfer of the AR flux to the rapid open flux, $\tau^{\rm r}_{\rm AR}$, are free parameters of the model (together with $\alpha$ governing the change of the slope $m$ with the SN, as described in Sect.~\ref{sec:emerge}).
These parameters are fixed by comparing total and open magnetic flux with the corresponding observations listed in Sect.~\ref{sec:data}.
To do this, we use the genetic algorithm PIKAIA \citep{Charbonneau:1995},
which searches for the set of parameters minimising the difference between the modelled and the reference data sets.
We minimise the sum of the reduced chi-squared values, $\chi^2$, taking the errors of the observations and the number of data points into account. In other words, we search for the maximum of $1/(\chi^2_\mathrm{total}
+ \chi^2_\mathrm{OMF})$; see \citet{VieiraSolanki2010} for details and \citet{Dasi:2016} for a discussion of uncertainties in the parameter fitting.
The best-fit values of the parameters are listed in Table~\ref{tbl:paras}.
For comparison, the last column of Table~\ref{tbl:paras} also lists the values
of the five free parameters that are already present in the VS2010 model, as obtained by \citet{Wu18_ssi} for the most recent version of the model. This version used the same ISN2.0 input record extended back with the data from \citet{Vaquero15} as done here.
The values of the parameters are very close in the two models,
except $\tau^{\rm s}_{\rm SSE}$, for which \citet{Wu18_ssi} obtained 20.6~years compared to our 10.15~years (10.11~years when using GSN).
Note, however, that \citet{Wu18_ssi} considered ERs rather than SSEs.
Interestingly, the values we obtain are close to the value of 10.08~years found to produce a best-fit by \citet{VieiraSolanki2010} and is within the range 10--90~years that is consistent with observation \citep[see][]{VieiraSolanki2010}.
The value of about 4~years for the decay time of the slow open flux $\tau^{\rm s}_{\rm open}$ is close to the radial decay term with a timescale of 5~years introduced into surface flux transport simulations by
\citet{Schrijver:2002} and \citet{Baumann:2006} to act on the unipolar fields at the solar poles. This was needed to reproduce the observed polar field reversals. Similarly, the $\tau^{\rm r}_{\rm open}$ obtained here (Table~\ref{tbl:paras}) is consistent with the estimate of the decay of closed flux carried by interplanetary coronal mass ejections, between 1 and 5 AU of between 38 and 55 days by \citet{Owens:2006}.
\begin{figure*
\centering
\includegraphics[width=17cm]{krivova-fig4}
\caption{
Open magnetic flux computed with the new (red) and old (blue) models, together with the empirical reconstruction from the geomagnetic aa-index \citep[][green]{Lockwood14_geo}.
Also shown are the reconstruction by \citet[][black solid line with shading marking the uncertainty]{Wu18_composite} from the decadal
INTCAL13 $^{14}$C data \citep{reimer13} and the in-situ measurements from \citet[][dotted blue line]{Owens:2017} since 1998. Panels (a) and (b) use ISN and GSN as input, respectively.
}
\label{fig:flux_compare_open}
\end{figure*}
\section{Reconstruction of the total and open magnetic flux}
\label{sec:resu}
The computed total magnetic flux from 1965 onward is plotted in Fig.~\ref{fig:flux_compare_total}.
Following \citet{Krivova2007},
to account for the flux undetected due to the limited spatial resolution of observations \citep[see][]{Krivova:2004}, the contribution of the ER flux was multiplied by a factor of 0.4 before adding it to the contribution from the ARs and the open flux.
Shown are means over the daily values for each Carrington rotation.
The new and old models are depicted in red and blue, respectively. Symbols show the observations: Wilcox Solar Observatory
(WSO, squares), National Solar Observatory at Kitt Peak (NSO KP, diamonds), and Mount Wilson Observatory (MWO, triangles). Each individual symbol represents the total photospheric magnetic flux over a given Carrington rotation.
To compute the $\chi^2$ value, we use the average of the three (1976--2002) or two (after 2002, see Sect.~\ref{sec:data}) datasets indicated by the black solid line.
The computed open magnetic flux is shown in Fig.~\ref{fig:flux_compare_open}.
The new model is shown in red, the old model is in blue,
and the reconstruction from the geomagnetic aa-index by \citet{Lockwood14_geo} is represented by the green line. Also shown are the in-situ measurements by \citet[][dotted blue line]{Owens:2017} since 1998 and a reconstruction (black solid line with shading indicating the uncertainty) of the open flux by \citet{Wu18_composite} from independent decadal INTCAL13 $^{14}$C data \citep{reimer13}.
Note that the underlying $^{14}$C data and thus also the OF reconstructed from them are decadal averages.
The agreement between our model and the $^{14}$C-based reconstruction is quite good.
Particularly impressive is the agreement in the level of the open flux during the Maunder minimum, which is where we expect to see the biggest improvement relative to the old model. We emphasise that this $^{14}$C-based record was not used to constrain our model.
(Note that in the old model, the GSN record was used without the data from \citet{Vaquero15} over the Maunder minimum, and the computed open flux in the old model is therefore essentially flat at the zero level.)
Interestingly, over the 19th and the first half of the 18th centuries, the GSN-based reconstruction is closer to the isotope-based open flux than the reconstruction from the ISN, which lies somewhat higher.
\begin{table*}
\centering
\caption{Comparison of the modelled open and total magnetic fluxes to observations and independent reconstructions, quantified through their relative differences in means (in \%) and reduced $\chi^2$ values (listed in brackets).
}
\begin{tabular}{l|cc|cc}
\hline
Input & \multicolumn{2}{c|}{ISN} & \multicolumn{2}{c}{GSN} \\
model version & old & new & old & new\\
\hline
$F_\mathrm{total}$ & $-$7.7 (0.037) & $-$9.9 (0.058) & $-$4.1 (0.037) & 0.8 (0.058) \\
$F_\mathrm{open, L14}$ & $-$2.2 (0.297) & 0.4 (0.176) & $-$4.6 (0.389) & $-$1.6 (0.219) \\
\hline\\[-3mm]
$F_\mathrm{open, Wu18b}$\tablefootmark{*} &
$-$15.1 (1.718) & 6.1 (0.630) & $-$48.5 (1.674) & $-$9.6 (0.230) \\
$F_\mathrm{open, O17}$ & $-$15.4 (0.655) & 3.8 (0.236) & $-$3.5 (0.495) & 5.8 (0.252) \\
\hline
\end{tabular}
\label{tbl:chi2}
\tablefoot{The top part of the table lists the data sets that were used for parameter fitting (the average of the total magnetic flux measurements and the open flux reconstruction by \citealt{Lockwood14_geo}), while the bottom part lists independent data sets that were not used for the optimisation (OF reconstruction from $^{14}$C data by \citealt{Wu18_composite} and the in-situ measurements by \citealt{Owens:2017}).\\
\tablefoottext{*} For decadally-averaged reconstructions. As $^{14}$C data used for the reconstruction by \citet{Wu18_composite} are decadal averages, only decadally-averaged values of the OF could be reconstructed. Thus, to compute the corresponding $\chi^2$ values, our reconstructions were re-sampled, too.}
\end{table*}
A quantitative comparison of the total and open magnetic fluxes resulting from the old and the new models with the observations and independent reconstructions is presented in
Table~\ref{tbl:chi2}. The table lists the relative difference in means and the $\chi^2$ values (in brackets) between the models and the data.
For the total magnetic flux, the results are quite similar for both versions of the model.
In both cases, the mean modelled total magnetic flux is somewhat closer to the observations when the GSN is used as input.
The absolute difference in the means is slightly higher or lower for the new model if ISN2.0 or GSN are used, respectively.
The new $\chi^2$ values are somewhat higher than in the old model.
This is, however, primarily due to fact that, by model design, the variability of the ER component on time scales shorter than the solar cycle in the old model was essentially smoothed out (see Eqs.~\ref{eq:emerg_act_old}--~\ref{eq:emerg_eph_old}) resulting in weaker short-term fluctuations than in the new model
(see Fig.~~\ref{fig:flux_compare_open}).
Thus, if we smooth the total flux from both models with a 3-months window before comparing, the $\chi^2$ values for the old model remain almost unchanged (0.036 for both ISN2.0 and GSN), while those for the new model decrease to 0.043 for ISN2.0 and 0.047 for GSN. In all cases, the $\chi^2$ values are quite low.
For the open flux, the new model provides a notably better fit than the old model to all three alternative datasets.
In all but one case, the absolute mean differences are significantly lower for the new model.
The only exception is the GSN-based reconstruction versus the in-situ data by \citet{Owens:2017}, for which the absolute mean difference is slightly lower for the old model.
However, the results are quite close for both versions of the model in this case. Note also that these data cover only a short recent period of time, over which the two models do not differ significantly.
The $\chi^2$ values are lower for the new model in all cases.
The fit is poorest for the reconstruction based on the decadal values of $^{14}$C.
Very recently, new $^{14}$C-based activity measures with annual resolution were published by \citet{brehm:2021}.
An application of our model to the dataset of \citet{brehm:2021} will be subject of a separate publication (Usoskin et al., submitted).
\section{Summary and discussion}
\label{sec:Summary}
We have revised the simple model describing the evolution of the Sun's global total and open magnetic flux, originally proposed by \citet{Solanki2000,Solanki2002} and improved by \citet{VieiraSolanki2010}. The new version of the model takes into account the observation that fluxes of magnetic features follow a single power law, including internetwork fields, ERs and ARs
(\citealt{Parnell_2009}; cf. \citealt{Anusha:2017}). It also takes into account the fact that emergence rates of magnetic bipoles with fluxes between $10^{16}$~Mx and $10^{23}$~Mx, i.e. from the smallest ERs (and large internetwork features) to the largest ARs, also follow a power law according to the analysis by \cite{Thornton-Parnell:2011}.
We assume that the difference in emergence rates between the maximum and the minimum of solar activity is adequately described by the varying power-law exponent, affecting magnetic features with fluxes $\phi > 10^{16}$\,Mx. Thus, for the smallest features
there is no change in emergence rate over the solar cycle, while the ratio of emergence rates during maximum to minimum increases steadily with increasing magnetic flux.
Thus, the number of emerging ARs varies between the solar maximum and minimum of cycle 21 by a factor of 8.3, while this factor is 2.5 for ERs and close to unity for internetwork fields. These values are consistent with the respective ratios found by \citet{Harvey-Zwaan:1993} and \citet{Harvey_PhD:1993}, while the fact that the flux in internetwork fields hardly changes over the cycle is in agreement with the results obtained by, e.g., \citet{Buehler+al:2013} and \citet{Lites:2014}.
Using the sunspot number time series as input, the model returns time series of the open and total magnetic flux of the Sun. These resulting time series reproduce the open magnetic flux between 1845 and 2010 reconstructed by \cite{Lockwood14_geo} and total magnetic flux averaged over individual Carrington rotations obtained by various observatories between 1976 and 2009.
The main novel feature of the results of the model is that, in contrast to the earlier versions of the model, the output open magnetic flux does not drop to essentially zero during the Maunder minimum when almost no sunspots were present for multiple decades, in agreement with open flux reconstructed from $^{14}$C data \citep[e.g.,][]{Wu18_composite}.
This significant improvement is a result of the model allowing for a non-zero emergence of magnetic flux in small-scale bipolar features (encompassing ERs and internetwork fields) even during extended periods when sunspots were not present, e.g. during the Maunder minimum and other, similar grand minima \citep[e.g.,][]{Usoskin:2007}.
Even with this major update, the model still has some room for further improvements.
Because it uses sunspot numbers as input (the only data of solar activity available prior to the middle of the 19th century), it cannot properly treat variations in solar activity that are not reflected in the number of sunspots. This is particularly evident during grand minima. During such times sunspots are only occasionally visible, whereas cosmogenic isotopes continue to display cyclic variations. This suggests that, in the context of the present model, the slope $m$ continues to vary in a cyclical manner and can go lower than the lowest value we have obtained ($m(0)= -3.952$ for ISN2.0 and $m(0)= -3.677$ for GSN).
In the old version of the model the ER emergence was constructed as a smooth, sinusoidal function.
In the new model, ER (and internetwork) emergence rate closely follows that of sunspots and has therefore the same temporal resolution as the input sunspot number series.
As a consequence, the new model does not feature temporal lags or shifts to the corresponding sunspot cycle, as in the old model.
However, it does account for the finding by \citet{Harvey:1994}
that small-scale features (in her study ERs) belonging to a given cycle start emerging at a relatively high rate well before the sunspot cycle starts.
The internetwork is independent of sunspot emergence and is, thus, not associated with a particular sunspot cycle. These regions are simply a result of the dynamo not having completely switched off even at times when there are no sunspots visible.
In this way, the overlap between neighbouring solar magnetic cycles is naturally introduced. This was the main feature of the original model of \citet{Solanki2000} responsible for the change in the level of the Sun's open magnetic flux from one solar minimum to another, first noticed by \citet{Lockwood1999}.
At the same time, the new model allows for the dynamo to continue working and produce activity cycles during grand minima, which are sufficiently strong to modulate cosmic rays (see, e.g., Fig.~\ref{fig:flux_compare_open} for the evolution of the open flux during the Maunder minimum) and hence influence the production of cosmogenic isotopes \citep{Beer+al:1998,Fligge+al:1999,Usoskin2001}, but are too weak to produce more than occasional sunspots.
The higher level of the magnetic flux during periods of very low solar activity (e.g. during the Maunder and the Dalton minima, see Fig.~\ref{fig:flux_compare_open})
will presumably lead to a weaker secular variation of the total solar irradiance (TSI) and in particular the rise of the TSI since the Maunder minimum \citep[cf.,][]{Yeo:2020}, an important parameter for understanding the influence of solar irradiance variability on long-term climate trends \citep[e.g.,][]{Gray:2010,Solanki:2013}. The influence of the revised magnetic flux time series on TSI will be the subject of a future investigation.
Another important application of this model will be in reconstructing total magnetic flux and sunspot numbers from production rates of cosmogenic isotopes. Such an application will require the model to be inverted, as has successfully been done with the older version of the model \citep[e.g.,][]{Usoskin2003,Usoskin2004,Solanki2004,usoskin_AA_16,Wu18_composite}. Such work is ongoing and will be the subject of a follow-on publication.
\begin{acknowledgements}
We thank Mathew Owens for the helpful comments and stimulating discussion.
NAK and SKS acknowledge support by the German Federal Ministry of Education and Research (Project No. 01LG1909C). SKS received support from the Ministry of Education of Korea through the BK21 plus program of the National Research Foundation.
BH was supported by the International Max-Planck Research School (IMPRS) for Solar System Science at the University of Göttingen.
IU acknowledges a partial support from the Academy of Finland (Projects ESPERA no. 321882).
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
Although its negative influence and its weaponizing usage is known for ages \cite{soll2016long}, \emph{fake news} (in non-political context also known as \emph{false news}~\cite{wardle2017information}) and its negative impact was globally recognized, during the U.S. elections in 2016, as one of the major challenges for the society of the 21st century \cite{activities2017intentions,howell2013digital}. Importantly, it was used to promote both political campaigns in the election. Beside the promotion of political campaigns~\cite{barstow2008behind,allcott2017social}, fake news occurs with various purposes or due to various circumstances, e.g. to destabilize governments in third countries, accidentally due to unconscious misinterpretation of facts~\cite{quandt2019fake} or as a worthwhile revenue stream based on advertisement~\cite{waldrop2017news,kshetri2017economics,braun2019fake}.
Following~\cite{quandt2019fake} the term \emph{fake news} is used twice (a) to discredit and downgrade media and journalism; and (b) to summarize various forms of wrong, misguided, or fabricated information. Throughout this manuscript we are speaking about (b) when discussing fake news. Fake news articles, as just described, are to a large extend published, maintained, circulated and promoted in social media~\cite{wingfield2016google,koohikamali2017information,bovet2019influence,shu2019studying}.
On a high level, two strategies of potential interventions have been highlighted~\cite{lazer2018science}, (i) empowering of individuals to evaluate and assess fake news and (ii) structural changes preventing exposure of fake news to individuals. Most likely, machine learning based intervention strategies can be categorized into the second class of strategies. However, our intent is to propose an approach, while mainly based on state-of-the-art machine learning methodology, that can be categorized into the first class. Because we believe that the most sustainable strategy to fight the impact of fake news is to empower individuals to evaluate and assess fake news, we propose to assess content with evidence from reliable news sources supporting the hypotheses in the articles. We leave the final decision to the user which helps to improve acceptance because no actual censorship is happening. However, a quantitative statistical evaluation is still possible by simply adding a threshold from cross-validation experiments on top of the mechanism.
In summary our contribution can be structured into the following aspects:
\begin{itemize}
\item Modular system for the comparison of news articles from various sources.
\item Unsupervised approach for verification of a queried article and its content.
\item Automatic querying for supporting articles using News API.
\item An intuitive user interface which allows to individualize the collection of reliable sources and to receive visual feedback for the queried article.
\end{itemize}
The outline of the paper is as follows: Section \ref{sec:related_work} summarizes the main related work, highlighting prior approaches towards verification. In section \ref{sec:building_blocks}, we outline the relevant machine learning building blocks of our approach. Section \ref{sec:appoach_details} introduces our system for news verification and describes the workflow, architecture and user interface. The deployment of the solution architecture is described in section \ref{sec:architecture} as well as a discussion of the approach in general and its advantages in section \ref{sec:discussion}. Finally, in section \ref{sec:conclusion} we summarize the approach.
\section{Related Work}\label{sec:related_work}
Following the line of argumentation of \cite{zhou2020survey}, approaches to identify fake news can be structured into four major categories: knowledge-based, style-based, propagation-based and source-based
Propagation-based analyses are concerned with how fake news spread online which is mostly formulated as a binary classification problem. The input can be either a news cascade \cite{castillo2011information} or a self-defined graph \cite{jin2016news}.
For style-based analysis, the writing style is assessed according to malicious intent. Perez et al.~\cite{umich} point out stylistic biases that exists in text in order to automate fake news detection. Source-based approaches assess the credibility of a news source \cite{newsguard, baly2018predicting} while knowledge-based approaches compare news content with known facts \cite{botnevik2020brenda}.
According to the scheme from Zhou et al.~\cite{zhou2020survey}, our proposed approach combines two categories of fake news detection: source-based and knowledge-based analysis. For both, we highlight prior work.
The most prevalent source-based approach is to rate news sources on their credibility. Traditional source-based approaches are Web ranking algorithms which rely on website credibility to improve search results for user queries \cite{page1999pagerank}. Two current resources for news publisher credibility are MediaBias/FactCheck \cite{mediabiasfactcheck} and NewsGuard~\cite{newsguard}, a browser extension that displays ratings of news websites. The ratings are manually curated by journalists. Since the ratings go through a manual review process the list of rated websites is prone to be incomplete and quickly outdated. Therefore, recent efforts aim for automating source reliability ratings. Based on \emph{expert-}features including for example web-traffic, the existence of a verified Twitter account or textual information, the authors of~\cite{baly2018predicting} classify the news sources in a supervised manner using a Support Vector Machine.
Another approach to evaluate credibility of the knowledge is proposed by Esteves et al. \cite{jens}. The proposed approaches are based on supervised learning to automatically extract source reputation cues and to compute a credibility factor.
A further approach that also taps into style-based methods is to analyse text and metadata in the article. Rashkin et al. \cite{rashkin} assess the reliability of entire news articles by predicting whether the document originates from a website classified as hoax, satire or propaganda by comparing the language of real news with those three categories to find linguistic characteristics of untrustworthy text. Wang et al. showed that significant improvements can be achieved for fine-grained fake news detection when meta-data is combined with text \cite{wang}.
Knowledge-based approaches mostly tackle the process of fact-checking. Several fact-check-ing organizations such as CORRECTIV~\cite{correctiv}, PolitiFact~\cite{politifact} and Snopes~\cite{snopes} operate by manually verifying claims (see \cite{zhou2020survey} for more expert-based fact-checking websites). A drawback of manual verification is that it may reach readers too late. An approach tackling this issue was recently published~\cite{botnevik2020brenda} where the authors approached fact-checking with machine learning methods and focuse on claim verification.
Another related approach is presented in \cite{fever} where the authors introduce a Fact Extraction and VERification (FEVER) Shared Task. The aim is to classify whether a claim is factual or not by retrieving evidence from Wikipedia.
Both works treat the task as a classification problem, and a critical challenge with this approach is that we can not guarantee that the system is able to give suggestions to very recent claims.
An approach geared towards misinformation detection for social media treating exactly this challenge is presented in~\cite{hossain2020covid} by including a retrieval step.
While the authors still include classification as a second step, i.e. for stance detection, we completely omit any supervised task and focus on retrieval and an expert-knowledge-based scoring.
\section{Building Blocks of the Approach}
\label{sec:building_blocks}
As presented in more detail in section \ref{sec:appoach_details} we propose an evidence retrieval approach to handle fake news instead of classifying content. The learning task is formulated as an unsupervised machine learning problem. The evidence supporting the hypothesis of the queried article is gathered from a collection of reliable news sources which we provide to the user. It is individualized by selecting an arbitrary number of sources out of a curated list of reliable news sources for evidence-gathering purposes.
Technically, we propose a two-step process:
\begin{enumerate}
\item \emph{Aggregation}-step: Extract information from the given article and query for similar content from reliable sources
\item \emph{Refining}-step: Narrow the supporting evidence down by calculating the semantic distance of the text with the collection that was retrieved in step 1.
\end{enumerate}
In the following subsection we briefly introduce the most relevant machine learning concepts, forming the basis of the proposed approach. To calculate the semantic distance of news articles we rely on distributed word embeddings and the Word Mover's distance.
\subsection{Word Embedding}
Mikolov et al. \cite{mikolov2013efficient} proposed the Word2Vec algorithm to learn vector representations of words. The method is based on the distributional hypothesis \cite{harris54,Firth_Papers57} that words get their meaning from the context in which they appear. Mikolov et al. propose two different variations of the Word2Vec algorithm, both typically trained on large text corpora. The Continuous Bag-of-Words Model (CBOW) and the Continuous Skip-gram Model (skip-gram) which predict target words from source context words and source context words from target words, respectively. Specifically, they propose a shallow neural network architecture, which trains continuous word vectors representations to maximize the log probability of neighboring words in a corpus. For a given sequence of words $w_1, w_2, ..., w_N$, it models the probability of this particular sequence as follows
\begin{equation}
\frac{1}{N}\sum_{n=1}^{N}\sum_{j\in nb(n)} \log p(w_j\mid w_n)
\end{equation}
Here, $nb(n)$ is the set of neighboring words of the word $w_n$. The unsupervised training is done by optimizing the maximum likelihood of a corpus of sentences (sequences of words) such that the word embeddings capture the semantic information of words and relations between them, given a particular context. In their original work~\cite{mikolov2013efficient}, the authors approximated the objective above by more efficiently trainable objectives.
A flaw of Word2Vec is its inability to infer continuous representations for words not seen during training. Especially in domains such as news, new vocabulary can emerge rapidly. A simple way to account for that is to incorporate morphological information about words in the text representations. Bojanowski et al. \cite{bojanowski2016enriching} proposed \emph{fastText}, an extension of the skip-gram model, which learns word representations by including sub-word information. This is achieved by not only representing words with vectors but also the sub-word parts they consist of, bag of character \textit{n}-grams. Word vector representations are built as the sum of their sub-word and their own representation.
In this work, we experimented with two embedding models, Word2Vec and fastText embeddings.
Although there are by far more than two approaches available in the literature (also more advanced approaches like Transformers~\cite{vaswani2017attention}), see \cite{li2018word,almeida2019word} for comprehensive reviews, we focus on those because they can be efficiently implemented on standard hardware and are well-established in the NLP community.
Nevertheless, there is freedom in experimenting with other word embeddings as it only requires a change of the distance threshold. Hence, the approach can be easily adapted to support news verification for different languages.
\subsection{Word Mover's Distance}
Earth mover's distance (EMD), also known as the Wasserstein distance, is a distance measure between two probability distribution. Kusner et al. \cite{kusner2015word} proposed a version of EMD applicable to language models, the Word mover's distance (WMD) which evaluates the distance between two documents represented in a continuous space using word embeddings such as the aforementioned Word2Vec and fastText embeddings. For any two documents \textit{A} and \textit{B}, WMD is defined as the minimum cost of transforming document \textit{A} into document \textit{B}. Each document is represented by the relative frequencies of its words relative to the total number of words of the document, i.e., for the \textit{j}th word in the document,
\begin{equation}
d_{A,j} = count(j)/\mid A \mid
\end{equation}
where $\mid A \mid$ is the total word count of document A and $count(j)$ is number of occurrences of the word with vocabulary index $j$. The \textit{j}th word is represented by its corresponding word embedding, say $\mathbf{v}_j \in \mathbb{R}^{n}$. The $n$-dimensional word embeddings are obtained from a pre-trained model, e.g. Word2Vec or fastText. The distance between two words can easily be measured using Euclidean distance,
\begin{equation}
\delta(i,j) =\lVert \mathbf{v}_i - \mathbf{v}_j \rVert
\end{equation}
Based on this choice, the Word mover's distance is defined to be the solution of the following linear program,
\begin{equation}
\begin{aligned}
WMD (A,B) =\quad & \min_{\mathbf{T}\geq0}\sum_{i=1}^{V}\sum_{j=1}^{V}\mathbf{T}_{i,j} \delta(i,j)\\
\textrm{such that}\quad & \sum_{i=1}^{V}\mathbf{T}_{i,j}=d_{A,j} \\
\textrm{and}\quad & \sum_{j=1}^{V}\mathbf{T}_{i,j}=d_{A,i} \\
\end{aligned}
\end{equation}
Here, $\mathbf{T} \in \mathbb{R}^{V \times V}$ is a non-negative matrix, where $\textbf{T}_{i,j}$ denotes how much of word \textit{i} in document \textit{A} is assigned to tokens of word \textit{j} in document \textit{B}.
Empirically, WMD has reported improved performance on many real world classification tasks as demonstrated in \cite{kusner2015word}. The WMD has intriguing properties. The distance between two documents can be broken down and represented as the sparse distances between few individual words. The distance metric is also hyper-parameter free. The most important feature is that it incorporates the semantic information encoded in the word embedding space and is agnostic to arbitrary word embedding models.
\section{A Retrieval-Based Approach Supporting Fake News Identification Methods}\label{sec:appoach_details}
We constructed a pipelined system which helps in extracting semantically similar articles from reliable news sources. Its core is the analysis of the credibility of news articles based on the overall evidence collected from a set of automatically retrieved articles published by reliable news sources.
Figure \ref{fig:system} gives an overview of the components and workflow. The system consists of three components: a news content extractor, a search engine query and a content analyzer. All of these components can easily be exchanged and extended depending on the language and the list of reliable news sources.
\begin{figure}
\centering
\includegraphics[scale=0.78]{workflow_color_newnames.pdf}
\caption{Overview of the system's workflow with aggregation and refining step.}
\label{fig:system}
\end{figure}
\subsection{Article Extractor}
Given a link to an article that should be verified the article extractor component extracts information from the link such as the publication date, the article title, its authors and the textual content. For most of the news sources, the python library Newspaper3k\footnote{https://newspaper.readthedocs.io/en/latest/} is suitable and manages to extract all of the above information. However, at least for some sources, we built our own article extractors by parsing the HTML page of the article and extracting certain tags.
We extract relevant keywords and entities from the title and body of the article using the Newspaper API keyword extractor. The goal of this step is to get as much relevant information characterizing the article as possible. This extracted information is used in building an automated query.
\subsection{Querying the Google News API}
To obtain news articles from reliable news sources we use a query component which queries the Google News API. Using the keywords and entities we obtained in the step before, we construct a query. We structure the query so that we can filter the articles based on date, number of requested news articles from the sources, location and language. The API returns ten article links based on the search criterion from every source selected. The number of articles returned by the API can be changed based on individual requirements and computation power. The article extractor component extracts the content of the articles obtained from the search API.
The system offers six news sources and we can easily add new sources or remove existing ones from the list. Automatic querying used here is different from manual news search using search engines as we aggregate news based on dates, keywords extracted from the article, and selected reliable sources.
\subsection{Content Analysis: Semantic Distance Analysis}
The content analysis component computes the semantic distance between the query article and the articles returned by the query component. Before computing the distance score, we clean the article titles and bodies by removing stop words and special symbols and computing their bag of n-grams representations. The semantic distance of articles is calculated using word embeddings and the WMD. For the word embeddings, we experimented with different word embeddings such as fastText and the pre-trained Google news embeddings. The quality of the word embeddings depends on the size of training data, thus, we use pre-trained word embeddings.
Since the original WMD is computationally expensive, we approximate the distance by using the Regularized Wasserstein distance proposed by \cite{regWMD} and only keep the five closest articles. The five articles with the least distance are then selected for computation with the original WMD. The WMD returns a distance score for each remaining article from the individual sources. The smaller the distance, the more related the articles are. Only articles that are below a predefined threshold are considered as similar to the given article. We set the distance threshold by empirically checking the distances of a couple of articles. Similar news articles, i.e. articles that fall below the distance threshold, are then displayed with a message that closely related articles were found. If the system does not return similar articles the reader is informed that the given article is potentially fake.
Our prototype was exemplary tested on a small set of articles. A systematic evaluation with a self-curated dataset and the FakeNewsNet \cite{data} dataset is planned.
The semantic distance analysis in our approach is based on unsupervised models which in turn make the system highly adaptable to different languages. We just need to replace the word embeddings and adapt the threshold.
Furthermore, the unsupervised nature renders the approach agnostic to concept drifts which means that the machine learning task is independent of the hypotheses in a text.
\section{Architecture and Deployment}\label{sec:architecture}
\begin{figure*}
\centering
\includegraphics[scale=0.9]{Architecture_FND_cropped.png}
\caption{The technical architecture of the Fake News Detector with three docker components.}
\label{fig:architecture}
\end{figure*}
To showcase our approach we build a "Fake News Detector" system. The Fake News Detector system
consists of a few technical components. Its technical architecture is based on a set of docker components (see Figure \ref{fig:architecture}). In detail, there are the following three docker components:
\begin{enumerate}
\item A container with simple django running the python code of our application and serving the frontend.
\item A container serving the data for the backend - the model container.
\item A container that includes data pre-processed by several NLTK functions.
\end{enumerate}
The Fake News Detector application can be accessed via web UI (see Figure \ref{fig:ui}). In the UI, we can insert a link of a news article to be verified, in this example, we want to verify an article titled \textit{Gatorade banned and fined \$300k for bad-mouthing water}. Next, we select the news sources to check and match against. By clicking on the verification button the analysis process is started at the backend running all the components. After the analysis, the results are shown as a list of potentially matching articles. If no matching articles are found after the analysis, a message is displayed that the article might be potentially fake. In the example shown in Figure \ref{fig:ui}, we selected all six sources. The system queried against all sources and analysed the potentially matching articles using semantic similarity and found that CNN has published a similar news article during the same time frame.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{checkArticles_cropped.jpg}
\caption{The User Interface of the Fake News Detector. In this example, the query article is titled \textit{Gatorade banned and fined \$300k for bad-mouthing water}. The user pastes the URL into the text box, selects sources and receives a matching article from the source CNN. }
\label{fig:ui}
\end{figure*}
\section{Discussion and Future Work}\label{sec:discussion}
In the previous sections, we have addressed various benefits of following such an unsupervised approach. These benefits are the plasticity of the system, its modularity and user-driven decision making. In this section, we discuss several challenges and insights for the future work.
One challenge for the demonstrator is to deal with very recent news. The system will not be able to collect semantically similar articles from other reliable news sources that might not have published yet on the subject. Here, date as well as publishing time become important. For future work, we propose to highlight this fact to the user to avoid perceiving recent news as potentially fake.
Another challenge in analyzing news articles is novel words which are out-of-vocabulary. For these words no word vectors exist which in turn affects the WMD score. The system can handle this scenario but the performance may be impaired. To overcome this we are crawling news from major news sources every hour to build a text dataset. We propose to update word embedding models frequently to avoid an increased number of out-of-vocabulary words in recent news articles.
Recently, Yokoi et al. \cite{yokoi2020word}, proposed an improvement on Word Mover's Distance, namely Word Rotator's Distance (WRD) which measures the degree of semantic overlap between two texts using word alignment. This approach is designed such that the norm (a proxy for the importance of word) and angle of word vectors (a proxy for dissimilarity between words) correspond to the probability mass and transportation cost in EMD, respectively. This approach outperforms the WMD in several semantic textual similarity tasks. This alternative might help in improving the document distance threshold and is subject for our planned evaluation.
We also propose to use the approach of document similarity for related use cases where we see potential in two directions.
The first direction is helping fact-checkers by providing adequate evidence to verify hypotheses. By providing them similar content, e.g. evidence in the form of news but also scientific articles, the system can support their task.
Second, reviewers in several domains, e.g. medical health news review, need to determine how comprehensible a text document is. By comparing a text to scientific or more simple language it is possible to provide a comprehensibility score.
Since the system alone does not guarantee news verification or falsification we recommend considering combining it with fact-checking methods.
\section{Conclusion}\label{sec:conclusion}
We presented a system to find semantically similar articles to a given news article from selected reliable sources. For the system, we propose an evidence retrieval approach to handle fake news instead of treating it as a classification task. This way, we aid the users in finding supporting evidence and, thus, manual search work can be reduced.
The benefits of our system are that (i) it is unsupervised and therefore agnostic to concept drifts, (ii) it gives the user decision power and (iii) it is modular, i.e. the system can be easily adapted to other languages, extended and improved with further components.
\begin{acknowledgments}
This research has been partly funded by the Federal Ministry of Education and Research of Germany as part of the Competence Center for Machine Learning ML2R (01|S18038B).
T. Wirtz contributed as part of the Fraunhofer Center for Machine Learning within the Fraunhofer Cluster for Cognitive Internet Technologies.
\end{acknowledgments}
|
\section{Introduction}
Real-world machine learning approaches not only must be able to perfectly learn the desired concept (\textit{e.g.}, training data samples), but also should be adaptable to learn new data classes continually and gradually. For instance, robots usually face new knowledge (\textit{i.e.}, tasks) which are essential to be learned constantly \cite{mi_generalized_2020}. In such settings, Class Incremental Learning (CIL) is indispensable to properly deal with data from unknown classes. It is of considerable importance especially when the Artificial Intelligence (AI) agent is confronted with the lack of computational or memory capacity \cite{belouadah_scail_2020}. Therefore, the unavailability of the previous data instances is probable and needs to be addressed in order to have an effective method in this field of research. CIL has drawn the attention of many researchers in recent years due to its crucial role in real-world AI applications.
\begin{figure*}[t]
\center
\includegraphics[width=1\linewidth,scale=0.1]{ClaRe-Overview.pdf}
\caption{The outline of our proposed method, ClaRe, in the training phase which includes both $\mathcal{D}$ and $\mathcal{R}$ networks. $\mathcal{D}$ consists of a feature extractor ($\mathcal{D}_{F}$) alongside a FC network ending with a soft-max layer on top forming $\mathcal{D}_{C}$. $\mathcal{D}$ has a duty to determine which class the input data belongs to. Moreover, $\mathcal{R}$ intends to generate the old data via remembering the representations have been stored so far. Three goals are pursued during the training process: (1) minimizing the reconstruction error, \textit{i.e.}, $X \approx X'$, (2) enforcing the distributions returned by $\mathcal{D}_{\text{F}}$ to be close to a standard normal distribution ($\mathcal{N}(\mu_{z},\sigma_{z})$) and (3) enabling $\mathcal{D}$ to precisely distinguish between data instances of different classes. All of the distributions are conditioned with $c$ variable to uniquely specify the class of input samples. This approach improves the performance of $\mathcal{D}$ while $\mathcal{R}$ is capable of simulationary generating more real data to overcome the catastrophic forgetting as the most challenging problem of CIL.
}
\label{fig:overview}
\end{figure*}
Previously proposed state-of-the-arts can be fallen into three major categories \cite{zhao_maintaining_2020}: (1) parameter control, (2) knowledge distillation and (3) rehearsal. The key idea of parameter control strategy is to confine the learned parameters (\textit{i.e.}, weights) of the old model in confronting with learning new tasks \cite{huang_densely_2017}. Since the importance of parameters is difficult to measure accurately in a series of tasks, this strategy usually performs poorly in CIL. Knowledge distillation is another solution which is widely-used to cope with the catastrophic forgetting in the CIL problem. In this category, the main knowledge is transferred to a student model from a teacher model by facing new data. Whereas, the capabilities of the old model are appropriately preserved \cite{li_learning_2018}. The rehearsal strategy attempts to alleviate the catastrophic forgetting by utilizing the original data to reproduce the training set. Most of the methods belonging to this strategy store a limited number of original training samples to retrain the deep neural network which should learn new classes \cite{hou_learning_2019}. Generating original instances based on a generative process \cite{wu_memory_2018, zhu_unpaired_2017}, \textit{e.g.}, Generative Adversarial Networks (GANs) \cite{goodfellow_generative_2014} that become a hot research topic in AI \cite{kang_effective_2020}, is another approach of the rehearsal strategy. However, generative models prefer producing previous samples to storing them and thus this approach highly relies on the quality of the generated data. Moreover, training an additional generative model simultaneously is considered expensive for an algorithm. Note that some other solutions by the combination of the above-mentioned strategies are presented by researchers. For example, \cite{rebuffi_icarl_2017} has applied both the distillation and the rehearsal strategies. Also, \cite{castro_end_2018} has exploited these two strategies along with using a balanced fine-tuning to alleviate class imbalance.
To overcome the aforementioned challenges such as catastrophic forgetting, imbalance data and memory limitation, we propose a novel method, ClaRe, to effectively tackle the issue of catastrophic forgetting when acquiring and accumulating new knowledge is necessary. The key idea of ClaRe is to remember a lightweight representation for each class that must be learned over time while storing none of the original data in each increment is required at all. Accordingly, our proposed method is not restricted by the memory limitation. Basically, ClaRe is one neural network constituting of two relevant sub-modules, one of which is responsible for detecting the correct class of each input sample and the other memorizes precise representations for each of classes the model is confronted with. These networks support each other to play their roles in the best way. The proposed solution is very close and related to the rehearsal strategy with the approach of using generative models which is considered very promising for CIL tasks. Although ClaRe is also implementable by the conditional GAN, we merely exploit one neural network formed by two sub-modules that are trained simultaneously. Whereas using conditional GAN necessitates training two different neural networks in order to concurrently learn the concept of the input data in addition to a classifier for the task of classification.
In a nutshell, the most important contributions of ClaRe are: (1) solving the catastrophic forgetting problem caused by unavailability of the original training data, (2) providing generalization due to producing different data instances of one class, (3) imbalance class issue does not exist any longer owing to the optional number of simulated data that can be generated by our model, (4) overcoming the probable memory limitation in so many real-world applications and (5) results confirm the effectiveness of ClaRe, since it behaves very stable after adding new classes (data samples).
\section{Proposed Method}
ClaRe is a practical and effective solution for incrementally and continually learning new knowledge in such a way that there is no need to store the old data belonging to classes have been learned thus far. To this end, the main concept of original classes should be memorized. Generally, the proposed method is composed of two sub-networks: (1) $\mathcal{D}$iscriminator and (2) $\mathcal{R}$emembering network. $\mathcal{D}$ tends to correctly differentiate the input data samples and it also helps $\mathcal{R}$ to accurately memorize the representations. At the same time, The other network ($\mathcal{R}$) attempts to remember the concept of the leaned classes prior to the beginning of each training phase (except the first stage) and to memorize the representation of each class during the training process. Inspired by the Conditional Variational Auto-Encoder (CVAE) \cite{sohn_learning_2015}, these two neural networks are jointly learned on available training samples. When the trained model faces new classes, data instances following the distribution of the previous class data are generated by $\mathcal{R}$. $\mathcal{D}$ is subsequently retrained on the generated samples along with the data of new classes. A sketch of ClaRe is depicted in Fig. \ref{fig:overview}. $\mathcal{D}$ and $\mathcal{R}$ networks are explained in the following subsections.
\subsection{$\mathcal{D}$ Network}
Distinguishing between samples of various classes is the main duty of $\mathcal{D}$. This network comprises two major parts, a Feature extractor (F) and a Classifier (C). From now on, F and C are denoted by $\mathcal{D}_{F}$ and $\mathcal{D}_{C}$, respectively. $\mathcal{D}_{F}$ is constituted of two layers with the sizes of $512 \times 1$ and $256 \times 1$. $\mathcal{D}_{F}$ is trained in interaction with $\mathcal{R}$ to improve the quality of the representations should be stored by our model for each class. $\mathcal{D}_{C}$ consists of a Fully-Connected (FC) layer with the size of $256 \times class\_no$ which ends with a soft-max layer on top. $class\_no$ is the total number of classes including both old and new ones.
\begin{table*}
\centering
\renewcommand{\arraystretch}{1.1}
\caption{Classification accuracy of our method, ClaRe, in the groups of $g = 1,~2~or~5$ classes at a time.}
\label{tab:classification_accuracy}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{$g$} & \multicolumn{10}{c|}{$Accuracy~(\%)$} \\
\cline{2-11}
\multirow{2}{*}{} & Class 0 & Class 1 & Class 2 & Class 3 & Class 4 & Class 5 & Class 6 & Class 7 & Class 8 & Class 9 \\
\noalign{\hrule height 1.5pt}
\hline
1 & 100 & 99.9 & 98.6 & 95.2 & 93.4 & 89.5 & 87.6 & 83.5 & 81.3 & 78.6 \\
\hline
2 & \multicolumn{2}{c|}{99.9} & \multicolumn{2}{c|}{96.7} & \multicolumn{2}{c|}{95.6} & \multicolumn{2}{c|}{91.8} & \multicolumn{2}{c|}{88.2} \\
\hline
5 & \multicolumn{5}{c|}{96.3} & \multicolumn{5}{c|}{92.7} \\
\hline
\end{tabular}
\vspace{-3mm}
\end{table*}
\subsection{$\mathcal{R}$ Network}
$\mathcal{R}$ is a two-layer neural network with the sizes of $256 \times1 $ and $512 \times 1$. The input layer of $\mathcal{R}$ is imposed to follow a specific type of normal distribution. Note that the last FC layer of $\mathcal{D}_{F}$ is equivalent to the input layer of $\mathcal{R}$. In this way of approaching the problem, $\mathcal{R}$ is capable of generating the original samples that have already been classified by $\mathcal{D}$ (see the subsection \ref{sec:training} for more details about the training process). These two neural networks as a whole, \textit{i.e.}, $\mathcal{R} + \mathcal{D}_{F}$, can be considered as an encoder-decoder neural network which is trained to generate previously seen but unavailable instances by the learned model. Nevertheless, applying a minor alternation to the encoding-decoding process is necessary to introduce some regularization of the latent representation. Consequently, an input is encoded as a distribution over the latent space instead of a single point. In this case, data generation can be done by taking a point randomly from that latent space and decode it to get new content. It is worth mentioning that, generalizability which is one of the most important features for simulationary generating previous class data, is also met by taking this solution, since for each class generated samples are not the same.
\subsection{Training ClaRe: $\mathcal{D} + \mathcal{R}$}
\label{sec:training}
$\mathcal{D}_{F}$ and $\mathcal{R}$ are trained jointly and simultaneously. An overview of ClaRe in the training stage is illustrated in Fig. \ref{fig:overview}. As stated previously, inspired by CVAEs, $\mathcal{D}_{F}$ and $\mathcal{R}$ acts as an encoder-decoder network. As a consequence, $\mathcal{D}_{F}$ is able to precisely learn the concept of different input classes and $\mathcal{R}$ memorizes their representations during the training procedure. In this case, $\mathcal{R}$ is able to remember the old unavailable data and then generates them just before the start of the training stage to facilitate this process as well as enhance the detection accuracy. Note that the training process of $\mathcal{D}_{F} + \mathcal{R}$ at the first step slightly differs from the rest. From the second step on, we need to simulate the prior training data that is not the case in the first phase. Therefore, they are able to generate appropriate amount of data. This helps us to cope with the imbalance data problem. The training of $\mathcal{D} + \mathcal{R}$ is done by optimizing the
\begin{align}
\label{eq:total_loss}
\mathcal{L}_{Total} = \mathcal{L}_{\mathcal{D}} + \mathcal{L}_{\mathcal{D}_{F}+\mathcal{R}}
\end{align}
The loss function that is minimized during the training of $\mathcal{D}_{F} + \mathcal{R}$ (\textit{i.e.}, $\mathcal{L}_{\mathcal{D}_{F} + \mathcal{R}}$) is constituted of a reconstruction term on the final layer of $\mathcal{R}$ that makes the encoding-decoding scheme as efficient as possible and a regularization term on the latent space, that has a tendency to regularize the organization of the latent representation by making distributions returned by $\mathcal{D}_{F}$, close to a standard normal distribution. This regularization term is expressed as the Kullback-Leibler (KL) divergence between the returned distribution and a standard Gaussian. In fact, we need to have control on data generation process to create data form a specific class. To this end, the label of the input data must be taken into account. Hence, $\mathcal{D}_{F}$ models the latent space ($z$) based on the input data ($X$) and another variable ($c$), \textit{i.e.}, all of the distributions are conditioned with a variable $c$. Accordingly, for each possible value of $c$, there is a $\mathcal{R}(z|X,c)$. $\mathcal{L}_{\mathcal{D}_{F}+\mathcal{R}}$ is calculated by
\begin{equation}
\small
\begin{split}
\label{eq:vae_loss}
\log & \mathcal{R}(X|c) - KL(\mathcal{D}_{F}(z|X,c)\|\mathcal{R}(z|X,c)) = \\
& \underbrace{\mathbb{E}_{z \sim \mathcal{D}_{F}} [\log{\mathcal{R}(X|z,c)}]}_{\text{Reconstruction Error}} + \underbrace{KL(\mathcal{D}_{F}(z|X,c)\|\mathcal{R}(z|c))}_{\text{Regularization}}
\end{split}
\end{equation}
Where $z \sim \mathcal{N}(\mu_{z},\sigma_{z}) = \mathcal{D}_{F}(X)$ and $c$ is a one-hot encoding to determine each class individually. When the training procedure is completed successfully, $\mathcal{D}$ is able to accurately classify the input instances and $\mathcal{R}$ can effectively simulate the old data from the second phase of training onward.
\subsection{Testing ClaRe: $\mathcal{D}$}
For the evaluation of our work, merely $\mathcal{D}$ including both major parts, \textit{i.e.}, $\mathcal{D}_{F}$ and $\mathcal{D}_{C}$, is exploited to classify the input data which is present in the test set. Note that remaining modules are not useful until the next training stage when the model is confronted with samples from new classes. Thus, our model should be retrained on both new data and generated samples.
\section{Experiments}
In this section, a thorough evaluation is carried out to show the high performance of ClaRe. The scenario and the setting for assessing our approach are provided in the subsection \ref{sec:res}. ClaRe has been implemented using Pytorch\footnote{https://pytorch.org} on GOOGLE COLAB\footnote{https://colab.research.google.com}.
The results confirm that the proposed method outperforms the other state-of-the-arts.
\subsection{Experimental Setup}
\label{sec:res}
ClaRe is evaluated on the MNIST benchmark \cite{mnist}. Generally, learning classes incrementally by a reasonable step is very common among the proposed methods for the evaluation process. Therefore, our model learns all 10 classes in groups of $g = 1,~2~or~5$ classes at a time. For example, when $g = 1$ in the test scenarios, our model faces new classes one by one, \textit{i.e.}, one class is added in each increment. The evaluation metric is the multi-class classification accuracy on the test set while the best results are reported in each experiment.
\vspace{-3mm}
\subsection{Experimental Results}
Table \ref{tab:classification_accuracy} shows the final results of our method. As can be seen, ClaRe is very stable against facing new knowledge or tasks owing to the very little degradation on accuracy after learning new class data. The reported results show the high functionality of $\mathcal{R}$ to generate old samples belonging to the prior classes. The insensitivity of our work to the increase of classes makes ClaRe more scalable and also more reusable.
To show the superiority of our proposed method, we compare ClaRe with the baselines and also the other stat-of-the-arts which has been presented by \cite{wu_memory_2018}. As upper and lower bounds, joint training with all data (\textit{i.e.}, non-sequential) and sequential fine tuning are considered. Additionally, there are two other state-of-the-art methods to make a comparison with our work comprising the adaptation of elastic weight consolidation to conditional GANs proposed by \cite{seff_continual_2017} and the deep generative replay module of \cite{shin_continual_2017} implemented as an unconditional GAN followed by a classifier to predict the label. sequential fine tuning completely forgets previous tasks in the dataset, while we can observe different degrees of forgetting in the other solutions. Table \ref{tab:avg_accuracy_comparison} shows the average classification accuracy of each baseline alongside the other approaches including ClaRe. Obviously, the accuracy of our work is lower than joint training, since it shows the best possible result (upper bound) for CIL task in our test scenario while it outperforms the other state-of-the-arts by a large margin. Results confirm the high performance of ClaRe for both 5 and 10 sequential tasks. The term of sequential tasks refers to learning classes incrementally. Indeed, classes are added one by one in each increment and then the accuracy is calculated. Finally, the average value is reported for 5 and 10 classes. As it is clear in this table, our method is very close to the upper bound especially when the first five data classes are considered.
\begin{table}
\small
\centering
\renewcommand{\arraystretch}{1.1}
\caption{Average classification accuracy comparison with baselines along with the other state-of-the-art methods for 5 and 10 sequential tasks. The best result is boldface underlined.}
\label{tab:avg_accuracy_comparison}
\begin{tabular}{ccc}
\hline
\multirow{3}{*}{Method} & \multicolumn{2}{c}{$Accuracy~(\%)$} \\
\multirow{3}{*}{} & 5 tasks & 10 tasks \\
\multirow{3}{*}{} & (0-4) & (0-9) \\
\noalign{\hrule height 1.5pt}
Joint training (upper bound) & 97.7 & 96.9 \\
Sequential fine Tuning (lower bound) & 19.9 & 10.1 \\
\hline
Elastic weight consolidation \cite{seff_continual_2017} & 70.6 & 77.0 \\
Deep generative replay \cite{shin_continual_2017} & 90.4 & 85.4 \\
ClaRe (ours) & \textbf{\underline{97.4}} & \textbf{\underline{90.8}} \\
\hline
\end{tabular}
\vspace{-4mm}
\end{table}
\section{Complexity}
As previously discussed, our proposed method is also implementable using conditional GAN. Therefore, we need to concurrently train two individual networks, generator and discriminator, in competing with each other. Moreover, training a classifier is essential to classify the input data. Instead of training three different modules, ClaRe is merely constituted of one neural network which includes two sub-modules, $\mathcal{D}$ and $\mathcal{R}$. Although ClaRe has an outstanding performance in both providing the model with high quality generated old samples and classifying the input data, the process of simulating original instances is time-consuming. Designing a module to effectively learn the feature space of the input data improves the model in terms of accuracy as well as the time complexity. In this case, confronting with new knowledge necessitates modifying the feature space so that it embraces the concept of new classes. However, introducing such modules can be considered very challenging and needs to be thoroughly investigated as a promising future work.
\section{Conclusion}
In this paper, we propose a practical and novel method, ClaRe, to effectively address the catastrophic forgetting problem when our model faces new knowledge or unseen concepts. ClaRe is composed of two correlated sub-modules, $\mathcal{D}$iscriminator and $\mathcal{R}$emembering network. Furthermore, $\mathcal{D}$ consists of two major parts, a feature extractor and a classifier which are known as $\mathcal{D}_{F}$ and $\mathcal{D}_{C}$, respectively. Note that $\mathcal{D}_{F}$ and $\mathcal{R}$ are trained jointly. $\mathcal{D}$ is responsible for distinguishing between the existing classes while $\mathcal{R}$ remembers the representations of previous data classes to appropriately generate old samples. The reported results confirm the superiority of our method compared to the other state-of-the-arts.
\bibliographystyle{ieee_fullname}
|
\section*{Abstract}
As cities expand, human mobility has become a central focus of urban planning and policy making to make cities more inclusive and sustainable. Initiatives such as the ``15-minutes city'' have been put in place to shift the attention from monocentric city configurations to polycentric structures, increasing the availability and diversity of local urban amenities. Ultimately they expect to increase local walkability and increase mobility within residential areas.
While we know how urban amenities influence human mobility at the city level, little is known about spatial variations in this relationship.
Here, we use mobile phone, census, and volunteered geographical data to measure geographic variations in the relationship between origin-destination flows and local urban accessibility in Barcelona.
Using a Negative Binomial Geographically Weighted Regression model, we show that, globally, people tend to visit neighborhoods with better access to education and retail. Locally, these and other features change in sign and magnitude through the different neighborhoods of the city in ways that are not explained by administrative boundaries, and that provide deeper insights regarding urban characteristics such as rental prices.
In conclusion, our work suggests that the qualities of a 15-minutes city can be measured at scale, delivering actionable insights on the polycentric structure of cities, and how people use and access this structure.
\section*{Introduction}
Rates of urbanization have been rapidly increasing across all world's regions over the last century~\cite{united20182018}. Currently more than half of the global population live in urban areas and increasingly in highly-dense cities~\cite{united20182018}. Cities are projected to absorb more than two-thirds of the growth in global population, with 68 percent of the world’s population living in urban areas by 2050~\cite{desa2019world}. Making cities inclusive, safe, resilient, and sustainable has thus become a global priority~\cite{griggs2013sustainable}.
Human mobility within cities is recognised as a key area of policy intervention to achieve these goals. Prior research has established a connection between mobility within cities and urban form~\cite{gordon1989influence, muniz2005urban}. As cities expand, average travel time and distance have increased giving rise to a debate regarding the optimal city size and structure~\cite{ewing2015compactness, kirkley2018betweenness}. Urban sprawl is considered to be detrimental to city livability and sustainability, inducing longer within-city travel~\cite{batty2003traffic}, increased private car usage~\cite{brueckner2000urban, glaeser2004sprawl}, congestion~\cite{bento2005effects}, obesity~\cite{ewing2003relationship}, and pollution~\cite{anderson1996urban}. We also know that longer commutes are more prevalent in certain populations, particularly highly educated, affluent, male and dual-earner populations~\cite{green1999longer, white1986sex}. Yet distinctive patterns exist in less developed and unequal societies in which the poorest segments of the population tend to commute longer distances~\cite{lucas2012transport, lucas2016transport}.
To promote more localised mobility patterns, current planning strategies have used smart growth interventions to move away from monocentric city structures to polycentric configurations~\cite{brezzi2015assessing, veneri2015urban}, with ideas such as a 15-minutes city gaining salience in policy circles. In essence, interventions promoting polycentricity seek to increase the offer of local amenities, such as schools, public transport options, health care facilities, food outlets, jobs, recreational areas, and retail shops, creating sustainable, inclusive, and walkable local areas within a small radius~\cite{veneri2015urban, artmann2019smart}. The exploration of local accessibility, regional accessibility and transit access opportunities is of utmost importance for the improvement of livability in cities. Previous studies suggest the need to focus on planning for pedestrian-scale neighborhoods to enhance local accessibility, adapt travel behaviors, and reduce car ownership and vehicle miles traveled~\cite{krizek2003residential, Zhang2020}.
Existing research indicates that spatial concentration of job, public services and commercial activity influences human mobility in cities, giving rise to destination hotspots or multiple centres~\cite{roth2011structure, louail2014mobile, bassolas2019hierarchical}. Yet, less is known about how much the availability and diversity of local amenities delivers the goal of retaining local population locally, and the ways in which the relationship between local amenities and mobility intensity varies geographically and interacts with the socio-demographic profile of resident populations.
To address these gaps, we aim to model how local amenities contribute to origin-destination flows of the resident population in a city; to quantify the extent of spatial variability in this influence; and, to assess emerging patterns associated with the socio-demographic profile of local populations in Barcelona, Spain. Traditionally, surveys and travel diaries are used to solve these tasks. While valuable, these data sources are expensive, infrequent, offer low population coverage, lack statistical validity at fine geographical scales, and become available with a lag of one or two years after collection~\cite{hidalgo2008dynamics, tao2018travel}. Mobile phone data has emerged as a novel source to capture human mobility~\cite{gonzalez2008understanding}. While they present shortcomings, mobile phone data are generated at unprecedented temporal frequency and geographical granularity. They record data for the selected population of mobile phone users and raise concerns of personal privacy and confidentiality. Ensuring anonymization, mobile phone data enable developing a more refined spatio-temporal and contemporary understanding of human mobility~\cite{de2013unique}.
A collaboration between public and private entities, the Barcelona city council and Vodafone, has enabled us to access aggregated mobile phone data of Barcelona to capture its local patterns of human mobility. Barcelona provides a unique study area to investigate the geographical variability in the way urban amenities influence local patterns of mobility. Over the last two decades, and more recently pushed by the COVID-19 pandemic, various initiatives have been put in motion to deliver the concept of self-sufficient neighborhoods, or the 15-minutes city, in which local areas have the required offer of urban amenities to be accessed within a short distance in a residential area (one example of this is the distribution of food markets in the city~\cite{provansal1992els}). Yet, little empirical evidence exists as to the extent that local urban amenities reduce the intensity and distance of mobility. We draw from volunteered geographical information data from OpenStreetMap (OSM) to measure the availability and accessibility to local urban amenities or Points of Interest (POI), jointly with data from the Spanish population census to characterize the socio-demographic composition of local populations. Then, to assess the influence of urban amenities on mobility patterns, we first use a Negative Binomial (NB) regression model to identify the key system-wide urban amenities relating to low mobility intensity (or high retention rates) and how mobility patterns vary across population groups. Second, we apply a NB Geographically Weighted Regression (GWR) model to assess the extent and direction of spatial variation in these relationships. The result of this analysis is a set of vectors that characterizes the accessibility and mobility of each neighborhood, which may be different from the global city patterns. We also demonstrate how these vectors open further local analyses, by studying their relationship with rental prices in the city.
The paper is structured around three main sections. First, we provide a description of the data sources to capture mobility, accessibility to urban amenities and characterize the socio-demographic composition of the local resident population. We also discuss the statistical modeling approach employed to measure the global and local relationship between urban amenities and mobility flows. Second, we present and discuss the results and their implications. Third, we summarize the key insights from our analysis and discuss potential avenues for future work.
\section*{Materials and methods}
With the objective of characterizing the mobility of the residents in a city, here we introduce the datasets and methods used in this paper. Then, we describe how to estimate accessibility to several types of amenities and services within each neighborhood, and how these measurements, jointly with census and mobile phone data, enable the identification of local patterns. Finally, as an example application, we use these patterns to explore the relationship between mobility and rental prices.
\subsection*{City data}
The Barcelona municipality, part of the Barcelona Metropolitan Area, is divided into 73 neighborhoods.
The city data consist of the census available from the \emph{Ajuntament de Barcelona} (city council) and the \emph{Instituto Nacional de Estadísticas} (INE, National Statistics Institute). For each neighborhood we have the following mean values for 2018: a number that characterizes income and social well-being denoted Human Development Index (HDI), age, population, share of population registered as women, share of immigrant population, mean rental price, and mean rental price per square meter (see Fig~\ref{fig:bcn_neighborhoods} for their spatial distributions).
The distributions of each variable reveal large spatial heterogeneity in income, rental price, touristic attractiveness (defined as log-transformed visitation density from mobile phone data during August 2018, see next subsection), and immigration ratio. These variations suggest that a local approach to understand mobility may be suitable.
\begin{figure}[!h]
\includegraphics[width=\linewidth]{Fig1.png}
\caption{{\bf City data for the neighborhoods of Barcelona.}
Choropleth maps of neighborhood features in Barcelona.
a) Mean socio-economic characterization, Human Development Index (HDI; unitless, from 0 to 10). b) Mean age (years). c) Resident population count. d) Touristic attractiveness. e) Population percentage registered as women. f) Ratio between immigrants and nationals. g) Mean rental price in Euros. h) Mean rental price per square meter in Euros.
Administrative boundaries are sourced from CartoBCN under a CC BY 4.0 license, with permission from Ajuntament de Barcelona, original copyright 2020.
}
\label{fig:bcn_neighborhoods}
\end{figure}
\subsection*{Aggregated mobile phone data flows}
The geographically aggregated mobile phone data were provided by the Barcelona city council, originally collected by Vodafone, one of the largest mobile phone operators in Spain (estimated 25\% to 30\% market share). The dataset contains the aggregated number of visitors with registered active mobile phone operations within the city limits in four-hour time frames, daily.
Visitors were segregated by the operator in the following types: residents per neighborhood in the Barcelona municipality, commuters from nearby cities, regional tourists, national tourists, and foreign tourists.
Our work focuses on residents, for whom we can map the patterns in destination visits according to their neighborhood of origin.
We focus our analysis on the {\em most neutral} month (i.e., neither winter or summer breaks, nor containing too many holidays or special events). The selected month is June for two reasons. First, it is a neutral month with no special events (see S1 Appendix). Second, discarding special months, it presents the best goodness-of-fit for the mobility models (details on the next section).
In geographical terms, our dataset contains two types of geolocation: while most of the city data (previous subsection) is associated to neighborhoods or political boundaries, the visitor counts are aggregated in a regular grid (212 cells of $500~m^2$, see details in Ref.~\cite{graells2020measuring}), thus defining a bipartite origin-destination (OD) graph between neighborhoods of residential origin and cells of visitor destination. To work with a neighborhood-to-neighborhood network, we assigned the number of visitors in grid cells to their corresponding neighborhoods by weighting the total number of visitors by the intersection of each cell with the neighborhoods. Note that the weight is calculated by the area of the intersection, not by its population, as the mobility data contains visitor counts to each cell. This count is not necessarily related to the static population of each neighborhood or cell.
We are interested in the average mobility within a month, as we seek to find generalizable insights. To do so, we modelled the average daily origin-to-destination mobility count.
There is a remarkable number of null OD pairs (61\% on average), which may be explained due to anonymization purposes in the data: if a cell contained less than 50 people in a given time-frame, it was reported as 0. This implies that some infrequent flows have zeroes in both observations for business days and weekends. We removed those OD pairs, and only kept zeroes when either business days/weekends was non null, as this may characterize work-centric areas, among other situations. The filtered amount of zeroes was reduced to 53\% on average (see the obtained distribution of OD counts in Fig~S2 in S1 Appendix).
For illustration purposes, we estimated an overall importance of origin and destination for each neighborhood, defined as the square root of the mean of the number of incoming/outcoming flows in business days (see Fig~\ref{fig:neighborhoods_and_grids} (a) and (b)).
The overall destination popularity highlights the historical center, but also the port area, which has almost no inhabitants, but receives many visits due to workers, and people going to/from the airport. We also estimated a popularity score for August 2018 (summer holidays) considering only roaming-plan visitors, and considered it as a touristic attractiveness score for each neighborhood due to being the month with the highest tourist influx in the city (see its spatial distribution in Fig~\ref{fig:bcn_neighborhoods} (d)).
The resulting network has a few noticeable details, such as the predominance of flows between contiguous neighborhoods, and two potential communities that split the city into a north cluster, and a cluster that comprises the downtown and south areas, including the low populated sectors around the port, and the highway connecting with the airport (see Fig~\ref{fig:neighborhoods_and_grids} (c)).
\begin{figure}[!h]
\begin{adjustwidth}{-2.25in}{0in}
\includegraphics[width=\linewidth]{Fig2.png}
\caption{{\bf Origin-destination flows between neighborhoods in Barcelona.}
a) Map of average importance as origin area (neighborhoods). b) Map of average importance as destination area (neighborhoods) for local people in Barcelona. c) Node-link diagram of the flow network, with links color coded by the flow importance (log-transformed, monthly average in business days). d) The official reference neighborhood identifiers of the city.
Administrative boundaries are sourced from CartoBCN under a CC BY 4.0 license, with permission from Ajuntament de Barcelona, original copyright 2020.}
\label{fig:neighborhoods_and_grids}
\end{adjustwidth}
\end{figure}
\subsection*{Accessibility in 15 minutes}
Accessibility can be interpreted in different ways depending on the context. Commonly, it is an assessment of the availability of services for older, disabled or disadvantaged people~\cite{Schmocker2008}; as well as reachability to certain places, workplaces or leisure activities during certain times of the day via specific types of transportation~\cite{RoyalSociety}. %
It can be measured in a variety of ways~\cite{RoyalSociety}, from multiple points of reference that include the individual perspective, the place perspective, or both by appropriate population segmentation strategies~\cite{Weng19}.
Here we focus on the walking accessibility, which has gained increased attention since Carlos Moreno kick-started the 15-minutes city concept in 2019~\cite{Moreno}, and is fueled~\cite{c40} by the potential advantages for urban health~\cite{Alfonzo}, sustainability~\cite{Capasso}, and overall quality of life~\cite{nuvolati}.
Note that there are alternatives to this proposal. For instance, given that other modes of transportation (such as bicycles or public transport) afford varying levels of accessibility~\cite{Boisjoly,texasassesment}, we could integrate them in the definition. Another option would be to tune the somewhat arbitrary 15-minutes threshold (e.g., Portland pursues accessibility within 20 minutes~\cite{portland}), or weighting amenity types against demographic groups~\cite{Weng19}. We prefer our choice of metric for being clear and straightforward, which will help us understand and communicate how our methods work. Thus, we define the accessibility $A$ to an amenity category $c$ at neighborhood $i$ as follows:
$$
A_{ic} = \frac{\sum_{x \in i} \text{number of } c \text{ amenities reachable by } x \text{ within a 15-minutes walk}}{\text{Population of neighborhood } i},
$$
where $x \in i$ denotes an inhabitant of neighborhood $i$.
The census data contains an approximate address for each inhabitant, which enables the measurement of how many amenities are reachable within a 15-minutes walk.
To quantify the accessibility of each neighborhood, we downloaded an OSM~\cite{OpenStreetMap} database dump of 2018 to identify POIs in the city, as well as the walkable street network of the city. OSM is a Volunteered Geographical Information site where anyone can enter and edit geographical information. It has a similar (and sometimes greater and up-to-date) coverage and accuracy to commercial maps~\cite{haklay2010good}, and it is freely available.
We calculated the accessibility metric for the following amenity categories: \emph{education} (schools, universities, etc.), \emph{entertainment} (indoor places where people pay for access), \emph{finance} (banks and similar), \emph{food} (restaurants, bars, cafes, etc.), \emph{government} facilities, \emph{health} facilities (hospitals, primary care centers, medical offices, etc.), \emph{professional} services, \emph{recreational} areas (parks, outdoors, etc.), \emph{religion} venues, \emph{retail} (malls, supermarkets, shopping venues), and \emph{public transport} (bus stops, metro stations, bike stations, etc.).
We also estimated the diversity of reachable amenities per inhabitant, as diversity has shown to be an important predictor in characteristics and development of places~\cite{eagle2010network,quercia2018diversity}. The diversity is defined as the Shannon Entropy:
$$
H = - \sum_{c \in C} p_{c} \log p_{c},
$$
where $p_c$ is the frequency of an amenity of category $c \in C$, and $C$ is the set of amenity categories, excluding \emph{public transport}.
The spatial distribution of these metrics showcase how the central neighborhoods concentrate a greater amount of amenities, which could indicate a monocentric city. However, the diversity of amenities follows an opposite trend, hinting that the city provides (or aims to provide) access to several types of venues and services in its entire administrative area (see their spatial distributions in Fig~\ref{fig:amenity_distribution}).
\begin{figure}[!h]
\includegraphics[width=\linewidth]{Fig3.png}
\caption{{\bf Distribution of mean number of reachable amenities (and its diversity) per neighborhood.}
Each map shows the spatial distribution of the accessibility to the following type of amenities: a) education. b) entertainment. c) finance. d) food. e) government. f) health. g) professional. h) recreation. i) religion. j) retail. k) public transport. The accessibility value per neighborhood corresponds to the mean of the accessibility values of the neighborhood inhabitants. Panel l) shows the diversity of reachable amenities in each neighborhood.
Administrative boundaries are sourced from CartoBCN under a CC BY 4.0 license, with permission from Ajuntament de Barcelona, original copyright 2020.}
\label{fig:amenity_distribution}
\end{figure}
For each neighborhood in the city, we now have several metrics related to its population, how this population moves in the city from origins to destinations, and the amenities that are reachable from a person's place of usual residence. We now set out to estimate global mobility patterns in the city.
\subsection*{Global mobility patterns from a gravity model}
At city scale, the most common strategy to model flows is to fit a form of gravity model~\cite{zipf1946p}.
Gravity models are a simple phenomenological approach that allows us to quantify the movement between two zones of different sizes taking into account only three parameters: the population of the origin and destination regions, and the distance between them.
The reasoning behind this is that mobility between two regions should be proportional to their population sizes, but inversely proportional to the distance between the regions, to account for the inconvenience of moving between distant locations.
Gravity models are not the only way to model mobility. One of the most recent models is the radiation model~\cite{simini2012universal}, which is parameter free and provides better prediction performance for commuting flows between cities and counties. In contrast to the gravity model, originally proposed as a metaphor, the radiation model has coherent units and can be derived from human behavior. However, despite these positive traits, the gravity model is still preferred by urban modelers and practitioners. Being parameter free, the radiation model makes it difficult to establish the relationship between mobility flows and place-specific features. Moreover, the gravity model is flexible to incorporate additional explanatory variables. This flexibility may explain its comparatively better performance in inferring intra-city flows~\cite{liang2013unraveling,palchykov2014inferring}. As such, understanding why the gravity model works, and how it relates to human behavior and other theory-based models is an active area of research~\cite{hong2019gravity}.
For these reasons, we work with gravity models as it offers improved flexibility and performance in our context.
A gravity model is typically defined as follows:
$$
T_{ij} = \frac{G M_i^\alpha M_j^\beta}{f(d_{ij})^\gamma},
$$
where $T_{ij}$ is the number of visitors at destination area $j$ from origin area $i$, $G$ is a gravity constant, $M_i$ is a measure of mass for area $i$ (usually its population), and $f$ is a deterrence function based on the distance $d_{ij}$ between areas $i$ and $j$. A common and well-behaved way to find the parameters of this model is by adjusting a Generalized Linear Model (GLM), with a Negative Binomial (NB) distribution for count data. This distribution has been used before to model area influx in mobile phone data~\cite{graells2017effect,beiro2018shopping}, as it controls for over-dispersion. The model is specified as follows:
$$
\mathbb{E}[T_{ij}] = \exp[\log(G) + \alpha \log(M_i) + \beta \log(M_j) - \gamma \log(f(d_{ij}))].
$$
This GLM is fitted by maximizing the log-likelihood function.
To account for the socio-demographic and accessibility attributes of each neighborhood, we generalize the model equation into:
$$
\mathbb{E}[T_{ij}] = \exp[\beta_0 + \sum_k \beta_{kij} x_{kij}],
$$
where $x_{kij}$ is the $k$th-feature of the flow from neighborhood $i$ to neighborhood $j$. These features include:
\begin{itemize}
\item Population of origin and destination areas, and the distance that separates them.
\item Touristic attractiveness for origin and destination areas.
\item The difference in accessibility between origin $i$ and destination $j$: $\Delta A_{ijc} = A_{jc} - A_{ic}$, where $c$ is an amenity category, $A_{jc}$ is the accessibility to $c$ venues at neighborhood $j$ (the definition of $A_{ic}$ is analogous). A positive value indicates that the people from neighborhood $i$ moves toward a neighborhood $j$ with better accessibility in the corresponding category.
\item Similarly, we estimate the difference in amenity diversity, $\Delta \text{amenity diversity}_{ij}$.
\item The difference in socio-demographic characterization $\Delta \text{HDI}_{ij}$ between two areas.
\item The difference in the fraction of the population that is women ($\Delta \%\text{women}_{ij}$).
\item The difference in the ratio of population between immigrants and nationals ($\Delta \text{immigrant}_{ij}$).
\item The difference in mean age of the destination and origin areas, $\Delta \text{age}_{ij}$,
\item A binary variable $\text{weekend}$ indicating if the visitor count from $i$ to $j$ corresponds to a weekend day (1) or not (0).
\item A binary variable $\text{to other neighborhood}$ indicating if $i \neq j$ in the corresponding $T_{ij}$.
\end{itemize}
Before fitting the model, we log-transform the distance and population variables (including touristic attractiveness).
The results of fitting the regression model using these features allows us to measure how the different factors are related to the mobility in the city, as it is usually done in related works.
\subsection*{Local mobility patterns using Geographically Weighted Regression (GWR)}
If our premise --that a city is a system of many internal cities-- is correct, we should observe that a global model misses peculiarities for each sub-system of the city. We would expect that these cities were different due to their local cultures, location within the city, and other factors. For instance, distance as a factor may not be equally relevant for everyone, as some people may need to travel longer distances to work. Sometimes the factor may have opposite effects, i.e., a factor that is associated to out-neighborhood mobility in a neighborhood is associated to in-neighborhood mobility in another; for example, accessibility to entertainment amenities might attract people from all over the city to a particular area. All these situations might not be well captured in a global model, because it would try to enforce a constant factor for the entire city.
To account for local variations that affect mobility within the city, the model should include a notion of geographical position in its parameters.
One effective approach to this are Geographically Weighted Regression (GWR) models~\cite{brunsdon1996geographically}. GWR is a kernel regression approach to geographically varying phenomena that performs several GLM regressions~\cite{Fotheringham2002}, one per spatial unit under analysis. Each of these regressions weights the observations in the data (in our case, origin-destination flows) according to the distance to the spatial unit in question, through a kernel function that estimates the weight according to a given bandwidth. The bandwidth can be defined either based on domain knowledge or using model-selection criteria~\cite{comber2020}.
A GWR model is expressed as follows:
$$
\mathbb{E}[T_{ij}] = \exp[\beta_{0ij} + \sum_k \beta_{(k)}(\mathbf{u}_{ij}) x_{kij}],
$$
where $\beta_{(k)}(\mathbf{u}_{ij})$ is a vector of geographically varying parameters at a given position $\mathbf{u}_{ij}$ of $T_{ij}$, and $\beta_{0ij}$ is the intercept corresponding to $T_{ij}$. Since we work with origin-destination flows, there are three alternatives to define $\mathbf{u}_{ij}$: the origin of each OD pair, its destination, or its midpoint~\cite{blainey2013extending}. We chose to consider the origin position given our problem definition: we are interested in patterns at the origin neighborhoods, i.e., what motivates people to move from or to stay at a given area. Patterns in destination neighborhoods may be of use to estimate land use patterns, which is out of the scope of this paper, or the mirror question of what attracts people to a given area. OD mid-points make little sense in this context as two orthogonal flows may share their midpoint without having anything else in common. In contrast, the origin position allows agglomeration of all OD pairs starting at our unit of analysis.
The outcome of applying GWR is a vector of regression factors for every OD pair, which can be aggregated by neighborhood. By definition, all OD pairs with the same origin position will have the same regression results because they use the same spatial kernel. We interpret the resulting vectors as a characterization of each neighborhood's mobility and accessibility.
If we observe little spatial variation in the coefficient vectors this would refute our city of cities assumption. On the other hand, significant spatial variation of the coefficient vectors would indeed support the idea of multiple cities within a city, each one with their own mobility patterns.
\subsection*{Relationship between local mobility and rental prices}
To exemplify how local patterns of mobility can improve our understanding of cities, we explore how these patterns relate to an existing problem in Barcelona, namely, the surge in rental prices. The cost of life in a city is deeply attached to this value, and Barcelona has recently faced over-tourism~\cite{milano2019overtourism} and gentrification processes~\cite{anguelovski2018assessing}.
The problem is well researched, in terms of which local factors of neighborhoods are related to sudden surges in price and tourist accommodation, and its consequences on housing prices~\cite{blanco2018barcelona,garcia2019short,lagonigro2020understanding}.
Typical pricing modeling uses hedonic regression~\cite{halvorsen1981choice}. In addition to housing unit attributes, neighborhood and population features are also included in such models, as well as the spatial dependency of the issue under analysis.
Up to now, the phenomena has been explored from a static perspective where neighborhoods only have contiguity relationships and land use designations. However, how people connect neighborhoods through their mobility is also important. With relationships like these, urban planners and the city council would have more information to consider in their plans to improve quality of life in the city.
Here we propose to measure this relationship using correlation. The Spearman correlation coefficient $\rho$ between two vectors is defined as:
\begin{equation*}
\rho(X, Y) = \frac{\text{cov}(X_R, Y_R)}{\sigma_{X_R} \sigma_{Y_R}},
\end{equation*}
where \textit{cov} is the covariance between two vectors, and $\sigma$ is the standard deviation of a vector. In this case, the functions are applied to the rank-transformed versions of $X$ and $Y$, denoted $X_R$ and $Y_R$. The value of $\rho$ lies in $[-1, 1]$, and it is interpreted as follows: a value of $\rho = 0$ shows no relation between the two variables; a value of $\rho = 1$ shows a perfect positive relation (i.e., an increase in one variable is reflected with an equivalent increase in the other); and a value of $\rho = -1$ shows a perfect negative relation where increases in one variable are related to decrements in the other. Thus, we calculate the $\rho$ coefficients between the rental attributes of each neighborhood and its local mobility variables, and analyze both their magnitude and sign to find which local patterns are related with higher or lower rental princes in the city.
\section*{Results}
Being a vibrant city, Barcelona has several periods throughout the year that influence how people move within it. For instance, the Mobile World Congress (represented by the highest spike in February in Fig~S1 in S1 Appendix) brings many foreigners and national visitors into the city, and changes to the transportation network are common to accommodate the increased demand on transportation. Another example is that an important fraction of the resident population leaves the city during holiday periods, especially in August (see also Fig~S1 in S1 Appendix).
This seasonality suggests that we need to define a typical or average month to analyze mobility. Hence, for every month of the year, we estimated the average origin-destination flow between neighborhoods at business days and weekends, and fitted a global model on them. Our criteria to select a month to analyze was to compare the Akaike Information Criterion corrected (AICc) of each model. The AICc is a measurement of information loss, as such, a model with lower AICc is preferred. The corrected version of AIC accounts for multiple model comparisons and their sample size. In this setting, the months with the lowest AICc are August, December and June, in that order (see Fig~S3 in S1 Appendix). However, considering seasonality, the two months with lowest AICc values (August and December) exhibit a decay of local population due to the summer holidays and the Christmas break, and thus, do not represent a typical month of city life. Hence, we chose the third lowest value: June. This month will be used for the rest of the analysis.
In the global model the two most important factors are: \textit{log(distance)} ($\beta = -2.11$), with a negative effect, i.e., longer distances between neighborhoods imply smaller flows; and \textit{to other neighborhood} ($\beta = 1.56$), with a positive effect, implying that flows from a neighborhood tend to go to other neighborhoods (see all factors in Fig~\ref{fig:global_model_results}). Attributes such as $\Delta$~\textit{education}, $\Delta$~\textit{retail}, and $\Delta$~\textit{finance} are positive. To interpret these values, recall that $\Delta$ variables were defined as the difference between a destination value minus an origin value. Thus, a positive factor can be interpreted as follows: at the city level, people tend to visit neighborhoods with greater access to \textit{education} and \textit{retail} than their own. On the contrary, there is a tendency to travel less to neighborhoods when they have greater accessibility to \textit{professional} and \textit{health} amenities. The former type of venue offers services provided by professionals, and may not be remarkably different between neighborhoods. The health amenities focus on local people by design of the public health system, CatSalut. These insights are expected and coherent with the literature and the urban design of Barcelona.
\begin{figure}[!h]
\includegraphics[width=\linewidth]{Fig4.png}
\caption{{\bf Global gravity model factors.}
Positive values indicate that the corresponding factor is associated with an increase in flows from one neighborhood to another.}
\label{fig:global_model_results}
\end{figure}
Before adjusting the local model, we needed to select a kernel function and a bandwidth for the ratio of influence between neighborhoods and their flows.
The tested kernels were exponential and bi-square distance weight decay, and the bandwidth ranged between 750 meters and 2 kilometers (with specific values tested using bisection search). We kept the parameters from the model with the lowest AICc~\cite{oshan2019mgwr}. The final local model was fitted using 750 meters as bandwidth and an adaptive bisquare kernel.
To check model coherence, we noted that the model residuals present a symmetric distribution around zero in both cases, global and local (see Fig~S4 in S1 Appendix).
We report global and local model statistics and their factor values in Table~\ref{table:model_results}.
The local model AICc is considerably lower than that for the global model, suggesting that our local model can better explain human mobility flows in Barcelona. The GWR results reveal wide geographic variability in the relationship between human mobility flows and local factors. GWR regression estimates range from negative to positive and display relatively high standard deviations. Only five variables are consistently negative ($\log$(\textit{distance}), \textit{in weekend} and $\log$(\textit{destination population})), or positive (\textit{to other neighborhood} and $\log$(\textit{destination touristic attractiveness})). These estimates show the expected sign and are consistent with our global model estimates. However, the extent of the relationship between these factors and human mobility flows varies across Barcelona (Fig~\ref{fig:gwr_model_results}). Distance, for example, seems to exert a prominently large impact on deterring mobility in western and central-western neighborhoods but has a less pronounced influence on coastal northern and southern neighborhoods.
\begin{table}[!ht]
\begin{adjustwidth}{-2.25in}{0in}
\centering
\caption{
{\bf Regression Results in the Global Model (Negative Binomial Regression) and the Local Model (GWR with GLM Negative Binomial Regression).}}
\begin{tabular}{|p{4.1cm}|rrrr|rrrrr|}
\hline
{} & \multicolumn{4}{c|}{\textbf{Global Model} (AICc = 10544.52)} & \multicolumn{5}{c|}{\textbf{Local Model (GWR)} (AICc = 5625.56)}\\
Variable & $\beta$ & SE & t(Est/SE) & $p$-value & Mean & STD & Min & Median & Max \\
\thickhline
Intercept & 3.638 & 0.300 & 12.128 & $<$ 0.001 & -2.059 & 7.122 & -19.758 & -0.857 & 22.533 \\
$\Delta$ education & 1.061 & 0.167 & 6.340 & $<$ 0.001 & 1.188 & 1.078 & -1.887 & 1.331 & 3.325 \\
$\Delta$ entertainment & -0.139 & 0.196 & -0.707 & 0.479 & 0.106 & 1.741 & -3.360 & -0.122 & 4.884 \\
$\Delta$ finance & 0.409 & 0.231 & 1.766 & 0.077 & 0.512 & 1.462 & -2.213 & 0.731 & 3.488 \\
$\Delta$ food & 0.213 & 0.245 & 0.870 & 0.384 & -1.907 & 1.710 & -6.852 & -1.932 & 1.363 \\
$\Delta$ government & -0.147 & 0.084 & -1.743 & 0.081 & -0.027 & 0.388 & -0.740 & 0.006 & 0.950 \\
$\Delta$ health & -0.691 & 0.186 & -3.709 & $<$ 0.001 & -0.589 & 0.906 & -3.071 & -0.434 & 1.161 \\
$\Delta$ professional & -0.696 & 0.313 & -2.219 & 0.026 & -0.137 & 2.083 & -5.224 & -0.083 & 4.444 \\
$\Delta$ public transport & -0.666 & 0.090 & -7.419 & $<$ 0.001 & -0.456 & 0.507 & -1.624 & -0.339 & 0.527 \\
$\Delta$ recreation & -0.177 & 0.149 & -1.190 & 0.234 & -0.467 & 0.883 & -3.002 & -0.566 & 1.334 \\
$\Delta$ religion & 0.249 & 0.100 & 2.498 & 0.012 & 0.131 & 0.736 & -1.082 & -0.012 & 1.679 \\
$\Delta$ retail & 0.672 & 0.126 & 5.347 & $<$ 0.001 & 1.296 & 0.913 & -0.920 & 1.270 & 3.400 \\
$\Delta$ amenity diversity & 0.199 & 0.094 & 2.128 & 0.033 & -0.146 & 0.401 & -1.316 & -0.062 & 1.029 \\
$\Delta$ HDI & -0.082 & 0.030 & -2.728 & 0.006 & -0.041 & 0.100 & -0.297 & -0.036 & 0.166 \\
$\Delta$ age & 0.031 & 0.012 & 2.615 & 0.009 & 0.003 & 0.068 & -0.143 & 0.022 & 0.128 \\
$\Delta$ \%women & 0.063 & 0.028 & 2.217 & 0.027 & 0.042 & 0.152 & -0.310 & 0.054 & 0.328 \\
$\Delta$ \%immigrants & 0.345 & 0.056 & 6.194 & $<$ 0.001 & 0.192 & 0.259 & -0.386 & 0.193 & 0.729 \\
log(distance) & -2.106 & 0.041 & -51.802 & $<$ 0.001 & -2.328 & 0.452 & -3.173 & -2.312 & -1.335 \\
log(DST population) & -0.392 & 0.027 & -14.390 & $<$ 0.001 & -0.300 & 0.145 & -0.704 & -0.293 & -0.005 \\
log(DST touristic attractiveness) & 0.132 & 0.021 & 6.449 & $<$ 0.001 & 0.279 & 0.110 & 0.025 & 0.320 & 0.544 \\
log(SRC population) & 0.259 & 0.031 & 8.229 & $<$ 0.001 & 0.663 & 0.738 & -2.210 & 0.617 & 2.054 \\
log(SRC touristic attractiveness) & 0.166 & 0.020 & 8.230 & $<$ 0.001 & 0.178 & 0.387 & -0.997 & 0.111 & 1.199 \\
to other neighborhood & 1.560 & 0.105 & 14.865 & $<$ 0.001 & 1.781 & 0.364 & 0.987 & 1.794 & 3.269 \\
in weekend & -0.923 & 0.032 & -28.661 & $<$ 0.001 & -1.034 & 0.169 & -1.394 & -1.066 & -0.573 \\
\hline
{Local Model Info.} & \multicolumn{9}{c|}{Adaptive bisquare kernel, Bandwidth=769m, Adj. critical t value (95\%)=2.95, DF=4380.72}\\
\hline
\end{tabular}
\begin{flushleft} Summary table for global and local regression models. Acronyms and abbreviations: AICc (Akaike Information Criterion corrected), SRC (source neighborhood), DST (destination neighborhood), HDI (human development index). In the local model, \textit{Mean} refers to the mean $\beta$ value of all coefficients for all regressions performed within the model. The other statistical properties are defined analogously.
\end{flushleft}
\label{table:model_results}
\end{adjustwidth}
\end{table}
A much wider extent of geographic variability exists in the relationship between human mobility and other factors, and for many neighborhoods, this relationship contrasts with that captured by the global model. For instance, the global model indicates that differences in educational and retail amenities between neighborhoods greatly contribute to shaping mobility. Yet, our GWR estimates indicate that while disparities in accessibility to educational facilities are strongly associated with mobility out of certain central and northern neighborhoods, and with reduced mobility in others, this characteristic is more prominent in southern and north-western neighborhoods. Similarly, GWR estimates indicate that while unequal access to retail amenities tend to promote mobility from central and eastern coastal neighborhoods, they are negatively associated with mobility in western and northern areas of Barcelona. In addition, our results also reveal how geographic differences in the composition of the resident population is associated with mobility. For example, our GWR estimates indicate that large differences in age profile tend to trigger mobility from inland eastern neighborhoods, suggesting that areas with an older population profile tend to experience higher levels of mobility.
\begin{figure}[!h]
\begin{adjustwidth}{-2.25in}{0in}
\includegraphics[width=0.95\linewidth]{Fig5.png}
\caption{{\bf Spatial distribution of the local regression factors from the GWR model.}
Each choropleth map uses a color scale with red tones to depict positive values, and blue tones to depict negative values. Factors: a) Distance. b) Neighborhood Population. c) Destination Population. d) Neighborhood Touristic Attractiveness. e) Destination Touristic Attractiveness. f) $\Delta$~\emph{education}. g) $\Delta$~\emph{entertainment}. h) $\Delta$~\emph{finance}. i) $\Delta$~\emph{food}. j) $\Delta$~\emph{government}. k) $\Delta$~\emph{health}. l) $\Delta$~\emph{professional}. m) $\Delta$~\emph{recreation}. n) $\Delta$~\emph{religion}. o) $\Delta$~\emph{retail}. p) $\Delta$~\emph{public transport}. q) $\Delta$~\emph{neighborhood amenity diversity}. r) $\Delta$~\emph{neighborhood Human Development Index} (HDI). s) $\Delta$~\emph{neighborhood mean age}. t) $\Delta$~\emph{\% women population}. u) weekend flag (binary). v) visiting other neighborhood (binary). w) $\Delta$~\emph{ratio immigrants to nationals}.
Administrative boundaries are sourced from CartoBCN under a CC BY 4.0 license, with permission from Ajuntament de Barcelona, original copyright 2020.}
\label{fig:gwr_model_results}
\end{adjustwidth}
\end{figure}
Polycentric city strategies are associated with highly walkable spaces within local neighborhoods and reduced long-distance mobility to other neighborhoods. Key to inform such strategies is identification of the set of urban amenities influencing mobility from neighborhoods contributing to low local levels of resident population retention. To identify the set of urban amenities underpinning local levels of mobility in Barcelona, we created a heatmap that shows the accessibility factors from the GWR and highlights the significant ones, such that these can provide local actionable insights to experts and policy makers.
Fig~\ref{fig:heatmap_GWR_coefficients} reveals that a small but distinctive set of urban amenities tend to shape the local patterns of human mobility in most neighborhoods across Barcelona, such as retail, education, and food amenities. The majority of neighborhoods have differences in accessibility to urban amenities, and these tend to significantly influence mobility from different neighborhoods. Differences in accessibility to retail outlets stands out as the most consistent factor linked to increased levels of mobility in many neighborhoods. Interestingly, the reduced levels of mobility effect from differences in food amenity accessibility is probably explained because of the large concentration of food amenities in the center of the city, meaning that the citizens from these neighborhoods do not tend to move to the center of the city. Differences in accessibility to entertainment from neighborhoods from the north-western inner part of the city also increase levels of mobility, similarly to the differences in food amenities, but having a positive effect on the levels of mobility as the data suggests that these people tend to move to the center of the city. Lastly, differences in accessibility to religion amenities from north-western neighborhoods also stand out increasing levels of mobility.
\begin{figure}[!h]
\begin{adjustwidth}{-2.25in}{0in}
\includegraphics[width=\linewidth]{Fig6.png}
\caption{{\bf Heatmap of GWR coefficients by neighborhood and variable.} This heatmap shows the results from the GWR model, allowing for the identification of clusters of neighborhoods and variables. The colormap is diverging, negative values are blue, positive values are red, and values close to 0 are white. Variables were sorted by their standard deviation and neighborhoods were ordered and clustered using ward linkage distance. The cells with an overlaid black square are statistically significant at the 0.05 level.}
\label{fig:heatmap_GWR_coefficients}
\end{adjustwidth}
\end{figure}
Finally, we explored how the local patterns of mobility allow us to find novel perspectives regarding urban phenomena. We estimated the correlation between relative neighborhood rental prices and the neighborhood characterization, both from its direct features and those derived from the local model of mobility.
Many of the neighborhood attribute correlations with rental prices are significant (see Fig~\ref{fig:mobility_rent_correlations} (a)). The exceptions are access to \emph{religion}, to \emph{public transport}, the amenity diversity, and the demographic variables related to population size, age, gender and migration. The significant correlations are positive (ranging from access to health venues with $\rho = 0.44$ to access to food venues with $\rho = 0.70$), which is expected, as better accessibility in a neighborhood is a positive trait.
\begin{figure}[!h]
\includegraphics[width=\linewidth]{Fig7.png}
\caption{{\bf Correlations between neighborhoods' attributes and mobility factors with relative rental prices.} Grey bars represent non-statistically significant correlations, magenta bars represent statistically significant correlations; significance was determined with Bonferroni-corrected $p$-values (significance level 0.05).
a) Spearman correlation between neighborhood features and rental price per square meter. b) Spearman correlation between neighborhood local mobility factors and rental price per square meter.}
\label{fig:mobility_rent_correlations}
\end{figure}
There are ten variables from the local features that have significant correlations with rental prices, five positive and five negative (see Fig~\ref{fig:mobility_rent_correlations} (b)).
The positive correlations are:
$\Delta$~\emph{\% women} ($\rho = 0.51$),
$\Delta$~\emph{food} ($\rho = 0.48$),
$\log$(\textit{origin population}) ($\rho = 0.47$),
$\log$(\textit{destination touristic attractiveness}) ($\rho = 0.46$),
and $\Delta$~\emph{retail} ($\rho = 0.44$).
Of these variables, three are related to tourism (access to food and retail venues, and the attractiveness of the destination area), which showcases that neighborhoods where its inhabitants go to shop and dine are more expensive. These four variables can be linked to gentrification processes, suggesting that places that are not necessarily touristic but allow to reach touristic places are more expensive for housing, in coherence with over-tourism in the city~\cite{garcia2019short}.
The positive association with population size needs more exploration, as it needs to take into account the different types of residential density.
Regarding the strong positive correlation with the $\Delta$~\emph{\% women} factor, note that according to the spatial distribution of this factor (see Fig~\ref{fig:gwr_model_results}~(t)), there are sign differences between the city core (positive) and its periphery (negative), which is different from the neighborhood attribute of fraction of population that is women (see Fig~\ref{fig:bcn_neighborhoods}~(e)). We interpret this difference in relationship to access to safe places by women, and other potential gender gaps in urban mobility~\cite{gauvin2019gender}.
The negative correlations include the following factors: $\log$(\textit{destination population}) ($\rho = -0.57$), which means that neighborhoods where its inhabitants tend to travel to densely populated areas have lower rental prices; $\Delta$~\emph{finance} ($\rho = -0.46$) and $\Delta$~\emph{health} ($\rho = -0.41$), showing that areas where inhabitants move to other neighborhoods for these amenities are less expensive with respect to housing. Lastly, the $\Delta$~\emph{entertainment} correlation ($\rho = -0.37$) shows that neighborhoods where its residents move to other neighborhoods for amenities linked to nightlife and tourism in Barcelona have lower rental prices.
In summary, these results provide evidence on how Barcelona is not a monocentric city, but rather a polycentric city with diverse functional areas. Distinct groups of factors influence different regions and their local mobility, and it is the combination of amenities with the different ``small-cities'' within the city that define the many centers of Barcelona. Note that the actual shape of these multiple cities is not defined; such a task is out of the scope of this paper, however, a potential way to achieve this segmentation is through these local representations. As we observed, the local variability in mobility factors may inform urban planners and policy makers regarding the functioning of the city and its on-going phenomena. Besides, even though some variables are not necessarily significant with respect to origin-destination flow inference, we find these are relevant as they present variability and are informative with respect to other issues.
\section*{Discussion and Conclusions}
We sought to improve our understanding of the relationship between human mobility and urban amenities in the context of polycentric and 15min city planning strategies. Particularly we aimed to measure the extent of geographic variability in this relationship and how it was associated with the socio-demographic profile of local population. We integrated new (mobile phone and OSM data) and traditional (census data) forms of data to model and quantify the relationship between human mobility flows, and urban amenities and local population composition.
Our results provided evidence of wide geographic variability in the relationship between human mobility and accessibility to local urban amenities. We showed that a distinctive set of local urban amenities contribute to shape local patterns of human mobility at the neighborhood level in Barcelona. Our findings revealed that the local patterns of human mobility in most neighborhoods in Barcelona are explained by differences in accessibility in one or two types of urban amenities. Unequal access to food outlets emerged as the most consistent factor linked to reduced levels of mobility in most neighborhoods.
We also presented evidence of significant differences between global and local patterns. We showed that accessibility indicators of urban amenities displaying statistically significant relationships with human mobility at the global level do not necessarily explain human mobility patterns at the local level. For example, our results indicate that accessibility to educational and retail amenities displays a positive association with human mobility at the global level; yet, wide geographic variability exists at the local level. Accessibility to education amenities exerts a strong impact on promoting mobility in a few central and northern neighborhoods of Barcelona, but its effect is negligible in other neighborhoods. Similarly, accessibility to retail amenities is prominent shaping mobility in coastal central neighborhoods but relatively unimportant in other areas across Barcelona.
Additionally, we also presented evidence on the urban amenities linked to reduced human mobility levels. Our results, for example, revealed that differences in accessibility to professional, health and public transport amenities are negatively associated with human mobility in specific neighborhoods. These results suggest that greater accessibility to these amenities in these neighborhoods tend to deter mobility to neighborhoods with significantly lower accessibility levels for these services, and \textit{vice versa}.
Together, our findings contribute to inform spatially targeted urban policy interventions: first, by providing evidence on the extent of geographic variability in the relationship between urban amenities and human mobility; and, second, by identifying the set of neighborhoods and urban amenities associated with high and low mobility levels, or resident population retention rates. Ultimately our findings can effectively inform urban policy interventions seeking to reduce local levels of mobility and move closer towards the realization of a 15-minutes city or polycentric city by identifying the specific set of urban amenities requiring intervention in each neighborhood.
To exemplify how local mobility models help to understand urban phenomena, we compared how the found patterns by our model differ when analyzing a current problem that affects Barcelona and many other cities in Europe: the rise of rental prices, which are linked to over-tourism and gentrification processes. One of the key novel insights we obtained is that {\em better access} to tourist attractions can push rental prices upwards as easily as {\em having} tourist attraction. A plausible explanation is that the presence of tourists drives the price down, but closeness to the attraction is still valuable.
Our work is scoped by the market share of the mobile phone operator, which is likely to be biased toward specific socio-economic and demographic groups. Given that we used aggregated, anonymized data, we cannot measure and therefore correct any biases emerging from the data. Since the usage of mobile phone data in data-for-good projects is an active area of research~\cite{oliver2020mobile}, we expect that new methods and other datasets will enable mitigating these biases.
We devise four main lines of future work:
to integrate the several types of visitors in a common framework, to include other variables into the model, the definition of time-aware metrics, and multi-city analysis.
First, we focused on the origin and destination patterns of Barcelona residents. However, the city's floating population has a non-negligible share of commuters and tourists that should be taken into account in future iterations of the work. This line presents challenges as those visitors do not have an origin to map directly into the model.
Second, even though we have included population and demographic variables into our model, there may be a need for other variables.
We focused on the importance of quantity and diversity of local amenities, but there may be other qualities that are relevant; if there are multiple 15-minutes cities, it is likely that some of them present better amenities than others, thus, attracting visitors. The availability of user-generated content, such as reviews and check-ins, may provide additional insights to deepen the analysis on the importance of urban amenities shaping these patterns of mobility.
Additionally, urban disciplines such as transportation may need to incorporate factors such as employment density, fine-grained population data, distribution of mode of transportation usage, among others. Some of the factors under study address these variables indirectly, such as the population counts and the socio-economic characterization, however, for urban planning practice they may need to be explicit.
Third, incorporating time into the model would enable the measurement of effects of public policy changes and urban interventions. There are emerging methods that extend GWR with temporal awareness~\cite{gmd-2019-292} that could be used in future analyses.
Finally, to advance on the path to creating inclusive and sustainable cities, quantitative methods combined with fine-grained datasets are required to understand and compare different cities. This will allow researchers and urban planners to distinguish between specific results from a case study and systematic subdivisions in cities that need to be made visible and measurable.
\section*{Acknowledgments}
This project has received funding from the European Union's Hori\-zon 2020 research and innovation programme under grant agree\-ment No. 857191 (IoTwins project). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Eduardo Graells-Garrido was partially funded by ANID Fondecyt de Iniciaci\'on 11180913.
We acknowledge the following software libraries used in the analysis:
numpy~\cite{harris2020array}, scipy~\cite{virtanen2020scipy},
matplotlib~\cite{hunter2007matplotlib}, seaborn,
PySAL~\cite{rey2010pysal}, mgwr~\cite{oshan2019mgwr}, pandas~\cite{mckinney2011pandas}, and geopandas.
Map data used in the accessibility analysis includes copyrighted data from OpenStreetMap contributors.
We thank the anonymous reviewers for their suggestions to improve this paper.
Finally, we thank the people from \emph{Ajuntament de Barcelona} and Vodafone for providing access to the data and for useful discussions.
\nolinenumbers
|
\section{Introduction}\label{intro}
Generalization bounds for learning provide a theoretical guarantee on the performance of a learning algorithm on unseen data. The goal of such bounds is to provide
control of the error on unseen data with pre-specified confidence. In certain situations, such bounds may also help in designing new learning algorithms. With the great success of deep neural networks (DNNs)
on a variety of machine learning tasks, it is natural to try to construct theoretical bounds for such models. Recent years have seen several notable efforts to construct these bounds.
One classical approach, that of Probably-Approximately-Correct (PAC) learning, aims to develop generalization bounds that hold (with high probability) {\it uniformly} over all hypotheses $h$ (parametrized with parameters $\theta$) in some hypothesis class $\mathcal{H}$ and over all data distributions $D$ in some class of distributions $\mathcal{D}$. The well-known VC-dimension bounds \cite{Vapnik} fall in this category. Since these are uniform bounds, they are \emph{worst case} and very general. However, the flip side is that they tend to be very loose and not useful in practical situations. For example, the VC dimension of neural networks is typically bounded by the number of parameters of the network. Since modern DNNs have millions of parameters, these uniform VC dimension bounds are typically not useful at all.
Better prospects for developing useful and non-trivial bounds arise when one develops \emph{data-dependent} and \emph{non-uniform} bounds. For example, there has been lot of recent work on bounds based on \emph{Rademacher complexity} \cite{Neyshabur2015, Bartlett2017}. These bounds are dependent on the particular data distribution $D$ ($\in \mathcal{D}$) but still depend uniformly on the entire hypothesis class $\mathcal{H}$. Recent work on computing these bounds for modern DNNs still gives rise to bounds that may not be useful for at least some classes of modern DNNs \cite{Dziugaite2017}.
Another data-dependent approach to developing useful generalization bounds is that of the \emph{PAC-Bayesian} framework \cite{McAllester1999, Langford2002, Seeger2003, Catoni2007}. In this case, one starts with some 'prior' distribution $\pi_0$ (that is independent of the data) over the hypothesis space $\mathcal{H}$, and then constructs a 'posterior' distribution $\nu$ over $\mathcal{H}$ after training the model on some training dataset. Note that the 'prior' and 'posterior' do not refer to the corresponding terms in Bayesian probability in the strict sense. Here, the posterior distribution $\nu$ can be thought of as being determined by the training data and the training algorithm in general. The PAC-Bayes framework studies an averaging of the generalization error according to the posterior $\nu$ (the so-called Gibbs classifier). The bounds thus obtained are valid for all posterior distributions $\nu$ over the hypothesis space $\mathcal{H}$, but the bound itself depends on the posterior $\nu$. Thus, these bounds are \emph{non-uniform} in the hypothesis space as well. In other words, the precise numerical bounds are valid for parameters in the neighborhood of those learned by the \underline{actual} learning algorithm. Thus, these have the potential to provide tight and useful bounds for complex models like DNNs.
A point to note is that the bulk of the work in statistical learning theory within each of the above approaches has dealt with providing bounds for the \emph{inductive} learning setting. In this case, the goal is to learn a labeling rule from a finite set of labeled training examples, and this rule is then used to label new (test) examples. In many realistic situations, however, one has a \emph{full sample}, which consists of a training set of labeled examples, together with an unlabeled set of points which needs to be labeled.This is the \emph{transductive} setting, where one is interested only in labeling the unlabeled set as accurately as possible. In this case, it is plausible that if one leverages the information of the test data points appropriately, one may be able to derive tighter risk bounds. Transductive learning bounds are valid for the particular dataset at hand. If the dataset is augmented with new data points, the bounds have to be computed again in general. In this work, we show that it is possible to get tight bounds on the full sample risk by using the method of Hilbert coresets, a coreset defined in terms of a Hilbert norm on the space of loss functions (a more precise description will follow later). This work builds on the work of Campbell and Broderick \cite{Campbell2017, Campbell2018} and applies it to the transductive learning framework. Our principal contributions in this paper are as follows:
\begin{itemize}
\item We compute non-vacuous and tight upper bounds on the \emph{expected full sample risk} for multi-class classification tasks for a variety of state-of-the-art hypothesis classes (DNN architectures) and datasets, and compare these results with that obtained from explicit training algorithms. Our bounds are valid for any data distribution that lies in a class of distributions $\mathcal{D}$. In particular, we study the class of distributions $\mathcal{D}$: \emph{Convex hull of a probability simplex} $\mathcal{D}_{CH}^{(K)}$. The calculation of the bounds for the expected full sample risk requires knowledge of the full dataset along with the labels.
\vspace{0.3cm}
\item The developed bounds have nice properties. First, these bounds are \emph{non-uniform} in the sense that although the bounds are valid for all distributions over $\mathcal{H}$, the precise bound depends on particular distribution over $\mathcal{H}$, i.e. the precise neighborhood of the parameters $\theta_{\star}$ learned by the training algorithm.
\vspace{0.3cm}
\item Second, although the bounds we develop technically apply to a \emph{stochastic} DNN, appropriate choice of data-independent priors and data-dependent posteriors in many practical examples gives rise to Gibbs classifiers that are highly clustered around the classifier learned by the original deterministic DNN (with parameters $\theta_{\star}$ learned by the training algorithm, see results in Section \ref{results}). Thus, although there is no guarantee, the bounds can become effectively \emph{deterministic} in many practical cases.
\vspace{0.3cm}
\item Finally, the numerical bounds \emph{decrease} with an increase in the
number of training samples. This behavior has been tested for a variety of DNNs and holds up across changes in depth and width of the DNN as well as
different mini-batch size used for training. This is in contrast with certain existing bounds which may even \emph{increase} with an increase in the number of
training samples \cite{Bartlett2017, Golowich2017, Neyshabur2018}; see \cite{Nagarajan2019} for a discussion.
\end{itemize}
Transductive bounds have received much less attention than inductive bounds in general. Generalization bounds for the transductive setting have been developed within the PAC-Bayesian framework in \cite{derbeko, begin14}. Our bounds are not PAC-Bayes bounds in the strict sense. Our methods do provide a bound on the expected full sample risk as in the PAC-Bayes case, however this bound is computed using Hilbert coresets which is a novel approach. Also, the bounds developed in \cite{derbeko, begin14} are applicable for a uniform distribution on the full sample and have only been applied to simple hypothesis classes. On the other hand, the bounds developed in this paper are applicable to more general data distributions, such as those that lie in the convex hull of the probability simplex of $K$ probability vectors. Finally, to our knowledge our paper is the first one to develop meaningful and tight upper bounds within a transductive framework for state-of-the-art DNNs.
\section{Problem Formulation}\label{formulation}
We start with considering an arbitrary input space $\mathcal{X}$ and binary output space $\mathcal{Y} = \{-1,+1\}$. For simplicity, we focus on the binary classification case; however our results are easily generalizable to the multi-class classification case as well. Consistent with the original transductive setting of Vapnik \cite{Vapnik1998}, we consider a \emph{full sample} $Z$ of $N$ examples (input-output pairs) that is assumed to be drawn i.i.d from an unknown distribution $D$ on $\mathcal{X}\times \mathcal{Y}$:
\begin{equation} Z:= \{(x_n,y_n)\}_{n=1}^N \end{equation} For simplicity we will refer to the probability distribution as $D \equiv \{p_n\} = \{P(x_n, y_n)\}; n=1,2,..,N$.
$D$ is assumed to lie in a class of data distributions $\mathcal{D}$. As mentioned previously, we consider the class of distributions where $\mathcal{D}$ is the \emph{convex hull of a probability simplex} $\mathcal{D}_{CH}^{(K)}(\{\hat{p}^{(i)}\})_{i \in [K]}$ defined by the $K$ probability vectors $(\{\hat{p}^{(i)}\}); i \in [K]$. A training dataset $Z^{(S)} =\{(x_i, y_i)_{i=1}^{N_S}\} \in \mathcal{X}\times\mathcal{Y}$, that contains $N_S$ input-output pairs is obtained from the full sample by sampling without replacement. Given $Z^{(S)}$ and the unlabeled set $U_S$ consisting of $N-N_S$ inputs, the goal in transductive learning is to learn a classifier that correctly classifies the unlabeled inputs of the set $U_S$, i.e. minimize the risk of the classifier on $U_S$.
We consider in this work general loss functions (including unbounded loss functions) $l: \mathcal{Y} \times \mathcal{Y} \rightarrow \mathbf{R}$, and a (discrete or continuous) set $\mathcal{H}$ of hypotheses $h: \Theta \times \mathcal{X} \rightarrow \mathcal{Y}$ where $\theta$ lies in some parameter space $\Theta$. Although the bounds we develop apply to general loss functions, most of the relevant applications of our bounds for classification apply to bounded, and in particular \emph{zero-one}, loss functions. We provide a bound on the \emph{full sample risk} of a hypothesis $h \in \mathcal{H}$, which is given by:
\begin{equation}\label{sample-risk}
R(h; \theta) = \mathbf{E}_{Z \sim D}\;l(h(\theta, x), y)
\end{equation}
Inspired by the PAC-Bayesian approach, we study an averaging of the full sample risk according to a posterior distribution $\nu$ over $\mathcal{H}$ that is dependent on the data and the training algorithm in general.
Since $\mathcal{H}$ is parametrized by $\theta \in \Theta$, by a slight abuse of notation we denote the distribution on the parameters $\theta$ by $\nu(\theta)$ as well. So, we aim to provide bounds on the \emph{expected full sample risk} $R(\nu)$:
\begin{equation} \label{expected-sample-risk}
R(\nu) = \mathbf{E}_{\theta \sim \nu}\,R(h; \theta) = \mathbf{E}_{\theta \sim \nu}\; \mathbf{E}_{Z \sim D} \;l(h(\theta, x), y) = \mathbf{E}_{\theta \sim \nu} \left(\sum_{n=1}^N \, p_n\, l_n(\theta) \right) : = \mathbf{E}_{\theta \sim \nu} \, L(\theta)
\end{equation}
where the individual loss for a data point is defined as $l_n(\theta):= l(h(\theta, x_n),y_n)$ and $L(\theta):= \sum_{n=1}^N\,p_n\,l_n(\theta)$ is the total loss function of the full sample. Both $L(\theta)$ and $l_n(\theta), n\in [N]$ are functions of the parameters $\theta$. We aim to develop upper bounds on $R(\nu)$ using the Hilbert coresets approach, which we summarize next.
\subsection{Summary of Hilbert Coresets Approach}\label{coresets-summary}
As described above, we consider a dataset $Z = \{(x_n,y_n)\}_{n=1}^N \sim D$, parameters $\theta \in \Theta$, and a \emph{cost function} $L: \Theta \rightarrow \mathbf{R}$ that is additively decomposable into a set of functions $L_n(\theta)$, i.e. $L(\theta):= \sum_{n=1}^N\, L_n(\theta)$. By comparing to (\ref{expected-sample-risk}) and the discussion following that, we can make the identification:
\begin{equation}
L_n(\theta) := p_n\,l_n (\theta)
\end{equation}
Now, an \emph{$\epsilon$-coreset} is given by a weighted dataset with nonnegative weights $w_n, n= 1,2...,N$, only a small number $m$ of which are non-zero, such that the weighted cost function $L(w, \theta) := \sum_{n=1}^N\,w_n\,L_n(\theta) $ satisfies: \begin{equation} \label{coresets-init} |L(w, \theta) - L(\theta)| \leq \epsilon\, L(\theta)\hspace{0.5cm} \forall \theta \in \Theta \end{equation} for $\epsilon > 0$.
If we consider a bounded uniform norm for functions of the form $f: \Theta \rightarrow \mathbf{R}$, which is weighted by the total cost $L(\theta)$ as in
\begin{equation} \label{uni-norm} || f || := \sup_{\theta \in \Theta} \; \left| \frac{f(\theta)}{L(\theta)} \right|, \end{equation} we can rewrite the problem of finding the best coreset of size $m$ as:
\begin{equation} \label{coresets-2}
\min_{w}\;|| L(w)-L ||^2 \hspace{0.3cm} \mathrm{s.t.} \sum_{n=1}^N \, \mathbf{I}\,[w_n > 0] \leq m; \;\; w \geq 0
\end{equation}
The norm $||f||$ of any function $f: \Theta \rightarrow \mathbf{R}$ gets rid of its dependence on $\theta$, thus there is no dependence on $\theta$ in the norm in (\ref{coresets-2}). The key insight of Campbell \emph{et al} \cite{Campbell2017, Campbell2018} was to approach the problem from a geometrical point of view. In particular, the idea is to construct coresets for cost functions in a \emph{Hilbert space}, i.e. in which the cost functions $L_n(\theta)$ (defined in the beginning of this subsection) are endowed with a norm corresponding to an inner product. The inner product provides a notion of directionality to the various $L_n$ and allows us to use more general norms than the uniform norm in (\ref{uni-norm}) (which we will see in section \ref{coresets-fw}). This in turn provides us with two distinct benefits: \begin{itemize} \item{It allows one to choose coreset points intelligently based on the residual cost approximation error}. \item{Theoretical guarantees on approximation quality incorporate the error in the approximation via the (mis)alignment of $L_n$ relative to the total cost $L$.} \end{itemize}
\subsection{Hilbert Coresets via Frank-Wolfe}\label{coresets-fw}
Campbell {\it et al} \cite{Campbell2017, Campbell2018} developed various algorithms for computing Hilbert coresets; here we focus on the Frank-Wolfe algorithm since it produces exponential convergence for moderate values of the coreset size $m$. Furthermore, it can be elegantly tied to risk bounds and computed explicitly in many interesting cases. In this case, the problem in (\ref{coresets-2}) is reformulated by replacing the cardinality constraint on the weights with a polytope constraint on $w$:
\begin{eqnarray} \label{FW-init}
& & \min_{w}\;|| L(w)-L ||^2 \\
& s.t.& \sum_{n=1}^N w_n ||L_n|| = \sum_{n=1}^N ||L_n||; \;\; w \geq 0 \nonumber
\end{eqnarray} One can rewrite the objective function $|| L(w)-L ||^2$ as $J(w) := (w-1)^T\,K\,(w-1)$ with the matrix $\hat{K}$ having elements defined in terms of the inner product, $\hat{K}_{n,m} := \inner {L_n} {L_m}$ (which will be defined shortly). This optimization problem is then solved by applying the Frank-Wolfe algorithm for $m$ iterations to find the coresets $w$\cite{Campbell2017}. The size of the coreset (number of non-zero coreset weights) is smaller than $m$ (typically much smaller), see details in Algorithm 2 of \cite{Campbell2017}. Finally, it can be shown that the Frank-Wolfe algorithm satisfies:
\begin{equation} \label{fw-bound0}
|| L(w) - L || \leq \frac{\sigma\, \eta\, \bar{\eta}\, \beta }{\sqrt{\eta^2\,(m-1)+\bar{\eta}^2\, \beta^{-2(m-2)}}},
\end{equation} where $\sigma := \sum_{n=1}^N\,||L_n||$. The remaining quantities $\{\eta,\,\bar{\eta}, \beta\}$ can be computed from all the data ($\{x_n,y_n\}; n = 1,2..., N$) and the $\nu$-weighted $L2$-norm, see \cite{Campbell2017}. Variants of these quantities (which will be relevant for us later) will be defined precisely in section \ref{coresets-bounds-fw}. In particular $0 < \beta < 1$, and for moderately large $m$, the second term in the denominator dominates giving rise to an exponential dependence on $m$ ($ \sim \beta^m$) .
The above bound on the coreset approximation error for the cost $L$ is valid for any choice of norm. However, for our purposes we will be concerned with the $\nu$-weighted $L2$-norm defined by:
\begin{equation} \label{l2-norm} ||L(w)-L||^2_{\nu, 2} := \mathbf{E}_{\theta \sim \nu}\,[L(w,\theta)-L(\theta)]^2, \end{equation} with an induced inner product:
\begin{equation} \label{innerprod} \langle L_n, L_m \rangle := \mathbf{E}_{\theta \sim \nu} [L_n(\theta)\,L_m(\theta)] = K_{n,m}, \end{equation}
where $K_{n,m}$ was defined below (\ref{FW-init}). In the following, we will see that an upper bound on a variant of $||L-L(w)||$ can be connected to an upper bound on the full sample risk.
\section{Risk Bounds via Frank-Wolfe Hilbert Coresets}\label{coresets-bounds-fw}
Leveraging the Frank-Wolfe Hilbert coreset construction scheme, we provide an upper bound on the expected full sample risk which is valid for all posteriors $\nu \in \mathcal{H}$. However, the precise bound is dependent on the posterior $\nu$. Furthermore, the bound is independent of the data-distribution over the set of data-distributions $\mathcal{D}$. The result can be stated as follows:
\begin{theorem} \label{theorem1}
Given a full sample $Z$ of size $N$, an unknown distribution on the data belonging to $\mathcal{D}^{(K)}_{CH}(\{\hat{p}^{(i)}\})_{i \in [K]}$, a hypothesis space $\mathcal{H}: \Theta \times \mathcal{X} \rightarrow \mathcal{Y}$, a loss function $l: \mathcal{Y} \times \mathcal{Y} \rightarrow \mathbf{R}$, for any prior distribution $\pi_0$ on $\mathcal{H}$, then the output $\{\tilde{w}, p\}$ obtained after $m$ iterations of applying \emph{Algorithm 1} satisfies:
\begin{eqnarray}
\forall \nu \,\mathrm{on} \,\mathcal{H},\, \forall D \in \mathcal{D}_{CH^{(K)}}, \hspace{0.5cm} |R(\nu)| &\leq& ||L(\tilde{w})||_{\nu, 2} + ||L(\tilde{w}) - L||_{\nu, 2} \label{theorem1-ineqA} \\
&\leq& ||L(\tilde{w})||_{\nu, 2} + \frac{\hat{\sigma}\,\hat{\eta} \,\hat{\bar{\eta}} \,\hat{\beta}} {\sqrt{\hat{\bar{\eta}}^2\, \hat{\beta}^{-2(m-1)} + \hat{\eta}^2 \,(m-1)}}, \label{theorem1-ineqB}
\end{eqnarray}
\end{theorem}
where $L(\theta)$ is defined as in (\ref{expected-sample-risk}), $L (\tilde{w}, \theta ) := \sum_{n=1}^N\, \tilde{w}_n\,l_n(\theta); \tilde{w}_n := p_n\,w_n$, and the quantities $\{ \hat{\sigma}, \hat{\eta},\,\hat{\bar{\eta}}, \hat{\beta}\}$ are defined as follows:
\begin{eqnarray} \label{consts}
\hat{\sigma} &:=& \max_{i \in [K]}\,\sigma^{(i)};\;\hat{\eta} := \max_{i \in [K]}\, \eta^{(i)};\;\hat{\beta} := \max_{i \in [K]}\,\beta^{(i)}\nonumber \\
\sigma^{(i)} &:=& \sum_{n=1}^N\,\hat{p}^{(i)}_n\,|| l_n ||; \;
{\eta^{(i)}}^2 = 1 - \frac{||L^{(i)}||^2} {(\sigma^{(i)})^2}; \; {\beta^{(i)}}^2 = 1 - \frac{(r^{(i)})^2}{(\sigma^{(i)})^2 \hat{\bar{\eta}}^2}; \;\; i \in [K]\nonumber \\
\hat{\bar{\eta}}^2 &:=& \max_{n,m} \norm{ \frac{l_n}{|| l_n||} - \frac {l_m} {|| l_m ||} };\;L^{(i)}(\theta):= \sum_{n=1}^N \hat{p}^{(i)}_n l_n(\theta),
\end{eqnarray}
with $r^{(i)}$ corresponding to the shortest distance from the probability vector $\hat{p}^{(i)}$ ($i^{th}$ vertex of the convex hull $\mathcal{D}_{CH}^{(K)}$) to the relative boundary of the feasible region of $\tilde{w}$. All of the quantities above as well as the quantities needed in \emph{Algorithm 1} can be computed from the full sample ($\{x_n,y_n\}; n = 1,2..., N$), the $\nu$-weighted $L2$-norm (\ref{l2-norm}), and inner product (\ref{innerprod}).
\begin{table}
\centering
\scalebox{1.3}{%
\begin{tabular}{|l|}
\hline
\textbf{Algorithm 1 (FW)}\\
\hline
\bf{Given}: $\{l_n (\theta)\}_{n=1}^N,\, m, \langle . \rangle, \, \mathcal{D}_{CH}^{(K)}(\{\hat{p}^{(i)}\})_{i \in [K]}$ \\
\textbf{for} $i$ in $\{ 1, 2, ..., K \}$ \textbf{do} \\
$\quad p_0 \leftarrow \hat{p}^{(i)} \qquad \qquad \qquad \qquad \qquad \qquad$ \textcolor{blue}{ $\triangleleft$ initialize $p$ to $i^{th}$ vertex in $\mathcal{D}_{CH}^{(K)}(\{\hat{p}^{(i)}\})_{i \in [K]}$}\\
$\quad \forall n \in [N]$, compute $||l_n||$ and $\sigma^{(i)}$ from (\ref{consts}) \quad \textcolor{blue}{$\triangleleft$ compute norms} \\
$\quad f_0 \leftarrow \argmax_{n \in [N]} \langle L^{(i)}, \frac{l_n}{||l_n||} \rangle$ \qquad \qquad \qquad \textcolor{blue}{$\triangleleft$ select vertex in $\tilde{w}$-polytope greedily} \\
$\quad \tilde{w}_0^{(i)} \leftarrow \frac{\sigma^{(i)}} {|| l_{f_0}||} \mathbf{1}_{f_0}$ \qquad \qquad \qquad \qquad \qquad \textcolor{blue}{$\triangleleft$ initialize $\tilde{w}$ (for $p = \hat{p}^{(i)}$) with full weight on $f_0$}\\
\quad \textbf{for} $t$ in $\{ 1, 2, ..., m-1 \}$ \textbf{do} \qquad \qquad \qquad\textcolor{blue}{$\triangleleft$ carry out the following for $m-1$ iterations for each $\hat{p}^{(i)}$}\\
$\quad \quad f_t \leftarrow \argmax_{n \in [N]} \inner {L^{(i)} - L(\tilde{w}^{(i)}_{t-1})}{\frac{l_n}{||l_n ||}}$ \qquad \textcolor{blue}{$\triangleleft$ find FW index at $t^{th}$ iteration}\\
$\quad \quad \gamma_t \leftarrow \frac {\inner {\frac{\sigma^{(i)}} {|| l_{f_t}||} l_{f_t} - L(\tilde{w}^{(i)}_{t-1})} {L^{(i)} - L(\tilde{w}^{(i)}_{t-1})}} {\norm{\frac{\sigma^{(i)}} {|| l_{f_t}||} l_{f_t} - L(\tilde{w}^{(i)}_{t-1})}^2}$ \qquad \qquad \qquad\textcolor{blue}{$\triangleleft$ closed from line search for step-size $\gamma$ at $t^{th}$ iteration}\\
$\quad \quad \tilde{w}^{(i)}_t \leftarrow (1-\gamma_t ) \tilde{w}^{(i)}_{t-1} + \gamma_t \frac{\sigma^{(i)}} {\tilde{\sigma}_{f_t}} \mathbf{1}_{f_t}$ \qquad \qquad \quad \textcolor{blue}{$\triangleleft$ add or reweight $f_t^{th}$ data point }\\
\quad \textbf{end for}\\
\textbf{end for}\\
\textbf{return} $\{\tilde{w}, p\} = \{\tilde{w}^{(I)}_{m-1}, \hat{p}^{(I)}\} = \argmax_{i \in [K]}\,\tilde{J}(\tilde{w}^{(i)}_{m-1}, \hat{p}^{(i)})$\\
\hline
\end{tabular}}
\end{table}\label{algo1}
The proof of the first inequality in (\ref{theorem1}) can be obtained in a straightforward manner from Jensen's inequality. In particular, applying Jensen's inequality to the convex function $\phi(.):= (.)^2$, one has:
\begin{equation} \phi(\mathbf{E}_{\theta \sim \nu}(.)) \leq \mathbf{E}_{\theta \sim \nu}\,[\phi(.)]\end{equation} Thus, choosing, $(.)$ as $ L(\tilde{w}, \theta) - L(\theta)$ and $L(\tilde{w}, \theta)$ respectively, one gets:
\begin{eqnarray}
| \left[ \mathbf{E}_{\theta \sim \nu}\, (L(\tilde{w}, \theta) - L(\theta)) \right] | &\leq & (\mathbf{E}_{\theta \sim \nu}\, (L(\tilde{w}, \theta) - L(\theta))^2)^{1/2} = || L(\tilde{w}, \theta) - L(\theta) ||_{\nu, 2}, \nonumber \\
|\left[ \mathbf{E}_{\theta \sim \nu}\, (L(\tilde{w}, \theta) \right] | &\leq & (\mathbf{E}_{\theta \sim \nu}\, (L(\tilde{w}, \theta) )^2)^{1/2} = || L(\tilde{w}, \theta) - L(\theta) ||_{\nu, 2}
\end{eqnarray}
\begin{eqnarray}
\implies | R(\nu) | &=& |\mathbf{E}_{\theta \sim \nu}\, L(\theta)| \leq |\left[ \mathbf{E}_{\theta \sim \nu}\, (L(\tilde{w}, \theta) \right] | + | \left[ \mathbf{E}_{\theta \sim \nu}\, (L(\tilde{w}, \theta) - L(\theta)) \right] | \nonumber \\ &\leq& || L(\tilde{w}, \theta) ||_{\nu, 2} + || L(\tilde{w}, \theta) - L(\theta) ||_{\nu, 2}
\end{eqnarray} where we have used the definition of the expected full sample risk from (\ref{expected-sample-risk}).
The details of the proof of the second inequality in (\ref{theorem1}) are much more involved and are provided in the appendix. Here we point out some salient features. First we can write
\begin{equation} \label{objective2} || L(\tilde{w}, \theta) - L(\theta) ||^2_{\nu, 2} = (\tilde{w}-p)^T \tilde{K} (\tilde{w} -p)\,\;\; \mathrm{with}\;\; \tilde{K}_{n,m} := \langle l_n, l_m\rangle, \end{equation} For notational simplicity we denote the RHS above
as $ \tilde{J}(\tilde{w},p)$, which we interpret as an objective function to optimize. The above problem is similar to the Hilbert coreset problem posed in (\ref{FW-init}) with $J(w)$ in (\ref{FW-init}) replaced
by a slightly different quantity $\tilde{J}(\tilde{w}, p)$. In contrast to (\ref{FW-init}), the objective function $\tilde{J}(\tilde{w}, p)$ in (\ref{objective2}) depends on two vectors $\{\tilde{w}, p\}$. Since the goal is to provide a bound for all $D \in \mathcal{D}_{CH}^{(K)}$, one can consider the following optimization problem by adapting the problem in (\ref{FW-init}) to this case:
\begin{eqnarray} \label{wp-constraints}
& & \max_{D = \{p_n\} \in \mathcal{D}_{CH}^{(K)}}\, \min_{\tilde{w}}\,(\tilde{w}-p)^T \tilde{K} (\tilde{w}-p) \\
&s.t. & \sum_{n=1}^N \tilde{w}_n ||l_n|| = \sum_{n=1}^N p_n ||l_n||, \, \tilde{w} \geq 0 \nonumber \\
&s.t. & p \in \mathcal{D}_{CH}^{(K)}(\{\hat{p}^{(i)}\})_{i \in [K]},\; \sum_{n=1}^N p_n = 1, p \geq 0. \nonumber
\end{eqnarray} Please refer to the appendix for the full proof. An important point to note about the result in Theorem \ref{theorem1} is that the upper bound on the full sample risk denoted by the RHS in (\ref{theorem1-ineqB}) approximately goes like $\hat{\beta}^m$ even for moderate number of Frank-Wolfe iterations $m$.
Theorem (\ref{theorem1}) is valid for the $\nu$-weighted $L2$-norm. However, this norm and the associated inner product involves computing an expectation over the parameters $\theta$ (see (\ref{l2-norm}) and (\ref{innerprod})) which does not admit a closed form evaluation in general. The expectation can be estimated in an unbiased manner by the method of random projections, as in \cite{Campbell2017, Campbell2018, Rahimi2008}. In particular, we construct a random projection of each vector $l_n(\theta); n \in [N]$ into a finite ($J$) dimensional vector space using $J$ samples $\theta \sim \nu(\theta)$:
\begin{equation} \label{rand-proj}u_n = \sqrt{\frac{1}{J}}[l_n(\theta_1), l_n(\theta_2), ..., l_n(\theta_J)]^T, n \in [N],\end{equation}
which serves as finite dimensional approximation of $L_n(\theta)$. In particular, for all $n, m \in [N]$, we have: \begin{equation} \langle l_n, l_m\rangle \approx u_n^T\,u_m.\end{equation} Leveraging similar techniques as in \cite{Campbell2017}, we can construct \emph{Algorithm 2}, and arrive at the following result.
\begin{table}
\centering
\scalebox{1.3}{%
\begin{tabular}{|l|}
\hline
\textbf{Algorithm 2 (RP-FW)}\\
\hline
\bf{Given}: $\{l_n(\theta)\}_{n=1}^N, \nu (\theta), J, m, \mathcal{D}_{CH}^{(K)}(\{\hat{p}^{(i)}\})_{i \in [K]}$ \\
$\cdot$ $\nu (\theta)$: $\{\theta_j\}^J_{j=1} \sim \nu (\theta) $ \qquad \qquad \qquad \qquad \qquad \qquad \textcolor{blue}{$\triangleleft$ take $J$ samples from posterior $\nu(\theta)$}\\
$\cdot$ $u_n = \sqrt{\frac{1}{J}} [l_n(\theta_1),...,l_n(\theta_J)]^T$ $\forall n \in [N]$ \quad \qquad \qquad \textcolor{blue}{$\triangleleft$ construct random projection vector for $l_n(\theta)$ using (\ref{rand-proj})}\\
$\cdot$ \textbf{return} $\{\tilde{w},p\} = FW[u_n, m, (.)^T (.), \mathcal{D}^{(K)}_{CH} (\{\hat{p}^{(i)}\})_{i \in [K]}]$ \; \textcolor{blue}{$\triangleleft$ construct coreset with random projection vectors $u_n$}\\
\qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \textcolor{blue}{by applying \emph{Algorithm 1} with inner product $\langle u_n, u_m\rangle = u_n^T\,u_m$}\\
\hline
\end{tabular}}
\end{table}\label{algo2}
\begin{theorem} \label{theorem2}
Given the general setup of theorem \ref{theorem1}, for loss functions $l$ such that $l_n (\mu)\,l_m(\mu)$ (with $\mu \sim \nu(\theta) $) is sub-Gaussian with parameter $\xi^2$, for $0 < \delta < 1$, for $\mathcal{D} = \mathcal{D}^{(K)}_{CH} (\{\hat{p}^{(i)}\})_{i \in [K]}$, the output of \emph{Algorithm 2} satisfies with probability $> 1- \delta$:
\begin{eqnarray} \label{ineq2}
|R(\nu)| &\leq& ||L(\tilde{w}) ||_{\nu, 2} + || L(\tilde{w}) - L ||_{\nu, 2} \nonumber \\
&\leq& \left(|| u(\tilde{w})||^2_{\nu, 2} + || \tilde{w} ||_1^2 \,\sqrt{\frac{2\xi^2}{J}\log{\frac{2N^2}{\delta}}}\right)^{1/2} + \left(|| u(\tilde{w}) - u ||^2_{\nu, 2} + || \tilde{w} - p||_1^2 \,\sqrt{\frac{2\xi^2}{J}\log{\frac{2N^2}{\delta}}}\right)^{1/2}, \label{ineq2-A}\\
&\leq& \left(|| u(\tilde{w})||^2_{\nu, 2} + || \tilde{w} ||_1^2 \,\sqrt{\frac{2\xi^2}{J}\log{\frac{2N^2}{\delta}}}\right)^{1/2} + \left( \frac{\hat{\sigma}\,\hat{\eta} \,\hat{\bar{\eta}} \,\hat{\beta}}{\sqrt{\hat{\bar{\eta}}^2\, \hat{\beta}^{-2(m-1)} + \hat{\eta}^2 \,(m-1)}} + || \tilde{w} - p||_1^2 \,\sqrt{\frac{2\xi^2}{J}\log{\frac{2N^2}{\delta}}}\right)^{1/2}, \label{ineq2-B}
\end{eqnarray}
The quantities $\{ \hat{\sigma}, \hat{\eta},\,\hat{\bar{\eta}}, \hat{\beta}\}$ are defined as before. The sub-Gaussianity parameter $\xi$ can be estimated from the loss functions. For the $0-1$ loss function, a simple bound is $\xi \leq 1/2$. Also note that $\{\tilde{w}, p\} = \{\tilde{w}_{m-1}^{(I)}, \hat{p}^{(I)}\}$ where $I \in [K]$ is the index that maximizes $\tilde{J}(\tilde{w}^{(i)}_{M-1}, \hat{p}^{(i)})$ (see \emph{Algorithm 1}). The details of the proof can be found in the appendix.
\end{theorem}
\section{Experiments and Instantiation of Bounds}\label{expts}
In this section, we present our experimental setup for computing Hilbert coreset bounds for three datasets: i) Binary MNIST, and ii) CIFAR-100. MNIST is a very well known image classification dataset and is usually treated as a multiclass problem. We follow the results of \cite{Dziugaite2017} and create a binary classification dataset out of it by labeling the digits 0-4 with class `-1' and digits 5-9 with class `1'. The CIFAR-100 datasets is also a well known dataset with hundred different classes of images.
The distribution on the data in each case is assumed to be a uniform distribution for simplicity, with $p_n = \frac{1}{N}$ for a full sample of size $N$. In other words, the class of data distributions, e.g. the convex hull $\mathcal{D}_{CH}^{(K)}$ is simply a point. This can be extended in a straightforward manner to cases where $\mathcal{D}_{CH}^{(K)}$ using the methods described in the appendix. Finally, the loss function $l$ used for the computation of bounds is the standard 0-1 loss used in classification. Further experimental details for each of these datasets are provided below.
\begin{subsection}{Binary MNIST}\label{mnist}
The MNIST dataset is obtained from Keras. This dataset is split into the training set (60000 images) and test set (10000 images).
We convert this into a binary classification dataset as discussed above. The images are preprocessed by scaling them to float values in the range $[0,1]$.
Keras is used to construct a simple neural network binary classifier, similar to that in \cite{Dziugaite2017}.
In particular, the hypothesis class $\mathcal{H}$ for this dataset is taken to be a feed-forward neural network which has three hidden layers each with 600 nodes and a ReLu activation function.
The output layer of the network uses a sigmoid activation function. The weights at each node are initialized according to a truncated normal distribution with mean $0$ and standard deviation $0.04$.
The biases of the first hidden layer are 0.1, whereas the biases of the second hidden layer and output layer are initialized to 0.
The model is trained on different random subsets of the MNIST dataset using the cross-entropy loss (however the bounds will be provided with respect to the 0-1 loss).The first checkpoint uses 500 data points from the training set. The second checkpoint used 3000 data points, including the 500 from the first checkpoint.The third checkpoint uses 10000 data points, including the 3000 already used, and the final checkpoint uses 30000 data points. The models were trained for 120 epochs for each checkpoint.
\end{subsection}
\begin{subsection}{CIFAR-100}\label{cifar100}
For this dataset, we use a state-of-the art neural network architecture as the hypothesis class and a training algorithm is described in Xie et. al. \cite{xie2020muscle}.
The reason for doing this is the following. As a performer on DARPA Learning with Less Labeling (LwLL) program, the goal of our team was to develop novel theoretical bounds
for other performers who are working on the model-building and training of new algorithms for efficient learning of classifiers with fewer labels. As part of this exercise, we approached the authors in \cite{xie2020muscle}
(who are part of the Information Sciences Institute (ISI) team) for the training models developed by them and developed theoretical bounds for those models (the associated hypothesis classes $\mathcal{H}$). As we will show below, we obtained tight non-vacuous upper bounds for such hypothesis classes as well, indicating the power and diversity of the proposed approach.
We summarize the training algorithm used in \cite{xie2020muscle} for completeness. \cite{xie2020muscle} uses mutual-information based unsupervised $\&$ semi-supervised Concurrent learning (MUSCLE), a hybrid learning approach that uses mutual information to combine both unsupervised and semisupervised learning. MUSCLE can be used as a standalone training scheme for neural networks, and can also be
incorporated into other learning approaches. It is shown that the proposed hybrid model outperforms state of the art for the CIFAR-100 dataset as well as others. The basic idea of MUSCLE is to combine combine
the loss from (semi)-supervised learning methods including real and pseudo labels with a loss term arising from the mutual information between two different transformed versions of the input sample ($x_i$). Furthermore, a consistency loss that measures the consistency of the prediction on an input sample and an augmented sample, can be added as a regularization term. Please refer to \cite{xie2020muscle} and references therein for further details about the training algorithm and neural network models (hypothesis classes).
The model is trained with different random subsets of the data, at the 500, 2500, 4000 and 10000 checkpoints.
\end{subsection}
\subsection{Choice of Prior and Posterior Distributions} \label{prior-posterior}
As explained in section \ref{intro}, the theoretical bounds we develop are stochastic bounds. So, one has to choose an appropriate prior distribution (any distribution independent of the data) and a posterior distribution (that depends on the training data and the training algorithm in general). In our analysis, we choose the prior distribution to be $\pi_0 = N(0, \sigma^2\cdot \mathbf{I})$. Inspired by the analysis of Banerjee {\it et al} \cite{Banerjee2020}, we use a training data-dependent posterior $\nu_S = N(\theta^{\dag}, \Sigma^S_{\theta^{\dag}})$, where $\theta^{\dag}$ denotes the parameters \emph{learned} by training the models (in the previous subsections) on the training set $S$, and the covariance $\Sigma_{\theta^{\dag}}$ is defined as follows:
\begin{eqnarray} \label{hessian}
\Sigma^S_{\theta^{\dag}} &=& diag(\sigma_j^2); \; \sigma_j^{-2} := \max\,[H^S_{j,j}(\theta^{\dag}), \sigma^{-2}], \;\;j= 1,2,..,dim(\Theta) \nonumber\\
\mathrm{where}\;\; H^S(\theta^{\dag}) &:=& \frac{1}{N_S} \sum_{i=1}^{N_S} \, \nabla^2\,l(y_i, \phi(\theta^{\dag}, x_i))\;\;\mathrm{(Diagonal\; Hessian)}
\end{eqnarray}
The anisotropy in the posterior $\nu_S$ can be understood as follows. For directions $j$ in parameter space where $H^S_{j,j}(\theta^{\dag})$ is \emph{larger} than $\sigma^{-2}$, the posterior variance is determined by the Hessian, and will be small enough such that $\theta_j$ will not deviate far from the learned parameters $\theta^{\dag}_j$ when sampling from the posterior $\nu_S$. On the other hand, for directions $j$ in parameter space where $H^S_{j,j}(\theta^{\dag})$ is \emph{smaller} than $\sigma^{-2}$, the posterior variance $\sigma_j^2$ is the same as that of the prior $\sigma^2$. Thus, the posterior standard deviation is upper bounded by the prior standard deviation and different weights and biases are assumed to be independent. One can optimize over the prior standard deviation $\sigma$ to get a good bound.
\subsection{Bound Calculation Workflow}\label{workflow}
Given the data points $(x_n, y_n), n=1,2,..,N_S$ of the training set checkpoint $S$ of a particular dataset, a trained model (with the learned parameters $\theta^{\dag}$) for that checkpoint, and a prior distribution $\pi_0= N(0, \sigma^2\cdot \mathbf{I})$ with an appropriately chosen $\sigma$, the workflow for computing the coreset bounds corresponding to the training set $S$ is as follows:
\begin{itemize}
\item We compute the diagonal Hessian of the (cross-entropy) loss for the trained model at the trained parameters $H^S_{j,j}(\theta^{\dag})$ for the training set $S$ via the methods described in \cite{Sagun2017}. Then we compute the posterior variance($\sigma_j^2$) as described in section \ref{prior-posterior}. This gives the data dependent posterior $\nu_S$ for the training checkpoint $S$.
\item Given $\nu_S$, we compute an approximation to the inner product $\langle \cdot \rangle$ by the method of \emph{random features projection} \cite{Rahimi2008, Campbell2017} as in \emph{Algorithm 2} which converge to the true inner products {\it with high probability} as the number of samples from the posterior ($J$) increases (see Theorem \ref{theorem2}).
\item From the inner products computed with the posterior for the training set $S$, and using the full dataset of size $N$, we compute all the constants $\{ \hat{\sigma}, \hat{\eta},\,\hat{\bar{\eta}}, \hat{\beta}\}$ in Theorems \ref{theorem1} and \ref{theorem2} and then implement \emph{Algorithm 1} with number of iterations $m$ equal to training set checkpoint size $N_S$ to find the coreset weights $\tilde{w}^{(I)}_{m-1}$ and the appropriate probability vector $\hat{p}^{(I)}$.
\item Finally, using the coreset weights and the inner product $(.)^T\,(.)$, we compute \emph{coreset upper bound-I} (first term in the RHS of (\ref{ineq2-A})) as well as \emph{coreset upper bound-II} (RHS of (\ref{ineq2-B})).
\end{itemize}
The above workflow was carried out for all different training checkpoints (of increasing sizes) for both datasets mentioned in sections \ref{mnist} and \ref{cifar100}.
\subsection{Experimental Results}\label{results}
\begin{figure}[h]
\begin{tabular}{ll}
\includegraphics[scale=0.6]{str_binarymnist_plots.png}
&
\includegraphics[scale=0.6]{isi_cifar100_plots.png}
\end{tabular}
\caption{\textbf{Left}: \footnotesize{\emph{Coreset Upper Bound - I} (RHS of (\ref{ineq2-A}), shown in blue) and \emph{Coreset Upper Bound -II} (RHS of (\ref{ineq2-B}), shown in red) on the expected full sample risk for the three-layer ReLU DNN model for four training checkpoints of the Binary MNIST dataset with prior standard deviation $\sigma = 10^{-4}$, based on Theorem \ref{theorem2}. The expected empirical risk (EER) of the trained model on the full sample is shown in green.}
\textbf{Right}: \footnotesize{\emph{Coreset Upper Bound - I} (RHS of (\ref{ineq2-A}), shown in blue) and \emph{Coreset Upper Bound -II} (RHS of (\ref{ineq2-B}), shown in red) on the expected full sample risk for the Xie \emph{et al} \cite{xie2020muscle} model for four training checkpoints of the CIFAR-100 dataset with prior standard deviation $\sigma = 10^{-4}$, based on Theorem \ref{theorem2}. The expected empirical risk (EER) of the trained model on the full sample is shown in green.}
}\label{plots}
\end{figure}
For our numerical results, we choose $J$, the number of samples from the posterior (see section \ref{workflow}), as 7000. We experimented with other values of $J$ and found that the numerical results for the bounds were stable with $J \gtrsim 7000$. We also experimented with the choice of the prior standard deviation $\sigma$. As discussed in section \ref{prior-posterior}, the posterior standard deviation is upper bounded by $\sigma$. So, we would like to choose a low-value of $\sigma$ to come up with an \emph{effectively deterministic} bound. After some experimentation, we chose a value of $\sigma = 10^{-4}$ that seemed to provide good bounds which were effectively deterministic in the sense that the expected empirical risk (EER) is very close to the empirical risk. For each dataset, we use the number of Frank-Wolfe iterations $m$ (in \emph{Algorithm 1}) equal to the size of the training set checkpoint $S$. We find that the size of the coreset (the number of non-zero weights) is \emph{much smaller} than the number of Frank-Wolfe iterations or the size of the training set $S$. For example, for the CIFAR-100 dataset, the coreset size is only $\sim150$, around two orders of magnitude smaller than the size of the largest training set checkpoint $S = 10000$. This is also consistent with results in \cite{Campbell2017, Campbell2018}.
The left plot in Figure \ref{plots} shows that the bounds for the MNIST model are very tight. For example, the empirical expected full sample risk at the first checkpoint with 500 samples (0.49) is quite close to the \emph{coreset upper bound-I} (0.51) and the \emph{coreset upper bound-II} (0.56). We get tight non-vacuous bounds for the DNN hypothesis class used in section \ref{mnist} for all training checkpoints. Also, the coreset upper bounds obtained for the training checkpoint with 30000 samples ($0.145\, \& \,0.154$) are better than that obtained in \cite{Dziugaite2017} for 55000 samples (0.207). As discussed earlier, we use the same model and training procedure as \cite{Dziugaite2017} to allow a fair comparison.
The right plot in Figure \ref{plots} shows the results for the CIFAR-100 model with the state-of-the-art neural network architecture and training algorithm as in \cite{xie2020muscle}. For this model, the \emph{coreset upper bound-I} is non-vacuous at all training checkpoints, while the \emph{coreset upper bound-II} is non-vacuous for training set sizes larger than around 8000. Thus we are able to compute meaningful and effectively deterministic upper bounds for the full sample risk for complex datasets (like CIFAR-100) and state-of-the-art hypothesis classes and training algorithms, as in \cite{xie2020muscle} in section \ref{cifar100}.
The constants in the bounds in Theorems \ref{theorem1} and \ref{theorem2} depend on the posterior distribution $\nu$. As discussed in section \ref{workflow}, the posterior distribution $\nu_S$ is computed for each training checkpoint $S$. Therefore, the constants in the bounds plotted in Figure \ref{plots} also change with each training checkpoint $S$. An important point to note is that the bounds for both datasets/models decrease significantly with the size of the training set $S$. This is in contrast to some other recent bounds for DNNs that may even increase with the size of the training set \cite{Bartlett2017, Golowich2017, Neyshabur2018}, see discussion in \cite{Nagarajan2019}.
Finally, as pointed out in the beginning of section \ref{expts}, we have assumed a uniform distribution on the data for simplicity implying that the class of data distributions $\mathcal{D}$ is trivial. However, it is straightforward to extend these results to non-trivial classes like $\mathcal{D}_{CH}^{(K)}$, using the methods described in the appendix. The expected empirical full sample risk will still be upper bounded by the bounds computed for these data distribution classes as long as the uniform distribution belongs to $\mathcal{D}_{CH}^{(K)}$.
\begin{section}{Conclusions}
In this paper, we have developed the formalism for computing upper bounds on the \emph{expected full sample risk} for supervised classification tasks within a transductive setting using the method of Hilbert coresets. We compute explicit numerical bounds for complex datasets and complex hypothesis classes (such as state-of-the-art DNNs) and compare these to the performance of modern training algorithms. The bounds obtained are tight and meaningful. The bounds we develop are stochastic,
however they are effectively deterministic for appropriate choice of priors. The bounds are \emph{non-uniform} in the hypothesis space, since the precise numerical values of the bound depend on the learned weights of the training algorithm ($\theta_{\star}$). Finally, the bounds we develop decrease significantly with the size of the training set ($S$) as can be seen from section \ref{results}.
There are a couple of avenues for future research. First, it should be possible to derive generalization bounds for other classes of data distributions $\mathcal{D}$, such as in which the probability vector for the full sample lies inside an ellipsoid. Second, as mentioned in section \ref{results}, the coreset size is much smaller than the size of the training set $S$ for a given training checkpoint. This suggests that it might be possible to develop a model-free (agnostic) active learning approach leveraging the coresets framework. For example, given a prior and an initial (presumably small) set of labeled data, one can use the coresets approach to construct an initial posterior on the hypothesis space. Using this posterior, one can make predictions on the unlabeled data and compute a \emph{pseudo risk} for the entire dataset. Then one could use this \emph{pseudo-risk} to construct the next set of coreset weights and update the coreset weighted risk. At each step, one should also update the posterior appropriately. The goal would be to show that this procedure converges and provides an upper bound on the expected true risk. The hope is that this should provide much tighter generalization and sample complexity bounds.
\end{section}
\section{Acknowledgement}
The material in this paper is based upon work supported by the United States Air Force and DARPA under Contract No. FA8750-19-C-0205 for the DARPA I20 \emph{Learning with Less Labeling (LwLL)} program. S.D., P.K., and R.K.P are grateful to Prof. T. Broderick for providing inspiration to the ideas in this paper as well as providing feedback on the manuscript. The authors would like to particularly thank Prof. V. Chandrasekaran and Prof. A. Willsky for insightful discussions and comments on the manuscript.
|
\section{Introduction}
Radioisotopes production for medical applications has led to the development and improvement of cyclotrons, which currently play a remarkable role in the sustainability of the radionuclide supply~\cite{oliver-1}, especially for positron emission tomography (PET) scans. The progress on the cyclotron technology over the last decades ~\cite{schmor} has resulted in very mature, compact and versatile products, which are commercially available in the market~\cite{smirnov}. As a result of this progress, the number of cyclotron facilities for medical purposes is expanding worldwide~\cite{IAEA-06}, offering alternatives for the production of medical radioisotopes such as $^{99m}$Tc~\cite{boschi}, traditionally produced in nuclear reactors.
Cyclotrons for biomedical radionuclide production are typically based on the acceleration of light ions up to tens of MeV. Hydrogen anions are selected in most of the present facilities as the accelerated light ion beam because of the simplification of the beam extraction from the cyclotron to the target. For the production of H$^-$ beam, the use of internal ion sources has demonstrated to be the most cost-effective and compact solution~\cite{ehlers,an}. Extraction systems based on thin stripper foils results in an effective mechanism able to modify the extracted ion energy, while avoiding any turn separation requirement at the extraction point~\cite{jasna,kleeven}. In addition, some last-generation cyclotrons integrate superconducting magnets to provide intense magnetic fields with advantages in power consumption, size, weight, and therefore in compactness.
Nevertheless, the use of H$^-$ internal ion sources has as a drawback the deterioration of the vacuum level in the acceleration chamber, especially in the central region~\cite{ehlers,yang}. Part of the H$_2$ gas injected continuously in the ion source to produce the H$^-$ beam is extracted from the ion source slit together with the beam, configuring the main source of residual gas of the vacuum chamber. Since the binding energy of the hydrogen anions ($<\!1$~eV)~\cite{lykke} is low, the interaction between the H$_2$ residual gas and the H$^-$ beam plays an important role in the transport of the beam along the cyclotron. In addition, the high magnetic field produces a strong electric field in the rest-frame of the H$^-$, which could induce the detachment of the second slightly bound electron limiting the lifetime of high-energy ions. These interactions reduce the survival rate of the beam and consequently the total beam transmission. As we will discuss in the following sections, they represent a non-negligible source of sustained and uncontrolled losses that could damage the vacuum chamber walls and/or activate the chamber, increasing the maintenance time and reducing the availability. Therefore, the magnitude of the impact of these issues for compact cyclotrons~\cite{papash,Zhang,pradhan} imposes a dedicated study, which could be also applied to other type of accelerators~\cite{carneiro,plum}.
In this paper, firstly, we review H$^-$ beam stripping interactions from residual gas and magnetic fields. We then describe the implementation of those physical processes into the beam dynamics code \textsc{OPAL}\xspace~\cite{opal}. And finally, we use this implementation to evaluate the impact and optimize the beam dynamics of the AMIT cyclotron~\cite{cyclomed}, which is a cutting-edge compact accelerator aimed to produce on-demand short-lived radioisotopes to enhance medical nuclear imaging. This is achieved by an autonomous and small-sized design in order to provide an alternative to the centralized radioisotope production.
\section{Beam stripping interactions}
The stripping interactions are stochastic processes in which each individual particle could experience the interaction according to a certain probability. Let's assume incident particles on a homogeneous medium are subjected to a process with a mean free path $\lambda$ between interactions. Hence, the probability of suffering an interaction before reaching a path length $x$ is~\cite{leo}:
\begin{equation}
P(x) \;=\; 1 - e^{-x/\lambda},
\label{eq:prob_1}
\end{equation}
where $P(x)$ represents the cumulative interaction probability of the process.
When the interaction occurs between a beam colliding with particles of a material, the mean free path is described in terms of the cross section, $\sigma$, and the particle density of interacting centers of the material, $n$:
\begin{equation}
\lambda = \frac{1}{n\,\sigma}.
\end{equation}
Likewise, the mean free path could be connected with the mean time of interaction or lifetime, $\tau$:
\begin{equation}
\tau=\frac{\lambda}{\beta c}=\frac{1}{n\sigma\beta c},
\end{equation}
where $\beta$ is the ratio of the velocity of the particles over the speed of light, $c$.
\subsection{Residual gas stripping}
\label{sec:residual}
Residual beam interactions with the molecules of the remaining gas in the acceleration region entail the most important source of losses in a H$^-$ compact cyclotron.
Assuming a beam flux incident in a gas with density $N$ (number of the gas molecules per unit volume under the vacuum condition), the beam fraction lost per unit of traveled length will be, in compliance with Eq.~\eqref{eq:prob_1}:
\begin{equation}
f_g=1-e^{-N\sigma x},
\label{eq:fraction_lost_gas}
\end{equation}
where the gas density, $N$, is easily obtained from the ideal gas equation. If several types of molecules are considered in the residual gas as well as different physical processes of comparable significance, the total mean free path is given by:
\begin{equation}
\frac{1}{\lambda_{total}}=\sum_j \frac{1}{\lambda_j}=N_{total}\cdot\sigma_{total}=\sum_jN_j\,\sigma^{\,j}_{total}=\sum_j\left(\sum_iN_j\,\sigma^{\,j}_{i}\right),
\end{equation}
where $N_j$ is the particle density of each gas component of the residual gas, and $\sigma_i$ represents the cross section of each of the physical phenomena.
In case of H$^-$ ion beams the second electron is bound with a low energy of $\varepsilon=0.754195(19)$~eV~\cite{lykke}. Hence, there is a relevant probability of electron detachment during the acceleration process through interaction with the residual gas.
The experimental cross section for charge-transfer interactions of hydrogen atoms and ions has been measured since the 1950s for interactions with different gases. The data have been published and compiled in different reports~\cite{Allison,barnett-2,Nakai,phelpsH2,barnett-chebyshev,phelpsAr}. Moreover, theoretical studies for high energy cross sections have been developed based on an extension of the Bethe theory for the total inelastic cross section in the Born approximation~\cite{Gillespie1,Gillespie2,Gillespie3}.
In addition, different analytic methods to fit the cross section data have been formulated. One of them~\cite{barnett-chebyshev} is based on fitting the recommended cross sections as function of the projectile energy, $E$, using least-squares method with Chebyshev polynomials:
\begin{equation}
\ln{[\sigma(E)]}=\frac{1}{2}a_0 + \sum_{i=1}^{k} a_i \cdot T_i(X),
\label{eq:barnett}
\end{equation}
\begin{equation}
X=\frac{(\ln{E}-\ln{E_{min}})-(\ln{E_{max}}-\ln{E})}{\ln{E_{max}}-\ln{E_{min}}},
\end{equation}
\vspace{-2pt}
\noindent where $T_i$ are the Chebyshev orthogonal polynomials, $a_i$ ($i=0,1,...,k$) denote adjustable parameters relative to each reaction, $k$ is the smallest number of coefficients providing an accurate fit and $E_{min}$ and $E_{max}$ are parameters that limit the region of analytic representation of the cross section. This procedure facilitates the data interpolation in a restricted measured energy range; however, it cannot be used for extrapolation, because it often shows non-physical behavior just outside the considered energy interval.
A more robust method~\cite{Nakai} makes use of analytic expressions that approximate low-energy and high-energy asymptotic trends. It is based on a semi-empirical expression for inelastic collision cross sections~\cite{Green} developing functional forms to fit a compiled set of experimental data using the two-step least-squared method. For the considered reactions, the analytical function formula is given by the general expression:
\begin{equation}
\sigma_{qq'} = \sigma_0 \left[ f(E_1) + a_7\!\cdot\!f(E_1/a_8) \right],
\label{eq:CS_Nakai}
\end{equation}
where $qq'$ represents any combination of initial and final charge states of the particle, $\sigma_0$ is a convenient cross section unit ($\sigma_0 = 1\!\cdot\!10^{-16}\;\text{cm}^2$), whereas $f(E)$ and $E_1$ are given by:
\begin{equation}
f(E) = \frac{ a_1\!\cdot\!\left(\!\displaystyle\frac{E}{E_R}\!\right)^{\!a_2} }{ 1+\left(\!\displaystyle\frac{E}{a_3}\!\right)^{\!a_2+a_4}\!\!+\left(\!\displaystyle\frac{E}{a_5}\!\right)^{\!a_2+a_6} },
\label{eq:CS_Nakai2}
\end{equation}
\vspace{-2mm}
\begin{equation}
E_1 = E_0 - E_{th},
\label{eq:CS_Nakai4}
\end{equation}
being $E_0$ the energy of the incident particle in keV, $E_{th}$ is the threshold energy of the reaction in keV, the symbols $a_i$ denote adjustable parameters and $E_R$ is the Rydberg energy multiplied by the ratio of the atomic hydrogen mass, $m_{H}$, to the electron mass, $m_e$:
\begin{equation}
E_R = hcR_{\infty}\!\cdot\!\frac{m_H}{m_e} = \frac{m_He^4}{8\varepsilon_0^2h^2},
\label{eq:CS_Nakai3}
\end{equation}
where $h$ is the Planck constant, $c$ is the speed of light in vacuum, $R_{\infty}$ is the Rydberg constant, $e$ is elementary charge and $\varepsilon_0$ is the vacuum electric permittivity.
This analytic expression has been improved for reactions of hydrogen ions with hydrogen gas~\cite{TabataShi} by means of linear combinations of $f(E)$, taking into account additional experimental data and considering more setting parameters. The enhancement of the function makes it possible to extrapolate the cross section data to some extent.
\subsection{Electromagnetic stripping}
\label{sec:lorentz}
H$^-$ ions traveling in a high-magnetic field might undergo single-electron detachment reactions due to the opposite bending force experienced by electrons and nucleus according to their electric charge. The field component orthogonal to the velocity of the particles produces an electric field, $\mathcal{E}$, in the ion's rest-frame according to the Lorentz transformation ($\mathcal{E}\!=\!\gamma\beta c B$). The strength of this effect is a function of the energy of the ions and the magnetic field, $B$. This effect, called electromagnetic or Lorentz stripping, is relevant just for H$^-$ beams due to the low binding energy of the second electron. Only single-electron detachment processes are expected~\cite{Furman} because of the high binding energy of the first electron ($13.598434600291(12)$ eV~\cite{Mohr}). The fraction of H$^-$ beam particles dissociated by the electromagnetic field after a traveled distance $L$ can be evaluated through the interaction probability (see Eq.~\eqref{eq:prob_1}):
\begin{equation}
f=1-e^{\,-\,L/\beta c\gamma\tau}=1-e^{\,-\,t/\gamma\tau},
\end{equation}
where $\tau$ is the particle lifetime in the rest frame of the ions.
The process can be analyzed from the decay of an atomic system in a weak and static electric field. For sufficiently high fields, the potential perceived by the electrons is modified, decreasing at some distance below the binding energies. The dissociation of the atomic system occurs by tunneling through the potential barrier into a decay channel giving rise to the ground state of the daughter atom. A theoretical study based on the calculation of the electric dissociation rate directly from the formal theory of decay have obtained an expression for the of H$^-$ ion lifetime~\cite{Scherk}:
\begin{equation}
\tau= \frac{4m_ez_T}{S_0\mathcal{N}^2\hslash\,(1+p)^2\left(1-\displaystyle\frac{1}{2k_0z_t}\right)}\,\cdot\,\exp\!{\left(\frac{4k_0z_T}{3}\right)}
\label{lifetime}
\end{equation}
where $z_T=\varepsilon/eE$ is the outer classical turning radius, $\varepsilon$ is the electron binding energy, $S_0$ is the spectroscopic coefficient of finding the daughter atom in the ground state, $p$ is the polarization of the ionic wave function, $k_0$ is a parameter derived from the ionic wave function and determined from the relation $k_0^2\!=\!2m_e\varepsilon/\hslash^2$ and $\mathcal{N}$ is a normalization factor given by:
\begin{equation}
\mathcal{N}=\frac{[2k_0(k_0+\alpha)(2k_0+\alpha)]^{1/2}}{\alpha},
\end{equation}
being $\alpha$ a parameter for the ionic potential function.
This lifetime expression can be parametrized in terms of the electric field, $\mathcal{E}$ and the atomic properties:
\begin{equation}
\label{lifetime_}
\tau =
\frac{\sqrt{2\varepsilon m_e}\,\alpha^2}{S_0(1+p)^2\,e\left(\negthickspace\displaystyle\frac{\sqrt{2\varepsilon m_e}}{\hslash}\!+\!\alpha\negmedspace\right)\negmedspace\nms\left(\negthickspace\displaystyle\frac{\sqrt{8\varepsilon m_e}}{\hslash}\!+\!\alpha\negmedspace\right)\negmedspace\nms\left(\nts1\!-\!\displaystyle\frac{\sqrt{2}\,e\,\hslash\,\mathcal{E}}{4\sqrt{m_e}\,\varepsilon^{3/2}}\right)}\,\frac{1}{\mathcal{E}}\cdot \exp{\negmedspace\left[\frac{2}{3}\!\left(\frac{\displaystyle8m_e}{\displaystyle e^2\,\hslash^2}\right)^{\negthickspace\nts1/2}\!\frac{\varepsilon^{3/2}}{\mathcal{E}}\right]},
\end{equation}
\noindent where $m_e$ is the electron mass, $e$ is the elementary charge and $\hbar$ is the reduced Planck constant. The values of the atomic parameters are already known: $p~=~0.0126$~\cite{Scherk}, $S_0=0.783\,(5)$~\cite{Keating} and $\alpha=3.806\!\cdot\!10^{10}\;\text{m}^{-1}$~\cite{Tietz}. This expression can be written in a simpler way as:
\begin{equation}
\tau=\frac{a_F}{(1-\eta \mathcal{E})\,\mathcal{E}}\,\cdot\,\exp{\!\left(\frac{b_F}{\mathcal{E}}\right)},
\label{lifetime_P2}
\end{equation}
\begin{equation}
a_F=\frac{\sqrt{2\varepsilon m_e}\,\alpha^2}{S_0\,(1+p)^2\,e\left(\negthickspace\displaystyle\frac{\sqrt{2\varepsilon m_e}}{\hslash}\!+\!\alpha\negmedspace\right)\negmedspace\nms\left(\negthickspace\displaystyle\frac{\sqrt{8\varepsilon m_e}}{\hslash}\!+\!\alpha\negmedspace\right)},
\end{equation}
\begin{equation}
b_F=\frac{2}{3}\left(\frac{\displaystyle8m_e}{\displaystyle e^2\,\hslash^2}\right)^{\negmedspace\nms1/2}\!\varepsilon^{3/2} = \frac{4}{3}\frac{\varepsilon^{3/2}}{\sqrt{e\,\hslash\,\mu_B}},
\end{equation}
\begin{equation}
\eta=\frac{\sqrt{2}\,e\,\hslash}{4\sqrt{m_e}\,\varepsilon^{3/2}} = \sqrt{\frac{m_e}{2}}\,\frac{\mu_B}{\varepsilon^{3/2}},
\end{equation}
\vspace{0.5mm}
\noindent where $\mu_B$ is the Bohr magneton. The parameters depend on fundamental constants and ion properties, taking values of $a_F=2.65\,(2)\cdot 10^{-6}$ s V/m, $b_F=4.4741\,(2)\cdot 10^{9}$ V/m and $\eta=1.49007\,(5)\cdot 10^{-10}$ m/V.
Apart from that, prior to this theoretical approach, a simpler expression for the rest-frame lifetime was obtained experimentally~\cite{Stinson,Jason,Keating} as a function of some setting parameters ($A_1$, $A_2$):
\begin{equation}
\tau=\frac{A_1}{\mathcal{E}}\,\cdot\,\exp{\!\left(\frac{A_2}{\mathcal{E}}\right)}.
\end{equation}
This expression has been widely used in previous studies as a first approximation to determine the influence of the magnetic field in a H$^-$ beam. However, it could be derived from the theoretical lifetime (Eqs.~\eqref{lifetime_} or~\eqref{lifetime_P2}) taking into account some approximations: disregarding the polarization factor, $p$, and considering $\eta E \!<\!<\!1$. Thus, a relationship for the experimental constants $A_1$ and $A_2$ as a function of the binding energy can be found as:
\begin{equation}
A_1(\varepsilon)=C_1\,\frac{\varepsilon}{S_0N^2},
\end{equation}
\vspace{-3mm}
\begin{equation}
A_2(\varepsilon)=C_2\,\varepsilon^{3/2},
\end{equation}
where $C_1$ and $C_2$ are constants independent of the atomic structure that can be referred to as fundamental constants~\cite{nist,codata} by means of the theoretical expression under the considered approximation:
\begin{equation}
C_1=\frac{4\,m_e}{e\,\hslash}=\frac{2}{\mu_B}=2.157\cdot 10^{23} \;\text{A}^{-1}\,\text{m}^{-2},
\end{equation}
\begin{equation}
C_2=\frac{4}{3}\frac{1}{\sqrt{e\,\hslash\,\mu_B}}=1.065\cdot 10^{38}\;\;\text{T}^{1/2}\text{A}^{-1/2}\,\text{J}^{-1}\,\text{s}^{-1}.
\end{equation}
Hence, the parameters of the experimental mean lifetime expression can be reassessed: $A_1=2.714\,(17)\cdot10^{-6}$ s V/m and $A_2=4.47407\,(17)\cdot10^{9}$ V/m, in good agreement with previous values~\cite{Jason,Keating}. It is important to emphasize that these parameters have always been determined by experimental adjustments. Therefore, their assignment based on fundamental constants provides an important enhancement for the study of electromagnetic stripping.
The evaluation of the lifetime for a H$^-$ beam shows that the Lorentz stripping becomes relevant for high-energy beams immersed in high-magnetic fields. Hence, this process has to be assessed and quantified in the design of an accelerator in order to prevent unexpected beam losses.
\section{\textsc{OPAL}\xspace computational model for beam stripping}
\textsc{OPAL}\xspace (Object Oriented Parallel Accelerator Library) is a versatile open source Particle-In-Cell (PIC) code developed for large-scale charged-particle optics simulations in accelerators and beam lines~\cite{opal}. It is organized into two different flavors, one of them (\textsc{OPAL-cycl}) exclusively dedicated to cyclotrons. The particles are evolved in time by either a fourth-order Runge-Kutta method or a second-order leapfrog according to the collision-less Vlasov-Poisson equation.
One of the more attractive features of \textsc{OPAL}\xspace is the capability to perform beam interactions with matter through Monte Carlo simulations. The physical reactions considered by the code have been recently extended with the implementation of beam stripping reactions~\cite{calvo}. This new feature enhances the potential and versatility of OPAL, providing the capability to extend beam dynamics studies to cyclotrons where this could be a relevant issue.
The beam stripping algorithm evaluates the interaction of hydrogen ions with the residual gas and the electromagnetic fields. In the first case, the cross sections of the processes are estimated according to the energy by means of analytical functions (see Sec.~\ref{sec:residual}). The implementation allows the user to set the pressure, temperature and composition of the residual gas, which could be selected for the calculations as either molecular hydrogen (H$_2$) or dry air in the usual proportion~\cite{picard}. For precise simulations, a 2D pressure field map from an external file can be imported into \textsc{OPAL}\xspace, providing more realistic vacuum conditions. Concerning electromagnetic stripping, the electric dissociation lifetime of H$^-$ ions is evaluated through the theoretical formalism (see Sec.~\ref{sec:lorentz}). In both instances, the individual probability at each integration step for every particle is assessed.
A stochastic evaluation method through an uniformly generated random number is used to evaluate if a physical reaction occurs. In case of interaction, it will be stripped and removed from the beam, or optionally transformed to a secondary heavy particle, compliant with the occurred physical phenomena. In this case, the secondary particle will continue its movement in agreement with the charge-to-mass ratio. Fig.~\ref{fig:flowchart} summarizes the iterative steps evaluated by the algorithm until the end of the accelerating process, or until the particle is removed from the beam.
\begin{figure}[h]
\centering
\includegraphics [width=0.55\textwidth] {OPALBSTP_flowchart.pdf}
\caption{\textsc{OPAL}\xspace beam stripping physics algorithm flowchart.}
\label{fig:flowchart}
\end{figure}
\section{Beam stripping assessment}
\subsection{\textsc{AMIT}\xspace cyclotron}
\label{sec:amitCyclotron}
\begin{figure}[!t]
\centering
\includegraphics[width=8.6cm]{Interior_cyclotron_1_text4.jpg}
\caption{Top view of the \textsc{AMIT}\xspace cyclotron 3D model. The main components, the center and the axes of the global reference frame and the scale are indicated.}
\label{fig:AMIT_3D}
\end{figure}
The \textsc{AMIT}\xspace (Advanced Molecular Imaging Technologies) cyclotron is a superconducting weak focusing cyclotron designed for single-dose production of $^{18}$F and $^{11}$C radionuclides. It is a Lawrence-type machine accelerating a H$^-$ beam current of $10\,\mu$A up to $8.5$~MeV~\cite{oliver}. Its compact configuration (Fig.~\ref{fig:AMIT_3D}) is aimed to contribute to the deployment in hospitals for on-site production of short-life radioisotopes. The very compact design of the accelerator is achieved by the combination of: 1) a high magnetic field (central value of $4$~T) created by a superconducting magnet of NbTi~\cite{garcia-tabares}, and 2) with the use of an internal Penning Ionization Gauge (PIG)-type ion source. A radiofrequency system~\cite{gavela-2} of $60.134$~MHz in a one $180\degree$ dee configuration, at the end of a quarter wave coaxial resonator, provides the electric field for the beam acceleration. The required final energy imposes a $60$~kV accelerating peak voltage. An independent vacuum chamber, adapted to the magnet aperture, holds the accelerating system, the ion source, the extraction system, and the beam diagnostics. The stripping-based extraction provides the final proton beam, making use of a high-efficiency mechanism through thin carbon foils. The evaluation of the extraction process resulted in total charge transfer with low energy losses for the optimized foil thickness~\cite{calvo_phd}. The stripper is installed in a movable mechanical structure, enabling the energy tuning of the resultant beam.
\subsection{Vacuum system}
The vacuum system of the \textsc{AMIT}\xspace cyclotron has been designed with the goal of achieving at least an absolute residual gas pressure below $10^{-5}$~hPa in the acceleration chamber. The system has to deal with the compactness of the accelerator that limits the space available to install the vacuum pumps considering the fringe magnetic field and the radiation environment among other issues. The independent acceleration chamber of $700$~mm length, $52$~mm height, and $280$~mm width is made of copper (surface area of $0.26$~m$^2$) and steel (surface area of $0.5$~m$^2$). It hosts the internal ion source, aligned with the axial direction, and the RF cavity with a $12$~mm free height inside the dee electrode.
The vacuum system has been simulated with Molflow+~\cite{molflow-1,molflow} to evaluate different configurations of the vacuum pumps, taking into account the pumping system capacity, the impact on the whole accelerator size, the maintenance requirements, as well the interactions with other cyclotron subsystems. From those simulations, a configuration with two diffusion pumps connected to the cylindrical resonator was chosen as a preferred option. The pumps system has the capability to achieve an ultimate pressure below $10^{-7}\!$~hPa, but due to the continuous gas injection from the ion source ($0-10$~sccm), the pressure level in the chamber during operation is much higher (around $10^{-4}-10^{-5}\!$~hPa).
Molflow+ considers steady state molecular flow of an ideal gas for the simulations. Therefore, the mean free path of the residual gas particles or molecules is considerably greater than the dimensions of the vacuum vessel. The vacuum level is determined by the throughput, depending on the pressure gradient and the conductance. Vacuum calculation for the \textsc{AMIT}\xspace chamber has been performed with Molflow+ considering H$_2$ residual gas, a conductance of $0.70\;\text{l}/\text{s}$ and a throughput of $4.3\cdot10^{-3}\;\text{Pa}\!\cdot\!\text{m}^3/\text{s}\,= 2.546\;\text{sccm}$.
The results are easily scalable if other input parameters have to be considered.
\begin{figure}[!b]
\centering
\subfigure[]
{\includegraphics [width=8.6cm] {PressureFieldmapGapx.pdf}} $\;$
\subfigure[]
{\includegraphics[width=8.6cm]{PressureFieldmapGapy_dee.pdf}}
\caption{Pressure level along $x$-coordinate at the gap center (a) and along $y$-coordinate at $x=0$ (b). The figures are referred to the cyclotron center (see Fig.~\ref{fig:AMIT_3D}). Thus, $y$-positive values in subfigure (b) correspond to the dee region, whereas the negative part represents the dummy-dee area and the rest of the vacuum chamber. The red line represents the raw data from Molflow+ simulations and the blue line is a spline fit.}
\label{fig:PressureFieldmapXY}
\end{figure}
The simulation results (see Fig.~\ref{fig:PressureFieldmapXY}) show an average value of the pressure of $7\cdot10^{-5}$~hPa. However, the vacuum level is worse in the central region ($\sim 1.7\cdot10^{-4}$~hPa), because in front of the ion source slit the gas flux from the source and the low conductance increases the pressure significantly. Furthermore, the dee region has a slightly larger pressure due to the gas evacuation limitations through the opening holes at the back of the structure. Fig.~\ref{fig:PressureFieldmapXY} illustrates the pressure level in the vacuum chamber along the center of the accelerating gap (at $y=0$) and along the perpendicular direction passing through the center of the machine (at $x=0$). Given this vacuum level, a relevant effect of the beam stripping interactions in the beam transmission through the cyclotron is expected.
\subsection{\textsc{AMIT}\xspace beam stripping analysis}
The compact design of the \textsc{AMIT}\xspace cyclotron requires a careful analysis of the beam stripping interactions in order to evaluate their contribution to the final beam losses. The implementation of these particle-matter interactions into \textsc{OPAL}\xspace has allowed the assessment of both, the electromagnetic stripping and the residual gas interactions. The beam dynamics simulations with beam stripping interactions have been performed under the nominal accelerations conditions of the cyclotrons as it has been exposed in the previous section~\ref{sec:amitCyclotron}.
\subsubsection{Electromagnetic stripping}
A preliminary calculation shows that a $4$~T magnetic field for the maximum achievable energy of $8.5$~MeV in the \textsc{AMIT}\xspace cyclotron corresponds to a beam-rest-frame electric field of $E\!=\!160$~MV/m. This entails a marginal beam fraction loss per unit length of $1.42\cdot10^{-6}$~m$^{-1}$ due to electromagnetic stripping obtained from the theoretical evaluation. Therefore, Lorentz stripping is not expected to have any noteworthy contribution to the stripping losses for the current cyclotron configuration due to the low energy of the beam. The \textsc{OPAL}\xspace simulations have corroborated the negligible contribution of electromagnetic stripping to the beam losses, and consequently it can be considered as a source of second order losses.
\subsubsection{Gas stripping}
\begin{figure}[!t]
\centering
\subfigure[]
{\includegraphics [width=8.6cm] {BSTP-XYdist1-H2.pdf}} \qquad
\subfigure[]
{\includegraphics [width=8.6cm] {BSTP-XYdist3-H2.pdf}}
\caption{(a) Spatial distributions in the cyclotron plane of the beam losses produced by gas stripping. (b) Normalized distribution over the total amount of stripping losses along different directions.}
\label{fig:BSTP-hist_spatial}
\end{figure}
The gas stripping losses along the acceleration process in \textsc{AMIT}\xspace cyclotron has been characterized through \textsc{OPAL}\xspace simulations including the pressure field map obtained from Molflow+ results. The average rate of stripping losses estimated by the simulation reaches $(14.6\pm1.5)\,\%$ under the considered vacuum conditions. The uncertainty of the result is due to the stochastic error associated with stripping losses. Fig.~\ref{fig:BSTP-hist_spatial} illustrates the spatial characterization of the beam stripping losses along the cyclotron. They are predominantly produced in the first stages of the acceleration, some millimeters after the ion source, for two main reasons. Firstly, the pressure in the central region is one order of magnitude higher in comparison with the rest of the chamber. Secondly, the cross section of H$^-$ interactions with the gas molecules is maximal for the energy range $(0.1-300)$~keV of the beam in the central region. Nevertheless, the stripping interactions persist along the beam acceleration at a lower level given the long traveled distances until the extraction.
\begin{figure}[!t]
\centering
\includegraphics [width=8.6cm] {BSTP-Edist3-Radius-H2.pdf}
\caption{Histogram of normalized energy distribution of the losses. The normalization is over the total amount of stripping losses. Solid orange line with error bars represents the radial location of particles corresponding to each histogram bin associated.}
\label{fig:BSTP-hist}
\end{figure}
The energy characterization of the beam stripping losses is presented in Fig.~\ref{fig:BSTP-hist}. It is manifest that the main part of stripped particles induces low-energy losses ($<\!1$ MeV). However, the prolongation of the stripping interactions along the entire beam path induces high-energy losses, which will entail the continuous damage and degradation of several components in the long term and the increase of the radioactive activation by secondary particles of energy above $2.5$~MeV.
Given the relevance of the beam stripping reactions caused by the interactions with the residual gas in a compact cyclotron, it is essential to look for some improvements to minimize their impact. The vacuum level in the acceleration chamber is directly related to the amount of neutral gas injected, depending on the input gas flow in the ion source and on the size of the chimney slit. The modifications of these parameters involve a linear variation of the pressure of the vacuum chamber under the consideration of a steady state molecular flow from the source. Thus, they can be adjusted to reduce the stripping losses in the \textsc{AMIT}\xspace cyclotron as well as to modify the injected current.
\begin{figure}[!b]
\centering
\subfigure[]
{\includegraphics [width=8.6cm] {Losses-GasLoad-H2.pdf} \label{fig:Losses-gasLoad-H2} } $\;$
\subfigure[]
{\includegraphics [width=8.6cm] {RelativeCurrent2-GasLoad.pdf} \label{fig:gasLoad-RelativeCurrent} }
\caption{(a) Stripping losses rate as a function of the gas load of the chimney. (b) Relative final beam current from stripping losses simulations considering the experimental results of the ion source extraction measurements. To facilitate the data analysis the red dashed line represents the normalized injected current over the maximum from the experimental measurements~\cite{calvo-2}.}
\label{fig:Losses-gasLoad}
\end{figure}
The stripping loss rate as a function of the gas load has been obtained (Fig.~\ref{fig:Losses-gasLoad-H2}), considering the actual flow range provided by the gas supply system of the ion source ($0-10$~sccm). The gas load is of great relevance in the fulfillment of the cyclotron beam requirements, because it influences directly the final beam current through two associated effects. On the one hand, the gas load influences on the injected beam. In accordance with the experimental results obtained from the measurements of the \textsc{AMIT}\xspace ion source in a test bench facility~\cite{calvo-2}, the injected beam current presents a non-linear behavior with the gas flow rate. The production and the neutralization of negative ions take place simultaneously by different reaction mechanisms inside the plasma source~\cite{kuo}. Then, a balance between the gas ionization rate inside the chimney and the survival of H$^-$ ions led to a maximum extracted beam current at $4$~sccm of gas load for different operating conditions of the ion source. In spite of the fact that these experimental results were obtained under DC extraction conditions, the behavior of the extracted current from the source must be proportional in RF conditions, and therefore, it is acceptable to use these measurements to quantify the expected current in the cyclotron. On the other hand, higher gas load implies higher stripping losses by the increment of the pressure level. The increase of the beam losses can even reduce the final beam current below the minimum needed for the radioisotope production. Therefore, the optimum solution will require a balance between both effects. The results of the optimum gas flow rate to maximize the injected current in combination with the limited beam stripping losses allow to establish a gas load of $3.5$~sccm to achieve a maximized final beam current (Fig.~\ref{fig:gasLoad-RelativeCurrent}), slightly lower than the optimum when only ion source current is considered.
In addition to the gas load, the stripping losses can be optimized by the adjustment of the chimney slit size, which modifies the conductance. From the designed nominal size of $6$~mm height and $2$~mm width, different slit areas have been considered from $0.6$~mm$^2$ to $2.4$~mm$^2$, modifying both the height (between $6$, $5$, $4$ and $3$~mm), and the width (between $0.2$, $0.3$ and $0.4$~mm). A linear increase of the pressure level in the vacuum chamber with the slit size has been considered. Multiparticle simulations have been performed to evaluate its effect on the beam stripping losses along the cyclotron, observing, as expected, a linear increase with the size of the slit (Fig.~\ref{fig:Losses-ChimneyArea}) both for changes in height or width.
\begin{figure}[!b]
\centering
\subfigure[]
{\includegraphics [width=8.6cm] {Slit_BeamStripping3D.pdf} \label{fig:Losses-ChimneyArea}}
\subfigure[]
{\includegraphics [width=8.6cm] {SlitArea_RelativeCurrent_Width-fit.pdf} \label{fig:RelativeCurrent-ChimneyArea}}
\caption{(a) Beam stripping losses rate as a function of the slit height and width. The color surface mesh is interpolated from the results of the simulations. (b) Normalized relative current as a function of the slit area. The values are normalized over the expected current from the nominal slit ($6\!\times\!0.2$~mm$^2$). The solid red line is a linear fit to the data weighted over the stochastic error associated with the stripping losses. The results are grouped by color and symbol according to the slit width, with each point representing a different slit height.}
\label{fig:Losses-Area}
\end{figure}
The slit size also has an impact on the injected current, the beam size and the associated losses due to axial deviations. First, considering a linear relationship between the injected current and slit area, a net increase on the final beam current will be expected as the area is modified in the considered range, in spite of the rise of beam stripping with the residual gas. Moreover, the wider the beam width, the greater the losses in the central region due to particle collisions with the puller. Fig.~\ref{fig:RelativeCurrent-ChimneyArea} represents the relative beam current at the end of the cyclotron in relation to the expected current from the nominal slit. The significant increase in the final current is clearly noticeable. Therefore, it can be concluded that a larger slit width than the nominal will be an advantageous enhancement on the final beam current. A deviation from the linear trend is observed due to the increase of losses in the central region due to impact with the puller. Lastly, a slight reduction in the height of the slit entails the complete axial beam transmission, reached for a slit height of $\leq\!5.5$~mm. Larger slit height leads to particle collisions with the top and bottom of the electrode in the medium-high energy range, that must be avoided to prevent the activation of the accelerator. Thus, the maximum slit height must be set below this value in order to avoid these beam losses.
Based on the results obtained, the analysis of the beam losses for different slit sizes allows to establish the range of dimensions that provide an optimization of the final current without an excessive rise of the stripping losses. It should be mentioned that optimizing the slit dimensions shall also consider the effect on the plasma meniscus and the extraction electric field modification. These effects will be analyzed in future studies.
The results obtained in this section will provide a starting point during the commissioning of the cyclotron to validate the simulations against experimental data, balancing all the related parameters to optimize the final beam current.
\section{Conclusions}
Beam stripping interactions with residual gas and electromagnetic fields (Lorentz stripping) have been studied in detail in order to analyze the effect on the cyclotron beam transmission through the implementation of these physical interactions into the beam dynamics code \textsc{OPAL}\xspace. These interactions constitute one of the main drawbacks in H$^-$ compact cyclotrons with internal ion source, due to their influence on the beam transmission, the high potential for damaging the cyclotron in-vessel components and the associated radioactive environment. The importance of the gas stripping has been corroborated by the application to the \textsc{AMIT}\xspace cyclotron. Thanks to this assessment and optimization of the ion source parameters and the vacuum system has been obtained. Firstly, the control of the gas load has been studied taking into account previous experimental measurements of the current extracted from the ion source. The balance of beam production at the ion source and the increase of stripping losses has been evaluated. Secondly, we assessed the effects of the ion source slit size and the injected beam current on the stripping losses and on the central region and the beam axial deviations. Bearing that in mind, stripping interactions have to be considered during the beam dynamics simulations of future compact H$^-$ cyclotrons, as one of the key effects for the optimization of the nominal parameters. The \textsc{OPAL}\xspace code has been upgraded and validated as a powerful tool for this use.
\section*{Acknowledgement}
The authors thank PSI and OPAL team for supporting this research. All simulations were performed using the HPC resources provided by the computing facilities of Extremadura Research Centre for Advanced Technologies (CETA-CIEMAT), funded by the European Regional Development Fund (ERDF). CETA-CIEMAT belongs to CIEMAT and the Government of Spain. This work was partially supported by the Spanish Ministry of Economy and Competitiveness under project FPA2016-78987-P.
|
\section{Introduction}
The increase of mobile data traffic and the internet's demand impose a high
spectral efficiency, low latency, and massive connectivity requirements
towards fifth-generation (5G) wireless networks and beyond. In this context,
direct conversion transceivers employing quadrature up/down conversion of
radio-frequency (RF) signals have attracted considerable attention owing to
their potential in reducing power consumption and cost. In addition, they
request neither external intermediate frequency filters nor image rejection
filters \cite{trans1} and \cite{trans2}. Nevertheless, such transceivers
suffer from inevitable RF front-end imperfections because of component
mismatches and manufacturing defects that degrade the performance of
wireless communication systems (WCS) \cite{imperf1}. Furthermore, the
amplitude and/or phase mismatch between the in-phase (I) and
quadrature-phase (Q) at the transmitter (Tx) or/and the receiver (Rx)
branches, so-called I/Q imbalance (IQI) and known also as zero intermediate
frequency, is one of the major performance-limiting impairments causing
self-interference and degrading the reliability, particularly in high-rate
WCS \cite{imperf2}. To this end, there have been several works modeling IQI,
studying its impact on system performance, and tackling the problem. For
instance, in \cite{work1}-\cite{workfin} and the citations referred therein,
the IQI influence on the system performance has been extensively assessed.
Moreover, in \cite{algo1}-\cite{algo2} and \cite{calib1}-\cite{calib3}, as
well as the references therein, various algorithms for IQI parameters'
estimation and calibration have been proposed, respectively. Even though,
such IQI mitigating mechanisms fail to completely eliminate hardware
impairments. As a consequence, residual hardware impairments (RHI), which
are added to the transmitted or/and the received signals, degrade further
the performance of WCS \cite{RHI1}-\cite{RHIfin}. On the other hand, several
fading models have been proposed over years to model either the fading or
the joint shadowing/fading phenomena. Therefore, dealing with a generalized
distribution modeling such behaviors is of paramount importance in providing
unified expressions for WCS performance criteria. Towards this end, the
Fox's H-function (FHF) random variable (RV), has been proposed as a unified
distribution subsuming most fading environment models. Precisely, the
probability density function (PDF) of various models can be expressed in
terms of a (i) single FHF (e.g., $\alpha$-$\mu$, Gamma-Gamma (GG),
Generalized Bessel-K, Fisher-Snedecor $\mathcal{F}$), (ii) finite sum of
FHFs (e.g., M\'alaga $\mathcal{M}$), and (iii) infinite summation of FHFs
(e.g., shadowed $\kappa$-$\mu$) \cite{AccessOurs}. Moreover, FHF has various
powerful properties, namely the (i) product, quotient, and power of FHF RVs
are FHF RV \cite{HFOX}, (ii) sum of FHF RVs can be tightly approximated by
an FHF \cite{WCL}-\cite{IRS}, and (iii) numerous complex transforms of FHFs,
e.g. Laplace and Mellin transforms, can be written as FHFs \cite{kilbas}. By
leveraging such properties, the performance analysis, and physical layer
security of intelligent reflecting surface aided cooperative WCS with
spatial diversity over FHF fading channels in presence of co-channel
interference (CCI), jamming signals, or/and eavesdroppers can be
straightforwardly investigated \cite{IRS}, \cite{Hwork3}.\newline
On the other hand, the instantaneous symbol error probability (SEP) is a
crucial performance metric quantifying the instantaneous communication
reliability. Importantly, such a metric is often provided in complicated
integral form depending on the employed modulation technique \cite{gqfunc}-%
\cite{talambura}. Therefore, the closed-form of the average SEP (ASEP)
remains a big challenge for numerous communication systems due to the
mathematical intractability of both the end-to-end (e2e) fading model and
the SEP expression. Obviously, obtaining accurate and simpler bounds or
approximate expressions for the ASEP is strongly depending on the SEP's
ones. To this end, it has been shown in \cite{yassineaccess} that the use of
the Trapezoidal rule technique leads to a simple and tight approximate
expression for the SEP's integral form. In fact, the mathematical
tractability of such an approximation lies in its exponential form, making
from the ASEP approximate expression a summation of Laplace transforms of
the SNR's PDF.
\subsection{Related work}
Recently, the impact of RF impairments on the performance of a WCS has been
the focus of various works. In \cite{weib}, assuming IQI at the receiver
only, the effects of the IQI on a generalized frequency division
multiplexing system under Weibull fading channels was investigated by
evaluating the ASER for $M$-ary quadrature amplitude modulation ($M$-QAM),
while in \cite{iqidqpsk}, the IQI effects on the bit error rate of Gray
coded differential quadrature phase-shift keying (GC-DQPSK) over Rayleigh
fading channel ware evaluated. The IQI impact on the outage probability of
both single- and multi- carrier systems over cascaded Nakagami-$m$ fading
channels have been quantified in \cite{multicarier}, whereas in \cite{gama},
a free-space optical system using subcarrier quadrature phase-shift keying
(QPSK) is assessed for Gamma-Gamma fading channels with IQI at the receiver.
On the other hand, the performance analysis of systems suffering from RHI
has attracted considerable attention \cite{RHI1}-\cite{RHIfin}.
Specifically, tight closed-form and asymptotic outage probability (OP)
expressions for an impaired RHI cognitive amplify-and-forward (AF)
multi-relay cooperative network undergoing Rayleigh fading model ware
derived in \cite{RHI1}. Likewise, OP and \textcolor{black}{CC} for an impaired
RHI WCS with CCI and imperfect channel state information (CSI) were
investigated in \cite{RHI2}. Very recently, OP of an impaired RHI hybrid
satellite-terrestrial relay network was analyzed in \cite{RHI3}, while the
performance of non-orthogonal multiple access-based AF relaying network
subject to RHI and Nakagami-$m$ fading model was investigated in \cite%
{RHIfin}.
\subsection{Motivation and contributions}
Capitalizing on the above, the RF impairments have a paramount impact on the
performance of wireless communication systems. Although the performance of
WCSs, under the assumption of perfect CSI at the receiver, is treated in
most of the existing works. Nonetheless, such a hypothesis is not practical
as the receiver requires often complex channel estimation algorithms
providing, in some cases, uncertain values, which would result in
performance degradation. To overcome this limitation, differential
modulation, among other solutions, can be employed particularly for
low-power wireless systems \cite{abouie}. In \cite{owc}, it is shown that
employing GC-DQPSK in the inter-satellite optical wireless system enhances
the performance compared to other modulation techniques. Furthermore, such a
scheme simplifies the detection as the channel estimation and tracking
becomes needless, which reduces the cost and complexity of the receiver.
To the best of the authors' knowledge, the joint effect of both IQI and RHI
on the performance of a WCS undergoing FHF fading channel has not been
addressed in the open technical literature. Motivated by this, in this
paper, we analyze the performance of an impaired IQI/RHI WCS experiencing
FHF fading model. Specifically, by deriving the cumulative distribution
function (CDF) expression of the signal-to-interference-plus-noise-ratio
(SINR), closed-form and asymptotic expressions for the OP,
\textcolor{black}{channel capacity (CC) under constant power with optimum rate
adaptation (ORA) policy}, and ASEP for different modulation schemes are
presented for the proposed impaired IQI/RHI WCS. Pointedly, the key
contributions of this paper can be summarized as follows:
\begin{itemize}
\item Expressions of the SINR per symbol at the receiver's input impaired by
both IQI and RHI are presented for different cases: (i) Tx Impaired by both
IQI and RHI, (ii) Rx Impaired by both IQI and RHI, and (iii) both Tx and Rx
are impaired.
\item Based on the statistical properties of the SINR, novel closed-form and
asymptotic expressions for the OP as well as for CC are derived.
\item A simple exponential approximation for the SEP for various coherent
and non-coherent used modulation techniques is proposed, based on which,
closed-form and asymptotic expressions for the ASEP are derived.
\item Finally, to gain further insights into the system's performance, the
derived expressions are evaluated for both $\alpha-\mu$ fading and M\'alaga $%
\mathcal{M}$ turbulence channels with pointing errors (TCPE).
\end{itemize}
\subsection{Organization}
The rest of this paper is structured as follows. Section II is dedicated to
the statistical properties of the considered WCS by retrieving the PDF and
CDF of SINR, while closed-form and asymptotic expressions for the OP, CC,
and ASEP are derived in Section III. All the analytical results are
retrieved for two particular cases of FHF RV, namely $\alpha-\mu$ and
M\'alaga $\mathcal{M}$ distributions in Section IV. In Section V, Numerical
results and insightful discussions are presented. Finally, some conclusions are drawn in Section VI.
\subsection{Notations}
For the sake of clarity, the notations used throughout the paper are
summarized in Table \ref{notations}.
\begin{table}[tbh]
\captionsetup{font=small}
\caption{Symbols and notations.}
\label{notations}\centering
\begin{tabular}{p{1.5cm}|p{6cm}|p{3.5cm}|p{5.5cm}}
\hline
\centering \textit{Symbol} & \centering \textit{Meaning} & \centering \textit{Symbol} & \qquad \qquad
\qquad \qquad \textit{Meaning} \\ \hline\hline
\centering$f_{X}(\cdot )$ & Probability density function of the random
variable $X$ & \centering$\gamma _{id}$ & SNR for ideal RF end-to-end \\
\hline
\centering$H_{p,q}^{m,n}(\cdot )$ & Univariate Fox's H-function & \centering$%
\gamma _{\varkappa }$ & SINR of impaired RF end-to-end \\ \hline
\centering$L^{\left( \ell \right) }$ & Complex contour ensuring the
convergence of the $\ell $th Mellin-Barnes integral & \centering$F_{X}(\cdot
)$ & Cumulative distribution function of a random variable $X$ \\ \hline
\centering$s$ & Ideal transmitter signal & \centering$P_{\text{out}}(\cdot )$
& Outage probability $\left( \text{OP}\right) $ \\ \hline
\centering$\varkappa $ & Transmitter (t) or receiver (r) & \centering$%
C^{\varkappa }$ & Channel capacity (CC) \\ \hline
\centering$s_{\varkappa }$ & Transmitted or received signal & \centering$%
H_{p_{1},q_{1}:p_{2},q_{2}:p_{3},q_{3}}^{m_{1},n_{1}:m_{2},n_{2}:m_{3},n_{3}}(\cdot )
$ & Bivariate Fox's H-function \\ \hline
\centering$G_{\varkappa }$ & IQI gain & \centering$H\left( \gamma \right) $
& Symbol error probability $\left( \text{SEP}\right) $ \\ \hline
\centering IRR$_{\varkappa }$ & Image rejection ratio & \centering$P_{s}$ &
Average symbol error probability $\left( \text{ASEP}\right) $ \\ \hline\hline
\end{tabular}%
\end{table}
\section{System and Signal model}
Consider a WCS, which consists of a single transmitter and receiver and
subject to IQI and RHI impairments. In our setup, both Tx and Rx are
equipped with a single antenna and the flat fading channel is modeled by a
generalized SFHF RV. The SNR's PDF can be expressed as
\begin{equation}
f_{\gamma }\left( \gamma \right) =\sum_{\ell =1}^{L}\psi _{\ell }{\small H}%
_{p{\small ,q}}^{m{\small ,n}}\left( \phi _{\ell }\gamma \left\vert
\begin{array}{c}
\left( a_{i}^{(\ell )},A_{i}^{(\ell )}\right) _{i=1:p} \\
\left( b_{i}^{(\ell )},B_{i}^{(\ell )}\right) _{i=1:q}%
\end{array}%
\right. \right) , \label{pdf}
\end{equation}%
where $\psi _{\ell }$ and $\phi _{\ell }$ are any two positive constants
satisfying $\int_{0}^{\infty }f_{\gamma }\left( \gamma \right) d\gamma =1$, $%
\text{i.e., }$%
\begin{equation}
\sum_{\ell =1}^{L}E_{\ell }\frac{\psi _{\ell }}{\phi _{\ell }}=1,
\label{Condition}
\end{equation}%
with
\begin{gather}
E_{\ell }=\frac{\prod\limits_{i=1}^{m}\Gamma \left( \mathcal{B}_{i}^{(\ell
)}\right) \prod\limits_{i=1}^{n}\Gamma \left( 1-\mathcal{A}_{i}^{(\ell
)}\right) }{\prod\limits_{i=n+1}^{p}\Gamma \left( \mathcal{A}_{i}^{(\ell
)}\right) \prod\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}_{i}^{(\ell
)}\right) },\ell =1..L, \label{Ell} \\
\mathcal{A}_{i}^{(\ell )}=a_{i}^{(\ell )}+A_{i}^{(\ell )}, \label{mathA} \\
\mathcal{B}_{i}^{(\ell )}=b_{i}^{(\ell )}+B_{i}^{(\ell )}, \label{mathB}
\end{gather}%
and
\begin{equation}
{\small H}_{p{\small ,q}}^{m{\small ,n}}\left( \phi _{\ell }x\left\vert
\begin{array}{c}
\left( a_{i}^{(\ell )},A_{i}^{(\ell )}\right) _{i=1:p} \\
\left( b_{i}^{(\ell )},B_{i}^{(\ell )}\right) _{i=1:q}%
\end{array}%
\right. \right) =\frac{1}{2\pi j}\int_{\mathcal{L}_{v}^{\left( _{\ell
}\right) }}\frac{\prod\limits_{i=1}^{m}\Gamma \left( b_{i}^{(\ell
)}+B_{i}^{(\ell )}v\right) \prod\limits_{i=1}^{n}\Gamma \left(
1-a_{i}^{(\ell )}-A_{i}^{(\ell )}v\right) }{\prod\limits_{i=n+1}^{p}\Gamma
\left( a_{i}^{(\ell )}+A_{i}^{(\ell )}v\right)
\prod\limits_{i=m+1}^{q}\Gamma \left( 1-b_{i}^{(\ell )}-B_{i}^{(\ell
)}v\right) }x^{-v}dv. \label{Hffox}
\end{equation}%
denotes the univariate FHF, $j=\sqrt{-1}$, $\mathcal{L}_{v}^{\left( _{\ell
}\right) }$ is an appropriately chosen complex contour, ensuring the
convergence of the $\ell $th above Mellin-Barnes integral, $\Gamma \left(
.\right) $ denotes the Euler Gamma function \cite{integraltable}, $%
a_{i}^{(\ell )}$ and $b_{i}^{(\ell )}$ are real-valued numbers, whereas $%
A_{i}^{(\ell )}$ and $B_{i}^{(\ell )}$ are two real positive numbers.
\subsection{RF impairments baseband model}
\subsubsection{IQI impairment model}
The time-domain baseband representation of the IQI-impaired signal, at
either the transmitter or the receiver, can be expressed as \cite{imperf1}
\begin{equation}
s_{\varkappa ,IQI}=G_{\varkappa ,1}s+G_{\varkappa ,2}s^{\ast },
\end{equation}%
where $\varkappa $ denotes $t$ or $r$ for the transmitter (Tx) or the
receiver (Rx) node, respectively, $s$ represents the ideal baseband signal
under perfect in-phase and quadrature-phase (I/Q) matching. Further, the IQI
parameters $G_{\varkappa ,1}$ and $G_{\varkappa ,2}$ are expressed,
respectively, as
\begin{equation}
G_{\varkappa ,1}=\frac{1+g_{\varkappa }e^{\varepsilon _{\varkappa }\varphi
_{\varkappa }}}{2}, \label{G1X}
\end{equation}%
\begin{equation}
G_{\varkappa ,2}=\frac{1-g_{\varkappa }e^{-\varepsilon _{\varkappa }\varphi
_{\varkappa }}}{2}, \label{G2X}
\end{equation}%
where $\varepsilon _{t}=+j,$ and $\varepsilon _{r}=-j$ with $\ j=\sqrt{-1}$,
while $g_{\varkappa }$ and $\varphi _{\varkappa }$ represent the gain and
phase mismatch for the node $\varkappa $, respectively. Particularly, for
perfect I/Q matching, $g_{\varkappa }=1$ and $\varphi _{\varkappa }=0$,
leading to $G_{\varkappa ,1}=1$ and $G_{\varkappa ,2}=0.$ On the other hand,
the corresponding image rejection ratio $\left( \text{IRR}\right) $, which
determines the amount of image frequency attenuation, is linked to the IQI
parameters as follow
\begin{equation}
IRR_{\varkappa }=\frac{\left\vert G_{\varkappa ,1}\right\vert ^{2}}{%
\left\vert G_{\varkappa ,2}\right\vert ^{2}}. \label{IRRdef}
\end{equation}%
Substituting (\ref{G1X}) and (\ref{G2X}) into (\ref{IRRdef}), yields
\begin{equation}
IRR_{\varkappa }=\frac{g_{\varkappa }^{2}+2g_{\varkappa }\cos \left( \varphi
_{\varkappa }\right) +1}{g_{\varkappa }^{2}-2g_{\varkappa }\cos \left(
\varphi _{\varkappa }\right) +1}. \label{IRRdef2}
\end{equation}%
For practical analog RF front-end electronics, the IRR's value is typically
in the range $\left[ 30\text{ dB},40\text{ dB}\right] $ indicating a
possible combination of 0.2 to 0.6 dB of gain mismatch and $1^{\circ}$ to $5^{\circ}$ of
phase imbalance \cite{Rfmicro}.
\subsubsection{RHI model}
In \cite{imperf1} and \cite{studer}, it has been shown that the complex
normal distribution fits accurately the residual distortion noise behavior,
with zero mean and average power proportional that of the signal of
interest, i.e.,
\begin{equation}
\eta_{\varkappa}\sim\mathcal{CN}\left(0,\kappa_{\varkappa}^{2}P_{\varkappa}%
\right),
\end{equation}
where $\kappa_{\varkappa}^{2}$ is a proportionality parameter describing the
residual additive impairments severity at node $\varkappa,$ while $P_{t}$
and $P_{r}$ refer to the transmit and receive power, respectively. Note that
$\kappa_{t}=\kappa_{r}=0$ is corresponding to ideal hardware on both sides
case.
\subsection{Received Signal Model}
Under the Assumption of single antennas at both Tx and Rx, we provide in
this subsection the expression of the received signal in various scenarios
of Tx and/or Rx impairments.
\subsubsection{Ideal scenario}
The received signal with ideal Tx and Rx RF front-end is represented as
\begin{equation}
r_{id}=h\mathrm{s}+\mathrm{n}, \label{idealreceived signal}
\end{equation}%
where $h$ and $n$ denote the channel coefficient and the additive complex
Gaussian receiver's noise, respectively. Moreover, in this case, the
instantaneous signal-to-noise ratio (SNR) per symbol at the receiver input
is expressed as
\begin{equation}
\gamma _{id}=\frac{P_{s}}{N_{0}}\left\vert h\right\vert ^{2},
\label{idealSNR}
\end{equation}%
where $P_{s}$ is the energy per transmitted symbol and $N_{0}$ represents
the single-sided power spectral density noise. In what follows, we will focus
on deriving the expression of the signal-to-interference-plus-noise-ratio
(SINR) in which Tx and/or Rx suffers from both IQI and RHI impairments.
\subsubsection{Tx impaired by both IQI and RHI}
In this scenario, it is assumed that Tx experiences IQI and RHI, while the
Rx RF front-end is ideal (i.e., without RHI nor IQI). Thus, the transmitted
signal can be represented as
\begin{eqnarray}
\mathrm{s}_{t,imp} &=&\mathrm{s}_{t,IQI}+\eta _{t},\notag \\
&=&G_{t,1}\mathrm{s}+G_{t,2}\mathrm{s}^{\ast }+\eta _{t}.
\end{eqnarray}
Consequently, the received signal can be expressed as
\begin{eqnarray}
y &=&h\mathrm{s}_{t,imp}+\mathrm{n},\notag \\
&=&hG_{t,1}\mathrm{s}+hG_{t,2}\mathrm{s}^{\ast }+h\eta _{t}+\mathrm{n}.
\end{eqnarray}
It follows that the instantaneous SINR can be evaluated using (\ref{idealSNR}%
), as
\begin{equation}
\gamma _{t}=\frac{\left\vert G_{t,1}\right\vert ^{2}}{\left\vert
G_{t,2}\right\vert ^{2}+\kappa _{t}^{2}+\frac{1}{\gamma _{id}}}.
\label{snrTxonly}
\end{equation}
\subsubsection{Rx impaired by both IQI and RHI}
In a similar manner, the Rx experiences here both IQI and RHI, while the Tx
RF front-end is supposed ideal. To this end, the received signal is given as
\begin{eqnarray}
y &=&G_{r,1}\left( h\mathrm{s}+\mathrm{n}\right) +G_{r,2}\left( h^{\ast }\mathrm{s}^{\ast }+\mathrm{n}^{\ast
}\right) +\eta _{r} \notag \\
&=&G_{r,1}h\mathrm{s}+G_{r,2}h^{\ast }\mathrm{s}^{\ast }+G_{r,1}\mathrm{n}+G_{r,2}\mathrm{n}^{\ast }+\mathrm{\eta} _{r}.
\label{yRXonly}
\end{eqnarray}
Thus, the corresponding instantaneous SINR can be simplified using (\ref%
{idealSNR}) to
\begin{equation}
\gamma _{r}=\frac{\left\vert G_{r,1}\right\vert ^{2}}{\left\vert
G_{r,2}\right\vert ^{2}+\kappa _{r}^{2}+\frac{\Lambda }{\gamma _{id}}},
\label{snrRxonly}
\end{equation}
with%
\begin{equation}
\Lambda =\left\vert G_{r,1}\right\vert ^{2}+\left\vert G_{r,2}\right\vert
^{2}.
\end{equation}
\subsubsection{Both Tx and Rx are impaired}
In this case, both Tx and Rx experience IQI and RHI. As a result, the
equivalent baseband received signal can be written as
\begin{eqnarray}
y &=&G_{r,1}\left[ h\left( G_{t,1}\mathrm{s}+G_{t,2}\mathrm{s}^{\ast }+\eta _{t}\right)+\mathrm{n}%
\right]+G_{r,2}\left[ h\left( G_{t,1}\mathrm{s}+G_{t,2}\mathrm{s}^{\ast }+\eta _{t}\right)+\mathrm{n}\right]
^{\ast }+\eta _{r}\notag \\
&=&\left[ \zeta _{11}h+\zeta _{22}h^{\ast }\right]\mathrm{s}+\left[ \zeta
_{12}h+\zeta _{21}h^{\ast }\right]\mathrm{s}^{\ast }+G_{r,1}h\eta _{t}+G_{r,2}h^{\ast }\eta _{t}^{\ast }+G_{r,1}\mathrm{n}+G_{r,2}\mathrm{n}^{\ast }+\eta _{r},
\label{yTXRX}
\end{eqnarray}
where $\zeta _{ij}$ are the elements of the following matrix%
\begin{equation}
\mathbf{\zeta }=\left(
\begin{array}{cc}
G_{r,1}G_{t,1} & G_{r,1}G_{t,2} \\
G_{r,2}G_{t,1}^{\ast } & G_{r,2}G_{t,2}^{\ast }%
\end{array}%
\right) . \label{matrix}
\end{equation}%
The corresponding instantaneous SINR is then expressed as
\begin{equation}
\gamma _{t/r}=\frac{\left\vert \zeta _{11}h+\zeta _{22}h^{\ast }\right\vert
^{2}P_{s}}{\left\vert \zeta _{12}h+\zeta _{21}h^{\ast }\right\vert
^{2}P_{s}+\Lambda \left\vert h\right\vert ^{2}\kappa _{t}^{2}P_{s}+\Lambda
\sigma ^{2}+\kappa _{r}^{2}P_{r}}. \label{snrTXRX}
\end{equation}%
On the other hand, one can check from $\left( \ref{IRRdef}\right) $
and $\left( \ref{matrix}\right) $ that
\begin{equation}
\left\vert \zeta _{11}\right\vert \left\vert \zeta _{22}\right\vert
=\left\vert \zeta _{12}\right\vert \left\vert \zeta _{21}\right\vert =\frac{%
\left\vert G_{r,1}\right\vert ^{2}\left\vert G_{t,1}\right\vert ^{2}}{\sqrt{%
IRR_{r}IRR_{t}}}.
\end{equation}%
As mentioned above, IRR$_{\varkappa }$ practically lies in the range $\left[
30\text{ dB}-40\text{ dB}\right] $ and the gains $\left\vert G_{\varkappa
,1}\right\vert \leq 1,$ it follows that $\left\vert \zeta _{11}\right\vert
\left\vert \zeta _{22}\right\vert $ as well as $\left\vert \zeta
_{12}\right\vert \left\vert \zeta _{21}\right\vert $ are relatively smaller$.
$ As a result, $\gamma _{t/r}$ can be accurately approximated by
\begin{equation}
\gamma _{t/r}\approx \frac{\left\vert \zeta _{11}\right\vert ^{2}+\left\vert
\zeta _{22}\right\vert ^{2}}{\left\vert \zeta _{12}\right\vert
^{2}+\left\vert \zeta _{21}\right\vert ^{2}+\Lambda \kappa _{t}^{2}+\kappa
_{r}^{2}+\frac{\Lambda }{\gamma _{id}}}.
\end{equation}%
Subsequently, using $\left( \ref{snrTxonly}\right)$, $\left( \ref{snrRxonly}\right) $, and
$\left( \ref{snrTXRX}\right) $, the instantaneous SINR can be expressed for
all cases in a unified form as
\begin{equation}
\gamma _{\varkappa }=\frac{\omega _{\varkappa }}{\varrho _{\varkappa }+\frac{%
\tau _{\varkappa }}{\gamma _{id}}}, \label{SINR}
\end{equation}%
where $\omega _{\varkappa }$, $\varrho _{\varkappa }$ and $\tau _{\varkappa }
$ are summarized in Table \ref{parametresIQI}.
\begin{table}[tbh]
\centering
\captionsetup{font=small}
\caption{$\protect\omega _{\varkappa }$, $\protect\varrho _{\varkappa }$ and
$\protect\tau _{\varkappa }$ per impairment.}$\centering%
\begin{tabular}{c|c|c|c}
\hline\hline
\backslashbox{$\varkappa$}{\textit{Parameters}} & $\omega _{\varkappa }$ & $%
\varrho _{\varkappa }$ & $\tau _{\varkappa }$ \\ \hline
$t$ & $\left\vert G_{t,1}\right\vert ^{2}$ & $\left\vert G_{t,2}\right\vert
^{2}+\kappa _{t}^{2}$ & $1$ \\ \hline
$r$ & $\left\vert G_{r,1}\right\vert ^{2}$ & $\left\vert G_{r,2}\right\vert
^{2}+\kappa _{r}^{2}$ & $\Lambda $ \\ \hline
$t/r$ & $\left\vert \zeta _{11}\right\vert ^{2}+\left\vert \zeta
_{22}\right\vert ^{2}$ & $\left\vert \zeta _{12}\right\vert ^{2}+\left\vert
\zeta _{21}\right\vert ^{2}+\kappa _{r}^{2}+\Lambda \kappa _{t}^{2}$ & $%
\Lambda $ \\ \hline\hline
\end{tabular}%
$%
\label{parametresIQI}
\end{table}
\begin{remark}
One can notice from (\ref{SINR}), that the SINR is upper bounded by $\frac{%
\omega _{\varkappa }}{\varrho _{\varkappa }}$. On the other hand, for ideal
RF front e2e, namely (i) ideal RHI $($i.e., $\kappa _{\varkappa }=0),$ and
(ii) ideal IQI $($i.e., $G_{\varkappa ,1}=1$, $G_{\varkappa ,2}=0)$, the
aforesaid parameters can be simplified to $\omega _{\varkappa }=1,$ $\varrho
_{\varkappa }=0,$ and $\tau _{\varkappa }$ $=1.$ \label{remark1}
\end{remark}
\section{Performance Analysis}
\subsection{Outage Probability}
\subsubsection{Exact analysis}
For a fixed impairment $\varkappa $, and SINR threshold $\gamma _{\text{th}%
}^{(\varkappa )}$, the OP can be defined as the probability the SINR fall
below $\gamma _{\text{th}}^{(\varkappa )}.$ Leveraging Remark \ref{remark1},
such a metric can be defined as
\begin{equation}
P_{\text{out}}^{\left( \varkappa \right) }=F_{\gamma _{\varkappa }}\left(
\gamma _{\text{th}}^{(\varkappa )}\right) ,\gamma _{\text{th}}^{(\varkappa
)}\leq \frac{\omega _{\varkappa }}{\varrho _{\varkappa }},
\label{Poutgeneral}
\end{equation}%
with
\begin{equation}
F_{\gamma _{\varkappa }}\left( \gamma \right) =F_{\gamma _{id}}\left( \frac{%
\tau _{\varkappa }\gamma }{\omega _{\varkappa }-\varrho _{\varkappa }\gamma }%
\right) ,
\end{equation}%
and $F_{\gamma _{id}}\left( .\right) $ denotes the CDF of ideal e2e SNR and
which can be evaluated relying to (\ref{pdf}) as
\begin{eqnarray}
F_{\gamma _{id}}\left( \gamma \right) &=&1-\int_{\gamma }^{\infty
}f_{\gamma _{id}}\left( t\right) dt \notag \\
&=&1-\sum_{\ell =1}^{L}\int_{\gamma }^{\infty }\psi _{\ell }{\small H}_{p%
{\small ,q}}^{m{\small ,n}}\left( \phi _{\ell }t\left\vert
\begin{array}{c}
\left( a_{i}^{\left( \ell \right) },A_{i}^{\left( \ell \right) }\right) _{i=%
{\small 1}:p} \\
\left( b_{i}^{\left( \ell \right) },B_{i}^{\left( \ell \right) }\right) _{i=%
{\small 1}:q}%
\end{array}%
\right. \right) dt \notag \\
&=&1-\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}{\small H}_{p+1%
{\small ,q+1}}^{m+1{\small ,n}}\left( \phi _{\ell }\gamma \left\vert
\begin{array}{c}
\Psi ^{\left( \ell \right) },\left( 1,1\right) \\
\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\right. \right) , \label{CDFideal}
\end{eqnarray}%
with
\begin{gather}
\Psi ^{\left( \ell \right) }=\left\{ \left( \mathcal{A}_{i}^{\left( \ell
\right) },A_{i}^{\left( \ell \right) }\right) _{i=1:n};\left( \mathcal{A}%
_{i}^{\left( \ell \right) },A_{i}^{\left( \ell \right) }\right)
_{i=n+1:p}\right\} , \\
\text{and} \notag \\
\Upsilon ^{\left( \ell \right) }=\left\{ \left( \mathcal{B}_{i}^{\left( \ell
\right) },B_{i}^{\left( \ell \right) }\right) _{i=1:m_{{}}};\left( \mathcal{B%
}_{i}^{\left( \ell \right) },B_{i}^{\left( \ell \right) }\right)
_{i=m+1:q_{{}}}\right\} ,
\end{gather}%
and $\mathcal{A}_{i}^{\left( \ell \right) }$, and $\mathcal{B}_{i}^{\left(
\ell \right) }$ are defined in (\ref{mathA}) and (\ref{mathB}),
respectively. Hence for $\gamma \geq 0$,
\begin{equation}
F_{\gamma _{\varkappa }}\left( \gamma \right) =1-\sum_{\ell =1}^{L}\frac{%
\psi _{\ell }}{\phi _{\ell }}{\small H}_{p+1{\small ,q+1}}^{m+1{\small ,n}%
}\left( \frac{\phi _{\ell }\tau _{\varkappa }\gamma }{\omega _{\varkappa
}-\varrho _{\varkappa }\gamma }\left\vert
\begin{array}{c}
\Psi ^{\left( \ell \right) },\left( 1,1\right) \\
\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\right. \right) . \label{CDFIQI}
\end{equation}
\subsubsection{Asymptotic analysis}
It is worthy to mention that the scale $\phi _{\ell }$ is usually inversely
proportional to the average SNR $\overline{\gamma }_{id}$ \cite[Tables II-V]%
{tables}$.$ Hence, at high SNR values, $\phi _{\ell }$ tends to $0$, and
then the FHF given in $\left( \ref{CDFIQI}\right) $, can be asymptotically
approximated in the high SNR regime as
\begin{equation}
{\small H}_{p+1{\small ,q+1}}^{m+1{\small ,n}}\left( \frac{\phi _{\ell }\tau
_{\varkappa }\gamma }{\omega _{\varkappa }-\varrho _{\varkappa }\gamma }%
\left\vert
\begin{array}{c}
\Psi ^{\left( \ell \right) },\left( 1,1\right) \\
\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\right. \right) \sim E_{\ell }-\sum_{i=1}^{m}\mathcal{F}_{\ell ,i}\left(
\frac{\phi _{\ell }\tau _{\varkappa }\gamma }{\omega _{\varkappa }-\varrho
_{\varkappa }\gamma }\right) ^{\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}}. \label{ApproFHF}
\end{equation}%
where $E_{\ell }$ is being defined in (\ref{Ell}) and
\begin{equation}
\mathcal{F}_{\ell ,i}=\frac{\prod\limits_{k=1,k\neq i}^{m}\Gamma \left(
\mathcal{B}_{k}^{\left( \ell \right) }-\mathcal{B}_{i}^{\left( \ell \right) }%
\frac{B_{k}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}\right)
\prod\limits_{k=1}^{n}\Gamma \left( 1-\mathcal{A}_{k}^{\left( \ell \right) }+%
\mathcal{B}_{i}^{\left( \ell \right) }\frac{A_{k}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}\right) }{\mathcal{B}_{i}^{\left( \ell \right)
}\prod\limits_{k=n+1}^{p}\Gamma \left( \mathcal{A}_{k}^{\left( \ell \right)
}-\mathcal{B}_{i}^{\left( \ell \right) }\frac{A_{k}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}\right) \prod\limits_{k=m+1}^{q}\Gamma \left( 1-%
\mathcal{B}_{k}^{\left( \ell \right) }+\mathcal{B}_{i}^{\left( \ell \right) }%
\frac{B_{k}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}\right) }.
\label{Fli}
\end{equation}%
, such that $\mathcal{B}_{k}^{\left( \ell \right) }-\mathcal{B}_{i}^{\left( \ell
\right) }\frac{B_{k}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}$
and $\mathcal{A}_{k}^{\left( \ell \right) }-\mathcal{B}_{i}^{\left( \ell
\right) }\frac{A_{k}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}$
are non-negative integers.
Now, by substituting $\left( \ref{ApproFHF}%
\right) $ into $\left( \ref{CDFIQI}\right) $ and using (\ref{Condition}),
the OP can be asymptotically approximated for high SNR values as
\begin{eqnarray}
P_{\text{out}}^{\left( \varkappa \right) } &\sim &1-\sum_{\ell =1}^{L}\frac{%
\psi _{\ell }}{\phi _{\ell }}\left[ E_{\ell }-\sum_{i=1}^{m}\mathcal{F}%
_{\ell ,i}\left( \frac{\phi _{\ell }\tau _{\varkappa }\gamma _{\text{th}%
}^{\left( \varkappa \right) }}{\omega _{\varkappa }-\varrho _{\varkappa
}\gamma _{\text{th}}^{\left( \varkappa \right) }}\right) ^{\frac{\mathcal{B}%
_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}}\right] , \notag
\\
&\sim &\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\sum_{i=1}^{m}%
\mathcal{F}_{\ell ,i}\left( \frac{\phi _{\ell }\tau _{\varkappa }\gamma _{%
\text{th}}^{\left( \varkappa \right) }}{\omega _{\varkappa }-\varrho
_{\varkappa }\gamma _{\text{th}}^{\left( \varkappa \right) }}\right) ^{\frac{%
\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}}.
\label{asymptopoutnonideal}
\end{eqnarray}
\begin{remark}
For ideal RF front e2e ( i.e., $\omega _{\varkappa }=\tau
_{\varkappa }=1$ and $\varrho _{\varkappa }=0)$, the asymptotic
expression of OP, for any arbitrary values of $\gamma _{\text{th}}$, can be
reexpressed as
\begin{equation}
P_{\text{out}}^{\left( id\right) }\sim \sum_{\ell =1}^{L}\frac{\psi _{\ell }%
}{\phi _{\ell }}\sum_{i=1}^{m}\mathcal{F}_{\ell ,i}\left( \phi _{\ell
}\gamma _{\text{th}}\right) ^{\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}}. \label{asymptopoutideal}
\end{equation}
\end{remark}
\subsection{Channel Capacity}
\subsubsection{Exact analysis}
The ergodic channel capacity, for a given impairment $\varkappa $ in bits/s,
under constant power with ORA policy over any composite fading or shadowing
channel is expressed as
\begin{equation}
\mathcal{C}^{\left( \varkappa \right) }=\frac{1}{\log (2)}\int_{0}^{\frac{%
\omega _{\varkappa }}{\varrho _{\varkappa }}}\log (1+\gamma )\frac{\partial
F_{\gamma _{\varkappa }}\left( \gamma \right) }{\partial \gamma }d\gamma .
\label{ORAexpression}
\end{equation}
\begin{proposition}
\label{ORAexact} The
closed-form expression of the CC under ORA policy can be
written as
\begin{IEEEeqnarray}{rCl}
\mathcal{C}^{\left( \varkappa \right) }=\frac{1}{\log \left( 2\right) }%
\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}{\small H}_{1,0:p%
{\small ,q+1:1,2}}^{0,1:m+1{\small ,n:1,1}}\left( \frac{\phi _{\ell }\tau
_{\varkappa }}{\varrho _{\varkappa }},\frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}\left\vert
\begin{array}{c}
\left( {\small 1};{\small 1},{\small 1}\right) \\
-%
\end{array}%
\right\vert
\begin{array}{c}
\Psi ^{\left( \ell \right) } \\
\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1,1\right) ;- \\
\left( 1,1\right) ;\left( 0,1\right)%
\end{array}%
\right. \right). \label{ORAcapa}
\end{IEEEeqnarray}
where ${\small H}_{p_{1},q_{1}:p_{2},q_{2}:p_{3},q_{3}}^{m_{1}{\small ,n}_{1}%
{\small :m}_{2}{\small ,n}_{2}{\small :m}_{3},n_{3}}\left( .\right) $
denotes the Bivariate FHF \cite[Eq. (2.57)]{mathai}.
\end{proposition}
\begin{proof}
The proof is provided in Appendix A.
\end{proof}
\begin{corollary}
\textcolor{black}{For ideal RF e2e, the expression of the CC under
ORA policy can be reduced to}
\begin{equation}
\mathcal{C}^{\left( id\right) }=\frac{1}{\log \left( 2\right) }\sum_{\ell
=1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}{\small H}_{p+2{\small ,q+2}}^{m+2%
{\small ,n+1}}\left( \phi _{\ell }\left\vert
\begin{array}{c}
\left( 0,1\right) ,\Psi ^{\left( \ell \right) },\left( 1,1\right) \\
\left( 0,1\right) ,\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\right. \right). \label{idealcapa}
\end{equation}%
\label{crolaryidealcapa}
\end{corollary}
\begin{proof}
The proof is provided in Appendix B.
\end{proof}
\subsubsection{Asymptotic analysis}
By plugging $\left( \ref{ApproFHF}\right) $ in $\left( \ref{EQORAappedix}%
\right) $, the asymptotic expression of the \textcolor{black}{CC} in high SNR
regime can be written as
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &\sim &\log _{2}\left( 1+\frac{%
\omega _{\varkappa }}{\varrho _{\varkappa }}\right) \sum_{\ell =1}^{L}\frac{%
\psi _{\ell }}{\phi _{\ell }}E_{\ell }-\sum_{\ell =1}^{L}\frac{\psi _{\ell }%
}{\log \left( 2\right) \phi _{\ell }}\sum_{i=1}^{m}\mathcal{F}_{\ell ,i}\int_{0}^{\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}}\frac{1%
}{1+\gamma }\left( \frac{\phi _{\ell }\tau _{\varkappa }\gamma }{\omega
_{\varkappa }-\varrho _{\varkappa }\gamma }\right) ^{\frac{\mathcal{B}%
_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}}d\gamma , \notag
\\
&\overset{(a)}{\sim }&\log _{2}\left( 1+\frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}\right) -\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\log \left(
2\right) \phi _{\ell }}\sum_{i=1}^{m}\mathcal{F}_{\ell ,i}\int_{0}^{\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}}\frac{1%
}{1+\gamma }\left( \frac{\phi _{\ell }\tau _{\varkappa }\gamma }{\omega
_{\varkappa }-\varrho _{\varkappa }\gamma }\right) ^{\frac{\mathcal{B}%
_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}}d\gamma .
\end{eqnarray}%
where step $\left( a\right) $ holds using $\left( \ref{Condition}\right) $.
By using \cite[Eqs. (2.9.5),(2.9.15)]{kilbas} and \cite[3.194 Eq. (1)]%
{integraltable}, one obtains
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &\sim &\log _{2}\left( 1+\frac{%
\omega _{\varkappa }}{\varrho _{\varkappa }}\right) -\frac{\omega
_{\varkappa }}{\log (2)\varrho _{\varkappa }}\sum_{\ell =1}^{L}\frac{\psi
_{\ell }}{\phi _{\ell }}\sum_{i=1}^{m}\frac{\mathcal{F}_{\ell ,i}}{\Gamma
\left( \frac{\mathcal{B}_{i}}{Bi}\right) }\left( \frac{\phi _{\ell }\tau
_{\varkappa }}{\varrho _{\varkappa }}\right) ^{\frac{\mathcal{B}_{i}^{\left(
\ell \right) }}{B_{i}^{\left( \ell \right) }}}\left( \frac{1}{2\pi j}\right)
^{2}\int_{\mathcal{L}_{s}^{\left( \ell \right) }}\int_{\mathcal{L}%
_{v}^{\left( \ell \right) }}\Gamma \left( s\right) \notag \\
&&\times \frac{\Gamma \left( 1-s\right) \Gamma \left( v\right) \Gamma \left(
\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}%
-v\right) \Gamma \left( 1+\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}-s-v\right) }{\Gamma \left( 2+\frac{\mathcal{B}%
_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}-s-v\right) }%
\left( \frac{\omega _{\varkappa }}{\varrho _{\varkappa }}\right) ^{-s}\left(
-1\right) ^{-v}dsdv. \label{ZZ}
\end{eqnarray}%
which can be equivalently expressed using the bivariate FHF as
\begin{IEEEeqnarray}{rCl}
\mathcal{C}^{\left( \varkappa \right) } &\sim &\log _{2}\left( 1+\frac{%
\omega _{\varkappa }}{\varrho _{\varkappa }}\right) -\frac{\omega
_{\varkappa }}{\log (2)\varrho _{\varkappa }}\sum_{\ell =1}^{L}\frac{\psi
_{\ell }}{\phi _{\ell }}\sum_{i=1}^{m}\frac{\mathcal{F}_{\ell ,i}}{\Gamma
\left( \frac{\mathcal{B}_{i}}{Bi}\right) }\left( \frac{\phi _{\ell }\tau
_{\varkappa }}{\varrho _{\varkappa }}\right) ^{\substack{ \frac{\mathcal{B}%
_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }} \\ }} \notag \\
&&\times {\small H}_{1,1:1{\small ,1:1,1}}^{0,1:1{\small ,1:1,1}}\left(
\frac{\omega _{\varkappa }}{\varrho _{\varkappa }},-1\left\vert
\begin{array}{c}
\left( {\small -}\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left(
\ell \right) }};{\small 1},{\small 1}\right) ;- \\
-;\left( -1{\small -}\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }};{\small 1},{\small 1}\right)
\end{array}%
\right\vert
\begin{array}{c}
\left( 0,1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1-\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell
\right) }},1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\right. \right) . \label{asymptoORA}
\end{IEEEeqnarray}
\begin{remark}\label{remarkCapaceiling}
\begin{itemize}\item[]
\item As $\phi _{\ell }$ is inversely proportional to $\overline{\gamma }%
_{id}$ and based on $\left( \ref{Condition}\right) $,
\textcolor{black}{it
follows} that $\frac{\psi _{\ell }}{\phi _{\ell }}\neq \infty .$ Thus, it
can be clearly seen from $\left( \ref{asymptoORA}\right)$ that
\begin{equation}
\lim_{\overline{\gamma }_{id}\rightarrow \infty }\mathcal{C}^{\left(
\varkappa \right) }\mathcal{=C}^{\left( \varkappa ,\infty \right) }=\log
_{2}\left( 1+\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}\right) ,
\label{capaceiling}
\end{equation}%
which demonstrates that the capacity channel for non-ideal RF front end
(i.e., $\varrho _{\varkappa }\neq 0$) has a ceiling, depending exclusively
on the impairments parameters and is irrespective of the fading severity
parameters, that can't be crossed by increasing the SNR.
\item The asymptotic expression of the CC under ideal case can be obtained
straightforwardly using \cite[Eq. (1.8.2)]{kilbas} \textcolor{black}{along}
with (\ref{idealcapa}).
\end{itemize}
\end{remark}
\subsection{Average Symbol Error Probability}
In this section, a tight approximate expression for the ASEP of various
modulation schemes is derived. We start first by obtaining a simple
exponential-based approximate expression for the SEP using the Trapezoidal
integration rule \cite{yassineaccess}.
Let $\mathcal{H}\left( \gamma \right)$ denote the SEP for a given modulation
scheme. Table \ref{SEPmodulations} outlines the SEP's integral-based
expression for different modulation techniques.
\begin{table}[tbh]
\captionsetup{font=small}
\caption{Integral form of the SEP for numerous modulation schemes.}
\label{SEPmodulations}\centering%
\begin{tabular}{c|c}
\hline
\textit{Modulation scheme} & $\mathcal{H}\left( \gamma \right) $ \\
\hline\hline
$M$-PSK & $\mathcal{D}\left( 1,\frac{M-1}{M},\sin ^{2}\left( \frac{\pi }{M}%
\right) ,\gamma \right) $ \\ \hline
$M$-QAM & $\mathcal{D}\left( \frac{4\left( \sqrt{M}-1\right) }{\sqrt{M}},%
\frac{1}{2},\frac{3}{2\left( M-1\right) },\gamma \right) -\mathcal{D}\left(
4\left( \frac{\sqrt{M}-1}{\sqrt{M}}\right) ^{2},\frac{1}{4},\frac{3}{2\left(
M-1\right) },\gamma \right) $ \\ \hline
$M$-DPSK & $\mathcal{E}\left( 1,\frac{M-1}{M},\sin ^{2}\left( \frac{\pi }{M}%
\right) ,\cos \left( \frac{\pi }{M}\right) ,\gamma \right) $ \\ \hline
GC-DQPSK & $\mathcal{E}\left( \frac{1}{2},1,1,-\frac{1}{\sqrt{2}},\gamma
\right) $ \\ \hline\hline
\end{tabular}%
\end{table}
\par
with
\begin{equation}
\mathcal{D}\left( a,b,c,\gamma \right) =\frac{a}{\pi }\int_{0}^{b\pi }\exp
\left( -\frac{c\gamma }{\sin ^{2}\left( t\right) }\right) dt, \label{Dfunc}
\end{equation}%
and
\begin{equation}
\mathcal{E}\left( a,b,c,d,\gamma \right) =\frac{a}{\pi }\int_{0}^{b\pi }\exp
\left( -\frac{c\gamma }{1+d\cos \left( t\right) }\right) dt. \label{Efunc}
\end{equation}%
\newline
It follows that the ASEP can be calculated by averaging $\mathcal{H}\left(
\gamma \right) $ over the statistics of the involved fading channel as
\begin{equation}
P_{s}^{\left( \varkappa \right) }=\int_{0}^{\infty }f_{\gamma _{\varkappa
}}\left( \gamma \right) \mathcal{H}\left( \gamma \right) d\gamma .
\label{generalASEP}
\end{equation}%
In the sequel, \textcolor{black}{a} tight approximate expression for the SEP
per each considered modulation scheme is obtained, relying on Table \ref%
{SEPmodulations} alongside the Trapezoidal rule. Leveraging these results,
accurate approximate and asymptotic expressions for ASEP are provided using $%
\left( \ref{pdf}\right) $ jointly with $\left( \ref{generalASEP}\right) $.
\begin{proposition}
The SEP for the considered modulation schemes can be tightly approximated by
\begin{equation}
\mathcal{H}\left( \gamma \right) \simeq \sum_{n=0}^{N}\theta _{n}\exp \left(
-\delta _{n}\gamma \right) , \label{SEPbin}
\end{equation}%
where $\theta _{n}$ and $\delta _{n}$ are summarized in Tables \ref%
{diracandteta1} and \ref{diracandteta2} depending on the %
\textcolor{black}{employed} modulation scheme. Note that an even positive
number $N$ is required to evaluate such coefficients for $M$-QAM modulation
technique.
\begin{table}[tbh]
\captionsetup{font=small,width=0.8\columnwidth}
\caption{The coefficients $\protect\theta _{n}$ and $\protect%
\delta _{n}$ for $M$-PSK, $M$-DPSK and GC-DQPSK modulations.}
\label{diracandteta1}\renewcommand{\arraystretch}{1} \centering%
\begin{tabular}[b]{c|c|c|c}
\hline
\textit{Modulation} & \multicolumn{2}{c|}{$\theta _{n}$} & %
\multirow{2}{*}{\textcolor{black}{$\delta _{n}\ (0\leq n\leq N)$}} \\ \cline{2-3}
\textit{scheme} & $n=0$, $N$ & $1\leq n\leq N-1$ & \\ \hline\hline
$M$-PSK & $\frac{M-1}{2NM}$ & $\frac{M-1}{NM}$ & $\frac{\sin ^{2}\left(
\frac{\pi }{M}\right) }{\sin ^{2}\left( \frac{n\left( M-1\right) \pi }{NM}%
\right) }$ \\ \hline
$M$-DPSK & $\frac{M-1}{2NM}$ & $\frac{M-1}{NM}$ & $\frac{\sin ^{2}\left(
\frac{\pi }{M}\right) }{1+\cos \left( \frac{\pi }{M}\right) \cos \left(
\frac{n\left( M-1\right) \pi }{NM}\right) }$ \\ \hline
GC-DQPSK & $\frac{1}{4N}$ & $\frac{1}{2N}$ & $\frac{1}{1-\frac{1}{\sqrt{2}}%
\cos \left( \frac{n\pi }{N}\right) }$ \\ \hline\hline
\end{tabular}%
\end{table}
\begin{table}[tbh]
\captionsetup{font=small}
\caption{The coefficients $\protect\theta _{n}$ and $\protect%
\delta _{n}$ for $M$-QAM modulation.}
\label{diracandteta2}\renewcommand{\arraystretch}{1} \centering%
\begin{tabular}[b]{c|c|c|c|c|c|c}
\hline
\textit{Modulation} & \multicolumn{5}{c|}{$\theta _{n}$} & %
\multirow{2}{*}{\textcolor{black}{$\delta _{n}\ (0\leq n\leq N)$}} \\ \cline{2-6}
\textit{scheme} & $n=0$ & $1\leq n\leq \frac{N}{2}-1$ & $n=\frac{N}{2}$ & $%
\frac{N}{2}+1\leq n\leq N-1$ & $n=N$ & \\ \hline\hline
$M$-QAM & $\frac{\sqrt{M}-1}{NM}$ & $2\frac{\sqrt{M}-1}{NM}$ & $\frac{M-1}{NM%
}$ & $2\frac{\sqrt{M}-1}{N\sqrt{M}}$ & $2\frac{\sqrt{M}-1}{N\sqrt{M}}$ & $%
\frac{3}{2\left( M-1\right) \sin ^{2}\left( \frac{n\pi }{2N}\right) }$ \\
\hline\hline
\end{tabular}%
\end{table}
\end{proposition}
\begin{proof}
Importantly, for a given positive number $N$, a real-valued $\varphi $, and
an arbitrary function $f$, the following definite integral is known to be
accurately approximated using the Trapezoidal rule as
\begin{equation}
\int_{0}^{N\varphi }f\left( t\right) dt=\frac{\varphi }{2}\left[
g_{0}+g_{N}+2\sum\limits_{n=1}^{N-1}g_{n}\right] . \label{TrapzEq}
\end{equation}%
where $g_{0}=f\left( 0\right) $ and $g_{n}=f\left( n\varphi \right) .$ By
setting $\varphi =\frac{b\pi }{N}$ and \textcolor{black}{the function $f$ in} $%
\left( \ref{TrapzEq}\right) $ to either $\mathcal{D}$ and $\mathcal{E}$
defined in $\left( \ref{Dfunc}\right) $ or $\left( \ref{Efunc}\right) $,
respectively, along with Table \ref{SEPmodulations}'s coefficients, and
performing some algebraic operations, both Tables $\ref{diracandteta1},$ and
$\ref{diracandteta2},$ can be easily obtained, which concludes the proof.
\end{proof}
\begin{remark}
Of note, the greater $N$ is, the accurate the approximation is. However, the
computational complexity becomes higher with the increase of such a number.
To this end, it is necessary to look for an accuracy-complexity tradeoff.
Numerically, we checked that above $N\geq 5$, the relative error becomes
negligible \cite{yassineaccess}. Owing to this fact, $N$ is set to $5$ for
the modulation techniques summarized in Table $\ref{diracandteta1}$ and to $%
6 $ (i.e. an even number above $5$) for $M$-QAM.
\end{remark}
\subsubsection{Exact Analysis}
\begin{proposition}
\label{propASEPexact}
The ASEP for the four considered modulation techniques can be tightly
approximated as
\begin{IEEEeqnarray}{rCl}
P_{s}^{\left( \varkappa \right) } &\simeq &\sum_{n=0}^{N}\theta _{n}\left[
1-\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}{\small H}_{1,0:p%
{\small ,q+1:0,2}}^{0,1:m+1{\small ,n:1,0}}\left( \frac{\phi _{\ell }\tau
_{\varkappa }}{\varrho _{\varkappa }},\frac{\delta _{n}\omega _{\varkappa }}{%
\varrho _{\varkappa }}\left\vert
\begin{array}{c}
\left( {\small 1};{\small 1},{\small 1}\right) \\
-%
\end{array}%
\right\vert
\begin{array}{c}
\Psi ^{\left( \ell \right) } \\
\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\left\vert
\begin{array}{c}
- \\
\left( 1,1\right) ;\left( 0,1\right)
\end{array}%
\right. \right) \right] . \notag \\
&& \label{ASEPMARY}
\end{IEEEeqnarray}
\end{proposition}
\begin{proof}
The proof is provided in Appendix C.
\end{proof}
\begin{corollary}
\label{corolidealasep}
For \textcolor{black}{ideal RF e2e}, the above approximate ASEP's expression
can be simplified in terms of a univariate FHF instead as
\begin{equation}
P_{s}^{\left( \varkappa \right) }\simeq \sum_{n=1}^{N}\theta _{n}-\sum_{\ell
=1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\sum_{n=1}^{N}\theta _{n}{\small H}%
_{p+2{\small ,q+1}}^{m+1{\small ,n+1}}\left( \frac{\phi _{\ell }}{\delta _{n}%
}\left\vert
\begin{array}{c}
\left( 0,1\right) ,\Psi ^{\left( \ell \right) },\left( 1,1\right) \\
\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\right. \right) . \label{idealASEP}
\end{equation}
\end{corollary}
\begin{proof}
The proof is provided in Appendix D.
\end{proof}
\subsubsection{Asymptotic analysis}
\begin{proposition}
\label{propasympAsep}
the asymptotic expression of the ASEP can be expressed as
\begin{IEEEeqnarray}{rCl}
P_{s}^{\left( \varkappa \right) } &\sim &\sum_{n=1}^{N}\theta _{n}\exp
\left( -\frac{\delta _{n}\omega _{\varkappa }}{\varrho _{\varkappa }}\right)
+\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}\sum_{\ell =1}^{L}\frac{%
\psi _{\ell }}{\phi _{\ell }}\sum_{n=1}^{N}\theta _{n}\delta
_{n}\sum_{i=1}^{m}\frac{\mathcal{F}_{\ell ,i}\left( \frac{\phi _{\ell }\tau
_{\varkappa }}{\omega _{\varkappa }}\right) }{\Gamma \left( \frac{\mathcal{B}%
_{i}}{B_{i}}\right) }^{\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}} \notag \\
&&\times {\small H}_{1,1:0{\small ,1:1,1}}^{0,1:1{\small ,0:1,1}}\left(
\frac{\delta _{n}\omega _{\varkappa }}{\varrho _{\varkappa }},-1\left\vert
\begin{array}{c}
\left( {\small -}\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left(
\ell \right) }};{\small 1},{\small 1}\right) \\
-;\left( -1{\small -}\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }};{\small 1},{\small 1}\right)
\end{array}%
\right\vert
\begin{array}{c}
- \\
\left( 0,1\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1-\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell
\right) }},1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\right. \right) . \label{AsymptoASEP}
\end{IEEEeqnarray}
\end{proposition}
\begin{proof}
the proof is provided in appendix E.
\end{proof}
\begin{remark}
\begin{itemize}
\item[]
\item Similarly to (\ref{capaceiling}), for \textcolor{black}{non-ideal RF e2e}%
, the ASEP admits the following ceiling that can't be crossed by the
increase of SNR as%
\begin{equation}
\lim_{_{\overline{\gamma }_{id}\rightarrow \infty }}P_{s}^{\left( \varkappa
\right) }=P_{s}^{\left( \varkappa ,\infty \right) }=\sum_{n=0}^{N}\theta
_{n}\exp \left( -\frac{\delta _{n}\omega _{\varkappa }}{\varrho _{\varkappa }%
}\right) . \label{valofASEPatinfinity}
\end{equation}%
Moreover, it can be seen that this ceiling is irrespective of the fading
severity parameters, whereas it depends on both $M$ (as $\delta _{n}$ and $%
\theta _{n}$ are expressed in terms of the modulation %
\textcolor{black}{parameter} $M$) and the impairments parameters.
Particularly, when $\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}$
approaches infinity, $P_{s}^{\left( \varkappa \right) }$ tends to $0$. On
the other hand, it can be easily shown that $\theta _{n}$ and $\delta _{n}$
are increasing and decreasing functions with $M$, respectively. As a result,
$\theta _{n}\exp \left( -\frac{\delta _{n}\omega _{\varkappa }}{\varrho
_{\varkappa }}\right) $ is an increasing function \textcolor{black}{with $M$}.
Thus the smaller the value of $M$ is, the smaller the ASEP is, leading to
the system performance enhancement.
\item For ideal case, the ASEP's asymptotic expression can be easily
obtained directly from (\ref{idealASEP}) with the help of \cite[Eq. (1.8.2)]%
{kilbas}.
\end{itemize}
\end{remark}
\section{Applications}
It is worthwhile to note that the aforementioned derived expressions for
various metrics can be simplified significantly depending on the fading amplitude distribution simplicity. As mentioned before, \textcolor{black}{numerous} well-known PDF of recent fading distributions can be expressed in terms of FHF. In this
section, two generalized fading models are considered:
\subsection{$\protect\alpha-\protect\mu$ fading}
Let's $\alpha $ and $\mu $ denote two real numbers that reflect the
non-linearity of the propagation medium and the clustering of the multipath
waves, \textcolor{black}{respectively}. The SNR's PDF of $\alpha -\mu $ fading
model can be obtained by setting $L=1$, $\psi _{1}=$ $\frac{\mu _{{}}^{\frac{%
2}{\alpha }}}{\overline{\gamma }\Gamma \left( \mu \right) }$, $\phi _{1}=%
\frac{\mu ^{\frac{2}{\alpha }}}{\overline{\gamma }}$, $m=q=1,n=p=0$, $%
b_{1}^{\left( 1\right) }=$ $\mu -\frac{2}{\alpha }$, and $B_{1}^{\left(
1\right) }=\frac{2}{\alpha }$ in (\ref{pdf}). Therefore, the closed-form and
asymptotic expressions for the three considered metrics are summarized in
Tables \ref{closedalpha} and \ref{asympalpha}, respectively, where $\Psi
\left( \cdot \right) $ denotes the Digamma function which is defined as the
logarithmic derivative of the Gamma function \cite{digam} and
\begin{equation}
\Theta _{\varkappa }=\frac{\omega _{\varkappa }\mu ^{\mu }}{\varrho
_{\varkappa }\Gamma \left( \frac{\alpha \mu }{2}\right) \Gamma \left( 1+\mu
\right) }\left( \frac{\tau _{\varkappa }}{\varrho _{\varkappa }\overline{%
\gamma }_{id}}\right) ^{\frac{\alpha \mu }{2}}.
\end{equation}
\begin{table}[tbh]
\captionsetup{font=small}
\caption{Closed-form expressions for $\protect\alpha -\protect\mu $ fading
model.}\renewcommand{\arraystretch}{1.2} \centering%
\begin{tabular}{c|c|c}
\hline
\parbox[t]{2mm}{\multirow{5}{*}{\rotatebox[origin=c]{90}{\small{%
\textit{Ideal RF e2e}}}}} & OP & $1-\frac{1}{\Gamma \left( \mu \right) }%
H_{1,2}^{2,0}\left( \frac{\mu ^{\frac{2}{\alpha }}}{\overline{\gamma }_{id}}%
\gamma _{\text{th}}\left\vert
\begin{array}{c}
-;\left( 1,1\right) \\
\left( 0,1\right) ,\left( \mu ,\frac{2}{\alpha }\right) ;-%
\end{array}%
\right. \right) $ \\ \cline{2-3}
& CC & $\frac{1}{\log (2)\Gamma \left( \mu \right) }H_{2,3}^{3,1}\left(
\frac{\mu ^{\frac{2}{\alpha }}}{\overline{\gamma }_{id}}\left\vert
\begin{array}{c}
\left( 0,1\right) ;\left( 1,1\right) \\
\left( 0,1\right) ,\left( 0,1\right) ,\left( \mu ,\frac{2}{\alpha }\right) ;-%
\end{array}%
\right. \right) $ \\ \cline{2-3}
& ASEP & $\sum_{n=0}^{N}\theta _{n}\left[ 1-\frac{1}{\Gamma \left( \mu
\right) }H_{2,2}^{2,1}\left( \frac{\mu ^{\frac{2}{\alpha }}}{\delta _{n}%
\overline{\gamma }_{id}}\left\vert
\begin{array}{c}
\left( 0,1\right) ;\left( 1,1\right) \\
\left( 0,1\right) ,\left( \mu ,\frac{2}{\alpha }\right) ;-%
\end{array}%
\right. \right) \right] $ \\ \hline\hline
\parbox[t]{2mm}{\centering
\multirow{5}{*}{\rotatebox[origin=c]{90}{\small{\textit{Non-ideal RF e2e}}}}}
& OP & $1-\frac{1}{\Gamma \left( \mu \right) }H_{1,2}^{2,0}\left( \frac{\mu
^{\frac{2}{\alpha }}}{\overline{\gamma }_{id}}\frac{\tau _{\varkappa }\gamma
_{\text{th}}}{\omega _{\varkappa }-\varrho _{\varkappa }\gamma _{\text{th}}}%
\left\vert
\begin{array}{c}
-;\left( 1,1\right) \\
\left( 0,1\right) ,\left( \mu ,\frac{2}{\alpha }\right) ;-%
\end{array}%
\right. \right) $ \\ \cline{2-3}
& CC & $\frac{1}{\log (2)\Gamma \left( \mu \right) }%
H_{1,0:0,2:1,2}^{0,1:2,0:1,1}\left( \frac{\mu ^{\frac{2}{\alpha }}\tau
_{\varkappa }}{\overline{\gamma }_{id}\varrho _{\varkappa }},\frac{\omega
_{\varkappa }}{\varrho _{\varkappa }}\left\vert
\begin{array}{c}
\left( 1;1,1\right) \\
-%
\end{array}%
\right\vert
\begin{array}{c}
- \\
\left( 0,1\right) ,\left( \mu ,\frac{2}{\alpha }\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1,1\right) ;- \\
\left( 1,1\right) ;\left( 0,1\right)
\end{array}%
\right. \right) $ \\ \cline{2-3}
& ASEP & $\sum_{n=0}^{N}\theta _{n}\left[ 1-\frac{1}{\Gamma \left( \mu
\right) }H_{1,0:0,2:0,2}^{0,1:2,0:1,0}\left( \frac{\mu ^{\frac{2}{\alpha }%
}\tau _{\varkappa }}{\overline{\gamma }_{id}\varrho _{\varkappa }},\frac{%
\delta _{n}\omega _{\varkappa }}{\varrho _{\varkappa }}\left\vert
\begin{array}{c}
\left( 1;1,1\right) \\
-%
\end{array}%
\right\vert
\begin{array}{c}
- \\
\left( 0,1\right) ,\left( \mu ,\frac{2}{\alpha }\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
- \\
\left( 1,1\right) ;\left( 0,1\right)
\end{array}%
\right. \right) \right] $ \\ \hline
\end{tabular}%
\label{closedalpha}
\end{table}
\begin{table}[tbh]
\captionsetup{font=small}
\caption{Asymptotic expressions for $\protect\alpha -\protect\mu $ fading
model.}
\label{asympalpha}\centering\renewcommand{\arraystretch}{2}%
\begin{tabular}{c|c|c}
\hline
\parbox[t]{2mm}{\multirow{3}{*}{\rotatebox[origin=c]{90}{\small{%
\textit{Ideal RF e2e}}}}} & OP & $\frac{\mu ^{\mu -1}}{\Gamma \left( \mu
\right) }\left( \frac{\gamma _{\text{th}}}{\overline{\gamma }_{id}}\right) ^{%
\frac{\alpha \mu }{2}} $ \\ \cline{2-3}
& CC & $\frac{1}{\log (2)}\left[ \frac{2}{\alpha }\Psi \left( \mu \right)
-\log \left( \frac{\mu ^{\frac{2}{\alpha }}}{\overline{\gamma }_{id}}\right) %
\right] $ \\ \cline{2-3}
& ASEP & $\frac{\alpha \Gamma \left( \frac{\alpha \mu }{2}\right) \mu ^{\mu }%
}{2\Gamma \left( \mu \right) }\sum_{n=0}^{N}\theta _{n}\left( \frac{1}{%
\delta _{n}\overline{\gamma }_{id}}\right) ^{\frac{\alpha \mu }{2}}$ \\
\hline\hline
\parbox[t]{2mm}{\multirow{5}{*}{\rotatebox[origin=c]{90}{\small{%
\textit{Non-ideal RF e2e}}}}} & OP & $\frac{\mu ^{\mu -1}}{\Gamma \left( \mu
\right) }\left( \frac{\tau _{\varkappa }\gamma _{\text{th}}}{\left( \omega
_{\varkappa }-\varrho _{\varkappa }\gamma _{\text{th}}\right) \overline{%
\gamma }_{id}}\right) ^{\frac{\alpha \mu }{2}}$ \\ \cline{2-3}
& CC & $C^{\left( \varkappa ,\infty \right) }-\frac{\Theta _{\varkappa }}{%
\log (2)}H_{1,1:1,1:1,1}^{0,1:1,1:1,1}\left( \frac{\omega _{\varkappa }}{%
\varrho _{\varkappa }},-1\left\vert
\begin{array}{c}
\left( -\frac{\alpha \mu }{2};1,1\right) ;- \\
-;\left( -1-\frac{\alpha \mu }{2};1,1\right)%
\end{array}%
\right\vert
\begin{array}{c}
\left( 0,1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1-\frac{\alpha \mu }{2},1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\right. \right) $ \\ \cline{2-3}
& ASEP & $P_{s}^{\left( \varkappa ,\infty \right) }+\Theta _{\varkappa
}\sum_{n=1}^{N}\theta _{n}\delta _{n}H_{1,1:0,1:1,1}^{0,1:1,0:1,1}\left(
\frac{\delta _{n}\omega _{\varkappa }}{\varrho _{\varkappa }},-1\left\vert
\begin{array}{c}
\left( -\frac{\alpha \mu }{2};1,1\right) \\
-;\left( -1-\frac{\alpha \mu }{2};1,1\right)%
\end{array}%
\right\vert
\begin{array}{c}
- \\
\left( 0,1\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1-\frac{\alpha \mu }{2},1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\right. \right) $ \\ \hline
\end{tabular}%
\end{table}
\subsection{M\'alaga $\mathcal{M}$ \textcolor{black}{TCPE}}
According to \cite{faissal}, for M\'alaga $\mathcal{M}$ TCPE, the SNR's \ PDF
can be expressed in terms of FHF by setting $L=\beta ,$ $\phi _{\ell }=\frac{%
B^{\prime r}}{\mu _{r}}$, $\psi _{\ell }=\xi ^{2}A^{\prime }\lambda _{\ell
}\phi _{\ell },$ $m=q=3$, $n=0$, $p=1,$ $b_{1}^{\left( \ell \right) }=\xi
^{2}-r,$ $b_{2}^{\left( \ell \right) }=\alpha -r,$ $b_{3}^{\left( \ell
\right) }=\ell -r$ and $B_{1}^{\left( \ell \right) }=B_{2}^{\left( \ell
\right) }=B_{3}^{\left( \ell \right) }=r$ while $a_{1}^{\left( \ell \right)
}=\xi ^{2}-r+1$ and $\ A_{1}^{\left( \ell \right) }=r$ in $\left( \ref{pdf}%
\right) $.
With $\beta $ is a natural number denoting the amount of fading, $r$ is a
parameter defining the detection technique (i.e., $r=1,2$ refers to
heterodyne detection and IM/DD technique, respectively), while $\mu _{r}$
denotes the average SNR when $r=1$ and the average electrical SNR when $r=2.$
Further, $A^{\prime },B^{\prime }$, and $\lambda _{\ell }$ are positive real
parameters given by%
\begin{equation*}
A^{\prime }\triangleq \frac{\left( 1+\frac{\Omega ^{\prime }}{\kappa \beta }%
\right) ^{1-\beta }}{\Gamma \left( \alpha \right) },\text{ and }B^{\prime
}\triangleq \frac{\xi ^{2}\alpha \beta \left( \kappa +\Omega ^{\prime
}\right) }{\left( \xi ^{2}+1\right) \left( \kappa \beta +\Omega ^{\prime
}\right) },\text{ }\lambda _{\ell }\triangleq \left( _{\ell -1}^{\beta
-1}\right) \frac{1}{\left( \ell -1\right) !}\left( \frac{\Omega ^{\prime }}{%
\kappa }\right) ^{\ell -1},
\end{equation*}
with $\alpha $ is a positive number related to the effective number of
large-scale cells of the scattering process, and $\xi $ accounts for the ratio
between the equivalent beam radius at the receiver to the pointing error
dis-placement standard deviation. Further, $\kappa =2b_{0}\left( 1-\rho \right) $
denotes the scattering component's average received power, $2b_{0}$ is the
average power of the total scatter components, $0\leq \rho \leq 1$
represents the amount of scattering power coupled to the LOS component, $%
\Omega ^{\prime }=\Omega +2b_{0}\rho $ denotes the average power from the
coherent contributions and $\Omega $ is the LOS component's average power.
Therefore, the closed-form and asymptotic expressions for the three
considered metrics are summarized in Tables \ref{closedmalaga} and \ref%
{asympmalaga}, respectively, where $\Xi _{\ell }\triangleq \left\{ \left( \xi ^{2},r\right) ,\left(
\alpha ,r\right) ,\left( \ell ,r\right) \right\}$, while $\mathcal{L}\left( y\right)$, and $\mathcal{G}\left( y\right)$ are given by (\ref{LLL}) and (\ref{GGG}), respectively.
\begin{table}[tbh]
\captionsetup{font=small,format=hang
\caption{Closed-form expressions under M\'alaga $\mathcal{M}$ turbulence with
the presence of pointing errors.}
\centering\renewcommand{\arraystretch}{1.5}
\begin{tabular}{c|c|c}
\hline
\parbox[t]{2mm}{\multirow{5}{*}{\rotatebox[origin=c]{90}{\centering
\small{\textit{Ideal RF e2e}}}}} & OP & $1-\sum_{\ell =1}^{\beta }\frac{\psi
_{\ell }}{\phi _{\ell }}H_{2,4}^{4,0}\left( \phi _{\ell }\gamma _{\text{th}%
}\left\vert
\begin{array}{c}
-;\left( \xi ^{2}+1,r\right) ,\left( 1,1\right) \\
\left( 0,1\right) ,\Xi _{\ell };-%
\end{array}%
\right. \right) $ \\ \cline{2-3}
& CC & $\sum_{\ell =1}^{\beta }\frac{\psi _{\ell }}{\log \left( 2\right)
\phi _{\ell }}H_{3,5}^{5,1}\left( \phi _{\ell }\left\vert
\begin{array}{c}
\left( 0,1\right) ;\left( \xi ^{2}+1,r\right) ,\left( 1,1\right) \\
\left( 0,1\right) ,\left( 0,1\right) ,\Xi _{\ell };-%
\end{array}%
\right. \right) $ \\ \cline{2-3}
& ASEP & $\sum_{n=0}^{N}\theta _{n}\left[ 1-\sum_{\ell =1}^{\beta }\frac{%
\psi _{\ell }}{\phi _{\ell }}H_{3,4}^{4,1}\left( \frac{\phi _{\ell }}{\delta
_{n}}\left\vert
\begin{array}{c}
\left( 0,1\right) ;\left( \xi ^{2}+1,r\right) ,\left( 1,1\right) \\
\left( 0,1\right) ,\Xi _{\ell };-%
\end{array}%
\right. \right) \right] $ \\ \hline\hline
\parbox[t]{2mm}{\multirow{5}{*}{\rotatebox[origin=c]{90}{\small{%
\textit{Non-ideal RF e2e}}}}} & OP & $1-\sum_{\ell =1}^{\beta }\frac{\psi
_{\ell }}{\phi _{\ell }}H_{2,4}^{4,0}\left( \phi _{\ell }\frac{\tau
_{\varkappa }\gamma _{\text{th}}}{\omega _{\varkappa }-\varrho _{\varkappa
}\gamma _{\text{th}}}\left\vert
\begin{array}{c}
-;\left( \xi ^{2}+1,r\right) ,\left( 1,1\right) \\
\left( 0,1\right) ,\Xi _{\ell };-%
\end{array}%
\right. \right) $ \\ \cline{2-3}
& CC & $\frac{1}{\log (2)}\sum_{\ell =1}^{\beta }\frac{\psi _{\ell }}{\phi
_{\ell }}H_{1,0:1,4:1,2}^{0,1:4,0:1,1}\left( \frac{\phi _{\ell }\tau
_{\varkappa }}{\varrho _{\varkappa }},\frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}\left\vert
\begin{array}{c}
\left( 1;1,1\right) \\
-%
\end{array}%
\right\vert
\begin{array}{c}
-;\left( \xi ^{2}+1,r\right) \\
\left( 0,1\right) ,\Xi _{\ell };-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1,1\right) ;- \\
\left( 1,1\right) ;\left( 0,1\right)
\end{array}%
\right. \right) $ \\ \cline{2-3}
& ASEP & $\sum_{n=0}^{N}\theta _{n}\left[ 1-\sum_{\ell =1}^{\beta }\frac{%
\psi _{\ell }}{\phi _{\ell }}H_{1,0:1,4:0,2}^{0,1:4,0:1,0}\left( \frac{\phi
_{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }},\frac{\delta _{n}\omega
_{\varkappa }}{\varrho _{\varkappa }}\left\vert
\begin{array}{c}
\left( 1;1,1\right) \\
-%
\end{array}%
\right\vert
\begin{array}{c}
-;\left( \xi ^{2}+1,r\right) \\
\left( 0,1\right) ,\Xi _{\ell };-%
\end{array}%
\left\vert
\begin{array}{c}
- \\
\left( 1,1\right) ;\left( 0,1\right)
\end{array}%
\right. \right) \right] $ \\ \hline
\end{tabular}%
\label{closedmalaga}
\end{table}
\begin{equation}
\mathcal{L}\left( y\right) =\frac{\omega _{\varkappa }}{\varrho _{\varkappa
}\Gamma \left( y\right) }H_{1,1:1,1:1,1}^{0,1:1,1:1,1}\left( \frac{\omega
_{\varkappa }}{\varrho _{\varkappa }},-1\left\vert
\begin{array}{c}
\left( -y;1,1\right) ;- \\
-;\left( -1-y;1,1\right)
\end{array}%
\right\vert
\begin{array}{c}
\left( 0,1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1-y,1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\right. \right),\label{LLL}
\end{equation}
\begin{equation}
\mathcal{G}\left( y\right) =\frac{\omega _{\varkappa }}{\varrho _{\varkappa
}\Gamma \left( y\right) }H_{1,1:0,1:1,1}^{0,1:1,0:1,1}\left( \frac{\delta
_{n}\omega _{\varkappa }}{\varrho _{\varkappa }},-1\left\vert
\begin{array}{c}
\left( -y;1,1\right) \\
-;\left( -1-y;1,1\right)%
\end{array}%
\right\vert
\begin{array}{c}
- \\
\left( 0,1\right) ;-%
\end{array}%
\left\vert
\begin{array}{c}
\left( 1-y,1\right) ;- \\
\left( 0,1\right) ;-%
\end{array}%
\right. \right).\label{GGG}
\end{equation}
\begin{table}[tbh]
\captionsetup{font=small
\caption{Asymptotic expressions under M\'alaga $\mathcal{M}$ turbulence with
the presence of pointing errors.}
\label{asympmalaga}\centering\renewcommand{\arraystretch}{1.5}%
\begin{tabular}{c|c|c}
\hline
\parbox[t]{2mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{\small{\textit{Ideal RF e2e}}}}} & OP & $\sum_{\ell =1}^{\beta }\frac{\psi _{\ell }}{\phi _{\ell }}r%
\left[
\begin{array}{c}
\frac{\Gamma \left( \alpha -\xi ^{2}\right) \Gamma \left( k-\xi ^{2}\right)
}{\xi ^{2}}\left( \phi _{\ell }\gamma _{\text{th}}\right) ^{\frac{\xi ^{2}}{r%
}} \\
+\frac{\Gamma \left( \ell -\alpha \right) }{\alpha \left( \xi ^{2}-\alpha
\right) }\left( \phi _{\ell }\gamma _{\text{th}}\right) ^{\frac{\alpha }{r}}+%
\frac{\Gamma \left( \alpha -\ell \right) }{\ell \left( \xi ^{2}-\ell \right)
}\left( \phi _{\ell }\gamma _{\text{th}}\right) ^{\frac{\ell }{r}}%
\end{array}%
\right] $ \\ \cline{2-3}
& CC & $\sum_{\ell =1}^{\beta }\frac{\psi _{\ell }}{\log \left( 2\right)
\phi _{\ell }}\frac{\Gamma \left( \alpha \right) \Gamma \left( \ell \right)
}{\xi ^{2}}\left[ r\Psi \left( \alpha \right) +r\Psi \left( \ell \right) -%
\frac{r}{\xi ^{2}}-\log \left( \phi _{\ell }\right) \right] $ \\ \cline{2-3}
& ASEP & $r\sum_{\ell =1}^{\beta }\frac{\psi _{\ell }}{\phi _{\ell }}%
\sum_{n=0}^{N}\theta _{n}\left[
\begin{array}{c}
\frac{\Gamma \left( 1+\frac{\xi ^{2}}{r}\right) \Gamma \left( \alpha -\xi
^{2}\right) \Gamma \left( \ell -\xi ^{2}\right) }{\xi ^{2}}\left( \frac{\phi
_{\ell }}{\delta _{n}}\right) ^{\frac{\xi ^{2}}{r}} \\
+\frac{\Gamma \left( 1+\frac{\alpha }{r}\right) \Gamma \left( \ell -\alpha
\right) }{\alpha \left( \xi ^{2}-\alpha \right) }\left( \frac{\phi _{\ell }}{%
\delta _{n}}\right) ^{\frac{\alpha }{r}}+\frac{\Gamma \left( 1+\frac{\ell }{r%
}\right) \Gamma \left( \alpha -\ell \right) }{\ell \left( \xi ^{2}-\ell
\right) }\left( \frac{\phi _{\ell }}{\delta _{n}}\right) ^{\frac{\ell }{r}}%
\end{array}%
\right] $ \\ \hline\hline
\parbox[t]{2mm}{\multirow{5}{*}{\rotatebox[origin=c]{90}{\small{\textit{Non-ideal RF e2e}}}}} & OP & $\sum_{\ell =1}^{\beta }r\frac{\psi _{\ell }}{\phi _{\ell }}%
\left[
\begin{array}{c}
\frac{\Gamma \left( \alpha -\xi ^{2}\right) \Gamma \left( \ell -\xi
^{2}\right) }{\xi ^{2}}\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho
_{\varkappa }}\right) ^{\frac{\xi ^{2}}{r}}\left( \frac{\gamma _{\text{th}}}{%
\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}-\gamma _{\text{th}}}%
\right) ^{\frac{\xi ^{2}}{r}}\text{ } \\
+\frac{\Gamma \left( \ell -\alpha \right) }{\alpha \left( \xi ^{2}-\alpha
\right) }\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }}%
\right) ^{\frac{\alpha }{r}}\left( \frac{\gamma _{\text{th}}}{\frac{\omega
_{\varkappa }}{\varrho _{\varkappa }}-\gamma _{\text{th}}}\right) ^{\frac{%
\alpha }{r}}+\frac{\Gamma \left( \alpha -\ell \right) }{\ell \left( \xi
^{2}-\ell \right) }\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho
_{\varkappa }}\right) ^{\frac{\ell }{r}}\left( \frac{\gamma _{\text{th}}}{%
\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}-\gamma _{\text{th}}}%
\right) ^{\frac{\ell }{r}}%
\end{array}%
\right] $ \\ \cline{2-3}
& CC & $\mathcal{C}^{\left( \varkappa ,\infty \right) }-\sum_{\ell
=1}^{\beta }\frac{\psi _{\ell }}{\log \left( 2\right) \phi _{\ell }}\left[
\begin{array}{c}
\frac{\Gamma \left( \alpha -\xi ^{2}\right) \Gamma \left( \ell -\xi
^{2}\right) }{\xi ^{2}}\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho
_{\varkappa }}\right) ^{\xi ^{2}}\mathcal{L}\left( \xi ^{2}\right) +\frac{%
\Gamma \left( \ell -\alpha \right) }{\alpha \left( \xi ^{2}-\alpha \right) }%
\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }}\right)
^{\alpha }\mathcal{L}\left( \alpha \right) \\
+\frac{\Gamma \left( \alpha -\ell \right) }{\ell \left( \xi ^{2}-\ell
\right) }\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }}%
\right) ^{\ell }\mathcal{L}\left( \ell \right)%
\end{array}%
\right] $ \\ \cline{2-3}
& ASEP & $P_{s}^{\left( \varkappa ,\infty \right) }+\sum_{\ell =1}^{\beta }%
\frac{\psi _{\ell }}{\phi _{\ell }}\sum_{n=0}^{N}\theta _{n}\delta _{n}\left[
\begin{array}{c}
\frac{\Gamma \left( \alpha -\xi ^{2}\right) \Gamma \left( \ell -\xi
^{2}\right) }{\xi ^{2}}\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho
_{\varkappa }}\right) ^{\xi ^{2}}\mathcal{G}\left( \xi ^{2}\right) \\
+\frac{\Gamma \left( \ell -\alpha \right) }{\alpha \left( \xi ^{2}-\alpha
\right) }\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }}%
\right) ^{\alpha }\mathcal{G}\left( \alpha \right) +\frac{\Gamma \left(
\alpha -\ell \right) }{\ell \left( \xi ^{2}-\ell \right) }\left( \frac{\phi
_{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }}\right) ^{\ell }\mathcal{G}%
\left( \ell \right)%
\end{array}%
\right] $ \\ \hline
\end{tabular}%
\end{table}
\section{Numerical Results}
In this section, we evaluate and illustrate the effects of the RF impairments
on the performance of a WCS subject to either $\alpha-\mu$ fading
or M\'alaga $\mathcal{M}$ turbulence channel with the presence of pointing errors under heterodyne technique detection. To this end, the OP, CC, and ASEP are illustrated, assuming $IRR_{\varkappa}=20$ dB, $g_{\varkappa}<1$, $\varphi=3
{^{\circ}}$, and $\kappa_{t}=\kappa_{r}=0.2$. All the considered scenarios are treated, \textcolor{black}{namely (i) Tx impaired by both IQI and RHI, (ii) Rx impaired by both IQI and RHI, and (iii) both Tx and Rx are impaired.}
It is noted that in all figures, the \textcolor{black}{analytical} results are
shown \textcolor{black}{in} continuous lines, the markers are referring to the simulation ones, whereas the dashed lines illustrate the asymptotic curves.
Figures 1 and 2 present
the OP versus the normalized SNR for \textcolor{black}{all considered scenarios}. \textcolor{black}{It can be seen that} the OP degradation is more severe when the IQI is considered at Tx rather than Rx. \textcolor{black}{This can be explained as the noise is scaled by $\frac{1+g_{\varkappa}^{2}}{2}$
when Rx impaired only by IQI as can be ascertained in (\ref{yRXonly})} which doesn't exceed $1$ as $g_{\varkappa }<1$. \textcolor{black}{Furthermore, one} can
notice a minor performance loss caused by RHI impairments. \newline Figures
3 and 4 illustrate the effects of RF impairments on the \textcolor{black}{CC}. \textcolor{black}{As expected,} the IQI and RF impairments have a detrimental effect, \textcolor{black}{which} becomes more defective when the RHI impairments \textcolor{black}{are}
added, especially for high SNR \textcolor{black}{values}. Further, it can be
seen that the presence of the RF impairments \textcolor{black}{leads to a steady of the CC above a certain threshold, as discussed in Remark \ref{remarkCapaceiling}}.\newline Finally, to quantify the impact of the RF impairments
on the ASEP, we depict in Figures 5-18 the ASEP for GC-DQPSK, $M$-PSK, $M$-DPSK, and $M$-QAM \textcolor{black}{modulation techniques}. It is first observed that the RF impairments
exhibit different effects on \textcolor{black}{all the considered modulation schemes}.
For instance, it can be notice\textcolor{black}{d that the effect of the RF impairments on the ASEP is minor in the case of GC-DQPSK compared to \textcolor{black}{other }$M$-ary modulation techniques}. On the other hand, the effects such impairments \textcolor{black}{on the ASEP} become more detrimental
with the increasing of the modulation parameter $M$, while for a fixed value
of $M$, the impact these impairments on $M$-DPSK modulation is \textcolor{black}{more significant} compared \textcolor{black}{to the remaining considered} modulation schemes.
\section{Conclusion}
In this paper, a general framework for
\textcolor{black}{the analysis of OP,
CC, and ASEP} of a WCS fading channel in the
presence of both IQI and RHI at the RF e2e was developed. \textcolor{black}{Precisely}, three realistic %
\textcolor{black}{possible} cases were \textcolor{black}{then} considered,
\textcolor{black}{namely (i) Tx impaired by both IQI and RHI, (ii) Rx impaired
by both IQI and RHI, and (iii) both Tx and Rx are impaired.} The
corresponding metrics' closed-form, approximate, and asymptotic expressions were evaluated and useful insights into the overall system
performance we provided. Particularly, new tight approximate expression of the SEP for
numerous $M$-ary coherent and non-coherent modulation schemes were derived,
based on which, tight approximate and asymptotic expressions for
the ASEP were deduced. Numerical simulations confirmed analytical results for all considered impairments and fading models. As a result, we demonstrated that all metrics have ceilings that can't be crossed for any arbitrary fading model and entire SNR range. Moreover, it was shown that the RF impairments have a detrimental impact on the WCS performance and that such an influence becomes severe with the increase of modulation
\textcolor{black}{parameter $M$. Noticeably, the $M$-DPSK is the most sensitive modulation scheme to the RF impairments. Essentially, the RF impairments must be taken
into consideration in the efficient design of a WCS as its performance in high
SNR regime is irrespective of the fading model, while it is significantly
impacted by the RF impairments parameters.}
\clearpage
\begin{figure}[tbp]
\vspace{-1.2cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{OPalpha.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{$P_{\text{out}}$ versus the normalized SNR over $\alpha -\mu $ fading channel \\with $\alpha=2.3$ and $\mu=2.$ }
\label{figOpalpha}
\end{minipage}
\hfill %
\hspace{0.2cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{OPmalaga.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption {$P_{\text{out}}$ versus the normalized SNR over M\'alaga $\mathcal{M}$ turbulence channel\\ with $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{figOPmalaga}
\end{minipage}
\hfill
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm} \includegraphics[width=10cm, height=7cm]{capaalpha.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{Channel capacity versus SNR over $\alpha -\mu $ fading channel \\with $\alpha=3$ and $\mu=2.3 $.}
\label{figcapalpha}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{capamalaga.eps}
\vspace*{-8mm}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{Channel capacity versus SNR over M\'alaga $\mathcal{M}$ turbulence channel\\ with $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{figcapamalaga}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{dqpskalpha.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{ASEP for GC-DQPSK over $\alpha -\mu $ fading channel with\\ $\alpha =2.3$ and $\mu =2$.}
\label{figdqpskalpha}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{dqpskmalaga.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{ASEP for GC-DQPSK over M\'alaga $\mathcal{M}$ turbulence channel with\\ $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{figdqpskmalaga}
\end{minipage}
\end{figure}
\newpage
\begin{figure}[tbp]
\vspace{-1.2cm}
\par
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{8mpskalpha.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{ASEP for $8-$PSK over $\alpha -\mu $ fading channel with\\$\alpha =2.3$ and $\mu =2.$}
\label{fig8mpskalpha}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{16mpskalpha.eps}
\vspace*{-8mm}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{ASEP for $16-$PSK over $\alpha -\mu $ fading channel with\\$\alpha =2.3$ and $\mu =2.$}
\label{fig16mpskalpha}
\end{minipage}
\hfill
\par
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1.5cm} \includegraphics[width=10cm, height=7cm]{8mpskmalaga.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{ASEP for $8-$PSK over M\'alaga $\mathcal{M}$ turbulence channel with\\ $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{fig8mpskmalga}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{16mpskmalaga.eps}
\vspace*{-8mm}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{ASEP for $16-$PSK over M\'alaga $\mathcal{M}$ turbulence channel with\\ $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{fig16mpskmalga}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{8mdpskalpha.eps}
\vspace*{-8mm}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{ASEP for $8-$DPSK over $\alpha -\mu $ fading channel with\\$\alpha =2.3$ and $\mu =2.$}
\label{8mdpskalpha}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{16mdpskalpha.eps}
\vspace*{-8mm}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{ASEP for $16-$DPSK over $\alpha -\mu $ fading channel with\\$\alpha =2.3$ and $\mu =2.$}
\label{fig16mdpskalpha}
\end{minipage}
\end{figure}
\newpage
\begin{figure}[tbp]
\vspace{-1.2cm}
\par
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{8mdpskmalaga.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{ASEP for $8-$DPSK over M\'alaga $\mathcal{M}$ turbulence channel with\\ $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{fig8mdpskmalaga}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{16mdpskmalaga.eps}
\vspace*{-8mm}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{ASEP for $16-$DPSK over M\'alaga $\mathcal{M}$ turbulence channel with\\ $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{fig16mdpskmalaga}
\end{minipage}
\hfill
\par
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1.2cm} \includegraphics[width=10cm, height=7cm]{8qamalpha.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{ASEP for $8-$QAM over $\alpha -\mu $ fading channel with\\$\alpha =2.3$ and $\mu =2.$}
\label{fig8qamalpha}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{16qamalpha.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt
\caption{ASEP for $16-$QAM over $\alpha -\mu $ fading channel with\\$\alpha =2.3$ and $\mu =2.$}
\label{fig16qamalpha}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{8qammalaga.eps}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{ASEP for $8-$QAM over M\'alaga $\mathcal{M}$ turbulence channel with\\ $\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{fig8qammalaga}
\end{minipage}
\hfill \hspace{0.1cm}
\begin{minipage}[t]{0.55\columnwidth}
\hspace{-1cm}\includegraphics[width=10cm, height=7cm]{16qammalaga.eps}
\vspace*{-8mm}
\captionsetup{format=plain,font=scriptsize,skip=0pt,margin=-0.2cm}
\caption{ASEP for $16-$QAM over M\'alaga $\mathcal{M}$ turbulence channel with\\$\alpha =2.296,\beta =2,\Omega =1,\rho =0.596,\Omega ^{\prime }=1.3265$ and $\xi=3.85$.}
\label{fig16qammalaga}
\end{minipage}
\end{figure}
\clearpage
\newpage
\mbox{}
\section*{Appendix A: Proof of proposition \protect\ref{ORAexact}}
\begin{flushleft}
By setting $U^{\prime }=\frac{\partial F_{\gamma _{\varkappa }}\left( \gamma
\right) }{\partial \gamma },$ $V=\log (1+\gamma )$, and using the
integration by parts in $\left( \ref{ORAexpression}\right) $ with the help
of $\left( \ref{CDFIQI}\right) $, one obtains
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &=&\frac{1}{\log \left( 2\right) }%
\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{0}^{\frac{\omega
_{\varkappa }}{\varrho _{\varkappa }}}\frac{1}{1+\gamma }{\small H}_{p+1{\small ,q+1}}^{m+1{\small ,n}}\left( \frac{\phi
_{\ell }\tau _{\varkappa }\gamma }{\omega _{\varkappa }-\varrho _{\varkappa
}\gamma }\left\vert
\begin{array}{c}
\Psi ^{\left( \ell \right) } \\
\left( 0,1\right) ,\Upsilon ^{\left( \ell \right) }%
\end{array}%
\right. \right) d\gamma. \label{EQORAappedix}
\end{eqnarray}
by plugging (\ref{Hffox}) into (\ref{EQORAappedix}), we get
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &=&\frac{1}{\log \left( 2\right) }%
\frac{1}{2\pi j}\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{%
\mathcal{L}_{v}^{\left( \ell \right) }}\frac{\prod\limits_{i=1}^{m}\Gamma
\left( \mathcal{B}_{i}^{\left( \ell \right) }+B_{i}^{\left( \ell \right)
}v\right) \prod\limits_{i=1}^{n}\Gamma \left( 1-\mathcal{A}_{i}^{\left( \ell
\right) }-A_{i}^{\left( \ell \right) }v\right) }{v\prod\limits_{i=n+1}^{p}%
\Gamma \left( \mathcal{A}_{i}^{\left( \ell \right) }+A_{i}^{\left( \ell
\right) }v\right) \prod\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}%
_{i}^{\left( \ell \right) }-B_{i}^{\left( \ell \right) }v\right) } \notag \\
&&\times \left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }}%
\right) ^{-v}\underbrace{\int_{0}^{\frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}}\gamma ^{-v}\left( 1+\gamma \right) ^{-1}\left( \frac{\omega
_{\varkappa }}{\varrho _{\varkappa }}-\gamma \right) ^{v}d\gamma }_{\mathcal{%
I}_{\varkappa }}dv. \label{ORAdetail}
\end{eqnarray}%
Further, the inner integral $\mathcal{I}_{\varkappa }$ can be evaluated
using \cite[Eq.(3.197.8)]{integraltable} as
\begin{equation}
\mathcal{I}_{\varkappa }\mathcal{=}\frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}\Gamma \left( 1+v\right) \Gamma \left( 1-v\right) \text{ }%
_{2}F_{1}\left( 1,1-v;2;-\frac{\omega _{\varkappa }}{\varrho _{\varkappa }}%
\right) , \label{innerintegral}
\end{equation}%
where $_{2}F_{1}\left( .,.;.;.\right) $ denotes the Gauss hypergeometric
function (GHF) \cite[Eq. (9.100)]{integraltable}.\newline
Now plugging $\left( \ref{innerintegral}\right) $ into $\left( \ref%
{ORAdetail}\right) $ and using the integral Mellin-Barnes representation of
the GHF \cite[Eq. (9.113)]{integraltable}, yields
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &=&\frac{1}{\log (2)}\left( \frac{1}{%
2\pi j}\right) ^{2}\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}%
\int_{\mathcal{L}_{v}^{\left( \ell \right) }}\int_{\mathcal{L}_{w}^{\left(
\ell \right) }}\frac{\Gamma \left( {-v-w}\right) \Gamma \left( v\right)
\prod\limits_{i=1}^{m}\Gamma \left( \mathcal{B}_{i}^{\left( \ell \right)
}+B_{i}^{\left( \ell \right) }v\right) \prod\limits_{i=1}^{n}\Gamma \left( 1-%
\mathcal{A}_{i}^{\left( \ell \right) }-A_{i}^{\left( \ell \right) }v\right)
}{\prod\limits_{i=n+1}^{p}\Gamma \left( \mathcal{A}_{i}^{\left( \ell \right)
}+A_{i}^{\left( \ell \right) }v\right) \prod\limits_{i=m+1}^{q}\Gamma \left(
1-\mathcal{B}_{i}^{\left( \ell \right) }-A_{i}^{\left( \ell \right)
}v\right) } \notag \\
&&\times \frac{\Gamma \left( 1+w\right) \Gamma \left( {-w}\right) }{\Gamma
\left( 1-w\right) }\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho
_{\varkappa }}\right) ^{-v}\left( \frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}\right) ^{-w}dvdw, \label{ORRA}
\end{eqnarray}%
which concludes the proof of Proposition 1.
\end{flushleft}
\section*{Appendix B: Proof of corollary \protect\ref{crolaryidealcapa}}
\begin{flushleft}
By replacing $\omega _{\varkappa }=\tau _{\varkappa }=1$ in $\left( \ref%
{ORRA}\right) $, the CC under ORA policy can be simplified to%
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &=&\frac{1}{\log (2)}\left( \frac{1}{%
2\pi j}\right) ^{2}\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}%
\int_{\mathcal{L}_{v}^{\left( \ell \right) }}\int_{\mathcal{L}_{w}^{\left(
\ell \right) }}\frac{\Gamma \left( -v-w\right) \Gamma \left( v\right)
\prod\limits_{i=1}^{m}\Gamma \left( \mathcal{B}_{i}^{\left( \ell \right)
}+B_{i}^{\left( \ell \right) }v\right) }{\prod\limits_{i=n+1}^{p}\Gamma
\left( \mathcal{A}_{i}^{\left( \ell \right) }+A_{i}^{\left( \ell \right)
}v\right) } \notag \\
&&\times \frac{\prod\limits_{i=1}^{n}\Gamma \left( 1-\mathcal{A}_{i}^{\left(
\ell \right) }-A_{i}^{\left( \ell \right) }v\right) }{\prod%
\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}_{i}^{\left( \ell \right)
}-B_{i}^{\left( \ell \right) }v\right) }\frac{\Gamma \left( -w\right) \Gamma
\left( 1+w\right) }{\Gamma \left( 1-w\right) }\phi _{\ell }^{-v}\left( \frac{%
1}{\varrho _{\varkappa }}\right) ^{-w-v}dvdw. \notag
\end{eqnarray}%
Using the change of variable $s=$ $-w-v,$ we get
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &=&\frac{1}{\log (2)}\frac{1}{2\pi j}%
\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{\mathcal{L}%
_{v}^{\left( \ell \right) }}\frac{\Gamma \left( v\right)
\prod\limits_{i=1}^{m}\Gamma \left( \mathcal{B}_{i}^{\left( \ell \right)
}+B_{i}^{\left( \ell \right) }v\right) \prod\limits_{i=1}^{n}\Gamma \left( 1-%
\mathcal{A}_{i}^{\left( \ell \right) }-A_{i}^{\left( \ell \right) }v\right)
}{\prod\limits_{i=n+1}^{p}\Gamma \left( \mathcal{A}_{i}^{\left( \ell \right)
}+A_{i}^{\left( \ell \right) }v\right) \prod\limits_{i=m+1}^{q}\Gamma \left(
1-\mathcal{B}_{i}^{\left( \ell \right) }-B_{i}^{\left( \ell \right)
}v\right) }\phi _{\ell }^{-v} \notag \\
&&\times \underbrace{\frac{1}{2\pi j}\int_{\mathcal{L}_{s}^{\left( \ell
\right) }}\frac{\Gamma \left( s+v\right) \Gamma \left( 1-s-v\right) }{\Gamma
\left( 1+s+v\right) }\Gamma \left( s\right) \varrho _{\varkappa }^{-s}ds}_{%
\mathcal{K}_{\varkappa }}dv
\end{eqnarray}%
For significantly smaller values of $\varrho _{\varkappa }$ (i.e., $\varrho
_{\varkappa }<<1$), the condition \cite[Eq. (1.2.15)]{kilbas} is satisfied.
It follows that the inner integral $\mathcal{K}_{\varkappa }$ can be
expressed as an infinite summation of the residues evaluated at the left poles.
Subsequently, relying on \cite[Theorem 1.3]{kilbas}, $\mathcal{K}_{\varkappa
}$ can be expressed as
\begin{equation}
\mathcal{K}_{\varkappa }\mathcal{=}\sum_{l=0}^{\infty }\frac{\left(
-1\right) ^{l}\Gamma \left( -l+v\right) \Gamma \left( 1+l-v\right) }{%
l!\Gamma \left( 1-l+v\right) }\varrho _{\varkappa }^{l}. \label{K-residu}
\end{equation}%
Particularly, for $\varrho _{\varkappa }=0$ (i.e., case of ideal RF e$2$e), $%
\left( \ref{K-residu}\right) $ is reduced to
\begin{equation}
\mathcal{K}_{\varkappa }\mathcal{=}\frac{\Gamma \left( v\right) \Gamma
\left( 1-v\right) }{\Gamma \left( 1+v\right) }. \label{K-res22}
\end{equation}%
Thus, the CC under ORA policy and ideal RF e$2$e can be expressed as%
\begin{eqnarray}
\mathcal{C}^{\left( \varkappa \right) } &=&\frac{1}{\log (2)}\sum_{\ell
=1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\left( \frac{1}{2\pi j}\right)
\int_{\mathcal{L}_{v}^{\left( \ell \right) }}\frac{\Gamma \left( v\right)
\Gamma \left( v\right) \Gamma \left( 1-v\right) }{\Gamma \left( 1+v\right) }
\notag \\
&&\times \frac{\prod\limits_{i=1}^{m}\Gamma \left( \mathcal{B}_{i}^{\left(
\ell \right) }+B_{i}^{\left( \ell \right) }v\right)
\prod\limits_{i=1}^{n}\Gamma \left( 1-\mathcal{A}_{i}^{\left( \ell \right)
}-A_{i}^{\left( \ell \right) }v\right) }{\prod\limits_{i=n+1}^{p}\Gamma
\left( \mathcal{A}_{i}^{\left( \ell \right) }+A_{i}^{\left( \ell \right)
}v\right) \prod\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}_{i}^{\left(
\ell \right) }-B_{i}^{\left( \ell \right) }v\right) }\phi _{\ell }^{-v}dv.
\end{eqnarray}%
This concludes the proof.
\end{flushleft}
\section*{Appendix C: Proof of proposition \protect\ref{propASEPexact}}
\begin{flushleft}
\textcolor{black}{By substituting $\left( \ref{SEPbin}\right)$ into the
following equation}
\begin{equation}
P_{s}^{\left( \varkappa \right) }=\int_{0}^{\frac{\omega _{\varkappa }}{%
\varrho _{\varkappa }}}\mathcal{H}\left( \gamma \right) \frac{\partial
F_{\gamma _{\varkappa }}\left( \gamma \right) }{\partial \gamma }d\gamma ,
\label{PSexpression}
\end{equation}%
one \textcolor{black}{gets a tight approximate formula}
\begin{equation}
P_{s}^{\left( \varkappa \right) }\simeq \sum_{n=0}^{N}\theta _{n}\mathcal{I}%
_{n}^{\left( \varkappa \right) }, \label{Psaprox}
\end{equation}%
with
\begin{equation}
\mathcal{I}_{n}^{\left( \varkappa \right) }=\int_{0}^{\frac{\omega
_{\varkappa }}{\varrho _{\varkappa }}}\exp \left( -\delta _{n}\gamma \right)
\frac{\partial F_{\gamma _{\varkappa }}\left( \gamma \right) }{\partial
\gamma }d\gamma . \label{integral_n}
\end{equation}%
By setting $U^{\prime }=\frac{\partial F_{\gamma _{\varkappa }}\left( \gamma
\right) }{\partial \gamma }$, $V=\exp \left( -\delta _{n}\gamma \right) $,
and using $\left( \ref{CDFIQI}\right) $, \textcolor{black}{one can obtain}
\begin{eqnarray}
\mathcal{I}_{n}^{\left( \varkappa \right) } &=&1-\delta _{n}\sum_{\ell
=1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{0}^{\frac{\omega _{\varkappa
}}{\varrho _{\varkappa }}}\exp \left( -\delta _{n}\gamma \right) \notag \\
&&\times {\small H}_{p+1{\small ,q+1}}^{m+1{\small ,n}}\left( \frac{\phi
_{\ell }\tau _{\varkappa }\gamma }{\omega _{\varkappa }-\varrho _{\varkappa
}\gamma }\left\vert
\begin{array}{c}
\left( \mathcal{A}_{i}^{\left( \ell \right) },A_{i}^{\left( \ell \right)
}\right) _{i=1:p},\left( 1,1\right) \\
\left( 0,1\right) ,\left( \mathcal{B}_{i}^{\left( \ell \right)
},B_{i}^{\left( \ell \right) }\right) _{i=1:q_{{}}}%
\end{array}%
\right. \right) d\gamma . \notag \\
&& \label{EqIntegral-n}
\end{eqnarray}%
\textcolor{black}{Subsequently,} $\left( \ref{EqIntegral-n}\right) $ can be
rewritten with the help of $\left( \ref{Hffox}\right) $ as
\begin{eqnarray}
\mathcal{I}_{n}^{\left( \varkappa \right) } &=&1-\frac{\delta _{n}}{2\pi j}%
\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{\mathcal{L}%
_{v}^{\left( \ell \right) }}\frac{\prod\limits_{i=1}^{m}\Gamma \left(
\mathcal{B}_{i}^{\left( \ell \right) }+B_{i}^{\left( \ell \right) }v\right)
\prod\limits_{i=1}^{n}\Gamma \left( 1-\mathcal{A}_{i}^{\left( \ell \right)
}-A_{i}^{\left( \ell \right) }v\right) }{v\prod\limits_{i=n+1}^{p}\Gamma
\left( \mathcal{A}_{i}^{\left( \ell \right) }+A_{i}^{\left( \ell \right)
}v\right) \prod\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}_{i}^{\left(
\ell \right) }-B_{i}^{\left( \ell \right) }v\right) } \notag \\
&&\times \left( \frac{\phi _{\ell }\tau _{\varkappa }}{\varrho _{\varkappa }}%
\right) ^{-v}\underbrace{\int_{0}^{\frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}}\gamma ^{-v}\left( \frac{\omega _{\varkappa }}{\varrho
_{\varkappa }}-\gamma \right) ^{v}\exp \left( -\delta _{n}\gamma \right)
d\gamma }_{\mathcal{J}_{\varkappa }}dv. \label{integ2}
\end{eqnarray}%
\textcolor{black}{whereas} the inner integral $\mathcal{J}_{\varkappa }$ can
be evaluated using \cite[Eq. (3.383.1)]{integraltable} as
\begin{equation}
\mathcal{J}_{\varkappa }\mathcal{=}\Gamma \left( 1+v\right) \frac{\omega
_{\varkappa }}{\varrho _{\varkappa }}\text{ }_{1}F_{1}\left( 1-v;2;-\frac{%
\delta _{n}\omega _{\varkappa }}{\varrho _{\varkappa }}\right) ,
\label{J_integral}
\end{equation}%
where $_{1}F_{1}\left( .;.;.\right) $ denotes the confluent hypergeometric function \cite[ Eq. (9.21)]{integraltable}. \textcolor{black}{Lastly,}
plugging $\left( \ref{J_integral}\right) $ into $\left( \ref{integ2}\right) $
and \textcolor{black}{using} \cite[Eqs. (2.9.14)]{kilbas} %
\textcolor{black}{along} with (\ref{Hffox}), yields
\begin{eqnarray}
\mathcal{I}_{n}^{\left( \varkappa \right) } &=&1-\left( \frac{1}{2\pi j}%
\right) ^{2}\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{%
\mathcal{L}_{v}^{\left( \ell \right) }}\int_{\mathcal{L}_{s}^{\left( \ell
\right) }}\Gamma \left( -v-s\right) \frac{\Gamma \left( v\right)
\prod\limits_{i=1}^{m}\Gamma \left( \mathcal{B}_{i}^{\left( \ell \right)
}+B_{i}^{\left( \ell \right) }v\right) }{\prod\limits_{i=n+1}^{p}\Gamma
\left( \mathcal{A}_{i}^{\left( \ell \right) }+A_{i}^{\left( \ell \right)
}v\right) } \notag \\
&&\times \frac{\prod\limits_{i=1}^{n}\Gamma \left( 1-\mathcal{A}_{i}^{\left(
\ell \right) }-A_{i}^{\left( \ell \right) }v\right) }{\prod%
\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}_{i}^{\left( \ell \right)
}-B_{i}^{\left( \ell \right) }v\right) }\frac{\Gamma \left( 1+s\right) }{%
\Gamma \left( 1-s\right) }\left( \frac{\phi _{\ell }\tau _{\varkappa }}{%
\varrho _{\varkappa }}\right) ^{-v}\left( \frac{\delta _{n}\omega
_{\varkappa }}{\varrho _{\varkappa }}\right) ^{-s}dvds, \label{Integral_n}
\end{eqnarray}
%
\textcolor{black}{which concludes the proof.}
\end{flushleft}
\section*{Appendix D: Proof of corollary \protect\ref{corolidealasep}}
\begin{flushleft}
By setting $\omega _{\varkappa }=\tau _{\varkappa }=1,\left( \ref{Integral_n}%
\right) $ can be rewritten using the change of variable $w=-s-v$ as
\begin{eqnarray}
\mathcal{I}_{n}^{\left( \varkappa \right) } &=&1-\frac{1}{2\pi j}\sum_{\ell
=1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{\mathcal{L}_{v}^{\left( \ell
\right) }}\frac{\Gamma \left( v\right) \prod\limits_{i=1}^{m}\Gamma \left(
\mathcal{B}_{i}^{\left( \ell \right) }+B_{i}^{\left( \ell \right) }v\right)
\prod\limits_{i=1}^{n}\Gamma \left( 1-\mathcal{A}_{i}^{\left( \ell \right)
}-A_{i}^{\left( \ell \right) }v\right) }{\prod\limits_{i=n+1}^{p}\Gamma
\left( \mathcal{A}_{i}^{\left( \ell \right) }+A_{i}^{\left( \ell \right)
}v\right) \prod\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}_{i}^{\left(
\ell \right) }-B_{i}^{\left( \ell \right) }v\right) } \notag \\
&&\times \left( \frac{\phi _{\ell }}{\delta _{n}}\right) ^{-v}\underbrace{%
\frac{1}{2\pi j}\int_{\mathcal{L}_{w}^{\left( \ell \right) }}\Gamma \left(
w\right) \frac{\Gamma \left( 1-w-v\right) }{\Gamma \left( 1+w+v\right) }%
\left( \frac{\varrho _{\varkappa }}{\delta _{n}}\right) ^{w}dw}_{\mathcal{M}%
_{\varkappa }}dv. \label{Integrall-n}
\end{eqnarray}%
\textcolor{black}{Again, for significantly smaller values of $\varrho _{\varkappa }$ (i.e., $\varrho _{\varkappa }<<1$) and similarly to $\mathcal{K}_{\varkappa }$, the inner integral $\mathcal{M}_{\varkappa }$ can
be expressed as}
\begin{equation}
\mathcal{M}_{\varkappa }\mathcal{=}\sum_{l=0}^{\infty }\frac{\left(
-1\right) ^{l}\Gamma \left( 1+l-v\right) }{l!\Gamma \left( 1-l+v\right) }%
\left( \frac{\varrho _{\varkappa }}{\delta _{n}}\right) ^{l}.
\label{M-residu}
\end{equation}%
Specifically, for $\varrho _{\varkappa }=0$, $\left( \ref{M-residu}\right) $
is reduced to
\begin{equation}
\mathcal{M}_{\varkappa }\mathcal{=}\frac{\Gamma \left( 1-v\right) }{\Gamma
\left( 1+v\right) }. \label{M-res2}
\end{equation}%
\textcolor{black}{As a result, $\left( \ref{Integrall-n}\right)$ can be
reexpressed by incorporating $\left( \ref{M-res2}\right)$ as}
\begin{eqnarray}
\mathcal{I}_{n}^{\left( \varkappa \right) } &=&1-\frac{1}{2\pi j}\sum_{\ell
=1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\int_{\mathcal{L}_{v}^{\left( \ell
\right) }}\frac{\Gamma \left( v\right) \prod\limits_{i=1}^{m}\Gamma \left(
\mathcal{B}_{i}^{\left( \ell \right) }+B_{i}^{\left( \ell \right) }v\right)
}{\Gamma \left( 1+v\right) \prod\limits_{i=n+1}^{p}\Gamma \left( \mathcal{A}%
_{i}^{\left( \ell \right) }+A_{i}^{\left( \ell \right) }v\right) } \notag \\
&&\times \frac{\Gamma \left( 1-v\right) \prod\limits_{i=1}^{n}\Gamma \left(
1-\mathcal{A}_{i}^{\left( \ell \right) }-A_{i}^{\left( \ell \right)
}v\right) }{\prod\limits_{i=m+1}^{q}\Gamma \left( 1-\mathcal{B}_{i}^{\left(
\ell \right) }-B_{i}^{\left( \ell \right) }v\right) }\left( \frac{\phi
_{\ell }}{\delta _{n}}\right) ^{-v}dv.
\end{eqnarray}%
\textcolor{black}{Thus, the proof is concluded.}
\end{flushleft}
\section*{Appendix E: Proof of proposition \protect\ref{propasympAsep}}
\begin{flushleft}
By plugging $\left( \ref{ApproFHF}\right) $ in $\left( \ref{EqIntegral-n}%
\right) $ and using \cite[Eq. (2.9.4)]{kilbas}, yields
\textcolor{black}{the
following approximate expression in the high SNR regime}
\begin{eqnarray}
\mathcal{I}_{n}^{\left( \varkappa \right) } &\sim &1-\sum_{\ell =1}^{L}\frac{%
\psi _{\ell }}{\phi _{\ell }}E_{\ell }\left[ 1-\exp \left( -\frac{\delta
_{n}\omega _{\varkappa }}{\varrho _{\varkappa }}\right) \right] +\delta
_{n}\sum_{\ell =1}^{L}\frac{\psi _{\ell }}{\phi _{\ell }}\sum_{i=1}^{m}%
\mathcal{F}_{\ell ,i}\left( \frac{\phi _{\ell }\tau _{\varkappa }}{\omega
_{\varkappa }}\right) ^{\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}} \notag \\
&&\times \frac{1}{2\pi j}\int_{\mathcal{L}_{s}^{\left( \ell \right) }}\Gamma
\left( s\right) \left( \delta _{n}\right) ^{-s}\int_{0}^{\frac{\omega
_{\varkappa }}{\varrho _{\varkappa }}}\gamma ^{\frac{\mathcal{B}_{i}^{\left(
\ell \right) }}{B_{i}^{\left( \ell \right) }}-s}\left( 1-\frac{\varrho
_{\varkappa }}{\omega _{\varkappa }}\gamma \right) ^{\frac{\mathcal{B}%
_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}}d\gamma ds.\label{HH}
\end{eqnarray}
Now, using $\left( \ref{Hffox}\right) $, \cite[Eq. (2.9.15)]{kilbas}, and
\cite[Eq. (1.3.194)]{integraltable}, one obtains
\begin{eqnarray}
\mathcal{I}_{n}^{\left( \varkappa \right) } &\sim &\exp \left( -\frac{\delta
_{n}\omega _{\varkappa }}{\varrho _{\varkappa }}\right) +\delta
_{n}\sum_{\ell =1}^{L}\frac{\psi _{\ell }\omega _{\varkappa }}{\phi _{\ell
}\varrho _{\varkappa }}\sum_{i=1}^{m}\frac{\mathcal{F}_{\ell ,i}\left( \frac{%
\phi _{\ell }\tau _{\varkappa }}{\omega _{\varkappa }}\right) ^{\frac{%
\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}}}{%
\Gamma \left( \frac{\mathcal{B}_{i}}{B_{i}}\right) }\left( \frac{1}{2\pi j}%
\right) ^{2}\int_{\mathcal{L}_{s}^{\left( \ell \right) }}\int_{\mathcal{L}%
_{v}^{\left( \ell \right) }}\Gamma \left( s\right) \Gamma \left( v\right)
\notag \\
&&\times \frac{\Gamma \left( 1+\frac{\mathcal{B}_{i}^{\left( \ell \right) }}{%
B_{i}^{\left( \ell \right) }}-s-v\right) }{\Gamma \left( 2+\frac{\mathcal{B}%
_{i}^{\left( \ell \right) }}{B_{i}^{\left( \ell \right) }}-s-v\right) }%
\Gamma \left( \frac{\mathcal{B}_{i}^{\left( \ell \right) }}{B_{i}^{\left(
\ell \right) }}-v\right) \left( \frac{\delta _{n}\omega _{\varkappa }}{%
\varrho _{\varkappa }}\right) ^{-s}\left( -1\right) ^{-v}dsdv.\label{RRR}
\end{eqnarray}%
Finally, substituting (\ref{RRR}) into (\ref{Psaprox}), yields (\ref%
{AsymptoASEP}), which concludes the proof.
\end{flushleft}
|
\chapter{Prelude}
This thesis presents three main results in geometric analysis. Chapter 2 is concerned with the evolution of figure-eight curves by curve-shortening flow. Chapter 3 presents our work on the geometric flow on space curves that is point-wise curvature preserving. We find all stationary solutions of this flow and show that the stationary solutions corresponding to helices are $L^2$ linearly stable. Most of this chapter has been published as an article in Archiv der Mathematik \cite{COI2}, but we have added a section here on numerical solutions of the key partial differential equation. Chapter 4 presents an abridgment of our work on a family of Lie groups that, when equipped with canonical left-invariant metrics, interpolates between the Sol geometry and a model of Hyperbolic space. The complete version of our work will appear in Experimental Mathematics, and a preprint can be found on the ArXiv as well \cite{COI}.
The common thread running through our work is the geometric analysis of differential equations. That is to say: the obtainment of information about geometric objects by the analysis of (usually nonlinear) differential equations. Deep and difficult work on the Ricci flow and the behavior of minimal submanifolds are some of the most famous parts of geometric analysis, but we will deal with simpler and more concrete problems in this thesis.
\begin{center}
\textbf{Curve-Shortening Flow on Figure-Eight Curves}
\end{center}
Let $C(t): S^1\cross [0,T) \to \mathbb{R}^2$ be a family of immersed curves in the plane. We say that the family is evolving according to curve shortening flow (CSF) if and only if at any $(u,t) \in S^1\cross [0,T)$ we have
$$\frac{\partial C}{\partial t} = k N$$
where $k$ is the curvature and $N$ is the unit normal vector of the immersed curve $C(\cdot,t)$. For any given smoothly immersed initial curve $C(0)$, the existence of a corresponding family of curves $C(t)$ undergoing CSF is guaranteed, and the proof of short-term existence can be found in \cite{GH}. Due to the work of Gage, Hamilton, and Grayson in \cite{GH} and \cite{G}, we know that CSF shrinks any smoothly embedded initial curve to a point. We also know that, after the appropriate rescaling, any embedded curve becomes a circle in the limit. We will start this chapter by presenting some basic lemmas about the CSF, proven first in \cite{ANG}, \cite{GH}, or \cite{G}.
Later in the chapter, we focus on the asymptotic results about balanced figure-eight curves (having equal-area lobes) obtained by Grayson in \cite{G3}. Namely, we will present Grayson's proof that the isoperimetric ratio of a balanced figure-eight blows up in the singularity. Then, we will present Angenent's main result from \cite{ANG} that characterizes the blowup set of collapsing lobes, and we will apply Angenent's result to figure-eight curves. We finish this chapter with a discussion of a related problem communicated from Richard Schwartz concerning the evolution of figure-eight curves that have only one inflection point and are four-fold symmetric. These curves, which we term \textbf{Concinnous}, will be the main object of our study in the latter portion of Chapter 2. We will offer a method towards proving the following result, which describes the limiting shape of a concinnous figure-eight curve after evolution by CSF and an affine rescaling. Qualitatively, if we rescale the curve so that it has aspect ratio equal to $1$, the shape will limit to that of a bowtie. We summarize our main result below:
\begin{theorem*}
Concinnous figure-eight curves converge to a point under CSF, and, after affine-rescaling, concinnous curves converge to a bow-tie shape.
\end{theorem*}
The proof in this chapter is incomplete. For example, we need to also show that the results in \cite{ANG} (where strictly convex immersed curves are studied) still apply to the case of the figure-eight. However, we are preparing another paper which will address these issues.
\begin{center}
\textbf{The Curvature Preserving Flow on Space Curves}
\end{center}
Substantial work has been done towards understanding geometric flows on curves immersed in Riemannian Manifolds. For example, the author of \cite{H} found that the vortex filament flow is equivalent to the non-linear Schr\"odinger equation, which enabled the discovery of explicit soliton solutions. Recently, geometric evolutions that are integrable, in the sense of admitting a Hamiltonian structure, have been of interest. The authors of \cite{I} and \cite{BSW} analyze the integrability of flows in Euclidean space and Riemannian Manifolds, respectively. In this chapter, we study the following geometric flow for curves in $\mathbb{R}^3$ with strictly positive torsion that preserves arc-length and curvature:
$$X_t=\frac{1}{\sqrt{\tau}}\textbf{B}.$$
Hydrodynamic and magnetodynamic motions related to this geometric evolution equation have been considered, as mentioned in \cite{SR}. The case when curvature is constant demonstrates a great deal of structure. After rescaling so that the curvature is identically $1$, the evolution equation for torsion is given by:
$$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big).$$
This flow has been studied since at least the publication of \cite{SR}. The authors of \cite{SR} show that the above equation, which they term the \textit{extended Dym equation}, is equivalent to the m$^2$KDV equation. In addition, they present auto-B\"acklund transformations and compute explicit soliton solutions. In Chapter 3, we continue the investigation of the curvature-preserving geometric flow.
We will characterize the flow $X_t=\frac{1}{\sqrt{\tau}}\textbf{B}$ as the unique flow on space curves that is both curvature and arc-length preserving, and we will provide another proof that this flow is equivalent to the m$^2$KDV equation. Afterwards, we will prove the global existence and uniqueness of solutions of the flow (for periodic $C^\infty$ initial data).
Then, we will concern ourselves with the stationary solutions to the geometric flow in the case of constant curvature. Namely, we derive the linearization of the evolution equation for torsion around explicit stationary solutions, and prove $L^2(\mathbb{R})$ stability of the linearization in the case of constant torsion (or for helices).
In addition to our work in \cite{COI2}, we present some numerical work done with Mathematica related to the evolution equation for torsion.
\begin{center}
\textbf{An Interpolation from Sol to Hyperbolic Space}
\end{center}
As mentioned earlier, this chapter presents an abbreviation of our work in \cite{COI}, whose complete version will appear in Experimental Mathematics. We study a one-parameter family of homogeneous Riemannian 3-manifolds that interpolates between three Thurston geometries: Sol, $\mathbb{H}^2\cross\mathbb{R}$, and $\mathbb{H}^3$. Sol is quite strange from a geometric point of view. For example, it is neither rotationally symmetric nor isotropic, and, since Sol has sectional curvature of both signs, the Riemannian exponential map is singular. In this article, we attempt to show that Sol's peculiarity can be slowly untangled by an interpolation of geometries until we reach $\mathbb{H}^2\cross\mathbb{R}$, which has a qualitatively "better" behavior than Sol. The interpolation continues on to $\mathbb{H}^3$, but we will not spend much time with that part of the family.
Our family of Riemannian 3-manifolds arises from a one-parameter family of solvable Lie groups equipped with canonical left-invariant metrics. We denote the groups by $G_\alpha$ with $-1\leq\alpha\leq 1$. Each $G_\alpha$ is the semi-direct product of $\mathbb{R}$ with $\mathbb{R}^2$, with the following operation on $\mathbb{R}^3$:
$$(x,y,z)\ast(x',y',z')=(x'e^z+x, y'e^{-\alpha z}+y, z'+z).$$
Then $\mathbb{R}^3$, which is only the underlying set, can be equipped with the following left-invariant metric:
$$ds^2=e^{-2z}dx^2+e^{2\alpha z}dy^2+dz^2.$$
These $G_\alpha$ groups, when endowed with the canonical metric, perform our desired interpolation: linking the familiar and the unfamiliar. It is a natural question to analyze what happens in between.
For positive $\alpha$, it happens that typical geodesics starting at the origin in $G_\alpha$ spiral around certain cylinders. For each such geodesic, there is an associated period that determines how long it takes for it to spiral exactly once around. We denote the function determining the period by $P$, and we will show that it is a function of the initial tangent vector. We call a geodesic segment $\gamma$ of length $T$ $\textit{small, perfect,}$ or $large$ whenever $T<P_\gamma,$ $T=P_\gamma,$ or $T>P_\gamma,$ respectively. Our primary aim is the following conjecture:
\begin{conjecture*}
A geodesic segment in $G_\alpha$ is a length minimizer
if and only if it is small or perfect.
\end{conjecture*}
The above conjecture is already known for $G_1$, or Sol, and was proven in \cite{MS}. In this article, we will reduce the general conjecture to obtaining bounds on the derivative of the period function by proving the \textbf{Bounding Box Theorem}. In particular, we will define a certain curve, and a key step in our would-be proof of the main conjecture is that a portion of this curve should be the graph of a monotonically decreasing function. The main obstacle to proving the whole conjecture is this monotonicity result. We are able to show the desired monotonicity for the group $G_{1/2}$ with our \textbf{Monotonicity Theorem} because we have found an explicit formula for the period function in this case. The group $G_{1/2}$ maximizes scalar curvature in our family, which offers more credence that it is truly a "special case" along with Sol. Thus, our main theorem is a proof of the conjecture for $\alpha=1/2$:
\begin{theorem*}
A geodesic segment in $G_{1/2}$ is a length minimizer
if and only if it is small or perfect.
\end{theorem*}
We have sufficient information about the period function for the group $G_{1/2}$ to extend the main result from \cite{MS}, obtaining a characterization of the cut locus and its consequences for geodesic spheres.
\begin{center}
\textbf{Acknowledgments}
\end{center}
I would first like to thank my advisor, Professor Richard Schwartz, for his encouragement and guidance during the last four years. He has taught me mathematics and a great way of seeing the beauty in mathematics.
I would like to thank all of my professors at Brown University for their openness, friendliness, and erudition in mathematics. I would especially like to express gratitude to Professors Benoit Pausader, Georgios Daskalopoulos, and Jeremy Kahn.
Lastly, I would like to thank my parents for their love and resolute support, ever since I was born.
\chapter{Curve-Shortening Flow on Figure-Eight Curves}
\section{Preliminaries}
Let $C(t): S^1\cross [0,T) \to \mathbb{R}^2$ be a family of immersed curves in the plane. We say that the family is evolving according to curve shortening flow (CSF) if and only if at any $(u,t) \in S^1\cross [0,T)$ we have
$$\frac{\partial C}{\partial t} = k N$$
where $k$ is the curvature and $N$ is the unit normal vector of the immersed curve $C(\cdot,t)$. A curvature function $k$ determines a plane curve up to Euclidean isometries. Thus, it suffices to understand the evolution of the curvature in order to understand the evolution of the curve.
The partial differential equation governing the evolution of curvature $k$ is given by:
\begin{equation}\frac{\partial k}{\partial t} = \frac{\partial^2 k}{\partial s^2} + k^3\end{equation}
where the derivatives in equation $(1)$ are with respect to arc-length. Remembering that the curve is shortening in length, we see that equation $(1)$ is a non-linear PDE. In many applications, however, it is more useful to use a modification of CSF that preserves some geometric quantity. Examples include the area preserving flow, the flow fixing $x$-coordinates, and the flow fixing tangent angles. First let us review the flow fixing $x$-coordinates.
We choose cartesian coordinates in the plane and rotate the evolution to fix $x$-coordinates. This yields a different flow that has the same point-set curves as solutions, but a different evolution equation for curvature. Equations describing this flow are found in the following lemma, proven in \cite{G} inter alia.
\newpage
\begin{lemma}[\cite{G}]
Choose coordinates so that $C(t)$ is locally a graph: $(x,y(x,t))$. If $'$ denotes differentiation w.r.t. $x$ and if $\theta$ is the angle that tangents to the curve make with the $x$ axis (i.e. $\theta(x,t)=\arctan{y'(x,t)}$), then
$$\frac{\partial y}{\partial t} = \frac{y''}{1+y'^2} \quad \textrm{ and }\quad \frac{\partial \theta}{\partial t} =\theta''\cdot \cos^2{\theta}$$
\end{lemma}
\begin{proof}
Since
$$\frac{\partial C}{\partial t} = k N$$
the speed of the evolution of the curve in its normal direction is $k$. Adjusting for the speed in the vertical direction (after having chosen coordinates) gets us
$$\frac{\partial y}{\partial t} = k \sec{\theta}.$$
Away from vertical tangents, we know that
$$k = \frac{y''}{(1+y'^2)^{3/2}} \quad\textrm{ and } \quad \sec{\theta} = \sqrt{1+y'^2}$$
so we have derived the evolution equation for $y$. The equation for $\theta$ is also derived easily and is strictly parabolic away from vertical tangents.
\end{proof}
\begin{figure}[h!]
\centering
\includegraphics[width=1\textwidth]{Figure1}
\caption{The local geometry of the $x$-coordinate preserving modification.}
\end{figure}
We also need to understand how the area enclosed by a curve decays with CSF.
\begin{lemma}[Also found in \cite{G}]
Let $A$ be the area bounded by a curve, the $x$-axis, and two vertical lines that intersect the curve at $x=a$ and $x=b$. Then
$$\frac{dA}{dt}= \int_a^b k\hspace{3pt} ds$$
\end{lemma}
\begin{proof}
This follows by differentiation of an integral:
$$\frac{dA}{dt} = \frac{d}{dt} \int_a^b y\hspace{3pt}dx = \int_a^b \frac{y''}{1+y'^2} dx = \int_a^b k \hspace{3pt}ds.$$
\end{proof}
Another modification to CSF is obtained by rotating to fix tangents to the curve (i.e. fixing $\theta$), which can be done away from inflection points. Using the chain rule, one can show that determining the evolution by CSF is equivalent to solving the following PDE:
\begin{lemma}[See \cite{ANG} or \cite{GH}]
For the CSF modification that fixes tangents, the evolution equation for curvature is
$$\frac{\partial k}{\partial t} = k^2 \frac{\partial^2 k}{\partial \theta^2}+k^3$$
\end{lemma}
We will be particularly interested in the evolution by CSF of the simplest non-embedded plane curves: figure-eights.
\begin{definition}
A smooth curve $C$ immersed in the plane is a \textbf{Figure-Eight} if and only if
\begin{itemize}
\item It has exactly one double point.
\item It has total rotation number $\int_C k ds =0$.
\end{itemize}
\end{definition}
\begin{definition}
A figure-eight curve $C$ is called \textbf{Balanced} if and only if its lobes bound regions of equal area.
\end{definition}
\begin{definition}
A balanced figure-eight curve is termed \textbf{Concinnous} if and only if
\begin{itemize}
\item Its double point is its only inflection point.
\item It is reflection symmetric about two perpendicular axes intersecting at its double point.
\end{itemize}
The axis of symmetry that only intersects the curve at its double point is called the \textbf{Major Axis}, and the other is called the \textbf{Minor Axis}.
\end{definition}
As might be expected, more stringent conditions on our figure-eight curves make it easier to prove interesting results, whence the justification for the previous definitions.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Figure3}
\caption{Some typical examples for each class of figure-eight curve}
\end{figure}
\newpage
\section{Balanced Figure-Eights}
Figure-eights that are not balanced evolve until one lobe collapses to a point. On the other hand, balanced figure-eights will evolve until the enclosed area converges to zero \cite{G3}. There are two possibilities here: either the balanced figure-eight converges to a point or to some line segment. It is still unknown whether all balanced figure-eights behave in the same manner.
\begin{conjecture}[Grayson]
Every balanced figure-eight converges to a point under the curve-shortening flow.
\end{conjecture}
For a figure-eight, the function $\theta$ is single-valued, and the image of the figure-eight in the $x\theta$-plane is a closed curve with well defined extrema. Moreover, because the evolution of $\theta$ is strictly parabolic (cf. Lemma 1), $\theta_{max}$ is strictly decreasing and $\theta_{min}$ is strictly increasing. By comparing the rate at which the extrema of $\theta$ approach each other with the diminution of the enclosed area, we can get a quantitative statement about the shape of a balanced figure-eight under CSF. The key to the proof of the main theorem in this section is the oft-employed maximum principle for strictly parabolic evolution equations.
\begin{figure}[H]
\centering
\includegraphics[width=0.55\textwidth]{Figure2}
\caption{The function $\theta(x,t)$ gives us an evolution of closed curves, evolving according to the equation in Lemma 1.}
\end{figure}
\begin{lemma}[The Maximum Principle as used in \cite{G}]
Suppose $F(x,t): [0,\epsilon]\cross [t_0, t_0 +\epsilon] \to\mathbb{R}$ satisfies a strictly parabolic evolution w.r.t time. Assume also that $F$ is not identically zero at $t_0$, $F(x,t_0)\geq 0$ for all $x$, and that $F(0,t), F(\epsilon, t) \geq 0$ for all $t$. Then
$$F(x,t)>0 \quad \textrm{ for all } \quad (x,t) \in (0,\epsilon) \cross (t_0, t_0 + \epsilon]$$
\end{lemma}
\begin{theorem}[Grayson, \cite{G3}]
When $C(0)$ is a balanced figure-eight, the isoperimetric ratio converges to $\infty$ in the singularity.
\end{theorem}
\begin{proof}
We shall use the modification of CSF that preserves $x$-coordinates. We also choose cartesian coordinates so that $\theta_{min}(t=0)$ is zero and the double point is at the origin. Without loss of generality, the initial curve encloses unit area. Now, the interior angle at the double point of our figure-eight is bounded between $0$ and $\pi$, so by Lemma 2, we know that the total area of our curve decays like:
\begin{equation}-4\pi \leq \frac{dA}{dt} \leq -2\pi.\end{equation}
Thus, after $1/(8\pi)$ seconds of evolution, our curve will still have a total area greater than $1/2$.
Suppose that the isoperimetric ratio of our curve at $t=0$ satisfies\\ $L^2/A=L^2 < M$. Then $|x| < \sqrt{M}/2$, and, since the curve is shortening in length, this remains true for all time. In our effort to apply the maximum principle to the evolution equation for $\theta$, we will compare $\theta$ with a function whose range is a subset of $[0,\pi/2)$. For the purposes of this theorem, we consider the solution $f(x,t)$ to the following (linear) initial value problem:
$$\frac{\partial f}{\partial t} = \frac{1}{2}\cdot \frac{\partial^2 f}{\partial x^2} \quad\textrm{and} \quad f(x,0) = \begin{cases} 0, & |x|< \sqrt{M} \\ \pi/4, & |x| \geq M\end{cases}$$
Explicitly, we have
$$f(x,t)=\frac{1}{8} \pi \left(\text{erfc}\left(\frac{\sqrt{M}-x}{\sqrt{2}
\sqrt{t}}\right)+\text{erfc}\left(\frac{\sqrt{M}+x}{\sqrt{2} \sqrt{t}}\right)\right)$$
where $\text{erfc}$ is the complementary error function. The range of $f(x,t)$ is contained in the interval $[0,\pi/4]$ and begins below the graph of $\theta(x,t=0)$. Thus, since the maximum principle applies for $\theta \in [0,\pi/4]$, we get that
$$\theta_{min}(t) > f(0,t)=\frac{\pi}{4} \text{erfc}\left(\frac{\sqrt{M}}{\sqrt{2} \sqrt{t}}\right)$$
for all further time. In particular, after $1/(8\pi)$ seconds, $\theta_{min}$ has increased by at least $\delta=\frac{\pi}{4}\text{erfc}\big(2\sqrt{M\pi}\big)$.
Given our starting, unit area curve, we also suppose that the isoperimetric ratio stays bounded during the evolution, i.e. $\frac{L^2}{A}(t) <M$ for all time. Since the area is converging to zero, the area halves infinitely many times, but we have already shown that every time the area halves, $\theta_{min}$ increases by at least a fixed positive amount $\delta$, which would mean that $\theta_{min}$ converges to $\infty$, an obvious contradiction.
\end{proof}
Before we continue, we remark that our particular choice of $f(x,t)$ for the comparison function was only a matter of convenience. We could just have easily chosen some other step function's evolution, so long as its range were contained in $[0,\pi/2]$.
\section{Concinnous Figure-Eights}
The following is an interesting and useful proposition of Grayson and Angenent, and more can be learned about its proof and consequences in \cite{DHW}.
\begin{proposition}
The limit of a balanced figure-eight curve in the singularity is contained in the closure of the set of (pre-singular) double points.
\end{proposition}
This is of course in the same vein as Grayson's conjecture, for the proposition ensures that the limiting set is either a point or line segment. For us, this also means that balanced figure-eight curves with sufficient symmetry (like our concinnous curves) definitely converge to a point.
\begin{corollary}
Concinnous figure-eight curves converge to a point under the curve-shortening flow.
\end{corollary}
The natural question to ask now: what is the asymptotic shape of the curve? The work done in \cite{HA} guarantees that (any) figure-eight curve cannot evolve in a self-similar manner. Theorem 1, on the other hand, tells us that the isoperimetric ratio blows up in the singularity. Thus, the area preserving flow will have the shape limit to two increasingly long and narrow slits. In this section, we will obtain a precise description of the blow-up set of a concinnous figure-eight by applying a theorem proven by Angenent in \cite{ANG}.
Here we will mainly employ the modification of CSF that fixes tangent directions, previously mentioned in Lemma 3. The key is that, away from the only inflection point, we can parametrize our (strictly convex) concinnous curves by their tangent angle. Also, by symmetry, we need only examine the behavior of one lobe to get information about the other.
Choose cartesian coordinates so that the double point of our concinnous figure-eight curve is at the origin and such that the minor axis of symmetry coincides with the $x$-axis. Now, if $2\alpha(t)$ is the interior angle at the double point, then the domain for $\theta$, our tangent angle parametrization, is
$$D_1(t)= \big(-\alpha(t), \pi +\alpha(t)\big)$$
and the length of the interval $D_1(t)=2\alpha(t)+\pi$ is bounded as:
$$\pi<|D_1(t)| <2\pi.$$
Before we continue, we make precise what we mean by "blow-up set":
\begin{definition}
The \textbf{Blowup Set} is
$$\Sigma = \{ \theta\in D_1 | \lim_{t\to T} k(\theta, t) = \infty\} $$
\end{definition}
Angenent's Theorem is a very precise description of the blowup set whenever we have a collapsing lobe, which occurs, for example, in the case of balanced figure-eights and symmetric cardioids. Angenent colorfully calls the limiting shape a grim-reaper noose, because after rescaling so that $k_{max}(t)\equiv 1$, the blow-up set converges uniformly to a grim-reaper curve (i.e. $k(\theta)= \cos{\theta}$). Quantitatively, we have
\begin{theorem}[Angenent, \cite{ANG}]
If the blowup set $\Sigma$ has length less than $2\pi$, then for some $\beta\in D_1(t)$, we have
$$\Sigma = [\beta-\pi/2, \beta + \pi/2],$$
and
$$\lim_{t\to T} \frac{k(\beta+\phi, t)}{k_{max}(t)}= \cos{\phi}$$
uniformly for $\phi \in [-\pi/2, \pi/2]$.
\end{theorem}
For our concinnous figure-eights, we have an even more precise description of $\Sigma$. Indeed, after choosing cartesian coordinates so that the double point is at the origin and the $x$-axis is the minor symmetry axis of the curve, we know that $k_{max}(t)$ must be attained at $x_{max}(t)$ (i.e. at the end of the curve). The following corollary easily follows:
\begin{corollary}
The blowup set of a concinnous figure-eight curve is $\Sigma = [0, \pi]$ and
$$\lim_{t\to T} \frac{k(\phi,t)}{k_{max}(t)}= \sin{\phi}$$
uniformly for $\phi\in[0,\pi]$.
\end{corollary}
Qualitatively, the far end of a figure-eight becomes a grim-reaper curve (turned sideways) after the appropriate rescaling.
\newpage
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{Figure4}
\caption{We are comparing the flow of a portion of the $\theta$ curve with the red step-function in the proof of Theorem 1. }
\end{figure}
\newpage
\section{The Affine Rescaling}
Professor Richard Schwartz has suggested to us another rescaling of curve-shortening flow that is especially suited for concinnous figure-eights: an "affine" rescaling that keeps the curve in a "bounding box" of side-length $1$. Typical affine transformations do \textit{not} commute with the CSF, unlike isometries of the plane, so we are performing the affine transformation at the end of the evolution and seeing what the result should be. We \textit{cannot} affinely rescale, afterwards continue the CSF, and still get the same result.
In this section, we will use both modifications of CSF. As usual, we choose cartesian coordinates so that the double point is at the origin and the minor axis of symmetry is the $x$-axis. Also, by the symmetry of a concinnous figure eight, we need only study the evolution of the portion of the curve that lies in the positive quadrant (i.e. $1/4$ of the curve). In particular, if we are using the $x-$coordinate preserving flow, we are studying the evolution of a multi-variable function $y(x,t)$ that satisfies
$$\frac{\partial y}{\partial t} = \frac{y''}{1+y'^2}$$
as in Lemma 1, with the initial function $y(x,0)$ corresponding to a concave curve having a vertical tangent at $x_{max}(0)$. The domain of this function is $D(t)=[0, x_{max}(t)]$, the range is $R(t)=[0,y_{max}(t)]$, and we choose the orientation of the curve so that $y''<0$ on $D(t)$. Both $x_{max}(t)$ and $y_{max}(t)$ are monotonically decreasing; we also define $x^*(t)$ to be the function of time that satisfies
$$y(x^*(t), t) =y_{max}(t).$$
Now we can make our affine rescaling explicit. We are dilating the $x$ direction by $1/x_{max}(t)$ and the $y$ direction by $1/y_{max}(t)$, which evidently leads to a curve contained within a unit square bounding box.
In the previous section, we applied Angenent's theorem on collapsing lobes to concinnous figure-eights and saw that the far end converges uniformly to a grim-reaper curve turned sideways. When we perform the affine rescaling on the grim reaper portion of the curve, we get a vertical line-segment because of its position on the far end and because the diameter is bounded in the $y$ direction and unbounded in the $x$ direction. In the framework of this section, this means that
\begin{proposition}
$$\lim_{t\to T} \frac{x^*(t)}{x_{max}(t)}=1.$$
\end{proposition}
What about the rest of the curve (after affine rescaling)? Numerical evidence, communicated to us from Richard Schwartz, suggests that the resulting shape should be shaped like a bow-tie. In one quadrant, this means that our curve limits to two line segments intersecting in the corner of the bounding box. This is our main result in Chapter 2:
\begin{theorem}[Main Claim]
After affine-rescaling, concinnous figure-eights converge to a bow-tie shape.
\end{theorem}
\section{Towards our Main Claim}
The proof in this chapter is incomplete. For example, we need to also show that the results in \cite{ANG} (Angenent only studied strictly convex immersed curves) still apply to the case of the figure-eight. However, we are preparing another paper which will address these issues and provide a complete proof of our main claim.
This next lemma gives us a goal to aim for.
\begin{lemma}
The concinnous figure-eight converges to a bow-tie if
$$\lim_{t\to T}\bigg( y_{max}(t)\cdot\frac{dx_{max}}{dt}(t)+x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)=-\pi.$$
\end{lemma}
\begin{proof}
After applying the affine rescaling and before the blow-up time, our concinnous figure-eight remains convex except at the double point. The limiting curve, after affine rescaling, will still have absolute curvature that is everywhere greater than or equal to zero. By Proposition 2, the far end of the figure-eight becomes a line segment after the affine rescaling (alternatively, it limits to the vertical edge of the bounding box). The only curve from the origin to the corner of the bounding box that encloses an area equal to $1/2$ and has $|k|\geq0$ is the obvious line segment. Thus, to show that our concinnous figure-eight converges to a bow-tie, it suffices to show that
$$\lim_{t\to T} \frac{x_{max}(t)y_{max}(t)}{2A(t)} = 1$$
where $A(t)$ is the area enclosed by one quarter of the curve and the axes of symmetry. Using Lemma 2, we see that
$$\frac{dA}{dt}(t)= -(\alpha(t)+\pi/2)$$
but since the blow-up set $\Sigma= [0,\pi]$, we know $\lim_{t\to T} \alpha(t)=0$.
Finally, by L'Hospital's rule, we have
$$\lim_{t\to T} \frac{x_{max}(t)y_{max}(t)}{2A(t)} = 1$$
if we can show that
$$\lim_{t\to T} \frac{y_{max}(t)\cdot\frac{dx_{max}}{dt}(t)+x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)}{-2(\alpha(t)+\pi/2)} = 1$$
which is equivalent to showing
$$\lim_{t\to T}\bigg( y_{max}(t)\cdot\frac{dx_{max}}{dt}(t)+x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)=-\pi.$$
\end{proof}
We continue by proving some technical lemmas (we are using the $x$-coordinate preserving flow in the next few).
\begin{lemma}
$$\frac{d x_{max}}{dt}(t)=k(x_{max}(t),t) \quad \textrm{ and } \quad \frac{d y_{max}}{dt}(t)=k(x^*(t),t)$$
\end{lemma}
\begin{proof}
By the chain rule:
$$\frac{d y_{max}}{dt}(t)=\frac{d}{dt}y(x^*(t),t)=\frac{\partial y}{\partial x}(x,t) \frac{d x^*}{dt}(t)+\frac{\partial y}{\partial t}(x,t)= y''(x^*(t),t)=k(x^*(t),t)$$
The proof is identical for the derivative of $x_{max}(t)$.
\end{proof}
\begin{lemma}
$$\lim_{t\to T}\frac{\frac{dy_{max}}{dt}(t)}{\frac{dx_{max}}{dt}(t)}=0$$
\end{lemma}
\begin{proof}
By Corollary 2, we know that
$$\lim_{t\to T} \frac{k(x^*(t),t)}{k(x_{max}(t),t)}=0$$
so the result follows immediately from the previous lemma.
\end{proof}
\begin{lemma}
$$\lim_{t\to T} \frac{y_{max}(t)}{x_{max}(t)}=0$$
\end{lemma}
\begin{proof}
Use L'Hospital's rule and the previous lemma.
\end{proof}
\begin{lemma}
$$\lim_{t\to T} \bigg(-y_{max}(t)\cdot \frac{dx_{max}}{dt}(t)\bigg)= \frac{\pi}{2}$$
\end{lemma}
\begin{proof}
Corollary 2 states that the far end of our concinnous figure-eight converges in $C^2$ to a Grim Reaper curve, after rescaling so that the maximum curvature is identically equal to $1$. However, from the same result, we also know
$$k_{max}(t)= \sup_{x\in D(t)} \abs{k(x,t)} = \abs{k(x_{max}(t),t)}.$$
Since the Grim Reaper is turned on its side, has finite vertical diameter equal to $\pi$, and is reflection-symmetric across the minor axis of our figure eight, the desired result follows immediately.
\end{proof}
The rest of the proof will be split into two main parts. In the first part, we will prove that
$$\limsup_{t\to T} \bigg(-x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)\leq\frac{\pi}{2}$$
by employing some integral estimates and our characterization of the Blowup Set in Corollary 2. In the second part we will use the Maximum Principle to show that
$$\liminf_{t\to T} \bigg(-x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)\geq\frac{\pi}{2}.$$
Together with Lemma 5, these results prove our main theorem.
\begin{center}
\textbf{First Part}
\end{center}
To use the full strength of Corollary 2, we need to use the CSF modification that preserves tangent angles. We recall that the evolution equation for curvature is then given by
$$\frac{\partial k}{\partial t} = k^2 \frac{\partial^2 k}{\partial \theta^2}+k^3$$
and the line element is $ds=d\theta/k$. We are now in a position to get integral formulae for $x_{max}(t), y_{max}(t)$, and their derivatives.
\begin{lemma}
$$x_{max}(t) = -\int_{-\alpha(t)}^{\pi/2} \frac{\cos\theta}{k(\theta,t)}d\theta \quad \textrm{ and } \quad y_{max}(t) = -\int_{0}^{\pi/2} \frac{\sin\theta}{k(\theta,t)}d\theta.$$
\end{lemma}
\begin{proof} This is obvious as long as we keep track of how our curve is parametrized by $\theta$ and if we remember that $ds=d\theta/k$.
\end{proof}
\begin{lemma}
$$\frac{dy_{max}}{dt}(t)=k(\theta=0,t)$$
\end{lemma}
\begin{proof}
Directly differentiating the integral equation in Lemma 10 gets us
$$\frac{dy_{max}}{dt}(t)=\int_0^{\pi/2}(k_{\theta \theta} + k)\sin(\theta) d\theta.$$
Since
$$(k_{\theta \theta} + k)\sin\theta=\frac{\partial}{\partial \theta}\big(k_\theta \sin\theta -k\cos \theta\big)$$
we have
$$\frac{dy_{max}}{dt}(t)=\big(k_\theta \sin\theta -k\cos \theta\big)\bigg|_{\theta=0}^{\theta=\pi/2}=k(\theta=0,t)$$
because curvature is maximized at the far end of our figure eight (i.e. at $\theta=\pi/2$).
\end{proof}
We are now equipped to accomplish our goal for Part 1.
\begin{lemma}
$$\limsup_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) \leq\frac{\pi}{2}$$
\end{lemma}
\begin{proof}
Using the two preceding lemmas, we have
$$-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)=\int_{-\alpha(t)}^{\pi/2} \frac{k(0,t)\cos\theta}{k(\theta,t)}d\theta$$
so
$$\abs{-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)}\leq (\pi/2+\alpha(t))\cdot\sup_{\theta\in[-\alpha(t),\pi/2]}\abs{\frac{k(0,t)}{k(\theta,t)}}.$$
From Corollary 2, which describes the blow-up set $[0,\pi]$, we get
$$\lim_{t\to T} \sup_{\theta\in[-\alpha(t),\pi/2]}\abs{\frac{k(0,t)}{k(\theta,t)}}=1$$
therefore, since $\lim_{t\to T} \alpha(t)=0$ as well, we have
$$\limsup_{t\to T} \abs{-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)}\leq \frac{\pi}{2}$$
as desired.
\end{proof}
\begin{center}
\textbf{Second Part}
\end{center}
Recall that our goal for this section is to prove
$$\liminf_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) \geq \frac{\pi}{2}.$$
In what follows, we use the modification of CSF that preserves $x-$coordinates. As usual, this means that we are looking at the evolution of one quarter of the concinnous figure-eight, viewed as the graph of the concave function $y(x,t)$. This function evolves with time according to
$$\frac{\partial y}{\partial t} = \frac{y''}{1+y'^2}.$$
Consider the function
$$Y(x,t):= \frac{y(x,t)}{x_{max}(t)}.$$
Using the quotient rule, we see that $Y(x,t)$ satisfies the following evolution equation:
$$\frac{\partial Y}{\partial t}(x,t) = \frac{Y''(x,t)}{1+y'(x,t)^2}-\frac{Y(x,t)\cdot \frac{dx_{max}}{dt}(t)}{x_{max}(t)}.$$
Away from vertical tangents, the evolution of $Y$ is strictly parabolic. In particular, by applying the Maximum Principle, we see that local maxima of $Y$ are strictly decreasing. Thus:
$$\frac{d}{dt}\bigg(\frac{y_{max}(t)}{x_{max}(t)}\bigg)<0$$
so we get, for all $t$,
$$-x_{max}(t)\frac{dy_{max}}{dt}(t)> -y_{max}(t)\frac{dx_{max}}{dt}(t)$$
and taking the limit:
$$\liminf_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) \geq \lim_{t\to T} \bigg(-y_{max}(t)\frac{dx_{max}}{dt}(t)\bigg)=\frac{\pi}{2}.$$
As a consequence, and using the results from Part 1, we have
$$\lim_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) = \lim_{t\to T} \bigg(-y_{max}(t)\frac{dx_{max}}{dt}(t)\bigg)=\frac{\pi}{2}$$
therefore, by Lemma 5, our Main Theorem is proven. Q.E.D.
\chapter{The Curvature Preserving Flow on Space Curves}
\section{Preliminaries}
We recall the following standard computation:
\begin{lemma}[cf. \cite{BSW}, \cite{I}]
Let $\gamma(t)$ be a family of smooth curves immersed in $\mathbb{R}^3$ and let $X(s,t)$ be a parametrization of $\gamma(t)$ by arc-length. Consider the following geometric evolution equation:
\begin{equation}X_t = h_1 \textbf{T} +h_2 \textbf{N}+h_3 \textbf{B}\end{equation}
where $\{\textbf{T},\textbf{N},\textbf{B}\}$ is the Frenet-Serret Frame and where we denote the curvature and torsion by $\kappa$ and $\tau$, respectively. Let $h_1, h_2, h_3$ be arbitrary smooth functions of $\kappa$ and $\tau$ on $\gamma(t)$. If the evolution is also arc-length preserving, then the evolution equations of $\kappa$ and $\tau$ are
$$\begin{pmatrix} \kappa_t \\ \tau_t \end{pmatrix}=P\begin{pmatrix} h_3 \\ h_1 \end{pmatrix}$$
where $P$ is
$$\begin{pmatrix} -\tau D_s - D_s\tau & D_s^2\frac{1}{\kappa}D_s-\frac{\tau^2}{\kappa}D_s+D_s\kappa\\
D_s\frac{1}{\kappa}D_s^2-D_s\frac{\tau^2}{\kappa}+\kappa D_s& D_s(\frac{\tau}{\kappa^2}D_s+D_s\frac{\tau}{\kappa^2})D_s+\tau D_s+D_s\tau\end{pmatrix}.$$
\end{lemma}
The next theorem follows without difficulty.
\begin{theorem}
Up to a rescaling, a geometric evolution of curves immersed in $\mathbb{R}^3$, as in equation $(3)$, is both curvature and arc-length preserving if and only if its evolution evolution is equivalent to
\begin{equation}
X_t = \frac{1}{\sqrt{\tau}}\textbf{B}.
\end{equation}
\end{theorem}
\begin{proof}
Let $X_t$ be a curvature and arc-length preserving geometric flow. The tangential component of $X_t$ in equation $(3)$ provides no interesting geometric information; it amounts to a re-parametrization of the curve. Thus, we may assume that $h_1=0$, so, since $X_t$ is arc-length preserving, $h_2=0$ as well. Lemma 13 gives us that
$$\kappa_t = -\tau D_s (h_3)-D_s(\tau h_3)=-2\tau D_s (h_3)-h_3D_s(\tau).$$
Since $X_t$ is curvature preserving, we must have $\kappa_t=0$, or
$$D_s(\log{h_3})=D_s(\log{\tau^{-1/2}}).$$
Integrating, we see that
$$h_3=\frac{c}{\sqrt{\tau}}$$
where $c$ is a constant. Therefore, up to a rescaling, the evolution $X_t$ must be precisely as in $(4)$.
\end{proof}
Unfortunately, the evolution in $(4)$ only makes sense if $\gamma(t)$ has strictly positive torsion (or strictly negative torsion, with the flow $X_t=\frac{1}{\sqrt{-\tau}}\textbf{B}$). This motivates the following definition.
\begin{definition}
We call a smooth curve immersed in $\mathbb{R}^3$ a positive curve and only if it has strictly positive torsion.
\end{definition}
Fortunately, there are many interesting positive curves. For example, some knots admit a parametrization with constant curvature and strictly positive torsion, and there exist closed curves with constant (positive) torsion. Henceforth, we will only consider positive curves.
The partial differential equation governing the evolution of $\tau$ follows below:
\begin{lemma}
For the geometric flow given in equation $(4)$, the evolution equations of curvature and torsion are $\kappa_t =0$ and
\begin{equation}
\tau_t = \kappa D_s(\tau^{-1/2})+D_s\bigg(\frac{D_s^2(\tau^{-1/2})-\tau^{3/2}}{\kappa}\bigg)
\end{equation}
\end{lemma}
The equation for torsion is reminiscent of the Rosenau-Hyman family of equations, which are studied, inter alia, in \cite{HEER} and \cite{LW}. The authors of \cite{SR} call the evolution of the torsion, in the case of constant curvature, the \textit{extended Dym equation} for its relationship with the Dym equation (a rescaling and limiting process converts the evolution for torsion to the Dym equation).
As discussed in \cite{BSW}, the condition for the flow $X_t$ from Lemma 13 to be the gradient of a functional is for the Frechet derivative of $(h_3,h_1)$ to be self-adjoint. In general, this does not occur for the flow under our consideration in equation $(4)$, so it cannot be integrable in the sense of admitting a Hamiltonian structure (Indeed, when $\kappa$ is not constant, it is not even formally integrable in the sense of \cite{MSS}). Nevertheless, the case of constant curvature exhibits a great deal of structure, which makes the study of the evolution equation in $(5)$ worthwhile.
\section{Constant Curvature}
When $\kappa$ is constant, we may rescale the curves $\gamma(t)$ so that $\kappa \equiv 1$. In this way, the evolution of torsion becomes
\begin{equation}
\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big)
\end{equation}
\subsection{Equivalence with the m$^2$KDV Equation}
We recall the notion of "equivalence" of two partial differential equations from \cite{CFA}:
\begin{definition}
Two partial differential equations are equivalent if one can be obtained from the other by a transformation involving the dependent variables or the introduction of a potential variable.
\end{definition}
The authors in \cite{CFA} discuss a general method of transforming quasilinear partial differential equation, such as the evolution of $\tau$ in $(6)$, to semi-linear equations. By applying their algorithm, we obtain another proof of the following theorem, first demonstrated in \cite{SR}.
\begin{theorem}[\cite{SR}]
The evolution equation for torsion, in the case of constant curvature, which is given by
$$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big),$$
is equivalent to the m$^2$KDV equation. Thus, it is a completely integrable evolution equation.
\end{theorem}
\begin{proof}
First, let $\tau=v^2$, so that $(6)$ becomes
$$
2vv_t=D_s\bigg(\frac{1}{v}-v^3+D_s^2\big(\frac{1}{v}\big)\bigg)=-\frac{v_s}{v^2}-3v^2v_s-\frac{v_{sss}}{v^2}-\frac{6v_s^3}{v^4}+\frac{6v_sv_{ss}}{v^3}$$
or, in a simpler form:
\begin{equation}
v_t=D_s\bigg(\frac{1}{4v^2}-\frac{3v^2}{4}+\frac{3v_s^2}{4v^4}-\frac{v_{ss}}{2v^3}\bigg)
\end{equation}
A potentiation, $v=w_s$ followed by a simple change of variables $(t\rightarrow -t/2)$ yields
\begin{equation}
w_t=-\frac{1}{2w_s^2}+\frac{3w_s^2}{2}-\frac{3w_{ss}^2}{2w_s^4}+\frac{w_{sss}}{w_s^3}.
\end{equation}
Equation $(8)$ is fecund territory for a pure hodograph transformation, as used, for example, in \cite{CFA}. Let $\tilde{t}=t$, $\xi=w(s,t)$, and $s=\eta(\xi,\tilde{t})$. The resulting equation, after a simple computation, is
\begin{equation}
\eta_{\tilde{t}}=\eta_{\xi\xi\xi}-\frac{3\eta_{\xi\xi}^2}{2\eta_\xi}+\frac{\eta_\xi^3}{2}-\frac{3}{2\eta_\xi}.
\end{equation}
We rename the variables to the usual variables of space and time: $s$ and $t$; in addition, we anti-potentiate the equation by letting $\eta_s=z$. This makes equation $(9)$ equivalent to
\begin{equation}
z_t = z_{sss}-\frac{3}{2}\bigg(\frac{z_s^2}{z}\bigg)_s+\frac{3z^2z_s}{2}+\frac{3z_s}{2z^2}.
\end{equation}
Lastly, if we let $q=\sinh(z/2)$ in equation (10), a transformation also discussed in \cite{CD}, we get the m$^2$KDV equation
\begin{equation}q_t=q_{sss}-\frac{3}{2}\bigg(\frac{qq_s^2}{1+q^2}\bigg)_s+6q^2q_s\end{equation}
which finishes the proof of the theorem.
\end{proof}
Equations $(6)$ and $(7)$ give us the first two integrals of motion of this flow:
$$\int \sqrt{\tau} ds \textrm{ and } \int \tau ds.$$
The rest can be found by pulling back the m$^2$KDV invariants; we note that these invariants were obtained in a different way by the authors of \cite{SR}. Our next step is to prove the long term existence of our geometric flow using Theorem 5.
\begin{corollary}
Let $\tau_0\in C_{per}^\infty([0,2\pi])$ be a strictly positive and periodic function. We can solve the evolution equation for torsion with initial data $\tau_0$ and get the unique solution $\tau(t)\in C_{per}^\infty([0,2\pi])$ that is strictly positive for all time $t\geq 0$.
\end{corollary}
\begin{proof}
In order to prove this corollary, it suffices to show that we can reconstruct the solution of $(6)$ by solving the m$^2$KDV equation $(11)$ instead. The long term existence for solutions to the m$^2$KDV equation will imply the same for $(6)$, so all that remains to prove is that every differential transformation we used in the proof of Theorem 5 is truly "invertible" in the sense that no singularities arise. We use the same notation for our differential transformations as before.
For the proof of this corollary, it is more convenient to work with an equivalent form of the m$^2$KDV equation called the Calogero-Degasperis-Fokas (CDF) equation \cite{CFA}, obtained by letting $z=e^u$ in equation $(10)$:
$$
u_t=u_{sss}-\frac{u_s^3}{2}+\frac{3u_s}{2}\big(e^{2u}+e^{-2u}\big).
$$
Beginning with our initial torsion, $\tau_0$, we let $v_0=\sqrt{\tau_0}$ and then
$$w_0(s)=\int_0^s v_0(\tilde{s})d\tilde{s}.$$
By construction $w_0\in C^\infty([0,2\pi])$, is not periodic, but satisfies $w_0'(s)>0$ for all $s\in[0,2\pi]$. Thus, $w_0$ has a global inverse $\eta_0(\xi)$ and we also have $\eta_0'(\xi) := z_0(\xi)>0$ for all $\xi$ in the domain of $\eta_0$, which is some interval of the form $[0,M]$. Finally, let $u_0=\log(z_0)$, which is well defined and in $C^\infty([0,M])$ because $z_0>0$ and $z_0\in C^\infty([0,M])$. The CDF equation is globally well-posed for smooth functions on a compact interval with periodic boundary conditions (see \cite{CD}, \cite{HEER}), so we get a solution $u(\xi,t)$ defined for all time with $u(\xi,0)=u_0(\xi)$ and $u(0,t)=u(M,t)$ for all $t$. Our next step is to reconstruct the solution for $(6)$ using $u(t)$.
First, let $z(t)=e^{u(t)}$, which will be a strictly positive, periodic, $C^\infty$ function for all time. Then, let
$$\eta(\xi,t)=\int_0^\xi z(\zeta, t) d\zeta \quad \textrm{ for all } \xi\in[0,M]$$
which, since $\int z(\zeta,t) d\zeta$ is a time-independent quantity, satisfies
$$\eta(0,t)=0 \quad \textrm{and}\quad \eta(M,t)=2\pi$$
for all time. Obviously $\eta_\xi(\xi, t)>0$ for all $(\xi,t)$, so the hodograph transformation $\tilde{t}=t$, $\xi=w(s,t)$, and $s=\eta(\xi,\tilde{t})$ makes sense. We recover a $C^\infty$ function $w(s,t)$ with the property that
$$w(0,t)=0 \quad\textrm{and} \quad w(2\pi,t)=M$$
for all time, and up to a trivial change of coordinates, we have
$$0<v(s,t)=\frac{\partial w}{\partial s}(s,t_0)\in C^\infty_{per}([0,2\pi])$$
for any fixed time $t_0\geq 0$. Lastly, $\tau(s,t)=v(s,t)^2$ solves equation $(6)$, is in $C^\infty_{per}([0,2\pi])$, and satisfies $\tau(s,0)=\tau_0(s)$. Thus, we have reconstructed our desired solution of $(6)$ using the completely integrable CDF equation instead.
\end{proof}
\subsection{Stationary Solutions}
Helices, with constant curvature and constant torsion, are the obvious stationary solutions. In what follows, we find the rest.
\begin{theorem}
The stationary solutions of
$$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big)$$
are given by the following integral formula
$$\int\frac{du}{\sqrt{C+2Au-u^2-u^{-2}}}=s+k$$
where $\tau(s)=u(s)^{-2}$ and where $A,k$ and $C$ are appropriate real constants. When $A=0$ we get an explicit formula for the solutions:
$$\tau(s)=\frac{2}{C\pm\sqrt{(-4+C^2)}\cdot\sin(2(s+k))},$$
with $k$ and $C\geq 2$ real constants.
\end{theorem}
\begin{proof}
Let $\tau=u^{-2}$, then, after integrating once, we must examine the following ordinary differential equation (where $A$ is a constant):
\begin{equation}
A=u-\frac{1}{u^3}+D_s^2(u)
\end{equation}
Since equation $(12)$ is autonomous, we may proceed with a reduction of order argument. Let $w(u)=D_s(u)$ so that $D_s^2(u)=wD_u(w)$ by the chain rule. This substitution gives us the first order equation:
\begin{equation}
wD_u(w)=A-u+\frac{1}{u^3}
\end{equation}
or
$$D_u(w^2)=2A-2u+\frac{2}{u^3}.$$
Integrating, we get
$$D_s(u)=w(u)=\sqrt{C+2Au-u^2-u^{-2}}$$
which is a separable differential equation. So, the stationary solutions of equation $(6)$ are given by the following integral formula:
\begin{equation}
\int\frac{du}{\sqrt{C+2Au-u^2-u^{-2}}}=s+k
\end{equation}
for appropriate constants $C, A, \textrm{ and } k$.
It would be pleasant to have explicit solutions, and this occurs in the case when $A=0$, which is more easily handled. Equation $(13)$ above becomes
$$D_s(u)=w(u)=\sqrt{C-u^2-u^{-2}}$$
which is a differential equation that can be solved with the aid of Mathematica or another computer algebra system. The result is
\begin{equation}
u(s)=\sqrt{\frac{C\pm\sqrt{(-4+C^2)}\cdot\sin(2(s+k))}{2}}
\end{equation}
Where $C,k$ are real constants and $C\geq 2$. The corresponding torsion is:
$$\tau(s)=\frac{2}{C\pm\sqrt{(-4+C^2)}\cdot\sin(2(s+k))}$$
\end{proof}
Integrating $\tau$ and $\kappa$ as above using the Frenet-Serret equations will yield the corresponding stationary curves, up to a choice of the initial Frenet-Serret frame and isometries of $\mathbb{R}^3$.
\subsection{$L^2(\mathbb{R})$ Linear Stability of Helices}
First, we derive the linearization of the evolution for torsion around the stationary solutions corresponding to helices. The linearization of equation $(6)$ at any stationary solution $\tau_0$ is obtained by letting $\tau(s,t) = \tau_0(s,t)+\epsilon w(s,t)$, substituting into equation $(6)$, dividing by $\epsilon$ and then taking the limit as $\epsilon\rightarrow 0$. This is nothing more than the Gateaux derivative of our differential operator at $\tau_0$ in the direction of $w$. Alternatively, one may think of $w$ as the first-order approximation for solutions of $(6)$ near the stationary solution. We can perform this operation when $\tau_0$ is given by an explicit formula, but for brevity's sake, we only mention here the linearization around helices when $\tau_0$ is constant. A short calculation yields
\begin{proposition}
The linearization of the evolution equation $(6)$ around the stationary solutions of constant torsion is
\begin{equation}w_t+2w_s+\frac{1}{2}w_{sss}=0.\end{equation}
\end{proposition}
In what follows, we show the $L^2(\mathbb{R})$ linear stability of the the constant torsion stationary solution. First we recall the definition of linear stability:
\begin{definition}
A stationary solution $\phi$ of a nonlinear PDE is called $L^2(\mathbb{R})$ linearly stable when $v = 0$ is a stable solution of the corresponding linearized PDE with respect to the $L^2(\mathbb{R})$ norm and whenever $v_{t=0}$ is in $L^2(\mathbb{R})$.
\end{definition}
To work towards this, we need to use test functions from the Schwartz Space $\mathcal{S}(\mathbb{R})$, so we first recall that
$$\mathcal{S}(\mathbb{R}):=\{ f\in C^{\infty}(\mathbb{R}) \textrm{ s.t. } \sup_{x\in\mathbb{R}} (1+\abs{x})^N\abs{\partial^{\alpha}f(x)}<\infty, \textrm{ for all } N, \alpha \}.$$
Intuitively, functions in $\mathcal{S}(\mathbb{R})$ are smooth and rapidly decreasing. For more on distributions and the Schwartz Space, review \cite{F}.
The rest of this section is devoted to proving:
\begin{theorem}
Helices correspond to $L^2(\mathbb{R})$ linearly stable stationary solutions of $$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big).$$
\end{theorem}
\begin{proof}
Helices correspond to the constant torsion stationary solutions, so we analyze the linearized PDE in (16):
$$w_t+2w_s+\frac{1}{2}w_{sss}=0.$$
Henceforth, $w_0(s)$ denotes the initial data and $w(s,t)$ denotes the respective solution to the above, linear PDE. Thus, to prove the theorem, it suffices to show: for every $\epsilon$, there exists a $\delta$ such that if $w_0\in L^2(\mathbb{R})$ and $\|w_0\|_2<\delta$, then $\|w(t)\|_2<\epsilon$ for all $t\geq0$.
We follow the standard process of finding weak solutions to linear PDEs via the Fourier transform. Moreover, we know by the Plancherel Theorem that we can extend the Fourier transform by density and continuity from $\mathcal{S}(\mathbb{R})$ to an isomorphism on $L^2(\mathbb{R})$ with the same properties. Hence, it suffices to prove the desired stability result for initial data in $\mathcal{S}(\mathbb{R})$.
Let $F_t(\xi)=e^{4 i \pi ^3 \xi^3 t-4 i \pi \xi t}$. We notice that since $F_t$ is a bounded continuous function for all $t\geq 0$, it can be considered a tempered distribution (or a member of $\mathcal{S}'(\mathbb{R})$, the continuous linear functionals on $\mathcal{S}(\mathbb{R})$), so its inverse Fourier transform makes sense.
Indeed, we can let $B_t(s)=\mathcal{F}^{-1}(F_t(\xi))\in \mathcal{S}'(\mathbb{R})$ and, again, we denote $w_0\in \mathcal{S}(\mathbb{R})$ to be our initial data. Let
$$w(t)=B_t \ast w_0$$
so that $w(t)$ is a $C^{\infty}$ function with at most polynomial growth for all of its derivatives (see \cite{F}). Moreover, $w(t)$ satisfies equation $(16)$ in the distributional sense, as can be checked by taking the Fourier Transform. Lastly, since the Fourier Transform is a unitary isomorphism, it follows that
$$\lim_{t\rightarrow 0} w(t) = w_0$$
in the distribution topology of $\mathcal{S}'(\mathbb{R})$. Hence, $w(t)$ is the weak solution to equation $(16)$ with initial data $w_0\in\mathcal{S}(\mathbb{R})$. In our final step, we use the Plancherel Theorem and the fact that $\mathcal{F}(B_t)=F_t$ is a continuous function with $\|F_t\|_\infty =1$ for all $t\geq0$ to get:
$$\|w(t)\|_2=\|B_t \ast w_0 \|_2 =\|F_t \cdot \mathcal{F}(w_0)\|_2\leq \|F_t\|_\infty \|\mathcal{F}(w_0)\|_2= \|\mathcal{F}(w_0)\|_2=\|w_0\|_2.$$
From the inequality above, the desired $L^2(\mathbb{R})$ stability for initial data in $\mathcal{S}(\mathbb{R})$ follows forthrightly.
\end{proof}
\subsection{Numerical Rendering of a Stationary Curve}
We provide here the figures obtained from a numerical integration of the Frenet-Serret equations on Mathematica for the following choice of torsion:
$$\tau_1(s)=\frac{2}{3+\sqrt{5}\sin(2s)}$$
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{PLOTTOR}
\caption{This is the graph of the torsion $\tau_1$ over two periods.}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{XYPROJ}
\caption{This is the projection of the curve corresponding to $\tau_1$ into the $xy$ plane. The projections into the other planes look very similar.}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.35\textwidth]{TOPDOWN}
\caption{This is a top-down view of the same curve, now exhibiting an almost trefoil shape.}
\end{figure}
\section{Numerical Solutions to the Torsion Evolution Equation}
Mathematica offers a built-in function, \textbf{NDSolve}, which can numerically solve both ordinary and partial differential equations. In our case, the PDE in question is a highly non-linear, time-dependent evolution, and Mathematica will usually employ the "method of lines" to get a numerical solution. The method of lines involves a discretization of the spatial variable that reduces the problem to one of integrating a system of ODE's, and its employment in numerical integration of PDE's is documented for at least the past five decades \cite{SC}.
The benefits of examining numerical solutions are manifold. First, we are able to try different initial data for torsion and see if the evolution behaves nicely; second, we can observe whether our stationary solutions demonstrate stability by slightly perturbing the initial data; third, we can experiment with the case of non-constant curvature, which is difficult to investigate with analytic methods. In this section, we pursue all three directions with several examples in each case.
\subsection{Constant Curvature with Different Choices for Initial Torsion}
In this subsection, the curvature of our positive curves will always be identically equal to $1$. In our experiments, we have observed that initial torsion that is "too close" to zero leads to numerical problems (as might be expected), so we tend to choose our torsion to be at least $\gg1$. Our first example is the evolution of $2\pi$ periodic, smooth initial data that is quite far from vanishing anywhere:
$$\tau(s,t=0)=10+\frac{\sin(s)}{2}$$
We numerically solve equation $(6)$ for $\tau(s,t)$ and plot the graph of the result.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{FirstTor}
\caption{The is the graph for $(s,t)\in [0,2\pi] \cross [0,5]$}
\end{figure}
\newpage
Already, we see a wave structure traveling along some direction (not parallel to the axes) while preserving its structure, for the most part. Quasi-periodicity is behavior typical of integrable partial differential equations, and this is evident for our flow as well. We direct the reader to the following figure, which depicts how our initial sine-wave is (almost) seen again at the times $t= 1.32, 2.64, 3.96$.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{FirstTorWave}
\caption{The is the graph of $\tau$ at times $t=0, 1.32, 2.64,$ and $3.96$.}
\end{figure}
Our next example is the evolution of the following initial torsion data:
$$\tau(s,t=0) = 10 +\sin{s}+\cos{s}.$$
Once again, we numerically solve and plot the graph of the result.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{SecondTor}
\caption{The is the graph for $(s,t)\in [0,2\pi] \cross [0,4]$}
\end{figure}
Although we can once again see the propagation of a wavefront, in this case the solution at fixed times seems to split into two different waves with different velocities.
If we plot $\tau$ at the quasi periods $t=0, 1.26, 2.53,$ and $3.685$, we get the subsequent graphs:
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{SecondTorWave}
\caption{The is the graph of $\tau$ at times $t=0, 1.26, 2.53,$ and $3.685$.}
\end{figure}
We may conclude that the evolution equation $(6)$ demonstrates some typical behavior for non-linear wave equations.
\subsection{Numerical Evidence for the Stability of Helices}
In Section 2.3 of this chapter, we have proven that helix stationary solutions of the equation $(6)$ are $L^2(\mathbb{R})$ linearly stable. In this section, we will examine small perturbations of the helix solutions and observe that there is some evidence that helices have non-linear stability as well.
To that end, let
$$\tau_0(s)=1+\frac{\sin(s)}{100}$$
and let $\tau(s,t)$ be the solution to the initial value problem with $\tau(s,0)=\tau_0(s)$. Let $\| \cdot \|_2$ be the usual norm on $L^2(\mathbb{T})=L^2([0,2\pi]/_{per})$. The function $\tau_0$ adds only a small initial perturbation to the helix solution $\tau \equiv 1$. Indeed, we can compute
$$\| \tau(s,0)-1\|_2 \approx 0.0177245$$
Now we define the function
$$S(t) := \| \tau(\cdot,t)-1\|_2$$
which measures how much our solution for $\tau$ will deviate from the the constant torsion of a helix. Numerical integration of our (numerical) solution for $\tau(s,t)$ allows us to graph $S(t)$ and see whether the solution is non-linearly stable in $L^2(\mathbb{T})$ norm.
Below is an approximation of $S(t)$ for $t\in [0,50]$:
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{Stable}
\end{figure}
Barring numerical errors that occur when integrating differential equations over long time intervals, it appears that the solution $\tau(s,t)$ stays relatively close to $1$ in $L^2(\mathbb{T})$ norm, since $S(t)\leq S(0)$ for most of the time interval. We can also glean that asymptotic stability is unlikely, since it appears that the difference, $S(t)$, does not decrease to zero in time. We might conjecture that helix solutions are non-linearly stable, but the evidence for this is certainly not definitive.
\subsection{Solutions with Non-Constant Curvature}
We recall equation $(5)$, which is the evolution equation for the torsion with arbitrary curvature:
$$\tau_t = \kappa D_s(\tau^{-1/2})+D_s\bigg(\frac{D_s^2(\tau^{-1/2})-\tau^{3/2}}{\kappa}\bigg)$$
Our goal in this subsection is to numerically integrate the above inhomogeneous PDE for a certain choice of curvature and initial torsion. We will choose our curvature to be periodic, $\kappa(s) = 2+\cos{s}$ and our initial torsion will be $\tau(s,0)=1+\frac{\sin{s}}{10}$. To observe how the inhomogeneous PDE differs from the PDE with constant coefficients, we also compare the solution found with the same initial torsion but with constant curvature equal to $2$ (the average value of our choice of $\kappa(s)$ over the interval $[0,2\pi]$). The results of numerically solving the PDE are depicted on the next page.
Obtaining analytic results on the inhomogeneous PDE is obstructed by the inconvenient algebraic nature of the PDE (e.g. it is not even formally integrable). It would be interesting to tackle the inhomogeneous case, especially if one could obtain such results regardless of the choice of curvature.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{Alone}
\caption{The graph of the numerical solution for $\tau(s,t)$ with $(s,t)\in[0,2\pi]\cross[0,1]$.}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{Together}
\caption{The graphs of the numerical solutions for $\tau(s,t)$ with $(s,t)\in[0,2\pi]\cross[0,1]$ and with the blue graph corresponding to the solution with constant curvature equal to $2$. Observe that the solution to the inhomogeneous PDE more rapidly differs from the initial torsion.}
\end{figure}
\section{Computer Code}
The graphs in this chapter were rendered in Mathematica, which we also used to numerically solve the PDE. The code we typically used to perform the latter is similar to the following:
\begin{lstlisting}[language=Mathematica]
NDSolve[{D[ta[s, ti], ti] == (Cos[s] + 2)*D[ta[s, ti]^(-1/2), s] +
D[(D[ta[s, ti]^(-1/2), {s, 2}] - ta[s, ti]^(3/2))/(Cos[s] + 2),
s], ta[s, 0] == 1 + Sin[s]/10,
ta[0, ti] == ta[2 Pi, ti]}, {ta}, {s, 0, 2 Pi}, {ti, 0, 1},
MaxStepSize -> 0.01]
\end{lstlisting}
Afterwards, we can recover the associated curve (after choosing initial conditions) by integrating the Frenet-Serret equations like in:
\begin{lstlisting}[language=Mathematica]
Manipulate[
eqns = {t'[s] == \[Kappa][s] n[s],
n'[s] == -\[Kappa][s] t[s] - \[Tau][s] b[s],
b'[s] == \[Tau][s] n[s], r'[s] == t[s], t[0] == t0, n[0] == n0,
b[0] == b0, r[0] == r0};
\[Kappa][s_] := Sin[s] + 5;
\[Tau][s_] := ta[s, ti] /. s1;
{t0, n0, b0} = Orthogonalize[{{1, 0, 0}, {0, 1, 0}, {0, 0, 1}}];
r0 = {0, 0, 0};
sol = First@NDSolve[eqns, {r, t, n, b}, {s, 0, 2 Pi}];, {ti, 0, 4}]
\end{lstlisting}
\chapter{An Interpolation from Sol to Hyperbolic Space}
\section{The Basic Structure}
In this section, we collect some basic facts about all of the $G_\alpha$ groups. The main idea is that a fruitful way of analyzing the geometry of these Lie groups is to first understand the geodesic flow, and this is the setting in which we will continue our analysis for the remainder of this paper.
The principal object of our study will be a one-parameter family of three-dimensional Lie groups, whose Lie algebras are of Type VI in Bianchi's classification, as elaborated in \cite{LB}. For ease of computation, we can also construct our groups as certain subgroups of $GL_3(\mathbb{R})$, and to that end we let
$$G_\alpha=\bigg\{ \begin{pmatrix} e^z&0&x \\ 0&e^{-\alpha z}&y\\0&0&1\end{pmatrix}\bigg| x,y,z\in\mathbb{R}\bigg\} \textrm{ for all } -1\leq\alpha\leq1$$
We can consider each $G_\alpha$ as a matrix group or, equivalently, as $\mathbb{R}^3$ with the following group law:
$$(x,y,z)\ast(x',y',z')=(x'e^z+x, y'e^{-\alpha z}+y, z'+z).$$
Then, $\mathbb{R}^3$ with this group law has the following left-invariant metric
$$ds^2=e^{-2z}dx^2+e^{2\alpha z}dy^2+dz^2.$$
The Lie algebra of $G_\alpha,$ which we denote $\frak{g}_\alpha$, has the following orthonormal basis:
$$\bigg\{ X=\begin{pmatrix} 0&0&1 \\ 0&0&0\\0&0&0\end{pmatrix}\quad Y=\begin{pmatrix} 0&0&0 \\ 0&0&1\\0&0&0\end{pmatrix}\quad Z=\begin{pmatrix} 1&0&0 \\ 0&-\alpha&0\\0&0&0\end{pmatrix}\bigg\}$$
or
\begin{equation}X=e^z \frac{\partial}{\partial x}, Y=e^{-\alpha z}\frac{\partial}{\partial y}, Z=\frac{\partial}{\partial z}.\end{equation}
So, the structure equations are:
\begin{equation}[X,Y]=0\quad [Y,Z]=\alpha Y\quad [X,Z]=-X\end{equation}
The behavior of $\frak{g}_\alpha$ coincides with the Lie algebras of Type VI (in Bianchi's Classification) when $0<|\alpha|<1$ and we have three limiting cases: $\alpha=1,$ which is a Bianchi group of type VI$_{0}$, $\alpha=0,$ which is a Bianchi group of type III, and $\alpha=-1$, which is a Bianchi group of type V. The intermediate cases are not unimodular, unlike the limiting cases, and that may also be of interest. As a consequence of the classification done in \cite{LB}, no two of the Lie algebras $\frak{g}_\alpha$ are isomorphic, hence
\begin{proposition}[Bianchi, \cite{LB}]
No two of the $G_\alpha$ are Lie group isomorphic.
\end{proposition}
One might ask what occurs if we let the parameter $|\alpha|>1$, and the answer is simple. In this case, the Lie algebra will be isomorphic to that of one of our $G_\alpha$ groups with $|\alpha|<1$. Since we restrict ourselves to simply connected Lie groups (indeed, Lie groups that are diffeomorphic to $\mathbb{R}^3$), this means that the corresponding Lie groups are isomorphic as well.
Now we recall an essential fact from Riemannian geometry. Given a smooth manifold with a Riemannian metric, there exists a unique torsion-free connection which is compatible with the metric, which is called the Levi-Civita connection. This can be easily proven with the following lemma, the proof of which may be found in \cite{K}.
\begin{lemma}[Koszul Formula]
Let $\nabla$ be a torsion-free, metric connection on a Riemannian manifold $(M,g)$. Then, for any vector fields $X,Y,$ and $Z,$ we have:
$$2g(\nabla_X Y, Z)= X(g(Y,Z))+Y(g(X,Z))-Z(g(X,Y))+$$
$$+g([X,Y],Z)-g([X,Z],Y)-g([Y,Z],X)$$
\end{lemma}
Using the Koszul Formula, we can easily compute the Levi-Civita connection $\nabla$ for each $G_\alpha$ from equation $(18)$. We get:
\begin{proposition}
The Levi-Civita connection of $G_\alpha$, with its left-invariant metric, is completely determined by
$$\begin{pmatrix} \nabla_X X && \nabla_X Y&&\nabla_X Z\\ \nabla_Y X && \nabla_Y Y&&\nabla_Y Z\\ \nabla_Z X && \nabla_Z Y&&\nabla_Z Z\end{pmatrix}=\begin{pmatrix} Z && 0&&-X\\ 0 && -\alpha Z&&\alpha Y\\ 0 && 0&&0\end{pmatrix}$$
where $\{X,Y,Z\}$ is the orthonormal basis of the Lie algebra, as in $(1)$.
\end{proposition}
The coordinate planes play a special role in the geometry of $G_\alpha.$ Each group (which is diffeomorphic to $\mathbb{R}^3$) has three foliations by the $XZ,$ $YZ,$ and $XY$ planes. It will be worthwhile to compute the curvatures of these surfaces in $G_\alpha$. The sectional curvature can be computed easily from the Levi-Civita Connection, while the extrinsic (Gaussian) curvature and mean curvature are computed using the Weingarten equation. Lastly, for surfaces in a Riemannian 3-manifold we have the relation
\begin{equation} Intrinsic = Extrinsic + Sectional \end{equation}
from Gauss' Theorema Egregium. For details, see the first chapter in \cite{N}. Straightforward computations yield the following proposition and its immediate consequences.
\begin{proposition} The relevant curvatures of the coordinate planes:
\begin{center}
\begin{tabular}{|c c c c c|}
\hline
Plane& Sectional & Intrinsic & Extrinsic (Gaussian) & Mean \\
\hline\hline
XY& $\alpha$ & 0 & $-\alpha$ & $(1-\alpha)/2$ \\
\hline
XZ& -1 & -1 & 0 &0 \\
\hline
YZ& $-\alpha^2$ & $-\alpha^2$ & 0 &0\\
\hline
\end{tabular}
\end{center}
\end{proposition}
\begin{corollary}
The $XY$ plane is a minimal surface (having vanishing mean curvature) in $G_\alpha$ if and only if $\alpha=1$, and the $XZ$ and $YZ$ planes are minimal for all $\alpha$. Also, the $XY$ plane is always a constant-mean-curvature surface.
\end{corollary}
We will later strengthen this corollary and get that the $XZ$ and $YZ$ planes are geodesically embedded.
It can be easily seen that $G_1$ is a model for the Sol geometry, $G_0$ is a model of $\mathbb{H}^2\cross\mathbb{R}$, and $G_{-1}$ is a model of $\mathbb{H}^3,$ or hyperbolic space. We can also compute the Ricci and scalar curvatures as well and we get that the scalar curvature of $G_\alpha$ is $S_\alpha=2\alpha-2-2\alpha^2$. $S_\alpha$ attains its maximum over the family at $S_{1/2}=-3/2$, and the minimum is attained at $S_{-1}=-6$ (as might be expected). Therefore, the group $G_{1/2}$ may also be considered a special case: the member of the interpolation that maximizes scalar curvature. Moreover, $S_\alpha$ is symmetric in the positive side of the family, in the sense that for all non-negative $\alpha$, $S_{\alpha}=S_{1-\alpha}$.
Other self-evident properties of the coordinate planes could be stated, but we let the reader find these himself.
Now, we turn our attention to the geodesic flow of $G_\alpha$. Rather than attempting to derive analytic formulas for the geodesics from the geodesic equation, as done for Sol in \cite{T}, we restrict the geodesic flow to the unit-tangent bundle and consider the resulting vector field. The idea of restricting the geodesic flow to $S(G_1)$ for Sol was first explored by Grayson in his thesis \cite{G2} and then used by Richard Schwartz and the current author to characterize the cut locus of the origin of Sol in \cite{MS}. We recall that the cut locus of a point $p$ in a Riemannian manifold $(M,g)$ is the locus of points on geodesics starting at $p$ where the geodesics cease to be length minimizing.
Now, we extend the previous ideas to the other $G_\alpha$ groups. Indeed, consider $\frak{g}_\alpha$ and let $S(G_\alpha)$ be the unit sphere centered at the origin in $\mathfrak{g}_\alpha.$ Suppose that $\gamma(t)$ is a geodesic parametrized by arc length such that $\gamma(0)$ is the identity of $G$. Then, we can realize the development of $\gamma'(t)$, the tangent vector field along $\gamma$, as a curve on $S(G_\alpha),$ which will be the integral curve of a vector field on $S(G_\alpha)$ denoted by $\Sigma_\alpha.$ We can compute the vector field $\Sigma_\alpha$ explicitly.
\begin{proposition}
For the group $G_\alpha$ the vector field $\Sigma_\alpha$ is given by
$$\Sigma_{\alpha}(x,y,z)= (xz, -\alpha yz, \alpha y^2-x^2)$$
\end{proposition}
\begin{proof}
Since we are dealing with a homogeneous space (a Lie group) it suffices to examine the infinitesimal change of $V=\gamma'(0)=(x,y,z).$ We remark that parallel translation along $\gamma$ preserves $\gamma'$ because we have a geodesic, but parallel translation does not preserve the constant (w.r.t. the left-invariant orthonormal frame) vector field $V=\gamma'(0)$ along $\gamma$. Indeed, the infinitesimal change in the constant vector field $V$ as we parallel translate along $\gamma$ is precisely the covariant derivative of $V$ with respect to itself, or $\nabla_V V.$ Then, our vector field on $S(G_\alpha)$ is precisely:
$$\Sigma_\alpha=\nabla_V (\gamma' - V)=\nabla_V \gamma' -\nabla_V V=-\nabla_V V.$$
We also remark that since $V$ is a constant vector field, $\Sigma_\alpha$ is determined completely by the Levi-Civita connection that we previously computed, and this computation is elementary.
\end{proof}
We remark where the equilibria points of $\Sigma_\alpha$ are, since these correspond to straight-line geodesics. When $0<\alpha\leq 1,$ the equilibria are
$$\bigg(\pm\sqrt{\frac{\alpha}{1+\alpha}},\pm\sqrt{\frac{1}{1+\alpha}},0\bigg) \textrm{ and } (0,0,\pm1).$$
When $\alpha=0$, the set $\{X=0\}\cap S(G_\alpha)$ is an equator of equilibria, and when $\alpha<0,$ the only equilibria are at the poles.
A glance at $\Sigma_\alpha$ gets us our promised strengthening of Corollary 4:
\begin{corollary}
The $XZ$ and $YZ$ planes are geodesically embedded. The $XY$ is never geodesically embedded, even when it is a minimal surface (i.e. for $\alpha=1$).
\end{corollary}
Consider the complement of the union of the two planes $X=0$ and $Y=0$ in $\frak{g}_\alpha$. This is the union of four connected components, which we call \textit{sectors}. Since the $XZ$ and $YZ$ planes are geodesically embedded, we have
\begin{corollary}
The Riemannian exponential map, which we denote by $E$, preserves each sector of $\frak{g}_\alpha$. In particular, if $(x,y,z)\in \frak{g}_\alpha$ is such that $x,y>0,$ then $E(x,y,z)=(a,b,c)$ with $a,b>0$.
\end{corollary}
We will use Corollary 6 often and without mentioning it. We make the following key observation about $\Sigma_\alpha$.
\begin{proposition}
The integral curves of $\Sigma_\alpha$ are precisely the level sets of the function $H(x,y,z)=\abs{x}^\alpha y$ on the unit sphere.
\end{proposition}
\begin{proof}
Without loss of generality, we consider the positive sector. We recall that the symplectic gradient is the analogue in symplectic geometry of the gradient in Riemannian geometry. In the case of the sphere with standard symplectic structure, the symplectic gradient is defined by taking the gradient of the function $H$ (on the sphere) and rotating it 90 degrees counterclockwise. Doing this computation for $H,$ yields $\nabla_{sym}H=x^{\alpha-1}\cdot\Sigma_\alpha(x,y,z).$ Since this vector field is the same as the structure field up to a scalar function, the desired property follows.
\end{proof}
\begin{remark}
$\Sigma_\alpha$ is a Hamiltonian system in these coordinates if and only if $\alpha=1,$ i.e. for Sol ($G_1$)
\end{remark}
We finish this section with a conjecture that, if true, provides some connection between the groups in the $G_\alpha$ family, for $\alpha \in [-1,1]$. We recall that the \textit{volume entropy}, $h$, of a homogeneous Riemannian manifold $(M,g)$ is a measure of the volume growth in $M$. We can define
$$h(M,g):=\lim_{R\rightarrow \infty} \frac{\log(\textrm{Vol } B(R))}{R}$$
where $B(R)$ is a geodesic ball of radius $R$ in $M$. Since $G_{-1}$ is a model of Hyperbolic space and $G_{1}$ is Sol, we know that $h(G_{-1})=2$ and $h(G_{1})=1$ (see \cite{S}). Based on this, we conjecture that:
\begin{conjecture}
$h(G_\alpha)$ is a monotonically decreasing function of $\alpha$ for $\alpha \in [-1,1]$.
\end{conjecture}
\section{The Positive Alpha Family}
In this section, we start to explore the positive $\alpha$ side of the family. These geometries exhibit common behaviors such as geodesics always lying on certain cylinders, spiraling around in a "periodic-drift" manner. A natural way to classify vectors in the Lie algebra is by how much the associated geodesic segment under the Riemannian exponential map spirals around its associated cylinder. This classification allows us to discern how the exponential map behaves with great detail.
\subsection{Grayson Cylinders and Period Functions}
More than a few of our theorems in Section 2 may be considered generalizations of results in \cite{G2} and \cite{MS}. To begin our analysis, we study the \textit{Grayson Cylinders} of the $G_\alpha$ groups.
\begin{definition}
We call the level sets of $H(x,y,z)=|x|^\alpha y$ that are closed curves \textbf{loop level sets}.
\end{definition}
The proof of the following theorem follows a method first used in \cite{G2} for Sol.
\begin{theorem}[The Grayson Cylinder Theorem]
Any geodesic with initial tangent vector on the same loop level set as
$$\bigg(\beta\sqrt{\frac{\alpha}{1+\alpha}}, \frac{\beta}{\sqrt{1+\alpha}}, \sqrt{1-\beta^2}\bigg), \beta\in[0,1]$$
lies on the cylinder given by
$$w^2+e^{2z}+\frac{1}{\alpha}e^{-2\alpha z}=\frac{1+\alpha}{\alpha}\cdot\frac{1}{\beta^2}$$
where $w=x-y\sqrt{\alpha}$. We call these cylinders "Grayson Cylinders".
\end{theorem}
If we consider Grayson Cylinders as regular surfaces in $\mathbb{R}^3$ with the ordinary Euclidean metric, then a simple derivation of their first and second fundamental forms reveals that they are surfaces with Gaussian curvature identically equal to zero. Hence, they are locally isometric to ordinary cylinders, and, because they are also diffeomorphic to ordinary cylinders, Grayson Cylinders are in fact isometric to ordinary cylinders, for all choices of $\alpha$ and $\beta$.
It is easier to gauge the shape of a Grayson Cylinder by looking at its projection onto the planes normal to the line $x-y\sqrt{\alpha}$, or, alternatively, as the implicit plot of a function of the two variables $w$ and $z$ defined in the statement of Theorem 8. It appears that as $\alpha$ is fixed, the Grayson Cylinders limit to two "hyperbolic slabs" as $\beta$ goes to zero. Alternatively, as $\beta$ is fixed and $\alpha$ varies, it appears that one side of the Grayson Cylinder is ballooning outwards. In Figures 1 and 2, we have some examples generated with the Mathematica code provided at the end of this chapter.
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA1}
\caption{$\alpha=1$ and $\beta=1/2$.}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA25}
\caption{$\alpha=1/4$ and $\beta=1/2$.}
\end{subfigure}
\caption{Slices of Grayson Cylinders with varying $\alpha$}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA121}
\caption{$\alpha=1/2$ and $\beta=1/2$.}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA122}
\caption{$\alpha=1/2$ and $\beta=1/5$.}
\end{subfigure}
\caption{Slices of Grayson Cylinders with varying $\beta$}
\end{figure}
We denote loop level sets by $\lambda$. Each loop level set has an associated period, $P_\lambda,$ which is the time it takes for a flowline to go exactly once around $\lambda$, and it suffices to compute the period at one vector in a loop level set to know $P_\lambda$. We can compare $P_\lambda$ to the length $T$ of a geodesic segment $\gamma$ associated to a flowline that starts at some point of $\lambda$ and flows for time $T.$ We call $\gamma$ $\textit{small, perfect,}$ or $large$ whenever $T<P_\lambda,$ $T=P_\lambda,$ or $T>P_\lambda,$ respectively. It seems that this "classification" of geodesic segments in $G_\alpha$ is ideal. For instance, it was shown in \cite{MS} that a geodesic in Sol is length-minimizing if and only if it is small or perfect. We now derive an integral formula for $P_\lambda$ and simplify the integral in two special cases.
\begin{proposition}
Let $\lambda$ be the loop level set associated to the vector
$$V_\beta=\bigg(\beta\sqrt{\frac{\alpha}{1+\alpha}}, \frac{\beta}{\sqrt{1+\alpha}}, \sqrt{1-\beta^2}\bigg),$$
then
$$P_\lambda(\beta) = \int_{-t_1}^{t_0} \frac{2dt}{\sqrt{1-\frac{\beta^2}{\alpha+1}(\alpha e^{2t}+e^{-2\alpha t})}}$$
where $t_0$ and $t_1$ are the times it takes to flow from $V_\beta$ to the equator of $S(G_\alpha)$ in the direction of, and opposite to the flow of $\lambda$, respectively.
\end{proposition}
\begin{remark}
The times $t_0$ and $t_1$ are precisely when the flat flow lines hit the unit circle, or when
\begin{equation}\alpha e^{2t_0}+e^{-2\alpha t_0}=\frac{\alpha+1}{\beta^2} \textrm{ and } \alpha e^{-2t_1}+e^{2\alpha t_1}=\frac{\alpha+1}{\beta^2}\end{equation}
\end{remark}
Stephen Miller helped us to numerically compute the period function for any choice of positive $\alpha$. An explicit formula for the period function in Sol ($G_1$) was derived in \cite{MS} and \cite{T}. It is:
$$P_\lambda(\beta)=\frac{4}{\sqrt{1+\beta^2}}\cdot K\bigg(\frac{1-\beta^2}{1+\beta^2}\bigg)$$
where $K(m)$ is the complete elliptic integral of the first kind, with the parameter as in Mathematica.
A closed-form expression of $P_\lambda$ can also be obtained for $G_{1/2}$. Since elliptic integrals have been studied extensively and many of their properties are well-known, the following expression allows us to analyze $P_\lambda$ more easily.
\begin{corollary}
When $\alpha=1/2,$ or for the group $G_{1/2}$, the period function is given by
$$P_\lambda(\beta)=\frac{4\sqrt{3}}{\beta\sqrt{e^{t_0-t_1}+2e^{t_1}}}\cdot K\bigg(\frac{2(e^{t_1}-e^{-t_0})}{e^{t_0-t_1}+2e^{t_1}}\bigg)$$
where
$$t_0=\log\bigg(\frac{1}{\beta}\cdot\bigg(\frac{1}{(-\beta^3 + \sqrt{-1 + \beta^6})^{\frac{1}{3}}} + (-\beta^3 +
\sqrt{-1 + \beta^6})^{\frac{1}{3}}\bigg)\bigg)$$
and
$$t_1=\log\bigg(\frac{1}{2}\bigg(\frac{1}{\beta^2}+\frac{1}{\beta^4(-2+\frac{1}{\beta^6}+\frac{2\sqrt{-1+\beta^6}}{\beta^3})^{\frac{1}{3}}}+(-2+\frac{1}{\beta^6}+\frac{2\sqrt{-1+\beta^6}}{\beta^3})^{\frac{1}{3}}\bigg)\bigg)$$
\end{corollary}
Here we state an essential fact, which is forthrightly supplied to us by the above expression of $P_\lambda(\beta)$ in terms of an elliptic integral. We have:
\begin{proposition}
$$\frac{d}{d\beta}\big(P_\lambda(\beta)\big)<0$$
when $\alpha=1$ and $\alpha=1/2$. Moreover, for $\alpha=1$,
$$\lim_{\beta\to 1} P_\lambda(\beta)=\pi\sqrt{2}$$
and for $\alpha=1/2$,
$$\lim_{\beta\to 1} P_\lambda(\beta)=2\pi.$$
\end{proposition}
We could not find a similar formula for $P_\lambda$, when $\alpha$ is not $1$ or $1/2$, in terms of elliptic integrals or hypergeometric functions. This is unfortunate, as Proposition 10 is vital to the method here and in \cite{MS} to characterize the cut locus of $G_1$ and $G_{1/2}$. However, we can still numerically compute the period function as presented at the end of this chapter, and this allows us to conjecture:
\begin{conjecture} Let $P(\beta)$ be the period function in $G_\alpha$, then
$$\frac{d}{d\beta}\big(P_\lambda(\beta)\big)<0$$
and
$$\lim_{\beta\to 1} P(\beta)=\frac{\pi\sqrt{2}}{\sqrt{\alpha}}.$$
\end{conjecture}
This conjecture would lend something quantitative to the idea that the bad behavior (or the cut locus) of $G_\alpha$ dissipates at infinity as Sol interpolates to $\mathbb{H}^2\cross\mathbb{R}$. With Stephen Miller's Mathematica code, we get the following numerical evidence for Conjecture 3.
\begin{center}
\begin{tabular}{|c c c|}
\hline
$\alpha$ & Numerical Value of $P(\alpha,\beta=.999)$ & $\pi\sqrt{2}/\sqrt{\alpha}$ \\
\hline\hline
0.1 & 14.0792 & 14.0496 \\
\hline
0.2 & 9.94735 & 9.93459 \\
\hline
0.3 & 8.11985 & 8.11156 \\
\hline
0.4 & 7.03114 & 7.02481 \\
\hline
0.5 & 6.28842 & 6.28319 \\
\hline
0.6 & 5.7403 & 5.73574 \\
\hline
0.7 & 5.31436 & 5.31026 \\
\hline
0.8 & 4.97106 & 4.96729 \\
\hline
0.9 & 4.68673 & 4.68321 \\
\hline
1. & 4.44622 & 4.44288 \\
\hline
\end{tabular}
\end{center}
\subsection{Concatenation and Some Other Useful Facts}
An important property which extends from Sol to $G_\alpha$ for $0<\alpha<1$ is that the loop level sets are symmetric with respect to the plane $Z=0.$ This simple observation allows the technique of $\it{concatenation},$ essential to the analysis in \cite{MS}, to be replicated for all of the $G_\alpha$ groups, when $\alpha>0$. For the interested reader, Richard Schwartz's Java program \cite{S2} uses concatenation to generate geodesics and geodesic spheres in Sol, and a modified version of this program can generate the spheres and geodesics in any $G_\alpha$ group as well as in other Lie groups, such as Nil. As an illustration of the power of this technique in numerical simulations, we present in Figure 3 the geodesic spheres of radius around $5$ in four different geometries. The spheres are presented from the same angle, the purple line is the $z$ axis, and the red lines are the horizontal axes. The salient phenomenon is that one "lobe" of the sphere is contracting as $\alpha$ goes to zero. Qualitatively, this corresponds to the dissipation of the "bad" behavior (or the cut locus) of $G_\alpha$ as $\alpha$ tends to zero, since the amount of shear diminishes.
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUP1}
\caption{In Sol, or the group $G_1$.}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUP34}
\caption{In the group $G_{3/4}$.}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUP12}
\caption{In the group $G_{1/2}$}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUPH2R}
\caption{In the group $G_0$, or $\mathbb{H}^2\cross \mathbb{R}$.}
\end{subfigure}
\caption{Geodesic spheres of radius around $5$ in four different geometries.}
\end{figure}
Each flowline $\lambda$ of our vector field corresponds to a segment of a geodesic $\gamma$. Let $T$ be the time it takes to trace out $\lambda,$ then $T$ is exactly the length of $\gamma$ since we always take unit-speed geodesics. Let $L_\lambda$ be the far endpoint of $\gamma$ and consider the equally spaced times
$$0=t_0<t_1<\ldots <t_n=T$$
with corresponding points $\lambda_0, \ldots, \lambda_n$ along $\lambda.$ Then we have
$$L_\lambda=\lim_{n\to\infty} (\epsilon_n \lambda_0)\ast\ldots\ast(\epsilon_n \lambda_n),\quad \epsilon_n=T/(n+1)$$
where $\ast$ is the group law in $G_\alpha.$ The above equation is well-defined because the underlying space of both $G_\alpha$ and its Lie algebra is $\mathbb{R}^3.$ We will also use the notation $\lambda=a|b$ to indicate that we are splitting $\lambda$ into two sub-trajectories, $a$ and $b.$
The above equation yields:
$$L_\lambda=L_a \ast L_b$$
when $\lambda=a|b.$
We set $\epsilon_n \lambda_j=(x_{n,j},y_{n,j},z_{n,j})$.
Vertical displacements commute in $G_\alpha,$ therefore the third coordinate of
the far endpoint of $\gamma$ is given by
$$\lim_{n \to \infty} \sum_{j=0}^n z_{n,j}.$$
From this, and the symmetry of the flow lines with respect to $Z=0,$ we get the following lemmas. Since these results are an immediate extension of the results in \cite{MS}, we provide only sketches of their proofs.
\begin{lemma}
If the map $(x,y,z) \to (x,y,-z)$ exchanges the
two endpoints of the flowline $\lambda$ then the endpoints of the geodesic segment $\gamma$
both lie in the plane $Z=0$ and $L_\lambda$ is a
horizontal translation. In this case we call
$\lambda$ {\it symmetric\/}.
\end{lemma}
\begin{proof}
Since $\lambda$ is symmetric, the sum $\lim_{n \to \infty} \sum_{j=0}^n z_{n,j}$ vanishes, so the total vertical displacement is zero.
\end{proof}
\begin{lemma}
If $\lambda$ is not symmetric then we can write
$\lambda=a|b|c$ where $a,c$ are either symmetric
or empty, and $b$ lies entirely above or entirely
below the plane $Z=0$. Since $L_a$ and
$L_c$ are horizontal translations -- or just
the identity in the empty cases -- and
$L_b$ is not such a translation, the endpoints of $\lambda$ are
not in the same horizontal plane.
\end{lemma}
\begin{lemma} If $\lambda=a|b$, where both $a$ and $b$ are
symmetric, then both endpoints of
$\gamma$ lie in the plane $Z=0$. We can do this whenever
$\lambda$ is one full period of a loop
level set. Hence, a perfect
geodesic segment has both endpoints in the same
horizontal plane.
\end{lemma}
\begin{proof}
Since $\lambda=a|b$, we know that $L_\lambda=L_a\ast L_b$. Since $L_a$ and $L_b$ are horizontal translations by Lemma 16, it follows that $L_\lambda$ stays in the same $Z=0$ plane.
\end{proof}
\begin{lemma} If $\lambda_1$ and $\lambda_2$ are full trajectories of the same
loop level set, then
we can write $\lambda_1=a|b$ and $\lambda_2=b|a$, which leads to
$L_{\lambda_2}=L_a^{-1}L_{\lambda_1}L_a$. Working this out with the group law in $G_\alpha$ gives:
$(a_1 e^{-z}, b_1 e^{\alpha z}, 0)=(a_2, b_2, 0)$, where $(x,y,z)=L_a$ and $(a_i, b_i, 0)=L_{\lambda_i}$.
In particular, we have $a_1^\alpha b_1=a_2^\alpha b_2$ and we call the function (of the flowlines)
$H_\lambda=\sqrt{|a_1^\alpha b_1|}$ the {\it holonomy invariant\/} of
the loop level set $\lambda.$
\end{lemma}
Let $E$ be the Riemannian
exponential map.
We call $V_+=(x,y,z)$ and $V_-=(x,y,-z)$, vectors in the Lie algebra, {\it partners\/}.
The symmetric trajectories discussed in Lemma 16 have endpoints
which are partners.
Note that if $V_+$ and $V_-$ are partners, then
one is perfect if and only if the other one is, because they lie on the same loop level set. The next two facts are generalizations of results from \cite{MS} about Sol, and, in particular, Corollary 8 proves half of our main conjecture.
\begin{theorem}
If $V_+$ and $V_-$ are perfect partners, then
$E(V_+)=E(V_-)$.
\end{theorem}
\begin{proof}
Let $\lambda_{\pm}$ be the trajectory which
makes one circuit around the loop level set starting
at $U_{\pm}$. As above, we write the flowlines as
$\lambda_+=a|b$ and
$\lambda_-=b|a$. Since $V_+$ and $V_-$ are partners, we can take
$a$ and $b$ both to be symmetric.
But then the elements
$L_a$, $L_b$, $L_{\lambda_1}$, $L_{\lambda_2}$
all preserve the plane $Z=0$ and hence mutually commute, by Lemma 16.
By Lemma 19, we have
$L_{\lambda_+}=L_{\lambda_-}$.
But $E(V_{\pm})=L_{\lambda_{\pm}}$.
\end{proof}
\begin{corollary}
A large geodesic segment is not a length minimizer.
\end{corollary}
\begin{proof}
If this is false then, by shortening our geodesic, we
can find a perfect geodesic segment $\gamma$, corresponding
to a perfect vector $V=(x,y,z)$, which is a unique
geodesic minimizer without conjugate points. If $z \not =0$
we immediately contradict Theorem 9.
If $z=0$, we consider the variation,
$\epsilon \to \gamma(\epsilon)$ through same-length
perfect geodesic segments
$\gamma(\epsilon)$ corresponding to
the vector $V_{\epsilon}=(x_{\epsilon},y_{\epsilon},\epsilon)$.
The vectors $V_{\epsilon}$ and $V_{-\epsilon}$ are partners, so
$\gamma(\epsilon)$ and $\gamma(-\epsilon)$
have the same endpoint. Hence, this variation
corresponds to a conjugate point on $\gamma$
and again we have a contradiction.
\end{proof}
The next step is to analyze what happens for small and perfect geodesics. To begin, we point out another consequence of Theorem 9. Let $M$ be the set of vectors in the Lie algebra of $G_\alpha$ associated to small geodesic segments and let $\partial M$ be the set of vectors associated to perfect geodesic segments. Lastly, let $\partial_0 M$ be the intersection of $\partial M$ with the plane $Z=0.$ Since $E$ identifies perfect partner vectors, we have a vanishing Jacobi field at each point of $\partial_0 M.$ However, we still have:
\begin{proposition}
$dE$ is nonsingular in $\partial M - \partial_0 M$
\end{proposition}
Another useful consequence of what we have heretofore shown is that the Holonomy function (defined in Lemma 19) is monotonically increasing. The following result will be useful later, when we analyze the behavior of $E$ on the set of perfect vectors. More precisely, we have:
\begin{proposition}
Let $P$ be the unique period associated to a flowline $\lambda$. We know that the holonomy $H_\lambda$ is an invariant of the flowline, so it is a function of $P$. Moreover, $H_\lambda$ varies monotonically with the flowlines. Explicitly, we have $\frac{dH}{dP}(P)>0$.
\end{proposition}
To begin our analysis of the small geodesic segments, we prove an interesting generalization of the \textit{Reciprocity Lemma} from \cite{MS}. If $V$ is a perfect vector, then $E(V)$ will lie in the $Z=0$ plane by Lemma 18; however, we can get something better.
\begin{theorem}
Let $V=(x,y,z)$ be a perfect vector. There exists a number $\mu\neq 0$ such that $E(V)=\mu(\alpha y,x,0).$
\end{theorem}
\subsection{Symmetric Flowlines}
We now introduce another technique introduced first in \cite{MS}: the emphasis on symmetric flow lines, which is justified by our previous lemmas. We introduce the following sets in $\frak{g}_\alpha$ and $G_\alpha$:
\begin{itemize}
\item Let $M, \partial M \subset \frak{g}_\alpha$ be the set of small and perfect vectors, as previously defined.
\item Let $\Pi$ be the $XY$ plane in $\frak{g}_\alpha$ and $\tilde{\Pi}$ be the $XY$ plane in $G_\alpha$.
\item Let $\partial_0 M=\partial M \cap \Pi$.
\item Let $M^{symm} \subset M$ be those small vectors which correspond to symmetric flowlines.
\item Let $\partial_0 N=E(\partial_0 M)$.
\item Let $\partial N$ be the complement, in $\tilde{\Pi}$, of the
component of $\tilde{\Pi}-\partial_0 N$ that contains the origin.
\item Let $N=G_\alpha-\partial N$.
\item For any set $A$ in either $\frak{g}_\alpha$ or $G_\alpha,$ we denote $A_+$ to be the elements of $A$ in the positive sector, where $x,y>0$.
\end{itemize}
Our underlying goal is to show that $\partial N$ is the cut locus of the origin in $G_\alpha$. This has already been done for Sol ($G_1$) in \cite{MS}, and we shall prove the same for $G_{1/2}$. Thus, although the notation we use here is suggestive of certain topological relationships (e.g. is $\partial N$ the topological boundary of $N$?), we are only able to prove these relationships for $G_{1/2}$ in the present paper. Reflections across the $XZ$ and $YZ$ planes are isometries in every $G_\alpha$, so proving something for the positive sector (where $x,y>0$) proves the same result for every sector. This is useful in simplifying many proofs.
A first step towards proving that the cut locus is $\partial N$ is to show that $$E(M) \cap \partial N =\emptyset,$$ or, intuitively, that the exponential map "separates" small and perfect vectors. The following lemma is a step towards this.
\begin{lemma}
If $E(M) \cap \partial N \not = \emptyset$, then
$E(M_+^{symm}) \cap \partial N_+ \not = \emptyset$.
\end{lemma}
With this lemma in hand, we should analyze the symmetric flowlines in detail in order to prove that $E(M_+^{symm}) \cap \partial N_+= \emptyset$. Symmetric flowlines are governed by a certain system of nonlinear ordinary differential equations. Let $\Theta_P^+$ denote those points in the
(unique in the positive sector) loop level set of period $P$ having
all coordinates positive. Every element
of $M_+^{{\rm symm\/}}$ corresponds to a small symmetric
flowline starting in $\Theta_P^+$.
\newline
\newline
{\bf The Canonical Parametrization:\/}
The set $\Theta_P^+$ is an open arc.
We fix a period $P$ and we set $\rho=P/2$.
Let $p_0=(x(0),y(0),0) \in \Theta_P \cap \Pi$ be the point with $x(0)>y(0)$. The initial value $x(0)$ varies from $\sqrt{(\alpha+1)/\alpha}$ to $1$.
We then let
\begin{equation}
p_t=(x(t),y(t),z(t))
\end{equation}
be the point on $\Theta_P^+$ which we reach after time $t \in (0,\rho)$ by
flowing {\it backwards\/} along the structure field
$\Sigma$. That is
\begin{equation}
\label{backwards}
\frac{dp}{dt}=
(x',y',z')=-\Sigma(x,y,z)=(-xz,+\alpha yz,x^2-\alpha y^2).
\end{equation}
Henceforth, we use the notation $x'$ to stand for $dx/dt$, etc.
\newline
\newline
{\bf The Associated Flowlines:\/}
We let $\hat{p}_t$ be the partner of $p_t$, namely
\begin{equation}
\hat{p}_t=(x(t),y(t),-z(t)).
\end{equation}
We let $\lambda_t$ be the small symmetric
flowline having endpoints $p_t$ and $\hat{p}_t$.
Since the structure field $\Sigma$
points downward at $p_0$, the symmetric flowline $\lambda_t$
starts out small and increases all the way to a perfect
flowline as $t$ increases from $0$ to $\rho$.
We call the limiting perfect flowline $\lambda_{\rho}$.
\newline
\newline
\noindent
{\bf The Associated Plane Curves:\/}
Let $V_t \in M_+^{{\rm symm\/}}$ be the vector
corresponding to $\lambda_t$. (Recall that $E(V_t)=L_{\lambda_t}$)
Define
\begin{equation}
\Lambda_P(t):=E(V_t)=(a(t),b(t),0) \hskip 30 pt t \in (0,\rho].
\end{equation}
These plane curves are in $\tilde{\Pi}$ because they are endpoints of symmetric flowlines, and they will be among our main objects of interest in what follows. In Figure 4, we present a collection of the plane curves (colored blue) for $\alpha=1/2$ with the choice of $x(0)$ varying from $0.6$ to $0.95$ at intervals of $0.05$. We also include the initial value $x_0=1/\sqrt{3}$, which corresponds to the straight geodesic segment in $G_{1/2}$. The black curve is an approximation of $\partial_0 N_+$, or endpoints of perfect flowlines, which are the right-hand endpoints of each $\Lambda_P$ curve.
\begin{figure}[h!]
\centering
\includegraphics[width=1\textwidth]{SYMMPLOT1}
\caption{The image of $\Lambda_P$ over the interval $(0,\rho]$ for varying $x_0$ and $\partial_0 N_+$. }
\end{figure}
\begin{lemma}
\label{endpoint}
$\Lambda_P(\rho) \in \partial_0 N_+$, and $0<b(\rho)<a(\rho)$.
\end{lemma}
We have $E(M_+^{{\rm symm\/}}) \cap \partial N_+=\emptyset$ provided that
\begin{equation}
\label{goal}
\Lambda_P(0,\rho) \cap \partial N_+=\emptyset,
\textrm{ for all periods } P.
\end{equation}
So all we have to do is establish Equation 25. Let $B_P$ be the rectangle in the $XY$ plane with vertices
$$(0,0,0), (0,b(\rho),0), (a(\rho),0,0), \textrm{ and } (a(\rho),b(\rho),0).$$
Our first step in proving Equation 25 is to contain the image of $\Lambda_P$ with the following theorem, which we will prove to be true for each $G_\alpha$ group:
\begin{theorem}[The Bounding Box Theorem]
$\Lambda_P(0,\rho) \subset {\rm interior\/}(B_P)$ for all $P$.
\end{theorem}
The Bounding Triangle Theorem serves a similar role in \cite{MS} for Sol $(G_1)$, but it cannot be generalized to any other $G_{\alpha}$ group. It states that $\Lambda_P(0,\rho)$ is contained inside the triangle with vertices $(0,0,0), (a(\rho),0,0)$, and $(a(\rho),b(\rho),0)$. In Figure 6, we depict the image of a single plane curve $\Lambda_P$ for $\alpha=1/2$ and $x_0=0.99945$, which illustrates the failure of the Bounding Triangle Theorem in the other Lie groups.
Now, if we could also manage to show ${\rm interior\/}(B_P)\cap \partial N_+ =\emptyset$, we would finish proving Equation 25. Since the Bounding Box Theorem is not as powerful as the Bounding Triangle theorem of \cite{MS}, we need more information about $\partial N_0$ to prove Equation 25 than was needed in \cite{MS}. We succeed in performing this second step for the group $G_{1/2}$ by getting bounds on the derivative of the period function (using its expression in terms of an elliptic integral in that case). The necessary ingredient that we get is
\begin{theorem*}[The Monotonicity Theorem]
For $\alpha=1/2$, $\partial_0 N_+$ is the graph of a non-increasing function (in Cartesian coordinates).
\end{theorem*}
\subsection{Proof of the Main Results for $G_{1/2}$}
For $G_{1/2}$, assuming that the Bounding Box and Monotonicity Theorems are true, we can proceed to characterize the cut locus of the identity.
First, we prove equation $(25)$:
\begin{theorem}
For the group $G_{1/2}$ we have, for all $P$,
$$\Lambda_P(0,\rho) \cap \partial N_+ = \emptyset$$
\end{theorem}
\begin{proof}
By the Bounding Box Theorem we know that $\Lambda_P(0,\rho) \subset {\rm interior\/}(B_P)$ for all $P$. By the Monotonicity Theorem, we know that $\partial_0 N_+$ is the graph of a decreasing function in Cartesian coordinates, so we conclude that $\partial N_+$ is disjoint from ${\rm interior\/}(B_P)$ for all $P$. Our desired result holds.
\end{proof}
The above theorem, combined with Lemma 20 gets us:
\begin{corollary}
\label{smallperfect}
$$E(M) \cap \partial N=\emptyset$$
\end{corollary}
The rest of our argument for showing that the cut locus of $G_{1/2}$ is $\partial N$ follows exactly as in \cite{MS}.
Let $E$ be Riemannian exponential map.
Let $\mathcal M$ be the component
of $\partial M_+-\partial_0M_+$ which contains
vectors with all coordinates positive.
Let ${\mathcal N\/}=\partial N_+ - \partial_0 N_+$. We first prove a few lemmas.
\begin{lemma}
The map $E$ is injective on $\mathcal{M}$.
\end{lemma}
\begin{proof}
Let $V_1$ and $V_2$ be two vectors in $\mathcal{M}$ such that $E(V_1)=E(V_2)$. We also let $U_1=E(V_1)$ and $U_2=E(V_2)$ and denote the $j^{th}$ coordinate of $U_i$ as $U_{ij}$ and likewise for $\frac{V_i}{\|V_i\|}$. Since $U_1$ and $U_2$ have the same holonomy invariant and since the holonomy is monotonic with respect to choice of flowline, it follows that $\frac{V_1}{\|V_1\|}$ and $\frac{V_2}{\|V_2\|}$ lie on the same loop level set in $S(G_{1/2})$. Thus, $V_{11}V_{12}^2=V_{21}V_{22}^2$. By the Reciprocity Lemma, and since $U_1=U_2$, we get
$$\frac{V_{12}}{V_{11}}=\frac{V_{22}}{V_{21}}.$$
We can now conclude that $V_{11}=V_{21}$ and $V_{12}=V_{22}$. Since $\|V_1\|=\|V_2\|$, we get $V_1=V_2$, finishing the proof.
\end{proof}
\begin{lemma}
\label{SP0}
$E({\mathcal M\/}) \subset \mathcal N$.
\end{lemma}
\begin{proof}
The map $E$ is injective
on ${\mathcal M\/} \cup \partial_0 M_+$, by the previous lemma. At the same
time, $E(\partial_0 M_+)=\partial_0 N_+$.
Hence
\begin{equation}
\label{alternative}
E({\mathcal M\/}) \subset \Pi - \partial_0 N_+.
\end{equation}
By definition, $\mathcal N$ is one of the components of the
$\Pi-\partial_0 N_+$. Therefore,
since $\mathcal M$ is connected, the image
$E({\mathcal M\/})$ is either contained in $\mathcal N$ or disjoint from $\mathcal N$.
Since the sets are evidently not disjoint (large perfect vectors land far away from the identity and near the line $x=y/\sqrt{2}$), we have containment.
\end{proof}
\begin{corollary}
\label{SP}
$E(\partial M) \cap E(M)=\emptyset$.
\end{corollary}
\begin{proof}
Up to symmetry, every vector in $\partial M$ lies either in
$\mathcal M$ or in $\partial_0 M_+$.
By definition, $E(\partial_0 M)=\partial_0 N \subset \partial N$. So,
by the previous result, we have
$E(\partial M) \subset \partial N$.
By Corollary \ref{smallperfect} we have
$E(M) \cap \partial N=\emptyset$.
Combining these two statements gives the result.
\end{proof}
\begin{theorem}
\label{minimi}
Perfect geodesic segments are length minimizing.
\end{theorem}
\begin{proof}
Suppose $V_1 \in \partial M$ and
$E(V_1)=E(V_2)$ for some $V_2$ with $\|V_2\|< \|V_1\|$.
By symmetries of $G_{1/2}$ and the flowlines, we can assume that both $V_1$ and $V_2$ are in the positive sector and that their third coordinates are also positive.
By Corollary 8, we have $V_2 \in M \cup \partial M$.
By Corollary \ref{SP} we have $V_2 \in \mathcal{M}$.
But then $V_1=V_2$, by Lemma 22, which contradicts $\|V_2\|< \|V_1\|$.
\end{proof}
The results above identify $\partial N$ as the cut locus of the
identity of $G_{1/2}$ just as obtained in \cite{MS} for Sol. We can summarize by saying
\begin{theorem}
A geodesic segment in $G_{1/2}$ is a length
minimizer if and only if it is small or perfect.
\end{theorem}
In addition, small geodesic segments are
unique length minimizers and they have no conjugate points.
Hence, using standard results about the cut locus, as in \cite{N}, we get that $E: M \to N$
is an injective, proper, local diffeomorphism. This implies
that $E: M \to N$ is also surjective and hence a diffeomorphism.
Moreover, $E: \partial_0 M_+ \to \partial_0 N_+$ is a diffeomorphism, by similar considerations. Results about the geodesic spheres in $G_{1/2}$ follow immediately, as in \cite{MS} for Sol, by "sewing up" $\partial M$ in a 2-1 fashion with $E$.
In particular, we have:
\begin{corollary}
Geodesic spheres in $G_{1/2}$ are always topological spheres.
\end{corollary}
The rest of this chapter is devoted to proving our technical results: the Bounding Box Theorem and the Monotonicity Theorem. We will prove the Bounding Box Theorem in full generality, i.e. for all $\alpha\in (0,1]$. However, we only manage to prove the Monotonicity Theorem for $G_{1/2}$, where we have an expression of the period in terms of an elliptic integral. It is our opinion that either an expression for $P$ in terms of hypergeometric functions exists for general $\alpha$ or a thorough analysis of the (novel?) integral function in Proposition 9 can be done to demonstrate the monotonicity results required. Regardless, our Bounding Box Theorem does half of the work necessary to finish the proof of our main conjecture: for all $G_\alpha$ groups, a geodesic segment is length minimizing if and only if it is small or perfect.
We reiterate that the necessary step to prove our conjecture is to show the Monotonicity Theorem holds for general $G_\alpha$ and that there is encouraging numerical evidence supporting this proposition. We plan to investigate this last step and prove our main conjecture in the future.
\begin{center}
\begin{figure}[H]
\vspace{1.5in}
\includegraphics[width=1\textwidth]{PNPlot}
\caption{Here, for $G_{1/2}$, we have plotted points on $\partial_0 N_+$, as $x_0$ varies from $0.6$ to $0.98$ in increments of $0.02$.}
\end{figure}
\end{center}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{BoundingT}
\caption{This depicts a the image of $\Lambda_P$ for $\alpha=1/2$ and $x_0=0.99945$ over the interval $(0,\rho]$.}
\end{figure}
\section{Proof of the Bounding Box Theorem}
We now study the system of ODE's that governs the behavior of $x,y,z,a,$ and $b$ as in equation $(22)$.
We write
$\lambda_{t+\epsilon}=u|\lambda_t|v$,
where $u$ is the flowline
connecting $p_{t+\epsilon}$ to $p_t$ and $v$ is the flowline
connecting $\hat{p}_t$ to $\hat p_{t+\epsilon}$.
We have
$$
(a',b',0)=\Lambda'_P(t)=\lim_{\epsilon \to 0} \frac{\Lambda_P(t+\epsilon)-\Lambda(t)}{\epsilon},
$$
$$
\Lambda_P(t+\epsilon) \approx (\epsilon x, \epsilon y,\epsilon z) *
(a,b,0) * ( \epsilon x, \epsilon y,-\epsilon z).
$$
The approximation is true up to order $\epsilon^2$ and
$(*)$ denotes multiplication in $G_\alpha$.
A direct calculation gives
\begin{equation} a'=2x+az \textrm{ and } b'=2y-\alpha bz.\end{equation}
Simply from its differential equation, it is evident that $a'>0$ on $(0,\rho)$. This implies that
$\Lambda_P(t)$ is the graph of a function for $t\in (0,\rho)$, hence $\Lambda_P(t)$
avoids the vertical sides of $B_P$. This is the first, easy step in proving the Bounding Box Theorem.
To finish the proof, it would suffice to show that $\Lambda_P(t)$ also avoids the horizontal sides of $B_P$, which amounts to proving that $b'(t)>0$ for all $t\in(0,\rho]$. A priori, it is not evident that $b'>0$ in this interval. For example, the function $b$ may start out concave. Also, after the half-period $\rho$, $b'$ may actually be negative. However, the remarkable fact that $b'>0$ in $(0,\rho]$ for all choices of $\alpha$ and $x_0$ is also true, and we demonstrate this fact in what follows.
Once again, we collect the ODE's of interest to us together:
$$x'=-xz\quad y'=\alpha y z\quad z'=x^2-\alpha y^2\quad b'=2y-\alpha b z$$ from which we compute
\begin{equation}z''=-2z(x^2+\alpha y^2)\quad b''=\alpha b(\alpha z^2-z').\end{equation}
\begin{lemma}
To show that $b'>0$ in $(0,\rho),$ it suffices to show that $b'>0$ whenever $b''<0$ in the interval $(0,\rho).$
\end{lemma}
We observe that $z''<0$ everywhere in $(0,\rho)$. Also, whenever $b''<0,$ we have that $z'>\alpha z^2>0$. We first get an inequality regarding the function $z(t):$
\begin{lemma}
$$2\alpha \int_0^t z(s) ds\geq t\alpha z(t)$$
\end{lemma}
We now recall the Log-Convex Version of Hermite-Hadamard, proven first in \cite{GPP}:
\begin{proposition}[\cite{GPP}]
If $f$ is log-convex on $[a,b]$ then
$$\frac{1}{b-a}\int_a^b f(s) ds\leq \frac{f(b)-f(a)}{\log f(b)-\log f(a)}.$$
\end{proposition}
Let's return to one of our initial ODE's: $y'=\alpha yz.$ Dividing by $y,$ integrating, and multiplying by 2, we get:
\begin{equation}2\log y(t) -2\log y(0)= 2\alpha \int_0^t z(s) ds. \end{equation}
Since $z'>0$ whenever $b''<0,$ we get that $y^2$ is log-convex whenever $b''<0.$ We can now get:
\begin{lemma} For all $t$ where $b''(t)<0,$ we have
$$\int_0^t y(s)^2 ds \leq t\cdot\frac{y(t)^2}{2\log y(t)/y(0)}$$
\end{lemma}
We are now in a position to prove the Bounding Box Theorem:
\begin{proof}
By Lemma 24, it suffices to show that $b'>0$ whenever $b''<0.$ If we integrate the ODE for $b$, we get
\begin{equation} b(t)=\frac{2}{y(t)}\int_0^t y(s)^2 ds.\end{equation}
Differentiating, we want to show $y(t)^3-y'(t)\int_0^t y(s)^2 ds\geq 0$ whenever $b''<0.$ Equivalently (we can divide by $y,y'$ since they are always strictly greater than $0$):
$$ \int_0^t y(s)^2 ds\leq \frac{y(t)^3}{y'(t)}.$$
By Lemma 26 it suffices to show
$$t\cdot\frac{y(t)^2}{2\log y(t)/y(0)}\leq\frac{y(t)^3}{y'(t)}$$
or, by cancelling some terms and taking the reciprocal,
$$2\log y(t)/y(0) \geq t\cdot \frac{y'(t)}{y(t)}.$$
Since $y'=\alpha yz$, this is equivalent to
$$2\alpha \int_0^t z(s) ds\geq t\alpha z(t)$$
which is nothing but the inequality of Lemma 25.
\end{proof}
Finally, since $b'>0, \Lambda_P(t)$ is an increasing function, so $\Lambda_P(t)$ avoids the vertical sides of $B_P$ when $t\in(0,\rho)$. This, along with the previously stated fact that $\Lambda_P(t)$ avoids the horizontal sides of $B_P$, finishes the proof of the Bounding Box Theorem.
\section{Proof of the Monotonicity Theorem}
\subsection{Endpoints of Symmetric Flowlines}
Henceforth, we view $\partial_0 N_+$ as the following parametrized curve in the $XY$ plane. Denote $x(0)=x_0$, then
$$\partial_0 N_+= \{ (a_{x_0}(P(x_0)/2),b_{x_0}(P(x_0)/2),0)\}, \textrm{ as } x_0 \textrm{ varies in } \bigg(\sqrt{\frac{\alpha}{1+\alpha}},1\bigg).$$
To prove Equation $25$ in general it would suffice, by using the Bounding Box Theorem, to prove $B_P\cap \partial_+ N=\emptyset$, and, to prove the latter statement, it suffices to show that $\partial_0 N_+$ is the graph of a decreasing function in Cartesian coordinates. This involves differentiating our ODE's with respect to the initial value $x_0$. Let $\bar{x}$ denote $dx(t,x_0)/dx_0$, etc. Then we get
$$\bar{x}'=-x\bar{z}-z\bar{x},\quad \bar{y}'=\alpha y\bar{z}+\alpha z\bar{y},\quad \bar{z}'=2x\bar{x}-2\alpha y \bar{y},$$
$$\bar{a}'=2\bar{x}+a\bar{x}+x\bar{a}, \quad \bar{b}'=2\bar{y}-\alpha\bar{y}b-\alpha y\bar{b}.$$
Since $x^2+y^2+z^2=1$ for all $t, x_0,$ it follows that
\begin{equation}x\bar{x}+y\bar{y}+z\bar{z}=0\end{equation}
always, and we can get a similar equation for the time derivative. Now, we prove some very useful propositions:
\begin{proposition}
$ax-\alpha by = 2z$ for all $t$ and $x_0$.
\end{proposition}
Since the above equality is true for all $t$ and $x_0$ we can differentiate with respect to $x_0$ and get
\begin{corollary}
$a\bar{x}+x\bar{a}-\alpha b\bar{y}-\alpha y \bar{b}=2\bar{z}$ for all $t$ and $x_0$.
\end{corollary}
Now we prove:
\begin{proposition}
$x\bar{a}+y\bar{b}=0$ for all $t$ and $x_0$.
\end{proposition}
Corollary 12 and Proposition 14 combine to get us the following useful expressions for $\bar{a}$ and $\bar{b}$.
\begin{corollary}
We have
$$\bar{a}=\frac{1}{x(1+\alpha)}\bigg(2\bar{z}+\alpha b\bar{y}-a\bar{x}\bigg)$$
and
$$\bar{b}=-\frac{1}{y(1+\alpha)}\bigg(2\bar{z}+\alpha b\bar{y}-a\bar{x}\bigg).$$
\end{corollary}
\subsection{The Monotonicity Theorem for $G_{1/2}$}
To get our desired results about $\partial_0 N_+$, we must look at a particular case of our one-parameter family, where we have more information about the derivative of $P(x_0)$, courtesy of the expression of $P$ in terms of an elliptic integral. Everything we have heretofore shown is, however, applicable to every $G_\alpha$ with $0<\alpha\leq 1$.
Although we restrict ourselves to $G_{1/2}$, the methods presented here could just as well be applied to $G_1$, where we also have an explicit formula for the period function.
In this section, we show that $\partial_0 N_+$ is the graph of a non-increasing function in Cartesian coordinates (for $G_{1/2}$) by using properties of the period function. This finishes the proof of Equation 25, which in turn allows us to prove our main theorem. For encouragement, we refer the reader back to Figure 5, where we can see that $\partial_0 N_+$ is indeed the graph of a non-increasing function in Cartesian coordinates. It will be relatively easy to show that $\partial_0N_+$ is the graph of a Cartesian function, but the proof that $\partial_0N_+$ is a non-increasing function will be more involved. For example, we will first need to show that $\partial_0N_+$ limits to the line $b=4$ as $x_0\to 1$.
Recall that $P(\beta)$ is decreasing with respect to $\beta$ in the case when $\alpha=1$ or $1/2$. Here, we change variables for the period function from $\beta$ to $x_0\in (1/\sqrt{3},1)$ and get:
\begin{proposition}
$$\frac{d}{dx_0}\big(P(x_0)\big) >0$$
\end{proposition}
Since $z$ always vanishes at the half period, we have
\begin{proposition}
For any initial value $x_0$, we have
$$\bar{z}+(\frac{1}{2}\frac{dP}{dx_0})z'=0$$
at the time $t=P(x_0)/2$. Also, since $z'<0$ at the half period, we get that
$$\bar{z}(P(x_0)/2)>0, \quad \forall x_0.$$
\end{proposition}
From equation $(31)$ and the fact that $x'$ and $y'$ always vanish at the half-period, we have
\begin{proposition}
$$\bar{y}(P(x_0)/2)=\frac{1}{2\sqrt{x_0}}>0 \quad \textrm{ and } \quad \bar{x}(P(x_0)/2)=-2x_0<0$$
\end{proposition}
We are ready to get some information about $\partial_0 N_+$, beginning with:
\begin{corollary}
$\partial_0 N_+$ is the graph of a function in Cartesian coordinates.
\end{corollary}
\begin{proof}
This is equivalent to showing that
$$\frac{d}{dx_0}a_{x_0}(P(x_0)/2)>0.$$
The chain rule gets us
$$\frac{d}{dx_0}a_{x_0}(P(x_0)/2)=\bar{a}(P(x_0)/2)+(\frac{1}{2}\frac{dP(x_0)}{dx_0})a'(P(x_0)/2)$$
$$=\bigg(\frac{2}{3x}\bigg(2\bar{z}+\frac{1}{2}b\bar{y}-a\bar{x}\bigg)+(x+az/2)\frac{dP}{dx_0}\bigg)\bigg\rvert_{t=P(x_0)/2}$$
By the previous propositions, we know that all the terms above are positive at $P(x_0)/2$, whence the desired result.
\end{proof}
As with showing that $b'>0$ in the interval $(0,\rho)$, things are more difficult with the function $b$. We need three lemmas first. The proof of the following may also suggest that $\alpha=1/2$ is a "special case"; nevertheless, the situation is different than for Sol. Richard Schwartz proves a similar limit in \cite{S} for Sol (in which case, the limit is $2$), but his method uses an additional symmetry of the flow lines that we cannot use here.
\begin{lemma}For $\alpha=1/2$,
$$\lim_{x_0\to 1} b_{x_0}(P(x_0)/2)=4$$
\end{lemma}
More generally, we have the following conjecture
\begin{conjecture} Let
$$L(\alpha):=\lim_{x_0\to 1} b_{x_0,\alpha}(P(x_0,\alpha)/2), \textrm{ defined for all } \alpha \in (0,1].$$
We conjecture that $L(\alpha)$ is monotonically decreasing from $\alpha=0$ to $\alpha=1$ and that $\lim_{\alpha\to 0} L(\alpha)=\infty$. In fact, we also conjecture that
$$L(\alpha)=\frac{2}{\alpha}.$$
\end{conjecture}
The next technical lemma is quite wearisome to prove. We direct the reader to Figure 7, which provides numerical evidence for its veracity.
\begin{lemma}
For $\alpha=1/2,$ we have
$$G(x_0):=\frac{dP}{dx_0}(x_0)-\pi\bigg(\frac{1}{2\sqrt{x_0}}+\frac{2x_0\sqrt{x_0}}{1-x_0^2}\bigg)<0,\quad \forall x_0 \in \bigg(\frac{1}{\sqrt{3}},1\bigg)$$
\end{lemma}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{DPBOUND}
\caption{The graph of the function $G(x_0)$.}
\end{figure}
The penultimate step towards the Monotonicity Theorem:
\begin{lemma}
$$\frac{d}{dx_0}b_{x_0}(P(x_0)/2)<0, \textrm{ whenever } b_{x_0}(P(x_0)/2)>\pi$$
\end{lemma}
\begin{proof}
By the chain rule we know
$$\frac{d}{dx_0}b_{x_0}(P(x_0)/2)=\bigg(-\frac{2}{3y}\bigg(2\bar{z}+\frac{1}{2}b\bar{y}-a\bar{x}\bigg)+(y-\frac{1}{4}bz)\frac{dP}{dx_0}\bigg)\bigg\rvert_{t=P(x_0)/2}.$$
The $bz$ term is zero at the half period, and we can simplify further. The above is less than zero if and only if
$$(2x^2+2y^2)\frac{dP}{dx_0}<b\bar{y}-2a\bar{x} \textrm{ at } t=P(x_0)/2.$$
However we know $x^2+y^2=1$ at the half period, and we may also employ the Reciprocity Lemma. This yields that our desired result is equivalent to showing
$$\frac{dP}{dx_0}(x_0)<\big(b(\bar{y}-\frac{y}{x}\bar{x})\big)\big\rvert_{t=P(x_0)/2}.$$
By hypothesis, $b_{x_0}(P(x_0)/2)>\pi$, and we can use equation $(31)$. This means it suffices to show
$$\frac{dP}{dx_0}(x_0)<\pi\bigg(\frac{1}{2\sqrt{x_0}}+\frac{2x_0\sqrt{x_0}}{1-x_0^2}\bigg),\quad \forall x_0 \in \bigg(\frac{1}{\sqrt{3}},1\bigg)$$
which is nothing but Lemma 28.
\end{proof}
Now we are ready to prove the Monotonicity Theorem.
\begin{theorem}[The Monotonicity Theorem]
For $\alpha=1/2$, $\partial_0 N_+$ is the graph of a non-increasing function (in Cartesian coordinates).
\end{theorem}
\begin{proof}
This is equivalent to showing
$$\frac{d}{dx_0}b_{x_0}(P(x_0)/2)\leq0$$
or, as in the proof of the previous lemma,
$$\frac{dP}{dx_0}(x_0)\leq\big(b(\bar{y}-\frac{y}{x}\bar{x})\big)\big\rvert_{t=P(x_0)/2}.$$
We now show that $b_{x_0}(P(x_0)/2)\geq4$ always. If this is true, then the theorem follows by Lemma 29. Otherwise, assume that $b_{x_0}(P(x_0)/2)<4$ for some choice of $x_0$. By Lemma 27, $b_{x_0}(P(x_0)/2)$ limits to $4$ as $x_0$ tends to $1$, so $b_{x_0}(P(x_0)/2)$ must eventually be greater than $\pi$. Let $x_0'$ be the last initial value where $b(P/2)\leq \pi$. By Lemma 29, $b_{x_0}(P(x_0)/2)$ will be strictly decreasing in $x_0$, which is a contradiction. Hence, $b>4$ always and the theorem is proven.
\end{proof}
\section{Computer Code}
Here we present the Mathematica code that generates the figures we have presented. In addition, we present the Mathematica code written by Stephen Miller that allows numerical computation of the period function for arbitrary positive $\alpha$. We note that the whole of the program \cite{S2}, used to generate Figure 3, is available online on Richard Schwartz's website.
\subsection{Figures 1 and 2}
This gets us the implicit plots in the plane:
\begin{lstlisting}[language=Mathematica]
Manipulate[
ContourPlot[(1/a) E^(2*a*z) + E^(-2 z) +
w^2 == (1+a)/(a*b^2), {w,-10,10}, {z,-10,10},
PlotRange -> All, AspectRatio -> Automatic, PlotPoints -> 50],
{b,0.001,1}, {a,0.001,1}]
\end{lstlisting}
A 3-D rendering can be obtained with:
\begin{lstlisting}[language=Mathematica]
Manipulate[
ContourPlot3D[(1/a) E^(2*a*z) +
E^(-2 z) + (x - Sqrt[a]*y)^2 == (1 + a)/(a*b^2), {x, -10,
10}, {y,-10,10}, {z,-10,10}, PlotRange -> All], {b,0.01,
1}, {a,0.05,1}]
\end{lstlisting}
\newpage
\subsection{Figure 4 and 5}
The following renders the symmetric flowline associated to any choice of $x_0 \in (\frac{1}{\sqrt{3}}, 1)$ for the $G_{1/2}$ group. Moreover, this provides numerical evidence that the unwieldy expression for the period function in $G_{1/2}$ we presented earlier is indeed correct. Figure 5 just adds the straight line from the origin to the endpoint of the flowline to show that the "Bounding Triangle Theorem" does not work in general.
\begin{lstlisting}[language=Mathematica]
(*This is the period function for the G_{1/2} group,
as in Corollary 3.4 *)
P[B_]:= (*as in the statement of the corollary *)
(*This is the period function,
after the change of variables from \beta to x0*)
L1[x0_] := P[((3*Sqrt[3]/2)*(x0-x0^3))^(1/3)]
(*This numerically solves the fundamental system of ODE's *)
Manipulate[
s1 = NDSolve[{x'[t] == -x[t]*z[t], y'[t] == (1/2)*y[t]*z[t],
z'[t] == -(1/2)*y[t]^2 + x[t]^2, a'[t] == 2*x[t] + a[t]*z[t],
b'[t] == 2 y[t] - (1/2)*b[t]*z[t], z[0] == 0, a[0] == 0,
b[0] == 0, x[0] == x0, y[0] == Sqrt[1 - x0^2]}, {x, y, z, a,
b}, {t, 0, Re[L1[x0]]/2}], {x0, 1/Sqrt[3], 1}]
(*This plots the flowline*)
ParametricPlot[{a[t], b[t]} /. s1, {t, 0,Re[L1[x0]/2]},
PlotRange -> All]
\end{lstlisting}
\newpage
\subsection{Figure 6}
We can use Stephen Miller's program to compute the period for arbitrary alpha, or we can "find" it by inspection, preferably looking at the function $z$.
\begin{lstlisting}[language=Mathematica]
(*This numerically solves our ODE's, d is alpha here*)
Manipulate[
s = NDSolve[{x'[t] == -x[t]*z[t], y'[t] == d*y[t]*z[t],
z'[t] == -d*y[t]^2 + x[t]^2, a'[t] == 2*x[t] + a[t]*z[t],
b'[t] == 2 y[t] - d*b[t]*z[t], z[0] == 0, a[0] == 0, b[0] == 0,
x[0] == c1, y[0] == Sqrt[1 - c1^2]}, {x, y, z, a, b}, {t, 0,
20}], {c1, Sqrt[d/(d + 1)], 1}, {d, 0, 1}]
(*We can use this to find \rho by inspection*)
Manipulate[z[t] /. s, {t, 0,20}]
(*This plots the derivative of b*)
Plot[b'[t] /. s, {t, 0, 15.99}, AspectRatio -> Automatic,
AxesOrigin -> {0, 0}]
\end{lstlisting}
\subsection{Figure 7}
This code provides the numerical evidence for Lemma 28.
\begin{lstlisting}[language=Mathematica]
L1[x0_]:= (*As before*)
(*First we evaluate the derivative*)
D[L1[x0],x0]
(*Then, we plot G(x0)*)
Plot
{x0,1/Sqrt[3.],1},AxesOrigin -> {1/Sqrt[3.], 0}]
\end{lstlisting}
\newpage
\subsection{Computing the General Period Function}
Stephen Miller helped us with this code. It allows us to compute the period function for any choice of $\alpha$ and $\beta$. Using it, it appears that the monotonicity results we would like are indeed true for arbitrary $\alpha$, a promising sign for our Main Conjecture.
\begin{lstlisting}[language=Mathematica]
(*Defining the integrand*)
integrand[t_, A_, B_] = 2/
Sqrt[1 - B^2/(A +
1) (A Exp[2 t] + Exp[-2 A t])]
(*Numerically finding the endpoints of integration *)
endpoints[A_, B_] :=
Sort[Log[Flatten[
y /. NSolve[
1 - B^2/(A + 1) (A y + y^-A) == 0,
y, 20]]]/2]
(*Numerical integration*)
p[A_, B_] :=
NIntegrate[integrand[t, A, B],
Join[{t}, endpoints[A, B]]]
(*Generates the table presented earlier*)
Table[{a, p[a, .999], Pi*Sqrt[2/a]}, {a, 0.1, 1, 0.1}]
\end{lstlisting}
\chapter{Prelude}
This thesis presents three main results in geometric analysis. Chapter 2 is concerned with the evolution of figure-eight curves by curve-shortening flow. Chapter 3 presents our work on the geometric flow on space curves that is point-wise curvature preserving. We find all stationary solutions of this flow and show that the stationary solutions corresponding to helices are $L^2$ linearly stable. Most of this chapter has been published as an article in Archiv der Mathematik \cite{COI2}, but we have added a section here on numerical solutions of the key partial differential equation. Chapter 4 presents an abridgment of our work on a family of Lie groups that, when equipped with canonical left-invariant metrics, interpolates between the Sol geometry and a model of Hyperbolic space. The complete version of our work will appear in Experimental Mathematics, and a preprint can be found on the ArXiv as well \cite{COI}.
The common thread running through our work is the geometric analysis of differential equations. That is to say: the obtainment of information about geometric objects by the analysis of (usually nonlinear) differential equations. Deep and difficult work on the Ricci flow and the behavior of minimal submanifolds are some of the most famous parts of geometric analysis, but we will deal with simpler and more concrete problems in this thesis.
\begin{center}
\textbf{Curve-Shortening Flow on Figure-Eight Curves}
\end{center}
Let $C(t): S^1\cross [0,T) \to \mathbb{R}^2$ be a family of immersed curves in the plane. We say that the family is evolving according to curve shortening flow (CSF) if and only if at any $(u,t) \in S^1\cross [0,T)$ we have
$$\frac{\partial C}{\partial t} = k N$$
where $k$ is the curvature and $N$ is the unit normal vector of the immersed curve $C(\cdot,t)$. For any given smoothly immersed initial curve $C(0)$, the existence of a corresponding family of curves $C(t)$ undergoing CSF is guaranteed, and the proof of short-term existence can be found in \cite{GH}. Due to the work of Gage, Hamilton, and Grayson in \cite{GH} and \cite{G}, we know that CSF shrinks any smoothly embedded initial curve to a point. We also know that, after the appropriate rescaling, any embedded curve becomes a circle in the limit. We will start this chapter by presenting some basic lemmas about the CSF, proven first in \cite{ANG}, \cite{GH}, or \cite{G}.
Later in the chapter, we focus on the asymptotic results about balanced figure-eight curves (having equal-area lobes) obtained by Grayson in \cite{G3}. Namely, we will present Grayson's proof that the isoperimetric ratio of a balanced figure-eight blows up in the singularity. Then, we will present Angenent's main result from \cite{ANG} that characterizes the blowup set of collapsing lobes. We finish this chapter with a discussion of a related problem communicated from Richard Schwartz concerning the evolution of figure-eight curves that have only one inflection point, are four-fold symmetric, and have additional monotonicity assumptions on curvature. These curves, which we term \textbf{Concinnous}, will be the main object of our study in the latter portion of Chapter 2. We will offer a method towards proving the following result, which describes the limiting shape of a concinnous figure-eight curve after evolution by CSF and an affine rescaling. Qualitatively, if we rescale the curve so that it has aspect ratio equal to $1$, the shape will limit to that of a bowtie. In a subsequent paper, Schwartz and I prove the following theorem:
\begin{theorem*}
Concinnous figure-eight curves converge to a point under CSF, and, after affine-rescaling, concinnous curves converge to a bow-tie shape.
\end{theorem*}
In this thesis, I will sketch some ideas in the proof of this result.
\begin{center}
\textbf{The Curvature Preserving Flow on Space Curves}
\end{center}
Substantial work has been done towards understanding geometric flows on curves immersed in Riemannian Manifolds. For example, the author of \cite{H} found that the vortex filament flow is equivalent to the non-linear Schr\"odinger equation, which enabled the discovery of explicit soliton solutions. Recently, geometric evolutions that are integrable, in the sense of admitting a Hamiltonian structure, have been of interest. The authors of \cite{I} and \cite{BSW} analyze the integrability of flows in Euclidean space and Riemannian Manifolds, respectively. In this chapter, we study the following geometric flow for curves in $\mathbb{R}^3$ with strictly positive torsion that preserves arc-length and curvature:
$$X_t=\frac{1}{\sqrt{\tau}}\textbf{B}.$$
Hydrodynamic and magnetodynamic motions related to this geometric evolution equation have been considered, as mentioned in \cite{SR}. The case when curvature is constant demonstrates a great deal of structure. After rescaling so that the curvature is identically $1$, the evolution equation for torsion is given by:
$$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big).$$
This flow has been studied since at least the publication of \cite{SR}. The authors of \cite{SR} show that the above equation, which they term the \textit{extended Dym equation}, is equivalent to the m$^2$KDV equation. In addition, they present auto-B\"acklund transformations and compute explicit soliton solutions. In Chapter 3, we continue the investigation of the curvature-preserving geometric flow.
We will characterize the flow $X_t=\frac{1}{\sqrt{\tau}}\textbf{B}$ as the unique flow on space curves that is both curvature and arc-length preserving, and we will provide another proof that this flow is equivalent to the m$^2$KDV equation. Afterwards, we will prove the global existence and uniqueness of solutions of the flow (for periodic $C^\infty$ initial data).
Then, we will concern ourselves with the stationary solutions to the geometric flow in the case of constant curvature. Namely, we derive the linearization of the evolution equation for torsion around explicit stationary solutions, and prove $L^2(\mathbb{R})$ stability of the linearization in the case of constant torsion (or for helices).
In addition to our work in \cite{COI2}, we present some numerical work done with Mathematica related to the evolution equation for torsion.
\begin{center}
\textbf{An Interpolation from Sol to Hyperbolic Space}
\end{center}
As mentioned earlier, this chapter presents an abbreviation of our work in \cite{COI}, whose complete version will appear in Experimental Mathematics. We study a one-parameter family of homogeneous Riemannian 3-manifolds that interpolates between three Thurston geometries: Sol, $\mathbb{H}^2\cross\mathbb{R}$, and $\mathbb{H}^3$. Sol is quite strange from a geometric point of view. For example, it is neither rotationally symmetric nor isotropic, and, since Sol has sectional curvature of both signs, the Riemannian exponential map is singular. In this article, we attempt to show that Sol's peculiarity can be slowly untangled by an interpolation of geometries until we reach $\mathbb{H}^2\cross\mathbb{R}$, which has a qualitatively "better" behavior than Sol. The interpolation continues on to $\mathbb{H}^3$, but we will not spend much time with that part of the family.
Our family of Riemannian 3-manifolds arises from a one-parameter family of solvable Lie groups equipped with canonical left-invariant metrics. We denote the groups by $G_\alpha$ with $-1\leq\alpha\leq 1$. Each $G_\alpha$ is the semi-direct product of $\mathbb{R}$ with $\mathbb{R}^2$, with the following operation on $\mathbb{R}^3$:
$$(x,y,z)\ast(x',y',z')=(x'e^z+x, y'e^{-\alpha z}+y, z'+z).$$
Then $\mathbb{R}^3$, which is only the underlying set, can be equipped with the following left-invariant metric:
$$ds^2=e^{-2z}dx^2+e^{2\alpha z}dy^2+dz^2.$$
These $G_\alpha$ groups, when endowed with the canonical metric, perform our desired interpolation: linking the familiar and the unfamiliar. It is a natural question to analyze what happens in between.
For positive $\alpha$, it happens that typical geodesics starting at the origin in $G_\alpha$ spiral around certain cylinders. For each such geodesic, there is an associated period that determines how long it takes for it to spiral exactly once around. We denote the function determining the period by $P$, and we will show that it is a function of the initial tangent vector. We call a geodesic segment $\gamma$ of length $T$ $\textit{small, perfect,}$ or $large$ whenever $T<P_\gamma,$ $T=P_\gamma,$ or $T>P_\gamma,$ respectively. Our primary aim is the following conjecture:
\begin{conjecture*}
A geodesic segment in $G_\alpha$ is a length minimizer
if and only if it is small or perfect.
\end{conjecture*}
The above conjecture is already known for $G_1$, or Sol, and was proven in \cite{MS}. In this article, we will reduce the general conjecture to obtaining bounds on the derivative of the period function by proving the \textbf{Bounding Box Theorem}. In particular, we will define a certain curve, and a key step in our would-be proof of the main conjecture is that a portion of this curve should be the graph of a monotonically decreasing function. The main obstacle to proving the whole conjecture is this monotonicity result. We are able to show the desired monotonicity for the group $G_{1/2}$ with our \textbf{Monotonicity Theorem} because we have found an explicit formula for the period function in this case. The group $G_{1/2}$ maximizes scalar curvature in our family, which offers more credence that it is truly a "special case" along with Sol. Thus, our main theorem is a proof of the conjecture for $\alpha=1/2$:
\begin{theorem*}
A geodesic segment in $G_{1/2}$ is a length minimizer
if and only if it is small or perfect.
\end{theorem*}
We have sufficient information about the period function for the group $G_{1/2}$ to extend the main result from \cite{MS}, obtaining a characterization of the cut locus and its consequences for geodesic spheres.
\begin{center}
\textbf{Acknowledgments}
\end{center}
I would first like to thank my advisor, Professor Richard Schwartz, for his encouragement and guidance during the last four years. He has taught me mathematics and a great way of seeing the beauty in mathematics.
I would like to thank all of my professors at Brown University for their openness, friendliness, and erudition in mathematics. I would especially like to express gratitude to Professors Benoit Pausader, Georgios Daskalopoulos, and Jeremy Kahn.
Lastly, I would like to thank my parents for their love and resolute support, ever since I was born.
\chapter{Curve-Shortening Flow on Figure-Eight Curves}
\section{Preliminaries}
Let $C(t): S^1\cross [0,T) \to \mathbb{R}^2$ be a family of immersed curves in the plane. We say that the family is evolving according to curve shortening flow (CSF) if and only if at any $(u,t) \in S^1\cross [0,T)$ we have
$$\frac{\partial C}{\partial t} = k N$$
where $k$ is the curvature and $N$ is the unit normal vector of the immersed curve $C(\cdot,t)$. A curvature function $k$ determines a plane curve up to Euclidean isometries. Thus, it suffices to understand the evolution of the curvature in order to understand the evolution of the curve.
The partial differential equation governing the evolution of curvature $k$ is given by:
\begin{equation}\frac{\partial k}{\partial t} = \frac{\partial^2 k}{\partial s^2} + k^3\end{equation}
where the derivatives in equation $(1)$ are with respect to arc-length. Remembering that the curve is shortening in length, we see that equation $(1)$ is a non-linear PDE. In many applications, however, it is more useful to use a modification of CSF that preserves some geometric quantity. Examples include the area preserving flow, the flow fixing $x$-coordinates, and the flow fixing tangent angles. First let us review the flow fixing $x$-coordinates.
We choose cartesian coordinates in the plane and rotate the evolution to fix $x$-coordinates. This yields a different flow that has the same point-set curves as solutions, but a different evolution equation for curvature. Equations describing this flow are found in the following lemma, proven in \cite{G} inter alia.
\newpage
\begin{lemma}[\cite{G}]
Choose coordinates so that $C(t)$ is locally a graph: $(x,y(x,t))$. If $'$ denotes differentiation w.r.t. $x$ and if $\theta$ is the angle that tangents to the curve make with the $x$ axis (i.e. $\theta(x,t)=\arctan{y'(x,t)}$), then
$$\frac{\partial y}{\partial t} = \frac{y''}{1+y'^2} \quad \textrm{ and }\quad \frac{\partial \theta}{\partial t} =\theta''\cdot \cos^2{\theta}$$
\end{lemma}
\begin{proof}
Since
$$\frac{\partial C}{\partial t} = k N$$
the speed of the evolution of the curve in its normal direction is $k$. Adjusting for the speed in the vertical direction (after having chosen coordinates) gets us
$$\frac{\partial y}{\partial t} = k \sec{\theta}.$$
Away from vertical tangents, we know that
$$k = \frac{y''}{(1+y'^2)^{3/2}} \quad\textrm{ and } \quad \sec{\theta} = \sqrt{1+y'^2}$$
so we have derived the evolution equation for $y$. The equation for $\theta$ is also derived easily and is strictly parabolic away from vertical tangents.
\end{proof}
\begin{figure}[h!]
\centering
\includegraphics[width=1\textwidth]{Figure1}
\caption{The local geometry of the $x$-coordinate preserving modification.}
\end{figure}
We also need to understand how the area enclosed by a curve decays with CSF.
\begin{lemma}[Also found in \cite{G}]
Let $A$ be the area bounded by a curve, the $x$-axis, and two vertical lines that intersect the curve at $x=a$ and $x=b$. Then
$$\frac{dA}{dt}= \int_a^b k\hspace{3pt} ds$$
\end{lemma}
\begin{proof}
This follows by differentiation of an integral:
$$\frac{dA}{dt} = \frac{d}{dt} \int_a^b y\hspace{3pt}dx = \int_a^b \frac{y''}{1+y'^2} dx = \int_a^b k \hspace{3pt}ds.$$
\end{proof}
Another modification to CSF is obtained by rotating to fix tangents to the curve (i.e. fixing $\theta$), which can be done away from inflection points. Using the chain rule, one can show that determining the evolution by CSF is equivalent to solving the following PDE:
\begin{lemma}[See \cite{ANG} or \cite{GH}]
For the CSF modification that fixes tangents, the evolution equation for curvature is
$$\frac{\partial k}{\partial t} = k^2 \frac{\partial^2 k}{\partial \theta^2}+k^3$$
\end{lemma}
We will be particularly interested in the evolution by CSF of the simplest non-embedded plane curves: figure-eights.
\begin{definition}
A smooth curve $C$ immersed in the plane is a \textbf{Figure-Eight} if and only if
\begin{itemize}
\item It has exactly one double point.
\item It has total rotation number $\int_C k ds =0$.
\end{itemize}
\end{definition}
\begin{definition}
A figure-eight curve $C$ is called \textbf{Balanced} if and only if its lobes bound regions of equal area.
\end{definition}
\begin{definition}
A balanced figure-eight curve is termed \textbf{Concinnous} if and only if
\begin{itemize}
\item Its double point is its only inflection point.
\item It is reflection symmetric about two perpendicular axes intersecting at its double point.
\item Suppose we parametrize so that $k>0$ in a lobe. Then, $k_\theta>0$ in the interior of that lobe and $k_{\theta\theta}<0$ in the interior of the appropriate half-lobe.
\end{itemize}
The axis of symmetry that only intersects the curve at its double point is called the \textbf{Major Axis}, and the other is called the \textbf{Minor Axis}.
\end{definition}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Figure3}
\caption{Some typical examples for each class of figure-eight curve}
\end{figure}
\newpage
\section{Balanced Figure-Eights}
Figure-eights that are not balanced evolve until one lobe collapses to a point. On the other hand, balanced figure-eights will evolve until the enclosed area converges to zero \cite{G3}. There are two possibilities here: either the balanced figure-eight converges to a point or to some line segment. It is still unknown whether all balanced figure-eights behave in the same manner.
\begin{conjecture}[Grayson]
Every balanced figure-eight converges to a point under the curve-shortening flow.
\end{conjecture}
For a figure-eight, the function $\theta$ is single-valued, and the image of the figure-eight in the $x\theta$-plane is a closed curve with well defined extrema. Moreover, because the evolution of $\theta$ is strictly parabolic (cf. Lemma 1), $\theta_{max}$ is strictly decreasing and $\theta_{min}$ is strictly increasing. By comparing the rate at which the extrema of $\theta$ approach each other with the diminution of the enclosed area, we can get a quantitative statement about the shape of a balanced figure-eight under CSF. The key to the proof of the main theorem in this section is the oft-employed maximum principle for strictly parabolic evolution equations.
\begin{figure}[H]
\centering
\includegraphics[width=0.55\textwidth]{Figure2}
\caption{The function $\theta(x,t)$ gives us an evolution of closed curves, evolving according to the equation in Lemma 1.}
\end{figure}
\begin{lemma}[The Maximum Principle as used in \cite{G}]
Suppose $F(x,t): [0,\epsilon]\cross [t_0, t_0 +\epsilon] \to\mathbb{R}$ satisfies a strictly parabolic evolution w.r.t time. Assume also that $F$ is not identically zero at $t_0$, $F(x,t_0)\geq 0$ for all $x$, and that $F(0,t), F(\epsilon, t) \geq 0$ for all $t$. Then
$$F(x,t)>0 \quad \textrm{ for all } \quad (x,t) \in (0,\epsilon) \cross (t_0, t_0 + \epsilon]$$
\end{lemma}
\begin{theorem}[Grayson, \cite{G3}]
When $C(0)$ is a balanced figure-eight, the isoperimetric ratio converges to $\infty$ in the singularity.
\end{theorem}
\begin{proof}
We shall use the modification of CSF that preserves $x$-coordinates. We also choose cartesian coordinates so that $\theta_{min}(t=0)$ is zero and the double point is at the origin. Without loss of generality, the initial curve encloses unit area. Now, the interior angle at the double point of our figure-eight is bounded between $0$ and $\pi$, so by Lemma 2, we know that the total area of our curve decays like:
\begin{equation}-4\pi \leq \frac{dA}{dt} \leq -2\pi.\end{equation}
Thus, after $1/(8\pi)$ seconds of evolution, our curve will still have a total area greater than $1/2$.
Suppose that the isoperimetric ratio of our curve at $t=0$ satisfies\\ $L^2/A=L^2 < M$. Then $|x| < \sqrt{M}/2$, and, since the curve is shortening in length, this remains true for all time. In our effort to apply the maximum principle to the evolution equation for $\theta$, we will compare $\theta$ with a function whose range is a subset of $[0,\pi/2)$. For the purposes of this theorem, we consider the solution $f(x,t)$ to the following (linear) initial value problem:
$$\frac{\partial f}{\partial t} = \frac{1}{2}\cdot \frac{\partial^2 f}{\partial x^2} \quad\textrm{and} \quad f(x,0) = \begin{cases} 0, & |x|< \sqrt{M} \\ \pi/4, & |x| \geq M\end{cases}$$
Explicitly, we have
$$f(x,t)=\frac{1}{8} \pi \left(\text{erfc}\left(\frac{\sqrt{M}-x}{\sqrt{2}
\sqrt{t}}\right)+\text{erfc}\left(\frac{\sqrt{M}+x}{\sqrt{2} \sqrt{t}}\right)\right)$$
where $\text{erfc}$ is the complementary error function. The range of $f(x,t)$ is contained in the interval $[0,\pi/4]$ and begins below the graph of $\theta(x,t=0)$. Thus, since the maximum principle applies for $\theta \in [0,\pi/4]$, we get that
$$\theta_{min}(t) > f(0,t)=\frac{\pi}{4} \text{erfc}\left(\frac{\sqrt{M}}{\sqrt{2} \sqrt{t}}\right)$$
for all further time. In particular, after $1/(8\pi)$ seconds, $\theta_{min}$ has increased by at least $\delta=\frac{\pi}{4}\text{erfc}\big(2\sqrt{M\pi}\big)$.
Given our starting, unit area curve, we also suppose that the isoperimetric ratio stays bounded during the evolution, i.e. $\frac{L^2}{A}(t) <M$ for all time. Since the area is converging to zero, the area halves infinitely many times, but we have already shown that every time the area halves, $\theta_{min}$ increases by at least a fixed positive amount $\delta$, which would mean that $\theta_{min}$ converges to $\infty$, an obvious contradiction.
\end{proof}
Before we continue, we remark that our particular choice of $f(x,t)$ for the comparison function was only a matter of convenience. We could just have easily chosen some other step function's evolution, so long as its range were contained in $[0,\pi/2]$.
\section{Concinnous Figure-Eights}
The following is an interesting and useful proposition of Grayson and Angenent, and more can be learned about its proof and consequences in \cite{DHW}.
\begin{proposition}
The limit of a balanced figure-eight curve in the singularity is contained in the closure of the set of (pre-singular) double points.
\end{proposition}
This is of course in the same vein as Grayson's conjecture, for the proposition ensures that the limiting set is either a point or line segment. For us, this also means that balanced figure-eight curves with sufficient symmetry (like our concinnous curves) definitely converge to a point.
\begin{corollary}
Concinnous figure-eight curves converge to a point under the curve-shortening flow.
\end{corollary}
The natural question to ask now: what is the asymptotic shape of the curve? The work done in \cite{HA} guarantees that (any) figure-eight curve cannot evolve in a self-similar manner. Theorem 1, on the other hand, tells us that the isoperimetric ratio blows up in the singularity. Thus, the area preserving flow will have the shape limit to two increasingly long and narrow slits.
Here we will mainly employ the modification of CSF that fixes tangent directions, previously mentioned in Lemma 3. The key is that, away from the only inflection point, we can parametrize our (strictly convex) concinnous curves by their tangent angle. Also, by symmetry, we need only examine the behavior of one lobe to get information about the other.
Choose cartesian coordinates so that the double point of our concinnous figure-eight curve is at the origin and such that the minor axis of symmetry coincides with the $x$-axis. Now, if $2\alpha(t)$ is the interior angle at the double point, then the domain for $\theta$, our tangent angle parametrization, is
$$D_1(t)= \big(-\alpha(t), \pi +\alpha(t)\big)$$
and the length of the interval $D_1(t)=2\alpha(t)+\pi$ is bounded as:
$$\pi<|D_1(t)| <2\pi.$$
Before we continue, we make precise what we mean by "blow-up set":
\begin{definition}
The \textbf{Blowup Set} is
$$\Sigma = \{ \theta\in D_1 | \lim_{t\to T} k(\theta, t) = \infty\} $$
\end{definition}
Angenent's Theorem is a very precise description of the blowup set whenever we have a collapsing lobe, which occurs, for example, in the case of balanced figure-eights and symmetric cardioids. Angenent colorfully calls the limiting shape a grim-reaper noose, because after rescaling so that $k_{max}(t)\equiv 1$, the blow-up set converges uniformly to a grim-reaper curve (i.e. $k(\theta)= \cos{\theta}$). Quantitatively, we have
\begin{theorem}[Angenent, \cite{ANG}]
If the blowup set $\Sigma$ has length less than $2\pi$, then for some $\beta\in D_1(t)$, we have
$$\Sigma = [\beta-\pi/2, \beta + \pi/2],$$
and
$$\lim_{t\to T} \frac{k(\beta+\phi, t)}{k_{max}(t)}= \cos{\phi}$$
uniformly for $\phi \in [-\pi/2, \pi/2]$.
\end{theorem}
For our concinnous figure-eights, we have an even more precise description of $\Sigma$. Indeed, after choosing cartesian coordinates so that the double point is at the origin and the $x$-axis is the minor symmetry axis of the curve, we know that $k_{max}(t)$ must be attained at $x_{max}(t)$ (i.e. at the end of the curve). We are able to get the following result in a future paper, written with Richard Schwartz:
\begin{corollary}
The blowup set of a concinnous figure-eight curve is $\Sigma = [0, \pi]$ and
$$\lim_{t\to T} \frac{k(\phi,t)}{k_{max}(t)}= \sin{\phi}$$
uniformly for $\phi\in[0,\pi]$.
\end{corollary}
Qualitatively, the far end of a figure-eight becomes a grim-reaper curve (turned sideways) after the appropriate rescaling.
\newpage
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{Figure4}
\caption{We are comparing the flow of a portion of the $\theta$ curve with the red step-function in the proof of Theorem 1. }
\end{figure}
\newpage
\section{The Affine Rescaling}
Professor Richard Schwartz has suggested to us another rescaling of curve-shortening flow that is especially suited for concinnous figure-eights: an "affine" rescaling that keeps the curve in a "bounding box" of side-length $1$. Typical affine transformations do \textit{not} commute with the CSF, unlike isometries of the plane, so we are performing the affine transformation at the end of the evolution and seeing what the result should be. We \textit{cannot} affinely rescale, afterwards continue the CSF, and still get the same result.
In this section, we will use both modifications of CSF. As usual, we choose cartesian coordinates so that the double point is at the origin and the minor axis of symmetry is the $x$-axis. Also, by the symmetry of a concinnous figure eight, we need only study the evolution of the portion of the curve that lies in the positive quadrant (i.e. $1/4$ of the curve). In particular, if we are using the $x-$coordinate preserving flow, we are studying the evolution of a multi-variable function $y(x,t)$ that satisfies
$$\frac{\partial y}{\partial t} = \frac{y''}{1+y'^2}$$
as in Lemma 1, with the initial function $y(x,0)$ corresponding to a concave curve having a vertical tangent at $x_{max}(0)$. The domain of this function is $D(t)=[0, x_{max}(t)]$, the range is $R(t)=[0,y_{max}(t)]$, and we choose the orientation of the curve so that $y''<0$ on $D(t)$. Both $x_{max}(t)$ and $y_{max}(t)$ are monotonically decreasing; we also define $x^*(t)$ to be the function of time that satisfies
$$y(x^*(t), t) =y_{max}(t).$$
Now we can make our affine rescaling explicit. We are dilating the $x$ direction by $1/x_{max}(t)$ and the $y$ direction by $1/y_{max}(t)$, which evidently leads to a curve contained within a unit square bounding box.
In the previous section, we saw Angenent's theorem on collapsing nooses and a related result for concinnous figure-eights. The far end converges to a grim-reaper curve. When we perform the affine rescaling on the grim reaper portion of the curve, we get a vertical line-segment because of its position on the far end and because the diameter is bounded in the $y$ direction and unbounded in the $x$ direction. In the framework of this section, this means that
\begin{proposition}
$$\lim_{t\to T} \frac{x^*(t)}{x_{max}(t)}=1.$$
\end{proposition}
What about the rest of the curve (after affine rescaling)? Numerical evidence, communicated to us from Richard Schwartz, suggests that the resulting shape should be shaped like a bow-tie. In one quadrant, this means that our curve limits to two line segments intersecting in the corner of the bounding box. This is our main claim in Chapter 2:
\begin{theorem}[Main Claim]
After affine-rescaling, concinnous figure-eights converge to a bow-tie shape.
\end{theorem}
\section{Towards our Main Claim}
The "proof" in this chapter is just a sketch. For example, we need to also show that the results in \cite{ANG} (Angenent only studied strictly convex immersed curves) still apply to the case of the figure-eight. However, we are preparing another paper which will address these issues and provide a complete proof of our main claim.
This next lemma gives us a goal to aim for.
\begin{lemma}
The concinnous figure-eight converges to a bow-tie if
$$\lim_{t\to T}\bigg( y_{max}(t)\cdot\frac{dx_{max}}{dt}(t)+x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)=-\pi.$$
\end{lemma}
\begin{proof}
After applying the affine rescaling and before the blow-up time, our concinnous figure-eight remains convex except at the double point. The limiting curve, after affine rescaling, will still have absolute curvature that is everywhere greater than or equal to zero. By Proposition 2, the far end of the figure-eight becomes a line segment after the affine rescaling (alternatively, it limits to the vertical edge of the bounding box). The only curve from the origin to the corner of the bounding box that encloses an area equal to $1/2$ and has $|k|\geq0$ is the obvious line segment. Thus, to show that our concinnous figure-eight converges to a bow-tie, it suffices to show that
$$\lim_{t\to T} \frac{x_{max}(t)y_{max}(t)}{2A(t)} = 1$$
where $A(t)$ is the area enclosed by one quarter of the curve and the axes of symmetry. Using Lemma 2, we see that
$$\frac{dA}{dt}(t)= -(\alpha(t)+\pi/2)$$
but since the blow-up set $\Sigma= [0,\pi]$, we know $\lim_{t\to T} \alpha(t)=0$.
Finally, by L'Hospital's rule, we have
$$\lim_{t\to T} \frac{x_{max}(t)y_{max}(t)}{2A(t)} = 1$$
if we can show that
$$\lim_{t\to T} \frac{y_{max}(t)\cdot\frac{dx_{max}}{dt}(t)+x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)}{-2(\alpha(t)+\pi/2)} = 1$$
which is equivalent to showing
$$\lim_{t\to T}\bigg( y_{max}(t)\cdot\frac{dx_{max}}{dt}(t)+x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)=-\pi.$$
\end{proof}
We continue by proving some technical lemmas (we are using the $x$-coordinate preserving flow in the next few).
\begin{lemma}
$$\frac{d x_{max}}{dt}(t)=k(x_{max}(t),t) \quad \textrm{ and } \quad \frac{d y_{max}}{dt}(t)=k(x^*(t),t)$$
\end{lemma}
\begin{proof}
By the chain rule:
$$\frac{d y_{max}}{dt}(t)=\frac{d}{dt}y(x^*(t),t)=\frac{\partial y}{\partial x}(x,t) \frac{d x^*}{dt}(t)+\frac{\partial y}{\partial t}(x,t)= y''(x^*(t),t)=k(x^*(t),t)$$
The proof is identical for the derivative of $x_{max}(t)$.
\end{proof}
\begin{lemma}
$$\lim_{t\to T}\frac{\frac{dy_{max}}{dt}(t)}{\frac{dx_{max}}{dt}(t)}=0$$
\end{lemma}
\begin{proof}
By Corollary 2, we know that
$$\lim_{t\to T} \frac{k(x^*(t),t)}{k(x_{max}(t),t)}=0$$
so the result follows immediately from the previous lemma.
\end{proof}
\begin{lemma}
$$\lim_{t\to T} \frac{y_{max}(t)}{x_{max}(t)}=0$$
\end{lemma}
\begin{proof}
Use L'Hospital's rule and the previous lemma.
\end{proof}
\begin{lemma}
$$\lim_{t\to T} \bigg(-y_{max}(t)\cdot \frac{dx_{max}}{dt}(t)\bigg)= \frac{\pi}{2}$$
\end{lemma}
\begin{proof}
Corollary 2 states that the far end of our concinnous figure-eight converges in $C^2$ to a Grim Reaper curve, after rescaling so that the maximum curvature is identically equal to $1$. However, from the same result, we also know
$$k_{max}(t)= \sup_{x\in D(t)} \abs{k(x,t)} = \abs{k(x_{max}(t),t)}.$$
Since the Grim Reaper is turned on its side, has finite vertical diameter equal to $\pi$, and is reflection-symmetric across the minor axis of our figure eight, the desired result follows immediately.
\end{proof}
The rest of the proof will be split into two main parts. In the first part, we will prove that
$$\limsup_{t\to T} \bigg(-x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)\leq\frac{\pi}{2}$$
by employing some integral estimates and our characterization of the Blowup Set in Corollary 2. In the second part we will use the Maximum Principle to show that
$$\liminf_{t\to T} \bigg(-x_{max}(t)\cdot\frac{dy_{max}}{dt}(t)\bigg)\geq\frac{\pi}{2}.$$
Together with Lemma 5, these results prove our main theorem.
\begin{center}
\textbf{First Part}
\end{center}
To use the full strength of Corollary 2, we need to use the CSF modification that preserves tangent angles. We recall that the evolution equation for curvature is then given by
$$\frac{\partial k}{\partial t} = k^2 \frac{\partial^2 k}{\partial \theta^2}+k^3$$
and the line element is $ds=d\theta/k$. We are now in a position to get integral formulae for $x_{max}(t), y_{max}(t)$, and their derivatives.
\begin{lemma}
$$x_{max}(t) = -\int_{-\alpha(t)}^{\pi/2} \frac{\cos\theta}{k(\theta,t)}d\theta \quad \textrm{ and } \quad y_{max}(t) = -\int_{0}^{\pi/2} \frac{\sin\theta}{k(\theta,t)}d\theta.$$
\end{lemma}
\begin{proof} This is obvious as long as we keep track of how our curve is parametrized by $\theta$ and if we remember that $ds=d\theta/k$.
\end{proof}
\begin{lemma}
$$\frac{dy_{max}}{dt}(t)=k(\theta=0,t)$$
\end{lemma}
\begin{proof}
Directly differentiating the integral equation in Lemma 10 gets us
$$\frac{dy_{max}}{dt}(t)=\int_0^{\pi/2}(k_{\theta \theta} + k)\sin(\theta) d\theta.$$
Since
$$(k_{\theta \theta} + k)\sin\theta=\frac{\partial}{\partial \theta}\big(k_\theta \sin\theta -k\cos \theta\big)$$
we have
$$\frac{dy_{max}}{dt}(t)=\big(k_\theta \sin\theta -k\cos \theta\big)\bigg|_{\theta=0}^{\theta=\pi/2}=k(\theta=0,t)$$
because curvature is maximized at the far end of our figure eight (i.e. at $\theta=\pi/2$).
\end{proof}
We are now equipped to accomplish our goal for Part 1.
\begin{lemma}
$$\limsup_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) \leq\frac{\pi}{2}$$
\end{lemma}
\begin{proof}
Using the two preceding lemmas, we have
$$-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)=\int_{-\alpha(t)}^{\pi/2} \frac{k(0,t)\cos\theta}{k(\theta,t)}d\theta$$
so
$$\abs{-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)}\leq (\pi/2+\alpha(t))\cdot\sup_{\theta\in[-\alpha(t),\pi/2]}\abs{\frac{k(0,t)}{k(\theta,t)}}.$$
From Corollary 2, which describes the blow-up set $[0,\pi]$, we get
$$\lim_{t\to T} \sup_{\theta\in[-\alpha(t),\pi/2]}\abs{\frac{k(0,t)}{k(\theta,t)}}=1$$
therefore, since $\lim_{t\to T} \alpha(t)=0$ as well, we have
$$\limsup_{t\to T} \abs{-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)}\leq \frac{\pi}{2}$$
as desired.
\end{proof}
\begin{center}
\textbf{Second Part}
\end{center}
Recall that our goal for this section is to prove
$$\liminf_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) \geq \frac{\pi}{2}.$$
In what follows, we use the modification of CSF that preserves $x-$coordinates. As usual, this means that we are looking at the evolution of one quarter of the concinnous figure-eight, viewed as the graph of the concave function $y(x,t)$. This function evolves with time according to
$$\frac{\partial y}{\partial t} = \frac{y''}{1+y'^2}.$$
Consider the function
$$Y(x,t):= \frac{y(x,t)}{x_{max}(t)}.$$
Using the quotient rule, we see that $Y(x,t)$ satisfies the following evolution equation:
$$\frac{\partial Y}{\partial t}(x,t) = \frac{Y''(x,t)}{1+y'(x,t)^2}-\frac{Y(x,t)\cdot \frac{dx_{max}}{dt}(t)}{x_{max}(t)}.$$
Away from vertical tangents, the evolution of $Y$ is strictly parabolic. In particular, by applying the Maximum Principle, we see that local maxima of $Y$ are strictly decreasing. Thus:
$$\frac{d}{dt}\bigg(\frac{y_{max}(t)}{x_{max}(t)}\bigg)<0$$
so we get, for all $t$,
$$-x_{max}(t)\frac{dy_{max}}{dt}(t)> -y_{max}(t)\frac{dx_{max}}{dt}(t)$$
and taking the limit:
$$\liminf_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) \geq \lim_{t\to T} \bigg(-y_{max}(t)\frac{dx_{max}}{dt}(t)\bigg)=\frac{\pi}{2}.$$
As a consequence, and using the results from Part 1, we have
$$\lim_{t\to T} \bigg(-x_{max}(t)\cdot \frac{dy_{max}}{dt}(t)\bigg) = \lim_{t\to T} \bigg(-y_{max}(t)\frac{dx_{max}}{dt}(t)\bigg)=\frac{\pi}{2}$$
therefore, by Lemma 5, our Main Theorem is proven. Q.E.D.
\chapter{The Curvature Preserving Flow on Space Curves}
\section{Preliminaries}
We recall the following standard computation:
\begin{lemma}[cf. \cite{BSW}, \cite{I}]
Let $\gamma(t)$ be a family of smooth curves immersed in $\mathbb{R}^3$ and let $X(s,t)$ be a parametrization of $\gamma(t)$ by arc-length. Consider the following geometric evolution equation:
\begin{equation}X_t = h_1 \textbf{T} +h_2 \textbf{N}+h_3 \textbf{B}\end{equation}
where $\{\textbf{T},\textbf{N},\textbf{B}\}$ is the Frenet-Serret Frame and where we denote the curvature and torsion by $\kappa$ and $\tau$, respectively. Let $h_1, h_2, h_3$ be arbitrary smooth functions of $\kappa$ and $\tau$ on $\gamma(t)$. If the evolution is also arc-length preserving, then the evolution equations of $\kappa$ and $\tau$ are
$$\begin{pmatrix} \kappa_t \\ \tau_t \end{pmatrix}=P\begin{pmatrix} h_3 \\ h_1 \end{pmatrix}$$
where $P$ is
$$\begin{pmatrix} -\tau D_s - D_s\tau & D_s^2\frac{1}{\kappa}D_s-\frac{\tau^2}{\kappa}D_s+D_s\kappa\\
D_s\frac{1}{\kappa}D_s^2-D_s\frac{\tau^2}{\kappa}+\kappa D_s& D_s(\frac{\tau}{\kappa^2}D_s+D_s\frac{\tau}{\kappa^2})D_s+\tau D_s+D_s\tau\end{pmatrix}.$$
\end{lemma}
The next theorem follows without difficulty.
\begin{theorem}
Up to a rescaling, a geometric evolution of curves immersed in $\mathbb{R}^3$, as in equation $(3)$, is both curvature and arc-length preserving if and only if its evolution evolution is equivalent to
\begin{equation}
X_t = \frac{1}{\sqrt{\tau}}\textbf{B}.
\end{equation}
\end{theorem}
\begin{proof}
Let $X_t$ be a curvature and arc-length preserving geometric flow. The tangential component of $X_t$ in equation $(3)$ provides no interesting geometric information; it amounts to a re-parametrization of the curve. Thus, we may assume that $h_1=0$, so, since $X_t$ is arc-length preserving, $h_2=0$ as well. Lemma 13 gives us that
$$\kappa_t = -\tau D_s (h_3)-D_s(\tau h_3)=-2\tau D_s (h_3)-h_3D_s(\tau).$$
Since $X_t$ is curvature preserving, we must have $\kappa_t=0$, or
$$D_s(\log{h_3})=D_s(\log{\tau^{-1/2}}).$$
Integrating, we see that
$$h_3=\frac{c}{\sqrt{\tau}}$$
where $c$ is a constant. Therefore, up to a rescaling, the evolution $X_t$ must be precisely as in $(4)$.
\end{proof}
Unfortunately, the evolution in $(4)$ only makes sense if $\gamma(t)$ has strictly positive torsion (or strictly negative torsion, with the flow $X_t=\frac{1}{\sqrt{-\tau}}\textbf{B}$). This motivates the following definition.
\begin{definition}
We call a smooth curve immersed in $\mathbb{R}^3$ a positive curve and only if it has strictly positive torsion.
\end{definition}
Fortunately, there are many interesting positive curves. For example, some knots admit a parametrization with constant curvature and strictly positive torsion, and there exist closed curves with constant (positive) torsion. Henceforth, we will only consider positive curves.
The partial differential equation governing the evolution of $\tau$ follows below:
\begin{lemma}
For the geometric flow given in equation $(4)$, the evolution equations of curvature and torsion are $\kappa_t =0$ and
\begin{equation}
\tau_t = \kappa D_s(\tau^{-1/2})+D_s\bigg(\frac{D_s^2(\tau^{-1/2})-\tau^{3/2}}{\kappa}\bigg)
\end{equation}
\end{lemma}
The equation for torsion is reminiscent of the Rosenau-Hyman family of equations, which are studied, inter alia, in \cite{HEER} and \cite{LW}. The authors of \cite{SR} call the evolution of the torsion, in the case of constant curvature, the \textit{extended Dym equation} for its relationship with the Dym equation (a rescaling and limiting process converts the evolution for torsion to the Dym equation).
As discussed in \cite{BSW}, the condition for the flow $X_t$ from Lemma 13 to be the gradient of a functional is for the Frechet derivative of $(h_3,h_1)$ to be self-adjoint. In general, this does not occur for the flow under our consideration in equation $(4)$, so it cannot be integrable in the sense of admitting a Hamiltonian structure (Indeed, when $\kappa$ is not constant, it is not even formally integrable in the sense of \cite{MSS}). Nevertheless, the case of constant curvature exhibits a great deal of structure, which makes the study of the evolution equation in $(5)$ worthwhile.
\section{Constant Curvature}
When $\kappa$ is constant, we may rescale the curves $\gamma(t)$ so that $\kappa \equiv 1$. In this way, the evolution of torsion becomes
\begin{equation}
\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big)
\end{equation}
\subsection{Equivalence with the m$^2$KDV Equation}
We recall the notion of "equivalence" of two partial differential equations from \cite{CFA}:
\begin{definition}
Two partial differential equations are equivalent if one can be obtained from the other by a transformation involving the dependent variables or the introduction of a potential variable.
\end{definition}
The authors in \cite{CFA} discuss a general method of transforming quasilinear partial differential equation, such as the evolution of $\tau$ in $(6)$, to semi-linear equations. By applying their algorithm, we obtain another proof of the following theorem, first demonstrated in \cite{SR}.
\begin{theorem}[\cite{SR}]
The evolution equation for torsion, in the case of constant curvature, which is given by
$$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big),$$
is equivalent to the m$^2$KDV equation. Thus, it is a completely integrable evolution equation.
\end{theorem}
\begin{proof}
First, let $\tau=v^2$, so that $(6)$ becomes
$$
2vv_t=D_s\bigg(\frac{1}{v}-v^3+D_s^2\big(\frac{1}{v}\big)\bigg)=-\frac{v_s}{v^2}-3v^2v_s-\frac{v_{sss}}{v^2}-\frac{6v_s^3}{v^4}+\frac{6v_sv_{ss}}{v^3}$$
or, in a simpler form:
\begin{equation}
v_t=D_s\bigg(\frac{1}{4v^2}-\frac{3v^2}{4}+\frac{3v_s^2}{4v^4}-\frac{v_{ss}}{2v^3}\bigg)
\end{equation}
A potentiation, $v=w_s$ followed by a simple change of variables $(t\rightarrow -t/2)$ yields
\begin{equation}
w_t=-\frac{1}{2w_s^2}+\frac{3w_s^2}{2}-\frac{3w_{ss}^2}{2w_s^4}+\frac{w_{sss}}{w_s^3}.
\end{equation}
Equation $(8)$ is fecund territory for a pure hodograph transformation, as used, for example, in \cite{CFA}. Let $\tilde{t}=t$, $\xi=w(s,t)$, and $s=\eta(\xi,\tilde{t})$. The resulting equation, after a simple computation, is
\begin{equation}
\eta_{\tilde{t}}=\eta_{\xi\xi\xi}-\frac{3\eta_{\xi\xi}^2}{2\eta_\xi}+\frac{\eta_\xi^3}{2}-\frac{3}{2\eta_\xi}.
\end{equation}
We rename the variables to the usual variables of space and time: $s$ and $t$; in addition, we anti-potentiate the equation by letting $\eta_s=z$. This makes equation $(9)$ equivalent to
\begin{equation}
z_t = z_{sss}-\frac{3}{2}\bigg(\frac{z_s^2}{z}\bigg)_s+\frac{3z^2z_s}{2}+\frac{3z_s}{2z^2}.
\end{equation}
Lastly, if we let $q=\sinh(z/2)$ in equation (10), a transformation also discussed in \cite{CD}, we get the m$^2$KDV equation
\begin{equation}q_t=q_{sss}-\frac{3}{2}\bigg(\frac{qq_s^2}{1+q^2}\bigg)_s+6q^2q_s\end{equation}
which finishes the proof of the theorem.
\end{proof}
Equations $(6)$ and $(7)$ give us the first two integrals of motion of this flow:
$$\int \sqrt{\tau} ds \textrm{ and } \int \tau ds.$$
The rest can be found by pulling back the m$^2$KDV invariants; we note that these invariants were obtained in a different way by the authors of \cite{SR}. Our next step is to prove the long term existence of our geometric flow using Theorem 5.
\begin{corollary}
Let $\tau_0\in C_{per}^\infty([0,2\pi])$ be a strictly positive and periodic function. We can solve the evolution equation for torsion with initial data $\tau_0$ and get the unique solution $\tau(t)\in C_{per}^\infty([0,2\pi])$ that is strictly positive for all time $t\geq 0$.
\end{corollary}
\begin{proof}
In order to prove this corollary, it suffices to show that we can reconstruct the solution of $(6)$ by solving the m$^2$KDV equation $(11)$ instead. The long term existence for solutions to the m$^2$KDV equation will imply the same for $(6)$, so all that remains to prove is that every differential transformation we used in the proof of Theorem 5 is truly "invertible" in the sense that no singularities arise. We use the same notation for our differential transformations as before.
For the proof of this corollary, it is more convenient to work with an equivalent form of the m$^2$KDV equation called the Calogero-Degasperis-Fokas (CDF) equation \cite{CFA}, obtained by letting $z=e^u$ in equation $(10)$:
$$
u_t=u_{sss}-\frac{u_s^3}{2}+\frac{3u_s}{2}\big(e^{2u}+e^{-2u}\big).
$$
Beginning with our initial torsion, $\tau_0$, we let $v_0=\sqrt{\tau_0}$ and then
$$w_0(s)=\int_0^s v_0(\tilde{s})d\tilde{s}.$$
By construction $w_0\in C^\infty([0,2\pi])$, is not periodic, but satisfies $w_0'(s)>0$ for all $s\in[0,2\pi]$. Thus, $w_0$ has a global inverse $\eta_0(\xi)$ and we also have $\eta_0'(\xi) := z_0(\xi)>0$ for all $\xi$ in the domain of $\eta_0$, which is some interval of the form $[0,M]$. Finally, let $u_0=\log(z_0)$, which is well defined and in $C^\infty([0,M])$ because $z_0>0$ and $z_0\in C^\infty([0,M])$. The CDF equation is globally well-posed for smooth functions on a compact interval with periodic boundary conditions (see \cite{CD}, \cite{HEER}), so we get a solution $u(\xi,t)$ defined for all time with $u(\xi,0)=u_0(\xi)$ and $u(0,t)=u(M,t)$ for all $t$. Our next step is to reconstruct the solution for $(6)$ using $u(t)$.
First, let $z(t)=e^{u(t)}$, which will be a strictly positive, periodic, $C^\infty$ function for all time. Then, let
$$\eta(\xi,t)=\int_0^\xi z(\zeta, t) d\zeta \quad \textrm{ for all } \xi\in[0,M]$$
which, since $\int z(\zeta,t) d\zeta$ is a time-independent quantity, satisfies
$$\eta(0,t)=0 \quad \textrm{and}\quad \eta(M,t)=2\pi$$
for all time. Obviously $\eta_\xi(\xi, t)>0$ for all $(\xi,t)$, so the hodograph transformation $\tilde{t}=t$, $\xi=w(s,t)$, and $s=\eta(\xi,\tilde{t})$ makes sense. We recover a $C^\infty$ function $w(s,t)$ with the property that
$$w(0,t)=0 \quad\textrm{and} \quad w(2\pi,t)=M$$
for all time, and up to a trivial change of coordinates, we have
$$0<v(s,t)=\frac{\partial w}{\partial s}(s,t_0)\in C^\infty_{per}([0,2\pi])$$
for any fixed time $t_0\geq 0$. Lastly, $\tau(s,t)=v(s,t)^2$ solves equation $(6)$, is in $C^\infty_{per}([0,2\pi])$, and satisfies $\tau(s,0)=\tau_0(s)$. Thus, we have reconstructed our desired solution of $(6)$ using the completely integrable CDF equation instead.
\end{proof}
\subsection{Stationary Solutions}
Helices, with constant curvature and constant torsion, are the obvious stationary solutions. In what follows, we find the rest.
\begin{theorem}
The stationary solutions of
$$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big)$$
are given by the following integral formula
$$\int\frac{du}{\sqrt{C+2Au-u^2-u^{-2}}}=s+k$$
where $\tau(s)=u(s)^{-2}$ and where $A,k$ and $C$ are appropriate real constants. When $A=0$ we get an explicit formula for the solutions:
$$\tau(s)=\frac{2}{C\pm\sqrt{(-4+C^2)}\cdot\sin(2(s+k))},$$
with $k$ and $C\geq 2$ real constants.
\end{theorem}
\begin{proof}
Let $\tau=u^{-2}$, then, after integrating once, we must examine the following ordinary differential equation (where $A$ is a constant):
\begin{equation}
A=u-\frac{1}{u^3}+D_s^2(u)
\end{equation}
Since equation $(12)$ is autonomous, we may proceed with a reduction of order argument. Let $w(u)=D_s(u)$ so that $D_s^2(u)=wD_u(w)$ by the chain rule. This substitution gives us the first order equation:
\begin{equation}
wD_u(w)=A-u+\frac{1}{u^3}
\end{equation}
or
$$D_u(w^2)=2A-2u+\frac{2}{u^3}.$$
Integrating, we get
$$D_s(u)=w(u)=\sqrt{C+2Au-u^2-u^{-2}}$$
which is a separable differential equation. So, the stationary solutions of equation $(6)$ are given by the following integral formula:
\begin{equation}
\int\frac{du}{\sqrt{C+2Au-u^2-u^{-2}}}=s+k
\end{equation}
for appropriate constants $C, A, \textrm{ and } k$.
It would be pleasant to have explicit solutions, and this occurs in the case when $A=0$, which is more easily handled. Equation $(13)$ above becomes
$$D_s(u)=w(u)=\sqrt{C-u^2-u^{-2}}$$
which is a differential equation that can be solved with the aid of Mathematica or another computer algebra system. The result is
\begin{equation}
u(s)=\sqrt{\frac{C\pm\sqrt{(-4+C^2)}\cdot\sin(2(s+k))}{2}}
\end{equation}
Where $C,k$ are real constants and $C\geq 2$. The corresponding torsion is:
$$\tau(s)=\frac{2}{C\pm\sqrt{(-4+C^2)}\cdot\sin(2(s+k))}$$
\end{proof}
Integrating $\tau$ and $\kappa$ as above using the Frenet-Serret equations will yield the corresponding stationary curves, up to a choice of the initial Frenet-Serret frame and isometries of $\mathbb{R}^3$.
\subsection{$L^2(\mathbb{R})$ Linear Stability of Helices}
First, we derive the linearization of the evolution for torsion around the stationary solutions corresponding to helices. The linearization of equation $(6)$ at any stationary solution $\tau_0$ is obtained by letting $\tau(s,t) = \tau_0(s,t)+\epsilon w(s,t)$, substituting into equation $(6)$, dividing by $\epsilon$ and then taking the limit as $\epsilon\rightarrow 0$. This is nothing more than the Gateaux derivative of our differential operator at $\tau_0$ in the direction of $w$. Alternatively, one may think of $w$ as the first-order approximation for solutions of $(6)$ near the stationary solution. We can perform this operation when $\tau_0$ is given by an explicit formula, but for brevity's sake, we only mention here the linearization around helices when $\tau_0$ is constant. A short calculation yields
\begin{proposition}
The linearization of the evolution equation $(6)$ around the stationary solutions of constant torsion is
\begin{equation}w_t+2w_s+\frac{1}{2}w_{sss}=0.\end{equation}
\end{proposition}
In what follows, we show the $L^2(\mathbb{R})$ linear stability of the the constant torsion stationary solution. First we recall the definition of linear stability:
\begin{definition}
A stationary solution $\phi$ of a nonlinear PDE is called $L^2(\mathbb{R})$ linearly stable when $v = 0$ is a stable solution of the corresponding linearized PDE with respect to the $L^2(\mathbb{R})$ norm and whenever $v_{t=0}$ is in $L^2(\mathbb{R})$.
\end{definition}
To work towards this, we need to use test functions from the Schwartz Space $\mathcal{S}(\mathbb{R})$, so we first recall that
$$\mathcal{S}(\mathbb{R}):=\{ f\in C^{\infty}(\mathbb{R}) \textrm{ s.t. } \sup_{x\in\mathbb{R}} (1+\abs{x})^N\abs{\partial^{\alpha}f(x)}<\infty, \textrm{ for all } N, \alpha \}.$$
Intuitively, functions in $\mathcal{S}(\mathbb{R})$ are smooth and rapidly decreasing. For more on distributions and the Schwartz Space, review \cite{F}.
The rest of this section is devoted to proving:
\begin{theorem}
Helices correspond to $L^2(\mathbb{R})$ linearly stable stationary solutions of $$\tau_t = D_s\big(\tau^{-1/2}-\tau^{3/2}+D_s^2(\tau^{-1/2})\big).$$
\end{theorem}
\begin{proof}
Helices correspond to the constant torsion stationary solutions, so we analyze the linearized PDE in (16):
$$w_t+2w_s+\frac{1}{2}w_{sss}=0.$$
Henceforth, $w_0(s)$ denotes the initial data and $w(s,t)$ denotes the respective solution to the above, linear PDE. Thus, to prove the theorem, it suffices to show: for every $\epsilon$, there exists a $\delta$ such that if $w_0\in L^2(\mathbb{R})$ and $\|w_0\|_2<\delta$, then $\|w(t)\|_2<\epsilon$ for all $t\geq0$.
We follow the standard process of finding weak solutions to linear PDEs via the Fourier transform. Moreover, we know by the Plancherel Theorem that we can extend the Fourier transform by density and continuity from $\mathcal{S}(\mathbb{R})$ to an isomorphism on $L^2(\mathbb{R})$ with the same properties. Hence, it suffices to prove the desired stability result for initial data in $\mathcal{S}(\mathbb{R})$.
Let $F_t(\xi)=e^{4 i \pi ^3 \xi^3 t-4 i \pi \xi t}$. We notice that since $F_t$ is a bounded continuous function for all $t\geq 0$, it can be considered a tempered distribution (or a member of $\mathcal{S}'(\mathbb{R})$, the continuous linear functionals on $\mathcal{S}(\mathbb{R})$), so its inverse Fourier transform makes sense.
Indeed, we can let $B_t(s)=\mathcal{F}^{-1}(F_t(\xi))\in \mathcal{S}'(\mathbb{R})$ and, again, we denote $w_0\in \mathcal{S}(\mathbb{R})$ to be our initial data. Let
$$w(t)=B_t \ast w_0$$
so that $w(t)$ is a $C^{\infty}$ function with at most polynomial growth for all of its derivatives (see \cite{F}). Moreover, $w(t)$ satisfies equation $(16)$ in the distributional sense, as can be checked by taking the Fourier Transform. Lastly, since the Fourier Transform is a unitary isomorphism, it follows that
$$\lim_{t\rightarrow 0} w(t) = w_0$$
in the distribution topology of $\mathcal{S}'(\mathbb{R})$. Hence, $w(t)$ is the weak solution to equation $(16)$ with initial data $w_0\in\mathcal{S}(\mathbb{R})$. In our final step, we use the Plancherel Theorem and the fact that $\mathcal{F}(B_t)=F_t$ is a continuous function with $\|F_t\|_\infty =1$ for all $t\geq0$ to get:
$$\|w(t)\|_2=\|B_t \ast w_0 \|_2 =\|F_t \cdot \mathcal{F}(w_0)\|_2\leq \|F_t\|_\infty \|\mathcal{F}(w_0)\|_2= \|\mathcal{F}(w_0)\|_2=\|w_0\|_2.$$
From the inequality above, the desired $L^2(\mathbb{R})$ stability for initial data in $\mathcal{S}(\mathbb{R})$ follows forthrightly.
\end{proof}
\subsection{Numerical Rendering of a Stationary Curve}
We provide here the figures obtained from a numerical integration of the Frenet-Serret equations on Mathematica for the following choice of torsion:
$$\tau_1(s)=\frac{2}{3+\sqrt{5}\sin(2s)}$$
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{PLOTTOR}
\caption{This is the graph of the torsion $\tau_1$ over two periods.}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{XYPROJ}
\caption{This is the projection of the curve corresponding to $\tau_1$ into the $xy$ plane. The projections into the other planes look very similar.}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.35\textwidth]{TOPDOWN}
\caption{This is a top-down view of the same curve, now exhibiting an almost trefoil shape.}
\end{figure}
\section{Numerical Solutions to the Torsion Evolution Equation}
Mathematica offers a built-in function, \textbf{NDSolve}, which can numerically solve both ordinary and partial differential equations. In our case, the PDE in question is a highly non-linear, time-dependent evolution, and Mathematica will usually employ the "method of lines" to get a numerical solution. The method of lines involves a discretization of the spatial variable that reduces the problem to one of integrating a system of ODE's, and its employment in numerical integration of PDE's is documented for at least the past five decades \cite{SC}.
The benefits of examining numerical solutions are manifold. First, we are able to try different initial data for torsion and see if the evolution behaves nicely; second, we can observe whether our stationary solutions demonstrate stability by slightly perturbing the initial data; third, we can experiment with the case of non-constant curvature, which is difficult to investigate with analytic methods. In this section, we pursue all three directions with several examples in each case.
\subsection{Constant Curvature with Different Choices for Initial Torsion}
In this subsection, the curvature of our positive curves will always be identically equal to $1$. In our experiments, we have observed that initial torsion that is "too close" to zero leads to numerical problems (as might be expected), so we tend to choose our torsion to be at least $\gg1$. Our first example is the evolution of $2\pi$ periodic, smooth initial data that is quite far from vanishing anywhere:
$$\tau(s,t=0)=10+\frac{\sin(s)}{2}$$
We numerically solve equation $(6)$ for $\tau(s,t)$ and plot the graph of the result.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{FirstTor}
\caption{The is the graph for $(s,t)\in [0,2\pi] \cross [0,5]$}
\end{figure}
\newpage
Already, we see a wave structure traveling along some direction (not parallel to the axes) while preserving its structure, for the most part. Quasi-periodicity is behavior typical of integrable partial differential equations, and this is evident for our flow as well. We direct the reader to the following figure, which depicts how our initial sine-wave is (almost) seen again at the times $t= 1.32, 2.64, 3.96$.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{FirstTorWave}
\caption{The is the graph of $\tau$ at times $t=0, 1.32, 2.64,$ and $3.96$.}
\end{figure}
Our next example is the evolution of the following initial torsion data:
$$\tau(s,t=0) = 10 +\sin{s}+\cos{s}.$$
Once again, we numerically solve and plot the graph of the result.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{SecondTor}
\caption{The is the graph for $(s,t)\in [0,2\pi] \cross [0,4]$}
\end{figure}
Although we can once again see the propagation of a wavefront, in this case the solution at fixed times seems to split into two different waves with different velocities.
If we plot $\tau$ at the quasi periods $t=0, 1.26, 2.53,$ and $3.685$, we get the subsequent graphs:
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{SecondTorWave}
\caption{The is the graph of $\tau$ at times $t=0, 1.26, 2.53,$ and $3.685$.}
\end{figure}
We may conclude that the evolution equation $(6)$ demonstrates some typical behavior for non-linear wave equations.
\subsection{Numerical Evidence for the Stability of Helices}
In Section 2.3 of this chapter, we have proven that helix stationary solutions of the equation $(6)$ are $L^2(\mathbb{R})$ linearly stable. In this section, we will examine small perturbations of the helix solutions and observe that there is some evidence that helices have non-linear stability as well.
To that end, let
$$\tau_0(s)=1+\frac{\sin(s)}{100}$$
and let $\tau(s,t)$ be the solution to the initial value problem with $\tau(s,0)=\tau_0(s)$. Let $\| \cdot \|_2$ be the usual norm on $L^2(\mathbb{T})=L^2([0,2\pi]/_{per})$. The function $\tau_0$ adds only a small initial perturbation to the helix solution $\tau \equiv 1$. Indeed, we can compute
$$\| \tau(s,0)-1\|_2 \approx 0.0177245$$
Now we define the function
$$S(t) := \| \tau(\cdot,t)-1\|_2$$
which measures how much our solution for $\tau$ will deviate from the the constant torsion of a helix. Numerical integration of our (numerical) solution for $\tau(s,t)$ allows us to graph $S(t)$ and see whether the solution is non-linearly stable in $L^2(\mathbb{T})$ norm.
Below is an approximation of $S(t)$ for $t\in [0,50]$:
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{Stable}
\end{figure}
Barring numerical errors that occur when integrating differential equations over long time intervals, it appears that the solution $\tau(s,t)$ stays relatively close to $1$ in $L^2(\mathbb{T})$ norm, since $S(t)\leq S(0)$ for most of the time interval. We can also glean that asymptotic stability is unlikely, since it appears that the difference, $S(t)$, does not decrease to zero in time. We might conjecture that helix solutions are non-linearly stable, but the evidence for this is certainly not definitive.
\subsection{Solutions with Non-Constant Curvature}
We recall equation $(5)$, which is the evolution equation for the torsion with arbitrary curvature:
$$\tau_t = \kappa D_s(\tau^{-1/2})+D_s\bigg(\frac{D_s^2(\tau^{-1/2})-\tau^{3/2}}{\kappa}\bigg)$$
Our goal in this subsection is to numerically integrate the above inhomogeneous PDE for a certain choice of curvature and initial torsion. We will choose our curvature to be periodic, $\kappa(s) = 2+\cos{s}$ and our initial torsion will be $\tau(s,0)=1+\frac{\sin{s}}{10}$. To observe how the inhomogeneous PDE differs from the PDE with constant coefficients, we also compare the solution found with the same initial torsion but with constant curvature equal to $2$ (the average value of our choice of $\kappa(s)$ over the interval $[0,2\pi]$). The results of numerically solving the PDE are depicted on the next page.
Obtaining analytic results on the inhomogeneous PDE is obstructed by the inconvenient algebraic nature of the PDE (e.g. it is not even formally integrable). It would be interesting to tackle the inhomogeneous case, especially if one could obtain such results regardless of the choice of curvature.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{Alone}
\caption{The graph of the numerical solution for $\tau(s,t)$ with $(s,t)\in[0,2\pi]\cross[0,1]$.}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{Together}
\caption{The graphs of the numerical solutions for $\tau(s,t)$ with $(s,t)\in[0,2\pi]\cross[0,1]$ and with the blue graph corresponding to the solution with constant curvature equal to $2$. Observe that the solution to the inhomogeneous PDE more rapidly differs from the initial torsion.}
\end{figure}
\section{Computer Code}
The graphs in this chapter were rendered in Mathematica, which we also used to numerically solve the PDE. The code we typically used to perform the latter is similar to the following:
\begin{lstlisting}[language=Mathematica]
NDSolve[{D[ta[s, ti], ti] == (Cos[s] + 2)*D[ta[s, ti]^(-1/2), s] +
D[(D[ta[s, ti]^(-1/2), {s, 2}] - ta[s, ti]^(3/2))/(Cos[s] + 2),
s], ta[s, 0] == 1 + Sin[s]/10,
ta[0, ti] == ta[2 Pi, ti]}, {ta}, {s, 0, 2 Pi}, {ti, 0, 1},
MaxStepSize -> 0.01]
\end{lstlisting}
Afterwards, we can recover the associated curve (after choosing initial conditions) by integrating the Frenet-Serret equations like in:
\begin{lstlisting}[language=Mathematica]
Manipulate[
eqns = {t'[s] == \[Kappa][s] n[s],
n'[s] == -\[Kappa][s] t[s] - \[Tau][s] b[s],
b'[s] == \[Tau][s] n[s], r'[s] == t[s], t[0] == t0, n[0] == n0,
b[0] == b0, r[0] == r0};
\[Kappa][s_] := Sin[s] + 5;
\[Tau][s_] := ta[s, ti] /. s1;
{t0, n0, b0} = Orthogonalize[{{1, 0, 0}, {0, 1, 0}, {0, 0, 1}}];
r0 = {0, 0, 0};
sol = First@NDSolve[eqns, {r, t, n, b}, {s, 0, 2 Pi}];, {ti, 0, 4}]
\end{lstlisting}
\chapter{An Interpolation from Sol to Hyperbolic Space}
\section{The Basic Structure}
In this section, we collect some basic facts about all of the $G_\alpha$ groups. The main idea is that a fruitful way of analyzing the geometry of these Lie groups is to first understand the geodesic flow, and this is the setting in which we will continue our analysis for the remainder of this paper.
The principal object of our study will be a one-parameter family of three-dimensional Lie groups, whose Lie algebras are of Type VI in Bianchi's classification, as elaborated in \cite{LB}. For ease of computation, we can also construct our groups as certain subgroups of $GL_3(\mathbb{R})$, and to that end we let
$$G_\alpha=\bigg\{ \begin{pmatrix} e^z&0&x \\ 0&e^{-\alpha z}&y\\0&0&1\end{pmatrix}\bigg| x,y,z\in\mathbb{R}\bigg\} \textrm{ for all } -1\leq\alpha\leq1$$
We can consider each $G_\alpha$ as a matrix group or, equivalently, as $\mathbb{R}^3$ with the following group law:
$$(x,y,z)\ast(x',y',z')=(x'e^z+x, y'e^{-\alpha z}+y, z'+z).$$
Then, $\mathbb{R}^3$ with this group law has the following left-invariant metric
$$ds^2=e^{-2z}dx^2+e^{2\alpha z}dy^2+dz^2.$$
The Lie algebra of $G_\alpha,$ which we denote $\frak{g}_\alpha$, has the following orthonormal basis:
$$\bigg\{ X=\begin{pmatrix} 0&0&1 \\ 0&0&0\\0&0&0\end{pmatrix}\quad Y=\begin{pmatrix} 0&0&0 \\ 0&0&1\\0&0&0\end{pmatrix}\quad Z=\begin{pmatrix} 1&0&0 \\ 0&-\alpha&0\\0&0&0\end{pmatrix}\bigg\}$$
or
\begin{equation}X=e^z \frac{\partial}{\partial x}, Y=e^{-\alpha z}\frac{\partial}{\partial y}, Z=\frac{\partial}{\partial z}.\end{equation}
So, the structure equations are:
\begin{equation}[X,Y]=0\quad [Y,Z]=\alpha Y\quad [X,Z]=-X\end{equation}
The behavior of $\frak{g}_\alpha$ coincides with the Lie algebras of Type VI (in Bianchi's Classification) when $0<|\alpha|<1$ and we have three limiting cases: $\alpha=1,$ which is a Bianchi group of type VI$_{0}$, $\alpha=0,$ which is a Bianchi group of type III, and $\alpha=-1$, which is a Bianchi group of type V. The intermediate cases are not unimodular, unlike the limiting cases, and that may also be of interest. As a consequence of the classification done in \cite{LB}, no two of the Lie algebras $\frak{g}_\alpha$ are isomorphic, hence
\begin{proposition}[Bianchi, \cite{LB}]
No two of the $G_\alpha$ are Lie group isomorphic.
\end{proposition}
One might ask what occurs if we let the parameter $|\alpha|>1$, and the answer is simple. In this case, the Lie algebra will be isomorphic to that of one of our $G_\alpha$ groups with $|\alpha|<1$. Since we restrict ourselves to simply connected Lie groups (indeed, Lie groups that are diffeomorphic to $\mathbb{R}^3$), this means that the corresponding Lie groups are isomorphic as well.
Now we recall an essential fact from Riemannian geometry. Given a smooth manifold with a Riemannian metric, there exists a unique torsion-free connection which is compatible with the metric, which is called the Levi-Civita connection. This can be easily proven with the following lemma, the proof of which may be found in \cite{K}.
\begin{lemma}[Koszul Formula]
Let $\nabla$ be a torsion-free, metric connection on a Riemannian manifold $(M,g)$. Then, for any vector fields $X,Y,$ and $Z,$ we have:
$$2g(\nabla_X Y, Z)= X(g(Y,Z))+Y(g(X,Z))-Z(g(X,Y))+$$
$$+g([X,Y],Z)-g([X,Z],Y)-g([Y,Z],X)$$
\end{lemma}
Using the Koszul Formula, we can easily compute the Levi-Civita connection $\nabla$ for each $G_\alpha$ from equation $(18)$. We get:
\begin{proposition}
The Levi-Civita connection of $G_\alpha$, with its left-invariant metric, is completely determined by
$$\begin{pmatrix} \nabla_X X && \nabla_X Y&&\nabla_X Z\\ \nabla_Y X && \nabla_Y Y&&\nabla_Y Z\\ \nabla_Z X && \nabla_Z Y&&\nabla_Z Z\end{pmatrix}=\begin{pmatrix} Z && 0&&-X\\ 0 && -\alpha Z&&\alpha Y\\ 0 && 0&&0\end{pmatrix}$$
where $\{X,Y,Z\}$ is the orthonormal basis of the Lie algebra, as in $(1)$.
\end{proposition}
The coordinate planes play a special role in the geometry of $G_\alpha.$ Each group (which is diffeomorphic to $\mathbb{R}^3$) has three foliations by the $XZ,$ $YZ,$ and $XY$ planes. It will be worthwhile to compute the curvatures of these surfaces in $G_\alpha$. The sectional curvature can be computed easily from the Levi-Civita Connection, while the extrinsic (Gaussian) curvature and mean curvature are computed using the Weingarten equation. Lastly, for surfaces in a Riemannian 3-manifold we have the relation
\begin{equation} Intrinsic = Extrinsic + Sectional \end{equation}
from Gauss' Theorema Egregium. For details, see the first chapter in \cite{N}. Straightforward computations yield the following proposition and its immediate consequences.
\begin{proposition} The relevant curvatures of the coordinate planes:
\begin{center}
\begin{tabular}{|c c c c c|}
\hline
Plane& Sectional & Intrinsic & Extrinsic (Gaussian) & Mean \\
\hline\hline
XY& $\alpha$ & 0 & $-\alpha$ & $(1-\alpha)/2$ \\
\hline
XZ& -1 & -1 & 0 &0 \\
\hline
YZ& $-\alpha^2$ & $-\alpha^2$ & 0 &0\\
\hline
\end{tabular}
\end{center}
\end{proposition}
\begin{corollary}
The $XY$ plane is a minimal surface (having vanishing mean curvature) in $G_\alpha$ if and only if $\alpha=1$, and the $XZ$ and $YZ$ planes are minimal for all $\alpha$. Also, the $XY$ plane is always a constant-mean-curvature surface.
\end{corollary}
We will later strengthen this corollary and get that the $XZ$ and $YZ$ planes are geodesically embedded.
It can be easily seen that $G_1$ is a model for the Sol geometry, $G_0$ is a model of $\mathbb{H}^2\cross\mathbb{R}$, and $G_{-1}$ is a model of $\mathbb{H}^3,$ or hyperbolic space. We can also compute the Ricci and scalar curvatures as well and we get that the scalar curvature of $G_\alpha$ is $S_\alpha=2\alpha-2-2\alpha^2$. $S_\alpha$ attains its maximum over the family at $S_{1/2}=-3/2$, and the minimum is attained at $S_{-1}=-6$ (as might be expected). Therefore, the group $G_{1/2}$ may also be considered a special case: the member of the interpolation that maximizes scalar curvature. Moreover, $S_\alpha$ is symmetric in the positive side of the family, in the sense that for all non-negative $\alpha$, $S_{\alpha}=S_{1-\alpha}$.
Other self-evident properties of the coordinate planes could be stated, but we let the reader find these himself.
Now, we turn our attention to the geodesic flow of $G_\alpha$. Rather than attempting to derive analytic formulas for the geodesics from the geodesic equation, as done for Sol in \cite{T}, we restrict the geodesic flow to the unit-tangent bundle and consider the resulting vector field. The idea of restricting the geodesic flow to $S(G_1)$ for Sol was first explored by Grayson in his thesis \cite{G2} and then used by Richard Schwartz and the current author to characterize the cut locus of the origin of Sol in \cite{MS}. We recall that the cut locus of a point $p$ in a Riemannian manifold $(M,g)$ is the locus of points on geodesics starting at $p$ where the geodesics cease to be length minimizing.
Now, we extend the previous ideas to the other $G_\alpha$ groups. Indeed, consider $\frak{g}_\alpha$ and let $S(G_\alpha)$ be the unit sphere centered at the origin in $\mathfrak{g}_\alpha.$ Suppose that $\gamma(t)$ is a geodesic parametrized by arc length such that $\gamma(0)$ is the identity of $G$. Then, we can realize the development of $\gamma'(t)$, the tangent vector field along $\gamma$, as a curve on $S(G_\alpha),$ which will be the integral curve of a vector field on $S(G_\alpha)$ denoted by $\Sigma_\alpha.$ We can compute the vector field $\Sigma_\alpha$ explicitly.
\begin{proposition}
For the group $G_\alpha$ the vector field $\Sigma_\alpha$ is given by
$$\Sigma_{\alpha}(x,y,z)= (xz, -\alpha yz, \alpha y^2-x^2)$$
\end{proposition}
\begin{proof}
Since we are dealing with a homogeneous space (a Lie group) it suffices to examine the infinitesimal change of $V=\gamma'(0)=(x,y,z).$ We remark that parallel translation along $\gamma$ preserves $\gamma'$ because we have a geodesic, but parallel translation does not preserve the constant (w.r.t. the left-invariant orthonormal frame) vector field $V=\gamma'(0)$ along $\gamma$. Indeed, the infinitesimal change in the constant vector field $V$ as we parallel translate along $\gamma$ is precisely the covariant derivative of $V$ with respect to itself, or $\nabla_V V.$ Then, our vector field on $S(G_\alpha)$ is precisely:
$$\Sigma_\alpha=\nabla_V (\gamma' - V)=\nabla_V \gamma' -\nabla_V V=-\nabla_V V.$$
We also remark that since $V$ is a constant vector field, $\Sigma_\alpha$ is determined completely by the Levi-Civita connection that we previously computed, and this computation is elementary.
\end{proof}
We remark where the equilibria points of $\Sigma_\alpha$ are, since these correspond to straight-line geodesics. When $0<\alpha\leq 1,$ the equilibria are
$$\bigg(\pm\sqrt{\frac{\alpha}{1+\alpha}},\pm\sqrt{\frac{1}{1+\alpha}},0\bigg) \textrm{ and } (0,0,\pm1).$$
When $\alpha=0$, the set $\{X=0\}\cap S(G_\alpha)$ is an equator of equilibria, and when $\alpha<0,$ the only equilibria are at the poles.
A glance at $\Sigma_\alpha$ gets us our promised strengthening of Corollary 4:
\begin{corollary}
The $XZ$ and $YZ$ planes are geodesically embedded. The $XY$ is never geodesically embedded, even when it is a minimal surface (i.e. for $\alpha=1$).
\end{corollary}
Consider the complement of the union of the two planes $X=0$ and $Y=0$ in $\frak{g}_\alpha$. This is the union of four connected components, which we call \textit{sectors}. Since the $XZ$ and $YZ$ planes are geodesically embedded, we have
\begin{corollary}
The Riemannian exponential map, which we denote by $E$, preserves each sector of $\frak{g}_\alpha$. In particular, if $(x,y,z)\in \frak{g}_\alpha$ is such that $x,y>0,$ then $E(x,y,z)=(a,b,c)$ with $a,b>0$.
\end{corollary}
We will use Corollary 6 often and without mentioning it. We make the following key observation about $\Sigma_\alpha$.
\begin{proposition}
The integral curves of $\Sigma_\alpha$ are precisely the level sets of the function $H(x,y,z)=\abs{x}^\alpha y$ on the unit sphere.
\end{proposition}
\begin{proof}
Without loss of generality, we consider the positive sector. We recall that the symplectic gradient is the analogue in symplectic geometry of the gradient in Riemannian geometry. In the case of the sphere with standard symplectic structure, the symplectic gradient is defined by taking the gradient of the function $H$ (on the sphere) and rotating it 90 degrees counterclockwise. Doing this computation for $H,$ yields $\nabla_{sym}H=x^{\alpha-1}\cdot\Sigma_\alpha(x,y,z).$ Since this vector field is the same as the structure field up to a scalar function, the desired property follows.
\end{proof}
\begin{remark}
$\Sigma_\alpha$ is a Hamiltonian system in these coordinates if and only if $\alpha=1,$ i.e. for Sol ($G_1$)
\end{remark}
We finish this section with a conjecture that, if true, provides some connection between the groups in the $G_\alpha$ family, for $\alpha \in [-1,1]$. We recall that the \textit{volume entropy}, $h$, of a homogeneous Riemannian manifold $(M,g)$ is a measure of the volume growth in $M$. We can define
$$h(M,g):=\lim_{R\rightarrow \infty} \frac{\log(\textrm{Vol } B(R))}{R}$$
where $B(R)$ is a geodesic ball of radius $R$ in $M$. Since $G_{-1}$ is a model of Hyperbolic space and $G_{1}$ is Sol, we know that $h(G_{-1})=2$ and $h(G_{1})=1$ (see \cite{S}). Based on this, we conjecture that:
\begin{conjecture}
$h(G_\alpha)$ is a monotonically decreasing function of $\alpha$ for $\alpha \in [-1,1]$.
\end{conjecture}
\section{The Positive Alpha Family}
In this section, we start to explore the positive $\alpha$ side of the family. These geometries exhibit common behaviors such as geodesics always lying on certain cylinders, spiraling around in a "periodic-drift" manner. A natural way to classify vectors in the Lie algebra is by how much the associated geodesic segment under the Riemannian exponential map spirals around its associated cylinder. This classification allows us to discern how the exponential map behaves with great detail.
\subsection{Grayson Cylinders and Period Functions}
More than a few of our theorems in Section 2 may be considered generalizations of results in \cite{G2} and \cite{MS}. To begin our analysis, we study the \textit{Grayson Cylinders} of the $G_\alpha$ groups.
\begin{definition}
We call the level sets of $H(x,y,z)=|x|^\alpha y$ that are closed curves \textbf{loop level sets}.
\end{definition}
The proof of the following theorem follows a method first used in \cite{G2} for Sol.
\begin{theorem}[The Grayson Cylinder Theorem]
Any geodesic with initial tangent vector on the same loop level set as
$$\bigg(\beta\sqrt{\frac{\alpha}{1+\alpha}}, \frac{\beta}{\sqrt{1+\alpha}}, \sqrt{1-\beta^2}\bigg), \beta\in[0,1]$$
lies on the cylinder given by
$$w^2+e^{2z}+\frac{1}{\alpha}e^{-2\alpha z}=\frac{1+\alpha}{\alpha}\cdot\frac{1}{\beta^2}$$
where $w=x-y\sqrt{\alpha}$. We call these cylinders "Grayson Cylinders".
\end{theorem}
If we consider Grayson Cylinders as regular surfaces in $\mathbb{R}^3$ with the ordinary Euclidean metric, then a simple derivation of their first and second fundamental forms reveals that they are surfaces with Gaussian curvature identically equal to zero. Hence, they are locally isometric to ordinary cylinders, and, because they are also diffeomorphic to ordinary cylinders, Grayson Cylinders are in fact isometric to ordinary cylinders, for all choices of $\alpha$ and $\beta$.
It is easier to gauge the shape of a Grayson Cylinder by looking at its projection onto the planes normal to the line $x-y\sqrt{\alpha}$, or, alternatively, as the implicit plot of a function of the two variables $w$ and $z$ defined in the statement of Theorem 8. It appears that as $\alpha$ is fixed, the Grayson Cylinders limit to two "hyperbolic slabs" as $\beta$ goes to zero. Alternatively, as $\beta$ is fixed and $\alpha$ varies, it appears that one side of the Grayson Cylinder is ballooning outwards. In Figures 1 and 2, we have some examples generated with the Mathematica code provided at the end of this chapter.
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA1}
\caption{$\alpha=1$ and $\beta=1/2$.}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA25}
\caption{$\alpha=1/4$ and $\beta=1/2$.}
\end{subfigure}
\caption{Slices of Grayson Cylinders with varying $\alpha$}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA121}
\caption{$\alpha=1/2$ and $\beta=1/2$.}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{GCA122}
\caption{$\alpha=1/2$ and $\beta=1/5$.}
\end{subfigure}
\caption{Slices of Grayson Cylinders with varying $\beta$}
\end{figure}
We denote loop level sets by $\lambda$. Each loop level set has an associated period, $P_\lambda,$ which is the time it takes for a flowline to go exactly once around $\lambda$, and it suffices to compute the period at one vector in a loop level set to know $P_\lambda$. We can compare $P_\lambda$ to the length $T$ of a geodesic segment $\gamma$ associated to a flowline that starts at some point of $\lambda$ and flows for time $T.$ We call $\gamma$ $\textit{small, perfect,}$ or $large$ whenever $T<P_\lambda,$ $T=P_\lambda,$ or $T>P_\lambda,$ respectively. It seems that this "classification" of geodesic segments in $G_\alpha$ is ideal. For instance, it was shown in \cite{MS} that a geodesic in Sol is length-minimizing if and only if it is small or perfect. We now derive an integral formula for $P_\lambda$ and simplify the integral in two special cases.
\begin{proposition}
Let $\lambda$ be the loop level set associated to the vector
$$V_\beta=\bigg(\beta\sqrt{\frac{\alpha}{1+\alpha}}, \frac{\beta}{\sqrt{1+\alpha}}, \sqrt{1-\beta^2}\bigg),$$
then
$$P_\lambda(\beta) = \int_{-t_1}^{t_0} \frac{2dt}{\sqrt{1-\frac{\beta^2}{\alpha+1}(\alpha e^{2t}+e^{-2\alpha t})}}$$
where $t_0$ and $t_1$ are the times it takes to flow from $V_\beta$ to the equator of $S(G_\alpha)$ in the direction of, and opposite to the flow of $\lambda$, respectively.
\end{proposition}
\begin{remark}
The times $t_0$ and $t_1$ are precisely when the flat flow lines hit the unit circle, or when
\begin{equation}\alpha e^{2t_0}+e^{-2\alpha t_0}=\frac{\alpha+1}{\beta^2} \textrm{ and } \alpha e^{-2t_1}+e^{2\alpha t_1}=\frac{\alpha+1}{\beta^2}\end{equation}
\end{remark}
Stephen Miller helped us to numerically compute the period function for any choice of positive $\alpha$. An explicit formula for the period function in Sol ($G_1$) was derived in \cite{MS} and \cite{T}. It is:
$$P_\lambda(\beta)=\frac{4}{\sqrt{1+\beta^2}}\cdot K\bigg(\frac{1-\beta^2}{1+\beta^2}\bigg)$$
where $K(m)$ is the complete elliptic integral of the first kind, with the parameter as in Mathematica.
A closed-form expression of $P_\lambda$ can also be obtained for $G_{1/2}$. Since elliptic integrals have been studied extensively and many of their properties are well-known, the following expression allows us to analyze $P_\lambda$ more easily.
\begin{corollary}
When $\alpha=1/2,$ or for the group $G_{1/2}$, the period function is given by
$$P_\lambda(\beta)=\frac{4\sqrt{3}}{\beta\sqrt{e^{t_0-t_1}+2e^{t_1}}}\cdot K\bigg(\frac{2(e^{t_1}-e^{-t_0})}{e^{t_0-t_1}+2e^{t_1}}\bigg)$$
where
$$t_0=\log\bigg(\frac{1}{\beta}\cdot\bigg(\frac{1}{(-\beta^3 + \sqrt{-1 + \beta^6})^{\frac{1}{3}}} + (-\beta^3 +
\sqrt{-1 + \beta^6})^{\frac{1}{3}}\bigg)\bigg)$$
and
$$t_1=\log\bigg(\frac{1}{2}\bigg(\frac{1}{\beta^2}+\frac{1}{\beta^4(-2+\frac{1}{\beta^6}+\frac{2\sqrt{-1+\beta^6}}{\beta^3})^{\frac{1}{3}}}+(-2+\frac{1}{\beta^6}+\frac{2\sqrt{-1+\beta^6}}{\beta^3})^{\frac{1}{3}}\bigg)\bigg)$$
\end{corollary}
Here we state an essential fact, which is forthrightly supplied to us by the above expression of $P_\lambda(\beta)$ in terms of an elliptic integral. We have:
\begin{proposition}
$$\frac{d}{d\beta}\big(P_\lambda(\beta)\big)<0$$
when $\alpha=1$ and $\alpha=1/2$. Moreover, for $\alpha=1$,
$$\lim_{\beta\to 1} P_\lambda(\beta)=\pi\sqrt{2}$$
and for $\alpha=1/2$,
$$\lim_{\beta\to 1} P_\lambda(\beta)=2\pi.$$
\end{proposition}
We could not find a similar formula for $P_\lambda$, when $\alpha$ is not $1$ or $1/2$, in terms of elliptic integrals or hypergeometric functions. This is unfortunate, as Proposition 10 is vital to the method here and in \cite{MS} to characterize the cut locus of $G_1$ and $G_{1/2}$. However, we can still numerically compute the period function as presented at the end of this chapter, and this allows us to conjecture:
\begin{conjecture} Let $P(\beta)$ be the period function in $G_\alpha$, then
$$\frac{d}{d\beta}\big(P_\lambda(\beta)\big)<0$$
and
$$\lim_{\beta\to 1} P(\beta)=\frac{\pi\sqrt{2}}{\sqrt{\alpha}}.$$
\end{conjecture}
This conjecture would lend something quantitative to the idea that the bad behavior (or the cut locus) of $G_\alpha$ dissipates at infinity as Sol interpolates to $\mathbb{H}^2\cross\mathbb{R}$. With Stephen Miller's Mathematica code, we get the following numerical evidence for Conjecture 3.
\begin{center}
\begin{tabular}{|c c c|}
\hline
$\alpha$ & Numerical Value of $P(\alpha,\beta=.999)$ & $\pi\sqrt{2}/\sqrt{\alpha}$ \\
\hline\hline
0.1 & 14.0792 & 14.0496 \\
\hline
0.2 & 9.94735 & 9.93459 \\
\hline
0.3 & 8.11985 & 8.11156 \\
\hline
0.4 & 7.03114 & 7.02481 \\
\hline
0.5 & 6.28842 & 6.28319 \\
\hline
0.6 & 5.7403 & 5.73574 \\
\hline
0.7 & 5.31436 & 5.31026 \\
\hline
0.8 & 4.97106 & 4.96729 \\
\hline
0.9 & 4.68673 & 4.68321 \\
\hline
1. & 4.44622 & 4.44288 \\
\hline
\end{tabular}
\end{center}
\subsection{Concatenation and Some Other Useful Facts}
An important property which extends from Sol to $G_\alpha$ for $0<\alpha<1$ is that the loop level sets are symmetric with respect to the plane $Z=0.$ This simple observation allows the technique of $\it{concatenation},$ essential to the analysis in \cite{MS}, to be replicated for all of the $G_\alpha$ groups, when $\alpha>0$. For the interested reader, Richard Schwartz's Java program \cite{S2} uses concatenation to generate geodesics and geodesic spheres in Sol, and a modified version of this program can generate the spheres and geodesics in any $G_\alpha$ group as well as in other Lie groups, such as Nil. As an illustration of the power of this technique in numerical simulations, we present in Figure 3 the geodesic spheres of radius around $5$ in four different geometries. The spheres are presented from the same angle, the purple line is the $z$ axis, and the red lines are the horizontal axes. The salient phenomenon is that one "lobe" of the sphere is contracting as $\alpha$ goes to zero. Qualitatively, this corresponds to the dissipation of the "bad" behavior (or the cut locus) of $G_\alpha$ as $\alpha$ tends to zero, since the amount of shear diminishes.
\begin{figure}[H]
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUP1}
\caption{In Sol, or the group $G_1$.}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUP34}
\caption{In the group $G_{3/4}$.}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUP12}
\caption{In the group $G_{1/2}$}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=0.9\linewidth]{GROUPH2R}
\caption{In the group $G_0$, or $\mathbb{H}^2\cross \mathbb{R}$.}
\end{subfigure}
\caption{Geodesic spheres of radius around $5$ in four different geometries.}
\end{figure}
Each flowline $\lambda$ of our vector field corresponds to a segment of a geodesic $\gamma$. Let $T$ be the time it takes to trace out $\lambda,$ then $T$ is exactly the length of $\gamma$ since we always take unit-speed geodesics. Let $L_\lambda$ be the far endpoint of $\gamma$ and consider the equally spaced times
$$0=t_0<t_1<\ldots <t_n=T$$
with corresponding points $\lambda_0, \ldots, \lambda_n$ along $\lambda.$ Then we have
$$L_\lambda=\lim_{n\to\infty} (\epsilon_n \lambda_0)\ast\ldots\ast(\epsilon_n \lambda_n),\quad \epsilon_n=T/(n+1)$$
where $\ast$ is the group law in $G_\alpha.$ The above equation is well-defined because the underlying space of both $G_\alpha$ and its Lie algebra is $\mathbb{R}^3.$ We will also use the notation $\lambda=a|b$ to indicate that we are splitting $\lambda$ into two sub-trajectories, $a$ and $b.$
The above equation yields:
$$L_\lambda=L_a \ast L_b$$
when $\lambda=a|b.$
We set $\epsilon_n \lambda_j=(x_{n,j},y_{n,j},z_{n,j})$.
Vertical displacements commute in $G_\alpha,$ therefore the third coordinate of
the far endpoint of $\gamma$ is given by
$$\lim_{n \to \infty} \sum_{j=0}^n z_{n,j}.$$
From this, and the symmetry of the flow lines with respect to $Z=0,$ we get the following lemmas. Since these results are an immediate extension of the results in \cite{MS}, we provide only sketches of their proofs.
\begin{lemma}
If the map $(x,y,z) \to (x,y,-z)$ exchanges the
two endpoints of the flowline $\lambda$ then the endpoints of the geodesic segment $\gamma$
both lie in the plane $Z=0$ and $L_\lambda$ is a
horizontal translation. In this case we call
$\lambda$ {\it symmetric\/}.
\end{lemma}
\begin{proof}
Since $\lambda$ is symmetric, the sum $\lim_{n \to \infty} \sum_{j=0}^n z_{n,j}$ vanishes, so the total vertical displacement is zero.
\end{proof}
\begin{lemma}
If $\lambda$ is not symmetric then we can write
$\lambda=a|b|c$ where $a,c$ are either symmetric
or empty, and $b$ lies entirely above or entirely
below the plane $Z=0$. Since $L_a$ and
$L_c$ are horizontal translations -- or just
the identity in the empty cases -- and
$L_b$ is not such a translation, the endpoints of $\lambda$ are
not in the same horizontal plane.
\end{lemma}
\begin{lemma} If $\lambda=a|b$, where both $a$ and $b$ are
symmetric, then both endpoints of
$\gamma$ lie in the plane $Z=0$. We can do this whenever
$\lambda$ is one full period of a loop
level set. Hence, a perfect
geodesic segment has both endpoints in the same
horizontal plane.
\end{lemma}
\begin{proof}
Since $\lambda=a|b$, we know that $L_\lambda=L_a\ast L_b$. Since $L_a$ and $L_b$ are horizontal translations by Lemma 16, it follows that $L_\lambda$ stays in the same $Z=0$ plane.
\end{proof}
\begin{lemma} If $\lambda_1$ and $\lambda_2$ are full trajectories of the same
loop level set, then
we can write $\lambda_1=a|b$ and $\lambda_2=b|a$, which leads to
$L_{\lambda_2}=L_a^{-1}L_{\lambda_1}L_a$. Working this out with the group law in $G_\alpha$ gives:
$(a_1 e^{-z}, b_1 e^{\alpha z}, 0)=(a_2, b_2, 0)$, where $(x,y,z)=L_a$ and $(a_i, b_i, 0)=L_{\lambda_i}$.
In particular, we have $a_1^\alpha b_1=a_2^\alpha b_2$ and we call the function (of the flowlines)
$H_\lambda=\sqrt{|a_1^\alpha b_1|}$ the {\it holonomy invariant\/} of
the loop level set $\lambda.$
\end{lemma}
Let $E$ be the Riemannian
exponential map.
We call $V_+=(x,y,z)$ and $V_-=(x,y,-z)$, vectors in the Lie algebra, {\it partners\/}.
The symmetric trajectories discussed in Lemma 16 have endpoints
which are partners.
Note that if $V_+$ and $V_-$ are partners, then
one is perfect if and only if the other one is, because they lie on the same loop level set. The next two facts are generalizations of results from \cite{MS} about Sol, and, in particular, Corollary 8 proves half of our main conjecture.
\begin{theorem}
If $V_+$ and $V_-$ are perfect partners, then
$E(V_+)=E(V_-)$.
\end{theorem}
\begin{proof}
Let $\lambda_{\pm}$ be the trajectory which
makes one circuit around the loop level set starting
at $U_{\pm}$. As above, we write the flowlines as
$\lambda_+=a|b$ and
$\lambda_-=b|a$. Since $V_+$ and $V_-$ are partners, we can take
$a$ and $b$ both to be symmetric.
But then the elements
$L_a$, $L_b$, $L_{\lambda_1}$, $L_{\lambda_2}$
all preserve the plane $Z=0$ and hence mutually commute, by Lemma 16.
By Lemma 19, we have
$L_{\lambda_+}=L_{\lambda_-}$.
But $E(V_{\pm})=L_{\lambda_{\pm}}$.
\end{proof}
\begin{corollary}
A large geodesic segment is not a length minimizer.
\end{corollary}
\begin{proof}
If this is false then, by shortening our geodesic, we
can find a perfect geodesic segment $\gamma$, corresponding
to a perfect vector $V=(x,y,z)$, which is a unique
geodesic minimizer without conjugate points. If $z \not =0$
we immediately contradict Theorem 9.
If $z=0$, we consider the variation,
$\epsilon \to \gamma(\epsilon)$ through same-length
perfect geodesic segments
$\gamma(\epsilon)$ corresponding to
the vector $V_{\epsilon}=(x_{\epsilon},y_{\epsilon},\epsilon)$.
The vectors $V_{\epsilon}$ and $V_{-\epsilon}$ are partners, so
$\gamma(\epsilon)$ and $\gamma(-\epsilon)$
have the same endpoint. Hence, this variation
corresponds to a conjugate point on $\gamma$
and again we have a contradiction.
\end{proof}
The next step is to analyze what happens for small and perfect geodesics. To begin, we point out another consequence of Theorem 9. Let $M$ be the set of vectors in the Lie algebra of $G_\alpha$ associated to small geodesic segments and let $\partial M$ be the set of vectors associated to perfect geodesic segments. Lastly, let $\partial_0 M$ be the intersection of $\partial M$ with the plane $Z=0.$ Since $E$ identifies perfect partner vectors, we have a vanishing Jacobi field at each point of $\partial_0 M.$ However, we still have:
\begin{proposition}
$dE$ is nonsingular in $\partial M - \partial_0 M$
\end{proposition}
Another useful consequence of what we have heretofore shown is that the Holonomy function (defined in Lemma 19) is monotonically increasing. The following result will be useful later, when we analyze the behavior of $E$ on the set of perfect vectors. More precisely, we have:
\begin{proposition}
Let $P$ be the unique period associated to a flowline $\lambda$. We know that the holonomy $H_\lambda$ is an invariant of the flowline, so it is a function of $P$. Moreover, $H_\lambda$ varies monotonically with the flowlines. Explicitly, we have $\frac{dH}{dP}(P)>0$.
\end{proposition}
To begin our analysis of the small geodesic segments, we prove an interesting generalization of the \textit{Reciprocity Lemma} from \cite{MS}. If $V$ is a perfect vector, then $E(V)$ will lie in the $Z=0$ plane by Lemma 18; however, we can get something better.
\begin{theorem}
Let $V=(x,y,z)$ be a perfect vector. There exists a number $\mu\neq 0$ such that $E(V)=\mu(\alpha y,x,0).$
\end{theorem}
\subsection{Symmetric Flowlines}
We now introduce another technique introduced first in \cite{MS}: the emphasis on symmetric flow lines, which is justified by our previous lemmas. We introduce the following sets in $\frak{g}_\alpha$ and $G_\alpha$:
\begin{itemize}
\item Let $M, \partial M \subset \frak{g}_\alpha$ be the set of small and perfect vectors, as previously defined.
\item Let $\Pi$ be the $XY$ plane in $\frak{g}_\alpha$ and $\tilde{\Pi}$ be the $XY$ plane in $G_\alpha$.
\item Let $\partial_0 M=\partial M \cap \Pi$.
\item Let $M^{symm} \subset M$ be those small vectors which correspond to symmetric flowlines.
\item Let $\partial_0 N=E(\partial_0 M)$.
\item Let $\partial N$ be the complement, in $\tilde{\Pi}$, of the
component of $\tilde{\Pi}-\partial_0 N$ that contains the origin.
\item Let $N=G_\alpha-\partial N$.
\item For any set $A$ in either $\frak{g}_\alpha$ or $G_\alpha,$ we denote $A_+$ to be the elements of $A$ in the positive sector, where $x,y>0$.
\end{itemize}
Our underlying goal is to show that $\partial N$ is the cut locus of the origin in $G_\alpha$. This has already been done for Sol ($G_1$) in \cite{MS}, and we shall prove the same for $G_{1/2}$. Thus, although the notation we use here is suggestive of certain topological relationships (e.g. is $\partial N$ the topological boundary of $N$?), we are only able to prove these relationships for $G_{1/2}$ in the present paper. Reflections across the $XZ$ and $YZ$ planes are isometries in every $G_\alpha$, so proving something for the positive sector (where $x,y>0$) proves the same result for every sector. This is useful in simplifying many proofs.
A first step towards proving that the cut locus is $\partial N$ is to show that $$E(M) \cap \partial N =\emptyset,$$ or, intuitively, that the exponential map "separates" small and perfect vectors. The following lemma is a step towards this.
\begin{lemma}
If $E(M) \cap \partial N \not = \emptyset$, then
$E(M_+^{symm}) \cap \partial N_+ \not = \emptyset$.
\end{lemma}
With this lemma in hand, we should analyze the symmetric flowlines in detail in order to prove that $E(M_+^{symm}) \cap \partial N_+= \emptyset$. Symmetric flowlines are governed by a certain system of nonlinear ordinary differential equations. Let $\Theta_P^+$ denote those points in the
(unique in the positive sector) loop level set of period $P$ having
all coordinates positive. Every element
of $M_+^{{\rm symm\/}}$ corresponds to a small symmetric
flowline starting in $\Theta_P^+$.
\newline
\newline
{\bf The Canonical Parametrization:\/}
The set $\Theta_P^+$ is an open arc.
We fix a period $P$ and we set $\rho=P/2$.
Let $p_0=(x(0),y(0),0) \in \Theta_P \cap \Pi$ be the point with $x(0)>y(0)$. The initial value $x(0)$ varies from $\sqrt{(\alpha+1)/\alpha}$ to $1$.
We then let
\begin{equation}
p_t=(x(t),y(t),z(t))
\end{equation}
be the point on $\Theta_P^+$ which we reach after time $t \in (0,\rho)$ by
flowing {\it backwards\/} along the structure field
$\Sigma$. That is
\begin{equation}
\label{backwards}
\frac{dp}{dt}=
(x',y',z')=-\Sigma(x,y,z)=(-xz,+\alpha yz,x^2-\alpha y^2).
\end{equation}
Henceforth, we use the notation $x'$ to stand for $dx/dt$, etc.
\newline
\newline
{\bf The Associated Flowlines:\/}
We let $\hat{p}_t$ be the partner of $p_t$, namely
\begin{equation}
\hat{p}_t=(x(t),y(t),-z(t)).
\end{equation}
We let $\lambda_t$ be the small symmetric
flowline having endpoints $p_t$ and $\hat{p}_t$.
Since the structure field $\Sigma$
points downward at $p_0$, the symmetric flowline $\lambda_t$
starts out small and increases all the way to a perfect
flowline as $t$ increases from $0$ to $\rho$.
We call the limiting perfect flowline $\lambda_{\rho}$.
\newline
\newline
\noindent
{\bf The Associated Plane Curves:\/}
Let $V_t \in M_+^{{\rm symm\/}}$ be the vector
corresponding to $\lambda_t$. (Recall that $E(V_t)=L_{\lambda_t}$)
Define
\begin{equation}
\Lambda_P(t):=E(V_t)=(a(t),b(t),0) \hskip 30 pt t \in (0,\rho].
\end{equation}
These plane curves are in $\tilde{\Pi}$ because they are endpoints of symmetric flowlines, and they will be among our main objects of interest in what follows. In Figure 4, we present a collection of the plane curves (colored blue) for $\alpha=1/2$ with the choice of $x(0)$ varying from $0.6$ to $0.95$ at intervals of $0.05$. We also include the initial value $x_0=1/\sqrt{3}$, which corresponds to the straight geodesic segment in $G_{1/2}$. The black curve is an approximation of $\partial_0 N_+$, or endpoints of perfect flowlines, which are the right-hand endpoints of each $\Lambda_P$ curve.
\begin{figure}[h!]
\centering
\includegraphics[width=1\textwidth]{SYMMPLOT1}
\caption{The image of $\Lambda_P$ over the interval $(0,\rho]$ for varying $x_0$ and $\partial_0 N_+$. }
\end{figure}
\begin{lemma}
\label{endpoint}
$\Lambda_P(\rho) \in \partial_0 N_+$, and $0<b(\rho)<a(\rho)$.
\end{lemma}
We have $E(M_+^{{\rm symm\/}}) \cap \partial N_+=\emptyset$ provided that
\begin{equation}
\label{goal}
\Lambda_P(0,\rho) \cap \partial N_+=\emptyset,
\textrm{ for all periods } P.
\end{equation}
So all we have to do is establish Equation 25. Let $B_P$ be the rectangle in the $XY$ plane with vertices
$$(0,0,0), (0,b(\rho),0), (a(\rho),0,0), \textrm{ and } (a(\rho),b(\rho),0).$$
Our first step in proving Equation 25 is to contain the image of $\Lambda_P$ with the following theorem, which we will prove to be true for each $G_\alpha$ group:
\begin{theorem}[The Bounding Box Theorem]
$\Lambda_P(0,\rho) \subset {\rm interior\/}(B_P)$ for all $P$.
\end{theorem}
The Bounding Triangle Theorem serves a similar role in \cite{MS} for Sol $(G_1)$, but it cannot be generalized to any other $G_{\alpha}$ group. It states that $\Lambda_P(0,\rho)$ is contained inside the triangle with vertices $(0,0,0), (a(\rho),0,0)$, and $(a(\rho),b(\rho),0)$. In Figure 6, we depict the image of a single plane curve $\Lambda_P$ for $\alpha=1/2$ and $x_0=0.99945$, which illustrates the failure of the Bounding Triangle Theorem in the other Lie groups.
Now, if we could also manage to show ${\rm interior\/}(B_P)\cap \partial N_+ =\emptyset$, we would finish proving Equation 25. Since the Bounding Box Theorem is not as powerful as the Bounding Triangle theorem of \cite{MS}, we need more information about $\partial N_0$ to prove Equation 25 than was needed in \cite{MS}. We succeed in performing this second step for the group $G_{1/2}$ by getting bounds on the derivative of the period function (using its expression in terms of an elliptic integral in that case). The necessary ingredient that we get is
\begin{theorem*}[The Monotonicity Theorem]
For $\alpha=1/2$, $\partial_0 N_+$ is the graph of a non-increasing function (in Cartesian coordinates).
\end{theorem*}
\subsection{Proof of the Main Results for $G_{1/2}$}
For $G_{1/2}$, assuming that the Bounding Box and Monotonicity Theorems are true, we can proceed to characterize the cut locus of the identity.
First, we prove equation $(25)$:
\begin{theorem}
For the group $G_{1/2}$ we have, for all $P$,
$$\Lambda_P(0,\rho) \cap \partial N_+ = \emptyset$$
\end{theorem}
\begin{proof}
By the Bounding Box Theorem we know that $\Lambda_P(0,\rho) \subset {\rm interior\/}(B_P)$ for all $P$. By the Monotonicity Theorem, we know that $\partial_0 N_+$ is the graph of a decreasing function in Cartesian coordinates, so we conclude that $\partial N_+$ is disjoint from ${\rm interior\/}(B_P)$ for all $P$. Our desired result holds.
\end{proof}
The above theorem, combined with Lemma 20 gets us:
\begin{corollary}
\label{smallperfect}
$$E(M) \cap \partial N=\emptyset$$
\end{corollary}
The rest of our argument for showing that the cut locus of $G_{1/2}$ is $\partial N$ follows exactly as in \cite{MS}.
Let $E$ be Riemannian exponential map.
Let $\mathcal M$ be the component
of $\partial M_+-\partial_0M_+$ which contains
vectors with all coordinates positive.
Let ${\mathcal N\/}=\partial N_+ - \partial_0 N_+$. We first prove a few lemmas.
\begin{lemma}
The map $E$ is injective on $\mathcal{M}$.
\end{lemma}
\begin{proof}
Let $V_1$ and $V_2$ be two vectors in $\mathcal{M}$ such that $E(V_1)=E(V_2)$. We also let $U_1=E(V_1)$ and $U_2=E(V_2)$ and denote the $j^{th}$ coordinate of $U_i$ as $U_{ij}$ and likewise for $\frac{V_i}{\|V_i\|}$. Since $U_1$ and $U_2$ have the same holonomy invariant and since the holonomy is monotonic with respect to choice of flowline, it follows that $\frac{V_1}{\|V_1\|}$ and $\frac{V_2}{\|V_2\|}$ lie on the same loop level set in $S(G_{1/2})$. Thus, $V_{11}V_{12}^2=V_{21}V_{22}^2$. By the Reciprocity Lemma, and since $U_1=U_2$, we get
$$\frac{V_{12}}{V_{11}}=\frac{V_{22}}{V_{21}}.$$
We can now conclude that $V_{11}=V_{21}$ and $V_{12}=V_{22}$. Since $\|V_1\|=\|V_2\|$, we get $V_1=V_2$, finishing the proof.
\end{proof}
\begin{lemma}
\label{SP0}
$E({\mathcal M\/}) \subset \mathcal N$.
\end{lemma}
\begin{proof}
The map $E$ is injective
on ${\mathcal M\/} \cup \partial_0 M_+$, by the previous lemma. At the same
time, $E(\partial_0 M_+)=\partial_0 N_+$.
Hence
\begin{equation}
\label{alternative}
E({\mathcal M\/}) \subset \Pi - \partial_0 N_+.
\end{equation}
By definition, $\mathcal N$ is one of the components of the
$\Pi-\partial_0 N_+$. Therefore,
since $\mathcal M$ is connected, the image
$E({\mathcal M\/})$ is either contained in $\mathcal N$ or disjoint from $\mathcal N$.
Since the sets are evidently not disjoint (large perfect vectors land far away from the identity and near the line $x=y/\sqrt{2}$), we have containment.
\end{proof}
\begin{corollary}
\label{SP}
$E(\partial M) \cap E(M)=\emptyset$.
\end{corollary}
\begin{proof}
Up to symmetry, every vector in $\partial M$ lies either in
$\mathcal M$ or in $\partial_0 M_+$.
By definition, $E(\partial_0 M)=\partial_0 N \subset \partial N$. So,
by the previous result, we have
$E(\partial M) \subset \partial N$.
By Corollary \ref{smallperfect} we have
$E(M) \cap \partial N=\emptyset$.
Combining these two statements gives the result.
\end{proof}
\begin{theorem}
\label{minimi}
Perfect geodesic segments are length minimizing.
\end{theorem}
\begin{proof}
Suppose $V_1 \in \partial M$ and
$E(V_1)=E(V_2)$ for some $V_2$ with $\|V_2\|< \|V_1\|$.
By symmetries of $G_{1/2}$ and the flowlines, we can assume that both $V_1$ and $V_2$ are in the positive sector and that their third coordinates are also positive.
By Corollary 8, we have $V_2 \in M \cup \partial M$.
By Corollary \ref{SP} we have $V_2 \in \mathcal{M}$.
But then $V_1=V_2$, by Lemma 22, which contradicts $\|V_2\|< \|V_1\|$.
\end{proof}
The results above identify $\partial N$ as the cut locus of the
identity of $G_{1/2}$ just as obtained in \cite{MS} for Sol. We can summarize by saying
\begin{theorem}
A geodesic segment in $G_{1/2}$ is a length
minimizer if and only if it is small or perfect.
\end{theorem}
In addition, small geodesic segments are
unique length minimizers and they have no conjugate points.
Hence, using standard results about the cut locus, as in \cite{N}, we get that $E: M \to N$
is an injective, proper, local diffeomorphism. This implies
that $E: M \to N$ is also surjective and hence a diffeomorphism.
Moreover, $E: \partial_0 M_+ \to \partial_0 N_+$ is a diffeomorphism, by similar considerations. Results about the geodesic spheres in $G_{1/2}$ follow immediately, as in \cite{MS} for Sol, by "sewing up" $\partial M$ in a 2-1 fashion with $E$.
In particular, we have:
\begin{corollary}
Geodesic spheres in $G_{1/2}$ are always topological spheres.
\end{corollary}
The rest of this chapter is devoted to proving our technical results: the Bounding Box Theorem and the Monotonicity Theorem. We will prove the Bounding Box Theorem in full generality, i.e. for all $\alpha\in (0,1]$. However, we only manage to prove the Monotonicity Theorem for $G_{1/2}$, where we have an expression of the period in terms of an elliptic integral. It is our opinion that either an expression for $P$ in terms of hypergeometric functions exists for general $\alpha$ or a thorough analysis of the (novel?) integral function in Proposition 9 can be done to demonstrate the monotonicity results required. Regardless, our Bounding Box Theorem does half of the work necessary to finish the proof of our main conjecture: for all $G_\alpha$ groups, a geodesic segment is length minimizing if and only if it is small or perfect.
We reiterate that the necessary step to prove our conjecture is to show the Monotonicity Theorem holds for general $G_\alpha$ and that there is encouraging numerical evidence supporting this proposition. We plan to investigate this last step and prove our main conjecture in the future.
\begin{center}
\begin{figure}[H]
\vspace{1.5in}
\includegraphics[width=1\textwidth]{PNPlot}
\caption{Here, for $G_{1/2}$, we have plotted points on $\partial_0 N_+$, as $x_0$ varies from $0.6$ to $0.98$ in increments of $0.02$.}
\end{figure}
\end{center}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{BoundingT}
\caption{This depicts a the image of $\Lambda_P$ for $\alpha=1/2$ and $x_0=0.99945$ over the interval $(0,\rho]$.}
\end{figure}
\section{Proof of the Bounding Box Theorem}
We now study the system of ODE's that governs the behavior of $x,y,z,a,$ and $b$ as in equation $(22)$.
We write
$\lambda_{t+\epsilon}=u|\lambda_t|v$,
where $u$ is the flowline
connecting $p_{t+\epsilon}$ to $p_t$ and $v$ is the flowline
connecting $\hat{p}_t$ to $\hat p_{t+\epsilon}$.
We have
$$
(a',b',0)=\Lambda'_P(t)=\lim_{\epsilon \to 0} \frac{\Lambda_P(t+\epsilon)-\Lambda(t)}{\epsilon},
$$
$$
\Lambda_P(t+\epsilon) \approx (\epsilon x, \epsilon y,\epsilon z) *
(a,b,0) * ( \epsilon x, \epsilon y,-\epsilon z).
$$
The approximation is true up to order $\epsilon^2$ and
$(*)$ denotes multiplication in $G_\alpha$.
A direct calculation gives
\begin{equation} a'=2x+az \textrm{ and } b'=2y-\alpha bz.\end{equation}
Simply from its differential equation, it is evident that $a'>0$ on $(0,\rho)$. This implies that
$\Lambda_P(t)$ is the graph of a function for $t\in (0,\rho)$, hence $\Lambda_P(t)$
avoids the vertical sides of $B_P$. This is the first, easy step in proving the Bounding Box Theorem.
To finish the proof, it would suffice to show that $\Lambda_P(t)$ also avoids the horizontal sides of $B_P$, which amounts to proving that $b'(t)>0$ for all $t\in(0,\rho]$. A priori, it is not evident that $b'>0$ in this interval. For example, the function $b$ may start out concave. Also, after the half-period $\rho$, $b'$ may actually be negative. However, the remarkable fact that $b'>0$ in $(0,\rho]$ for all choices of $\alpha$ and $x_0$ is also true, and we demonstrate this fact in what follows.
Once again, we collect the ODE's of interest to us together:
$$x'=-xz\quad y'=\alpha y z\quad z'=x^2-\alpha y^2\quad b'=2y-\alpha b z$$ from which we compute
\begin{equation}z''=-2z(x^2+\alpha y^2)\quad b''=\alpha b(\alpha z^2-z').\end{equation}
\begin{lemma}
To show that $b'>0$ in $(0,\rho),$ it suffices to show that $b'>0$ whenever $b''<0$ in the interval $(0,\rho).$
\end{lemma}
We observe that $z''<0$ everywhere in $(0,\rho)$. Also, whenever $b''<0,$ we have that $z'>\alpha z^2>0$. We first get an inequality regarding the function $z(t):$
\begin{lemma}
$$2\alpha \int_0^t z(s) ds\geq t\alpha z(t)$$
\end{lemma}
We now recall the Log-Convex Version of Hermite-Hadamard, proven first in \cite{GPP}:
\begin{proposition}[\cite{GPP}]
If $f$ is log-convex on $[a,b]$ then
$$\frac{1}{b-a}\int_a^b f(s) ds\leq \frac{f(b)-f(a)}{\log f(b)-\log f(a)}.$$
\end{proposition}
Let's return to one of our initial ODE's: $y'=\alpha yz.$ Dividing by $y,$ integrating, and multiplying by 2, we get:
\begin{equation}2\log y(t) -2\log y(0)= 2\alpha \int_0^t z(s) ds. \end{equation}
Since $z'>0$ whenever $b''<0,$ we get that $y^2$ is log-convex whenever $b''<0.$ We can now get:
\begin{lemma} For all $t$ where $b''(t)<0,$ we have
$$\int_0^t y(s)^2 ds \leq t\cdot\frac{y(t)^2}{2\log y(t)/y(0)}$$
\end{lemma}
We are now in a position to prove the Bounding Box Theorem:
\begin{proof}
By Lemma 24, it suffices to show that $b'>0$ whenever $b''<0.$ If we integrate the ODE for $b$, we get
\begin{equation} b(t)=\frac{2}{y(t)}\int_0^t y(s)^2 ds.\end{equation}
Differentiating, we want to show $y(t)^3-y'(t)\int_0^t y(s)^2 ds\geq 0$ whenever $b''<0.$ Equivalently (we can divide by $y,y'$ since they are always strictly greater than $0$):
$$ \int_0^t y(s)^2 ds\leq \frac{y(t)^3}{y'(t)}.$$
By Lemma 26 it suffices to show
$$t\cdot\frac{y(t)^2}{2\log y(t)/y(0)}\leq\frac{y(t)^3}{y'(t)}$$
or, by cancelling some terms and taking the reciprocal,
$$2\log y(t)/y(0) \geq t\cdot \frac{y'(t)}{y(t)}.$$
Since $y'=\alpha yz$, this is equivalent to
$$2\alpha \int_0^t z(s) ds\geq t\alpha z(t)$$
which is nothing but the inequality of Lemma 25.
\end{proof}
Finally, since $b'>0, \Lambda_P(t)$ is an increasing function, so $\Lambda_P(t)$ avoids the vertical sides of $B_P$ when $t\in(0,\rho)$. This, along with the previously stated fact that $\Lambda_P(t)$ avoids the horizontal sides of $B_P$, finishes the proof of the Bounding Box Theorem.
\section{Proof of the Monotonicity Theorem}
\subsection{Endpoints of Symmetric Flowlines}
Henceforth, we view $\partial_0 N_+$ as the following parametrized curve in the $XY$ plane. Denote $x(0)=x_0$, then
$$\partial_0 N_+= \{ (a_{x_0}(P(x_0)/2),b_{x_0}(P(x_0)/2),0)\}, \textrm{ as } x_0 \textrm{ varies in } \bigg(\sqrt{\frac{\alpha}{1+\alpha}},1\bigg).$$
To prove Equation $25$ in general it would suffice, by using the Bounding Box Theorem, to prove $B_P\cap \partial_+ N=\emptyset$, and, to prove the latter statement, it suffices to show that $\partial_0 N_+$ is the graph of a decreasing function in Cartesian coordinates. This involves differentiating our ODE's with respect to the initial value $x_0$. Let $\bar{x}$ denote $dx(t,x_0)/dx_0$, etc. Then we get
$$\bar{x}'=-x\bar{z}-z\bar{x},\quad \bar{y}'=\alpha y\bar{z}+\alpha z\bar{y},\quad \bar{z}'=2x\bar{x}-2\alpha y \bar{y},$$
$$\bar{a}'=2\bar{x}+a\bar{x}+x\bar{a}, \quad \bar{b}'=2\bar{y}-\alpha\bar{y}b-\alpha y\bar{b}.$$
Since $x^2+y^2+z^2=1$ for all $t, x_0,$ it follows that
\begin{equation}x\bar{x}+y\bar{y}+z\bar{z}=0\end{equation}
always, and we can get a similar equation for the time derivative. Now, we prove some very useful propositions:
\begin{proposition}
$ax-\alpha by = 2z$ for all $t$ and $x_0$.
\end{proposition}
Since the above equality is true for all $t$ and $x_0$ we can differentiate with respect to $x_0$ and get
\begin{corollary}
$a\bar{x}+x\bar{a}-\alpha b\bar{y}-\alpha y \bar{b}=2\bar{z}$ for all $t$ and $x_0$.
\end{corollary}
Now we prove:
\begin{proposition}
$x\bar{a}+y\bar{b}=0$ for all $t$ and $x_0$.
\end{proposition}
Corollary 12 and Proposition 14 combine to get us the following useful expressions for $\bar{a}$ and $\bar{b}$.
\begin{corollary}
We have
$$\bar{a}=\frac{1}{x(1+\alpha)}\bigg(2\bar{z}+\alpha b\bar{y}-a\bar{x}\bigg)$$
and
$$\bar{b}=-\frac{1}{y(1+\alpha)}\bigg(2\bar{z}+\alpha b\bar{y}-a\bar{x}\bigg).$$
\end{corollary}
\subsection{The Monotonicity Theorem for $G_{1/2}$}
To get our desired results about $\partial_0 N_+$, we must look at a particular case of our one-parameter family, where we have more information about the derivative of $P(x_0)$, courtesy of the expression of $P$ in terms of an elliptic integral. Everything we have heretofore shown is, however, applicable to every $G_\alpha$ with $0<\alpha\leq 1$.
Although we restrict ourselves to $G_{1/2}$, the methods presented here could just as well be applied to $G_1$, where we also have an explicit formula for the period function.
In this section, we show that $\partial_0 N_+$ is the graph of a non-increasing function in Cartesian coordinates (for $G_{1/2}$) by using properties of the period function. This finishes the proof of Equation 25, which in turn allows us to prove our main theorem. For encouragement, we refer the reader back to Figure 5, where we can see that $\partial_0 N_+$ is indeed the graph of a non-increasing function in Cartesian coordinates. It will be relatively easy to show that $\partial_0N_+$ is the graph of a Cartesian function, but the proof that $\partial_0N_+$ is a non-increasing function will be more involved. For example, we will first need to show that $\partial_0N_+$ limits to the line $b=4$ as $x_0\to 1$.
Recall that $P(\beta)$ is decreasing with respect to $\beta$ in the case when $\alpha=1$ or $1/2$. Here, we change variables for the period function from $\beta$ to $x_0\in (1/\sqrt{3},1)$ and get:
\begin{proposition}
$$\frac{d}{dx_0}\big(P(x_0)\big) >0$$
\end{proposition}
Since $z$ always vanishes at the half period, we have
\begin{proposition}
For any initial value $x_0$, we have
$$\bar{z}+(\frac{1}{2}\frac{dP}{dx_0})z'=0$$
at the time $t=P(x_0)/2$. Also, since $z'<0$ at the half period, we get that
$$\bar{z}(P(x_0)/2)>0, \quad \forall x_0.$$
\end{proposition}
From equation $(31)$ and the fact that $x'$ and $y'$ always vanish at the half-period, we have
\begin{proposition}
$$\bar{y}(P(x_0)/2)=\frac{1}{2\sqrt{x_0}}>0 \quad \textrm{ and } \quad \bar{x}(P(x_0)/2)=-2x_0<0$$
\end{proposition}
We are ready to get some information about $\partial_0 N_+$, beginning with:
\begin{corollary}
$\partial_0 N_+$ is the graph of a function in Cartesian coordinates.
\end{corollary}
\begin{proof}
This is equivalent to showing that
$$\frac{d}{dx_0}a_{x_0}(P(x_0)/2)>0.$$
The chain rule gets us
$$\frac{d}{dx_0}a_{x_0}(P(x_0)/2)=\bar{a}(P(x_0)/2)+(\frac{1}{2}\frac{dP(x_0)}{dx_0})a'(P(x_0)/2)$$
$$=\bigg(\frac{2}{3x}\bigg(2\bar{z}+\frac{1}{2}b\bar{y}-a\bar{x}\bigg)+(x+az/2)\frac{dP}{dx_0}\bigg)\bigg\rvert_{t=P(x_0)/2}$$
By the previous propositions, we know that all the terms above are positive at $P(x_0)/2$, whence the desired result.
\end{proof}
As with showing that $b'>0$ in the interval $(0,\rho)$, things are more difficult with the function $b$. We need three lemmas first. The proof of the following may also suggest that $\alpha=1/2$ is a "special case"; nevertheless, the situation is different than for Sol. Richard Schwartz proves a similar limit in \cite{S} for Sol (in which case, the limit is $2$), but his method uses an additional symmetry of the flow lines that we cannot use here.
\begin{lemma}For $\alpha=1/2$,
$$\lim_{x_0\to 1} b_{x_0}(P(x_0)/2)=4$$
\end{lemma}
More generally, we have the following conjecture
\begin{conjecture} Let
$$L(\alpha):=\lim_{x_0\to 1} b_{x_0,\alpha}(P(x_0,\alpha)/2), \textrm{ defined for all } \alpha \in (0,1].$$
We conjecture that $L(\alpha)$ is monotonically decreasing from $\alpha=0$ to $\alpha=1$ and that $\lim_{\alpha\to 0} L(\alpha)=\infty$. In fact, we also conjecture that
$$L(\alpha)=\frac{2}{\alpha}.$$
\end{conjecture}
The next technical lemma is quite wearisome to prove. We direct the reader to Figure 7, which provides numerical evidence for its veracity.
\begin{lemma}
For $\alpha=1/2,$ we have
$$G(x_0):=\frac{dP}{dx_0}(x_0)-\pi\bigg(\frac{1}{2\sqrt{x_0}}+\frac{2x_0\sqrt{x_0}}{1-x_0^2}\bigg)<0,\quad \forall x_0 \in \bigg(\frac{1}{\sqrt{3}},1\bigg)$$
\end{lemma}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{DPBOUND}
\caption{The graph of the function $G(x_0)$.}
\end{figure}
The penultimate step towards the Monotonicity Theorem:
\begin{lemma}
$$\frac{d}{dx_0}b_{x_0}(P(x_0)/2)<0, \textrm{ whenever } b_{x_0}(P(x_0)/2)>\pi$$
\end{lemma}
\begin{proof}
By the chain rule we know
$$\frac{d}{dx_0}b_{x_0}(P(x_0)/2)=\bigg(-\frac{2}{3y}\bigg(2\bar{z}+\frac{1}{2}b\bar{y}-a\bar{x}\bigg)+(y-\frac{1}{4}bz)\frac{dP}{dx_0}\bigg)\bigg\rvert_{t=P(x_0)/2}.$$
The $bz$ term is zero at the half period, and we can simplify further. The above is less than zero if and only if
$$(2x^2+2y^2)\frac{dP}{dx_0}<b\bar{y}-2a\bar{x} \textrm{ at } t=P(x_0)/2.$$
However we know $x^2+y^2=1$ at the half period, and we may also employ the Reciprocity Lemma. This yields that our desired result is equivalent to showing
$$\frac{dP}{dx_0}(x_0)<\big(b(\bar{y}-\frac{y}{x}\bar{x})\big)\big\rvert_{t=P(x_0)/2}.$$
By hypothesis, $b_{x_0}(P(x_0)/2)>\pi$, and we can use equation $(31)$. This means it suffices to show
$$\frac{dP}{dx_0}(x_0)<\pi\bigg(\frac{1}{2\sqrt{x_0}}+\frac{2x_0\sqrt{x_0}}{1-x_0^2}\bigg),\quad \forall x_0 \in \bigg(\frac{1}{\sqrt{3}},1\bigg)$$
which is nothing but Lemma 28.
\end{proof}
Now we are ready to prove the Monotonicity Theorem.
\begin{theorem}[The Monotonicity Theorem]
For $\alpha=1/2$, $\partial_0 N_+$ is the graph of a non-increasing function (in Cartesian coordinates).
\end{theorem}
\begin{proof}
This is equivalent to showing
$$\frac{d}{dx_0}b_{x_0}(P(x_0)/2)\leq0$$
or, as in the proof of the previous lemma,
$$\frac{dP}{dx_0}(x_0)\leq\big(b(\bar{y}-\frac{y}{x}\bar{x})\big)\big\rvert_{t=P(x_0)/2}.$$
We now show that $b_{x_0}(P(x_0)/2)\geq4$ always. If this is true, then the theorem follows by Lemma 29. Otherwise, assume that $b_{x_0}(P(x_0)/2)<4$ for some choice of $x_0$. By Lemma 27, $b_{x_0}(P(x_0)/2)$ limits to $4$ as $x_0$ tends to $1$, so $b_{x_0}(P(x_0)/2)$ must eventually be greater than $\pi$. Let $x_0'$ be the last initial value where $b(P/2)\leq \pi$. By Lemma 29, $b_{x_0}(P(x_0)/2)$ will be strictly decreasing in $x_0$, which is a contradiction. Hence, $b>4$ always and the theorem is proven.
\end{proof}
\section{Computer Code}
Here we present the Mathematica code that generates the figures we have presented. In addition, we present the Mathematica code written by Stephen Miller that allows numerical computation of the period function for arbitrary positive $\alpha$. We note that the whole of the program \cite{S2}, used to generate Figure 3, is available online on Richard Schwartz's website.
\subsection{Figures 1 and 2}
This gets us the implicit plots in the plane:
\begin{lstlisting}[language=Mathematica]
Manipulate[
ContourPlot[(1/a) E^(2*a*z) + E^(-2 z) +
w^2 == (1+a)/(a*b^2), {w,-10,10}, {z,-10,10},
PlotRange -> All, AspectRatio -> Automatic, PlotPoints -> 50],
{b,0.001,1}, {a,0.001,1}]
\end{lstlisting}
A 3-D rendering can be obtained with:
\begin{lstlisting}[language=Mathematica]
Manipulate[
ContourPlot3D[(1/a) E^(2*a*z) +
E^(-2 z) + (x - Sqrt[a]*y)^2 == (1 + a)/(a*b^2), {x, -10,
10}, {y,-10,10}, {z,-10,10}, PlotRange -> All], {b,0.01,
1}, {a,0.05,1}]
\end{lstlisting}
\newpage
\subsection{Figure 4 and 5}
The following renders the symmetric flowline associated to any choice of $x_0 \in (\frac{1}{\sqrt{3}}, 1)$ for the $G_{1/2}$ group. Moreover, this provides numerical evidence that the unwieldy expression for the period function in $G_{1/2}$ we presented earlier is indeed correct. Figure 5 just adds the straight line from the origin to the endpoint of the flowline to show that the "Bounding Triangle Theorem" does not work in general.
\begin{lstlisting}[language=Mathematica]
(*This is the period function for the G_{1/2} group,
as in Corollary 3.4 *)
P[B_]:= (*as in the statement of the corollary *)
(*This is the period function,
after the change of variables from \beta to x0*)
L1[x0_] := P[((3*Sqrt[3]/2)*(x0-x0^3))^(1/3)]
(*This numerically solves the fundamental system of ODE's *)
Manipulate[
s1 = NDSolve[{x'[t] == -x[t]*z[t], y'[t] == (1/2)*y[t]*z[t],
z'[t] == -(1/2)*y[t]^2 + x[t]^2, a'[t] == 2*x[t] + a[t]*z[t],
b'[t] == 2 y[t] - (1/2)*b[t]*z[t], z[0] == 0, a[0] == 0,
b[0] == 0, x[0] == x0, y[0] == Sqrt[1 - x0^2]}, {x, y, z, a,
b}, {t, 0, Re[L1[x0]]/2}], {x0, 1/Sqrt[3], 1}]
(*This plots the flowline*)
ParametricPlot[{a[t], b[t]} /. s1, {t, 0,Re[L1[x0]/2]},
PlotRange -> All]
\end{lstlisting}
\newpage
\subsection{Figure 6}
We can use Stephen Miller's program to compute the period for arbitrary alpha, or we can "find" it by inspection, preferably looking at the function $z$.
\begin{lstlisting}[language=Mathematica]
(*This numerically solves our ODE's, d is alpha here*)
Manipulate[
s = NDSolve[{x'[t] == -x[t]*z[t], y'[t] == d*y[t]*z[t],
z'[t] == -d*y[t]^2 + x[t]^2, a'[t] == 2*x[t] + a[t]*z[t],
b'[t] == 2 y[t] - d*b[t]*z[t], z[0] == 0, a[0] == 0, b[0] == 0,
x[0] == c1, y[0] == Sqrt[1 - c1^2]}, {x, y, z, a, b}, {t, 0,
20}], {c1, Sqrt[d/(d + 1)], 1}, {d, 0, 1}]
(*We can use this to find \rho by inspection*)
Manipulate[z[t] /. s, {t, 0,20}]
(*This plots the derivative of b*)
Plot[b'[t] /. s, {t, 0, 15.99}, AspectRatio -> Automatic,
AxesOrigin -> {0, 0}]
\end{lstlisting}
\subsection{Figure 7}
This code provides the numerical evidence for Lemma 28.
\begin{lstlisting}[language=Mathematica]
L1[x0_]:= (*As before*)
(*First we evaluate the derivative*)
D[L1[x0],x0]
(*Then, we plot G(x0)*)
Plot
{x0,1/Sqrt[3.],1},AxesOrigin -> {1/Sqrt[3.], 0}]
\end{lstlisting}
\newpage
\subsection{Computing the General Period Function}
Stephen Miller helped us with this code. It allows us to compute the period function for any choice of $\alpha$ and $\beta$. Using it, it appears that the monotonicity results we would like are indeed true for arbitrary $\alpha$, a promising sign for our Main Conjecture.
\begin{lstlisting}[language=Mathematica]
(*Defining the integrand*)
integrand[t_, A_, B_] = 2/
Sqrt[1 - B^2/(A +
1) (A Exp[2 t] + Exp[-2 A t])]
(*Numerically finding the endpoints of integration *)
endpoints[A_, B_] :=
Sort[Log[Flatten[
y /. NSolve[
1 - B^2/(A + 1) (A y + y^-A) == 0,
y, 20]]]/2]
(*Numerical integration*)
p[A_, B_] :=
NIntegrate[integrand[t, A, B],
Join[{t}, endpoints[A, B]]]
(*Generates the table presented earlier*)
Table[{a, p[a, .999], Pi*Sqrt[2/a]}, {a, 0.1, 1, 0.1}]
\end{lstlisting}
|
\section{Introduction}\label{sec:intro}
Neural networks (NNs) are well-established models in the field of machine learning. Given their flexibility and capability in addressing complex tasks where a large amount of data is at the disposal, \textit{deep} NNs are the state-of-art of many classification issues, including image recognition \cite{Pelletier19}, speech analysis \cite{Deng13} and medical diagnostics \cite{kermany2018identifying}; for a detailed treatment of various NNs architectures, as well as of related deep learning topics, we refer e.g. to \cite{Goodfellow16}.
In general, a network depends on a set of weights that are optimized in such a way that an objective function, made of a loss function and a regularization term, is minimized \cite{girosi1995regularization}. Loss functions measure the possible discrepancy between the NN prediction and a label encoding the information about the event occurrence. Loss design is a crucial aspect of machine learning theory \cite{rosasco2004loss} and, over the years, several functions have been proposed, typically with the aim of accounting for the specific properties of the problem under analysis \cite{Janocha16,Lin17,Liu16,Zhu20}. Among these choices, Cross Entropy (CE) and its generalizations \cite{Good52,Lu19,Zhang18} possess a solid theoretical background that can be traced back to information theory, they being related to the notions of \textit{entropy} and \textit{Kullback-Leibler divergence}, and their minimization corresponds to the maximization of the likelihood under a Bernoulli model (see e.g. \cite[Section 2.8]{Murphy12}).
While the network is optimized by minimizing a certain loss, the classification results are then evaluated by considering some well-established \textit{metrics} or \textit{skill scores}, which are often chosen according to the particular task and derived from the elements of the so-called \textit{confusion matrix} (CM). Although loss minimization and score maximization are two intertwined concepts, yet a direct maximization of a score during the training of the network is to be avoided, since the score is typically a discontinuous function with respect to the predictions given by the model. Therefore, other strategies have been taken into consideration.
For example, in \cite{Huang19} the authors propose an adaptive loss alignment with respect to the evaluation metric via a reinforcement learning approach. A loss function that \textit{approximates} the \textit{0-1 loss} is proposed in \cite{Singh10}. In \cite{Rezatofighi19}, a loss that generalizes the \textit{Jaccard index} is discussed for object detection issues. In order to address the lack of regularity of the score, \textit{fuzzy} or \textit{probabilistic} confusion matrices have been examined e.g. in \cite{Koco13}, where the minimization of the off-diagonal elements of a probabilistic confusion matrix is considered; in \cite{Wang13}, where a confusion entropy metric is constructed; and in \cite{Trajdos16}, where local fuzzy CMs have been employed.
This fuzzy/probabilistic approach was inspirational for the framework introduced in the present paper, which also presents similarities with the main idea investigated in \cite{Ohsaki17} in the context of kernel logistic regression, and which has to be intended as the theoretically-founded generalization of the approach carried out, e.g., in \cite{Zhang17}. Indeed, our study addresses the problem of filling the gap between score maximization and loss minimization by constructing a class of Score-Oriented Loss (SOL) functions out of a classical CM. The main idea consists in treating the threshold that influences the entries of the confusion matrix not as a fixed value, but as a random variable. Therefore, by considering the expected value of the matrix, we build losses that are indeed derivable with respect to the weights of the model. Furthermore, we theoretically motivate and then numerically show that the probability distribution chosen \textit{a priori} for the threshold severely influences the outcome of the \textit{a posteriori} score maximization. More in details, the optimal threshold is very likely to be placed in the \textit{concentration areas} of the \textit{a priori} distribution. As a consequence, using a SOL function related to a certain skill score, we automatically obtain an optimized value without performing any further \textit{a posteriori} procedure.
The paper is organized as follows. In Section \ref{sec:costruzione_loss}, starting from a classical confusion matrix, we construct the theoretical foundation of our SOL functions and we investigate their properties. Section \ref{sec:results} describes two possible choices for the probability distribution associated to the threshold and illustrates their application to two experimental forecasting problems. Finally, our conclusions are offered in Section \ref{sec:conclusions}.
\section{Construction and properties of score-oriented losses}\label{sec:costruzione_loss}
Let $\Omega\subset\mathbb{R}^m$ and let $\mathcal{X}\subset \Omega$ be a set of data, $m\in\mathbb{N}_{\ge 1}$. Let $\mathcal{Y}$ be a finite set of $d$ \textit{labels} or \textit{classes} to be learned, where $d\in\mathbb{N}_{\ge 2}$ and the labels (or classes) are integer encoded or \textit{one-hot} encoded in practice \cite{Harris13}. Suppose that each element in $\mathcal{X}$ is uniquely assigned to a class in $\mathcal{Y}$. Then, the supervised classification problem, where $d=2$, consists of constructing a function $g$ on $\Omega$ by \textit{learning} from the labeled data set $\mathcal{X}$, in such a way that $g$ \textit{models} the data-label relation between elements in $\Omega$ and labels in $\mathcal{Y}$. In order to achieve such result, one usually defines a \textit{loss} function $\ell(g(\boldsymbol{x}),y)$ that measures the possible discrepancy between the prediction $g(\boldsymbol{x})$ given by the model, $\boldsymbol{x}\in\mathcal{X}$, and its true label $y\in\mathcal{Y}$. In this view, an NN builds the function $g$ by minimizing the loss on the training set $X$, as well as focusing on its generalization capability when predicting possible unseen test samples in $\Omega$.
More specifically, the network produces the output
\begin{equation*}
\hat{y}(\boldsymbol{x},\boldsymbol{w})\coloneqq (\sigma\circ h)(\boldsymbol{x},\boldsymbol{w})\in [0,1],\; \boldsymbol{x}\in\Omega,
\end{equation*}
where $h(\boldsymbol{x},\boldsymbol{w})$ denotes the outcome of the \textit{input} and \textit{hidden} layers, being $\boldsymbol{w}$ the vector (matrix) of weight parameters, and $\sigma$ is the \textit{sigmoid} activation function defined as $\sigma(h)=(1+e^{-h})^{-1}$.
During the training procedure, the weights of the network are adjusted via \textit{backpropagation} in such a way that a certain objective function is minimized, i.e. we consider the following minimization problem
\begin{equation}\label{eq:obj_fun}
\min_{\boldsymbol{w}}\ell(\hat{y}(\boldsymbol{x},\boldsymbol{w}),y)+\lambda R(\boldsymbol{w}),\; \boldsymbol{x}\in\Omega,
\end{equation}
where $\ell$ is a loss function and $R(\boldsymbol{w})$ is a possible regularization term that is controlled by a parameter $\lambda\in\mathbb{R}_{>0}$. The outcomes of this regression problem can be clustered by means of some thresholding process in order to construct a classifier whose predictive effectiveness is assessed by means of specific skill scores. We now introduce a one-parameter family of CMs from which we will derive a set of probabilistic skill scores. This approach will allow the introduction of a corresponding set of probabilistic loss functions and, accordingly, an {\em{a priori}} optimization of the scores.
\subsection{Confusion matrices with probabilistic thresholds}
Let us consider a batch of predictions-labels $\mathcal{S}=\mathcal{S}_n(\boldsymbol{w})=\{(\hat{y}(\boldsymbol{x}_i,\boldsymbol{w}),y_i)\}_{i=1,\dots,n}$, where $y_i\in\{0,1\}$ is the true label associated to the element $\boldsymbol{x}_i\in\mathcal{X}$ and $\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})\in[0,1]$ is the prediction for $\boldsymbol{x}_i$ given by the model. Furthermore, let $n^+$ and $n^-$ be the number of elements in $\mathcal{S}_n(\boldsymbol{w})$ whose true class $y_i$, $i=1,\dots,n$, is the \textit{positive} ($\{y=1\}$) or the \textit{negative} ($\{y=0\}$), respectively. Therefore, $n^++n^-=n$.
Let $\tau\in\R,\;\tau\in(0,1),$ and let
\begin{equation*}
\mathbbm{1}_{x}(\tau)=\mathbbm{1}_{\{x>\tau\}}=\begin{dcases} 0 & \textrm{if } x\le \tau,\\
1 & \textrm{if }x>\tau,\end{dcases}\quad x\in(0,1),
\end{equation*}
be the indicator function. The classical confusion matrix is defined as
\begin{equation*}
\mathrm{CM}(\tau)=\mathrm{CM}(\tau,\mathcal{S}_n(\boldsymbol{w})) =
\begin{pmatrix}
\mathrm{TN}(\tau,\mathcal{S}_n(\boldsymbol{w})) & \mathrm{FP}(\tau,\mathcal{S}_n(\boldsymbol{w})) \\
\mathrm{FN}(\tau,\mathcal{S}_n(\boldsymbol{w})) & \mathrm{TP}(\tau,\mathcal{S}_n(\boldsymbol{w}))
\end{pmatrix},
\end{equation*}
where the elements are
\begin{equation}\label{eq:mat_defi}
\begin{split}
& \mathrm{TN}(\tau,\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{(1-y_i)\mathbbm{1}_{\{\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})<\tau\}}},\;\mathrm{TP}(\tau,\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{y_i\mathbbm{1}_{\{\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})>\tau\}}},\vspace{5pt}\\
& \mathrm{FP}(\tau,\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{(1-y_i)\mathbbm{1}_{\{\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})>\tau\}}},\;\mathrm{FN}(\tau,\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{y_i\mathbbm{1}_{\{\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})<\tau\}}}.
\end{split}
\end{equation}
In what follows, we let $\tau$ be a continuous random variable whose probability density function (pdf) $f$ is supported in $[a,b]\subseteq[0,1]$, $a,b\in\R$. Furthermore, we denote as $F$ the cumulative density function (cdf)
\begin{equation*}
F(x)=\int_{a}^{x}{f(\xi)\mathrm{d}\xi},\quad x\le b.
\end{equation*}
We recall that
\begin{equation*}
\mathbb{E}_{\tau}[\mathbbm{1}_{x}(\tau)]=\int_{a}^{b}{\mathbbm{1}_{\{x>\xi\}}f(\xi)\mathrm{d}\xi}=\int_{a}^{x}{f(\xi)\mathrm{d}\xi}=F(x).
\end{equation*}
This leads to the following definition.
\begin{definition}
Let $\mathcal{S}_n(\boldsymbol{w})$ be a batch of predictions-labels and let $\mathrm{CM}(\tau,\mathcal{S}_n(\boldsymbol{w}))$ be a confusion matrix as defined in \eqref{eq:mat_defi}. Moreover, let $\tau$ be a continuous random variable on $(0,1)$ with cdf $F$. We define the expected confusion matrix as
\begin{equation*}
\overline{\mathrm{CM}}_{F}=\overline{\mathrm{CM}}_F(\mathcal{S}_n(\boldsymbol{w}))=
\begin{pmatrix}
\overline{\mathrm{TN}}_F(\mathcal{S}_n(\boldsymbol{w})) & \overline{\mathrm{FP}}_F(\mathcal{S}_n(\boldsymbol{w})) \\
\overline{\mathrm{FN}}_F(\mathcal{S}_n(\boldsymbol{w})) & \overline{\mathrm{TP}}_F(\mathcal{S}_n(\boldsymbol{w}))
\end{pmatrix},
\end{equation*}
where
\begin{equation}\label{eq:fcm_entries}
\begin{split}
& \overline{\mathrm{TN}}_F(\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{(1-y_i)(1-F(\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})))},\; \overline{\mathrm{TP}}_F(\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{y_iF(\hat{y}(\boldsymbol{x}_i,\boldsymbol{w}))},\vspace{5pt}\\
& \overline{\mathrm{FP}}_F(\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{(1-y_i)F(\hat{y}(\boldsymbol{x}_i,\boldsymbol{w}))},\vspace{5pt}\; \overline{\mathrm{FN}}_F(\mathcal{S}_n(\boldsymbol{w})) = \sum_{i=1}^n{y_i(1-F(\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})))}.
\end{split}
\end{equation}
\end{definition}
\begin{remark}
We observe that $\overline{\mathrm{CM}}_F$ is indeed the expected value of the matrix $\mathrm{CM}(\tau)$ with respect to $\tau$, meaning that
\begin{equation}
\mathbb{E}_{\tau}[\mathrm{CM}(\tau)]\coloneqq \begin{pmatrix}
\mathbb{E}_{\tau}[\mathrm{TN}(\tau,\mathcal{S}_n(\boldsymbol{w}))] & \mathbb{E}_{\tau}[\mathrm{FP}(\tau,\mathcal{S}_n(\boldsymbol{w}))] \\
\mathbb{E}_{\tau}[\mathrm{FN}(\tau,\mathcal{S}_n(\boldsymbol{w}))] & \mathbb{E}_{\tau}[\mathrm{TP}(\tau,\mathcal{S}_n(\boldsymbol{w}))]
\end{pmatrix}=\overline{\mathrm{CM}}_F.
\end{equation}
Moreover, the sum by rows is preserved, that is for any $\tau\in(0,1)$
\begin{equation*}
\big\lVert \overline{\mathrm{CM}}_F(\mathcal{S}_n(\boldsymbol{w})) \big\lVert_{\infty}=\big\lVert \mathrm{CM}(\tau,\mathcal{S}_n(\boldsymbol{w})) \big\lVert_{\infty}=\max\{n^-,n^+\}.
\end{equation*}
Hence, all considered matrices are contained in the set
\begin{equation}\label{eq:def_c}
\mathcal{C}(\mathcal{S}_n)\coloneqq \{\mathrm{A}\in \mathcal{M}^+_{2}\:|\: \mathrm{A}_{11}+\mathrm{A}_{12} = n^-,\;\mathrm{A}_{21}+\mathrm{A}_{22} = n^+\},
\end{equation}
where $\mathrm{A}=(\mathrm{A}_{ij})_{i,j=1,2}$ and $\mathcal{M}^+_{2}$ is the set of $2\times 2$ positive-valued matrices.
\end{remark}
Therefore, we can prove the following results by considering well-known concentration inequalities:
\begin{proposition}\label{prp:expected_matrix}
Let $\mathcal{S}_n\coloneqq\mathcal{S}_n(\boldsymbol{w})$ be a batch of predictions-labels, $n\in\mathbb{N}$, and let $\tau$ be a continuous random variable on $(0,1)$ with cdf $F$. Then,
letting $\mathrm{N}\in\{\mathrm{TN},\mathrm{FP}\}$, $\mathrm{P}\in\{\mathrm{TP},\mathrm{FN}\}$ and $\varepsilon>0$, we have
\begin{equation}\label{eq:thx_mc}
\begin{split}
& \mathbb{P}(|\mathrm{N}(\tau,\mathcal{S}_n)-\overline{\mathrm{N}}_F(\mathcal{S}_n)|\ge\varepsilon)\le 2\: \mathrm{exp}\bigg(-\frac{2\varepsilon^2}{(n^-)^2}\bigg),\\
& \mathbb{P}(|\mathrm{P}(\tau,\mathcal{S}_n)-\overline{\mathrm{P}}_F(\mathcal{S}_n)|\ge\varepsilon)\le 2\: \mathrm{exp}\bigg(-\frac{2\varepsilon^2}{(n^+)^2}\bigg).
\end{split}
\end{equation}
Moreover, denoting as $\mathrm{tr}(\mathrm{CM})=\mathrm{TN}+\mathrm{TP}$ the trace of the confusion matrix, we get
\begin{equation}\label{eq:thx_ho}
\mathbb{P}(|\mathrm{tr}(\mathrm{CM}(\tau,\mathcal{S}_n)) - \mathrm{tr}(\overline{\mathrm{CM}}_F(\mathcal{S}_n))|\ge\varepsilon)\le 2\: \mathrm{exp}\bigg(-\frac{2\varepsilon^2}{(n^-)^2+(n^+)^2}\bigg).
\end{equation}
\end{proposition}
\begin{proof}
In order to prove \eqref{eq:thx_mc} it is sufficient to consider the McDiarmid's inequality (see e.g. \cite[Theorem 4.5, p.88]{ShaweTaylor04}) and to observe that
\begin{equation*}
\sup_{\tau_1,\tau_2\in(0,1)}{|\mathrm{N}(\tau_1,\mathcal{S}_n)-\mathrm{N}(\tau_2,\mathcal{S}_n)|}=n^-,\; \sup_{\tau_1,\tau_2\in(0,1)}{|\mathrm{P}(\tau_1,\mathcal{S}_n)-\mathrm{P}(\tau_2,\mathcal{S}_n)|}=n^+.
\end{equation*}
Then, we can apply the Hoeffding's inequality \cite{Hoeffding63} to $\mathrm{TN}(\tau,\mathcal{S}_n)+\mathrm{TP}(\tau,\mathcal{S}_n)$, thus obtaining \eqref{eq:thx_ho}.
\end{proof}
\subsection{Derived scores and loss functions}
A skill score is a function mapping the space of confusion matrices onto $\R$. In the classical setting, the skill score is denoted as $s:M_{2,2}(\mathbb{N}) \rightarrow \R$ and is constructed on $\mathrm{CM}(\tau,\mathcal{S}_n(\boldsymbol{w}))$ for a fixed value of $\tau \in (0,1)$. In the case of confusion matrices with probabilistic threshold, the score explicitly depends on the cumulative distribution function $F$ adopted and is therefore denoted as ${\overline{s}}_F:M_{2,2}(\R) \rightarrow \R$. In this latter case, we can introduce a class of Score-Oriented Loss (SOL) functions as follows.
\begin{definition}\label{def:scoloss}
Let $\mathcal{S}_n(\boldsymbol{w})$ be a batch of predictions-labels and
let $s(\tau,\mathcal{S}_n(\boldsymbol{w}))=s(\tau)$ be a score calculated upon $\mathrm{CM}(\tau,\mathcal{S}_n(\boldsymbol{w}))$, $\tau\in(0,1)$. Indicating with ${\overline{s}}_F(\mathcal{S}_n(\boldsymbol{w}))={\overline{s}}_F$ the score $s$ calculated upon $\overline{\mathrm{CM}}_F(\mathcal{S}_n(\boldsymbol{w}))$, a Score-Oriented Loss (SOL) function related to $s$ is defined as
\begin{equation*}
\ell_{\overline{s}}(F,\mathcal{S}_n(\boldsymbol{w})))\coloneqq - {\overline{s}}_F(\mathcal{S}_n(\boldsymbol{w}))).
\end{equation*}
\end{definition}
As an example, taking the \textit{accuracy} score
\begin{equation*}
\textrm{acc}=\frac{\textrm{TP}+\textrm{TN}}{\textrm{TP}+\textrm{TN}+\textrm{FP}+\textrm{FN}},
\end{equation*}
we get
\begin{equation*}
\begin{split}
\overline{\textrm{acc}}_F(\mathcal{S}_n(\boldsymbol{w}))) & \coloneqq \frac{\sum_{i=1}^n{y_iF(\hat{y}(\boldsymbol{x}_i,\boldsymbol{w}))+(1-y_i)(1-F(\hat{y}(\boldsymbol{x}_i,\boldsymbol{w})))}}{n}
\end{split}
\end{equation*}
and then the loss function $\ell_{\overline{\textrm{acc}}}(F,\mathcal{S}_n(\boldsymbol{w}))\coloneqq - \overline{\textrm{acc}}_F(\mathcal{S}_n(\boldsymbol{w})))$.
\begin{remark}
By setting $0/0\coloneqq 0$ and by virtue of the fundamental theorem of calculus, we observe that $\ell_{\overline{s}}$ is derivable with respect to the weights vector $\boldsymbol{w}$, which is a fundamental property for its usage as loss function.
\end{remark}
We now prove the following
\begin{theorem}\label{thm:linea}
Let $\mathcal{S}_n\coloneqq\mathcal{S}_n(\boldsymbol{w})$ be a batch of predictions-labels, $n\in\mathbb{N}$, and let $\tau$ be a continuous random variable on $(0,1)$ with cdf $F$. Then, if $s$ is linear with respect to the entries of the confusion matrix, we have
\begin{equation}\label{eq:score_expected}
\mathbb{E}_{\tau}[s(\tau,\mathcal{S}_n)]=\overline{s}_F(\mathcal{S}_n).
\end{equation}
Moreover, letting $\mathrm{supp}(s(\tau,\mathcal{S}_n))=[a,b]$, $a,b\in\mathbb{R}$, $a<b$, and $\varepsilon>0$, we obtain
\begin{equation}\label{eq:score_proba}
\mathbb{P}(|s(\tau,\mathcal{S}_n)-\overline{s}_F(\mathcal{S}_n)|\ge\varepsilon)\le 2\: \mathrm{exp}\bigg(-\frac{2\varepsilon^2}{(b-a)^2}\bigg).
\end{equation}
\end{theorem}
\begin{proof}
Due to the linearity of the expected value, the results in Proposition \ref{prp:expected_matrix} imply \eqref{eq:score_expected}. As a consequence, the McDiarmid's inequality can be applied in order to obtain \eqref{eq:score_proba}.
\end{proof}
\begin{remark}\label{rem:rems}
Suppose that the score is not linear with respect to the entries of the vectorized confusion matrix $\mathrm{vec}(\mathrm{CM}(\tau))=(\mathrm{TN}(\tau),\mathrm{FP}(\tau),\mathrm{FN}(\tau),\mathrm{TP}(\tau))$. By looking at the scores $s$ and $\overline{s}_F$ as functions of $\mathrm{vec}(\mathrm{CM}(\tau))$, i.e. $s(\tau)=s(\mathrm{vec}(\mathrm{CM}(\tau)))$ and $\overline{s}_F=s(\mathrm{vec}(\overline{\mathrm{CM}}_F))$, we have the Taylor approximation (see e.g. \cite{Benaroya05})
\begin{equation*}
s(\mathrm{vec}(\mathrm{CM}(\tau)))\approx s(\mathrm{vec}(\overline{\mathrm{CM}}_F))+\sum_{|\boldsymbol{\alpha}|\ge 1}{\frac{D^{\boldsymbol{\alpha}}s(\mathrm{vec}(\overline{\mathrm{CM}}_F))}{\boldsymbol{\alpha}!}(\mathrm{vec}(\mathrm{CM}(\tau)-\overline{\mathrm{CM}}_F))^{\boldsymbol{\alpha}}},
\end{equation*}
which implies
\begin{equation*}
\mathbb{E}_{\tau}[s(\tau)]\approx \overline{s}_F+\sum_{|\boldsymbol{\alpha}|\ge 1}{\frac{D^{\boldsymbol{\alpha}}\overline{s}_F}{\boldsymbol{\alpha}!}\mathbb{E}_{\tau}[(\mathrm{vec}(\mathrm{CM}(\tau)-\overline{\mathrm{CM}}_F))^{\boldsymbol{\alpha}}]},
\end{equation*}
where $\boldsymbol{\alpha}\in\mathbb{N}^4$ denotes the classical multi-index notation.
\end{remark}
Recalling \eqref{eq:obj_fun}, let $T(\boldsymbol{w})\coloneqq \ell(\mathcal{S}(\boldsymbol{w}))+\lambda R(\boldsymbol{w})$ be the target function to minimize during the training of the network, where $\ell$ is a generic loss function. When performing \textit{a posteriori} maximization of a score $s$, letting $\boldsymbol{w}^{\star}=\argmin_{\boldsymbol{w}}T(\boldsymbol{w})$, we compute
\begin{equation*}
\max_{\tau\in(0,1)}s(\tau,\mathcal{S}(\boldsymbol{w}^{\star})).
\end{equation*}
In our setting with SOL functions, in view of Definition \ref{def:scoloss} and Remark \ref{rem:rems}, the training minimization problem results in
\begin{equation*}
\min_{\boldsymbol{w}}\ell_{\overline{s}}(F,\mathcal{S}(\boldsymbol{w})))+\lambda R(\boldsymbol{w})\approx \max_{\boldsymbol{w}} \mathbb{E}_{\tau}[s(\tau,\mathcal{S}(\boldsymbol{w}))]-\lambda R(\boldsymbol{w}),
\end{equation*}
where the equality is achieved under the assumptions of Theorem \ref{thm:linea}. Therefore, the maximization of the expected value of the score is included in the training process. Moreover, we observe that while a direct maximization of the score $s(\tau^{\star},\mathcal{S}(\boldsymbol{w}))$ for a fixed $\tau^{\star}\in(0,1)$ is not possible due to the lack of derivability, this procedure can be approximated by considering for $\tau$ a probability distribution highly concentrated around the mean value $\tau^{\star}$. In Section \ref{sec:results}, we review two different probability distributions that express different concentration properties.
The following result concerns the mean absolute deviation (mad) \cite{Geary35} of a Lipschitz continuous score $s$ from its expected value $\overline{s}_F$.
\begin{theorem}\label{thm:lippa}
Let $\mathcal{S}_n$ be a batch of predictions-labels, $n\in\mathbb{N}$, and let $\tau$ be a continuous random variable on $(0,1)$ with cdf $F$. Assume that $s$ is a Lipschitz continuous function with respect to the entries of $\mathrm{CM}(\tau,\mathcal{S}_n)$ on the set $\mathcal{C}(\mathcal{S}_n)$. Then,
\begin{equation*}
\mathbb{E}_{\tau}[|s(\tau,\mathcal{S}_n)-{\overline{s}}_F(\mathcal{S}_n)|]\le \frac{1}{2}\: K_{s}(J_F+\mathrm{mad}_{\tau}[\mathrm{TN}(\tau)]+\mathrm{mad}_{\tau}[\mathrm{TP}(\tau)]),
\end{equation*}
where
\begin{equation}\label{eq:lip_cons}
K_s\coloneqq \sup_{\mathrm{A}\in\mathcal{C}(\mathcal{S}_n)}\lVert \nabla s(\mathrm{A}_{11},\mathrm{A}_{12},\mathrm{A}_{21},\mathrm{A}_{22}) \lVert_{\infty},
\end{equation}
with $\mathcal{C}(\mathcal{S}_n)$ as defined in \eqref{eq:def_c},
\begin{equation}\label{eq:weird_cons}
J_F\coloneqq \mathbb{E}_{\tau}[\:|\:|\mathrm{TN}(\tau)-\overline{\mathrm{TN}}_F|-|\mathrm{TP}(\tau)-\overline{\mathrm{TP}}_F|\:|\:],
\end{equation}
and $mad$ indicates the mean absolute deviation.
\end{theorem}
\begin{proof}
As in Remark \ref{rem:rems}, we consider $s$ as a function of the vectorized confusion matrix $\mathrm{vec}(\mathrm{CM}(\tau))=(\mathrm{TN}(\tau),\mathrm{FP}(\tau),\mathrm{FN}(\tau),\mathrm{TP}(\tau))$, and thus
\begin{equation*}
{\overline{s}}_F=s(\overline{\mathrm{TN}}_F,\overline{\mathrm{FP}}_F,\overline{\mathrm{FN}}_F,\overline{\mathrm{TP}}_F).
\end{equation*}
The Lipschitz continuity of the score $s$ yields to
\begin{equation}\label{eq:lippi}
\begin{split}
&|s(\mathrm{TN}(\tau),\mathrm{FP}(\tau),\mathrm{FN}(\tau),\mathrm{TP}(\tau))-s(\overline{\mathrm{TN}}_F,\overline{\mathrm{FP}}_F,\overline{\mathrm{FN}}_F,\overline{\mathrm{TP}}_F)|\le\\
& K_{s}\lVert (\mathrm{TN}(\tau),\mathrm{FP}(\tau),\mathrm{FN}(\tau),\mathrm{TP}(\tau))-(\overline{\mathrm{TN}}_F,\overline{\mathrm{FP}}_F,\overline{\mathrm{FN}}_F,\overline{\mathrm{TP}}_F)\lVert_{\infty},
\end{split}
\end{equation}
where $K_{s}$ is the Lipschitz constant defined in \eqref{eq:lip_cons}. Then, since
\begin{equation*}
\mathrm{CM}(\tau,\mathcal{S}_n),\overline{\mathrm{CM}}_F(\mathcal{S}_n)\in\mathcal{C}(\mathcal{S}_n),
\end{equation*}
we have
\begin{equation*}
\begin{split}
& \lVert (\mathrm{TN}(\tau),\mathrm{FP}(\tau),\mathrm{FN}(\tau),\mathrm{TP}(\tau))-(\overline{\mathrm{TN}}_F,\overline{\mathrm{FP}}_F,\overline{\mathrm{FN}}_F,\overline{\mathrm{TP}}_F)\lVert_{\infty}=\\
& \max\{|\mathrm{TN}(\tau)-\overline{\mathrm{TN}}_F|,|\mathrm{TP}(\tau)-\overline{\mathrm{TP}}_F|\}=\\
& \frac{|\mathrm{TN}(\tau)-\overline{\mathrm{TN}}_F|+|\mathrm{TP}(\tau)-\overline{\mathrm{TP}}_F|+|\:|\mathrm{TN}(\tau)-\overline{\mathrm{TN}}_F|-|\mathrm{TP}(\tau)-\overline{\mathrm{TP}}_F|\:|}{2}.
\end{split}
\end{equation*}
We proceed by taking the expected value of both sides of \eqref{eq:lippi}. Then, in order to conclude the proof, it is sufficient to observe that
\begin{equation*}
\mathbb{E}_{\tau}[|\mathrm{TN}(\tau)-\overline{\mathrm{TN}}_F|]=\mathrm{mad}_{\tau}[\mathrm{TN}(\tau)],\quad \mathbb{E}_{\tau}[|\mathrm{TP}(\tau)-\overline{\mathrm{TP}}_F|]=\mathrm{mad}_{\tau}[\mathrm{TP}(\tau)],
\end{equation*}
and to define $J_F$ as in \eqref{eq:weird_cons}.
\end{proof}
\begin{remark}
Concerning the bound presented in Theorem \ref{thm:lippa}, we observe that the Lipschitz constant $K_s$ is decreasing as $n$ gets larger. Indeed, the larger the number of samples in the batch, the smaller the variation of the score under a little alteration of the entries of the confusion matrix.
For example, we have
\begin{equation*}
\nabla \mathrm{acc}(\mathrm{TN},\mathrm{FP},\mathrm{FN},\mathrm{TP})=\frac{1}{n^2}(n-\mathrm{TN}-\mathrm{TP}, n-\mathrm{TN}-\mathrm{TP},-\mathrm{TN}-\mathrm{TP},-\mathrm{TN}-\mathrm{TP})
\end{equation*}
which yields to $K_{\mathrm{acc}}=1/n$.
\end{remark}
\section{Applications to experimental forecasting problems}\label{sec:results}
The implementation of a machine learning approach based on SOL functions requires first the choice of the probability distribution function associated to the threshold $\tau$. In this paper we will utilize two distributions supported in $[0,1]$.
Specifically, if $\tau\sim\mathcal{U}(0,1)$, i.e. it is a random variable uniformly distributed on $(0,1)$, its pdf and cdf are $f_{\mathcal{U}}(x)= 1$ and $F_{\mathcal{U}}(x)=x$ respectively, $x\in(0,1)$ (see Figure \ref{fig:distributions}, top row). This well-known distribution is characterized by a high variance and does not favor any particular value of the threshold. Therefore, with respect to a classical confusion matrix \eqref{eq:mat_defi}, in $\overline{\mathrm{CM}}_{F_{\mathcal{U}}}$ the indicator functions of the outputs are replaced by the outputs themselves. Although its probabilistic derivation from the uniform distribution is not trivial, from an empirical viewpoint the confusion matrix $\overline{\mathrm{CM}}_{F_{\mathcal{U}}}$ is probably the most straightforward way to extend a classical confusion matrix to having non-integer valued entries \cite{Lawson14}.
Let us now consider $\mu\in (0,1)$ and let $\delta\in\R,\;\delta>0$ be such that $[\mu-\delta,\mu+\delta]\subseteq[0,1]$. We take $\tau\sim\mathcal{C}(\mu,\delta)$, i.e. it is a random variable distributed according to the \textit{raised cosine} distribution, whose related pdf and cdf on (0,1) are
\begin{equation*}
\begin{split}
& f_{\mathcal{C}(\mu,\delta)}(x)=\frac{1}{2\delta}\bigg(1+\cos\bigg(\frac{x-\mu}{\delta}\pi\bigg)\bigg)\mathbbm{1}_{\{\mu-\delta<x<\mu+\delta\}},\\
& F_{\mathcal{C}(\mu,\delta)}(x)=\frac{1}{2}\bigg(1+\frac{x-\mu}{\delta}+ \frac{1}{\pi}\sin\bigg(\frac{x-\mu}{\delta}\pi\bigg)\bigg)\mathbbm{1}_{\{\mu-\delta<x<\mu+\delta\}}+\mathbbm{1}_{\{x\ge\mu+\delta\}}.
\end{split}
\end{equation*}
In this case $\mathbb{E}[\tau]=\mu$ and $\mathrm{Var}[\tau]=\delta^2(\pi^2-6)/(3\pi^2)$. We plotted $f_{\mathcal{C}(\mu,\delta)}$ and $F_{\mathcal{C}(\mu,\delta)}$ in Figure \ref{fig:distributions}, bottom row. Differently from the uniform case, the raised cosine distribution can potentially drive the model to be maximized for a restricted interval of threshold values around $\mu$, since we can achieve a high concentration by choosing a \textit{small} $\delta$.
\begin{comment}
\begin{figure}[htbp]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.46\linewidth]{uniform_density.pdf} &
\includegraphics[width=0.46\linewidth]{uniform_cumulative.pdf} \\
\includegraphics[width=0.46\linewidth]{raised_density.pdf} &
\includegraphics[width=0.46\linewidth]{raised_cumulative.pdf}
\end{tabular}
\caption{Top row: The functions $f_{\mathcal{U}}(x)$ (left panel) and $F_{\mathcal{U}}(x)$ (right panel), $x\in(0,1)$ for the uniform distribution. Bottom row: the functions $f_{\mathcal{C}(\mu,\delta)}(x)$ (left panel) and $F_{\mathcal{C}(\mu,\delta)}(x)$ (right panel), $x\in(0,1)$, for $\mu=0.5$ and $\delta=0.2$, for the raised cosine distribution.}
\label{fig:distributions}
\end{figure}
\end{comment}
\begin{figure}[htbp]
\centering
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{uniform_density.pdf}
\caption{The function $f_{\mathcal{U}}$.}
\end{subfigure}
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{uniform_cumulative.pdf}
\caption{The function $F_{\mathcal{U}}$.}
\end{subfigure}
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{raised_density.pdf}
\caption{The function $f_{\mathcal{C}(\mu,\delta)}$.}
\end{subfigure}
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{raised_cumulative.pdf}
\caption{The function $F_{\mathcal{C}(\mu,\delta)}$.}
\end{subfigure}
\caption{Top row: The functions $f_{\mathcal{U}}(x)$ (left panel) and $F_{\mathcal{U}}(x)$ (right panel), $x\in(0,1)$ for the uniform distribution. Bottom row: the functions $f_{\mathcal{C}(\mu,\delta)}(x)$ (left panel) and $F_{\mathcal{C}(\mu,\delta)}(x)$ (right panel), $x\in(0,1)$, for $\mu=0.5$ and $\delta=0.2$, for the raised cosine distribution.}
\label{fig:distributions}
\end{figure}
We now test the performances of SOL functions relying on these two distribution functions, when applied to two actual binary classification problems.
A \textsc{Python} implementation of the SOLs is available for the scientific community at $\texttt{https://github.com/cesc14/SOL}\:.$
\subsection{Optimized \textit{a posteriori} thresholds and \textit{a priori} distributions}
We first considered the \textit{Adult} training dataset \cite{Kohavi96}, where the task is to predict whether a person can make more than 50,000 dollars per year based on census data. First, we performed the following pre-processing steps:
\begin{enumerate}
\item
We removed the samples containing missing values.
\item
We dropped the feature \textit{education}, which is highly correlated to the feature \textit{education-num}. Indeed, both features concern the education level of the people in the dataset.
\item
The possible values in the \textit{native-country} column, which reports the country of origin of the people in the data set, were restricted to $1$ in place of \textit{United-States} and $0$ for all other countries.
\item
All categorical features were \textit{one-hot} encoded. Finally, all features were standardized.
\end{enumerate}
The resulting elements were then described by $48$ numerical features. The data set consisted of 30162 samples, where 7508 people made more than 50,000 dollars per year, while 22654 did not. As for the network architecture, we used a fully-connected feed-forward NN with an input layer ($50$ ReLu neurons), two hidden layers ($20$ and $5$ ReLu neurons) and a sigmoid output unit. During the training phase, $1/3$ of training set was used as a validation set. We trained the network on $500$ epochs and we imposed an early stopping condition when no improvement in the validation loss was obtained after $30$ epochs.
The following experiments were devoted to show how the choice of the distribution of the threshold $\tau$ affected the result of the \textit{a posteriori} score maximization performed by varying the threshold in $(0,1)$. In doing so, we considered the SOL related to the \textit{f$_1$-score} \cite{lipton2014optimal}, which is the harmonic mean of precision and recall, i.e.
\begin{equation*}
\textrm{f$_1$-score}=\frac{2\textrm{TP}}{2\textrm{TP}+\textrm{FP}+\textrm{FN}}.
\end{equation*}
We observe that this score does not satisfy the linearity assumption in Theorem \ref{thm:linea} (see also Remark \ref{rem:rems}).
We tested the uniform distribution $\mathcal{U}$ as well as the raised cosine distribution $\mathcal{C}(\mu,\delta)$ with different values of the parameters $\mu,\delta$. More precisely, for each combination of $(\mu,\delta)\in \{0.3,0.5,0.7\}\times \{0.1,0.2,0.3\}$
we randomly selected $4/5$ elements in the data set for training the NN
and we performed the \textit{a posteriori} score maximization on the training set.
We repeated this procedure $100$ times.
In Table \ref{tab:prima}, we present the obtained results. We report that in some cases the training optimization process was unsuccessful, that is the validation loss got stuck from the beginning of the training, and no actual learning took place. Therefore, we did not include such cases in the computation of the other entries of the table.
Let $\eta\in\mathbb{N}$, $\eta\le100$, be the number of successful training processes. Figure \ref{fig:adult} compares the probability distribution chosen \textit{a priori} for the threshold as in the left panel of Figure \ref{fig:distributions} (dashed blue line) to the \textit{empirical} probability distribution $f_{\tau^{\star}}$ (red histograms) derived from the optimal thresholds $\tau^{\star}_1,\dots,\tau^{\star}_{\eta}$ over the $\eta$ successful runs. We modelled this distribution via normalized histograms. Moreover, we also displayed the average optimal threshold (red cross) resulting from the $\eta$ runs. The horizontal solid red line indicates the standard deviation related to the optimal thresholds. These experiments show that the optimal threshold is indeed heavily influenced by the chosen distribution, which effectively drives the optimization process to prefer a certain interval of threshold values.
\begin{table}[h]
\centering
\resizebox{0.99\textwidth}{!}{
\begin{tabular}{|c | c | c c c c c c |}
\hline
& \multirow{3}{*}{Uniform} & \multicolumn{6}{c||}{Raised cosine}\\
\cline{3-8}
& & \multicolumn{2}{c|}{$\mu=0.3$} & \multicolumn{2}{c|}{$\mu=0.5$} & \multicolumn{2}{c||}{$\mu=0.7$} \\
& & $\delta=0.1$ & \multicolumn{1}{c|}{$\delta=0.3$} & $\delta=0.1$ & \multicolumn{1}{c|}{$\delta=0.3$} & $\delta=0.1$ & \multicolumn{1}{c||}{$\delta=0.3$} \\
\hline
\#success & $100$ & $90$ & \multicolumn{1}{c|}{$100$} & $100$ & \multicolumn{1}{c|}{$100$} & $83$ & $100$\\
\#out-of-support & $0$ & $8$ & \multicolumn{1}{c|}{$13$} & $14$ & \multicolumn{1}{c|}{$9$} & $13$ & $13$ \\
\#epochs & $119.46_{(\pm 43.95 )}$ & $125.21_{(\pm 44.39 )}$ & \multicolumn{1}{c|}{$121.05_{(\pm 46.14 )}$} & $124.07_{(\pm 42.27 )}$ & \multicolumn{1}{c|}{$131.12_{(\pm 48.89 )}$} & $124.00_{(\pm 44.97 )}$ & $123.08_{(\pm 38.51 )}$\\
$\tau^{\star}$ & $0.53_{(\pm 0.36 )}$ & $0.31_{(\pm 0.06 )}$ & \multicolumn{1}{c|}{$0.35_{(\pm 0.22 )}$} & $0.50_{(\pm 0.07 )}$ & \multicolumn{1}{c|}{$0.52_{(\pm 0.19 )}$} & $0.70_{(\pm 0.08 )}$ & $0.66_{(\pm 0.20 )}$\\
f$_1$-score($\tau^{\star})$ & $0.7510_{(\pm 0.01 )}$ & $0.7578_{(\pm 0.01 )}$ & \multicolumn{1}{c|}{$0.7556_{(\pm 0.01 )}$} & $0.7557_{(\pm 0.01 )}$ & \multicolumn{1}{c|}{$0.7571_{(\pm 0.01 )}$} & $0.7549_{(\pm 0.01 )}$ & $0.7546_{(\pm 0.01 )}$\\
\hline
\end{tabular}
}
\caption{Outcomes of the application of the SOL functions in the case of a neural network trained on the {\em{Adult}} database. The chosen SOL relies on the $f_1$-score and the threshold $\tau$ is drawn according to the uniform and raised cosine distributions.}
\label{tab:prima}
\end{table}
\begin{figure}[htbp]
\centering
$$
\begin{array}{c}
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{fig_kde_uniform_nofailure_hist.pdf}
\caption{Uniform}
\end{subfigure}\\
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{fig_kde_cosine_0.3_0.1_nofailure_hist.pdf}
\caption{Raised cosine: $\mu=0.3$, $\delta=0.1$.}
\end{subfigure}
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{fig_kde_cosine_0.3_0.3_nofailure_hist.pdf}
\caption{Raised cosine: $\mu=0.3$, $\delta=0.3$.}
\end{subfigure}\\
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{fig_kde_cosine_0.5_0.1_nofailure_hist.pdf}
\caption{Raised cosine: $\mu=0.5$, $\delta=0.1$.}
\end{subfigure}
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{fig_kde_cosine_0.5_0.3_nofailure_hist.pdf}
\caption{Raised cosine: $\mu=0.5$, $\delta=0.3$.}
\end{subfigure}\\
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{fig_kde_cosine_0.7_0.1_nofailure_hist.pdf}
\caption{Raised cosine: $\mu=0.7$, $\delta=0.1$.}
\end{subfigure}
\begin{subfigure}[t]{.4\linewidth}
\centering
\includegraphics[width=\linewidth]{fig_kde_cosine_0.7_0.3_nofailure_hist.pdf}
\caption{Raised cosine: $\mu=0.7$, $\delta=0.3$.}
\end{subfigure}
\end{array}
$$
\caption{Comparison between theoretical and empirical distribution functions for the threshold. Dashed blue lines: uniform distribution and raised cosine distribution for different values of $\mu$ and $\delta$. Red histograms: the optimal thresholds provided by $100$ runs of the network when the loss is driven by $f_1$-score. Red crosses and red horizontal bars: average values of the histograms with standard deviations.}
\label{fig:adult}
\end{figure}
\subsection{SOLs versus CE: addressing strongly imbalanced classes}
We then considered the problem of predicting high levels of pollution by using the Beijing air quality data set from the University of California at Irvine (UCI), available at $https://archive.ics.uci.edu/ml/datasets/Beijing+PM2.5+Data$.
The data set contains PM2.5 pollution
data and meteorological data including dew point, temperature, pressure, wind direction, wind speed and the cumulative number of hours of snow and rain. Data are collected every hour for five years, sourced from the data interface released by the US Embassy in Beijing \cite{Liang2015}.
We first fixed a level of PM2.5 concentration equal to $400$ and we labelled the feature vector containing the meteorological data and the PM2.5 concentration at time $T$ with class $1$ if the PM2.5 concentration at time $T+1$ is over the fixed level.
The data set ranges from 01/01/2010 to 12/31/2014: we considered $13104$ samples in training (about one year and a half) and $4320$ samples in testing (about half a year) by taking the testing period right after the training period. In order to guarantee statistical robustness we repeated $100$ times the construction of training and test sets, by shifting the training period and test period of $5$ days each time. The data set is strongly imbalanced since the mean rate of positive samples in the training set is about $1.41\%$ and in the testing set is about $1.19\%$. During the training phase, $1/3$ of the training set was used as validation set.
We used a fully-connected feed-forward NN with an input layer ($15$ ReLu neurons), a hidden layer ($8$ ReLu neurons) and a sigmoid output unit. We trained on $1000$ epochs and we imposed an early stopping condition when no improvement in the validation loss was obtained after $50$ epochs. As for loss functions, we considered the SOL functions related to the \textit{True Skill Statistic} (TSS) \cite{allouche2006assessing} and \textit{Critical Success Index} (CSI) \cite{schaefer1990critical},
which are very common skill scores adopted in meteorology applications and particularly appropriate in the case of strongly imbalanced data set. They are defined as
\begin{equation*}
\textrm{TSS}=\frac{\textrm{TP}}{\textrm{TP}+\textrm{FN}}+\frac{\textrm{TN}}{\textrm{TN}+\textrm{FP}}-1,\quad
\textrm{CSI}=\frac{\textrm{TP}}{\textrm{TP}+\textrm{FP}+\textrm{FN}}.
\end{equation*}
We point out that the TSS score satisfies the linearity assumption in Theorem \ref{thm:linea}, while the CSI does not.
Considering both the uniform and raised Cosine distributions $C(\mu,\delta)$ (in the case $\mu=0.5$ and $\delta=0.1,0.3$), we compared the proposed SOLs against the CE loss. For each run we performed the \textit{a posteriori} score maximization on the training set in order to compute the optimum threshold. In Table \ref{tab:poll_tss} we reported the performances on both training and test sets by evaluating the skill scores with both the usual threshold $0.5$ and the optimum threshold $\tau^{\star}$ selected by the \textit{a posteriori} maximization. As in the previous application, we did not include cases in which the training optimization process was unsuccessful, and we reported the number of success on $100$ runs.
Table \ref{tab:poll_tss} shows that using the SOL related to a given skill score leads to better results with respect to the ones obtained by using the CE loss, in the case of both the training and the test sets. In particular, the skill scores are largely satisfactory by setting $\tau=0.5$ without the need to perform the a posteriori score maximization. In the case where the raised cosine distribution is implemented, this is due to the capability of the SOL function of driving the optimization process to promote thresholds in a neighborhood of the parameter $\mu$ (equal to $0.5$ in these experiments). In the case of the uniform distribution, no particular values of the threshold are promoted in the optimization process and this leads to results with high skill scores independently from the choice of the threshold.
Finally, we notice that the mean number of epochs to reach the convergence condition is lower for SOL functions with respect to the CE loss: in particular such a number is much lower for the SOL function in which the raised cosine distribution is chosen with $\delta=0.1$.
\begin{table}[htb!]
\centering
\resizebox{0.99\textwidth}{!}{
\begin{tabular}{|c | c c c c c | c c|}
\hline
& \multicolumn{5}{c|}{Training} & \multicolumn{2}{c|}{Test} \\
\cline{2-8}
Loss & \#success &
\#epoch &
TSS &
$\tau^{\star}$ &
TSS($\tau^{\star}$) & TSS & TSS($\tau^{\star}$) \\
\hline
CE & 100 & $1000.00_{(\pm 0.00 )}$ & $0.6130_{(\pm 0.26 )}$ & $0.06_{(\pm 0.03 )}$ & $0.9302_{(\pm 0.01 )}$ & $0.5180_{(\pm 0.31 )}$ & $0.8608_{(\pm 0.19 )}$ \\
TSS SOL ($\mathcal{U}$) & 100 & $898.11_{(\pm 209.45 )}$ & $0.9515_{(\pm 0.02 )}$ & $0.65_{(\pm 0.15 )}$ & $0.9579_{(\pm 0.01 )}$ & $0.8634_{(\pm 0.20 )}$ & $0.8681_{(\pm 0.19 )}$ \\
TSS SOL ($\mathcal{C}(0.5,0.1)$) & 87 & $499.29_{(\pm 327.05 )}$ & $0.9517_{(\pm 0.02 )}$ & $0.52_{(\pm 0.02 )}$ & $0.9568_{(\pm 0.02 )}$ & $0.8699_{(\pm 0.18 )}$ & $0.8649_{(\pm 0.18 )}$ \\
TSS SOL ($\mathcal{C}(0.5,0.3)$) & 94 & $794.95_{(\pm 292.17 )}$ & $0.9561_{(\pm 0.01 )}$ & $0.57_{(\pm 0.06 )}$ & $0.9624_{(\pm 0.01 )}$ & $0.8641_{(\pm 0.20 )}$ & $0.8617_{(\pm 0.20 )}$ \\
TSS SOL ($\mathcal{C}(0.5,0.5)$) & 99 & $877.49_{(\pm 231.40 )}$ & $0.9562_{(\pm 0.01 )}$ & $0.61_{(\pm 0.08 )}$ & $0.9616_{(\pm 0.01 )}$ & $0.8676_{(\pm 0.19 )}$ & $0.8653_{(\pm 0.19 )}$ \\
\hline
\end{tabular}
}\vspace{0.5cm}
\resizebox{0.99\textwidth}{!}{
\begin{tabular}{|c | c c c c c | c c|}
\hline
& \multicolumn{5}{c|}{Training} & \multicolumn{2}{c|}{Test} \\
\cline{2-8}
Loss & \#success &
\#epoch &
CSI &
$\tau^{\star}$ &
CSI($\tau^{\star}$) & CSI & CSI($\tau^{\star}$) \\
\hline
CE & 100 & $1000.00_{(\pm 0.00 )}$ & $0.5222_{(\pm 0.22 )}$ & $0.44_{(\pm 0.09 )}$ & $0.6201_{(\pm 0.06 )}$ & $0.3583_{(\pm 0.23 )}$ & $0.4345_{(\pm 0.18 )}$ \\
CSI SOL ($\mathcal{U}$) & 100 & $697.03_{(\pm 258.69 )}$ & $0.6718_{(\pm 0.12 )}$ & $0.62_{(\pm 0.14 )}$ & $0.6962_{(\pm 0.07 )}$ & $0.4514_{(\pm 0.17 )}$ & $0.4613_{(\pm 0.17 )}$ \\
CSI SOL ($\mathcal{C}(0.5,0.1)$) & 91 & $398.26_{(\pm 221.41 )}$ & $0.6769_{(\pm 0.12 )}$ & $0.51_{(\pm 0.02 )}$ & $0.6913_{(\pm 0.09 )}$ & $0.4454_{(\pm 0.19 )}$ & $0.4427_{(\pm 0.21 )}$ \\
CSI SOL ($\mathcal{C}(0.5,0.3)$) & 100 & $528.04_{(\pm 251.06 )}$ & $0.6720_{(\pm 0.12 )}$ & $0.53_{(\pm 0.05 )}$ & $0.6951_{(\pm 0.06 )}$ & $0.4486_{(\pm 0.18 )}$ & $0.4582_{(\pm 0.18 )}$ \\
CSI SOL ($\mathcal{C}(0.5,0.5)$) & 100 & $618.63_{(\pm 267.47 )}$ & $0.6698_{(\pm 0.14 )}$ & $0.55_{(\pm 0.08 )}$ & $0.6955_{(\pm 0.08 )}$ & $0.4530_{(\pm 0.19 )}$ & $0.4582_{(\pm 0.19 )}$ \\
\hline
\end{tabular}
}
\caption{Comparison of binary crossentropy and SOL functions in the case of a neural network trained on the pollution dataset from UCI. The chosen SOL relies on the TSS (top table) and CSI (bottom table) scores and the threshold $\tau$ is drawn according to the uniform and raised cosine distributions. In columns "TSS" and "CSI" the scores are computed with respect to the threshold $\tau=0.5$.
}
\label{tab:poll_tss}
\end{table}
\section{Conclusions}\label{sec:conclusions}
In this paper, we discussed a new approach for the automatic maximization of a chosen score in classification tasks. After providing a detailed theoretical construction, we showed the effectiveness as well as the consistency of our framework. In particular, we described how this approach naturally allows the introduction of loss functions whose explicit form depends on the form of the skill scores defined on probabilistic confusion matrices. The performances of these loss functions during the training phase of supervised networks are tested in the case of two experimental forecasting problems.
We point out that the SOL functions that we have here introduced in the binary classification setting can be extended to the multi-class case, i.e. for $d>2$, in a rather straightforward fashion. In fact, the idea is to use a standard one-versus-rest approach in which $d$ different $2 \times 2$ matrices $\mathrm{CM}_1(\tau_1,\mathcal{S}(\boldsymbol{w})),\dots,\mathrm{CM}_d(\tau_d,\mathcal{S}(\boldsymbol{w}))$ are derived in the case of $\tau_1,\dots,\tau_d$ different random variables. In this way, the investigation of the multi-class case reduces to one of the binary setting.
As far as future work is concerned, we are currently following two paths.
From a methodological viewpoint, we think that the SOL functions approach should be extended to the case of value-weighted skill scores that allow the assessment of prediction errors on the basis of their impact on the decision making process \cite{guastavino2021}.
From the viewpoint of applications, our main interest is to exploit this approach in the framework of space weather \cite{camporeale2018machine} and, in particular, for the prediction of extreme solar flaring storms, whereby the training sets of solar magnetograms used for training the networks are strongly imbalanced against highly energetic events \cite{benvenuto2020machine}.
\section{Acknowledgements}
SG is financially supported by a regional grant of the `Fondo Sociale Europeo', Regione Liguria. FM is supported by the financial contribution from the agreement ASI-INAF n.2018-16-HH.0. MP and CC acknowledges this same agreement. This research has been partly accomplished within the Rete ITaliana di Approssimazione (RITA) and with the support of GNCS-IN$\delta$AM.
\bibliographystyle{siamplain}
|
\section{Introduction to $p$-Rational Number Fields}
In 2016, R. Greenberg described a method (in \cite{GRE16}) to construct Galois extensions of $\mathbb{Q}$ with Galois group isomorphic to an open subgroup of $GL_{n}(\mathbb{Z}_p)$ (for various values of $n$ and primes $p$). His method is based on the (conjectured) existence of \emph{$p$-rationnal fields} (as defined thereafter) of prescribed Galois group. More specifically, Greenberg conjectured that, for any odd prime $p$ and any natural integer $t$, there exists a $p$-rational number field $K$ whose Galois group over $\mathbb{Q}$ is $(\mathbb{Z} / 2 \mathbb{Z})^t$ (\cite{GRE16} Conjecture 4.8).
~\\
We begin by giving the reader a quick overview of the notion of $p$-rational fields.
Let $K$ be a number field (i.e. a finite extension of $\mathbb{Q}$), and $p$ a prime number. In the following, the various mathematical objects being defined will depend on $K$ and $p$, even if it does not appear in the notation most of the time.
An extension $L/K$ is said to be ``$p$-ramified'' if the extension is unramified outside the places above $p$. Let $M$ be the maximal $p$-ramified pro-$p$-extension of $K$, and $G = \mathrm{Gal}(M/K)$. We also denote by $M^{\mathrm{ab}}$ the maximal abelian $p$-ramified pro-$p$-extension of $K$, and $G^{\mathrm{ab}} = \mathrm{Gal}(M^{\mathrm{ab}} / K)$. One can notice that $G^{\mathrm{ab}}$ is the abelianization of $G$, and $M^{\mathrm{ab}}$ is the subfield of $M$ fixed by $G^{\mathrm{ab}}$. We also set $\widetilde{K}$ to denote the compositum of all $\mathbb{Z}_p$-extensions of $K$ (i.e. extensions of $K$ whose Galois group over $K$ is isomorphic to the additive group $\mathbb{Z}_p$) : as every $\mathbb{Z}_p$-extension is abelian and $p$-ramified, $\widetilde{K}$ is contained in $M^{\mathrm{ab}}$.
It is known, by class field theory, that $G^{\mathrm{ab}}$ is a $\mathbb{Z}_p$-module of rank $1 + r_2 + \delta$ (see \cite{MOV88} Chap II. \S 1 for example), where $\delta \geq 0$ is the defect in $p$ of the Leopoldt's conjecture for the number field $K$, and $r_2$ is the number of pairs of complex embeddings of $K$. Thus, there is an isomorphism of $\mathbb{Z}_p$-modules $G^{\mathrm{ab}} \simeq \mathbb{Z}_{p}^{1 + r_2 + \delta} \times \mathfrak{X}$,
where $\mathfrak{X}$ is the $\mathbb{Z}_p$-torsion sub-group of $G^{\mathrm{ab}}$. Using $\widetilde{K}$, we have : $\mathrm{Gal}(\widetilde{K}/K) \simeq \mathbb{Z}_{p}^{1 +r_2 + \delta}$ and $\mathrm{Gal}(M^{\mathrm{ab}} / \widetilde{K}) \simeq \mathfrak{X}$.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\[
\xymatrix{
M \\
M^{\mathrm{ab}} \ar@{-}[u] \\
\widetilde{K} \ar@{-}[u] \\
K \ar@{-}[u]
\ar@/^2pc/@{-}[uuu]^{\substack{\text{maximal} \\ p\text{-ramified} \\ \text{pro-}p\text{-extension} \\ \mathrm{Gal}(M/K) = G}}
\ar@/_1pc/@{-}[uu]_{\substack{\text{maximal} \\ \text{abelian} \\ p\text{-ramified} \\ \text{pro-}p\text{-extension} \\ \mathrm{Gal}(M^{\mathrm{ab}} /K) = G^{\mathrm{ab}} }}
}
\]
\end{subfigure}
\hspace{2.5cm}
\begin{subfigure}[b]{0.3\textwidth}
\[
\xymatrix{
M \\
M^{\mathrm{ab}} \ar@{-}[u] \\
\widetilde{K} \ar@{-}[u] \ar@/^1pc/@{-}[u]^{\mathfrak{X}} \\
K \ar@{-}[u]
\ar@/^1pc/@{-}[u]^{\mathbb{Z}_{p}^{1 + r_2 + \delta}}
\ar@/_1pc/@{-}[uu]_{\mathbb{Z}_{p}^{1 + r_2 + \delta} \times \mathfrak{X} }
}
\]
\end{subfigure}
\end{figure}
With the notations above, we have :
\begin{theorem}[\cite{MOV88} Prop. 1] \label{theodefi}
The following conditions are equivalent :
\begin{enumerate}
\item $K$ satisfies the Leopoldt's Conjecture in $p$ (i.e. $\delta = 0$) and $\mathfrak{X}$ is trivial.
\item $G^{\mathrm{ab}} = \mathrm{Gal}(M^{\mathrm{ab}} / K) \simeq \mathbb{Z}_{p}^{1 + r_2}$, i.e. $G^{\mathrm{ab}}$ is a free $\mathbb{Z}_p$-module of rank $1 + r_2$.
\item $G = \mathrm{Gal}(M / K)$ is a free pro-$p$-group with $1 + r_2$ generators.
\end{enumerate}
\end{theorem}
\begin{definition}[$p$-Rational Field] \label{defiprat} A number field $K$ is said to be ``$p$-rational'' if it satisfies any (and therefore all) of the conditions of Theorem~\ref{theodefi}.
\end{definition}
We can give some fundamental examples of $p$-rational fields :
\begin{enumerate}
\item The field of rational numbers $\mathbb{Q}$ is $p$-rational for every prime $p$.
\item The cyclotomic field $\mathbb{Q}(\zeta_{p^n})$ is $p$-rational for $p$ a regular prime (i.e. $p$ a prime such that $p$ does not divide the class number of $\mathbb{Q}(\zeta_p)$).
\item An imaginary quadratic field $K$ is $p$-rational for all (but a finite number of) primes $p$ (see \S 2).
\end{enumerate}
\begin{remark}
In the most general case, G. Gras conjectured in \cite{GRA16} that any number field $K$ is $p$-rational for all but a finite number of primes $p$.
\end{remark}
\begin{remark}
There exists algorithmic methods to determine if a given number field $K$ is $p$-rational for a given prime $p$, based on underlying profound algebraic results : see \cite{GRA19} for some useful \textsc{Pari/Gp} algorithms.
\end{remark}
We will use the following criterion of $p$-rationality for abelian number fields :
\begin{proposition}[\cite{GRE16} Proposition 3.6] \label{critsub}
Let $K$ be an abelian number field. Suppose that $[K : \mathbb{Q}]$ is not divisible by $p$. Then $K$ is $p$-rational if and only if every cyclic extension of $\mathbb{Q}$ contained in $K$ is $p$-rational.
\end{proposition}
\begin{remark} A special case, which will be of constant use throughout this paper, is the following : if $K$ is a multiquadratic number field (i.e. $\mathrm{Gal}(K/\mathbb{Q}) \simeq (\mathbb{Z} / 2 \mathbb{Z})^{t}$) and $p \neq 2$ is prime, then $K$ is $p$-rational if and only if all quadratic subfields of $K$ are $p$-rational.
In the most simple case, let $\mathbb{Q}(\sqrt{n})$ and $\mathbb{Q}(\sqrt{m})$ be two distinct quadratic number fields and let $p \neq 2$ a prime number. As stated previously, the biquadratic field $\mathbb{Q}(\sqrt{n}, \sqrt{m})$ is $p$-rational if and only if the quadratic subfields $\mathbb{Q}(\sqrt{n})$, $\mathbb{Q}(\sqrt{m})$ and $\mathbb{Q}(\sqrt{nm})$ are all $p$-rational. We can notice that it is not enough for $\mathbb{Q}(\sqrt{n})$ and $\mathbb{Q}(\sqrt{m})$ to be $p$-rational in order for their compositum $\mathbb{Q}(\sqrt{n}, \sqrt{m})$ to be $p$-rational as well. For example, the quadratic fields $\mathbb{Q}(\sqrt{2})$ and $\mathbb{Q}(\sqrt{19})$ are $5$-rational. Nonetheless, their compositum $\mathbb{Q}(\sqrt{2},\sqrt{19})$ is not $5$-rational, because its quadratic subfield $\mathbb{Q}(\sqrt{38})$ is not. (One can test the $5$-rationality of these fields using the \textsc{Pari/Gp} algorithm of \cite{GRA19}).
\end{remark}
The notion of $p$-rational fields has been investigated by many authors since (at least) the 1980's. They were used -- at first -- to exhibit non-abelian number fields satisfying the Leopoldt's conjecture.
In 2016, Greenberg revisited the notion in \cite{GRE16} and conjectured the following :
\begin{conjecture}[\cite{GRE16}, Greenberg, 2016] \label{conjgre} ~\\ For any odd prime $p$ and any natural integer $t$, there exists a $p$-rational number field $K$ whose Galois group over $\mathbb{Q}$ is $(\mathbb{Z} / 2 \mathbb{Z})^t$.
\end{conjecture}
This conjecture, and in particular the cases of quadratic and biquadratic fields, has been lately invastigated by many authors, such as Barbulescu and Ray (\cite{BR}, 2019), Gras (\cite{GRA19}, 2019), Assim and Bouazzaoui (\cite{AB19}, 2020), Benmerieme and Movahhedi (\cite{BM20}, 2021). They proved this conjecture for quadratic et biquadratic fields and all odd primes $p$ (the cases $t=1$ and $t=2$).
\section{Quadratic and Biquadratic $p$-Rational Number Fields}
\subsection{Quadratic $p$-rational number fiels}
\subsubsection{Criteria of $p$-rationality for quadratic numbers fields}
For the special (and elementary) case of quadratic fields, one has special criteria of $p$-rationality, such as the following :
\begin{proposition}[\cite{GRE16} Prop. 4.1 or \cite{BM20} Coro. 2.6] \label{critqua}
Suppose that $K$ is a quadratic field and that either $p \geq 5$ or that $p = 3$ and is unramified in $K / \mathbb{Q}$.
\begin{enumerate}
\item Assume that $K$ is imaginary. Then $K$ is $p$-rational if and only if the $p$-Hilbert class field of $K$ is contained in the anti-cyclotomic $\mathbb{Z}_p$-extension of $K$. In particular, if $h_K$ is not divisible by $p$, then $K$ is $p$-rational.
\item Assume that $K$ is real. Let $\varepsilon_0$ be the fundamental unit of $K$. Then $K$ is $p$-rational if and only if $p \nmid h_K$ and $\varepsilon_0$ is not a $p$-th power in the completion $K_v$, for at least a place $v \mid p$.
\end{enumerate}
\end{proposition}
\begin{remark} The assumption on $p$ only guarantees that $\mu_p \not\subset K_v$ for each place $v$ of $K$ dividing $p$. The previous proposition is proved in two different ways in \cite{GRE16} and \cite{BM20}. In \cite{BM20}, the authors also give alternative criteria of $p$-rationality for real quadratic fields.
\end{remark}
\subsubsection{Imaginary Case}
For $p$ an odd prime, we know the condition $p \nmid h_K$ is a sufficient condition of $p$-rationality when $K$ is quadratic imaginary and $p \geq 5$ (or $p=3$ and $p$ not ramified in $K/\mathbb{Q}$) (Propositon~\ref{critqua}). We can state important corollaries :
\begin{corollary}
Let $K$ be an imaginary quadratic number field. Then the set of primes $p$ for which $K$ is not $p$-rational is finite, and is contained in the set of divisors of $h_K$. In particular, an imaginary quadratic number field is $p$-rational for all but a finite number of primes $p$.
\end{corollary}
\begin{corollary}
For every odd prime $p$, the number field $\mathbb{Q}(i)$ is $p$-rational.
\end{corollary}
In general, to prove $p \nmid h_K$ for a given imaginary quadratic field, we use :
\begin{lemma}[\cite{LOU05} \S 1, Prop.2] \label{lemmlou}
If L is an imaginary quadratic field and $-d$ is its discriminant (with $d > 0$), we have
\[ h_L \leq \frac{w_L \cdot \sqrt{d_L}}{4\pi} (\log(d_L) + 1+ \gamma - \log(\pi)) \]
with
\[ w_L = \begin{cases}
6 & \text{if } d = 3 \\
4 & \text{if } d = 4 \\
2 & \text{if } d \geq 5
\end{cases} \]
and $\gamma$ the Euler constant. Moreover, we have $1+ \gamma - \log(\pi) \leq \frac{1}{2}$.
\end{lemma}
\begin{remark}
We can also show that, for every odd prime $p$, there even exists \emph{infinitely many} $p$-rationnal imaginary quadratic fields. To prove it, we can use a result by Hartung (\cite{HAR74}) proving there exists infinitely many imaginary quadratic fields $K$ such that $p$ does not divide the class number $h_K$.
The reader may notice Hartung proves this theorem only in the case $p=3$. But, as stated in the article itself, the method can be easily adapted to prove the theorem for every odd prime $p$.
\end{remark}
\subsubsection{Real Case}
The real case is much more complicated : units of these real quadratic fields play their part in the complexity of the setting.
\begin{example}
Some examples of real quadratic $p$-rational fields may even be found in the litterature prior to Greenberg's statement, for example in \cite{BYE01}, where the author does not use the terminology of $p$-rationality. However, for every prime $p \geq 5$, \cite{BYE01} Prop. 3.1 exhibits a real quadratic number field satisfying conditions which are sufficient to assert the $p$-rationality. The author even proves that there exists infinitely many real quadratic fields verifying these conditions, and such fields are $p$-rational. Thus, for every prime $p \geq 5$, there exists infinitely many $p$-rational real quadratic fields.
\end{example}
Many examples of $p$-rational real quadratic fields may be found in the litterature. In particular, we will use the following real quadratic fields :
\begin{proposition}[\cite{BM20} Prop. 4.4] \label{propmova}
For all prime $p \geq 5$, the real quadratic fields $\mathbb{Q}(\sqrt{p(p-2)})$, $\mathbb{Q}(\sqrt{p(p+2)})$ and $\mathbb{Q}(\sqrt{(p-2)(p+2)})$ are $p$-rational.
\end{proposition}
\begin{remark} Notice that the proof is made possible because, for all these fields, we have a simple explicit formula for their fundamental unit. We will use the $p$-rationality of these fields to prove our main theorem. \end{remark}
\subsection{Biquadratic $p$-Rational Number Fields}
Using known $p$-rational quadratic number fields, it is possible to construct biquadratic $p$-rational number fields (under certain conditions).
In 2020, Benmerieme and Movahhedi proved in \cite{BM20} the following :
\begin{proposition}[\cite{BM20} Prop. 4.4]
For all prime $p \geq 5$, the real biquadratic field \[ \mathbb{Q}(\sqrt{p(p-2)}, \sqrt{p(p+2)}) \] is $p$-rational.
\end{proposition}
\begin{remark} In \cite{BM20}, the authors prove this proposition by considering every quadratic subfield of $\mathbb{Q}(\sqrt{p(p-2)}, \sqrt{p(p+2)})$ : specifically, they prove $\mathbb{Q}(\sqrt{p(p+2)})$, $\mathbb{Q}(\sqrt{p(p-2)})$ and $\mathbb{Q}(\sqrt{(p+2)(p-2)})$ are $p$-rational for all prime $p \geq 5$ (see Proposition \ref{propmova}).
\end{remark}
\begin{remark} In a similar way, we could exhibit other $p$-rational biquadratic fields, for $p \geq 5$, such as $\mathbb{Q}(\sqrt{-p},\sqrt{-(p+2)})$ (\cite{BM20} Prop. 4.2), or $\mathbb{Q}(\sqrt{-(p-1)}, \sqrt{-(p+1)})$ for $p \equiv 3 \pmod{4}$ (using the $p$-rationality of $\mathbb{Q}(\sqrt{p^2 -1})$ as proved in this case by Byeon in \cite{BYE01}).
\end{remark}
\section{Triquadratic Rational Number Fields}
We state our main result :
\begin{theorem}
There exists infinitely many primes $p$ such that \[ \mathbb{Q}(\sqrt{p(p+2)}, \sqrt{p(p-2)},i) \] is $p$-rational.
\end{theorem}
\begin{proof}
To prove that $K = \mathbb{Q}(\sqrt{p(p+2)}, \sqrt{p(p-2)},i)$ is $p$-rational for infinitely many primes $p$, we will prove there exists infinitely many primes $p$ such that the following quadratic subfields are all $p$-rationnal :
\begin{multicols}{2}
\begin{enumerate}
\item $K_1 = \mathbb{Q}(\sqrt{p(p+2)})$
\item $K_2 = \mathbb{Q}(\sqrt{p(p-2)})$
\item $K_3 = \mathbb{Q}(\sqrt{(p+2)(p-2)})$
\item $K_4 = \mathbb{Q}(i)$
\item $K_5 = \mathbb{Q}(i\sqrt{p(p+2)})$
\item $K_6 = \mathbb{Q}(i\sqrt{p(p-2)})$
\item $K_7 = \mathbb{Q}(i\sqrt{(p+2)(p-2)})$.
\end{enumerate}
\end{multicols}
Indeed, we already know $K_4$ is $p$-rationnal for all odd prime $p$. We also know $K_1$, $K_2$ et $K_3$ are $p$-rational for all prime $p \geq 5$ , according to Proposition \ref{propmova}.
Then, we use the analytic proposition proved apart in the next section (Proposition~\ref{propan})~: for $A > 0$, there exists infinitely many primes $p$ such that $p-2$ has a square factor larger than $(\log p)^A$ and $p+2$ has a square factor larger than $(\log p)^A$ as well. In particular, we can set $A = 2$, and we get, for all such $p$~:
\begin{align*}
\mathrm{disc}(K_5) & \leq \frac{4p(p+2)}{(\log p)^{4}} \\
\mathrm{disc}(K_6) & \leq \frac{4p(p-2)}{(\log p)^{4}} \\
\mathrm{disc}(K_7) & \leq \frac{4(p+2)(p-2)}{(\log p)^{8}}
\end{align*}
From now on, we'll only consider primes $p \geq 5$ lying in the infinite set of primes described previously, which we will denote by $\mathcal{P}$.
Using Lemma~\ref{lemmlou}, for every $p \in \mathcal{P}$, we get :
\allowdisplaybreaks
\begin{align*}
h(K_5) & \leq \frac{6}{4\pi} \sqrt{ \frac{4p(p+2)}{(\log p)^4} } \left( \log \left( \frac{4p(p+2)}{(\log p)^4} \right) + \frac{1}{2} \right) \ll \frac{p}{\log(p)} \\
h(K_6) & \leq \frac{6}{4\pi} \sqrt{ \frac{4p(p-2)}{(\log p)^4} } \left( \log \left( \frac{4p(p-2)}{(\log p)^4} \right) + \frac{1}{2} \right) \ll \frac{p}{\log(p)} \\
h(K_7) & \leq \frac{6}{4\pi} \sqrt{ \frac{4(p+2)(p+2)}{(\log p)^8} } \left( \log \left( \frac{4(p+2)(p+2)}{(\log p)^8} \right) + \frac{1}{2} \right) \ll \frac{p}{(\log p)^3}
\end{align*}
with absolute and effective implied constants.
Then, as $p \in \mathcal{P}$ tends to infinity, $p$ gets larger than $h(K_5)$, $h(K_6)$ and $h(K_7)$.
Thus, if $p$ is a big enough prime in $\mathcal{P}$, then $p \nmid h(K_5)$, $p \nmid h(K_6)$ and $p \nmid h(K_7)$, and these fields are $p$-rational.
In conclusion, there exists infinitely many primes $p$ such that all quadratic subfields of $K = \mathbb{Q}(\sqrt{p(p+2)},\sqrt{p(p-2)},i)$ are $p$-rational, so $K$ itself is $p$-rational for infinitely many primes $p$.
\end{proof}
\section{Proof of the Analytic Proposition}
Now, we are going to prove the analytic proposition we previously used to prove our main theorem :
\begin{proposition} \label{propan} Let $A > 0$.
There exists infinitely many primes $p$ such that : there exists $m,n \in \mathbb{N}$ with
\[ \begin{cases} n^2 \mid p-2 \\
m^2 \mid p+2 \\
(\log p)^A < n \\
(\log p)^A < m
\end{cases} \]
In other words, there exists infinitely many primes $p$ such that : $p-2$ has a square factor larger than $(\log p)^A$, and $p+2$ has a square factor larger than $(\log p)^A$.
\end{proposition}
Notations : for positive integers $m,n$, we set
\begin{align*}
G(m) & : = \left\{ p \text{ prime} \left| \begin{array}{l}
p \equiv -2 \pmod{m^2} \\
(\log p)^A < m \\ \end{array} \right. \right\} \\
H(n) & := \left\{ p \text{ prime} \left| \begin{array}{l}
p \equiv 2 \pmod{n^2} \\
(\log p)^A < n \\ \end{array} \right. \right\}
\end{align*}
and
\[ I(m,n) := G(m) \cap H(n) = \left\{ p \text{ prime} \left| \begin{array}{l} p \equiv - 2 \pmod{m^2} \\
p \equiv 2 \pmod{n^2} \\
(\log p)^A < m\\
(\log p)^A < n \\ \end{array} \right. \right\} \]
In order to prove the previous proposition, it is sufficient to show that the sum
\[ \sum_{\substack{p \text{ prime} \\ 3 \leq p < x}} \log(p) \left( \sum_{\substack{m \in \mathbb{N} \\ m < \sqrt{x+2} \\ p \in G(m) }} 1 \right) \times \left( \sum_{\substack{n \in \mathbb{N} \\ n < \sqrt{x-2} \\ p \in H(n) }} 1 \right)
\]
tends to infinity as $x \rightarrow + \infty$.
\subsection{First lower bound}
\allowdisplaybreaks
First, we have
\begin{align}
& \sum_{\substack{p \text{ prime} \\ 3 \leq p < x}} \log(p) \left( \sum_{\substack{m \in \mathbb{N} \\ m < \sqrt{x+2} \\ p \in G(m) }} 1 \right) \times \left( \sum_{\substack{n \in \mathbb{N} \\ n < \sqrt{x-2} \\ p \in H(n) }} 1 \right) \label{somme1} \\
& = \sum_{\substack{p \text{ prime} \\ 3 \leq p < x}} \sum_{\substack{m,n \in \mathbb{N} \\ m < \sqrt{x+2} \\ n < \sqrt{x-2} \\ p \in I(m,n)}} \log(p) \\
& = \sum_{\substack{m,n \in \mathbb{N} \\ m < \sqrt{x+2} \\ n < \sqrt{x-2} }} \sum_{\substack{p \text{ prime} \\ 3 \leq p < x \\ p\in I(m,n)}} (\log p) \\
& \geq \sum_{\substack{m,n \in \mathbb{N} \\ m < \sqrt{x+2} \\ n < \sqrt{x-2} \\ (m,n) = 1 \\ m,n \text{ odd}}} \sum_{\substack{p \text{ prime} \\ 3 \leq p < x \\ p \in I(m,n)}} \log(p) \label{somme2}
\end{align}
We obtain a first lower bound by restricting the first sum (on $m,n \in \mathbb{N}$) with the conditions $(m,n)=1$ and $m,n$ odd. We add the first condition in order to use the Chinese remainder theorem : as $n$ and $m$ are coprime, so are $m^2$ and $n^2$, and there exists $a_{m,n} \in \mathbb{Z}$ such that, for all $k \in \mathbb{Z}$, we have :
\[ \begin{cases} k \equiv 2 & [n^2] \\ k \equiv -2 & [m^2] \end{cases} \quad \iff \quad k \equiv a_{m,n} [m^2 n^2] \]
Then, if $m$ and $n$ are coprime, we get
\[ I(m,n) = \left\{ p \text{ prime} \left| \begin{array}{l} p \equiv a_{m,n} \pmod{m^2 n^2} \\ (\log p)^A < m \\ (\log p)^A < n \end{array} \right. \right\} \]
Moreover, the condition ``$m,n$ odd'' ensures that $a_{m,n}$ and $m^2 n^2$ are coprime.
~\\
Let $B>0$ such that $A <B$. For $x$ big enough (in particular, such that $\sqrt{x-2}$ is larger than $(\log x)^B$), we may bound~\eqref{somme2} from below by
\begin{equation} \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1 \\ m,n \text{ odd}}} \sum_{\substack{ 3 \leq p < x \\ p \in I(m,n) }} \log(p) \label{somme3}
\end{equation}
We notice that :
\[ \begin{cases} p < x \\ (\log p)^A < n \\ (\log p)^A < m \end{cases} \iff \begin{cases} (\log p)^A < (\log x)^A \\ (\log p)^A < n \\ (\log p)^A < m \end{cases} \iff (\log p)^A < \min\left( (\log x)^A, n, m \right) \]
Then, if $(\log x)^A < m,n$ (as in~\eqref{somme3}), we get $\min((\log x)^A,n,m) = (\log(x))^A$, and the previous conditions are equivalent to $(\log p)^A < (\log x)^A$, i.e. $p < x$.
Thus~\eqref{somme3} is equal to
\begin{equation} \label{somme4}
\sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \sum_{\substack{ 3 \leq p < x \\ p \equiv a_{m,n} [m^2 n^2] }} \log(p)
\end{equation}
which is a convenient lower bound of our first sum~\eqref{somme1}.
\subsection{Siegel-Walfisz's Theorem}
\begin{definition}
For $(a,q)=1$, we define
\[ \theta(x;q,a) = \sum_{\substack{ p \text{ prime} \\ p < x \\ p \equiv a [q] }} \log(p) \qquad \text{et} \qquad \psi(x;a,q) = \sum_{\substack{ n < x \\ n \equiv a [q] }} \Lambda(n) \]
with $\Lambda(n) = \begin{cases} \log(p) & \text{if } n = p^k \text{ with } p \text{ prime} \\
0 & \text{else} \end{cases}$
\end{definition}
The prime number theorem gives an equivalent at infinity for $\theta$ and $\psi$. An estimation of the error term is given by the following theorem :
\begin{theorem}[\textsc{Siegel-Walfisz}, \cite{IWAKOW} Corollary 5.29] Let $a,q \in \mathbb{N}$, $q \geq 1$, $(a,q) = 1$. Let $C > 0$. For $q \ll \log(x)^C$, the following holds
\[ \psi(x;q,a) = \frac{x}{\varphi(q)} + O\left( \frac{x}{(\log x)^C} \right) \]
for every $x \geq 2$.
The implied constant depends only on $C$.
\end{theorem}
Later in the proof, we will need this result using $\theta$ instead of $\psi$.
First, we check that $\displaystyle \psi(x;q,a) = \theta(x;q,a) + O \left( x^\frac{1}{2} \log(x)^2 \right)$ (see \cite{HW} XXII. Theorem 413 p.452)
Then, the previous theorem may be adapted in the following way :
\begin{corollary} \label{coroSW} Let $a,q \in \mathbb{N}$, $q \geq 1$, $(a,q) = 1$.
Let $C > 0$. For $q \ll \log(x)^C$, the following holds
\[ \theta(x;q,a) = \frac{x}{\varphi(q)} + O\left( \frac{x}{(\log x)^C} \right) \]
for every $x \geq 2$. The implied constant depends only on $C$.
\end{corollary}
\subsection{Equivalent at infinity}
We want to prove sum~\eqref{somme1} tends to infinity as $x \rightarrow + \infty$. We bounded~\eqref{somme1} from below by~\eqref{somme4}, which again can be bounded from below (thanks to the previous notations) by :
\begin{equation}
\sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \left[ \theta \left( x , a_{m,n}, m^2 n^2 \right) - \log(2) \right]
\end{equation}
Remark : the definition of $\theta$ that we previously gave also involves $p=2$, while our initial sum did not, which explains the appearance of a $\log(2)$ term, which makes no difference whatsoever in the end.
~\\
By Corollary \ref{coroSW}, if we set $C > 4B$, we have $m^2 n^2 \leq (\log x)^{4B} < (\log x)^C$, and we get
\begin{equation}
\sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \theta \left( x , a_{m,n}, m^2 n^2 \right) =
\sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \left[ \frac{x}{\varphi(m^2 n^2)} + E_{m,n}(x) \right]
\end{equation}
with $\displaystyle E_{m,n}(x) \ll \frac{x}{(\log x)^C}$, the implied constant depending only on $C$.
We can notice that we are able to use these estimates as we chose $m,n$ odd, so that the residue $a_{m,n}$ and the modulus $m^2 n^2$ are coprime.
\newpage
Thus, we need to prove :
\begin{enumerate} \item[(i)] that the main term
\[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{x}{\varphi(m^2 n^2)} \]
tends tends to infinity as $x \rightarrow + \infty$, and
\item[(ii)] that the growth of the main term is not impeached by the error term~:
\[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} E_{m,n}(x) - \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \log(2) \]
\end{enumerate}
\subsection{Estimation of the Main Term}
We are going to prove that
\[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{x}{\varphi(m^2 n^2)} \underset{x \rightarrow \infty}{\longrightarrow} \infty \]
by bounding this term below and give an equivalent at infinity of this lower bound. First, we use the inequality $\varphi(m^2 n^2) \leq m^2 n^2$, so that
\begin{equation}
\sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{x}{\varphi(m^2 n^2)} \geq \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{x}{m^2 n^2}
\end{equation}
Then, we get
\begin{align*}
& \displaystyle \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{1}{m^2 n^2} \\
= & \displaystyle \sum_{\substack{ (\log x)^A < (2m+1),(2n+1) < (\log x)^B \\ ((2m+1),(2n+1))=1}} \frac{1}{(2m+1)^2 (2n+1)^2} \\
= & \displaystyle \sum_{ (\log x)^A < (2m+1), (2n+1) < (\log x)^B } \frac{1}{(2m+1)^2 (2n+1)^2} \\
& \qquad - \underbrace{ \sum_{\substack{ (\log x)^A < (2m+1),(2n+1) < (\log x)^B \\ ((2m+1),(2n+1))> 1}} \frac{1}{(2m+1)^2 (2n+1)^2} }_{(*)} \\
\geq & \displaystyle \sum_{ \frac{(\log x)^A - 1}{2} < m,n < \frac{(\log x)^B - 1}{2} } \frac{1}{(2m+1)^2 (2n+1)^2} \\
& \qquad - \underbrace{\sum_{2 \leq d \leq (\log x)^B} \sum_{\substack{ (\log x)^A < (2m+1), (2n+1) < (\log x)^B \\ d \mid 2m+1 \\ d \mid 2n + 1}} \frac{1}{(2m+1)^2 (2n+1)^2}}_{(**)} \\
\geq & \displaystyle \sum_{ \frac{(\log x)^A - 1}{2} < m,n < \frac{(\log x)^B - 1}{2} } \frac{1}{(2m+1)^2 (2n+1)^2} \\
& \qquad - \underbrace{ \sum_{2 \leq d \leq (\log x)^B} \sum_{\substack{ k,k' \text{ odd} \\ (\log x)^A < dk, dk < (\log x)^B }} \frac{1}{(dk)^2 (dk')^2} }_{(***)} \\
\end{align*}
Note that we have $(*) \leq (**) \leq (***)$, which explains the inequalities.
Thus, the latter is equal to
\[ \left( \sum_{ \frac{(\log x)^A - 1}{2} < n < \frac{(\log x)^B - 1}{2} } \frac{1}{ (2n+1)^2} \right)^2 - \sum_{2 \leq d \leq (\log x)^B} \frac{1}{d^4} \left( \sum_{\substack{ \frac{(\log x)^A - d}{2d} < n < \frac{(\log x)^B - d}{2d} }} \frac{1}{(2n+1)^2} \right)^2 \]
We now want to bound the last quantity from below. To do this, one uses lower and upper bounds which we get by comparisons beetwen series and integrals.
Specifically, we use the inequality
$\displaystyle \sum_{n=a}^{b} \frac{1}{(2n+1)^2} \geq \int_{a}^{b+1} \frac{\mathrm{d}t}{(2t+1)^2}$
so that we get
\[ \sum_{ \frac{(\log x)^A -1}{2} < n < \frac{(\log x)^B - 1}{2} } \frac{1}{(2n+1)^2}
\geq \frac{1}{2 \left( (\log x)^A + 2 \right)} - \frac{1}{2 (\log x)^B}
\]
In the same way, we use the inequality $\displaystyle \sum_{n=a}^{b} \frac{1}{(2n+1)^2} \leq \int_{a}^{b} \frac{\mathrm{d}t}{(2t+1)^2} + \frac{1}{(2a+1)^2}$
to get
\[
\displaystyle \sum_{\substack{ \frac{(\log x)^A - d}{2d} < n < \frac{(\log x)^B - d}{2d} }} \frac{1}{(2n+1)^2} \leq \displaystyle \frac{d}{2(\log x)^A } - \frac{d}{2(\log x)^B} + \frac{d^2}{(\log x)^{2A}}
\]
Using the previous inequalities, we have
\allowdisplaybreaks
\begin{align*} & \displaystyle \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{1}{m^2 n^2} \\
\geq\;& \displaystyle \left( \sum_{ \frac{(\log x)^A - 1}{2} < n < \frac{(\log x)^B - 1}{2} } \frac{1}{ (2n+1)^2} \right)^2 \\
& \qquad - \sum_{2 \leq d \leq (\log x)^B} \frac{1}{d^4} \left( \sum_{\substack{ \frac{(\log x)^A - d}{2d} < n < \frac{(\log x)^B-d}{2d} }} \frac{1}{n^2} \right)^2 \\
\geq\; & \displaystyle \left( \frac{1}{2 \left( (\log x)^A + 2 \right)} - \frac{1}{2 (\log x)^B} \right)^2 \\
& \qquad - \sum_{2 \leq d \leq (\log x)^B} \frac{1}{d^4} \left( \frac{d}{2(\log x)^A } - \frac{d}{2(\log x)^B} + \frac{d^2}{(\log x)^{2A}} \right)^2 \\
\geq\;& \displaystyle \frac{1}{4 \left( (\log x)^A + 2 \right)^2 } + \frac{1}{4 (\log x)^{2B} } - \frac{1}{2 \left( (\log x)^A + 2 \right) (\log x)^B } \\
& \displaystyle - \sum_{2 \leq d \leq (\log x)^B} \frac{1}{d^2} \left( \frac{1}{4(\log x)^{2A} } + \frac{1}{4(\log x)^{2B}} + \frac{d^2}{(\log x)^{4A}} + \frac{d}{(\log x)^{3A}} \right) \\
\geq\;& \displaystyle \frac{1}{4 \left( (\log x)^A + 2 \right)^2 } + \frac{1}{4 (\log x)^{2B} } - \frac{1}{2 \left( (\log x)^A + 2 \right) (\log x)^B } \\
& \displaystyle - \left( \sum_{2 \leq d \leq (\log x)^B} \frac{1}{d^2} \right) \left( \frac{1}{4(\log x)^{2A} } + \frac{1}{4(\log x)^{2B}} \right) \\
& \displaystyle - \left( \sum_{2 \leq d \leq (\log x)^B} \frac{1}{d} \right) \frac{1}{(\log x)^{3A}} - \left( \sum_{2 \leq d \leq (\log x)^B} 1 \right) \frac{1}{(\log x)^{4A}} \\
\geq\;& \displaystyle \frac{1}{4 \left( (\log x)^A + 2 \right)^2 } + \frac{1}{4 (\log x)^{2B} } - \frac{1}{2 \left( (\log x)^A + 2 \right) (\log x)^B } \\
& \displaystyle - (\zeta(2) - 1) \left( \frac{1}{4(\log x)^{2A} } + \frac{1}{4(\log x)^{2B}} \right) \\
& \displaystyle - (1 + \log(\log(x)^B) ) \frac{1}{(\log x)^{3A}} - \frac{(\log x)^B}{(\log x)^{4A}} \\
\end{align*}
and, if we take $A < B < 2A$, the latter is finally equivalent to $\displaystyle \frac{2 - \zeta(2)}{4} \frac{1}{(\log x)^{2A}}$
Thus, we proved that
\[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{1}{m^2 n^2} \]
is larger than a quantity which is equivalent at infinity to
\[ \frac{2 - \zeta(2)}{4} \frac{1}{(\log x)^{2A}} \]
Subsequently, the main term
\[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ impairs}}} \frac{x}{\varphi(m^2 n^2)} \]
is larger than a quantity which is equivalent at infinity to
\begin{equation}
\frac{2 - \zeta(2)}{4} \frac{x}{(\log x)^{2A}} \label{mino}
\end{equation}
Thus, the main term tends to infinity as $x \rightarrow + \infty$, at least as fast as~\eqref{mino}.
\subsection{Error Terms}
Remember that $E_{m,n}(x) = \left| \theta(x; a, m^2 n^2 ) - \frac{x}{\varphi(m^2 n^2)} \right|$. According to Siegel-Walfisz, for a given $C > 0$ and every $q \leq (\log x)^C$, the following holds
\[ \left| \theta(x; a, q ) - \frac{x}{\varphi(q)} \right| = O \left( \frac{x}{(\log x)^C} \right) \]
Then, if $4B < C$, for every big enough $x$, we know that $m^2 n^2 < (\log x)^C$ as soon as $(\log x)^A < m,n < (\log x)^B$. So each $E_{m,n}(x)$ is dominated by $\displaystyle \frac{x}{(\log x)^C}$ in the sum \[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} E_{m,n}(x)\]
Considering that we get at most $(\log x)^{2B}$ terms, we have~:
\[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} E_{m,n}(x) \ll (\log x)^{2B} \frac{x}{(\log x)^C} \]
Noticing that we chose $C$ such that $C > 4B$, we have $C > 2A + 2B$, i.e. $2B - C < - 2A$ and we can conclude that
\[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} E_{m,n}(x) \ll \frac{x}{(\log x)^{2A}} \]
The last part of the error term, namely \[ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \log(2) \] is at most $(\log x)^{2B}$ times $\log(2)$, so it may be bounded above by $(\log x)^{2B} \log(2)$, which is also dominated by$ \frac{x}{(\log x)^{2A}}$.
\subsection{Conclusion}
We proved
\[ \sum_{\substack{p \text{ prime} \\ 3 \leq p < x}} \log(p) \left( \sum_{\substack{m \in \mathbb{N} \\ m < \sqrt{x+2} \\ p \in G(m) }} 1 \right) \times \left( \sum_{\substack{n \in \mathbb{N} \\ n < \sqrt{x-2} \\ p \in H(n) }} 1 \right)
\]
was bounded below by
\[ \underbrace{\sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \frac{x}{\varphi(m^2 n^2)} }_{A(x)} + \underbrace{ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} E(x)}_{O(A(x))} + \underbrace{ \sum_{\substack{ (\log x)^A < m,n < (\log x)^B \\ (m,n)=1\\ m,n \text{ odd}}} \log(2) }_{O(A(x))} \]
with $A(x) \rightarrow \infty$, showing that our first sum tends to infinity as $x \rightarrow \infty$. \quad $\square$
\section{Open questions}
\begin{enumerate}
\item In order to prove Greenberg's conjecture for every integer $t$ and every odd prime $p$, one could try at first to exhibit $p$-rational totally real triquadratic number fields. This would require to exhibit a well chosen set of real quadratic fields, whose fundamental units can be explicitally found.
\item It is possible, being given a $p$-rational totally real multiquadratic field $K$ of order $2^t$ for every odd prime $p$, that we could adapt our analytic proposition to prove $K(i)$ is $p$-rationnal for an infinity of prime $p$ (probably under certain restricting conditions).
\item The analytic proposition may be improved : for example, one may consider a finite collection of positive integers $m_i$ (and relatively prime residues $r_i$) and search for primes $p$ such that
\[ \forall i,\quad p \equiv r_i \pmod{m_{i}^{2}} \quad \text{and}\quad m_{i} > (\log p)^A \]
One may also give a proper estimation of the number of such $p$, rather than the lower bound we gave.
In appendix, we prove an alternate (stronger) version (under GRH) of the analytic proposition \ref{propan}.
\end{enumerate}
\begin{appendix}
\section{A stronger analytic proposition under GRH}
For $A > 0$, we proved there exists infinitely many primes $p$ such that $(p+2)$ and $(p-2)$ both admit square factors larger than $(\log x)^A$. We can prove, under GRH, a stronger result. More specifically :
\begin{proposition}
Let $\varepsilon < \frac{1}{8} $. Suppose GRH holds true for $L(s,\chi)$ with $\chi \pmod{q}$. There exists infinitely many primes $p$ for which there exists $m,n \in \mathbb{N}^{\ast}$, such that
\[ \begin{cases}
n^2 \mid p-2 \\
m^2 \mid p+2 \\
p^{\varepsilon} < n \\
p^{\varepsilon} < m
\end{cases} \]
\end{proposition}
\begin{proof} The proof is similar to the previous one.
In the same way, we consider the sum
\[
\sum_{\substack{p \text{ prime} \\ 3 \leq p < x}} \log(p) \Bigg( \sum_{\substack{m \in \mathbb{N} \\ m < \sqrt{x+2} \\ m^2 \mid p + 2 \\ p^\varepsilon < m }} 1 \Bigg) \times \Bigg( \sum_{\substack{ n \in \mathbb{N} \\ n < \sqrt{x-2} \\ n^2 \mid p -2 \\ p^\varepsilon < n }} 1 \Bigg)
\]
Then, using similar calculations as before, we show that the previous sum is bounded below by
\[
\sum_{\substack{m,n \in \mathbb{N} \\ m < \sqrt{x+2} \\ n < \sqrt{x-2} \\ (m,n)=1 \\ m,n \text{ odd}}} \sum_{\substack{ p \text{ prime} \\ 3 \leq p < x \\ p \equiv a_{m,n} [m^2 n^2] \\ p < m^{1/\varepsilon} \\ p < n^{1 / \varepsilon}}} \log(p)
\]
which is larger (for $x$ big enough) than
\begin{equation} \label{sommea}
\sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} \sum^{p \text{ prime}}_{\substack{p \equiv a_{m,n} [m^2n^2] \\ 3 \leq p < x}} \log(p)
\end{equation}
for $\varepsilon < \alpha < \frac{1}{2}$.
~\\
Then, under GRH for $L(s,\chi)$ with $\chi \pmod{q}$, we have (see \cite{IWAKOW} \S 17.1) :
\[ \theta(x; m^2 n^2, a_{m,n}) = \sum_{\substack{p \text{ prime} \\ p \equiv a_{m,n} [m^2n^2] \\ p < x}} \log(p) = \frac{x}{\varphi(m^2 n^2)} + O \left( x^{1/2} (\log x)^2 \right)
\]
So~\eqref{sommea} is greater than
\[ \sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} \left( \frac{x}{\varphi(m^2 n^2)} + E_{m,n} - \log(2) \right)
\]
with $E_{m,n} \ll x^{1/2} (\log x)^2$.
~\\
First, we show the main term
\[
\sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} \frac{x}{\varphi(m^2 n^2)}
\]
is larger than a quantity equivalent to $\displaystyle \frac{2 - \zeta(2)}{4} x^{1 - 2 \varepsilon}$ at infinity.
Indead, we have
\begin{align*}
\sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1\\ m,n \text{ odd}}} \frac{1}{m^2 n^2} & = \left( \sum_{ \frac{x^\varepsilon - 1}{2} < n < \frac{x^\alpha - 1}{2} } \frac{1}{ (2n+1)^2} \right)^2 \\
& \quad - \sum_{2 \leq d \leq x^\alpha} \frac{1}{d^4} \left( \sum_{\substack{ \frac{x^\varepsilon - d}{2d} < n < \frac{x^\alpha - d}{2d} }} \frac{1}{(2n+1)^2} \right)^2 \\
& \geq \left( \frac{1}{2(x^\varepsilon +2)} - \frac{1}{2x^\alpha} \right)^2 - \sum_{2 \leq d \leq x^\alpha} \frac{1}{d^4} \left( \frac{d}{2(x^\varepsilon - 2d)} - \frac{d}{2x^\alpha} \right)^2 \\
& \geq \frac{1}{4(x^\varepsilon + 2)^2} - \frac{1}{2(x^\varepsilon +2)x^\alpha} + \frac{1}{4 x^{2\alpha}} \\
& \quad - \frac{\zeta(2)-1}{4 (x^\varepsilon - 4)^2} + \frac{\zeta(2)-1}{2(x^\varepsilon - 2x^\alpha)x^\alpha} - \frac{\zeta(2) -1}{4x^{2\alpha}}
\end{align*}
Therefore we can consider the first error term $\sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} E_{m,n}(x)$ with $E_{m,n}(x) \ll x^{1/2} (\log x)^2$ (the implied constant being independent of $E_{m,n}$). As we get at most $x^{2\alpha}$ terms which are $O \left( x^{1/2} (\log x)^2 \right)$ (under GRH), this error term is $O \left( x^{\frac{1}{2} + 2 \alpha} (\log x)^2 \right)$.
Thus, we take $\alpha$ such that $\varepsilon < \alpha < \frac{1}{4} - \varepsilon$ (justifying the fact that we took $\varepsilon < \frac{1}{8}$), so that $\frac{1}{2} + 2 \alpha < 1 - 2 \varepsilon$. Then, this first error term is $O \left( \frac{2 -\zeta(2)}{4} x^{1 - 2\varepsilon} \right)$.
Finally, the second error term, namely
$\sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} (\log 2)$
is at most $x^{2 \alpha}$ times $\log(2)$, and may be bounded above by $x^{2 \alpha} \log(2) = O \left( \frac{2 - \zeta(2) }{4} x^{1 - 2 \varepsilon} \right)$ (as $\alpha$ has been set previously such that $2 \alpha < \frac{1}{2} < 1 - 2\varepsilon$)
In conclusion, we bounded below our first sum by :
\[
\underbrace{\sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} \frac{x}{\varphi(m^2 n^2)}}_{A(x)} + \underbrace{\sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} E_{m,n} + \sum_{\substack{ x^\varepsilon < m,n < x^\alpha \\ (m,n)=1 \\ m,n \text{ odd}}} \log(2)}_{O(A(x))}
\]
with $A(x) \rightarrow + \infty$, which is sufficient to prove the proposition.
\end{proof}
\end{appendix}
|
\section{Introduction}
Infection with the SARS-CoV-2 virus represents a
serious health risk. A number of studies have by now established that this risk increases exponentially with age (see \cite{levin2020assessing} for a survey and metastudy). This applies both to the risk of dying and
to the risk of a severe course of the illness
requiring hospitalisation \cite{CDC_age_risk}.
Vaccination
campaigns have therefore prioritized the elderly
(along with health workers). Countries with active
outbreaks have in general introduced NPIs to restrict
mobility and reduce the number of personal contacts.\cite{chen2020tracking}
However, NPIs have a high social and economic cost.
This cost arises both through the limitations to
economic activity \cite{cutler2020covid} and the
health costs from those catching the virus and
becoming infected \cite{gros2020containment}.
It is
thus key to understand when NPIs can be lifted safely.
This has lead to a race between mass scale vaccination campaigns
and the disease, which continued to spread and evolve.
\cite{dye2020scale} provides an analysis of the scale
and the spread of the disease in 2020. \cite{zheng2021mathematical} provides a mathematical model.
\cite{gnanvi2021reliability} provides an overview
of the statistical models used to forecast the
spread of the disease. A notable application of statistical modelling to the European experience is provided by \cite{agosto2021monitoring} and (using
a Poisson process) \cite{agosto2020poisson}.
Moreover, new variants have appeared \cite{bauer2021relaxing}, with new and potentially
more infectious strands \cite{davies2021estimated}.
The ultimate aim of vaccination campaigns is to
achieve `herd' or population immunity, i.e.\ a
state in which the share of the population that
has been immunized by infection or vaccination
brings down the effective reproduction number
$R\textsubscript{eff}$ below unity without further
NPIs needed \cite{linka2020reproduction}. However,
vaccination takes time, mostly due to the difficulties
in ramping up the production of vaccines \cite{gros2021incentives},
but also in part also due to bottlenecks
in distribution and implementation.
A core issue for policy makers therefore is
the point during the vaccination campaign
at which NPIs can be lifted.
We provide a general rule to answer this question.
Our proposed rule is based only on observable data
and is the result of combining two simple relationships
\begin{itemize}
\item
The age-dependency of the case fatality and the
hospitalization rate, which has been established
to increase approximately
exponentially with age \cite{levin2020assessing}.
\item
The population structure for the elderly, which is
to first order linear at the top, which means that
the size of age cohorts increases gradually top-down
from the maximum age (about 100 years)
\cite{grundy2017population}.
\end{itemize}
We concentrate on observable outcomes like death or
hospitalisation because their number constitutes
a key determinant for the imposition of lock-downs
\cite{ferraresi2020great} and other NPIs, which
come with severe economic and social costs in terms
of lost output and employment \cite{dreger2021lockdowns}
as well as indirect impacts on health and well-being \cite{Brodeur}.
The combination of these two elements leads
to a simple rule
of thumb to recognise the 'sweet point' at
which NPIs can
be loosened while still 'flattening the curve'.
We then add a third element, namely
\begin{itemize}
\item
The functional dependency of daily vaccinations rates,
which are observed to increase in most countries
approximately linearly over time.
\end{itemize}
This third element allows one to predict
the stage in the vaccination process at
which the 'sweet point' will be reached.
We concentrate throughout mostly on the
case of Europe, because this is the region
in which a resurgence of infections coincided
with an, initially at least, sluggish vaccination campaign.
\section{Modeling framework}
\begin{figure}[!t]
\centering
\includegraphics[width=0.95\columnwidth]{fig_ageDistribution.pdf}
\newline
\includegraphics[width=0.95\columnwidth]{fig_ageQuadratic.pdf}
\caption{{\bf Age distributions.}
(Top) Fractional yearly age cohorts for
Germany (DE), Ireland (IE), Italy (IT) and
EU27 (EU). Also shown is a linear interpolation
for the EU data (black line), for ages 60 and
above. Data from \cite{eurostat21}.
(Bottom) Five-year age cohorts for India
and Nigeria. A quadratic interpolation to
the data for Nigeria is also provided.
}
\label{fig_ageDistribution}
\end{figure}
\subsection{Age-dependency of the health risks of Covid-19 infections}
We concentrate on two observable health risks:
death and hospitalisation
It has been widely documented that the risk to
die from a Covid-19 infection rises strongly
with age. A meta-study suggests an exponential
relationship \cite{levin2020assessing}, which
can be parameterized as
\begin{equation}
\mathrm{IFR}\approx 0.01\mbox{e}^{-7.529+0.121*a}\,
\sim \mbox{e}^{a/a_0}
\qquad\quad
a_0 = 8.26
\label{IFR}
\end{equation}
where $a\in[0,100]$ is the age cohort. The infection
fatality rate $\mathrm{IFR}\in[0,1]$ is very high
for $a=100$, namely $\mathrm{IFR}(100)=0.93$.
The constant $a_0$ denotes the half life age difference
in terms of mortality. To be more precise, for an age
difference of $8.26$ years the risk increases by a
factor of $e=2.78$. The risk doubles for an
age difference of 5.7 years.\footnote{Statistics
from the CDC of the US show that the over 85 years
old have a Covid-19 mortality rate 7,900 times higher
than that of the age group 5-17. This translates
also into a doubling age difference of about 8
years (= (85-11)/ln(7900)\,\cite{CDC_age_risk}.}
Limiting the strain on health systems has been another constant concern of policy makers. The risks of requiring hospitalisation is also age specific. Statistics
from the US CDC show that people aged 85 years
and older face a risk of hospitalisation from
Covid-19 infections about 95 times higher
than that of the age group 5-17. This translates
also into a doubling age difference of about 16
years (= (85-11)/ln(95), which we denote as $a_{0h}$ about double of that of the fatality risk.
In the following we will concentrate on decision rules designed to limit fatalities, but we will also refer to consequence of this somewhat different age dependency when the aim is to limit the pressure on the health system;
\subsection{Age pyramid}
We present the age pyramid for a range of selected countries
in Fig.~\ref{fig_ageDistribution},
where the age cohorts are given as a share of the
entire population. One notices that the age pyramid
closes generically quadratically at the very top. For
EU countries,he relevant range for the quadratic
dependency is however restricted, applying only to
ages around 85 and above.
Most Covid-19 vaccination campaigns started by
following top-down strategies to varying degrees
\cite{castro2021prioritizing}. For the time being we concentrate on EU countries, for which
the respective age pyramids can be approximated
linearly, as shown in Fig.~\ref{fig_ageDistribution}.
This approximation is intended as an overall fit
to ages 60 and above. The case of countries with
younger populations, like India and Nigeria, will
be treated later. \cite{ritchie2019age}
In the following we set the maximum age to zero,
counting down from 100. The actual age is then
$100-x$. The age density, denoted by $\rho(x)$,
varies by cohorts. Countries with population
pyramids closing linearly at the top are
described by
\begin{equation}
\rho(x) = px,
\qquad\quad
\int_0^{\delta\!A} \rho(x)\,dx = v,
\qquad\quad
v = \frac{p(\delta\!A)^2}{2}\,,
\label{rho_x}
\end{equation}
where $v$ is the number of people (relative to
the total population) vaccinated top-down to
an age difference $\delta\!A$.
Fitting to aggregate European demographic data yields
$p\approx0.00033$, as shown in
Fig.~\ref{fig_ageDistribution}, with little
difference across the major EU member countries.
This value for $p$ refers to the case that
$\rho(x)$ is measured relative to the total
population, in terms of percentiles of cohorts by year. For example, with this value of $\rho(x)$ the share of the 70-year-olds in the total population is equal to about 0.01 $=0.00033*30$.
The linear approximation holds well down to
60 years. Below this age, the size of the cohort
no longer increases (and even falls in some countries,
like Italy). Here we concentrate on the age cohorts
from 60 years up, which are the ones subject to the
highest mortality risk, constituting the largest
proportion of the overall loss of life. The case
of Germany illustrates this proposition. Taking
into account the combined effect of (\ref{IFR})
and the age distribution, as presented in
Fig.~\ref{fig_ageDistribution}, one finds that
about 1.5 million people above the age of sixty die in
the hypothetical scenario that the entire population
would be eventually infected with
SARS-CoV-2. By contrast, the
fatality count would include only 75 thousand
below sixty, i.e. by a factor of twenty. We thus feel
justified concentrating our analysis on the age
cohorts above 60, for which the population pyramid
is approximately linear.
People over 60 years of age account
for about 26 percent of the total population of
the EU, with their shares ranging from 20 percent
in the case of Ireland to 29 percent in the case
of Italy. This implies that vaccinating about one
fourth of the population will avoid 95
percent of fatalities ($19/20)$.
This calculation based solely on age obviously
represents an approximation.
Due to vaccine hesitancy, non-responders
and people with contraindications to
the vaccine, the uptake among the
elderly could be less than 100 percent. But
these factors are also present among all age
groups, reducing thus the overall benefit from
a vaccination campaign, but not necessarily
the advantage of age-sensitive vaccination.
Vaccine hesitancy is in particular likely to be
lower among the elderly, implying that the
share of the benefits from offering vaccination
to the elderly first might be even higher than
the 95\% suggested on demographic
considerations alone. A factor suggesting
otherwise may however be `long Covid'
\cite{altmann2021decoding}.
There is also evidence that immunity wanes more
quickly at higher ages \cite{hansen2021assessment},
implying that the re-infection risk is higher
for the elderly. This effect, however, plays out on a time scale beyond that of most current vaccination campaigns.
\begin{figure}[!t]
\centering
\includegraphics[width=0.95\columnwidth]{fig_dailyVaccinations.pdf}
\caption{{\bf Growth of daily vaccination rates.}
Per million daily vaccinations in Italy, France, Germany
and the United States (USA).
The growth is roughly linear, modulo in part substantial
fluctuations. Data smoothed over seven days, from \cite{OWID_covid19}.
}
\label{fig_dailyVaccinations}
\end{figure}
\section{Flattening the health cost curve}
Putting the two basic elements
-- the exponential age-dependency of the case fatality rate,
the linear functionality of population pyramid, --
together, we proceed to calculate the 'sweet spot' at
which the vaccination campaign can stabilise health costs.
We concentrate on the \emph{growth} of fatalities
because this is the key concern for policy makers.
But we also show that our approach can easily
be used to address the generic concern 'flatten
the curve', i.e.\
to prevent an explosive increase in hospitalizations
which could overwhelm health systems.
Throughout this section we take the growth of infections as given. In section \ref{Vaccination hesitancy} we discuss how the results are modified by the feedback between vaccination and infections.
\subsection{Vaccination and infection fatality rate}
We assume that full vaccination provides
a high level of protection against severe
illness and death, as confirmed not only by
trial data \cite{mahase2020covid}, but also
by real world
application \cite{aran2021estimating,dagan2021bnt162b2}.
For our model we furthermore assume that vaccination
is allocated strictly by age, starting with
the oldest. In practice the situation is
more complicated. Firstly, because a substantial
fraction of the available vaccine is reserved
for potential spreaders in most countries
\cite{castro2021prioritizing}, independent of
their age. Secondly, one needs to distinguish
between people having received one or two shots.
Both effects could be incorporated into the
framework developed here. In order to clarify
the mechanisms at work, we study in the
following the idealized situation that
`vaccinated' implies full protection.
\begin{comment}
\color{blue}
NEEDED?
We define as elderly the cohorts for which the
linear approximation (\ref{rho_x}) is valid.
We first evaluate the risk $R_v\in[0,R_{\rm max}]$
to die from a Covid-19 infection for elderly
aged $100-\delta\!A$ and above. Given that the
IFR (\ref{IFR}) decreases exponentially with
decreasing age, as $R_{\rm max}\exp(-x/a_0)$,
we find
\begin{equation}
\frac{R_v}{R_{\rm max}}
= \frac{p}{v}\int_0^{\delta\!A} x\,{\rm e}^{-x/a_0}dx
=\frac{pa_0^2}{v}\left[1-
\left(1+\frac{\delta\!A}{a_0}\right){\rm e}^{-\delta\!A/a_0}\right]\,,
\label{R_v}
\end{equation}
where $R_{\rm max}$ is the maximal mortality rate,
taken here as the value of the IFR at age 100 (which is close to 0.9 as mentioned above).
For the evaluation of the integral in (\ref{R_v})
we used
$a_0^2d[(x/a_0+1)\exp(-x/a_0)]/dx = -x\exp(-x/a_0)$.
We may express the right-hand side of (\ref{R_v})
in terms of the population density parameter
$p$, namely as
\begin{equation}
\frac{R_v(p)}{R_{\rm max}} =
\frac{pa_0^2}{v}\left[1-
\left(1+\frac{\delta\!A}{a_0}\right){\rm e}^{-\delta\!A/a_0}\right],
\qquad\quad
\delta\!A=\sqrt{2v/p}\,,
\label{R_v_p}
\end{equation}
where the percentage of top-down vaccinated $v$
is fixed when comparing two countries with
different $p$. It is evident that $R_v(p)$ is
a non-linear function of $p$, with the dominant
term being however a linear contribution.
\color{black}
\end{comment}
\section{Medical balancing condition}
In a first step we derive a medical balancing
condition that is generically valid, viz for
all types of population pyramids. In a second
step we will make use of the fact that population
pyramids may often be well approximated linearly
at the top. As before, we do not work with the nominal
age $a\in[0,100]$, but with the relative age $x$,
as measured top-down, $a=100-x$, denoting by
$\rho(x)$ the respective population density. The
fraction $v$ of people vaccinated top-down
to an age difference $\delta\!A$ is then
\begin{equation}
v = \int_0^{\delta\!A} \rho(x)dx,
\qquad\quad
\frac{dv}{dt} = \rho(\delta\!A)\frac{d(\delta\!A)}{dt}\,,
\label{v_delta_A_dt}
\end{equation}
where included also a differential relation
with regard to the vaccination speed $\dot v$.
Above expression reduces to
$v=p(\delta\!A)^2/2$ for a population pyramid
closing linearly at the top, viz
for $\rho(x)=px$.
A key variable for policy makers is the medical load, i.e. the number of patients requiring hospitalisation
and maybe even intensive care or who die.
We start by considering the mortality risk. The
medical load or cost for society at any given
state of the vaccination
campaign can then be expressed as as the number
of people who risk dying each time period:
\begin{equation}
C^{\rm med} = \left[\int_{\delta\!A}^\infty
\rho(x)\,\mbox{e}^{-x/a_0}dx\right]I(t)\,,
\label{C_med_general}
\end{equation}
where it is was assumed that fully vaccinated people
will not get seriously ill anymore.
Daily Covid-19 fatalities do not increase when
$C^{\rm med}$ remains stable, viz when the
balancing condition $dC^{\rm med}/dt=0$ holds.
Setting the time derivative of the right-hand
side of (\ref{C_med_general}) to zero leads
to the condition that:
\begin{equation}
\frac{d\delta\!A}{dt}\left[
\rho(\delta\!A)\,\mbox{e}^{-\delta\!A/a_0}
\right] = \left[\int_{\delta\!A}^\infty
\rho(x)\,\mbox{e}^{-x/a_0}dx\right]
\frac{\dot I}{I}\,.
\label{balancing_condition_general_00}
\end{equation}
With the expression for $\dot v$ given in
(\ref{v_delta_A_dt}), the left-hand side of
(\ref{balancing_condition_general_00})
reduces to $\dot{v}\exp(-\delta\!A/a_0)$.
We then have
\begin{equation}
\dot{v} = \mbox{e}^{\delta\!A/a_0}
\left[\int_{\delta\!A}^\infty
\rho(x)\,\mbox{e}^{-x/a_0}dx\right]
\frac{\dot I}{I}\,.
\label{balancing_condition_general}
\end{equation}
This expression is valid for all
population densities $\rho(x)$, in
particular also for the tabulated
population density of a given country.
\subsection{Linear population pyramid}
Evaluating the integral in
(\ref{balancing_condition_general})
for the linear case, when $\rho(x)=px$,
one has
\begin{eqnarray}
\nonumber
p\int_{\delta\!A}^\infty x\,\mbox{e}^{-x/a_0}dx
&=& -(pa_0^2) \left[1+\frac{x}{a_0}\right]
\mbox{e}^{-x/a_0}\Big|_{\delta\!A}^\infty
\\ &=& (pa_0^2) \left[1+\frac{\delta\!A}{a_0}\right]
\mbox{e}^{-\delta\!A/a_0}\,.
\label{integral_linear}
\end{eqnarray}
With
\begin{equation}
v_0 = \frac{a_0^2p}{2},
\qquad\quad
v = \frac{p\delta\!A^2}{2},
\qquad\quad
\frac{\delta\!A}{a_0} = \sqrt{\frac{v}{v_0}}
\label{relations}
\end{equation}
one then finds that
(\ref{balancing_condition_general})
reduces to
\begin{equation}
\frac{\dot{v}}{v_0} = 2
\left(1+\sqrt{\frac{v}{v_0}}\,\right)\frac{\dot{I}}{I}\,,
\label{balancing_linear}
\end{equation}
where $\dot{I}/I$ is the (relative) increase
of the incidence.
\section{Consequences of the balancing condition}
\subsection{The sweet spot for an old country}
When the population pyramid is linear at
the top (as it is typically for ageing societies),
the balancing condition (\ref{balancing_linear})
leads to a simple rule of thumb, given that $v_0$
is equal to about 0.01, or one percent:
\begin{quote}
{\em``For every proportional increase $\dot I/I$
of the incidence, one needs to vaccinate an
additional percentage of \emph{at least}
twice that amount in order to outrun the virus.''}
\end{quote}
This lower bound holds for $v\to0$, becoming
larger when vaccination progresses.
The balancing condition thus provides a simple
decision rule based on observable data (infections
and vaccinations). Once this 'sweet spot' has been
reached, NPIs can be gradually loosened without risking
an increase in fatalities.
\subsection{The sweet spot: hospitalizations}
The balancing conditions derived above describe
the point at which the number of fatalities
(per unit of time) stops increasing. One key
parameter used to derive these results was $a_0$,
which denotes the half live age difference in
terms of mortality. For the risk of hospitalization
this parameter, $a_{0h}$, about twice as high,
but the functional form remains the same.
For a linear age pyramid this implies that the
evolution of hospitalisations can be described
by the same balancing condition,
\begin{equation}
\frac{\dot{v}}{v_{0h}} = 2
\left(1+\sqrt{\frac{v}{v_{0h}}}\,\right)\frac{\dot{I}}{I}\,,
\label{balancing_linear_h}
\end{equation}
where the subscript $_h$ indicates that the value
is calculated with $a_{0h}$ instead of $a_0$.
This changes two of the three relationships
in equation (\ref{relations}):
\begin{equation}
v_{0h} = \frac{a_{0h}^2p}{2},
\qquad\quad
v = \frac{p\delta\!A^2}{2},
\qquad\quad
\frac{\delta\!A}{a_{0h}} = \sqrt{\frac{v}{v_{0h}}}
\label{relations_h}
\end{equation}
The (relative) increase of the incidence,
$\dot{I}/I$ and $v$ are not affected.
Comparing (\ref{relations_h}) and (\ref{relations})
shows that $v_{0h}/v_0$ is equal to $a_{0h}/a_0$.
It was documented above that the mean doubling age
for hospitalizations is about two times larger than
that for deaths (16 instead of 8 years). This
implies that $v_{0h}$ is about four times larger
than $v_0$. It follows that at low vaccination
rates ($v$ small) the balancing condition is four
times more stringent if the aim is to keep
hospitalizations from increasing than if the
aim is to keep fatalities in check, i.e.\
the decision rule would become:
\begin{quote}
{\em``For every proportional increase $\dot I/I$
of the incidence, one needs to vaccinate an
additional percentage of \emph{at least}
eight times that amount in order to keep the
pressure on hospitals constant.''}
\end{quote}
A policy that takes into account the need to limit
the pressure on health systems would thus loosen
NPIs much later than a policy which concentrates only
on fatalities. The difference is a factor of four at
low vaccination rates, but it diminishes as a higher
proportion is vaccinated (as $v$ increases) because the
higher value of $a_{0h}$ enters the denominator of
the second term in the brackets on the right hand side
of equation (\ref{balancing_linear_h}).
\subsection{The sweet spot for a young country}
As documented below, younger countries have a
top of the population pyramid which closes
quadratically. In this case the size of the
cohorts counting from age 100 down can be described by:
\begin{equation}
\rho(x)=qx^2,
\quad\quad
v = \frac{q\delta\!A^3}{3},
\quad\quad
\frac{\delta\!A}{a_0} = \left(\frac{3v}{a_0^3q}\right)^{1/3},
\quad\quad
v_1 \equiv \frac{a_0^3q}{3}
\label{rho_xx}
\end{equation}
and
\begin{equation}
q\int_{\delta\!A}^\infty x^2\,\mbox{e}^{-x/a_0}dx
= -(qa_0^3)
\left[2+\frac{2x}{a_0}+\frac{x^2}{a_0^2}\right]
\mbox{e}^{-x/a_0}\Big|_{\delta\!A}^\infty,
\label{integral_quadratic}
\end{equation}
one obtains the balancing condition
\begin{equation}
\frac{\dot v}{v_1} = 3\left[
2 + 2\left(\frac{v}{v_1}\right)^{1/3}
+ \left(\frac{v}{v_1}\right)^{2/3}
\right]\frac{\dot I}{I}
\label{balancing_quadratic}
\end{equation}
A fit to the population pyramid of Nigeria
yields $q=4\cdot10^{-6}$, viz $v_1=a_0^3q/3=0.0002$,
see (\ref{IFR}) and Fig.~\ref{fig_ageDistribution}.
The reference fraction of vaccinated is hence
exceedingly small, $v_1\sim 0.02\%$
A generalization
to the case of a quadratic population pyramid,
as given by (\ref{balancing_quadratic}), yields
a factor $6v_1/v_0\approx 0.12$ in the limit
$v\to0$, instead of two (given that $v_0$ corresponds
to about one percent). At the start, vaccination
campaigns have hence strong control capabilities in
countries with young populations, viz when the
population pyramid is quadratic at the top. This
advantage decreases rapidly, however, with
the progress of vaccination. To be concrete, dividing
equation (\ref{balancing_linear}) by
(\ref{balancing_quadratic}) one obtains a factor
of $1.7$ at $v=0.1$, when ten percent of the
population has been vaccinated top-down.
\subsection{Connection with the evolution of the pandemic}
The rate of increase in the number of infected will
be affected by the fraction of the population already
vaccinated.
The evidence regarding the impact of vaccines
on the spread of infections is less clear than the impact of vaccination on the risk of death or a severe course
\cite{lipsitch2021interpreting}. Some results point to a reduction in transmission of 40 percent \cite{harris2021effect} whereas others report a much higher impact among the vulnerable population in long-term care facilities \cite{de2021high}.
Any impact of vaccination on infectiousness
would not change the balancing condition
(\ref{balancing_linear}), which applies in
general. Moreover, with vaccination reducing
infectiousness, whatever the size of the impact, the growth of the disease spread slows down, ceteris
paribus, (to a lower value for $\dot I/I$),
making it easier to reach the point where the curve `flattens'.
The relative increase $\dot I/I$ appearing
on the right-hand side of (\ref{balancing_linear})
could be evaluated using a dynamic epidemiological
model. For the SIR model \cite{gros2015complex},
the simplest case, we have
$\dot I/I = gS-\lambda$, where $g$ and $\lambda$ are
the reproduction and the recovery rates, respectively.
The reservoir of susceptibles $S$ decreases with the
fraction of the population that is immunized, either
because of having been infected, or due to being
vaccinated. We leave this approach for further
studies, concentrating here on data-driven considerations.
The dynamics of the growth of a pandemic is
studied in a large body of literature. One of
the first to employ the
canonical model was \cite{ferguson2020imperial}. \cite{agosto2021monitoring} employ statistical
models to predict the contagion curve and the associated
reproduction rate using a Poisson process. See also
\cite{harvey2020time} for time series models and
\cite{shinde2020forecasting} for a survey.
We thus conclude that the position of the
'sweet spot' in terms of observed infections is not
changed if one takes into account the impact of
vaccinations on the spread of the disease. With
continuing vaccinations the sweet spot becomes
a turning point because once it has been
reached, vaccination rates could plateau,
but the condition for deaths to fall would
remain fulfilled, as additional vaccinations
reduce the growth of infection, ${\dot I}/{I}$,
which is the
determining factor on the right-hand side of
(\ref{balancing_linear}).
\section{Vaccination hesitancy}
\label{Vaccination hesitancy}
In deriving (\ref{balancing_condition_general}),
we assumed full vaccination of everybody aged
above $a=100-\delta\!A$. In practice, there will
be a certain hesitancy $h(x)\in[0,1]$, meaning
that a fraction $h(x)$ remains non-vaccinated,
even after vaccines have become available to the
age cohort $x$. The age specific indicator $h=h(x)$
is used here as a general term, including people
that are not willing to be vaccinated, together
with the fraction of the population that cannot
be vaccinated due to medical contra-indications.
People not developing an immune response despite
being fully vaccinated are also subsumed under
$h(x)$.
In Europe, the fraction $h(x)$ of not vaccinated
starts close to zero in most countries for the
very old, remaining in general below 20 per cent
down to age 60.\footnote{\href{https://www.ecdc.europa.eu/en/publications-data/data-covid-19-vaccination-eu-eea}{https://www.ecdc.europa.eu/en/publications-data/data-covid-19-vaccination-eu-eea}}
It is straightforward to generalize the derivation of
the medical balancing condition to the case $h(x)>0$.
For this the term:
\begin{equation}
c_0I(t)\int_0^{\delta\!A} h(x)\rho(x)\mbox{e}^{-x/a_0}dx
\label{C_Med_term_hesitancy}
\end{equation}
needs to be added to the medical cost $C^{\rm med}$,
see (\ref{C_med_general}). One finds
\begin{equation}
\dot{v}\big(1-h(\delta\!A)\big) = \mbox{e}^{\delta\!A/a_0}
\left[
\int_{\delta\!A}^\infty \rho(x)\,\mbox{e}^{-x/a_0}dx +
\int_0^{\delta\!A} h(x)\rho(x)\,\mbox{e}^{-x/a_0}dx
\right]
\frac{\dot I}{I}\,,
\label{balancing_condition_hesitancy}
\end{equation}
which reduces to (\ref{balancing_condition_general})
when $h(x)\to0$. Clearly, it becomes more difficult
to retain medical balancing when vaccine hesitancy
is large, which is however not the case for the elderly.
For the case of age-independent hesitancy, $h(x)\equiv h$,
one can evaluate both the first and the second
integral on the right-hand side of
(\ref{balancing_condition_hesitancy}). The latter is
\begin{eqnarray}
\nonumber
hp\int_0^{\delta\!A} x\,\mbox{e}^{-x/a_0}dx
&=& -h(pa_0^2) \left[1+\frac{x}{a_0}\right]
\mbox{e}^{-x/a_0}\Big|_0^{\delta\!A}
\\ &=& - h(pa_0^2) \left[1+\frac{\delta\!A}{a_0}\right]
\mbox{e}^{-\delta\!A/a_0} + h(p a_0)^2\,.
\label{integral_linear_hesitancy}
\end{eqnarray}
The first term on the right-hand side is $(-h)$
times the obtained for the case without hesitancy.
All together we hence have
\begin{equation}
\frac{\dot{v}}{v_0} =
2 \left(1+\sqrt{\frac{v}{v_0}}\,\right)\frac{\dot{I}}{I}
+
2h\frac{\mbox{e}^{\delta\!A/a_0}}{1-h}\,\frac{\dot{I}}{I}
= 2\left[1+\sqrt{\frac{v}{v_0}}+
h\frac{\mbox{e}^{\delta\!A/a_0}}{1-h}\, \right]\frac{\dot{I}}{I}\,,
\label{balancing_linear_hesitancy}
\end{equation}
which generalizes (\ref{balancing_linear}).
Vaccine hesitancy thus introduces an explicit
age dependency, given that cohorts of age $a$
are defined via $a=100-\delta\!A$. The
basic message of the balancing without vaccine
hesitancy thus remains, namely that it takes a
vaccination campaign that is at least two times
quicker than the percentage spread of the disease.
\section{Evolution of vaccination campaigns}
It is not possible to vaccinate the entire
population instantly, because vaccines have
to be first mass-produced and then distributed.
This is illustrated in Fig.~\ref{fig_dailyVaccinations},
where the daily vaccination rates are shown for
a range of selected countries. Daily rates may vary,
in particular for smaller countries, when
larger batches are delivered from abroad.
But for larger entities,
like the EU or the US, the trend is linear.
Overall, vaccination rates can be expected to
track deliveries, with eventual organizational
problems leading only to temporary delays.
Over the course of several months, the
daily vaccination rates shown in
Fig.~\ref{fig_dailyVaccinations} rise
roughly linearly during the initial states
of the campaign,
in line with steadily increasing production
capacities.\footnote{This linear ramp-up has
been predicted \cite{gros2021covideconomicsrace}.
The reason is that Covid-19 vaccine were
ordered ahead of their approval in large batches.
But ramping up production capacities implies
substantial adjustment costs. Minimizing these adjustment
costs, subject to fulfilling the order within a
contracted time period, leads to a linear ramping-up of
production capacities \cite{gros2021incentives}.}
Given these considerations, and the data presented
in Fig.~\ref{fig_dailyVaccinations}, we assume that
the number of daily jabs, viz the vaccination
rate, increases linearly. The fraction of
the population $v$ vaccinated top down
increases then with the square of time $t$,
\begin{equation}
v=v(t) = \frac{1}{2}\left(\frac{t}{t_0}\right)^2
= \frac{p}{2}(\delta\!A)^2,
\qquad\quad
\delta\!A =\frac{t}{t_0\sqrt{p}}\equiv a_0g_v t\,.
\label{v_t_t1}
\end{equation}
The parameter $t_0$ denotes the time
needed to vaccinate one half of the population.
Given a linear increase of cohort sizes with age, one finds
that the age of the youngest cohort that can be
fully vaccinated, denoted $\delta\!A$,
falls linearly over time, see (\ref{rho_x}).
The factor $a_0$ in the last definition is
the characteristic age determining the
exponential functionality of the IFR, as
defined by (\ref{IFR}).
An order of magnitude estimate for the length
of vaccination campaigns, $t_0$, can be evaluated
from available data. For example, in Israel
it took about 10 weeks (from the beginning of
January of 2021 to mid-March 2021) to fully vaccinate
half the population, resulting in an estimate
of $t_0 = 10$ (weeks). In the EU, only about 5
percent of the population was fully vaccinated
over the same period corresponding to an estimate of
$t_0 =10\sqrt{10}\approx32$ (weeks).\footnote{The
time needed to vaccinate the entire population, i.e.\
to the point $v=1$, is equal $\sqrt{2} t_0$. The parameter
$t_0$ thus does not denote the full length of the
vaccination campaign, but the time needed to vaccinate
50\% of the population. At that point more than 99\%
of the fatalities can be avoided and NPIs can be lifted.
$t_0$ provides thus a good parametrization of the
effective length of a vaccination campaign.}
The balancing condition (\ref{balancing_linear})
for the case of linear population density
can be used to write the sweet spot at which
medical costs are kept from rising, in terms
of the two structural parameters $a_0$ and $p$,
\begin{equation}
\frac{t}{t_0} =
\left(a_0^2p+\frac{t}{t_0}a_0\sqrt{p}\,\right)t_0\frac{\dot{I}}{I},
\qquad\quad
v_0 = \frac{a_0^2p}{2}\,,
\label{delta_C_p_a0}
\end{equation}
see (\ref{IFR}) and (\ref{rho_x}).\footnote{We
concentrate on the general analytical solution in
order to avoid having to make too many specific
assumptions about the way the pandemic spreads.
For a more detailed, structural approach
see \cite{shen2021projected}.}
As above, one could substitute the parameter $a_0$ with
$a_{0h}$ if the aim is to keep hospitalisations under
control. Given that equation (\ref{delta_C_p_a0})
contains the square of the parameter it follows that it
will take about 4 times as long to reach the sweet spot
in terms of hospitalizations than in terms of fatalities.
\section{Discussion}
A key aim for policy makers grappling with a
continuing outbreak, even when an increasing proportion
of the population is being vaccinated, is to
`flatten the curve', i.e.\
to keep hospitalizations and fatalities from
rising exponentially \cite{moore2021vaccination}.
We identified a balancing condition, or 'sweet spot',
at which the health costs remain constant as the
most vulnerable are being vaccinated first.
The balancing condition can be interpreted as indicating
the path of relaxing NPIs along which infections can
still increase, but fatalities or death remain under control.
Our aim is not to provide detailed
epidemiological modeling and simulations.
Instead, we have shown that three key factors can
be combined into a simple formula that
determines the impact of vaccination campaigns
with regard to the time evolution of medical
costs. First, the mortality risk from a Covid-19
infection (and that of hospitalisation) increases
exponentially with age. Second, the sizes
of age cohorts decrease from the top of the
population pyramid.
Third, vaccination proceeds at an increasing speed.
We find considerable differences across
two dimensions: the form of the age pyramid
and the measure of health costs: hospitalisations
or fatalities.
\begin{itemize}
\item
Age pyramid: Older countries need to vaccinate
more quickly than younger ones. The difference
between a typical European country and a high
fertility country in Sub-Saharan Africa can be
a factor of 6.
\item
Deaths or hospitalisations: For an old country,
aiming at keeping hospitalisations constant entails
a requirement of a vaccination rate about 4 times
larger than if the aim is to keep fatalities
constant. For a young country the difference
is even larger.
\end{itemize}
Our discussion has focused on the case of Europe, but the formula we derive holds generally and can take into account the wide differences one observes in the speed of vaccination campaigns and the age structure of the population. Moreover, our approach is general enough to accommodate the emergence of different strains of the virus which might increase contagion and/or the risk of severe courses. For example, the impact of the diffusion of the so-called delta variant, whose diffusion has recently been modelled \cite{gilmour2021preliminary}, would show up in a higher rate of growth infections, which would need a correspondingly higher speed of vaccination to offset its higher contagiousness.
\subsection{Limitations}
In our framework, `vaccinated' implies full immunity,
which is attained for most Covid-19 vaccines only
after the second jab. A reduced levels of immunity,
like 95\%, is equivalent to an equivalent degree of
vaccine hesitancy, which we also discuss.
Note that only a proportion of the jabs, and not all,
is administered following a strict age criterion
\cite{cook2021impact}.
Vaccine hesitancy and other factors, such as
waning immunity with age, can reduce the overall
effect of vaccination campaigns. Moreover, there
are other, less age specific costs of the disease,
like `long Covid' \cite{sudre2021attributes}.
Incorporating these factors would refine the
model, at the same time making it necessary
to estimate an increased number of parameters.
See in this regard, e.g.,
\cite{RKIdritteWelle} and \cite{bosetti2021race}
for recent contributions using state-of-the-art
epidemiological models.
We have concentrated on the fact that vaccination
basically eliminates the risk of death \cite{mateo2021risk}.
Vaccination reduces however also the spread of
the virus \cite{rossman2021patterns}.
This provides an additional element which
increases the importance of vaccination speed.
However, this element becomes significant primarily
in later stages of a vaccination campaign,
beyond the point when the vulnerable groups
have been vaccinated.
\section*{Acknowledgements}
This research has been supported by the European
Union's Horizon 2020 research and
innovation program, PERISCOPE:
Pan European Response to the Impacts
of Covid-19 and future Pandemics and
Epidemics, under grant agreement
No.\ 101016233, H2020-SC1-PHE CORONAVIRUS-2020-2-RTD.
\section*{Data Availability}
The electronic supporting material
contains the data analyzed.
\bibliographystyle{unsrt}
|
\section{A comparative study of the two control approaches}
\label{sec:comparison}
\subsection{The MPC Controller}
The MPC paradigm combines three key components.
The first is a dynamic model of the ego vehicle,
allowing the MPC Controller to predict the vehicle states in a given horizon for changing inputs.
The vehicle state is denoted by the vector $x = [p, v, a]^{T}$,
where $p$ is the vehicle position, $v$ is the speed and $a$ is the acceleration.
Similarly, the state of vehicle ahead is $x_a = [p_a, v_a, a_a]^{T}$.
The relative distance between the two vehicles is then $d = p_a - p$.
We further require that the acceleration is buffered as follows:
\begin{equation}
\dot{a} = \frac{u - a}{\tau}
\end{equation}
where the control stimulus $u$ is the desired acceleration,
and $\tau$ is the time constant of the actuator lag that captures the inertial characteristics of the vehicle actuator.
The vehicle dynamics model is described by the following equation.
\begin{equation}
\dot{x} = A_\tau \cdot x + B_\tau \cdot u
\end{equation}
where
\begin{equation}
A_\tau = \begin{bmatrix} 0 & 1 & 0 \\ 0 & 0 & 1 \\ 0 & 0 & -\frac{1}{\tau} \end{bmatrix},
B_\tau = \begin{bmatrix} 0 \\ 0 \\ \frac{1}{\tau} \end{bmatrix}
\end{equation}
In order to enhance the stability of the system under the constant time sampling control method,
we discretize the vehicle dynamics model \cite{dorf2011modern}.
If $\Delta t$ is the discretization pace,
the discrete longitudinal dynamics model of the ego vehicle at time instant $t_k$ is given as follows:
\begin{equation}
x(t_{k+1}) = A_{d} \cdot x(t_k) + B_{d} \cdot u (t_k)
\end{equation}
where
\begin{equation}
\begin{aligned}
A_{d} &= e^{A_{\tau} \Delta t} = \begin{bmatrix}
1 & \Delta t & \tau^2 \left(e^{-\frac{\Delta t}{\tau}} - 1 \right) + \Delta t \tau \\
0 & 1 & \tau \left(1 - e^{-\frac{\Delta t}{\tau}}\right) \\
0 & 0 & e^{-\frac{\Delta t}{\tau}} \\
\end{bmatrix} \\
B_{d} &= \int_0^{\Delta t} e^{A_{\tau} t} dt \cdot B_{\tau} = \begin{bmatrix}
\tau^2 \left(1 - e^{-\frac{\Delta t}{\tau}}\right) + \frac{\Delta t^2}{2} - \Delta t \tau \\
\tau \left(e^{-\frac{\Delta t}{\tau}} - 1 \right) + \Delta t\\
1 - e^{-\frac{\Delta t}{\tau}}
\end{bmatrix}
\end{aligned}
\end{equation}
In addition to the state of the ego vehicle,
the MPC Controller also estimates the position of the vehicle ahead in order to compute predictions.
We assume that in each MPC prediction horizon, the vehicle ahead decelerates at a constant rate $a_{a}$.
Thus, the position of the vehicle ahead before stopping can be estimated as follows.
\begin{equation}
p_{a}(t_k) = p_{a}(t_0) + v_{a}(t_0) \cdot (t_k - t_0) + 1/2 \cdot a_{a} \cdot (t_k - t_0)^2
\end{equation}
The second key component is a cost function,
which describes the expected behavior of the ego vehicle, in order to minimize the relative distance.
Optimization consists in finding the best possible inputs that minimize the cost function.
The cost function for time horizon $h$ is modeled as a standard quadratic function
and the optimization problem is formulated as follows:
\begin{equation}
\label{optimization-equation}
argmin(u^{*}(\cdot), \sum_{k=1}^{h} (x_{opt}^{T} Q x_{opt} + r u^2) )
\end{equation}
%
where $Q$ is the weighting matrix for the state vector and $r$ is the weight for the control stimulus.
The new state variable $x_{opt}$ is used to get the relative distance as close as possible to the constant $d_{c}$
and the speed of the ego vehicle to $v_a$.
It is defined by
\begin{equation}
\label{x-opt}
x_{opt} = x_{a} - x - [d_{c}, 0, 0 ]^T
\end{equation}
The weighting matrix is a diagonal matrix $Q=diag[q_p , q_v , q_a]^T$ ,
where $q_p$, $q_v$ and $q_a$ are weighting parameters for vehicle position, speed and acceleration respectively.
By adjusting the values of these weighting parameters, we can configure preference of the MPC control tendency.
For instance, by enlarging the value of $q_p$ we force MPC to drive the ego vehicle closer to the vehicle ahead,
so that to reduce the relative distance.
Additionally while performing the optimization,
the MPC Controller enforces the following constraints on the minimum or maximum values of speed and acceleration of the vehicle:
\begin{equation}
\begin{cases}
v_{min} \leq v \leq v_{max} \\
u_{min} \leq u \leq u_{max}
\end{cases}
\end{equation}
where $u_{min}$, $u_{max}$ (and $v_{min}$, $v_{max}$) are user-specified parameters for control stimulus and speed, respectively.
The third component of the MPC Controller is the optimization algorithm
for solving this linear quadratic programming problem.
For this purpose, we use the open source Python library, cvxopt \cite{andersen2011interior}.
\begin{figure} [H]
\centering
\includegraphics[width=0.6\hsize]{figs/mpc.png}
\caption{Architecture of the MPC controller}
\label{mpc}
\end{figure}
Fig.\ref{mpc} shows the architecture of the MPC Controller.
The controller executes an iterative process optimizing the predictions of vehicle states while manipulating inputs for a given horizon.
The predictions are based on the specified kinematic model of the vehicle.
For each control cycle at time $t_k$,
the controller takes as input the current states of the ego vehicle and of the vehicle ahead,
and computes the future states of the ego vehicle to predict the optimal control stimuli $u^{*}$
minimizing the cost function in the interval [$t_k$, $t_{k + h}$], where $h$ is the prediction horizon.
The MPC Controller chooses the first element in the sequence as the control stimulus for the ego vehicle,
and repeats the cycle at time $t_{k + 1}$.
A key advantage of MPC policy is flexibility in achieving complex objectives and implementing multiple constraints
when performing optimizations.
\subsection{The Safe Controller}
\label{safe-avoidance}
In a recent work \cite{wang2020safe},
we propose a correct-by-design safe and efficient controller for autonomous vehicles.
The controller minimizes the distance between the ego vehicle and the vehicle ahead while preserving safety
for both nominal and out-of-nominal conditions.
For nominal conditions, the Safe Controller is based on the relative speed $v_{e} - v_{a}$ between the ego vehicle and the vehicle ahead.
It computes the target speed $v_{safe}$ for moderate nominal acceleration and deceleration rates to enhance passenger comfort.
For out-of-nominal conditions, the Safe Controller computes a target speed $v_{max}$
taking into account only the speed of the ego vehicle and the needed braking distance.
The braking distance is computed for a maximal deceleration rate that is much larger than the nominal deceleration to cope with dangerous situations,
e.g., sudden stops of the vehicle ahead caused by accidents.
The controller always keeps the speed $v_{safe} \leq v_{max}$ to make sure that in all circumstances safety is preserved.
We briefly review the general design principle that we specialize for nominal and out-of-nominal conditions.
The safe control policy relies on the following three functions.
\begin{itemize}
\item The function $d(t)$ gives the relative distance at time $t$ between the ego vehicle and the vehicle ahead,
which is either stopped or moving in the same direction.
\item The braking function $B(v, v')$ gives the distance travelled by the ego vehicle,
when braking from initial speed $v$ to target speed $v'$.
When the target speed is $v' = 0$ (i.e, the ego vehicle brakes to a stop),
this function is abbreviated as $B(v)$ for simplicity.
\item The accelerating function $A(v, v')$ gives the distance travelled by the ego vehicle,
when accelerating from initial speed $v$ to target speed $v'$.
\end{itemize}
We make no specific assumptions about the implementation of the accelerating and braking functions,
e.g., whether acceleration is constant or variable.
We simply require that the following properties hold:
\begin{itemize}
\item $B(v, v') = 0$ and $A(v, v') = 0$ if and only if $v = v'$.
\item Additivity property:
$B(v, v_1) + B(v_1, v_2) = B(v, v_2)$ and $A(v, v_1) + A(v_1, v_2) = A(v, v_2)$.
\item Strict monotonicity:
$B(v, v_1) < B(v, v_2) $ and $A(v, v_1) < A(v, v_2) $ if $v_1 < v_2$.
\end{itemize}
The basic idea for avoiding collision is to moderate the speed of the vehicle and
anticipate the changes of the relative distance so as to have enough space and time to adjust and brake.
For any time $t$, the vehicle only needs to keep track of the distance $d(t)$ and
check in real-time whether $d(t)$ is greater than the minimal safe braking distance $B(v_t)$ for the current speed $v_t$.
It starts braking as soon as $d(t)$ reaches the minimal safe braking distance.
In this way, it is guaranteed that if the obstacles ahead do not move in the opposite direction, no collision would happen.
We consider that the vehicle speed can change between a finite set of increasing levels $v_0, v_1, ..., v_n$,
where $n$ is a constant, $v_0 = 0$ and $v_n$ is the limit speed of the vehicle.
The triggering of acceleration and braking from one level to another is controlled according to the free distance ahead
and based on the bounds computed as follows, for each speed level $v_i, i\in [1, n]$,
\begin{itemize}
\item $B_{i} = B(v_i)$ is the minimal safe braking distance needed for the vehicle to fully stop from speed $v_i$;
\item $D_{i} = A(v_{i-1}, v_{i}) + B(v_i)$ is the minimal safe distance needed for the vehicle
to accelerate from speed $v_{i-1}$ to $v_i$ and then brake from $v_i$ to stop.
\end{itemize}
We show in \cite{wang2020safe} that the following function specifies the highest safe speed level $v$
as a function of the current speed $v_t$ of the vehicle and the distance $d$,
provided that their initial values $v_0$ and $d_0$ satisfy the condition $B(v_0) \leq d_0$.
\begin{equation*}
v = Control(d, v_t)
\end{equation*}
\begin{equation*}
v =
\begin{cases}
v_{i+1} & \text{when ~~ $v_t = v_i \wedge d = D_{i+1}$} \\
v_{i-1} & \text{when ~~ $v_t = v_i \wedge d = B_{i}$} \\
v_{i} & \text{when ~~ $v_t = v_i \wedge D_{i+1} > d > B_{i}$}
\end{cases}
\end{equation*}
Fig.\ref{ledder} illustrates the principle for $n=4$ speed levels.
As the value of $d$ increases, the speed of the vehicle climbs up levels.
Safety is preserved by construction.
The vehicle can accelerate to a higher level,
if it can safely and efficiently use the available distance by combining acceleration and decelaration;
in particular braking to a lower level if the distance reaches the bound for safe braking.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{figs/ledder.png}
\caption{Illustration of the collision avoidance principle for $n=4$}
\label{ledder}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{figs/scheduler.png}
\caption{Automaton modelling the collision avoidance principle}
\label{automaton}
\end{figure}
Fig.\ref{automaton} provides a scheme for the computation of $Control(d, v_t)$ in the form of an extended automaton.
The control states correspond to speed levels $v_0, ..., v_n$.
The transitions model instantaneous acceleration and braking steps
triggered by conditions involving the distance $d$ and the pre-computed bounds $B_{i}$ and $D_{i}$.
If the control location is $v_{i}$ and the distance is equal to the minimal safe acceleration distance (i.e., $d = D_{i+1}$),
then the automaton moves to location $v_{i+1}$ after the speed is increased to $v_{i+1}$.
If the distance reaches the minimal safe braking distance (i.e., $d = B_{i}$),
then the automaton moves to location $v_{i-1}$ after the speed is decreased to $v_{i-1}$.
Given that $B_{i} = B_{i-1} + B(v_i, v_{i-1})$,
after braking to $v_{i-1}$ there is still enough space for safe braking.
If none of the triggering conditions holds,
then the distance $d$ is such that $B_i < d < D_{i+1}$.
The automaton stays at location $v_i$ and the speed remains unchanged.
Details about the implementation of this control principle can be found in \cite{wang2020safe}.
In our context, the speed $v_{safe}$ is computed for nominal conditions
considering that $v_t$ = $v_e$ - $v_a$ and that the braking and acceleration functions are defined for moderate rates.
Hence, $v_{safe}$ = Control( $d$, $v_e$ - $v_a$).
For out-of-nominal conditions,
the maximal speed $v_{max}$ of the ego vehicle is $v_{max}$ = Max\{ $v ~ | ~ B_{max}(v_e) \leq d$ \},
where $B_{max}$ is the deceleration function for some maximal deceleration rate.
To make sure that collision is avoided in any case,
the controller compares the speeds $v_{safe}$ and $v_{max}$ and
when $v_{max}$ is reached, a command is issued for emergency braking.
Compared with the MPC Controller,
the Safe Controller can guarantee safety and efficiency
since at any time it chooses the speed minimizing the relative distance depending on simple criteria.
On the contrary, the MPC Controller applies more involved computation
trying to estimate the future states of both vehicles according to the kinematic model of the vehicle,
which often requires some intelligent algorithms (e.g., genetic computation).
Thus, it may require computationally expensive optimization techniques.
On the contrary, the Safe Controller is computationally cheap and
can be easily implemented in real time without additional costs.
\subsection{Evaluation of the two control approaches}
\subsubsection{Experimental setting and evaluation criteria}
\
\vspace{0.1cm}
\newline
We implement the MPC Controller and the Safe Controller in the Carla (version 0.9.8) simulator \cite{Dosovitskiy17}.
For the Safe Controller, the acceleration/deceleration in the nominal setting is taken $3 ~ m/s^2$
and the deceleration rate $12 ~ m/s^2$ in the out-of-nominal setting.
The speed levels are from the set $\{0, 4, 8, 12,16,20,24,28,32\} (m/s)$.
The control policy for the computation of $v_{safe}$ is based on the relative speed $v_e$ - $v_a$.
For the MPC Controller, the prediction horizon is set to 10 steps.
The constant $d_c$ in Equation \ref{x-opt} is $20$ m.
The weighting matrix $Q$ is set to $diag [50, 400, 1]^{T}$.
Time lag $\tau$ is 0.3, and control stimulus weight $r$ is set to 1.
To evaluate the efficiency of the two controllers,
we carry out a set of comprehensive experiments performed on a Windows 10 PC with AMD R5 3500 and NVIDIA GTX 1660 SUPER.
We consider both nominal and out-of-nominal scenarios.
\begin{enumerate}
\item In a nominal scenario, the speed of the vehicle ahead is described by the function
$v_{a}(t) = A \sin(\frac{2 \pi}{T} t) + v_{a, 0}$, where $v_{a, 0}$ is taken equal to $12 ~ m/s$.
In the experiments, we consider three different values of $A$, i.e., $\{6, 9, 12\} ~ m/s$,
and three different values of $T$, i.e., $ \{10, 20, 30\} ~ s$.
\item In an out-of-nominal scenario, the vehicle ahead brakes suddenly and stops.
We assume that the ego vehicle does not know when the sudden braking may occur.
\end{enumerate}
In all scenarios the ego vehicle and the vehicle ahead move in the same lane in the same direction.
The initial speed of the ego vehicle is set to $0 ~ m/s$.
The initial relative distance between the two vehicles is set to $10 ~ m$.
We check whether the MPC Controller violates safety for nominal and out-of-nominal scenarios.
For nominal scenarios,
we evaluate the efficiency of the two controllers
with respect to the following three criteria defined for simulation time $t_{sim}$.
\begin{enumerate}
\item Performance is measured as the ratio of average speed of the ego vehicle
with respect to that of the vehicle ahead, i.e.,
\begin{equation*}
\mathcal{M}_{p} = \frac{\int_0^{t_{sim}} v_{e}(t) dt}{\int_0^{t_{sim}} v_{a}(t) dt}
\end{equation*}
where $v_{e}$ is the speed of the ego vehicle and $v_{a}$ is the speed of the vehicle ahead.
\item Road occupancy is defined as the ratio of the space occupied by the vehicles over the total available space.
For our case with two vehicles,
we consider $1/d$ as a measure of the occupancy, where $d$ is their relative distance.
For a simulation in time interval $[0, t_{sim}]$, it is given by the formula:
\begin{equation*}
\mathcal{M}_{o} = \frac{1}{t_{sim}} \int_0^{t_{sim}} 1/d(t) dt
\end{equation*}
The higher the value $\mathcal{M}_{o}$, the higher the occupancy.
Note that the measure uses $d$, the distance between the two vehicles without taking into account any safety margin.
%
\item Comfort means that the variations of acceleration are close to its average value.
We consider that it is measured as the reciprocal of the acceleration variance, i.e.,
\begin{equation*}
\mathcal{M}_{c} = ( \frac{1}{t_{sim}}\int_0^{t_{sim}} (a(t) - \bar{a})^2 dt ) ^{-1}
\end{equation*}
where $a$ is the acceleration of the ego vehicle, and $\bar{a} = \frac{1}{t_{sim}} \int_0^{t_{sim}} a(t) dt$.
Note that the higher the value $\mathcal{M}_{c}$, the higher the comfort level.
\end{enumerate}
\subsubsection{Evaluation of the two controllers}
\
\vspace{0.2cm}
\newline
\textbf{Nominal scenarios}
Fig. \ref{velocity-plot} shows the speed of the ego vehicle for the two controllers
and the maximal safe speed $v_{max}$ for the nominal scenarios.
Note that the MPC Controller closely follows the speed of the vehicle ahead,
in particular when the period of the speed function increases, e.g., T = 30 s.
However, the speed of the MPC Controller cannot avoid unsafe situations as shown in Fig.\ref{velocity-plot}
when the orange line ($v_{mpc}$) crosses the red line ($v_{max}$).
On the contrary, the Safe Controller is less sensitive to speed changes of the vehicle ahead
and allows smaller speed variation due to safety constraints.
Despite these constraints, the performance measured as the average speed has not been sacrificed.
The upmost part of Table.\ref{metrics-10} provides performance metrics showing that in most cases,
the Safe Controller produces higher average speeds than the MPC Controller and thus maintains slightly higher ratios.
Nonetheless, the differences increase when the amplitude of the speed function becomes larger, e.g., A = 12.
Fig. \ref{distance-plot} provides the relative distance for the two controllers.
It shows that the amplitude variation for the MPC Controller is much smaller,
especially when the period and the amplitude of the speed function are larger.
This is because the MPC Controller favors speed tracking.
On the contrary, the Safe Controller maintains a smaller relative distance on average than the MPC Controller,
since its control policy focuses on distance minimization.
This observation is confirmed by the higher occupancy metrics for the Safe Controller provided in Table.\ref{metrics-10}.
The comfort metrics provided by Table. \ref{metrics-10} show that
comfort for the Safe Controller is much higher than for the MPC Controller
when the period and the amplitude of the speed function are small.
The reason is that the Safe Controller is less sensitive to the speed changes of the vehicle ahead
and avoids alternating changes of acceleration and deceleration.
Nonetheless, for larger periods, e.g. T = 30, the Safe Controller is less comfortable.
\begin{figure} [H]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/section2_v.eps}
\caption{Speed for the Safe Controller (in green), the MPC Controller (in orange)
and the maximal speed $v_{max}$ (in red) for nominal scenarios.
Note the unsafe situations where $v_{mpc} > v_{max}$ .}
\label{velocity-plot}
\end{figure}
\vspace{-1cm}
\begin{figure} [H]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/section2_d.eps}
\caption{Relative distance for the Safe Controller (in green) and the MPC Controller (in orange) for nominal scenarios.}
\label{distance-plot}
\end{figure}
\vspace{-1cm}
\begin{table} [H]
\centering
\includegraphics[width=0.6\hsize]{figs/metrics-10.png}
\caption{Efficiency metrics for the Safe Controller (in green) and the MPC Controller (in orange) for nominal scenarios
when the initial relative distance is 10 m.}
\label{metrics-10}
\end{table}
\noindent
\textbf{Out-of-nominal scenarios}
Fig. \ref{velocity-break-plot} and Fig. \ref{distance-break-plot}
show the speed and the relative distance for the emergency scenarios, respectively.
Note that both controllers react to a sudden brake of the vehicle ahead.
The Safe Controller always maintains a safe distance between the two vehicles, thus avoiding collision.
On the contrary, the MPC Controller is unsafe in two out of nine cases,
in particular, when the period of the speed function becomes large.
For instance, in Fig. \ref{distance-break-plot}
we can see that a collision occurs after 40 seconds when T= 30 s and A = 12 m/s
(the blue dashed line marks the beginning of the braking).
\vspace{-0.5cm}
\begin{figure} [H]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/section2_v_m12.eps}
\caption{Speed for the Safe Controller (in green) and the MPC Controller (in orange) in out-of-nominal scenarios
when the vehicle ahead suddenly brakes.}
\label{velocity-break-plot}
\end{figure}
\vspace{-1.5cm}
\begin{figure} [H]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/section2_d_m12.eps}
\caption{Relative distance for the Safe Controller (in green) and the MPC Controller (in orange)
when the vehicle ahead suddenly brakes at the moment indicated by the dashed blue line.}
\label{distance-break-plot}
\end{figure}
\vspace{-1cm}
\iffalse
The speed levels and the corresponding braking and accelerating distances are shown in TABLE \ref{parameters}.
\begin{table}[htp]
\caption{Braking and accelerating distances for the chosen speed levels}
\begin{center}
\begin{tabular}{c|ccccccccc}
speed level (m/s) & 0 & 4 & 8 & 12 & 16 & 20 & 24 & 28 & 32 \\ \hline
braking distance (m) & 0 & 5 & 13 & 27 & 45 & 70 & 99 & 134 & 174 \\ \hline
accelerating distance (m) & 0 & 10 & 18 & 32 & 50 & 75 & 104 & 139 & 179 \\
\end{tabular}
\end{center}
\label{parameters}
\end{table}
\fi
\iffalse
The assumptions made by each controller are shown in the following Table \ref{assumptions}.
\begin{table}[htp]
\caption{Inputs, outputs and assumptions of each controller}
\begin{center}
\begin{tabular}{ c | c | c | c }
& inputs & outputs & assumptions \\
\hline
MPC controller & front vehicle velocity $v_{f}$ & & $v_{f} \geq 0$, \\
& front vehicle position $p_{f}$ & acceleration $u^*$ & worst case braking $b$ \\
\hline
Safe controller & free distance ahead $F(t)$ & target speed level $v$ & $v_{f} \geq 0$, \\
& & & worst case braking $b$ \\
\end{tabular}
\end{center}
\label{assumptions}
\end{table}
\fi
\section{Conclusion and discussion}
\label{sec:conclusion}
We propose a method for building a hybrid safe-by-construction and efficient collision avoidance controller.
The controller integrates a MPC Controller, a discrete Safe Controller
and a Switch that combines the outputs of the two controllers to generate stimuli that are safe and efficient.
We show experimentally that the Hybrid Controller besides guaranteeing safety, ensures high efficiency
because it "takes the best" from each one of the integrated controllers.
The MPC Controller seeks policies that reduce both the relative speed and the relative distance
while in nominal scenarios the Safe Controller seeks minimization of the relative distance.
We show that this hybrid control policy ensures a very good efficiency
measured by three criteria: performance, road occupancy and comfort.
We adopt a pragmatic and progressive methodology
based on the comparative evaluation of the two constituent controllers for both nominal and out-of-nominal scenarios.
The evaluation provides a good insight on the merits of the respective control principles which motivates the design of the Hybrid Controller.
The experimental results confirm the feasibility and the practical relevance of hybrid controllers for safe and efficient driving.
A key and original lesson from our results is that the Safe Controller is not simply a monitor that takes over in critical situations.
It also significantly contributes to efficiency applying a control policy that nicely complements the MPC policy.
The experimental results show that the interplay between the dynamics of discrete and continuous controllers
pursuing complementary objectives can be surprisingly rich.
Its study may lead to more elaborated and enhanced hybrid policies.
In future work we will investigate our hybrid control principle
by replacing the MPC Controller with other types of adaptive controllers, e.g., machine-learning-based controllers \cite{learning2016}.
\section{Hybrid collision avoidance control}
\label{sec:hybrid}
\subsection{Design and implementation of the Hybrid Controller}
The above comparative evaluation confirms that
in most cases the MPC Controller is slightly better in terms of performance and occupancy,
while it fails to be safe for out-of-nominal scenarios when the vehicle ahead abruptly stops.
As expected the Safe Controller preserves safety in all scenarios and assures more comfortable driving.
We have explained the design principle of the Hybrid Controller in the Introduction (Fig.\ref{combination-arch}).
It consists of the Safe Controller and the MPC Controller running in parallel
and a Switch deciding which one of the control stimuli takes effect.
The Switch produces the target speed of the ego vehicle taking care that it never exceeds the maximal safe speed $v_{max}$
computed as a function of the distance $d$ and the maximal deceleration of the ego vehicle.
For constant braking rate $a_{max}$, we have $v_{max} = (2\cdot a_{max} \cdot d) ^{1/2}$.
In the Hybrid Controller under the constraint $v \leq v_{max}$,
the Switch selects the highest speed between $v_{mpc}$ and $v_{safe}$ so as to achieve the best performance.
Hence, it continuously applies the following rules to select the target speed $v$ of the Hybrid Controller: \\
\textbf{if} $v_{safe} \leq v_{mpc} \leq v_{max} $ \textbf{then} $v := v_{mpc}$ \\
\textbf{else if} $v_{mpc} \leq v_{safe}$ \textbf{then} $v := v_{safe}$ \\
\textbf{else if} $v_{max} \leq v_{mpc}$ \textbf{then} $v := v_{max}$
Notice that in all cases the rules prevent the target speed $v$ from exceeding $v_{max}$.
In nominal conditions it can happen that $v_{mpc} \leq v_{safe}$.
This is the case when the distance $d$ is large enough and the speed of the vehicle ahead is decreasing.
The speed $v_{safe}$ can exceed $v_{mpc}$ as the Safe Controller focuses on minimizing the relative distance
while the MPC Controller tracks the speed of the vehicle ahead.
An important difference from similar works dealing with hybrid controllers \cite{provebecorrectacc,2017Adaptive}
is that our Safe Controller contributes not only to safety but also to a large extent to performance
and comfort and even in some cases to improve occupancy.
This observation is confirmed by experimental results provided in the next section.
Notice that this hybrid control principle can be applied
by replacing in our architecture the MPC Controller by other adaptive controllers,
including controllers based on machine learning \cite{learning2016}.
\subsection{Evaluation of the Hybrid Controller}
We consider both nominal and out-of-nominal scenarios as before and adopt the same experimental settings.
For nominal scenarios, we compare the efficicency of the three controllers.
Furthermore, for out-of-nominal scenarios we consider additional braking rates of the vehicle ahead.
\vspace{0.25cm}
\noindent
\textbf{Nominal scenarios}
Fig.\ref{combine-velocity-plot} depicts the speed of the ego vehicle for the three controllers in nominal scenarios.
It shows that the Hybrid Controller can also track the speed of the vehicle ahead closely, taking the best from the two controllers.
Notice that the purple line can be above the orange line,
for instance during simulation time around 20 seconds when T = 30 and A =9.
The upmost part of Table.\ref{metrics-hybrid-10} provides results comparing the performance of the three controllers.
Note that the Hybrid Controller outperforms the two other controllers.
Fig.\ref{combine-distance-plot} depicts the relative distance for the three controllers.
As expected the distance maintained by the Hybrid Controller is in general smaller
taking advantage of the strength of the Safe Controller for minimizing the relative distance.
This is also shown in the middle part of Table \ref{metrics-hybrid-10},
which compares the occupancy for the three controllers.
The Hybrid Controller achieves higher occupancy than the other two controllers.
The comfort metrics provided in the bottom part of Table \ref{metrics-hybrid-10},
show that when the period and the amplitude of the speed function are small,
the Safe Controller produces the most comfortable driving policies.
While when they become larger,
the Hybrid Controller is better than the Safe Controller and is slightly outperformed by the MPC Controller.
Table \ref{time-percentage} shows time percentages corresponding to the application
by the Hybrid Controller of the MPC policy, the nominal safe policy and the out-of-nominal safe policy.
We can see that the MPC Controller contributes more than the other two,
while the contribution of the Safe Controller is non-negligible.
The maximal safe speed $v_{max}$ is applied to a very small percentage of cases to ensure safety.
\vspace{0.25cm}
\noindent
\textbf{Out-of-nominal scenarios}
Fig.\ref{combine-velocity-break-plot} and Fig.\ref{combine-distance-break-plot}
provide results for the three controllers in the out-of-nominal scenarios
where the vehicle ahead suddenly brakes with rate $12 ~ m/s^2$.
Note that the MPC Controller becomes unsafe for increasing amplitude and period of the speed of the vehicle ahead.
There are two out of nine settings where the MPC control policy results in collision.
In Annex A, additional experiments for braking rates $4 ~ m/s^2$ and $8 ~ m/s^2$ are provided.
%
They show that the MPC Controller is safe for all scenarios with braking rate $4 ~ m/s^2$
while it is fails to be safe in one out of nine scenarios with braking rate $8 ~ m/s^2$.
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/v.eps}
\caption{Speed for the Hybrid Controller (in purple), the MPC Controller (in orange), the Safe Controller (in green)
when the speed of the vehicle ahead is a sinusoidal function (in blue).}
\label{combine-velocity-plot}
\end{figure}
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/d.eps}
\caption{Relative distance for the Hybrid Controller (in purple),
the MPC Controller (in orange) and the Safe Controller (in green)
when the speed of the vehicle ahead is a sinusoidal function.}
\label{combine-distance-plot}
\end{figure}
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/v_m12.eps}
\caption{Speed of the ego vehicle for the Hybrid Controller (in purple),
the MPC Controller (in orange) and the Safe Controller (in green)
when the vehicle ahead brakes with rate $12 ~ m/s^{2}$.}
\label{combine-velocity-break-plot}
\end{figure}
\vspace{-1.5cm}
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/d_m12.eps}
\caption{Relative distance for the Hybrid Controller (in purple),
the MPC Controller (in orange) and the Safe Controller (in green)
when the vehicle ahead brakes with rate $12 ~ m/s^{2}$.}
\label{combine-distance-break-plot}
\end{figure}
\vspace{-1.5cm}
\begin{table} [H]
\centering
\includegraphics[width=0.7\hsize]{figs/metrics-hybrid-10.png}
\caption{Efficiency metrics for the Hybrid Controller (in purple),
the Safe Controller (in green) and the MPC Controller (in orange) for nominal scenarios. }
\label{metrics-hybrid-10}
\end{table}
\vspace{-1.5cm}
\begin{table} [H]
\centering
\includegraphics[width=0.7\hsize]{figs/time-percentage.png}
\caption{Percentage of time in use for the Safe Controller,
the MPC Controller and the maximal safe speed $v_{max}$ for nominal scenarios. }
\label{time-percentage}
\end{table}
\iffalse
\begin{figure}[h]
\centering
\includegraphics[scale = 0.5]{figs/safe-regions.png}
\caption{Illustraction of the vehicle movement based on the hybrid control}
\label{rta-arch}
\end{figure}
\begin{algorithm}[h]
\caption{Safe and efficient collision avoidance algorithm}
\begin{algorithmic}[1]
\State Input: state $x$ of the ego vehicle, state $x_{f}$ of the front vehicle
\State Output: acceleration $a$ for the ego vehicle
\State Initialization
\While{true}
\If{Mode = MPCController}
\State $a(t)$ = NominalController($x(t)$, $x_{f}(t)$)
\If{NominalControllerUnsafe($a(t)$, $x(t)$, $x_{f}(t)$)}
\State Mode = SafeController
%
\EndIf
\EndIf
\If{Mode = SafeController}
\State $a(t)$ = SafeController($x(t)$, $x_{f}(t)$)
\If{NominalControllerSafe($a(t)$, $x(t)$, $x_{f}(t)$)}
\State Mode = MPCController
\EndIf
\EndIf
\State $t = t + \triangle t$
\EndWhile
\end{algorithmic}
\label{algorithm1}
\end{algorithm}
\fi
\section{Introduction}
It is widely believed that the deployment of autonomous vehicles can improve not only the traffic efficiency, but also its safety.
Collision avoidance, as a fundamental safety requirement for autonomous vehicle control,
plays a crucial role in guaranteeing traffic safety and reducing the number of vehicle crashes,
given the fact that more than 50 percent of the total amount of vehicle crashes are rear-end collisions
\footnote{http://www.ntsb.gov/safety/safetystudies/Documents/SIR1501.pdf}.
A variety of approaches and frameworks has been investigated for collision avoidance control.
The underlying assumptions vary largely with the level of modeling of the vehicle dynamics and the nature of the controller stimuli.
Control-based techniques typically focus on collision avoidance
for adaptive cruise control \cite{park2009obstacle,mpc2014}
taking into account the impact of perception uncertainty and accuracy of vehicle models \cite{li2013,robust2014}.
They allow achieving optimality for specific tasks or scenarios without providing strict safety guarantees.
Model Predictive Control (MPC) \cite{mpc2014,survey2000},
as a prominent optimal control approach has been widely used for vehicle control
because it allows handling multiple constraints in a receding horizon.
Nonetheless, MPC relies on the use of optimization algorithms and by its nature cannot guarantee safety, namely collision avoidance.
Furthermore, depending on the optimization algorithms and the dynamic model of the vehicles,
it may also result in high computational complexity because of heavy iterative calculations.
A different line of works focus on safety using formal methods.
These apply a variety of techniques including reachability analysis \cite{loos2011adaptive,atva18motionplanner},
Responsibility-Sensitive Safety (RSS) model \cite{shalev2017formal},
logic-based controller synthesis \cite{nilsson2015correct,sadraddini2017provably},
as well as the design of safety supervision mechanisms for specific scenarios \cite{korssen2017systematic,krook2019design}.
%
The basic principle of safe collision avoidance control, as formalized and implemented in our previous work \cite{wang2020safe},
is to keep a safe distance with the preceding vehicles such that in any case
the ego vehicle has enough space to brake and avoid collisions.
Although these results can guarantee correctness by construction,
they lead to solutions that privilege strict safety at the expense of efficiency.
Designing a collision avoidance controller for autonomous vehicles
that meets both efficiency and safety requirements remains a non-trivial problem.
The two types of requirements are antagonistic as efficiency implies conflicting properties
such as performance, i.e., maximization of the average speed,
road occupancy, i.e., keeping the inter-vehicle distance as small as possible
and comfort, i.e., no sudden speed changes.
In search of solutions seeking compromises between efficiency and safety,
a few works adopt a hybrid approach combining continuous and discrete control dynamics.
The continuous controller is supervised by an automaton that takes over to handle critical situations.
Hybrid approaches rely on a principle of "division of roles" often applied in systems engineering
that distinguishes between nominal operating conditions and out-of-nominal ones \cite{2018Robust}.
The continuous controller has parameters tuned to achieve given goals for nominal operation
while the discrete controller deals with out-of-nominal situations.
For example, \cite{provebecorrectacc,2017Adaptive} propose a switch-control approach,
where a MPC Controller is safeguarded by an emergency maneuver activated to avoid collision
when the ego vehicle is in a critical situation.
Nonetheless, the emergency maneuver only takes care of safety
without investigating possible trade-offs between safety and efficiency of the switch-control policy.
We design and experimentally evaluate a safe-by-construction Hybrid Controller
that proves to be efficient for the three mentioned criteria.
The architecture of the Hybrid Controller is shown in Fig.\ref{combination-arch}.
It results from the integration into a single architecture of a nominal
MPC Controller and the discrete Safe Controller presented in \cite{wang2020safe}.
The two controllers running in parallel receive the speed $v_e$ and $v_a$ of the ego vehicle
and the vehicle ahead respectively, as well as their distance $d$,
and compute the target speeds $v_{mpc}$ and $v_{safe}$ respectively.
The control policies for computing $v_{mpc}$ and $v_{safe}$ adopt nominal conditions.
In particular, $v_{safe}$ is a safe speed under the assumption
that the speed of the vehicle ahead is a continuous function and the deceleration does not exceed some limit
corresponding to normal driving conditions.
In addition to $v_{safe}$, the Safe Controller provides a speed $v_{max}$
that is the maximal safe speed for out-of-nominal conditions when the vehicle ahead suddenly stops, e.g. in case of accident.
This speed is computed as a function of the relative distance between the ego vehicle and the car ahead
with the maximum deceleration rate of the ego vehicle.
The Hybrid Controller uses a Switch selecting between the three speeds $v_{mpc}$, $v_{safe}$ and $v_{max}$
to optimize efficiency criteria while preventing the speed of the ego vehicle to exceed $v_{max}$.
We show that the combined use of $v_{mpc}$, $v_{safe}$ and $v_{max}$ ensures both efficiency and safety in nominal conditions
and moreover safety is preserved in out-of-nominal situations.
\begin{figure}[h]
\centering
\includegraphics[scale = 0.45]{figs/combination-diagram.png}
\caption{The architecture of the Hybrid Controller}
\label{combination-arch}
\end{figure}
Our solution is inspired by the Simplex architecture principle \cite{simplex2001},
for runtime assurance of safety-critical systems.
The architecture uses a Decision Module that switches control from a high-performance
but unverified (hence potentially unsafe) Advanced Controller
to a verified-safe Baseline Controller if some safety violation is imminent.
Nonetheless, in our solution the Safe Controller contributes not only to out-of-nominal situations
but also to some nominal situations where it proves to be more efficient than the MPC Controller.
Hence, not only the Hybrid Controller is safe but also efficiency gains from the synergistic collaboration are substantial.
Of course, the alternation of roles between the MPC and the Safe Controller should be implemented
so as to avoid sudden changes of the kinematic state of the vehicle.
In particular, care should be taken to avoid jerk (i.e., abrupt changes of acceleration) that might cause passenger discomfort.
Additionally, the paper proposes a pragmatic methodology for the comparative evaluation of the three controllers
for two types of scenarios:
1) nominal scenarios where the speed of the vehicle ahead is a known continuous function; and
2) out-of-nominal scenarios where the vehicle ahead abruptly brakes.
For nominal scenarios, the three controllers are evaluated against three efficiency criteria.
\begin{itemize}
\item The first criterion is performance that measures how much close the speed of the ego vehicle can get to the speed of the vehicle ahead.
For a period of time it can be defined as the ratio of the average speed of the ego vehicle with respect to the average speed of the vehicle ahead.
This ratio is less than one if the distance between the two vehicles is initially zero.
\item The second criterion is road occupancy that measures how much close can get the ego vehicle to the vehicle ahead in collision-free scenarios.
\item The third criterion measures passenger comfort that decreases as the standard deviation of the acceleration increases.
\end{itemize}
The paper is organized as follows.
Section 2 presents the design principles for the MPC Controller and the Safe Controller,
as well as a comparative study for the two types of scenarios
and the three efficiency criteria using the Carla simulator.
Section 3 presents the design and the implementation of the Hybrid Controller
and its experimental evaluation.
Section 4 emphasizes the feasibility and the practical relevance of hybrid controllers
for safe and efficient driving and then outlines directions for future work.
\section*{Annex A}
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/v_m4.eps}
\caption{Speed of the ego vehicle for the Hybrid Controller (in purple),
the MPC Controller (in orange) and the Safe Controller (in green)
when the vehicle ahead brakes with rate $4 ~ m/s^{2}$.}
\label{combine-velocity-plot-m4}
\end{figure}
\vspace{-0.5cm}
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/d_m4.eps}
\caption{Relative distance for the Hybrid Controller (in purple),
the MPC Controller (in orange) and the Safe Controller (in green)
when the vehicle ahead brakes with rate $4 ~ m/s^{2}$.}
\label{combine-distance-plot-m4}
\end{figure}
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/v_m8.eps}
\caption{Speed of the ego vehicle for the Hybrid Controller (in purple),
the MPC Controller (in orange) and the Safe Controller (in green)
when the vehicle ahead brakes with rate $8 ~ m/s^{2}$.}
\label{combine-velocity-break-plot-m8}
\end{figure}
\begin{figure} [!htb]
\centering
\includegraphics[width=0.7\hsize]{expr/submite_3_15/eps_03_15_d_10/d_m8.eps}
\caption{Relative distance for the Hybrid Controller (in purple),
the MPC Controller (in orange) and the Safe Controller (in green)
when the vehicle ahead brakes with rate $8 ~ m/s^{2}$.}
\label{combine-distance-break-plot-m8}
\end{figure}
\end{document}
|
\section{Introduction}
Oilfield planning and operations related decisions such as well placement and production settings have direct impact on cost and revenue for any given project. Hence, field development planning is done to maximize fluid recovery or net present value of cash flow, and is often accomplished using reservoir simulation studies. Simulation results are used to compare the performance of various well configuration and controls. Optimization algorithms are used in a workflow to search for well settings which maximize return on investment [1]. Search involves running thousands of reservoir simulation run for a range of possible solutions and identifying a few solutions of interest for further deliberations.
Optimization algorithms used for such automated workflow are mainly divided in two categories: point based and population based. In this work, population based evolutionary optimization approach is used for oilfield well control optimization problem as these algorithms are relatively simple to program and more effective for a large scale, highly non-linear, and multi-modal problems. Moreover, these algorithms do not require gradient calculations and can also be used for discrete problems. Two popular evolutionary optimization algorithms, genetic algorithm (GA) and particle swarm algorithm (PSO), are used in standalone mode for solving oil well control optimization problem. Next two optimization methods, GA and PSO, are used with another popular optimization method, covariance matrix adaptation – evolution strategy (CMA-ES), in hybrid mode. CMA-ES is also a gradient-free evolutionary optimization algorithm and has been successfully used in different problem settings [6].
Computational requirement of such oilfield well design and control study is quite high, and hence commercial cloud computing is used to handle workload. Parallel programming is used to distribute computational workload among nodes of a large cluster. Two different computational setup are used, one in which a single cluster with a large number of cores is used, and another in which multiple mid-sized clusters are used with a common file server.
The main contributions of this work are:
\begin{itemize}
\item Hybrid evolutionary algorithms, GA-CMA-ES and PSO-CMA-ES, are proposed for oilfield well control optimization problem.
\item Hybrid algorithms are tested using well control optimization study of large Olympus benchmark case.
\item Robust optimization is done using an ensemble of 10 simulation model realizations, and results are compared with that from a single realization case.
\item Parallel distribution of computational workload on cloud is done in two ways, one between nodes of large cluster and another between nodes of multiple mid-sized clusters with a common file server.
\end{itemize}
\section{Related Works}
Udy {\it et al.} summarized various approaches for oilfield production optimization and compared algorithms including adjoint methods, particle swarm optimization, simulated annealing and genetic algorithms [1]. Tanaka {\it et al.} compared three different optimization approaches GA, PSO and ensemble-optimization (EnOpt) for Olympus benchmark case [3]. Harb {\it et al.} used hybrid evolutionary optimization method, black hole particle swarm optimization (BHPSO), for simultaneously optimizing well count, location, type, and trajectory [4]. Fonesca {\it et al.} used a hybrid of ensemble optimization EnOpt and gradient free covariance matrix adaptation (CMA) method for well control problem [5], while Bouzarkouna {\it et al.} used CMA-ES method for optimal well placement under uncertainty [2].
\section{Approach}
Olympus benchmark case is a synthetic simulation model based on a typical North Sea offshore oilfield, and is used in this work to compare the effectiveness of different optimization algorithms [11]. Field is a 9km by 3km by 50m in size with minor faults and limited aquifer support. Subsurface uncertainty is captured by 50 realizations with different permeability, porosity, net-to-gross and initial water saturation, of which only 10 random realizations are used in this work to reduce computational requirement. GA and PSO optimization algorithms are first tested for ensemble of 10 realizations, and then going forward a representative realization is used for the remaining studies. A realization in such studies stands for a random simulation model, while ensemble of realizations is often used to capture subsurface uncertainty.
Simulation model has 11 production wells and 7 injection wells which are all on bottom-hole pressure (BHP) control (Figure~\ref{olympus_model}). For the optimization study, each well’s BHP control setting is changed independently at four different pre-defined times during the 20 year simulation period. Thus, there are total 72 different control variables defined as continuous real parameter which can take any value within a given range.
Two different objective functions are used, one as the weighted sum of cumulative fluid production/injection while second as the net present value of discounted cash-flow [1, 14]. This is done to have a rigorous analysis by encoding varying degree of non-linearity among the two objective functions. Weighted sum of cumulative fluid (WCF) is defined as:
\begin{equation}
WCF = Q_{op} - 0.1 * (Q_{wp}+Q_{wi})
\end{equation}
where $WCF$ is weighted sum, $Q_{op}$ is total oil produced, $Q_{wp}$ is total water produced, and $Q_{wi}$ is total water injected. Net present value (NPV) of discounted cash-flow incorporates additional degree of non-linearity and is defined as:
\begin{equation}
NPV = \sum_{f=1}^{N_f}\sum_{\tau=1}^{N_\tau}\frac{Q_{f,\tau}C_f}{(1+d_f)^{\tau-1}}
\end{equation}
where $N_{f}$ is the number of fluid component, $N_{\tau}$ is number of time intervals, $Q_{f,\tau}$ is the volume component produced/injected in time interval $\tau$, $C_{f}$ is the price/cost while $d_{f}$ is discount factor for component $f$.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{sim_model}
\caption{Olympus simulation model overview: initial oil saturation and well locations}
\label{olympus_model}
\end{figure}
Parallelization is easy for evolutionary optimization algorithms as objective function calculation for population samples can be done in parallel [3]. This calculation includes making the reservoir simulation run for the population sample (or simulation case), and then determining the required value. Message-Passing Interface (MPI) is used in this study to distribute workload among nodes and thus speed-up calculations. For a large cluster setting, master node assigns simulation cases to different worker nodes of cluster and then accumulates simulation results. A number of small clusters can also be used to parallalize workload if a large cluster is not available or is not efficient. For such a setting, master node on main cluster uses a common file server accessible by smaller clusters which in turn make simulation runs.
\section{Algorithm}
Three different gradient-free, evolutionary optimization algorithms used in this study are genetic algorithm (GA), particle swarm optimization (PSO), and covariance matrix adaptation-evolution strategy (CMA-ES).
Genetic algorithm (GA) has been a popular optimization algorithm and uses three basic steps for population refinement, namely selection, crossover and mutation [9]. For a real-coded genetic algorithm (RCGA) used in this study, real variables themselves are used for population refinement and objective function calculation, as compared to binary encoding done for binary-coded genetic algorithm (BCGA) [10]. In this work, RCGA is used to obviate the need of encoding and decoding control variables. Crossover and mutation steps of GA help the algorithm in exploring complex and multi-modal solution space effectively.
Particle swarm optimization (PSO) uses collective intelligence based on particle position and velocity. It is the exchange of information among particles that leads to the emergence of global patterns [7, 8, 13]. Sample update is done using inertia, personal-best, and global-best values along with random components. It is essential to properly tune hyperparameters for both GA and PSO algorithms to find good solutions and help optimization run converge. Based on domain expertise, the algorithms can be made more exploratory or more exploitative by tuning hyperparameters. The hyperparameters can also be varied during an optimization run to have more exploration in the beginning and more exploitation towards the end of the optimization run.
Covariance matrix adaptation - evolution strategy (CMA-ES), on the other hand, is a stochastic population-based optimization approach where population is updated using weighted mean of samples from the previous generation and a random draw from a multivariate normal distribution [5, 6]. Next, parameters of multivariate normal distribution are updated using objective function evaluations. CMA-ES attempts to learn a second-order model of the underlying objective function and follows a natural gradient approximation of the expected fitness, similar to the approximation of the inverse Hessian matrix in Quasi-Newton methods [6, 18]. For multi-modal functions, where local optima cannot be interpreted as perturbations to an underlying convex (unimodal) topology, its performance can strongly decrease.
\begin{table}[t]
\caption{Solving Rastrigin test function using GA and PSO}
\label{rastrigin}
\centering
\begin{tabular}{ccccc}
\toprule
Dimensions & Population size & Iterations & Best cost (GA) & Best cost (PSO)\\
\midrule
2 & 40 & 100 & 0.99 & 7.7e-10 \\
50 & 40 & 100 & 172.1 & 322.0 \\
50 & 100 & 200 & 97.8 & 158.7 \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[b]
\centering
\includegraphics[scale=0.75]{pso_comparison}
\caption{Cost history for solving Rastrigin test function using PSO algirithm}
\label{pso_rastrigin}
\end{figure}
In continuous domains, the volume of search space grows exponentially with dimensions, and is known as the curse of dimensionality. This phenomenon causes a population size that is capable of adequate search space coverage in low dimensional search spaces to degrade into sparse coverage in high dimensional search spaces. GA and PSO algorithms work well for a problem with smaller number of control variables (or dimensions), but have problem converging when the number of control variables increase [12, 17]. To demonstrate this, both GA and PSO are used for solving Rastrigin function with varied number of dimensions. Rastrigin function has global minimum at zeros for any given number of dimensions and the associated cost is also zero. Cost is calculated as:
\begin{equation}
f = 10d + \sum_{i=1}^{d}[x_i^2-10\cos (2\pi x_i)]
\end{equation}
where $d$ is the number of dimensions, $x_i$ is the value along dimension $i$, and $f$ is the cost. As the dimensionality increases, both GA and PSO require larger population size and iterations, but still do not manage to converge to the global minimum (Table~\ref{rastrigin}, Figure~\ref{pso_rastrigin}). While CMA-ES is not very effective for multi-modal problems, it has been found to have excellent convergence properties for unimodal problems. Thus once GA and PSO algorithms have established the region of interest, CMA-ES can be used to achieve convergence or further improvement. In general, effectiveness of any optimization algorithm depends on the problem class under consideration, and is commonly known as the no free lunch theorem for optimization [19].
In this work, GA and PSO are each first used with ensemble of 10 realizations as well as a single representative realization to establish their efficacy with such optimization problems. Going forward, optimization studies are done only for single representative model in order to cut down on computational cost. For using two optimization methods in a hybrid mode, one optimization algorithm is first allowed to run for a number of generations, and then second algorithm takes over. Two hybrid approaches, GA-CMA-ES and PSO-CMA-ES, are tested and compared with standalone GA and PSO algorithms [14, 15, 16, 18].
\section{Results}
Olympus model has 11 producer wells and 7 injector wells, all on bottom-hole pressure (BHP) constraint (Figure~\ref{olympus_model}). Constraints can be changed independently for each well at four predefined times giving a total of 72 control variables. Well control settings for an ensemble of 10 realizations as well as a single representative realization are first optimized using GA, and then steps are repeated using PSO. Table~\ref{optim_param} shows related parameter values for the optimization runs. A discount factor of 0.08 was used for all fluid components in NPV calculation. For ensemble case, all 10 realizations (or simulation models) are simulated for any given population sample and results are averaged. As can be seen in Figures~\ref{ol_wcf} and~\ref{ol2_wcf}, both GA and PSO algorithms are able to optimize well control settings and yield higher weighted cumulative fluid as compared to the default case. As sampling is randomized in both the algorithms, every optimization run will proceed in unique fashion. Thus multiple optimization runs are required for comparing different algorithms. Going forward, only single representative reservoir model is optimized to cut down on computational requirement.
\begin{table}[t]
\caption{Optimization study parameters}
\label{optim_param}
\centering
\begin{tabular}{lr||lr}
\toprule
Population size & 40 & Oil price & \$ 40 per barrel \\
Iterations (GA or PSO) & 100 & Produced water cost & \$ 4 per barrel \\
Iterations (Hybrid) & 150 & Injected water cost & \$ 2 per barrel \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[b]
\centering
\begin{minipage}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{olympus}
\caption{GA and PSO based ensemble optimization run with WCF objective}
\label{ol_wcf}
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{olympus2_wcf}
\caption{GA and PSO based single realization optimization run with WCF objective}
\label{ol2_wcf}
\end{minipage}
\end{figure}
To further validate algorithm efficacy, objective function is next changed to net present value (NPV) of discounted cash-flow, and the results for GA and PSO based optimization runs are shown in Figure~\ref{ol2_npv}. Again algorithms seem to be effective in handling such simulation based oil production optimization problem. As compared to weighted sum of cumulative fluid (WCF) based optimization, NPV based optimization yields muted increase over default/base case because of the discounting.
Hybrid optimization run is made by transferring the resulting population from one algorithm to the next as its starting population for further improvement. Total 6000 simulation runs are made for such studies, out of which first 4000 runs (or 100 generations) are done using primary algorithm followed by 2000 runs (or 50 generations) using secondary algorithm. Two different hybrid optimization settings are: GA followed by CMA-ES (GA-CMA-ES) and PSO followed by CMA-ES (PSO-CMA-ES). Once region of interest is found using GA or PSO, CMA-ES algorithm further refines the solution as it is a better algorithm for local exploitation. Figures~\ref{ol2_hybrid_wcf} and~\ref{ol2_hybrid_npv} show that CMA-ES is able to further improve upon the solution, and thus GA-CMA-ES and PSO-CMA-ES approach of hybrid optimization is better as compared to standalone GA or PSO. There is a marked jump in objective function value after 4000 runs when CMA-ES takes over from GA or PSO. It needs to be mentioned that CMA-ES does not have good exploratory mechanism for multi-modal solution space, and hence it does not perform as good as GA or PSO in standalone mode. Domain expertise is crucial for success of such optimization studies as their design would vary depending on underlying solution space and objective function.
\begin{figure}[t]
\centering
\includegraphics[scale=0.65]{olympus2_npv}
\caption{GA and PSO based single realization optimization run with NPV objective}
\label{ol2_npv}
\end{figure}
\begin{figure}[b]
\centering
\begin{minipage}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{olympus2_hybrid_wcf}
\caption{Hybrid GA-CMA-ES and PSO-CMA-ES based single realization optimization run with WCF objective}
\label{ol2_hybrid_wcf}
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{olympus2_hybrid_npv}
\caption{Hybrid GA-CMA-ES and PSO-CMA-ES based single realization optimization run with NPV objective}
\label{ol2_hybrid_npv}
\end{minipage}
\end{figure}
Population based algorithms can be easily programmed in parallel mode with objective function for each sample being calculated using a separate node. In the calculation of objective function for a given population sample, compute intensive step is making the reservoir simulation run. Either a large cluster can be used to divide workload among cores, or multiple mid-size clusters can be used with a common file server. By reading from and writing to a common file server (Figure~\ref{config}), multiple clusters can be used to carry-out compute intensive simulation runs, thus reducing turn-around time.
\section{Conclusion}
Evolutionary algorithms are efficient at solving oilfield well control optimization problem. Two such algorithms are used in hybrid mode to explore multi-modal solution space and attain convergence. GA-CMA-ES and PSO-CMA-ES hybrid optimization approach perform better as compared to using GA or PSO for such optimization problem. For the hybrid optimization, GA or PSO algorithm is first used to explore solution space and identify region of interest, and then CMA-ES is used to further refine the solution. Domain expertise is crucial for success of such optimization studies as their design would vary depending on solution space and objective function. Cloud computing and parallel programming are used to handle the computational workload. Two different cluster settings, one with single large cluster and other with multiple mid-sized clusters, have been used to handle high computational workload.
\begin{ack}
Research was made possible by generous cloud credits provided by Amazon Web Services, DigitalOcean, Google Cloud and IBM Cloud. Olympus benchmark reservoir model was made available by the Netherlands Organization for Applied Scientific Research (TNO).
\end{ack}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{config}
\caption{Compute resource setup with multiple clusters and common file-server}
\label{config}
\end{figure}
\section*{References}
\medskip
\small
[1] Udy, J., Hansen, B., Maddux, S., Petersen, D., Heilner, S., Stevens, K., Lignell, D.\ \& Hedengren, J.D.\ (2017) Review of field development optimization of waterflooding, EOR, and well placement focusing on history matching and optimization algorithms. {\it Processes} {\bf 5}(34).
[2] Bouzarkouna, Z., Ding, D.Y.\ \& Auger, A.\ (2012) Well placement optimization under uncertainty with CMA-ES using the neighborhood. {\it $13^{th}$ European Conference on the Mathematics of Oil Recovery}.
[3] Tanaka, S., Wang, Z., Dehghani, K., He, J., Velusamy, B.\ \& Wen, X.\ (2018) Large scale field development optimization using high performance parallel simulation and cloud computing technology. {\it SPE Annual Technical Conference and Exhibition}.
[4] Harb, A., Kassem, H.\ \& Ghorayeb, K.\ (2019) Black hole particle swarm optimization for well placement optimization. {\it Computational Geosciences}, {\bf 24} (6):1979-2000.
[5] Fonseca, R.M., Leeuwenburgh, O., Van den Hof, P.M.J.\ \& Jansen, J.D.\ (2013) Improving the ensemble optimization method through covariance matrix adaptation (CMA-EnOpt). {\it SPE Reservoir Simulation Symposium}.
[6] Hansen, N.\ (2016) The CMA evolution strategy: a tutorial. arXiv preprint \href{https://arxiv.org/abs/1604.00772}{arXiv:1604.00772}.
[7] Wang, D., Tan, D.\ \& Liu, L.\ (2018) Particle swarm optimization algorithm: an overview. {\it Soft Computing}, {\bf 22}:387-408.
[8] Sengupta, S., Basak, S.\ \& Peters, R.A.\ (2019) Particle swarm optimization: a survey of historical and recent developments with hybridization perspectives. {\it Machine Learning and Knowledge Extraction}, {\bf 1}:157-191.
[9] Katoch, S., Chauhan, S.S.\ \& Kumar, V.\ (2021) A review on genetic algorithm: past, present, and future. {\it Multimedia Tools and Applications}, {\bf 80}:8091-8126.
[10] Chuang, Y.C., Chen, C.T.\ \& Hwang, C.\ (2015) A real-coded genetic algorithm with a direction-based crossover operator. {\it Information Sciences}, {\bf 305}:320-348.
[11] Fonseca, R.M., Rossa, E.D., Emerick, A.A., Hanea, R.G.\ \& Jansen, J.D.\ (2020) Introduction to the special issue: overview of OLYMPUS optimization benchmark challenge. {\it Computational Geosciences}, {\bf 24} (6):1933-1941.
[12] Chen, S., Montgomery, J.\ \& Bolufé-Röhler, A.\ (2015) Measuring the curse of dimensionality and its effects on particle swarm optimization and differential evolution. {\it Appl Intell}, {\bf 42}:514-526.
[13] Zhang, Y., Wang, S.\ \& Ji, G.\ (2015) A comprehensive survey on particle swarm optimization algorithm and its applications. {\it Mathematical Problems in Engineering}, {\bf 2015}.
[14] Baumann, E.J.M, Dale, S.I.\ \& Bellout, M.C.\ (2020) FieldOpt: A powerful and effective programming framework tailored for field development optimization. {\it Computers and Geosciences}, {\bf 135}.
[15] Xu, P., Luo, W., Lin, X., Qiao, Y.\ \& Zhu, T.\ (2019) Hybrid of PSO and CMA-ES for global optimization. {\it IEEE Congress on Evolutionary Computation}.
[16] Santos, J., Ferreira, A.\ \& Flintsch, G.\ (2019) An adaptive hybrid genetic algorithm for pavement management. {\it International Journal of Pavement Engineering}, {\bf 20} (3):266-286.
[17] Clerc, M.\ \& Kennedy, J.\ (2002) The particle swarm - explosion, stability, and convergence in a multidimensional complex space. {\it IEEE Transactions on Evolutionary Computation}, {\bf 6} (1):58-73.
[18] Muller, C.L., Baumgartner, B.\ \& Sbalzarini, I.F.\ (2009) Particle swarm CMA evolution strategy for the optimization of multi-funnel landscapes. {\it IEEE Congress on Evolutionary Computation}, pp. 2685-2692.
[19] Wolpert, D.H.,\ \& Macready, W.G.\ (1997) No free lunch theorems for optimization. {\it IEEE Transactions on Evolutionary Computation}, {\bf 1} (1):67-82.
\end{document}
|
\section{History of the subject}\label{intro}
Let $\mathbb N=\{1,2,\ldots\}$ and $\mathbb N_0=\mathbb N\cup\{0\}.$
The Hurwitz zeta function is defined by the following series
\begin{equation}\label{E-2}
\zeta(z,q)=\sum_{n=0}^\infty\frac1{(n+q)^z},\quad\text{Re}(z)>1,~q\neq0,-1,-2,\ldots
\end{equation}
and it can be analytically continued to the entire $z$-plane except for a simple pole at $z=1$ with residue 1 (see \cite{Cohen} and \cite{SC}).
Special cases of $\zeta(z,q)$ include the Riemann zeta function
$$\zeta(z,1)=\zeta(z)=\sum_{n=1}^\infty\frac1{n^z}$$
and
$$\zeta\left(z,\frac12\right)=2^z\sum_{n=0}^\infty\frac1{(2n+1)^z}=(2^z-1)\zeta(z).$$
It is known that (\cite[p. 76, Corollary 9.6.10]{Cohen})
\begin{equation}\label{Bernoulli}\zeta(1-n,q)=-\frac1n B_n(q)\end{equation}
for $n\in\mathbb N,$ where $B_n(q)$ denotes the $n$th Bernoulli polynomial which
is defined as the coefficient of
$t^n$ in the generating function
\begin{equation*}
\frac{te^{qt}}{e^t-1}=\sum_{n=0}^{\infty}B_{n}(q) \frac{t^n}{n!}.
\end{equation*}
In particular, $B_{n}(0)=B_{n}$ is the $n$th Bernoulli number.
The Bernoulli numbers and polynomials arise from Bernoulli's calculations of power sums in 1713, that is,
$$
\sum_{j=0}^{m}j^{n}=\frac{B_{n+1}(m+1)-B_{n+1}}{n+1}
$$
(see \cite[p. 5, (2.2)]{Sun}).
The derivatives of Hurwitz zeta function $\zeta(z,q)$ at non-positive integers have many applications in number theory and mathematical physics
(see \cite{EE1} and \cite{Gross}). In 1986, by using Watson's Lemma and Laplace's method, Elizalde \cite{EE} gave an asymptotic expansion for the first derivatives
\begin{equation}\label{E-1}
\zeta'(-n,q)\equiv \left. \frac{\partial}{\partial z}\zeta(z,q)\right|_{z=-n}, \quad n\in\mathbb N_0.
\end{equation}
It is also interesting to refer \cite{WZ} and \cite{EE0,Ru,seri} for some researches on $\zeta'(z,q)$ at $z=0$ and at $-z\in\mathbb{N}$, respectively. In 1984, to study an analogue of the formula of Chowla and Selberg for real quadratic fields, Deninger \cite{De} investigated the second derivative $\zeta''(0,q)$.
In \cite{EE0} (also see \cite[Chapter 2]{EE1}), Elizalde further obtained some recursive formulas for the higher order derivatives, and presented explicit formulas for the case of small $m$ and small $-z\in\mathbb N$.
In this note we shall consider the alternating Hurwitz (or Hurwitz-type Euler) zeta function
which is defined by the following series
\begin{equation}\label{E-zeta-def}
\zeta_E(z,q)=\sum_{n=0}^\infty\frac{(-1)^n}{(n+q)^{z}},\quad\text{Re}(z)>0,~q\neq0,-1,-2,\ldots
\end{equation}
(see \cite[(3.1)]{Choi}, \cite[(1.3)]{HKK} and \cite[(1.1)]{WZ}). We see that $\zeta_E(z,q)$ can be analytically
continued to the entire $z$-plane without any pole and it satisfies the following identities:
\begin{equation}\label{J-1}
\zeta_E(z,q+1)+\zeta_E(z,q)=q^{-z},
\end{equation}
\begin{equation}\label{J-2}
\frac{\partial}{\partial q}\zeta_E(z,q)=-z\zeta_E(z+1,q).
\end{equation}
In analogue with (\ref{Bernoulli}), it is known that (\cite[p. 41, (3.8)]{WZ} and \cite[p. 520, (3.20)]{Choi})
\begin{equation}\label{Euler}\zeta_{E}(-n,q)=\frac{1}{2}E_n(q)\end{equation}
for $n\in\mathbb N_0,$ where $E_n(q)$ denotes the $n$th Euler polynomial which
is defined as the coefficient of
$t^n$ in the generating function
\begin{equation}\label{Euler-polynomial}
\frac{2e^{qt}}{e^t+1}=\sum_{n=0}^{\infty}E_{n}(q) \frac{t^n}{n!}.
\end{equation}
The Euler polynomials were introduced by Euler who calculated the alternating power sums, that is,
\begin{equation*}
\sum_{j=0}^{m}(-1)^{j}j^{n}=\frac{(-1)^{{m}}E_{n}(m+1)+E_{n}(0)}{2}
\end{equation*}
(see \cite[p. 5, (2.3)]{Sun}).
In algebraic number theory, the zeta function $\zeta_E(z,q)$ represents a partial zeta function of cyclotomic fields in one version of Stark's conjectures (see \cite[p. 4249, (6.13)]{HK-G}),
and its special case, the alternative series,
\begin{equation}~\label{Riemann-Euler}
\zeta_{E}(z,1)=\zeta_{E}(z)=\sum_{n=1}^{\infty}\frac{(-1)^{n-1}}{n^{z}},
\end{equation}
is also a particular case of Witten's zeta functions in mathematical physics (see \cite[p. 248, (3.14)]{Min}).
In the following we will denote the $m$th derivative of $\zeta_E(z,q)$ with respect to its first argument by $\zeta_E^{(m)}$ and $\zeta_E\equiv \zeta_E^{(0)},\zeta_E'\equiv \zeta_E^{(1)}$ and $\zeta_E''\equiv \zeta_E^{(2)}$.
Inspiring by the works of Elizalde \cite{EE0} and Rudaz \cite{Ru}, using the Boole summation formula,
we provide asymptotic expansions for the higher order derivatives of the alternating Hurwitz zeta function (\ref{E-zeta-def}) with respect to its first argument
$$\zeta_{E}^{(m)}(z,q)\equiv\frac{\partial^m}{\partial z^m}\zeta_E(z,q)$$
(see Section \ref{2.2}).
\section{Boole summation formula}
In this section, as a preliminary we state the Boole summation formula. It will serve as a main tool for our approach.
Then as an application, we shall prove an asymptotic series expansion of $\zeta_{E}(z,q)$ (see Theorem \ref{thm1}).
But at first, we need to recall the definition of quasi--periodic Euler functions $\overline E_n(x),$ see \cite{HKK2016} for some details on their properties and applications.
Let $n\in\mathbb N_0.$ If $0\leq x <1$, we write
\begin{equation}\label{ae-ft-1}
\overline E_n(x)=E_n(x),
\end{equation}
where $E_n(x)$ denotes the $n$th Euler polynomials.
We may extend this definition to all $x\in\mathbb{R}$ by setting
\begin{equation}\label{ae-ft}
\overline E_n(x+1)=-\overline E_n(x).
\end{equation}
In this way, $\overline E_{n}(x)$ becomes a quasi--periodic function on $\mathbb{R}$, and it has the following Fourier expansion
$$\overline E_n(x)=\frac{4n!}{\pi^{n+1}}\sum_{k=0}^\infty\frac{\sin((2k+1)\pi x-\frac12\pi n)}{(2k+1)^{n+1}},$$
where $0\leq x<1$ if $n\in\mathbb N$ and $0<x<1$ if $n=0$ (see \cite[p.~805, 23.1.16]{AS}).
Now using $\overline E_{n}(x)$ we state the Boole summation formula. It is obtained by Boole, but a similar one has been shown by Euler previously (see \cite[Theorem 1.2]{CanDa}, \cite[Lemma 2.1]{KMS} and \cite[24.17.1--2]{NIST}).
\begin{lemma}[Boole summation formula]\label{BSF}
Let $\alpha,\beta$ and $N$ be integers such that $\alpha<\beta$ and $N\in\mathbb N.$ If $f^{(N+1)}(t)$ is absolutely integrable over $[\alpha,\beta].$
Then
\begin{equation}\label{Boole}
\begin{aligned}
2\sum_{n=\alpha}^{\beta-1}(-1)^nf(n)& = \sum_{k=0}^{N}\frac{E_k(0)}{k!}\left((-1)^{\beta-1}f^{(k)}(\beta)+(-1)^{\alpha}f^{(k)}(\alpha) \right) \\
&\quad+
\frac1{N!}\int_{\alpha}^{\beta}\overline E_{N}(-t)f^{(N+1)}(t)dt,
\end{aligned}
\end{equation}
where $\overline E_{n}(t),n\in\mathbb N_0,$ is the $n$th quasi-periodic Euler functions.
\end{lemma}
From the above lemma we get the following asymptotic expansion. It should be noted that in the sense of asymptotic expansions,
the infinite series
here need to be understood as a sum over $k$
terminating at some arbitrarily chosen finite $N,$ together with an integral formula for the remainder term $R_{N+1}$
(see \cite[(7)]{EE}).
\begin{lemma}\label{lem1}
Let $C^{\infty}[0,1]$ be the space of infinitely differentiable real-valued functions on the interval $[0,1]$. For $f(t)\in C^{\infty}[0,1]$, we have the following
asymptotic expansion $$f(0)=\frac12\Delta(0)-\frac14\Delta'(0)-\frac12\sum_{k=2}^\infty\frac{(-1)^k}{k!}E_k(0)\Delta^{(k)}(0),$$
where $\Delta(t)=f(t+1)+f(t).$
\end{lemma}
\begin{proof}
By setting $\alpha=0$ and $\beta=1$ in (\ref{Boole}),
we have
\begin{equation}\label{Boole-2}
\begin{aligned}
f(0)& = \frac{1}{2}\sum_{k=0}^{N}\frac{E_k(0)}{k!}\left(f^{(k)}(0)+f^{(k)}(1) \right) \\
&\quad+
\frac1{2 N!}\int_{0}^{1}\overline E_{N}(-t)f^{(N+1)}(t)dt.
\end{aligned}
\end{equation}
Then noticing that $\Delta(t)=f(t+1)+f(t)$ and $E_{0}(0)=1$, $E_{1}(0)=-\frac{1}{2}$, $E_{k}(0)=0$ if any even $k\geq2$ (e.g. \cite[p. 5, Corollary 1.1]{Sun}),
we get
\begin{equation}\label{Boole-2}
\begin{aligned}
f(0)=\frac12\Delta(0)-\frac14\Delta'(0)-\frac12\sum_{k=2}^{N}\frac{(-1)^k}{k!}E_k(0)\Delta^{(k)}(0)+R_{N+1},
\end{aligned}
\end{equation}
where $$R_{N+1}=\frac1{2N!}\int_{0}^{1}\overline E_{N}(-t)f^{(N+1)}(t)dt.$$
Finally in considering the above remark,
we can state the asymptotic expansion
$$f(0)=\frac12\Delta(0)-\frac14\Delta'(0)-\frac12\sum_{k=2}^\infty\frac{(-1)^k}{k!}E_k(0)\Delta^{(k)}(0),$$
which is the desired result.
\end{proof}
By Lemma \ref{lem1} we have the following asymptotic series expansion of $\zeta_{E}(z,q)$ (see Section \ref{proofs} for its proof).
\begin{theorem}\label{thm1}
For all $z\in\mathbb C$ and $q\neq0,-1,-2,\ldots,$ we have the following
asymptotic expansion
\begin{equation}\label{main1} \zeta_E(z,q)=\frac12 q^{-z}+\frac14zq^{-z-1}-\frac12\sum_{k=2}^\infty\frac1{k!}E_k(0)(z)_kq^{-z-k},\end{equation}
as $q\to\infty$, where
$$(z)_k=z(z+1)\cdots(z+k-1)=\frac{\Gamma(z+k)}{\Gamma(z)}$$
is Pochhammer's symbol (rising factorial function). Here $\Gamma(z)$ is the classical Euler gamma function.
\end{theorem}
\begin{remark} We refer to V.A. Zorich's book \cite[p. 593, Definition 2]{zorich} for general concept of asymptotic expansions.\end{remark}
\begin{remark}
In a recent article \cite{HK2022}, by using the alternating Hurwitz zeta function
(\ref{E-zeta-def}),
we defined modified Stieltjes constants $\tilde\gamma_k(q)$ from
the following Taylor expansion,
\begin{equation}\label{l-s-con}
\zeta_E(z,q)=\sum_{k=0}^\infty\frac{(-1)^k\tilde\gamma_k(q)}{k!}(z-1)^k.
\end{equation}
Then by using the asymptotic expansion
(\ref{main1}), we obtained an asymptotic expansion of $\tilde\gamma_k(q),$ see \cite[Theorem 3.17]{HK2022} and its proof.
\end{remark}
Theorem \ref{thm1} implies the following known results on the special values of $\zeta_E(z,q)$ at $z=-n$ for $n\in\mathbb N_0.$
Its proof will be given in Section \ref{proofs}.
\begin{corollary}[{\cite[(3.8)]{WZ}}]\label{cor1}
For all $n\in\mathbb N_0$ we have
$$\zeta_E(-n,q)=\frac12 E_n(q),$$
where $E_n(q)$ denotes the $n$th Euler polynomials.
In particular, when $n=0,$ we obtain $\zeta_E(0,q)=1/2.$
\end{corollary}
\section{Main results} \label{2.2}
In this section, we state our main results, that is, we shall provide asymptotic expansions for the first and higher order derivatives of the alternating Hurwitz zeta function $\zeta_{E}(z,q)$.
Their proofs will be shown in Section \ref{proofs}.
\begin{theorem}\label{thm2}
For all $z\in\mathbb C$ and $q\neq0,-1,-2,\ldots,$ we have the following
asymptotic expansion $$\zeta_E'(z,q)=\frac14 q^{-z-1}-\zeta_E(z,q)\log q-\frac12\sum_{k=2}^\infty E_k(0)\sum_{j=0}^{k-1}\frac{(z)_j}{j!(k-j)}q^{-z-k},$$
as $q\to\infty$.\end{theorem}
Now setting $z=-n$ for integer $n\in\mathbb N_0$ in Theorem \ref{thm2}. Using Corollary \ref{cor1} and
$$(-n)_j=(-1)^j\frac{\Gamma(n+1)}{\Gamma(n+1-j)}=(-1)^j\frac{n!}{(n-j)!},$$
we obtain the asymptotic expansion
\begin{equation}\label{d-sum}
\begin{aligned}
\zeta_E'(-n,q)&=\frac14 q^{n-1}-\frac12E_n(q)\log q \\
&\quad-\frac12\sum_{k=2}^\infty E_k(0)\sum_{j=0}^{\min(n,k-1)}\binom nj\frac{(-1)^j}{k-j}q^{n-k},
\end{aligned}
\end{equation}
as $q\to\infty$.
In particular, setting $n=0$ and 1 in (\ref{d-sum}) respectively, we have the following asymptotic expansions
\begin{equation}\label{n=0}
\zeta_E'(0,q)=-\frac12\log q +\frac14q^{-1}-\frac12\sum_{k=2}^\infty\frac{E_k(0)}{k}q^{-k},
\end{equation}
\begin{equation}\label{n=1}
\zeta_E'(-1,q)=\frac14-\frac12\left(q-\frac12\right)\log q+\frac12\sum_{k=2}^\infty\frac{E_k(0)}{k(k-1)}q^{-(k-1)},
\end{equation}
as $q\to\infty$.
Furthermore, let $n\geq2.$ Then it is not difficult to see that the following double sum identity holds
\begin{equation}\label{s-sum}
\sum_{k=2}^\infty a(k)\sum_{j=0}^{\min(n,k-1)}b_n(k,j)=\sum_{k=2}^na(k)\sum_{j=0}^{k-1}b_n(k,j)+\sum_{k=n+1}^\infty a(k)\sum_{j=0}^nb_n(k,j).
\end{equation}
Using (\ref{d-sum}), (\ref{s-sum}) and the following identity (see \cite[p. 2833, (16)]{Ru})
\begin{equation}\label{ru16}
\sum_{j=0}^n\binom nj\frac{(-1)^j}{k-j}=\frac{(-1)^nn!}{k(k-1)\cdots(k-n)},
\end{equation}
we arrive at the following corollary.
\begin{corollary}\label{cor2}
For all $n\geq2$ we have the following
asymptotic expansion
$$\begin{aligned}
\zeta_E'(-n,q)&=\frac14 q^{n-1}-\frac12E_n(q)\log q \\
&\quad-\frac12\sum_{k=2}^n E_k(0)\sum_{j=0}^{k-1}\binom nj\frac{(-1)^j}{k-j}q^{n-k} \\
&\quad+(-1)^{n+1}\frac{n!}2\sum_{k=n+1}^\infty\frac{E_k(0)}{k(k-1)\cdots(k-n)}q^{n-k},
\end{aligned}$$
as $q\to\infty$.
\end{corollary}
In particular, setting $n=2$ and 3 in Corollary \ref{cor2}, we have the following expansions
\begin{equation}\label{n=2}
\zeta_E'(-2,q)=\frac14q-\frac12(q^2-q)\log q-\sum_{k=3}^\infty\frac{E_k(0)}{k(k-1)(k-2)}q^{-(k-2)},
\end{equation}
\begin{equation}\label{n=3}
\begin{aligned}
\zeta_E'(-3,q)&=-\frac{11}{48}+\frac14 q^2-\frac12\left(q^3-\frac32 q^2+\frac14\right)\log q \\
&\quad+3\sum_{k=4}^\infty\frac{E_k(0)}{k(k-1)(k-2)(k-3)}q^{-(k-3)},
\end{aligned}
\end{equation}
as $q\to\infty$,
since $E_2(0)=0$.
In the following, we will employ the usual convention that an empty sum is taken to be zero. For example,
if $m=1,$ then we understand that $\sum_{k=1}^{m-1}=0.$
Denote by $\log^jq=(\log q)^j$ for $j\geq1.$ Using Theorem \ref{thm1}, we may represent $\zeta_E^{(m)}(z,q)$, the higher order derivatives of the alternating Hurwitz zeta functions, in terms of the lower orders
from the following asymptotic expansion.
\begin{theorem}\label{thm3}
For all $z\in\mathbb C,q\neq0,-1,-2,\ldots$ and $m\geq2,$ we have the following
asymptotic expansion $$\zeta_E^{(m)}(z,q)=-\sum_{j=1}^m\binom mj\zeta_E^{(m-j)}(z,q)\log^j q-\frac12\sum_{k=2}^\infty c_{k,m}(z)q^{-z-k},$$
as $q\to\infty$, where
\begin{equation}\label{ckm}
c_{k,m}(z)=E_k(0)\left\{\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1}\frac1{j_1-j_2}
\cdots\sum_{j_m=0}^{j_{m-1}-1}\frac{(z)_{j_m}}{j_m!(j_{m-1}-j_m)}\right\}.
\end{equation}
\end{theorem}
If we restrict ourselves to the particular values $z=-n,n\in\mathbb N_0,$ then we obtain the following result.
\begin{corollary}\label{cor3}
For all $n\in\mathbb N_0$ and $m\geq2,$ we have the following
asymptotic expansion $$\zeta_E^{(m)}(-n,q)=-\sum_{j=1}^m\binom mj\zeta_E^{(m-j)}(-n,q)\log^j q-\frac12\sum_{k=2}^\infty c_{k,m}(-n)q^{n-k},$$
as $q\to\infty$, where
$$c_{k,m}(-n)=E_k(0)\left\{\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1}\frac1{j_1-j_2}
\cdots\sum_{j_m=0}^{\min(n,j_{m-1}-1)}\binom{n}{j_m}\frac{(-1)^{j_m}}{j_{m-1}-j_m}\right\}.$$
\end{corollary}
\begin{remark}
We remark here that the formula in Theorem \ref{thm2} for $\zeta_E'(z,q)$ cannot be obtained from Theorem \ref{thm3} simply by setting $m=1$ as
it contains an additional term in $q^{-z-1}.$
\end{remark}
\section{Proofs of the results}\label{proofs}
In this section, we prove Theorems \ref{thm1}, Corollary \ref{cor1}, Theorems \ref{thm2} and \ref{thm3} respectively.
\subsection*{Proof of Theorem \ref{thm1}.}
We give a proof of Theorem \ref{thm1} using Proposition \ref{lem1}.
For $t\in\mathbb{R}$, put $f(t)=\zeta_E(z,q+t)$ and then from (\ref{J-1}) we immediately have
\begin{equation}\label{delta}
\begin{aligned}
\Delta(0)&=f(1)+f(0)=q^{-z}, \\
\Delta'(0)&=\frac{d}{dt}\left[f(t+1)+f(t)\right]_{t=0} \\
&=\frac{d}{dt}\left[\zeta_E(z,q+t+1)+\zeta_E(z,q+t)\right]_{t=0} \\
&=\frac{d}{dt}\left[(q+t)^{-z}\right]_{t=0} \\
&=-zq^{-z-1}
\end{aligned}
\end{equation}
and \begin{equation}\label{delta1}
\Delta^{(k)}(0)=(-1)^k(z)_kq^{-z-k},
\end{equation}
where $k\in\mathbb N_0.$
Substituting (\ref{delta}) and (\ref{delta1}) into Proposition \ref{lem1}, we complete the proof.
\subsection*{Proof of Corollary \ref{cor1}.}
By putting $z=-n,n\in\mathbb N_0$ in Theorem \ref{thm1}, we have the following formula for $\zeta_E(-n,q)$:
$$\begin{aligned}
\zeta_E(-n,q)&=\frac12 q^{n}-\frac14nq^{n-1}-\frac12\sum_{k=2}^\infty\frac1{k!}E_k(0)(-n)_kq^{n-k} \\
&=\frac12 q^{n}-\frac14nq^{n-1}+\frac12\sum_{k=2}^\infty(-1)^{k+1}\binom nkE_k(0)q^{n-k} \\
&=\frac12 q^{n}-\frac14nq^{n-1}+\frac12\sum_{k=2}^n\binom nkE_k(0)q^{n-k} \\
&\quad(\text{by using $E_{k}(0)=0$ if any even $k\geq2$}) \\
&=\frac12\sum_{k=0}^n\binom nkE_k(0)q^{n-k},
\end{aligned}$$
where we used $E_0(0)=1$ and $E_1(0)=-1/2.$ The result follows by noticing that $$E_n(q)=\sum_{k=0}^n\binom nkE_k(0)q^{n-k}$$ (see \cite[p. 804, 23.1.7]{AS} and \cite[p. 3, (1.4)]{Sun}).
\subsection*{Proof of Theorem \ref{thm2}.}
To prove this, we need the following lemma.
\begin{lemma}\label{lem2}
For any nonnegative integer $k,$ denote by
$$\binom {z}{k}=\frac{z(z-1)\cdots(z-k+1)}{k!}.$$
Then we have
\begin{equation}\label{first}
\frac{d}{dz}\binom{-z}{k}=\sum_{j=0}^{k-1}\frac{(-1)^{k-j}}{k-j}\binom{-z}{j}
\end{equation}
and
\begin{equation}\label{second}
\frac{d}{dz}\left\{\frac{(z)_k}{k!}\right\}=\sum_{j=0}^{k-1}\frac{(z)_j}{j!(k-j)}.
\end{equation}
\end{lemma}
For the reader’s convenience, we briefly review the proof of Lemma \ref{lem2}.
\begin{proof}[Proof of Lemma \ref{lem2}]
(\ref{first}) follows directly from the induction.
The case $k=1$ is clear. It is well-known that
$$\binom{-z}{k+1}=\binom{-z-1}{k}+\binom{-z-1}{k+1}.$$
Hence, by differentiating it with respect to $z,$ we obtain
\begin{equation}\label{binom}
\begin{aligned}
\frac{d}{dz}\binom{-z}{k+1}&=\frac{d}{dz}\left\{\binom{-z-1}{k}+\binom{-z-1}{k+1}\right\} \\
&=\sum_{j=0}^{k-1}\frac{(-1)^{k-j}}{k-j}\binom{-z-1}{j}+
\sum_{j=0}^{k}\frac{(-1)^{k-j+1}}{k-j+1}\binom{-z-1}{j}.
\end{aligned}
\end{equation}
Since
$$\begin{aligned}
\sum_{j=0}^{k}\frac{(-1)^{k-j+1}}{k-j+1}\binom{-z-1}{j}
&=\sum_{j=0}^{k-1}\frac{(-1)^{k-j}}{k-j}\binom{-z-1}{j+1}+\frac{(-1)^{k+1}}{k+1},
\end{aligned}$$
(\ref{binom}) becomes
\begin{equation}\label{binom2}
\begin{aligned}
\frac{d}{dz}\binom{-z}{k+1}
&=\sum_{j=0}^{k-1}\frac{(-1)^{k-j}}{k-j}\left\{\binom{-z-1}{j}+\binom{-z-1}{j+1}\right\}+\frac{(-1)^{k+1}}{k+1} \\
&=\sum_{j=0}^{k-1}\frac{(-1)^{k-j}}{k-j}\binom{-z}{j+1}+\frac{(-1)^{k+1}}{k+1} \\
&=\sum_{j=1}^{k}\frac{(-1)^{k-j+1}}{k-j+1}\binom{-z}{j}+\frac{(-1)^{k+1}}{k+1} \\
&=\sum_{j=0}^{k}\frac{(-1)^{k-j+1}}{k-j+1}\binom{-z}{j}.
\end{aligned}
\end{equation}
Thus (\ref{first}) follows.
To see (\ref{second}), note that
$$\binom{-z}{k}=\frac{(-1)^k(z)_k}{k!}$$
and it then follows directly from the first part.
\end{proof}
Now we are at the position to prove Theorem \ref{thm2}.
Setting $$f(t)=\zeta_E'(z,q+t)=\frac{\partial}{\partial z}\zeta_E(z,q+t).$$
From (\ref{J-1}) we immediately have
\begin{equation}\label{zeta-p-1}
\begin{aligned}
\Delta(t)&=\frac{\partial}{\partial z}\left[\zeta_E(z,q+t+1)+\zeta_E(z,q+t)\right] \\
&=\frac{\partial}{\partial z}(q+t)^{-z} \\
&=\frac{\partial}{\partial z}\left[e^{-z\log(q+t)}\right] \\
&=-(q+t)^{-z}\log(q+t)
\end{aligned}
\end{equation}
and setting $t=0$ in (\ref{zeta-p-1}), we find that
\begin{equation}\label{zeta-p-2}
\begin{aligned}
\Delta(0)=-q^{-z}\log q.
\end{aligned}
\end{equation}
By (\ref{zeta-p-1}), we have
\begin{equation}\label{zeta-p-3}
\begin{aligned}
\Delta'(t)=-(q+t)^{-z-1}+z(q+t)^{-z-1}\log(q+t).
\end{aligned}
\end{equation}
Setting $t=0$ in (\ref{zeta-p-3}), we obtain
\begin{equation}\label{zeta-p-4}
\begin{aligned}
\Delta'(0)=-q^{-z-1}+zq^{-z-1}\log q.
\end{aligned}
\end{equation}
Similarly,
\begin{equation}\label{zeta-p-5}
\begin{aligned}
\Delta^{(2)}(t)&=(z+1)(q+t)^{-z-2}+z(q+t)^{-z-2} \\
&\quad-z(z+1)(q+t)^{-z-2}\log(q+t)
\end{aligned}
\end{equation}
and
\begin{equation}\label{zeta-p-6}
\begin{aligned}
\Delta^{(3)}(t)&=-(z+1)(z+2)(q+t)^{-z-3}-z(z+2)(q+t)^{-z-3} \\
&\quad-z(z+1)(q+t)^{-z-3}+z(z+1)(z+2)(q+t)^{-z-3}\log(q+t).
\end{aligned}
\end{equation}
This suggests the following general formula for positive integer $k\geq2,$
\begin{equation}\label{zeta-p-k}
\begin{aligned}
\Delta^{(k)}(t)&=(-1)^k\sum_{j=0}^{k-1}\frac{(z)_k}{z+j}(q+t)^{-z-k} \\
&\quad+(-1)^{k+1}(z)_k(q+t)^{-z-k}\log(q+t).
\end{aligned}
\end{equation}
It is easy to see that
$$\sum_{j=0}^{k-1}\frac{(z)_k}{z+j}=\frac{d}{dz}[(z)_k],$$
so that
\begin{equation}\label{zeta-p-k-1}
\begin{aligned}
\Delta^{(k)}(t)
&=(-1)^k\frac{d}{dz}[(z)_k](q+t)^{-z-k} \\
&\quad+(-1)^{k+1}(z)_k(q+t)^{-z-k}\log(q+t) \\
&=(-1)^kk!\sum_{j=0}^{k-1}\frac{(z)_j}{j!(k-j)}(q+t)^{-z-k} \\
&\quad+(-1)^{k+1}(z)_k(q+t)^{-z-k}\log(q+t),
\end{aligned}
\end{equation}
the last equality follows from Lemma \ref{lem2}. Setting $t=0$ in (\ref{zeta-p-k-1}), we get
\begin{equation}\label{zeta-p-k-1-0}
\begin{aligned}
\Delta^{(k)}(0)
&=(-1)^kk!\sum_{j=0}^{k-1}\frac{(z)_j}{j!(k-j)}q^{-z-k} +(-1)^{k+1}(z)_kq^{-z-k}\log q.
\end{aligned}
\end{equation}
From (\ref{zeta-p-2}), (\ref{zeta-p-4}) and (\ref{zeta-p-k-1-0}), Proposition \ref{lem1} immediately implies that
\begin{equation}\label{zeta-fi-dir}
\begin{aligned}
\zeta_E'(z,q)&=-\frac12q^{-z}\log q+\frac14\left( q^{-z-1}- zq^{-z-1}\log q\right) \\
&\quad-\frac12\sum_{k=2}^\infty E_k(0)\sum_{j=0}^{k-1}\frac{(z)_j}{j!(k-j)}q^{-z-k} \\
&\quad+\frac12\sum_{k=2}^\infty\frac1{k!}E_k(0)(z)_kq^{-z-k}\log q.
\end{aligned}
\end{equation}
From Theorem \ref{thm1}, we have
\begin{equation}\label{thm1-eq}
\begin{aligned}
\frac12\sum_{k=2}^\infty\frac1{k!}E_k(0)(z)_kq^{-z-k}\log q
=\left(\frac12 q^{-z}+\frac14 z q^{-z-1}-\zeta_E(z,q)\right)\log q.
\end{aligned}
\end{equation}
Substituting (\ref{thm1-eq}) into (\ref{zeta-fi-dir}), we obtain
$$\zeta_E'(z,q)=\frac14 q^{-z-1}-\zeta_E(z,q)\log q-\frac12\sum_{k=2}^\infty E_k(0)\sum_{j=0}^{k-1}\frac{(z)_j}{j!(k-j)}q^{-z-k},$$
which completes the proof.
\subsection*{Proof of Theorem \ref{thm3}.}
First, we extend the proof of Theorem \ref{thm2} to obtain an asymptotic expansion representing the higher order derivatives of
the alternating Hurwitz (or Hurwitz-type Euler) zeta function in terms of the lower orders.
Recall our notation
\begin{equation}\label{any-der}
\zeta_{E}^{(m)}(z,q)\equiv\frac{\partial^m}{\partial z^m}\zeta_E(z,q)\quad (m\geq2).
\end{equation}
Now, repeating, for $\zeta_E''(z,q),$ the same procedure just used for $\zeta_E'(z,q)$ in the proof of Theorem \ref{thm2}, we arrive at the following equality \begin{equation}\label{thm1-eq-re}
\begin{aligned}
\zeta_E''(z,q)&=-\frac14 q^{-z-1}\log q-\zeta_E'(z,q)\log q \\
&\quad+\frac12\sum_{k=2}^\infty E_k(0)\sum_{j_1=0}^{k-1}\frac{(z)_{j_1}}{j_1!(k-j_1)}q^{-z-k}\log q \\
&\quad-\frac12\sum_{k=2}^\infty E_k(0)\sum_{j_1=0}^{k-1}\frac{1}{k-j_1}
\sum_{j_2=0}^{j_1-1}\frac{(z)_{j_2}}{j_2!(j_1-j_2)}q^{-z-k}.
\end{aligned}
\end{equation}
In fact, the above identity (\ref{thm1-eq-re}) follows directly from Theorem \ref{thm2} and Lemme \ref{lem2}.
From Theorem \ref{thm2}, we have
\begin{equation}\label{thm2-eq}
\begin{aligned}
\frac12\sum_{k=2}^\infty E_k(0)&\sum_{j_1=0}^{k-1}\frac{(z)_{j_1}}{j_1!(k-j_1)}q^{-z-k}\log q \\
&=\left(\frac14 q^{-z-1}-\zeta_E(z,q)\log q-\zeta_E'(z,q)\right)\log q.
\end{aligned}
\end{equation}
Substituting (\ref{thm2-eq}) into (\ref{thm1-eq-re}), we get
\begin{equation}\label{two-comp}
\begin{aligned}
\zeta_E''(z,q)&=-2\zeta_E'(z,q)\log q -\zeta_E(z,q)\log^2 q \\
&\quad-\frac12\sum_{k=2}^\infty E_k(0)\sum_{j_1=0}^{k-1}\frac{1}{k-j_1}
\sum_{j_2=0}^{j_1-1}\frac{(z)_{j_2}}{j_2!(j_1-j_2)}q^{-z-k}.
\end{aligned}
\end{equation}
With some additional effort, the following general recurrence can be derived, which yields an asymptotic expansion for a higher order derivative
of the alternating Hurwitz zeta function in terms of the lower orders:
\begin{equation}\label{m-der-pf}
\zeta_E^{(m)}(z,q)=-\sum_{j=1}^m\binom mj\zeta_E^{(m-j)}(z,q)\log^j q-\frac12\sum_{k=2}^\infty c_{k,m}(z)q^{-z-k},
\end{equation}
where $z\in\mathbb C,q\neq0,-1,-2,\ldots,$ $m\geq2,$ and
$c_{k,m}(z)$ being given by (\ref{ckm}).
The proof of (\ref{m-der-pf}) proceeds by an induction on $m.$
The case $m=2$ is clear by (\ref{two-comp}).
Also it is easily seen that
\begin{equation}\label{pf-1}
\begin{aligned}
\frac{\partial}{\partial z}
c_{k,m}(z)
&=E_k(0)\left\{\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1}\frac1{j_1-j_2}
\cdots\sum_{j_m=0}^{j_{m-1}-1}\frac1{j_{m-1}-j_m}\frac{\partial}{\partial z}\left[\frac{(z)_{j_m}}{j_m!}\right]\right\} \\
&\quad(\text{recall (\ref{ckm})}) \\
&=E_k(0)\left\{\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1}\frac1{j_1-j_2}
\cdots\sum_{j_{m+1}=0}^{j_{m}-1}\frac{(z)_{j_{m+1}}}{j_{m+1}!(j_{m}-j_{m+1})}\right\} \\
&\quad(\text{by Lemma \ref{lem2}}) \\
&=c_{k,m+1}(z)
\end{aligned}
\end{equation}
and
\begin{equation}\label{pf-2}
\begin{aligned}
\frac12\sum_{k=2}^\infty c_{k,m}(z)q^{-z-k}\log q=-\sum_{j=1}^m\binom mj\zeta_E^{(m-j)}(z,q)\log^{j+1} q
-\zeta_E^{(m)}(z,q)\log q,
\end{aligned}
\end{equation}
where we used the inductive hypothesis (see (\ref{m-der-pf})).
By using (\ref{m-der-pf}), (\ref{pf-1}) and (\ref{pf-2}), we obatin
$$
\begin{aligned}
\zeta_E^{(m+1)}(z,q)&=\frac{\partial}{\partial z}\zeta_E^{(m)}(z,q)\\&=-\frac{\partial}{\partial z}\left\{\sum_{j=1}^m\binom mj\zeta_E^{(m-j)}(z,q)\log^j q+\frac12\sum_{k=2}^\infty c_{k,m}(z)q^{-z-k}\right\} \\
&\quad(\text{by the inductive hypothesis (\ref{m-der-pf})}) \\
&=-\sum_{j=1}^m\binom mj\zeta_E^{(m-j+1)}(z,q)\log^j q +\frac12\sum_{k=2}^\infty c_{k,m}(z)q^{-z-k}\log q \\
&\quad-\frac12\sum_{k=2}^\infty \frac{\partial}{\partial z}\left[c_{k,m}(z)\right]q^{-z-k}
\\
&=-\sum_{j=1}^m\binom mj\zeta_E^{(m-j+1)}(z,q)\log^j q - \sum_{j=1}^m\binom mj\zeta_E^{(m-j)}(z,q)\log^{j+1} q \\
&\quad-\zeta_E^{(m)}(z,q)\log q-\frac12\sum_{k=2}^\infty \frac{\partial}{\partial z}\left[c_{k,m}(z)\right]q^{-z-k} \\
&\quad(\text{by (\ref{pf-2})}) \\
&=-\binom{m+1}1\zeta_E^{(m)}(z,q)\log q \\
&\quad-\sum_{j=1}^{m-1}\left\{\binom{m}{j+1} +\binom m{j}\right\} \zeta_E^{(m-j)}(z,q)\log^{j+1}q \\
&\quad -\binom{m+1}{m+1}\zeta_E(z,q)\log^{m+1} q
-\frac12\sum_{k=2}^\infty c_{k,m+1}(z)q^{-z-k} \\
&\quad(\text{by (\ref{pf-1})})\\
&=-\sum_{j=1}^{m+1}\binom {m+1}j\zeta_E^{(m-j+1)}(z,q)\log^j q-\frac12\sum_{k=2}^\infty c_{k,m+1}(z)q^{-z-k}
\end{aligned}
$$
and this completes the inductive argument.
\section{Concluding remarks}
In this section, as an example, we shall write the asymptotic series for the second derivative of $\zeta_{E}(z,q)$ at non-positive values of $z$ explicitly.
For an integer $n\in\mathbb N_0,$ it follows from Corollary \ref{cor3} with $m=2$ that
\begin{equation}\label{k=2}
\begin{aligned}
\zeta_E^{''}(-n,q)
&=-2\zeta_E'(-n,q)\log q-\zeta_E(-n,q)\log^2q \\
&\quad-\frac12\sum_{k=2}^\infty E_k(0)\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{\min(n,j_1-1)} \binom{n}{j_2}
\frac{(-1)^{j_2}}{j_1-j_2}q^{n-k} .
\end{aligned}
\end{equation}
Now setting $n=0,1$ in (\ref{k=2}) respectively, we have
\begin{equation}\label{k=2 with n=0}
\begin{aligned}
\zeta_E^{''}(0,q)
&=\frac12\log^2 q -\frac12q^{-1}\log q \\
&\quad+2\sum_{k=3}^\infty E_k(0)\left[\frac{\log q}{k}-\frac12\sum_{j=1}^{k-1}\frac1{j(k-j)}\right]q^{-k} ,
\end{aligned}
\end{equation}
and
\begin{equation}\label{k=2 with n=1}
\begin{aligned}
\zeta_E^{''}(-1,q)
&=\frac12 q\log^2 q-\frac12\left(\log q+\frac12\log^2 q\right) \\
&\quad+\frac12\sum_{k=3}^\infty E_k(0)\left[\sum_{j=2}^{k-1}\frac1{(k-j)j(j-1)}-\frac1{k-1}-\frac{2\log q}{k(k-1)} \right]q^{-k+1} .
\end{aligned}
\end{equation}
For $n\geq2,$ separating the last term in the sum (\ref{k=2}), we have
\begin{equation}\label{k=2-ne}
\begin{aligned}
\zeta_E^{''}(-n,q)
&=-2\zeta_E'(-n,q)\log q-\zeta_E(-n,q)\log^2q \\
&\quad-\frac12\sum_{k=2}^n E_k(0)\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1} \binom{n}{j_2}
\frac{(-1)^{j_2}}{j_1-j_2}q^{n-k} \\
&\quad-\frac12\sum_{k=n+1}^\infty E_k(0)\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{\min(n,j_1-1)} \binom{n}{j_2}
\frac{(-1)^{j_2}}{j_1-j_2}q^{n-k}.
\end{aligned}
\end{equation}
Thus, using (\ref{s-sum}) and (\ref{ru16}), we can rewrite (\ref{k=2-ne}) as
\begin{equation}\label{k=2-ne1}
\begin{aligned}
\zeta_E^{''}(-n,q)
&=-2\zeta_E'(-n,q)\log q-\zeta_E(-n,q)\log^2q \\
&\quad-\frac12\sum_{k=2}^n E_k(0)\sum_{j_1=0}^{k-1}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1} \binom{n}{j_2}
\frac{(-1)^{j_2}}{j_1-j_2}q^{n-k} \\
&\quad-\frac12\sum_{k=n+1}^\infty E_k(0)\sum_{j_1=0}^{n}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1} \binom{n}{j_2}
\frac{(-1)^{j_2}}{j_1-j_2}q^{n-k} \\
&\quad-\frac12\sum_{k=n+1}^\infty E_k(0)\sum_{j_1=n+1}^{k-1}\frac{(-1)^nn!}{(k-j_1)j_1(j_1-1)\cdots(j_1-n)}
q^{n-k} .
\end{aligned}
\end{equation}
When $n\geq2,$ substituting the results of Corollaries \ref{cor1} and \ref{cor2} into (\ref{k=2-ne1}) it is seen that
\begin{equation}\label{k=2-fin}
\begin{aligned}
&\zeta_E^{''}(-n,q) \\
&=\frac12 E_n(q)\log^2 q-\frac12 q^{n-1}\log q \\
&\quad+\sum_{k=2}^n E_k(0)\sum_{j_1=0}^{k-1}
\left[\binom{n}{j_1}\frac{(-1)^{j_1}\log q}{k-j_1} -\frac1{2(k-j_1)}\sum_{j_2=0}^{j_1-1} \binom{n}{j_2}
\frac{(-1)^{j_2}}{j_1-j_2}\right]q^{n-k} \\
&\quad+\sum_{k=n+1}^\infty E_k(0)\left[
\frac{(-1)^nn!\log q}{k(k-1)\cdots(k-n)}
-\frac12\sum_{j_1=0}^{n}\frac1{k-j_1}\sum_{j_2=0}^{j_1-1} \binom{n}{j_2}
\frac{(-1)^{j_2}}{j_1-j_2}\right. \\
&\qquad\qquad\qquad\qquad\left.-\frac12\sum_{j_1=n+1}^{k-1}\frac{(-1)^nn!}{(k-j_1)j_1(j_1-1)\cdots(j_1-n)}\right]
q^{n-k} .
\end{aligned}
\end{equation}
Below there is a list of the first few results for $n=2,3$ from (\ref{k=2-fin}):
\begin{equation}\label{k=2 with n=2}
\begin{aligned}
\zeta_E^{''}(-2,q)
&=\frac12q^2\log^2 q-\frac12(\log q+\log^2 q)q +\frac14\left(\frac{\log q}{3}-\frac32\right)q^{-1} \\
&\quad+\sum_{k=5}^\infty E_k(0)\left[
\frac{2\log q}{k(k-1)(k-2)}-\frac12\left(\frac1{k-1}+\frac{5}{2(k-2)}\right)\right. \\
&\quad\qquad\qquad\qquad-\left.\sum_{j=3}^{k-1}\frac1{(k-j)j(j-1)(j-2)} \right]q^{-k+2}
\end{aligned}
\end{equation}
and
\begin{equation}\label{k=2 with n=3}
\begin{aligned}
\zeta_E^{''}(-3,q)
&=\frac12q^3\log^2 q-\frac12\left(\frac32\log q^2+\log q\right)q^2
\\
&\quad +\left(\frac18\log^2 q +\frac{11}{24}\log q +\frac14\right) +\frac18\left(\frac15\log q+\frac16\right)q^{-2} \\
&\quad+\sum_{k=7}^\infty E_k(0)\left[
\sum_{j=4}^{k-1}\frac3{(k-j)j(j-1)(j-2)(j-3)} \right. \\
&\quad\qquad\qquad\qquad
-\frac{6\log q}{k(k-1)(k-2)(k-3)} \\
&\quad\qquad\qquad\qquad
\left. -\frac12\left(\frac1{k-1}-\frac5{2(k-2)}+\frac{11}{6(k-3)}\right)
\right]q^{-k+3} .
\end{aligned}
\end{equation}
|
\section{Introduction}
With the rapid increase of online social networks (OSNs) such as Facebook and Twitter, large amount of data is being generated daily. A valuable mining area of network data is composed when OSNs are modeled into nodes and edges. Identifying key nodes in such networks is the basis of major applications such as viral marketing \cite{jalili2017information}, controlling epidemic spreading \cite{wang2017vaccination}, and determining sources of misinformation \cite{azzimonti2018social}.
Designing centrality measures is a main approach to quantify node influence. Numerous centrality measures exploiting various properties of the network topology have been developed \cite{lu2016vital}. Information exploited can be either in the neighborhood of the node or concerning all the topological structure of the network. The former called local centrality measures are less computationally expensive as compared to the later called global centrality measures. However, local centrality measures, usually, aren't as much as accurate as global centrality measures. Recent works tend to combine both local and global measures \cite{sciarra2018change,ibnoulouafi2018m}.
Real-world OSNs often exhibit a community structure in which groups of nodes are closely connected to each other and sparsely connected to nodes in other communities \cite{girvan2002community, jebabli2015user}. Community structure has major implications on the dynamics of the network \cite{cherifi2019community}. To this end, researchers have taken classical centrality measures a step further to incorporate community structure information \cite{hwang2006bridging, ghalmane2019immunization, guimera2005functional, tulu2018identifying,gupta2015community, chakraborty2016immunization, kumar2018efficient, ghalmane2019centrality}. Community-aware centrality measures can be divided into two groups. The former explicitly rely on the community structure. They incorporate information about the type of links in a community (intra-community links and inter-community links). The latter targets ``bridges" that lie between communities without extracting the community structure information.
As classical centrality measures neglect the community structure, this raises a key question. Do community-aware centrality measures provide distinctive information about the members within OSNs when compared to classical centrality measures? Previous works have studied the relationship between classical centrality measures \cite{li2015correlation, oldham2019consistency, shao2018rank, Landherr2010, grando2016analysis} and between classical and hierarchy measures \cite{rajeh2020interplay}.
Nonetheless, to our knowledge, there is no previous work on the relationship between classical and community-aware centrality measures on OSNs. To fill this gap, here, 5 classical and 5 community-aware centrality measures are used in a comparative evaluation involving 8 real-world OSN. The community structure of the networks is extracted using the Infomap \cite{rosvall2008maps} community detection algorithm. Then, Kendall's Tau correlation and RBO similarity are calculated on all the possible combinations between the classical and community-aware centrality measures. Two groups of community-aware centrality measures can be seen. The first group provides distinctive information when compared against classical centrality measures and is consistent across the networks under study. It includes Bridging centrality, Community Hub-Bridge, and Participation Coefficient. The second group shows varying correlation and similarity on networks. It includes Community-based Mediator and Number of Neighboring Communities.
The paper is organized as follows. Classical and community-aware centrality measures alongside basic definitions are provided in section \ref{sec:Def}. The datasets and tools are provided in section \ref{sec:mm}. Experimental results are discussed in section \ref{sec:res}. Finally, the conclusion and future works are provided in section \ref{sec:conc}.
\section{Preliminaries and Definitions}
\label{sec:Def}
In this section preliminaries and definitions used throughout the rest of the paper are given.
\begin{itemize}
\item Consider a undirected and unweighted OSN as $G(V,E)$ where $V$ is the set of nodes
and $E \subseteq V \times V$ is the set of edges and $N=|V|$ is the total size of the network.
Nodes represent individuals and edges represent
social links between these individuals. The semantics of the social links depend on the platform of the OSN.
\item Consider $A = (a_{i,j})$ as the adjacency matrix showing connectivity of the network
$G$ such that $a_{i,j}= 1$, if node $i$ is connected to node $j$ and $a_{i,j}= 0$, otherwise.
\item Let the neighborhood of any node $i$ be defined as the set $\mathcal{N}_p(i)={\{j \in V:
(i,j) \in E\}}$ at length $p$, where $p=1,2,...,D$. $D$ is the diameter of $G$. Accordingly, two
nodes are neighbors of order $A^p$ if there's a minimal path connecting them at $p$ steps.
\item Let $C$ be the set of communities $C=\{c_1, c_2, ..., c_k\}$. The intra-community links are obtained from the graph $G_l$ where all inter-community links of the nodes are removed.
The inter-community links are obtained from the graph $G_g$ where all intra-community links of the nodes are removed.
\end{itemize}
\subsection{Classical Centrality Measures}
Following are the definitions of the 5 most popular centrality measures used in the study.
\subsubsection{Degree Centrality} is simply the total number connections a node has in the network. It is defined as follows:
\begin{equation}
\alpha_{d}(i)=\sum_{j=1}^{N}a_{ij}
\end{equation}
where $a_{ij}$ is obtained from $A^{1}$, 1-step neighborhood ($p$=1).
\subsubsection{Betweenness Centrality} captures the number of times a node falls between the shortest paths linking other node pairs. It is defined as follows:
\begin{equation}
\alpha_b(i)=\sum_{s,t\neq i }{\frac{\sigma_{i}(s,t)}{\sigma(s,t)} }
\end{equation}
where $\sigma(s,t)$ is the number of shortest paths between nodes $s$ and $t$ and $\sigma_{i}(s,t)$ is the number of shortest paths between nodes $s$ and $t$ that pass through node $i$.
\subsubsection{Closeness Centrality} is inverse of sum of geodesic distances to every other node from a given node. It is defined as follows:
\begin{equation}
\alpha_c(i)=\frac{N-1}{\sum_{j=1}^{N-1}d(i,j)}
\end{equation}
where $d(i,j)$ is the shortest-path distance between node $i$ and $j$.
\subsubsection{Katz Centrality} is based on how many nodes a node is connected to and also to the connectivity of its neighbors . It is defined as follows:
\begin{equation}
\alpha_k(i)= \sum_{p=1} \sum_{j=1} s^p a^p_{ij}
\end{equation}
where $a^p_{ij}$ is the connectivity of node $i$ with respect to all the other nodes at $A^p$ and $s^p$ is the attenuation factor where $s \in$ [0,1].
\subsubsection{PageRank Centrality} quantifies a node's importance similarly to Katz centrality with an additional layer based on a random surfer. It is defined as follows:
\begin{equation}
\alpha_p(i)=\frac{1-d}{N} + d \sum_{j \in \mathcal{N}_1(i)} \frac{\alpha_p(j)}{k_j}
\end{equation}
where $\alpha_p(i)$ and $\alpha_p(j)$ are the PageRank centralities of node $i$ and node $j$, respectively, $\mathcal{N}_1(i)$ is the set of direct neighbors of node $i$, $k_j$ is the number of links from node $j$ to node $i$, and $d$ is the damping parameter where $d \in$ [0,1], set to 0.85 in the experiments.
\subsection{Community-aware Centrality Measures}
Following are the definitions of the 5 community-aware measures of centrality used:
\subsubsection{Number of Neighboring Communities (NNC) \cite{ghalmane2019immunization}} is based on the number of communities a node can reach in one hop. For a node in community $c_k \subset C$, it is defined as follows:
\begin{equation}
\beta_{NNC}(i) = \sum_{c_l \subset C \backslash c_k}^{N} \bigvee_{j \in c_l} a_{ij}
\end{equation}
where $\bigvee_{j \in c_l} a_{ij} = 1$ when node $i$ is connected to at least one node $j$ in community $c_l$.
\subsubsection{Community Hub-Bridge (CHB) \cite{ghalmane2019immunization}} assumes a node simultaneously can act as a hub and a bridge. It combines the intra-community and inter- community links by weighting the former with the community size and the latter with the number of neighboring communities. For a node in community $c_k \subset C$, it is defined as follows:
\begin{equation}
\beta_{CHB}(i) = h_i(c_k) + b_i(c_k)
\end{equation}
where hub influence is given by $h_i(c_k)= |c_k| \times k_i^{intra}$
and bridge influence is given by $b_i(c_k) = \beta_{NNC}(i) \times k_i^{inter}$.
\subsubsection{Participation Coefficient (PC) \cite{guimera2005functional}} is based on the intra-community and inter-community links distribution. The more the links of a node are distributed across different communities, the higher its participation coefficient. It is defined as follows:
\begin{equation}
\beta_{PC}(i) = 1 - \sum_{c=1}^{N_c}
\left(
\frac{k_{i,c}}{k_i^{tot}}
\right)^2
\end{equation}
where $N_c$ is the total number of communities, $k_{i,c}$ is the number of links node $i$ has in a given community $c$ (can be inter-community or intra-community links), and $k_i^{tot}$ is the total degree of node $i$.
\subsubsection{Community‐based Mediator (CBM) \cite{tulu2018identifying}} takes into consideration the intra-community and inter-community ratio of a node, then it incorporates a random walker and entropy based on the ratio of the different link types. It is defined as follows:
\begin{equation}
\beta_{CBM}(i) = H_i \times \frac{k_i^{tot}}{\sum_{i=1}^{N} k_i}
\end{equation}
where $H_i = [-\sum \rho_i^{intra} log(\rho_i^{intra})] + [- \sum \rho_i^{inter} log(\rho_i^{inter})]$ is the entropy of node $i$ based on its $\rho^{intra}$ and $\rho^{inter}$ which represent the density of the communities a node links to (either its community or external communities), $k_i^{tot}$ is the total degree of node $i$, and $\sum_{i=1}^{N} k_i$ is the total degrees in the network.
\subsubsection{Bridging Centrality (BC) \cite{hwang2006bridging}} extracts node bridges by using betweenness centrality and bridging coefficient. The bridging coefficient quantifies the proximity of a node to high degree nodes. It is defined as follows:
\begin{equation}
\beta_{BC}(i) = \alpha_b(i) \times \mathbb{B}(i)
\end{equation}
where $\alpha_b(i)$ is the classical betweenness centrality of node $i$
and $\mathbb{B}(i)= \frac{k_i^{-1}}{\sum_{j \in \mathcal{N}_1(i)} k_j^{-1}}$ is the bridging coefficient where $\mathcal{N}_1(i)$ is the set of direct neighbors of node $i$.
\section{Datasets and Materials}
\label{sec:mm}
In this section, the 8 real-world online social networks are briefly discussed, alongside the tools applied. Table \ref{tab:table1} reports the basic topological characteristics of the networks. Note that the mixing parameter $\mu$ is defined as the proportion of inter-community links to the total
links in a given network. It is calculated after the community structure is uncovered by the community detection algorithm.
\subsection{Data}
\subsubsection{FB Ego} this network (ego-facebook) is collected from participants using Facebook. Nodes represent users on Facebook and edges represent online friendships \cite{rossi2015network}.
\subsubsection{FB Princeton} this network (socfb-Princeton12) is collected from Facebook among students at Princeton University. Nodes represent users on Facebook and edges represent online friendships \cite{rossi2015network}.
\subsubsection{FB Caltech} this network (socfb-Caltech36) is collected from the Facebook application among students at Caltech University. Nodes represent users on Facebook and edges represent online friendships \cite{rossi2015network}.
\subsubsection{FB Politician Pages} this network (fb-pages-politician) is collected from Facebook pages. Nodes represent politician pages from different countries created on Facebook and edges represent mutual likes among them \cite{rossi2015network}.
\subsubsection{Retweetes Copenhagen} this network (rt-twitter-copen) is collected from Twitter. Nodes are users on Twitter tweeting in parallel to the United Nations conference in Copenhagen about climate change and edges represent retweets among the users \cite{rossi2015network}.
\subsubsection{DeezerEU} this network (deezer\_europe) is obtained form Deezer, a platform for music streaming. Nodes are Deezer European users and edges represent online friendships \cite{rozemberczki2020characteristic}.
\subsubsection{Hamsterster} this network (petster-friendships-hamster) is obtained from an online social pet network hamsterster.com. Nodes represent users and edges represent friendships among them.
\cite{kunegis2014handbook}.
\subsubsection{PGP} this network (arenas-pgp) is obtained from the web of trust. Nodes are users using the Pretty Good Privacy (PGP) algorithm and edges represent secure information sharing among them \cite{kunegis2014handbook}.
\begin{table*}[t!]
\centering
\caption{Basic topological properties of the real-world networks. \textit{N} is the total numbers of nodes. E is the number of edges. $<k>$ is the average degree. $<d>$ is the average shortest path. $\nu$ is the density. $\zeta$ is the transitivity (also called global clustering coefficient). $k_{nn}(k)$ is the assortativity (also called degree correlation coefficient). $Q$ is the modularity. $\mu$ is the mixing parameter. * indicates the topological properties of the largest connected component of the network in case it is disconnected.
}
\label{tab:table1}
\begin{tabular}{lcccccccccc}
\hline
Network & $N$ & $E$ & $<k>$ & $<d>$ & $\nu$ & $\zeta$ & $k_{nn}(k)$ & $Q$ & $\mu$\\
\hline
Retweets Copenhagen & 761 & 1,029 & 2.70 & 5.35 & 0.003 & 0.060 & -0.099 & 0.695 &
0.287 \\
FB Caltech* & 762 & 16,651 & 43.70 & 2.23 & 0.057 & 0.291 & -0.066 & 0.389 & 0.410\\
Hamsterster* & 1,788 & 12,476 & 13.49 & 3.45 & 0.007 & 0.090 & -0.088 & 0.391 & 0.298 \\
FB Ego & 4,039 & 88,234 & 43.69 & 3.69 & 0.010 & 0.519 & 0.063 & 0.814 & 0.077\\
FB Politician Pages & 5,908 & 41,729 & 14.12 & 4.66 & 0.002 & 0.301 & 0.018 & 0.836 &
0.111 \\
FB Princeton* & 6,575 & 293,307 & 89.21 & 2.67 & 0.013 & 0.163 & 0.090 & 0.417 & 0.365 \\
PGP & 10,680 & 24,316 & 4.55 & 7.48 & 0.0004 & 0.378 & 0.238 & 0.813 & 0.172 \\
DeezerEU & 28,281 & 92,752 & 6.55 & 6.44 & 0.002 & 0.095 & 0.104 & 0.565 & 0.429\\
\hline
\end{tabular}
\end{table*}
\subsection{Tools}
\subsubsection{Kendall's Tau Correlation} is used to assess the relationship for all possible combinations between classical and community-aware centrality measures. Assume that $R(\alpha)$ and $R(\beta)$ are the ranking lists of a classical centrality and a community-aware centrality, respectively. The correlation value resulted [-1,+1] reveals the degree of ordinal association between the two given sets of ranks. If $R(\alpha_i) > R(\alpha_j)$ and $R(\beta_i) > R(\beta_j)$ or $R(\alpha_i) < R(\alpha_j)$ and $R(\beta_i) < R(\beta_j)$, node pair ($i,j$) is concordant. If $R(\alpha_i) > R(\alpha_j)$ and $R(\beta_i) < R(\beta_j)$ or $R(\alpha_i) < R(\alpha_j)$ and $R(\beta_i) > R(\beta_j)$, node pair ($i,j$) is discordant. If $R(\alpha_i) = R(\alpha_j)$ and/or $R(\beta_i) = R(\beta_j)$, node pair ($i,j$) is neither concordant nor discordant. It is defined as follows:
\begin{equation}
\tau_b(R(\alpha),R(\beta))=\frac{n_{c}-n_{d}}{\sqrt{(n_{c}+n_{disc}+u)(N_{c}+N_{d}+v)}}
\label{eq2}
\end{equation}
where $n_{c}$ and $n_{d}$ stand for the number of concordant and discordant pairs, respectively, and $u$ and $v$ hold the number of tied pairs in sets $R(\alpha)$ and $R(\beta)$, respectively.
\subsubsection{Rank-Biased Overlap (RBO) \cite{webber2010similarity}}
is capable of placing more emphasis on the top nodes between the two ranked lists $R(\alpha)$ and $R(\beta)$ of classical and community-aware centrality measures. Its value ranges between [0,1]. It is defined as follows:
\begin{equation}
RBO(R(\alpha), R(\beta)) = (1-p){\sum_{d=1}^{\infty} p^{(d-1)}} \frac{|R(\alpha_{d}) \cap
R(\beta_{d})|}{d}
\end{equation}
where $p$ dictates ``user persistence'' and the weight to the top ranks, $d$ is the depth reached on sets $R(\alpha)$ and $R(\beta)$, and $|R(\alpha_{d}) \cap R(\beta_{d})|/{d}$ is the proportion of the similarity overlap at depth $d$. Note that $p$ is set to 0.9 in the experiments.
\subsubsection{Infomap Community Detection Algorithm \cite{rosvall2008maps}} is based on compression of information. The idea is that a random walker on a network is likely to stay longer inside a given community and shorter outside communities. Accordingly, using Huffman coding, each community is defined by a unique codeword and nodes inside communities are defined by other codewords that can be reused in different communities. The optimization algorithm minimizes the coding resulted by the path of the random walker, achieving a concise map of the community structure.
\section{Experimental Results}
\label{sec:res}
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=1\linewidth, height=2.6 in]{CorrelationsWithoutMU.eps}
\end{center}
\caption{{Heatmaps of the Kendall's Tau correlation ($\tau_b$) of real-world networks across the various combinations between classical ($\alpha$) and community-aware ($\beta$) centrality measures. The classical centrality measures are: $\alpha_d$ = Degree, $\alpha_b$ = Betweenness, $\alpha_c$ = Closeness, $\alpha_k$ = Katz, $\alpha_p$ = PageRank. The
community-aware centrality measures are: $\beta_{BC}$ = Bridging centrality, $\beta_{CHB}$ = Community Hub-Bridge, $\beta_{PC}$ = Participation Coefficient, $\beta_{CBM}$ = Community-based Mediator, $\beta_{NNC}$ = Number of Neighboring Communities.}}
\label{fig:corrHeatmapsMain}
\vspace{-0.4cm}
\end{figure*}
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=1\linewidth, height=2.6 in]{SimilarityWithoutMU.eps}
\end{center}
\caption{{Heatmaps of the RBO similarity at $p$=0.9 of real-world networks across the various combinations between classical ($\alpha$) and community-aware ($\beta$) centrality measures.The classical centrality measures are: $\alpha_d$ = Degree, $\alpha_b$ = Betweenness, $\alpha_c$ = Closeness, $\alpha_k$ = Katz, $\alpha_p$ = PageRank. The
community-aware centrality measures are: $\beta_{BC}$ = Bridging centrality, $\beta_{CHB}$ = Community Hub-Bridge, $\beta_{PC}$ = Participation Coefficient, $\beta_{CBM}$ = Community-based Mediator, $\beta_{NNC}$ = Number of Neighboring Communities.}}
\label{fig:simHeatmapsMain}
\vspace{-0.4cm}
\end{figure*}
In this section the results of the experiments performed on the real-world networks are reported. The first set of experiments involves calculating Kendall's Tau correlation coefficient for all possible combinations between classical and community-aware centrality measures. The second experiment involves calculating the RBO similarity across all the combinations.
\subsection{Correlation Analysis}
Kendall Tau's correlation is applied on each network given all of the possible combinations between the 5 classical and 5 community-aware centrality measures. The 25 different combinations of the Kendall Tau's correlation for the 8 OSNs are reported in figure \ref{fig:corrHeatmapsMain}. The Kendall's Tau values range from -0.17 to 0.83. Low correlation from -0.17 to 0.3 is characterized by the dark purple color of the heatmaps. Medium correlation from 0.3 to 0.6 is characterized by the fuchsia color. High correlation above 0.6 is characterized by the light pink color.
Networks' heatmaps are arranged from low correlation (FB Ego) to medium-high (DeezerEU) correlation between classical and community-aware centrality measures. Heatmaps show that there are different behaviors among the community-aware centrality measures under study when they are compared to classical centrality measures. Specifically, Bridging centrality
($\beta_{BC}$), Community Hub-Bridge ($\beta_{CHB}$) and Participation Coefficient ($\beta_{PC}$) show consistency in their low correlation with classical centrality measures. On the other hand, Community-based Mediator ($\beta_{CBM}$) and Number of Neighboring Communities ($\beta_{NNC}$) vary across networks. In FB Ego, FB Politician Pages, and PGP, the correlation values are in the low to medium range, while in Hamsterster, FB Princeton, FB Caltech, and DeezerEU they are in the medium to high range.
Note that in Retweets Copenhagen network, the community-aware centrality measures show high correlation with the classical centrality measures degree and betweenness but low to medium correlation with the others. This is with the exception of Community Hub-Bridge ($\beta_{CHB}$) which shows low correlation with all classical centrality measures.
This experiment aims to answer the main research question, that is, do community-aware centrality measures provide distinctive information about the members within OSNs when compared to classical centrality measures? Results show that community-aware centrality measures indeed provide different information from that of classical centrality measures to the members within OSNs. Nonetheless, Bridging centrality ($\beta_{BC}$), Community Hub-Bridge ($\beta_{CHB}$), and Participation Coefficient ($\beta_{PC}$) show consistency in providing distinctive information to the members of 8 networks at hand. They always show low correlation. While Community-based Mediator ($\beta_{CBM}$) and Number of Neighboring Communities ($\beta_{NNC}$) show discrepancy in their behavior from one network to another.
\subsection{Similarity Analysis}
As top nodes are more important than bottom nodes in centrality assessment, RBO is calculated. Moreover, high correlation doesn't necessarily mean high similarity. This is more obvious when ties exist among the rankings of a set. Figure \ref{fig:simHeatmapsMain} shows the RBO similarity heatmaps of the 8 OSNs. The RBO values range from 0 to 0.86. Low similarity from 0 to 0.3 is characterized by the dark purple color. Medium similarity from 0.3 to
0.6 is characterized by the fuchsia color. High similarity over 0.6 is characterized by the light pink color. For comparison purposes, the networks are arranged in the same order as in figure \ref{fig:corrHeatmapsMain}.
Inspecting the heatmaps, Bridging centrality ($\beta_{BC}$) shows almost no similarity with all other classical centrality measures. To a less extent come Community Hub-Bridge ($\beta_{CHB}$) and Participation Coefficient ($\beta_{PC}$) community-aware centrality measures. For these community-aware centralities, the low similarity is consistent across the networks. Community-based Mediator ($\beta_{CBM}$) and Number of Neighboring
Communities ($\beta_{NNC}$) change from one network to another. For example, taking the RBO similarity of the combination ($\alpha_d, \beta_{NNC}$) in DeezerEU, it is equal to 0.68 while in FB Princeton it is equal to 0.04.
This experiment shows consistency with the previous experiment. Indeed, Bridging centrality ($\beta_{BC}$), Community Hub-Bridge ($\beta_{CHB}$), and Participation Coefficient ($\beta_{PC}$) community-aware centrality measures show the lowest similarity to classical centrality measures and their behavior is consistent across the 8 OSNs under study. This case
is similar to the case under Kendall Tau's correlation. However, RBO is more extreme than Kendall's Tau correlation, where low values of similarity can be seen. This is simply due to the RBO definition accounting for ranks. When a group of nodes acquires the same rank, as RBO moves from depth $d$ to $d+1$, the group of tied nodes occurring at $d$ are surpassed and hence account less to the similarity between the two ranked lists.
Referring back to the main research question, indeed, community-aware and classical centrality measures do not convey the same information. Nonetheless, these measures can be divided into two groups. The first group has consistent low similarity with the classical centrality measures while the second group has varying similarity across the networks.
\section{Conclusion}
\label{sec:conc}
Communities have major consequences on the dynamics of a network. Humans tend to form communities within their social presence according to one or many similarity criteria. In addition to that, humans tend to follow other members manifesting power, influence, or popularity, resulting in dense community structures. Centrality measures aim to identify the key members
within OSNs, which is crucial for a lot of strategic applications. However, these measures are agnostic to the community structure. Newly developed centrality measures account for the existence of communities.
Most works have been conducted on classical centrality measures on online social networks. In this work, we shed the light on the relationship between classical and community-aware centrality measures in OSNs. Using 8 real-world OSNs from different platforms, their community structure is uncovered using Infomap. Then, for each network, 5 classical and 5 community-aware centrality measures are calculated. After that, correlation and similarity evaluation between all possible classical and community-aware centrality measures is conducted. Results show that globally these two types of centrality do not convey the same information. Moreover,
community-aware centrality measures exhibit two behaviors. The first set (Bridging centrality, Community Hub-Bridge, and Participation Coefficient) exhibit low correlation and low similarity for all the networks under study. The second set (Community-based Mediator and Number of Neighboring
Communities) shows varying correlation and similarity across networks.
Results of this study suggest that community-aware centrality measures are worth looking into when searching for key members in OSNs, as they provide different information from classical centrality measures. This work opens future research directions. Further study will investigate the effect of network topology on the relationship between classical and community-aware
centrality measures and whether results are consistent using different community detection algorithms.
\bibliographystyle{unsrt}
|
\section{Introduction}
Graph neural networks (GNNs) have enabled significant advances on graph-structured data~\cite{kipf2016semi,velivckovic2017graph} and are widely used in many applications like node classification, graph classification, and recommendation systems. However, existing works show that they are vulnerable towards adversarial attacks~\cite{10.1145/3219819.3220078,zugner_adversarial_2019,10.1145/3394520} like unnoticeable perturbations, which is still a critical challenge in employing GNNs in safety-critical applications.
Albeit various studies have been proposed to ensure the robustness of the graph neural networks against adversarial attacks~\cite{tang2020transferring,zhu2019robust,10.1145/3394486.3403049,ijcai2019-669}, the significance of adversarial augmentations has been ignored, especially under unsupervised learning setting. Recently, self-supervised learning has achieved remarkable performances on graph-structured data, like DGI~\cite{velivckovic2018deep}, GraphCL~\cite{NEURIPS2020_3fe23034}, etc. These works use pairs of augmentations on unlabeled graphs to define a classification task for pretext learning of graph representations. Also, GraphCL~\cite{NEURIPS2020_3fe23034} has found that contrastive learning with randomly generated graph augmentations can somehow increase the robustness; however, we argue that such randomly generated samples are not the optimal choice to achieve the robustness of representations and adversarial augmentations can perform provably much better.
Thereby, we present a novel adversarial self-supervised learning framework to learn robust graph representations.
We introduce adversarial samples into the input.
Primairly, both supervised and unsupervised approaches can be used to generate adversarial samples. For example, Metattack~\cite{zugner_adversarial_2019}, a supervised adversarial attack method, can be directly applied.
Besides, we further propose an unsupervised method to generated adversarial graphs, which uses unsupervised contrastive loss as the target of Metattack to generete adversarial samples. After generating perturbed graphs, we maximize the similarity between representations of the clean graph and the adversarial attacked graph to suppress distortions caused by adversarial perturbations. This will result in representations that are robust against adversarial attacks.
We refer to this novel adversarial self-supervised graph representation learning method as \textbf{Graph} \textbf{A}dversarial \textbf{C}ontrastive \textbf{L}earning (GraphACL). To the best of our knowledge, this is the first attempt to use adversarial samples to increase the robustness of graph representations based on contrastive learning. We also build a theoretical framework to analyze the robustness of graph contrastive learning based on the Information Bottleneck Principle. To verify the effectiveness of GraphACL, we conduct experiments on public academic dataset, Cora, Citeseer, Pubmed under both targeted attack~(i.e., Netattack) and global attack~(i.e., Metattack). Experimental results suggest that GraphACL outperforms DGI and other baselines significantly, thus proving our method can learn robust representations under various graph adversarial attacks.
\begin{figure*}[hbtp]
\centering
\includegraphics[width=\textwidth]{fig/framework.pdf}
\caption{Graph Adversarial Contrasive Learning Framework}
\label{fig:framework}
\end{figure*}
In summary, our contributions are as follows:
\begin{itemize}
\item We propose GraphACL, a general framework to use self-supervised graph contrastive learning with adversarial samples to learn robust graph representations.
\item We theoretically prove that our method could improve the robustness of graph representation learning from the perspective of information theory.
\item We present an unsupervised graph adversarial attack method that use meta-gradient to poison the graph structure to maximize the contrastive loss between clean and perturbed graphs.
\item We conduct extensive experiments to demonstrate the effectiveness of our proposed GraphACL under various types of adversarial attacks, which indicates that GraphACL can significantly improve the performance of previous methods both in evasive and poisoning settings.
\end{itemize}
\section{Methodologies}
\subsection{Graph Adversarial Attack}
In this subsection, we will formulate the classic optimization problem of graph adversarial attack.
Let $G=(A,X)$ be an attribute graph with adjacency matrix $A\in\{0,1\}^{N\times N}$ and attribute matrix $X\in\mathbb{R}^{N\times D}$, where $N$ is the number of nodes and $D$ is the dimension of the node feature vector. Considering a semi-supervised node classification task, where labels of the nodes $\mathcal{V}_L\in \mathcal{V}$ are given. Each node is assigned as one class in $\mathcal{C}\in\{c_1,...,c_k\}$. The goal of adversarial attack can be mathematically formulated as a bilevel optimization problem
\begin{equation}
\begin{aligned}
&\mathop{\text{max}}_{{G}_{attack}\in\Phi(G)}\mathcal{L}(f_{\theta^*}({G}_{attack}))\\ & s.t. \ \ \theta^*=\mathop{arg min}_\theta \mathcal{L}(f_\theta({G}))
\end{aligned}
\end{equation}
\noindent where $\Phi(G)$ is the space of perturbation on the input graph, $\mathcal{L}$ is the cross entropy by default and $f_\theta(\cdot)$ is the surrogate model.
Based on whether to re-train the model on the attacked graph, the attack type is categorized by poisoning attack and evasive attack. Poisoning attack requires re-training while evasive attack does not.
\subsection{Graph Adversarial Contrastive Learning Framework}
As illustrated in Figure~\ref{fig:framework}, we now present our framework to learn robust representations via adversarial contrastive training.
Firstly, we conduct adversarial generation on the perturbed graph. Then, we use the input graph and an adversarial augmented graph as different views of the same graph. A shared encoder like GCN encodes multi-views of the graph and then outputs respective local representations $f_1(G)$ and $f_2(G_{adv})$, where $G$ is the input graph and $G_{adv}$ is the adversarial augmentation. $f_1(\cdot)$ and $f_2(\cdot)$ are encoders that can be the same or different with their unshared projection layers. Crossed local-global information maximization is implemented by maximizing the information between local representations of the input graph and global representations of the adversarial graph, vice versa. The GraphACL framework is modified on DGI framework by additionally introducing an adversarial augmented view of the
input graph. The other omitted settings are the same with DGI, and negative samples are also used. Therefore, the improvement of GraphACL over DGI is of our concern.
\begin{figure*}[htbp]
\centering
\includegraphics[width=.6\textwidth]{fig/attack_su.png}
\caption{Generation of graph adversarial augmentations under supervised loss.}
\label{fig:attack_su}
\end{figure*}
When several labels are known, adversarial augmentation ${G}_{adv}$ could be obtained by supervised generation method.
The process is illustrated in Figure \ref{fig:attack_su} Then
we use the contrastive learning objective to maximize the similarity between input examples ${G}$ and their instance-wise adversarial augmentation ${G}_{adv}$. Then we can formulate our Graph Adversarial Contrastive Learning objective as follow:
\begin{small}
\begin{equation}
\begin{aligned}
\label{eq:sup}
\mathcal{L}_{GACL}^{sup} = &\mathop{\min}_{f_1,f_2} (L_{cl}^{self}( f_1({G}), f_1({G})_{global}) \\
&+ \alpha L_{cl}^{adv}( f_1({G}), f_2({G}_{adv})_{global}) \\
&+ \beta L_{cl}^{adv}( f_1({G})_{global}, f_2({G}_{adv}))),
\end{aligned}
\end{equation}
\end{small}where $L_{cl}$ is contrastive loss that is negative mutual information essentially and $\alpha$ balances between contrastive loss $L_{cl}^{self}$ and $L_{cl}^{adv}$. Similar to DGI \cite{velivckovic2018deep}, $f(\cdot)_{global}$ is the global representation of the whole graph.
\begin{figure*}[htbp]
\centering
\includegraphics[width=.8\textwidth]{fig/attack_unsu.png}
\caption{Generation of graph adversarial augmentations under unsupervised loss.}
\label{fig:attack_unsu}
\end{figure*}
If no label information is given, the unsupervised adversarial training strategy of Graph Adversarial Contrastive Learning objective can be formulated as:
\begin{small}
\begin{equation}
\begin{aligned}
\label{eq:unsup}
\mathcal{L}_{GACL}^{unsup} = &\mathop{\min}_{f_1,f_2} \mathop{\max}_{g} (L_{cl}^{self}( f_1({G}), f_1({G})_{global})\\
&+ \alpha L_{cl}^{adv}( f_1({G}), f_2(g({G}))_{global}) \\
&+ \beta L_{cl}^{adv}( f_1({G})_{global}, f_2(g({G})))),\\
\end{aligned}
\end{equation}
\end{small}
where $g(\cdot)$ is adversarial samples generation function. The process is illustrated in Figure \ref{fig:attack_unsu}
The detailed procedure is presented as \textbf{Algorithm \ref{al:1}}. Different ways of generating adversarial augmentations are formulated in the next section of theoretical analysis.
\begin{algorithm}[h]
\caption{The Procedure of One Iteration in GraphACL}
\label{al:1}
\begin{algorithmic}[1]
\Require
Input Graph $G=(A,X)$;
\Ensure
$f_1(\cdot)$, $f_2(\cdot)$: graph encoders;
\If{use supervised adversarial augmentation}
\State generate adversarial graph $\hat{G}$ based on Eq.(\ref{adv-obj});
\Else
\State generate adversarial graph $\hat{G}$ based on Eq.(\ref{adv-obj-un});
\EndIf
\State Generate node representations of input graph $f_1(G)$;
\State Generate node representations of adversarial augmented graph $f_2(G_{adv})$;
\State Generate global representation of input graph $f_1(G)_g$ by mean pooling;
\State Generate global representation of adversarial augmented graph $f_2(G_{adv})_g$ by mean pooling;
\State Compute contrastive loss by Eq.(\ref{eq:sup}) or Eq.(\ref{eq:unsup});
\State Back propagate gradients and update $f_1(\cdot)$ and $f_2(\cdot)$;\\
\Return $f_1(\cdot)$ and $f_2(\cdot)$;
\end{algorithmic}
\end{algorithm}
\section{Theoretical Analysis on Graph Adversarial Contrastive Learning}
In this section, we first formulate the Information Bottleneck Principle in graph self-supervised learning and achieve the related objective function. Then, we illustrate the generation of adversarial augmentations in Figure~\ref{fig:attacks}. Finally, we derive desirable lower bounds of the information bottleneck and formulate the objective function according to supervised and unsupervised adversarial augmentations.
\subsection{Information Bottleneck Principle for Graph Self-supervised Learning}
The Information Bottleneck (IB) \cite{tishby2000the,tishby2015deep} provides an essential principle for representation learning from the perspective of information theory, which is an optimal representation need to contain minimal yet sufficient information for downstream tasks. It encourages the representation to involve as much information about the target as possible to obtain high prediction accuracy, and discard redundant information that is irrelevant to the target. In graph representation learning, each graph ${G}(A, X)$ contains information of both the graph structure $A\in\mathbb{R}^{N\times N}$ and node features $X\in\mathbb{R}^{N\times d}$. Applying IB to graph self-supervised learning, we desire to learn an optimal graph representation $Z$, which is informative about the original graph ${G} \in {{\mathcal{G}}}$ , but invariant to its augmentations $\widehat{G} \in {\widehat{\mathcal{G}}}$. This principle can be formulated as follows:
\begin{equation}\label{ibloss}
{\mathcal{L}}_{\mathrm{IB}} \triangleq \beta I(Z, \widehat{\mathcal{G}})- I\left(Z, \mathcal{G}\right),
\end{equation}
where $I(\cdot,\cdot)$ denotes mutual information between variables and $\beta>0$ is a hyper-parameter to control the trade-off between preserving information and being invariant to distortions.
We use $\widehat{G}$ to represent different views of the corresponding graph $G$. For the first term of ${\mathcal{L}}_{\mathrm{IB}}$, we utilize an upper bound proved in \cite{cheng2020club} to derive a tractable bound of the mutual information between $Z$ and $\widehat{G}$:
\begin{equation}\label{term2}
\begin{aligned}
I(Z, \widehat{\mathcal{G}}) \leq & \sum_{z \in {Z}} \sum_{\widehat{G} \in {\widehat{\mathcal{G}}}} p(\widehat{G}, z) \log (p(z \mid \widehat{G})) \\
& -\sum_{z \in {Z}} \sum_{\widehat{G} \in {\widehat{\mathcal{G}}}} p(\widehat{G}) p(z) \log (p(z \mid \widehat{G})),
\end{aligned}
\end{equation}
Also, the mutual information between $Z$ and $\mathcal{G}$ can be written as
\begin{equation}
\begin{aligned}
I(Z , \mathcal{G}) = & \sum_{z \in {Z}} \sum_{G \in {\mathcal{G}}} p(z, G) \log \frac{p(z, G)}{p(z) p(G)} \\
= & \sum_{z \in {Z}} \sum_{G \in {\mathcal{G}}} p(z, G) \log p(G | z) + H(\mathcal{G}). \\
\end{aligned}
\end{equation}
The entropy term $H(\mathcal{G})$ could be dropped which results in
\begin{equation}\label{term1}
\begin{aligned}
I(Z, \mathcal{G}) & \geq \sum_{z \in {Z}} \sum_{G \in {\mathcal{G}}} p(z, G) \log p(G | z).\\
\end{aligned}
\end{equation}
By combining Eq.(\ref{ibloss}), Eq.(\ref{term2}) and Eq.(\ref{term1}), we can minimize the upper bound of IB by:
\begin{equation}
\begin{aligned}
\hat{\mathcal{L}}_{\mathrm{IB}} =
&\beta \frac{1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} [\log p(z_{j}^{i} \mid \widehat{G}_{j}^{i})\\
&- \frac{1}{M} \sum_{k=1}^{M} \log p(z_{k}^{i} \mid \widehat{G}_{j}^{i})] \\
& - \frac{ 1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} \log p(G^{i} \mid z_{j}^{i}),\\
\end{aligned}
\label{loss}
\end{equation}
where $N$ is the number of original graphs, $M$ is the number of augmentations of each original input graph. In graph contrastive learning, $p(z_{j}^{i} \mid \widehat{G}_{j}^{i})$ can be viewed as an encoder $f_{\theta}$: $\widehat{\mathcal{G}} \rightarrow Z$.
We assume
$p(z_{k}^{i} \mid z_{j}^{i}, \widehat{G}_{j}^{i}) = p(z_{k}^{i} \mid z_{j}^{i})$,
which means the representation $z_{k}^{i}$ of an augmented graph cannot depend directly on another augmented graph $\widehat{G}_{j}^{i}$. Also, since the function $f_{\theta}$ is deterministic, we have
\begin{equation}
\begin{aligned}
p(z_{k}^{i} \mid \widehat{G}_{j}^{i}) = & \sum p(z_{k}^{i} \mid z_{j}^{i}, \widehat{G}_{j}^{i})p(z_{j}^{i} \mid \widehat{G}_{j}^{i}) \\
= & p(z_{k}^{i} \mid z_{j}^{i})p(z_{j}^{i} \mid \widehat{G}_{j}^{i}) = p(z_{k}^{i} \mid z_{j}^{i})
\end{aligned}
\end{equation}
Further, Eq.(\ref{loss}) can be written as:
\begin{equation}
\begin{aligned}
\hat{\mathcal{L}}_{\mathrm{IB}} =& \beta \frac{ 1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} [ - \frac{1}{M}\sum_{k=1}^{M} \log p(z_{k}^{i} \mid z_{j}^{i})] \\
& - \frac{ 1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} \log p(G^{i} \mid z_{j}^{i}),
\end{aligned}
\label{loss1}
\end{equation}
where $p(z_{k}^{i} | z_{j}^{i})$ could be viewed as a similarity measurement between representations of different augmentations. Eventually, we formulate the problem with IB and obtain a general objective function of graph self-supervised learning as Eq.(\ref{loss1}). Intuitively, the objective function motivates GNN to increase the averaged similarity of representations between different augmentations, thus making the learned representations invariant and robust to various different views.
\subsection{Generation of Supervised Graph Adversarial Augmentations}\label{loss-sup}
The generation of supervised graph adversarial augmentation is schematically shown in Figure~\ref{fig:attack_su}, which utilizes previous graph adversarial attack methods like Metattack~\cite{zugner_adversarial_2019}. Primarily, a surrogate model is applied to the perturbed graph to generate predictions. Then the supervised loss is computed by cross entropy. Finally, we use the gradient to modify the structure of the original graph to generate adversarial samples.
Suppose $G$ is the original graph and its node labels are $Y$, we consider a softmax regression layer between $Z$ and $Y$. The posterior class probabilities can be written as:
\begin{equation}
\begin{aligned}
P_{Y \mid Z}(y \mid z) &=\frac{e^{w_{y}^{T} z}}{\sum_{k} e^{w_{k}^{T} z}}\\
&= \frac{e^{w_{y}^{T} f_{\theta}({G})}}{\sum_{k} e^{w_{k}^{T} f_{\theta}({{G}})}}\\
\end{aligned}
\end{equation}
where $\mathcal{W}=\left\{w_{y}\right\}_{y=1}^{k}$ is the vector of classification parameters for class $y$ and $\theta$ is the parameter of the encoder $f_\theta(\cdot)$, which are learned by minimizing the cross-entropy loss
\begin{equation}
L_{c e}( {G},Y ; \mathcal{W}, \theta)=-\log \frac{e^{w_{y}^{T} f_{\theta}({G})}}{\sum_{k} e^{w_{k}^{T} f_{\theta}({G})}}.
\end{equation}
Given the learned encoder and classifier, an optimal perturbation for $G$ is generated by maximizing the cross-entropy loss:
\begin{equation} \label{adv-obj}
\begin{aligned}
&G_{adv*}=\arg \max _{G_{adv}} L_{c e}(G_{adv}, y ; \mathcal{W}, \theta) \\
&\quad \text { s.t. } G_{adv} \in \Phi\left(G\right),
\end{aligned}
\end{equation}
where $\Phi\left(G\right)$ means the space of perturbation on the original graph. Then we can further formulate a constrained optimization problem as following
\begin{equation}
\begin{aligned}
& G_{adv} \in \Phi\left(X\right) \\
\text { s.t. } & \mathcal{Q}\left(G, G_{adv}\right)<\epsilon
\end{aligned}
\end{equation}
where $\mathcal{Q}(\cdot)$ represents a distance measurement function, $\epsilon$ is a parameter for imperceptible perturbation evaluation.
\begin{equation}
\begin{split}
&\ \frac{1}{K} \sum_{k=1}^{M} \log p[f_{\theta}(\widehat{G}) \mid f_{\theta}(G)]\\
= &\ \frac{1}{K} \sum_{k=1}^{M} \log{ \sum p[f_{\theta}(\widehat{G}) \mid f_{\theta}(G), y)p(y \mid f_{\theta}(G)]}\\
= &\ \frac{1}{K} \sum_{k=1}^{M} \log \sum p[f_{\theta}(\widehat{G}) \mid y)p(y \mid f_{\theta}(G))]\\
> &\ \log [\sum p( f_{\theta}({G}_{adv*}) \mid y)p(y \mid f_{\theta}(G))]\\
= &\ \log p[ f_{\theta}({G}_{adv*}) \mid f_{\theta}(G)],
\end{split}
\end{equation}
Since similarity between clean graph representations and adversarial augmentation representations becomes a lower bound of the averaged similarity between representations on the original graph and all augmentations.
\subsection{Generation of Unsupervised Graph Adversarial Augmentations}\label{loss-unsup}
The unsupervised graph adversarial augmentation generation is schematically in Figure~\ref{fig:attack_unsu}.
Graph $i$ is mapped into an example pair $(\widehat{G}^{i}_{k}, \widehat{G}^{i}_{j})$. Graph contrastive learning is performed through maximizing the agreement between an positive pair. Eq.(\ref{loss1}) tells us that if we want to get a more robust representation, we need to increase $\frac{1}{K}\sum_{k=1}^{M} \log p(z_{k}^{i} | z_{j}^{i})$. Similar to supervised situation, there is a lower bound $\frac{1}{K} \sum_{k=1}^{M} \log p[f_{\theta}(\widehat{G}^{i}_{k}) | f_{\theta}(\widehat{G}^{i}_{j}) ] > \log p[ f_{\theta}({G}_{adv*}^{i})| f_{\theta}(\widehat{G}^{i}_{j})]$.
The choice of ${G}_{adv*}$ could be formulate as a two-stage alternative optimization problem: one is self-supervised learning, the other is adversarial attack or generation of adversarial augmentations. A generative function $g(\cdot)$ is introduced to denote the generation of adversarial samples ${G}_{adv*}$. For example, the generation function $g(\cdot)$ can be the same as it in Metattack. In the first stage, adversarial samples can be generated by ${G}_{adv} = g({G})$, which is further considered as an augmentation or a different view of the perturbed ${G}$. Hence, self-supervised learning is conducted to maximize the mutual information between different views by optimizing corresponding encoders $f_1$ and $f_2$.
Then, given the encoders, we can optimize $g(\cdot)$ by using adversarial attacks to minimize self-supervised loss and obtain a new adversarial graph.
Still using Metattack as an example, $g(\cdot)$ is optimized by attacking the gradient of the self-supervised loss.
Finally, we formulate the two-stage of unsupervised training strategy -- adversarial attack and self-supervised learning as an underlying min-max objective function in the following:
\begin{equation}
\min _{g} \max _{f_{1}, f_{2}} I(f_1({G}), f_2(g({G})))
\end{equation}
In practice, as shown in Figure~\ref{fig:attack_unsu}, the adversarial unsupervised training strategy is modified on DGI framework:
\begin{equation}
\label{adv-obj-un}
\begin{aligned}
\min _{g} \max _{f_{1}, f_{2}} \{ &I[f_1({G})_{global}, f_2(g({G}))] \\
&+ I[f_1({G}), f_2(g({G}))_{global}] \}.
\end{aligned}
\end{equation}
\section{Experiments}
\subsection{Experimental Settings}
To evaluate the robustness of different models against adversarial attacks, we conduct experiments on the following benchmarks with Netattack \cite{pmlr-v97-bojchevski19a} and Metattack \cite{zugner_adversarial_2019}, where adversarial augmentations are generated by supervised or unsupervised contrastive loss, respectively. Netattack works based on boolean features; therefore, the features in each dataset are preprocessed to be 0 or 1. We follow the experimental settings in Netattack \cite{pmlr-v97-bojchevski19a,NEURIPS2020_3fe23034,wu2020graph} exactly: We test the classification accuracy of the 40 selected target nodes: 10 nodes with the highest margin of classification, which is most likely to be classified correctly; 10 nodes with the lowest margin of classification but still classified correctly, which may be easily attacked; 20 other random nodes. Each perturbation denotes a filp on a boolean feature or a modification on an edge related to the node. Robustness experiments are evaluated on a clean graph and corrupted graphs with a number of perturbations from 1 to 4. For Metattack, we use the standard Metattack setting with a perturbation rate of 0.05 and 0.2 on clean graphs to generate modified graphs. Then we test the node classification accuracy on the modified graph. In this experiment, the adversarial augmentations are generated by our proposed unsupervised contrastive loss.
We evaluate two types of robustness tasks, including evasive and poisoning. We include the baselines such as GCN, RGCN, GAT; the results are cited from \cite{wu2020graph}. Previous work has not included the pre-trained models in attack experiments, and DGI is now considered to be compared with since we desire to evaluate the impact of adversarial augmentations. GIB \cite{wu2020graph} is one of the previous SOTA on these experiments; however, it is not related to our comparison on whether to use adversarial augmentations nor unsupervised pretraining. To verify the impact of introducing the adversarial augmentation, we focus on the improvements of GraphACL over DGI and GCN. For Metattack, we only evaluate the model's performance on the evasive task.
Datasets are summarized in the supplementary materials. The results in Table \ref{results} and \ref{meta-results} denote averaged classification accuracy and standard deviation over 5 random seeds. GACL is short for GraphACL.
We denote DGI and GraphACL as pre-trained methods, which are unsupervised pre-trained with only the graph and features and without any other information in downstream tasks. No previous study includes pre-trained methods in robustness experiments; however, we find it effective to defend unknown attacks with pretraining. Our main hypothesis is that the adversarial augmentations will help the model learn more robust representations, which is confirmed by comparing DGI and GraphACL. Thereby, we highlight the best results in DGI, GCL and GraphACL in bold, surpassing all other results except few special cases.
\subsection{Robustness Evaluation under Netattack}
\begin{small}
\begin{table}
\centering
\caption{Classification Accuracy(\%) under Netattack over 5 random seeds}
\label{results}
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccccc|cccc}
\hline
\multirow{2}{*}{} & \multirow{2}{*}{\textbf{Model}} & \multirow{2}{*}{\textbf{Clean}} & \multicolumn{4}{c}{\textbf{Evasive}} & \multicolumn{4}{c}{\textbf{Poisoning}} \\ \cline{4-11}
& & \textbf{} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} \\ \hline
\multirow{5}{*}{\rotatebox{90}{\textbf{Cora}}} & \textbf{GCN} & 80.0\scriptsize{$\pm$7.87} & 51.5\scriptsize{$\pm$4.87} & 38.0\scriptsize{$\pm$6.22} & 31.0\scriptsize{$\pm$2.24} & 26.0\scriptsize{$\pm$3.79} & 47.5\scriptsize{$\pm$7.07} & 39.5\scriptsize{$\pm$2.74} & 30.0\scriptsize{$\pm$5.00} & 26.5\scriptsize{$\pm$3.79} \\
& \textbf{RGCN} & 80.0\scriptsize{$\pm$4.67} & 49.5\scriptsize{$\pm$6.47} & 36.0\scriptsize{$\pm$5.18} & 30.5\scriptsize{$\pm$3.25} & 25.5\scriptsize{$\pm$2.09} & 46.5\scriptsize{$\pm$5.75} & 35.5\scriptsize{$\pm$3.70} & 29.0\scriptsize{$\pm$3.79} & 25.5\scriptsize{$\pm$2.73} \\
& \textbf{GAT} & 77.8\scriptsize{$\pm$3.97} & 48.0\scriptsize{$\pm$8.73} & 39.5\scriptsize{$\pm$5.70} & 36.5\scriptsize{$\pm$5.48} & 32.5\scriptsize{$\pm$5.30} & 50.5\scriptsize{$\pm$5.70} & 38.0\scriptsize{$\pm$5.97} & 33.5\scriptsize{$\pm$2.85} & 26.0\scriptsize{$\pm$3.79} \\
& \textbf{DGI} & 82.5\scriptsize{$\pm$4.33} & 62.0\scriptsize{$\pm$4.81} & 46.0\scriptsize{$\pm$3.79} & 34.0\scriptsize{$\pm$5.18} & {27.5}\scriptsize{$\pm$3.06} & 62.5\scriptsize{$\pm$3.54} & 43.5\scriptsize{$\pm$3.79} & {31.5}\scriptsize{$\pm$6.75} & 26.5\scriptsize{$\pm$4.18} \\
& \textbf{GCL} & 64.4\scriptsize{$\pm$4.27} & 53.8\scriptsize{$\pm$5.20} & 38.8\scriptsize{$\pm$8.54} & 25.6\scriptsize{$\pm$3.15} & 18.8\scriptsize{$\pm$3.23} & 41.9\scriptsize{$\pm$6.25} & 33.1\scriptsize{$\pm$8.00} & 28.8\scriptsize{$\pm$7.77} & 23.1\scriptsize{$\pm$5.54} \\
& \textbf{GACL} & 82.0\scriptsize{$\pm$3.26} & \textbf{67.5}\scriptsize{$\pm$5.00} & \textbf{46.0}\scriptsize{$\pm$5.76} & \textbf{35.5}\scriptsize{$\pm$4.81} & \textbf{29.0}\scriptsize{$\pm$6.75} & \textbf{64.5}\scriptsize{$\pm$5.70} & \textbf{44.0}\scriptsize{$\pm$7.42} & \textbf{33.5}\scriptsize{$\pm$6.02} & \textbf{27.5}\scriptsize{$\pm$5.30} \\ \hline
\multirow{5}{*}{\rotatebox{90}{\textbf{Citeseer}}} & \textbf{GCN} & 71.8\scriptsize{$\pm$6.94} & 42.5\scriptsize{$\pm$7.07} & 27.5\scriptsize{$\pm$6.37} & 18.0\scriptsize{$\pm$3.26} & 15.0\scriptsize{$\pm$2.50} & 29.0\scriptsize{$\pm$7.20} & 20.5\scriptsize{$\pm$1.12} & 17.5\scriptsize{$\pm$1.77} & 13.0\scriptsize{$\pm$2.09} \\
& \textbf{RGCN} & 73.5\scriptsize{$\pm$8.40} & 41.5\scriptsize{$\pm$7.42} & 24.5\scriptsize{$\pm$6.47} & 18.5\scriptsize{$\pm$6.52} & 13.0\scriptsize{$\pm$1.11} & 31.0\scriptsize{$\pm$5.48} & 19.5\scriptsize{$\pm$2.09} & 13.5\scriptsize{$\pm$2.85} & 5.00\scriptsize{$\pm$1.77} \\
& \textbf{GAT} & 72.3\scriptsize{$\pm$8.38} & {49.0}\scriptsize{$\pm$9.12} & 33.0\scriptsize{$\pm$5.97} & 22.0\scriptsize{$\pm$4.81} & 18.0\scriptsize{$\pm$3.26} & 38.0\scriptsize{$\pm$5.12} & 23.5\scriptsize{$\pm$4.87} & 16.5\scriptsize{$\pm$4.54} & 12.0\scriptsize{$\pm$2.09} \\
& \textbf{DGI} & 78.5\scriptsize{$\pm$5.76} & 64.0\scriptsize{$\pm$4.18} & 49.5\scriptsize{$\pm$4.47} & 36.5\scriptsize{$\pm$5.18} & 30.5\scriptsize{$\pm$5.97} & 57.5\scriptsize{$\pm$4.68} & 40.0\scriptsize{$\pm$7.70} & 31.0\scriptsize{$\pm$2.24} & {25.5}\scriptsize{$\pm$5.70} \\
& \textbf{GCL} & 70.0\scriptsize{$\pm$7.36} & 59.4\scriptsize{$\pm$6.57} & 47.5\scriptsize{$\pm$5.40} & 36.3\scriptsize{$\pm$6.29} & 32.5\scriptsize{$\pm$4.08} & 50.6\scriptsize{$\pm$5.54} & 38.3\scriptsize{$\pm$6.77} & 38.1\scriptsize{$\pm$4.27} & 26.9\scriptsize{$\pm$8.00}\\
& \textbf{GACL} & 77.5\scriptsize{$\pm$3.06} & \textbf{66.0}\scriptsize{$\pm$3.79} & \textbf{53.0}\scriptsize{$\pm$5.70} & \textbf{46.0}\scriptsize{$\pm$4.87} & \textbf{37.0}\scriptsize{$\pm$1.12} & \textbf{63.5}\scriptsize{$\pm$4.18} & \textbf{41.0}\scriptsize{$\pm$3.79} & \textbf{40.0}\scriptsize{$\pm$9.19} & \textbf{30.5}\scriptsize{$\pm$5.97} \\ \hline
\multirow{5}{*}{\rotatebox{90}{\textbf{Pubmed}}} & \textbf{GCN} & 82.6\scriptsize{$\pm$6.98} & 39.5\scriptsize{$\pm$4.81} & 32.0\scriptsize{$\pm$4.81} & 31.0\scriptsize{$\pm$5.76} & 31.0\scriptsize{$\pm$5.76} & 36.0\scriptsize{$\pm$4.18} & 32.5\scriptsize{$\pm$6.37} & 31.0\scriptsize{$\pm$5.76} & 28.5\scriptsize{$\pm$5.18} \\
& \textbf{RGCN} & 79.0\scriptsize{$\pm$5.18} & 39.5\scriptsize{$\pm$5.70} & 33.0\scriptsize{$\pm$4.80} & 31.5\scriptsize{$\pm$4.18} & 30.0\scriptsize{$\pm$5.00} & 38.5\scriptsize{$\pm$4.18} & 31.5\scriptsize{$\pm$2.85} & 29.5\scriptsize{$\pm$3.70} & 27.0\scriptsize{$\pm$3.70} \\
& \textbf{GAT} & 78.6\scriptsize{$\pm$6.70} & 41.0\scriptsize{$\pm$8.40} & 33.5\scriptsize{$\pm$4.18} & 30.5\scriptsize{$\pm$4.47} & 31.0\scriptsize{$\pm$4.18} & 39.5\scriptsize{$\pm$3.26} & 31.0\scriptsize{$\pm$4.18} & 30.0\scriptsize{$\pm$3.06} & 35.5\scriptsize{$\pm$5.97} \\
& \textbf{DGI} & 79.0\scriptsize{$\pm$7.20} & 40.5\scriptsize{$\pm$5.86} & 31.0\scriptsize{$\pm$4.54} & 29.5\scriptsize{$\pm$3.71} & 28.0\scriptsize{$\pm$2.74} & 40.0\scriptsize{$\pm$4.81} & 31.0\scriptsize{$\pm$3.79} & 28.5\scriptsize{$\pm$4.18} & 28.0\scriptsize{$\pm$4.68} \\
& \textbf{GCL} & 67.5\scriptsize{$\pm$7.07} & \textbf{45.3}\scriptsize{$\pm$1.77} & \textbf{35.8}\scriptsize{$\pm$6.29} & 28.3\scriptsize{$\pm$5.10} & 28.1\scriptsize{$\pm$4.26} & 40.25\scriptsize{$\pm$1.77} & 33.25\scriptsize{$\pm$6.61} & \textbf{30.3}\scriptsize{$\pm$1.44} & 19.4\scriptsize{$\pm$3.75}\\
& \textbf{GACL} & 83.0\scriptsize{$\pm$5.42} & 43.0\scriptsize{$\pm$5.42} & 34.0\scriptsize{$\pm$5.18} & \textbf{30.0}\scriptsize{$\pm$3.06} & \textbf{28.5}\scriptsize{$\pm$4.18} & \textbf{41.0}\scriptsize{$\pm$3.79} & \textbf{34.0}\scriptsize{$\pm$2.85} & 29.5\scriptsize{$\pm$3.71} & \textbf{28.5}\scriptsize{$\pm$4.18} \\ \hline
\end{tabular}
}
\end{table}
\end{small}
\textbf{Netattack.}
For experiments on Cora, GraphACL with pretraining and adversarial augmentations, outperforms all previous methods like GCN remarkably. In evasive experiments, GraphACL surpasses GCN by 16.0\%, RGCN by 18.0\%, GAT by 19.5\% and DGI by 5.5\% on average on the task with one perturbation. Also, GraphACL achieves 17.0\% and 2.0\% improvements over GCN and DGI respectively, when being poisoned on one perturbation case. When the number of perturbations gets larger, the averaged results of GraphACL are still a bit higher.
GraphACL also achieves significant improvements on both evasive and poisoning experiments on Citeseer. Note that many nodes in Citeseer have few degrees, thus making the attack much harder to defend. When the number of perturbations is 1, GraphACL surpasses DGI and GCN by 2.0\% and 23.5\% on evasive tasks, 3.5\% and 34.5\% on poisoning tasks. Especially, when the number of perturbations is 3, GraphACL surpasses DGI by 9.5\% on evasive tasks and 9.0\% on poisoning tasks. All results of GraphACL surpass DGI and GCN a lot on both evasive and poisoning tasks. We attribute the success to the added views of adversarial augmentations, which makes the model more defensive to the unseen attacks on graphs like Citeseer.
Additionally, GraphACL achieves the best on the clean graph and obtains similar results to GraphCL on Pubmed both in evasive setting and poisoning setting. When the number of perturbations is 1, GraphACL improves the averaged accuracy on GCN, RGCN, GAT, and DGI by 3.5\%, 3.5\%, 2.0\%, and 2.5\%, respectively on evasive tasks.
\subsection{Robustness Evaluation under Metattack}
\textbf{Metattack.}
We also evaluate the performances of DGI and GraphACL on Cora and Citeseer under Metattack in Table~\ref{meta-results}. We use our proposed unsupervised attack method in Figure~\ref{fig:attack_unsu} to generate graph adversarial samples for GraphACL. On the second row, clean denotes evaluation on the clean graph after Metattack, while 0.05 and 0.2 denotes the perturbed rate on the graph. The first column denotes the the perturbed rate in GraphACL training, which is related to adversarial generation. When this rate is 0.000, the method is indeed DGI. While the rate increases, the perturbation gets stronger. As information bottleneck demonstrated, there would be a desired representation containing sufficient information with less nuisance to get more robust performance. The results in Table \ref{meta-results} prove the same idea. The performance gets higher first, achieves a peak, and then goes down when the perturbed rate increases. Surprisingly, we find that the optimal results are all related to GraphACL with 0.030 rate of perturbation. If the perturbation rate is much larger, the graph is corrupted too much to maintain sufficient information, thus resulting in poor performance. The best performances of GraphACL all outperform DGI. When the evalutaion graph is more perturbed, which means the robust representation is much more needed, the improvements get much higher. In Cora, the best GraphACL achieves 1.4\%, 2.2\% and 3.5\% higher in performance than DGI on clean, 0.05-perturbed, 0.2-perturbed graph respectively. In Citeseer, the best GraphACL achieves 2.0\%, 5.4\% and 3.2\% higher in performance than DGI on clean, 0.05-perturbed, 0.2-perturbed graph respectively.
\begin{table*}[hbtp]
\centering
\caption{Classification Accuracy(\%) under Metattack over 5 random seeds}
\begin{tabular}{ccccccc}
\hline
\textbf{Dataset} & \multicolumn{3}{c}{\textbf{Cora}} & \multicolumn{3}{c}{\textbf{Citeseer}} \\ \hline
\textbf{Rate \textbackslash{}Task} & \textbf{Clean} & \textbf{0.05} & \textbf{0.2} & \textbf{Clean} & \textbf{0.05} & \textbf{0.2} \\ \hline
\textbf{DGI} & 75.2\scriptsize{$\pm$2.71} & 73.8\scriptsize{$\pm$2.48} & 71.5\scriptsize{$\pm$2.77} & 67.5\scriptsize{$\pm$3.44} & 64.9\scriptsize{$\pm$4.18} & 65.9\scriptsize{$\pm$4.10} \\
\textbf{GraphACL-0.001} & 75.4\scriptsize{$\pm$2.71} & 74.7\scriptsize{$\pm$3.23} & 73.1\scriptsize{$\pm$3.56} & 69.1\scriptsize{$\pm$1.47} & 69.0\scriptsize{$\pm$1.56} & 68.9\scriptsize{$\pm$1.63} \\
\textbf{GraphACL-0.010} & 76.6\scriptsize{$\pm$2.21} & 75.7\scriptsize{$\pm$2.47} & 74.1\scriptsize{$\pm$2.55} & 70.2\scriptsize{$\pm$1.90} & 69.9\scriptsize{$\pm$2.01} & 68.8\scriptsize{$\pm$2.04} \\
\textbf{GraphACL-0.020} & 76.2\scriptsize{$\pm$2.22} & 74.8\scriptsize{$\pm$2.93} & 74.9\scriptsize{$\pm$3.84} & 68.6\scriptsize{$\pm$3.59} & 68.0\scriptsize{$\pm$3.83} & 66.4\scriptsize{$\pm$3.28} \\
\textbf{GraphACL-0.030} & \textbf{76.8}\scriptsize{$\pm$0.82} & \textbf{76.0}\scriptsize{$\pm$1.40} & \textbf{75.0}\scriptsize{$\pm$1.70} & \textbf{70.5}\scriptsize{$\pm$1.88} & \textbf{70.3}\scriptsize{$\pm$1.87} & \textbf{69.1}\scriptsize{$\pm$2.40} \\
\textbf{GraphACL-0.040} & 73.7\scriptsize{$\pm$3.95} & 73.3\scriptsize{$\pm$3.90} & 72.4\scriptsize{$\pm$3.96} & 69.2\scriptsize{$\pm$2.50} & 68.8\scriptsize{$\pm$2.54} & 67.3\scriptsize{$\pm$3.21} \\
\textbf{GraphACL-0.050} & 74.4\scriptsize{$\pm$4.11} & 73.8\scriptsize{$\pm$4.67} & 72.8\scriptsize{$\pm$5.55} & 69.8\scriptsize{$\pm$1.70} & 69.5\scriptsize{$\pm$1.88} & 68.2\scriptsize{$\pm$2.35} \\ \hline
\end{tabular}
\label{meta-results}
\end{table*}
\subsection{Perturbation Rate Sensitivity for Adversarial Samples}
\begin{figure*}[htbp]
\centering
\subfigure{\includegraphics[width=0.333\textwidth]{fig/Cora-evasive.png}}
\subfigure{\includegraphics[width=0.333\textwidth]{fig/citeseer-evasive.png}}
\subfigure{\includegraphics[width=0.309\textwidth]{fig/pubmed-evasive.png}}
\subfigure{\includegraphics[width=0.333\textwidth]{fig/Cora-poison.png}}
\subfigure{\includegraphics[width=0.333\textwidth]{fig/citeseer-poison.png}}
\subfigure{\includegraphics[width=0.309\textwidth]{fig/pubmed-poison.png}}
\caption{Results on different perturbation rates for adversarial samples}
\label{pert_rate}
\end{figure*}
In Figure \ref{pert_rate}, we evaluate GraphACL with different perturbation rates for adversarial samples on Cora, Citeseer and Pubmed. Experiments are conducted over 5 random seeds. The solid line denotes GraphACL with zero perturbation rate served as a baseline, which means no augmentation is included, i.e., DGI. The dotted line denotes GraphACL with the least positive perturbation rate, which usually performs the worst within GraphACL. The other three lines in dash-dot style related to GraphACL with different suitable perturbation rates. Their performances are similar, which means our method is not sensitive to the perturbation rate in a reasonable range. The difference is mainly based on the dataset. Within the range, there is a best perturbation rate for GraphACL to improve the performance over baselines up to 8\%.
\section{Related work}
\subsection{Adversarial Attack and Defense on Graph Data}
The first graph adversarial attack is proposed by Zugner et al. to generate adversarial graph data using an efficient greedy search method ~\cite{10.1145/3219819.3220078}. The generated graph can be used to fool GNN or other traditional graph learning methods. Then some methods are proposed to attack the topological structure by adding or removing edges according to the gradient of a surrogate model. Xu et al. proposed an optimization-based attack method based on the gradient of the surrogate model~\cite{ijcai2019-550}. Zugner et al. presented to use meta-gradient to guide the perturbation of graph adjacency matrix ~\cite{zugner_adversarial_2019}. Wu et al. argued that integrated gradients can better reflect the effect of perturbing certain features or edges ~\cite{ijcai2019-669}. Also, Bojchevski et al. took DeepWalk~\cite{10.1145/2623330.2623732} as base method using eigen-decomposition and genetic algorithm based strategy to attack the network embedding ~\cite{pmlr-v97-bojchevski19a}.
To ensure the robustness under adversarial attack, many methods have been proposed to defense GNN models~\cite{wang2019adversarial,DBLP:journals/corr/abs-1903-05994,8924766,DBLP:journals/corr/abs-1911-04429,zhu2019robust,wu2020graph}. Wang et al. thought the vulnerabilities of graph neural networks are related to the aggregation layer and the perceptron layer ~\cite{wang2019adversarial}. To address these two disadvantages, they propose an adversarial training framework with a modified GNN model to improve the robustness of GNNs. Chen et al. proposed different defense strategies based on adversarial training for target and global adversarial attack with smoothing distillation and smoothing cross-entropy loss function \cite{DBLP:journals/corr/abs-1903-05994}. Feng et al. proposed a method of adversarial training for the attack on node features with a graph adversarial regularizer which encourages the model to
generate similar predictions on the perturbed target node and its connected nodes ~\cite{8924766}. Wang et al. pointed out that the values of perturbation could be continuous or even negative ~\cite{DBLP:journals/corr/abs-1911-04429}. Zhu et al. proposed to use Gaussian distribution to increase the robustness of Graph Convolutional Network ~\cite{zhu2019robust}. Wu et al. applied the information bottleneck principle on semi-supervised learning settings~\cite{wu2020graph} to defense the targeted node attack~\cite{10.1145/3219819.3220078}.
\subsection{Self-supervised Graph Representation Learning}
Self-supervised contrastive learning~\cite{pmlr-v119-chen20j,He_2020_CVPR} showed significant performance on graph-structured data. The contrastive learning approach usually needs to generate augmented graph sample pairs of the original graph. Then the similarity between the representation of augmented graph pairs is minimized to learn graph representations. Veli{\v{c}}kovi{\'c} et al. proposed to maximize the information between local and global graph representations to learn node representations~\cite{velivckovic2018deep}. Zhu et al. proposed various augmentation strategies to generate augmented graph samples ~\cite{zhu2020deep}. Hassani et al. introduced multi-view contrastive learning ~\cite{pmlr-v119-hassani20a} that maximizes the information between graph and its diffusion versions~\cite{NEURIPS2019_23c89427}. Qiu et al. used an anonymous random walk to generate augmented subgraphs from a large graph and minimize the similarity between the paired subgraphs and maximize the similarity between subgraphs and negative samples ~\cite{10.1145/3394486.3403168}. Although various methods have been proposed to use self-supervised contrastive learning to learn graph representations, few works
considered the quality of augmentation samples.
\section{Conclusion and Discussion}
To summarize, we introduce adversarial augmentations into graph self-supervised representation learning and propose a novel Graph Adversarial Contrastive Learning (GraphACL) framework. Theoretically, we obtain an upper bound of the Information Bottleneck loss function for graph contrastive learning. With adversarial augmentations, our method could result in a much tighter bound and more robust representations. Based on the theoretical analysis, we formulate the GraphACL framework and present relative objective functions in both supervised and unsupervised settings. To verify the empirical performance, we conduct experiments on classic benchmarks attacked by Netattack or Metattack. GraphACL outperforms DGI and other baselines on both evasive and poisoning tasks, thus proving itself a more robust way of graph representation learning. The analysis of different perturbation rates also indicates that our method is not sensitive to the rate. Albeit our model is built on top of Deep graph infomax (DGI), our theory can be easily extended to other models by combining adversarial learning and graph self-supervised learning together.
\bibliographystyle{splncs04}
\section{Introduction}
Graph neural networks (GNNs) have enabled significant advances on graph-structured data~\cite{kipf2016semi,velivckovic2017graph} and are widely used in many applications like node classification, graph classification, and recommendation systems. However, existing works show that they are vulnerable towards adversarial attacks~\cite{10.1145/3219819.3220078,zugner_adversarial_2019,10.1145/3394520} like unnoticeable perturbations, which is still a critical challenge in employing GNNs in safety-critical applications.
Albeit various studies have been proposed to ensure the robustness of the graph neural networks against adversarial attacks~\cite{tang2020transferring,zhu2019robust,10.1145/3394486.3403049,ijcai2019-669}, the significance of adversarial augmentations has been ignored, especially under unsupervised learning setting. Recently, self-supervised learning has achieved remarkable performances on graph-structured data, like DGI~\cite{velivckovic2018deep}, GraphCL~\cite{NEURIPS2020_3fe23034}, etc. These works use pairs of augmentations on unlabeled graphs to define a classification task for pretext learning of graph representations. Also, GraphCL~\cite{NEURIPS2020_3fe23034} has found that contrastive learning with randomly generated graph augmentations can somehow increase the robustness; however, we argue that such randomly generated samples are not the optimal choice to achieve the robustness of representations and adversarial augmentations can perform provably much better.
Thereby, we present a novel adversarial self-supervised learning framework to learn robust graph representations.
We introduce adversarial samples into the input.
Primairly, both supervised and unsupervised approaches can be used to generate adversarial samples. For example, Metattack~\cite{zugner_adversarial_2019}, a supervised adversarial attack method, can be directly applied.
Besides, we further propose an unsupervised method to generated adversarial graphs, which uses unsupervised contrastive loss as the target of Metattack to generete adversarial samples. After generating perturbed graphs, we maximize the similarity between representations of the clean graph and the adversarial attacked graph to suppress distortions caused by adversarial perturbations. This will result in representations that are robust against adversarial attacks.
We refer to this novel adversarial self-supervised graph representation learning method as \textbf{Graph} \textbf{A}dversarial \textbf{C}ontrastive \textbf{L}earning (GraphACL). To the best of our knowledge, this is the first attempt to use adversarial samples to increase the robustness of graph representations based on contrastive learning. We also build a theoretical framework to analyze the robustness of graph contrastive learning based on the Information Bottleneck Principle. To verify the effectiveness of GraphACL, we conduct experiments on public academic dataset, Cora, Citeseer, Pubmed under both targeted attack~(i.e., Netattack) and global attack~(i.e., Metattack). Experimental results suggest that GraphACL outperforms DGI and other baselines significantly, thus proving our method can learn robust representations under various graph adversarial attacks.
\begin{figure*}[hbtp]
\centering
\includegraphics[width=\textwidth]{fig/framework.pdf}
\caption{Graph Adversarial Contrasive Learning Framework}
\label{fig:framework}
\end{figure*}
In summary, our contributions are as follows:
\begin{itemize}
\item We propose GraphACL, a general framework to use self-supervised graph contrastive learning with adversarial samples to learn robust graph representations.
\item We theoretically prove that our method could improve the robustness of graph representation learning from the perspective of information theory.
\item We present an unsupervised graph adversarial attack method that use meta-gradient to poison the graph structure to maximize the contrastive loss between clean and perturbed graphs.
\item We conduct extensive experiments to demonstrate the effectiveness of our proposed GraphACL under various types of adversarial attacks, which indicates that GraphACL can significantly improve the performance of previous methods both in evasive and poisoning settings.
\end{itemize}
\section{Methodologies}
\subsection{Graph Adversarial Attack}
In this subsection, we will formulate the classic optimization problem of graph adversarial attack.
Let $G=(A,X)$ be an attribute graph with adjacency matrix $A\in\{0,1\}^{N\times N}$ and attribute matrix $X\in\mathbb{R}^{N\times D}$, where $N$ is the number of nodes and $D$ is the dimension of the node feature vector. Considering a semi-supervised node classification task, where labels of the nodes $\mathcal{V}_L\in \mathcal{V}$ are given. Each node is assigned as one class in $\mathcal{C}\in\{c_1,...,c_k\}$. The goal of adversarial attack can be mathematically formulated as a bilevel optimization problem
\begin{equation}
\begin{aligned}
&\mathop{\text{max}}_{{G}_{attack}\in\Phi(G)}\mathcal{L}(f_{\theta^*}({G}_{attack}))\\ & s.t. \ \ \theta^*=\mathop{arg min}_\theta \mathcal{L}(f_\theta({G}))
\end{aligned}
\end{equation}
\noindent where $\Phi(G)$ is the space of perturbation on the input graph, $\mathcal{L}$ is the cross entropy by default and $f_\theta(\cdot)$ is the surrogate model.
Based on whether to re-train the model on the attacked graph, the attack type is categorized by poisoning attack and evasive attack. Poisoning attack requires re-training while evasive attack does not.
\subsection{Graph Adversarial Contrastive Learning Framework}
As illustrated in Figure~\ref{fig:framework}, we now present our framework to learn robust representations via adversarial contrastive training.
Firstly, we conduct adversarial generation on the perturbed graph. Then, we use the input graph and an adversarial augmented graph as different views of the same graph. A shared encoder like GCN encodes multi-views of the graph and then outputs respective local representations $f_1(G)$ and $f_2(G_{adv})$, where $G$ is the input graph and $G_{adv}$ is the adversarial augmentation. $f_1(\cdot)$ and $f_2(\cdot)$ are encoders that can be the same or different with their unshared projection layers. Crossed local-global information maximization is implemented by maximizing the information between local representations of the input graph and global representations of the adversarial graph, vice versa. The GraphACL framework is modified on DGI framework by additionally introducing an adversarial augmented view of the
input graph. The other omitted settings are the same with DGI, and negative samples are also used. Therefore, the improvement of GraphACL over DGI is of our concern.
\begin{figure*}[htbp]
\centering
\includegraphics[width=.6\textwidth]{fig/attack_su.png}
\caption{Generation of graph adversarial augmentations under supervised loss.}
\label{fig:attack_su}
\end{figure*}
When several labels are known, adversarial augmentation ${G}_{adv}$ could be obtained by supervised generation method.
The process is illustrated in Figure \ref{fig:attack_su} Then
we use the contrastive learning objective to maximize the similarity between input examples ${G}$ and their instance-wise adversarial augmentation ${G}_{adv}$. Then we can formulate our Graph Adversarial Contrastive Learning objective as follow:
\begin{small}
\begin{equation}
\begin{aligned}
\label{eq:sup}
\mathcal{L}_{GACL}^{sup} = &\mathop{\min}_{f_1,f_2} (L_{cl}^{self}( f_1({G}), f_1({G})_{global}) \\
&+ \alpha L_{cl}^{adv}( f_1({G}), f_2({G}_{adv})_{global}) \\
&+ \beta L_{cl}^{adv}( f_1({G})_{global}, f_2({G}_{adv}))),
\end{aligned}
\end{equation}
\end{small}where $L_{cl}$ is contrastive loss that is negative mutual information essentially and $\alpha$ balances between contrastive loss $L_{cl}^{self}$ and $L_{cl}^{adv}$. Similar to DGI \cite{velivckovic2018deep}, $f(\cdot)_{global}$ is the global representation of the whole graph.
\begin{figure*}[htbp]
\centering
\includegraphics[width=.8\textwidth]{fig/attack_unsu.png}
\caption{Generation of graph adversarial augmentations under unsupervised loss.}
\label{fig:attack_unsu}
\end{figure*}
If no label information is given, the unsupervised adversarial training strategy of Graph Adversarial Contrastive Learning objective can be formulated as:
\begin{small}
\begin{equation}
\begin{aligned}
\label{eq:unsup}
\mathcal{L}_{GACL}^{unsup} = &\mathop{\min}_{f_1,f_2} \mathop{\max}_{g} (L_{cl}^{self}( f_1({G}), f_1({G})_{global})\\
&+ \alpha L_{cl}^{adv}( f_1({G}), f_2(g({G}))_{global}) \\
&+ \beta L_{cl}^{adv}( f_1({G})_{global}, f_2(g({G})))),\\
\end{aligned}
\end{equation}
\end{small}
where $g(\cdot)$ is adversarial samples generation function. The process is illustrated in Figure \ref{fig:attack_unsu}
The detailed procedure is presented as \textbf{Algorithm \ref{al:1}}. Different ways of generating adversarial augmentations are formulated in the next section of theoretical analysis.
\begin{algorithm}[h]
\caption{The Procedure of One Iteration in GraphACL}
\label{al:1}
\begin{algorithmic}[1]
\Require
Input Graph $G=(A,X)$;
\Ensure
$f_1(\cdot)$, $f_2(\cdot)$: graph encoders;
\If{use supervised adversarial augmentation}
\State generate adversarial graph $\hat{G}$ based on Eq.(\ref{adv-obj});
\Else
\State generate adversarial graph $\hat{G}$ based on Eq.(\ref{adv-obj-un});
\EndIf
\State Generate node representations of input graph $f_1(G)$;
\State Generate node representations of adversarial augmented graph $f_2(G_{adv})$;
\State Generate global representation of input graph $f_1(G)_g$ by mean pooling;
\State Generate global representation of adversarial augmented graph $f_2(G_{adv})_g$ by mean pooling;
\State Compute contrastive loss by Eq.(\ref{eq:sup}) or Eq.(\ref{eq:unsup});
\State Back propagate gradients and update $f_1(\cdot)$ and $f_2(\cdot)$;\\
\Return $f_1(\cdot)$ and $f_2(\cdot)$;
\end{algorithmic}
\end{algorithm}
\section{Theoretical Analysis on Graph Adversarial Contrastive Learning}
In this section, we first formulate the Information Bottleneck Principle in graph self-supervised learning and achieve the related objective function. Then, we illustrate the generation of adversarial augmentations in Figure~\ref{fig:attacks}. Finally, we derive desirable lower bounds of the information bottleneck and formulate the objective function according to supervised and unsupervised adversarial augmentations.
\subsection{Information Bottleneck Principle for Graph Self-supervised Learning}
The Information Bottleneck (IB) \cite{tishby2000the,tishby2015deep} provides an essential principle for representation learning from the perspective of information theory, which is an optimal representation need to contain minimal yet sufficient information for downstream tasks. It encourages the representation to involve as much information about the target as possible to obtain high prediction accuracy, and discard redundant information that is irrelevant to the target. In graph representation learning, each graph ${G}(A, X)$ contains information of both the graph structure $A\in\mathbb{R}^{N\times N}$ and node features $X\in\mathbb{R}^{N\times d}$. Applying IB to graph self-supervised learning, we desire to learn an optimal graph representation $Z$, which is informative about the original graph ${G} \in {{\mathcal{G}}}$ , but invariant to its augmentations $\widehat{G} \in {\widehat{\mathcal{G}}}$. This principle can be formulated as follows:
\begin{equation}\label{ibloss}
{\mathcal{L}}_{\mathrm{IB}} \triangleq \beta I(Z, \widehat{\mathcal{G}})- I\left(Z, \mathcal{G}\right),
\end{equation}
where $I(\cdot,\cdot)$ denotes mutual information between variables and $\beta>0$ is a hyper-parameter to control the trade-off between preserving information and being invariant to distortions.
We use $\widehat{G}$ to represent different views of the corresponding graph $G$. For the first term of ${\mathcal{L}}_{\mathrm{IB}}$, we utilize an upper bound proved in \cite{cheng2020club} to derive a tractable bound of the mutual information between $Z$ and $\widehat{G}$:
\begin{equation}\label{term2}
\begin{aligned}
I(Z, \widehat{\mathcal{G}}) \leq & \sum_{z \in {Z}} \sum_{\widehat{G} \in {\widehat{\mathcal{G}}}} p(\widehat{G}, z) \log (p(z \mid \widehat{G})) \\
& -\sum_{z \in {Z}} \sum_{\widehat{G} \in {\widehat{\mathcal{G}}}} p(\widehat{G}) p(z) \log (p(z \mid \widehat{G})),
\end{aligned}
\end{equation}
Also, the mutual information between $Z$ and $\mathcal{G}$ can be written as
\begin{equation}
\begin{aligned}
I(Z , \mathcal{G}) = & \sum_{z \in {Z}} \sum_{G \in {\mathcal{G}}} p(z, G) \log \frac{p(z, G)}{p(z) p(G)} \\
= & \sum_{z \in {Z}} \sum_{G \in {\mathcal{G}}} p(z, G) \log p(G | z) + H(\mathcal{G}). \\
\end{aligned}
\end{equation}
The entropy term $H(\mathcal{G})$ could be dropped which results in
\begin{equation}\label{term1}
\begin{aligned}
I(Z, \mathcal{G}) & \geq \sum_{z \in {Z}} \sum_{G \in {\mathcal{G}}} p(z, G) \log p(G | z).\\
\end{aligned}
\end{equation}
By combining Eq.(\ref{ibloss}), Eq.(\ref{term2}) and Eq.(\ref{term1}), we can minimize the upper bound of IB by:
\begin{equation}
\begin{aligned}
\hat{\mathcal{L}}_{\mathrm{IB}} =
&\beta \frac{1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} [\log p(z_{j}^{i} \mid \widehat{G}_{j}^{i})\\
&- \frac{1}{M} \sum_{k=1}^{M} \log p(z_{k}^{i} \mid \widehat{G}_{j}^{i})] \\
& - \frac{ 1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} \log p(G^{i} \mid z_{j}^{i}),\\
\end{aligned}
\label{loss}
\end{equation}
where $N$ is the number of original graphs, $M$ is the number of augmentations of each original input graph. In graph contrastive learning, $p(z_{j}^{i} \mid \widehat{G}_{j}^{i})$ can be viewed as an encoder $f_{\theta}$: $\widehat{\mathcal{G}} \rightarrow Z$.
We assume
$p(z_{k}^{i} \mid z_{j}^{i}, \widehat{G}_{j}^{i}) = p(z_{k}^{i} \mid z_{j}^{i})$,
which means the representation $z_{k}^{i}$ of an augmented graph cannot depend directly on another augmented graph $\widehat{G}_{j}^{i}$. Also, since the function $f_{\theta}$ is deterministic, we have
\begin{equation}
\begin{aligned}
p(z_{k}^{i} \mid \widehat{G}_{j}^{i}) = & \sum p(z_{k}^{i} \mid z_{j}^{i}, \widehat{G}_{j}^{i})p(z_{j}^{i} \mid \widehat{G}_{j}^{i}) \\
= & p(z_{k}^{i} \mid z_{j}^{i})p(z_{j}^{i} \mid \widehat{G}_{j}^{i}) = p(z_{k}^{i} \mid z_{j}^{i})
\end{aligned}
\end{equation}
Further, Eq.(\ref{loss}) can be written as:
\begin{equation}
\begin{aligned}
\hat{\mathcal{L}}_{\mathrm{IB}} =& \beta \frac{ 1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} [ - \frac{1}{M}\sum_{k=1}^{M} \log p(z_{k}^{i} \mid z_{j}^{i})] \\
& - \frac{ 1}{NM} \sum_{i=1}^{N} \sum_{j=1}^{M} \log p(G^{i} \mid z_{j}^{i}),
\end{aligned}
\label{loss1}
\end{equation}
where $p(z_{k}^{i} | z_{j}^{i})$ could be viewed as a similarity measurement between representations of different augmentations. Eventually, we formulate the problem with IB and obtain a general objective function of graph self-supervised learning as Eq.(\ref{loss1}). Intuitively, the objective function motivates GNN to increase the averaged similarity of representations between different augmentations, thus making the learned representations invariant and robust to various different views.
\subsection{Generation of Supervised Graph Adversarial Augmentations}\label{loss-sup}
The generation of supervised graph adversarial augmentation is schematically shown in Figure~\ref{fig:attack_su}, which utilizes previous graph adversarial attack methods like Metattack~\cite{zugner_adversarial_2019}. Primarily, a surrogate model is applied to the perturbed graph to generate predictions. Then the supervised loss is computed by cross entropy. Finally, we use the gradient to modify the structure of the original graph to generate adversarial samples.
Suppose $G$ is the original graph and its node labels are $Y$, we consider a softmax regression layer between $Z$ and $Y$. The posterior class probabilities can be written as:
\begin{equation}
\begin{aligned}
P_{Y \mid Z}(y \mid z) &=\frac{e^{w_{y}^{T} z}}{\sum_{k} e^{w_{k}^{T} z}}\\
&= \frac{e^{w_{y}^{T} f_{\theta}({G})}}{\sum_{k} e^{w_{k}^{T} f_{\theta}({{G}})}}\\
\end{aligned}
\end{equation}
where $\mathcal{W}=\left\{w_{y}\right\}_{y=1}^{k}$ is the vector of classification parameters for class $y$ and $\theta$ is the parameter of the encoder $f_\theta(\cdot)$, which are learned by minimizing the cross-entropy loss
\begin{equation}
L_{c e}( {G},Y ; \mathcal{W}, \theta)=-\log \frac{e^{w_{y}^{T} f_{\theta}({G})}}{\sum_{k} e^{w_{k}^{T} f_{\theta}({G})}}.
\end{equation}
Given the learned encoder and classifier, an optimal perturbation for $G$ is generated by maximizing the cross-entropy loss:
\begin{equation} \label{adv-obj}
\begin{aligned}
&G_{adv*}=\arg \max _{G_{adv}} L_{c e}(G_{adv}, y ; \mathcal{W}, \theta) \\
&\quad \text { s.t. } G_{adv} \in \Phi\left(G\right),
\end{aligned}
\end{equation}
where $\Phi\left(G\right)$ means the space of perturbation on the original graph. Then we can further formulate a constrained optimization problem as following
\begin{equation}
\begin{aligned}
& G_{adv} \in \Phi\left(X\right) \\
\text { s.t. } & \mathcal{Q}\left(G, G_{adv}\right)<\epsilon
\end{aligned}
\end{equation}
where $\mathcal{Q}(\cdot)$ represents a distance measurement function, $\epsilon$ is a parameter for imperceptible perturbation evaluation.
\begin{equation}
\begin{split}
&\ \frac{1}{K} \sum_{k=1}^{M} \log p[f_{\theta}(\widehat{G}) \mid f_{\theta}(G)]\\
= &\ \frac{1}{K} \sum_{k=1}^{M} \log{ \sum p[f_{\theta}(\widehat{G}) \mid f_{\theta}(G), y)p(y \mid f_{\theta}(G)]}\\
= &\ \frac{1}{K} \sum_{k=1}^{M} \log \sum p[f_{\theta}(\widehat{G}) \mid y)p(y \mid f_{\theta}(G))]\\
> &\ \log [\sum p( f_{\theta}({G}_{adv*}) \mid y)p(y \mid f_{\theta}(G))]\\
= &\ \log p[ f_{\theta}({G}_{adv*}) \mid f_{\theta}(G)],
\end{split}
\end{equation}
Since similarity between clean graph representations and adversarial augmentation representations becomes a lower bound of the averaged similarity between representations on the original graph and all augmentations.
\subsection{Generation of Unsupervised Graph Adversarial Augmentations}\label{loss-unsup}
The unsupervised graph adversarial augmentation generation is schematically in Figure~\ref{fig:attack_unsu}.
Graph $i$ is mapped into an example pair $(\widehat{G}^{i}_{k}, \widehat{G}^{i}_{j})$. Graph contrastive learning is performed through maximizing the agreement between an positive pair. Eq.(\ref{loss1}) tells us that if we want to get a more robust representation, we need to increase $\frac{1}{K}\sum_{k=1}^{M} \log p(z_{k}^{i} | z_{j}^{i})$. Similar to supervised situation, there is a lower bound $\frac{1}{K} \sum_{k=1}^{M} \log p[f_{\theta}(\widehat{G}^{i}_{k}) | f_{\theta}(\widehat{G}^{i}_{j}) ] > \log p[ f_{\theta}({G}_{adv*}^{i})| f_{\theta}(\widehat{G}^{i}_{j})]$.
The choice of ${G}_{adv*}$ could be formulate as a two-stage alternative optimization problem: one is self-supervised learning, the other is adversarial attack or generation of adversarial augmentations. A generative function $g(\cdot)$ is introduced to denote the generation of adversarial samples ${G}_{adv*}$. For example, the generation function $g(\cdot)$ can be the same as it in Metattack. In the first stage, adversarial samples can be generated by ${G}_{adv} = g({G})$, which is further considered as an augmentation or a different view of the perturbed ${G}$. Hence, self-supervised learning is conducted to maximize the mutual information between different views by optimizing corresponding encoders $f_1$ and $f_2$.
Then, given the encoders, we can optimize $g(\cdot)$ by using adversarial attacks to minimize self-supervised loss and obtain a new adversarial graph.
Still using Metattack as an example, $g(\cdot)$ is optimized by attacking the gradient of the self-supervised loss.
Finally, we formulate the two-stage of unsupervised training strategy -- adversarial attack and self-supervised learning as an underlying min-max objective function in the following:
\begin{equation}
\min _{g} \max _{f_{1}, f_{2}} I(f_1({G}), f_2(g({G})))
\end{equation}
In practice, as shown in Figure~\ref{fig:attack_unsu}, the adversarial unsupervised training strategy is modified on DGI framework:
\begin{equation}
\label{adv-obj-un}
\begin{aligned}
\min _{g} \max _{f_{1}, f_{2}} \{ &I[f_1({G})_{global}, f_2(g({G}))] \\
&+ I[f_1({G}), f_2(g({G}))_{global}] \}.
\end{aligned}
\end{equation}
\section{Experiments}
\subsection{Experimental Settings}
To evaluate the robustness of different models against adversarial attacks, we conduct experiments on the following benchmarks with Netattack \cite{pmlr-v97-bojchevski19a} and Metattack \cite{zugner_adversarial_2019}, where adversarial augmentations are generated by supervised or unsupervised contrastive loss, respectively. Netattack works based on boolean features; therefore, the features in each dataset are preprocessed to be 0 or 1. We follow the experimental settings in Netattack \cite{pmlr-v97-bojchevski19a,NEURIPS2020_3fe23034,wu2020graph} exactly: We test the classification accuracy of the 40 selected target nodes: 10 nodes with the highest margin of classification, which is most likely to be classified correctly; 10 nodes with the lowest margin of classification but still classified correctly, which may be easily attacked; 20 other random nodes. Each perturbation denotes a filp on a boolean feature or a modification on an edge related to the node. Robustness experiments are evaluated on a clean graph and corrupted graphs with a number of perturbations from 1 to 4. For Metattack, we use the standard Metattack setting with a perturbation rate of 0.05 and 0.2 on clean graphs to generate modified graphs. Then we test the node classification accuracy on the modified graph. In this experiment, the adversarial augmentations are generated by our proposed unsupervised contrastive loss.
We evaluate two types of robustness tasks, including evasive and poisoning. We include the baselines such as GCN, RGCN, GAT; the results are cited from \cite{wu2020graph}. Previous work has not included the pre-trained models in attack experiments, and DGI is now considered to be compared with since we desire to evaluate the impact of adversarial augmentations. GIB \cite{wu2020graph} is one of the previous SOTA on these experiments; however, it is not related to our comparison on whether to use adversarial augmentations nor unsupervised pretraining. To verify the impact of introducing the adversarial augmentation, we focus on the improvements of GraphACL over DGI and GCN. For Metattack, we only evaluate the model's performance on the evasive task.
Datasets are summarized in the supplementary materials. The results in Table \ref{results} and \ref{meta-results} denote averaged classification accuracy and standard deviation over 5 random seeds. GACL is short for GraphACL.
We denote DGI and GraphACL as pre-trained methods, which are unsupervised pre-trained with only the graph and features and without any other information in downstream tasks. No previous study includes pre-trained methods in robustness experiments; however, we find it effective to defend unknown attacks with pretraining. Our main hypothesis is that the adversarial augmentations will help the model learn more robust representations, which is confirmed by comparing DGI and GraphACL. Thereby, we highlight the best results in DGI, GCL and GraphACL in bold, surpassing all other results except few special cases.
\subsection{Robustness Evaluation under Netattack}
\begin{small}
\begin{table}
\centering
\caption{Classification Accuracy(\%) under Netattack over 5 random seeds}
\label{results}
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccccc|cccc}
\hline
\multirow{2}{*}{} & \multirow{2}{*}{\textbf{Model}} & \multirow{2}{*}{\textbf{Clean}} & \multicolumn{4}{c}{\textbf{Evasive}} & \multicolumn{4}{c}{\textbf{Poisoning}} \\ \cline{4-11}
& & \textbf{} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} \\ \hline
\multirow{5}{*}{\rotatebox{90}{\textbf{Cora}}} & \textbf{GCN} & 80.0\scriptsize{$\pm$7.87} & 51.5\scriptsize{$\pm$4.87} & 38.0\scriptsize{$\pm$6.22} & 31.0\scriptsize{$\pm$2.24} & 26.0\scriptsize{$\pm$3.79} & 47.5\scriptsize{$\pm$7.07} & 39.5\scriptsize{$\pm$2.74} & 30.0\scriptsize{$\pm$5.00} & 26.5\scriptsize{$\pm$3.79} \\
& \textbf{RGCN} & 80.0\scriptsize{$\pm$4.67} & 49.5\scriptsize{$\pm$6.47} & 36.0\scriptsize{$\pm$5.18} & 30.5\scriptsize{$\pm$3.25} & 25.5\scriptsize{$\pm$2.09} & 46.5\scriptsize{$\pm$5.75} & 35.5\scriptsize{$\pm$3.70} & 29.0\scriptsize{$\pm$3.79} & 25.5\scriptsize{$\pm$2.73} \\
& \textbf{GAT} & 77.8\scriptsize{$\pm$3.97} & 48.0\scriptsize{$\pm$8.73} & 39.5\scriptsize{$\pm$5.70} & 36.5\scriptsize{$\pm$5.48} & 32.5\scriptsize{$\pm$5.30} & 50.5\scriptsize{$\pm$5.70} & 38.0\scriptsize{$\pm$5.97} & 33.5\scriptsize{$\pm$2.85} & 26.0\scriptsize{$\pm$3.79} \\
& \textbf{DGI} & 82.5\scriptsize{$\pm$4.33} & 62.0\scriptsize{$\pm$4.81} & 46.0\scriptsize{$\pm$3.79} & 34.0\scriptsize{$\pm$5.18} & {27.5}\scriptsize{$\pm$3.06} & 62.5\scriptsize{$\pm$3.54} & 43.5\scriptsize{$\pm$3.79} & {31.5}\scriptsize{$\pm$6.75} & 26.5\scriptsize{$\pm$4.18} \\
& \textbf{GCL} & 64.4\scriptsize{$\pm$4.27} & 53.8\scriptsize{$\pm$5.20} & 38.8\scriptsize{$\pm$8.54} & 25.6\scriptsize{$\pm$3.15} & 18.8\scriptsize{$\pm$3.23} & 41.9\scriptsize{$\pm$6.25} & 33.1\scriptsize{$\pm$8.00} & 28.8\scriptsize{$\pm$7.77} & 23.1\scriptsize{$\pm$5.54} \\
& \textbf{GACL} & 82.0\scriptsize{$\pm$3.26} & \textbf{67.5}\scriptsize{$\pm$5.00} & \textbf{46.0}\scriptsize{$\pm$5.76} & \textbf{35.5}\scriptsize{$\pm$4.81} & \textbf{29.0}\scriptsize{$\pm$6.75} & \textbf{64.5}\scriptsize{$\pm$5.70} & \textbf{44.0}\scriptsize{$\pm$7.42} & \textbf{33.5}\scriptsize{$\pm$6.02} & \textbf{27.5}\scriptsize{$\pm$5.30} \\ \hline
\multirow{5}{*}{\rotatebox{90}{\textbf{Citeseer}}} & \textbf{GCN} & 71.8\scriptsize{$\pm$6.94} & 42.5\scriptsize{$\pm$7.07} & 27.5\scriptsize{$\pm$6.37} & 18.0\scriptsize{$\pm$3.26} & 15.0\scriptsize{$\pm$2.50} & 29.0\scriptsize{$\pm$7.20} & 20.5\scriptsize{$\pm$1.12} & 17.5\scriptsize{$\pm$1.77} & 13.0\scriptsize{$\pm$2.09} \\
& \textbf{RGCN} & 73.5\scriptsize{$\pm$8.40} & 41.5\scriptsize{$\pm$7.42} & 24.5\scriptsize{$\pm$6.47} & 18.5\scriptsize{$\pm$6.52} & 13.0\scriptsize{$\pm$1.11} & 31.0\scriptsize{$\pm$5.48} & 19.5\scriptsize{$\pm$2.09} & 13.5\scriptsize{$\pm$2.85} & 5.00\scriptsize{$\pm$1.77} \\
& \textbf{GAT} & 72.3\scriptsize{$\pm$8.38} & {49.0}\scriptsize{$\pm$9.12} & 33.0\scriptsize{$\pm$5.97} & 22.0\scriptsize{$\pm$4.81} & 18.0\scriptsize{$\pm$3.26} & 38.0\scriptsize{$\pm$5.12} & 23.5\scriptsize{$\pm$4.87} & 16.5\scriptsize{$\pm$4.54} & 12.0\scriptsize{$\pm$2.09} \\
& \textbf{DGI} & 78.5\scriptsize{$\pm$5.76} & 64.0\scriptsize{$\pm$4.18} & 49.5\scriptsize{$\pm$4.47} & 36.5\scriptsize{$\pm$5.18} & 30.5\scriptsize{$\pm$5.97} & 57.5\scriptsize{$\pm$4.68} & 40.0\scriptsize{$\pm$7.70} & 31.0\scriptsize{$\pm$2.24} & {25.5}\scriptsize{$\pm$5.70} \\
& \textbf{GCL} & 70.0\scriptsize{$\pm$7.36} & 59.4\scriptsize{$\pm$6.57} & 47.5\scriptsize{$\pm$5.40} & 36.3\scriptsize{$\pm$6.29} & 32.5\scriptsize{$\pm$4.08} & 50.6\scriptsize{$\pm$5.54} & 38.3\scriptsize{$\pm$6.77} & 38.1\scriptsize{$\pm$4.27} & 26.9\scriptsize{$\pm$8.00}\\
& \textbf{GACL} & 77.5\scriptsize{$\pm$3.06} & \textbf{66.0}\scriptsize{$\pm$3.79} & \textbf{53.0}\scriptsize{$\pm$5.70} & \textbf{46.0}\scriptsize{$\pm$4.87} & \textbf{37.0}\scriptsize{$\pm$1.12} & \textbf{63.5}\scriptsize{$\pm$4.18} & \textbf{41.0}\scriptsize{$\pm$3.79} & \textbf{40.0}\scriptsize{$\pm$9.19} & \textbf{30.5}\scriptsize{$\pm$5.97} \\ \hline
\multirow{5}{*}{\rotatebox{90}{\textbf{Pubmed}}} & \textbf{GCN} & 82.6\scriptsize{$\pm$6.98} & 39.5\scriptsize{$\pm$4.81} & 32.0\scriptsize{$\pm$4.81} & 31.0\scriptsize{$\pm$5.76} & 31.0\scriptsize{$\pm$5.76} & 36.0\scriptsize{$\pm$4.18} & 32.5\scriptsize{$\pm$6.37} & 31.0\scriptsize{$\pm$5.76} & 28.5\scriptsize{$\pm$5.18} \\
& \textbf{RGCN} & 79.0\scriptsize{$\pm$5.18} & 39.5\scriptsize{$\pm$5.70} & 33.0\scriptsize{$\pm$4.80} & 31.5\scriptsize{$\pm$4.18} & 30.0\scriptsize{$\pm$5.00} & 38.5\scriptsize{$\pm$4.18} & 31.5\scriptsize{$\pm$2.85} & 29.5\scriptsize{$\pm$3.70} & 27.0\scriptsize{$\pm$3.70} \\
& \textbf{GAT} & 78.6\scriptsize{$\pm$6.70} & 41.0\scriptsize{$\pm$8.40} & 33.5\scriptsize{$\pm$4.18} & 30.5\scriptsize{$\pm$4.47} & 31.0\scriptsize{$\pm$4.18} & 39.5\scriptsize{$\pm$3.26} & 31.0\scriptsize{$\pm$4.18} & 30.0\scriptsize{$\pm$3.06} & 35.5\scriptsize{$\pm$5.97} \\
& \textbf{DGI} & 79.0\scriptsize{$\pm$7.20} & 40.5\scriptsize{$\pm$5.86} & 31.0\scriptsize{$\pm$4.54} & 29.5\scriptsize{$\pm$3.71} & 28.0\scriptsize{$\pm$2.74} & 40.0\scriptsize{$\pm$4.81} & 31.0\scriptsize{$\pm$3.79} & 28.5\scriptsize{$\pm$4.18} & 28.0\scriptsize{$\pm$4.68} \\
& \textbf{GCL} & 67.5\scriptsize{$\pm$7.07} & \textbf{45.3}\scriptsize{$\pm$1.77} & \textbf{35.8}\scriptsize{$\pm$6.29} & 28.3\scriptsize{$\pm$5.10} & 28.1\scriptsize{$\pm$4.26} & 40.25\scriptsize{$\pm$1.77} & 33.25\scriptsize{$\pm$6.61} & \textbf{30.3}\scriptsize{$\pm$1.44} & 19.4\scriptsize{$\pm$3.75}\\
& \textbf{GACL} & 83.0\scriptsize{$\pm$5.42} & 43.0\scriptsize{$\pm$5.42} & 34.0\scriptsize{$\pm$5.18} & \textbf{30.0}\scriptsize{$\pm$3.06} & \textbf{28.5}\scriptsize{$\pm$4.18} & \textbf{41.0}\scriptsize{$\pm$3.79} & \textbf{34.0}\scriptsize{$\pm$2.85} & 29.5\scriptsize{$\pm$3.71} & \textbf{28.5}\scriptsize{$\pm$4.18} \\ \hline
\end{tabular}
}
\end{table}
\end{small}
\textbf{Netattack.}
For experiments on Cora, GraphACL with pretraining and adversarial augmentations, outperforms all previous methods like GCN remarkably. In evasive experiments, GraphACL surpasses GCN by 16.0\%, RGCN by 18.0\%, GAT by 19.5\% and DGI by 5.5\% on average on the task with one perturbation. Also, GraphACL achieves 17.0\% and 2.0\% improvements over GCN and DGI respectively, when being poisoned on one perturbation case. When the number of perturbations gets larger, the averaged results of GraphACL are still a bit higher.
GraphACL also achieves significant improvements on both evasive and poisoning experiments on Citeseer. Note that many nodes in Citeseer have few degrees, thus making the attack much harder to defend. When the number of perturbations is 1, GraphACL surpasses DGI and GCN by 2.0\% and 23.5\% on evasive tasks, 3.5\% and 34.5\% on poisoning tasks. Especially, when the number of perturbations is 3, GraphACL surpasses DGI by 9.5\% on evasive tasks and 9.0\% on poisoning tasks. All results of GraphACL surpass DGI and GCN a lot on both evasive and poisoning tasks. We attribute the success to the added views of adversarial augmentations, which makes the model more defensive to the unseen attacks on graphs like Citeseer.
Additionally, GraphACL achieves the best on the clean graph and obtains similar results to GraphCL on Pubmed both in evasive setting and poisoning setting. When the number of perturbations is 1, GraphACL improves the averaged accuracy on GCN, RGCN, GAT, and DGI by 3.5\%, 3.5\%, 2.0\%, and 2.5\%, respectively on evasive tasks.
\subsection{Robustness Evaluation under Metattack}
\textbf{Metattack.}
We also evaluate the performances of DGI and GraphACL on Cora and Citeseer under Metattack in Table~\ref{meta-results}. We use our proposed unsupervised attack method in Figure~\ref{fig:attack_unsu} to generate graph adversarial samples for GraphACL. On the second row, clean denotes evaluation on the clean graph after Metattack, while 0.05 and 0.2 denotes the perturbed rate on the graph. The first column denotes the the perturbed rate in GraphACL training, which is related to adversarial generation. When this rate is 0.000, the method is indeed DGI. While the rate increases, the perturbation gets stronger. As information bottleneck demonstrated, there would be a desired representation containing sufficient information with less nuisance to get more robust performance. The results in Table \ref{meta-results} prove the same idea. The performance gets higher first, achieves a peak, and then goes down when the perturbed rate increases. Surprisingly, we find that the optimal results are all related to GraphACL with 0.030 rate of perturbation. If the perturbation rate is much larger, the graph is corrupted too much to maintain sufficient information, thus resulting in poor performance. The best performances of GraphACL all outperform DGI. When the evalutaion graph is more perturbed, which means the robust representation is much more needed, the improvements get much higher. In Cora, the best GraphACL achieves 1.4\%, 2.2\% and 3.5\% higher in performance than DGI on clean, 0.05-perturbed, 0.2-perturbed graph respectively. In Citeseer, the best GraphACL achieves 2.0\%, 5.4\% and 3.2\% higher in performance than DGI on clean, 0.05-perturbed, 0.2-perturbed graph respectively.
\begin{table*}[hbtp]
\centering
\caption{Classification Accuracy(\%) under Metattack over 5 random seeds}
\begin{tabular}{ccccccc}
\hline
\textbf{Dataset} & \multicolumn{3}{c}{\textbf{Cora}} & \multicolumn{3}{c}{\textbf{Citeseer}} \\ \hline
\textbf{Rate \textbackslash{}Task} & \textbf{Clean} & \textbf{0.05} & \textbf{0.2} & \textbf{Clean} & \textbf{0.05} & \textbf{0.2} \\ \hline
\textbf{DGI} & 75.2\scriptsize{$\pm$2.71} & 73.8\scriptsize{$\pm$2.48} & 71.5\scriptsize{$\pm$2.77} & 67.5\scriptsize{$\pm$3.44} & 64.9\scriptsize{$\pm$4.18} & 65.9\scriptsize{$\pm$4.10} \\
\textbf{GraphACL-0.001} & 75.4\scriptsize{$\pm$2.71} & 74.7\scriptsize{$\pm$3.23} & 73.1\scriptsize{$\pm$3.56} & 69.1\scriptsize{$\pm$1.47} & 69.0\scriptsize{$\pm$1.56} & 68.9\scriptsize{$\pm$1.63} \\
\textbf{GraphACL-0.010} & 76.6\scriptsize{$\pm$2.21} & 75.7\scriptsize{$\pm$2.47} & 74.1\scriptsize{$\pm$2.55} & 70.2\scriptsize{$\pm$1.90} & 69.9\scriptsize{$\pm$2.01} & 68.8\scriptsize{$\pm$2.04} \\
\textbf{GraphACL-0.020} & 76.2\scriptsize{$\pm$2.22} & 74.8\scriptsize{$\pm$2.93} & 74.9\scriptsize{$\pm$3.84} & 68.6\scriptsize{$\pm$3.59} & 68.0\scriptsize{$\pm$3.83} & 66.4\scriptsize{$\pm$3.28} \\
\textbf{GraphACL-0.030} & \textbf{76.8}\scriptsize{$\pm$0.82} & \textbf{76.0}\scriptsize{$\pm$1.40} & \textbf{75.0}\scriptsize{$\pm$1.70} & \textbf{70.5}\scriptsize{$\pm$1.88} & \textbf{70.3}\scriptsize{$\pm$1.87} & \textbf{69.1}\scriptsize{$\pm$2.40} \\
\textbf{GraphACL-0.040} & 73.7\scriptsize{$\pm$3.95} & 73.3\scriptsize{$\pm$3.90} & 72.4\scriptsize{$\pm$3.96} & 69.2\scriptsize{$\pm$2.50} & 68.8\scriptsize{$\pm$2.54} & 67.3\scriptsize{$\pm$3.21} \\
\textbf{GraphACL-0.050} & 74.4\scriptsize{$\pm$4.11} & 73.8\scriptsize{$\pm$4.67} & 72.8\scriptsize{$\pm$5.55} & 69.8\scriptsize{$\pm$1.70} & 69.5\scriptsize{$\pm$1.88} & 68.2\scriptsize{$\pm$2.35} \\ \hline
\end{tabular}
\label{meta-results}
\end{table*}
\subsection{Perturbation Rate Sensitivity for Adversarial Samples}
\begin{figure*}[htbp]
\centering
\subfigure{\includegraphics[width=0.333\textwidth]{fig/Cora-evasive.png}}
\subfigure{\includegraphics[width=0.333\textwidth]{fig/citeseer-evasive.png}}
\subfigure{\includegraphics[width=0.309\textwidth]{fig/pubmed-evasive.png}}
\subfigure{\includegraphics[width=0.333\textwidth]{fig/Cora-poison.png}}
\subfigure{\includegraphics[width=0.333\textwidth]{fig/citeseer-poison.png}}
\subfigure{\includegraphics[width=0.309\textwidth]{fig/pubmed-poison.png}}
\caption{Results on different perturbation rates for adversarial samples}
\label{pert_rate}
\end{figure*}
In Figure \ref{pert_rate}, we evaluate GraphACL with different perturbation rates for adversarial samples on Cora, Citeseer and Pubmed. Experiments are conducted over 5 random seeds. The solid line denotes GraphACL with zero perturbation rate served as a baseline, which means no augmentation is included, i.e., DGI. The dotted line denotes GraphACL with the least positive perturbation rate, which usually performs the worst within GraphACL. The other three lines in dash-dot style related to GraphACL with different suitable perturbation rates. Their performances are similar, which means our method is not sensitive to the perturbation rate in a reasonable range. The difference is mainly based on the dataset. Within the range, there is a best perturbation rate for GraphACL to improve the performance over baselines up to 8\%.
\section{Related work}
\subsection{Adversarial Attack and Defense on Graph Data}
The first graph adversarial attack is proposed by Zugner et al. to generate adversarial graph data using an efficient greedy search method ~\cite{10.1145/3219819.3220078}. The generated graph can be used to fool GNN or other traditional graph learning methods. Then some methods are proposed to attack the topological structure by adding or removing edges according to the gradient of a surrogate model. Xu et al. proposed an optimization-based attack method based on the gradient of the surrogate model~\cite{ijcai2019-550}. Zugner et al. presented to use meta-gradient to guide the perturbation of graph adjacency matrix ~\cite{zugner_adversarial_2019}. Wu et al. argued that integrated gradients can better reflect the effect of perturbing certain features or edges ~\cite{ijcai2019-669}. Also, Bojchevski et al. took DeepWalk~\cite{10.1145/2623330.2623732} as base method using eigen-decomposition and genetic algorithm based strategy to attack the network embedding ~\cite{pmlr-v97-bojchevski19a}.
To ensure the robustness under adversarial attack, many methods have been proposed to defense GNN models~\cite{wang2019adversarial,DBLP:journals/corr/abs-1903-05994,8924766,DBLP:journals/corr/abs-1911-04429,zhu2019robust,wu2020graph}. Wang et al. thought the vulnerabilities of graph neural networks are related to the aggregation layer and the perceptron layer ~\cite{wang2019adversarial}. To address these two disadvantages, they propose an adversarial training framework with a modified GNN model to improve the robustness of GNNs. Chen et al. proposed different defense strategies based on adversarial training for target and global adversarial attack with smoothing distillation and smoothing cross-entropy loss function \cite{DBLP:journals/corr/abs-1903-05994}. Feng et al. proposed a method of adversarial training for the attack on node features with a graph adversarial regularizer which encourages the model to
generate similar predictions on the perturbed target node and its connected nodes ~\cite{8924766}. Wang et al. pointed out that the values of perturbation could be continuous or even negative ~\cite{DBLP:journals/corr/abs-1911-04429}. Zhu et al. proposed to use Gaussian distribution to increase the robustness of Graph Convolutional Network ~\cite{zhu2019robust}. Wu et al. applied the information bottleneck principle on semi-supervised learning settings~\cite{wu2020graph} to defense the targeted node attack~\cite{10.1145/3219819.3220078}.
\subsection{Self-supervised Graph Representation Learning}
Self-supervised contrastive learning~\cite{pmlr-v119-chen20j,He_2020_CVPR} showed significant performance on graph-structured data. The contrastive learning approach usually needs to generate augmented graph sample pairs of the original graph. Then the similarity between the representation of augmented graph pairs is minimized to learn graph representations. Veli{\v{c}}kovi{\'c} et al. proposed to maximize the information between local and global graph representations to learn node representations~\cite{velivckovic2018deep}. Zhu et al. proposed various augmentation strategies to generate augmented graph samples ~\cite{zhu2020deep}. Hassani et al. introduced multi-view contrastive learning ~\cite{pmlr-v119-hassani20a} that maximizes the information between graph and its diffusion versions~\cite{NEURIPS2019_23c89427}. Qiu et al. used an anonymous random walk to generate augmented subgraphs from a large graph and minimize the similarity between the paired subgraphs and maximize the similarity between subgraphs and negative samples ~\cite{10.1145/3394486.3403168}. Although various methods have been proposed to use self-supervised contrastive learning to learn graph representations, few works
considered the quality of augmentation samples.
\section{Conclusion and Discussion}
To summarize, we introduce adversarial augmentations into graph self-supervised representation learning and propose a novel Graph Adversarial Contrastive Learning (GraphACL) framework. Theoretically, we obtain an upper bound of the Information Bottleneck loss function for graph contrastive learning. With adversarial augmentations, our method could result in a much tighter bound and more robust representations. Based on the theoretical analysis, we formulate the GraphACL framework and present relative objective functions in both supervised and unsupervised settings. To verify the empirical performance, we conduct experiments on classic benchmarks attacked by Netattack or Metattack. GraphACL outperforms DGI and other baselines on both evasive and poisoning tasks, thus proving itself a more robust way of graph representation learning. The analysis of different perturbation rates also indicates that our method is not sensitive to the rate. Albeit our model is built on top of Deep graph infomax (DGI), our theory can be easily extended to other models by combining adversarial learning and graph self-supervised learning together.
\bibliographystyle{splncs04}
|
\section{Introduction}
Jets in \pp\ and heavy-ion collisions arise from hard-scattered (high-$Q^{2}$) quarks and gluons of the incoming beams. In vacuum, a highly virtual parton generated in such interaction comes on-shell by radiating gluons, resulting in a jet shower. Studying jet properties in \pp\ collisions provides the opportunity to explore the perturbative and non-perturbative QCD effects in vacuum. In addition, the comparison between data and different QCD-based Monte Carlo (MC) event generators helps to constrain model parameters. In heavy-ion collisions, a highly energetic parton---while traversing through the Quark-Gluon Plasma (QGP)---interacts with the colored medium and loses its energy via medium-induced gluon radiation. This phenomenon is known as the jet quenching\cite{Cunqueiro:2021wls}. Suppression of jet yield, modification of jet shape and substructure, and jet deflection in the QGP are the manifestations of jet quenching in heavy-ion collisions. \\
In these proceedings we present recent jet measurements by the STAR experiment at RHIC, addressing jet substructure in \pp\ collisions and jet quenching in Au+Au collisions.
\section{Jet measurements in \pp\ collisions}
\label{Sect:pp}
The fragmentation and evolution of a hard-scattered parton is described by the Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) splitting kernels~\cite{Dokshitzer:1977sg,Gribov:1972ri,Altarelli:1977zs}. This splitting probability depends on the momentum fraction of the split and the opening angle, and can be studied in \pp\ collisions. In QCD, higher order corrections contribute to the jet mass and substructure observables. To study these observables in STAR, jets are studied by clustering charged tracks from the time projection chamber and neutral particles from the barrel electromagnetic calorimeter towers using the anti-\ensuremath{k_\mathrm{T}}\ jet reconstruction algorithm~\cite{Salam:2010nqg} with different resolution parameters (\ensuremath{R}) between 0.2 and 0.6. Furthermore, such measurements in \pp\ collisions provide a baseline for the similar measurements in heavy-ion collisions to study the modification of parton shower in the finite-temperature QCD medium.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Zg.eps}
\caption{The $z_{\mathrm g}$\ distributions for different $p_{\rm T,jet}$ in \pp\ collisions at \ensuremath{\sqrt{s}}\ = 200 GeV\cite{STAR:2020ejj}.}
\label{fig:zg}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Rg.eps}
\caption{The $R_{\mathrm g}$\ distributions for different $p_{\rm T,jet}$ in \pp\ collisions at \ensuremath{\sqrt{s}}\ = 200 GeV\cite{STAR:2020ejj}.}
\label{fig:Rg}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{JetMassA.pdf}
\includegraphics[width=0.7\textwidth]{JetMassb.pdf}
\caption{Jet mass distributions for different $p_{\rm T,jet}$ and jet \ensuremath{R}\ in \pp\ collisions at \ensuremath{\sqrt{s}}\ = 200 GeV\cite{STAR:2021lvw}.}
\label{fig:Jetmass}
\end{figure}
\subsection{SoftDrop jet grooming}
The SoftDrop jet-grooming\cite{Larkoski:2015lea,Dasgupta:2013ihk,Larkoski:2014wba} algorithm helps to study jet substructure by suppressing soft large-angle radiations. In this procedure, the soft and wide-angle radiations are removed sequentially from the jet de-clustering tree. This is achieved using the Cambridge/Aachen (C/A) clustering algorithm\cite{Salam:2010nqg} by de-clustering jet branching history with removing the soft branch until it satisfies the condition:
\begin{equation}
z_{g} = \frac{ {\rm min} (p_{\rm T,1}, p_{\rm T,2})}{p_{\rm T,1}+p_{\rm T,2}} > z_{\rm cut} \big( \frac{R_{\rm g}}{R}\big)^{\beta}.
\end{equation}
In Eq. (1), indices 1 and 2 represent the two sub-jets of splitting. The radius ($R$) is the distance in pseudorapidity and azimuthal angle space between two sub-jets, and $R_{\mathrm g}$\ is the groomed jet radius. The SoftDrop threshold $z_{\rm cut}=0.1$, and angular exponent $\beta=0$ are used for this de-clustering procedure for infrared and collinear (IRC) safety\cite{Larkoski:2014wba}.
\\
At RHIC, the first fully corrected SoftDrop observables, $z_{\mathrm g}$\ and $R_{\mathrm g}$ , are measured in \pp\ collisions at $\sqrt{s}$ = 200 GeV by the STAR experiment for inclusive jets with \ensuremath{R}\ = 0.2, 0.4, and 0.6, and 15 $<$ \ensuremath{p_\mathrm{T,jet}} $<$ 60 \ensuremath{\mathrm{GeV/}c} \cite{STAR:2020ejj}. Figures~\ref{fig:zg} and \ref{fig:Rg} show the distributions of $z_{\mathrm g}$\ and $R_{\mathrm g}$, respectively. The shape of $z_{\mathrm g}$\ distributions indicates no \ensuremath{p_\mathrm{T,jet}}\ dependence above 30 \ensuremath{\mathrm{GeV/}c}, and they are more asymmetric than the DGLAP splitting function for a leading order quark emitting a gluon. $R_{\mathrm g}$\ distributions reveal a narrowing with increasing \ensuremath{p_\mathrm{T,jet}}, and the splitting is asymmetric at high \ensuremath{p_\mathrm{T,jet}}. The STAR-tuned\cite{STAR:2019yqm} PYTHIA-6 with Perugia 2012 well describes the jet substructure observables at this energy. The comparisons with MC event generator predictions help further study different hadronization models for the higher-order QCD corrections at RHIC energy.
\subsection{Jet mass}
The mass of quark or gluon jets is sensitive to the fragmentation of highly virtual parent partons. The SoftDrop grooming procedure removes soft and wide-angle radiations from jets making the groomed jets less sensitive to the higher order QCD corrections. Jet mass is defined as the four-momentum sum of jet constituents, $M= \large| \sum_{i \in \rm jet} p_{i}\large| = \sqrt{E^{2}-{ {\textbf {\it p}}}^{2}}$. Here $E$ and ${\textbf {\it p}}$ are the energy and three-momentum of the jet, respectively.
Studying both ungroomed and groomed jets, and comparing to different MC event generators can provide information on different pQCD effects and fragmentation. The STAR experiment has reported the first fully corrected ungroomed ($M$) and groomed ($M_{\rm g}$) mass distributions of inclusive jets for several values of \ensuremath{R}\ at $\sqrt{s}=200$ GeV as shown in Fig~\ref{fig:Jetmass}\cite{STAR:2021lvw}. These jets are selected within the range of $30 < \ensuremath{p_\mathrm{T,jet}} < 40 $ \ensuremath{\mathrm{GeV/}c}. It is observed that the mean and width of the jet mass increases with increasing \ensuremath{R}\ due to the inclusion of wide-angle radiation. The same trend is also seen with growing \ensuremath{p_\mathrm{T,jet}}\ that increases the radiation phase space. The groomed jet mass distribution gets shifted to a smaller value than that of ungroomed mass due to the reduction of soft radiation in the SoftDrop grooming procedure. The LHC-tuned PYTHIA-8 and HERWIG-7 EE4C MC event generators over- and under-predicts the jet mass at RHIC, respectively, whereas the STAR-tuned PYTHIA-6 quantitatively describes the data. This observation is similar to that for the $R_{\mathrm g}$\ observable as discussed in the previous subsection. These measurements serve as a reference for future jet mass measurements in heavy-ion collisions at RHIC.
\section{Jet quenching measurements in Au+Au collisions}
\label{Sect:AuAu}
The STAR experiment has reported measurements of jet quenching using observable high-\ensuremath{p_\mathrm{T}}\ hadron suppression\cite{STAR:2002ggv} and dihadron correlations\cite{STAR:2002svs,STAR:2006vcp}.The hadronic measurements have limited information on the underlying mechanism of jet quenching due to the final-state effects in heavy-ion collisions. Over the last few years, the application of jet reconstruction algorithms and the development of methods for rigorous correction of uncorrelated background in heavy-ion collisions enable us to study jet quenching in more detail using fully reconstructed jets.\\
The first measurements of inclusive jet, semi-inclusive hadron+jet, and preliminary results of \ensuremath{\gamma_\mathrm{dir}}+jet and \ensuremath{\pi^0}+jet measurements have been reported by the STAR experiment. The measurement techniques and their results are discussed in this section.
\subsection{Inclusive jet suppression}
Jet measurements in heavy-ion collisions are complicated due to the presence of large uncorrelated background. For the inclusive jet spectrum measurements in STAR, jets are reconstructed using anti-\ensuremath{k_\mathrm{T}}\ algorithm\cite{Salam:2010nqg} with an additional requirement of a high-\ensuremath{p_\mathrm{T}}\ hadronic constituent ($p_{\rm T,lead}^{\rm min}$), in order to identify jets from hard scattering processes. The selection of $p_{\rm T,lead}^{\rm min}$ should satisfy the following criteria: i) it must be sufficiently high so that contributions from combinatorial jets are negligible; ii) the probability of multiple constituents with \ensuremath{p_\mathrm{T}}\ $\geq p_{\rm T,lead}^{\rm min}$ is negligible; iii) this $p_{\rm T,lead}^{\rm min}$ cut does not introduce a selection bias on the jet population within the considered $p_{\rm T,jet}$ range. Using this technique, the first fully corrected inclusive jet spectrea in central and peripheral Au+Au collisions at \ensuremath{\sqrt{s_\mathrm{NN}}}\ = 200 GeV have been reported with $p_{\rm T,lead}^{\rm min}$=5 \ensuremath{\mathrm{GeV/}c}\ \cite{STAR:2020xiv}. \\
The nuclear modification factor (\ensuremath{R_\mathrm{AA}}) is defined as the ratio of inclusive charged jet yield in central Au+Au collisions to its cross sections in \pp\ collisions scaled by the nuclear thickness factor $\langle T_{\mathrm{AA}} \rangle$ of central collisions. Similarly, \ensuremath{R_\mathrm{CP}}\ is defined considering 60-80\% peripheral collisions as a reference instead of \pp\ collisions. For \ensuremath{R_\mathrm{AA}}, PYTHIA is used as a vacuum reference, hence it is labeled as \ensuremath{R^{\rm Pythia}_\mathrm{AA}}. Figure~\ref{fig:InclJetRAa} shows the \ensuremath{R^{\rm Pythia}_\mathrm{AA}}\ as a function of $p_{\rm T,jet}^{\rm ch}$ for inclusive jets with $\ensuremath{R}=0.2, 0.3$ and 0.4 within 15 $<p_{\rm T,jet}^{\rm ch}<$ 30 \ensuremath{\mathrm{GeV/}c}. Strong suppression is observed in 0-10\% central \ensuremath{{\rm Au+Au}}\ collisions, and no jet \ensuremath{R}\ dependence of the suppression is seen. Different theory calculations\cite{Vitev:2009rd,Chien:2015hda,Casalderrey-Solana:2016jvj,He:2015pra,Ke:2018jem}are consistent with the data. The \ensuremath{R_\mathrm{CP}}\ shows strong and similar suppressions for $\ensuremath{R}=0.2$ and 0.3 jets as shown in Fig~\ref{fig:InclJetRCP}. The yields of inclusive charged hadrons and jets show a comparable level of suppression within the same \ensuremath{p_\mathrm{T}}\ interval in central Au+Au collisions at \ensuremath{\sqrt{s_\mathrm{NN}}}\ = 200 GeV. In addition, the comparison between central Au+Au collisions at RHIC and central Pb+Pb collisions at the LHC---although within different \ensuremath{p_\mathrm{T}}\ intervals---show a similar magnitude of suppression for charged hadrons and jets yields.
The medium-induced broadening of the inclusive jet is measured by taking the ratio of inclusive jet yields for $\ensuremath{R}=0.2$ and 0.4. No significant modification of the transverse jet profile due to jet quenching is observed for the inclusive jet population in central Au+Au collisions at $\ensuremath{\sqrt{s_\mathrm{NN}}}\ = 200$ GeV and is consistent with the LHC data. The ongoing full jet analysis will access the inclusive jet suppression measurement at higher \ensuremath{p_\mathrm{T,jet}}\ at RHIC energies.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{Fig19.eps}
\caption{\ensuremath{R^{\rm Pythia}_\mathrm{AA}}\ as a function of \ensuremath{p_\mathrm{T,jet}}\ in 0-10\% central Au+Au collisions and for different jet \ensuremath{R}\ at \ensuremath{\sqrt{s_\mathrm{NN}}}\ = 200 GeV\cite{STAR:2020xiv}.}
\label{fig:InclJetRAa}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{Fig16.eps}
\caption{\ensuremath{R_\mathrm{CP}}\ as a function of \ensuremath{p_\mathrm{T,jet}}\ in Au+Au collisions and for different jet \ensuremath{R}\ at \ensuremath{\sqrt{s_\mathrm{NN}}}\ = 200 GeV\cite{STAR:2020xiv}.}
\label{fig:InclJetRCP}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{DirPhoJet_1520_R0205_Pythia6_WTheory_HP2020.pdf}
\caption{\ensuremath{\gamma_\mathrm{dir}}+jet \ensuremath{{I}_{AA}}\ as a function of \ensuremath{p_\mathrm{T,jet}}\ for \ensuremath{R}\ = 0.2 and 0.5 at \ensuremath{\sqrt{s_\mathrm{NN}}}\ = 200 GeV\cite{Sahoo:2020kwh}.}
\label{fig:DirPhoIAA}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{PTShift_All_HP2020_Pythia6and8_HP2020_Aug12.pdf}
\caption{$-\Delta p_{\rm T,jet}$ for different observables measured at RHIC and the LHC\cite{Sahoo:2020kwh}.}
\label{fig:PtShift}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.1\textwidth]{STARLHCHardProbes_v5}
\caption{The kinematic coverages of the STAR hard probe measurements (past, current, and future projection) are shown and compared to the LHC (published) measurements\cite{STARBUR2225}.}
\label{fig:STARBUR}
\end{figure}
\subsection{Semi-inclusive \ensuremath{\gamma_\mathrm{dir}}+jet and hadron+jet suppression}
The STAR experiment published\cite{STAR:2017hhs} the measurement of semi-inclusive distribution of reconstructed recoil jets from a high-\ensuremath{p_\mathrm{T}}\ trigger hadron (h+jet) in Au+Au collisions. In this measurement, the uncorrelated background contribution is mitigated by using a novel mixed-event (ME) technique. It is found that the contributions from multi-parton interactions in the recoil acceptance are negligible. \\
The semi-inclusive h+jet measurement enables us to perform similar \ensuremath{\gamma_\mathrm{dir}}+jet and \ensuremath{\pi^0}+jet measurements in STAR by combining the method used in the previous \ensuremath{\gamma_\mathrm{dir}}+hadron and \ensuremath{\pi^0}+hadron correlation measurements\cite{STAR:2016jdz}. In this measurement, \ensuremath{\gamma_\mathrm{dir}}\ and \ensuremath{\pi^0}\ are within the trigger energy ranges of 9 $<$ \ensuremath{E_\mathrm{T}^\mathrm{trig}} $<$ 20 GeV and 9 $<$ \ensuremath{E_\mathrm{T}^\mathrm{trig}} $<$ 15 GeV, respectively. However, in these proceedings only \ensuremath{\gamma_\mathrm{dir}}+jet results with 15 $<$ \ensuremath{E_\mathrm{T}^\mathrm{trig}} $<$ 20 GeV range is discussed, which is the highest \ensuremath{E_\mathrm{T}^\mathrm{trig}}\ range measured in this analysis. Recoil jets are measured using anti-\ensuremath{k_\mathrm{T}}\ algorithm with \ensuremath{R}\ = 0.2 and 0.5. The same ME technique as in h+jet paper is applied to subtract uncorrelated jet background. The unfolding procedure, as in h+jet paper, is applied to correct for the detector effects and heavy-ion background fluctuations. Finally, the ratio of recoil jet yield in central Au+Au collisions to that of PYTHIA-8, \ensuremath{I^\mathrm{PYTHIA}_{AA}}, is presented for the aforementioned jet radii. The \ensuremath{I^\mathrm{PYTHIA}_{AA}}\ of \ensuremath{\gamma_\mathrm{dir}}+jet for \ensuremath{R}\ = 0.2, within 15 $<$ \ensuremath{E_\mathrm{T}^\mathrm{trig}} $<$ 20 GeV, shows a stronger suppression than that of \ensuremath{R}\ = 0.5 as shown in Fig.~\ref{fig:DirPhoIAA}, which hints at a potential \ensuremath{R}\ dependence of recoil jet suppression at RHIC. The upcoming results with the \pp\ data as a reference will be reported in the final publication of this measurement. Precision measurement with extended recoil jet \ensuremath{p_\mathrm{T}^\mathrm{jet,ch}}\ range for \ensuremath{\gamma_\mathrm{dir}}+jet is planned with the upcoming RHIC runs.
\subsection{Charged jet \ensuremath{p_\mathrm{T,jet}}\ spectrum shift: RHIC vs. LHC}
Jet suppression is commonly reported via \ensuremath{R_\mathrm{AA}}\ and \ensuremath{{I}_{AA}}\ as a function of \ensuremath{p_\mathrm{T,jet}}. These observables convolute the effect of energy loss with the shape of the jet \ensuremath{p_\mathrm{T,jet}}\ spectrum. In order to deconvolute this, a \ensuremath{p_\mathrm{T,jet}}\ shift ($-\Delta p_{\rm T,jet}$ in Fig.~\ref{fig:PtShift}) is measured for a quantitative comparison of jet energy loss from different observables. The STAR h+jet\cite{STAR:2017hhs}, inclusive jet, preliminary \ensuremath{\gamma_\mathrm{dir}}+jet, and \ensuremath{\pi^0}+jet measurements, and ALICE h+jet\cite{ALICE:2015mdb} measurements report values of $-\Delta p_{\rm T,jet}$, although within different kinematic ranges. While comparing these values as shown in Fig.~\ref{fig:PtShift}, an indication of smaller in-medium energy loss at RHIC than the LHC is seen.
\section{STAR ongoing measurements and upcoming data-taking plan}
\label{Sect:Future}
There are several ongoing jet measurements in STAR to study the QGP medium properties in heavy-ion collisions, such as full jet reconstruction to extend the \ensuremath{p_\mathrm{T,jet}}\ reach, jet fragmentation function, jet shape, heavy-flavor jet, and recoil jet azimuthal angular correlation with trigger particles.
The STAR experiment plans to take high statistics data of Au+Au collisions at \ensuremath{\sqrt{s_\mathrm{NN}}}\ = 200 GeV in 2023 and 2025, and \pp\ collision data at $\sqrt{s} = 200$ GeV along with $p$+A collisions in 2024 with 28 cryo-weeks for each year\cite{STARBUR2225}.
The kinematic coverages of various measurements related to hard probes using these datasets are presented in Fig~\ref{fig:STARBUR}.
These datasets are crucial for studying the inner-workings of QGP utilizing precision jet measurements.
\section{Summary}
\label{Sect:Summary}
The STAR experiment has recently reported important results on jet substructure observables in \pp\ collisions, and various jet quenching observables in heavy-ion collisions to study QGP properties. Several other jet measurements are ongoing and will be presented in the near future. Besides, STAR's upcoming data-taking (during 2023-2025 RHIC runs) is crucial for the precision jet measurements with large kinematic coverages, whereas high statistics \pp\ data (2024 RHIC run) are important for providing high precision references.
\section*{Acknowledgement}
NRS is supported by the Fundamental Research Funds of Shandong University and NNSF of China:12050410235.
\bibliographystyle{ws-ijprai}
|
\section{Introduction}
Several time-harmonic wave propagation phenomena can be modeled by variations of the Helmholtz equation.
We are interested in developing tools for the numerical simulation of linear acoustic propagation at a fixed frequency in a subsonic flow in three dimensions.
Such phenomena can be modeled by the convected Helmholtz equation for the acoustic potential:
\begin{equation}\label{CHinit}
-\nabla\cdot(\rho(\nabla\phi-(\mathbf M\cdot\nabla\phi)\mathbf M+\textrm i \kappa \phi\mathbf M))
-\rho(\kappa^2\phi+\textrm i \kappa \mathbf M\cdot\nabla\phi)=0,
\end{equation}
where $\rho$ is the real-valued fluid density and $\mathbf M$ is the vector-valued fluid velocity, both assumed to be depending on the space variable, while $\kappa$ is the wavenumber.
The density $\rho$ is naturally assumed to be positive. We are interested in the subsonic regime, so the Mach number $|\mathbf M|$ is assumed to be no more than $1$.
For comparison the standard Helmholtz equation then corresponds to the case of a fluid with velocity equal to zero and a constant density.
The methods of interest here are the so-called quasi-Trefftz methods that were introduced as an extension of Trefftz methods for problems of wave propagation in inhomogeneous media.
In a search for bounds on the solutions to boundary value problems, Trefftz introduced in the 1920s the idea to leverage trial functions satisfying the governing equation \cite{Trefftz,transl}.
Since then, this idea has lead to the development of several numerical methods, generally referred to as Trefftz methods.
In the present work, Trefftz methods refer to a class of numerical methods falling within the category of Galerkin methods, and specifically relying on functions that satisfy the governing equation, both for the derivation of their weak formulation and for the discretization of this formulation.
%
Implementing Trefftz methods requires bases of
exact solutions to the governing equation - called Trefftz functions - in order to discretize the weak formulation.
This then limits their application since exact solutions are not known for many equations. Nevertheless these methods are particularly popular in the field of wave propagation, as circular, spherical, plane or even evanescent waves are some natural and common candidates Trefftz functions, see for instance \cite{moiolaTH,LuoT,survey} and references therein.
Early works on such methods include the introduction \cite{uwvf,uwvfart} and study \cite{Huttuwvf,buffam,gabg,Hutt1,Hutt2,Hutt3} of the ultra-weak variational formulation, as well as more general wave-based methods \cite{PWDGh,PWDGp,PWDGhp}.
More recent works include \cite{TVEM,TVEMnum} focusing on Trefftz Virtual Element Methods, and \cite{barucond,cong} focusing on conditioning aspects.
Existing extensions to space-time problems include work on acoustics and elasto-acoustics \cite{STTDGac,STTDGea}, as well as Friedrichs systems \cite{Morel1,Morel2}.
Intuitively, the advantage of Trefftz methods relies on their PDE-dependent function spaces: since Trefftz functions solve exactly the PDE, they can be expected to have better approximation properties than non-PDE dependent function spaces. In practice, local approximation properties of the discrete spaces are a corner stone of convergence properties of Galerkin methods. Typically these would be stated for a space $\mathbb V$ of functions to be approximated by functions in a discrete space $\mathbb V_h$ locally on some region $\Omega$ as:
$$
\forall u \in \mathbb V,
\exists u_a\in\mathbb V_h \text{ s. t. }
\forall \varx\in\Omega,
|u(\varx)-u_a(\varx)|\leq C | \varx-\pG|^{n}
$$
for some order of approximation $n$.
This is precisely where the advantage of Trefftz functions over standard polynomial spaces can be emphasized, as spaces of Trefftz functions require less degrees of freedom to achieve a given order of accuracy than standard polynomial spaces.
It is however important to keep in mind that these approximation properties of spaces of Trefftz functions hold only for the approximation of exact solutions to the governing PDE (as will be illustrated in \Cref{thm:approx}),
whereas approximation properties of standard polynomial spaces hold for the approximation of smooth enough function that do not necessarily solve the governing PDE.
Bases of Trefftz functions are also known to suffer from ill-conditioning issues in certain regimes, these aspects are beyond the scope of this work.
In the general context of wave propagation, the application of Trefftz methods in their standard form to problems of propagation through inhomogeneous media is similarly limited since one more time global exact solutions are not known for most variable-coefficient equations.
Quasi-Trefftz methods, relying on approximate solutions - called quasi-Trefftz functions - rather than exact solutions to the governing equation, were introduced to extend Trefftz methods to problems governed by variable-coefficient equations. They were first introduced in \cite{LMD} under the name of Generalized Plane Wave (GPW) methods for 2D problems governed by the Helmholtz equation.
The original idea behind the GPW concept was to retain the oscillating behavior of a plane wave (PW) while allowing for some extra degrees of freedom to be adapted to the varying PDE coefficient, and this is where their name came from. Initially this was performed via the introduction of Higher Order Terms (HOT) in the phase of a PW as follows:
\begin{equation}\label{eq:c/GPW}
\left\{\begin{array}{l}
\varphi(\varx)= \exp( i \widetilde{\kappa}
\mathbf k\cdot\varx+\text{HOT})\\
\left[-\Delta-\kappa^2\epsilon(\varx)\right]\varphi(\varx)\approx0,
\end{array}\right.
\text{ instead of }
\left\{\begin{array}{l}
\phi(\varx)=\exp i\kappa\mathbf k\cdot\varx\\
\left[-\Delta-\kappa^2\right]\phi=0,
\end{array}\right.
\text{ for any unit vector } \mathbf k\in\mathbb R^3,
\end{equation}
where $\kappa$ is the wavenumber of the PW while $\widetilde{\kappa}
$ can be interpreted as the local wavenumber of the GPW.
A procedure to construct a basis of such GPWs was proposed in \cite{LMinterp}, together with a study of the approximation properties of the basis.
A systematic procedure to study these approximation properties was introduced in \cite{IGS}, and used in \cite{AbGPW} on a new type of GPWs.
The idea of the associated GPW-based Galerkin method was presented in \cite{LMD}, a proof of convergent for a variant was studied in \cite{LMM}. The method was applied a problem of mode conversion for wave propagation in plasmas in \cite{LMmc}.
Other works on quasi-Trefftz methods include
\cite{Yuan} on the convergence of a GPW Discontinuous Galerkin method for anisotropic Helmholtz problems,
\cite{Morel2}, on linear transport problems,
as well as \cite{IGMS} on time-dependent wave propagation problems.
Quasi-Trefftz methods rely on function spaces of approximate solutions to the governing equation, as opposed to exact solutions, and this is their fundamental property. In our work, we define this approximation as a local property in the sense of a Taylor expansion.
Given the partial differential operator $\mathcal L$ of the governing equation, and any point $\pG$ in the domain of interest, we consider functions $\varphi$ with the following property:
\begin{equation}
\label{eq:TaylExp}
\mathcal L \varphi(\varx)=O(|\varx - \pG|^q),
\end{equation}
for some parameter $q$ providing some flexibility in the desired order of approximation with respect to the distance $|\varx - \pG|$. In other words the degree $q-1$ Taylor polynomial of the image of $\varphi$ through the operator $\mathcal L$ is zero.
In the context of Discontinuous Galerkin methods, with function spaces of local functions defined element-wise on a computational mesh, then if \eqref{eq:TaylExp} holds within each element, with $\pG$ in the element, then the remainder can be described as $O(h^q)$ where $h$ denotes the mesh size, as $|\varx - \pG|\leq h$ for all $\varx$ in the element.
There the goal is to establish so-called $h$-convergence properties, that is in the regime $h\to 0$.
Note that, in order to prove their convergence, quasi-Trefftz methods so far include in their weak formulations a stabilization term to handle the non-zero remainder in the quasi-Trefftz property \eqref{eq:TaylExp}, see \cite{LMM,Yuan,IGMS}.
\subsection{Central results}
Our goal is to address here the fundamental question of basis functions at the centre of quasi-Trefftz methods for the three-dimensional convected Helmholtz equation: the actual construction of basis functions is fundamental to the discretization stage, and therefore to the implementation of the methods, while the approximation properties of the discrete space are a fundamental element in the proof of convergence of the methods.
Since quasi-Trefftz functions satisfy a local quasi-Trefftz property \eqref{eq:TaylExp},
more precisely, the goal of this work is twofold:
\begin{enumerate}
\item developing algorithms for the construction of {\bf local} basis functions for quasi-Trefftz function spaces for the partial differential operator of the three-dimensional convected Helmholtz equation, $\mathbb V_h$, guaranteeing a limited computational cost for the practical construction,
\item studying the {\bf local} approximation property of the resulting spaces $\mathbb V_h$ in the following sense ; given $n\in\mathbb N$, there is a space $\mathbb V_h$ satisfying:
\begin{equation}\label{IntPb}
\forall
\text{ satisfying the governing PDE,
\exists u_a\in\mathbb V_h \text{ s. t. }
\forall \varx\in\mathbb R^3,
|u(\varx)-u_a(\varx)|\leq C | \varx-\pG|^{n+1}.
\end{equation}
\end{enumerate}
Inspired by classical PWs $\exp \Lambda \cdot\Big(\varx-\pG \Big)$, we will focus on three different families of quasi-Trefftz functions:
\begin{itemize}
\item phase-based GPWs, following the original ansatz proposed in \cite{LMD} via the introduction of higher order terms in the phase of a PW, see \eqref{eq:c/GPW},
\item amplitude-based GPWs, following the ansatz proposed in \cite{AbGPW} via the introduction of higher order terms in the amplitude of a PW,
\item purely polynomial quasi-Trefftz functions, which so far we have only used for time-dependent wave propagation in \cite{IGMS}.
\end{itemize}
In each of the wave-based cases the ansatz is an extension of cases studied previously in two-dimensions, whereas the situation is different for the polynomial case. This is the first time that polynomial quasi-Trefftz functions are proposed for time-harmonic problems.
We will pursue the announced goals for these three families of quasi-Trefftz functions, highlighting the similarities and differences between the three cases. The fundamental contribution of this work is to show that these three families of quasi-Trefftz functions achieve the approximation properties with exactly the same number of degrees of freedom as their Trefftz function (wave-based) counterpart do for the constant-coefficient cases studied in the literature.
It is important to note that general time-harmonic wave-propagation equations have no exact polynomial solution, in other words there exist no polynomial Trefftz function in this case. However, there are more quasi-Trefftz functions than Trefftz functions, since the former are defined by a less restrictive constraint, and as we will see there exist polynomial quasi-Trefftz functions.
\subsection{
Preliminaries}
Throughout this article, we will use the following notation.
The canonical basis of $\mathbb R^3$ or $(\mathbb N_0)^3$ is denoted $\{e_k,k\in\{1,2, 3\}\}$, and $|\cdot |$ denotes the euclidean norm on $\mathbb R^3$.
The set of positive integers is denoted $\mathbb N$
and
the set of non-negative integers is denoted $\mathbb N_0 :=\mathbb N\cup \{0\}$,
while the zero multi-index is denoted $\mathbf 0 = (0,0,0)$.
Multi-indices in $(\mathbb N_0)^3$ are denoted $\alpha,\beta,i,j$, the sum of multi-indices is defined as $i+j = (i_1+j_1,i_2+j_2,i_3+j_3)$ for all $i$ and $j $ in $(\mathbb N_0)^3$ while $|\cdot |$ also denotes the length for a multi-index, that is $|i|=i_1+i_2+i_3$ for all $i\in(\mathbb N_0)^3$,
and for the sake of compactness
$i\leq j$ means that $i_k\leq j_k$ for $k\in\{1,2,3\}$,
$i< j$ means that $i\neq j$ and $i_k\leq j_k$ for $k\in\{1,2,3\}$,
and the linear order $\prec$ on $\mathbb N_0^3$ is defined by
$$
\forall (\mu,\nu)\in\left(\mathbb N_0^3\right)^2,
\mu\prec\nu
\Leftrightarrow
\left\{
\begin{array}{l}
|\mu|<|\nu|\ , \text{ or }\\
|\mu|=|\nu| \text{ and } \mu_1<\nu_1\ , \text{ or }\\
|\mu|=|\nu|,\ \mu_1=\nu_1\text{ and } \mu_2<\nu_2.
\end{array}
\right.
$$
The generic point of interest in the domain of the equation is denoted $\pG$.
The coefficient of a Taylor expansion in the neighborhood of $\pG$, for any $n\in\mathbb N$, are denoted $T_ f[\varb] = \frac1{\varb!}\partial_{\varx}^{\varb} f(\pG)$ for all function $f\in\mathcal C^n$ at $\pG$ with $\varb\in\mathbb N_0^3$, $|\varb|\leq n$. We chose to avoid an unnecessary explicit mention of $\pG$ in the $T$ notation because all Taylor expansion will be performed at $\pG$.
\begin{rmk}\label{rmk:SimpRules}
Given $n\in\mathbb N_0$ and $\pG\in\mathbb R^3$, a few simple derivative rules can then be expressed in a compact way as follows:
$$
\left\{\begin{array}{l}
\displaystyle
T_{fg}[\varb] = \sum_{\gamma\leq \varb} T_ f[\varb-\gamma] T_ g[\gamma]
\text{ if }f\text{ and }g\text{ are }\mathcal C^n\text{ at }\pG\text{ with }|\varb|\leq n,
\\
\displaystyle
T_{fgh}[\varb]
= \sum_{\gamma\leq \varb} \sum_{\eta\leq \gamma} T_ f[\varb-\gamma] T_ g[\eta] T_ h[\gamma-\eta],
\\
\displaystyle
T_{\partial_\varx^\alpha f}[\beta]=\frac{(\alpha+\varb)!}{{\beta}!} T_{f}[\alpha+\varb]
\text{ if }f\text{ is }\mathcal C^n\text{ at }\pG\text{ with }|\alpha|+|\varb|\leq n,
\\
T_{(X-\pG)^i}[\varb] = 1(\varb-i).
\end{array}\right.
$$
\end{rmk}
\begin{dfn}
A linear partial differential operator of order $2$, in three dimensions, with a given set of complex-valued functions $c = \{\funcci ; i\in\mathbb N_0^3, |i|\leq 2\}$ will be denoted hereafter as
$$ \mathcal L_{c} := \sum_{i\in\mathbb N_0^3; |i|\leq 2} \funcci\left( \varx \right) \partial_{\varx}^i,$$
where $\varx=(x_1,x_2,x_3)\in\mathbb R^3$ and $\partial_{\varx}^i = \partial_{x_1}^{i_1} \partial_{x_2}^{i_2} \partial_{x_3}^{i_3}$.
\end{dfn}
We will make use of the fact that the set $\{i\in\mathbb N_0^3, |i|\leq 2\}$ can be split as:
$$
\{i\in\mathbb N_0^3, |i|\leq 2\}
=
\{2e_k, 1\leq k\leq 3\}
\cup
\{e_k+e_{k'}, 1\leq k<k'\leq 3\}
\cup
\{e_k, 1\leq k\leq 3\}
\cup
\{ \mathbf 0 \}
$$
For instance, in the case of the convected Helmholtz equation, the variable coefficients of the partial differential operator can then be defined as follows:
$$
\left\{\begin{array}{ll}
\displaystyle c_{2e_k}=\rho\left((\mathbf M_k)^2-1\right) &\text{for }1\leq k\leq 3,\\
\displaystyle c_{e_k+e_l}=\rho\mathbf M_k\mathbf M_l &\text{for }1\leq k<k'\leq 3,\\
\displaystyle c_{e_k} = \rho \sum_l \mathbf M_l\partial_{x_l}\mathbf M_k +\nabla\cdot(\rho\mathbf M)\mathbf M_k-\partial_{x_k}\rho-2\textrm i\kappa\rho\mathbf M_k &\text{for }1\leq k\leq 3,\\
\displaystyle c_ {\mathbf 0} = -\textrm i \kappa \nabla\cdot(\rho\mathbf M)-\rho\kappa^2.
\end{array}\right.
$$
Beyond the convected Helmholtz equation, the work proposed in this article relies on a set of minimal hypotheses for the partial differential operator $\mathcal L_c$.
The first aspect will lead to the well-posedness of a subproblem in the construction of quasi-Trefftz functions, whereas the second will lead to the construction of a set of linearly independent quasi-Trefftz functions.
We gather the two in the following Hypothesis.
\begin{hyp}
\label{hyp:PDop}
Given a point $\pG\in\mathbb R^3$ and a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$,
the functions are
assumed to be $\mathcal C^\infty$ at the point $\pG$, with
$c_{2e_1}(\pG)\neq 0$ and the matrix defined by:
$$
\mathcal C:=\begin{bmatrix}
\coefc{2e_1}[\mathbf 0] & \frac12 \coefc{e_1+e_2}[\mathbf 0] & \frac12 \coefc{e_1+e_3}[\mathbf 0] \\
\frac12 \coefc{e_1+e_2}[\mathbf 0] & \coefc{2e_2}[\mathbf 0] & \frac12 \coefc{e_2+e_3}[\mathbf 0] \\
\frac12 \coefc{e_1+e_3}[\mathbf 0] &\frac12 \coefc{e_2+e_3[\mathbf 0] } & \coefc{2e_3}[\mathbf 0]
\end{bmatrix}
$$
is non-singular.
As a consequence,
there are two real matrices, an orthogonal matrix $\Pmat$ and a non-singular diagonal matrix $\Dmat$, depending only on the set of coefficients $c$ evaluated at $\pG$, such that $\Cmat = \Pmat\Dmat\Pmat^T$.
\end{hyp}
\noindent
The convected Helmholtz operator satisfies \Cref{hyp:PDop} according to the following comments.
\begin{itemize}
\item The density $\rho(\pG)$ is positive, and there is at least one index $k\in\{1,2,3\}$ such that $\mathbf M_k(\pG)\neq 1$ since the Mach number is assumed to be no more than $1$, $|\mathbf M(\pG)|<1$.
Hence $\rho(\pG)\left((\mathbf M_k)^2(\pG)-1\right)\neq 0$, in other words in particular we indeed have $c_{2e_1}\neq 0$.
\item The matrix $\mathcal C$ defined by:
$$
\begin{array}{l}
\begin{bmatrix}
\coefc{2e_1}[\mathbf 0] & \frac12 \coefc{e_1+e_2}[\mathbf 0] & \frac12 \coefc{e_1+e_3}[\mathbf 0] \\
\frac12 \coefc{e_1+e_2}[\mathbf 0] & \coefc{2e_2}[\mathbf 0] & \frac12 \coefc{e_2+e_3}[\mathbf 0] \\
\frac12 \coefc{e_1+e_3}[\mathbf 0] &\frac12 \coefc{e_2+e_3 }[\mathbf 0] & \coefc{2e_3}[\mathbf 0]
\end{bmatrix}
\\
=
\rho(\pG)\begin{bmatrix}
\left( \mathbf M_1(\pG)\right)^2 -1& \frac12 \mathbf M_1(\pG)\mathbf M_2(\pG)& \frac12\mathbf M_1(\pG)\mathbf M_3(\pG)\\
\frac12\mathbf M_1(\pG)\mathbf M_2(\pG)&\left( \mathbf M_2(\pG)\right)^2-1& \frac12\mathbf M_2(\pG)\mathbf M_3(\pG)\\
\frac12\mathbf M_1(\pG)\mathbf M_3(\pG) &\frac12 \mathbf M_2(\pG)\mathbf M_3(\pG)&\left( \mathbf M_3(\pG)\right)^2-1
\end{bmatrix}
,
\end{array}
$$
is related to the second order terms in the partial differential operator,
and the density $\rho(\pG)$ is positive.
Under the assumption that $|\mathbf M(\pG)|<1$ it can be shown that $1/\rho(\pG)\mathcal C$ is non-singular, see appendix \ref{app:eigval}.
Hence since the Mach number is assumed to be no greater than $1$, the matrix $\mathcal C$ is indeed non-singular.
\end{itemize}
Finally, $\left\{\lambda_i,\mu_i,\nu_i, i\in(\mathbb N_0)^3\right\}$ denote complex polynomial coefficients, and, for a given integer $d$, we denote the corresponding polynomials:
$$
P:=\sum_{i\in\mathbb N_0^3, |i|\leq d} \lambda_i \mathbf X^i,\
Q:=\sum_{i\in\mathbb N_0^3, |i|\leq d} \mu_i \mathbf X^i \text{ and }
R:=\sum_{i\in\mathbb N_0^3, |i|\leq d} \nu_i \mathbf X^i, \text{ where }\mathbf X^i = X_1^{i_1} X_2^{i_2} X_3^{i_3}.$$
As a convention, when referring to {\it a polynomial of degree at most equal to $d$} we include the zero polynomial.
\section{Three types of quasi-Trefftz functions}
\label{sec:QTfams}
The quasi-Trefftz property for a function is a property of the image of this function under the action of the partial differential operator. Two fundamental aspects of this property are related to its statement in terms of a Taylor expansion approximation: (i) the fact that it is a local property, and (ii) the fact that it allows for a choice in the desired order of approximation.
Hence, as it relies on enforcing the quasi-Trefftz property, the construction of quasi-Trefftz functions is performed at a given point $\pG$ and constructed functions satisfy the property at a given order of approximation $q$.
The first question is obviously that of the existence of quasi-Trefftz functions.
A second question is nevertheless equally important for the efficient implementation of quasi-Trefftz methods, that of the computational cost of the practical construction of quasi-Trefftz bases.
Indeed, the construct the quasi-Trefftz functions is only a pre-computation to the discretization of a quasi-Trefftz weak formulation, therefore its computational cost must be acceptable compared to that of the assembly of the discrete matrix and the resolution of the linear system.
The former question will be addressed by the derivation of an algorithm for the construction of quasi-Trefftz functions.
The latter question will be settled by the precise steps of the construction algorithm, as the algorithm only applies explicit closed formulas while it requires no numerical resolution of any system.
In this section, we present three types of quasi-Trefftz functions.
\begin{itemize}
\item The original GPW ansatz, namely $\exp \left( \Lambda \cdot\Big(\varx-\pG \Big) + \text{HOT}\right)$, introduced higher order terms in the phase of a PW. The general form of this ansatz can be described as $\exp P(\varx-\pG)$, for some polynomial $P$.
\item In comparison, a new ansatz was proposed \cite{AbGPW} via the introduction of higher order terms in the amplitude of a PW as $(1+\text{HOT}) \exp \Lambda \cdot\Big(\varx-\pG \Big)$. Therefore such an ansatz has the general form $Q(\varx-\pG) \exp \Lambda \cdot\Big(\varx-\pG \Big)$, for some polynomial $Q$ and some $\Lambda$.
\item Moreover, we propose here to consider purely polynomial quasi-Trefftz functions, described as $R(\varx-\pG)$ for some polynomial $R$.
\end{itemize}
To guarantee the existence of such quasi-Trefftz function, we will study the existence of polynomials ($P$, $Q$ or $R$) such that the associated ansatz satisfies the desired quasi-Trefftz property \eqref{eq:TaylExp}.
To do so we will conveniently reformulate the problem to evidence properties of the resulting system, underlining the shared common structure of these systems. These properties will appear to be central to the construction of quasi-Trefftz functions.
\subsection{Amplitude-based GPWs}
Initially, the abstract problem of construction of an amplitude-based GPW can be written, for a given $q\in\mathbb N$, as:
\begin{equation}
\label{eq:AbProb
\left\{
\begin{array}{l}
\text{Find } (\polQ, \Lambda)\in\mathbb C[X_1,X_2,X_3] \times\mathbb C^3 \text{ such that }\\
J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big) \text{ satisfies } \\
\mathcal L_{c} J(\varx) = O(|\varx-\pG|^q)
\end{array}
\right.
\end{equation}
The unknowns here are of two types, either polynomial, for $\polQ$, or scalar, for the three components of $\Lambda$, and the specific role of $\Lambda$ will be highlighted throughout the discussion.
To formulate a more concrete problem, we focus on the action of the differential operator $\mathcal L_{c}$ on the ansatz. If $J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big)$ with $\polQ\in\mathbb C[X_1,X_2,X_3]$ and $ \Lambda\in\mathbb C^3$, then
$$
\begin{array}{rl}
\mathcal L_{c} J(\varx)
=&\displaystyle
\left(
\sum_{k=1}^3
\funcc{2e_k}(\varx)\left(\partial_\varx^{2e_k}\polQ(\varx-\pG) + 2 \Lambda_k \partial_{\varx}^{e_k} \polQ(\varx-\pG) + \Lambda_k^2 \polQ(\varx-\pG)\right) \right.
\\ &\displaystyle
+
\sum_{1\leq k<k'\leq 3}
\funcc{e_k+e_k'}(\varx)\Big(\partial_{\varx}^{e_k+e_{k'}} \polQ(\varx-\pG) +\Lambda_k\partial_{\varx}^{e_{k'}} \polQ(\varx-\pG) +\Lambda_{k'}\partial_{\varx}^{e_{k}} \polQ(\varx-\pG) + \Lambda_k\Lambda_{k'} \polQ(\varx-\pG)\Big)
\\ &\displaystyle\left.
+
\sum_{k=1}^3
\funcc{e_k} (\varx)\Big(\partial_{\varx}^{e_k} \polQ(\varx-\pG) +\Lambda_k \polQ(\varx-\pG) \Big)
+
\funcczr(\varx) \polQ(\varx-\pG)
\right) \exp \Lambda \cdot\Big(\varx-\pG \Big) .
\end{array}
$$
We can then define the partial differential operator
$$
\begin{array}{rl}
\mathcal L_{c}^{Am,\Lambda} \polQ (\varx):= &
\displaystyle
\sum_{k=1}^3
\funcc{2e_k}(\varx)\left(\partial_\varx^{2e_k}\polQ(\varx-\pG) + 2 \Lambda_k \partial_{\varx}^{e_k} \polQ(\varx-\pG) + \Lambda_k^2 \polQ(\varx-\pG)\right)
\\&\displaystyle
+
\sum_{1\leq k<k'\leq 3}
\funcc{e_k+e_k'}(\varx)\Big(\partial_{\varx}^{e_k+e_{k'}} \polQ(\varx-\pG) +\Lambda_k\partial_{\varx}^{e_{k'}} \polQ(\varx-\pG) +\Lambda_{k'}\partial_{\varx}^{e_{k}} \polQ(\varx-\pG) + \Lambda_k\Lambda_{k'} \polQ(\varx-\pG)\Big)
\\& \displaystyle
+
\sum_{k=1}^3
\funcc{e_k} (\varx)\Big(\partial_{\varx}^{e_k} \polQ(\varx-\pG) +\Lambda_k \polQ(\varx-\pG) \Big)
+
\funcczr(\varx) \polQ(\varx-\pG)
\end{array}
$$
to emphasize that
$$
\mathcal L_{c} J(\varx)
= \Big(\mathcal L_{c}^{Am,\Lambda} \polQ (\varx)\Big) \exp \Lambda \cdot\Big(\varx-\pG \Big),
$$
where for $ \Lambda\in\mathbb C^3$ the exponential term is locally bounded.
As a result, for $J$ to satisfy the desired property $\mathcal L_{c} J(\varx) = O(|\varx-\pG|^q)$, it is then sufficient for $\Lambda$ and $\polQ$ to satisfy $\mathcal L_{c}^{Am,\Lambda} \polQ(\varx)= O(|\varx-\pG|^q)$.
Therefore we will formulate a new problem for construction of GPWs
\begin{equation}
\label{prob:Ab}
\left\{
\begin{array}{l}
\text{Find } (\polQ, \Lambda)\in\mathbb C[X_1,X_2,X_3] \times\mathbb C^3 \text{ such that }\\
\mathcal L_{c}^{Am,\Lambda} \polQ(\varx)
= O(|\varx-\pG|^q)
\\\text{then }J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big),
\end{array}
\right.
\end{equation}
and any solution $J$ to \eqref{prob:Ab} will be solution to the initial problem \eqref{eq:AbProb}.
We can now express a concrete problem in terms of scalar equations and scalar unknowns thanks to \Cref{rmk:SimpRules},
the equations corresponding to cancelling the Taylor expansion coefficients of $\mathcal L_{c}^{Am,\Lambda} \polQ$ for orders from 0 to $q-1$ and the unknowns corresponding to all the free parameters defining the GPW namely the $\mu$s and $\Lambda$.
Indeed, \eqref{prob:Ab} can equivalently be stated as follows as long as $d\geq q+1$:
\begin{equation}
\label{prob:Abref
\left\{
\begin{array}{l}
\text{Find } \Lambda\in\mathbb C^3 \text{ and } \{\mu_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq d\} \text{ satisfying } \forall \varb \text{ such that } |\varb|<q \\
\displaystyle
\sum_{\gamma\leq \varb}\left(
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma]\left((\gamma_k+2)(\gamma_k+1)\mu_{\gamma+2e_k} + 2 \Lambda_k (\gamma_k+1)\mu_{\gamma+e_k} + \Lambda_k^2\mu_{\gamma}\right) \right.
\\\displaystyle
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma]\Big((\gamma_k+1)(\gamma_{k'}+1)\mu_{\gamma+e_k+e_{k'}} +\Lambda_k(\gamma_{k'}+1)\mu_{\gamma+e_{k'}} +\Lambda_{k'}(\gamma_{k}+1)\mu_{\gamma+e_k} + \Lambda_k\Lambda_{k'}\mu_\gamma\Big)
\\ \displaystyle\left.
+
\sum_{k=1}^3
\coefc{e_k}[\varb-\gamma]\Big((\gamma_{k}+1)\mu_{\gamma+e_k} +\Lambda_k \mu_\gamma \Big)
+
\coefczr[\varb-\gamma] \mu_\gamma\right) = 0,
\end{array}
\right.
\end{equation}
if $\{\mu_i\}$ is the set of polynomial coefficients of $\polQ$.
The choice $d\geq q+1$ simply ensures that all equations have the same structure. Indeed, for instance, there would be no $\mu_{\varb+2e_k}$ term in $T_{\mathcal L_{c}^{Am,\Lambda} \polQ}[\beta]$ for $|\beta|=q$ if we chose $d<q+1$.
Hence we will always consider the case:
\begin{center}
\fbox{ $d\geq q+1$ }
\end{center}
\subsection{Phase-based GPWs}
Similarly, the abstract problem of construction of a phase-based GPW can initially be written, for a given $q\in\mathbb N$, as:
\begin{equation}
\label{eq:PbProb
\left\{
\begin{array}{l}
\text{Find a polynomial } \polP\in\mathbb C[X_1,X_2,X_3] \text{ such that }\\
G(\varx) := \exp \polP(\varx-\pG) \text{ satisfies } \\
\mathcal L_{c} G(\varx) = O(|\varx-\pG|^q).
\end{array}
\right.
\end{equation}
Thanks to the definition of the partial differential operator
$$
\begin{array}{rl}
\mathcal L_{c}^{Ph} \polP (\varx):= &
\displaystyle
\sum_{k=1 }^3
\funcc{2e_k}(\varx)\left(\partial_{\varx}^{2e_k}\polP(\varx-\pG) + \left(\partial_{\varx}^{e_k} \polP(\varx-\pG)\right)^2\right)
\\&\displaystyle
+
\sum_{1\leq k<k'\leq 3}
\funcc{e_k+e_k'}(\varx)\Big(\partial_{\varx}^{e_k+e_{k'}}\polP(\varx-\pG) + \partial_{\varx}^{e_k}\polP(\varx-\pG)\partial_{\varx}^{e_{k'}} \polP(\varx-\pG)\Big)
\\&\displaystyle
+
\sum_{k=1}^3
\funcc{e_k}(\varx) \partial_{\varx}^{e_k} \polP(\varx-\pG)
+
\funcczr(\varx),
\end{array}
$$
we can easily verify that
$$
\mathcal L_{c} G(\varx)
= \Big(\mathcal L_{c}^{Ph} \polP (\varx)\Big) \exp \polP(\varx-\pG).
$$
Hence any solution $G$ to the problem:
\begin{equation}
\label{prob:Pb}
\left\{
\begin{array}{l}
\text{Find a polynomial } \polP\in\mathbb C[X_1,X_2,X_3] \text{ such that }\\
\mathcal L_{c}^{Ph}\polP(\varx)
= O(|\varx-\pG|^q)
\\\text{then }G(\varx) := \exp \polP(\varx-\pG)
\end{array}
\right.
\end{equation}
will also be a solution to the initial problem \eqref{eq:PbProb}.
In terms of scalar unknowns and equations,
as long as $d\geq q+1$, this is then equivalent to:
\begin{equation}
\label{prob:Pbref
\left\{
\begin{array}{l}
\text{Find } \{\lambda_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq d\} \text{ satisfying } \forall \varb \text{ such that } |\varb|<q \\
\displaystyle
\sum_{\gamma\leq \varb}\left(
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma]\left((\gamma_k+2)(\gamma_k+1)\lambda_{\gamma+2e_k} + \sum_{\eta\leq \gamma}(\gamma_{k}-\eta_{k} +1)\lambda_{\gamma-\eta+e_{k}}(\eta_k +1)\lambda_{\eta+e_k}\right) \right.
\\\displaystyle
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma]\left((\gamma_k+1)(\gamma_{k'}+1)\lambda_{\gamma+e_k+e_{k'}} + \sum_{\eta\leq \gamma}(\gamma_{k'}-\eta_{k'} +1)\lambda_{\gamma-\eta+e_{k'}}(\eta_k +1)\lambda_{\eta+e_k}\right)
\\ \displaystyle\left.
+
\sum_{k=1}^3
\coefc{e_k}[\varb-\gamma](\gamma_{k}+1)\lambda_{\gamma+e_k} \right)
+
\coefczr[\varb]= 0,
\end{array}
\right.
\end{equation}
if $\{\lambda_i\}$ is the set of polynomial coefficients of $\polP$.
Similarly here the choice $d\geq q+1$ simply ensures that all equations have the same structure.
\subsection{Polynomial functions}
The abstract problem of construction of a purely polynomial quasi-Trefftz function can simply be written, for a given $q\in\mathbb N$, as:
\begin{equation}
\label{eq:PolProb
\left\{
\begin{array}{l}
\text{Find a polynomial } \polR\in\mathbb C[X_1,X_2,X_3] \text{ such that } \\
H(\varx) := \polR(\varx-\pG) \text{ satisfies } \\
\mathcal L_{c} H(\varx) = O(|\varx-\pG|^q).
\end{array}
\right.
\end{equation}
In terms of scalar unknowns and equations, as long as $d\geq q+1$ to ensure again that all equations have the same structure, this is equivalent to:
\begin{equation}
\label{prob:Polref
\left\{
\begin{array}{l}
\text{Find } \{\nu_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq d\} \text{ satisfying } \forall \varb \text{ such that } |\varb|<q \\
\displaystyle
\sum_{\gamma\leq \varb}\left(
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma](\gamma_k+2)(\gamma_k+1)\nu_{\gamma+2e_k} \right.
\\\displaystyle
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma](\gamma_k+1)(\gamma_{k'}+1)\nu_{\gamma+e_k+e_{k'}}
\\ \displaystyle\left.
+
\sum_{k=1}^3
\coefc{e_k}[\varb-\gamma](\gamma_{k}+1)\nu_{\gamma+e_k}
+
\coefczr[\varb-\gamma]\nu_\gamma \right) = 0,
\end{array}
\right.
\end{equation}
if $\{\nu_i\}$ is the set of polynomial coefficients of $\polR$.
\subsection{Common structure}
These systems share common aspects but also exhibit differences.
We will leverage the former to derive very similar construction algorithms for the three types of quasi-Trefftz functions.
The unknowns in both systems \eqref{prob:Abref}, \eqref{prob:Pbref} and \eqref{prob:Polref} include the $(d+1)(d+2)(d+3)/6$ polynomial coefficients, respectively $\{\mu_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq d\}$,$\{\lambda_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq d\}$, and $\{\nu_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq d\}$, while only in the amplitude-based case there are three additional scalar unknowns, $\Lambda\in\mathbb C^3$. In the polynomial case the system is linear, whereas in both GPW cases the systems are non-linear. However in the amplitude-based case the non-linear terms are limited to products of one $\mu_i$ and powers of $\Lambda_k$s.
Besides, each system has $q(q+1)(q+2)/6$ equations, and we will now describe their common layer structure. A close inspection of the equations reveals an underlying structure linked to the unknowns' and equations' multi-indices. Indeed, for any equation $\beta$, unknowns $\mu_i$ or $\lambda_i$ with $|i|\leq|\beta|+1$ may appear in non-linear terms, whereas unknowns $\mu_i$, $\lambda_i$ or $\nu_i$ with $|i|=|\beta|+2$ can only appear in linear terms. This is summarized in the following two tables.
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|c|}
\hline
Amplitude-based & Phase-based& Indices&Comments\\\hline\hline
$\Lambda_{k}\mu_{\gamma+e_k}$& $\lambda_{\gamma-\eta+e_{k}}\lambda_{\eta+e_k}$&
$\eta\leq \gamma\leq \varb$ and $1\leq k\leq 3$&$\begin{array}{c}|\gamma+e_{k }|\leq |\varb|+1\\|\eta+e_{k }|\leq|\varb|+1\\|\gamma-\eta+e_{k }|\leq|\varb|+1\end{array}$
\\\hline
$ \Lambda_k^2\mu_{\gamma}$ &&
$\gamma\leq \varb$ and $1\leq k\leq 3$&$|\gamma|\leq |\varb|$
\\\hline
$\begin{array}{c}\Lambda_{k}\mu_{\gamma+e_{k'}}\\ \Lambda_{k'}\mu_{\gamma+e_k} \end{array}$&$\lambda_{\gamma-\eta+e_{k'}}\lambda_{\eta+e_k}$&
$\eta\leq \gamma\leq \varb$ and $1\leq k<k'\leq 3$&$\begin{array}{c}|\gamma+e_{k }|\leq|\varb|+1\\|\gamma+e_{k^{'} }|\leq|\varb|+1\\|\gamma-\eta+e_{k^{'} }|\leq|\varb|+1\\|\eta+e_{k^{'} }|\leq|\varb|+1\end{array}$
\\\hline
$\Lambda_k\Lambda_{k'}\mu_\gamma$& &
$\gamma\leq \varb$ and $1\leq k<k'\leq 3$&$|\gamma|\leq|\varb|$
\\\hline
$\Lambda_k \mu_\gamma$& & $ \gamma\leq \varb$ and $1\leq k\leq 3$&$|\gamma|\leq|\varb|$
\\\hline
\end{tabular}}\\
{Non-linear terms in amplitude-base and phase-base \Cref{prob:Abref,prob:Pbref} for a given $\varb$.
}
\end{center}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|c|c|}
\hline
Ampl.-based & Phase-based&Polynomial& Indices&Comments\\\hline\hline
$\mu_{\gamma+2e_k} $&
$\lambda_{\gamma+2e_k} $&
$\nu_{\gamma+2e_k} $&
$ \gamma\leq \varb$ and $1\leq k\leq 3$&$|\gamma+2e_k|=|\varb|+2$ if $\gamma=\varb$\\
&&&& otherwise $|\gamma+2e_k|\leq|\varb|+1$
\\\hline
$\mu_{\gamma+e_k+e_k'}$&$\lambda_{\gamma+e_k+e_k'}$& $\nu_{\gamma+e_k+e_k'}$&
$\gamma\leq \varb$ and $1\leq k<k'\leq 3$&$|\gamma+e_k+e_k'|=|\varb|+2$ if $\gamma=\varb$\\
&&&& otherwise $|\gamma+2e_k|\leq|\varb|+1$
\\\hline
$\mu_{\gamma+e_k}$&$\lambda_{\gamma+e_k}$ & $\nu_{\gamma+e_k}$&
$ \gamma\leq \varb$ and $1\leq k\leq 3$&$|\gamma+e_k|\leq|\varb|+1$
\\\hline
$\mu_{\gamma}$&& $\nu_{\gamma}$&
$ \gamma\leq \varb$&$|\gamma|\leq|\varb|+1$
\\\hline
\end{tabular}}\\
{Linear terms in amplitude-base, phase-base and polynomial \Cref{prob:Abref,prob:Pbref,prob:Polref} for a given $\varb$.}
\end{center}
We have seen that in each
case choosing $d\geq q+1$ ensures that all equations of the system share a common structure. It is straightforward to see from these tables that none of the $\mu$, $\lambda$ or $\nu$ unknowns with indices $i\in\mathbb N_0^3$ such that $|i|>q+1$ appear in the system since $|\varb|\leq q-1$. Hence these unknowns are not constrained by the system: their values do not affect the system - and therefore neither do they affect the quasi-Trefftz property - even though they would of course affect the definition of the corresponding quasi-Trefftz function.
As a result, it is sufficient to seek a polynomial $P$, $Q$ and $R$ of degree $d$ satisfying:
\begin{center}
\fbox{ $d=q+1$ }
\end{center}
As we can see from the {\it Comments} columns of the previous tables, all non-linear terms involve unknowns with multi-indices of length at most equal to $|\varb|+1$, while the only unknowns with multi-indices of length $|\varb|+2$ are $\mu_{\varb+e_k+e_k'}$, $\lambda_{\varb+e_k+e_k'}$ or $\nu_{\varb+e_k+e_k'}$ with $1\leq k\leq k'\leq 3$.
The length of multi-indices then plays an important role in the structure of the system.
Hence, in the index space $(\mathbb N_0)^3$, we describe the set of multi-indices of a given length $\ell$ as a layer, as illustrated in \Cref{fig:IndexLayer}.
In order to take advantage of the systems layer structure, we will now split the sets of equations and unknowns according to their multi-index lengths.
\newcommand{\draw[dashed,thick]}{\draw[dashed,thick]}
\newcommand{\NodeThreeD}[3]{\draw (#1,#2,#3) node[circle,fill,inner sep=2pt] {};
\draw[dashed,thick](#1,#2,0)--(#1,#2,#3);\draw[dashed,thick](0,#2,0)--(#1,#2,0);\draw[dashed,thick](#1,0,0)--(#1,#2,0);
\draw[dashed,thick](#1,0,#3)--(#1,#2,#3);\draw[dashed,thick](0,0,#3)--(#1,0,#3);\draw[dashed,thick](#1,0,0)--(#1,0,#3);
\draw[dashed,thick](0,#2,#3)--(#1,#2,#3);\draw[dashed,thick](0,#2,0)--(0,#2,#3);\draw[dashed,thick](0,0,#3)--(0,#2,#3);}
\begin{figure}[htb]\centering
\resizebox{.49\linewidth}{!}{
\begin{tikzpicture}
[rotate around x=-90,rotate around y=0,rotate around z=-30,grid/.style={very thin,gray}]
\foreach \x in {0,1,...,7}
\foreach \y in {0,1,...,7}
\foreach \z in {0,1,...,7}
{
\draw[grid,lightgray] (\x,0,0) -- (\x,7,0);
\draw[grid,lightgray] (0,\y,0) -- (7,\y,0);
\draw[grid,lightgray] (0,\y,0) -- (0,\y,7);
\draw[grid,lightgray] (0,0,\z) -- (0,7,\z);
}
\draw [->] (0,0) -- (8,0,0) node [right] {$\indn_{1}$};
\draw [->] (0,0) -- (0,8,0) node [above] {$\indn_{2}$};
\draw [->] (0,0) -- (0,0,8) node [below left] {$\indn_3$};
\draw(0,0)node[left]{$(0,0,0)$};
\draw[fill=blue,opacity=.5] (0,7,0)--(0,0,7)--(7,0,0);
\NodeThreeD313;
\draw[] (3,1,3) node[color=white,circle,fill] {};
\draw (3,0,0) node[below left] {$3$};
\draw (0,1,0) node[above left] {$1$};
\draw (0,0,3) node[left] {$3$};
\draw (0,0,7) node[left] {$7$};
\draw (0,7,0) node [above] {$7$};
\draw (7,0,0) node[below left] {$7$};
\foreach \x in {0,1,...,7}
{
\pgfmathsetmacro{\int}{7-\x}
\foreach \y in {0,...,\int}
{
\draw (\x,\y,7-\x-\y) node[circle,fill=blue,inner sep=2pt] {};
}
}
\end{tikzpicture}
}
\caption{Illustration of a multi-index layer in $(\mathbb N_0)^3$. The layer $\{\indn\in\mathbb N_0^3, |\indn|=\ell\}$ for $\ell =7$ is represented in blue. All elements in the layer are represented as blue dots, the element $\indn=(3,1,3)$ is highlighted in white.}
\label{fig:IndexLayer}
\end{figure}
Let's consider, for $\ell\in\mathbb N_0$ with $\ell<q$, the subset of equations corresponding to $|\varb|=\ell$. From our previous observations we know that all the terms involving unknowns with multi-index length equal to $\ell+2$, namely $\mu_{\varb+e_k+e_k'}$, $\lambda_{\varb+e_k+e_k'}$ or $\nu_{\varb+e_k+e_k'}$ with $1\leq k\leq k'\leq 3$, are linear.
Hence if unknowns with a shorter multi-index -- and the $\Lambda$ unknowns in the amplitude-based case -- were already known, it would suggest, for $\ell\in\mathbb N_0$ with $\ell<q$, to define a linear underdetermined subsystem with:
\begin{itemize}
\item $\frac{(\ell+1)(\ell+2)}{2}$ equations, namely the equations corresponding to $\varb$ with $|\varb|=\ell$,
\item $\frac{(\ell+3)(\ell+4)}{2}$ unknowns, namely the unknowns $\{ \mu_i,|i|=\ell+2 \}$, $\{ \lambda_i,|i|=\ell+2 \}$ or $\{ \nu_i,|i|=\ell+2 \}$,
\item a right hand side depending on $\{ \mu_i,|i|<\ell+2 \}\cup \Lambda$, $\{ \lambda_i,|i|<\ell+2 \}$, or $\{ \nu_i,|i|<\ell+2 \}$.
\end{itemize}
To ensure that the right hand side is known at each layer $\ell$, it is then only natural to proceed layer by layer for increasing values of $\ell$ from $0$ to $q-1$.
The construction of a solution to the initial system, \eqref{prob:Abref}, \eqref{prob:Pbref} or \eqref{prob:Polref}, then boils down to the successive construction of a solution to each subsystem.
In each case, a set of subsystems gathers $\displaystyle 1/2 \sum_{\ell=0}^{q-1} (\ell+1)(\ell+2) = q(q+1)(q+2)/6$ equations, so that is exactly the full set of equations of the initial system.
From the point of view of unknowns the situation is different. Aside from the $\displaystyle 1/2\sum_{\ell=0}^{q-1} (\ell+3)(\ell+4)= q(q^2+9q+26)/6$ unknowns appearing in the combined subsystems,
we immediately notice that the unknowns $\{ \mu_i,|i|\leq 1 \}$, $\Lambda$, $\{ \lambda_i,|i|\leq 1 \}$, and $\{ \nu_i,|i|\leq 1 \}$ do not belong to any set of subsystem unknowns, but only appear in right hand sides of the subsystems.
So the subsystems' solvability won't be affected by these terms, yet their values need to be fixed in order for the subsystem's right hand sides to be known.
The construction of a solution to the initial system will hence start from setting the values of $\{ \mu_i,|i|\leq 1 \}$ and $\Lambda$, the values of $\{ \lambda_i,|i|\leq 1 \}$ or the values of $\{ \nu_i,|i|\leq 1 \}$ before turning to the hierarchy of subsystems for increasing values of $\ell$ from $0$ to $q-1$.
Not only do the amplitude and phase based cases share the same layer structure, but their subsystems also share the same structure. Indeed, independently of the case, for a given layer $\ell$, the subsystem reads as
\begin{equation}
\label{gensubsys
\left\{
\begin{array}{l}
\text{Find } \{\xi_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|=\ell+2\} \text{ satisfying } \forall \varb \text{ such that } |\varb|=\ell \\
\displaystyle
\sum_{k=1}^3
(\varb_k+2)(\varb_k+1)\coefc{2e_k}[\mathbf 0]\xi_{\varb+2e_k}
\\\displaystyle
+
\sum_{1\leq k<k'\leq 3}
(\varb_k+1)(\varb_{k'}+1)\coefc{e_k+e_{k'}}[\mathbf 0]\xi_{\varb+e_k+e_{k'}} = \mathsf B_\varb,
\end{array}
\right.
\end{equation}
where the right hand side $B$ depends not only on the case but also on variable coefficients of the PDE.
As a consequence, the study of existence of solutions to these subsystems is independent of the case.
These subsystems will be the backbone of the construction algorithm for both families of GPWs.
\begin{rmk}
\label{rem:pol}
Gathering unknowns according to the length of their index, $|i|$,
is related to splitting the unknowns from the polynomial $P$, $Q$ or $R$ according to the total degree of each monomial:
\begin{equation*}
P = \sum_{\ell' = 0}^{q+1}
\left(\sum_
|i| = \ell'} \lambda_i \mathbf X^i \right),\
Q = \sum_{\ell' = 0}^{q+1}
\left(\sum_
|i| = \ell'} \mu_i \mathbf X^i \right),
\text{ or }
R = \sum_{\ell' = 0}^{q+1}
\left(\sum_
|i| = \ell'} \nu_i \mathbf X^i \right),
\end{equation*}
and the $\ell$th subsystem is related to certain derivatives of homogeneous polynomials of degree $\ell'=\ell+2$.
\end{rmk}
The subsystems are linear and underdetermined. Their right hand sides depend on the Taylor expansion coefficients of the set of complex-valued PDE coefficients $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$, as well as other unknowns hopefully previously computed.
Let's now turn to the question of existence of solutions to each subsystem.
\subsection{Subsystems}
To proceed, following \Cref{rem:pol}, we will denote by $\CoPo= \mathbb C[\mathbf X]$ the space of complex polynomials in three variables, $\mathbf X = (X,Y,Z)$, and by $\HoPo{d}\subset \CoPo$ the space of homogeneous polynomials of degree $d$.
In order to prove the existence of a solution to each linear subsystem, we therefore introduce the partial differential operator
\begin{equation*}
\begin{array}{rccc}
\Delta_{c,\ell}: &\HoPo{\ell+2}&\rightarrow &\HoPo{\ell}\\
&f &\mapsto& \Delta_c f
\end{array}
\end{equation*}
where, given the set of complex-valued PDE coefficients $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$, the linear operator $\Delta_c$ is defined on $\mathbb A$ by
\begin{equation*}
\left\{
\begin{array}{l}
\begin{array}{rl}
\displaystyle
\Delta_c \left[\sum_{i\in(\mathbb N_0)^3,|i|<2} \xi_i \mathbf X^i\right] &\displaystyle
=
0
\end{array}
\\
\begin{array}{rl}
\displaystyle
\text{if } d\geq 2,
\Delta_c \left[\sum_{i\in(\mathbb N_0)^3, |i|\leq d} \xi_i \mathbf X^i\right] &\displaystyle
=
\sum_{|\varb|=d-2}\left( \sum_{k=1}^3 (\varb_k+2)(\varb_k+1) \coefc{2e_k}[\mathbf 0] \xi_{\varb+2e_k} \right.
\\&\displaystyle \left.
+
\sum_{1\leq k< k'\leq 3} (\varb_k+1)(\varb_{k'}+1) \coefc{e_k+e_{k'}}[\mathbf 0] \xi_{\varb+e_k+e_{k'}}
\right)\mathbf X^\varb
\end{array}
\end{array}
\right.
\end{equation*}
The existence of solutions to \eqref{gensubsys} is equivalent to the surjectivity of the operator $\Delta_{c,\ell}$. While $\dim \HoPo{\ell+2} = (\ell+3)(\ell+4)/2 $ and $\dim \HoPo{\ell}= (\ell+1)(\ell+2)/2$, unfortunately, it is not simple here to express explicitly the kernel of $\Delta_{c,\ell}$ to find its dimension, unlike for the 2D Laplacian operator. However, we can evidence the operator's full-rank by exhibiting the echelon structure of a particular matrix representation of $\Delta_{c,\ell}$ in appropriate bases.
To describe such a matrix, we will use the canonical bases of $\HoPo{L}$ for $L\in\{\ell,\ell+2\}$, namely $\{ \mathbf X^i, i\in\mathbb N_0^3 \text{ and }|i|=L \}$, and we will number the columns for $L=\ell+2$ and the rows for $L=\ell$ according to the linear order $\prec$ defined for multi-indices $i\in\mathbb N_0^3$ with a given length $L$ by
$$\forall (i,j)\in\left(\mathbb N_0^3\right)^2,i\prec j\text{ if }
\left\{\begin{array}{l}
i_1<j_1 ; \text{ or}\\
i_1=j_1 \text{ and }i_2<j_2,
\end{array}\right.
$$
while $|i|=L$ and $|j|=L$, so $i_3 = L-i_1-i_2$ and similarly $j_3=L-j_1-j_2$. On each row $\varb$ of the matrix, the only non-zero terms are
\begin{equation*}
{\renewcommand{\arraystretch}{1.7}
\begin{array}{|c|c|}
\hline
\text{Column index} & \text{Matrix entry}
\\\hline
\beta+2e_k \text{ with } 1\leq k \leq 3 & (\varb_k+2)(\varb_k+1) \coefc{2e_k}[\mathbf 0]
\\\hline
\varb+e_k+e_{k'} \text{ with } 1\leq k< k'\leq 3 & (\varb_k+1)(\varb_{k'}+1) \coefc{e_k+e_{k'}}[\mathbf 0]
\\\hline
\end{array}
}\
\end{equation*}
One can easily verify that
\begin{equation}
\label{eq:orderIndices}
\varb + 2e_3
\prec \varb + e_2+e_3
\prec \varb + 2e_2
\prec \varb + e_1+e_3
\prec \varb + e_1+e_2
\prec \varb + 2e_1,
\end{equation}
therefore the last nonzero entry on each row $\varb$ stands in column $\varb+2e_1$, and the entry is $ (\varb_1+2)(\varb_1+1) \coefc{2e_1}[\mathbf 0] $. So, under the simple assumption that $ \coefc{2e_1}[\mathbf 0] \neq 0$ (as included in \Cref{hyp:PDop}), the last nonzero entries for any two rows $\varb$ and $\widetilde \varb$ with $\varb \prec \widetilde \varb$ are respectively in columns $\varb+2e_1$ and $\widetilde \varb+2e_1$. Since $\varb+2e_1 \prec \widetilde \varb+2e_1$ it shows the echelon structure of the matrix, which implies that the matrix has full rank.
Back to Subsystem \eqref{gensubsys}, this guarantees the existence of solutions for any right hand side $B\in \mathbb C^{(\ell+1)(\ell+2)/2}$.
\begin{rmk}
\label{rmk:nonSym}
By deciding for a numbering scheme we broke the symmetry between the three component indices. Of course, under the assumption that $ \coefc{2e_2}[\mathbf 0] \neq 0$, or $ \coefc{2e_3}[\mathbf 0] \neq 0$, we could choose an index numbering scheme for which the corresponding matrix would have a similar echelon structure.
\end{rmk}
The space spanned by the $\frac{(\ell+3)(\ell+4)}{2}$ columns of the matrix is of dimension $\frac{(\ell+1)(\ell+2)}{2}$. The columns displaying a $ (\varb_1+2)(\varb_1+1) \coefc{2e_1}[\mathbf 0] $ entry form a linearly independent set, since they are in echelon form. Their indices are $\{\varb+2e_1,\varb\in(\mathbb N_0)^3,|\varb|=\ell\}$, which is equivalent to $\{i\in(\mathbb N_0)^3,|i|=\ell+2, i_1>1 \}$ as represented in \Cref{fig:layer}.
Hence the indices of the remaining columns, i.e. the columns which do not display a $ (\varb_1+2)(\varb_1+1) \coefc{2e_1}[\mathbf 0] $ entry, are simply $\{i\in(\mathbb N_0)^3,|i|=\ell+2, i_1\leq1 \}$.
Accordingly, for a given right hand side, we can take advantage of the echelon structure to compute a solution to the subsystem \eqref{gensubsys}, by simply fixing first the values of the $\{\xi_{i}, i\in(\mathbb N_0)^3,|i|=\ell+2, i_1\leq1\}$ unknowns, and then solving by substitution the resulting square triangular system for the remaining unknowns $\{\xi_{i}, i\in(\mathbb N_0)^3,|i|=\ell+2, i_1>1\}$. See \Cref{Algo:SubS}.
\begin{figure}
\centering
\resizebox{.8\linewidth}{!}{
\begin{tikzpicture}
\def1{1}
\def0.866025{0.866025}
\def0.5{0.5}
\def-0.5{-0.5}
\def0.866025{0.866025}
\def\lev{5
\pgfmathsetmacro{\levt}{\lev+2}
\def9{0}
\pgfmathsetmacro{\iniy}{1}
\pgfmathsetmacro{\iendb}{9+0.5*\levt}
\pgfmathsetmacro{\jendb}{\iniy+0.866025*\levt}
\pgfmathsetmacro{\ienda}{9-0.5*\levt}
\pgfmathsetmacro{\jenda}{\iniy+ 0.866025*\levt}
\pgfmathsetmacro{\axdirx}{0.5}
\pgfmathsetmacro{\axdiry}{0.866025}
\pgfmathsetmacro{\aydirx}{-0.5}
\pgfmathsetmacro{\aydiry}{0.866025}
\pgfmathsetmacro{\fina}{\lev+1}
\foreach \x in {0,...,\fina}
{
\pgfmathsetmacro{\xs}{9+\x*\axdirx
\pgfmathsetmacro{\ys}{\iniy+\x*\axdiry}
\pgfmathsetmacro{\xe}{\xs+(\levt-\x)*\aydirx}
\pgfmathsetmacro{\ye}{\ys+(\levt-\x)*\aydiry}
\draw[thick,{color=orange}] (\xs,\ys) -- (\xe,\ye);
}
\foreach \y in {0,...,\fina}
{
\pgfmathsetmacro{\xs}{9+\y*\aydirx
\pgfmathsetmacro{\ys}{\iniy+\y*\aydiry}
\pgfmathsetmacro{\xe}{\xs+(\levt-\y)*\axdirx}
\pgfmathsetmacro{\ye}{\ys+(\levt-\y)*\axdiry}
\draw[thick,dotted,{color=red}] (\xs,\ys) -- (\xe,\ye);
}
\foreach \z in {0,...,\fina}
{
\pgfmathsetmacro{\xs}{9+(\levt-\z)*\axdirx
\pgfmathsetmacro{\ys}{\iniy+(\levt-\z)*\axdiry}
\pgfmathsetmacro{\xe}{9-(\levt-\z)*\axdirx}
\pgfmathsetmacro{\ye}{\ys}
\draw[thick,dashed,{color=brown}] (\xs,\ys) -- (\xe,\ye);
}
\pgfmathsetmacro{\betax}{3}
\pgfmathsetmacro{\betay}{1}
\pgfmathsetmacro{\betaz}{1}
\pgfmathsetmacro{\xs}{9 +(\betax+2)*\axdirx+(\betay+0)*\aydirx}
\pgfmathsetmacro{\ys}{\iniy+ (\betax+2)*\axdiry+(\betay+0)*\aydiry}
\draw[] (\xs,\ys) node[color=blue,diamond,fill,inner sep=4pt] {};
\draw[] (\xs,\ys) node[color=black,circle,fill,inner sep=2pt] {};
\pgfmathsetmacro{\xs}{9 +(\betax+0)*\axdirx+(\betay+2)*\aydirx}
\pgfmathsetmacro{\ys}{\iniy+ (\betax+0)*\axdiry+(\betay+2)*\aydiry}
\draw[] (\xs,\ys) node[color=black,circle,fill,inner sep=2pt] {};
\pgfmathsetmacro{\xs}{9 +(\betax+0)*\axdirx+(\betay+0)*\aydirx}
\pgfmathsetmacro{\ys}{\iniy+ (\betax+0)*\axdiry+(\betay+0)*\aydiry}
\draw[] (\xs,\ys) node[color=black,circle,fill,inner sep=2pt] {};
\pgfmathsetmacro{\xs}{9 +(\betax+1)*\axdirx+(\betay+0)*\aydirx}
\pgfmathsetmacro{\ys}{\iniy+ (\betax+1)*\axdiry+(\betay+0)*\aydiry}
\draw[] (\xs,\ys) node[color=black,circle,fill,inner sep=2pt] {};
\pgfmathsetmacro{\xs}{9 +(\betax+0)*\axdirx+(\betay+1)*\aydirx}
\pgfmathsetmacro{\ys}{\iniy+ (\betax+0)*\axdiry+(\betay+1)*\aydiry}
\draw[] (\xs,\ys) node[color=black,circle,fill,inner sep=2pt] {};
\pgfmathsetmacro{\xs}{9 +(\betax+1)*\axdirx+(\betay+1)*\aydirx}
\pgfmathsetmacro{\ys}{\iniy+ (\betax+1)*\axdiry+(\betay+1)*\aydiry}
\draw[] (\xs,\ys) node[color=black,circle,fill,inner sep=2pt] {};
\pgfmathsetmacro{\x}{9+6*\axdirx
\pgfmathsetmacro{\xm}{\x-0.025cm
\pgfmathsetmacro{\xp}{\x-0.005cm
\pgfmathsetmacro{\ya}{\iniy+3*\axdiry
\pgfmathsetmacro{\yb}{\iniy+2*\axdiry
\pgfmathsetmacro{\yc}{\iniy+1*\axdiry
\draw[thick,{color=orange}] (\xm,\ya) -- (\xp,\ya);
\draw (\x,\ya) node[anchor=west,font=\huge,{color=orange}] {$i_1$ constant};
\draw[thick,{color=red},dotted] (\xm,\yb) -- (\xp,\yb);
\draw (\x,\yb) node[anchor=west,font=\huge,{color=red}] {$i_2$ constant};
\draw[thick,{color=brown},dashed] (\xm,\yc) -- (\xp,\yc);
\draw (\x,\yc) node[anchor=west,font=\huge,{color=brown}] {$i_3$ constant};
\def9{9}
\pgfmathsetmacro{\iendb}{9+0.5*\levt}
\pgfmathsetmacro{\jendb}{\iniy+0.866025*\levt}
\pgfmathsetmacro{\ienda}{9-0.5*\levt}
\pgfmathsetmacro{\jenda}{\iniy+ 0.866025*\levt}
\pgfmathsetmacro{\axdirx}{0.5}
\pgfmathsetmacro{\axdiry}{0.866025}
\pgfmathsetmacro{\aydirx}{-0.5}
\pgfmathsetmacro{\aydiry}{0.866025}
\pgfmathsetmacro{\fina}{\lev+1}
\foreach \x in {0,...,\fina}
{
\pgfmathsetmacro{\xs}{9+\x*\axdirx
\pgfmathsetmacro{\ys}{\iniy+\x*\axdiry}
\pgfmathsetmacro{\xe}{\xs+(\levt-\x)*\aydirx}
\pgfmathsetmacro{\ye}{\ys+(\levt-\x)*\aydiry}
\draw[thick,{color=orange}] (\xs,\ys) -- (\xe,\ye);
}
\foreach \y in {0,...,\fina}
{
\pgfmathsetmacro{\xs}{9+\y*\aydirx
\pgfmathsetmacro{\ys}{\iniy+\y*\aydiry}
\pgfmathsetmacro{\xe}{\xs+(\levt-\y)*\axdirx}
\pgfmathsetmacro{\ye}{\ys+(\levt-\y)*\axdiry}
\draw[thick,dotted,{color=red}] (\xs,\ys) -- (\xe,\ye);
}
\foreach \z in {0,...,\fina}
{
\pgfmathsetmacro{\xs}{9+(\levt-\z)*\axdirx
\pgfmathsetmacro{\ys}{\iniy+(\levt-\z)*\axdiry}
\pgfmathsetmacro{\xe}{9-(\levt-\z)*\axdirx}
\pgfmathsetmacro{\ye}{\ys}
\draw[thick,dashed,{color=brown}] (\xs,\ys) -- (\xe,\ye);
}
\pgfmathsetmacro{\finaa}{\lev+2}
\foreach \xx in {2,...,\finaa}
{
\pgfmathsetmacro{\finb}{\lev+2-\xx}
\foreach \yy in {0,...,\finb}
{
\pgfmathsetmacro{\xs}{9 +\xx*\axdirx+\yy*\aydirx}
\pgfmathsetmacro{\ys}{\iniy+ \xx*\axdiry+\yy*\aydiry}
\draw (\xs,\ys) node[circle,fill=blue,diamond,inner sep=2pt] {}
}}
\pgfmathsetmacro{\x}{9+6*\axdirx
\pgfmathsetmacro{\xm}{\x-0.025cm
\pgfmathsetmacro{\xp}{\x-0.005cm
\pgfmathsetmacro{\ya}{\iniy+3*\axdiry
\end{tikzpicture}
}
\caption{Two representations of a layer of indices $i$ of the unknowns of the $\ell$th subsystem for $\ell = 5$, corresponding to the blue layer in \Cref{fig:IndexLayer}. Each grid point corresponds to one index $i\in(\mathbb N_0)^3$ with $|i|=\ell+2$.
Left: Indices of the unknowns involved in equation $\varb = (3,1,1)$ are highlighted with black circles.
The index $\varb+2e_1 = (5,1,1)$ is highlighted as a blue diamond.
Right: Indices corresponding to $\varb+2e_1 $ for all $\varb\in(\mathbb N_0)^3$ such that $|\varb|=\ell$.
}
\label{fig:layer}
\end{figure}
\subsection{Construction of
quasi-Trefftz function
}
Given a point $\pG\in\mathbb R^3$ and a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ satisfying \Cref{hyp:PDop}, we can now turn back to the construction of solutions to Systems \eqref{prob:Abref}, \eqref{prob:Pbref} and \eqref{prob:Polref}, and hence the construction of quasi-Trefftz functions.
\Cref{Algo:SubS} summarizes one way to compute $\xi= \{\xi_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|=\ell+2\}$ solution of a subsytem \eqref{gensubsys} for a given right hand side $\mathsf B^\ell= \{\mathsf B^\ell_{\varb}\in \mathbb C;\varb\in(\mathbb N_0)^3, |\varb|=\ell\}$.
Remarkably, \Cref{Algo:PbS,Algo:AbS,Algo:PolS} build solutions to the non-linear problems \eqref{prob:Abref}, \eqref{prob:Pbref} and \eqref{prob:Polref} while relying exclusively on explicit closed formulas.
\begin{algorithm}[H]
\caption{$\xi=$ solve\_subsystem$\left(\ell,\mathsf B^\ell,\left\{\coefc{e_k+e_{k'}}[\mathbf 0], 1\leq k\leq k'\leq 3\right\}\right)$}
\label{Algo:SubS}
\begin{algorithmic}[1]
\State Fix $\{\xi_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|=\ell+2, i_1\in\{0,1\}\}$
\For{$\varb_1 \gets 0$ to $ \ell$}
\For{$\varb_2 \gets 0$ to $\ell-\varb_1$}
\State $\varb := (\varb_1,\varb_2,\ell-\varb_1-\varb_2)$
\State
$\displaystyle
\begin{array}{rr}
\xi_{\varb+2e_1} := &\displaystyle
\frac{1}{(\varb_1+2)(\varb_1+1)\coefc{2e_1}[\mathbf 0]}\Bigg(\mathsf {B^\ell_{\varb}}-\sum_{k=2}^3
(\varb_k+2)(\varb_k+1)\coefc{2e_k}[\mathbf 0]\xi_{\varb+2e_k}
\\&\displaystyle
-\sum_{1\leq k<k'\leq 3}
(\varb_k+1)(\varb_{k'}+1)\coefc{e_k+e_{k'}}[\mathbf 0]\xi_{\varb+e_k+e_{k'}}
\Bigg)
\end{array}
$
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
Thanks to \Cref{Algo:SubS}, \Cref{Algo:AbS,Algo:PbS} will compute a solution to systems \eqref{prob:Abref} and \eqref{prob:Pbref} and then construct the associated GPWs $J$ and $G$, solutions to the initial problems \eqref{eq:AbProb} and \eqref{eq:PbProb}, while \Cref{Algo:PolS} will compute a solution to system \eqref{prob:Polref} and construct the associated polynomial quasi-Trefftz function $H$.
\begin{algorithm}[H]
\caption{Amplitude based}
\label{Algo:AbS}
\begin{algorithmic}[1]
\State Given $\pG\in\mathbb R^3$, $q\in\mathbb N$ and $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ satisfying \Cref{hyp:PDop}
\State Fix $\{\mu_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|\in\{0,1\}\}$ and $\Lambda\in\mathbb C^3$
\For{$\ell \gets 0$ to $q-1$}
\State Compute $\mathsf B^\ell= \{\mathsf B^\ell_{\varb}\in \mathbb C;\varb\in(\mathbb N_0)^3, |\varb|=\ell\}$ according to
$$
\begin{array}{rl}
\mathsf B^\ell_{\varb}=&
\displaystyle
-
\sum_{\gamma< \varb}
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma](\gamma_k+2)(\gamma_k+1)\mu_{\gamma+2e_k}\\
&\displaystyle
-\sum_{\gamma\leq \varb}\left(
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma]\left
2 \Lambda_k (\gamma_k+1)\mu_{\gamma+e_k} + \Lambda_k^2\mu_{\gamma}\right) \right.
\\&\displaystyle \phantom{+\sum_{\gamma\leq \varb}\Bigg(}
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma]\Big((\gamma_k+1)(\gamma_{k'}+1)\mu_{\gamma+e_k+e_{k'}} +\Lambda_k(\gamma_{k'}+1)\mu_{\gamma+e_{k'}}
\\&\displaystyle \phantom{+\sum_{\gamma\leq \varb}\Bigg(
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma]\Big(}
+\Lambda_{k'}(\gamma_{k}+1)\mu_{\gamma+e_k} + \Lambda_k\Lambda_{k'}\mu_\gamma\Big)
\\ &\displaystyle \phantom{+\sum_{\gamma\leq \varb}\Bigg(}\left.
+
\sum_{k=1}^3
\coefc{e_k}[\varb-\gamma]\Big((\gamma_{k}+1)\mu_{\gamma+e_k} +\Lambda_k \mu_\gamma \Big)
+
\coefczr[\varb-\gamma] \mu_\gamma\right)
\end{array}
$$
\State Compute $\mu^\ell= \{\mu_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|=\ell+2\}$ via \Cref{Algo:SubS}: $$\mu^\ell = \text{ solve\_subsystem}\left(\ell,\mathsf B^\ell,\left\{\coefc{e_k+e_{k'}}[\mathbf 0], 1\leq k\leq k'\leq 3\right\}\right)$$
\EndFor
\State $\displaystyle Q:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \mu_i \mathbf X^i $ with $\mathbf X^i = X_1^{i_1} X_2^{i_2} X_3^{i_3}$
\State $J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big)$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[H]
\caption{Phase based}
\label{Algo:PbS}
\begin{algorithmic}[1]
\State Given $\pG\in\mathbb R^3$, $q\in\mathbb N$ and $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ satisfying \Cref{hyp:PDop}
\State Fix $\{\lambda_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|\in\{0,1\}\}$
\For{$\ell \gets 0$ to $q-1$}
\State Compute $\mathsf B^\ell= \{\mathsf B^\ell_{\varb}\in \mathbb C;\varb\in(\mathbb N_0)^3, |\varb|=\ell\}$ according to
$$
\begin{array}{rl}
\mathsf B^\ell_{\varb}=&
\displaystyle
-
\sum_{\gamma< \varb}
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma](\gamma_k+2)(\gamma_k+1)\lambda_{\gamma+2e_k} \\&
\displaystyle
-\sum_{\gamma\leq \varb}\left(
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma]\left( \sum_{\eta\leq \gamma}(\gamma_{k}-\eta_{k} +1)\lambda_{\gamma-\eta+e_{k}}(\eta_k +1)\lambda_{\eta+e_k}\right) \right.
\\&\displaystyle \phantom{+\sum_{\gamma\leq \varb}\Bigg(}
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma]\Bigg((\gamma_k+1)(\gamma_{k'}+1)\lambda_{\gamma+e_k+e_{k'}}
\\&\displaystyle \phantom{+\sum_{\gamma\leq \varb}\Bigg(+\sum_{1\leq k<k'\leq 3}\coefc{e_k+e_{k'}}[\varb-\gamma]\Bigg(}
+ \sum_{\eta\leq \gamma}(\gamma_{k'}-\eta_{k'} +1)\lambda_{\gamma-\eta+e_{k'}}(\eta_k +1)\lambda_{\eta+e_k}\Bigg)
\\& \displaystyle \phantom{+\sum_{\gamma\leq \varb}\Bigg(}\left.
+
\sum_{k=1}^3
\coefc{e_k}[\varb-\gamma](\gamma_{k}+1)\lambda_{\gamma+e_k}
\right)
-
\coefczr[\varb]
\end{array}
$$
\State Compute $\lambda^\ell= \{\lambda_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|=\ell+2\}$ via \Cref{Algo:SubS}: $$\lambda^\ell = \text{ solve\_subsystem}\left(\ell,\mathsf B^\ell,\left\{\coefc{e_k+e_{k'}}[\mathbf 0], 1\leq k\leq k'\leq 3\right\}\right)$$
\EndFor
\State $\displaystyle P:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \lambda_i \mathbf X^i $ with $\mathbf X^i = X_1^{i_1} X_2^{i_2} X_3^{i_3}$
\State $G(\varx) :=\exp \polP(\varx-\pG)$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[H]
\caption{Polynomial}
\label{Algo:PolS}
\begin{algorithmic}[1]
\State Given $\pG\in\mathbb R^3$, $q\in\mathbb N$ and $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ satisfying \Cref{hyp:PDop}
\State Fix $\{\nu_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|\in\{0,1\}\}$
\For{$\ell \gets 0$ to $q-1$}
\State Compute $\mathsf B^\ell= \{\mathsf B^\ell_{\varb}\in \mathbb C;\varb\in(\mathbb N_0)^3, |\varb|=\ell\}$ according to
$$
\begin{array}{rl}
\mathsf B^\ell_{\varb}=&
\displaystyle
-
\sum_{\gamma< \varb}
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma](\gamma_k+2)(\gamma_k+1)\nu_{\gamma+2e_k}\\
&\displaystyle
-\sum_{\gamma\leq \varb}\left(
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma](\gamma_k+1)(\gamma_{k'}+1)\nu_{\gamma+e_k+e_{k'}} \right.
\\ &\displaystyle \phantom{+\sum_{\gamma\leq \varb}\Bigg(}\left.
+
\sum_{k=1}^3
\coefc{e_k}[\varb-\gamma](\gamma_{k}+1)\nu_{\gamma+e_k}
+
\coefczr[\varb-\gamma] \nu_\gamma\right)
\end{array}
$$
\State Compute $\nu^\ell= \{\nu_{i}\in \mathbb C;i\in(\mathbb N_0)^3, |i|=\ell+2\}$ via \Cref{Algo:SubS}: $$\nu^\ell = \text{ solve\_subsystem}\left(\ell,\mathsf B^\ell,\left\{\coefc{e_k+e_{k'}}[\mathbf 0], 1\leq k\leq k'\leq 3\right\}\right)$$
\EndFor
\State $\displaystyle R:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \nu_i \mathbf X^i $ with $\mathbf X^i = X_1^{i_1} X_2^{i_2} X_3^{i_3}$
\State $H(\varx) :=\polR(\varx-\pG)$
\end{algorithmic}
\end{algorithm}
Interestingly, the quasi-Trefftz property of the quasi-Trefftz functions, $J$, $G$ and $H$, built from \Cref{Algo:PbS,Algo:AbS,Algo:PolS} are satisfied independently of the fixed values throughout these algorithms: $\mathcal L_{c} J(\varx) = O(|\varx-\pG|^q)$, $\mathcal L_{c} G(\varx) = O(|\varx-\pG|^q)$ and $\mathcal L_{c} H(\varx) = O(|\varx-\pG|^q)$.
We will refer to the choice of these values as the initialization process.
However an appropriate choice of initialization will be crucial to prove approximation properties of the resulting sets of quasi-Trefftz functions.
\subsection{Initializatio
}
\label{sec:norm}
Keeping in mind the motivation for the design of GPWs, that is adding higher order terms either in the phase or the amplitude of a PW:
\begin{equation}
\label{eq:GPWint}
(1+\text{HOT})\exp \Lambda \cdot\Big(\varx-\pG \Big)
\text{ or }
\exp \Big[\Lambda \cdot\Big(\varx-\pG \Big)+\text{HOT}\Big],
\end{equation}
we now turn to the initialization process.
From \Cref{Algo:SubS,Algo:PbS,Algo:AbS}, the free parameters in the construction of a GPW are $\left\{\lambda_{i} \in\mathbb C \text{ for }i\in(\mathbb N_0)^3, |i|\leq q+1,i_1\in\{0,1\}\right\}$ for a Phase-based GPW, and for an Amplitude-based GPW $\left\{\mu_{i} \in\mathbb C\text{ for }i\in(\mathbb N_0)^3, |i|\leq q+1,i_1\in\{0,1\}\right\}$ plus $\Lambda\in\mathbb C^3$. In both cases, we follow the intuition that lead to the choice of ansatz \eqref{eq:GPWint} as a generalization of PW functions to build a family of GPWs. In order to do so, only a few free parameters are sufficient, corresponding to the linear terms in the phase, and except for the constant coefficient of the amplitude for an Amplitude-based GPW, we will naturally set the remaining parameters to zero to reduce the amount of computation associated with the construction of each GPW.
On the other hand, in the polynomial case, in order to prove the approximation properties of the resulting set of functions, besides the linear terms, we will also leverage the other free parameters to obtain a family of linearly independent quasi-Trefftz polynomials.
To summarize,
%
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|c|}
\hline
Amplitude based & Phase-based& Polynomial& Comment\\\hline\hline
$\Lambda\in\mathbb C^3$ & $\left\{\lambda_{i} \in\mathbb C;i\in(\mathbb N_0)^3, |i|=1\right\}$& $\left\{\nu_{i} \in\mathbb C;i\in(\mathbb N_0)^3, |i|=1\right\}$& $[\![1]\!]$ \\\hline
$\mu_{\mathbf 0}$& & $\nu_{\mathbf 0}$& Set to $1$
\\\hline
&$\lambda_{\mathbf 0}$& &Set to $0$
\\\hline
$\left\{\mu_{i} \in\mathbb C;i\in(\mathbb N_0)^3, |i|=1\right\}$& & & Set to $0$
\\\hline\hline
for $\ell$ from $0$ to $q-1$
&
for $\ell$ from $0$ to $q-1$
&
&
\\
$\{\mu_{i}\in \mathbb C;i\in(\mathbb N_0)^3,\qquad$
&
$\{\lambda_{i}\in \mathbb C;i\in(\mathbb N_0)^3,\qquad
&
& Set to $0$\\
$ |i|=\ell+2, i_1\in\{0,1\}\}$&$ |i|=\ell+2, i_1\in\{0,1\}\}$& &\\
\hline
&
&
for $\ell$ from $0$ to $q-1$
&
\\
&
&
$\{\nu_{i}\in \mathbb C;i\in(\mathbb N_0)^3,\qquad
&$[\![2]\!]$ \\
& &$ |i|=\ell+2, i_1\in\{0,1\}\}$&\\
\hline
\end{tabular}}\\
{Initialization summary
for the three types of quasi-Trefftz functions.
Top rows: corresponding to Step 1 in \Cref{Algo:PbS,Algo:AbS,Algo:PolS}.
Bottom rows: corresponding to Step 1 in \Cref{Algo:SubS}. }
\end{center}
In order to completely define our choices of quasi-Trefftz functions, it is then sufficient to describe how are chosen:
$[\![1]\!]$ the three parameters corresponding to linear terms,
and
$[\![2]\!]$ the other parameters in the polynomial case.
$[\![1]\!]$ In order to build not a single but rather a set of quasi-Trefftz functions, we now have three non-zero free parameters in each case, namely:
\begin{equation}
\label{eq:3params}
\Lambda=
\begin{bmatrix}\Lambda_1\\\Lambda_2\\\Lambda_3\end{bmatrix},
\begin{bmatrix}\lambda_{e_1}\\\lambda_{e_2}\\\lambda_{e_3}\end{bmatrix},
\begin{bmatrix}\nu_{e_1}\\\nu_{e_2}\\\nu_{e_3}\end{bmatrix} \text{ in }\mathbb C^3.
\end{equation}
In the constant-coefficient Helmholtz case, the matrix $\mathcal C$ introduced in \Cref{hyp:PDop} is the identity $I_3$ and it is then natural to fix these as $\cst\dir$, with $\cst = \mathrm i\kappa$ to obtain a PW exact solution since $(\cst\dir)^T (I_3\cst\dir)=-\kappa^2$ is independent of $\dir$. Yet in the general case, this matrix $\mathcal C$ is associated with anisotropy in the second order terms of the partial differential operator, and it is then natural to introduce (i) the orthonormal basis of eigenvectors of $\mathcal C$ via $\mathcal P$ and (ii) the anisotropic scaling by the eigenvalues of $\mathcal C$ via $\mathcal D$.
Hence for each quasi-Trefftz function, under \Cref{hyp:PDop}, we will fix these as $\cst \Pmat\Dmat^{-1/2}\dir$, where $\cst\in\mathbb C^*$ and $\dir\in\mathbb R^3$ with $|\dir|=1$.
To define a set of $p$ distinct -- and linearly independent as we will see later -- quasi-Trefftz functions, we will choose distinct directions $\{ \dir_l\in\mathbb R^3 \text{ for } l \text{ from } 1 \text{ to } p; |\dir_l|=1 \text{ with } \dir_L\neq\dir_k \text{ if }k\neq l \}$ while we will choose a common value for $\cst$ for each of the $p$ functions in the set.
Each direction $ \mathbf d_l$ will be defined by two angles $(\theta_l,\varphi_l)$ as follows:
$$
\mathbf d_l =
\begin{bmatrix}
\sin\theta_l\sin\varphi_l\\\sin\theta_l\cos\varphi_l\\\cos\theta_l
\end{bmatrix}.
$$
Under \Cref{hyp:PDop}, for any $X=\cst \Pmat\Dmat^{-1/2}\dir$, we can easily verify that:
$$
\begin{array}{rl} X^T( \Cmat X)& = \cst^2 (\Pmat\Dmat^{-1/2}\dir)^T(\Pmat\Dmat\Pmat^T)(\Pmat\Dmat^{-1/2}\dir)\\
& = \cst^2\dir^T\dir \\
&=\cst^2,
\end{array}
$$
or equivalently:
$$
\sum_{k=1}^3
\coefc{2e_k}[\mathbf 0]\left( X_{{k}}\right)^2
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\mathbf 0] X_{{k'}}X_{k}
=
\cst^2.
$$
As a result, this quantity does not depend on the direction $\dir$, but rather has the same value for the whole set of functions, as we discussed in the Helmholtz case. This crucial fact will be key to prove approximation properties of the quasi-Trefftz functions.
\begin{rmk}
As a result of this choice, the GPW functions boil down to classical PWs in the case of a constant coefficient Helmholtz equation.
\end{rmk}
$[\![2]\!]$ For the polynomial basis, we propose to fix the other parameters values as:
\begin{equation}
\label{eq:PolInit}
\forall i\in(\mathbb N_0)^3, |i|\geq 2, i_1\in\{0,1\}, \nu_i = (\nu_{e_1})^{i_1}(\nu_{e_2})^{i_2}(\nu_{e_3})^{i_3}.
\end{equation}
This particular choice will allow us to prove approximation properties of the corresponding polynomial quasi-Trefftz basis.
\section{Approximation properties}
The construction of quasi-Trefftz functions is based on Taylor expansions, it is therefore natural to use similar tools to study their approximation properties. The central idea here is precisely to approximate a given exact PDE solution $u$ by a linear combination $u_a$ of quasi-Trefftz functions by matching their respective Taylor expansions at $\pG$. Indeed, for any order of approximation $n$, we have:
\begin{equation}
\label{eq:TEapp}
\left\{\begin{array}{l}
\forall i\in\left(\mathbb N_0\right)^3, |i|\leq n,\\
\partial_i u (\pG) = \partial_i u_a(\pG),
\end{array}\right.
\Rightarrow u_a(x)-u(x) = O\left(\left|\varx-\pG\right|^{n+1}\right).
\end{equation}
This in turn leads to the convergence of $u-u_a$ in various norms of interest in the regime $\left|\varx-\pG\right|\rightarrow 0$, moreover higher order convergence follows from increasing the value of the order of approximation $n$ in the Taylor expansion.
Matching the Taylor expansion of a linear combination $u_a$ of quasi-Trefftz functions to that of a given function $u$ leads to a linear system.
\begin{itemize}
\item Each unknown is a weight of the desired linear combination, and is indexed $l$;
there are as many unknowns as there are functions in the quasi-Trefftz set.
\item Each equation corresponds to one Taylor expansion coefficient, and is indexed $i\in\left(\mathbb N_0\right)^3$ with $ |i|\leq n$;
there are $(n+1)(n+2)(n+3)/6$ equations.
\end{itemize}
The system's matrix can then be defined for a given list of quasi-trefftz function thanks to a numbering of the equations.
The entries of the system's matrix are the partial derivatives of quasi-Trefftz functions evaluated at $\pG$.
Hence, given any numbering of multi-indices $\mathcal N$, for any family of $p$ functions $\{b_l,\text{ for } l\in\mathbb N, l\leq p\}$, the $(\mathcal N(i),l)$ entry of the corresponding $\frac{(n+1)(n+2)(n+3)}{6} \times p$ matrix $M^{n,p}$ is:
\begin{equation}
\label{eq:genmat}
M^{n,p}_{\mathcal N(i),l} =
\frac{\partial_x^i b_l(\pG)}{i!} = T_{b_l}[i]
\end{equation}
Moreover, the value $p=(n+1)^2$ will be of particular interest in what follows. So to simplify the notation, the matrix corresponding to $p=(n+1)^2$ will be denoted with the superscript $[n]$, for instance $M^{[n]}$ instead of $M^{n,(n+1)^2}$.
For three families of quasi-Trefftz functions introduced in the previous section, we will use the following notation for the corresponding matrices.
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|}
\hline
Amplitude-based functions & Phase-based functions & Polynomial functions\\\hline
$\mathsf A}%{M\hspace{-.75em}A^{n,p}$ and $\mathsf A}%{M\hspace{-.75em}A^{[n]}$&
$\mathsf P}%{M\!\!\!P^{n,p}$ and $\mathsf P}%{M\!\!\!P^{[n]}$&
$\mathsf Q}%}{M\!\!\!Q^{n,p}$ and $\mathsf Q}%}{M\!\!\!Q^{[n]}$\\
\hline
\end{tabular}}
\end{center}
In order to evidence the structure of the linear system, the equations will be numbered as follows.
To leverage the loop structure of \Cref{Algo:PbS,Algo:AbS,Algo:PolS}, we want $\mathcal N$ to satisfy $|i|<|j|$ implies that $\mathcal N(i)<\mathcal N(j)$, hence we write it as:
$$
\forall i\in\left(\mathbb N_0\right)^3\text{ with } |i|\leq n,
\mathcal N(i) = \frac{|i|(|i|+1)(|i|+2)}{6}+\mathcal N_{|i|}(i),
$$
for some $\mathcal N_m$ providing a numbering of indices of length m.
For instance we can choose $\mathcal N_m$ to count indices according to the linear order $\prec$ within the level $m$ and in this case the numbering corresponds to $\mathcal N(i) = \sum_{j\prec i} 1$,
or choose $\mathcal N_m(i)=(i2+i3)(i2+i3+1)/2+i3+1$ for all $i$ such that $|i|=m$.
A road map was proposed in \cite{IGS} to prove approximation properties of
quasi-Trefftz functions. It can be summarized as follows:
\begin{enumerate}
\item\label{step:prelim} for each quasi-Trefftz function, express all the basic parameters in \Cref{Algo:PbS,Algo:AbS,Algo:PolS} in terms of the free parameters that are not set to $0$;
\item\label{step:idref} identify a reference case, here a classical PW case;
\item\label{step:propref} study useful properties of the reference matrix;
\item\label{step:link} establish a link between each of the quasi-Trefftz cases and the reference case;
\item\label{step:proof} prove the approximation properties of quasi-Trefftz bases.
\end{enumerate}
While \Cref{step:idref,step:propref,step:proof} are case-independent, \Cref{step:prelim,step:link} will be treated separately for each family of quasi-Trefftz functions.
These two key points rely on understanding how the entries of the linear system matrices depend on the initialization of our quasi-Trefftz functions, emphasizing their properties shared by corresponding entries on a given row as well as their differences.
Two important questions about these matrices concern their rank. (i) How large of a rank can they have? (ii) What particular choice of angles in the initialization parameters can guarantee the maximal rank? These will lead the choice (i) of how many different quasi-Trefftz functions to define, and (ii) of how to choose the initialization angles.
As a by-product, the resulting families of quasi-Trefftz functions will be proved to be linearly independent.
Given an order $n$ for the approximation property \eqref{eq:TEapp}, the order $q$ of the quasi-Trefftz property will be chosen to guarantee a similar construction for all the polynomial coefficients of quasi-Trefftz basis functions that will appear in the Taylor expansion \eqref{eq:TEapp}. This will require to set $q\geq n-1$, it is then sufficient to construct the quasi-Trefftz basis functions with the parameter $q$ satisfying:
\begin{center}
\fbox{ $ q=\max(n-1,1)$ }
\end{center}
This will be particularly helpful to describe all polynomial coefficients of the quasi-Trefftz basis functions in terms of the initialization parameters, see Section \ref{ssec:Relate}.
It seems important to underline the fundamental part that the interplay of \Cref{hyp:PDop} and the choice of initialization will play in the rest of this section.
\subsection{Preliminary results}
The goal is to investigate how the terms computed in \Cref{Algo:AbS,Algo:SubS,Algo:PbS},
namely $\mu_{\beta+2e_1}$ and $\lambda_{\beta+2e_1}$,
depend on the three free parameters from the initialization process \eqref{eq:3params}.
We will proceed by induction with respect to the level $\ell$. In each case the result will rely on a careful inspection of the right-hand side $B^\ell$ of the subsystems.
\subsubsection{For amplitude-based GPWs}
For an amplitude-based GPW, we focus on investigating properties of $\{\mu_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq q+1\}$.
Here, the three non-zero free parameters in the initialization procedure are $\Lambda_1,\Lambda_2,\Lambda_3$.
All $\mu$s computed from \Cref{Algo:AbS,Algo:SubS} clearly appear to be polynomials with respect to these three free parameters, that is they are elements of $\mathbb C[\Lambda_1, \Lambda_2, \Lambda_3]$.
Moreover, as first noted in \cite{LMinterp}, the initialization ensures that
$$
Q_N = 0, \text{ where }
Q_N(\Lambda_1,\Lambda_2,\Lambda_3):= \sum_{k=1}^3
\coefc{2e_k}[\mathbf 0]\left(\Lambda_k\right)^2
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\mathbf 0] \Lambda_{k'}\Lambda_k-\cst^2
$$
which turns our attention to elements of $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]/ (Q_N)$ instead.
We will therefore investigate how the other $\mu$s can be expressed in terms of the three free parameters, $\Lambda_1,\Lambda_2,\Lambda_3$.
\begin{lmm}\label{lmm:l2e1Ab}
Given $q\in\mathbb N$, a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any amplitude-based GPW associated to differential operator $\mathcal L_{c}$, $J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big)$ with $\displaystyle Q:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \mu_i \mathbf X^i $, constructed via \Cref{Algo:AbS,Algo:SubS}, with the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ a unit vector $\dir\in\mathbb S^2$.
Then $\mu_{2e_1}$ can be expressed as a polynomial of degree at most equal to $1$ in $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$, with coefficients depending on $\cst$ yet independent of $\dir$.
\end{lmm}
\begin{proof}
From the formulas in \Cref{Algo:SubS,Algo:AbS} for $\ell=0$ and $\varb=\mathbf 0$ we get:
$$\left\{
\begin{array}{l}
\mu_{2e_1} = \displaystyle
\frac{1}{2\coefc{2e_1}[\mathbf 0]}\Bigg(\mathsf {B^0_{\mathbf 0}}-\sum_{k=2}^3
2\coefc{2e_k}[\mathbf 0]\mu_{2e_k}-\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\mathbf 0]\mu_{e_k+e_{k'}}
\Bigg),\\
\mathsf B^0_{\mathbf 0} = \displaystyle-\sum_{k=1}^3 \coefc{2e_k}[\mathbf 0](2\Lambda_k\mu_{e_k}+\Lambda_k^2 \mu_\mathbf 0)
-\sum_{1\leq k < k'\leq 3} \coefc{e_k+e_{k'}}[\mathbf 0]\Big(\mu_{e_k+e_{k'}}+\Lambda_k \mu_{e_{k'}}+\Lambda_{k'} \mu_{e_{k}} + \Lambda_k\Lambda_{k'}\mu_\mathbf 0\Big)\\
\phantom{\mathsf B^0_{\mathbf 0} = }
-\displaystyle \sum_{k=1}^3\coefc{e_k}[\mathbf 0](\mu_{e_k}+\Lambda_k \mu_\mathbf 0)
- \coefc{\mathbf 0}[\mathbf 0]\mu_\mathbf 0,
\end{array}\right.
$$
so the initialization implies:
$$\left\{
\begin{array}{l}
\mu_{2e_1} = \displaystyle
\frac{1}{2\coefc{2e_1}[\mathbf 0]}\mathsf {B^0_{\mathbf 0}},\\
\mathsf B^0_{\mathbf 0} = \displaystyle-\sum_{k=1}^3 \coefc{2e_k}[\mathbf 0]\Lambda_k^2
-\sum_{1\leq k < k'\leq 3} \coefc{e_k+e_{k'}}[\mathbf 0] \Lambda_k\Lambda_{k'}
-\displaystyle \sum_{k=1}^3\coefc{e_k}[\mathbf 0]\Lambda_k
- \coefc{\mathbf 0}[\mathbf 0],\\
\phantom{\mathsf B^0_{\mathbf 0} }= -\displaystyle Q_N(\Lambda_1,\Lambda_2,\Lambda_3)-\cst^2
-\displaystyle \sum_{k=1}^3\coefc{e_k}[\mathbf 0]\Lambda_k
- \coefc{\mathbf 0}[\mathbf 0],
\end{array}\right.
$$
Therefore, since $ Q_N(\Lambda_1,\Lambda_2,\Lambda_3)=0$, we obtain
$$
\mu_{2e_1} =
\frac{1}{2\coefc{2e_1}[\mathbf 0]}
\left(-\cst^2
-\displaystyle \sum_{k=1}^3\coefc{e_k}[\mathbf 0]\Lambda_k
- \coefc{\mathbf 0}[\mathbf 0]\right),
$$
which proves the claim since $\cst$ is a fixed constant according to the initialization.
\end{proof}
\begin{prop}
\label{prop:Abmus}
Given $q\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any amplitude-based GPW, $J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big)$ with $\displaystyle Q:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \mu_i \mathbf X^i $, constructed via \Cref{Algo:AbS,Algo:SubS}, with the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ a unit vector $\dir\in\mathbb S^2$.
Then,
for all $\ell$ from 0 to $q-1$ and all $\beta\in(\mathbb N_0)^3$ such that $|\beta|=\ell$, $\mu_{\varb+2e_1}$
can be expressed as a polynomial of degree at most equal to
$|\varb|+1$
in $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$, with coefficients depending on $\cst$ yet independent of $\dir$.
\end{prop}
\begin{proof}
We will proceed by induction with respect to $\varb$ according to the linear order $\prec$, which corresponds to the order in which the $\mu_{\varb+2e_k} $ terms are computed in the algorithms.
The case $\varb=\mathbf 0$ is precisely \Cref{lmm:l2e1Ab}.
Assume $\varb\in(\mathbb N_0)^3$ with $0\prec\varb$ is such that the result holds for all $\varb'\prec \varb$: $\mu_{\varb'+2e_1}$ can be expressed as a polynomial of degree at most equal to $|\varb'|+1$ in $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$. Then $ \mu_{\varb+2e_1}$ is computed according to \Cref{Algo:SubS,Algo:AbS}. Hence, since
$$
2 \Lambda_k (\gamma_k+1)\mu_{\gamma+e_k} + \Lambda_k^2\mu_{\gamma} =
\Lambda_k(\gamma_{k'}+1)\mu_{\gamma+e_{k'}}
+\Lambda_{k'}(\gamma_{k}+1)\mu_{\gamma+e_k} + \Lambda_k\Lambda_{k'}\mu_\gamma
\text{ for } k'=k,
$$
we can gather these terms in a sum over $1\leq k\leq k'\leq 3$ and
$ \mu_{\varb+2e_1}$ can be written as:
\begin{equation}\label{eq:mb+2e1}
\begin{array}{ll}
\mu_{\varb+2e_1} = &\displaystyle
\frac{1}{(\varb_1+2)(\varb_1+1)\coefc{2e_1}[\mathbf 0]}\Bigg(
-\sum_{\gamma< \varb}
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma](\gamma_k+2)(\gamma_k+1)\mu_{\gamma+2e_k} \\&\displaystyle
-\sum_{\gamma\leq \varb}
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma](\gamma_k+1)(\gamma_{k'}+1)\mu_{\gamma+e_k+e_{k'}} \\&\displaystyle
-\sum_{\gamma\leq \varb}
\sum_{1\leq k\leq k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma]
\Big(\Lambda_k(\gamma_{k'}+1)\mu_{\gamma+e_{k'}} +\Lambda_{k'}(\gamma_{k}+1)\mu_{\gamma+e_k} + \Lambda_k\Lambda_{k'}\mu_\gamma\Big) \\&\displaystyle
-\sum_{\gamma\leq \varb}
\left(
\sum_{k=1}^3
\coefc{e_k}[\varb-\gamma]\Big((\gamma_{k}+1)\mu_{\gamma+e_k} +\Lambda_k \mu_\gamma \Big)
+
\coefczr[\varb-\gamma] \mu_\gamma
\right)
\\&\displaystyle-\sum_{k=2}^3
(\varb_k+2)(\varb_k+1)\coefc{2e_k}[\mathbf 0]\mu_{\varb+2e_k}
\\&\displaystyle
-\sum_{1\leq k<k'\leq 3}
(\varb_k+1)(\varb_{k'}+1)\coefc{e_k+e_{k'}}[\mathbf 0]\mu_{\varb+e_k+e_{k'}}
\Bigg).
\end{array}
\end{equation}
On the right hand side we observe that the $\mu$ terms fall in one of two categories as elements of $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$:
\begin{enumerate}
\item $\mu_i$ with $i_1\in\{0,1\}$, chosen in the initialization process, either $\mu_\mathbf 0=1$ or otherwise the terms set to zero,
\item $\mu_i$ with $i_1\geq 2$, computed at a previous iteration for $\varb'=i-2e_1\prec\varb$.
\end{enumerate}
The linear terms can be listed as follows.
\begin{center}
{\renewcommand{\arraystretch}{1.1}
\begin{tabular}{|c||c|c|c|c|c|c|}
\hline
Terms & Indices & $0$ & $1$ & $\mu_{\varb'+2e_1}$ \\\hline\hline
$\mu_{\gamma+2e_k}$&
$\begin{array}{c}
\gamma<\varb(\neq\mathbf 0),k\in\{1,2,3\}
\\ (\gamma+2e_k)_1\in\{0,1\}
\end{array}$
&
\ding{51}
& & \\\hline
$\mu_{\gamma+2e_k}$&
$\begin{array}{c}
\gamma<\varb(\neq\mathbf 0),k\in\{1,2,3\}
\\ (\gamma+2e_k)_1>1
\end{array}$
& & &
$ \varb' =\gamma+2(e_k-e_1) $ \\\hline
$\mu_{\gamma+e_k+e_{k'}}$ &
$\begin{array}{c}
\gamma\leq\varb,1\leq k<k'\leq 3
\\ (\gamma+e_k+e_{k'})_1\in\{0,1\}
\end{array}$
&
\ding{51}&&\\\hline
$\mu_{\gamma+e_k+e_{k'}}$ &
$\begin{array}{c}
\gamma\leq\varb,1\leq k<k'\leq 3
\\ (\gamma+e_k+e_{k'})_1>1
\end{array}$
& & &
$\varb'=\gamma+e_k+e_{k'}-2e_1$
\\\hline
$\mu_{\gamma+e_k}$ &
$\begin{array}{c}
\gamma\leq\varb,k\in\{1,2,3\}
\\ ({\gamma+e_k})_1\in\{0,1\}
\end{array}$
& \ding{51} & &
\\\hline
$\mu_{\gamma+e_k}$ &
$\begin{array}{c}
\gamma\leq\varb,k\in\{1,2,3\}
\\ ({\gamma+e_k})_1>1
\end{array}$
& & & $\varb'=\gamma+e_k-2e_1$
\\\hline
$\mu_{\gamma}$ &
$\gamma=\mathbf 0$
& & \ding{51} &
\\\hline
$\mu_{\gamma}$ &
$\begin{array}{c}
\gamma\leq\varb
\\ ({\gamma})_1\in\{0,1\},\gamma\neq\mathbf 0
\end{array}$
& \ding{51} & &
\\\hline
$\mu_{\gamma}$ &
$\begin{array}{c}
\gamma\leq\varb,
\\ ({\gamma})_1>1
\end{array}$
& & & $\varb'=\gamma-2e_1$
\\\hline
$\mu_{\varb+2e_k}$&
$\begin{array}{c}
k\in\{2,3\}
\\ (\varb+2e_k)_1\in\{0,1\}
\end{array}$
&
\ding{51}
& & \\\hline
$\mu_{\varb+2e_k}$&
$\begin{array}{c}
k\in\{2,3\}
\\ (\varb+2e_k)_1>1
\end{array}$
& & &
$ \varb' =\varb+2(e_k-e_1) $ \\\hline
$\mu_{\varb+e_k+e_{k'}}$ &
$\begin{array}{c}
1\leq k<k'\leq 3
\\ (\varb+e_k+e_{k'})_1\in\{0,1\}
\end{array}$
&
\ding{51}&&\\\hline
$\mu_{\varb+e_k+e_{k'}}$ &
$\begin{array}{c}
1\leq k<k'\leq 3
\\ (\varb+e_k+e_{k'})_1>1
\end{array}$
& & &
$\varb'=\varb+e_k+e_{k'}-2e_1$
\\\hline
\end{tabular}}
\end{center}
Hence the linear terms can be expressed as elements of $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$ as either $0$, $1$, or by induction hypothesis as a polynomial of degree at most equal to $|\varb'|+1$. Moreover, in this last case, the values of $\varb'$ identified in the previous table are such that:
$$
\left\{\begin{array}{l}
\gamma<\varb\Rightarrow |\gamma+2(e_k-e_1)|+1<|\varb|+1,\\
\gamma\leq\varb\Rightarrow |\gamma+e_k+e_{k'}-2e_1|+1\leq |\varb|+1,\\
\gamma\leq\varb\Rightarrow |\gamma+e_k-2e_1|+1\leq |\varb|,
\end{array}\right.
\text{ and }
\left\{\begin{array}{l}
|\varb+2(e_k-e_1)|+1=|\varb|+1,\\
|\varb+e_k+e_{k'}-2e_1|+1= |\varb|+1.
\end{array}\right.
$$
In summary all the linear terms in the right hand side of \eqref{eq:mb+2e1} can be expressed as polynomials of degree at most equal to $|\varb|+1$ in $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$.
As for the non-linear terms, they all appear for indices $\gamma\leq\varb$ and $1\leq k \leq k'\leq 3$, and they can be identified as follows.
\begin{center}
{\renewcommand{\arraystretch}{1.1}
\begin{tabular}{|c||c|c|c|c|c|c|c|c|}
\hline
Terms & Indices & $0$ & $1$& $\Lambda_k$ or $\Lambda_{k'}$ & $\mu_{\varb'+2e_1}$&Case \\\hline\hline
$\Lambda_k\mu_{\gamma+e_{k'}}$&
$(\gamma+e_{k'})_1\in\{0,1\}$& \ding{51} & &\ding{51}& & \\\hline
$\Lambda_k\mu_{\gamma+e_{k'}}$&
$ (\gamma+e_{k'})_1>1$& & &\ding{51} &$\varb'=\gamma+e_{k'}-2e_1$&1 \\\hline
$\Lambda_{k'}\mu_{\gamma+e_k}$ &
$(\gamma+e_k)_1\in\{0,1\}$ & \ding{51}&&\ding{51}&&\\\hline
$\Lambda_{k'}\mu_{\gamma+e_k}$ &
$(\gamma+e_k)_1>1$ & & &\ding{51}& $\varb'=\gamma+e_k-2e_1$ &2
\\\hline
$\Lambda_k\Lambda_{k'}\mu_\gamma$ &
$\gamma=\mathbf 0$ & & \ding{51} & \ding{51} \ding{51}&&3
\\\hline
$\Lambda_k\Lambda_{k'}\mu_\gamma$ &
$ ({\gamma})_1\in\{0,1\}$, $\gamma\neq\mathbf 0$& \ding{51} & & \ding{51} \ding{51}&&
\\\hline
$\Lambda_k\Lambda_{k'}\mu_\gamma$ &
$ ({\gamma})_1>1$& & & \ding{51} \ding{51}& $\varb'=\gamma-2e_1$ &4
\\\hline
$\Lambda_k\mu_\gamma$ &
$\gamma=\mathbf 0$ & & \ding{51} & \ding{51} &&5
\\\hline
$\Lambda_k\mu_\gamma$ &
$ ({\gamma})_1\in\{0,1\}$, $\gamma\neq\mathbf 0$& \ding{51} & & \ding{51} &&
\\\hline
$\Lambda_k\mu_\gamma$ &
$ ({\gamma})_1>1$& & & \ding{51} & $\varb'=\gamma-2e_1$ &6
\\\hline
\end{tabular}}
\end{center}
Hence, as elements of $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$, these non-linear terms can be expressed either as zero or by induction hypothesis as a polynomial of degree at most equal to:
\begin{itemize}
\item $|\beta'|+2 \leq |\varb|$ in cases 1 and 2,
\item $2$ in case 3,
\item $|\varb'|+3\leq |\varb|+1$ in case 4,
\item $1$ in case 5,
\item $|\varb'|+2\leq|\varb|$ in case 6.
\end{itemize}
In summary all the non-linear terms in the right hand side of \eqref{eq:mb+2e1} can be expressed as polynomials of degree at most equal to $|\varb|+1$ in $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$.
Therefore $ \mu_{\varb+2e_1}$ in \eqref{eq:mb+2e1} can be expressed as a polynomial of degree at most equal to $|\varb|+1$ in $\mathbb C[\Lambda_1,\Lambda_2,\Lambda_3]$. This concludes the proof.
\end{proof}
\subsubsection{For phase-based GPWs}
For a phase-based GPW, we focus on investigating properties of $\{\lambda_{i}\in \mathbb C,i\in\mathbb N_0^3, |i|\leq q+1\}$.
Here, the three non-zero free parameters in the initialization procedure are $\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}$.
All $\lambda$s computed from \Cref{Algo:PbS,Algo:SubS} clearly appear to be polynomials with respect to these three free parameters, that is they are elements of $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$.
Moreover, as first noted in \cite{LMinterp}, the initialization ensures that
$$
P_N = 0, \text{ where }
P_N(\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}):= \sum_{k=1}^3
\coefc{2e_k}[\mathbf 0]\left(\lambda_{e_k}\right)^2
+
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\mathbf 0] \lambda_{e_{k'}}\lambda_{e_k}-\cst^2
$$
which turns our attention to elements of $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]/ (P_N)$ instead.
We will therefore investigate how the other $\lambda$s can be expressed in terms of the three free parameters, $\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}$.
\begin{lmm}\label{lmm:l2e1Pb}
Given $q\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Conside any phase-based GPW associated to the partial differential operator $\mathcal L_{c}$, $G(\varx) :=\exp \polP(\varx-\pG)$ with $\displaystyle P:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \lambda_i \mathbf X^i $, constructed via \Cref{Algo:PbS,Algo:SubS}, with the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and a unit vector $\dir$.
Then $\lambda_{2e_1}$ can be expressed as a polynomial of degree at most equal to 1 in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$, with coefficients depending on $\cst$ yet independent of $\dir$.
\end{lmm}
\begin{proof}
Since $\lambda_{2e_1}$ is computed at iteration $\ell=0$ in \Cref{Algo:PbS}, according to \Cref{Algo:SubS}, we have:
$$
\lambda_{2e_1} =
\frac{1}{2\coefc{2e_1}[\mathbf 0]}\Bigg(\mathsf {B^0_{\mathbf 0}}-\sum_{k=2}^3
2\coefc{2e_k}[\mathbf 0]\lambda_{2e_k}
-\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\mathbf 0]\lambda_{e_k+e_{k'}}
\Bigg),
$$
so the initialization then implies:
\begin{equation}
\label{eq:lambda2e1Pb}
\lambda_{2e_1} =
\frac{1}{2\coefc{2e_1}[\mathbf 0]}\mathsf {B^0_{\mathbf 0}},
\end{equation}
By definition, the right hand side $B^0=\left[B^0_{\mathbf 0}\right]\in\mathbb C$ is:
$$
\begin{array}{rl}
B^0_{\mathbf 0}
&\displaystyle=
-\sum_{k=1}^3
\coefc{2e_k}[\mathbf 0]\left( \lambda_{e_{k}}\right)^2
-
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\mathbf 0]\Bigg(\lambda_{e_k+e_{k'}}
+ \lambda_{e_{k'}}\lambda_{e_k}\Bigg)
-
\sum_{k=1}^3
\coefc{e_k}[\mathbf 0]\lambda_{e_k}
-
\coefczr[\mathbf 0],\\
&\displaystyle=
-P_N(\lambda_{e_1},\lambda_{e_2},\lambda_{e_3})-\cst^2
-
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\mathbf 0]\lambda_{e_k+e_{k'}}
-
\sum_{k=1}^3
\coefc{e_k}[\mathbf 0]\lambda_{e_k}
-
\coefczr[\mathbf 0],
\end{array}
$$
so from the initialization, imposing all the $\lambda_{e_k+e_{k'}} $ to be zero as well as $P_N=0$, we can express the right hand side as:
\begin{equation}
\label{eq:B0Pb}
B^0_{\mathbf 0}
=
-\cst^2
-
\sum_{k=1}^3
\coefc{e_k}[\mathbf 0]\lambda_{e_k}
-
\coefczr[\mathbf 0].
\end{equation}
As a result, combining \eqref{eq:lambda2e1Pb} and \eqref{eq:B0Pb} yields:
$$
\lambda_{2e_1} =
\frac{1}{2\coefc{2e_1}[\mathbf 0]}\left( - \cst^2
-
\sum_{k=1}^3
\coefc{e_k}[\mathbf 0]\lambda_{e_k}
-
\coefczr[\mathbf 0]\right),
$$
which proves the claim since $\cst$ is a fixed constant according to the initialization.
\end{proof}
\begin{prop}
\label{prop:Pblambdas}
Given $q\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any phase-based GPW associated to the partial differential operator $\mathcal L_{c}$, $G(\varx) :=\exp \polP(\varx-\pG)$ with $\displaystyle P:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \lambda_i \mathbf X^i $, constructed via \Cref{Algo:PbS,Algo:SubS}, with the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and a unit vector $\dir$.
Then, for all $\ell$ from 0 to $q-1$ and all $\beta\in(\mathbb N_0)^3$ such that $|\beta|=\ell$, $\lambda_{\varb+2e_1}$ can be expressed as a polynomial of degree at most equal to $|\varb|+1$ in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$, with coefficients depending on $\cst$ yet independent of $\dir$.
\end{prop}
\begin{proof}
We will proceed by induction with respect to $\varb$ according to the linear order $\prec$, which corresponds to the order in which the $\lambda_{\varb+2e_k} $ terms are computed in the algorithms.
The case $\varb=\mathbf 0$ is precisely \Cref{lmm:l2e1Pb}.
Assume $\varb\in(\mathbb N_0)^3$ with $0\prec\varb$ is such that the result holds for all $\varb'\prec \varb$: $\lambda_{\varb'+2e_1}$ can be expressed as a polynomial of degree at most equal to $|\varb'|+1$ in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$. Then $ \lambda_{\varb+2e_1}$ is computed according to \Cref{Algo:SubS,Algo:PbS}. Hence, since
$$
(\gamma_{k}-\eta_{k} +1)\lambda_{\gamma-\eta+e_{k}}(\eta_k +1)\lambda_{\eta+e_k} =
(\gamma_{k'}-\eta_{k'} +1)\lambda_{\gamma-\eta+e_{k'}}(\eta_k +1)\lambda_{\eta+e_k}
\text{ for } k'=k,
$$
we can gather these terms in a sum over $1\leq k\leq k'\leq 3$ and
$ \lambda_{\varb+2e_1}$ can be written as:
\begin{equation}\label{eq:lb+2e1}
\begin{array}{rl}
\lambda_{\varb+2e_1} = &\displaystyle
\frac{1}{(\varb_1+2)(\varb_1+1)\coefc{2e_1}[\mathbf 0]}\Bigg(
-\sum_{\gamma< \varb}
\sum_{k=1}^3
\coefc{2e_k}[\varb-\gamma](\gamma_k+2)(\gamma_k+1)\lambda_{\gamma+2e_k}
\\&
\displaystyle
-\sum_{\gamma\leq \varb}
\sum_{1\leq k<k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma](\gamma_k+1)(\gamma_{k'}+1)\lambda_{\gamma+e_k+e_{k'}} \\&
\displaystyle
-\sum_{\gamma\leq \varb}
\sum_{1\leq k\leq k'\leq 3}
\coefc{e_k+e_{k'}}[\varb-\gamma]
\sum_{\eta\leq \gamma}(\gamma_{k'}-\eta_{k'} +1)\lambda_{\gamma-\eta+e_{k'}}(\eta_k +1)\lambda_{\eta+e_k}
\\&\displaystyle
-
\sum_{\gamma\leq \varb} \sum_{k=1}^3
\coefc{e_k}[\varb-\gamma](\gamma_{k}+1)\lambda_{\gamma+e_k}
-
\coefczr[\varb]
\\&\displaystyle
-\sum_{k=2}^3
(\varb_k+2)(\varb_k+1)\coefc{2e_k}[\mathbf 0]\lambda_{\varb+2e_k}
\\&\displaystyle
-\sum_{1\leq k<k'\leq 3}
(\varb_k+1)(\varb_{k'}+1)\coefc{e_k+e_{k'}}[\mathbf 0]\lambda_{\varb+e_k+e_{k'}}
\Bigg),
\end{array}
\end{equation}
On the right hand side we observe that the $\lambda$ terms fall in one of two categories as elements of $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$:
\begin{enumerate}
\item $\lambda_i$ with $i_1\in\{0,1\}$, chosen in the initialization process, either a $\lambda_{e_k}$ or otherwise the terms set to zero,
\item $\lambda_i$ with $i_1\geq 2$, computed at a previous iteration for $\varb'=i-2e_1\prec\varb$.
\end{enumerate}
The linear terms can be listed as follows.
\begin{center}
{\renewcommand{\arraystretch}{1.1}
\begin{tabular}{|c||c|c|c|c|c|}
\hline
Terms & Indices & $0$ & $\lambda_{e_k}$& $\lambda_{\varb'+2e_1}$ \\\hline\hline
$\lambda_{\gamma+2e_k}$&
$\begin{array}{c}
\gamma<\varb(\neq\mathbf 0),k\in\{1,2,3\}
\\ (\gamma+2e_k)_1\in\{0,1\}
\end{array}$
&
\ding{51}
& & \\\hline
$\lambda_{\gamma+2e_k}$&
$\begin{array}{c}
\gamma<\varb(\neq\mathbf 0),k\in\{1,2,3\}
\\ (\gamma+2e_k)_1>1
\end{array}$
& & &
$ \varb' =\gamma+2(e_k-e_1) $ \\\hline
$\lambda_{\gamma+e_k+e_{k'}}$ &
$\begin{array}{c}
\gamma\leq\varb,1\leq k<k'\leq 3
\\ (\gamma+e_k+e_{k'})_1\in\{0,1\}
\end{array}$
&
\ding{51}&&\\\hline
$\lambda_{\gamma+e_k+e_{k'}}$ &
$\begin{array}{c}
\gamma\leq\varb,1\leq k<k'\leq 3
\\ (\gamma+e_k+e_{k'})_1>1
\end{array}$
& & &
$\varb'=\gamma+e_k+e_{k'}-2e_1$
\\\hline
$\lambda_{\gamma+e_k}$ &
$\gamma=\mathbf 0,k\in\{1,2,3\}$
& & \ding{51} &
\\\hline
$\lambda_{\gamma+e_k}$ &
$\begin{array}{c}
\gamma\leq\varb,k\in\{1,2,3\}
\\ ({\gamma+e_k})_1\in\{0,1\},\gamma\neq\mathbf 0
\end{array}$
& \ding{51} & &
\\\hline
$\lambda_{\gamma+e_k}$ &
$\begin{array}{c}
\gamma\leq\varb,k\in\{1,2,3\}
\\ ({\gamma+e_k})_1>1
\end{array}$
& & & $\varb'=\gamma+e_k-2e_1$
\\\hline
$\lambda_{\varb+2e_k}$&
$\begin{array}{c}
k\in\{2,3\}
\\ (\varb+2e_k)_1\in\{0,1\}
\end{array}$
&
\ding{51}
& & \\\hline
$\lambda_{\varb+2e_k}$&
$\begin{array}{c}
k\in\{2,3\}
\\ (\varb+2e_k)_1>1
\end{array}$
& & &
$ \varb' =\varb+2(e_k-e_1) $ \\\hline
$\lambda_{\varb+e_k+e_{k'}}$ &
$\begin{array}{c}
1\leq k<k'\leq 3
\\ (\varb+e_k+e_{k'})_1\in\{0,1\}
\end{array}$
&
\ding{51}&&\\\hline
$\lambda_{\varb+e_k+e_{k'}}$ &
$\begin{array}{c}
1\leq k<k'\leq 3
\\ (\varb+e_k+e_{k'})_1>1
\end{array}$
& & &
$\varb'=\varb+e_k+e_{k'}-2e_1$
\\\hline
\end{tabular}}
\end{center}
Hence the linear terms can be expressed as elements of $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$ as either $0$, or a $\lambda_{e_k}$, or by induction hypothesis as a polynomial of degree at most equal to $|\varb'|+1$. Moreover, in this last case, the values of $\varb'$ identified in the previous table are such that:
$$
\left\{\begin{array}{l}
\gamma<\varb\Rightarrow |\gamma+2(e_k-e_1)|+1<|\varb|+1,\\
\gamma\leq\varb\Rightarrow |\gamma+e_k+e_{k'}-2e_1|+1\leq |\varb|+1,\\
\gamma\leq\varb\Rightarrow |\gamma+e_k-2e_1|+1\leq |\varb|,
\end{array}\right.
\text{ and }
\left\{\begin{array}{l}
|\varb+2(e_k-e_1)|+1=|\varb|+1,\\
|\varb+e_k+e_{k'}-2e_1|+1= |\varb|+1.
\end{array}\right.
$$
In summary all the linear terms in the right hand side of \eqref{eq:lb+2e1} can be expressed as polynomials of degree at most equal to $|\varb|+1$ in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$.
As for the non-linear terms, all of the form $\lambda_{\gamma-\eta+e_{k'}}\lambda_{\eta+e_k}$ with $\eta\leq\gamma\leq\varb$, they can be described as follows.
\begin{center}
{\renewcommand{\arraystretch}{1.1}
\begin{tabular}{|c|c|c|c|c|}
\hline
Indices & $0$ & $\lambda_{e_k}$ or $\lambda_{e_{k'}}$& $\lambda_{\varb'+2e_1}$ \\\hline\hline
$\eta=\gamma=\mathbf 0$
& &
\ding{51} \ding{51}&\\\hline
$\begin{array}{c}
\eta=\mathbf 0, \gamma\neq\mathbf 0
\\ ({\gamma-\eta+e_{k'}})_1\in\{0,1\}
\end{array}$ & \ding{51} & &\\\hline
$\begin{array}{c}
\eta=\mathbf 0
\\ ({\gamma-\eta+e_{k'}})_1>1
\end{array}$ & & \ding{51}&$\varb'=\gamma+e_{k'}-2e_1$ \\\hline
$\begin{array}{c}
\eta\neq\mathbf 0
\\ ({\eta+e_{k}})_1\in\{0,1\}
\end{array}$ &\ding{51}& &\\\hline
$\begin{array}{c}
({\eta+e_{k}})_1>1
\\ \eta=\gamma
\end{array}$& &\ding{51}&$\varb'=\eta+e_k-2e_1$ \\\hline
$\begin{array}{c}
({\eta+e_{k}})_1>1
\\ ({\gamma-\eta+e_{k'}})_1\in\{0,1\}
\end{array}$ & \ding{51}&& \\\hline
$\begin{array}{c}
({\eta+e_{k}})_1>1
\\ ({\gamma-\eta+e_{k'}})_1>1
\end{array}$ & & &$\begin{array}{c}\varb'=\eta+e_k-2e_1\\\varb''=\gamma-\eta+e_{k'}-2e_1\end{array}$ \\\hline
\end{tabular}}
\end{center}
Hence the non-linear terms can be expressed as elements of $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$ as either $0$, or a $\lambda_{e_k}\lambda_{e_{k'}}$, or by induction hypothesis as a polynomial of degree at most equal to $|\varb'|+2$ or $|\varb'|+|\varb''|+2$. Moreover, in this last two cases, the values of $\varb'$ and $\varb''$ identified in the previous table are such that:
$$
\left\{\begin{array}{l}
\gamma\leq\varb\Rightarrow |\gamma+e_{k'}-2e_1|+2\leq|\varb|+1\\
\eta\leq\varb\Rightarrow |\eta+e_k-2e_1|+2\leq |\varb|+1\\
\gamma\leq\varb\Rightarrow |\eta+e_k-2e_1|+ | \gamma-\eta+e_{k'}-2e_1 | +2\leq |\varb|
\end{array}\right.
$$
In summary all the non-linear terms in the right hand side of \eqref{eq:lb+2e1} can be expressed as polynomials of degree at most equal to $|\varb|+1$ in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$.
Therefore $ \lambda_{\varb+2e_1}$ in \eqref{eq:lb+2e1} can be expressed as a polynomial of degree at most equal to $|\varb|+1$ in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$. This concludes the proof.
\end{proof}
\subsection{The reference matrix}
Following our choice of initialization and the previous results presented in \Cref{prop:Abmus,prop:Pblambdas},
we consider
a set of $p$ directions $\{ \mathbf d_l\}_{1\leq l\leq p}$ defined by two angles $(\theta_l,\varphi_l)$ as follows:
$$
\mathbf d_l =
\begin{bmatrix}
\sin\theta_l\sin\varphi_l\\\sin\theta_l\cos\varphi_l\\\cos\theta_l
\end{bmatrix}.
$$
To describe the natural choice of reference case, we then define - for a common value $\cst\in\mathbb C$ and the matrices $\Pmat$ and $\Dmat$ from \Cref{hyp:PDop} - the functions:
$$
\chi_l:
\varx\mapsto \exp \cst \Pmat\Dmat^{-1/2}\dir_l\cdot\Big(\varx-\pG \Big),$$
as well as the associated matrix in $\mathbb C^{{(n+1)(n+2)(n+3)}/{6} \times p}$:
$$
\mathsf E}%{M\!\!\!E^{n,p}_{\mathcal N(i),l}=T_{\chi_l}[i].
$$
Each entry $(\mathcal N(i),l)$ being a multiple of $(\sin\varphi_l) ^{i_1} (\cos\varphi_l)^{i_2}(\sin\theta_l)^{i_1+i_2}(\cos\theta_l)^{i_3}$, hence we define the associated matrix in $\mathbb C^{{(n+1)(n+2)(n+3)}/{6} \times p}$:
$$
\mathsf R}%{M\!\!\!R_{\mathcal N(i),l}=\frac{(\sin\varphi_l) ^{i_1} (\cos\varphi_l)^{i_2}(\sin\theta_l)^{i_1+i_2}(\cos\theta_l)^{i_3}}{i!}.
$$
and show how they are related in the following result.
\begin{lmm}Consider any set of $p$ directions $\{ \mathbf d_l\}_{1\leq l\leq p}$ as well as the matrices $\Pmat$ and $\Dmat$ from \Cref{hyp:PDop}, together with the associated ${(n+1)(n+2)(n+3)}/{6} \times p$ complex matrices $\mathsf E}%{M\!\!\!E^{n,p}$ and $\mathsf R}%{M\!\!\!R^{n,p}$.
There exists a block diagonal
matrix $D^n\in \mathbb C^{ {(n+1)(n+2)(n+3)}/{6} \times {(n+1)(n+2)(n+3)}/{6} }$ such that $\mathsf E}%{M\!\!\!E^{n,p}=D^n\mathsf R}%{M\!\!\!R^{n,p}$ and $\mathsf R}%{M\!\!\!R^{n,p}=D^n\mathsf E}%{M\!\!\!E^{n,p}$, independently of their number $p$ of columns.
\end{lmm}
\begin{proof}
The entries of the reference matrix $\mathsf E}%{M\!\!\!E^{n,p}$ are:
$$
T_{\chi_l}[i] = \frac{1}{i!}
\left( \left( \cst \Pmat\Dmat^{-1/2} \dir_l \right)_1\right)^{i_1}
\left( \left( \cst \Pmat\Dmat^{-1/2} \dir_l \right)_2\right)^{i_2}
\left( \left( \cst \Pmat\Dmat^{-1/2} \dir_l \right)_3\right)^{i_3},
$$
and we consider row blocks corresponding to increasing values of $|i|$: for $m$ from $0$ to $n$ we denote by $\mathsf E}%{M\!\!\!E^{n,p}_{(m)}$ and $\mathsf R}%{M\!\!\!R^{n,p}_{(m)}$ the ${(m+1)(m+2)}/{2}\times p$ blocks of rows corresponding to all $i$ such that $|i|=m$, that is $\left(\mathsf E}%{M\!\!\!E^{n,p}_{(m)}\right)_{\mathcal N_m(i) l} = \left(\mathsf E}%{M\!\!\!E^{n,p}\right)_{\mathcal N(i)l}$.
As a reminder, $ \cst\in\mathbb C$ is a constant, while $\Pmat$ and $\Dmat$ depend only on the value of the PDE coefficients $c$ evaluated at $\pG$, hence the entries of the $3\times 3$ matrix $ \cst \Pmat\Dmat^{-1/2}$ are independent of $l$; moreover the matrix $\cst \Pmat\Dmat^{-1/2}$ is non-singular under \Cref{hyp:PDop}.
Thus, according to \Cref{lmm:prodsvecents} below for $A:=\cst \Pmat\Dmat^{-1/2}$, for all $m$ from $0$ to $n$ there exist an
${(m+1)(m+2)}/{2}\times{(m+1)(m+2)}/{2}$ matrix $\mathfrak C^{(m)}$, defined entry-wise by $\left(\mathfrak C^{(m)}\right)_{\mathcal N_m(i)\mathcal N_m(j)}=\mathfrak c_{ij}$, such that
$\mathsf E}%{M\!\!\!E^{n,p}_{(m)}=\mathfrak C^{(m)}\mathsf R}%{M\!\!\!R^{n,p}_{(m)}$
and $\mathsf R}%{M\!\!\!R^{n,p}_{(m)}=\mathfrak C^{(m)}\mathsf E}%{M\!\!\!E^{n,p}_{(m)}$.
Defining the block diagonal matrix $D^n:=\text{diag}\left(\mathfrak C^{(0)},\mathfrak C^{(1)},...,\mathfrak C^{(n)}\right)$,
we then obtain the desired property: $\mathsf E}%{M\!\!\!E^{n,p}=D^n\mathsf R}%{M\!\!\!R^{n,p}$ and $\mathsf R}%{M\!\!\!R^{n,p}=D^n\mathsf E}%{M\!\!\!E^{n,p}$.
\end{proof}
\begin{lmm}
\label{lmm:prodsvecents}
Let $A\in\mathbb C^{3\times 3}$ be non-singular. Then, for any vector $Y\in\mathbb C^3$, any product of powers of the entries of $AY$ with a total power $m\in \mathbb N$ can be written as a linear combination - depending only on $A$ but independent of $Y$ - of products of powers of the entries of $Y$ each one of the products having a total power equal to $m$. Mathematically speaking:
\begin{equation}
\label{eq:AYvsY}
\forall i\in(\mathbb N_0)^3,
\exists\ \{\mathfrak c_{ij}\in\mathbb C,j\in(\mathbb N_0)^3, |j|=|i|\} \text{ such that }
\forall Y\in\mathbb C^3,\
\prod_{k=1}^3 \big((AY)_k\big)^{i_k}
=
\sum_{|j|=|i|} \mathfrak c_{ij} \prod_{l=1}^3 (Y_l)^{j_l} .
\end{equation}
Moreover:
\begin{equation}
\label{eq:XvsA-1X}
\forall X\in\mathbb C^3,\
\prod_{k=1}^3 \big(X_k\big)^{i_k}
=
\sum_{|j|=|i|} \mathfrak c_{ij} \prod_{l=1}^3 \left((A^{-1}X)_l\right)^{j_l} .
\end{equation}
\end{lmm}
\begin{rmk}
The result still holds for $n\in\mathbb N$ and $A\in\mathbb C^{n\times n}$, the proof is more tedious as it requires the introduction of more indices. In this article we only use the $n=3$ case hence we do not prove the more general case.
\end{rmk}
\begin{proof}
Since, for $k\in\{1,2,3\}$ and any $3\times 3$ matrix $A$, we have:
$$
\big((AY)_k\big)^{i_k}
= \left(\sum_{l=1}^3 A_{kl}Y_l\right)^{i_k}
= \sum_{|j|=i_k} \frac{(i_k)!}{j!} \prod_{l=1}^3(A_{kl}Y_l)^{j_l}
= \sum_{|j|=i_k}\left( \frac{(i_k)!}{j!} \prod_{\tilde l=1}^3(A_{k\tilde l})^{j_{\tilde l}}\right) \prod_{l=1}^3(Y_l)^{j_l},
$$
then we get:
$$
\prod_{k=1}^3\big((AY)_k\big)^{i_k}
= \sum_{|j|=i_1}\sum_{|j'|=i_2}\sum_{|j''|=i_3} \frac{i!}{j!(j')!(j'')!}
\left(\prod_{\tilde l=1}^3(A_{1\tilde l})^{j_{\tilde l}}(A_{2\tilde l})^{j'_{\tilde l}}(A_{3\tilde l})^{j''_{\tilde l}}\right) \prod_{l=1}^3(Y_l)^{j_l+j'_l+j''_l}.
$$
This concludes the proof of \eqref{eq:AYvsY}, as $\displaystyle \sum_{l=1}^3\big(j_l+j'_l+j''_l\big) = |j|+|j'|+|j''| = i_1+i_2+i_3 = |i| $.
Moreover, if $A$ is non-singular, then, for any $X\in\mathbb C^3$, \eqref{eq:AYvsY} applied to $Y=A^{-1}X$ shows \eqref{eq:XvsA-1X}
\end{proof}
As a direct consequence we have the following results.
\begin{cor}
\label{cor:rkmatexp}
Given $(n,p)\in\mathbb N^2$ and any choice of $p$ directions $\{ \mathbf d_l\in\mathbb S^2\}_{1\leq l\leq p}$,
the rank of $\mathsf E}%{M\!\!\!E^{n,p}$ is equal to the rank of $\mathsf R}%{M\!\!\!R^{n,p}$.
\end{cor}
\begin{cor}
\label{cor:submatexp}
Similarly, any pair of sub-matrices of $\mathsf E}%{M\!\!\!E^{n,p}$ and $\mathsf R}%{M\!\!\!R^{n,p}$ corresponding to removing the \underline{same} sets of all rows $i$ with $|i|=m$ for a given $m$ - in particular removing all rows $i$ with $|i|>1$ - have the same rank.
\end{cor}
The next step is to study the rank of these matrices, and the following step will be to relate it to the rank of the quasi-Trefftz matrices $\mathsf A}%{M\hspace{-.75em}A$, $\mathsf P}%{M\!\!\!P$ and $\mathsf Q}%}{M\!\!\!Q$.
\subsection{Properties of the reference matrix}
We are interested here in the rank of the reference matrix, $\mathsf R}%{M\!\!\!R^{n,p}$, or equivalently the rank of the exponential matrix $\mathsf E}%{M\!\!\!E^{n,p}$ according to \Cref{cor:rkmatexp}, as well as the rank of some of their useful sub-matrices. In particular the rank's value depends on the number of columns $p$, and we will see that:
\begin{itemize}
\item independently of the number of columns $p$, the rank is at most $(n+1)^2$,
\item while there exist sets of $p=(n+1)^2$ directions that guarantee the rank of the corresponding matrix to be maximal, that is $(n+1)^2$.
\end{itemize}
To study this reference matrix, we first remark that its entries are a set of functions, on each row, evaluated at a set of points, on each column. The following result provides the maximum possible rank of such matrices depending on the dimension of the space generated by the function set.
\begin{lmm}
\label{lmm:matsfuncsevals}
Consider a set $\mathcal D$ of $N_f$ complex-valued functions defined on a domain $\Omega$ (in any dimension), denoted $d_k$, while the dimension of $span \ \mathcal D$ is $N_d<N_f$. We consider any $N_f\times N_p$ matrix $\mathsf D$ defined element wise by evaluating the elements of the function space $\mathcal D$ at a set of $N_p$ points $\{\Theta_l\in\Omega\}_{1\leq l\leq N_p}$, namely:
$$
\mathsf D_{kl} = d_k(\Theta_l).
$$
Then the rank of $\mathsf D$ is at most equal to $N_d$.
\footnote{
This argument was previously presented in our roadmap paper for a particular case defined by the function space $ \mathcal F_n=\{\theta\mapsto\cos^k \theta\sin^{K-k}\theta/(k!(K-k)!), 0\leq k\leq K\leq n\}$, containing $\# \mathcal F_n=(n+1)(n+2)/2$ functions and spanning a space of dimension $\dim span \mathcal F_n=2n+1$.
}
Moreover, if ${\mathcal B}=\{y_k, 1\leq k\leq N_b\}$ is any generating set of $span\ \mathcal D$, then the corresponding $N_b\times N_p$ matrix ${\mathsf B}$, namely:
$$
{\mathsf B}_{kl} = {y}_k(\Theta_l),
$$
has the same rank as $\mathsf D$.
\end{lmm}
\begin{proof}
Because of the number of elements in $\mathcal D$ and the dimension of its span, there exists a matrix $\mathsf C\in\mathbb C^{(N_f-N_d)\times N_f}$, of rank $N_f-N_d$ such that
$$
\forall k'\in\mathbb N, k'\leq N_f-N_d,\sum_{k=1}^{N_f}\mathsf C_{k'k}d_k = 0.
$$
In particular, independently of the number $N_p$ of columns of $\mathsf D$, this yields:
$$
\mathsf C\mathsf D = 0_{(N_f-N_d)\times N_p}.
$$
Hence the $N_p$ columns of $\mathsf D$ belong to the kernel of $\mathsf C$, which is of dimension $N_d$ according to the rank-nullity theorem.
So indeed the rank of $\mathsf D$ is at most equal to $N_d$.
Moreover, consider a set $\widetilde{\mathcal D}\subset\mathcal D$, being a generating set for $span\ \mathcal D$, and the submatrix of $\mathsf D$, denoted $\widetilde{\mathsf D}$ and obtained by keeping only the rows corresponding to each $d_k\in\widetilde{\mathcal D}$.
Since the space generated by the rows of $\mathsf D$ is the same as the space generated by the rows of $\widetilde{\mathsf D}$,
$\mathsf D$ and $\widetilde{\mathsf D}$ have the same rank.
Similarly, consider ${\mathcal B}=\{y_k, 1\leq k\leq N_b\}$ any generating set of $span\ \mathcal D$, then consider the corresponding matrix ${\mathsf B}$.
Each of its rows can be written as a linear combination of the rows of $\widetilde{\mathsf D}$, since $\widetilde{\mathcal D}$ is a generating set of $span\ \mathcal D$, while each row of $\widetilde{\mathsf D}$ can be written as a linear combination of the rows of ${\mathsf B}$, since ${\mathcal B}$ is also a generating set of $span\ \mathcal D$.
This proves the second claim.
\end{proof}
To address the particular case of the matrix $\mathsf R}%{M\!\!\!R$, we then define the following functions and function space:
$$
\forall i\in(\mathbb N_0)^3, \ f_i(\theta,\varphi) = (\sin\varphi) ^{i_1} (\cos\varphi)^{i_2}(\sin\theta)^{i_1+i_2}(\cos\theta)^{i_3}/i!
\text{ and }
\mathcal F_n :=\{f_i, i\in(\mathbb N_0)^3, |i|\leq n\},
$$
$$
\widetilde{\mathcal F}_n :=\{f_i, i\in(\mathbb N_0)^3, |i|\leq n,i_1\in\{0,1\}\}.
$$
The space $\mathcal F_n$ contains ${(n+1)(n+2)(n+3)}/{6} $ elements, let's now identify the dimension of $Span\mathcal F_n$.
We will make use of the following functions and function spaces:
$$
Z_l^m(\theta,\varphi) = e^{\textrm i m\varphi}(\sin\theta)^{|m|}(\cos\theta)^{l-|m|}
\text{ and }
\mathcal G_n :=\{Z_l^m,0\leq l\leq n,-l\leq m\leq l\},
$$
and for the spherical harmonics with
$
C_l^m = \sqrt{\frac{2l+1}{4\pi}\frac{(l-m)!}{(l+m)!}}
$ and the
Legendre polynomials $P_l^m$ (see appendix \ref{sec:SHrem}):
$$
Y_l^m(\theta,\varphi):=C_l^m P_l^m(\cos\theta)e^{\textrm i m\varphi}
\text{ and }
\mathcal H_n :=\{Y_l^m,0\leq l\leq n,-l\leq m\leq l\}.
$$
\begin{lmm}
\label{lmm:dimFn}
Given $n\in\mathbb N$, the spaces of trigonometric functions and spherical harmonics are such that:
\begin{itemize}
\item $Span\widetilde{\mathcal F}_n = Span\mathcal F_n = Span\mathcal G_n = Span\mathcal H_n$,
\item they are all of dimension $(n+1)^2$.
\end{itemize}
\end{lmm}
\begin{proof}
{\bf Step 1.}
One clearly sees that $\widetilde{\mathcal F}_n \subset {\mathcal F}_n$, we show that every element of ${\mathcal F}_n$ belongs to $Span\overline{\mathcal F}_n$.
Considering any function $f_i$ in $\mathcal F_n$, according to $i_1$ being even or odd, we will treat the two cases as follows.
If $i_1$ is even, since $(\sin\varphi )^{i_1}=(1-\cos^2\varphi) ^{i_1/2}$, then we have:
$$
f_i(\theta,\varphi)
= \sum_{k_\varphi = 0}^{i_1/2}
\begin{pmatrix} i_1/2\\k_\varphi \end{pmatrix}
(-1)^{k_\varphi}
(\cos\varphi)^{2k_\varphi+i_2}(\sin\theta)^{i_1+i_2}(\cos\theta)^{i_3}/i!,
$$
and since $(\sin\theta)^{i_1+i_2} = (\sin\theta)^{2k_\varphi+i_2}(1-\cos^2\theta)^{i_1/2-k_\varphi}$, we can write:
$$
f_i(\theta,\varphi)
= \sum_{k_\varphi = 0}^{i_1/2} \sum_{k_\theta = 0}^{i_1/2-k_\varphi}
\begin{pmatrix} i_1/2\\k_\varphi \end{pmatrix}
\begin{pmatrix} i_1/2-k_\varphi\\k_\theta \end{pmatrix}
(-1)^{k_\varphi+k_\theta}
\sin^0\varphi (\cos\varphi)^{2k_\varphi+i_2}(\sin\theta)^{2k_\varphi+i_2}(\cos\theta)^{2k_\theta+i_3}/i!.
$$
We then easily see that $0+(2k_\varphi+i_2) =2k_\varphi+i_2 $ while $(2k_\varphi+i_2)+(2k_\theta+i_3)\leq 2k_\varphi+i_2+i_1-2k_\varphi+i_3=|i|$, hence in particular $f_i\in Span\overline{\mathcal F}_n$.
If $i_1$ is odd, since $(\sin\varphi )^{i_1}=\sin\varphi(1-\cos^2\varphi) ^{(i_1-1)/2}$, then we have:
$$
f_i(\theta,\varphi)
= \sum_{k_\varphi = 0}^{(i_1-1)/2}
\begin{pmatrix} (i_1-1)/2\\k_\varphi \end{pmatrix}
(-1)^{k_\varphi}
\sin\varphi (\cos\varphi)^{2k_\varphi+i_2}(\sin\theta)^{i_1+i_2}(\cos\theta)^{i_3}/i!,
$$
and since $(\sin\theta)^{i_1+i_2} = (\sin\theta)^{2k_\varphi+1+i_2}(1-\cos^2\theta)^{(i_1-1)/2-k_\varphi}$, we can write:
$$
f_i(\theta,\varphi)
= \sum_{k_\varphi = 0}^{\frac{i_1-1}2} \sum_{k_\theta = 0}^{\frac{i_1-1}2-k_\varphi}
\begin{pmatrix} \frac{i_1-1}2\\k_\varphi \end{pmatrix}
\begin{pmatrix} \frac{i_1-1}2-k_\varphi\\k_\theta \end{pmatrix}
(-1)^{k_\varphi+k_\theta}
\sin\varphi (\cos\varphi)^{2k_\varphi+i_2}(\sin\theta)^{2k_\varphi+1+i_2}(\cos\theta)^{2k_\theta+i_3}/i!.
$$
We then easily verify that $1+(2k_\varphi+i_2) =2k_\varphi+1+i_2 $ while $(2k_\varphi+1+i_2)+(2k_\theta+i_3)\leq 2k_\varphi+1+i_2+i_1-1-2k_\varphi+i_3=|i|$, hence again $f_i\in Span\widetilde{\mathcal F}_n$.
This proves that $Span\ {\mathcal F}_n = Span\ \widetilde{\mathcal F}_n$.
{\bf Step 2.}
Let's start by considering any element $f_i$ of $\mathcal F_n$ to show that it belongs to $ Span\mathcal G_n$.
For all $i\in(\mathbb N_0)^3$ such that $|i|\leq n$, writing $\cos\varphi,\sin\varphi$ under their exponential form, and expanding their powers according to the binomial formula, we have:
$$
f_i(\theta,\varphi)
=
\frac1{i!(2\textrm i)^{i_1}2^{i_2}}(\sin\theta)^{i_1+i_2}\left(
\sum_{j_1=0}^{i_1}
\sum_{j_2=0}^{i_2}
(-1)^{j_1}
\begin{pmatrix} i_1\\j_1\end{pmatrix}
\begin{pmatrix} i_2\\j_2\end{pmatrix}
\exp\big( \textrm i (i_1-2j_1+i_2-2j_2)\varphi\big)
\right)
(\cos\theta)^{i_3}.
$$
In order to write each
$(\sin\theta)^{i_1+i_2}
\exp \big(\textrm i (i_1-2j_1+i_2-2j_2)\varphi\big)
(\cos\theta)^{i_3}$
term in this sum as a linear combination of $Z_l^m$s, the power of $\sin\theta$ must be the absolute value of the power of $\exp \textrm i \varphi$.
Therefore, defining $M_{i,j_1,j_2}:=\min( i_1-j_1+i_2-j_2,j_1+j_2)$, we can write $i_1+i_2$ independently of the sign of $i_1-2j_1+i_2-2j_2$ as:
$$
i_1+i_2 = |i_1-2j_1+i_2-2j_2| +2 M_{i,j_1,j_2}.
$$
This leads to
\begin{equation*}
\begin{array}{l}
(\sin\theta)^{i_1+i_2}
\exp \big(\textrm i (i_1-2j_1+i_2-2j_2)\varphi\big)
(\cos\theta)^{i_3}\\\displaystyle \phantom{(\sin)}
=(\sin\theta)^{ |i_1-2j_1+i_2-2j_2|}
(1-\cos^2\theta)^{M_{i,j_1,j_2}}
\exp\big( \textrm i (i_1-2j_1+i_2-2j_2)\varphi\big)
(\cos\theta)^{i_3}
\\\displaystyle \phantom{(\sin)}
=
\sum_{j_3=0}^{M_{i,j_1,j_2}}
\begin{pmatrix} M_{i,j_1,j_2}\\j_3\end{pmatrix}
(-1)^{j_3}
(\sin\theta)^{|i_1-2j_1+i_2-2j_2|}
(\cos\theta)^{2j_3+i_3}
\exp \big(\textrm i (i_1-2j_1+i_2-2j_2)\varphi\big),
\end{array}
\end{equation*}
so
\begin{equation*}
f_i(\theta,\varphi)
=
\frac1{i!(2\textrm i)^{i_1}2^{i_2}}
\sum_{j_1=0}^{i_1}
\sum_{j_2=0}^{i_2}
(-1)^{j_1}
\begin{pmatrix} i_1\\j_1\end{pmatrix}
\begin{pmatrix} i_2\\j_2\end{pmatrix}
\sum_{j_3=0}^{M_{i,j_1,j_2}}
\begin{pmatrix} M_{i,j_1,j_2}\\j_3\end{pmatrix}
(-1)^{j_3}
Z_{|i_1-2j_1+i_2-2j_2|+2j_3+i_3}^{|i_1-2j_1+i_2-2j_2|} (\theta,\varphi).
\end{equation*}
To verify that these $Z_l^m$s belong to the space $Span \mathcal G_n$, we must make sure that their indices satisfy $0\leq l\leq n$ and $|m|\leq l$. In fact:
\begin{itemize}
\item if $i_1-2j_1+i_2-2j_2\geq 0$ then
$$
\begin{array}{rl}
j_3\leq j_1+j_2&\Rightarrow i_1-2j_1+i_2-2j_2+2j_3+i_3\leq i_1+i_2+i_3,\\
i_1-2j_1+i_2-2j_2\geq 0&\Rightarrow i_1-2j_1+i_2-2j_2+2j_3+i_3\geq 0,\\
i_1-2j_1+i_2-2j_2= |i_1-2j_1+i_2-2j_2|&\Rightarrow |i_1-2j_1+i_2-2j_2|\leq i_1-2j_1+i_2-2j_2+2j_3+i_3.
\end{array}
$$
\item if $i_1-2j_1+i_2-2j_2 < 0$ then
$$
\begin{array}{rl}
j_3\leq i_1-j_1+i_2-j_2 &\Rightarrow 2j_1-i_1+2j_2-i_2+2j_3+i_3\leq i_1+i_2+i_3,\\
2j_1-i_1+2j_2-i_2\geq 0&\Rightarrow 2j_1-i_1+2j_2-i_2+2j_3+i_3\geq 0,\\
2j_1-i_1+2j_2-i_2= |i_1-2j_1+i_2-2j_2|&\Rightarrow |2j_1-i_1+2j_2-i_2|\leq 2j_1-i_1+2j_2-i_2+2j_3+i_3.
\end{array}
$$
\end{itemize}
To summarize, since $i_1+i_2+i_3=|i|$, the indices of each $Z_{|i_1-2j_1+i_2-2j_2|+2j_3+i_3}^{|i_1-2j_1+i_2-2j_2|} $ are such that:
$$\left\{
\begin{array}{l}
| i_1-2j_1+i_2-2j_2|+2j_3+i_3\leq |i|,\\
| i_1-2j_1+i_2-2j_2|+2j_3+i_3 \geq 0,\\
|i_1-2j_1+i_2-2j_2|\leq | i_1-2j_1+i_2-2j_2|+2j_3+i_3,
\end{array}\right.
$$
Hence, for $|i|\leq n$, $f_i$ can indeed be written as a linear combination of $Z_{l}^{m}$ with $0\leq l\leq n$ and $|m|\leq l$. In other words, each element of $\mathcal F_n$ belongs to $Span\mathcal G_n$
Besides, any element $Z_l^{\pm n}$ of $\mathcal G_n$ can also be written as a linear combination of $f_i$s. In fact, expanding $e^{\textrm i m\varphi}$ via the binomial formula we immediately see that for all $(l,m)\in\mathbb Z^2$ such that $0\leq l \leq n$ and $-l\leq m\leq l$ we have:
$$
Z_l^{\pm m}(\theta,\varphi) = \sum_{m' = 0}^{|m|} \begin{pmatrix}{|m|}\\m'\end{pmatrix} (\cos\varphi)^{m'}(\pm \textrm i\sin\varphi)^{|m|-m'}(\sin\theta)^{|m|}(\cos\theta)^{l-|m|},
$$
or equivalently
$$
Z_l^{\pm m}(\theta,\varphi)
= \sum_{m' = 0}^{|m|} (\pm \textrm i )^{|m|-m'}
(|m|)!(l-|m|)!f_{\left(|m|-m',m',l-|m|\right)},
\text{ where } \left|\left(|m|-m',m',l-|m|\right)\right| = l\leq n,
$$
hence each element of $\mathcal G_n$ belongs to $Span\mathcal F_n$.
As a consequence, $Span\mathcal G_n=Span\mathcal F_n$.
{\bf Step 3.}
Any function in $\mathcal H_n$ can be written
as
$$
\begin{array}{rl}
\text{if }m\geq0,\
Y_l^m(\theta,\phi)
&\displaystyle
=C_l^m \frac{(-1)^m}{2^ll!} (\sin\theta)^m e^{\textrm i m\phi}
\sum_{l'=\left\lceil\frac{l+m}2\right\rceil}^l
\begin{pmatrix} l\\l'\end{pmatrix} (-1)^{l-l'} \frac{(2l')!}{(2l'-l-m)!}(\cos\theta)^{2l'-l-m}
\\&\displaystyle
=C_l^m \frac{(-1)^m}{2^ll!}
\sum_{l'=\left\lceil\frac{l+m}2\right\rceil}^l
\begin{pmatrix} l\\l'\end{pmatrix} (-1)^{l-l'} \frac{(2l')!}{(2l'-l-m)!} Z_{2l'-l}^m
\end{array}
$$
and
$$
\begin{array}{rl}
\text{if }m<0,\
Y_l^m(\theta,\phi)
&\displaystyle
=C_l^m \frac{(l-|m|)!}{(l+|m|)!}\frac{1}{2^ll!} (\sin\theta)^{|m|} e^{\textrm i m\phi}
\sum_{l'=\left\lceil\frac{l+m}2\right\rceil}^l
\begin{pmatrix} l\\l'\end{pmatrix} (-1)^{l-l'} \frac{(2l')!}{(2l'-l-m)!}(\cos\theta)^{2l'-l-|m|}
\\&\displaystyle
=C_l^m \frac{(l-|m|)!}{(l+|m|)!}\frac{1}{2^ll!}
\sum_{l'=\left\lceil\frac{l+m}2\right\rceil}^l
\begin{pmatrix} l\\l'\end{pmatrix} (-1)^{l-l'} \frac{(2l')!}{(2l'-l-m)!}Z_{2l'-l}^m
\end{array}
$$
hence each element of $\mathcal H_n$ belongs to $Span\mathcal G_n$.
So $Span\mathcal H_n\subset Span\mathcal G_n$.
Moreover, by property of the spherical harmonics, these are linearly independent, therefore
$$
\dim \mathcal H_n = (n+1)^2.
$$
We have then shown that
$$
(n+1)^2\leq \dim Span\mathcal G_n.
$$
But the space $\mathcal G_n$ has $(n+1)^2$ elements, so $\dim Span\mathcal G_n=(n+1)^2$.
{\bf Conclusion}
The result follows trivially from combining the previous steps.
\end{proof}
As direct consequences of \Cref{lmm:dimFn,lmm:matsfuncsevals} as well as \Cref{cor:rkmatexp,cor:submatexp}, we get the following.
\begin{cor}
\label{cor:maxrk}
Given $(n,p)\in\mathbb N^2$,
for any choice of $p$ directions $\{ \mathbf d_l\}_{1\leq l\leq p}$3,
the rank of $\mathsf E}%{M\!\!\!E^{n,p}$, and hence the rank of $\mathsf R}%{M\!\!\!R^{n,p}$, cannot be larger than $(n+1)^2$.
Moreover the submatrices $\overline{\mathsf E}%{M\!\!\!E}^{n,p} $ and $\overline{\mathsf R}%{M\!\!\!R}^{n,p} $ have the same rank as $\mathsf R}%{M\!\!\!R^{n,p}$ as well.
\end{cor}
The next natural question is that of existence of a set of $(n+1)^2$ directions $\{ (\theta_l,\varphi_l)\in\mathbb R^2\}_{1\leq l\leq (n+1)^2}$ such that the corresponding matrices are of rank $(n+1)^2$.
Hence it is natural to now fix $p=(n+1)^2$, denoting hereafter the corresponding matrices $\mathsf E}%{M\!\!\!E^{[n]}$ and $\mathsf R}%{M\!\!\!R^{[n]}$. If such a set of directions exists, then increasing the value of $p$ will not increase the rank according to \Cref{cor:maxrk}.
\begin{theorem}
\label{thm:matmaxrk}
Given $n\in\mathbb N$, let $p=(n+1)^2$ directions on $\mathbb S^2$ be chosen as:
$$
\mathbf d_{l,m} = (\sin\theta_l\cos\varphi_{lm},\sin\theta_l\sin\varphi_{lm},\cos\theta_l)
$$
for all $l$ from $0$ to $n$ with $|m|\leq l$, where the $n+1$ colatitude angles $\{\theta_l\}_{0\leq l\leq n}\subset (0,\pi)$ are all different from each other, and the azimuths $\{\varphi_{lm}\}_{0\leq l\leq n,|m|\leq l}\subset [0,2\pi)$ satisfy $\varphi_{lm}\neq\varphi_{lm'}$ for every $m\neq m'$.
Then the two $\frac{(n+1)(n+2)(n+3)}{6} \times (n+1)^2$ matrices $\mathsf R}%{M\!\!\!R^{[n]} $ and $\mathsf E}%{M\!\!\!E^{[n]} $ are of rank $(n+1)^2$.
Moreover, their square sub-matrices of $\mathsf R}%{M\!\!\!R^{[n]} $ and $\mathsf E}%{M\!\!\!E^{[n]} $ corresponding to removing the rows $i$ such that $i_1>1$, denoted respectively $\overline{\mathsf R}%{M\!\!\!R}^{[n]} $ and $\overline{\mathsf E}%{M\!\!\!E}^{[n]} $, are also of rank $(n+1)^2$.
\end{theorem}
Here again, properties of spherical harmonics will be at the center of the proof.
\begin{proof}
Since $\mathcal H_n$ is a generating set for $Span\ \mathcal F_n$, then according to \Cref{lmm:matsfuncsevals} the rank of $\mathsf R}%{M\!\!\!R^{[n]}$ is that of the $(n+1)^2\times(n+1)^2$ matrix $\mathsf S_n$ defined element wise by evaluating the elements of the function space $\mathcal H_n$ at a set of $p=(n+1)^2$ points $\{(\theta_l,\varphi_{lm})\}_{0\leq l\leq n,|m|\leq l}\subset [0,2\pi)$.
As stated in Lemmas 3.4.1 and 3.4.2 from \cite{moiolaTH}, this matrix $\mathsf S_n$ is invertible. As a conclusion $\mathsf R}%{M\!\!\!R^{[n]} $ is indeed of rank $(n+1)^2$. As a direct consequence of \Cref{cor:rkmatexp}, $\mathsf E}%{M\!\!\!E^{[n]} $ is also of rank $(n+1)^2$.
Since $\overline{\mathcal F}_n$ is a generating set for $Span\ \mathcal F_n$ from \Cref{lmm:dimFn}, applying the second statement from \Cref{lmm:matsfuncsevals} then proves the second statement.
\end{proof}
\begin{rmk}
In terms of directions, the result is actually true for almost any set of points $\{ (\theta_l,\varphi_l)\in(0,\pi)\times[0,2\pi)\}_{0\leq l\leq n, |m|\leq l}$ as Lemma 3.4.1 from \cite{moiolaTH} states that the rank of the matrix $\mathsf S_n$ is $(n+1)^2$ for a dense open set of $(\mathbb S^2)^p$.
\end{rmk}
\subsection{Relating quasi-Trefftz and reference matrices}\label{ssec:Relate}
We are now interested in the rank of the quasi-Trefftz matrices, namely the $(n+1)(n+2)(n+3)/6\times p$ matrices $\mathsf A}%{M\hspace{-.75em}A$, $\mathsf P}%{M\!\!\!P$ and $\mathsf Q}%}{M\!\!\!Q$.
In order to leverage the properties of the reference matrix to prove properties of these quasi-Trefftz matrices,
the missing link is then to understand their relation to the reference matrix.
Because of their polynomial component, either in the phase, the amplitude, or the function itself, quasi-Trefftz functions have derivatives that share a common structure when evaluated at $\pG$ as long as their order is not higher than $q+1$. Even though higher order derivatives could also be studied, they would not share this common structure.
Hence we will start by studying the common properties of such derivatives of quasi-Trefftz functions.
Under the assumption that $q\geq n-1$, the entries in the quasi-Trefftz matrices are precisely such derivatives evaluated at a $\pG$. This will then allow us to establish a relation between the reference matrix and each of the quasi-Trefftz matrices.
Noticeably, neither the number of quasi-Trefftz functions chosen to construct each matrix, denoted $p$, nor the set of directions in the initialization come into play in this procedure: we establish relations between matrices independently of both. However, proving that the quasi-Trefftz functions can be constructed to guarantee that the corresponding quasi-Trefftz matrices have maximal rank will rely on an appropriate choices for $p$ and the set of directions.
\subsubsection{The GPW case}
To address the two GPW cases, it is natural to start from expressing the derivatives of GPW functions in terms of the three non-zero free parameters in the initialization, leveraging \Cref{prop:Pblambdas,prop:Abmus}.
\begin{prop}
\label{prop:AbME
Given $q\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any amplitude-based GPW associated to partial differential operator $\mathcal L_{c}$, $J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big)$ with $\displaystyle Q:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \mu_i \mathbf X^i $, constructed via \Cref{Algo:AbS,Algo:SubS}, with the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and a unit vector $\dir\in\mathbb S^2$.
Then, for all $j\in(\mathbb N_0)^3$ such that $|j|\leq q+1$, the difference $\partial_x^j J(\pG)/j!-(\Lambda_{1})^{j_1}(\Lambda_{2})^{j_2}(\Lambda_{3})^{j_3}/j!$ can be expressed as a polynomial in $\mathbb C[\Lambda_{1},\Lambda_{2},\Lambda_{3}]$, with degree smaller than $|j|$ and coefficients depending on $\cst$ yet independent of $\dir$.
\end{prop}
\begin{proof}
Since $J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big)$, we can express $\partial_\varx^j J$ thanks to Leibniz's rule as:
$$\forall x\in\mathbb R^3,
\partial_\varx^j J(\varx) = \sum_{\widetilde j\leq j}
\begin{pmatrix}j\\\widetilde j\end{pmatrix}
\Lambda^{j-\widetilde j} \partial_\varx^{\widetilde j}\polQ(\varx-\pG),
$$
and therefore
$$
\partial_\varx^j J(\pG) = \sum_{\widetilde j\leq j}
\frac{j!}{(j-\widetilde j)!}
\Lambda^{j-\widetilde j} \mu_{\widetilde j}.
$$
Let's consider the terms in this linear combination, starting from considering the possible degree of each individual term as a polynomial in $\mathbb C[\Lambda_{1},\Lambda_{2},\Lambda_{3}]$:
\begin{enumerate}
\item for $|\widetilde j|=0$, then $\Lambda^{j-\widetilde j} \mu_{\widetilde j}=(\Lambda_{1})^{j_1}(\Lambda_{2})^{j_2}(\Lambda_{3})^{j_3}$ is a polynomial of degree $|j|$,
\item for $|\widetilde j|=1$, then $\Lambda^{j-\widetilde j} \mu_{\widetilde j}=0$,
\item for $|\widetilde j|>1$, then $\Lambda^{j-\widetilde j} \mu_{\widetilde j}$ can be expressed as a polynomial of degree at most equal to $|j-\widetilde j|+|\widetilde j|-1<|j|$
according \Cref{prop:Abmus}.
\end{enumerate}
In the first case, the corresponding weight in the linear combination is precisely $j!/j!=1$, while all other terms can be expressed as a polynomial in $\mathbb C[\Lambda_{1},\Lambda_{2},\Lambda_{3}]$ with total degree smaller than $|j|$. This concludes the proof.
These polynomials have coefficients independent of $\dir$ as a consequence of
\Cref{prop:Abmus}.
\end{proof}
\begin{prop}
\label{prop:PbME
Given $q\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any phase-based GPW associated to differential operator $\mathcal L_{c}$, $G(\varx) :=\exp \polP(\varx-\pG)$ with $\displaystyle P:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \lambda_i \mathbf X^i $, constructed via \Cref{Algo:PbS,Algo:SubS}, with the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and a unit vector $\dir\in\mathbb S^2$.
Then, for all $j\in(\mathbb N_0)^3$ such that $|j|\leq q+1$, the difference $\partial_x^j G(\pG)/j!-(\lambda_{e_1})^{j_1}(\lambda_{e_2})^{j_2}(\lambda_{e_3})^{j_3}/j!$ can be expressed as a polynomial in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$, with degree smaller than $|j|$ and coefficients depending on $\cst$ yet independent of $\dir$.
\end{prop}
\begin{proof}
Since $G(\varx)=\exp \polP(\varx-\pG)$, we can express $\partial_\varx^j G$ thanks to Faa di Bruno's formula (see Appendix \ref{App:FdB}) as:
$$\forall x\in\mathbb R^3,
\partial_\varx^j G(\varx) =
\sum_{1\leq \widetilde m\leq |j|}
\exp(P(\varx-\pG)) \sum_{s=1}^{|j|}\sum_{p_s(j,\widetilde m)} j!\prod_{m=1}^s
\frac{1}{k_m!}
\left(\frac{1}{l_m!} \partial_x^{l_m} P(\varx-\pG) \right)^{k_m},
$$
where sets $p_s$ as well as indices $l_m$ and $k_m$ are as defined in \Cref{App:FdB},
and therefore
$$
\partial_\varx^j G(\pG) =
\sum_{1\leq \widetilde m\leq |j|}
\sum_{s=1}^{|j|}\sum_{p_s(j,\widetilde m)} j!\prod_{m=1}^s
\frac{1}{k_m!}
\left(\lambda_{l_m} \right)^{k_m}.
$$
Let's consider the terms in this linear combination, starting from considering the possible degree of each individual $\lambda_{l_m}$ term as a polynomial in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$:
\begin{enumerate}
\item if $|l_m|=1$, then $\lambda_{l_m}$ is a polynomial of degree $1$,
\item if $|l_m|>1$, then $\lambda_{l_m}$ can be expressed as a polynomial of degree $\leq|l_m|-1$ according \Cref{prop:Pblambdas}.
\end{enumerate}
Thus each $\prod_{m=1}^s \left(\lambda_{l_m} \right)^{k_m}$ can be expressed as a polynomial in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$ with total degree at most equal to:
$$
\sum_{|l_m| = 1}k_m +\sum_{|l_m|>1} k_m (|l_m|-1).
$$
Each partition of $j$ either does not or does contain any $l_m$ such that $|l_m|\neq 1$.
Accordingly, each $\prod_{m=1}^s \left(\lambda_{l_m} \right)^{k_m}$ falls into one of the two following categories, it can be expressed as a polynomial with total the degree:
\begin{enumerate}
\item either equal to $\sum_{m=1}^s k_m=\widetilde m$, when the partition contains only $l_m$ such that $|l_m|= 1$;
\item or smaller than $\sum_{m=1}^s k_m=\widetilde m\leq |j|$, when the partition contains at least one $l_m$ such that $|l_m|\neq 1$.
\end{enumerate}
In the first case, each $l_m$ belongs to $\{e_1,e_2,e_3\}$, and, since the partition must satisfy $\sum_{m=1}^s k_m l_m =j$, it corresponds to $s=3$ with $j=j_1e_1+j_2e_2+j_3e_3$. Hence $(\lambda_{e_1})^{j_1}(\lambda_{e_2})^{j_2}(\lambda_{e_3})^{j_3}$ is precisely of degree equal to $|j|$, with a weight of $\frac{j!}{j_1!j_2!j_3!} = 1$, whereas all other $\prod_{m=1}^s \left(\lambda_{l_m} \right)^{k_m}$ terms can be expressed as a polynomial in $\mathbb C[\lambda_{e_1},\lambda_{e_2},\lambda_{e_3}]$ with total degree smaller than $|j|$. This concludes the proof.
These polynomials have coefficients independent of $\dir$ as a consequence of \Cref{prop:Pblambdas}.
\end{proof}
Since \Cref{prop:AbME,prop:PbME} state similar relations between the derivatives of GPW functions on the one hand and the three initialization parameters on the other hand, we can now prove the common property of both families of GPWs.
\begin{prop}
Given $(n,p)\in\mathbb N^2$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any sets of phase-based and amplitude-based GPWs associated to partial differential operator $\mathcal L_{c}$, constructed via \Cref{Algo:PbS,Algo:AbS,Algo:SubS}, with $q=\max(n-1,1)$ and the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and any set of unit vectors $\{\dir_l\in\mathbb S^2,1\leq l\leq p\}$.
Then there exist square lower triangular matrices $\mathsf L_A}%{L\!\!A^n$ and $\mathsf L_P}%{L\!\!\!\!P^n$, whose diagonal coefficients are equal to $1$ and whose other non-zero coefficients depend only on (i) derivatives of the PDE coefficients $c$ evaluated at $\pG$ and (ii) the initialization parameter $\cst$, such that
\begin{equation*}
\mathsf A}%{M\hspace{-.75em}A^{n,p} = \mathsf L_A}%{L\!\!A^n \mathsf E}%{M\!\!\!E^{n,p} \text{ and } \mathsf P}%{M\!\!\!P^{n,p} = \mathsf L_P}%{L\!\!\!\!P^n \mathsf E}%{M\!\!\!E^{n,p}.
\end{equation*}
As a result, we have:
\begin{equation*}
rk\left(\mathsf A}%{M\hspace{-.75em}A^{n,p} \right)= rk\left( \mathsf E}%{M\!\!\!E^{n,p}\right) \text{ and } rk\left( \mathsf P}%{M\!\!\!P^{n,p}\right) = rk\left( \mathsf E}%{M\!\!\!E^{n,p}\right).
\end{equation*}
\end{prop}
\begin{proof}
This is a direct consequence of \Cref{prop:AbME,prop:PbME} and the choice of numbering for the matrix entries. Indeed:
\begin{itemize}
\item as a consequence of \Cref{prop:AbME,prop:PbME}, for all $j$ such that $|j|\leq q$ there exist coefficients denoted $l^A_{ij} $ and $l^P_{ij}$ for $i\in\mathbb N_0^3$ with $|i|<|j|$ satisfying for any vector $\dir$ and the corresponding amplitude- or phase-based GPW function, denoted respectively $J$ or $G$, for $|j|>0$:
$$
\partial_x^j J(\pG)=(\Lambda_{1})^{j_1}(\Lambda_{2})^{j_2}(\Lambda_{3})^{j_3}
+
\sum_{|i|< |j|}l^A_{ij} \frac{j!}{i!}(\Lambda_{1})^{i_3}(\Lambda_{2})^{i_2}(\Lambda_{3})^{i_3}
$$
$$
\partial_\varx^j G(\pG)= (\lambda_{e_1})^{j_1}(\lambda_{e_2})^{j_2}(\lambda_{e_3})^{j_3}
+
\sum_{|i|< |j|}l^P_{ij} \frac{j!}{i!}(\lambda_{e_1})^{i_3}(\lambda_{e_2})^{i_2}(\lambda_{e_3})^{i_3}
$$
while for $|j|=0$ both GPWs satisfy:
$$
J(\pG) = 1 \text{ and } G(\pG) = 1;
$$
\item we can then define two square matrices of size $(n+1)(n+2)(n+3)/6$, $\mathsf L_A}%{L\!\!A^n$ and $\mathsf L_P}%{L\!\!\!\!P^n$, by:
$$
\left(\mathsf L_A}%{L\!\!A^n\right)_{\mathcal N(j),\mathcal N(i)}=
\left\{\begin{array}{l}
1 \text{ if } i=j,\\
\frac{1}{i!}l^A_{ij} \text{ if } |i|<|j|,\\
0 \text{ otherwise},
\end{array}\right.
\text{ and similarly }
\left(\mathsf L_P}%{L\!\!\!\!P^n\right)_{\mathcal N(j),\mathcal N(i)}=
\left\{\begin{array}{l}
1 \text{ if } i=j,\\
\frac{1}{i!}l^P_{ij} \text{ if } |i|<|j| ,\\
0 \text{ otherwise},
\end{array}\right.
$$
both lower triangular matrices since $|i|<|j|$ implies that $\mathcal N(i)<\mathcal N(j)$,
\item
therefore, by definition of the matrices $\mathsf A}%{M\hspace{-.75em}A^{n,p}$, $\mathsf P}%{M\!\!\!P^{n,p}$ and $\mathsf R}%{M\!\!\!R^{n,p}$, for $j\in\left(\mathbb N_0\right)^3\times \mathbb N$ with $|j|\leq n$ and $l\leq p$, we have the following relations:
$$
\mathsf A}%{M\hspace{-.75em}A^{n,p}_{\mathcal N(j),l } = \sum_{k=1}^{\frac{(n+1)(n+2)(n+3)}6} \left(\mathsf L_A}%{L\!\!A^n\right)_{\mathcal N(j),k}\mathsf R}%{M\!\!\!R^{n,p}_{ k,l}
\text{ and }
\mathsf P}%{M\!\!\!P^{n,p}_{\mathcal N(j),l} = \sum_{k=1}^{\frac{(n+1)(n+2)(n+3)}6} \left(\mathsf L_P}%{L\!\!\!\!P^n\right)_{\mathcal N(j),k}\mathsf R}%{M\!\!\!R^{n,p}_{k,l }
$$
\end{itemize}
\end{proof}
Fixing $p=(n+1)^2$, we denote hereafter the corresponding $(n+1)(n+2)(n+3)/6 \times (n+1)^2$ matrices $\mathsf A}%{M\hspace{-.75em}A^{[n]}:=\mathsf A}%{M\hspace{-.75em}A^{n,(n+1)^2/2}$ and $\mathsf P}%{M\!\!\!P^{[n]}:=\mathsf P}%{M\!\!\!P^{n,(n+1)^2/2}$.
As a direct consequence of \Cref{thm:matmaxrk}, we then obtain the following result.
\begin{cor}
\label{cor:GPWmats}
Given $n\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Let $p=(n+1)^2$ directions on $\mathbb S^2$ be chosen as:
$$
\mathbf d_{l,m} = (\sin\theta_l\cos\varphi_{lm},\sin\theta_l\sin\varphi_{lm},\cos\theta_l)
$$
for all $l$ from $0$ to $n$ with $|m|\leq l$, where the $n+1$ colatitude angles $\{\theta_l\}_{0\leq l\leq n}\subset (0,\pi)$ are all different from each other, and the azimuths $\{\varphi_{lm}\}_{0\leq l\leq n,|m|\leq l}\subset [0,2\pi)$ satisfy $\varphi_{lm}\neq\varphi_{lm'}$ for every $m\neq m'$.
Consider any sets of phase-based and amplitude-based GPWs associated to partial differential operator $\mathcal L_{c}$, constructed via \Cref{Algo:PbS,Algo:AbS,Algo:SubS}, with $q=\max(n-1,1)$ the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and directions $\{\dir_{l,m},0\leq l\leq n, |m|\leq l\}$.
Then the corresponding matrices $\mathsf A}%{M\hspace{-.75em}A^{[n]} $ and $\mathsf P}%{M\!\!\!P^{[n]} $, of size ${(n+1)(n+2)(n+3)}/{6} \times (n+1)^2$ are of rank $(n+1)^2$.
\end{cor}
\subsubsection{The polynomial case}
To address the polynomial case, we consider again the corresponding Taylor expansion matrix, but in a different way that for the GPW case: thanks to the choice of initialization, only some of the rows of the matrix will be sufficient to study the desired rank properties.
\begin{prop}
\label{prop:Polnus}
Given $q\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any polynomial quasi-Trefftz function associated to partial differential operator $\mathcal L_{c}$, $H(\varx) :=\polR(\varx-\pG)$ with $\displaystyle R:=\sum_{i\in\mathbb N_0^3, |i|\leq q+1} \nu_i \mathbf X^i $, constructed via \Cref{Algo:PolS,Algo:SubS}, with the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and a unit vector $\dir$.
Then,
for all $j\in\left(\mathbb N_0\right)^3$ such that $|j|\leq q+1$ and $j_1\in\{0,1\}$ then $\partial_x^j H(\pG)/j!=(\nu_{e_1})^{j_1}(\nu_{e_2})^{j_2}(\nu_{e_3})^{j_3}$.
\end{prop}
\begin{proof}
Since $H(\varx)=\polR(\varx-\pG)$, we can express $\partial_\varx^j H$ directly as:
$$\forall x\in\mathbb R^3,
\partial_\varx^j H(\varx) =
\partial_\varx^j R(\varx-\pG),
$$
and therefore
$$
\partial_\varx^j H(\pG) =
j!\nu_j.
$$
The result is then a direct consequence of the initialization \eqref{eq:PolInit}.
\end{proof}
We can then turn to the rank of $\mathsf Q}%}{M\!\!\!Q^{n,p}$, which will naturally be related to the rank of the sub-matrix of $\mathsf E}%{M\!\!\!E^{n,p}$ consisting of its rows corresponding to $i$ with $i_1\in\{0,1\}$.
\begin{prop}
Given $(n,p)\in\mathbb N^2$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider any sets of polynomial quasi-Trefftz functions associated to partial differential operator $\mathcal L_{c}$, constructed via \Cref{Algo:PolS,Algo:SubS}, with $q=\max(n-1,1)$ and the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and any set of directions $\{\dir_l\in\mathbb S^2,1\leq l\leq p\}$. Then
$$
rk\left(\mathsf E}%{M\!\!\!E^{n,p}\right)\leq rk \left(\mathsf Q}%}{M\!\!\!Q^{n,p}\right).
$$
\end{prop}
\begin{proof}
If we denote by $ \overline{\mathsf R}%{M\!\!\!R}^{(n,p)}$ the sub-matrices of $\mathsf R}%{M\!\!\!R^{(n,p)}$ corresponding to removing the rows $i$ such that $i_1>1$, then, since $n\leq q+1$, it becomes clear as a consequence of \Cref{prop:Polnus} that there exists a matrix $\mathsf L_Q}%{L\!\!\!Q$ such that:
$$
\mathsf L_Q}%{L\!\!\!Q \mathsf Q}%}{M\!\!\!Q^{n,p} = \overline{\mathsf R}%{M\!\!\!R}^{(n,p)}, \text{ so } rk\left(\mathsf Q}%}{M\!\!\!Q^{n,p}\right)\geq rk\left( \overline{\mathsf R}%{M\!\!\!R}^{(n,p)}\right).
$$
Moreover from \Cref{cor:maxrk} we have $rk\left(\mathsf E}%{M\!\!\!E^{n,p}\right)=rk\left( \overline{\mathsf R}%{M\!\!\!R}^{(n,p)}\right)$.
This concludes the proof.
\end{proof}
Fixing $p=(n+1)^2$, we denote hereafter the corresponding matrix $\mathsf Q}%}{M\!\!\!Q^{[n]}:=\mathsf Q}%}{M\!\!\!Q^{n,(n+1)^2}$.
As a direct consequence of \Cref{thm:matmaxrk}, we then obtain the following result.
\begin{cor}
\label{cor:polmat}
Given $n\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Let $p=(n+1)^2$ directions on $\mathbb S^2$ be chosen as:
$$
\forall(l,m)\in(\mathbb N_0)^2, l\leq n, |m|\leq l,
\mathbf d_{l,m} = (\sin\theta_l\cos\varphi_{lm},\sin\theta_l\sin\varphi_{lm},\cos\theta_l),
$$
where the $n+1$ colatitude angles $\{\theta_l\}_{0\leq l\leq n}\subset (0,\pi)$ are all different from each other, and the azimuths $\{\varphi_{lm}\}_{0\leq l\leq n,|m|\leq l}\subset [0,2\pi)$ satisfy $\varphi_{lm}\neq\varphi_{lm'}$ for every $m\neq m'$.
Consider any set of polynomial quasi-Trefftz functions associated to partial differential operator $\mathcal L_{c}$, constructed via \Cref{Algo:PolS,Algo:SubS}, with $q=\max(n-1,1)$ and the initialization introduced in \Cref{sec:norm} with $\cst\in\mathbb C^*$ and directions $\{\dir_{l,m},0\leq l\leq n, |m|\leq l\}$.
Then the corresponding $\frac{(n+1)(n+2)(n+3)}{6} \times (n+1)^2$ matrix $\mathsf Q}%}{M\!\!\!Q^{[n]} $ is of rank $(n+1)^2$.
\end{cor}
\subsection{Quasi-Trefftz bases Approximation properties}
We finally want to show that the three spaces spanned by the quasi-Trefftz function sets introduced in \Cref{sec:QTfams}
satisfy the desired approximation property \eqref{IntPb},
via matching of $u_a$'s Taylor expansion to that of $u$ as announced in \eqref{eq:TEapp}.
\begin{theorem}\label{thm:approx}
Given $n\in\mathbb N$ and a point $\pG\in\mathbb R^3$, a set of complex-valued functions $c = \{\funcci,i\in\mathbb N_0^3, |i|\leq 2\}$ is assumed to satisfy \Cref{hyp:PDop}.
Consider the three quasi-Trefftz spaces associated to partial differential operator $\mathcal L_{c}$, defined as the spaces spanned by each of the three following sets:
\begin{itemize}
\item the set of amplitude-based GPWs, constructed via \Cref{Algo:AbS,Algo:SubS},
\item the set of phase-based GPWs, constructed via \Cref{Algo:PbS,Algo:SubS},
\item the set of polynomial functions, constructed via \Cref{Algo:PolS,Algo:SubS},
\end{itemize}
each of them constructed with $q=\max(n-1,1)$ and the initialization introduced in \Cref{sec:norm} for $\cst\in\mathbb C^*$ and $p=(n+1)^2$ directions
on $\mathbb S^2$.
If the set of directions is chosen as:
$$
\forall(l,m)\in(\mathbb N_0)^2, l\leq n, |m|\leq l,
\mathbf d_{l,m} = (\sin\theta_l\cos\varphi_{lm},\sin\theta_l\sin\varphi_{lm},\cos\theta_l),
$$
where the $n+1$ colatitude angles $\{\theta_l\}_{0\leq l\leq n}\subset (0,\pi)$ are all different from each other, and the azimuths $\{\varphi_{lm}\}_{0\leq l\leq n,|m|\leq l}\subset [0,2\pi)$ satisfy $\varphi_{lm}\neq\varphi_{lm'}$ for every $m\neq m'$,
then any of these three spaces, denoted $\mathbb V^G_h$, satisfies the following approximation property:
\begin{equation}\label{eq:approx}
\begin{array}{l}
\forall u\in \mathcal C^{\max(2,n)}(\Omega)
\text{ satisfying }\mathcal L_{c} u=0,
\exists u_a\in\mathbb V_h^{G},\\
\exists C\in\mathbb R \text{ s. t. }
\forall \varx\in\Omega,
\left\{ \begin{array}{l}
|u(\varx)-u_a(\varx)|\leq C \| \varx-\pG \|^{n+1},\\
\|\nabla u(\varx)-\nabla u_a(\varx)\|\leq C \| \varx-\pG \|^{n}.
\end{array}\right.
\end{array}
\end{equation}
The constant $C$ here depends on the domain $\Omega$, on the desired order $n$, on the PDE solution $u$ to be approximated, as well as on the PDE coefficients $c$.
It is to be noted that this
actually shows a convergence in the $H^1$ norm:
\begin{equation}
\forall
\text{ satisfying }\mathcal Lu=0,
\exists u_a\in\mathbb V_h,\exists C\in\mathbb R, \text{ s. t. }
\|u-u_a\|_{H^1(B(\pG,h))}\leq C h^{n}.
\end{equation}
\end{theorem}
\begin{proof}
It is sufficient to prove that, if $\mathsf M^{[n]}$ is any of the three $(n+1)(n+2)(n+3)/6 \times (n+1)^2$ quasi-Trefftz matrices, namely $\mathsf A}%{M\hspace{-.75em}A^{[n]}$, $\mathsf P}%{M\!\!\!P^{[b]}$ or $\mathsf Q}%}{M\!\!\!Q^{[n]}$, then the linear system defined by:
$$
\mathsf M^{[n]} \mathsf x = \mathsf F_n
$$
has a solution for any $\mathsf F_n$ in the vector space $\mathbb F_{n,\Lc}$ defined as:
$$
\displaystyle
\mathbb F_{n,\Lc} :=\left\{\mathsf F\in\mathbb C^{(n+1)(n+2)(n+3)/6},\exists v\in\mathcal C^{\max(2,n)}(\Omega)\text{ s.t. } \mathcal L_{c} v=0 \text{ and }\forall i\in\left(\mathbb N_0\right)^3, |i|\leq n,
\ \mathsf F_{\mathcal N(i)} = \partial_x^i v (\pG) /i!
\right\}.
$$
We first define a similar vector space by relaxing the Trefftz condition $\mathcal L_{c} v=0$ into a quasi-Trefftz condition:
$$
\begin{array}{l}
\displaystyle
\mathbb K_{n,\Lc} :=
\left\{
\mathsf K\in\mathbb C^{(n+1)(n+2)(n+3)/6},\exists v\in\mathcal C^{\max(2,n)}(\Omega)\text{ satisfying }\mathcal L_{c} v(\varx)=O\left(\|\varx-\pG\|^{n-1}\right)
\right.
\\\displaystyle
\phantom{
\mathfrak K :=
\mathfrak K :=
}
\left.
\text{ s.t. }\forall i\in\left(\mathbb N_0\right)^3, |i|\leq n, \ \mathsf K_{\mathcal N(i)} = \partial_x^i v (\pG) /i!
\right\}.
\end{array}
$$
It is then clear that $\mathbb F_{n,\Lc}\subset\mathbb K_{n,\Lc}$, but also that the range of $\mathsf M^{[n]}$ is also included in $\mathbb K_{n,\Lc}$.
Next we want to show that the dimension of $\mathbb K_{n,\Lc}$ is $(n+1)^2$. For any function $f\in\mathcal C^{\max(2,n)}(\Omega)$ satisfying $\mathcal L_{c} f(\varx)=O(\|\varx-\pG\|^{n-1})$, by Leibniz rule we have:
$$
\forall i\in\left(\mathbb N_0\right)^3, |i|\leq n-2,
\sum_{|j|\leq 2}\sum_{\tilde i\leq i} \begin{pmatrix} i\\\tilde i \end{pmatrix}\partial_x^{i-\tilde i} c_j\partial_x^{j+\tilde i} f (\pG) = 0.
$$
This shows that $\mathbb K_{n,\Lc}$ is a subset of the kernel of an $(n-1)n(n+1)/6\times (n+1)(n+2)(n+3)/6$ matrix $\mathsf R$, with the following properties:
$$
\forall i\in\left(\mathbb N_0\right)^3, |i|\leq n-2,
\left\{
\begin{array}{l}
\mathsf R_{\mathcal N(i)\mathcal N\left(i+2e_1\right)} = c_{2e_1}(\pG
,\\
\mathsf R_{\mathcal N(i)\mathcal N(j)}=0 \text{ if } j>i+2e_1.
\end{array}
\right.
$$
Therefore, since $ c_{2e_1}(\pG) \neq 0$ by \Cref{hyp:PDop}, choosing a numbering scheme $\mathcal N$ such that if $|i|=|j|$ then $i\prec j$ implies $\mathcal N(i)\leq \mathcal N(j)$ highlights the echelon structure of $\mathsf R$.
As a result, the echelon structure of $\mathsf R$ guarantees that it has maximal rank, namely $(n-1)n(n+1)/6$, while its kernel is of dimension
$$\frac{(n+1)(n+2)(n+3)}{6} -\frac{(n-1)n(n+1)}{6} = (n+1)^2. $$
Hence $\mathbb K_{n,\Lc}$ is a subset of a space of dimension $(n+1)^2$, but it also has a subset of the same dimension, namely the range of $\mathsf M^{[n]}$, so it is itself of dimension $(n+1)^2$.
This shows that the range of $\mathsf M^{[n]}$ and $\mathbb K_{n,\Lc}$ are the same space, and therefore any $\mathsf F\in\mathbb F_{n,\Lc}$ belongs to the range of $\mathsf M^{[n]}$. So this conclude the proof of the approximation of the function values.
Precisely because this proof relies on matching the Taylor expansions of $u$ and $u_a$, the result of approximation of the gradient can be obtained directly by taking derivatives of this Taylor expansion matching identity.
\end{proof}
\section{Conclusion}
Given a partial differential operator $\mathcal L$ and a parameter $q\in\mathbb N$, a quasi-Trefftz function $f$ is a function satisfying the following local property in the neighborhood of a given point $\pG$:
\begin{equation}
\label{eq:finalQTprop}
\forall \varx \text{ in a neighborhood of }\pG,\ \mathcal L f(\varx) = O\left(|\varx-\pG|^q\right).
\end{equation}
The work presented here may be summarized as follows.
\begin{itemize}
\item We introduced three families of quasi-Trefftz functions for a class of 3D PDEs including the convected Helmholtz equation.
Two of these, the GPWs, are generalizations of a PW ansatz, and were first introduced for a class of 2D problems.
The ansatz defining these two families are defined in a neighborhood of a given point $\pG$ under the following forms: for all $\varx$,
\begin{equation*}
\left\{
\begin{array}{l}
\displaystyle
J(\varx) := \polQ(\varx-\pG)\exp \Lambda \cdot\Big(\varx-\pG \Big)
\text{ for some polynomial } Q \text{ and some vector }\Lambda\in\mathbb C^3,\\
\displaystyle
G(\varx) := \exp \polP(\varx-\pG)
\text{ for some polynomial } P.
\end{array}
\right.
\end{equation*}
The third one is fully polynomial, and this is the first introduction and study of polynomial quasi-Trefftz functions. The corresponding ansatz is defined in a neighborhood of a given point $\pG$ under the following forms: for all $\varx$,
\begin{equation*}
H(\varx) := \polR(\varx-\pG)
\text{ for some polynomial } R.
\end{equation*}
\item We provide explicit algorithms to construct quasi-Trefftz functions belonging to each of these three families, by constructing the corresponding polynomial $P$, $Q$ or $R$. This is achieved by a careful study of the system obtained by setting to zero the degree-$(q-1)$ Taylor polynomial of the image of each ansatz through the operator $\mathcal L$.
It is then possible to choose adequately the degree of the polynomial $P$, $Q$ or $R$ in order to split this system into a hierarchy of linear triangular sub-system. Hence a solution to the system can be computed via an explicit formula by simply solving successively the subsystems by substitution.
Interestingly, some of the polynomial coefficients of $P$, $Q$ and $R$ are free in the resulting algorithms, and thanks to these sets of linearly independent quasi-Trefftz functions can be constructed.
As a consequence, beyond the construction of individual quasi-Trefftz functions, we can construct spaces of quasi-Trefftz functions.
\item
We prove that it is possible to construct quasi-Trefftz spaces $\mathbb V_h$ spanned by sets of such quasi-Trefftz functions enjoying high order approximation property for exact solutions of the PDE.
More precisely, given a given point $\pG$, \Cref{thm:approx} states that in order to achieve a given order of accuracy $n+1$ of local approximation property in the following sense:
\begin{equation}
\label{eq:cclapp}
\forall
\text{ satisfying the governing PDE,
\exists u_a\in\mathbb V_h \text{ s. t. }
\|u-u_a\|_{L^\infty(B(\pG,h))}\leq C h^{n+1},
\end{equation}
where $B(\pG,h)$ denotes the sphere centered at $\pG$ of radius $h$ in $\mathbb R^3$, it is sufficient to construct a quasi-Trefftz space $\mathbb V_h$ of dimension $p=(n+1)^2$ with basis functions satisfying the quasi-Trefftz property \eqref{eq:finalQTprop} with $q=\max(n-1,1)$.
For reference, reaching the same order of approximation in \eqref{eq:cclapp} using a standard polynomial space would yield a dimension $p=(n+1)(n+2)(n+3)/6$ (corresponding to the full space of polynomials of degrees at most equal to $n$).
For instance for $n=8$ the quasi-Trefftz space is of dimension $81$ while the polynomial space is of dimension $165$.
Besides, as noted in \Cref{thm:approx},
we actually show a convergence in $H^1$:
\begin{equation}
\forall
\text{ satisfying the governing PDE,
\exists u_a\in\mathbb V_h \text{ s. t. }
\|u-u_a\|_{H^1(B(\pG,h))}\leq C h^{n}.
\end{equation}
\end{itemize}
Future plans include :
\begin{itemize}
\item first testing the approximation properties of the three families of quasi-Trefftz functions, in particular from the perspective of initialization of the construction process, conditioning of the resulting basis and stability in the high-frequency regime,
\item
then
compare their performance with polynomial and wave-based bases - standardly used in the literature in aero-acoustics - in terms of accuracy, computing time and stability,
\item finally comparing performance of a parallel implementation of quasi-Trefftz methods on realistic industrial test cases.
\end{itemize}
|
\section{Introduction}
Complex networks, such as communication, biological, transportation, and contact networks, are widely analyzed. The daily production of data results in tremendously large real-world networks. Consequently, the analysis of such networks containing millions of nodes and billions of edges is more and more challenging, if not impossible, due to memory and time constraints. Thus, suitable extraction of the pertinent nodes and edges that preserve the essential information while reducing the size of the network is fundamental. Network backbones offer a way to do so. Two main research paths tackle this problem: coarse-grain backbones or filter-based backbones. In the former, one clusters together nodes sharing similarities to reduce the network size\cite{zeng2011coarse,zeng2019new}. In the latter, one removes nodes or edges from the network based on a given property \cite{coscia2017backboning, 10.1093/comnet/cnab021}.
The community structure is one of the significant properties in real-world networks. Indeed, it heavily determines their dynamics and their underlying functionalities \cite{boccaletti2006complex}. It is generally illustrated by dense regions of connected nodes that barely connect from one region to another. Communities can be non-overlapping or overlapping \cite{jebabli2018community}, hierarchical \cite{peel2020detectability}, and attributed \cite{atzmueller2021mining}. Community detection is one of the most prolific research areas in network science. It relies on numerous measures quantifying the quality of the community structure. Modularity is among the most popular\cite{newman2006modularity}. It compares the density of connections of the uncovered community structure with a similar random network. The higher the modularity, the higher the confidence in the tight community structure of the network.
Recent works have shown that one can exploit the community structure efficiently to extract backbones \cite{ghalmane2020extracting, ghalmane2021extracting}. Inspired by these works, we propose a filtering technique based on the preservation of the community structure of the network. It exploits the community structure using the concept of vitality. Vitality quantifies the contribution of a node to a given quality measure by removing this node and computing the variation of the quality measure. To assess its importance, we compute modularity as a quality measure, with and without the node in question. Then, one ranks the nodes from the lowest contribution on modularity to the highest. Subsequently, nodes with the lowest contribution are removed until one reaches the desired size of the network.
Comparative experimental evaluations are conducted on real-world weighted networks of different sizes and domains. The developed backbone extraction technique called ``modularity vitality backbone" is compared with the recently introduced community-based method ``overlapping nodes ego backbone" \cite{ghalmane2020extracting}. Results show that it is more effective in preserving the core information of the network and the community structure.
\newline
The main contributions of the paper summarize as follows:
\begin{itemize}
\item We propose a backbone filtering technique exploiting the community structure of networks.
\item Experiments with weighted networks show that it outperforms another alternative measure.
\item It can be easily adapted to any type of network (i.e., undirected, unweighted, and directed networks).
\end{itemize}
The remaining of the article is organized as follows. Section \ref{sec:relatedworks} discusses briefly the related works. Section \ref{sec:VCB} introduces the modularity vitality backbone. Sections \ref{sec:data} and \ref{sec:evaluationmeasures} present respectively the datasets and the evaluation measures used in this study. Section \ref{sec:experimentalresults} reports the results of the comparative evaluation. Section \ref{sec:discussion} discusses the results. Finally, section \ref{sec:conc} concludes the paper.
\section{Related Works}
\label{sec:relatedworks}
Backbones offer an ideal solution to the trade-off between preserving essential information in the network and reducing the network size. Backbone extraction studies concerns mainly two types of networks: mono-mode networks \cite{van2012high, cao2019motif, 10.1093/comnet/cnab021} and bipartite networks \cite{zweig2011systematic, neal2013identifying, neal2014backbone}. In this work, we consider mono-mode networks. Within this class of networks, there are two leading approaches for extracting backbones. The first is coarse-graining, and the second is filtering.
In coarse-graining methods, one group nodes with similar characteristics into a single node. For instance, authors in \cite{gfeller2007spectral} merge the nodes based on random walks. In the work of \cite{chen2016effectively}, authors use the $k$-nearest neighbors algorithm ($k$-NN) to group similar nodes based on the nearest higher-density-neighbor.
In filter-based methods, the goal is to remove redundant information by pruning nodes or edges in the network. Redundancy is assessed based on a statistical property.
Most of the works reported in the literature concern edge-filtering techniques. Serrano et al. propose the disparity filter. It uses a null model of the edge weights to preserve statistically significant edges \cite{serrano2009extracting}. Authors in \cite{goh2006skeleton} compute the betweenness centrality of edges and remove the ones that don't exceed a specific threshold. Authors in \cite{zhang2018extracting} use a combination of local and global information to extract the backbone. More precisely, they use the link weights to build the $h$-strength graph and the betweenness centrality to build the $h$-bridge graph. Then one obtains the backbone by merging $h$-strength and $h$-bridge. Simas et al. \cite{10.1093/comnet/cnab021} present the distance backbone based on the triangular organization of edges which preserves all shortest paths.
Node filtering techniques are less frequent. They rely on topological features to associate a score to the nodes. Nodes with the higher scores are then extracted \cite{DAI2018271, RePEc}.
In this line, in their recent work, Ghalmane et al. prune nodes based on the community structure characteristics of the network \cite{ghalmane2020extracting}. They propose two node-filtering techniques. The first one preserves the overlapping nodes and the hubs of the network. In contrast, the second conserves the overlapping nodes and their one-step neighbors to form the backbone. These two algorithms exhibit superior performances as compared to the popular disparity filter. These results illustrate the community structure's importance in preserving the core information in a network while reducing its size.
Inspired by these findings, we propose the ``modularity vitality backbone" algorithm. This node filtering technique also exploits the community structure of the network. It uses a measure of the node contribution to the modularity. Roughly speaking, nodes with the lowest contribution to the quality measure of the community structure are filtered. The remaining nodes form the backbone.
\section{Modularity Vitality Backbone}
\label{sec:VCB}
This section presents the vitality concept. Then, we briefly discuss various mesoscopic quality measures. We explain why we choose Newman's modularity as a quality measure. Finally, the algorithm of the proposed backbone extractor is given.
\subsubsection{Vitality Index.}
Let $G(V,E)$ be a simple and undirected graph where $V=\{v_1, v_2, ..., v_N\}$ is the set of nodes totaling $N=|V|$ and $E=\{(v_i, v_j) | v_i, v_j \in V\}$ is the set of edges. Denote $f(G)$ and $f(G\setminus \{u\})$ as two real-valued functions defined on the complete graph $G$ and on graph $G\setminus\{u\}$ without node or edge $u$. Then, the vitality index is the difference between both functions, defined as $\nu(G, u) = f(G) - f(G\setminus \{u\})$. The resulting value is a signed value, indicating the positive or negative contribution of the node or edge $u$ on graph $G$ \cite{koschutzki2005centrality}.
\subsubsection{Mesoscopic Quality Measures.}
There are numerous quality measures to characterize communities \cite{yang2015defining, leskovec2010empirical, rajeh2021characterizing}. Their goal is to answer how good is the community structure in a network. They use topological properties defined at the mesoscopic level. Let set $C=\{c_1, c_2,...,c_l,..., c_{n_c}\}$ represent $n_c = |C|$ communities of a graph $G$ and $f(c_l)$ represents a quality function of community $c_l$. One can categorize the quality functions into three main groups: \newline
\textbf{1. Based on internal connectivity:} such as internal density characterizing how densely connected the nodes are in a community compared to other communities. \newline
\textbf{2. Based on internal and external connectivity:} such as Flake-ODF measuring the fraction of nodes in a community with fewer internal edges than external ones. \newline
\textbf{3. Based on a network model:} such as Newman's modularity \cite{newman2006modularity} which assesses the difference between the real connections in the community $c_l$ and the random connections in the same community.
All of these quality functions characterize a single community. Hence, to quantify the quality of the overall community structure, one averages $f(c_l)$ over all the communities.
Newman's modularity is one of the most popular mesoscopic quality measures. Indeed, it is widely used in community detection algorithms as an optimization criterion \cite{clauset2004finding, brandes2007modularity, blondel2008fast}. This is the main reason why it is one of the well-accepted benchmarks for characterizing the community structure of the networks. Numerous extensions have also been proposed for modularity to account for networks with overlapping and hierarchical community structure \cite{chen2015fuzzy}. In this work, we use Newman's modularity as a quality measure to assess the vitality of nodes due to the following reasons: \newline
1. Modularity can naturally be extended to unweighted, undirected, and directed networks. \newline
2. Modularity vitality ensures that nodes that are the main contributors to the community structure are retained, regardless of their type (i.e., hub- or bridge-like).\newline
3. Previous works on modularity vitality centrality has proved to assign high scores to the most influential nodes \cite{magelinski2021measuring}.
\subsubsection{Algorithm.}
The ``modularity vitality backbone" is based on the vitality concept, where one can measure the contribution of a node or an edge using any quality measure computed on graphs. We use Newman's modularity as a quality measure. Nonetheless, one can opt for other quality measures to quantify the node and edge influence.
Newman's modularity enables us to differentiate between highly internally connected nodes (hubs) and nodes at the borders of the communities (bridges). Indeed, hubs increase the internal density of the communities. Therefore, they contribute positively to modularity. In contrast, bridges increase the connections between the communities. Consequently, they tend to decrease the modularity. As we choose to give equal importance to both types of nodes, we rank the nodes according to the absolute value of their modularity vitality score. It allows keeping nodes with the highest contribution to modularity, regardless of their role (i.e., hub nodes or bridge nodes). Then, one removes the nodes that barely contribute to modularity. The backbone extraction procedure is given in Algorithm 1.
\begin{algorithm}[!ht]
\DontPrintSemicolon
\KwInput{Graph $G(V,E)$, Community set: $C=\{c_1,c_2, ..., c_{n_c}\}$, Size $s$}
\KwOutput{Pruned graph $\hat{G}(\hat{V},\hat{E})$}
$Q(G) \leftarrow Modularity (G, C)$ \tcp*{Computing modularity vitality of nodes}
$D \leftarrow \varnothing$ \;
\For{$v \in V$}{
$Q(G \setminus \{v\}) \leftarrow \sum_{c \in C} \left[
\frac{|E_{c}^{in}|-|E_{v}^{in}|}{|E|-|E_{v}|} - \left(
\frac{2\left(|E_{c}^{in}| - |E_{v, c}^{in}|\right) + \left(|E_{c}^{out}| - |E_{v, c}^{out}| \right)}{2(|E|-|E_{v}|)} \right)^2\right]$\;
$\alpha(v) \leftarrow Q(G) - Q(G \setminus \{v\})$\;
$D[v] = |\alpha(v)|$
}
$D \leftarrow sort(D)$ \;
\While{$|V|> s$}
{
$\eta \leftarrow D.pop(v)$ \tcp*{Extracting the backbone}
$G \leftarrow G \setminus \eta$ \;
$V \leftarrow V \setminus \eta$
\If{$G$ is disconnected}
{
$G \leftarrow LCC(G)$
}
}
\caption{Modularity Vitality Backbone Extraction}
\end{algorithm}
Note that the vitality computation is not naively computed two times for each node. In such a case, the complexity can rapidly become prohibitive. Indeed, one computes instead the modularity variation reducing the computation's complexity to $O(|E|+Nn_c)$. It makes the vitality measure suitable for large-scale weighted networks. We also note that the symbol $|E|$ is extended to weighted networks.
\section{Datasets}
\label{sec:data}
We use a set of seven real-world networks originating from various domains (social, collaborative, ecological, and technological) in the experiments. The nodes and edges range from hundreds to thousands. We choose to integrate small networks in the experiments in order to get a better understanding on the filtering process. Table \ref{TableCharacteristcs} presents their basic topological characteristics. All the networks are freely available online\footnote{Aaron Clauset, Ellen Tucker, and Matthias Sainz, ``The Colorado Index of Complex Networks." https://icon.colorado.edu/ (2016).}\footnote{Tiago P. Peixoto, ``The Netzschleuder network catalogue and repository," https://networks.skewed.de/ (2020).}. As there is no ground truth available, we rely on the Louvain community detection algorithm to uncover their community structure \cite{blondel2008fast}.
\newline
\textbf{1. Zachary's Karate Club:} Nodes are members of a karate club and are connected if they are friends inside and outside the club. Edges are weighted by the relative interactions occurring between the members.\newline
\textbf{2. Wind Surfers:} Nodes are windsurfers in southern California in the fall of 1986. They are connected if they're friends. Edges are weighted based on the social closeness of the surfers to one another. \newline
\textbf{3. Madrid Train Bombing:} Nodes are terrorists in the train bombing of March 11, 2004, in Madrid. Edges represent contacts between the terrorists and are weighted based on the strength of their underlying relationship.\newline
\textbf{4. Les Mis\'erables:} Nodes are the characters in the novel ``Les Mis\'erables." Edges represent characters' co-appearances in the same chapter. They are weighted by the number of co-appearances.\newline
\textbf{5. Wiki Science:} Nodes are either applied, formal, natural, or social sciences Wikipedia pages. They're weighted by the cosine similarity between them. \newline
\textbf{6. Unicode Languages:} A bipartite network representing languages and countries. Weights represent the fraction of people in a given country having the literacy (reading and writing) of a specific language. \newline
\textbf{7. Scientific Collaboration:} Nodes are authors of articles in the ``Condensed Matter" category of arXiv. Edges represent co-authorship and are weighted by the number of joint papers among the authors.
\begin{table}[h]
\centering
\caption{Basic topological properties of the real-world networks under study. \textit{N} is the number of nodes. $|E|$ is the number of edges. $<k>$ is the average weighted degree. $\omega$ is the density. $\zeta$ is the transitivity. $k_{nn}(k)$ is the assortativity. $\epsilon$ is the efficiency. $Q$ is the weighted modularity of the network.}
\label{TableCharacteristcs}
\begin{tabular}{lcccccccccc}
\hline
Network & $N$ & $|E|$ & $<k>$ & $\omega $ & $\zeta$ & $k_{nn}(k)$ & $\epsilon$ & $Q$\\
\hline
Zachary's Karate Club & 33 & 77 & 13.59 & 0.139 & 0.256 & -0.476 & 0.492 & 0.444\\
Wind Surfers & 43 & 336 & 56.09 & 0.372 & 0.564 & -0.147 & 0.679 & 0.371\\
Madrid Train Bombing & 62 & 243 & 8.81 & 0.121 & 0.561 & 0.029 & 0.448 & 0.435 \\
Les Mis\'erables & 77 & 254 & 21.30 & 0.087 & 0.499 & -0.165 & 0.435 & 0.565 \\
Wiki Science & 687 & 6,523 & 7.35 & 0.028 & 0.469 & 0.244 & 0.323 & 0.631 \\
Unicode Languages & 868 & 1,255 & 0.697 & 0.003 & 0.00 & -0.171 & 0.255 & 0.772 \\
Scientific Collaboration & 16,726 & 47,594 & 9.23 & 0.0003 & 0.360 & 0.185 & 0.117 & 0.873\\
\hline
\end{tabular}
\end{table}
\section{Evaluation Measures}
\label{sec:evaluationmeasures}
We compare the effectiveness of the proposed backbone extraction technique and the ``overlapping hub ego backbone" based on four different evaluation measures classically used. \newline
\textbf{1. Average weighted degree:} The weighted degree of a node is the sum of the weights of all the edges connected to it. Hence, a higher average weighted degree backbone means that important nodes are kept in the graph, reflecting its connectedness. It is defined as follows:
\begin{equation}
<k> = \frac{1}{N}\sum_{i=1}^N k_i = \sum_{j \in \mathcal{N}(1)} w_{ij}
\end{equation}
where $\mathcal{N}(1)$ is the first-order neighborhood of node $i$. \newline
\textbf{2. Average link weight:} Links in the backbone preserve the information flow of the network. In other words, the higher the value of the links, the better the backbone in maintaining the core information of the graph. It is defined as follows:
\begin{equation}
<w> = \frac{1}{N}\sum_{i,j \in V} w_{ij}
\end{equation}
\newline
\textbf{3. Average betweenness:} Nodes with higher betweenness can disseminate information quickly. Hence, a backbone with higher average betweenness indicates that the speed of information dissemination is barely altered. It is defined as follows:
\begin{equation}
<b> = \frac{1}{N}\sum_{i=1}^N b_i = \sum_{i \neq s \neq t} \frac{\sigma_{s,t}^i}{\sigma_{s,t}}
\end{equation}
where $\sigma_{s,t}$ denotes the number of shortest paths between nodes $s$ and $t$ and $\sigma_{s,t}^i$ denotes the number of shortest paths between nodes $s$ and $t$ passing through node $i$. \newline
\textbf{4. Weighted Modularity:} Modularity assesses the quality of the community structure based on the difference between the actual and the expected fraction of edges in the communities. A backbone with higher modularity suggests that the community structure is less altered. It can be computed on unweighted and weighted networks \cite{newman2004analysis}. It is defined as follows:
\begin{equation}
Q = \frac{1}{2|E|} \sum_{i,j}\left[ A_{ij} - \frac{w_iw_j}{\sum_{i,j}w_{ij}}\right] \delta(c_i,c_j)
\end{equation}
where $A_{ij}$ is the weighted adjacency matrix of graph $G$ and $\delta(c_i,c_j)$ equals 1 if nodes $i$ and $j$ belong to the same community, otherwise it equals 0.
\section{Experimental Results}
\label{sec:experimentalresults}
The effectiveness of ``modularity vitality backbone" is compared with another community-aware backbone extraction technique recently introduced \cite{ghalmane2020extracting}. The authors propose two backbones in their work, namely ``overlapping nodes ego backbone" and ``overlapping nodes and hubs backbone." In their comparative evaluation, they show that both techniques perform favorably compared to the popular disparity filter.
Therefore, in this work, we restrict our comparison to the most effective: overlapping nodes ego backbone. Table \ref{TableResults} reports the experimental results for seven real-world networks under study. The backbones quality measures are the average weighted degree ($<k>$), the average link weight ($<w>$), the average betweenness ($<b>$), and the weighted modularity ($Q$). As in their paper, we fix the backbone size to 30\% of the original network.
Let's first discuss the average weighted degree. The higher its value, the better the backbone is in keeping the salient nodes maintaining its connectedness. Table \ref{TableResults} reports that the modularity vitality backbone outperforms overlapping nodes ego backbone in all of the networks under study. Moreover, the difference ranges from very small magnitudes to orders of magnitude higher. For example, in Wiki Science, the average weighted degree in the modularity vitality backbone is eleven times higher than the ``overlapping nodes ego backbone." On the contrary, in Scientific Collaboration, the difference is barely noticeable (0.02).
Let's turn to average link weight. The average link weight characterizes the relevance of the links kept in a backbone. Hence, the higher its value, the better the backbone is in preserving essential links. The results show that the modularity vitality backbone outperforms the overlapping nodes ego backbone in six out of the seven networks.
Now, we discuss the average betweenness. This measure indicates the amount of information flow that can pass through the nodes of a given backbone. The higher its value, the higher the efficiency of the backbone in information spreading. The modularity vitality backbone outperforms the overlapping nodes ego backbone in only one out of the seven networks. In Scientific Collaboration networks, their values are comparable. It indicates that the information spread within the modularity vitality backbone is not as efficient as overlapping nodes ego backbone. Note, however, that the differences between the two backbones are less pronounced.
Finally, we turn to weighted modularity. The higher the modularity of the backbone, the better the quality of its community structure. As reported in table \ref{TableResults}, the modularity vitality backbone outperforms the overlapping nodes ego backbone on all the networks under study. These results are not surprising. Indeed, the modularity vitality backbone prunes the nodes contributing less to the modularity of the network. Hence, it tends to preserve the modularity as pruning proceeds.
To summarize, the modularity vitality backbone exhibits a higher weighted modularity than the overlapping nodes ego backbone. It preserves essential nodes in the network (i.e., hubs and bridges), thus achieving a higher average weighted degree and average link weight. However, maintaining the community structure comes at a price of a lower average betweenness.
\begin{table}[t]
\centering
\caption{The computed values for the average weighted degree ($<k>$), link weight ($<w>$), and betweenness ($<b>$) alongside the weighted modularity ($Q$) of the backbones with 30\% of the initial size of the network. For brevity, MV stands for ``modularity vitality backbone," and OE stands for ``overlapping nodes ego backbone."}
\begin{tabular}{l|ll|ll|ll|ll}
\hline
\multirow{2}{*}{Network} & \multicolumn{2}{l|}{$<k>$} & \multicolumn{2}{l|}{$<w>$} & \multicolumn{2}{l|}{$<b>$} & \multicolumn{2}{l}{$Q$} \\
\cline{2-9}
& MV & OE & MV & OE & MV & OE & MV & OE \\
\hline
Zachary's Karate Club & \textbf{13.00} & 8.40 & \textbf{6.05} & 4.07 & 0.12 & \textbf{0.27} & \textbf{0.35} & 0.32 \\
Wind Surfers & \textbf{71.38} & 35.08 & \textbf{35.69} & 17.54 & 0.11 & \textbf{0.15} & \textbf{0.36} & 0.32 \\
Madrid Train Bombing & \textbf{8.53} & 3.90 & \textbf{4.26} & 1.95 & 0.09 & \textbf{0.14} & \textbf{0.38} & 0.17 \\
Les Mis\'erables & \textbf{39.48} & 19.08 & \textbf{19.74} & 9.54 & 0.08 & \textbf{0.14} & \textbf{0.49} & 0.48 \\
Wiki Science & \textbf{10.16} & 0.92 & \textbf{5.08} & 0.46 & 0.01 & \textbf{0.08} & \textbf{0.73} & 0.72 \\
Unicode Languages & \textbf{1.46} & 1.28 & \textbf{0.73} & 0.64 & \textbf{0.03} & 0.02 & \textbf{0.79} & 0.78 \\
Scientific Collaboration & \textbf{17.22} & 17.20 & 4.71 & \textbf{8.60} & 0.001 & 0.001 & \textbf{0.81} & 0.71 \\
\hline
\end{tabular}
\label{TableResults}
\end{table}
\section{Discussion}
\label{sec:discussion}
The constant increase of real-world networks size has prompted researchers to design a smaller yet accurate representation of networks. This problem is tackled either with coarse-graining or filter-based methods. A recent work by Ghalmane et al. \cite{ghalmane2020extracting} has shown interest in exploiting the modular structure of the network to deal with this issue. Building on this finding, we propose a new backbone extractor, ``modularity vitality backbone," that aims to preserve the quality of the community structure. Assigning a modularity vitality score to the nodes, it prunes those with a low contribution to the network's modularity.
We performed a comparative analysis with the recently introduced ``overlapping nodes ego backbone. These investigations on a set of seven real-world networks from various domains are globally at the advantage of the proposed technique. After pruning 70\% of the network, results show that the modularity vitality backbone maintains higher modularity than overlapping hubs ego backbone. This expected behavior comes with higher performance in terms of average node degree and average link weight. Nonetheless, information efficiency isn't guaranteed. Indeed, one can point out that in five out of seven networks, the modularity vitality backbone suffers from lower information efficiency.
It can be explained by how the modularity vitality backbone proceeds. Indeed, one removes nodes that barely affect modularity. Those nodes may have high betweenness, yet they do contribute much to the network's modular structure. If we consider the overlapping nodes ego backbone, it appears that nodes with high betweenness tend to be overlapping nodes or nodes near the overlaps. Consequently, they are preferred and kept in the backbone.
Nonetheless, they may not contribute to the modularity of the network as other less influential nodes. Another distinction lies in the fact that the modularity vitality backbone doesn't remove edges with low weights. On average, it has a higher number of links compared to the overlapping nodes ego backbone. Thus, it is normal to have lower average betweenness values due to the existence of those edges that play a role in maintaining the network's modularity. In other words, they play a role in showing a clearly defined community structure.
\begin{figure}[!ht]
\centering
\includegraphics[width=5in, height=4.5 in]{FigureDemonstration.png}
\caption{The backbone extraction of two networks: Zachary Karate Club on the left and Wind Surfers on the right. (A) represents the original network. (B) represents the modularity vitality backbone. (C) represents the overlapping nodes ego backbone. The different colors of the nodes correspond to the various communities uncovered using the Louvain community detection algorithm. The size of the nodes is proportional to their weighted degree. The size of the edges is proportional to their edge weight.}
\label{Fig1}
\end{figure}
To illustrate these differences, we refer to two small networks, namely Zachary Karate Club and Wind Surfers, given in Figure \ref{Fig1}. Let’s discuss Zachary Karate Club first, represented on the left side of Figure \ref{Fig1}. One can point out that the modularity vitality backbone represented in (B) has a clique-like topology. In contrast, the overlapping nodes ego backbone reported in (C) exhibits a star-like structure. It is a good illustration of why the modularity vitality backbone has higher modularity while ``overlapping nodes ego backbone” is characterized by higher information spreading efficiency. Moreover, one can note that the number of communities is the same as the original network using the proposed technique. In contrast, there are no more nodes from the green community in the overlapping nodes ego backbone.
Diving deeper, the modularity vitality backbone keeps node 24 in the blue community and node 5 in the green community. They are discarded by overlapping nodes ego backbone and replaced by node 13 from the yellow community and node 8 from the red community. If we look at the modularity vitality and betweenness scores of these nodes, we find that nodes 24 and 5 have high modularity but low betweenness. In contrast, nodes 13 and 8 have high betweenness and low modularity vitality scores.
The Wind Surfers network exhibits similar behavior. Indeed, the modularity vitality backbone shows a clear clique-like structure while a star-like structure emerges in the ``overlapping nodes ego backbone.” Overall, the modularity vitality backbone integrates more peripherical nodes while the overlapping nodes ego backbone tends to retain more nodes at the core of the communities.
Note that the proposed backbone extraction process can also integrate a further step reducing the number of edges. One may consider several strategies to do so. For instance, one may remove the links based on their weights in each community, preserving its connectedness. Another approach is to use the disparity filter to prune these edges. Additionally, one may prune edges in proportion to the size of the edge set in each community of the original network. So doing allows better preservation of the original community structure.
Additionally, preserving the nodes in the backbone according to the absolute value of the modularity vitality scores can be too brutal. One may integrate more information about the community structure, such as the community size, to better deal with the resolution limit issue \cite{fortunato2007resolution}.
\section{Conclusion}
\label{sec:conc}
Analyzing large-scale networks is essential to characterize their underlying topology and dynamics. However, the large size of networks hinders this process. Therefore, it is vital to remove redundant information from the network while keeping nodes and edges that preserve the relevant information. Backbones, whether coarse-grained or filter-based, tackle this problem.
Aware of the ubiquity of the modular structure of real-world networks, we propose a filter-based technique called ``modularity vitality backbone." The proposed algorithm aims to preserve the network's modularity as nodes are removed. This enables researchers to conduct studies on networks with smaller sizes yet maintain their dense regions, which in turn represent the main building blocks of the network. The proposed method extracts the backbone of real-world weighted networks after quantifying the contribution of the nodes to the overall modularity of the network. Based on these scores, one prunes the nodes that barely contribute to the network's modularity until one reaches the desired size of the backbone.
Experiments show that the modularity vitality backbone compares favorably with its alternative in terms of weighted modularity, average weighted degree, and average link weight. However, it doesn't necessarily keep the nodes contributing to the efficiency of information spreading. Instead, it preserves the nodes and their edges that strategically contribute to the modularity of the network. These results pave the way to developing a filtering backbone extractor dedicated to optimizing several quality measures simultaneously using the vitality framework.
In the short term, we plan to extend this preliminary work in various directions. Since modularity has known drawbacks, we plan to evaluate alternative mesoscopic quality measures. Moreover, we will develop the analysis using multiple mesoscopic and macroscopic evaluation measures. Additionally, we aim to investigate the influence of community detection algorithms on the backbone extraction process.
\bibliographystyle{unsrt}
\section{Introduction}
Complex networks, such as communication, biological, transportation, and contact networks, are widely analyzed. The daily production of data results in tremendously large real-world networks. Consequently, the analysis of such networks containing millions of nodes and billions of edges is more and more challenging, if not impossible, due to memory and time constraints. Thus, suitable extraction of the pertinent nodes and edges that preserve the essential information while reducing the size of the network is fundamental. Network backbones offer a way to do so. Two main research paths tackle this problem: coarse-grain backbones or filter-based backbones. In the former, one clusters together nodes sharing similarities to reduce the network size\cite{zeng2011coarse,zeng2019new}. In the latter, one removes nodes or edges from the network based on a given property \cite{coscia2017backboning, 10.1093/comnet/cnab021}.
The community structure is one of the significant properties in real-world networks. Indeed, it heavily determines their dynamics and their underlying functionalities \cite{boccaletti2006complex}. It is generally illustrated by dense regions of connected nodes that barely connect from one region to another. Communities can be non-overlapping or overlapping \cite{jebabli2018community}, hierarchical \cite{peel2020detectability}, and attributed \cite{atzmueller2021mining}. Community detection is one of the most prolific research areas in network science. It relies on numerous measures quantifying the quality of the community structure. Modularity is among the most popular\cite{newman2006modularity}. It compares the density of connections of the uncovered community structure with a similar random network. The higher the modularity, the higher the confidence in the tight community structure of the network.
Recent works have shown that one can exploit the community structure efficiently to extract backbones \cite{ghalmane2020extracting, ghalmane2021extracting}. Inspired by these works, we propose a filtering technique based on the preservation of the community structure of the network. It exploits the community structure using the concept of vitality. Vitality quantifies the contribution of a node to a given quality measure by removing this node and computing the variation of the quality measure. To assess its importance, we compute modularity as a quality measure, with and without the node in question. Then, one ranks the nodes from the lowest contribution on modularity to the highest. Subsequently, nodes with the lowest contribution are removed until one reaches the desired size of the network.
Comparative experimental evaluations are conducted on real-world weighted networks of different sizes and domains. The developed backbone extraction technique called ``modularity vitality backbone" is compared with the recently introduced community-based method ``overlapping nodes ego backbone" \cite{ghalmane2020extracting}. Results show that it is more effective in preserving the core information of the network and the community structure.
\newline
The main contributions of the paper summarize as follows:
\begin{itemize}
\item We propose a backbone filtering technique exploiting the community structure of networks.
\item Experiments with weighted networks show that it outperforms another alternative measure.
\item It can be easily adapted to any type of network (i.e., undirected, unweighted, and directed networks).
\end{itemize}
The remaining of the article is organized as follows. Section \ref{sec:relatedworks} discusses briefly the related works. Section \ref{sec:VCB} introduces the modularity vitality backbone. Sections \ref{sec:data} and \ref{sec:evaluationmeasures} present respectively the datasets and the evaluation measures used in this study. Section \ref{sec:experimentalresults} reports the results of the comparative evaluation. Section \ref{sec:discussion} discusses the results. Finally, section \ref{sec:conc} concludes the paper.
\section{Related Works}
\label{sec:relatedworks}
Backbones offer an ideal solution to the trade-off between preserving essential information in the network and reducing the network size. Backbone extraction studies concerns mainly two types of networks: mono-mode networks \cite{van2012high, cao2019motif, 10.1093/comnet/cnab021} and bipartite networks \cite{zweig2011systematic, neal2013identifying, neal2014backbone}. In this work, we consider mono-mode networks. Within this class of networks, there are two leading approaches for extracting backbones. The first is coarse-graining, and the second is filtering.
In coarse-graining methods, one group nodes with similar characteristics into a single node. For instance, authors in \cite{gfeller2007spectral} merge the nodes based on random walks. In the work of \cite{chen2016effectively}, authors use the $k$-nearest neighbors algorithm ($k$-NN) to group similar nodes based on the nearest higher-density-neighbor.
In filter-based methods, the goal is to remove redundant information by pruning nodes or edges in the network. Redundancy is assessed based on a statistical property.
Most of the works reported in the literature concern edge-filtering techniques. Serrano et al. propose the disparity filter. It uses a null model of the edge weights to preserve statistically significant edges \cite{serrano2009extracting}. Authors in \cite{goh2006skeleton} compute the betweenness centrality of edges and remove the ones that don't exceed a specific threshold. Authors in \cite{zhang2018extracting} use a combination of local and global information to extract the backbone. More precisely, they use the link weights to build the $h$-strength graph and the betweenness centrality to build the $h$-bridge graph. Then one obtains the backbone by merging $h$-strength and $h$-bridge. Simas et al. \cite{10.1093/comnet/cnab021} present the distance backbone based on the triangular organization of edges which preserves all shortest paths.
Node filtering techniques are less frequent. They rely on topological features to associate a score to the nodes. Nodes with the higher scores are then extracted \cite{DAI2018271, RePEc}.
In this line, in their recent work, Ghalmane et al. prune nodes based on the community structure characteristics of the network \cite{ghalmane2020extracting}. They propose two node-filtering techniques. The first one preserves the overlapping nodes and the hubs of the network. In contrast, the second conserves the overlapping nodes and their one-step neighbors to form the backbone. These two algorithms exhibit superior performances as compared to the popular disparity filter. These results illustrate the community structure's importance in preserving the core information in a network while reducing its size.
Inspired by these findings, we propose the ``modularity vitality backbone" algorithm. This node filtering technique also exploits the community structure of the network. It uses a measure of the node contribution to the modularity. Roughly speaking, nodes with the lowest contribution to the quality measure of the community structure are filtered. The remaining nodes form the backbone.
\section{Modularity Vitality Backbone}
\label{sec:VCB}
This section presents the vitality concept. Then, we briefly discuss various mesoscopic quality measures. We explain why we choose Newman's modularity as a quality measure. Finally, the algorithm of the proposed backbone extractor is given.
\subsection{Vitality Index.}
Let $G(V,E)$ be a simple and undirected graph where $V=\{v_1, v_2, ..., v_N\}$ is the set of nodes totaling $N=|V|$ and $E=\{(v_i, v_j) | v_i, v_j \in V\}$ is the set of edges. Denote $f(G)$ and $f(G\setminus \{u\})$ as two real-valued functions defined on the complete graph $G$ and on graph $G\setminus\{u\}$ without node or edge $u$. Then, the vitality index is the difference between both functions, defined as $\nu(G, u) = f(G) - f(G\setminus \{u\})$. The resulting value is a signed value, indicating the positive or negative contribution of the node or edge $u$ on graph $G$ \cite{koschutzki2005centrality}.
\subsection{Mesoscopic Quality Measures.}
There are numerous quality measures to characterize communities \cite{yang2015defining, leskovec2010empirical, rajeh2021characterizing}. Their goal is to answer how good is the community structure in a network. They use topological properties defined at the mesoscopic level. Let set $C=\{c_1, c_2,...,c_l,..., c_{n_c}\}$ represent $n_c = |C|$ communities of a graph $G$ and $f(c_l)$ represents a quality function of community $c_l$. One can categorize the quality functions into three main groups: \newline
\textbf{1. Based on internal connectivity:} such as internal density characterizing how densely connected the nodes are in a community compared to other communities. \newline
\textbf{2. Based on internal and external connectivity:} such as Flake-ODF measuring the fraction of nodes in a community with fewer internal edges than external ones. \newline
\textbf{3. Based on a network model:} such as Newman's modularity \cite{newman2006modularity} which assesses the difference between the real connections in the community $c_l$ and the random connections in the same community.
All of these quality functions characterize a single community. Hence, to quantify the quality of the overall community structure, one averages $f(c_l)$ over all the communities.
Newman's modularity is one of the most popular mesoscopic quality measures. Indeed, it is widely used in community detection algorithms as an optimization criterion \cite{clauset2004finding, brandes2007modularity, blondel2008fast}. This is the main reason why it is one of the well-accepted benchmarks for characterizing the community structure of the networks. Numerous extensions have also been proposed for modularity to account for networks with overlapping and hierarchical community structure \cite{chen2015fuzzy}. In this work, we use Newman's modularity as a quality measure to assess the vitality of nodes due to the following reasons: \newline
1. Modularity can naturally be extended to unweighted, undirected, and directed networks. \newline
2. Modularity vitality ensures that nodes that are the main contributors to the community structure are retained, regardless of their type (i.e., hub- or bridge-like).\newline
3. Previous works on modularity vitality centrality has proved to assign high scores to the most influential nodes \cite{magelinski2021measuring}.
\subsection{Algorithm.}
The ``modularity vitality backbone" is based on the vitality concept, where one can measure the contribution of a node or an edge using any quality measure computed on graphs. We use Newman's modularity as a quality measure. Nonetheless, one can opt for other quality measures to quantify the node and edge influence.
Newman's modularity enables us to differentiate between highly internally connected nodes (hubs) and nodes at the borders of the communities (bridges). Indeed, hubs increase the internal density of the communities. Therefore, they contribute positively to modularity. In contrast, bridges increase the connections between the communities. Consequently, they tend to decrease the modularity. As we choose to give equal importance to both types of nodes, we rank the nodes according to the absolute value of their modularity vitality score. It allows keeping nodes with the highest contribution to modularity, regardless of their role (i.e., hub nodes or bridge nodes). Then, one removes the nodes that barely contribute to modularity. The backbone extraction procedure is given in Algorithm 1.
\begin{figure}[!ht]
\centering
\includegraphics[width=1\columnwidth]{Alg1.png}
\caption{Algorithm 1}
\label{Fig1}
\end{figure}
Note that the vitality computation is not naively computed two times for each node. In such a case, the complexity can rapidly become prohibitive. Indeed, one computes instead the modularity variation reducing the computation's complexity to $O(|E|+Nn_c)$. It makes the vitality measure suitable for large-scale weighted networks. We also note that the symbol $|E|$ is extended to weighted networks.
\section{Datasets}
\label{sec:data}
We use a set of seven real-world networks originating from various domains (social, collaborative, ecological, and technological) in the experiments. The nodes and edges range from hundreds to thousands. We choose to integrate small networks in the experiments in order to get a better understanding on the filtering process. Table \ref{TableCharacteristcs} presents their basic topological characteristics. All the networks are freely available online\footnote{Aaron Clauset, Ellen Tucker, and Matthias Sainz, ``The Colorado Index of Complex Networks." https://icon.colorado.edu/ (2016).}\footnote{Tiago P. Peixoto, ``The Netzschleuder network catalogue and repository," https://networks.skewed.de/ (2020).}. As there is no ground truth available, we rely on the Louvain community detection algorithm to uncover their community structure \cite{blondel2008fast}.
\newline
\textbf{1. Zachary's Karate Club:} Nodes are members of a karate club and are connected if they are friends inside and outside the club. Edges are weighted by the relative interactions occurring between the members.\newline
\textbf{2. Wind Surfers:} Nodes are windsurfers in southern California in the fall of 1986. They are connected if they're friends. Edges are weighted based on the social closeness of the surfers to one another. \newline
\textbf{3. Madrid Train Bombing:} Nodes are terrorists in the train bombing of March 11, 2004, in Madrid. Edges represent contacts between the terrorists and are weighted based on the strength of their underlying relationship.\newline
\textbf{4. Les Mis\'erables:} Nodes are the characters in the novel ``Les Mis\'erables." Edges represent characters' co-appearances in the same chapter. They are weighted by the number of co-appearances.\newline
\textbf{5. Wiki Science:} Nodes are either applied, formal, natural, or social sciences Wikipedia pages. They're weighted by the cosine similarity between them. \newline
\textbf{6. Unicode Languages:} A bipartite network representing languages and countries. Weights represent the fraction of people in a given country having the literacy (reading and writing) of a specific language. \newline
\textbf{7. Scientific Collaboration:} Nodes are authors of articles in the ``Condensed Matter" category of arXiv. Edges represent co-authorship and are weighted by the number of joint papers among the authors.
\begin{table}[h]
\centering
\caption{Basic topological properties of the real-world networks under study. \textit{N} is the number of nodes. $|E|$ is the number of edges. $<k>$ is the average weighted degree. $\omega$ is the density. $\zeta$ is the transitivity. $k_{nn}(k)$ is the assortativity. $\epsilon$ is the efficiency. $Q$ is the weighted modularity of the network.}
\label{TableCharacteristcs}
\begin{tabular}{lcccccccccc}
\hline
Network & $N$ & $|E|$ & $<k>$ & $\omega $ & $\zeta$ & $k_{nn}(k)$ & $\epsilon$ & $Q$\\
\hline
Zachary's Karate Club & 33 & 77 & 13.59 & 0.139 & 0.256 & -0.476 & 0.492 & 0.444\\
Wind Surfers & 43 & 336 & 56.09 & 0.372 & 0.564 & -0.147 & 0.679 & 0.371\\
Madrid Train Bombing & 62 & 243 & 8.81 & 0.121 & 0.561 & 0.029 & 0.448 & 0.435 \\
Les Mis\'erables & 77 & 254 & 21.30 & 0.087 & 0.499 & -0.165 & 0.435 & 0.565 \\
Wiki Science & 687 & 6,523 & 7.35 & 0.028 & 0.469 & 0.244 & 0.323 & 0.631 \\
Unicode Languages & 868 & 1,255 & 0.697 & 0.003 & 0.00 & -0.171 & 0.255 & 0.772 \\
Scientific Collaboration & 16,726 & 47,594 & 9.23 & 0.0003 & 0.360 & 0.185 & 0.117 & 0.873\\
\hline
\end{tabular}
\end{table}
\section{Evaluation Measures}
\label{sec:evaluationmeasures}
We compare the effectiveness of the proposed backbone extraction technique and the ``overlapping hub ego backbone" based on four different evaluation measures classically used. \newline
\textbf{1. Average weighted degree:} The weighted degree of a node is the sum of the weights of all the edges connected to it. Hence, a higher average weighted degree backbone means that important nodes are kept in the graph, reflecting its connectedness. It is defined as follows:
\begin{equation}
<k> = \frac{1}{N}\sum_{i=1}^N k_i = \sum_{j \in \mathcal{N}(1)} w_{ij}
\end{equation}
where $\mathcal{N}(1)$ is the first-order neighborhood of node $i$. \newline
\textbf{2. Average link weight:} Links in the backbone preserve the information flow of the network. In other words, the higher the value of the links, the better the backbone in maintaining the core information of the graph. It is defined as follows:
\begin{equation}
<w> = \frac{1}{N}\sum_{i,j \in V} w_{ij}
\end{equation}
\newline
\textbf{3. Average betweenness:} Nodes with higher betweenness can disseminate information quickly. Hence, a backbone with higher average betweenness indicates that the speed of information dissemination is barely altered. It is defined as follows:
\begin{equation}
<b> = \frac{1}{N}\sum_{i=1}^N b_i = \sum_{i \neq s \neq t} \frac{\sigma_{s,t}^i}{\sigma_{s,t}}
\end{equation}
where $\sigma_{s,t}$ denotes the number of shortest paths between nodes $s$ and $t$ and $\sigma_{s,t}^i$ denotes the number of shortest paths between nodes $s$ and $t$ passing through node $i$. \newline
\textbf{4. Weighted Modularity:} Modularity assesses the quality of the community structure based on the difference between the actual and the expected fraction of edges in the communities. A backbone with higher modularity suggests that the community structure is less altered. It can be computed on unweighted and weighted networks \cite{newman2004analysis}. It is defined as follows:
\begin{equation}
Q = \frac{1}{2|E|} \sum_{i,j}\left[ A_{ij} - \frac{w_iw_j}{\sum_{i,j}w_{ij}}\right] \delta(c_i,c_j)
\end{equation}
where $A_{ij}$ is the weighted adjacency matrix of graph $G$ and $\delta(c_i,c_j)$ equals 1 if nodes $i$ and $j$ belong to the same community, otherwise it equals 0.
\section{Experimental Results}
\label{sec:experimentalresults}
The effectiveness of ``modularity vitality backbone" is compared with another community-aware backbone extraction technique recently introduced \cite{ghalmane2020extracting}. The authors propose two backbones in their work, namely ``overlapping nodes ego backbone" and ``overlapping nodes and hubs backbone." In their comparative evaluation, they show that both techniques perform favorably compared to the popular disparity filter.
Therefore, in this work, we restrict our comparison to the most effective: overlapping nodes ego backbone. Table \ref{TableResults} reports the experimental results for seven real-world networks under study. The backbones quality measures are the average weighted degree ($<k>$), the average link weight ($<w>$), the average betweenness ($<b>$), and the weighted modularity ($Q$). As in their paper, we fix the backbone size to 30\% of the original network.
Let's first discuss the average weighted degree. The higher its value, the better the backbone is in keeping the salient nodes maintaining its connectedness. Table \ref{TableResults} reports that the modularity vitality backbone outperforms overlapping nodes ego backbone in all of the networks under study. Moreover, the difference ranges from very small magnitudes to orders of magnitude higher. For example, in Wiki Science, the average weighted degree in the modularity vitality backbone is eleven times higher than the ``overlapping nodes ego backbone." On the contrary, in Scientific Collaboration, the difference is barely noticeable (0.02).
Let's turn to average link weight. The average link weight characterizes the relevance of the links kept in a backbone. Hence, the higher its value, the better the backbone is in preserving essential links. The results show that the modularity vitality backbone outperforms the overlapping nodes ego backbone in six out of the seven networks.
Now, we discuss the average betweenness. This measure indicates the amount of information flow that can pass through the nodes of a given backbone. The higher its value, the higher the efficiency of the backbone in information spreading. The modularity vitality backbone outperforms the overlapping nodes ego backbone in only one out of the seven networks. In Scientific Collaboration networks, their values are comparable. It indicates that the information spread within the modularity vitality backbone is not as efficient as overlapping nodes ego backbone. Note, however, that the differences between the two backbones are less pronounced.
Finally, we turn to weighted modularity. The higher the modularity of the backbone, the better the quality of its community structure. As reported in table \ref{TableResults}, the modularity vitality backbone outperforms the overlapping nodes ego backbone on all the networks under study. These results are not surprising. Indeed, the modularity vitality backbone prunes the nodes contributing less to the modularity of the network. Hence, it tends to preserve the modularity as pruning proceeds.
To summarize, the modularity vitality backbone exhibits a higher weighted modularity than the overlapping nodes ego backbone. It preserves essential nodes in the network (i.e., hubs and bridges), thus achieving a higher average weighted degree and average link weight. However, maintaining the community structure comes at a price of a lower average betweenness.
\begin{table}[t]
\centering
\caption{The computed values for the average weighted degree ($<k>$), link weight ($<w>$), and betweenness ($<b>$) alongside the weighted modularity ($Q$) of the backbones with 30\% of the initial size of the network. For brevity, MV stands for ``modularity vitality backbone," and OE stands for ``overlapping nodes ego backbone."}
\begin{tabular}{l|ll|ll|ll|ll}
\hline
\multirow{2}{*}{Network} & \multicolumn{2}{l|}{$<k>$} & \multicolumn{2}{l|}{$<w>$} & \multicolumn{2}{l|}{$<b>$} & \multicolumn{2}{l}{$Q$} \\
\cline{2-9}
& MV & OE & MV & OE & MV & OE & MV & OE \\
\hline
Zachary's Karate Club & \textbf{13.00} & 8.40 & \textbf{6.05} & 4.07 & 0.12 & \textbf{0.27} & \textbf{0.35} & 0.32 \\
Wind Surfers & \textbf{71.38} & 35.08 & \textbf{35.69} & 17.54 & 0.11 & \textbf{0.15} & \textbf{0.36} & 0.32 \\
Madrid Train Bombing & \textbf{8.53} & 3.90 & \textbf{4.26} & 1.95 & 0.09 & \textbf{0.14} & \textbf{0.38} & 0.17 \\
Les Mis\'erables & \textbf{39.48} & 19.08 & \textbf{19.74} & 9.54 & 0.08 & \textbf{0.14} & \textbf{0.49} & 0.48 \\
Wiki Science & \textbf{10.16} & 0.92 & \textbf{5.08} & 0.46 & 0.01 & \textbf{0.08} & \textbf{0.73} & 0.72 \\
Unicode Languages & \textbf{1.46} & 1.28 & \textbf{0.73} & 0.64 & \textbf{0.03} & 0.02 & \textbf{0.79} & 0.78 \\
Scientific Collaboration & \textbf{17.22} & 17.20 & 4.71 & \textbf{8.60} & 0.001 & 0.001 & \textbf{0.81} & 0.71 \\
\hline
\end{tabular}
\label{TableResults}
\end{table}
\section{Discussion}
\label{sec:discussion}
The constant increase of real-world networks size has prompted researchers to design a smaller yet accurate representation of networks. This problem is tackled either with coarse-graining or filter-based methods. A recent work by Ghalmane et al. \cite{ghalmane2020extracting} has shown interest in exploiting the modular structure of the network to deal with this issue. Building on this finding, we propose a new backbone extractor, ``modularity vitality backbone," that aims to preserve the quality of the community structure. Assigning a modularity vitality score to the nodes, it prunes those with a low contribution to the network's modularity.
We performed a comparative analysis with the recently introduced ``overlapping nodes ego backbone. These investigations on a set of seven real-world networks from various domains are globally at the advantage of the proposed technique. After pruning 70\% of the network, results show that the modularity vitality backbone maintains higher modularity than overlapping hubs ego backbone. This expected behavior comes with higher performance in terms of average node degree and average link weight. Nonetheless, information efficiency isn't guaranteed. Indeed, one can point out that in five out of seven networks, the modularity vitality backbone suffers from lower information efficiency.
It can be explained by how the modularity vitality backbone proceeds. Indeed, one removes nodes that barely affect modularity. Those nodes may have high betweenness, yet they do contribute much to the network's modular structure. If we consider the overlapping nodes ego backbone, it appears that nodes with high betweenness tend to be overlapping nodes or nodes near the overlaps. Consequently, they are preferred and kept in the backbone.
Nonetheless, they may not contribute to the modularity of the network as other less influential nodes. Another distinction lies in the fact that the modularity vitality backbone doesn't remove edges with low weights. On average, it has a higher number of links compared to the overlapping nodes ego backbone. Thus, it is normal to have lower average betweenness values due to the existence of those edges that play a role in maintaining the network's modularity. In other words, they play a role in showing a clearly defined community structure.
\begin{figure}[!ht]
\centering
\includegraphics[width=5in, height=4.5 in]{FigureDemonstration.png}
\caption{The backbone extraction of two networks: Zachary Karate Club on the left and Wind Surfers on the right. (A) represents the original network. (B) represents the modularity vitality backbone. (C) represents the overlapping nodes ego backbone. The different colors of the nodes correspond to the various communities uncovered using the Louvain community detection algorithm. The size of the nodes is proportional to their weighted degree. The size of the edges is proportional to their edge weight.}
\label{Fig1}
\end{figure}
To illustrate these differences, we refer to two small networks, namely Zachary Karate Club and Wind Surfers, given in Figure \ref{Fig1}. Let’s discuss Zachary Karate Club first, represented on the left side of Figure \ref{Fig1}. One can point out that the modularity vitality backbone represented in (B) has a clique-like topology. In contrast, the overlapping nodes ego backbone reported in (C) exhibits a star-like structure. It is a good illustration of why the modularity vitality backbone has higher modularity while ``overlapping nodes ego backbone” is characterized by higher information spreading efficiency. Moreover, one can note that the number of communities is the same as the original network using the proposed technique. In contrast, there are no more nodes from the green community in the overlapping nodes ego backbone.
Diving deeper, the modularity vitality backbone keeps node 24 in the blue community and node 5 in the green community. They are discarded by overlapping nodes ego backbone and replaced by node 13 from the yellow community and node 8 from the red community. If we look at the modularity vitality and betweenness scores of these nodes, we find that nodes 24 and 5 have high modularity but low betweenness. In contrast, nodes 13 and 8 have high betweenness and low modularity vitality scores.
The Wind Surfers network exhibits similar behavior. Indeed, the modularity vitality backbone shows a clear clique-like structure while a star-like structure emerges in the ``overlapping nodes ego backbone.” Overall, the modularity vitality backbone integrates more peripherical nodes while the overlapping nodes ego backbone tends to retain more nodes at the core of the communities.
Note that the proposed backbone extraction process can also integrate a further step reducing the number of edges. One may consider several strategies to do so. For instance, one may remove the links based on their weights in each community, preserving its connectedness. Another approach is to use the disparity filter to prune these edges. Additionally, one may prune edges in proportion to the size of the edge set in each community of the original network. So doing allows better preservation of the original community structure.
Additionally, preserving the nodes in the backbone according to the absolute value of the modularity vitality scores can be too brutal. One may integrate more information about the community structure, such as the community size, to better deal with the resolution limit issue \cite{fortunato2007resolution}.
\section{Conclusion}
\label{sec:conc}
Analyzing large-scale networks is essential to characterize their underlying topology and dynamics. However, the large size of networks hinders this process. Therefore, it is vital to remove redundant information from the network while keeping nodes and edges that preserve the relevant information. Backbones, whether coarse-grained or filter-based, tackle this problem.
Aware of the ubiquity of the modular structure of real-world networks, we propose a filter-based technique called ``modularity vitality backbone." The proposed algorithm aims to preserve the network's modularity as nodes are removed. This enables researchers to conduct studies on networks with smaller sizes yet maintain their dense regions, which in turn represent the main building blocks of the network. The proposed method extracts the backbone of real-world weighted networks after quantifying the contribution of the nodes to the overall modularity of the network. Based on these scores, one prunes the nodes that barely contribute to the network's modularity until one reaches the desired size of the backbone.
Experiments show that the modularity vitality backbone compares favorably with its alternative in terms of weighted modularity, average weighted degree, and average link weight. However, it doesn't necessarily keep the nodes contributing to the efficiency of information spreading. Instead, it preserves the nodes and their edges that strategically contribute to the modularity of the network. These results pave the way to developing a filtering backbone extractor dedicated to optimizing several quality measures simultaneously using the vitality framework.
In the short term, we plan to extend this preliminary work in various directions. Since modularity has known drawbacks, we plan to evaluate alternative mesoscopic quality measures. Moreover, we will develop the analysis using multiple mesoscopic and macroscopic evaluation measures. Additionally, we aim to investigate the influence of community detection algorithms on the backbone extraction process.
|
\section{Introduction}
Optimal transport offers a systematic approach to compare probability distributions by finding a transport plan (coupling) that minimizes the cost of transporting mass from one distribution to another. Optimal transport has been successfully applied in a wide range of fields, such as computer graphics \cite{solomon2015convolutional,solomon2014earth}, graph representation learning \cite{chen2020graph,petric2019got}, text classification \cite{yurochkin2019hierarchical}, and domain adaptation \cite{courty2016optimal,courty2014domain}, to name a few.
Despite the popularity of OT, existing OT formulations are mostly limited to {transporting scalar-valued mass.}
Applications such as diffusion tensor imaging \cite{le2001diffusion}, image set classification~\cite{huang2015log,harandi2014manifold},
anisotropic diffusion \cite{weickert1998anisotropic}, involve symmetric positive definite (SPD) matrix-valued distributions.
In brain imaging particularly, where the local diffusion of water molecules are encoded in a SPD matrix \cite{assaf2008diffusion}, being able to compare fields of such SPD matrices is crucial. This, however, requires a nontrivial generalization of existing optimal transport framework with careful construction of cost and transport plan.
In the quantum mechanics setting, existing works~\cite{jiang2012distances,carlen2014analog,chen2017matrix,chen2018vector} have explored geodesic formulation of the Wasserstein distance between vector and matrix-valued densities. In \cite{ning2013thesis,ning2014matrix}, the Monge-Kantorovich optimal mass transport problem has been studied for comparing matrix-valued power spectra measures. Recently, \cite{peyre2016quantum} proposed to solve an unbalanced optimal transport problem for SPD-valued distributions of unequal masses.
In this paper, we propose a general framework for solving the balanced OT problem between SPD-valued distributions, where the cost and the coupling are represented as block SPD matrices. We discuss a Riemannian manifold structure for the set of such block coupling matrices that allows to use the Riemannian optimization framework \cite{absil2009optimization, boumal2020intromanifolds} to solve various generalized OT problems. Specifically, our {contributions} are the following.
\begin{enumerate}
\item We introduce the general SPD matrix-valued balanced OT problem for block SPD marginals. We study the metric properties of the proposed formulation and discuss its relationship with the $2$-Wasserstein distance.
\item We propose a novel manifold structure for the set of block SPD coupling matrices, which generalizes the manifold structures studied in \cite{douik2019manifold,shi2021coupling,mishra2021manifold,mishra2019riemannian}. We discuss optimization-related ingredients like Riemannian metric, Riemannian gradient, Hessian, and retraction.
\item We extend our SPD-valued balanced OT formulation to block SPD Wasserstein barycenter and Gromov-Wasserstein OT.
\item We empirically illustrate the benefit of the proposed framework in domain adaptation, tensor-valued shape interpolation, and displacement interpolation between tensor fields.
\end{enumerate}
\paragraph{Organizations.}
We start with a brief review of optimal transport and Riemannian optimization in Section \ref{preliminary_sect}. In Section \ref{matrix_generalized_ot_sect}, we introduce the generalized SPD matrix-valued OT problem. Section \ref{manifold_geometry_sect} derives the necessary Riemannian optimization ingredients. Section \ref{sec:applications} develops two extensions to the matrix-valued OT problem. In Section \ref{Experiment_sect}, we test the proposed framework in various applications. In the appendix sections, we provide the proofs and show additional experiments.
\section{Preliminaries}
\label{preliminary_sect}
\subsection{Optimal transport}
\label{ot_preliminary_sect}
Consider two discrete measures supported on ${\mathbb R}^d$, $\mu = \sum_{i = 1}^m p_i \delta_{{\mathbf x}_i}$, $\nu = \sum_{j = 1}^n q_j \delta_{{\mathbf y}_j}$, where ${\mathbf x}_i, {\mathbf y}_j \in {\mathbb R}^d$. The weights ${\mathbf p} \in \Sigma_m, {\mathbf q} \in \Sigma_n$ are in probability simplex, i.e., $\sum_i p_i = \sum_j q_j = 1$, for $p_i, q_j \geq 0$ and $\delta_{\mathbf x}$ is the Dirac at ${\mathbf x}$. The $2$-Wasserstein distance between $\mu, \nu$ is given by solving the Monge-Kantorovich optimal transport problem:
\begin{equation}
\label{W22}
{\mathrm W}_2^2({\mathbf p},{\mathbf q}) = \min_{{\boldsymbol \gamma} \in \Pi({\mathbf p},{\mathbf q})} \sum_{i,j} \|{\mathbf x}_i-{\mathbf y}_j\|^2 \gamma_{i,j},
\end{equation}
where $\Pi({\mathbf p},{\mathbf q}) \coloneqq \{{\boldsymbol \gamma} \in {\mathbb R}^{m \times n} : {\boldsymbol \gamma} \geq 0, {\boldsymbol \gamma} {\mathbf 1} = {\mathbf p}, {\boldsymbol \gamma}^\top {\mathbf 1} = {\mathbf q} \}$ is the space of joint distribution between the source and
the target marginals. An optimal solution of (\ref{W22}) is referred to as an optimal transport plan (or coupling).
Recently, \cite{cuturi2013sinkhorn} proposed the Sinkhorn-Knopp algorithm~\cite{sinkhorn1964relationship,knight2008sinkhorn} for entropy-regularized OT formulation. For a recent survey of OT literature and related machine learning applications, please refer to~\cite{peyre2019computational}.
\subsection{Riemannian optimization
A matrix manifold $\mathcal{M}$ is a smooth subset of the ambient vector space ${\mathcal{V}}$ with local bijectivity to the Euclidean space. A Riemannian manifold is a manifold endowed with a Riemannian metric (a smooth, symmetric positive definite inner product structure $\langle \cdot, \cdot\rangle_x$) on every tangent space $T_x\mathcal{M}$. The induced norm on the tangent space is thus $\| u\|_x = \sqrt{\langle u,u \rangle_x}$.
The orthogonal projection operation for an embedded matrix manifold ${\rm P}_{x}: {\mathcal{V}} \xrightarrow{} T_x\mathcal{M}$ is a projection that is orthogonal with respect to the Riemannian metric $\langle \cdot, \cdot \rangle_x$. Retraction is a smooth map from tangent space to the manifold.
That is, for any $x \in \mathcal{M}$, retraction $R_x: T_x\mathcal{M} \xrightarrow{} \mathcal{M}$ such that 1) $R_x(0) = x$ and 2) $\mathrm{D} R_x(0)[u] = u$, where $\mathrm{D} f(x)[u]$ is the derivative of a function at $x$ along direction $u$.
The Riemannian gradient of a function $F:\mathcal{M} \xrightarrow{} {\mathbb R}$ at $x$, denoted as ${\mathrm{grad}} F(x)$ generalizes the notion of the Euclidean gradient, which is defined as the unique tangent vector satisfying $\langle {\mathrm{grad}} F(x) , u \rangle_x = \mathrm{D} F(x)[u] = \langle \nabla F(x) ,u\rangle_2$ for any $u \in T_x\mathcal{M}$. $\nabla F(x)$ is the Euclidean gradient and $\langle \cdot, \cdot\rangle_2$ denotes the Euclidean inner product. To minimize the function, Riemannian gradient descent \cite{absil2009optimization} and other first-order solvers apply retraction to update the iterates along the direction of negative Riemannian gradient while staying on the manifold, i.e., $x_{t+1} = R_{x_t}(-\eta \, {\mathrm{grad}} F(x_t))$, where $\eta$ is the step size.
Similarly, the Riemannian Hessian ${\mathrm{Hess}} F(x) : T_x\mathcal{M} \xrightarrow{} T_x\mathcal{M}$ is defined as the covariant derivative of Riemannian gradient. Popular second-order methods, such as trust regions and cubic regularized Newton's methods have been adapted to Riemannian optimization \cite{absil2007trust,agarwal2018adaptive}.
\section{Block SPD optimal transport}
\label{matrix_generalized_ot_sect}
Consider the same setting as in Section \ref{ot_preliminary_sect}, with two given measures $\mu = \sum_{i = 1}^m p_i \delta_{{\mathbf x}_i}$, $\nu = \sum_{j = 1}^n q_j \delta_{{\mathbf y}_j}$ and the marginals ${\mathbf p} \in \Sigma_m$ and ${\mathbf q} \in \Sigma_n$.
We consider \textit{lifting} the given probability measures to a space of matrix-valued measures where an associated \textit{mass} is represented as a positive semi-definite matrix. To this end, we convert the marginals ${\mathbf p}$ and ${\mathbf q}$ to block matrices ${\mathbf P}$ and ${\mathbf Q}$, where ${\mathbf P} \coloneqq \{[{\mathbf P}_i]_{m\times 1}:{\mathbf P}_i = p_i {\mathbf I}\}$ and ${\mathbf Q}\coloneqq\{[{\mathbf Q}_j]_{n\times 1}:{\mathbf Q}_j = q_i {\mathbf I}\}$. Here, $[\cdot]_{m \times n}$ denotes a collection of $mn$ matrices organized as a block matrix and ${\mathbf I}$ represents the identity matrix.
The cost of transporting a positive semi-definite (matrix) mass $\mathbf A$ from position ${\mathbf x}_i$ to ${\mathbf y}_j$ is parameterized by a (given) positive semi-definite matrix $\mathbf C_{i,j}$ and is computed as $\mathrm{tr}(\mathbf C_{i,j}\mathbf A)$. Under this setting, we propose a {block SPD matrix valued OT} problem as
\begin{equation}
{\rm MW}^2({\mathbf P}, {\mathbf Q}) \coloneqq \min_{{\mathbf \Gamma} \in {\boldsymbol \Pi}(m,n,d, {\mathbf P}, {\mathbf Q}) } \sum_{i,j} \mathrm{tr}( \mathbf C_{i,j} {\mathbf \Gamma}_{i,j}), \label{reformulated_mw}
\end{equation}
where ${\mathbf \Gamma} = [{\mathbf \Gamma}_{i,j}]_{m \times n}$ is a block-matrix coupling of size $m \times n$ and the set of such couplings are defined as ${\boldsymbol \Pi}(m,n,d,{\mathbf P}, {\mathbf Q}) \coloneqq \{ [{\mathbf \Gamma}_{i,j}]_{m \times n} : {\mathbf \Gamma}_{i,j} \in {\mathbb S}_{+}^d, \sum_j {\mathbf \Gamma}_{i,j} = {\mathbf P}_i, \sum_i {\mathbf \Gamma}_{i,j} = {\mathbf Q}_j, \forall i \in [m], j \in [n] \}$. Here ${\mathbb S}_{+}^d$ is used to denote the set of $d\times d$ positive semi-definite matrices and $\mathrm{tr}(\cdot)$ is the matrix trace. With a slight abuse of notation, let ${\rm MW}({\mathbf p}, {\mathbf q})\coloneqq {\rm MW}({\mathbf P}, {\mathbf Q})$ with ${\mathbf P},{\mathbf Q}$ defined above. Below we compare the generalized distance ${\rm MW}({\mathbf p}, {\mathbf q})$ to ${\rm W}_2({\mathbf p}, {\mathbf q})$ and analyze its metric properties.
\paragraph{Relationship with ${\rm W}_2({\mathbf p}, {\mathbf q})$.}
The $2$-Wasserstein distance ${\mathrm W}_2({\mathbf p},{\mathbf q})$ considers the transportation cost of a unit mass from ${\mathbf x}_i$ to ${\mathbf y}_j$ as $\|{\mathbf x}_i-{\mathbf y}_j\|^2$, which can be rewritten as $\mathrm{tr}{(({\mathbf x}_i -{\mathbf y}_j)({\mathbf x}_i - {\mathbf y}_j)^\top)}$. Hence, if we consider the cost for the proposed OT problem (\ref{reformulated_mw}) as $\mathbf C_{i,j} = ({\mathbf x}_i -{\mathbf y}_j)({\mathbf x}_i - {\mathbf y}_j)^\top$, then ${\rm MW}({\mathbf p}, {\mathbf q}) = {\mathrm W}_2({\mathbf p},{\mathbf q})$ where the optimal block-matrix coupling is ${\mathbf \Gamma}^* = {\boldsymbol \gamma}^* \otimes {\mathbf I}$ where ${\boldsymbol \gamma}^*$ is the optimal coupling for ${\rm W}_2({\mathbf p}, {\mathbf q})$ and $\otimes$ is the Kronecker product.
\paragraph{Metric properties of ${\rm MW}({\mathbf p}, {\mathbf q})$.}
In the following proposition, we formalize the conditions under which ${\rm MW}({\mathbf p}, {\mathbf q})$ is a valid distance metric.
\begin{proposition}
\label{mw_distance_prop}
Suppose the marginals ${\mathbf p}, {\mathbf q} \in \Sigma_n$ have the same support and the costs $\{\mathbf C_{i,j}\}_{i,j=1}^n$ satisfy
\begin{enumerate}
\item $\mathbf C_{i,j}=\mathbf C_{j,i}$ and
\item $\mathbf C_{i,j}\succ{\mathbf 0}$ for $i\neq j$ and $\mathbf C_{i,j}={\mathbf 0}$ for $i=j$,
\item $\forall (i,j,k)\in [n]^3,$ and $ \mathbf A\succeq {\mathbf 0},$ $\sqrt{\mathrm{tr}( \mathbf C_{i,j} \mathbf A)} \leq \sqrt{\mathrm{tr}( \mathbf C_{i,k} \mathbf A)} + \sqrt{\mathrm{tr}( \mathbf C_{j,k} \mathbf A)}$.
\end{enumerate}
Then, ${\rm MW}({\mathbf p}, {\mathbf q})$ is a metric between the marginals ${\mathbf p}$ and~${\mathbf q}$.
\end{proposition}
We remark that the conditions on $\mathbf C_{i,j}$ in Proposition \ref{mw_distance_prop} generalize the conditions required for ${\rm W}_2({\mathbf p}, {\mathbf q})$ in \eqref{W22} to be a metric. See for example \cite[Proposition 2.2]{peyre2019computational}. In Appendix \ref{cost_choice_appendix}, we discuss some particular constructions of the cost that satisfy the conditions.
The previous discussion on ${\rm MW}({\mathbf p}, {\mathbf q})$ motivates the generalized SPD matrix valued OT problem \eqref{reformulated_mw} where the marginals are arbitrary positive semi-definite matrix mass. That is, ${\mathbf P}\coloneqq\{[{\mathbf P}_i]_{m\times 1}:{\mathbf P}_i \in {\mathbb S}_{+}^d\}$ and ${\mathbf Q}\coloneqq\{[{\mathbf Q}_j]_{n\times 1}:{\mathbf P}_i \in {\mathbb S}_{+}^d\}$ and ${\mathbf P}$ and ${\mathbf Q}$ have the same total mass. Without loss of generality, we assume $\sum_i {\mathbf P}_i = \sum_j {\mathbf Q}_j = {\mathbf I}$. The problem is well-defined provided that the corresponding coupling constraint set ${\boldsymbol \Pi}(m,n,d,{\mathbf P}, {\mathbf Q})$ is non-empty. This is discussed in Section \ref{manifold_geometry_sect}.
In the next section, we analyze the coupling constraint set ${\boldsymbol \Pi}(m,n,d,{\mathbf P}, {\mathbf Q})$ and show that it can be endowed with a manifold structure. This allows to exploit the versatile Riemannian optimization framework to solve (\ref{reformulated_mw}) and any more general problem \cite{absil2009optimization}.
\section{{\Manname}}
\label{manifold_geometry_sect}
We propose the following Riemannian manifold structure,
termed as the {block SPD coupling manifold},
\begin{equation}\label{eq:manifold_proposed}
\begin{array}{l}
\mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})
\coloneqq \{ {\mathbf \Gamma} : {\mathbf \Gamma}_{i,j} \in {\mathbb S}_{++}^d, \sum_{j} {\mathbf \Gamma}_{i,j} = {\mathbf P}_i,
\sum_i {\mathbf \Gamma}_{i,j} ={\mathbf Q}_j \},
\end{array}
\end{equation}
where $\sum_i {\mathbf P}_i = \sum_j {\mathbf Q}_j = {\mathbf I}$. Particularly, we restrict ${\mathbf P}_i, {\mathbf Q}_j,{\mathbf \Gamma}_{i,j} \in {\mathbb S}_{++}^d$, the set of SPD matrices. This ensures that the proposed manifold $\mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})$ in (\ref{eq:manifold_proposed}) is the interior of the set ${\boldsymbol \Pi}(m,n,d,{\mathbf P}, {\mathbf Q})$.
For arbitrary SPD marginals ${\mathbf P}, {\mathbf Q}$, there is no guarantee that the set $\mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})$ defined in (\ref{eq:manifold_proposed}) is non-empty \cite{ning2014matrix}. Hence, we assume the following.
\begin{assumption}\label{assumption:existence}
The block-SPD marginals ${\mathbf P}$ and ${\mathbf Q}$ are chosen such that there exists at least one element in the set $\mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})$.
\end{assumption}
It should be noted that Assumption \ref{assumption:existence} can be easily satisfied for various types of block-SPD marginals such as uniform or arbitrary positive diagonal marginals.
\begin{algorithm}[t]
\caption{Riemannian optimization for solving (\ref{eq:manifold_problem}).}
\label{Riemannian_optimizer}
\begin{algorithmic}[1]
\STATE Initialize a feasible ${\mathbf \Gamma} \in \mathcal{M}_{m,n}^d$.
\WHILE{not converging}
\STATE Compute Riemannian gradient (and Hessian) at ${\mathbf \Gamma}$.
\STATE Compute the update step ${\boldsymbol \xi} \in T_{{\mathbf \Gamma}} \mathcal{M}_{m,n}^d$.
\STATE Update ${\mathbf \Gamma} \leftarrow R_{{\mathbf \Gamma}}({\boldsymbol \xi})$.
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\begin{proposition}
\label{smooth_set_prop}
Under Assumption \ref{assumption:existence}, the set $\mathcal{M}^d_{m,n}({\mathbf P}, {\mathbf Q})$ is smooth, i.e., differentiable.
\end{proposition}
Proposition \ref{smooth_set_prop} implies that we can endow $\mathcal{M}^d_{m,n}({\mathbf P}, {\mathbf Q})$ with a Riemannian structure (introduced in Section \ref{Riem_opt_sec}).
It should be emphasized that the proposed manifold $\mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})$ can be regarded as a generalization to existing manifold structures. For example, when $d =1$ and either $m=1$ or $n=1$, $\mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})$ reduces to the multinomial manifold of probability simplex \cite{sun2015heterogeneous}. When $d=1$ and $m, n \neq 1$, it becomes the so-called doubly stochastic manifold \cite{douik2019manifold} with uniform marginals or the more general matrix coupling manifold \cite{shi2021coupling}. When $d > 1$ and either $m =1$ or $n =1$, our proposed manifold simplifies to the simplex manifold of SPD matrices \cite{mishra2019riemannian}.
Below, we derive various optimization-related ingredients on $\mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})$ that allow optimization of an arbitrary differentiable objective function on the manifold. It follows the general treatment proposed by \cite{absil2009optimization}.
\subsection{Riemannian geometry and optimization on { $\mathcal{M}_{m,n}^d$}}
\label{Riem_opt_sec}
We consider the general problem
\begin{equation}
\min_{{\mathbf \Gamma} \in \mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})} F({\mathbf \Gamma}) \label{eq:manifold_problem}
\end{equation}
via the Riemannian optimization framework. It provides a principled class of optimization methods and computational tools for manifolds, both first order and second order, as long as the ingredients such as Riemannian metric, orthogonal projection, retraction, and Riemannian gradient (and Hessian) of a function are defined \cite{absil2009optimization, boumal2014manopt, boumal2020intromanifolds}. Conceptually, the Riemannian optimization framework treats (\ref{eq:manifold_problem}) as an ``unconstrained'' optimization problem over the constraint manifold $\mathcal{M}_{m,n}^d$ (omitted marginals ${\mathbf P}$, ${\mathbf Q}$ for clarity).
In Algorithm \ref{Riemannian_optimizer}, we outline the skeletal steps involved in optimization over $\mathcal{M}_{m,n}^d$, where the step ${\boldsymbol \xi}$ can be computed from different Riemannian methods. In Riemannian steepest descent, ${\boldsymbol \xi} = -\eta \, {\mathrm{grad}} F({\mathbf \Gamma})$, where ${\mathrm{grad}} F({\mathbf \Gamma})$ is the Riemannian gradient at ${\mathbf \Gamma}$ and $\eta$ is step size. Also, ${\boldsymbol \xi}$ is given by the ``conjugate'' direction of ${\mathrm{grad}} F({\mathbf \Gamma})$ in the Riemannian conjugate gradient method. And, for the Riemannian trust-region method, ${\boldsymbol \xi}$ computation involves minimizing a second-order approximation of the objective function in a trust-region ball
\cite{absil2009optimization}. Below, we show the computations of these ingredients.
\paragraph{Riemannian metric.}
The manifold $\mathcal{M}^d_{m,n}$ is a submanifold of the Cartesian product of $m \times n$ SPD manifold of size $d \times d$, which we denote as $\bigtimes_{m, n} {\mathbb S}_{++}^d$. The dimension of the manifold $\mathcal{M}^d_{m,n}$ is $(m-1)(n-1)d(d+1)/2$. The {tangent space} characterization of $\mathcal{M}_{m,n}^d$ at ${\mathbf \Gamma}$ is given by
\begin{align*}
&T_{\mathbf \Gamma} \mathcal{M}_{m,n}^d = \{ [{\mathbf U}_{i,j}]_{m \times n}: {\mathbf U}_{i,j} \in {\mathbb S}^d, \sum_{j} {\mathbf U}_{i,j} = {\mathbf 0}, \sum_{i} {\mathbf U}_{i,j} = {\mathbf 0}\},
\end{align*}
where ${\mathbb S}^d$ is the set of $d \times d$ symmetric matrices. We endow each SPD manifold with the affine-invariant Riemannian metric \cite{bhatia2009positive}, which induces a {Riemannian metric} for the product manifold $\mathcal{M}_{m,n}^d$ as
\begin{equation}
\langle {\mathbf U}, {\mathbf V} \rangle_{\mathbf \Gamma} = \sum_{i,j} \mathrm{tr}({\mathbf \Gamma}_{i,j}^{-1} {\mathbf U}_{i,j} {\mathbf \Gamma}^{-1}_{i,j} {\mathbf V}_{i,j}), \label{Riem_metric}
\end{equation}
for any ${\mathbf U}, {\mathbf V} \in T_{{\mathbf \Gamma}} \mathcal{M}_{m,n}^d$.
\paragraph{Orthogonal projection, Riemannian gradient, and Riemannian Hessian.}
As an embedded submanifold, the orthogonal projection plays a crucial role in deriving the Riemannian gradient (as orthogonal projection of the Euclidean gradient in the ambient space).
\begin{proposition}
\label{orthogonal_proj}
The orthogonal projection of any $\mathbf S \in \bigtimes_{m, n} {\mathbb S}^d$ to $T_{\mathbf \Gamma}\mathcal{M}_{m,n}^d$ with respect to the Riemannian metric \eqref{Riem_metric} is given by
\begin{equation*}
{\rm P}_{{\mathbf \Gamma}}(\mathbf S) = {\mathbf U}, \text{ with } {\mathbf U}_{i,j} = \mathbf S_{i,j} + {\mathbf \Gamma}_{i,j} ({\mathbf \Lambda}_i+ {\mathbf \Theta}_j) {\mathbf \Gamma}_{i,j},
\end{equation*}
where auxiliary variables ${\mathbf \Lambda}_i, {\mathbf \Theta}_j$ are solved from the system of matrix linear equations:
\begin{equation*}
\begin{cases}
-\sum_i \mathbf S_{i,j} = \sum_{i} {\mathbf \Gamma}_{i,j} ({\mathbf \Lambda}_i + {\mathbf \Theta}_j) {\mathbf \Gamma}_{i,j}, & \forall j \\
-\sum_j \mathbf S_{i,j} = \sum_{j} {\mathbf \Gamma}_{i,j} ({\mathbf \Lambda}_i + {\mathbf \Theta}_j) {\mathbf \Gamma}_{i,j}, & \forall i.
\end{cases}
\end{equation*}
\end{proposition}
Subsequently, the Riemannian gradient and Hessian are derived as the orthogonal projection of the gradient and Hessian from the ambient space.
\begin{proposition}
\label{Riem_grad_hess}
The Riemannian gradient and Hessian of $F: \mathcal{M}_{m\times n}^d \xrightarrow{} {\mathbb R}$ are derived as
\begin{align*}
{\mathrm{grad}} F({\mathbf \Gamma}) &= {\rm P}_{\mathbf \Gamma}([{\mathbf \Gamma}_{i,j} \{ \nabla F({\mathbf \Gamma}_{i,j}) \}_{\rm S} {\mathbf \Gamma}_{i,j}]_{m\times n}),\\
{\mathrm{Hess}} F({\mathbf \Gamma})[{\mathbf U}] &= {\rm P}_{\mathbf \Gamma}([\mathrm{D} {\mathrm{grad}} F({\mathbf \Gamma}_{i,j})[{\mathbf U}_{i,j}] - \{ {\mathbf U}_{i,j} {\mathbf \Gamma}_{i,j}^{-1} {\mathrm{grad}} F({\mathbf \Gamma}_{i,j}) \}_{\rm S}]_{m\times n}),
\end{align*}
where ${\mathbf U} \in T_{\mathbf \Gamma}\mathcal{M}_{m,n}^d$ and $\nabla F({\mathbf \Gamma}_{i,j})$ is the block partial derivative of $F$ with respect to ${\mathbf \Gamma}_{i,j}$. Here, $\mathrm{D} {\mathrm{grad}} F({\mathbf \Gamma}_{i,j})[{\mathbf U}_{i,j}]$ denotes the directional derivative of the Riemannian gradient ${\mathrm{grad}} F$ along ${\mathbf U}$ and $\{ \mathbf A\}_{\rm S} \coloneqq (\mathbf A + \mathbf A^\top)/2$.
\end{proposition}
\paragraph{Retraction and block matrix balancing algorithm.}
Retraction of $\mathcal{M}_{m, n}^d$ is given by a composition of two operations. The first operation is to ensure positive definiteness of the blocks in the coupling matrix. In particular, we use the exponential map associated with the affine-invariant metric on the SPD manifold ${\mathbb S}_{++}^d$ \cite{bhatia2009positive}. The second operation is to ensure that the summation of the row blocks and column blocks respect the block-SPD marginals. Given an initialized block SPD matrix $[\mathbf A_{i,j}]_{m \times n}$, where $\mathbf A_{i,j} \in {\mathbb S}_{++}^d$, the goal is to find a `closest' block SPD coupling matrix $\mathbf B \in \mathcal{M}_{m,n}^d$. The is achieved by alternatively normalizing the row and column blocks to the corresponding marginals. The procedure is outlined in Algorithm~\ref{bmb_algorithm}. The solution for the row and column normalization factors ${\mathbf R}_j, \mathbf L_i$, which are SPD matrices, are computed by solving the Riccati equation $\mathbf T {\mathbf X} \mathbf T = {\mathbf Y}$ for given ${\mathbf X}, {\mathbf Y} \in {\mathbb S}_{++}^d$. Here, $\mathbf T$ admits a unique solution \cite{bhatia2009positive,malago2018wasserstein}. Different from the scalar marginals case where the scaling can be expressed as a diagonal matrix, we need to symmetrically normalize each SPD block matrix. Algorithm~\ref{bmb_algorithm} is a generalization of the RAS algorithm for balancing non-negative matrices \cite{sinkhorn1967diagonal}, which is related to the popular Sinkhorn-Knopp algorithm \cite{sinkhorn1964relationship,knight2008sinkhorn}.
It should be noted that a similar matrix balancing algorithm has been introduced for positive operators \cite{gurvits2004classical,georgiou2015positive}, where the convergence is only established in limited cases. Algorithm~\ref{bmb_algorithm} is different from the quantum Sinkhorn algorithm proposed in \cite{peyre2016quantum} that applies to the {unbalanced} setting. Although we do not provide a theoretical convergence analysis for Algorithm \ref{bmb_algorithm}, we empirically observe quick convergence of this algorithm in various settings (see Appendix \ref{convergence_bmb_retr_appendix}).
\begin{algorithm}[t]
\caption{Block matrix balancing algorithm}
\label{bmb_algorithm}
\begin{algorithmic}[1]
\STATE \textbf{Input:} $[\mathbf A_{i,j}]_{m \times n}$, where $\mathbf A_{i,j} \in {\mathbb S}_{++}^d$.
\WHILE{not converging}
\STATE Find ${\mathbf R}_j \in {\mathbb S}_{++}^d$ such that $\sum_i {\mathbf R}_j \mathbf A_{i,j} {\mathbf R}_j = {\mathbf Q}_j$, $\forall j$.
\STATE Update $\mathbf A_{i,j} \xleftarrow{} {\mathbf R}_j \mathbf A_{i,j} {\mathbf R}_j$, $\forall j$.
\STATE Find $\mathbf L_i \in {\mathbb S}_{++}^d$ such that $\sum_j \mathbf L_i \mathbf A_{i,j} \mathbf L_i = {\mathbf P}_i$, $\forall i$.
\STATE Update $\mathbf A_{i,j} \xleftarrow{} \mathbf L_i \mathbf A_{i,j} \mathbf L_i, \forall i$.
\ENDWHILE
\STATE \textbf{Output:} The balanced matrix $\mathbf B = \mathbf A$.
\end{algorithmic}
\end{algorithm}
Based on Algorithm \ref{bmb_algorithm}, we define a retraction $R_{\mathbf \Gamma}({\mathbf U})$ at ${\mathbf \Gamma} \in \mathcal{M}_{m, n}^d$ for any ${\mathbf U} \in T_{\mathbf \Gamma} \mathcal{M}_{m,n}^d$ as
\begin{equation}
R_{\mathbf \Gamma}({\mathbf U}) = {\rm MBalance} \big([{\mathbf \Gamma}_{i,j} {\rm exp}({\mathbf \Gamma}_{i,j}^{-1} {\mathbf U}_{i,j})]_{m \times n} \big), \label{retr_define}
\end{equation}
where {\rm MBalance} calls the matrix balancing procedure in Algorithm~\ref{bmb_algorithm} and $\exp(\cdot)$ denotes the matrix exponential. The retraction proposed in \eqref{retr_define} is valid (i.e., satisfy the two conditions) for diagonal marginals and empirically we also see the retraction is well-defined for arbitrary block-SPD marginals. See Appendix \ref{convergence_bmb_retr_appendix} for more details.
\subsection{Convergence and computational complexity}
\label{convergence_complexity_sect}
\paragraph{Convergence of Riemannian optimization.}
Similar to Euclidean optimization, the necessary first-order optimality condition for any differentiable $F$ on $\mathcal{M}_{m,n}^d$ is ${\mathrm{grad}} F({\mathbf \Gamma}^*) = 0$, i.e., where the Riemannian gradient vanishes. We call such ${\mathbf \Gamma}^*$ the stationary point. The Riemannian methods are known to converge to a stationary point \cite{absil2009optimization,boumal2020intromanifolds} under standard assumptions. Additionally, we show the following.
\begin{theorem}
\label{Prop_Riem_opt_optimal}
Suppose the objective function of the problem $\min_{{\mathbf \Gamma} \in {\boldsymbol \Pi}(m,n,d,{\mathbf P}, {\mathbf Q})} F({\mathbf \Gamma})$ is (strictly) convex and the optimal solution ${\mathbf \Gamma}^*$ is positive definite, i.e., it lies in the interior of ${\boldsymbol \Pi}(m,n,d, {\mathbf P}, {\mathbf Q})$. Then, Riemannian optimization (Algorithm~\ref{Riemannian_optimizer}) for (\ref{eq:manifold_problem}) converges to the same global optimal solution ${\mathbf \Gamma}^*$.
\end{theorem}
Theorem \ref{Prop_Riem_opt_optimal} guarantees the quality of the solution obtained by Riemannian optimization for a class of objective functions which includes the SPD matrix-valued OT problem with convex regularization.
\paragraph{Computational complexity.}
The complexity of each iteration of the Riemannian optimization algorithm is dominated by the computations of retraction and the Riemannian gradient (and Hessian). These also make use of the orthogonal projection operation. All these operations cost $O(mnd^3)$. It should be noted that the number of parameters to be learned is $mnd^2$.
\section{Extensions of block SPD optimal transport}
In this section, we discuss extensions of the proposed SPD OT framework to Wasserstein barycenters and the Gromov-Wasserstein setting.
\label{sec:applications}
\subsection{Block SPD Wasserstein barycenter learning}
\label{block_spd_barycenter}
We consider the problem of computing the Wasserstein barycenter of a set of block SPD matrix-valued measures.
Let $\Delta_n({\mathbb S}_{++}^d) \coloneqq \{ {\mathbf P} = [{\mathbf P}_i]_{n \times 1} : {\mathbf P}_i\in{\mathbb S}_{++}^d, \sum_i {\mathbf P}_i = {\mathbf I} \}$ denotes the space of $n\times 1$ block SPD marginals.
Then, the Wasserstein barycenter $\bar{{\mathbf P}}$ of a set $\{{\mathbf P}^\ell\}_{\ell=1}^K \in \Delta_{n_\ell}({\mathbb S}_{++}^d)$ is computed as follows:
\begin{equation}
\bar{{\mathbf P}}= \argmin_{{{\mathbf P}} \in \Delta_{n}({\mathbb S}_{++}^d)} \sum_{\ell =1}^K \omega_\ell {\rm MW}^2_\epsilon({{\mathbf P}}, {\mathbf P}^\ell) , \label{w_barycenter}
\end{equation}
where the given non-negative weights satisfy $\sum_\ell \omega_\ell =1$.
It should be noted that we employ a regularized version of the proposed block SPD OT problem (\ref{reformulated_mw}) to ensure the differentiability of the objective function near boundary in~\eqref{w_barycenter}. The regularized block SPD OT problem is defined as
\begin{align}
&{\rm MW}^2_\epsilon({\mathbf P}, {\mathbf Q}) := \min_{{\mathbf \Gamma} \in \mathcal{M}_{m,n}^d({\mathbf P}, {\mathbf Q})} \sum_{i,j} \Big( \mathrm{tr}( \mathbf C_{i,j}{\mathbf \Gamma}_{i,j})+ \epsilon \, \Omega({\mathbf \Gamma}_{i,j}) \Big),\label{eqn:regMWOT}
\end{align}
where $\epsilon>0$ is the regularization parameter and $\Omega(\cdot)$ is a strictly convex regularization (e.g., entropic regularization) on the block SPD coupling matrices.
To solve for $\bar{{\mathbf P}}$ in (\ref{w_barycenter}), we consider Riemannian optimization on $\Delta_n({\mathbb S}_{++}^d)$, which has recently been studied in~\cite{mishra2019riemannian}.
The following result provides an expression for the Euclidean gradient of the objective function in problem~(\ref{w_barycenter}).
\begin{proposition}
\label{dual_var_solution_prop}
The Euclidean gradient of \eqref{w_barycenter} with respect to ${{\mathbf P}}_i$, for $i \in [n]$ is
\begin{equation*}
\sum_{\ell=1}^K \omega_\ell \nabla_{{{\mathbf P}}_i}{\rm MW}_\epsilon({{\mathbf P}}, {\mathbf P}^\ell) = - \sum_{\ell=1}^K \omega_\ell ({\mathbf \Lambda}^\ell_i)^*,
\end{equation*}
where $({\mathbf \Lambda}^\ell_i)^*$ is given by evaluating the orthogonal projection ${\rm P}_{({\mathbf \Gamma}^\ell)^*}(\nabla_{({\mathbf \Gamma}^\ell)^*} {\rm MW}_\epsilon)$, with $\nabla_{({\mathbf \Gamma}^\ell_{i,j})^*} {\rm MW}_\epsilon = \mathbf C_{i,j}^\ell + \epsilon \nabla \Omega(({\mathbf \Gamma}^\ell_{i,j})^*)$ and $({\mathbf \Gamma}^\ell)^*$ is the optimal coupling for ${\mathbf P}^\ell$. That is, $({\mathbf \Lambda}^\ell_i)^*$ is the auxiliary variable obtained during the solving of the system of matrix linear equations in Proposition \ref{orthogonal_proj}.
\end{proposition}
The algorithm for computing the barycenter in \eqref{w_barycenter} is outlined in Algorithm \ref{wb_algorithm} (Appendix \ref{geometry_spd_simplex_appendix}).
\subsection{Block SPD Gromov-Wasserstein discrepancy}
\label{sec:proposedGW}
The Gromov-Wasserstein (GW) distance~\cite{memoli2011gromov} generalizes the optimal transport to the case where the measures are supported on possibly different metric spaces ${\mathcal{X}}$ and ${\mathcal{Y}}$.
Let $\mathbf D^x\in{\mathbb R}^{m\times m}$ and $\mathbf D^y\in{\mathbb R}^{n\times n}$ represent the similarity (or distance) between elements in metric spaces ${\mathcal{X}}$ and ${\mathcal{Y}}$ respectively. Let ${\mathbf p}\in \Sigma_m$ and ${\mathbf q} \in \Sigma_n$ be the marginals corresponding to the elements in ${\mathcal{X}}$ and ${\mathcal{Y}}$, respectively.
Then, the GW discrepancy between the two distance-marginal pairs $(\mathbf D^x, {\mathbf p})$ and $(\mathbf D^y, {\mathbf q})$ is defined as
\begin{equation*}
\label{eqn:GW}
\begin{array}{ll}
{\rm GW} \big( (\mathbf D^x, {\mathbf p}), ( \mathbf D^y, {\mathbf q} ) \big)
\coloneqq \min_{{\boldsymbol \gamma} \in \Pi({\mathbf p}, {\mathbf q})} \sum_{i,i',j,j'} \mathcal{L}(D^x_{i,i'}, D^y_{j,j'}) \gamma_{i,j} \gamma_{i',j'},
\end{array}
\end{equation*}
where $D_{k,l}$ denotes the $(k,l)$-th element in the matrix $\mathbf D$ and $\mathcal{L}$ is a loss between the distance pairs. Common choices of $\mathcal{L}$ include the $L_2$ distance and the KL divergence.
We now generalize the GW framework to our setting where the marginals are SPD matrix-valued measures.
Let $(\mathbf D^x, {\mathbf P})$ and $(\mathbf D^y, {\mathbf Q})$ be two distance-marginal pairs, where the Dirac measures are given by $\sum_i {\mathbf P}_i \delta_{x_i}$, $\sum_j {\mathbf Q}_j \delta_{y_j}$ respectively, for $\{x_i\}_{i \in [m]} \subset {\mathcal{X}}, \{y_j\}_{j \in [n]} \subset {\mathcal{Y}}$. The marginals are tensor-valued with ${\mathbf P} \in \Delta_m({\mathbb S}_{++}^d)$, ${\mathbf Q} \in \Delta_n({\mathbb S}_{++}^d)$.
We define the SPD generalized GW discrepancy as
\begin{equation}\label{scalar_gw_ot}
{\rm MGW} \big( (\mathbf D^x, {\mathbf P}), ( \mathbf D^y, {\mathbf Q} ) \big) \coloneqq
\min\limits_{{\mathbf \Gamma} \in \mathcal{M}^d_{m \times n}} \sum\limits_{{i,i',} {j, j'}} \mathcal{L}\big( D^x_{i,i'},D^y_{j, j'} \big) \mathrm{tr}( {\mathbf \Gamma}_{i,j} {\mathbf \Gamma}_{i',j'}),
\end{equation}
where we use Riemannian optimization (Algorithm \ref{Riemannian_optimizer}) to solve problem~\eqref{scalar_gw_ot}.
\paragraph{Gromov-Wasserstein averaging of distance matrices.}
The GW formulation with scalar-valued probability measures has been used for averaging distance matrices~\cite{peyre2016gromov}. Building on \eqref{scalar_gw_ot}, we consider the problem of averaging distance matrices where the marginals are SPD-valued.
Let $\{(\mathbf D^\ell, {\mathbf P}^\ell)\}_{\ell = 1}^K$ with ${\mathbf P}^\ell \in \Delta_{n_\ell}({\mathbb S}_{++}^d)$, be a set of distance-marginal pairs on $K$ incomparable domains.
Suppose the barycenter marginals $\bar{{\mathbf P}} \in \Delta_{n}({\mathbb S}_{++}^d)$ are given, the goal is to find the average distance matrix $\bar{\mathbf D}$ by solving
\begin{equation}\label{gw_barycenter_problem}
\begin{array}{ll}
\bar{\mathbf D} = \argmin\limits_{\mathbf D \in {\mathbb S}^n : D_{i,j} \geq 0} \sum_{\ell=1}^K \omega_\ell \, {\rm MGW} \big( (\mathbf D, \bar{{\mathbf P}}), ( \mathbf D^\ell, {\mathbf P}^\ell ) \big),
\end{array}
\end{equation}
where the given weights satisfy $\sum_\ell \omega_\ell = 1$. Problem \eqref{gw_barycenter_problem} can be solved via a block coordinate descent method, that iteratively updates the couplings $\{ {\mathbf \Gamma}^\ell \}_{\ell=1}^K$ and the distance matrix $\bar{\mathbf D}$. The update of the coupling is performed via Algorithm \ref{Riemannian_optimizer}. For the update of the distance matrix, we show when the loss $\mathcal{L}$ is decomposable, including the case of $L_2$ distance or the KL divergence, the optimal $\bar{\mathbf D}$ admits a closed-form solution. This is a generalization of the result \cite[Proposition 3]{peyre2016gromov} to SPD-valued marginals.
\begin{proposition}
\label{average_distance_update}
Suppose the loss $\mathcal{L}$ can be decomposed as $\mathcal{L}(a, b) = f_1(a) + f_2(b) - h_1(a) h_2(b)$ with $f_1'/h_1'$ invertible, then \eqref{gw_barycenter_problem} has a closed form solution given by $\bar{D}_{i,i'} = \big(\frac{f_1'}{h_1'} \big)^{-1} \big(h_{i,i'} \big)$ with
\begin{equation*}
h_{i,i'} = \Big( \frac{\sum_{\ell=1}^K \omega_\ell \mathrm{tr} \big(\sum_j {\mathbf \Gamma}_{i,j}^\ell \sum_{j'} h_2( D^\ell_{j,j'}) {\mathbf \Gamma}^\ell_{i', j'} \big) }{\mathrm{tr}(\bar{{\mathbf P}}_i \bar{{\mathbf P}}_{i'})} \Big).
\end{equation*}
\end{proposition}
\section{Experiments}
\label{Experiment_sect}
In this section, we show the utility of the proposed framework in a number of applications. For empirical comparisons, we refer to our approaches, block SPD OT~\eqref{reformulated_mw}, the corresponding Wasserstein barycenter~\eqref{w_barycenter}, and block SPD Gromov-Wasserstein OT~\eqref{scalar_gw_ot}~\&~\eqref{gw_barycenter_problem}, collectively as RMOT (Riemannian optimized Matrix Optimal Transport).
For all the experiments, we use the Riemannian steepest descent method using the Manopt toolbox \cite{boumal2014manopt} in Algorithm~\ref{Riemannian_optimizer}.
The code can be found on \url{https://github.com/andyjm3/BlockSPDOT}.
\begin{figure}[!th]
\captionsetup{justification=centering}
\centering
\subfloat[MNIST \label{mnist_fig}]{\includegraphics[width = 0.32\textwidth, height = 0.24\textwidth]{Figures/domain_adaptation/mnist_da_result.pdf}}
\subfloat[Fashion MNIST \label{fmnist_fig}]{\includegraphics[width = 0.32\textwidth, height = 0.24\textwidth]{Figures/domain_adaptation/fashionmnist_da_result.pdf}}
\subfloat[Letters \label{letters_fig}]{\includegraphics[width = 0.32\textwidth, height = 0.24\textwidth]{Figures/domain_adaptation/letters_da_result.pdf}}
\caption{Domain adaptation and classification results for three datasets: MNIST \eqref{mnist_fig}, Fashion MNIST \eqref{fmnist_fig} and Letters~\eqref{letters_fig}. The skew ratio increases from uniform (uf) to $r=0.5$. For MNIST and Fashion MNIST, uf $=0.1$ and for Letters, uf $=1/26$. We observe that the proposed RMOT performs significantly better than the baselines.
}
\label{domain_adaptation_figure}
\end{figure}
\subsection{Domain adaptation}
We apply our OT framework to the application of unsupervised domain adaptation where the goal is to align the distribution of the source with the target for subsequent tasks.
Suppose we are given the source ${\mathbf p} \in \Sigma_m$ and target marginals ${\mathbf q} \in \Sigma_n$, along with samples $\{ {\mathbf X}_i\}_{i=1}^m, \{ {\mathbf Y}_j\}_{j = 1}^n$ from the source and target distributions. The samples are matrix-valued, i.e., ${\mathbf X}_i ,{\mathbf Y}_j \in {\mathbb R}^{d \times s}$. We define the cost as $\mathbf C_{i,j} = ({\mathbf X}_i - {\mathbf Y}_j)({\mathbf X}_i -{\mathbf Y}_j)^\top$. It should be noted that $\mathrm{tr}{(\mathbf C_{i,j})}=\|{\mathbf X}_i - {\mathbf Y}_j\|_{\rm F}^2$, which is the cost function under the $2$-Wasserstein OT setting~(\ref{W22}).
For domain adaptation, we first learn an optimal coupling between the source and target samples by solving the proposed OT problem \eqref{reformulated_mw} with marginals ${\mathbf P}, {\mathbf Q}$ constructed as ${\mathbf P} \coloneqq \{[{\mathbf P}_i]_{m\times 1}:{\mathbf P}_i = p_i {\mathbf I}\}$ and ${\mathbf Q}\coloneqq\{[{\mathbf Q}_j]_{n\times 1}:{\mathbf Q}_j = q_i {\mathbf I}\}$ (as discussed in Section \ref{matrix_generalized_ot_sect}). Finally, the source samples are projected to the target domain via barycentric projection.
With the optimal coupling ${\mathbf \Gamma}_{i,j}^*$, the barycentric projection of a source sample ${\mathbf X}_i$ is computed as
\begin{equation*}
\hat{{\mathbf X}}_i = \argmin\limits_{{\mathbf X}_i \in {\mathbb R}^{d\times s}} \sum_{i,j} \mathrm{tr} ( ({\mathbf X}_i - {\mathbf Y}_j)({\mathbf X}_i - {\mathbf Y}_j)^\top {\mathbf \Gamma}_{i,j}^* ) = {\mathbf P}_i^{-1}(\sum_j {\mathbf \Gamma}_{i,j}^* {\mathbf Y}_j).
\end{equation*}
The above formulation also works for structured samples. For instance, when the samples are SPD, i.e., ${\mathbf X}_i, {\mathbf Y}_j \in {\mathbb S}_{++}^d$, the projected source sample $\hat{{\mathbf X}}_i$ is now the solution to the matrix Lyapunov equation: $\{ {\mathbf P}_i \hat{{\mathbf X}}_i \}_{\rm S} = \{\sum_j {\mathbf \Gamma}^*_{i,j} {\mathbf Y}_j\}_{\rm S}$. Here, $\{ \mathbf A\}_{\rm S} = (\mathbf A + \mathbf A^\top)/2$.
\paragraph{Experimental setup.}
We employ domain adaptation to classify the test sets (target) of multiclass image datasets, where the training sets (source) have a different class distribution than the test sets.
Suppose we are given a training set $\{ {\mathbf X}_i\}_{i=1}^m$ and a test set $\{ {\mathbf Y}_j\}_{j=1}^n$ where ${\mathbf X}_i, {\mathbf Y}_j \in {\mathbb R}^{d \times s}$ are $s$ (normalized) image samples of the same class in $d$ dimension for each image set $i,j$. Instead of constructing the cost directly on the input space, which are not permutation-invariant, we first compute the sample covariances $\mathbf S_{x_i} = {\mathbf X}_i {\mathbf X}_i^\top/s$ and $\mathbf S_{y_j} = {\mathbf Y}_j {\mathbf Y}_j^\top/s$, $\forall i,j$. Now the cost between $i, j$ is given by $\mathbf C_{i,j} = (\mathbf S_{x_i} - \mathbf S_{y_j})(\mathbf S_{x_i} - \mathbf S_{y_j})^\top$. Once the block SPD matrix coupling is learnt, the $\mathbf S_{x_i}$ covarinaces are projected using the barycerntric projection to obtain $\hat{\mathbf S}_{x_i}, i \in [m]$. This is followed by nearest neighbour classification of $j$ based on the Frobenius distance $\| \hat{\mathbf S}_{x_i} - \mathbf S_{y_j} \|_{\rm F}, \forall i,j$. We set $p_i = 1/m$ and $q_j = 1/n$.
We compare the proposed {RMOT} (\ref{reformulated_mw}) with the following baselines: (i) sOT: the $2$-Wasserstein OT~(\ref{W22}) with the cost $c_{i,j} = \mathrm{tr}(\mathbf C_{i,j})=\|\mathbf S_{x_i} - \mathbf S_{y_j} \|_{\rm F}^2$ \cite{courty2016optimal}, and (ii) SPDOT: the $2$-Wasserstein OT (\ref{W22}) with the cost as the squared Riemannian geodesic distance between the SPD matrices $\mathbf S_{x_i}$ and $\mathbf S_{y_j}$ \cite{yair2019parallel}.
\textbf{Datasets.} We experiment on three multiclass image datasets - handwritten letters \cite{frey1991letter}, MNIST \cite{lecun1998gradient} and Fashion MNIST \cite{xiao2017fashion} - with various skewed distributions for the training set. MNIST and Fashion MNIST have $10$ classes, while Letters has $26$ classes.
Specifically, we fix the distribution of the test set to be uniform (with the same number of image sets per class). We increase the proportion of the a randomly chosen class in the training set to the ratio $r$, where $r=\{{\rm uf},0.1,0.2,0.3,0.4,0.5\}$ and ${\rm uf}$ is the ratio corresponding to the uniform distribution of all classes. We reduce the dimension of MNIST, fashion MNIST, and Letters by PCA to $d=5$ features. We set $s=d$, $m=250$, and $n=100$ for each dataset.
\textbf{Results.} Figure~\ref{domain_adaptation_figure} shows the classification accuracy on the three datasets.
We observe that the proposed RMOT outperforms sOT and SPDOT, especially in more challenging domain adaptation settings, i.e., higher skew ratios. This shows the usefulness of the non-trivial correlations learned by the SPD matrix valued couplings of {RMOT}.
\subsection{Tensor Gromov-Wasserstein distance averaging for shape interpolation}
We consider an application of the proposed block SPD Gromov-Wasserstein OT formulation (Section~\ref{sec:proposedGW}) for interpolating tensor-valued shapes. We are given two distance-marginal pairs $(\mathbf D^0, {\mathbf P}^0), (\mathbf D^1, {\mathbf P}^1)$ where $\mathbf D^0, \mathbf D^1 \in {\mathbb R}^{n \times n}$ are distance matrices computed from the shapes and ${\mathbf P}^0, {\mathbf P}^1$ are given tensor fields. The aim is to interpolate between the distance matrices with weights ${\boldsymbol \omega} = (t, 1-t), t \in [0,1]$. The interpolated distance matrix $\mathbf D^t$ is computed by solving \eqref{gw_barycenter_problem} via Riemannian optimization and Proposition \ref{average_distance_update}, with the barycenter tensor fields ${{\mathbf P}}^t$ given. Finally, the shape is recovered by performing multi-dimensional scaling to the distance matrix.
Figure \ref{GW_distance_averaging} presents the interpolated shapes with $n = 100$ sample points for the input shapes. The matrices $\mathbf D^0, \mathbf D^1$ are given by the Euclidean distance and we consider $L_2$ loss for $\mathcal{L}$. The input tensor fields ${\mathbf P}^0, {\mathbf P}^1$ are generated as uniformly random in (a), cross-oriented in (b) and smoothly varying in (c). The barycenter tensor fields are given by the linear interpolation of the inputs, i.e., ${\mathbf P}^t = (1-t) {\mathbf P}^0 + t {\mathbf P}^1$. The results show the proposed Riemannian optimization approach (Section~\ref{manifold_geometry_sect}) converges to reasonable stationary solutions for non-convex OT problems.
\begin{figure}[ht]
\centering
\begin{tabular}{@{}c@{ \hspace*{1.5pt} }c@{ \hspace*{2pt} }c@{ \hspace*{2pt} }c@{ \hspace*{2pt} }c@{ \hspace*{2pt} }c@{ \hspace*{2pt} }c@{ \hspace*{2pt} }}
\raisebox{.025\textwidth}{\footnotesize (a)}&
\includegraphics[width=.09\textwidth]{Figures/cat-dog-gw/cat-dog-input-1.pdf}&
\includegraphics[width=.09\textwidth]{Figures/cat-dog-gw/bary-1.pdf}&
\includegraphics[width=.09\textwidth]{Figures/cat-dog-gw/bary-2.pdf}&
\includegraphics[width=.09\textwidth]{Figures/cat-dog-gw/bary-3.pdf}&
\includegraphics[width=.09\textwidth]{Figures/cat-dog-gw/bary-4.pdf}&
\includegraphics[width=.09\textwidth]{Figures/cat-dog-gw/cat-dog-input-2.pdf}\\[5pt]
\raisebox{.025\textwidth}{\footnotesize (b)}&
\includegraphics[width=.09\textwidth]{Figures/annulus-star4-gw/input-r-1.pdf}&
\includegraphics[width=.09\textwidth]{Figures/annulus-star4-gw/bary-r-1.pdf}&
\includegraphics[width=.09\textwidth]{Figures/annulus-star4-gw/bary-r-2.pdf}&
\includegraphics[width=.09\textwidth]{Figures/annulus-star4-gw/bary-r-3.pdf}&
\includegraphics[width=.09\textwidth]{Figures/annulus-star4-gw/bary-r-4.pdf}&
\includegraphics[width=.09\textwidth]{Figures/annulus-star4-gw/input-r-2.pdf}\\[5pt]
\raisebox{.025\textwidth}{\footnotesize (c)}&
\includegraphics[width=.09\textwidth]{Figures/disk-star5-r/disk-star5-input-r-1.pdf}&
\includegraphics[width=.09\textwidth]{Figures/disk-star5-r/bary-r-1.pdf}&
\includegraphics[width=.09\textwidth]{Figures/disk-star5-r/bary-r-2.pdf}&
\includegraphics[width=.09\textwidth]{Figures/disk-star5-r/bary-r-3.pdf}&
\includegraphics[width=.09\textwidth]{Figures/disk-star5-r/bary-r-4.pdf}&
\includegraphics[width=.09\textwidth]{Figures/disk-star5-r/disk-star5-input-r-2.pdf}\\[3pt]
&{\footnotesize $t=0$} & {\footnotesize $t=0.2$} & {\footnotesize $t=0.4$} & {\footnotesize $t=0.6$} & {\footnotesize $t=0.8$} & {\footnotesize $t=1$} \\
\end{tabular}
\caption{Tensor-valued shape interpolation obtained using the proposed RMOT~(Section~\ref{sec:proposedGW}). }%
\label{GW_distance_averaging}
\end{figure}
\begin{figure*}[!th]
\captionsetup{justification=centering}
\centering
\subfloat[Linear interpolation]{\includegraphics[width = 0.19\textwidth, height = 0.12\textwidth]{Figures/dirac-pairs-smooth/linear-interp.pdf}}
\hspace*{3pt}
\subfloat[QOT ($\rho = 50$)]{\includegraphics[width = 0.19\textwidth, height = 0.12\textwidth]{Figures/dirac-pairs-smooth/QOT-ellipses_50.pdf}}
\hspace*{3pt}
\subfloat[QOT ($\rho = 100$)]{\includegraphics[width = 0.19\textwidth, height = 0.12\textwidth]{Figures/dirac-pairs-smooth/QOT-ellipses_100.pdf}}
\hspace*{3pt}
\subfloat[QOT ($\rho = 500$)]{\includegraphics[width = 0.19\textwidth, height = 0.12\textwidth]{Figures/dirac-pairs-smooth/QOT-ellipses_500.pdf}}
\hspace*{3pt}
\subfloat[{RMOT}]{\includegraphics[width = 0.19\textwidth, height = 0.12\textwidth]{Figures/dirac-pairs-smooth/BlockMOT-ellipses.pdf}} \\[5pt]
\subfloat[Input ($t = 0$) \label{input_1}]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/input-ellispses-1.pdf}}
\hspace*{0.5pt}
\subfloat[Input ($t=1$) \label{input_2}]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/input-ellispses-2.pdf}}
\hspace*{0.5pt}
\subfloat[QOT ($t=0.25$)]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/interpol-ellispses-2.pdf}}
\hspace*{0.5pt}
\subfloat[QOT ($t=0.5$)]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/interpol-ellispses-3.pdf}}
\hspace*{0.5pt}
\subfloat[QOT ($t=0.75$)]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/interpol-ellispses-4.pdf}}
\hspace*{0.5pt}
\subfloat[{RMOT} ($t=0.25$)]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/BlockMOT-interpol-ellispses-2.pdf}}
\hspace*{0.5pt}
\subfloat[{RMOT} ($t=0.5$)]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/BlockMOT-interpol-ellispses-3.pdf}}
\hspace*{0.5pt}
\subfloat[{RMOT} ($t=0.75$)]{\includegraphics[width = 0.12\textwidth, height = 0.12\textwidth]{Figures/2d-aniso-fields/BlockMOT-interpol-ellispses-4.pdf}}
\caption{Tensor field mass interpolation on 1-$d$ (top) and 2-$d$ (bottom) grids. On the top, each row corresponds to an interpolation where we show $7$ evenly-spaced interpolated tensor fields. On the bottom, the inputs are given in (f) and (g). We set $\rho = 100$ for QOT and show three evenly-spaced interpolated tensor fields.}
\label{interpolation_figure}
\end{figure*}
\subsection{Tensor field optimal transport mass interpolation}
We consider performing optimal transport and displacement interpolation between two tensor fields supported on regular 1-$d$ (or 2-$d$) grids~\cite{peyre2016quantum}.
We consider a common domain ${\mathcal{D}} = [0,1]$ (or $[0,1]^2$) with the cost defined as $\mathbf C_{i,j} = \| {\mathbf x}_i - {\mathbf y}_j \|^2 {\mathbf I}$ for ${\mathbf x}_i, {\mathbf y}_j \in {\mathcal{D}}$. The marginals ${\mathbf P}, {\mathbf Q}$ are given tensor fields. We first compute the balanced coupling ${\mathbf \Gamma}$ by solving an entropy regularized OT problem~(\ref{eqn:regMWOT}):
\begin{equation*}
\min_{{\mathbf \Gamma} \in \mathcal{M}_{m \times n}^d({\mathbf P}, {\mathbf Q})} \sum_{i,j} \Big( \mathrm{tr}(\mathbf C_{i,j}{\mathbf \Gamma}_{i,j}) - \epsilon \, H({\mathbf \Gamma}_{i,j}) \Big),
\end{equation*}
where the quantum entropy is defined as $H({\mathbf \Gamma}_{i,j}) \coloneqq - \mathrm{tr}({\mathbf \Gamma}_{i,j} \log({\mathbf \Gamma}_{i,j}) - {\mathbf \Gamma}_{i,j})$.
Then, the coupling is used to interpolate between the two tensor fields by generalizing the displacement interpolation \cite{mccann1997convexity} to SPD-valued marginals. Please refer to~\cite[Section~2.2]{peyre2016quantum} for more details. It should be noted that due to the balanced nature of our formulation, we do not need to adjust the couplings after matching as required in~\cite{peyre2016quantum}.
We compare interpolation results of the proposed (balanced) {RMOT} with both linear interpolation $(1-t){\mathbf P} + t {\mathbf Q}$ for $t \in [0,1]$ and the \textit{unbalanced} quantum OT (QOT) of \cite{peyre2016quantum}. The QOT solves the following problem with quantum KL regularization, i.e.,
\begin{equation*}
\begin{array}{ll}
\min\limits_{{\mathbf \Gamma}} \sum_{i,j} \Big( \mathrm{tr}(\mathbf C_{i,j} {\mathbf \Gamma}_{i,j}) - \epsilon \, H({\mathbf \Gamma}_{i,j}) + \rho \, {\rm KL}({\mathbf \Gamma} {\mathbbm{1}} | {\mathbf P}) + \rho \,{\rm KL}({\mathbf \Gamma}^\top {\mathbbm{1}} | {\mathbf Q}) \Big),
\end{array}
\end{equation*}
where ${\rm KL}({\mathbf P} | {\mathbf Q}) \coloneqq \sum_i \mathrm{tr} \big({\mathbf P}_i \log({\mathbf P}_i) - {\mathbf P}_i \log({\mathbf Q}_i) -{\mathbf P}_i +{\mathbf Q}_i \big)$ and ${\mathbf \Gamma} {\mathbbm{1}} \coloneqq [\sum_j ({\mathbf \Gamma}_{i,j})]_{m \times 1}$ and ${\mathbf \Gamma}^\top {\mathbbm{1}} \coloneqq [\sum_i ({\mathbf \Gamma}_{i,j})]_{n \times 1}$. For comparability, we set the same $\epsilon$ for both QOT and {RMOT}.
Figure \ref{interpolation_figure} compares the mass interpolation for both 1-$d$ (top) and 2-$d$ (bottom) grids.
For the 2-$d$ tensor fields, we further renders the tensor fields via a background texture where we perform anisotropic smoothing determined by the tensor direction.
In both the settings, we observe that the tensor fields generated from {RMOT} respect the marginal constraints more closely.
\subsection{Tensor field Wasserstein barycenter}
We also analyze the Wasserstein barycenters learned by the proposed RMOT approach and qualitatively compare with QOT barycenter \cite[Section 4.1]{peyre2016quantum}. We test on two tensor fields ($n = 4$) supported 2-$d$ grids.
Figure~\ref{barycenter_fig} compares barycenter from QOT (top) and {RMOT} (bottom) initialized from the normalized solution of QOT. We observe that the QOT solution is not optimal when the marginal constraint is enforced and the barycenter obtained does not lie in the simplex of tensors.
\begin{figure}[!t]
\captionsetup{justification=centering}
\centering
\subfloat{
\begin{minipage}[c][0.12\textwidth]{
0.12\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/input-1.pdf}
\end{minipage}}
\subfloat{
\begin{minipage}[c][0.12\textwidth]{
0.12\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/QOTbarycenter-2-5.pdf} \\[8pt]
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/MOTbarycenter-qot-2-5.pdf}
\end{minipage}}
\subfloat{
\begin{minipage}[c][0.12\textwidth]{
0.12\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/QOTbarycenter-3-5.pdf} \\[8pt]
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/MOTbarycenter-qot-3-5.pdf}
\end{minipage}}
\subfloat{
\begin{minipage}[c][0.12\textwidth]{
0.12\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/QOTbarycenter-4-5.pdf} \\[8pt]
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/MOTbarycenter-qot-4-5.pdf}
\end{minipage}}
\subfloat{
\begin{minipage}[c][0.12\textwidth]{
0.12\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{Figures/2d-bary-n2/input-2.pdf}
\end{minipage}}
\caption{Barycenter interpolation. From left to right $t = 0$ (input), $t= 0.25, 0.5, 0.75$ (barycenters), $t=1$ (input). The top row is QOT and the bottom is {RMOT}.}
\label{barycenter_fig}
\end{figure}
\section{Conclusion}
In this paper, we have discussed the balanced optimal transport (OT) problem involving SPD matrix-valued measures that is shown to be a natural generalization of $2$-Wasserstein formulation. It also preserves the metric property. For the SPD matrix-valued OT problem, the coupling matrix is a block matrix where each block is a symmetric positive definite matrix. The set of such coupling matrices can be endowed with Riemannian geometry, which enables optimization of both linear and non-linear objective functions. We have also shown how the SPD-valued OT setup extend many optimal transport problems to general SPD-valued marginals, including the Wasserstein barycenter and the Gromov-Wasserstein (GW) discrepancy. Experiments in a number of applications confirm the benefit of our approach.
In the appendix Section \ref{sec:appendix_projection}, we discuss the projected version of the SPD-valued OT problem. The robust version is shown to be a metric. As a future research direction, it would be interesting to explore such other properties. Another direction could be on learning of block coupling matrices with different constraints. In Section \ref{trace_mfd_appendix}, we discuss a constraint involving the matrix trace.
\bibliographystyle{amsalpha}
\input{main.bbl}
\newpage
|
\section{\label{s1}Introduction}
Colloidal suspensions are coated on substrates and upon drying produce various functional materials,
whose quality is determined by the microstructure composed of the colloidal particles.
During drying of colloidal suspensions,
the receding free surface (the liquid--air interface) induces the particles to form structures,
thereby causing a decrease in the drying rate.~\cite{dryc1, dryc2, dryc3}
The prediction and control of such drying characteristics is therefore required to improve material quality as well as to reduce drying time.
Figure~\ref{f1} schematically illustrates the structure formation corresponding to the drying characteristics of colloidal suspensions.
During the constant rate period, the drying rate is controlled by vapor diffusion from the free surface where the liquid evaporates.
The receding free surface sweeps the particles to form concentrated particle layers,
and then the falling rate period begins.
The condition that this structure formation occurs can be quantified by the particle drying P\'{e}clet number as $\mathrm{Pe} = \tau_\mathrm{D} / \tau_\mathrm{E} > 1$,
where $\tau_\mathrm{D}$ and $\tau_\mathrm{E}$ are the time scales of particle diffusion by the Brownian motion and the recession of the free surface, respectively.~\cite{clyr1, clyr2, sdry1}
The continuous evaporation requires liquid transport through the pores of the particle layer to the free surface.
With the growth of the particle layer, the permeation resistance of the particle layer increases.
The drying rate is thus controlled by the liquid transport through the particle layer and decreases gradually.
Such decrease in the drying rate has been quantitatively observed in experiments of the unidirectional drying of colloidal suspensions confined in a thin rectangular cell.
In these experiments, the drying rate can be evaluated by observing the decrease in the volume of the suspension in the drying cell.~\cite{udry1, udry2, udry3, udry4, udry5, udry6}
When the particle layer recedes with the free surface and grows to reach the substrate with further drying,
the free surface goes into the particle layer.
The drying rate is then controlled by vapor diffusion through the particle layer and will further decrease,
as observed for granular materials.
In this stage, liquid flow can be driven by a capillary pressure due to menisci formed among particles.
This liquid transport would suppress the decrease in the drying rate.~\cite{pors1, pors2}
We focus on the decrease in the drying rate due to the growth of concentrated particle layers.
The character of colloidal particles would affect the structure of the particle layers.
For example, in the drying of droplets of colloidal suspensions,
the morphology of the dried grains changes depending on interactions between the particles.~\cite{sdry2, sdry3}
As for the drying of binary colloidal mixtures containing particles of different sizes,
segregation of smaller particles to the top surface can occur depending on the mixing ratio, the particle size ratio, and drying rates.~\cite{seg1, seg2, seg3}
Such structural differences would be reflected in the drying characteristics.
In this study, we construct a model to investigate how the drying characteristics of colloidal suspensions is affected by interactions between the colloidal particles.
We first derive an analytical expression of the relationship between the drying rate and the permeation resistance of concentrated particle layers.
We then perform Langevin dynamics simulations that describe the Brownian motion of colloidal particles by stochastic differential equations.
The receding free surface is also considered in the simulations to induce the formation of concentrated particle layers.
During the simulations, the drying rate is evaluated from the permeation resistance of the particle layers by use of the derived analytical expression.
We assume the DLVO potential between charged colloidal particles and tune the potential from repulsive to attractive by varying ionic strength.
We show that the decrease in the drying rate is suppressed when the potential energy barrier is so low that the particles form aggregates.
We discuss how the interaction between particles affect the drying characteristics in terms of the structural differences of the particle layers.
\begin{figure}
\centering
\includegraphics[width=8cm]{fig1}
\caption{
Sketch of the drying characteristics of colloidal suspensions.
}
\label{f1}
\end{figure}
\section{\label{s2}Model}
\subsection{\label{s2-1}Drying rates}
\begin{figure}
\centering
\includegraphics[height=6cm]{fig2}
\caption{
Drying rate as a function of permeation resistance at different humidities.
The profiles are given by Eq.~(\ref{e2-1-4}) for $\varphi < 1$
and Eq.~(\ref{e2-1-11}) for $\varphi \rightarrow 1$.
}
\label{f2}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=5.5cm]{fig3}
\caption{
Characteristic thickness of concentrated particle layers as a function of particle diameter.
The values scaled by $1 \ \mathrm{mm}$ and the particle diameter $d$ are exhibited in the left and right axes, respectively.
The characteristic thickness $L_0 = R_0 / \rho_\mathrm{c}$ is calculated by Eqs.~(\ref{e2-1-7}) and (\ref{e2-1-12}) with setting the parameters:
$\eta = 1 \times 10^{-3} \ \mathrm{Pa \ s}$, $T = 300 \ \mathrm{K}$, $V_\mathrm{m} = 3 \times 10^{-29} \ \mathrm{m^3}$,
$p_0 = 3.5 \times 10^{3} \ \mathrm{Pa}$, $K = 8 \times 10^{-2} \ \mathrm{m/s}$, and $\Phi_\mathrm{c} = 0.64$.
From Eq.~(\ref{e2-1-6}), this parameter setting corresponds to $U_0 = 1 \times 10^{-6} \ \mathrm{m/s}$ for $\varphi = 0.5$.
}
\label{f3}
\end{figure}
The drying rate $U$ represents the decrease rate of liquid due to evaporation
and equals the mass transfer rate of the vapor from the free surface as
\begin{align}
U = \frac{KV_\mathrm{m}}{k_\mathrm{B} T}(p - p_\infty)
\label{e2-1-1},
\end{align}
where $K$ is the mass transfer coefficient of the vapor, $V_\mathrm{m}$ is the molecular volume of the liquid,
$T$ is the thermodynamic temperature, $k_\mathrm{B}$ is the Boltzmann constant.
Equation~(\ref{e2-1-1}) describes the diffusion of the vapor from the free surface to the ambient air stream.
The stagnant region in vicinity to the free surface is considered as the boundary layer
where the vapor diffuses.
Based on Fick's law,
the mass transfer rate is proportional to the difference in the vapor pressure of saturation $p$ and that of the ambient air $p_\infty$.
We then consider the drying of colloidal suspensions
where concentrated particle layers form as illustrated in Fig.~\ref{f1}.
Evaporation continues by liquid transport through the particle layer to the free surface.
This transport flow is originated by capillarity due to the curvature of the menisci formed among particles.
The permeation resistance of the particle layer induces a pressure drop $\Delta P$,
which is balanced by the capillary pressure.
Due to the pressure drop,
the vapor pressure at the free surface decreases from $p_0$ as
\begin{align}
p = p_0 \exp \left( -\frac{V_\mathrm{m} \Delta P}{k_\mathrm{B}T} \right)
\label{e2-1-2},
\end{align}
where $V_\mathrm{m} \Delta P$ gives the chemical potential drop of the liquid.
Equation~(\ref{e2-1-2}) is known as the Kelvin equation.
Since the liquid flow rate equals the drying rate due to mass balance,
the pressure drop is related to the drying rate by Darcy's law as
\begin{align}
\Delta P = \eta R U
\label{e2-1-3},
\end{align}
where $R$ is the permeation resistance of the particle layer, and $\eta$ is the liquid viscosity.
Combining Eqs.~(\ref{e2-1-1}) -- (\ref{e2-1-3}) yields an equation that relates the drying rate to the permeation resistance as
\begin{align}
\tilde{U} = \frac{e^{-(1 - \varphi) \tilde{R} \tilde{U}} - \varphi}{1 - \varphi}
\label{e2-1-4},
\end{align}
where the relative humidity is given by
\begin{align}
\varphi = \frac{p_\infty}{p_0}
\label{e2-1-5}.
\end{align}
In Eq.~(\ref{e2-1-4}), the drying rate and the permeation resistance are given as dimensionless quantities.
The drying rate $U$ is scaled by that without particle layers $U_0$ as
\begin{align}
\tilde{U} = \frac{U}{U_0}, \ \ \ \ \
U_0 = \frac{K V_\mathrm{m}}{k_\mathrm{B} T}(p_0 - p_\infty)
\label{e2-1-6}.
\end{align}
The permeation resistance $R$ is scaled by a characteristic resistance $R_0$ defined as
\begin{align}
\tilde{R} = \frac{R}{R_0}, \ \ \ \ \
R_0 = \frac{(k_\mathrm{B} T)^2}{\eta K V_\mathrm{m}^2 p_0}
= \frac{1}{\eta U_0} \frac{k_\mathrm{B} T}{V_\mathrm{m}}(1 - \varphi)
\label{e2-1-7}.
\end{align}
As compared to Eq.~(\ref{e2-1-3}),
the definition of $R_0$ indicates that
the flow rate through the particle layer with a permeation resistance of $R_0$ equals $U_0$
when a pressure of $(k_\mathrm{B} T/V_\mathrm{m}) (1 - \varphi)$ is applied.
In the previous study, Eq.~(\ref{e2-1-4}) was derived with ignoring humidity as $\varphi = 0$
and was approximately solved with assuming $V_\mathrm{m} \Delta P/ (k_\mathrm{B}T) \ll 1$,~\cite{udry5}
although this assumption only holds for $\tilde{R} \ll 1$.
The solution of Eq.~(\ref{e2-1-4}) is expressed by the Lambert $W$ function that satisfies $x = W(x e^x)$ as
\begin{align}
\tilde{U} = \frac{1}{1 - \varphi} \left[ \frac{1}{\tilde{R}} W(\tilde{R} e^{\varphi \tilde{R}}) - \varphi \right]
\label{e2-1-8}.
\end{align}
As described in Fig.~\ref{f2},
the drying rate decreases with an increase in the permeation resistance.
The asymptotic behavior is different depending on the humidity.
The asymptotic form for $\varphi > 0$ decreases proportional to $\tilde{R}^{-1}$ as
\begin{align}
\tilde{U} \sim -\frac{\ln \varphi}{1 - \varphi} \frac{1}{\tilde{R}}
\ \ \ \ (\tilde{R} \rightarrow \infty)
\label{e2-1-9},
\end{align}
while that for $\varphi = 0$ decreases more slowly as
\begin{align}
\tilde{U} \sim \frac{\ln \tilde{R}}{\tilde{R}}
\ \ \ \ (\tilde{R} \rightarrow \infty)
\label{e2-1-10}.
\end{align}
Combining Eqs.~(\ref{e2-1-3}) and (\ref{e2-1-9}) shows that
the pressure drop $\Delta P$ approaches to $-(k_\mathrm{B} T/V_\mathrm{m}) \ln \varphi$ as $\tilde{R} \rightarrow \infty$.
If this limiting pressure drop is larger than the maximum capillary pressure achieved by the maximum meniscus curvature,
the free surface will invade into the particle layer during drying.~\cite{cap}
Equation~(\ref{e2-1-8}) becomes a simple expression in the limit of $\varphi \rightarrow 1$ as
\begin{align}
\lim_{\varphi \rightarrow 1} \tilde{U} = \frac{1}{1 + \tilde{R}}
\label{e2-1-11}.
\end{align}
This expression indicates that the drying rate equals the rate of the liquid flow through the series resistance of $R$ and $R_0$
when a pressure of $(k_\mathrm{B} T/V_\mathrm{m}) (1 - \varphi)$ is applied.
From this expression, the characteristic resistance $R_0$ can be interpreted as the resistance of vapor transport in the boundary layer.
As noted in \ref{sA-2},
Eq.~(\ref{e2-1-11}) has the same form as the Ruth's filtration equation that describes the temporal variation of the flow rate in dead-end filtration.
The Ruth's filtration equation consider the series permeation resistance of the filter media and the cake deposited onto the filter.~\cite{filt}
When we assume that the particle layer is spatially uniform,
the permeation resistance is proportional to the thickness of the layer $L$ as $R = \rho_\mathrm{c} L$
with a constant resistivity $\rho_\mathrm{c}$.
We define a characteristic thickness as $L_0 = R_0/\rho_\mathrm{c}$.
The drying rate will decrease significantly when the particle layer grows as $L > L_0$.
The permeation resistivity can be evaluated by the Kozeny--Carman equation given by
\begin{align}
\rho_\mathrm{c} = \frac{180}{d^2} \frac{\Phi^2_\mathrm{c}}{(1 - \Phi_\mathrm{c})^3}
\label{e2-1-12},
\end{align}
where $d$ is the particle diameter, and $\Phi_\mathrm{c}$ is the volume fraction of the particles in the particle layer.
With assuming the uniformity of the particle layer,
the present model gives an analytical expression of the drying characteristics as derived in \ref{sA-2}.
As shown in Fig.~\ref{f3}, the characteristic thickness $L_0$ increases proportional to $d^2$.
We thus expect that the decrease in the drying rate will not be obvious for large particles.
In fact, in the previous experiments of unidirectional drying in confined systems,
the decrease in the drying rate was observed only for the particles of $d \lesssim 40 \ \mathrm{nm}$.~\cite{udry1, udry2, udry3, udry4, udry5, udry6}
The condition that the drying rate will obviously decrease during drying can be estimated as
$(\Phi_0 / \Phi_\mathrm{c}) H_0 > L_0$,
where $\Phi_0$ and $H_0$ are the initial particle volume fraction and thickness of the coating film of the colloidal suspension, respectively.
In this estimation, $(\Phi_0 / \Phi_\mathrm{c}) H_0$ equals the particle layer thickness when all the particles in the suspension form the particle layer.
When the particles of $d = 10 \ \mathrm{nm}$ form a particle layer with $\Phi_\mathrm{c} = 0.64$ (random close packing~\cite{rcp}),
the characteristic thickness is $L_0 = 4.3 \ \mathrm{mm}$.
The decrease in the drying rate will be obvious when the initial film thickness satisfies $H_0 > 27.5 \ \mathrm{mm}$ for $\Phi_0 = 0.1$.
\subsection{\label{s2-2}Formation of concentrated particle layers}
We consider a suspension containing spherical colloidal particles of a diameter $d$ and a mass $m$.
The suspension is coated on a plane substrate where we set the $z$-coordinate along the vertical direction.
The coating film is bounded by two parallel planes: the top free surface at $z = H$ and the bottom substrate at $z = 0$.
We describe the Brownian motion of the particles by the Langevin equations
and solve them numerically.
For the $i$-th particle,
the time evolution of the velocity $\boldsymbol{v}_i$ and the position $\boldsymbol{r}_i$ is expressed by~\cite{snap1, seg2, snap2, snap3}
\begin{align}
m \frac{\mathrm{d} \boldsymbol{v}_i}{\mathrm{d} t} &= -\zeta \boldsymbol{v}_i + \boldsymbol{F}_i^\mathrm{R}
+ \sum_j (\boldsymbol{F}_{ij}^\mathrm{cnt} + \boldsymbol{F}_{ij}^\mathrm{DLVO})
+ \boldsymbol{F}_i^\mathrm{cpl}, \nonumber \\
\frac{\mathrm{d} \boldsymbol{r}_i}{\mathrm{d} t} &= \boldsymbol{v}_i
\label{e2-2-1}.
\end{align}
As the influence of the ambient liquid,
the hydrodynamic drag $-\zeta \boldsymbol{v}_i$ and
the random force $\boldsymbol{F}_i^\mathrm{R}$ due to thermal fluctuations are considered,
while hydrodynamic interactions among the particles are neglected.
The hydrodynamic drag is expressed by Stokes' law as $\zeta = 3 \pi \eta d$.
The three components of the random force are given as the stochastic variables obeying
independent Gaussian distributions satisfying
\begin{align}
\langle \boldsymbol{F}_i^\mathrm{R}(t) \rangle = \boldsymbol{0}, \ \ \ \ \
\langle \boldsymbol{F}_i^\mathrm{R}(t) \boldsymbol{F}_i^\mathrm{R}(0) \rangle
= 2 k_\mathrm{B} T \zeta \delta(t) \boldsymbol{I}
\label{e2-2-2}.
\end{align}
The forces $-\zeta \boldsymbol{v}_i$ and $\boldsymbol{F}_i^\mathrm{R}$ cause the Brownian motion of the particle,
whose diffusion coefficient is given by the Stokes--Einstein relation $D = k_\mathrm{B} T / \zeta$.
The contact force $\boldsymbol{F}_{ij}^\mathrm{cnt}$ describes collisions between the particles
and is given by the Voigt model with the Herzian contact theory:~\cite{Herz}
\begin{align}
\boldsymbol{F}_{ij}^\mathrm{cnt}
= \left[ \frac{1}{3} E^\ast d^{1/2} \delta_{ij}^{3/2}
- \Gamma (\boldsymbol{v}_i - \boldsymbol{v}_j) \cdot \hat{\boldsymbol{n}}_{ij} \right] \hat{\boldsymbol{n}}_{ij}
\label{e2-2-3}.
\end{align}
The summation on the index $j$ includes the force from the substrate as well as the other particles.
The contact force acts in the direction normal to the tangent plane at the contact point.
This direction is expressed by the unit vector
$\hat{\boldsymbol{n}}_{ij} = \boldsymbol{r}_{ij}/|\boldsymbol{r}_{ij}|$
with $\boldsymbol{r}_{ij} = \boldsymbol{r}_i - \boldsymbol{r}_j$.
The normal relative displacement of the contact point is given by
$\delta_{ij} = \max ( 0, d - |\boldsymbol{r}_{ij}| )$.
The longitudinal elastic modulus $E^\ast$ is given by the Young's modulus $E$ and the Poisson ratio $\nu$ as
$E^\ast = E/(1 - \nu^2)$.
We assume that the damping coefficient is related to the elastic modulus as
$\Gamma = \lambda (m E^\ast d^{1/2} \delta_{ij}^{1/2} )^{1/2}$
to describe partially inelastic collisions.~\cite{cnt}
The contact force from the substrate is expressed by Eq.~(\ref{e2-2-3})
with substitution of $E^\ast \rightarrow 2^{1/2}E^\ast$ and $\delta_{ij} = \max ( 0, d/2 - z_i )$,
where $z_i = \hat{\boldsymbol{z}} \cdot \boldsymbol{r}_i$ is the $z$-coordinate of the $i$-th particle.
When an adhesive force $F_\mathrm{a}$ acts between the particles,
there is an equilibrium displacement given by $\delta_0 = [ 3 F_\mathrm{a} /(E^\ast d^{1/2}) ]^{2/3}$.
Considering small oscillation around $\delta_0$,
the elastic force is approximated by a linear restoring force with a spring constant of $E^\ast(d \delta_0)^{1/2}/2$.
Critical damping is realized when the factor of the damping coefficient $\Gamma$ satisfies $\lambda = 1$,
which we set in the present study.
The DLVO force $\boldsymbol{F}_{ij}^\mathrm{DLVO}$ describes the force acting between charged particles in liquid.
This force is given by the DLVO potential $V$ as~\cite{dlvo}
\begin{align}
\boldsymbol{F}_{ij}^\mathrm{DLVO}
= \left. -\frac{\mathrm{d} V (h)}{\mathrm{d} h} \right|_{h = h_{ij}} \hat{\boldsymbol{n}}_{ij}
\label{e2-2-4},
\end{align}
\begin{align}
V(h)
= -\frac{Ad}{24 h}
+ \pi \varepsilon_\mathrm{r} \varepsilon_0 d \psi_\mathrm{eff}^2 e^{-\kappa h}
\label{e2-2-5},
\end{align}
where $h$ is the surface separation between the particles.
The DLVO potential is the combination of the van der Waals attraction (first term) and the electric double layer repulsion (second term).
The magnitude of the van der Waals attraction is provided by the Hamaker constant $A$.
The electric double layer repulsion is characterized by the effective surface potential $\psi_\mathrm{eff}$ and the Debye parameter $\kappa$,
and they are given by
\begin{align}
\psi_\mathrm{eff} = \frac{4 k_\mathrm{B} T}{Z \mathrm{e}} \tanh \left( \frac{Z \mathrm{e} \psi_0}{4 k_\mathrm{B} T} \right), \ \ \
\kappa = \left( \frac{2 N_\mathrm{A} \mathrm{e}^2 I}{\varepsilon_\mathrm{r} \varepsilon_0 k_\mathrm{B} T} \right)^{1/2}
\label{e2-2-6},
\end{align}
where $\mathrm{e}$ is the elementary charge, $N_\mathrm{A}$ is the Avogadro constant, $\varepsilon_0$ is the electric constant,
$\varepsilon_\mathrm{r}$ is the relative permittivity, $Z$ is the ion valence, and $\psi_0$ is the surface electric potential.
The expression of $\psi_\mathrm{eff}$ in Eq.~(\ref{e2-2-6}) is that for valence symmetric electrolytes.
In this case, the ionic strength $I$ equals the electrolyte concentration multiplied by a factor of $Z^2$.
The inverse of the Debye parameter $\kappa^{-1}$ serves as a measure of the thickness of the electric double layer.
Note that the expression of Eq.~(\ref{e2-2-5}) is the approximate form for $\kappa^{-1} \ll h \ll d$, and we use it for simplicity.
We introduce a cut off distance $h_\mathrm{c}$ within which the DLVO force becomes constant,
and hence the adhesive force between the particles is given by $F_\mathrm{a} = (\mathrm{d} V/\mathrm{d} h)_{h = h_\mathrm{c}}$.
The cut off distance can be interpreted as the atomic length scale where the continuum picture, which is assumed in Eq.~(\ref{e2-2-5}), breaks down.~\cite{dlvo}
We calculate the DLVO force with setting $h_{ij} = \max (h_\mathrm{c}, |\boldsymbol{r}_{ij}|-d)$.
The capillary force $\boldsymbol{F}_i^\mathrm{cpl}$ acts to restore the particles
to the position satisfying an equilibrium contact angle with the free surface.
When the particles prefer to be perfectly wet,
the capillary force is expressed as
\begin{align}
\boldsymbol{F}_i^\mathrm{cpl} = -4 \pi \gamma \delta_i^\mathrm{fs} \left( 1 - \frac{\delta_i^\mathrm{fs}}{d} \right)
\hat{\boldsymbol{z}}
\label{e2-2-7},
\end{align}
where $\gamma$ is the surface tension of the liquid.
The protruding length of the particle from the free surface is given by $\delta_i^\mathrm{fs} = \max [ 0, d/2 - (H - z_i) ]$.
To describe the evaporation of the liquid,
the free surface moves in the $-z$ direction with a rate of $U$.
The particles are swept by the free surface through the capillary force,
thereby forming concentrated particle layers.
Although the ion concentration and the temperature can change during drying in real systems,
we assume that they are uniformly constant
and that the DLVO potential does not change.
According to Eq.~(\ref{e2-1-1}), the drying rate $U$ changes with reflecting the formation of concentrated particle layers.
We define the particle layer as the cluster of the particles moving with the free surface.
The clustering particles are detected as the particles contacting each other, i.e. $\delta_{ij} < 0$.
When one of the particles that compose the cluster satisfies $\delta_i^\mathrm{fs} < 0$, we regard the cluster as the particle layer.
The permeation resistance of the particle layer is calculated by integrating the resistivity of the cross-section of the computational domain vertical to the $z$-axis:
\begin{align}
R = \int_0^H \rho(z) \mathrm{d}z, \ \ \ \
\rho(z) = \frac{80}{[D_\mathrm{H} (z)]^2} \frac{S}{S_\mathrm{f}(z)}
\label{e2-2-8},
\end{align}
where $S$ is the cross-sectional area of the computational domain, $S_\mathrm{f}$ is that of the liquid flow path,
and $D_\mathrm{H}$ is the hydraulic diameter.
Since the cross-section of the flow path through the particle layer is quite complex shape,
the resistivity is evaluated by regarding the flow path as a cylindrical channel with a diameter of $D_\mathrm{H}$.~\cite{resist}
The evaluation of the resistivity in Eq.~(\ref{e2-2-8}) is an analogy to
the resistivity of a cylindrical channel whose diameter is $D$, i.e. $32/D^2$, which is derived from the Hagen--Poiseuille equation.
The factor $S/S_\mathrm{f}$ appears to consider the variation of the flow rate inversely proportional to the cross-sectional area of the flow path.
The factor 80 reflects the channel tortuosity and is selected such that the Kozeny--Carman equation Eq.~(\ref{e2-1-12}) can be derived from Eq.~(\ref{e2-2-8})
as indicated below.
The hydraulic diameter $D_\mathrm{H}$ is given by the cross-sectional area of the flow path $S_\mathrm{f}$
and the wetted perimeter of the cross-section $L_\mathrm{f}$ as~\cite{resist}
\begin{align}
D_\mathrm{H} = \frac{4 S_\mathrm{f}}{L_\mathrm{f}}
\label{e2-2-9},
\end{align}
\begin{align}
S_\mathrm{f}(z) = S - \frac{\pi}{4} d^2 \sum_{i \in P} \left[ 1 - \frac{4 (z - z_i)^2}{d^2} \right] \Theta \left( 1 - \frac{2|z - z_i|}{d} \right)
\label{e2-2-10},
\end{align}
\begin{align}
L_\mathrm{f}(z) = \pi d \sum_{i \in P} \Theta \left( 1 - \frac{2|z - z_i|}{d} \right)
\label{e2-2-11},
\end{align}
where the Heaviside step function $\Theta$ appears to indicate the region where the particles occupy: $|z_i - z| \leq d/2$.
The index in the summation $i \in P$ indicates the particles that compose the particle layer.
Assuming the uniform distribution of the particles with a volume fraction of $\Phi_\mathrm{c}$,
we obtain $S_\mathrm{f} = (1 - \Phi_\mathrm{c})S$ and $L_\mathrm{f} = 6\Phi_\mathrm{c} S /d$,
which lead to the Kozeny--Carman equation given by Eq.~(\ref{e2-1-12}).
\section{Results and discussion}
\begin{table}
\caption{
Time scales in the motion of particles.
}
\label{t1}
\centering
\scalebox{0.9}[0.9]{
\renewcommand\arraystretch{1.4}
\begin{tabular}{p{2.9cm}p{3.8cm}l}
\bhline{1.0pt}
\multicolumn{2}{l}{\raisebox{0.em}{Time scale}} &
\multicolumn{1}{l}{\raisebox{0.em}{Ratio to $\tau_\mathrm{hyd}$}}
\\
\bhline{1.0pt}
\multicolumn{3}{l}{Migration (Characteristic length: $d$)}\\
\hline
Evaporation & $\tau_\mathrm{E} = d/U_0$ & $ \ 25$\\
Diffusion & $\tau_\mathrm{D} = d^2/D$ & $ \ 1.0 \times 10^4 $\\
\hline
\multicolumn{3}{l}{Velocity change} \\
\hline
Hydrodynamic drag & $\tau_\mathrm{hyd} = m/\zeta$ & $ \ 1$\\
Contact force & $\tau_\mathrm{cnt} = (m/E^\ast)^{1/2} (d \delta_0)^{-1/4} $ & $ \ 1.0 \times 10^{-1}$\\
Capillary force & $\tau_\mathrm{cpl} = (m/\gamma)^{1/2}$ & $ \ 3.1 \times 10^{-1}$\\
\bhline{1.0pt}
\end{tabular}
}
\end{table}
\begin{table}
\caption{
Dimensionless Debye parameter $\kappa d$ and potential barrier $\Delta E$ of DLVO potentials
for different ionic strengths.
}
\label{t2}
\centering
\scalebox{0.9}[0.9]{
\renewcommand\arraystretch{1.4}
\begin{tabular}{p{1.8cm}rr}
\bhline{1.0pt}
$I / (\mathrm{mol/L})$ & $\ \ \kappa d \ \ \ $ & $\Delta E/ (k_\mathrm{B} T) $\\
\bhline{1.0pt}
\ 0.01 & 6.5 \ \ \ & $ 15.8 \ \ \ $\\
\ 0.03 & 11.3 \ \ \ & $ 13.7 \ \ \ $\\
\ 0.1 & 20.5 \ \ \ & $ 10.7 \ \ \ $\\
\ 0.3 & 35.6 \ \ \ & $ 7.5 \ \ \ $\\
\ 1 & 64.9 \ \ \ & $ 3.5 \ \ \ $\\
\bhline{1.0pt}
\end{tabular}
}
\end{table}
\begin{figure}
\centering
\includegraphics[width=7cm]{fig4}
\caption{
DLVO potentials for different ionic strengths, as functions of surface separation between particles.
}
\label{f4}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8cm]{fig5}
\caption{
Potential barrier (left axis) and the average contact number of particles for the initial configurations (right axis),
as functions of ionic strength.
Corresponding snapshots of the initial configurations are shown in upper row,
where the contact number of the particles is indicated by their color.
}
\label{f5}
\end{figure}
Numerical simulations are performed with varying the DLVO potential.
We investigate the effects of the interaction between the particles on the drying characteristics.
To describe a part of the coating film far from edges,
periodic boundary conditions are applied in the directions parallel to the substrate with a side length of $18d$.
In the initial coating film of the thickness of $50d$, the volume fraction of the particles is set as $\Phi_0 = 0.1$.
The particle motion described by Eq.~(\ref{e2-2-1}) contains multiple time scales as listed on Table~\ref{t1}.
The time integration of Eq.~(\ref{e2-2-1}) is performed by use of the velocity Verlet algorithm.
The time increment is set as $0.2\tau_\mathrm{cnt}$ to be smaller than the shortest time scale in Table~\ref{t1}, i.e. $\tau_\mathrm{cnt}$.
The drying P\'{e}clet number is $\mathrm{Pe} = \tau_\mathrm{D}/\tau_\mathrm{E} = 400$,
which corresponds to $U_0 = 0.44 \ \mathrm{m/s}$ for an aqueous suspension with
$d = 20 \ \mathrm{nm}$, $\eta = 1 \times 10^{-3} \ \mathrm{Pa \ s}$, and $T = 300 \ \mathrm{K}$.
This drying rate is larger by a factor of about $10^6$ than actual situations.
The drying P\'{e}clet number represents also the ratio of the particle diameter to the diffusion length of the particles concentrated by the receding free surface,~\cite{clyr1,seg3}
i.e. $\mathrm{Pe} = d/(D/U_0)$.
The current setting of $\mathrm{Pe} $ corresponds to reducing the length scale in the drying direction to perform simulations in smaller computational domains.
We correspondingly set the characteristic thickness given in Sec.~\ref{s2-1} as $L_0 / d = 2$,
which is about $10^{-6}$ times smaller than actual situations as demonstrated in Fig.~\ref{f3}.
The humidity is set as $\varphi = 0.5$.
The dimensionless parameters of the DLVO potential are set as
$A/(24 k_\mathrm{B} T) = 0.1$ and $\pi \varepsilon_\mathrm{r} \varepsilon_0 d \Psi_\mathrm{eff}^2 /(k_\mathrm{B} T) = 23$.
This setting corresponds to considering the silica particles in water with
$d = 20 \ \mathrm{nm}$, $T = 300 \ \mathrm{K}$, $\psi_0 = -50 \ \mathrm{mV}$, and $A = 1 \times 10^{-20} \ \mathrm{J}$.~\cite{hamak}
As shown in Table~\ref{t2} and Fig.~\ref{f4}, we change the other dimensionless parameter $\kappa d$ by changing the ionic strength according to Eq.~(\ref{e2-2-6}).
We define the potential barrier $\Delta E$ as the difference between the maximum and the local minimum (potential well) of the DLVO potential.
The potential barrier decreases with an increase in the ionic strength.
The cut off length $h_\mathrm{c}$ is set to satisfy $\delta_0 /d = 0.02$.
In the present simulations, we consider the DLVO force only between the particles with the separation of $h/d \leq 1$,
where the DLVO potential sufficiently decays since all the settings in Table~\ref{t2} satisfies $(\kappa d)^{-1} < 1$.
The initial configurations exhibited in Fig.~\ref{f5} are obtained as follows:
the particles are firstly distributed randomly so that the surface separations between the particles become more than that of the potential maximum,
and then the Brownian motion of the particles are solved during a time of $5 \tau_\mathrm{D}$.
The particles aggregate at $I \geq 0.3 \ \mathrm{mol/L}$ where $\Delta E/(k_\mathrm{B} T) \lesssim 10$ is satisfied.
The average contact number of the particles rises sharply accompanied by the aggregation.
\begin{figure}
\centering
\includegraphics[width=8cm]{fig6}
\caption{
Snapshots of the particle configurations during drying.
The ionic strengths are (a) $0.1 \ \mathrm{mol/L}$ and (b) $1 \ \mathrm{mol/L}$.
The particles composing the concentrated layers are indicated by the red color.
}
\label{f6}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=7cm]{fig7}
\caption{
Time variation of (a) coating film thickness and (b) drying rate for different ionic strengths.
The dotted line represents the time change of the film thickness containing no particles.
}
\label{f7}
\end{figure}
As shown in Fig.~\ref{f6},
the receding free surface sweeps the particles to form the concentrated particle layers.
Figure~\ref{f7} thus indicates that the drying rate decreases with the growth of the particle layers.
If there are no particles, the thickness of the liquid film decreases linearly with a constant drying rate.
The drying rate decreases most slowly at the highest ionic strength of $I = 1 \ \mathrm{mol/L}$ where the particles aggregate.
Comparing the cases of $I \leq 0.1 \ \mathrm{mol/L}$ where the particles are stably dispersed,
the drying rate is initially larger at the higher ionic strength of $I = 0.1\ \mathrm{mol/L}$,
while it finally becomes larger at the lower ionic strength of $I = 0.01\ \mathrm{mol/L}$.
This difference reflects the permeation resistance as discussed below.
\begin{figure}
\centering
\includegraphics[width=9cm]{fig8}
\caption{
Distributions of particle volume fraction at the times of $t/\tau_\mathrm{E} = 0, 5, 10, 30, 50$
for different ionic strengths.
The broken lines represent the position of the free surface.
}
\label{f8}
\end{figure}
Figure~\ref{f8} demonstrates the growth of the particle layers as the increase in the particle volume fraction beneath the free surface,
where the clear oscillatory profiles can be found except at $I = 1 \ \mathrm{mol/L}$.
This oscillatory profile indicates the stratification of the particles.
The DLVO potential barrier hinders the particles to contact each other,
thereby allowing the concentrated particles to rearrange to stratify.
When the potential barrier is too low so that the particles aggregate at $I = 1 \ \mathrm{mol/L}$,
the particles cannot easily rearrange and form loose particle layers,
thus the oscillatory profiles do not appear.
\begin{figure}
\centering
\includegraphics[width=7cm]{fig9}
\caption{
Time variation of (a) the permeation resistance and (b) the resistivity of concentrated particle layers for different ionic strengths.
The resistivity is scaled by that evaluated by the Kozeny--Carman equation Eq.~(\ref{e2-1-12}) with $\Phi_\mathrm{c} = 0.64$.
}
\label{f9}
\end{figure}
The permeation resistance increases with the growth of the concentrated particle layers as shown in Fig.~\ref{f9}(a).
The increase of the permeation resistance results in the decrease of the drying rate through Eq.~(\ref{e2-1-4}).
The larger permeation resistance is resulted from the larger particle volume fraction of the particle layers.
In the cases of $I \leq 0.1 \ \mathrm{mol/L}$,
the permeation resistance increases sharply at $t/\tau_\mathrm{E} \approx 5$,
which seems to correspond to the time when the first layer of the particle layer is formed.
In fact, this time is roughly evaluated as $t/\tau_\mathrm{E} = (\Phi_\mathrm{c} - \Phi_0)/\Phi_0$,
which equals $5.4$ with assuming $\Phi_\mathrm{c} = 0.64$.
This evaluation is derived from the particle mass balance of $(\Phi_\mathrm{c} - \Phi_0) d = \Phi_0 U_0 t$.
Comparing the cases of $I \leq 0.1 \ \mathrm{mol/L}$,
the permeation resistance is initially larger for the smaller ionic strength of $I = 0.01 \ \mathrm{mol/L}$.
In Fig.~\ref{f8}, the initial distributions at $t/\tau_\mathrm{E} = 0$ show that
the stronger repulsive interaction between the particles results in the larger particle volume fraction beneath the free surface,
and thus the permeation resistance becomes larger.
Such increased accumulation of particles near a wall by repulsive interactions between particles was also shown in the previous study.~\cite{wall}
However, the permeation resistance finally becomes larger at the larger ionic strength of $I = 0.1 \ \mathrm{mol/L}$,
corresponding to the larger particle volume fraction of the particle layer as demonstrated by the profiles at $t/\tau_\mathrm{E} \geq 10$ in Fig.~\ref{f8}.
This larger particle volume fraction can be attributed to the smaller potential barrier and potential range, which would allow the particles to be densely packed.
In Fig.~\ref{f9}(b), the average resistivity is evaluated as $\bar{\rho} = R / [2 (H - \langle z \rangle_P )]$,
where $\langle z \rangle_P$ is the mean $z$-coordinate of the particles that compose the particle layers,
and $2 (H - \langle z \rangle_P )$ gives an estimation of the thickness of the particle layers.
The resistivity also increases with time but converges in the cases of the stably dispersed particles at $I \leq 0.1 \ \mathrm{mol/L}$.
Although the particle volume fraction of the particle layer is not uniform especially near the free surface as shown in Fig.~\ref{f8},
the convergence of the resistivity suggests that the particle layer gradually become uniform with distance from the free surface.
However, such convergence is not observed at $I = 1 \ \mathrm{mol/L}$,
where the particle volume fraction of the particle layer continuously increases with time.
When we can observe a convergence of the resistivity in the Langevin dynamics simulations,
combining this evaluated resistivity and the continuum model given in \ref{sA-2}
would enable us to evaluate the drying characteristics of coating films whose thickness is larger than the simulation scale.
\begin{figure}
\centering
\includegraphics[width=8cm]{fig10}
\caption{
Drying rate and the particle volume fraction of concentrated particle layers at $t / \tau_\mathrm{E} = 50$, as functions of ionic strength.
}
\label{f10}
\end{figure}
Figure~\ref{f10} summarizes the effect of the interaction between the particles on the drying characteristics.
The particle volume fraction of the particle layer $\bar{\Phi}$ is evaluated
as the mean particle volume fraction in the region within the depth of $3d$ from the free surface, i.e. $[H-3d, H]$.
As long as the particles stably dispersed at $I \leq 0.1 \ \mathrm{mol/L}$,
decreasing the potential barrier by increasing the ionic strength results in
a little decrease in the drying rate with a little increase in the particle volume fraction of the particle layer.
On the other hand, when the particles aggregate at the higher ionic strength of $I > 0.1 \ \mathrm{mol/L}$,
we find obvious change that the particle layer becomes loose and the drying rate increases.
In the previous experimental study of drying of colloidal suspension droplets,
it was also indicated that the morphology of the dried grains can be changed
by forming aggregates, due to the decrease in the permeation resistance of the particle layers.~\cite{sdry3}
\section{Conclusions}
We constructed a model that can calculate the drying characteristics by use of Langevin dynamics simulations.
This model enables us to evaluate the drying characteristics of colloidal suspensions from the character of colloidal particles such as interactions.
The present calculation of this model shows that the formation of aggregates by attractive interactions suppresses the decrease in the drying rate.
This suppression is due to the loose structure of the concentrated particle layers.
The aggregation is thus preferable to reduce the drying time,
while the final dried structure would be loose and disordered.~\cite{snap3}
To improve the density and order of structures with reducing the drying time by aggregation,
further investigations considering the effects of morphology and strength of the aggregates would be required.
Note that the present finding will be true for filtration.
The aggregated particles are expected to form loose cake, thereby suppressing the decrease in the filtering flux.
Both filtration and drying are operations of solid--liquid separation where liquid is removed through an interface.
|
\section{Introduction}\label{sec-intro}
\subsection{Background: from toral automorphisms to finite automata}
In this paper, we are concerned with a generalization of the road (colouring) problem. That problem, posed by Adler-Goodwyn-Weiss in \cite{agw-77-tms} and solved by Trahtman in \cite{ant-09-ijm}, asked whether every strongly connected, aperiodic directed graph of constant out-degree is the underlying graph of a synchronizing deterministic finite automaton (DFA). Trahtman's road colouring theorem (\cref{thm-rct-aper} below) gives an affirmative answer, and \cite{bf-11-aaecc,bp-14-dam} give a generalization to periodic graphs (\cref{thm-prct}).
The motivation for the road problem comes from ergodic theory. Specifically, a weak form of the road colouring theorem was used to prove the main theorem of \cite{agw-77-tms}, which gives a criterion for measure-theoretic isomorphism of certain Markov chains. This was the first time the road problem was explicitly posed, although the real origin of the problem is the earlier paper \cite{aw-67-pnas}, which concerns the analogous isomorphism problem for hyperbolic automorphisms of the two-dimensional torus.
The present paper concerns a generalization of the road problem motivated by graph-theoretic invariants for a different, but related, isomorphism relation in ergodic theory. Specifically, Ashley-Marcus-Tuncel \cite{amt-95-etds} identified a graph-theoretic criterion for isomorphism of one-sided stationary Markov chains, implicit in \cite{bt-90-etds}, and gave a complete, effectively computable set of isomorphism invariants. They observed that a certain conjectural uniqueness result (the $O(G)$ conjecture, below in \S \ref{subsec-trans-stab-sync}) would, if proven, simplify the set of invariants, and proved the conjecture in a family of special cases. The conjecture reduces to a generalization of the road problem involving certain right-resolving graph homomorphisms (which we call \textit{right-resolvers}; see \S \ref{subsec-graph-homo-basic-defns}). For graphs of constant out-degree, these homomorphisms coincide with road colourings.
\subsection{Main ideas and contributions}
The main purpose of this paper is to present new results toward the $O(G)$ conjecture. There are two main ideas in the paper. The first idea concerns the \textit{stability relation} of a right-resolver, which was introduced in \cite{ckk-02-lncs} for DFAs or road colourings. For context, Kari \cite{jk-03-tcs} solved the road problem in the Eulerian case by finding, for a given graph, a road colouring with nontrivial (i.e. not merely diagonal) stability relation, then recursively finding a synchronizing road colouring of the strictly smaller quotient graph (in which the states are stability classes), and lifting it to the original graph. Trahtman's solution of the full road problem uses the same inductive strategy, paired with a more sophisticated technique for obtaining a colouring with a nontrivial stability relation.
By determining how the stability relation behaves with respect to composition of right-resolvers, we are able to study the aforementioned recursive lifting constructions more systematically, allowing us to apply them toward the $O(G)$ conjecture. In particular, we adapt Trahtman's proof of the road colouring theorem to cover a larger family of cases (\cref{thm-bfc-cyc-bunch}), in which the out-degrees of states are allowed to vary cyclically. To do so, we generalize, to the setting of right-resolvers, a sufficient condition for a road colouring to have a nontrivial stable pair, based on the idea of a function graph with a unique tallest tree. This condition is at the heart of all proofs of the road colouring theorem to date, and its importance has motivated detailed analysis \cite{ber-16-lncs}.
The second main idea in this paper is a graph property that we call \textit{bunchiness}, along with a weaker property called \textit{almost bunchiness}. Bunchy and almost bunchy graphs are characterized by the property that the right-resolvers they admit are unique up to automorphisms in a certain sense (\cref{prop-alm-bunchy-equiv}). We highlight the implicit role of bunchiness both in \cite{amt-95-etds} and in the road colouring literature, and prove the $O(G)$ conjecture for bunchy and almost bunchy graphs (\cref{thm-alm-og-exists}). Furthermore, we show that the fiber product of right-resolving homomorphisms satisfies a universal property (\cref{thm-univ-prop}) that further highlights the essential role of bunchy graphs.
Motivated by these results, we introduce a new conjecture, which we call the bunchy factor conjecture (see \S \ref{sec-conjec}), asserting essentially that the $O(G)$ conjecture can be proved using the stability approach that Trahtman used to prove the road colouring theorem, with bunchy graphs as the base of the recursion. Another way of articulating this conjecture is that the barrier to proving the $O(G)$ conjecture is our lack of a sufficiently general method of producing homomorphisms with nontrivial stability relation. The fact that the bunchy factor conjecture implies the $O(G)$ conjecture is made explicit in \cref{prop-bunchy-implies-o}, which relies primarily on \cref{prop-bunchy-same-o}, a uniqueness result that uses the universal property of the fiber product in an essential way.
\subsection{Organization of the paper}
In \S \ref{sec-graph-homo}, we recall and adapt standard material on graphs and homomorphisms. In \S \ref{sec-stab-sync}, we define the stability relation for a right-resolver, give its main structural properties, and relate it to synchronization. In \S \ref{sec-og-road}, we recall from \cite{amt-95-etds} the connection between the road problem and the $O(G)$ conjecture, and state our generalization of the road colouring theorem.
In \S \ref{sec-bunchy}, we introduce the concepts of bunchiness and almost bunchiness and present results involving them, including the $O(G)$ conjecture for bunchy and almost bunchy graphs and the universal property of the fiber product. In \S \ref{sec-conjec}, we pose the bunchy factor conjecture, which has several equivalent formulations, and discuss its relation to the $O(G)$ conjecture and the road problem. In \S \ref{sec-comput}, we give polynomial-time algorithms for construction and decision problems involving right-resolvers, and discuss the algorithmic implications of the $O(G)$ and bunchy factor conjectures.
The proofs of many results in \S \S \ref{sec-graph-homo}--\ref{sec-comput}, comprising the structural properties of graphs and right-resolvers, are deferred to \S \ref{sec-proofs}. The proof of our generalization of the road colouring theorem is deferred to \S \ref{sec-traht}.
\section{Graphs and graph homomorphisms}\label{sec-graph-homo}
\subsection{Basic definitions}\label{subsec-graph-homo-basic-defns}
We take all graphs to be finite and directed. A graph $G$ consists of a set $V(G)$ of states, or vertices, and a set $E(G)$ of edges, together with a pair of maps $s, t: E(G) \to V(G)$ giving the source and target of each edge. Loops (edges $e$ with $s(e) = t(e)$) and parallel edges (distinct edges $e,e'$ with $s(e) = s(e')$, $t(e) = t(e')$) are allowed. For $I \in V(G)$, we write $E_I(G) = s^{-1}(I)$ for the set of outgoing edges from $I$. We write $F(I) = t(E_I(G))$ for the set of follower states of $I$, and we write $E_{IJ}(G) = s^{-1}(I) \cap t^{-1}(J)$ for the set of edges from $I$ to $J$. We write $L(G)$ for the language of $G$, i.e. the set of finite edge paths in $G$, i.e. $e_1 e_2 \dots e_n$ where $t(e_i) = s(e_{i+1})$, $1 \leq i \leq n-1$. We also refer to elements of $L(G)$ as \textit{words}. The maps $s,t$ extend to $s,t: L(G) \to V(G)$ by $s(e_1 \dots e_n) = s(e_1)$, $t(e_1 \dots e_n) = t(e_n)$. We define $L_I(G) = \{ u \in L(G) \, | \, s(u) = I \}$, and $L_{IJ}(G) = \{ u \in L_I(G) \, | \, t(u) = J \}$. A \textit{cycle} in $G$ is a path $u \in L(G)$ with $s(u) = t(u)$, i.e. an element of $L_{II}(G)$ for some $I \in V(G)$.
A \textit{graph homomorphism} $\Phi: G \to H$ is a pair of maps $\Phi: E(G) \to E(H)$, $\partial \Phi: V(G) \to V(H)$ such that $s \circ \Phi = \partial \Phi \circ s$ and $t \circ \Phi = \partial \Phi \circ t$. If there is a surjective homomorphism from $G$ to $H$, then we say that $H$ is a \textit{factor} of $G$, and that $G$ is an \textit{extension} of $H$. Observe that every factor of a strongly connected graph is strongly connected. A graph homomorphism $\Phi: G \to H$ induces a map $L(G) \to L(H)$ (also written $\Phi$) in the obvious way.
A graph isomorphism is a homomorphism that is injective and surjective (i.e. on both edges and states), and an automorphism of a graph $G$ is an isomorphism from $G$ to itself. We denote the group of automorphisms of $G$ by $\mathrm{Aut}(G)$, and by $P(G)$ the (normal) subgroup of $\mathrm{Aut}(G)$ which acts trivially on states and permutes parallel edges. We generally identify isomorphic graphs, and use the symbol $=$ to denote isomorphism, except when discussing algorithms for deciding isomorphism, or confirming that the automorphism group of a given graph is trivial.
We now introduce the class of homomorphisms with which we are concerned.
\begin{defn}[right-resolver]\label{defn-rr}
Let $G,H$ be graphs. Let $\Phi: G \to H$ be a surjective graph homomorphism. We say that $\Phi$ is \textit{right-resolving}, or is a \textit{right-resolver}, if, for each $I \in V(G)$, the restriction $\Phi|_{E_I(G)} : E_I(G) \to E_{\partial \Phi(I)} (H)$ is a bijection. We denote the set of right-resolvers $G \to H$ by $\hom_R(G,H)$, and we write $H \leq_R G$ if $\hom_R(G,H) \neq \emptyset$.
\end{defn}
\begin{rmk}
The term ``right-resolving'' comes from symbolic dynamics, where the words are of primary importance and the actual graph is secondary. A graph homomorphism $\Phi: G \to H$ is right-resolving if and only if the associated map $\Phi: L(G) \to L(H)$ satisfies a certain condition on the symbols (edges) appearing to the right of a given symbol in a word. See \cite{amt-95-etds}, \S 8.2 for details if interested.
\end{rmk}
\begin{rmk}
The class of right-resolving graph homomorphisms is closed under composition. This reduces to the fact, applied to the outgoing edges from each state, that a composition of bijections is a bijection. This means that the relation $\leq_R$ is transitive, and since the graphs are finite, it is clearly antisymmetric, so it is indeed a partial order on the set of all graphs (really on the set of equivalence classes of graphs up to isomorphism).
\end{rmk}
The following lemma is evident but we state it explicitly for future reference.
\begin{lemma}\label{lemma-fiber-determ-img}
The image of a right-resolver is determined up to graph isomorphism by the partition of the domain into fibers. That is, if $H_1, H_2 \leq_R G$ via $\Phi_i \in \hom_R(G,H_i)$, and for any $I_1, I_2 \in V(G)$ we have $\partial \Phi_1(I_1) = \partial \Phi_1(I_2)$ if and only if $\partial \Phi_2(I_1) = \partial \Phi_2(I_2)$, then in fact the $H_i$ are isomorphic.
\end{lemma}
Note that the converse is not true: for a given $G,H$ with $H \leq_R G$, there may exist $\Phi, \Phi' \in \hom_R(G,H)$ with distinct partitions $\{ (\partial \Phi)^{-1}(I) \, | \, I \in V(H) \}$, $\{ (\partial \Phi')^{-1}(I) \, | \, I \in V(H) \}$ of $V(G)$. However, this cannot occur when $H$ is $\leq_R$-minimal:
\begin{thm}[\cite{amt-95-etds}, Theorem 3.2 and Corollary 3.3(a)]\label{thm-mg-sigma}
For any graph $G$, there exist a unique $\leq_R$-minimal graph $M(G) \leq_R G$ and a unique map $\Sigma_G : V(G) \to V(M(G))$ such that $\partial \Phi = \Sigma_G$ for any $\Phi \in \hom_R(G,M(G))$.
\end{thm}
The construction of $M(G)$ was first given in \cite{cg-70-acm}, though not in this notation. We discuss the proof of \cref{thm-mg-sigma} in \S \ref{subsec-right-air-mg}. The notion of $M(G)$, for a graph $G$, provides context for road colourings:
\begin{defn}[$M_D$ and road colourings]
For $D \geq 1$, let $M_D$ be the graph with a single state and $D$ self-loops. For a graph $G$ of constant out-degree $D$, a \textit{road colouring} of $G$ is a right-resolver $G \to M_D$.
\end{defn}
Note that each $M_D$ is $\leq_R$-minimal, and that $\hom_R(G,M_D)$ is nonempty if and only if $G$ has constant out-degree $D$, in which case $M(G) = M_D$.
\subsection{Subgraphs and connectedness}
A \textit{sink} in a graph $G$ is a state $I \in V(G)$ is a state with no outgoing edges, i.e. $F(I) = \emptyset$. We assume throughout that all graphs are sink-free; this is purely for convenience, as all of the results that do not require strong connectedness can be proved for graphs with sinks, with routine but tedious modifications to the proofs. We say that a graph $G$ is \textit{strongly connected}, or irreducible, if for any ordered pair $I, J \in V(G)$, there is a (directed) edge path in $G$ from $I$ to $J$, i.e. $L_{IJ}(G) \neq \emptyset$. Note that strongly connected graphs are sink-free. The \textit{period} of a strongly connected graph is the gcd of its cycle lengths.
A graph $H$ is a \textit{subgraph} of a graph $G$ if $E(H) \subseteq E(G)$, $V(H) \subseteq V(G)$, and the maps $s,t$ with respect to $H$ agree with their counterparts on $G$, restricted to $H$. An \textit{induced subgraph} of a graph $G$ is a subgraph $H$ such that $E_{IJ}(H) = E_{IJ}(G)$ for every $I,J \in V(G)$. A \textit{strong component} of a graph is a maximal strongly connected subgraph, i.e. a strongly connected subgraph that is not a proper subgraph of another strongly connected graph.
A \textit{principal subgraph} of a graph $G$ is a subgraph $H$ such that $E_I(H) = E_I(G)$ for every $I \in V(H)$. Note that every principal subgraph is induced. Note also that if $H$ is a principal subgraph of $G$, and $K$ is a principal subgraph of $H$, then $K$ is a principal subgraph of $G$. A \textit{principal component} is a strongly connected principal subgraph; note that the principal components are precisely the minimal principal subgraphs. In particular, any two principal components of a given graph have disjoint sets of states. The principal components of $G$ correspond to the sink states in the \textit{condensation} of $G$, which is the directed acyclic graph in which the states are the strong components, or maximal strongly connected subgraphs, of $G$, and with an edge $C_1 \to C_2$ in the derived graph if there is an edge $I_1 \to I_2$ for $I_i \in V(C_i)$ in $G$.
Let $G, H$ be graphs with $H \leq_R G$, and let $\Phi \in \hom_R(G,H)$. Let $K$ be a subgraph of $G$. Note, by the right-resolving property, that in order for $\Phi|_K: K \to H$ to be surjective, it is necessary and sufficient that $\partial \Phi|_{V(K)}: V(K) \to V(H)$ be surjective and that $K$ be a principal subgraph of $G$.
\begin{rmk}
The road problem and the $O(G)$ problem were both originally raised for strongly connected graphs, which is a natural restriction given the origins of both problems in the ergodic theory of stationary Markov chains. Moreover, strong connectedness is used in an important way in a lemma used to prove both the road colouring theorem and the almost bunchy case of the bunchy factor conjecture. This is why the $O(G)$ conjecture and the bunchy factor conjecture are stated only for strongly connected graphs.
However, it is quite natural from an automata-theoretic perspective, especially concerning computational complexity, to consider graphs that are not strongly connected. For instance, Eppstein \cite{de-90-siam} shows that it is NP-complete to determine the minimum length of a synchronizing word for a given synchronizing DFA, and his examples are not strongly connected; there are also several graph problems from symbolic dynamics \cite{fc-21-arxiv} that are NP-complete in general but have polynomial-time algorithms in the strongly connected case.
\end{rmk}
\section{Stability and synchronization}\label{sec-stab-sync}
\subsection{Transitions, stability, and synchronization}\label{subsec-trans-stab-sync}
A right-resolver on a graph $G$ induces transition maps on $V(G)$ in the standard way:
\begin{defn}[transition map]
Let $G, H$ be graphs with $H \leq_R G$. Let $\Phi \in \hom_R(G,H)$. For $I \in V(G)$ and $u \in L_{\partial \Phi(I)}(H)$, we write $I \cdot u$ for the terminal state $t(\gamma)$ of the unique $\gamma \in L_I(G)$ with $\Phi(\gamma) = u$. That is, $I \cdot u = t( (\Phi|_{L_I(G)} )^{-1}(u))$. We denote by $S_{\Phi}$ the set of maps of the form $I \mapsto I \cdot u$ with respect to $\Phi$.
\end{defn}
We now introduce the notion of a \textit{congruence} (see \cite{bpr-09-ca}, Chapter 1, or \cite{jk-03-tcs}, \S 3), of which we will see two important examples. The main example will be the stability relation, but we will also use a congruence in \cref{prop-construct-bg} to construct the maximal bunchy factor $B(G)$ of a given graph $G$.
\begin{defn}[congruences and quotients]
Let $G,H$ be graphs with $H \leq_R G$, let $\Phi \in \hom_R(G,H)$, and let $\sim$ be an equivalence relation on $V(G)$. We say that $\sim$ is a \textit{congruence} with respect to $\Phi$ if it is invariant under transitions, i.e. for all $I \in V(H)$, all $u \in L_I(H)$, and all $I_1, I_2 \in (\partial \Phi)^{-1}(I)$ with $I_1 \sim I_2$, we have $I_1 \cdot u = I_2 \cdot u$. We ``overload'' a congruence $\sim$ by defining it also on paths (in particular, edges), by saying that $\gamma_1 \sim_{\Phi} \gamma_2$, for $\gamma_1, \gamma_2 \in L(G)$, if $\Phi(\gamma_1) = \Phi(\gamma_2)$ and $s(\gamma_1) \sim_{\Phi} s(\gamma_2)$.
Define the \textit{quotient graph} $G/\sim$ by $V(G/\sim) = V(G)/\sim$, $E(G/\sim) = E(G)/\sim$, $s([e]_{\sim}) = [s(e)]_{\sim}$, and $t([e]_{\sim}) = [t(e)]_{\sim}$. There are right-resolvers $G \to K$, $K \to H$, which compose to $\Phi$, where the right-resolver $G \to K$ is the quotient map, and the right-resolver $K \to H$ takes a $\sim$ class to the image in $H$ of any of its representatives.
\end{defn}
Let $G,H$ be graphs with $H \leq_R G$, let $\Phi \in \hom_R(G,H)$, and let $\sim$ be a congruence on $G$ with respect to $\Phi$. Let $\Psi: E(G) \to E(K)$, $\partial \Psi: V(G) \to V(K)$ be the quotient maps. Let $\Delta : E(K) \to E(H)$, $\partial \Delta: V(K) \to V(H)$ be given by $\Delta([\cdot]_{\sim}) = \Phi(\cdot)$, $\partial \Delta([\cdot]_{\sim}) = \partial \Phi(\cdot)$ respectively. (This is not a global definition of $\Psi$, $\Delta$ as specific notation, although we will frequently use these symbols for such a factorization.) Note that we have $\Psi \in \hom_R(G, K)$ and $\Delta \in \hom_R(K, H)$, with $\Phi = \Delta \circ \Psi$. Moreover, we have $S_{\Delta} = S_{\Phi} / \sim$, in the following sense: for $I \in V(H)$, $I' \in (\partial \Phi)^{-1}(G)$, and $u \in L_I(H)$, we have $[I']_{\sim} \cdot u = [I' \cdot u]_{\sim}$, where the left-hand side denotes a mapping in $S_{\Delta}$, while the representative of the equivalence class on the right-hand side denotes a mapping in $S_{\Phi}$.
\begin{rmk}
The coarsest congruence, with respect to a right-resolver $\Phi \in \hom_R(G,H)$, is the total relation on the fibers, i.e. the relation $\bigsqcup_{I \in V(H)} ((\partial \Phi)^{-1}(I))^2$. The quotient of $G$ by this relation is simply $H$, with $\Phi$ as the quotient map.
\end{rmk}
\begin{defn}[stability relation for a right-resolver]
Let $G,H$ be graphs with $H \leq_R G$ and let $\Phi \in \hom_R(G,H)$. The \textit{stability relation} for $\Phi$, written $\sim_{\Phi}$, is the equivalence relation on $V(G)$ defined as follows: for $I \in V(H)$ and $I_1, I_2 \in (\partial \Phi)^{-1}(I)$, $I_1 \sim_{\Phi} I_2$ if and only if for all $u \in L_{I}(H)$, there exists $v \in L_{t(u)}(H)$ such that $I_1 \cdot uv = I_2 \cdot uv$.
\end{defn}
\begin{lemma}
Let $G,H$ be graphs with $H \leq_R G$ and let $\Phi \in \hom_R(G,H)$. The stability relation $\sim_{\Phi}$ is a congruence with respect to $\Phi$.
\end{lemma}
\begin{proof}
Let $I \in V(H)$ and $I_1, I_2 \in (\partial \Phi)^{-1}(I)$ with $I_1 \sim_{\Phi} I_2$. Let $u \in L_I(H)$ and let $v \in L_{t(u)}(H)$. Since $I_1 \sim_{\Phi} I_2$, there exists $w \in L_{t(v)}(H)$ such that $I_1 \cdot uvw = I_2 \cdot uvw$. Therefore $I_1 \cdot u \sim_{\Phi} I_2 \cdot u$, so $\sim_{\Phi}$ is indeed a congruence.
\end{proof}
We now define \textit{synchronizing} right-resolvers in terms of the stability relation, then show in \cref{prop-stab-sync-equiv} that, at least in the strongly connected case, this definition is equivalent to a more obvious notion of synchronization for a right-resolver.
\begin{defn}[synchronizer]
Let $G,H$ be graphs with $H \leq_R G$ and let $\Phi \in \hom_R(G,H)$. We say that $\Phi$ is \textit{synchronizing}, or is a \textit{synchronizer}, if each fiber $(\partial \Phi)^{-1}(I)$, $I \in V(H)$, is a $\sim_{\Phi}$ class. We denote the set of synchronizers $G \to H$ by $\hom_S(G,H)$, and we write $H \leq_S G$ if $\hom_S(G,H) \neq \emptyset$.
\end{defn}
Note that $\sim_{\Phi}$ depends on $\Phi$ only through $S_{\Phi}$. That is, if $\Phi, \Phi'$ are such that $S_{\Phi} = S_{\Phi'}$, then $\sim_{\Phi} = \sim_{\Phi'}$.
\begin{prop}\label{prop-stab-sync-equiv}
Let $G,H$ be graphs with $H \leq_R G$ and let $\Phi \in \hom_R(G,H)$. If $\Phi$ is synchronizing, then for every $I \in V(H)$, there exists $u \in L_I(H)$ with $|(\partial \Phi)^{-1}(I) \cdot u| = 1$. If $G$ is strongly connected, then the converse holds as well.
\end{prop}
\begin{proof}
First, suppose that $\Phi$ is synchronizing and let $I \in V(H)$. If $| (\partial \Phi)^{-1}(I) | = 1$, then we are done. Otherwise, since $(\partial \Phi)^{-1}(I)$ is a $\sim_{\Phi}$ class by assumption, there exists $u_1 \in L_I(H)$ with $|(\partial \Phi)^{-1}(I) \cdot u_1| < | (\partial \Phi)^{-1}(I)|$. Continuing inductively, we can produce a sequence of words $u_1, \dots, u_n$ such that $t(u_i) = s(u_{i+1})$ and $| (\partial \Phi)^{-1}(I) \cdot u_1 \cdots u_n | = 1$.
For the converse, suppose that $G$ is strongly connected. Let $I \in V(H)$. Let $u \in L_I(H)$ be such that $|(\partial \Phi)^{-1}(I) \cdot u| = 1$, and let $v \in L_I(H)$ be arbitrary. By strong connectedness, let $w \in L_{t(v)I}(H)$. Then $(\partial \Phi)^{-1}(I) \cdot vw \subseteq (\partial \Phi)^{-1}(I)$, so $(\partial \Phi)^{-1}(I) \cdot vwu \subseteq (\partial \Phi)^{-1}(I) \cdot u$. Therefore $|(\partial \Phi)^{-1}(I) \cdot vwu| = 1$. That is, for any $I_1, I_2 \in (\partial \Phi)^{-1}(I)$ and any $v \in L_{t(u)}(H)$, we have $wu$ such that $I_1 \cdot vwu = I_2 \cdot vwu$, so $I_1 \sim_{\Phi} I_2$. Therefore $(\partial \Phi)^{-1}(I)$ is a $\sim_{\Phi}$ class, and $\Phi$ is synchronizing.
\end{proof}
We now summarize the structure of stability, in the sense of its behaviour with respect to composition of right-resolvers. For the proof of \cref{thm-struct-sync-comp}, see \S \ref{subsec-stab-proofs}.
\begin{thm}\label{thm-struct-sync-comp}
Let $G,K,H$ be graphs with $H \leq_R K \leq_R G$. Let $\Psi \in \hom_R(G,K)$, $\Delta \in \hom_R(K,H)$, and let $\Phi = \Delta \circ \Psi$.
\begin{enumerate}
\item The $\sim_{\Psi}$ classes in $V(G)$ are the intersections of $\sim_{\Phi}$ classes with $\partial \Psi$ fibers. In particular, $\Psi$ is synchronizing if and only if every $\partial_{\Psi}$ fiber is contained in a $\sim_{\Phi}$ class.
\item If $K = G/\sim_{\Phi}$ and $\Psi$ is the quotient map for $\sim_{\Phi}$, then $\Psi$ is synchronizing and $\sim_{\Delta}$ is trivial.
\item If $\sim_{\Delta}$ is trivial, then $\sim_{\Phi} = \sim_{\Psi}$.
\item $\Phi$ is synchronizing if and only if both $\Psi$ and $\Delta$ are synchronizing.
\end{enumerate}
\end{thm}
The following observation follows immediately from \cref{thm-struct-sync-comp}(4).
\begin{cor}
The relation $\leq_S$ is transitive, and is thus a partial order on the class of graphs (again, really isomorphism classes of graphs), refining the partial order $\leq_R$.
\end{cor}
\begin{conj}[$O(G)$ conjecture]\label{main-conj}
Let $G$ be a strongly connected graph. Then the set of graphs $H$ with $H \leq_S G$ has a unique $\leq_S$-minimal element $O(G)$.
\end{conj}
\begin{rmk}
This remark is intended for readers interested in algebraic or categorical perspectives on automata theory. Recall that for a complete DFA, or road colouring $\Phi \in \hom_R(G,M_D)$, where $G$ is a graph of constant out-degree $D$, the set $S_{\Phi}$ of transition maps forms a transformation semigroup under composition. Indeed, a complete DFA is essentially a finite transformation semigroup together with a choice of generators; this perspective is taken explicitly in \cite{acs-17-prim,bf-11-aaecc} and mentioned in \cite{go-81-ijm}, the first paper on the road problem after \cite{agw-77-tms}.
For a general right-resolver $\Phi \in \hom_R(G,H)$, one could see $S_{\Phi}$ as the semigroup of transitions of a partial finite automaton (PFA), where a given transition is defined only on a single fiber. However, it is more helpful to see $S_{\Phi}$ as a \textit{semigroupoid} (equivalently, if the empty word is included, a small category). One reason is that, as we show in \S \ref{sec-comput}, it can decided in polynomial time whether $\Phi$ is synchronizing, and the length of a word synchronizing a given fiber is bounded by a polynomial in $|V(G)|$. This is in contrast to the high level of complexity typical of related problems in subset synchronization and synchronization of PFAs \cite{bfrs-21-stacs,vorel-16-ijfcs}.
The reader may verify as an exercise, generalizing Cayley's theorem or specializing the Yoneda lemma, that every finite semigroupoid is isomorphic to $S_{\Phi}$ for some graphs $G,H$ (although possibly with sinks) and some $\Phi \in \hom_R(G,H)$, with appropriate generalizations to the infinite case. Moreover, just as every group is a quotient of a free group, $S_{\Phi}$ is a quotient of the free semigroupoid $L(H)$.
\end{rmk}
\subsection{Sufficient conditions for stability}
We now give a pair of sufficient conditions for nontrivial stability, both of which are used in the proof of the road colouring theorem and one of which is also used in \S \ref{sec-bunchy} to obtain a right-resolver with nontrivial stability on an almost bunchy graph.
The first condition involves a special case of the operation known as in-amalgamation (\cite{lm-95-intro}, \S 2.4):
\begin{lemma}\label{lemma-amalg-stab}
Let $G,H$ be graphs with $H \leq_R G$ and let $\Phi \in \hom_R(G,H)$ be a right-resolver. Let $I \in V(H)$. Suppose that there exist $I_1, I_2 \in (\partial \Phi)^{-1}(I)$ such that $|E_{I_1 J}(G)| = |E_{I_2 J}(G)|$ for all $J \in V(G)$. Then there exists $\Phi' \in \hom_R(G,H)$ such that $I_1 \sim_{\Phi'} I_2$.
\end{lemma}
\begin{proof}
Let $F = F(I_1) = F(I_2)$. For each $J \in F$, choose a bijection $\Theta_J: E_{I_2 J}(G) \to E_{I_1 J}(G)$. Define $\Phi'$ as follows: $\partial \Phi' = \partial \Phi$, $\Phi'|_{E(G) \setminus E_{I_2}(G)} = \Phi|_{E(G) \setminus E_{I_2}(G)}$, and, for each $J \in F$, $\Phi'|_{E_{I_2 J}(G)} = \Phi|_{E_{I_1 J}(G)} \circ \Theta_J$. For any $a \in E_I(G)$, we then have $I_1 \cdot a = I_2 \cdot a$. Any $w \in L_I(H)$ is of the form $w = au$ with $a \in E_I(G)$, so $I_1 \cdot w = I_2 \cdot w$. Therefore $I_1 \sim_{\Phi'} I_2$.
\end{proof}
In \cref{lemma-amalg-stab}, the states $I_1, I_2$ are said to be \textit{in-amalgamated} by the operation $G \to G/\sim_{\Phi'}$; the inverse operation is known as \textit{in-splitting}. The lemma shows in particular that no fiber of a $\leq_S$-minimal graph $G$ over $M(G)$ has two states that can be in-amalgamated. Trahtman applies a special case of this fact to graphs of constant out-degree, and we follow his line of application; see the first paragraph of the proof of \cref{thm-bfc-cyc-bunch}, found in \S \ref{subsec-traht-app}.
The second sufficient condition, given in \cref{lemma-min-diff-stab}, concerns \textit{minimal images}:
\begin{defn}[minimal image]
Let $G,H$ be graphs with $H \leq_R G$, and let $\Phi \in \hom_R(G,H)$. A \textit{minimal image} is a set of the form $U = (\partial \Phi)^{-1}(I) \cdot u$ for some $I \in V(H)$ and $u \in L_I(H)$, such that $|U \cdot v| = |U|$ for any $v \in L_{t(u)}(H)$.
\end{defn}
\begin{rmk}
Let $G,H$ be graphs with $H \leq_R G$, and let $\Phi \in \hom_R(G,H)$. For any $I \in V(H)$, any $u \in L_{I}(H)$, and any $v \in L_{t(u)}(H)$, if $U = (\partial \Phi)^{-1}(I) \cdot u$, we clearly have $| U \cdot v | \leq_R | U |$. If there exists $v \in L_{t(u)}(H)$ such that this inequality is strict, then $|U|$ is not minimal, i.e. $U$ is not a minimal image. This is the reason for the term.
For a right-resolver $\Phi$ on a strongly connected graph, all minimal images have the same size, which is called the \textit{degree} of $\Phi$, and a word with minimal image is called a \textit{magic word}. See \cite{amt-95-etds}, \S 9.1 for a treatment of degrees, using symbolic dynamics. In this paper, we only need a small fragment of the theory of degree, which we establish in a self-contained way with no connectedness assumptions, using the properties of stability.
\end{rmk}
We use minimal images to give a criterion for stability that can be seen as a pairwise version of \cref{prop-stab-sync-equiv}.
\begin{lemma}\label{lemma-stab-sync-magic}
Let $G,H$ be graphs with $H \leq_R G$, and let $\Phi \in \hom_R(G,H)$. For $I \in V(H)$ and $I_1, I_2 \in (\partial \Phi)^{-1}(I)$, we have $I_1 \sim_{\Phi} I_2$ if and only $I_1 \cdot u = I_2 \cdot u$ for every word $u \in L_I(H)$ such that $(\partial \Phi)^{-1}(I) \cdot u$ is a minimal image.
\end{lemma}
\begin{proof}
Let $I \in V(H)$ and let $I_1, I_2 \in (\partial \Phi)^{-1}(I)$. First suppose that $I_1 \sim_{\Phi} I_2$, and let $u \in L_I(H)$. If $I_1 \cdot u \neq I_2 \cdot u$, then let $v \in L_{t(u)}(H)$ be such that $I_1 \cdot uv = I_2 \cdot uv$. Then $| (\partial \Phi)^{-1}(I) \cdot uv | < | (\partial \Phi)^{-1}(I) \cdot u|$, so $(\partial \Phi)^{-1}(I) \cdot u$ is not a minimal image. Contrapositively, if $(\partial \Phi)^{-1}(I) \cdot u$ is a minimal image, then $I_1 \cdot u = I_2 \cdot u$.
Conversely, suppose that $I_1 \cdot u = I_2 \cdot u$ for every word $u \in L_I(H)$ such that $(\partial \Phi)^{-1}(I) \cdot u$ is a minimal image. Let $r = \min_{u \in L_I(H)} |(\partial \Phi)^{-1}(I) \cdot u|$ and let $u \in L_I(H)$. We claim that there exists $w \in L_{t(u)}(H)$ with $I_1 \cdot uw = I_2 \cdot uw$. Indeed, if $I_1 \cdot u \neq I_1 \cdot u$, then $(\partial \Phi)^{-1}(I) \cdot u$ is not a minimal, image, so there exists $v \in L_{t(u)}(H)$ such that $|(\partial \Phi)^{-1}(I) \cdot uv | < |(\partial \Phi)^{-1}(I) \cdot u |$. We can thus inductively extend $v$ to obtain the desired $w$, so indeed $I_1 \sim_{\Phi} I_2$.
\end{proof}
The following easy observation about minimal images is the main reason that our results toward the bunchy factor conjecture (the generalized road colouring theorem and the related result for almost bunchy graphs) require strong connectedness.
\begin{lemma}\label{lemma-sc-min-img}
Let $G,H$ be strongly connected graphs with $H \leq_R G$, and let $\Phi \in \hom_R(G,H)$. Every minimal image for $\Phi$ has the same cardinality, and for every $I' \in V(G)$, there exists a minimal image $U$ with $I' \in U$.
\end{lemma}
In the proof of the generalized road colouring theorem, we apply \cref{lemma-sc-min-img} both directly and via \cref{lemma-min-diff-stab}. The proof of \cref{lemma-min-diff-stab} is adapted from the proof of Lemma 10.4.4 in \cite{bpr-09-ca}.
\begin{prop}\label{lemma-min-diff-stab}
Let $G, H$ be strongly connected graphs with $H \leq_R G$ and let $\Phi \in \hom_R(G,H)$. Let $I, J \in V(H)$ and let $u_1, u_2 \in L_{IJ}(H)$ be such that $U_i = (\partial \Phi)^{-1}(I) \cdot u_i$ are minimal images. If $|U_1 \Delta U_2| = 2$, say $U_1 \Delta U_2 = \{ J_1, J_2 \}$, then $J_1 \sim_{\Phi} J_2$.
\end{prop}
\begin{proof}
Let $r = |U_1| = |U_2|$. Suppose WLOG that $J_i \in U_i$, and let $U_0 = U_1 \cap U_2$, so that $U_i = U_0 \cup \{ J_i \}$. For any $v \in L_J(H)$, we have $(U_1 \cup U_2) \cdot v = (U_0 \cdot v) \cup (\{ J_1, J_2 \} \cdot v)$. We must have $|U_0 \cdot v| = |U_0| = r-1$ and $J_i \cdot v \notin U_0 \cdot v$, since otherwise we would have $|(\partial \Phi)^{-1}(I) \cdot u_i v| = |U_i \cdot v| < r$, contradicting the minimality assumption. Therefore $r-1 + | \{ J_1, J_2 \} \cdot v| = |(U_1 \cup U_2) \cdot v|$. Note that $|(U_1 \cup U_2) \cdot v| \in \{ r,r+1 \}$.
Let $v \in L_J(H)$ be such that $(\partial \Phi)^{-1}(J) \cdot v$ is a minimal image. By \cref{lemma-stab-sync-magic}, to show that $J_1 \sim_{\Phi} J_2$, we need to find $v \in L_J(H)$ with $| \{ J_1, J_2 \} \cdot v| = 1$, or equivalently $|(U_1 \cup U_2) \cdot v| \leq_R r$. By strong connectedness and \cref{lemma-sc-min-img}, we have $|(\partial \Phi)^{-1}(J) \cdot v| = r$. But $(U_1 \cup U_2) \cdot v \subseteq (\partial \Phi)^{-1}(I) \cdot v$, so indeed $|(U_1 \cup U_2) \cdot v| \leq_R r$.
\end{proof}
\section{The $O(G)$ conjecture and the road problem}\label{sec-og-road}
\subsection{Generalization of the road colouring theorem}
We first introduce the class of graphs involved in the theorem. A \textit{bunch} in a graph $G$ is a state $I \in V(G)$ with $|F(I)| = 1$. (This terminology, introduced in \cite{ant-09-ijm} and used also in \cite{bpr-09-ca}, is the origin of our term \textit{bunchy}, introduced in \S \ref{sec-bunchy}.) A strongly connected graph in which every state is a bunch is a \textit{cycle of bunches}. Let $M$ be a cycle of bunches with $V(M) = \{ I_0, \dots, I_{p-1} \}$, where $F(I_i) = \{ I_{i+1} \}$ (subscripts indexing states in a cycle of length $p$ should be read modulo $p$ throughout), and let $D_i = |E_{I_i}(M)|$. Note that $M$ is $\leq_R$-minimal if and only if the sequence of out-degrees $D_0, \dots, D_{p-1}$ is not a cyclic shift of a sequence obtained by concatenating a strictly shorter sequence with itself more than once.
Let $M$ be a $\leq_R$-minimal cycle of bunches. Let $O_{M,q}$ be the cycle of bunches in which the sequence of out-degrees consists of $q$ cyclic repetitions of $D_0, \dots, D_{p-1}$. Note that $O_{M,1} = M$. Observe that, for a strongly connected graph $G$ with $M=M(G)$ a cycle of bunches, if $q = \mathrm{per}(G)/\mathrm{per}(M)$ and $H$ is a cycle of bunches with $H \leq_S G$, then $H = O_{M,q}$. Let $O_{D,p} = O_{M_D,p}$ be the cycle of bunches of period $p$ and constant out-degree $D$. Note that $O_{D,1} = M_D$. For a strongly connected, aperiodic graph $G$ of constant out-degree $D$, a synchronizer $G \to M_D$ is precisely a synchronizing road colouring of $G$.
\begin{thm}[Trahtman, \cite{ant-09-ijm}]\label{thm-rct-aper}
Let $G$ be a strongly connected, aperiodic graph of constant out-degree $D$. Then $M_D \leq_S G$.
\end{thm}
\begin{thm}[Béal-Perrin \cite{bp-14-dam}, Budzban-Feinsilver \cite{bf-11-aaecc}]\label{thm-prct}
Let $G$ be a strongly connected graph of constant out-degree $D$ and period $p$. Then $O_{D,p} \leq_S G$.
\end{thm}
We prove the following generalization:
\begin{thm}\label{thm-bfc-cyc-bunch}
Let $G$ be a strongly connected graph such that $M(G)$ is a cycle of bunches. Let $q = \mathrm{per}(G)/\mathrm{per}(M(G))$. Then $O_{M(G),q} \leq_S G$.
\end{thm}
The proof (see \S \ref{sec-traht}) follows that of Theorems \ref{thm-rct-aper} and \ref{thm-prct}. The strategy is to show that if $G$ is not itself a cycle of bunches, then there exists $\Phi \in \hom_R(G,M(G))$ with $\sim_{\Phi}$ nontrivial, by constructing two minimal images that differ by a pair and applying \cref{lemma-min-diff-stab}. A very similar strategy is used to prove the bunchy factor conjecture for almost bunchy graphs (\cref{cor-alm-og-bunchy}), the (substantial) difference being the different techniques used to obtain the requisite pair of minimal images.
\subsection{The $O(G)$ conjecture implies the road colouring theorem}
We now recall the sense in which the $O(G)$ conjecture was first understood to relate to the road problem. Although the road colouring theorem clearly implies the $O(G)$ conjecture for strongly connected, aperiodic graphs of constant out-degree, the converse implication may not be apparent. Indeed, the $O(G)$ conjecture asserts that $O(G)$ is well-defined for every strongly connected graph, but does not immediately say how to compute $O(G)$, whereas the road colouring theorem explicitly specifies the form of $O(G)$ for the graphs $G$ to which it applies. However, the $O(G)$ conjecture does imply the road colouring theorem, via a key result from \cite{agw-77-tms}, for which we require a definition.
\begin{defn}[higher edge graph]\label{defn-higher-edge}
Let $G$ be a graph. For $k \geq 2$, the $k$-th \textit{higher edge graph} of $G$ is the graph $G^{[k]}$ with edge set consisting of edge paths $e_1 e_2 \cdots e_{k-1} e_k$ of length $k$ in $G$, and states given by $s(e_1 e_2 \cdots e_{k-1} e_k) = e_1 e_2 \cdots e_{k-1}$, $t(e_1 e_2 \cdots e_{k-1} e_k) = e_2 \cdots e_{k-1} e_k$. We define $G^{[1]} = G$.
\end{defn}
It is a standard result (\cite{lm-95-intro}, Chapter 2) that $G \leq_S G^{[k]}$ for any strongly connected graph $G$ and any $k \geq 1$. In our terminology, Adler-Goodwyn-Weiss showed the following:
\begin{lemma}[\cite{agw-77-tms}, Lemma 4]\label{lemma-agw-higher-edge}
Let $G$ be a strongly connected, aperiodic graph of constant out-degree $D$. Then for all sufficiently large $k$, we have $M_D \leq_S G^{[k]}$.
\end{lemma}
Together with an easy observation about partially ordered sets, the Adler-Goodwyn-Weiss result shows that the $O(G)$ conjecture implies the road colouring theorem.
\begin{lemma}\label{lemma-poset}
Let $(\mathcal{P}, \preceq)$ be a partially ordered set such that, for any $y \in \mathcal{P}$, there exists a unique $\preceq$-minimal element $O(y) \preceq y$. If $x \preceq y$, then $O(x) = O(y)$.
\end{lemma}
\begin{prop}[\cite{amt-95-etds}]\label{prop-amt-og-implies-road}
Suppose that the $O(G)$ conjecture is true. Let $G$ be a strongly connected, aperiodic graph of constant out-degree $D$. Then $M_D \leq_S G$.
\end{prop}
\begin{proof}
Let $k$ be large enough that $M_D \leq_S G^{[k]}$, by \cref{lemma-agw-higher-edge}. Then in fact $M_D = O(G^{[k]})$. Since $G \leq_S G^{[k]}$ as well, the result follows by \cref{lemma-poset}.
\end{proof}
\section{Bunchiness}\label{sec-bunchy}
In this section, we define and characterize the classes of bunchy and almost bunchy graphs, and demonstrate the importance of bunchy graphs to the structural properties of right-resolvers.
\subsection{Bunchy and almost bunchy graphs}
\begin{defn}[bunchy states and graphs]\label{defn-bunchy}
Let $G$ be a graph. We say that a state $I \in V(G)$ is \textit{bunchy} if $\Sigma_G|_{F(I)}: F(I) \to F(\Sigma_G(I)) \subseteq V(M(G))$ is a bijection. We say that $G$ is bunchy if every $I \in V(G)$ is bunchy. We say that $G$ is \textit{almost bunchy} if for each $I,J \in V(M)$, there exists at most one $I' \in \Sigma_G^{-1}(I)$ such that $|F(I') \cap \Sigma_G^{-1}(J)| \geq 2$.
\end{defn}
The following is evident but we state it explicitly for reference:
\begin{lemma}\label{lemma-bunchy-factor-princ}
The classes of bunchy and almost bunchy graphs are closed under right-resolvers. Moreover, if $G$ is a bunchy graph and $C$ is a principal subgraph of $G$ with $M(C) = M(G)$, then $C$ is also bunchy.
\end{lemma}
\begin{rmk}
We briefly discuss examples of bunchy and almost bunchy graphs. The only strongly connected bunchy graphs of constant out-degree are the cycles of bunches. For a given $\leq_R$-minimal cycle of bunches $M$ with sequence of out-degrees $D_0, \dots, D_{p-1}$, the only strongly connected bunchy graphs $G$ with $M(G) = M$ are the graphs $O_{M,q}$ introduced in the previous section.
A strongly connected almost bunchy (but not bunchy) graph of constant out-degree is a graph with a unique non-bunchy state, i.e. a state $I$ with $|F(I)| \geq 2$, together with a path from each element of $F(I)$ back to $I$. One example that has been considered in the literature is the graph $W_n$ studied in \cite{agv-10-lncs}, first discussed in \cite{wiel-50-mz}, and of interest due to its slow synchronization.
An almost bunchy graph $G$ can have at most $|V(M(G))|^2$ non-bunchy states, one for each ordered pair of $\Sigma_G$ fibers. One way to obtain an almost bunchy graph is to start with a bunchy graph and perform a sequence of in-splittings (recall \cref{lemma-amalg-stab}, and see also \cite{lm-95-intro}, \S 2.4), but not all in-splittings will preserve almost bunchiness, and not all almost bunchy graphs arise this way.
\end{rmk}
Bunchy and almost bunchy graphs are characterized in terms of automorphisms, with an important uniqueness consequence for the sets of transition maps induced on them by right-resolvers:
\begin{prop}\label{prop-alm-bunchy-equiv}
A graph $G$ is almost bunchy if and only if there is a unique right-resolver $G \to M(G)$ up to permutations of parallel edges: that is, if and only if, for any $\Phi_1, \Phi_2 \in \hom_R(G,M(G))$, there exist $\sigma \in P(G), \tau \in P(M(G))$ such that $\Phi_1 = \tau \circ \Phi_2 \circ \sigma$. Moreover, $G$ is bunchy if and only if we can take $\tau = \mathrm{id}$.
\end{prop}
\begin{prop}\label{cor-alm-stab-unique}
Let $G$ be an almost bunchy graph. Let $\Phi_1, \Phi_2 \in \hom_R(G,M(G))$. Then $S_{\Phi_1} = S_{\Phi_2}$. In particular, $\sim_{\Phi_1} \, = \, \sim_{\Phi_2}$.
\end{prop}
For the proofs of Propositions \ref{prop-alm-bunchy-equiv} and \ref{cor-alm-stab-unique}, see \S \ref{sec-bunchy-details}. The following definition is now justified.
\begin{defn}
Let $G$ be an almost bunchy graph. We denote by $\sim_G$ the unique relation on $V(G)$ with $\sim_G = \sim_{\Phi}$ for any $\Phi \in \hom_R(G,M(G))$.
\end{defn}
\subsection{Proof of the $O(G)$ conjecture in the bunchy case}
We now resolve the $O(G)$ conjecture in the almost bunchy case (which includes the bunchy case). This extends Corollary 4.3 in \cite{amt-95-etds}, which resolves the conjecture for graphs $G$ such that $M(G)$ has no parallel edges (so $G$ is trivially bunchy). The proof here is quite different from the proof in the no-parallel-edges case, and yields a polynomial-time algorithm (Algorithm \ref{alg-construct-stab}) for constructing $O(G)$.
\begin{thm}\label{thm-alm-og-exists}
Let $G$ be an almost bunchy graph and let $H \leq_S G$. If $H$ is $\leq_S$-minimal, then $H = G/\sim_G$. In particular, the set $\{ K \, | \, K \leq_S G \}$ has a unique $\leq_S$-minimal element $O = G/\sim_G$.
\end{thm}
\begin{proof}
Let $\Psi \in \hom_S(G,H)$ and $\Delta \in \hom_R(H,M)$. We can naturally identify $V(H) = V(G) / \sim_{\Psi}$ by \cref{lemma-fiber-determ-img}. The hypothesis that $H$ is $\leq_S$-minimal implies that $\sim_{\Delta}$ is trivial, so $\sim_{\Psi} \, = \, \sim_{\Delta \circ \Psi}$ by \cref{thm-struct-sync-comp}(3). By \cref{cor-alm-stab-unique}, we have $\sim_{\Delta \circ \Psi} \, = \, \sim_G$. We can thus identify $V(H)$ with $V(G) / \sim_G \, = V(G/ \sim_G)$. Thus $H = G/ \sim_G$ by a second application of \cref{lemma-fiber-determ-img}.
\end{proof}
In the strongly connected case, we can apply \cref{lemma-min-diff-stab}, which is also used in the proof of the road colouring theorem, to say more.
\begin{prop}\label{prop-bfc-alm}
Let $G$ be a strongly connected almost bunchy graph. If $G$ is not bunchy, then $\sim_G$ is nontrivial.
\end{prop}
\begin{proof}
Let $I,J \in V(M(G))$ and $I' \in \Sigma_G^{-1}(I)$ with $|F(I') \cap \Sigma_G^{-1}(J)| \geq 2$. Let $J_1, J_2 \in F(I') \cap \Sigma_G^{-1}(J)$, $J_1 \neq J_2$, and let $e_i \in E_{I' J_i}(G)$. Let $\Phi \in \hom_R(G,M(G))$ and let $a_i = \Phi(e_i)$. By strong connectedness and \cref{lemma-sc-min-img}, there exists a minimal image $U \subseteq \Sigma_G^{-1}(I)$ with $I' \in U$. Let $U_0 = U \setminus \{ I' \}$. Then $U_0 \cdot a_1 = U_0 \cdot a_2$ since $G$ is almost bunchy. Moreover, $J_i = I' \cdot a_i \notin U_0 \cdot a_i$ (otherwise, minimality would be contradicted). Thus $(U \cdot a_1) \Delta (U \cdot a_2) = \{ J_1, J_2 \}$. By \cref{lemma-min-diff-stab}, $J_1 \sim_{\Phi} J_2$.
\end{proof}
\begin{cor}\label{cor-alm-og-bunchy}
Let $G$ be a strongly connected almost bunchy graph. Then $O(G)$ is bunchy.
\end{cor}
\begin{proof}
If $|V(G)| = 1$ then the claim is clearly true. Suppose that the conclusion is true for all almost bunchy $G$ with $|V(G)| \leq_R N$, and let $G$ be almost bunchy with $|V(G)| = N+1$. If $G$ is bunchy, then $O(G)$ is clearly bunchy by \cref{lemma-bunchy-factor-princ}. If $G$ is not bunchy, then $\sim_{G}$ is nontrivial by \cref{prop-bfc-alm}, so $|V(G/\sim_{G})| \leq_R N$. Moreover, since $G/\sim_G \, \leq_S G$, it follows that $O(G) = O(G/\sim_G)$ is bunchy by the inductive hypothesis and \cref{lemma-bunchy-factor-princ}.
\end{proof}
\subsection{Universal property of the fiber product}\label{subsec-fiber-univ-prop}
We recall a standard construction known as the fiber product, and derive several new properties. Chief among these is the one exhibited in \cref{thm-univ-prop}, which is an analogue of the universal property often enjoyed by the fiber product, or pullback, in other categories (see e.g. \cite{er-17-dover}, Definition 3.1.15 and subsequent discussion).
\begin{defn}[fiber product]\label{defn-fiber}
Let $H_1, H_2,K$ be graphs and let $\Psi_i: H_i \to K$ be graph homomorphisms. The \textit{fiber product} of $\Psi_1, \Psi_2$ is the graph $P = H_1 \times_{\Psi_1, \Psi_2} H_2$ where
\begin{align*}
V(P) &= \bigsqcup_{I \in V(K)} (\partial \Psi_1)^{-1}(I) \times (\partial \Psi_2)^{-1}(I) \\
%
E(P) &= \{ (e_1, e_2) \, | \, e_i \in E(H_i), \, \Psi_1(e_1) = \Psi_2(e_2) \}
\end{align*}
together with the coordinate projections $\hat{\Psi}_i: P \to H_i$. We write $\Psi_P = \Psi_i \circ \Hat{\Psi}_i : P \to K$.
\end{defn}
\begin{rmk}
Observe that the $\Hat{\Psi}_i$ are surjective (respectively, right-resolving) when the $\Psi_i$ are surjective (respectively, right-resolving). Moreover, if $C$ is a principal subgraph of $P$ such that the restricted state maps $\partial \Hat{\Psi}_i|_{V(C)} : V(C) \to V(H_i)$ are surjective, then $H_i \leq_R C$, indeed $\Hat{\Psi}_i|_C \in \hom_R(C,H_i)$. In particular, this condition is satisfied if the $H_i$ are strongly connected and $C$ is a principal component of $P$.
\end{rmk}
\begin{rmk}
Often the convention is taken that $V(P) = V(H_1) \times V(H_2)$. However, all of the elements of the full Cartesian product that are not elements of $V(P)$, as we have defined it, would be isolated states, and in particular would be sinks. Our definition has the feature that the fiber product of two sink-free graphs (or rather, of two right-resolvers defined on such graphs) is also sink-free.
\end{rmk}
We now state the universal property of the fiber product. Compare with a similar diagram in \cite{amt-95-etds} (p. 289). See \S \ref{subsec-univ-prop-proof} for the proof.
\begin{thm}\label{thm-univ-prop}
Let $H_1, H_2$ be bunchy graphs with $M(H_1) = M(H_2) = M$. Let $\Psi_i \in \hom_R(H_i, M)$ be right-resolvers, and let $P = H_1 \times_{\Psi_1, \Psi_2} H_2$. Let $G$ be a common right-resolving extension of $H_1, H_2$ via $\Phi_i \in \hom_R(G, H_i)$. Then there exist a principal subgraph $C$ of $P$ and right-resolvers $\Delta_i \in \hom_R(G,C)$ such that $\Phi_i = \hat{\Psi}_i \circ \Delta_i$ and $\partial \Delta_1 = \partial \Delta_2$. In particular, $H_i \leq_R C \leq_R G$, with $\Hat{\Psi}_i|_C \in \hom_R(C,H_i)$.
\end{thm}
\begin{rmk}
The bunchiness hypothesis on the $H_i$ cannot be dropped, as the following construction illustrates. Let $G$ be a graph and let $\Phi_1, \Phi_2 \in \mathrm{Aut}(G)$. In the notation of the the theorem, we will take $H_1 = H_2 = G$. (Recall that any automorphism is right-resolving.) Let $M = M(G)$. Let $\Psi_i \in \hom_R(G, M)$ and $P = G \times_{\Psi_1, \Psi_2} G$. Let $C$ be a principal subgraph of $P$ with $\Hat{\Psi}_i|_V(C)$ surjective, and let $\Delta_i \in \hom_R(G,C)$ with $\Phi_i = \Hat{\Psi}_i|_C \circ \Delta_i$. Then $\Hat{\Psi}_i|_C$ and $\Delta_i$ are isomorphisms, since they compose to an isomorphism. In particular, since $\Psi_1 \circ \Hat{\Psi}_1|_C = \Psi_2 \circ \Hat{\Psi}_2|_C$, we have $\Psi_1 = \Psi_2 \circ \tau$ where $\tau = \Hat{\Psi}_2|_C \, \circ \left( \Hat{\Psi}_1|_C \right)^{-1}$ is an isomorphism. In other words, any two elements of $\hom_R(G, M)$ agree up to an automorphism of $G$. That condition always holds when $G$ is bunchy (see \cref{prop-alm-bunchy-equiv}), but fails in general.
\end{rmk}
We now give two applications of the universal property. The first, \cref{prop-bunchy-same-o}, is applied in \cref{prop-bunchy-implies-o}, which is the main motivation for the bunchy factor conjecture. See \S \ref{subsec-fiber-max-bunchy} for the proof of \cref{lemma-fiber-bunchy}.
\begin{lemma}\label{lemma-fiber-bunchy}
Let $H_1, H_2$ be bunchy graphs with $M(H_1) = M(H_2) = M$. Let $\Psi_i \in \hom_R(H_i, M)$, and let $P = H_1 \times_{\Psi_1, \Psi_2} H_2$. Then $P$ is bunchy. In particular, if $C$ is a principal subgraph of $P$ such that the restrictions $\partial \Hat{\Psi}_i|_{V(C)} : V(C) \to V(H_i)$ are surjective, then $C$ is bunchy.
\end{lemma}
\begin{prop}\label{prop-bunchy-same-o}
Let $G$ be a graph. Let $H_1, H_2 \leq_S G$ be bunchy. Then $O(H_1) = O(H_2)$, i.e. $G$ has at most one $\leq_S$-minimal bunchy synchronizing factor.
\end{prop}
\begin{proof}
Let $M = M(G)$, let $\Phi_i \in \hom_S(G,H_i)$, and let $\Psi_i \in \hom_R(H_i, M)$. Let $P = H_1 \times_{\Psi_1, \Psi_2} H_2$. By \cref{thm-univ-prop}, there is a principal subgraph $C$ of $P$ admitting $\Delta_i \in \hom_R(G,C)$ such that $\Phi_i = \hat{\Psi}_i|_C \circ \Delta_i$. Since each $\Phi_i$ is synchronizing, each restriction $\hat{\Psi}_i|_C$ is synchronizing as well, so $H_i \leq_S C$. Since $C$ is bunchy by \cref{lemma-fiber-bunchy}, we know that $O(C)$ is well-defined, and thus, by \cref{lemma-poset}, we have $O(H_1) = O(C) = O(H_2)$ as claimed.
\end{proof}
For the second application of the universal property, recall that the only strongly connected bunchy graphs of constant out-degree are the cycles of bunches. In particular, by the periodic road colouring theorem, for any strongly connected graph $G$ of constant out-degree $D$ and period $p$, the unique maximal bunchy right-resolving factor of $G$, namely $O_{D,p}$, is a synchronizing factor of $G$ (and is indeed equal to $O(G)$). We now show that every graph $G$ has a unique maximal bunchy right-resolving factor $B(G)$. The construction is similar to that of the auxiliary graph $\Tilde{G}$ in \cite{amt-95-etds}, \S 5. See \S \ref{subsec-fiber-max-bunchy} for the proof, as well as an explicit construction of $B(G)$ yielding a polynomial-time algorithm (Algorithm \ref{alg-construct-bg}).
\begin{prop}\label{prop-bg-exists}
Let $G$ be a graph.
\begin{enumerate}[label=(\arabic*)]
\item The set $\{ H \leq_R G \, | \, H \text{ is bunchy} \}$ has a unique $\leq_R$-maximal element $B=B(G)$.
\item Let $H \leq_R G$ be bunchy and $\Phi \in \hom_R(G,H)$. Then $\Phi$ factors through $B$, i.e. there exist $\Delta \in \hom_R(G,B)$, $\Theta \in \hom_R(B,H)$ such that $\Phi = \Theta \circ \Delta$.
\end{enumerate}
\end{prop}
\section{The $O(G)$ conjecture and bunchy synchronizing factors}\label{sec-conjec}
As we have seen, the $O(G)$ conjecture holds for strongly connected graphs $G$ such that $M(G)$ is a cycle of bunches, and for almost bunchy graphs (including the bunchy graphs) by \cref{thm-alm-og-exists}. Moreover, for strongly connected almost bunchy graphs, and strongly connected graphs that factor onto cycles of bunches, we know that there is a bunchy synchronizing factor, which we show inductively by assuming non-bunchiness and obtaining a right-resolver with a nontrivial stability relation. It seems plausible that, if the $O(G)$ conjecture is true, then it can be proven by a similar approach: assume non-bunchiness, find a right-resolver with nontrivial stability relation, recursively find a bunchy synchronizing factor, and apply \cref{prop-bunchy-same-o}. The next proposition gives several equivalent formulations of the hypothesis that this approach can be made to work. See \S \ref{subsec-fiber-fact-ext} for the proof.
\begin{prop}\label{prop-bfc-equiv-new}
The following statements are equivalent.
\begin{enumerate}[label=(\arabic*)]
\item Any strongly connected $\leq_S$-minimal graph is bunchy.
\item For any strongly connected graph $G$, there exists some bunchy $H \leq_S G$.
\item For any non-bunchy strongly connected graph $G$, there exists some $\Phi \in \hom_R(G,M(G))$ with $\sim_{\Phi}$ nontrivial.
\item For any strongly connected graph $G$, $B(G) \leq_S G$.
\end{enumerate}
\end{prop}
\begin{conj}[bunchy factor conjecture]
The assertions in \cref{prop-bfc-equiv-new} are true.
\end{conj}
\begin{prop}\label{prop-bunchy-implies-o}
The bunchy factor conjecture implies the $O(G)$ conjecture.
\end{prop}
\begin{proof}
Let $G$ be a strongly connected graph and let $A = \{ H \, | \, H \leq_S G, \, H \text{ is $\leq_S$-minimal} \}$. Clearly $|A| \geq 1$. By hypothesis, every element of $A$ is bunchy. By \cref{prop-bunchy-same-o}, $|A| \leq_R 1$, so $A$ has a single element, namely $O(G)$.
\end{proof}
Observe that the bunchy factor conjecture is a straightforward generalization of the road problem. As discussed above, the $O(G)$ conjecture was already known to imply the road colouring theorem, via the higher-edge result from \cite{agw-77-tms}. By contrast, the bunchy factor conjecture implies the road colouring theorem more directly, without reference to \cite{agw-77-tms}.
\section{Computing with right-resolvers}\label{sec-comput}
We now discuss the computational problems of constructing $O(G_1), O(G_2)$, for input graphs $G_1, G_2$ such that the $O(G_i)$ are known to exist, and deciding whether they are isomorphic. Although one could apply generic graph isomorphism algorithms, which are efficient in practice (Theorem 1 in \cite{miller-79-jcss} gives a polynomial-time reduction from directed to undirected graph isomorphism, and see \cite{babai-18-icm} for a survey of the state of the art), it is desirable to have a polynomial-time algorithm, in particular one that only uses constructions involved in the theory of right-resolvers and synchronization. We do not attempt detailed complexity analyses, noting only that all of the procedures we give can be easily seen to run in polynomial time.
\subsection{Basic routines}\label{subsec-comput-basic}
In \cite{amt-95-etds}, a polynomial-time algorithm is given for computing $M(G)$, along with a total ordering of $V(M(G))$ such that any graph isomorphism $M(G) \to M(H)$ must preserve the order of the states. Deciding whether $M(G)$, $M(H)$ are isomorphic is therefore no harder than constructing them.
There are also obvious polynomial-time procedures for constructing the fiber product of two right-resolvers, and for determining whether there is a path from one given state to another (in a graph that may not be strongly connected). With these basic routines, we can construct the stability relation of a right-resolver in polynomial time, as follows.
\begin{alg}\label{alg-construct-stab}
Construct the stability relation of a right-resolver.
\begin{enumerate}
\item Input: graphs $G, H$, and $\Phi \in \hom_R(G,H)$.
\item Construct $P = G \times_{\Phi, \Phi} G$.
\item Populate the set $U$ of states $(I_1, I_2) \in V(P)$ with no outgoing path to the diagonal in $V(P)$.
\item Populate and output: the set $\sim_{\Phi}$ of states $(I_1, I_2) \in V(P)$ with no outgoing path to $U$.
\end{enumerate}
\end{alg}
Recall that, by definition, $\Phi \in \hom_S(G,H)$ if and only if $H = G/\sim_{\Phi}$, i.e. the $\partial \Phi$-fibers are precisely the $\sim_{\Phi}$ classes. Since this is easy to check, Algorithm \ref{alg-construct-stab} can be used to decide whether $\Phi$ is synchronizing.
A similar procedure can be used to construct the maximum bunchy factor $B(G)$, following the construction in \cref{prop-construct-bg}:
\begin{alg}\label{alg-construct-bg}
Construct the maximum bunchy right-resolving factor of a graph.
\begin{enumerate}
\item Input: a graph $G$.
\item Construct $M(G)$, along with the quotient map $\Sigma_G: V(G) \to V(M)$.
\item Construct a graph $H$ with the following data:
\begin{align*}
V(H) &= V(G) \times V(G) \\
%
|E_{(I_1, I_2)(J_1, J_2)}(H)| &=
\begin{cases}
1, & \Sigma_G(I_1) = \Sigma_G(I_2), \Sigma_G(J_1) = \Sigma_G(J_2), \text{ and } J_i \in F(I_i) \\
%
0, & \text{otherwise}
\end{cases}
\end{align*}
\item Populate the set $\approx_0$ of pairs $(I_1, I_2)$ with a path in $H$ from the diagonal to $(I_1, I_2)$.
\item Construct and output: the transitive closure $\approx$ of the relation $\approx_0$.
\end{enumerate}
\end{alg}
Regarding step 4: referring to the description of $\approx_0$ in \cref{prop-construct-bg}, a path in $H$ from the diagonal to $(I_1, I_2)$ corresponds to a pair of paths $\gamma, \delta \in L(G)$ witnessing $I_1 \approx_0 I_2$.
\subsection{Decision procedures for common synchronizing factors and extensions}\label{subsec-comput-decide}
See \S \ref{subsec-fiber-fact-ext} for the proof of the following Proposition, which collects several similar statements relating common synchronizing factors and common synchronizing extensions. These results allow us to use fiber products of two graphs to decide questions about common factors of the graphs.
\begin{prop}\label{cor-air-fact-ext-mg}
Let $G_1,G_2$ be strongly connected graphs with $M(G_1) = M(G_2) = M$.
\begin{enumerate}
\item If $G_1, G_2$ have a common synchronizing factor, then there exist $\Phi_i \in \hom_R(G_i, M)$ and a principal component $C$ of $Q = G_1 \times_{\Phi_1, \Phi_2} G_2$ such that $\Hat{\Phi}_i|_C \in \hom_S(C,G_i)$.
\item Assume the $O(G)$ conjecture. Then the converse holds in (1). That is, suppose that there exist $\Phi_i \in \hom_R(G_i, M)$ and a principal component $C$ of $Q = G_1 \times_{\Phi_1, \Phi_2} G_2$, such that $\Hat{\Phi}_i|_C \in \hom_S(C,G_i)$. Then $G_1, G_2$ have a common synchronizing factor, specifically $O(G_1) = O(C) = O(G_2)$.
\item If the $G_i$ are bunchy, then the equivalence described in (1) and (2) holds unconditionally (i.e. without assuming any unproven conjectures).
\item Assume the bunchy factor conjecture. Then we have $O(G_1) = O(G_2)$ if and only if, for the essentially unique $\Phi_i \in \hom_R(B(G_i), M)$, there is a principal component $C$ of $P = B(G_1) \times_{\Phi_1, \Phi_2} B(G_2)$ such that $\Hat{\Phi}_i|_C \in \hom_S(C, B(G_i))$.
\end{enumerate}
\end{prop}
Ashley-Marcus-Tuncel give a polynomial-time algorithm for deciding whether two strongly connected graphs have a common strongly connected synchronizing extension. Their algorithm relies on the construction of a graph that they call $\Tilde{G}$, from an input graph $G$; see Theorem 5.2 and Remark 5.10 in \cite{amt-95-etds}. If the $O(G)$ conjecture is true, then by \cref{cor-air-fact-ext-mg}(2), the Ashley-Marcus-Tuncel algorithm also decides whether $O(G_1) = O(G_2)$ for $G_1, G_2$ strongly connected. Without assuming any unproven conjectures, a negative result from the algorithm shows, by \cref{cor-air-fact-ext-mg}(1), that $G_1, G_2$ have no common synchronizing factor, while a positive result is inconclusive.
For bunchy $G_1, G_2$, however, \cref{thm-alm-og-exists} and \cref{cor-air-fact-ext-mg}(3) show that it can be decided in polynomial time whether $O(G_1), O(G_2)$ are isomorphic without the Ashley-Marcus-Tuncel algorithm:
\begin{alg}\label{alg-decide-bunchy-og}
Decide whether $O(G_1), O(G_2)$ are isomorphic for $G_1, G_2$ strongly connected and bunchy.
\begin{enumerate}
\item Input: strongly connected bunchy graphs $G_1, G_2$ such that $M(G_1) = M(G_2) = M$.
\item Choose $\Phi_i \in \hom_R(G_i,M)$.
\item Construct $Q = G_1 \times_{\Phi_1, \Phi_2} G_2$.
\item For each principal component $C$, decide whether $\Hat{\Phi}_i|_C \in \hom_S(C, G_i)$, using Algorithm \ref{alg-construct-stab}. If so, halt, and return the result that $O(G_1) = O(G_2)$.
\item If no affirmative result is returned in step 4, then halt and return the result that $O(G_1) \neq O(G_2)$.
\end{enumerate}
\end{alg}
Regarding steps 4--5: \cref{cor-air-fact-ext-mg}(3) shows that we have $O(G_1) = O(G_2)$ if and only if $\Hat{\Phi}_i|_C \in \hom_S(C,G_i)$ for some principal component $C$ of $Q$. The same algorithm would work even if $G_1, G_2$ are not strongly connected but only bunchy, with ``principal component $C$'' replaced by ``principal subgraph $C$ such that each $\partial \Hat{\Phi}_i|_{V(C)}$ is surjective'', but it is not clear that the number of such subgraphs is bounded by a polynomial in $|V(G)|$.
Furthermore, if the bunchy factor conjecture is true, then isomorphism of $O(G_1), O(G_2)$ is equivalent, by \cref{prop-bfc-equiv-new}, to isomorphism of $O(B(G_1)), O(B(G_2))$, which can be decided without the use of the Ashley-Marcus-Tuncel $\Tilde{G}$ algorithm, using the following procedure. Note that this works even if we do not have an efficient way to find an element of $\hom_S(G_i, B(G_i))$.
\begin{alg}\label{alg-bfc-iso-og}
Decide isomorphism of $O(G_1), O(G_2)$ (assuming the bunchy factor conjecture).
\begin{enumerate}
\item Input: strongly connected graphs $G_1, G_2$.
\item Construct $B(G_1)$ and $B(G_2)$, using Algorithm \ref{alg-construct-bg}.
\item Decide whether $O(B(G_1)) = O(B(G_2))$, using Algorithm \ref{alg-decide-bunchy-og}.
\item Output: the value ``$O(B(G_1)) = O(B(G_2))$''.
\end{enumerate}
\end{alg}
Observe that the procedure of Algorithm \ref{alg-bfc-iso-og} is essentially what is described in \cref{cor-air-fact-ext-mg}(4).
\section{Proofs of structural results and additional details}\label{sec-proofs}
\subsection{Remarks on the proof of \cref{thm-mg-sigma}}\label{subsec-right-air-mg}
In this subsection, we revisit the first proof of \cref{thm-mg-sigma} given in \cite{amt-95-etds}. The proof of the uniqueness of $M(G)$, for a given graph $G$, is without issue, but the proof of the uniqueness of $\Sigma_G$ is not quite complete. That proof seems to assume that, for graphs $G, H$ with $H \leq_R G$, two right-resolvers $G \to H$ with distinct state maps must partition $V(G)$ differently. In general, this is false: per \cref{lemma-same-fiber-aut} below, counterexamples arise precisely when $\mathrm{Aut}(H)$ acts nontrivially on $V(H)$.
The second proof given in \cite{amt-95-etds}, constructing $V(M(G))$ by successive refinements of an initial state partition, can be made complete by observing that all of the state maps corresponding to the successive refinements are invariant under $\mathrm{Aut}(G)$.
\begin{prop}[Lemma 3.1 in \cite{amt-95-etds}]\label{prop-amt-right-meet}
Let $G, H_1, H_2$ be graphs with $H_i \leq_R G$ via $\Psi_i \in \hom_R(G, H_i)$, with partitions $\alpha_i = V(H_i)$ of $V(G)$. Then there is a graph $K \leq_R H_1, H_2$, with $V(K)$ equal to the finest common coarsening of $\alpha_1, \alpha_2$.
\end{prop}
\begin{cor}[Theorem 3.2 in \cite{amt-95-etds}]\label{cor-amt-mg-exists}
Let $G$ be a graph. Then there exists a unique $\leq_R$-minimal graph $M(G) \leq_R G$, with $V(M(G))$ given by the finest common coarsening of the partitions $\alpha = V(H)$ of $V(G)$, where $H \leq_R G$.
\end{cor}
We will show further that for any $\Phi_1, \Phi_2 \in \hom_R(G,M(G))$, we have $\partial \Phi_1 = \partial \Phi_2$. Note that $\partial \Phi_1, \partial \Phi_2$ at least have the same sets of fibers, since, if they did not, then by \cref{prop-amt-right-meet}, we could take their finest common coarsening, contradicting the minimality of $M(G)$. It is therefore enough to show that $\mathrm{Aut}(M(G))$ acts trivially on $M(G)$. For this, we need a lemma.
\begin{lemma}\label{lemma-same-fiber-aut}
Let $G,H$ be graphs with $H \leq_R G$, and let $\Phi_1, \Phi_2 \in \hom_R(G,H)$. Suppose that the $\partial \Phi_i$ have the same fibers, i.e. for any $I_1, I_2 \in V(G)$, we have $\partial \Phi_1(I_1) = \partial \Phi_1(I_2)$ if and only if $\partial \Phi_2(I_1) = \partial \Phi_2(I_2)$. Then there exists an automorphism $\tau \in \mathrm{Aut}(H)$ such that $\partial \Phi_2 = \partial(\tau \circ \Phi_1)$.
\end{lemma}
\begin{proof}
By the assumption of equal fibers, there exists a (unique) bijection $T: V(H) \to V(H)$ with $\partial \Phi_2 = T \circ \partial \Phi_1$. We need to find $\tau \in \mathrm{Aut}(H)$ with $\partial \tau = T$. Let $I \in V(H)$. For $J \in F(I)$, note, by the right-resolving hypothesis on the $\Phi_i$ and the choice of $T$, that $|E_{IJ}(H)| = |E_{T(I)T(J)}(H)|$. Letting $\tau|_{E_{IJ}(H)}: E_{IJ}(H) \to E_{T(I)T(J)}(H)$ be any bijection, we have $\tau \in \mathrm{Aut}(H)$ with $T = \partial \tau$ and $\partial \Phi_2 = \partial(\tau \circ \Phi_1)$.
\end{proof}
We now discuss the quotient of a graph by its automorphism group. Let $G$ be a graph. We will construct a graph $K = G/\mathrm{Aut}(G) \leq_R G$ as follows. Let $V(K)$ consist of the orbits in $V(G)$ under $\mathrm{Aut}(G)$. Let $I,J \in V(K)$ and $I_1, I_2 \in I$. We need to specify $|E_{IJ}(K)|$. We claim that $\sum_{J' \in J} |E_{I_1 J'}(G)| = \sum_{J' \in J} | E_{I_2 J'}(G)|$. Indeed, let $\tau \in \mathrm{Aut}(G)$ such that $I_2 = \partial \tau(I_1)$. For any $J' \in J \cap F(I_1)$, we have $\partial \tau(J') \in J \cap F(I_2)$, so $\sum_{J' \in J} |E_{I_1 J'}(G)| \leq_R \sum_{J' \in J} | E_{I_2 J'}(G)|$. Replacing $\tau$ with $\tau^{-1}$, we obtain equality. We can thus declare that $|E_{IJ}(K)| = \sum_{J' \in J} |E_{I_i J'}(G)|$. This edge count ensures that $K \leq_R G$.
In particular, if $\mathrm{Aut}(G)$ acts nontrivially on $V(G)$, then $K \neq G$, so $G$ is not $\leq_R$-minimal. It follows that, for a $\leq_R$-minimal graph $M$, $\mathrm{Aut}(M)$ acts trivially on $V(M)$. This observation, together with \cref{lemma-same-fiber-aut}, shows that for any $\Phi_1, \Phi_2 \in \hom_R(G,M(G))$, we have $\partial \Phi_1 = \partial \Phi_2$.
\subsection{Proof of \cref{thm-struct-sync-comp}}\label{subsec-stab-proofs}
We prove \cref{thm-struct-sync-comp} in several stages.
\begin{lemma}\label{lemma-stab-comp}
Let $G, K, H$ be graphs with $H \leq_R K \leq_R G$. Let $\Psi \in \hom_R(G,K)$, $\Delta \in \hom_R(K,H)$, and let $\Phi = \Delta \circ \Psi$. Then $\sim_{\Psi}$-classes are intersections of $\sim_{\Phi}$-classes with $\partial \Psi$-fibers. That is, for $I_1, I_2 \in V(G)$, if $I_1 \sim_{\Psi} I_2$, then $I_1 \sim_{\Phi} I_2$; conversely, if $I_1 \sim_{\Phi} I_2$ and moreover $\partial \Psi(I_1) = \partial \Psi(I_2)$, then $I_1 \sim_{\Psi} I_2$.
\end{lemma}
\begin{proof}
First suppose that $I_1 \sim_{\Psi} I_2$. Then in particular $\partial \Psi(I_1) = \partial \Psi (I_2) = \hat{I}$. Let $I = \partial \Delta(\hat{I})$ and $u \in L_I(H)$, and consider the unique $\lambda \in L_{\Hat{I}}(K)$ such that $\Delta(\lambda) = u$. Since $I_1 \sim_{\Psi} I_2$, there exists $\mu \in L_{t(\lambda)}(K)$ such that $I_1 \cdot \lambda \mu = I_2 \cdot \lambda \mu$. Letting $v = \Delta(\mu)$, we have $I_1 \cdot uv = I_2 \cdot uv$, so $I_1 \sim_{\Phi} I_2$.
For the converse, suppose that $I_1 \sim_{\Phi} I_2$ and $\partial \Psi(I_1) = \partial \Psi(I_2) = \Hat{I}$. Let $I = \Delta(\Hat{I})$, let $\lambda \in L_{\Hat{I}}(K)$, let $u = \Delta(\lambda)$, and let $v \in L_{t(u)}(H)$ be such that $I_1 \cdot uv = I_2 \cdot uv$. Consider the unique $\mu \in L_{t(\lambda)}(K)$ such that $\Delta(\mu) = v$. Then $I_1 \cdot \lambda \mu = I_2 \cdot \lambda \mu$, so indeed $I_1 \sim_{\Psi} I_2$.
\end{proof}
\cref{cor-comp-air-fiber} follows immediately from \cref{lemma-stab-comp}, and together they comprise \cref{thm-struct-sync-comp}(1).
\begin{cor}\label{cor-comp-air-fiber}
If $\Phi = \Delta \circ \Psi$ is a composition of right-resolvers, then $\Psi$ is synchronizing if and only if every $\partial \Psi$-fiber is contained in a $\sim_{\Phi}$-class.
\end{cor}
\begin{proof}[Proof of \cref{thm-struct-sync-comp}(2)]
By assumption, the $\sim_{\Phi}$ classes are precisely the $\partial \Psi$-fibers, so by \cref{lemma-stab-comp}, they are also the $\sim_{\Psi}$ classes. Thus indeed $\Psi \in \hom_S(G,K)$.
Moreover, let $I \in V(K)$, let $I'_1, I'_2 \in (\partial \Delta)^{-1}(I)$, and suppose that $I'_1 \sim_{\Delta} I'_2$. We claim that $I'_1 = I'_2$. Toward this end, we claim that $(\partial \Psi)^{-1}(\{ I'_1, I'_2 \})$ is a subset of a $\sim_{\Phi}$ class. Indeed, let $u \in L_I(H)$, and let $I_i \in (\partial \Psi)^{-1}(I'_i)$. Let $v \in L_{t(u)}(H)$ be such that $I'_1 \cdot uv = I'_2 \cdot uv = J'$. Let $J_i = I_i \cdot uv$. Then $J_1, J_2 \in (\partial \Psi)^{-1}(J')$. Since $\Psi$ is synchronizing, there exists $\gamma \in L_{J'}(K)$ with $J_1 \cdot \gamma = J_2 \cdot \gamma$. Let $w = \Delta(\gamma)$. Then $I_1 \cdot uvw = I_2 \cdot uvw$. Since $u \in L_I(H)$ was arbitrary, it follows that $I_1 \sim_{\Phi} I_2$ as claimed. Since, by assumption, a $\sim_{\Phi}$ class is precisely a $\partial \Psi$-fiber of a single state of $K$, we must have $I'_1 = I'_2$ as claimed. Therefore $\sim_{\Delta}$ is indeed trivial.
\end{proof}
\begin{lemma}\label{lemma-pullback-stab}
Let $G,K,H$ be graphs with $H \leq_R K \leq_R G$. Let $\Psi \in \hom_R(G,K)$, $\Delta \in \hom_R(K,H)$, and let $\Phi = \Delta \circ \Psi$. Let $I \in V(H)$, let $I'_1, I'_1 \in (\partial \Delta)^{-1}(I)$, and let $I_i \in (\partial \Psi)^{-1}(I'_i)$. If $I_1 \sim_{\Phi} I_2$, then $I'_1 \sim_{\Delta} I'_2$.
\end{lemma}
\begin{proof}
Suppose that $I_1 \sim_{\Phi} I_2$. Let $u \in L_I(H)$. Let $v \in L_{t(u)}(H)$ be such that $I_1 \cdot uv = I_2 \cdot uv$. Then, since $I'_i \cdot uv = \partial \Phi(I_i \cdot uv)$, we have $I'_1 \cdot uv = I'_2 \cdot uv$. Therefore $I'_1 \sim_{\Delta} I'_2$.
\end{proof}
\begin{proof}[Proof of \cref{thm-struct-sync-comp}(3)]
Suppose that $\sim_{\Delta}$ is trivial. Let $I \in V(H)$, let $I'_1, I'_2 \in (\partial \Delta)^{-1}(I)$ with $I'_1 \neq I'_2$, and let $I_i \in (\partial \Psi)^{-1}(I'_i)$. Since $\sim_{\Delta}$ is trivial, there exists $u \in L_I(H)$ such that, for every $v \in L_{t(u)}(H)$, we have $I'_1 \cdot uv \neq I'_2 \cdot uv$. This implies that $I_1 \cdot uv \neq I_2 \cdot uv$, so $I_1 \not\sim_{\Phi} I_2$ by \cref{lemma-pullback-stab}. It follows that each $\sim_{\Phi}$ class is contained inside a $\partial \Psi$ fiber. By \cref{lemma-stab-comp}, this shows that $\sim_{\Phi} = \sim_{\Psi}$.
\end{proof}
The final part of \cref{thm-struct-sync-comp} can be proved in the strongly connected case using symbolic dynamics, via the multiplicativity of degree under composition of right-resolvers; see \cite{lm-95-intro}, \S 9.1. Using the theory developed so far, we give a self-contained proof without the assumption of strong connectedness.
\begin{proof}[Proof of \cref{thm-struct-sync-comp}(4)]
Suppose that $\Psi, \Delta$ are synchronizing. Let $I \in V(H)$. Let $I_1, I_2 \in (\partial \Phi)^{-1}(I)$. We need to show that $I_1 \sim_{\Phi} I_2$. Let $u \in L_I(K)$. We need to find $v \in L_{t(u)}(H)$ such that $I_1 \cdot uv = I_2 \cdot uv$. Let $I'_i = \partial \Psi(I_i)$. Then $\partial \Delta(I'_1) = \partial \Delta(I'_2)$, so $I'_1 \sim_{\Delta} I'_2$ since $\Delta$ is synchronizing. Let $v_1 \in L_{t(u)}(H)$ be such that $I'_1 \cdot uv_1 = I'_2 \cdot uv_1$. Note that $I'_i \cdot uv_1 = \partial \Psi(I_i \cdot uv_1)$. Let $J' = I'_i \cdot uv_1$ and $J_i = I_i \cdot uv_1$. Then $\partial \Psi(J_1) = J' = \partial \Psi(J_2)$, so $J_1 \sim_{\Psi} J_2$ since $\Psi$ is synchronizing. Let $\gamma \in L_{J'}(K)$ be such that $J_1 \cdot \gamma = J_2 \cdot \gamma$. Let $v_2 = \Delta(\gamma)$. Then $I_i \cdot uv_1 v_2 = J_i \cdot \gamma$, so taking $v = v_1 v_2$, we have $I_1 \cdot uv = I_2 \cdot uv$. Thus $I_1 \sim_{\Phi} I_2$, so $\Phi$ is indeed synchronizing.
We prove the converse by the contrapositive. Suppose that $\Delta$ is not synchronizing. Then there exist $I \in V(H)$ and $I'_1, I'_2 \in (\partial \Delta)^{-1}(I)$ such that $I'_1 \not\sim_{\Delta} I'_2$. By \cref{lemma-pullback-stab}, there exist $I_i \in (\partial \Psi)^{-1}(I'_i)$ such that $I_1 \not\sim_{\Phi} I_2$. Thus $\Phi$ is not synchronizing. Similarly, suppose that $\Psi$ is not synchronizing. Then there exist $I' \in V(K)$ and $I_1, I_2 \in (\partial \Psi)^{-1}(I')$ such that $I_1 \not\sim_{\Psi} I_2$. By \cref{lemma-stab-comp}, $I_1 \not\sim_{\Phi} I_2$, so $\Phi$ is not synchronizing.
\end{proof}
\subsection{Proofs of Propositions \ref{prop-alm-bunchy-equiv} and \ref{cor-alm-stab-unique}}\label{sec-bunchy-details}
\begin{proof}[Proof of \cref{prop-alm-bunchy-equiv}]
Let $M = M(G)$. First, suppose that $G$ is almost bunchy. Let $\Phi_1, \Phi_2 \in \hom_R(G,M)$. For each $I,J \in V(M)$, and each $I' \in \Sigma_G^{-1}(I)$, let $A_{I',J} = \{ e \in E_{I'}(G) \, | \, \Sigma_G(t(e)) = J \}$. If there exists $I' \in \Sigma_G^{-1}(I)$ with $|F(I') \cap \Sigma_G^{-1}(J)| \geq 2$ (by almost bunchiness, there is at one most such $I'$ for any given $I,J$), then let $\tau_{IJ} \in P(M)$ be the permutation of parallel edges in $M$ given by
\begin{align*}
\tau_{IJ}|_{E(M) \setminus E_{IJ}(M)} &= \mathrm{id}|_{E(M) \setminus E_{IJ}(M)} \\
%
\tau_{IJ}|_{E_{IJ}(M)} &= \Phi_1|_{A_{I',J}} \circ (\Phi_2|_{A_{I',J}})^{-1}
\end{align*}
Distinct $\tau_{IJ}$ are permutations of disjoint sets and therefore commute. Let $\tau = \prod_{I,J \in V(M)} \tau_{IJ}$. (Note that if $G$ is bunchy, then we do not define any $\tau_{IJ}$ above, so the product is empty and $\tau = \mathrm{id}$.) Now, for each $I,J \in V(M)$ and each $I' \in \Sigma_G^{-1}(I)$ with $|F(I') \cap \Sigma_G^{-1}(J)| = 1$, let $\sigma_{IJ,I'} \in P(G)$ be given by
\begin{align*}
\sigma_{IJ,I'}|_{E(G) \setminus A_{I',J}} &= \mathrm{id}|_{E(G) \setminus A_{I',J}} \\
%
\sigma_{IJ,I'}|_{A_{I',J}} &= \Phi_2|_{A_{I',J}}^{-1} \circ \tau^{-1} \circ \Phi_1|_{A_{I',J}}
\end{align*}
All of the $\sigma_{IJ,I'}$ commute. Let $\sigma = \prod_{I,J \in V(M), \, I' \in \Sigma_G^{-1}(I)} \sigma_{IJ,I'}$. Then for $I \in V(M)$, $J \in F(I)$, and $I' \in \Sigma_G^{-1}(I)$,
\[
\tau \circ \Phi_2 \circ \sigma|_{A_{I',J}} = \tau \circ \Phi_2 \circ (\Phi_2|_{A_{I',J}}^{-1} \circ \tau^{-1} \circ \Phi_1|_{A_{I',J}} ) = \Phi_1|_{A_{I',J}}
\]
This concludes the proof in the ``only if'' direction.
For the ``if'' direction, which we prove in the contrapositive, suppose that $G$ is not almost bunchy. Let $I,J \in V(M)$ and $I_1, I_2 \in \Sigma_G^{-1}(I)$ such that $|F(I_i) \cap \Sigma_G^{-1}(J)| \geq 2$ for $i=1,2$. Let $e_{i,1}, e_{i,2} \in A_{I_i,J}$ be such that $t(e_{i,1}) \neq t(e_{i,2})$. Let $a_1, a_2 \in E_{IJ}(M)$, and let $\Phi_1, \Phi_2 \in \hom_R(G,M)$ be such that $\Phi_1(e_{1,j}) = a_j$ but $\Phi_2(e_{2,1}) = a_2$ and $\Phi_2(e_{2,2}) = a_1$. (The behaviour of $\Phi_i$ on $E(G) \setminus \{ e_{i,j} \}_{i,j=1,2}$ is irrelevant.) Then there do not exist $\sigma \in P(G)$, $\tau \in P(M)$ such that $\Phi_1 = \tau \circ \Phi_2 \circ \sigma$.
Finally, suppose that $G$ is almost bunchy, but not bunchy. Let $I,J \in V(M)$ and $I' \in \Sigma_G^{-1}(I)$ such that $|F(I') \cap \Sigma_G^{-1}(J)| \geq 2$. Let $e_1, e_2 \in A_{I', J}$ be such that $t(e_1) \neq t(e_2)$. Let $a_1, a_2 \in E_{IJ}(M)$, and let $\Phi_1, \Phi_2 \in \hom_R(G,M)$ be such that $\Phi_1(e_i) = a_i$, $\Phi_2(e_1) = a_2$, and $\Phi_2(e_2) = a_1$. Then there does not exist $\sigma \in P(G)$ such that $\Phi_1 = \Phi_2 \circ \sigma$.
\end{proof}
\begin{proof}[Proof of \cref{cor-alm-stab-unique}]
Let $M = M(G)$. Let $\sigma \in P(G)$, $\tau \in P(M)$ such that $\Phi_1 = \tau \circ \Phi_2 \circ \sigma$. It is enough to show that $S_{\Phi_1}, S_{\Phi_2}$ share a generating set $T$, which we now construct and examine. For $I,J \in V(M)$, if $J \notin F(I)$, let $T_{I,J,1} = T_{I,J,2} = \emptyset$. If $J \in F(I)$, then let $T_{I,J,i}$ be the set of maps $f_{a,i}: \Sigma_G^{-1}(I) \to \Sigma_G^{-1}(J)$ of the form $I' \mapsto t \circ (\Phi_i|_{E_{I'}})^{-1}(a)$, i.e. $I' \cdot f_{a,i} = I' \cdot a$ with respect to $\Phi_i$, where $a \in E_{IJ}(M)$. Clearly $T_{I,J,i}$ generates $S_{\Phi_i}$, in the sense that $S_{\Phi_i}$ is the smallest collection of maps closed under composition and containing the $T_{I,J,i}$ as $I,J$ range over $V(M)$.
We claim that $T_{I,J,1} = T_{I,J,2}$. Indeed, let $a \in E_{IJ}(M)$ and let $I' \in \Sigma_G^{-1}(I)$ such that $|F(I') \cap \Sigma_G^{-1}(J)| \geq 2$. It is enough to show that $f_{a,1} \in T_{I,J,2}$, as this will show that $T_{I,J,1} \subseteq T_{I,J,2}$, from which equality follows by symmetry. By almost bunchiness, there is at most one state $I' \in \Sigma_G^{-1}(I)$ with $|F(I') \cap \Sigma_G^{-1}(J)| \geq 2$. If there is no such state, then clearly $f_{a,1} = f_{a,2}$, so assume that such a state $I'$ exists. Observe that $I' \cdot f_{b, 2} = I' \cdot f_{a,1}$ where $b = \Phi_2 \circ (\Phi_1|_{E_{I'(G)}})^{-1} (a)$. Moreover, for every $I'' \in \Sigma_G^{-1}(I)$ with $I'' \neq I'$, we also have, by almost bunchiness, that $I'' \cdot f_{b,2} = I'' \cdot f_{a,1}$. Therefore $f_{a,1} = f_{b,2} \in T_{I,J,2}$, so indeed $T_{I,J,1} = T_{I,J,2}$ as claimed. Let $T_{I,J} = T_{I,J,i}$. Then $T_{I,J}$ generates both $S_{\Phi_1}, S_{\Phi_2}$, so indeed $S_{\Phi_1} = S_{\Phi_2}$.
\end{proof}
\subsection{Proof of \cref{thm-univ-prop}}\label{subsec-univ-prop-proof}
\begin{proof}[Proof of \cref{thm-univ-prop}]
Define $T: V(G) \to V(P)$ as follows: for $I' \in V(G)$, let $T(I') = (\partial \Phi_1(I'), \partial \Phi_2(I'))$. Note that $\partial \Psi_P \circ T = \partial(\Psi_i \circ \Phi_i) = \Sigma_G$ for each $i$. Let $C$ be the subgraph of $P$ induced by $T(V(G))$. Let $I' \in V(G)$ and $J' \in F(I')$; $I = \Sigma_G(I')$ and $J = \Sigma_G(J')$; and $I_i = \partial \Phi_i(I')$ and $J_i = \partial \Phi_i(J')$. Observe that $(I_1, I_2) = T(I')$ and $(J_1, J_2) = T(J')$.
As in the proof of \cref{prop-alm-bunchy-equiv}, let $A_{I',J} = \{ e \in E_{I'}(G) \, | \, \Sigma_G(t(e)) = J \}$. Since $H_i$ is bunchy, we have $F(I_i) \cap \Sigma_{H_i}^{-1}(J) = \{ J_i \}$, so $\Phi_i(A_{I', J}) = E_{I_i J_i}(H_i)$. It therefore makes sense to define
\[
\Delta_i|_{A_{I',J}} = (\Hat{\Psi}_i|_{E_{(I_1,I_2)(J_1, J_2)}(P)})^{-1} \circ \Phi_i|_{A_{I', J}}: A_{I',J} \to E_{(I_1, I_2)(J_1, J_2)}(P)
\]
Gluing these together, we obtain maps $\Delta_i: E(G) \to E(P)$. For $e \in A_{I',J}$, we have $s(\Delta_i(e)) = T(s(e))$ and $t(\Delta_i(e)) = T(t(e))$, so the $\Delta_i: G \to P$ are graph homomorphisms with $\partial \Delta_i = T$.
We now claim that $\Delta_i(E(G)) = E(C)$. Enumerate $E_{IJ}(M) = \{ a^{(1)}, \dots, a^{(k)} \}$, and let $e_i^{(j)} = (\Psi_i|_{E_{I_i J_i}(H_i)})^{-1}(a^{(j)})$. Then
\[
(e_1^{(j)}, e_2^{(j)}) \in E_{(I_1, I_2)(J_1, J_2)}(P) = E_{(I_1, I_2)(J_1, J_2)}(C)
\]
where equality holds because $(I_1, I_2), (J_1, J_2) \in V(C)$ and $C$ is an induced subgraph of $P$. Fixing $(I_1,I_2)$ and varying $J'$, thus varying $(J_1, J_2)$, the sets $E_{(I_1, I_2)(J_1, J_2)}(P)$ exhaust $E_{(I_1, I_2)}(P)$, so $E_{(I_1, I_2)}(C) = E_{(I_1, I_2)}(P)$. Since $\Delta_i|_{E_{I'}(G)}: E_{I'}(G) \to E_{(I_1, I_2)}(P)$ is surjective, we have $E_{(I_1, I_2)}(C) = \Delta_i(E_{I'}(G))$. Thus indeed $E(C) = \Delta_i(E(G))$.
Since we already know that $E_{(I_1, I_2)}(C) = E_{(I_1, I_2)}(P)$ for each $(I_1, I_2) \in V(C)$, it follows that $C$ is indeed a principal subgraph of $P$. Finally, we have
\[
\Hat{\Psi}_i \circ \Delta_i|_{A_{I',J}} = \Hat{\Psi}_i \circ (\Hat{\Psi}_i|_{E_{(I_1, I_2)(J_1, J_2)}(P)})^{-1} \circ \Phi_i|_{A_{I', J}} = \Phi_i|_{A_{I', J}}
\]
as claimed. This shows that $\Hat{\Psi}_i|_C: C \to H_i$ is surjective, so $H_i \leq_R C$.
\end{proof}
\subsection{Proofs of \cref{lemma-fiber-bunchy} and \cref{prop-bg-exists}, and construction of $B(G)$}\label{subsec-fiber-max-bunchy}
\begin{proof}[Proof of \cref{lemma-fiber-bunchy}]
Let $(I_1, I_2) \in V(P)$ with $I_i \in V(H_i)$. Let $e = (e_1, e_2), e' = (e_1', e_2') \in E_{(I_1, I_2)}(P)$ such that $t( \partial \Psi_P (e)) = t( \partial \Psi_P (e'))$. We must show that $t(e) = t(e')$. Toward that goal, note that for each $i =1,2$, we have
\begin{align*}
t(\partial \Psi_P(e)) &= \partial \Psi_P(t(e)) = \partial \Psi_i(t(e_i)) \\
%
t(\partial \Psi_P(e')) &= \partial \Psi_P(t(e')) = \partial \Psi_i(t(e'_i))
\end{align*}
Therefore $\partial \Psi_i(t(e_i)) = \partial \Psi_i(t(e'_i))$ for each $i$, since $t( \partial \Psi_P (e)) = t( \partial \Psi_P (e'))$ by hypothesis. But since the $H_i$ are bunchy, we in fact have $t(e_i) = t(e_i')$ for each $i$, so indeed $t(e) = t(e')$ as required. This shows that $\partial \Psi_P|_{F((I_1, I_2))}$ is a bijection, so $P$ is bunchy.
Let $C$ be a principal subgraph of $P$ such that $\partial \Hat{\Psi}_i|_{V(C)} : V(C) \to V(H_i)$ are surjective. Then $M(C) = M$. For each $J \in V(M)$, we have $\Sigma_C^{-1}(J) = \Sigma_P^{-1}(J) \cap V(C)$. Thus $C$ is bunchy.
\end{proof}
\begin{proof}[Proof of \cref{prop-bg-exists}]
To prove (1), let $H_1, H_2 \leq_R G$ be $\leq_R$-maximal among the bunchy right-resolving factors of $G$. Let $\Phi_i \in \hom_R(G,H_i)$ and $\Psi_i \in \hom_R(H_i, M)$. Let $P = H_1 \times_{\Psi_1, \Psi_2} H_2$. By \cref{thm-univ-prop}, there exist a principal subgraph $C$ of $P$, and $\Delta_i \in \hom_R(G,C)$, such that $\Phi_i = \Hat{\Psi}_i|_C \circ \Delta_i$, so $H_i \leq_R C$. By \cref{lemma-fiber-bunchy}, $C$ is bunchy, so by the maximality of the $H_i$, we have $H_1 = C = H_2$. This proves uniqueness, so we can take $B = H_i = C$.
To prove (2), let $H \leq_R G$ be bunchy and let $\Phi \in \hom_R(G,H)$. Let $\Phi' \in \hom(G,B)$, $\Psi \in \hom_R(H,M)$, $\Psi' \in \hom(B,M)$. Let $P = H \times_{\Phi, \Phi'} B$. By the universal property (\cref{thm-univ-prop}), there exist a principal subgraph $C$ of $P$ and $\Delta, \Delta' \in \hom_R(G,C)$ with $\Phi = \Hat{\Psi} \circ \Delta$ and $\Phi' = \Hat{\Psi}' \circ \Delta'$. Again by \cref{lemma-fiber-bunchy}, $C$ is bunchy, so by (1) and the fact that $B \leq_R C$, we have $B = C$. This proves (2) with $\Theta = \Hat{\Psi}'|_C$.
\end{proof}
We now present the construction of $B(G)$ described in Algorithm \ref{alg-construct-bg}.
\begin{prop}\label{prop-construct-bg}
Let $G$ be a graph. For $J_1, J_2 \in V(G)$ with $\Sigma_G(J_1) = \Sigma_G(J_2)$, write $J_1 \approx_0 J_2$ if there exist paths $\gamma = \gamma_1 \cdots \gamma_n, \delta = \delta_1 \cdots \delta_n \in L(G)$ where $\gamma_i, \delta_i \in E(G)$, such that $s(\gamma_1) = s(\delta_1)$, $t(\gamma_n) = J_1$, $t(\delta_n) = J_2$, and $\Sigma_G(t(\gamma_i)) = \Sigma_G(t(\delta_i))$ for each $i$. Let $\approx$ denote the transitive closure of $\approx_0$ and let $\Phi \in \hom_R(G,M(G))$. Then $\approx$ is a congruence with respect to $\Phi$, and $B(G) = G/\approx$.
\end{prop}
\begin{proof}
Let $M = M(G)$. Let $I \in V(M)$, $I_1, I_2 \in \Sigma_G^{-1}(I)$, and $J \in F(I)$. Suppose that $I_1 \approx I_2$. Let $J_i \in F(I_i) \cap \Sigma_G^{-1}(J)$. We need to show that $J_1 \approx J_2$. Let $I_1 = I^{(0)}, I^{(1)}, \dots, I^{(n)} = I_2 \in \Sigma_G^{-1}(I)$ with $I^{(j)} \approx_0 I^{(j+1)}$. Let $\gamma^{(j)}, \delta^{(j)} \in L(G)$ witness the relation $I^{(j)} \approx_0 I^{(j+1)}$. Choose $J^{(j)} \in F(I^{(j)}) \cap \Sigma_G^{-1}(J)$, with $J^{(0)} = J_1$ and $J^{(n)} = J_2$. Let $e^{(j)} \in E_{I^{(j)} J^{(j)}}(G)$. Then $\gamma^{(j)} e^{(j)}, \delta^{(j)} e^{(j+1)}$ witness $J^{(j)} \approx_0 J^{(j+1)}$. This shows that $J_1 \approx J_2$.
Let $\Phi \in \hom_R(G,M(G))$. As in the previous paragraph, let $I \in V(M)$, let $J \in F(I)$, and let $I_1, I_2 \in \Sigma_G^{-1}(I)$. Suppose that $I_1 \approx I_2$. Let $a \in E_{IJ}(M)$, let $J_i = I_i \cdot a$, and let $e_i = (\Phi|_{E_{I_i}(G)} )^{-1}(a)$. Then $t(e_i) = J_i$, so $J_1 \approx J_2$. Therefore $\approx$ is indeed a congruence for $\Phi$. It follows that $F([I_i]_{\approx}) \cap \Sigma_{G/\approx}^{-1}(J) = \{ [J_i]_{\approx} \}$, so indeed $G/\approx$ is bunchy.
To see that $G/\approx$ is $\leq_R$-maximal among the bunchy factors of $G$, let $H \leq_R G$ be bunchy and $\Psi \in \hom_R(G,H)$. If $I_1 \approx_0 I_2$ and $H$ is bunchy, then we must have $\partial \Psi(I_1) = \partial \Psi(I_2)$. Therefore the partition into $\partial \Psi$-fibers corresponds to an equivalence relation that coarsens the symmetric, reflexive relation $\approx_0$, and thus also coarsens the transitive closure $\approx$. Considering $V(H)$ as a partition of $V(G)$, we must therefore have $V(H) \preceq V(G)/\approx$, so $(G/\approx) \, = B(G)$ by the maximality of $B(G)$.
\end{proof}
\subsection{Proofs of Propositions \ref{prop-bfc-equiv-new} and \ref{cor-air-fact-ext-mg}}\label{subsec-fiber-fact-ext}
\begin{lemma}\label{lemma-princ-fiber-embed}
Let $G_1, G_2, H, K$ be graphs with $K \leq_R H \leq_R G_i$. Let $\Delta \in \hom_R(H,K)$. Let $\Psi_i \in \hom_R(G_i, H)$ and $P = G_1 \times_{\Psi_1, \Psi_2} G_2$. Let $\Phi_i = \Delta \circ \Psi_i$ and $Q = G_1 \times_{\Phi_1, \Phi_2} G_2$. Then, noting that $V(P) = V(Q)$ and $E(P) \subseteq E(Q)$, we have $\Hat{\Phi}_i|_P = \Hat{\Psi}_i$, and every principal subgraph of $P$ is a principal subgraph of $Q$.
\end{lemma}
\begin{proof}
Let $C$ be a principal subgraph of $P$. By the definition of the fiber product, we have $V(P) = V(Q) = V(G_1) \times V(G_2)$. Let $I \in V(H)$ and let $I_i \in (\partial \Psi_i)^{-1}(I)$. Suppose that $(I_1, I_2) \in V(C)$. We need to show that $E_{(I_1, I_2)}(C) = E_{(I_1, I_2)}(Q)$.
Since $E_{(I_1, I_2)}(C) = E_{(I_1, I_2)}(P)$ by the definition of a principal subgraph, it is enough to show that $E_{(I_1, I_2)}(P) = E_{(I_1, I_2)}(Q)$ for any $(I_1, I_2) \in V(P)$. Clearly $E_{(I_1, I_2)}(P) \subseteq E_{(I_1, I_2)}(Q)$, so it is enough to show that $|E_{(I_1, I_2)}(P)| = |E_{(I_1, I_2)}(Q)|$. To see this equality, note that since $\Delta \circ \Psi_P = \Phi_Q$, we have
\[
|E_{(I_1, I_2)}(P)| = |E_{\partial \Psi_P(I_1, I_2)}(H)| = |E_{\partial \Phi_Q(I_1,I_2)}(K)| = |E_{(I_1,I_2)}(Q)|
\]
where the equalities follow from the facts that $\Psi_P$, $\Delta$, and $\Phi_Q$ respectively are right-resolving. Therefore $E_{(I_1, I_2)}(P) = E_{(I_1, I_2)}(Q)$. This shows that $C$ is indeed a principal subgraph of $Q$. Moreover, for $e = (e_1, e_2) \in E(P)$, we have $\Hat{\Phi}_i(e) = e_i = \Hat{\Psi}_i(e)$, so indeed $\Hat{\Phi}_i|_P = \Hat{\Psi}_i$.
\end{proof}
\begin{lemma}\label{prop-air-fact-ext}
Let $G_1,G_2, K$ be graphs with $K \leq_S G_i$ via $\Phi_i \in \hom_S(G_i,K)$. Let $C$ be a principal subgraph of $P = G_1 \times_{\Phi_1, \Phi_2} G_2$ such that the $\partial \Hat{\Phi}_i|_{V(C)}: V(C) \to V(G_i)$ are surjective. Then $\Hat{\Phi}_i|_C \in \hom_S(C,G_i)$. In particular, $C$ is a common synchronizing extension of the $G_i$.
\end{lemma}
\begin{proof}
Let $I \in V(K)$ and let $I_i, I'_i \in V(G_i)$ with $\partial \Phi_i(I_i) = \partial \Phi_i(I'_i) = I$. Since the $\Phi_i$ are synchronizing, we have $I_i \sim_{\Phi_i} I'_i$. Suppose that $(I_1, I_2), (I'_1, I'_2) \in V(C)$. In order to show that $\Hat{\Phi}_i|_C \in \hom_S(C,G_i)$, we claim that $(I_1, I_2) \sim_{\Phi_P} (I'_1, I'_2)$. This will show that $\Phi_P|_C \in \hom_S(C,K)$. Since $\Phi_P|_C = \Phi_i \circ \Hat{\Phi}_i|_C$, it will then follow that $\Hat{\Phi}_i|_C \in \hom_S(C,G_i)$ by \cref{thm-struct-sync-comp}(4).
To prove this claim, let $u \in L_I(K)$. Since $I_1 \sim_{\Phi_1} I'_1$, there exists $v_1 \in L_{t(u)}(K)$ such that $I_1 \cdot uv_1 = I'_1 \cdot uv_1$. Similarly, since $I_2 \sim_{\Phi_2} I'_2$, there exists $v_2 \in L_{t(v_1)}(K)$ such that $I_2 \cdot uv_1 v_2 = I'_2 \cdot uv_1 v_2$. Then in particular $(I_1, I_2) \cdot u v_1 v_2 = (I'_1, I'_2) \cdot u v_1 v_2$. Since $u \in L_I(K)$ was arbitrary, we have $(I_1, I_2) \sim_{\Phi_P} (I'_1, I'_2)$ as claimed.
\end{proof}
\begin{proof}[Proof of \cref{cor-air-fact-ext-mg}(1)]
Let $K \leq_S G_i$ and $\Psi_i \in \hom_S(G_i,K)$. Let $P = G_1 \times_{\Psi_1, \Psi_2} G_2$ and let $C$ be a principal subgraph of $P$ such that $\partial \Hat{\Psi}_i|_{V(C)}: V(C) \to V(G_i)$ are surjective. By \cref{prop-air-fact-ext}, we have $\Hat{\Psi}_i|_C \in \hom_S(C,G_i)$. Let $\Delta \in \hom_R(K,M)$, let $\Phi_i = \Delta \circ \Psi_i$, and let $Q = G_1 \times_{\Phi_1, \Phi_2} G_2$. Then, by \cref{lemma-princ-fiber-embed}, $C$ is a principal subgraph of $Q$, with $\Hat{\Phi}_i|_C = \Hat{\Psi}_i|_C$. In particular, $\Hat{\Phi}_i|_C \in \hom_S(C,G_i)$.
\end{proof}
We now give an equivalent form of the $O(G)$ conjecture. Fragments of this result appear in \cref{cor-air-fact-ext-mg}. The logical structure of \cref{prop-og-equiv} (one statement is equivalent to the equivalence of four other statements) is unusual.
\begin{prop}\label{prop-og-equiv}
Let $\mathcal{F}$ be a family of graphs satisfying the following conditions:
\begin{enumerate}[label=(\roman*)]
\item If $G \in \mathcal{F}$ and $H \leq_R G$, then $H \in \mathcal{F}$.
\item Let $G_1, G_2, K \in \mathcal{F}$ with $K \leq_R G_i$. Let $\Phi_i \in \hom_R(G_i, K)$, let $P = G_1 \times_{\Phi_1, \Phi_2} G_2$, and let $C$ be a principal subgraph of $P$ such that the $\partial \Hat{\Phi}_i|_{V(C)}$ are surjective. Then $C \in \mathcal{F}$.
\end{enumerate}
Then the following assertions are equivalent.
\begin{enumerate}[label=(\arabic*)]
\item For any $G \in \mathcal{F}$, there exists a unique $\leq_S$-minimal graph $O(G) \leq_S G$.
\item For any $G_1, G_2 \in \mathcal{F}$, the following assertions are equivalent.
\begin{enumerate}[label=(\alph*)]
\item $O(G_1), O(G_2)$ exist and are equal.
\item $G_1, G_2$ have a common synchronizing factor.
\item $M(G_1) = M(G_2) = M$ and there exist $\Phi_i \in \hom_R(G_i,M)$ such that $\Hat{\Phi}_i \in \hom_S(C,G_i)$ for some principal subgraph $C$ of $Q = G_1 \times_{\Phi_1, \Phi_2} G_2$.
\item $G_1, G_2$ have a common synchronizing extension $K \in \mathcal{F}$.
\end{enumerate}
\end{enumerate}
\end{prop}
\begin{rmk}
The proof of \cref{prop-og-equiv} in fact shows that both (1) and (2) are equivalent to the assertion that (d) implies (b). Note that for $G_1 = G_2 = G$, (a) states that $O(G)$ is well-defined, while (b)--(d) are trivial.
\end{rmk}
\begin{rmk}
The $O(G)$ conjecture states that the equivalent statements (1) and (2) in \cref{prop-og-equiv} hold with $\mathcal{F}$ equal to the class of all strongly connected graphs.
\end{rmk}
\begin{proof}[Proof of \cref{prop-og-equiv}]
First, assume (2). Let $G$ be a graph and let $H_1, H_2 \leq_S G$ be $\leq_S$-minimal. Since $H_1, H_2$ have the common synchronizing extension $G$, they satisfy condition (d), so they also satisfy condition (b), i.e. there exists a common synchronizing factor $K \leq_S H_i$. But since the $H_i$ were assumed minimal, we must have $H_1 = K = H_2 = O(G)$.
Now, assume (1) and deduce (2) as follows. Trivially, (a) implies (b) and (c) implies (d). Moreover, (b) implies (c) by \cref{cor-air-fact-ext-mg}(1). Finally, assume (d). Suppose that $G_1, G_2$ have a common synchronizing extension $K$. Then $O(G_1) = O(K) = O(G_2)$, so (d) implies (a).
\end{proof}
\begin{proof}[Proof of \cref{cor-air-fact-ext-mg}(2)]
This is immediate from \cref{prop-og-equiv}, specifically, the equivalence of 2(b) and 2(c), with $\mathcal{F}$ taken to be the class of all strongly connected graphs.
\end{proof}
\begin{proof}[Proof of \cref{cor-air-fact-ext-mg}(3)]
By \cref{thm-alm-og-exists}, this is immediate from \cref{prop-og-equiv}, again via the equivalence of 2(b) and 2(c), but with $\mathcal{F}$ taken to be the class of strongly connected almost bunchy graphs.
\end{proof}
\begin{prop}\label{prop-bfc-equiv}
Let $\mathcal{F}$ be a family of graphs such that, if $G \in \mathcal{F}$ and $H \leq_R G$, then $H \in \mathcal{F}$. Then the following assertions are equivalent.
\begin{enumerate}[label=(\arabic*)]
\item Any $\leq_S$-minimal graph $H \in \mathcal{F}$ is bunchy.
\item For any $G \in \mathcal{F}$, there exists some bunchy $H \leq_S G$.
\item For any non-bunchy $G \in \mathcal{F}$, there exists some $\Phi \in \hom_R(G,M(G))$ with $\sim_{\Phi}$ nontrivial.
\item For any $G \in \mathcal{F}$, $B(G) \leq_S G$.
\end{enumerate}
\end{prop}
\begin{rmk}
Note that \cref{prop-bfc-equiv} is a more detailed version of \cref{prop-bfc-equiv-new}.
\end{rmk}
\begin{proof}
To see that (1) implies (2), let $G \in \mathcal{F}$ and consider the set $\{ H \, | \, H \leq_S G \}$. Being a finite partially ordered set, this set must have at least one minimal element. If (1) holds, then this minimal element is bunchy. Thus (2) holds.
To see that (2) implies (3), let $G \in \mathcal{F}$ be non-bunchy. There exists at least one $\leq_S$-minimal graph $H \leq_S G$. If (2) holds, then $G$ is not $\leq_S$-minimal, so $H \neq G$. Let $\Psi \in \hom_S(G,H)$. The $\partial \Psi$-fibers are precisely the $\sim_{\Psi}$-classes. Since $H \neq G$, the $\partial \Psi$-fibers are not merely singletons, so $\sim_{\Psi}$ is nontrivial. Let $M=M(G)$ and $\Delta \in \hom_R(H,M)$, and let $\Phi = \Delta \circ \Psi$. By \cref{lemma-stab-comp}, the $\sim_{\Phi}$-classes are unions of $\sim_{\Psi}$-classes, so in particular, $\sim_{\Phi}$ is nontrivial.
To see that (3) implies (1), let $G \in \mathcal{F}$ be non-bunchy. If (3) holds, then there exists $\Phi \in \hom_R(G,M(G))$ with $\sim_{\Phi}$ nontrivial. Then $G/\sim_{\Phi} \leq_S G$ and $G/\sim_{\Phi} \neq G$; in particular, $G$ is not $\leq_S$-minimal. This proves (1) in the contrapositive.
Finally, we show that (2) and (4) are equivalent. If (4) holds, then (2) holds since $B(G)$ is bunchy. Conversely, assume (2) and let $G \in \mathcal{F}$. Then there exists $H \leq_S G$ bunchy. Let $\Phi \in \hom_S(G,H)$ and let $B = B(G)$. By \cref{prop-bg-exists}(2) and \cref{thm-struct-sync-comp}(4), we have $\Phi = \Delta \circ \Theta$ for some $\Theta \in \hom_S(G,B)$, $\Delta \in \hom_S(B,H)$. Therefore (4) holds.
\end{proof}
\begin{cor}\label{cor-bunchy-implies-o}
Let $\mathcal{F}$ be a family of graphs such that, if $G \in \mathcal{F}$ and $H \leq_R G$, then $H \in \mathcal{F}$. Suppose that any $\leq_S$-minimal element of $\mathcal{F}$ is bunchy. Then $O(G)$ exists for any $G \in \mathcal{F}$.
\end{cor}
The proof of \cref{cor-bunchy-implies-o} is a trivial adaptation of the proof of \cref{prop-bunchy-implies-o}.
\begin{proof}[Proof of \cref{cor-air-fact-ext-mg}(4)]
Assuming the bunchy factor conjecture, the $O(G_i)$ are well-defined. By the equivalence of conditions (a),(c) in \cref{prop-bfc-equiv}, the hypothesis on the $B(G_i)$ is equivalent to the equality $O(B(G_1)) = O(B(G_2))$, which in turn is equivalent to the equality $O(G_1) = O(G_2)$ by \cref{prop-bfc-equiv}.
\end{proof}
\section{Proof of \cref{thm-bfc-cyc-bunch}, following Trahtman}\label{sec-traht}
In this section, we recall Trahtman's proof of the road colouring theorem and reformulate it in a form applicable to \cref{thm-bfc-cyc-bunch}.
\subsection{Systems of maps with unique tallest trees}\label{subsec-traht-utt}
Let $\{ V_i \}_{0 \leq_R i \leq_R p-1}$ be disjoint finite sets, with $p \in \mathbb{N}$. Let $a_i: V_i \to V_{i+1}$ be maps, with subscripts read modulo $p$. For $k \geq 0$, let $b_k = a_k a_{k+1} \cdots a_{p-1} a_0 \cdots a_{k-1}$. For $k \geq 0$ and $I \in V_k$, there exist $0 \leq_R \ell \leq_R p-1$, $m \geq 0$, $z \geq 1$, such that $I \cdot a_k \cdots a_{k+\ell-1} b_{k+\ell}^{m+z} = I \cdot a_k \cdots a_{k+\ell-1} b_{k+\ell}^{m}$. Consider the lexicographically minimal such $(\ell,m,z)$, and let $h(I) = mp+\ell$, $z(I) = z$. We refer to $\rho(I) = I \cdot a_k \cdots a_{k+\ell-1} b_{k+\ell}^{m+z}$ as the \textit{root} of $I$ and $h(I)$ as the \textit{height} of $I$. Let $h_{\max,k} = \max \{ h(I) \, | \, I \in V_k \}$, and let $h_k(J) = \max \{ h(I) \, | \, I \in V_k, \, \rho(I) = J \}$. Let $z_k = \mathrm{lcm} \{ z(I) \, | \, I \in V_k \}$. We say that the system $(V_i, a_i)_{0 \leq_R i \leq_R p-1}$ has a \textit{unique tallest tree} at $V_k$ if there is a unique $J$ with $h_k(J) = h_{\max,k}$. Note that the terms we have defined here still make sense even if the $V_i$ are not all pairwise disjoint, as long as any two are either equal or disjoint, since we can make them disjoint by replacing $V_i$ with $V_i \times \{ i \}$.
We now present our interpretation of a key step in Trahtman's proof, applying \cref{lemma-min-diff-stab} and closely following \cite{bpr-09-ca}.
\begin{lemma}\label{lemma-utt-stab}
Let $G,H$ be strongly connected graphs with $H \leq_R G$. Let $\Phi \in \hom_R(G,H)$. Let $I_0, \dots I_{p-1} \in V(H)$, not necessarily distinct, be such that $I_{i+1} \in F(I_i) \neq \emptyset$, and let $a_i \in E_{I_i I_{i+1}}(H)$ (with subscripts read modulo $p$). Suppose that the system $((\partial \Phi)^{-1}(I_i), a_i)_{0 \leq_R i \leq_R p-1}$ has a unique tallest tree, where we write $a_i$ for the map $I' \mapsto I' \cdot a_i$, $I' \in (\partial \Phi)^{-1}(I_i)$. Then $\sim_{\Phi}$ is nontrivial.
\end{lemma}
\begin{proof}
Suppose WLOG that the system has a unique tallest tree at $(\partial \Phi)^{-1}(I_0)$. Let $I' \in (\partial \Phi)^{-1}(I_0)$ be a state of maximal height $h(I) = h_{\max,0}$. Let $h_{\max,0} = mp + k$ where $k = z(I)$. Let $R = \rho(I)$. Note that $R \in (\partial \Phi)^{-1}(I_k)$.
By strong connectedness and \cref{lemma-sc-min-img}, let $U \subseteq (\partial \Phi)^{-1}(I_0)$ be a minimal image such that $I' \in U$. We claim that there is no other state $I'' \in U$ with $\rho(I'') = R$ and $I'' \neq I'$. Indeed, suppose that there is such a state $I''$. Then $I' \cdot a_0 \cdots a_{k-1} b_k^m = I'' \cdot a_0 \cdots a_{k-1} b_k^m = R$, so $| U \cdot a_0 \cdots a_{k-1} b_k^m | < |U|$, contradicting the minimality of $U$. This proves the claim. Let $U_0 = U \setminus \{ I' \}$. Then every element of $U_0$ has height strictly less than $H_{\max,0}$.
Let $u_1 = a_0 \cdots a_{k-1} b_k^{m-1} a_k \cdots a_{k-2}$, where the tail $a_k \cdots a_{k-2}$ includes each $a_i$ exactly once, other than $a_{k-1}$, and the subscripts are read modulo $p$. Let $u_2 = u_1 b_{k}^{z_k} = a_0 \cdots a_{k-1} b_{k-1}^{m + z_k} a_k \cdots a_{k-2}$. Let $J_i = I' \cdot u_i$ and $U_i = U \cdot u_i$. Observe that $J_1 \neq J_2$ by the assumed value of $h(I')$. However, since $I'$ is the unique element of $U$ with this maximal height, we have $U_1 \Delta U_2 = \{ J_1, J_2 \}$. Since the $U_i$ are minimal images, we have $J_1 \sim_{\Phi} J_2$ by \cref{lemma-min-diff-stab}.
\end{proof}
\subsection{Obtaining a right-resolver with a unique tallest tree}\label{subsec-traht-app}
Let $G$ be a graph. A \textit{total order colouring} is a total ordering of each edge set $E_I(H)$, i.e. a labeling of the edges of $G$ such that, if $|E_{I}(G)| = k$, then the edges in $E_{I}(G)$ are labeled bijectively by $\{ 0, \dots, k-1 \}$. Suppose that $M = M(G)$ is a cycle of bunches. Then, once a total order colouring of $M$ is fixed, total order colourings of $G$ correspond bijectively with right-resolvers $\Phi \in \hom_R(G,M)$.
Letting $V(M) = \{ I_0, \dots, I_{p-1} \}$, there is exactly one edge $a_i \in E_{I_i}(H)$ labeled $0$ for each $i$. This yields a subgraph $W$ of $G$ consisting of edges labeled $0$, which is a spanning subgraph of $G$ of constant out-degree $1$. Every graph of constant out-degree $1$ consists of a set of state-disjoint cycles, together with trees rooted on the cycles, directed toward their roots. The height of a tree is the maximum path length from a state in the tree to its root. Observe that the system $( (\partial \Phi)^{-1}(I_i), a_i )_{i=0}^{p-1}$ has a unique tallest tree, in the sense of mappings, if and only there is a unique tallest tree in $W$.
We now present our interpretation of the main technical lemma in the proof of the road colouring theorem (Lemma 10.4.6 in \cite{bpr-09-ca}). The following is not how the lemma is stated in \cite{bpr-09-ca}, but one can follow the proof and observe that it is equivalent.
\begin{lemma}[Trahtman]\label{lemma-traht-utt}
Let $G$ be a strongly connected graph such that $M(G)$ is a cycle of bunches. At least one of the following is true:
\begin{enumerate}
\item $G$ is itself a cycle of bunches.
\item $G$ has two distinct bunches whose outgoing edges have the same target.
\item $G$ admits a coarse total order colouring with a unique tallest tree.
\end{enumerate}
\end{lemma}
With this result, we can prove our generalization of the road colouring theorem.
\begin{proof}[Proof of \cref{thm-bfc-cyc-bunch}]
Let $M = M(G)$. The claim is trivially true if $|V(G)| = |V(M)|$, in which case $G=M$. Suppose that it is true for all $H$ with $|V(H)| \leq_R N$ and $M(H) = M$. Suppose that $|V(G)| = N+1$. If $G$ is bunchy, then we are done. If $G$ is not bunchy, but has two states that can be in-amalgamated, then they are stable for some $\Phi \in \hom_R(G,M)$ by \cref{lemma-amalg-stab}. Let $G' = G/\sim_{\Phi}$. Then $|V(G/\sim_{\Phi})| < |V(G)|$, so by the inductive hypothesis, there is some bunchy $H \leq_S G/\sim_{\Phi} \, \leq_S G$.
Now, suppose that $G$ does not have two states that can be in-amalgamated---in particular, $G$ does not have two distinct bunches whose outgoing edges have the same target. Since $G$ is not bunchy, it is in particular not a cycle of bunches, so by \cref{lemma-traht-utt}, it admits a total order colouring with unique tallest tree. As remarked above, this total order colouring corresponds to some right-resolver $\Phi \in \hom_R(G,M)$, which has $\sim_{\Phi}$ nontrivial by \cref{lemma-utt-stab}. Then once more $G/\sim_{\Phi}$ is strictly smaller than $G$. We can then apply \cref{prop-bfc-equiv} to conclude that any $\leq_S$-minimal $O$ with $M =M(O)$ is a cycle of bunches. If $O \leq_S G$, then $O=O_{M,p}$ for $p=\mathrm{per}(G)/\mathrm{per}(M)$.
\end{proof}
\section*{Acknowledgments}
Many thanks to Brian Marcus and Tom Meyerovitch for their patient, generous advice and supervision. Thanks also to Mike Boyle for suggesting (to Brian Marcus) a renewed attack on the $O(G)$ problem, shortly after the road problem was solved in 2007. The author was supported through NSERC Discovery Grant EJYR GR010163 (PI: Brian Marcus).
|
\section{Introduction}
Graph learning models are widely used in science, such as physics~\cite{bapst2020unveiling} and biochemistry~\cite{jumper2021highly}. In many such disciplines, building more accurate predictive models is typically not the only goal. It is often more crucial for scientists to discover the patterns from the data that induce certain predictions~\cite{cranmer2020discovering}. For example, identifying the functional groups in a molecule that yield its certain properties may provide insights to guide further experiments~\cite{wencel2013c}.
Recently, graph neural networks (GNNs) have become almost the de fato graph learning models due to their great expressive power~\cite{kipf2016semi, xu2018powerful}. However, their expressivity is often built upon a highly non-linear entanglement of irregular graph features. So, it is often quite challenging to figure out the patterns in the data that GNNs use to make predictions.
Many works have been recently proposed to extract critical data patterns for the prediction by interpreting GNNs in post-hoc ways~\cite{ying2019gnnexplainer,yuan2020xgnn,vu2020pgm,luo2020parameterized,schlichtkrull2021interpreting,yuan2021explainability,lin2021generative,henderson2021improving}.
They work on a pre-trained model and propose different types of combinatorial search methods to detect the subgraphs of the input data that affect the model predictions the most.
In contrast to the above post-hoc methods, inherently interpretable models have been rarely investigated for graph learning tasks. There are two main concerns regarding such models. First, the prediction accuracy and inherent interpretability of a model often forms a trade-off~\cite{du2019techniques}. Practitioners may not allow sacrificing prediction accuracy for better interpretability. Second, the attention mechanism, a widely-used technique to provide inherent interpretability, often cannot provide faithful interpretation~\cite{lipton2018mythos}. The rationale of the attention mechanism is to learn weights for different features during the model training, and the rank of the learned weights can be interpreted as the importance of certain features~\cite{bahdanau2014neural, xu2015show}. However, recent extensive evaluations in NLP tasks \cite{serrano2019attention,jain2019attention,mohankumar2020towards} have shown that the attention may not weigh the features that dominate the model output more than other features. In particular, for graph learning tasks, the widely-used graph attention models~\cite{velivckovic2018graph,li2015gated} seem unable to provide any reliable interpretation of the data~\cite{ying2019gnnexplainer,yu2020graph}.
Along another line of research, invariant learning~\cite{pearl2016causal,arjovsky2019invariant,chang2020invariant,krueger21a} has been proposed to provide inherent interpretability and better generalizability. They argue that the models na\"{i}vely trained over biased data may risk capturing spurious correlations between the input environment features and the labels, and thus suffer from severe generalization issues. So, they propose to train models that align with the causal relations between the signal features and the labels.
However, such training approaches to match causal relations typically have high computational complexity.
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[trim={0cm 0cm 0.8cm 0cm},clip,width=1\columnwidth]{img/arch.pdf}}
\vspace{-2mm}
\caption{The architecture of GSAT\xspace. $g_{\phi}$ encodes the input graph $G$ and learns stochastic attention $\alpha$ (from Bernoulli distributions) that randomly drop the edges and obtain a perturbed graph $G_S$. $f_{\theta}$ encodes $G_S$ to make predictions. GSAT\xspace does not constrain the size of $G_S$ but injects stochasticity to constrain information. The subgraph of $G_S$ with learnt reduced-stochasticity (edges with $p_e \rightarrow 1$) provides interpretation. GSAT\xspace is a unified model by adopting just one GNN for both $g_{\phi}$ and $f_{\theta}$. GSAT\xspace can be either trained from scratch or start from a pre-trained GNN predictor $f_{\theta}$.}
\label{fig:arch}
\end{center}
\vspace{-11mm}
\end{figure}
In this work, we are to address the above concerns by proposing \emph{Graph Stochastic Attention} (GSAT\xspace), a novel attention mechanism to build inherently interpretable and well generalizable GNNs. The rationale of GSAT\xspace roots in the notion of information bottleneck (IB)~\cite{tishby2000information,tishby2015deep}. We formulate the attention as an IB by injecting stochasticity into the attention to constrain the information flow from the input graph to the prediction~\cite{shannon1948mathematical}. Such stochasticity over the label-irrelevant graph components will be kept during the training while that over the label-relevant ones can automatically get reduced. This difference eventually provides model interpretation. By penalizing the amount of information from the input data, GSAT\xspace is also expected to be more generalizable.
Our study achieves the following observations and contributions. First, the IB principle frees GSAT\xspace from any potentially biased assumptions adopted in previous methods such as the size or the connectivity constraints on the detected graph patterns. Even when those assumptions are satisfied, GSAT\xspace still works the best without using such assumptions, while when those assumptions are not satisfied, GSAT\xspace achieves significantly better interpretation. See the sampled interpretation result visualizations in Fig.~\ref{fig:size-visual} and Fig.~\ref{fig:connect-visual}. Second, from the perspective of IB, all post-hoc interpretation methods are suboptimal. They essentially optimize a model without any information control and then perform a single-step projection to an information-controlled space, which makes the final interpretation performance sensitive to the pre-trained models. Third, by reducing the information from the input graph, GSAT\xspace can provably remove spurious correlations in the training data under certain assumptions and achieve better generalization. Fourth, if a pre-trained model is provided, GSAT\xspace may further improve both of its interpretation and prediction accuracy.
We evaluate GSAT\xspace in terms of both interpretability and label-prediction performance. Experiments over 8 datasets show that GSAT\xspace outperforms the state-of-the-art (SOTA) methods by up to $20\%$$\uparrow$ in interpretation AUC and $5\%$$\uparrow$ in prediction accuracy. Notably, GSAT\xspace achieves the SOTA performance on \emph{molhiv} on OGB~\cite{hu2020ogb} among the models that do not use manually-designed expert features.
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1.033\linewidth]{img/viz_mnist.pdf}}
\vspace{-1mm}
\centerline{\includegraphics[width=1.05\linewidth]{img/viz_mnist_gm.pdf}}
\end{center}
\vspace{-10mm}
\caption{Visualizing attention (normalized to $[0,1]$) of GSAT\xspace (second row) v.s. masks of GraphMask~\cite{schlichtkrull2021interpreting} (third row) on MNIST-75sp. The first row shows the ground-truth. Different digit samples contain interpretable subgraphs of different sizes, while GSAT\xspace is not sensitive to such varied sizes.}
\label{fig:size-visual}
\end{figure}
\begin{figure}[t]
\begin{center}
\vspace{-2mm}
\centerline{\includegraphics[trim={2cm 0cm 1cm 1.7cm},clip,width=1.0\linewidth]{img/viz_ba_3motifs.pdf}}
\end{center}
\vskip -10mm
\caption{Visualizing attention (normalized to $[0,1]$) of GSAT\xspace (first row) and masks of GraphMask~\cite{schlichtkrull2021interpreting} (second row) on a motif example, where graphs with three house motifs and graphs with two house motifs represent two classes. Samples may contain disconnected interpretable subgraphs, while GSAT\xspace detects them accurately. More details can be found in Appendix~\ref{appx:sup_exp}.}
\label{fig:connect-visual}
\vspace{-3mm}
\end{figure}
\section{Preliminaries}
As preliminaries, we define a few notations and concepts.
\textbf{Graph.} An attributed graph can be denoted as $G =(A,X)$ where $A$ is the adjacency matrix and $X$ includes node attributes. Let $V$ and $E$ denote the node set and the edge set, respectively. We focus on graph-level tasks: A training set of graphs with their labels $(G^{(i)}, Y^{(i)})$, $i=1,...,n$ are given, where each sample $(G^{(i)}, Y^{(i)})$ is assumed to be IID sampled from some unknown distribution $\mathbb{P}_{\mathcal{Y}\times\mathcal{G}}=\mathbb{P}_{\mathcal{Y}|\mathcal{G}}\mathbb{P}_{\mathcal{G}}$.
\textbf{Label-relevant Subgraph.} A label-relevant subgraph refers to the subgraph $G_S$ of the input graph $G$ that mostly indicates the label $Y$. For example, to determine the solubility of a molecule, the hydroxy group -OH is a positive-label-relevent subgraph, as if it exists, the molecule is often soluble to the water. Finding label-relevant subgraphs is a common goal of interpretable graph learning.
\textbf{Attention Mechanism.}
Attention mechanism has been widely used in interpretable neural networks for NLP and CV tasks~\cite{bahdanau2014neural, xu2015show, vaswani2017attention}.
However, GNNs with attention~\cite{velivckovic2018graph} often generate low-fidelity attention weights. As it learns multiple weights for every edge, it is far from trivial to combine those weights with the irregular graph structure to perform graph label-relevant feature selection.
There are two types of attention models: One normalizes the attention weights to sum to one~\cite{bahdanau2014neural}, while the other learns weights between $[0,1]$ without normalization~\cite{xu2015show}. As the counterparts in GNN models, GAT adopts the normalized one~\cite{velivckovic2018graph} while GGNN adopts the unnormalized one~\cite{li2015gated}. Our method belongs to the second category.
\textbf{Graph Neural Network.} GNNs are neural network models that encode graph-structured data into node representations or graph representations. They initialize each node feature representation with its attributes $h_v^{(0)}=X_v$ and then gradually update it by aggregating representations from its neighbors, i.e., $h_v^{(l+1)}\leftarrow q(h_v^{(l)}, \{h_{u}^{(l)}|u:(u,v)\in E\})$ where $q(\cdot)$ denotes a function implemented by NNs~\cite{gilmer2017neural}. Graph representations are often obtained via an aggregation (sum/mean) of node representations.
\textbf{Learning to Explain (L2X).}
L2X~\cite{chen2018learning} studies the feature selection problem in the regular feature space and proposed a mutual information (MI) maximization rule to select a fixed number of features. Specifically, let $I(a;b) \triangleq \sum_{a,b} \mathbb{P}(a,b)\log \frac{\mathbb{P}(a,b)}{\mathbb{P}(a)\mathbb{P}(b)}$ denote the MI between two random variables $a$ and $b$. Large MI indicates certain high correlation between two random variables. Hence, with input features $X\in\mathbb{R}^F$, L2X is to search a $k$-sized set of indices $S\subseteq\{1,2,...,F\}$, where $k=|S|<F$, such that the features in the subspace indexed by $S$ (denoted by $X_S$) maximizes the mutual information with the labels $Y$, i.e.,
\begin{align} \label{eq:L2X}
\max_{S\subseteq\{1,2,...,F\}}\; I(X_S;Y), \quad \text{s.t.} \; |S|\leq k.
\end{align}
Our model is inspired by L2X. However, as graph features and their interpretable counterparts are in an irregular space without a fixed dimension, directly applying L2X may achieve subpar performance in graph learning tasks. We propose to use information constraint instead in Sec.~\ref{sec:obj}.
Later, we will also use the \emph{entropy} defined as $H(a)$ $\triangleq - \sum_{a} \mathbb{P}(a)\log \mathbb{P}(a)$ and the \emph{KL-divergence} defined as $\text{KL}(\mathbb{P}(a)||\mathbb{Q}(a)) \triangleq \sum_{a} \mathbb{P}(a) \log \frac{\mathbb{P}(a)}{\mathbb{Q}(a)} $~\cite{cover1999elements}.
\section{Graph Learning Interpretation via GIB}
In this section, we will first propose the GIB-based objective for interpretable graph learning and point out the issues of post-hoc GNN interpretation methods.
\subsection{GIB-based Objective for Interpretation} \label{sec:obj}
Finding label-relevant subgraphs in graph learning tasks has unique challenges. As for the irregularity of graph structures, graph learning models often have to deal with the input graphs of various sizes. The critical subgraph patterns may be also of different sizes and be highly irregular. Consider the example of molecular solubility again, although the functional groups for positive solubility such as -OH, -NH$_2$ are of similar sizes, those for negative solubility range from small groups (e.g., -Cl) to extremely large ones (e.g. -C$_{10}$H$_9$). And, a molecule may contain multiple functional groups scattered in the graph that determine its properties.
Given these observations, it is not proper to just mimic the cardinality constraint used for a regular dimension space (Eq.~\eqref{eq:L2X}) and select subgraphs of certain sizes potentially with a connectivity constraint as done in~\cite{ying2019gnnexplainer}. Inspired by the graph information bottleneck (GIB) principle~\cite{wu2020graph,yu2020graph}, we propose to use information constraint instead to select label-relevant subgraphs, i.e., solving \vspace{-1mm}
\begin{align} \label{eq:GIB}
\max_{G_S} I(G_S ; Y), \text{s.t.} \; I(G_S;G) \leq \gamma, G_S\in \mathbb{G}_{sub}(G)
\end{align}
where $\mathbb{G}_{sub}(G)$ denotes the set of the subgraphs of $G$. Note that GIB does not impose any potentially biased constraints such as the size or the connectivity of the selected subgraphs. Instead, GIB uses the information constraint $I(G_S;G) \leq \gamma$ to select $G_S$ that inherits only the most indicative information from $G$ to predict the label $Y$ by maximizing $I(G_S ; Y)$. As thus, $G_S$ provides model interpretation.
\citet{yu2020graph} also considered using GIB to select subgraphs. However, we adopt a fundamentally different mechanism that we will provide a detailed comparison in Sec.~\ref{sec:comparison}.
\begin{figure}[t]
\vspace{-1mm}
\begin{center}
\centerline{\includegraphics[trim={0.5cm 0cm 2.5cm 0.5cm},clip,width=1.0\columnwidth]{img/projection.pdf}}
\end{center}
\vspace{-13mm}
\caption{Post-hoc methods just perform one-step projection to the information-constrained space, which is always suboptimal and the interpretation performance is sensitive to the pre-trained model.}
\vspace{-1mm}
\label{fig:projection}
\end{figure}
\begin{figure*}[t]
\vspace{-1mm}
\centering
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.49\linewidth]{img/ba_2motifs_xroc.pdf}
\includegraphics[width=0.49\linewidth]{img/ba_2motifs_loss.pdf}
\vspace*{-4mm}
\caption{Ba-2Motifs}
\label{fig:ba2}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.49\linewidth]{img/mutag_xroc.pdf}
\includegraphics[width=0.49\linewidth]{img/mutag_loss.pdf}
\vspace*{-4mm}
\caption{Mutag}
\label{fig:mutag}
\end{subfigure}
\vspace*{-2mm}
\caption{Issues of post-hoc interpretation methods. All methods are trained with $10$ random seeds; post-hoc methods are also provided with models pre-trained with different seeds.
Interpretation performance and the training losses of Eq.~\ref{eq:GIB} for GSAT\xspace and Eq.~\ref{eq:posthoc} for others are shown. We guarantee that all the pre-trained models are well-trained in their pre-training stage (Acc. $\sim$100\% Ba-2Motif, $\sim$90\% Mutag). }
\label{fig:post-hoc-fail}
\vspace*{-2mm}
\end{figure*}
\subsection{Issues of Post-hoc GNN Interpretation Methods} \label{sec:post-hoc}
Almost all previous GNN interpretation methods are post-hoc, such as GNNExplainer~\cite{ying2019gnnexplainer}, PGExplainer~\cite{luo2020parameterized} and GraphMask~\cite{schlichtkrull2021interpreting}.
Given a pre-trained predictor $f_{\theta}(\cdot): \mathcal{G}\rightarrow \mathcal{Y}$, they try to find out the subgraph $G_S$ that impacts the model predictions the most, while keeping the pre-trained model unchanged. This procedure essentially first maximizes the MI between $f_{\theta}(G)$ and $Y$ and obtains a model parameter\vspace{-1mm}
\begin{align} \label{eq:pretrain}
\tilde{\theta} \triangleq \arg\max_{\theta} I(f_{\theta}(G);Y),
\end{align}
and then optimizes a subgraph extractor $g_{\phi}$ via
\begin{align} \label{eq:posthoc}\vspace{-1mm}
\tilde{\phi} \triangleq\arg\max_{\phi} I(f_{\tilde{\theta}}(G_S);Y), \, \text{s.t.}\; G_S=g_{\phi}(G)\in \Omega.
\end{align}
where $\Omega$ implies a subset of the subgraphs $\mathbb{G}_{sub}(G)$ that satisfy some constraints, e.g., the cardinality constraint adopted by GNNExplainer and PGExplainer. Let us temporarily ignore the difference between different constraints and just focus on the optimization objective. The post-hoc objective Eq.~\eqref{eq:posthoc} and GIB (Eq.~\eqref{eq:GIB}) share some similar spirits. However, the post-hoc methods may not give or even approximate the optimal solution to Eq.~\eqref{eq:GIB} because $f_{\theta}\circ g_{\phi}$ is not jointly trained. From the optimization perspective, post-hoc methods just perform \emph{one-single step projection} (see Fig.~\ref{fig:projection}) from the model $f_{\tilde{\theta}}$ in an unconstrained space to $f_{\tilde{\theta}}\circ g_{\tilde{\phi}}$ in the information-constrained space $\Omega$ where the projection rule follows that the induced MI decrease $I(f_{\tilde{\theta}}(G);Y) - I(f_{\tilde{\theta}}(g_{\tilde{\phi}}(G));Y)$ gets minimized.
In practice, such a suboptimal behavior will yield two undesired consequences. First, $f_{\tilde{\theta}}$ may not fully extract the information from $G_S=g_{\phi}(G)$ to predict $Y$ during the optimization of Eq.~\eqref{eq:posthoc} because $f_{\tilde{\theta}}$ is originally trained to make $I(f_{\tilde{\theta}}(G);Y)$ approximate $I(G,Y)$ while $(G_S, Y) = (g_{\phi}(G), Y)$ follows a distribution different from $(G,Y)$. Therefore, $I(f_{\tilde{\theta}}(G_S);Y)$ may not well approximate $I(G_S;Y)$,
and thus may mislead the optimization of $g_{\phi}$ and disable $g_{\phi}$ to select $G_S$ that indeed indicates $Y$. GNNExplainer suffers from this issue over Ba-2Motif as shown in Fig.~\ref{fig:post-hoc-fail}: The training loss, $-I(f_{\tilde{\theta}}(G_S);Y)$ keeps high and the interpretation performance is subpar. It is possible to further decrease the training loss via a more aggressive optimization of $g_\phi$. However, the models may risk overfitting the data, which yields the second issue.
An aggressive optimization of $g_{\phi}$ may give a large \emph{empirical} MI $\hat{I}\left(f_{\tilde{\theta}}(g_{\phi}(G)); Y\right)$ (or a small training loss equivalently) by selecting features that help to distinguish labels for training but are essentially irrelevant to the labels or spuriously correlated with the labels in the population level.
Previous works have shown that label-irrelevant features are known to be discriminative enough to even identify each graph in the training dataset let alone the labels~\cite{suresh2021adversarial}. Empirically, we indeed observe such overfitting problems of all post-hoc methods over Mutag as shown in Fig.~\ref{fig:post-hoc-fail}, especially PGExplainer and GraphMask. In the first $5$ to $10$ epochs, these two models succeed in selecting good explanations while having a large training loss. Further training successfully decreases the loss (after $10$ epochs) but degenerates the interpretation performance substantially.
This might also be the reason why in the original literatures of these post-hoc methods, training over only a small number of epochs is suggested. However,
in practical tasks, it is hard to have the ground truth interpretation labels to verify the results and decide a trusty stopping criterion.
Another observation of Fig.~\ref{fig:post-hoc-fail} also matches our expectation: From the optimization perspective, post-hoc methods suffer from an initialization issue. Their interpretability can be highly sensitive to the pre-trained model $f_{\tilde{\theta}}$, as empirically demonstrated by the large variances in Fig.~\ref{fig:post-hoc-fail}. Only if the pre-trained $f_{\tilde{\theta}}$ approximates the optimal $f_{\theta^*}$, the performance can be roughly guaranteed. So, a joint training of $f_{\theta}\circ g_{\phi}$ according to the GIB principle Eq.~\eqref{eq:GIB} is typically needed.
\section{Stochastic Attention Mechanism for GIB} \label{sec:method}
In this section, we will first give a tractable variational bound of the GIB objective (Eq.~\eqref{eq:GIB}), and then introduce our model GSAT\xspace with the stochastic attention mechanism. We will further discuss how the stochastic attention mechanism improves both model interpretation and generalization.
\subsection{A Tractable Objective for GIB}
GSAT\xspace is to learn an extractor $g_{\phi}$ with parameter $\phi$ to extract $G_S\in \mathbb{G}_{\text{sub}}(G)$. $g_{\phi}$ blocks the label-irrelevant information in the data $G$ via injected stochasticity while allowing the label-relevant information kept in $G_S$ to make predictions. In GSAT\xspace, $g_{\phi}(G)$ essentially gives a distribution over $\mathbb{G}_{\text{sub}}(G)$. We also denote this distribution as $\mathbb{P}_{\phi}(G_S|G)$. Later, $g_{\phi}(G)$ and $\mathbb{P}_{\phi}(G_S|G)$ are used interchangeably.
Putting the constraint into the objective (Eq.\eqref{eq:GIB}), we obtain the optimization of $g_{\phi}$ via GIB, i.e., for some $\beta>0$,
\begin{align} \label{eq:GIB2}
\min_{\phi} -I(G_S ; Y) + \beta I(G_S;G),\, \text{s.t.}\,\, G_S \sim g_{\phi}(G).
\end{align}
Next, we follow~\citet{alemi2016deep, poole2019variational,wu2020graph} to derive a tractable variational upper bound of the two terms in Eq.~\eqref{eq:GIB2}. Detailed derivation is given in Appendix~\ref{appx:deriving}. For the term $I\left(G_S ; Y\right)$, we introduce a parameterized variational approximation $\mathbb{P}_{\theta}(Y|G_S)$ for $\mathbb{P}(Y|G_S)$. We obtain a lower bound:
\begin{align} \label{eq:predictor}
I\left(G_S ; Y\right) \geq \mathbb{E}_{G_S, Y} \left[ \log {\mathbb{P}_{\theta}(Y|G_S)} \right] + H(Y).
\end{align}
Note that $\mathbb{P}_{\theta}(Y|G_S)$ essentially works as the predictor $f_{\theta}:\mathcal{G}\rightarrow \mathcal{Y}$ with parameter $\theta$ in our model.
For the term $I(G_S;G)$, we introduce a variational approximation $\mathbb{Q}(G_S)$ for the marginal distribution $\mathbb{P}(G_S) = \sum_{G}\mathbb{P}_{\phi}(G_S|G)\mathbb{P}_{\mathcal{G}}(G)$. And, we obtain an upper bound:
\begin{align} \label{eq:extractor}
I\left(G_s ; G\right) \leq \mathbb{E}_{G}\left[\text{KL}(\mathbb{P}_{\phi}(G_S|G)||\mathbb{Q}(G_S)) \right]
\end{align}
Plugging in the above two inequalities, we obtain a variational upper bound of Eq.~\eqref{eq:GIB2} as the objective of GSAT\xspace:
\begin{align} \nonumber
&\min_{\theta,\phi}\,-\mathbb{E}\left[ \log {\mathbb{P}_{\theta}(Y|G_S)} \right] + \beta \mathbb{E}\left[\text{KL}(\mathbb{P}_{\phi}(G_S|G)||\mathbb{Q}(G_S)) \right],
\\ &\quad\quad \text{s.t.}\quad G_S \sim \mathbb{P}_{\phi}(G_S|G). \label{eq:proj}
\end{align}
Next, we specify $\mathbb{P}_{\theta}$ (aka $f_{\theta}$), $\mathbb{P}_{\phi}$ (aka $g_{\phi}$) and $\mathbb{Q}$ in GSAT\xspace.
\subsection{GSAT\xspace and Stochastic Attention Mechanism}
For clarity, we introduced the predictor $f_{\theta}$ and the extractor $g_{\phi}$ separately. Actually, GSAT\xspace is a unified model as $f_{\theta},\,g_{\phi}$ share the same GNN encoder except their last layers.
\textbf{Stochastic Attention via $\mathbb{P}_{\phi}$.} The extractor $g_{\phi}$ first encodes the input graph $G$ via the GNN into a set of node representations $\{h_v|v\in V\}$. For each edge $(u,v)\in E$, $g_{\phi}$ contains an MLP layer plus sigmoid that maps the concatenation $(h_u,h_v)$ into $p_{uv}\in [0,1]$. Then, for each forward pass of the training, we sample stochastic attention from Bernoulli distributions $\alpha_{uv}\sim \text{Bern}(p_{uv})$. To make sure the gradient w.r.t. $p_{uv}$ is computable, we apply the gumbel-softmax reparameterization trick~\cite{jang2016categorical}. The extracted graph $G_S$ will have an attention-selected subgraph as $A_S = \alpha \odot A$. Here $\alpha$ is the matrix with entries $\alpha_{uv}$ for $(u,v)\in E$ or zeros for the non-edge entries. $A$ is the adjacency matrix of $G$ and $\odot$ is entry-wise product. The distribution of $G_S$ given $G$ through the above procedure characterizes $\mathbb{P}_{\phi}(G_S|G)$, so $\mathbb{P}_{\phi}(G_S|G) = \prod_{u,v\in E}\mathbb{P}(\alpha_{uv}|p_{uv})$, where $p_{uv}$ is a function of $G$. This essentially makes the attention $\alpha_{uv}$ to be conditionally independent across different edges given the input graph $G$.
\textbf{Prediction via $\mathbb{P}_{\theta}$.} The predictor $f_{\theta}$ adopts the same GNN to encode the extracted graph $G_S$ to a graph representation, and finally passes such representation through an MLP layer plus softmax to model the distribution of $Y$. This procedure gives the variational distribution $\mathbb{P}_{\theta}(Y|G_S)$.
\textbf{Marginal Distribution Control via $\mathbb{Q}$.}
The bound Eq.\eqref{eq:extractor} is always true for any $\mathbb{Q}(G_S)$. We define $\mathbb{Q}(G_S)$ as follows. For every graph $G\sim \mathbb{P}_{\mathcal{G}}$ and every two directed node pair $(u,v)$ in $G$, we sample $\alpha_{uv}'\sim \text{Bern}(r)$ where $r\in[0,1]$ is a hyperparameter.
We remove all edges in $G$ and add all edges $(u,v)$ if $\alpha_{uv}'=1$.
Suppose the obtained graph is $G_S$. This procedure defines the distribution $\mathbb{Q}(G_S) = \sum_{G}\mathbb{P}(\alpha'|G)\mathbb{P}_{\mathcal{G}}(G)$. As $\alpha'$ is independent from the graph $G$ given its size $n$, $\mathbb{Q}(G_S) = \sum_{n}\mathbb{P}(\alpha'|n)\mathbb{P}_{\mathcal{G}}(G=n) = \mathbb{P}(n)\prod_{u,v=1}^n \mathbb{P}(\alpha_{uv}')$. The probability of an $n$-sized graph $\mathbb{P}(n)$ is a constant and thus will not affect the model. Note that our choice of $\mathbb{Q}(G_S)$ shares the similar spirit of using standard Gaussian as the latent distribution with variational auto-encoders~\cite{kingma2013auto}.
Using the above $\mathbb{P}_{\theta}$, the first term in Eq.\eqref{eq:proj} reduces to a standard cross entropy loss. Using $\mathbb{P}_{\phi}$ and $\mathbb{Q}$, the KL-divergence term becomes, for every $G\sim \mathbb{P}_{\mathcal{G}}$, $n$ as the size of $G$,
\begin{align}\label{eq:reg}
&\text{KL}(\mathbb{P}_{\phi}(G_S|G)||\mathbb{Q}(G_S)) = \\
&\sum_{(u,v)\in E} p_{uv} \log \frac{p_{uv}}{r} + \left(1-p_{uv}\right) \log \frac{1-p_{uv}}{1-r} + c(n,r). \nonumber
\end{align}
where $c(n,r)$ is a constant without any trainable parameters.
\subsection{The Interpretation Mechanism of GSAT\xspace}
The interpretability of GSAT\xspace essentially comes from the information control: GSAT\xspace decreases the information from the input graphs by injecting stochasticity via attention into $G_S$. In the training, the regularization term Eq.\eqref{eq:reg} would try to assign large stochasticity for all edges, yet driven by the classification loss $\min -I(G_S;Y)$ (equivalent to cross-entropy loss), GSAT\xspace can learn to reduce such stochasticity of the attention on the task-relevant subgraphs. So, it is not the entire $G_S$ but the part of $G_S$ with the stochasticity-reduced attention, aka $p_{uv} \rightarrow 1$, that provide model interpretation. Therefore, when GSAT\xspace provides interpretation, in practice, one can rank all edges according to $p_{uv}$ and use those top ranked ones (given a certain budget if needed) as the detected subgraph for interpretation.
The contribution of injecting stochasticity to the performance is so significant as shown in experiments (Table~\ref{table:ab-beta-noise-gin}), so is the contribution of our regularization term (Eq.~\eqref{eq:reg}) when we compare it with the sparsity-driven $\ell_1$-norm (Fig.~\ref{fig:reg-compare}).
GSAT\xspace is substantially different from previous methods, as we do not use any sparsity constraints such as $\ell_1$-norm~\cite{ying2019gnnexplainer,luo2020parameterized}, $\ell_0$-norm~\cite{schlichtkrull2021interpreting} or $\ell_2$-regression to $\{0,1\}$~\cite{yu2020graph} to select size-constrained (or connectivity-constrained) subgraphs. We actually observe that setting $r$ away from 0 in the marginal regularization (Eq.~\eqref{eq:reg}), i.e., pushing $G_S$ away from being sparse often provides more robust interpretation.
This matches our intuition that GIB by definition does not make any assumptions on the selected subgraphs but just constrains the information from the original graphs. Our experiments show that GSAT\xspace outperform baselines significantly without leveraging those assumptions in the optimization even if the label-relevant subgraphs satisfy these assumptions. If the label-relevant subgraphs are indeed disconnected or vary in sizes, the improvement of GSAT\xspace is expected to be even more.
\subsection{Further Comparison on Interpretation Mechanism} \label{sec:comparison}
PGExplainer and GraphMask also have stochasticity in their models~\cite{luo2020parameterized,schlichtkrull2021interpreting}. However, their main goal is to enable a gradient-based search over a discrete subgraph-selection space rather than control the information as GSAT\xspace does. Hence, they did not in principle derive the information regularization as ours (Eq.~\eqref{eq:reg}) but adopt sparsity constraints to extract a small subgraph $G_S$ directly used for interpretation.
IB-subgraph~\cite{yu2020graph} considers using GIB as the objective but does not inject any stochasticity to generate $G_S$, so its selected subgraph $G_S$ is a deterministic function of $G$. Specifically, IB-subgraph samples batches of graphs $G$ to estimate $I(G_S;G)$ and optimize a deterministic function $G_S=g_{\phi}(G)$ to minimize such MI estimation.
In this case $I(G_S;G) (= H(G_S) - H(G_S|G))$ reduces to the entropy $H(G_S)$,
which tends to give a small-sized $G_S$, because the space of small graphs is small and has a lower upper bound of the entropy.
By contrast, $G_{S} \sim g_{\phi}(G)$ is random in GSAT\xspace, and GSAT\xspace implements GIB mainly by increasing $H(G_S|G)$ via injecting stochasticity.
\subsection{Guaranteed Spurious Correlation Removal}
\label{sec:better-generalization}
GSAT\xspace can remove spurious correlations in the training data and has guaranteed interpretability. We may prove that if there exists a correspondence between a subgraph pattern $G_S^*$ and the label $Y$,
the pattern $G_S^*$ is the optimal solution of the GIB objective (Eq.~\eqref{eq:GIB}).
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[trim={0.6cm 0.2cm 0.1cm 0cm},clip,width=0.8\columnwidth]{img/sp_correlation.pdf}}
\end{center}
\vspace{-10mm}
\caption{$G_S^*$ determines $Y$. However, the environment features in $G\backslash G_S^*$ may contain spurious (backdoor) correlation with $Y$.}
\label{fig:sp_correlation}
\vspace{-1mm}
\end{figure}
\begin{theorem}
\label{thm:iboptim}
Suppose each $G$ contains a subgraph $G_S^*$ such that $Y$ is determined by $G_S^*$ in the sense that $Y=f(G_S^*)+\epsilon$ for some deterministic invertible function $f$ with randomness $\epsilon$ that is independent from $G$. Then, for any $\beta\in [0,1]$, $G_S=G_S^*$ maximizes the GIB $I\left(G_S ; Y\right) - \beta I \left(G_S;G\right) $, where $G_S\in \mathbb{G}_{\text{sub}}(G)$.
\end{theorem}
\begin{proof}
Consider the following derivation:
\begin{align*}
& I(G_S; Y) - \beta I(G_S; G) \\
= & I(Y;G,G_S) - I(G;Y|G_S) - \beta I(G_S; G) \\
= & I(Y;G,G_S) - (1-\beta) I(G;Y|G_S) - \beta I(G;G_S,Y) \\
= & I(Y;G) - (1-\beta) I(G;Y|G_S) - \beta I(G;G_S,Y) \\
= & (1-\beta) I(Y;G) - (1-\beta) I(G;Y|G_S) - \beta I(G;G_S|Y),
\end{align*}
where the third equality is because $G_S\in \mathbb{G}_{sub}(G)$, then $(G_S, G)$ holds no more information than $G$.
If $\beta\in[0,1]$, $G_S$ that maximizes $I(G_S, Y) - \beta I(G_S; G)$ can also minimize $(1-\beta) I(G;Y|G_S) + \beta I(G;G_S|Y)$. As $I(G;Y|G_S)\geq 0$, $I(G;G_S|Y)\geq 0$, the lower bound of $(1-\beta) I(G;Y|G_S) + \beta I(G;G_S|Y)$ is 0.
$G_S^*$ is the subgraph that makes $(1-\beta) I(G;Y|G_S^*) + \beta I(G;G_S^*|Y)= 0$. This is because (a) $Y=f(G_S^*)+\epsilon$ where $\epsilon$ is independent of $G$ so $I(G;Y|G_S^*) = 0$ and (b) $G_S^* = f^{-1}(Y-\epsilon)$ where $\epsilon$ is independent of $G$ so $I(G;G_S^*|Y) = 0$. Therefore, $G_S=G_S^*$ maximizes GIB $I\left(G_S ; Y\right) - \beta I \left(G_S;G\right)$, where $G_S\in \mathbb{G}_{\text{sub}}(G)$.
\end{proof}
Although $G_S^*$ determines $Y$, in the training dataset the data $G$ and $Y$ may have some spurious correlation caused by the environment~\cite{pearl2016causal,arjovsky2019invariant,chang2020invariant,krueger21a}. That is, $G\backslash G_S^*$ may have some correlation with the label, but this correlation is spurious and is not the true reason that determines its label (illustrated in Fig.~\ref{fig:sp_correlation}). A model trained over $G$ to predict $Y$ via just MI maximization may capture such spurious correlation. If such correlation is changed during the test phase, the model suffers from performance decay.
However, Theorem~\ref{thm:iboptim} indicates that GSAT\xspace by optimizing the GIB objective has the capability to address the above issue by only extracting $G_S^*$, which removes the spurious correlation and also provides guaranteed interpretability.
\vspace{-1mm}
\subsection{Fine-tuning and Interpreting a Pre-trained Model} \label{sec:finetune}
GSAT\xspace can also fine-tune and interpret a pre-trained GNN. Given a GNN $f_{\tilde{\theta}}$ pre-trained by $\max _{\theta} I(f_{\theta}(G) ; Y)$, GSAT\xspace can fine-tune it via $\max _{\theta,\phi} I(f_{\theta}(G_S) ; Y) - \beta I(G_S;G)$, $G_S\sim g_{\phi}(G)$ by initializing the GNN used in $g_{\phi}$ and $f_{\theta}$ as the one in the pre-trained model $f_{\tilde{\theta}}$.
We observe that this framework almost never hurts the original prediction performance (and sometimes even boosts it). Moreover, this framework often achieves better interpretation results compared with training the GNN from scratch.
\section{Other Related Works}
Besides the models \cite{ying2019gnnexplainer,luo2020parameterized,schlichtkrull2021interpreting,yu2020graph} that we have compared with in detail in Sec.~\ref{sec:post-hoc} and Sec.~\ref{sec:comparison}, we review some other interpretation methods here.
Most previous works on GNN interpretation are post-hoc~\cite{ribeiro2016model}. Some works strongly rely on the connectivity assumption and only search over the space of connected subgraphs for interpretation. They adopt either reinforcement learning~\cite{yuan2020xgnn} or Monte Carlo tree search~\cite{yuan2021explainability}. Other methods including PGM-Explainer~\cite{vu2020pgm} leveraging graphical models, Gem~\cite{lin2021generative} checking Granger causality and Graphlime~\cite{huang2020graphlime} using HSIC Lasso are only applied to node-level task interpretation. Some works check the gradients w.r.t. the input features to find important features~\cite{pope2019explainability, baldassarre2019explainability}.
Much fewer works have considered intrinsic interpretation. Recently, \citet{anonymous2022discovering} has proposed DIR to make the model avoid overfitting spurious correlations and only capture invariant rationales to provide interpretability. However, DIR needs to iteratively break graphs into subgraphs and assemble subgraphs into graphs during the model training, which is far more complicated than GSAT\xspace.
\section{Experiments} \label{experiments}
\begin{table*}[t]
\vspace{-2mm}
\caption{Interpretation Performance (AUC). The \underline{underlined} results highlight the best baselines. The \textbf{bold} font and \textbf{bold}$^{\dagger}$ font highlight when GSAT\xspace outperform the means of the best baselines based on the mean of GSAT\xspace and the mean-2*std of GSAT\xspace, respectively. }
\vspace{-0.1cm}
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lcccccc}
\toprule
& \multirow{2}{*}{Ba-2motifs} & \multirow{2}{*}{Mutag} & \multirow{2}{*}{MNIST-75sp} & \multicolumn{3}{c}{Spurious-motif} \\
& & & & $b=0.5$ & $b=0.7$ & $b=0.9$ \\
\midrule
GNNExplainer & $67.35\pm3.29$ & $61.98\pm5.45$ & $59.01\pm2.04$ & $62.62\pm1.35$ & $62.25\pm3.61$ & $58.86\pm1.93$ \\
PGExplainer & $84.59\pm9.09$ & $60.91\pm17.10$ & $69.34\pm4.32$ & $69.54\pm5.64$ & $72.33\pm9.18$ & $\underline{72.34}\pm2.91$ \\
GraphMask & $\underline{92.54}\pm8.07$ & $62.23\pm9.01$ & $\underline{73.10}\pm6.41$ & $72.06\pm5.58$ & $73.06\pm4.91$ & $66.68\pm6.96$ \\
IB-Subgraph & $86.06\pm28.37$ & $\underline{91.04}\pm6.59$ & $51.20\pm5.12$ & $57.29\pm14.35$ & $62.89\pm15.59$ & $47.29\pm13.39$ \\
DIR & $82.78\pm10.97$ & $64.44\pm28.81$ & $32.35\pm9.39$ & $\underline{78.15}\pm1.32$ & $\underline{77.68}\pm1.22$ & $49.08\pm3.66$ \\
\midrule
GIN+GSAT\xspace & $\mathbf{98.74}^\dagger\pm0.55$ & $\mathbf{99.60}^\dagger\pm0.51$ & $\mathbf{83.36}^\dagger\pm1.02$ & $\mathbf{78.45}\pm3.12$ & $74.07\pm5.28$ & $71.97\pm4.41$ \\
GIN+$\text{GSAT\xspace}^*$ & $\mathbf{97.43}^\dagger\pm1.77$ & $\mathbf{97.75}^\dagger\pm0.92$ & $\mathbf{83.70}^\dagger\pm1.46$ & $\mathbf{85.55}^\dagger\pm2.57$ & $\mathbf{85.56}^\dagger\pm1.93$ & $\mathbf{83.59}^\dagger\pm2.56$ \\
\midrule
PNA+GSAT\xspace & $\mathbf{93.77}\pm3.90$ & $\mathbf{99.07}^\dagger\pm0.50$ & $\mathbf{84.68}^\dagger\pm1.06$ & $\mathbf{83.34}^\dagger\pm2.17$ & $\mathbf{86.94}^\dagger\pm4.05$ & $\mathbf{88.66}^\dagger\pm2.44$ \\
PNA+$\text{GSAT\xspace}^*$ & $89.04\pm4.92$ & $\mathbf{96.22}^\dagger\pm2.08$ & $\mathbf{88.54}^\dagger\pm0.72$ & $\mathbf{90.55}^\dagger\pm1.48$ & $\mathbf{89.79}^\dagger\pm1.91$ & $\mathbf{89.54}^\dagger\pm1.78$ \\
\bottomrule
\label{table:Interpretation}
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vspace{-9mm}
\end{table*}
\begin{table*}[t]
\caption{Prediction Performance (Acc.). The \textbf{bold} font highlights the inherently interpretable methods that significantly outperform the corresponding backbone model, GIN or PNA, when the mean-1*std of a method $>$ the mean of its corresponding backbone model.}
\vspace{-0.1cm}
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lcccccc}
\toprule
& \multirow{2}{*}{MolHiv (AUC)} & \multirow{2}{*}{Graph-SST2} & \multirow{2}{*}{MNIST-75sp} & \multicolumn{3}{c}{Spurious-motif} \\
& & & & $b=0.5$ & $b=0.7$ & $b=0.9$ \\
\midrule
GIN & $76.69\pm1.25$ & $82.73\pm0.77$ & $95.74\pm0.36$ & $39.87\pm1.30$ & $39.04\pm1.62$ & $38.57\pm2.31$ \\
IB-subgraph & ${76.43}\pm2.65$ & $82.99\pm0.67$ & $93.10\pm1.32$ & $\mathbf{54.36}\pm7.09$ & $\mathbf{48.51}\pm5.76$ & $\mathbf{46.19}\pm5.63$ \\
DIR & $76.34\pm1.01$ & $82.32\pm0.85$ & $88.51\pm2.57$ & $\mathbf{45.49}\pm3.81$ & $41.13\pm2.62$ & $37.61\pm2.02$ \\
GIN+GSAT\xspace & $76.47\pm1.53$ & $82.95\pm0.58$ & $\mathbf{96.24}\pm0.17$ & $\mathbf{52.74}\pm4.08$ & $\mathbf{49.12}\pm3.29$ & $\mathbf{44.22}\pm5.57 $\\
GIN+$\text{GSAT\xspace}^*$ & $76.16\pm1.39$ & $82.57\pm0.71$ & $\mathbf{96.21}\pm0.14$ & $\mathbf{46.62}\pm2.95$ & $41.26\pm3.01$ & $39.74\pm2.20 $\\
\midrule
PNA (no scalars) & $78.91\pm1.04$ & $79.87\pm1.02$ & $87.20\pm5.61$ & $68.15\pm2.39$ & $66.35\pm3.34$ & $61.40\pm3.56 $ \\
PNA+GSAT\xspace & $\mathbf{80.24}\pm0.73$ & $\mathbf{80.92}\pm0.66$ & $\mathbf{93.96}\pm0.92$ & $68.74\pm2.24$ & $64.38\pm3.20$ & $57.01\pm2.95 $\\
PNA+$\text{GSAT\xspace}^*$ & $\mathbf{80.67}\pm0.95$ & $\mathbf{82.81}\pm0.56$ & $\mathbf{92.38}\pm1.44$ & $\mathbf{69.72}\pm1.93$ & $\mathbf{67.31}\pm1.86$ & $61.49\pm3.46 $\\
\bottomrule
\label{table:Generalization}
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -6mm
\end{table*}
We evaluate our method for both interpretability and prediction performance. We will compare our method with both state-of-the-art (SOTA) post-hoc interpretation methods and inherently interpretable models.
We will also compare with several invariant learning methods to demonstrate the ability of GSAT\xspace to remove spurious correlations. We briefly introduce datasets, baselines and experiment settings here, and more details can be found in Appendix~\ref{appx:setting}.
\vspace{-2mm}
\subsection{Datasets}
\textbf{Mutag}~\cite{debnath1991structure} is a molecular property prediction dataset. Following~\cite{luo2020parameterized}, -NO$_2$ and -NH$_2$ in mutagen graphs are labeled as ground-truth explanations.
\textbf{BA-2Motifs}~\cite{luo2020parameterized} is a synthetic dataset with binary graph labels. House motifs and cycle motifs give class labels and thus are regarded as ground-truth explanations for the two classes respectively.
\textbf{Spurious-Motif}~\cite{anonymous2022discovering} is a synthetic dataset with three graph classes. Each class contains a particular motif that can be regarded as the ground-truth explanation. Some spurious correlation between the rest graph components (other than the motifs) and the labels also exists in the training data.
The degree of such correlation is controlled by $b$, and we include datasets with $b=0.5$, $0.7$ and $0.9$.
\textbf{MNIST-75sp} ~\cite{knyazev2019understanding} is an image classification dataset, where each image in MNIST is converted to a superpixel graph. Nodes with nonzero pixel values provide ground-truth explanations. Note that the subgraphs that provide explanations are of different sizes in this dataset.
\textbf{Graph-SST2} ~\cite{socher2013recursive, yuan2020explainability} is a sentiment analysis dataset, where each text sequence in SST2 is converted to a graph. Following the splits in~\cite{anonymous2022discovering}, this dataset contains degree shifts and no ground-truth explanation labels. So, we only evaluate prediction performance and provide interpretation visualizations.
\textbf{OGBG-Molhiv} ~\cite{wu2018moleculenet, hu2020ogb} is a molecular property prediction datasets. We also evaluate GSAT\xspace on molbace, molbbbp, molclintox, moltox21 and molsider datasets from OGBG. As there are no ground truth explanation labels for these datasets, we only evaluate the prediction performance of GSAT\xspace.
\subsection{Baselines and Setup}
\textbf{Interpretability Baselines.} We compare interpretability with post-hoc methods GNNExplainer ~\cite{ying2019gnnexplainer}, PGExplainer ~\cite{luo2020parameterized}, GraphMask ~\cite{schlichtkrull2021interpreting}, and inherently interpretable models DIR \cite{anonymous2022discovering} and IB-subgraph ~\cite{yu2020graph}.
\textbf{Prediction Baselines.} We compare prediction performance with the backbone models GIN ~\cite{xu2018powerful} and PNA ~\cite{corso2020principal}, and inherently interpretable models DIR \cite{anonymous2022discovering} and IB-subgraph ~\cite{yu2020graph}.
\textbf{Invariant Learning Baselines.} We compare the ability to remove spurious correlations with invariant learning methods IRM~\cite{arjovsky2019invariant}, V-REx~\cite{krueger21a} and DIR \cite{anonymous2022discovering}. Baseline results yielded by empirical risk minimization (ERM) are also included.
\textbf{Metrics.} For interpretation evaluation, we report explanation ROC AUC following ~\cite{ying2019gnnexplainer, luo2020parameterized}. For prediction performance, we report classification ROC AUC for all OGBG datasets and report accuracy for all other datasets. All the results are averaged over 10 times tests with different random seeds. For the post-hoc methods, we do not cherry pick a pre-trained model. Instead, in each test, we interpret a model pre-trained independently that achieves the best validation performance.
\textbf{Setup.} Since we focus on graph classification tasks, GIN ~\cite{xu2018powerful} is used as the backbone model for both baselines and GSAT\xspace. We also apply PNA~\cite{corso2020principal} to further test the wide applicability of GSAT\xspace, for which we adopt the no-scalars version since the scalars used in PNA are essentially a type of attention, which may conflict with our method. GIN+GSAT denotes using GIN as the base GNN encoder of GSAT, and PNA+GSAT means replacing the GNN encoder with PNA.
In addition, we apply GSAT\xspace to fine-tune and interpret pre-trained models as described in Sec.~\ref{sec:finetune}, which is highlighted as $\text{GSAT\xspace}^*$.
In all the experiments, we use $r=0.7$ in Eq.~\eqref{eq:reg} by default or otherwise specified. Our studies have shown that GSAT\xspace is generally robust when $r\in[0.5, 0.9]$ (see Fig.~\ref{fig:reg-compare} later).
\subsection{Result Comparison and Analysis} \label{sec:results}
\textbf{Interpretability Results.} As shown in Table \ref{table:Interpretation}, our methods significantly outperform the baselines by 9\%$\uparrow$ on average and up to 20\%$\uparrow$. If we just compare among inherently interpretable models, the boost is even more significant. Moreover, GSAT\xspace also provides much stabler interpretation than the baselines as for the much smaller variance. \text{GSAT\xspace}$^*$ via fine-tuning a pre-trained model can often further boost the interpretation performance. Also, when the more expressive model PNA is used as the backbone, we find the posthoc methods are likely to suffer from the overfitting issue as explained in Sec.~\ref{sec:post-hoc}. However, GSAT\xspace does not suffer from that and can yield even better interpretation results.
Over Ba-2Motifs and Mutag, GNNExplainer and PGExplainer work worse than what reported in \cite{luo2020parameterized} as we do not cherry pick the pre-trained model. However, GSAT\xspace still significantly outperforms their reported performance in the Appendix \ref{appx:sup_exp}.
We also provide visualizations of the subgraphs discovered by GSAT\xspace in Appendix \ref{appx:visz}.
\textbf{Prediction Results.} As explained in Sec.~\ref{sec:better-generalization}, being trained via the GIB principle, GSAT\xspace is more generalizable and thus may achieve even better prediction performance.
As shown in Table \ref{table:Generalization}, GIN+GSAT\xspace significantly outperforms the backbone GIN over the Spurious-Motif datasets, where spurious correlation exists in the training data. For other datasets, GIN+GSAT\xspace can achieve comparable results, which matches our claim that GSAT\xspace provides interpretation without hurting the prediction. IB-subgraph, trained via the GIB principle, also achieves good prediction performance though its interpretability is poor (Table~\ref{table:Interpretation}).
When PNA is used, GSAT\xspace improves it by about $1-5\%$ on the datasets in the first three columns. Notably, $\text{GSAT\xspace}^*$ achieves the SOTA performance on \emph{molhiv} among all models that do not incorporate expert knowledge according to the \href{https://ogb.stanford.edu/docs/leader_graphprop/#ogbg-molhiv}{leaderboard}. Unexpectedly, PNA achieves very good performance on Spurious-Motif and \text{GSAT\xspace}$^*$ just slightly improves it. Our results on the other $5$ molecular datasets from OGBG are showed in Table \ref{table:5mol}, where GSAT\xspace and $\text{GSAT\xspace}^*$ mostly outperform PNA.
\textbf{Invariant Learning Results.} We note that DIR achieves a bit lower prediction performance in Table \ref{table:Generalization} than what reported in \cite{anonymous2022discovering} even after we extensively tune its parameters, which is probably due to the different backbone models used. Hence, we also compare with DIR by using their backbone model. And we include several invariant learning baselines reported in DIR to further demonstrate the ability of GSAT\xspace to remove spurious correlations. Results are shown in Table \ref{table:dir-acc}. GSAT\xspace significantly outperforms all invariant learning methods on spurious correlation removal, even without utilizing causality analysis, which further validates our claims in Sec.~\ref{sec:better-generalization}.
A comparison of interpretability of these models is shown in Table \ref{table:dir-prec5} in the appendix.
\begin{table}[t]
\vspace{-2mm}
\caption{Generalization ROC AUC on other OGBG-Mol datasets. The \textbf{bold} font highlights when GSAT\xspace outperforms PNA.}
\vspace{-0.1cm}
\begin{center}
\resizebox{\columnwidth}{!}{%
\begin{sc}
\begin{tabular}{lccccc}
\toprule
& molbace & molbbbp & molclintox & moltox21 & molsider \\
\midrule
PNA & $73.52\pm3.02$ & $67.21\pm1.34$ & $86.72\pm2.33$ & $75.08\pm0.64$ & $56.51\pm1.90$ \\
GSAT\xspace & $\mathbf{77.41}\pm2.42$ & $\mathbf{69.17}\pm1.12$ & $\mathbf{87.80}\pm2.36$ & $74.96\pm0.66$ & $\mathbf{57.58}\pm1.23$ \\
$\text{GSAT\xspace}^*$ & $73.61\pm1.59$ & $66.30\pm0.79$ & $\mathbf{89.26}\pm1.66$ & $\mathbf{75.71}\pm0.48$ & $\mathbf{59.19}\pm1.03$ \\
\bottomrule
\label{table:5mol}
\end{tabular}
\end{sc}
}
\end{center}
\vskip -10mm
\end{table}
\begin{table}[t]
\caption{Direct comparison (Acc.) with invariant learning methods on the ability to remove spurious correlations, by applying the backbone model used in~\cite{anonymous2022discovering}.}
\vspace{-4mm}
\begin{center}
\resizebox{\columnwidth}{!}{%
\begin{sc}
\begin{tabular}{lccc}
\toprule
Spurious-motif & $b = 0.5$ & $b=0.7$ & $b=0.9$ \\
\midrule
ERM & $39.69\pm1.73 $ & $38.93\pm1.74 $ & $33.61\pm1.02$ \\
V-REx & $39.43\pm2.69 $ & $39.08\pm1.56 $ & $34.81\pm2.04$ \\
IRM & $41.30\pm1.28 $ & $40.16\pm1.74 $ & $35.12\pm2.71$ \\
DIR & $45.50\pm2.15 $ & $43.36\pm1.64 $ & $39.87\pm0.56$ \\
GSAT\xspace & $\mathbf{53.27}^\dagger\pm5.12 $ & $\mathbf{56.50}^\dagger\pm3.96 $ & $\mathbf{53.11}^\dagger\pm4.64$ \\
$\text{GSAT\xspace}^*$ & $43.27\pm4.58 $ & $42.51\pm5.32 $ & $\mathbf{45.76}^\dagger\pm5.32$ \\
\bottomrule
\label{table:dir-acc}
\end{tabular}
\end{sc}
}
\end{center}
\vskip -0.7cm
\end{table}
\textbf{Ablation Studies.} We conduct ablation studies from three aspects: First, the importance of stochasticity in GSAT\xspace, where we replace the Bernoulli sampling procedure with setting attention $\alpha_{uv}=p_{uv}$ without stochasticity; Second, the importance of the information regularization term (Eq.~\eqref{eq:reg}), where we set its coefficient $\beta=0$ in Eq.~\eqref{eq:proj}; Third, the superiority of the information regularization term over the sparsity-driven term $\ell_1$-norm.
As shown in Table \ref{table:ab-beta-noise-gin}, the performance drops significantly when there is either no stochasticity or $\beta = 0$. Specifically, GSAT\xspace-NoStoch means applying deterministic attention $\in[0,1]$, which causes the most performance drop. GSAT\xspace-NoStoch-$\beta=0$ corresponds to using deterministic attention without the regularization term in Eq.~\eqref{eq:reg}, which causes the second most performance drop. GSAT\xspace-$\beta=0$ denotes applying stochastic attention with no regularization, which performs better than baselines but worse than original GSAT\xspace and suffers from large variance.
Overall, no stochasticity yields the biggest drop, which well matches our theory.
This also implies that directly using the deterministic attention mechanisms such as GAT~\cite{velivckovic2018graph} or GGNN~\cite{li2015gated} may not yield good interpretability.
Fig.~\ref{fig:reg-compare} shows that our information regularization term can achieve consistently better performance than the sparsity-driven $\ell_1$-norm regularization even when the grid search is used to tune hyperparameters. We also observe that when $r$ is close to 0, the results often get decreased or have higher variance. The best performance is often achieved when $r\in [0.5, 0.9]$, which matches our theory. More results on other datasets can be found in Fig.~\ref{fig:sp0709_l1} in the appendix.
\begin{table}[t]
\tiny
\vspace{-2mm}
\caption{Ablation study on $\beta$ and stochasticity in GSAT\xspace (GIN as the backbone model) on Spurious-Motif. We report both interpretation ROC AUC (top) and prediction accuracy (bottom).}
\vspace{-0.2cm}
\begin{center}
\begin{sc}
\begin{tabular}{lccc}
\toprule
Spurious-motif & $b = 0.5$ & $b=0.7$ & $b=0.9$ \\
\midrule
GSAT\xspace & $79.81\pm3.98 $ & $74.07\pm5.28 $ & $71.97\pm4.41$ \\
GSAT\xspace-$\beta=0$ & $66.00\pm11.04$ & $65.92\pm3.28 $ & $66.31\pm6.82$ \\
GSAT\xspace-NoStoch & $59.64\pm5.33 $ & $55.78\pm2.84 $ & $55.27\pm7.49$ \\
GSAT\xspace-NoStoch-$\beta=0$ & $63.37\pm12.33$ & $60.61\pm10.08$ & $66.19\pm7.76$ \\
\midrule
GIN & $39.87\pm1.30$ & $39.04\pm1.62$ & $38.57\pm2.31$ \\
GSAT\xspace & $51.86\pm5.51$ & $49.12\pm3.29$ & $44.22\pm5.57$ \\
GSAT\xspace-$\beta=0$ & $45.97\pm8.37$ & $49.67\pm7.01$ & $49.84\pm5.45$ \\
GSAT\xspace-NoStoch & $40.34\pm2.77$ & $41.90\pm3.70$ & $37.98\pm2.64$ \\
GSAT\xspace-NoStoch-$\beta=0$ & $43.41\pm8.05$ & $45.88\pm9.54$ & $42.25\pm9.77$ \\
\bottomrule
\label{table:ab-beta-noise-gin}
\end{tabular}
\end{sc}
\end{center}
\vspace{-6mm}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[trim={0.6cm 0.0cm 0.6cm 0.1cm},clip,width=0.494\linewidth]{img/reg_ab_05_x.pdf}
\includegraphics[trim={0.6cm 0.0cm 0.6cm 0.1cm},clip,width=0.494\linewidth]{img/reg_ab_05_clf.pdf}
\vspace*{-9mm}
\caption{Comparison between (a) using the information constraint in Eq.~\eqref{eq:reg} and (b) replacing it with $\ell_1$-norm. Results are shown for Spurious-Motif $b=0.5$, where $r$ is tuned from $0.9$ to $0.1$ and the coefficient of the $\ell_1$-norm $\lambda_1$ is tuned from $1e\text{-}5$ to $1$.}
\vspace{-0.2cm}
\label{fig:reg-compare}
\end{figure}
\section{Conclusion}
\emph{Graph Stochastic Attention} (GSAT\xspace) is a novel attention mechanism to build interpretable graph learning models. GSAT\xspace injects stochasticity to block label-irrelevant information and leverages the reduction of stochasticity to select label-relevant subgraphs. Such rationale is grounded by the information bottleneck principle. GSAT\xspace has many transformative characteristics. For example, it removes the sparsity, continuity or other potentially biased assumptions in graph learning interpretation without performance decay. It can also remove spurious correlation to better the model generalization. As a by-product, we also reveal a potentially severe issue behind post-hoc interpretation methods from the optimization perspective of information bottleneck.
\section{Supplementary Notations for Information Theory and Graph Neural Networks}
\textbf{Entropy.} Given a discrete random variable $a$, its entropy is defined as $H(a) \triangleq - \sum_{a} \mathbb{P}(a)\log \mathbb{P}(a)$. If $a$ is a continuous random variable, its differential entropy is defined as $H(a) \triangleq - \int_{a} \mathbb{P}(a)\log \mathbb{P}(a)da$.
\textbf{KL-Divergence}. Given two distributions $\mathbb{P}(x)$ and $\mathbb{Q}(x)$, KL-Divergence is used to measure the difference between $\mathbb{P}$ and $\mathbb{Q}$, and it is defined as $\text{KL}(\mathbb{P}(x)||\mathbb{Q}(x)) \triangleq \sum_{x}\mathbb{P}(x)\log \frac{\mathbb{P}(x)}{\mathbb{Q}(x)}$.
\textbf{Mutual Information.} Given two random variables $a$ and $b$, the mutual information (MI) $I(a;b)$ is a measure of the mutual dependence between them. MI quantifies the amount of information regarding one random variable if another random variable is known. Formally, $I(a;b) \triangleq \sum_{a,b} \mathbb{P}(a,b) \log \frac{\mathbb{P}(a,b)}{\mathbb{P}(a)\mathbb{P}(b)}$, where $\mathbb{P}(a,b)$ is the joint distribution and $\mathbb{P}(a)$, $\mathbb{P}(b)$ are the marginal distributions. By definition, $I(a,b) = \text{KL}(\mathbb{P}(a,b)||\mathbb{P}(a)\mathbb{P}(b)) = \sum_{a,b}\mathbb{P}(a,b)\log\mathbb{P}(a|b) - \sum_{b}\mathbb{P}(b)\log\mathbb{P}(b)= - H(a|b) + H(b)$.
\textbf{Graph Neural Networks (GNNs).} Given an $L$-layer GNN, let $h_v^{(l)}$ denote the node representation for node $v$ in the $i^{th}$ layer and $\mathcal{N}(v)$ denote a set of nodes adjacent to node $v$. Let $h_v^{(0)}$ be the node feature $X_v$.
Most GNNs follow a message passing scheme, where there are two main steps in each layer: (1) neighbourhood aggregation, $m_{v}^{(l)}=\operatorname{AGG}(\{h_{u}^{(l-1)} | u \in \mathcal{N}(v)\})$; (2) node representation update, $h_{v}^{(l)}=\operatorname{UPDATE}(m_{v}^{(l)}, h_{v}^{(l-1)})$. For graph classification tasks, after obtaining $h_v^{(L)}$ for each node, the graph representation is given by $h_G=\operatorname{POOL}(\{h_v^{(L)}| v \in V\})$ and $h_G$ will be used to make predictions. The above $\operatorname{AGG}$, $\operatorname{UPDATE}$, $\operatorname{POOL}$ are three functions. $\operatorname{AGG}$ and $\operatorname{POOL}$ are typically implemented via $\operatorname{SUM}$, $\operatorname{MEAN}$ and $\operatorname{MAX}$ while $\operatorname{UPDATE}$ is a fully connected (typically shallow) neural network. In some cases, edge representations may be in need, and they are often given by $h_{u,v}^{(l)} = \operatorname{CONCAT}(h_u^{(l)}, h_v^{(l)})$.
\section{Variational Bounds for the GIB Objective --- Eq.~\eqref{eq:predictor} and Eq.~\eqref{eq:extractor}} \label{appx:deriving}
From Eq.~\eqref{eq:GIB2}, the IB objective is:
\begin{align}
\min_{\phi} -I(G_S ; Y) + \beta I(G_S;G),\, \text{s.t.}\,\, G_S \sim g_{\phi}(G).
\end{align}
To optimize it, we introduce two variational bounds on the two terms, respectively.
For the first term $I\left(G_S ; Y\right)$, by definition:
\begin{equation}
I\left(G_S ; Y\right) = \mathbb{E}_{G_S, Y} \left[ \log \frac{\mathbb{P}(Y|G_S)}{\mathbb{P}(Y)} \right].
\end{equation}
Since $\mathbb{P}(Y|G_S)$ is intractable, we introduce a variational approximation $\mathbb{P}_{\theta}(Y|G_S)$ for it. Then, we obtain a lower bound for Eq.~\eqref{eq:predictor}:
\begin{align}
I\left(G_S ; Y\right) &= \mathbb{E}_{G_S, Y} \left[ \log \frac{\mathbb{P}_{\theta}(Y|G_S)}{ \mathbb{P}(Y)} \right] + \mathbb{E}_{G_S} \left[ \text{KL}( \mathbb{P}(Y|G_S) || \mathbb{P}_{\theta}(Y|G_S)) \right] \nonumber \\
& \geq \mathbb{E}_{G_S, Y} \left[ \log \frac{\mathbb{P}_{\theta}(Y|G_S)}{ \mathbb{P}(Y)} \right] \nonumber \\
& = \mathbb{E}_{G_S, Y} \left[ \log {\mathbb{P}_{\theta}(Y|G_S)} \right] + H(Y).
\end{align}
For the second term $I \left(G ; G_S\right)$, by definition:
\begin{equation}
I \left(G ; G_S\right) = \mathbb{E}_{G_S, G} \left[ \log \frac{\mathbb{P}(G_S|G)}{\mathbb{P}(G_S)} \right].
\end{equation}
Since $\mathbb{P}(G_S)$ is intractable, we introduce a variational approximation $\mathbb{Q}(G_S)$ for the marginal distribution $\mathbb{P}(G_S) = \sum_{G}\mathbb{P}_{\phi}(G_S|G)\mathbb{P}_{\mathcal{G}}(G)$. Then, we obtain an upper bound for Eq.~\eqref{eq:extractor}:
\begin{align}
I \left(G ; G_S\right) &= \mathbb{E}_{G_S, G} \left[ \log \frac{\mathbb{P}_{\phi}(G_S|G)}{\mathbb{Q}(G_S)} \right] - \text{KL} \left( \mathbb{P}(G_S) || \mathbb{Q}(G_S) \right) \nonumber \\
&\leq \mathbb{E}_{G} \left[ \text{KL} \left( \mathbb{P}_{\phi}(G_S|G) || \mathbb{Q}(G_S) \right) \right].
\end{align}
\section{Supplementary Experiments} \label{appx:setting}
\subsection{Details of the Datasets} \label{appx:datasets}
\textbf{Mutag}~\cite{debnath1991structure} is a molecular property prediction dataset, where nodes are atoms and edges are chemical bonds. Each graph is associated with a binary label based on its mutagenic effect. Following~\cite{luo2020parameterized}, -NO$_2$ and -NH$_2$ in mutagen graphs are labeled as ground-truth explanations.
\textbf{BA-2Motifs}~\cite{luo2020parameterized} is a synthetic dataset, where the base graph is generated by Barabási-Albert (BA) model. Each base graph is attached with a house-like motif or a five-node cycle motif. House motifs and cycle motifs give class labels and thus are regarded as ground-truth explanations for the two classes respectively.
\textbf{Spurious-Motif}~\cite{anonymous2022discovering} is a synthetic dataset with three graph classes. Following the notations in~\cite{anonymous2022discovering}, each graph consists of a base graph (tree/ladder/wheel denoted by $\bar{G}_S = 0, 1, 2$ respectively, with some abuse of notations) and a motif (cycle/house/crane denoted by $G_S = 0, 1, 2$, respectively, with some abuse of notations). The label is determined only by $G_S$, while there also exists spurious correlation between the label and $\bar{G}_S$. Specifically, to construct a graph in the training set, $G_S$ will be sampled uniformly, while $\bar{G}_S$ will be sampled with probability $\mathbb{P}(\bar{G}_S)$, where $\mathbb{P}(\bar{G}_S) = b$ if $\bar{G}_S=G_S$; otherwise $\mathbb{P}(\bar{G}_S) = (1-b) / 2$. So, $b$ is a parameter used to control the degree of such spurious correlation. When $b=1/3$, there is no spurious correlation. We include datasets with $b=0.5$, $b=0.7$ and $b=0.9$. Note that for testing data, the motifs and bases are randomly attached to each other, which can test if the model overfits the spurious correlation.
\textbf{MNIST-75sp} ~\cite{knyazev2019understanding} is a image classification dataset, where each image in MNIST is converted to a superpixel graph. Each node in the graph represents a superpixel and edges are formed based on spatial distance between superpixel centers. Node features are the coordinates of their centers of masses. Nodes with nonzero pixel values provide ground-truth explanations. Note that the subgraphs that provide explanations are of different sizes in this dataset.
\textbf{Graph-SST2} ~\cite{socher2013recursive, yuan2020explainability} is a sentiment analysis dataset, where each text sequence in SST2 is converted to a graph. Each node in the graph represents a word and edges are formed based on relationships between different words. We follow the dataset splits in~\cite{anonymous2022discovering} to create degree shifts in the training set, which can better test generalizability of models. Specifically, graphs with higher average node degree will be used to train and validate models, while graphs with fewer nodes will be used to test models. And this dataset contains no ground-truth explanation labels, so we only evaluate prediction performance here and provide interpretation visualizations in Appendix~\ref{appx:visz}.
\textbf{OGBG-Molhiv} ~\cite{wu2018moleculenet, hu2020ogb} is a molecular property prediction datasets, where nodes are atoms and edges are chemical bonds. A binary label is assigned to each graph according to whether a molecule inhibits HIV virus replication or not. We also evaluate GSAT\xspace on molbace, molbbbp, molclintox, moltox21 and molsider datasets from OGBG. As there are no ground truth explanation labels for these datasets, we only evaluate the prediction performance of GSAT\xspace.
\begin{table}[t]
\caption{Direct comparison with the interpretation ROC AUC of GNNExplainer and PGExplainer reported in~\cite{luo2020parameterized}, which are given a selected pre-trained model.}
\begin{center}
\begin{sc}
\begin{tabular}{lcc}
\toprule
& Ba-2motifs & Mutag \\
\midrule
GNNExplainer & $74.2$ & $72.7$ \\
PGExplainer & $92.6$ & $87.3$ \\
GSAT\xspace & $\mathbf{98.74}^\dagger\pm0.55 $ & $\mathbf{99.60}^\dagger\pm0.51$ \\
$\text{GSAT\xspace}^*$ & $\mathbf{97.43}^\dagger\pm0.02$ & $\mathbf{97.75}^\dagger\pm0.92$ \\
\bottomrule
\label{table:pge-auc}
\end{tabular}
\end{sc}
\end{center}
\vspace{-9mm}
\end{table}
\begin{table}[t]
\caption{Direct comparison with the interpretation precision@5 of DIR reported in \cite{anonymous2022discovering} based on the backbone model in \cite{anonymous2022discovering}.}
\begin{center}
\begin{sc}
\begin{tabular}{lccc}
\toprule
& \multicolumn{3}{c}{Spurious-motif} \\
& $b = 0.5$ & $b=0.7$ & $b=0.9$ \\
\midrule
GNNExplainer & $0.203\pm0.019$ & $0.167\pm0.039$ & $0.066\pm0.007$ \\
DIR & $0.255\pm0.016$ & $0.247\pm0.012$ & $0.192\pm0.044$ \\
GSAT\xspace & $\mathbf{0.519}^\dagger\pm0.022$ & $\mathbf{0.503}^\dagger\pm0.034$ & $\mathbf{0.416}^\dagger\pm0.081$ \\
$\text{GSAT\xspace}^*$ & $\mathbf{0.532}^\dagger\pm0.019$ & $\mathbf{0.512}^\dagger\pm0.011$ & $\mathbf{0.520}^\dagger\pm0.022$ \\
\bottomrule
\label{table:dir-prec5}
\end{tabular}
\end{sc}
\end{center}
\vspace{-9mm}
\end{table}
\begin{table}[t]
\caption{Ablation study on $\beta$ and stochasticity in GSAT (PNA as the
backbone model) on Spurious-Motif. We report both interpretation
ROC AUC (top) and prediction accuracy (bottom).}
\begin{center}
\begin{sc}
\begin{tabular}{lccc}
\toprule
Spurious-motif & $b = 0.5$ & $b=0.7$ & $b=0.9$ \\
\midrule
PNA+GSAT\xspace & $83.34\pm2.17 $ & $86.94\pm4.05 $ & $88.66\pm2.44$ \\
PNA+GSAT\xspace-$\beta=0$ & $82.01\pm6.43 $ & $78.88\pm6.74 $ & $80.53\pm5.03$ \\
PNA+GSAT\xspace-NoStoch & $79.72\pm3.86 $ & $76.36\pm2.57 $ & $80.21\pm3.76$ \\
PNA+GSAT\xspace-NoStoch-$\beta=0$ & $78.69\pm10.77$ & $78.97\pm13.95$ & $79.91\pm13.11$ \\
\midrule
PNA & $68.15\pm2.39 $ & $66.35\pm3.34 $ & $61.40\pm3.56$ \\
PNA+GSAT\xspace & $68.74\pm2.24 $ & $64.38\pm3.20 $ & $57.01\pm2.95$ \\
PNA+GSAT\xspace-$\beta=0$ & $59.68\pm7.28 $ & $58.03\pm11.84$ & $53.94\pm8.11$ \\
PNA+GSAT\xspace-NoStoch. & $51.92\pm11.17$ & $41.22\pm7.72 $ & $39.56\pm2.74$ \\
PNA+GSAT\xspace-NoStoch.-$\beta=0$ & $56.54\pm6.88 $ & $48.93\pm10.33$ & $45.82\pm9.60$ \\
\bottomrule
\label{table:ab-beta-noise-pna}
\end{tabular}
\end{sc}
\end{center}
\vspace{-9mm}
\end{table}
\begin{figure}[t]
\vskip 0.1in
\centering
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.49\linewidth]{img/reg_ab_07_x.pdf}
\includegraphics[width=0.49\linewidth]{img/reg_ab_07_clf.pdf}
\vspace*{-4mm}
\caption{Spurious-Motif, $b=0.7$}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\linewidth}
\centering
\includegraphics[width=0.49\linewidth]{img/reg_ab_09_x.pdf}
\vspace*{-4mm}
\includegraphics[width=0.49\linewidth]{img/reg_ab_09_clf.pdf}
\caption{Spurious-Motif, $b=0.9$}
\end{subfigure}
\vspace*{-2mm}
\caption{Ablation study on (a) using the info. constraint in Eq.~\eqref{eq:reg} and (b) replacing it with $\ell_1$-norm, where $r$ is tuned from $0.9$ to $0.1$ and the coefficient of the $\ell_1$-norm $\lambda_1$ is tuned from $1e\text{-}5$ to $1$.}
\label{fig:sp0709_l1}
\vspace*{-2mm}
\end{figure}
\subsection{Details on Hyperparameter Tuning}
\subsubsection{Backbone Models}
\textbf{Backbone Architecture.} We use a two-layer GIN ~\cite{xu2018powerful} with $64$ hidden dimensions and $0.3$ dropout ratio. We use the setting from ~\cite{corso2020principal} for PNA, which has 4 layers with $80$ hidden dimensions, $0.3$ dropout ratio, and no scalars are used. For OGBG-Mol datasets, we directly follow ~\cite{corso2020principal} using (mean, min, max, std) aggregators for PNA; yet we find PNA has convergence issues on other datasets when sum aggregator is not used. Hence, PNA uses (mean, min, max, std, sum) aggregators for all other datasets.
\textbf{Dataset Splits.} For Ba-2Motifs, we split it randomly into three sets ($80\%/10\%/10\%$). For Mutag, we split it randomly into $80\%/20\%$ to train and validate models, and following~\cite{luo2020parameterized} we use mutagen molecules with -NO$_2$ or -NH$_2$ as test data (because only these samples have explanation labels). For MNIST-75sp, we use the default splits given by \cite{knyazev2019understanding}; due to its large size in the graph setting, we also reduce the number of training samples following~\cite{anonymous2022discovering} to speed up training. For Graph-SST2, Spurious-Motifs and OGBG-Mol, we use the default splits given by \cite{yuan2020explainability} and \cite{anonymous2022discovering}. Following~\cite{corso2020principal}, edge features are not used for all OGBG-Mol datasets.
\textbf{Epoch.} We tune the number of epochs to make sure the convergence of all models. When GIN is used as the backbone model, MNIST-75sp and OGBG-Molhiv are trained for $200$ epochs, and all other datasets are trained for $100$ epochs. When PNA is used, Mutag and Ba-2Motifs are trained for $50$ epochs and all other datasets are trained for $200$ epochs. We report the performance of the epoch that achieves the best validation prediction performance and use the models that achieve such best validation performance as the pre-trained models. When multiple epochs achieve the same best performance, we report the one with the lowest validation prediction loss.
\textbf{Batch Size.} All datasets use a batch size of $128$; except for MNIST-75sp we use a batch size of $256$ to speed up training due to its large size in the graph setting.
\textbf{Learning Rate.} GIN uses $0.003$ learning rate for Spurious-Motifs and $0.001$ for all other datasets. PNA uses $0.01$ learning rate with scheduler following~\cite{corso2020principal}, $0.003$ learning rate for Graph-SST2 and Spurious-Motifs, and $0.001$ learning rate for all other datasets.
\subsubsection{GSAT\xspace}
\textbf{Basic Setting.} If not specified, GSAT\xspace uses the same settings mentioned for the backbone models. All Spurious-Motif datasets share the same hyperparameters, which are tuned based on $b=0.5$.
\textbf{Learning Rate.} When PNA is used, GSAT\xspace uses $0.001$ learning rate for all OGBG-Mol datasets; otherwise it uses the same learning rate as mentioned above.
\textbf{$r$ in Equation (\ref{eq:reg}).} Ba-2Motif and Mutag use $r=0.5$, and all other datasets use $r=0.7$. We find $r=0.7$ can generally provide great performance for all datasets. Inspired by curriculum learning~\cite{bengio2009curriculum}, $r$ will initially set to $0.9$ and gradually decay to the tuned value. We adopt a step decay, where $r$ will decay $0.1$ for every $10$ epochs.
\textbf{$\beta$ in Equation (\ref{eq:proj}).} $\beta$ is not tuned and is set to $\frac{1}{|E|}$ for all datasets.
\textbf{Temperature.} Temperature used in the Gumbel-softmax trick~\cite{jang2016categorical} is not tuned, and we use $1$ for all datasets.
\subsubsection{Baseline Interpretable Methods/Models}
\textbf{Basic Setting.} If not specified, baselines use the same settings mentioned for the backbone models. All Spurious-Motif datasets share the same hyperparameters, which are tuned based on $b=0.5$.
\textbf{GNNExplainer.} We tune the learning rate from $(1, 0.1, 0.01, 0.001)$ and the coefficient of the $\ell_1$-norm from $(0.1, 0.01, 0.001)$, based on validation interpretation ROC AUC. The coefficient of the entropy regularization term is set to the recommended value $1$. Again, in a real-world setting, post-hoc methods have no clear metric to tune hyper-parameters.
\textbf{PGExplainer.} We use the tuned recommended settings from~\cite{luo2020parameterized}, including the temperature, the coefficient of $\ell_1$-norm regularization and the coefficient of entropy regularization.
\textbf{GraphMask.} We use the recommended settings from~\cite{schlichtkrull2021interpreting}, including the temperature, gamma, zeta and the coefficient of $\ell_0$-norm regularization.
\textbf{DIR.} Causal ratio is tuned for Ba-2Motif and Mutag. Since the other datasets we use are the same, we use the recommended settings from~\cite{anonymous2022discovering}. However, even though datasets are the same, we find the same $\alpha$ specified in their source code do not work well in our setting. Hence, we tune $\alpha$ from $(10, 1, 0.1, 0.01, 0.001, 0.0001, 0.00001, 0.000001)$.
\textbf{IB-subgraph.} Due to the extreme inefficiency of IB-subgraph, we are only able to tune its mi-weight around the recommended value from $(2, 0.2, 0.02)$. And we use the default inner loop iterations and con-weight as specified in their source code. IB-subgraph needs $\sim$40 hours to train $100$ epochs for $1$ seed on Spurious-Motif and $\sim$150 hours for OGBG-Molhiv on a Quadro RTX 6000. By contrast, GSAT\xspace only needs $\sim$15 minutes to train $100$ epochs on OGBG-Molhiv.
\textbf{Random Seed.} All methods are trained with $10$ different random seeds; except for IB-subgraph we train it for $5$ different random seeds due to its inefficiency. For post-hoc methods, the pre-trained models are also trained with $10$ different random seeds instead of a fixed pre-trained model in~\cite{luo2020parameterized}. For inherently interpretable models, GSAT\xspace, IB-subgraph and DIR, we average the best epoch's performance according to their validation prediction performance. For post-hoc baselines, we average their last epoch's performance. For IB-subgraph, we stop training when there is no improvement over $20$ epochs to make the training possible on large datasets.
\subsection{Node/Edge Attention}
We also explore node-level attention, and we find it is especially useful for molecular datasets and datasets with large graph sizes. Hence, we use node-level attention for on Mutag, MNIST-75sp and OGBG-Mol datasets, and for all other datasets we use edge attention. Specifically, when node attention is used, the MLP layers in $\mathbb{P}_{\phi}$ will take as input the node embeddings and output $p_v$ for each $v \in V$. Then, the stochastic node attention is sampled for each node $\alpha_v \sim \text{Bern}(p_v)$. After that, $\alpha_{uv}$ is obtained by $\alpha_{uv} = \alpha_u \alpha_v$.
\subsection{Further Supplementary Experiments} \label{appx:sup_exp}
Fig.~\ref{fig:connect-visual} shows an experiment with disconnected critical subgraphs, where the dataset is generated in a similar way used to generate Ba-2Motifs. Specifically, each base graph is generated using the BA model and will be attached with two house motifs or three house motifs randomly. The number of house motifs represents the graph class. Both GSAT\xspace and GraphMask are trained with the same settings used on Ba-2Motifs.
Table~\ref{table:pge-auc} shows a direct comparison with PGExplainer and GNNExplainer between the interpretation ROC AUC reported in~\cite{luo2020parameterized} and the performance of GSAT\xspace. And GSAT\xspace still outperforms their methods significantly.
Table~\ref{table:dir-acc} and Table~\ref{table:dir-prec5} show direct comparisons with DIR, where we apply GSAT\xspace with the backbone model used in DIR. And GSAT\xspace still greatly outperforms their method.
Table~\ref{table:ab-beta-noise-pna} shows the ablation study on $\beta$ and stochasticity in GSAT\xspace, where PNA is the backbone model. Figure~\ref{fig:sp0709_l1} shows the ablation study of the information constraint introduced in Eq.~\eqref{eq:reg} on Spurious-Motif $b=0.7$ and $b=0.9$. We observe the same trends from these ablation studies as discussed in Sec.~\ref{sec:results}.
\section{Interpretation Visualization} \label{appx:visz}
We provide visualizations of the label-relevant subgraphs discovered by GSAT\xspace on eight datasets, as shown from Fig.~\ref{viz:mutag} to Fig.~\ref{viz:mnist}. The transparency of the edges shown in the figures represents the normalized attention weights learned by GSAT\xspace. The normalized attention weights are to rescale the learnt weights $\{p_{uv}|(u,v)\in E\}$ to $[0,1]$: For each graph, denote $p_{\min} = \min \{p_{uv}|(u,v)\in E\}$ and $p_{\max} = \max \{p_{uv}|(u,v)\in E\}$. We rescale the weights according to
\begin{align}\label{eq:normalization}
\hat{p}_{uv} = \frac{p_{uv}-p_{\min}}{p_{\max}-p_{\min}}
\end{align}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{img/viz_ba_2motifs.pdf}}
\vspace{-3mm}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for Ba-2Motifs. Nodes colored pink are ground-truth explanations, and each row represents a graph class.}
\label{viz:mutag}
\end{center}
\vspace{-5mm}
\end{figure}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{img/viz_mutag.pdf}}
\vspace{-3mm}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for Mutag. -NO$_2$ and -NH$_2$ are ground-truth explanations. We only present mutagen graphs as only these graphs are with ground-truth explanation labels.}
\end{center}
\vspace{-7mm}
\end{figure}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{img/viz_spmotif_0.5.pdf}}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for Spurious-Motif $b=0.5$. Nodes colored pink are ground-truth explanations, and each row represents a graph class.}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{img/viz_spmotif_0.7.pdf}}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for Spurious-Motif $b=0.7$. Nodes colored pink are ground-truth explanations, and each row represents a graph class.}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{img/viz_spmotif_0.9.pdf}}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for Spurious-Motif $b=0.9$. Nodes colored pink are ground-truth explanations, and each row represents a graph class.}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{img/viz_ogbg_molhiv.pdf}}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for OGBG-Molhiv. Each row represents a graph class.}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{img/viz_Graph-SST2.pdf}}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for Graph-SST2. The top two rows show sentences with negative sentiment, and the bottom two rows show sentences with positive sentiment.}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1.0\columnwidth]{img/viz_mnist.pdf}}
\caption{Visualizing label-relevant subgraphs discovered by GSAT\xspace for MNIST-75sp. The first row shows the raw images and the second row shows the normalized attention weights learned by GSAT\xspace.}
\label{viz:mnist}
\end{center}
\end{figure}
\section{Electronic Submission}
\label{submission}
Submission to ICML 2022 will be entirely electronic, via a web site
(not email). Information about the submission process and \LaTeX\ templates
are available on the conference web site at:
\begin{center}
\textbf{\texttt{http://icml.cc/}}
\end{center}
The guidelines below will be enforced for initial submissions and
camera-ready copies. Here is a brief summary:
\begin{itemize}
\item Submissions must be in PDF\@.
\item \textbf{New to this year}: If your paper has appendices, submit the appendix together with the main body and the references \textbf{as a single file}. Reviewers will not look for appendices as a separate PDF file. So if you submit such an extra file, reviewers will very likely miss it.
\item Page limit: The main body of the paper has to be fitted to 8 pages, excluding references and appendices; the space for the latter two is not limited. For the final version of the paper, authors can add one extra page to the main body.
\item \textbf{Do not include author information or acknowledgements} in your
initial submission.
\item Your paper should be in \textbf{10 point Times font}.
\item Make sure your PDF file only uses Type-1 fonts.
\item Place figure captions \emph{under} the figure (and omit titles from inside
the graphic file itself). Place table captions \emph{over} the table.
\item References must include page numbers whenever possible and be as complete
as possible. Place multiple citations in chronological order.
\item Do not alter the style template; in particular, do not compress the paper
format by reducing the vertical spaces.
\item Keep your abstract brief and self-contained, one paragraph and roughly
4--6 sentences. Gross violations will require correction at the
camera-ready phase. The title should have content words capitalized.
\end{itemize}
\subsection{Submitting Papers}
\textbf{Paper Deadline:} The deadline for paper submission that is
advertised on the conference website is strict. If your full,
anonymized, submission does not reach us on time, it will not be
considered for publication.
\textbf{Anonymous Submission:} ICML uses double-blind review: no identifying
author information may appear on the title page or in the paper
itself. \cref{author info} gives further details.
\textbf{Simultaneous Submission:} ICML will not accept any paper which,
at the time of submission, is under review for another conference or
has already been published. This policy also applies to papers that
overlap substantially in technical content with conference papers
under review or previously published. ICML submissions must not be
submitted to other conferences and journals during ICML's review
period.
Informal publications, such as technical
reports or papers in workshop proceedings which do not appear in
print, do not fall under these restrictions.
\medskip
Authors must provide their manuscripts in \textbf{PDF} format.
Furthermore, please make sure that files contain only embedded Type-1 fonts
(e.g.,~using the program \texttt{pdffonts} in linux or using
File/DocumentProperties/Fonts in Acrobat). Other fonts (like Type-3)
might come from graphics files imported into the document.
Authors using \textbf{Word} must convert their document to PDF\@. Most
of the latest versions of Word have the facility to do this
automatically. Submissions will not be accepted in Word format or any
format other than PDF\@. Really. We're not joking. Don't send Word.
Those who use \textbf{\LaTeX} should avoid including Type-3 fonts.
Those using \texttt{latex} and \texttt{dvips} may need the following
two commands:
{\footnotesize
\begin{verbatim}
dvips -Ppdf -tletter -G0 -o paper.ps paper.dvi
ps2pdf paper.ps
\end{verbatim}}
It is a zero following the ``-G'', which tells dvips to use
the config.pdf file. Newer \TeX\ distributions don't always need this
option.
Using \texttt{pdflatex} rather than \texttt{latex}, often gives better
results. This program avoids the Type-3 font problem, and supports more
advanced features in the \texttt{microtype} package.
\textbf{Graphics files} should be a reasonable size, and included from
an appropriate format. Use vector formats (.eps/.pdf) for plots,
lossless bitmap formats (.png) for raster graphics with sharp lines, and
jpeg for photo-like images.
The style file uses the \texttt{hyperref} package to make clickable
links in documents. If this causes problems for you, add
\texttt{nohyperref} as one of the options to the \texttt{icml2022}
usepackage statement.
\subsection{Submitting Final Camera-Ready Copy}
The final versions of papers accepted for publication should follow the
same format and naming convention as initial submissions, except that
author information (names and affiliations) should be given. See
\cref{final author} for formatting instructions.
The footnote, ``Preliminary work. Under review by the International
Conference on Machine Learning (ICML). Do not distribute.'' must be
modified to ``\textit{Proceedings of the
$\mathit{39}^{th}$ International Conference on Machine Learning},
Baltimore, Maryland, USA, PMLR 162, 2022.
Copyright 2022 by the author(s).''
For those using the \textbf{\LaTeX} style file, this change (and others) is
handled automatically by simply changing
$\mathtt{\backslash usepackage\{icml2022\}}$ to
$$\mathtt{\backslash usepackage[accepted]\{icml2022\}}$$
Authors using \textbf{Word} must edit the
footnote on the first page of the document themselves.
Camera-ready copies should have the title of the paper as running head
on each page except the first one. The running title consists of a
single line centered above a horizontal rule which is $1$~point thick.
The running head should be centered, bold and in $9$~point type. The
rule should be $10$~points above the main text. For those using the
\textbf{\LaTeX} style file, the original title is automatically set as running
head using the \texttt{fancyhdr} package which is included in the ICML
2022 style file package. In case that the original title exceeds the
size restrictions, a shorter form can be supplied by using
\verb|\icmltitlerunning{...}|
just before $\mathtt{\backslash begin\{document\}}$.
Authors using \textbf{Word} must edit the header of the document themselves.
\section{Format of the Paper}
All submissions must follow the specified format.
\subsection{Dimensions}
The text of the paper should be formatted in two columns, with an
overall width of 6.75~inches, height of 9.0~inches, and 0.25~inches
between the columns. The left margin should be 0.75~inches and the top
margin 1.0~inch (2.54~cm). The right and bottom margins will depend on
whether you print on US letter or A4 paper, but all final versions
must be produced for US letter size.
Do not write anything on the margins.
The paper body should be set in 10~point type with a vertical spacing
of 11~points. Please use Times typeface throughout the text.
\subsection{Title}
The paper title should be set in 14~point bold type and centered
between two horizontal rules that are 1~point thick, with 1.0~inch
between the top rule and the top edge of the page. Capitalize the
first letter of content words and put the rest of the title in lower
case.
\subsection{Author Information for Submission}
\label{author info}
ICML uses double-blind review, so author information must not appear. If
you are using \LaTeX\/ and the \texttt{icml2022.sty} file, use
\verb+\icmlauthor{...}+ to specify authors and \verb+\icmlaffiliation{...}+ to specify affiliations. (Read the TeX code used to produce this document for an example usage.) The author information
will not be printed unless \texttt{accepted} is passed as an argument to the
style file.
Submissions that include the author information will not
be reviewed.
\subsubsection{Self-Citations}
If you are citing published papers for which you are an author, refer
to yourself in the third person. In particular, do not use phrases
that reveal your identity (e.g., ``in previous work \cite{langley00}, we
have shown \ldots'').
Do not anonymize citations in the reference section. The only exception are manuscripts that are
not yet published (e.g., under submission). If you choose to refer to
such unpublished manuscripts \cite{anonymous}, anonymized copies have
to be submitted
as Supplementary Material via CMT\@. However, keep in mind that an ICML
paper should be self contained and should contain sufficient detail
for the reviewers to evaluate the work. In particular, reviewers are
not required to look at the Supplementary Material when writing their
review (they are not required to look at more than the first $8$ pages of the submitted document).
\subsubsection{Camera-Ready Author Information}
\label{final author}
If a paper is accepted, a final camera-ready copy must be prepared.
For camera-ready papers, author information should start 0.3~inches below the
bottom rule surrounding the title. The authors' names should appear in 10~point
bold type, in a row, separated by white space, and centered. Author names should
not be broken across lines. Unbolded superscripted numbers, starting 1, should
be used to refer to affiliations.
Affiliations should be numbered in the order of appearance. A single footnote
block of text should be used to list all the affiliations. (Academic
affiliations should list Department, University, City, State/Region, Country.
Similarly for industrial affiliations.)
Each distinct affiliations should be listed once. If an author has multiple
affiliations, multiple superscripts should be placed after the name, separated
by thin spaces. If the authors would like to highlight equal contribution by
multiple first authors, those authors should have an asterisk placed after their
name in superscript, and the term ``\textsuperscript{*}Equal contribution"
should be placed in the footnote block ahead of the list of affiliations. A
list of corresponding authors and their emails (in the format Full Name
\textless{}[email protected]\textgreater{}) can follow the list of affiliations.
Ideally only one or two names should be listed.
A sample file with author names is included in the ICML2022 style file
package. Turn on the \texttt{[accepted]} option to the stylefile to
see the names rendered. All of the guidelines above are implemented
by the \LaTeX\ style file.
\subsection{Abstract}
The paper abstract should begin in the left column, 0.4~inches below the final
address. The heading `Abstract' should be centered, bold, and in 11~point type.
The abstract body should use 10~point type, with a vertical spacing of
11~points, and should be indented 0.25~inches more than normal on left-hand and
right-hand margins. Insert 0.4~inches of blank space after the body. Keep your
abstract brief and self-contained, limiting it to one paragraph and roughly 4--6
sentences. Gross violations will require correction at the camera-ready phase.
\subsection{Partitioning the Text}
You should organize your paper into sections and paragraphs to help
readers place a structure on the material and understand its
contributions.
\subsubsection{Sections and Subsections}
Section headings should be numbered, flush left, and set in 11~pt bold
type with the content words capitalized. Leave 0.25~inches of space
before the heading and 0.15~inches after the heading.
Similarly, subsection headings should be numbered, flush left, and set
in 10~pt bold type with the content words capitalized. Leave
0.2~inches of space before the heading and 0.13~inches afterward.
Finally, subsubsection headings should be numbered, flush left, and
set in 10~pt small caps with the content words capitalized. Leave
0.18~inches of space before the heading and 0.1~inches after the
heading.
Please use no more than three levels of headings.
\subsubsection{Paragraphs and Footnotes}
Within each section or subsection, you should further partition the
paper into paragraphs. Do not indent the first line of a given
paragraph, but insert a blank line between succeeding ones.
You can use footnotes\footnote{Footnotes
should be complete sentences.} to provide readers with additional
information about a topic without interrupting the flow of the paper.
Indicate footnotes with a number in the text where the point is most
relevant. Place the footnote in 9~point type at the bottom of the
column in which it appears. Precede the first footnote in a column
with a horizontal rule of 0.8~inches.\footnote{Multiple footnotes can
appear in each column, in the same order as they appear in the text,
but spread them across columns and pages if possible.}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{icml_numpapers}}
\caption{Historical locations and number of accepted papers for International
Machine Learning Conferences (ICML 1993 -- ICML 2008) and International
Workshops on Machine Learning (ML 1988 -- ML 1992). At the time this figure was
produced, the number of accepted papers for ICML 2008 was unknown and instead
estimated.}
\label{icml-historical}
\end{center}
\vskip -0.2in
\end{figure}
\subsection{Figures}
You may want to include figures in the paper to illustrate
your approach and results. Such artwork should be centered,
legible, and separated from the text. Lines should be dark and at
least 0.5~points thick for purposes of reproduction, and text should
not appear on a gray background.
Label all distinct components of each figure. If the figure takes the
form of a graph, then give a name for each axis and include a legend
that briefly describes each curve. Do not include a title inside the
figure; instead, the caption should serve this function.
Number figures sequentially, placing the figure number and caption
\emph{after} the graphics, with at least 0.1~inches of space before
the caption and 0.1~inches after it, as in
\cref{icml-historical}. The figure caption should be set in
9~point type and centered unless it runs two or more lines, in which
case it should be flush left. You may float figures to the top or
bottom of a column, and you may set wide figures across both columns
(use the environment \texttt{figure*} in \LaTeX). Always place
two-column figures at the top or bottom of the page.
\subsection{Algorithms}
If you are using \LaTeX, please use the ``algorithm'' and ``algorithmic''
environments to format pseudocode. These require
the corresponding stylefiles, algorithm.sty and
algorithmic.sty, which are supplied with this package.
\cref{alg:example} shows an example.
\begin{algorithm}[tb]
\caption{Bubble Sort}
\label{alg:example}
\begin{algorithmic}
\STATE {\bfseries Input:} data $x_i$, size $m$
\REPEAT
\STATE Initialize $noChange = true$.
\FOR{$i=1$ {\bfseries to} $m-1$}
\IF{$x_i > x_{i+1}$}
\STATE Swap $x_i$ and $x_{i+1}$
\STATE $noChange = false$
\ENDIF
\ENDFOR
\UNTIL{$noChange$ is $true$}
\end{algorithmic}
\end{algorithm}
\subsection{Tables}
You may also want to include tables that summarize material. Like
figures, these should be centered, legible, and numbered consecutively.
However, place the title \emph{above} the table with at least
0.1~inches of space before the title and the same after it, as in
\cref{sample-table}. The table title should be set in 9~point
type and centered unless it runs two or more lines, in which case it
should be flush left.
\begin{table}[t]
\caption{Classification accuracies for naive Bayes and flexible
Bayes on various data sets.}
\label{sample-table}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lcccr}
\toprule
Data set & Naive & Flexible & Better? \\
\midrule
Breast & 95.9$\pm$ 0.2& 96.7$\pm$ 0.2& $\surd$ \\
Cleveland & 83.3$\pm$ 0.6& 80.0$\pm$ 0.6& $\times$\\
Glass2 & 61.9$\pm$ 1.4& 83.8$\pm$ 0.7& $\surd$ \\
Credit & 74.8$\pm$ 0.5& 78.3$\pm$ 0.6& \\
Horse & 73.3$\pm$ 0.9& 69.7$\pm$ 1.0& $\times$\\
Meta & 67.1$\pm$ 0.6& 76.5$\pm$ 0.5& $\surd$ \\
Pima & 75.1$\pm$ 0.6& 73.9$\pm$ 0.5& \\
Vehicle & 44.9$\pm$ 0.6& 61.5$\pm$ 0.4& $\surd$ \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
Tables contain textual material, whereas figures contain graphical material.
Specify the contents of each row and column in the table's topmost
row. Again, you may float tables to a column's top or bottom, and set
wide tables across both columns. Place two-column tables at the
top or bottom of the page.
\subsection{Theorems and such}
The preferred way is to number definitions, propositions, lemmas, etc. consecutively, within sections, as shown below.
\begin{definition}
\label{def:inj}
A function $f:X \to Y$ is injective if for any $x,y\in X$ different, $f(x)\ne f(y)$.
\end{definition}
Using \cref{def:inj} we immediate get the following result:
\begin{proposition}
If $f$ is injective mapping a set $X$ to another set $Y$,
the cardinality of $Y$ is at least as large as that of $X$
\end{proposition}
\begin{proof}
Left as an exercise to the reader.
\end{proof}
\cref{lem:usefullemma} stated next will prove to be useful.
\begin{lemma}
\label{lem:usefullemma}
For any $f:X \to Y$ and $g:Y\to Z$ injective functions, $f \circ g$ is injective.
\end{lemma}
\begin{theorem}
\label{thm:bigtheorem}
If $f:X\to Y$ is bijective, the cardinality of $X$ and $Y$ are the same.
\end{theorem}
An easy corollary of \cref{thm:bigtheorem} is the following:
\begin{corollary}
If $f:X\to Y$ is bijective,
the cardinality of $X$ is at least as large as that of $Y$.
\end{corollary}
\begin{assumption}
The set $X$ is finite.
\label{ass:xfinite}
\end{assumption}
\begin{remark}
According to some, it is only the finite case (cf. \cref{ass:xfinite}) that is interesting.
\end{remark}
\subsection{Citations and References}
Please use APA reference format regardless of your formatter
or word processor. If you rely on the \LaTeX\/ bibliographic
facility, use \texttt{natbib.sty} and \texttt{icml2022.bst}
included in the style-file package to obtain this format.
Citations within the text should include the authors' last names and
year. If the authors' names are included in the sentence, place only
the year in parentheses, for example when referencing Arthur Samuel's
pioneering work \yrcite{Samuel59}. Otherwise place the entire
reference in parentheses with the authors and year separated by a
comma \cite{Samuel59}. List multiple references separated by
semicolons \cite{kearns89,Samuel59,mitchell80}. Use the `et~al.'
construct only for citations with three or more authors or after
listing all authors to a publication in an earlier reference \cite{MachineLearningI}.
Authors should cite their own work in the third person
in the initial version of their paper submitted for blind review.
Please refer to \cref{author info} for detailed instructions on how to
cite your own papers.
Use an unnumbered first-level section heading for the references, and use a
hanging indent style, with the first line of the reference flush against the
left margin and subsequent lines indented by 10 points. The references at the
end of this document give examples for journal articles \cite{Samuel59},
conference publications \cite{langley00}, book chapters \cite{Newell81}, books
\cite{DudaHart2nd}, edited volumes \cite{MachineLearningI}, technical reports
\cite{mitchell80}, and dissertations \cite{kearns89}.
Alphabetize references by the surnames of the first authors, with
single author entries preceding multiple author entries. Order
references for the same authors by year of publication, with the
earliest first. Make sure that each reference includes all relevant
information (e.g., page numbers).
Please put some effort into making references complete, presentable, and
consistent, e.g. use the actual current name of authors.
If using bibtex, please protect capital letters of names and
abbreviations in titles, for example, use \{B\}ayesian or \{L\}ipschitz
in your .bib file.
\section*{Accessibility}
Authors are kindly asked to make their submissions as accessible as possible for everyone including people with disabilities and sensory or neurological differences.
Tips of how to achieve this and what to pay attention to will be provided on the conference website \url{http://icml.cc/}.
\section*{Software and Data}
If a paper is accepted, we strongly encourage the publication of software and data with the
camera-ready version of the paper whenever appropriate. This can be
done by including a URL in the camera-ready copy. However, \textbf{do not}
include URLs that reveal your institution or identity in your
submission for review. Instead, provide an anonymous URL or upload
the material as ``Supplementary Material'' into the CMT reviewing
system. Note that reviewers are not required to look at this material
when writing their review.
\section*{Acknowledgements}
\textbf{Do not} include acknowledgements in the initial version of
the paper submitted for blind review.
If a paper is accepted, the final camera-ready version can (and
probably should) include acknowledgements. In this case, please
place such acknowledgements in an unnumbered section at the
end of the paper. Typically, this will include thanks to reviewers
who gave useful comments, to colleagues who contributed to the ideas,
and to funding agencies and corporate sponsors that provided financial
support.
\nocite{langley00}
\subsubsection*{Acknowledgments}
We greatly thank the actionable suggestions given by reviewers. S. Miao and M. Liu are supported by the National Science Foundation (NSF) award HDR-2117997. P. Li is supported by the JPMorgan Faculty Award.
|
\section{Introduction}\label{S:0}
In this paper, we study the large deviations principle (LDP)~for~the incompressible Navier--Stokes (NS) system on the torus~${\mathbb T}^2={\mathbb R}^2/2\pi{\mathbb Z}^2$:
\begin{equation} \label{0.1}
\partial_t u-\nu\Delta u+\langle u,\nabla \rangle u+\nabla p=\eta(t,x), \quad\mathop{\rm div}\nolimits u=0, \quad x\in {\mathbb T}^2.
\end{equation}Here $u = (u_1(t,x), u_2(t,x))$ and~$p = p(t, x)$ are the unknown velocity field and pressure of the fluid,
$\nu>0$ is the viscosity, and $\eta$ is an external random force. We consider this system in the usual space
\begin{equation}\label{0.2}
H=\Bigg\{u\in L^2({\mathbb T}^2,{\mathbb R}^2): \, \int_{{\mathbb T}^2} u(x) {\textup d} x=0, \,\,\, \mathop{\rm div}\nolimits u=0 \,\, \text{in }{\mathbb T}^2 \Bigg\}
\end{equation}
endowed with the $L^2$-scalar product $\langle \cdot, \cdot\rangle$ and the corresponding norm $\|\cdot\|.$
Projecting the system \eqref{0.1} to the space~$H$, we eliminate the pressure term and obtain the evolution~equation
\begin{equation} \label{0.3}
\partial_t u-\nu \Delta u +\Pi(\langle u,\nabla \rangle u) = \Pi \eta,
\end{equation}where $\Pi$ is the Leray projection to $H$ in~$L^2({\mathbb T}^2,{\mathbb R}^2)$ (see Section~6 in Chapter~1 of \cite{lions1969}).
We assume that~$\eta$ is a white-in-time~noise of the form
\begin{equation}\label{0.4}
\eta(t,x)= \partial_t W(t,x),\quad W(t,x)=\sum_{l\in {\cal K}} b_lW_l(t) e_l(x),
\end{equation}where ${\cal K}\subset {\mathbb Z}^2_*$ is a finite set,
$\{b_l\}_{l\in {\cal K}}$ are non-zero real numbers, $\{W_l\}_{l\in {\cal K}} $ are independent standard
Brownian motions on a filtered probability space~$(\Omega,{\cal F}, \{{\cal F}_t\}, {\mathbb P})$ satisfying the usual conditions (see Definition 2.25 in~\cite{KS1991}), and
$$
e_l(x)= \begin{cases} l^\bot\cos\langle l,x\rangle & \text{if }l_1>0\text{ or } l_1=0,\, l_2>0, \\ l^\bot \sin\langle l,x\rangle & \text{if }l_1<0\text{ or } l_1=0,\, l_2<0 , \quad l =(l_1,l_2)\end{cases}
$$
with $l^\bot=(- l_2, l_1).$
In other words, ${\cal K}$ is the collection of the Fourier modes directly perturbed by the noise.~Under the above assumptions, the NS system~\eqref{0.3}
defines a family of Markov processes~$(u_t,{\mathbb P}_u)$ parametrised by the initial condition $u(0)=u \in H$.~The~ergodic properties of this family have been extensively studied in the literature. It~is~now well known that~$(u_t,{\mathbb P}_u)$ admits a unique and exponentially mixing stationary measure, provided that the set ${\cal K}$ is sufficiently large.~Under the condition that~${\cal K}$ contains all the determining modes, the ergodicity has~been~established in different settings in
the papers~\cite{FM-1995,KS-cmp2000,EMS-2001, KS-jmpa2002, BKL-2002}.~Later, it~was shown that the ergodicity remains true for much smaller set ${\cal K}$; see~the papers~\cite{HM-2006, HM-2011,FGRT-2015}
for the case when the noise is white-in-time and~\cite{KNS-2018, KNS-2019} for the case of a
general bounded noise.~The~reader is referred to the book~\cite{KS-book} for more references and for detailed description of different methods.
In this paper, we study the Donsker--Varadhan type LDP for the NS system~\eqref{0.3}.~This type of LDP has been extensively studied in the case of finite-dimensional diffusions and Markov processes in compact spaces;
see the papers~\cite{DV-1975}, the books~\cite{FW1984,DS1989,ADOZ00}, and the references therein.~The paper~\cite{Wu-2001} established a general criterion for Donsker--Varadhan type LDP for Markov processes that are strong Feller and irreducible. In that paper the criterion is applied to a class of stochastic damping Hamiltonian systems. There are only few papers considering the problem of LDP for randomly forced PDEs.~The first results are obtained in~\cite{gourcy-2007b, gourcy-2007a} in the case of the stochastic Burgers and NS equations with strong assumptions on~the decay of the coefficients~$\{b_l\}$.~Indeed, these papers use the criterion of~\cite{Wu-2001}, so they require some lower bounds for $\{b_l\}$ in order to guarantee the strong~Feller~property. These assumptions have been relaxed to the conditions~$b_l\neq 0$ for all~$l\in {\mathbb Z}^2_*$ and~$\sum_{l\in {\mathbb Z}^2_*}|l||b_l|^2<+\infty $
in the papers~\cite{JNPS-cpam2015, JNPS-2018}, where a family of dissipative PDEs is considered driven by a random kick-force. The proofs of these papers are based on a study~of the long-time behaviour of Feynman--Kac semigroup and a Kifer type criterion for the LDP.~Under similar non-degeneracy conditions, the local LDP is proved in~\cite{MN-2015} for the stochastic damped nonlinear wave equation, and the full LDP is proved in~\cite{VN-2019} for the stochastic NS system.~A~controllability approach is used in~\cite{JNPS-2019} to prove the LDP for the Lagrangian trajectories of the NS system. Recently,
the criterion of~\cite{Wu-2001} has been used in~\cite{WX-2018} in the case of SPDEs driven by stable type noises and in~\cite{WXX-2021} in the case of non-linear monotone SPDEs with white-in-time noise.
\smallskip
All the papers mentioned above establish the LDP under the assumption that the noise is non-degenerate, i.e., perturbs directly all the Fourier modes in the equation.~The goal of the present paper is to establish the LDP in the case of a highly degenerate noise, i.e., when only few Fourier modes are directly perturbed.~To formulate our main result, let us recall that a set ${\cal K}\subset {\mathbb Z}^2_*$
is a generator if any element of~${\mathbb Z}^2$
is a finite linear
combination of elements of~${\cal K}$ with integer coefficients. In~what follows, we assume that the following condition~is~satisfied.
\begin{description}
\item[\hypertarget{H}]{\bf (H)}
{\sl The set ${\cal K}\subset {\mathbb Z}^2_*$ in \eqref{0.4} is a finite symmetric (i.e., $-{\cal K}={\cal K}$) generator
that contains at least two non-parallel vectors~$m$ and~$n$ such that $|m|\neq |n|$.}
\end{description}
This is the condition under which the ergodicity of the NS system is established in~\cite{HM-2006, HM-2011} in the case of a white-in-time noise and in \cite{KNS-2018} in the case of a bounded noise. The set
$$
{\cal K}=\{(1,0), \, (-1,0), \, (1,1), \, (-1,-1)\}\subset {\mathbb Z}^2_*
$$ is an example satisfying this condition.
For any $u\in H$, let us define the family of occupation~measures \begin{equation} \label{0.5}
\zeta_t=\frac1t\int_{0}^{t}\delta_{u_s} {\textup d} s, \,\,\, t>0
\end{equation}on the probability space $(\Omega,{\cal F},{\mathbb P}_u)$, where
$\delta_v$ is the Dirac measure concentrated at $v\in H$.
\begin{mtheorem}
Under the Condition~\hyperlink{H}{\rm(H)}, the family~$\{\zeta_t,t>0\}$ satisfies the~LDP.
\end{mtheorem}
See Theorem~\ref{T:1.1} for more detailed formulation of this result.~The proof is carried out by using a criterion for the LDP developed in~\cite{JNPS-2018}
in a discrete-time setting and extended in~\cite{MN-2015} to the continuous-time.~According to that criterion, the LDP will be established if we show that the following five properties hold for the Feynman--Kac semigroup associated with the NS system~\eqref{0.3}:
growth~properties, existence of eigenvector, time-continuity, uniform irreducibility, and uniform Feller property.~The first three properties are verified in~\cite{VN-2019} and they hold no matter how degenerate is the noise. The~uniform irreducibility~property follows from the approximate controllability results obtained in~\cite{AS-2005, AS-2006}.~It is interesting to note that Condition~\hyperlink{H}{\rm(H)} is necessary and sufficient for the approximate controllability of the NS system if one uses controls acting via the
Fourier modes in ${\cal K}$.~The main technical difficulty of this paper is related to the verification of the uniform Feller property, which we carry out by developing
the Malliavin calculus analysis of the papers~\cite{MP-2006, HM-2006, HM-2011}.
More precisely, we derive the uniform Feller property from a gradient estimate for the
Feynman--Kac semigroup.~The proof of latter contains essential differences with respect to the situations studied in~\cite{MP-2006, HM-2006, HM-2011} because of the non-Markovian character of the Feynman--Kac semigroup.
\smallskip
This paper is organised as follows. In Section \ref{S:1}, we
explain how the Main Theorem is derived from the above-mentioned five properties. In Section \ref{S:2}, we recall some elements of Malliavin calculus, and in Section \ref{S:3}, we verify the uniform Feller property.
\subsubsection*{Acknowledgement}
Xuhui Peng would like to gratefully thank University of Macau for the hospitality, his research is supported in part by
by NNSFC (No. 12071123), Scientific Research
Project of Hunan Province Education Department (No. 20A329) and Program of Constructing Key Discipline in Hunan Province. Lihu Xu is supported in part by NNSFC grant (No.12071499), Macao S.A.R grant FDCT 0090/2019/A2 and University of Macau grant MYRG2018-00133-FST.
\subsubsection*{Notation}
In this paper, we use the following notation.
\smallskip
\noindent
$H$ is the space of divergence-free square-integrable vector fields on ${\mathbb T}^2$ with zero mean value (see~\eqref{0.2}).
It is endowed with the $L^2$-norm $\|\cdot\|$.
\smallskip
\noindent
$H^m= H^m({\mathbb T}^2, {\mathbb R}^2)\cap H$, where $H^m({\mathbb T}^2, {\mathbb R}^2)$ is the Sobolev space of order~$m\ge1$. We endow the space $H^m$ with the usual Sobolev norm~$\|\cdot \|_m$.
\smallskip
\noindent
$B_{H^m}(a,r)$ is the closed ball in $H^m$ of radius~$r>0$ centred at~$a$.~We write~$B_{H^m}(r)$ when~$a=0$.
\smallskip
\noindent
We consider the NS system in the vorticity formulation in the space of square integrable zero mean functions:
\begin{align}
\label{0.6}
\tilde{H}=\left\{w\in L^2({\mathbb T}^2,{\mathbb R}):\int_{{\mathbb T}^2}w(x){\textup d} x=0\right\}
\end{align}
equipped with the $L^2$-norm $ \|\cdot\|.$
Let $\tilde H^m= H^m({\mathbb T}^2, {\mathbb R})\cap \tilde H$, $m\ge1$ be endowed with the Sobolev norm denoted by $\|\cdot \|_m$.
\smallskip
\noindent
$L^\infty(H)$ is the space of bounded Borel-measurable functions $\psi:H\to{\mathbb R}$ with the norm $\|\psi\|_\infty=\sup_{u\in H}|\psi(u)|$.~$C_b(H)$ is the space of continuous functions $\psi\in L^\infty(H)$.~$C^1_b(H)$ is the space of functions~$\psi\in C_b(H)$ that are continuously Fr\'echet differentiable with bounded derivative.
\smallskip
\noindent
Let ${\mathfrak w}:H\to[1,+\infty]$ be a Borel-measurable function. Then $C_{\mathfrak w}(H)$ (resp.,~$L_{\mathfrak w}^\infty(H)$) is the space of continuous (resp., Borel-measurable) functions $\psi:H\to{\mathbb R}$ such that
$\|\psi\|_{L_{\mathfrak w}^\infty}=\sup_{u\in H} |\psi(u)|/{\mathfrak w}(u)<+\infty.$
\smallskip
\noindent
${\cal M}_+(H)$ is the collection of non-negative finite Borel measures on $H$ endowed~with the weak convergence topology.~For any~$\psi\in L^\infty(H)$ and $\mu\in {\cal M}_+(H)$, we write~$\langle \psi,\mu\rangle=\int_H \psi(u)\mu({\textup d} u).$
${\cal P}(H)$ is the subset of probability measures, and~${\cal P}_{\mathfrak w}(H)$ is the set of $\mu\in{\cal P}(H)$ such that $\langle {\mathfrak w},\mu\rangle<+\infty$.
\smallskip
\noindent
${\mathcal L}(X,Y)$ is the space of linear bounded operators between Banach spaces $X$ and~$Y$ endowed with the natural norm $\|\cdot\|_{{\mathcal L}(X,Y)}$.
\smallskip
\noindent
The letter $C$ is used to denote unessential constants that can change from line to line.
\section{Main results}\label{S:1}
\subsection{LDP and multiplicative ergodicity}\label{S:1.1}
Recall that a mapping $I : {\cal P}(H) \to [0, +\infty]$ is a good rate function if its level~sets
$$
\left\{\sigma\in{\cal P}(H) : I(\sigma) \le \alpha \right\}, \quad \alpha \ge0
$$ are compact. Moreover, if the effective domain of $I$, defined by
$$
D_I=\left\{\sigma\in{\cal P}(H):I(\sigma)<+\infty \right\},
$$ is not a singleton, we say that $I$ is a non-trivial good rate function.
For any~$\gamma>0$ and $M>0$, we set
$$
\Lambda(\gamma,M) : = \Bigg \{\nu \in {\cal P}(H) : \int_H e^{\gamma\|u\| ^2 }\nu ( {\textup d} u) \le M \Bigg\}.
$$
The following is a more detailed version of the Main Theorem formulated in the Introduction.
\begin{theorem} \label{T:1.1}
Assume that Condition~\hyperlink{H}{\rm(H)} is verified.~Then, for any~$\gamma>0$ and~$M>0$, the family of random probability measures~$\{\zeta_t,t>0\}$ defined by~\eqref{0.5} satisfies the LDP uniformly w.r.t. the initial measure $\nu\in \Lambda(\gamma,M)$. More~precisely, there is a non-trivial good rate function $I: {\cal P}(H)\to [0,+\infty]$ that does not depend on $\gamma$ and $M$ and satisfies the inequalities \begin{gather*}
\limsup_{t\to+\infty} \frac{1}{t} \log \sup_{\nu\in \Lambda(\gamma,M)} {\mathbb P}_\nu\left\{\zeta_t\in F\right\}\le -\inf_{\sigma\in F} I(\sigma),\\
\liminf_{t\to+\infty} \frac{1}{t} \log \inf_{\nu\in \Lambda(\gamma,M)}{\mathbb P}_\nu\left\{\zeta_t\in G\right\}\ge -\inf_{\sigma\in G}I(\sigma)
\end{gather*}for any closed set $F$ and any open set $G$ in ${\cal P}(H)$.
\end{theorem}
This theorem is derived from a multiplicative ergodic theorem for the NS system. To formulate that result, let us introduce the following weight functions:
\begin{align*}
{\mathfrak m}_\gamma(u)&=\exp\left(\gamma \|u\|^2\right), \quad \gamma>0,\\
{\mathfrak w}_m(u)&=1+\|u\|^{2m},\quad m\ge1,\,\, u\in H.
\end{align*}There is a constant $\gamma_0=\gamma_0({\mathfrak B}_{0})>0$, where ${\mathfrak B}_{0}=\sum_{l\in {\cal K}}b_l^2$, such that
\begin{align}
{\mathbb E}_u {\mathfrak m}_\gamma(u_t) &\le e^{-\gamma t} {\mathfrak m}_\gamma(u)+C,\label{1.1}
\\
{\mathbb E}_u {\mathfrak w}_m(u_t) &\le e^{-2m t} {\mathfrak w}_m(u)+C \label{1.2}
\end{align} for any $\gamma\in(0,\gamma_0)$, $m\ge1$, $u\in H$, and $t\ge0$,
where $C=C(m,\varkappa, {\mathfrak B}_{0})>0$ is a constant; e.g.,~see Proposition~2.4.9 in~\cite{KS-book} and Lemma~5.3 in~\cite{VN-2019} for a proof of these inequalities.
For any $V\in C_b(H),$
the Feynman--Kac semigroup associated with the Markov family $(u_t,{\mathbb P}_u)$ is defined by
$$
{\mathfrak P}_t^V\psi(u)={\mathbb E}_u \left\{ \exp\left(\int_{0}^tV(u_s) {\textup d} s\right) \psi(u_t)\right\}, \quad {\mathfrak P}_t^V: C_b(H)\to C_b(H);
$$ its dual is denoted by ${\mathfrak P}_t^{V*}:{\cal M}_+(H)\to {\cal M}_+(H)$.~From~\eqref{1.1} it~follows that~${\mathfrak P}_t^V$ maps the space $C_{{\mathfrak m}_\gamma}(H)$ into itself
for $\gamma\in(0,\gamma_0)$.
\begin{theorem} \label{T:1.2}
Assume that Condition~\hyperlink{H}{\rm(H)} is verified and $ V\in C_b^1(H)$.~Then there are constants $ m=m(V)\ge1$ and $\gamma=\gamma( {\mathfrak B}_{0})\in (0,\gamma_0)$ such that
there are unique eigenvectors $ h_V \in C_{{\mathfrak w}_m}(H)$ and $ \mu_V \in {\cal P}_{{\mathfrak m}_\gamma}(H)$ for the semigroups ${\mathfrak P}_t^V$ and ${\mathfrak P}_t^{V*}$ corresponding to an
eigenvalue $\lambda_V>0$, i.e,
$${\mathfrak P}^{V*}_t\mu_V=\lambda_V^t \mu_V, \quad\quad {\mathfrak P}^V_t h_V=\lambda_V^t h_V\quad \quad \text{for $t>0$},
$$ and normalised by $\langle h_V,\mu_V\rangle=1$.
For any $\psi\in C_{{\mathfrak m}_\gamma} (H),\, \nu\in {\cal P}(H)$, and~$R>0$, the following limits hold as~$t\to+\infty$:
\begin{align*}
\lambda_V^{-t}{\mathfrak P}_t^V\psi\to\,&\langle \psi,\mu_V\rangle h_V
\,\,\textup{in } C_b(B_H(R))\cap L^1(H,\mu_V), \\
\lambda^{-t}_V{\mathfrak P}_t^{V*}\nu \to\,&\langle h_V,\nu\rangle\mu_V
\,\,\textup{in }{\cal M}_+(H).
\end{align*}
Furthermore, for any $M>0$ and $\varkappa\in (0,\gamma)$,
$$
\lambda_V^{-t}\,{\mathbb E}_\nu\left\{\exp\bigg(\int_0^tV(u_s){\textup d} s \bigg) \psi(u_t)\right\}\to\langle \psi,\mu_V\rangle\,\langle h_V,\nu\rangle
$$
uniformly w.r.t. $\nu\in\Lambda(\varkappa,M)$ as $t\to+\infty$.
\end{theorem}
This theorem improves Theorem~1.1 in~\cite{VN-2019} in two directions. First, in this theorem, the noise is very degenerate, while in \cite{VN-2019} all the Fourier modes are assumed to be directly perturbed by the noise. Second, in the present situation, the class of functions~$V$ is larger, since the result in \cite{VN-2019} applies only to functions depending on finite-dimensional projection of~$u$.
Theorem~\ref{T:1.2} can be viewed as an improvement of Theorem~2.1 in~\cite{HM-2006}. Indeed, in the case~$V=0$, the Feynman--Kac semigroup reduces to the Markov semigroup with eigenvalue $\lambda_V=1$, eigenvector $h_V={\mathbf 1}$ (the function identically equal to~$1$~on~$H$), and the measure~$\mu_V =\mu$ is the unique stationary measure. The above limits imply that $\mu$ is mixing.
Theorem~\ref{T:1.1} is derived from Theorem~\ref{T:1.2} by using a Kifer type criterion
in unbounded spaces. Since this derivation is literally the same as in the non-degenerate case (see Section~1 in~\cite{VN-2019}), we do not give the details. The proof
of Theorem~\ref{T:1.2} is discussed in the next subsection.
\subsection{Proof of Theorem~\ref{T:1.2}}\label{S:1.2}
The proof of Theorem~\ref{T:1.2} is carried out by applying a result on large-time asymptotics of generalised Markov semigroups
established in \cite{JNPS-2018} in the discrete-time setting and extended in \cite{MN-2015} to the continuous-time.~Here we apply that result to the
Feynman--Kac semigroup ${\mathfrak P}_t^V$ and the associated kernel~$P_t^V(u,\Gamma)=({\mathfrak P}_t^{V*} \delta_u ) (\Gamma)$,~$u\in H$, $ \Gamma\in {\mathcal B}(H),$
where
$\delta_u$ is the Dirac measure concentrated at $u$.
By the regularising property of the NS system,
the measure~$P^V_{t}(u,\cdot)$ is concentrated on the space~$H^2$ for any $u\in H$ and $t>0$.~For any~$R>0$, let us
denote~$X_R=B_{H^2}(R)$, and let $V\in C_b(H)$ be arbitrary.~Then the following properties~hold.
\begin{description}
\item[Growth properties.] There are numbers $R_0>0$, $\gamma\in (0,\gamma_0)$, and $m\ge1$ such that the following quantities are finite:
\begin{equation} \label{1.3}
\sup_{t\ge0}
\frac{\|{\mathfrak P}^V_t{\mathfrak w}_m\|_{L_{{\mathfrak w}_m}^\infty}}{\|{\mathfrak P}^V_t{{\mathbf 1}}\|_{R_0}},\,\,\,\,
\sup_{t\ge0}
\frac{\|{\mathfrak P}_t^V{\mathfrak m}_\gamma\|_{L^\infty_{{\mathfrak m}_\gamma}}}{\|{\mathfrak P}_t^V{{\mathbf 1}}\|_{R_0}}, \,\,\, \,
\sup_{t\ge1}
\frac{\|{\mathfrak P}_t^V \Phi \|_{L^\infty_{{\mathfrak m}_\gamma}}}{\|{\mathfrak P}_t^V{{\mathbf 1}}\|_{R_0}},
\end{equation}
where~$\|\psi\|_R=\sup_{u\in X_R}|\psi(u)|$ and~$\Phi(u)=\|u\|_{H^2}^2$.
\item[Existence of an eigenvector.]For any $t>0$, there is a measure $\mu_{t,V}\in{\cal P}(H)$ and a number $\lambda_{t,V}>0$ such that ${\mathfrak P}_t^{V*}\mu_{t,V}=\lambda_{t,V}\mu_{t,V}$.~Moreover, for any numbers~$\varkappa\in (0,\gamma_0)$ and $n,m\ge1$, we have
\begin{gather*}
\int_H\! \left(\|u\|_{H^2}^{n}+{\mathfrak m}_\varkappa (u) \right) \mu_{t,V}({\textup d} u)<+\infty, \\
\|{\mathfrak P}_t^V{\mathfrak w}_m\|_{X_R}\int_{X_R^c}\!{\mathfrak w}_m(u)\mu_{t,V}({\textup d} u)\to0\,\,\textup{as } R\to+\infty.
\end{gather*}
\item[Time-continuity.]
For any $m\ge1$, $ \psi\in C_{{\mathfrak w}_m}(H) $, and $u\in H$,
the function $t\mapsto{\mathfrak P}^V_t \psi(u) $, ${\mathbb R}_+\to{\mathbb R}$ is continuous.
\item[Uniform irreducibility.] For
any $\rho, r, R> 0$, there are numbers $l=l(\rho,r,R)>0$ and~$p=p(V,\rho,r)>0$ such that
\begin{equation}\label{1.4}
P^V_l(u_0,B_{H}(\hat u,r))\ge p
\end{equation} for any $u_0\in X_R$ and $\hat u\in X_\rho$.
\item[Uniform Feller property.]
The family of functions $\{\|{\mathfrak P}_t^V{\mathbf 1}\|_R^{-1} {\mathfrak P}_t^V\psi,t\geq 0 \}$
is~uniformly equicontinuous\,\footnote{By uniform equicontinuity of $\{\|{\mathfrak P}_t^V{{\mathbf 1}}\|_R^{-1}{\mathfrak P}^V_t \psi, t\geq 0\}$ on $X_R$ we mean that for any $\varepsilon>0$, there is $\delta>0$ such that
$\|{\mathfrak P}^V_t{{\mathbf 1}}\|_R^{-1} \big|{\mathfrak P}_t^V\psi (u)-{\mathfrak P}_t^V\psi (u')\big|<\varepsilon$ for any $u,u'\in X_R$ with $\|u-u'\|< \delta $ and any $t\ge0$.
} on $X_R$ for any $V, \psi\in C_b^1(H)$ and $R\geq R_0.$
\end{description}
The first three of the above properties are established\,\footnote{These propositions and lemma in~\cite{VN-2019} are formulated in the case when $X_R=B_{H^1}(R)$ and the noise is non-degenerate. However, their proofs work in the setting
of the present paper without any change.} in Propositions~2.1 and~2.5 and Lemma~2.3 in~\cite{VN-2019}.~The proof of the uniform irreducibility is given below, and the uniform Feller property is established in Section \ref{S:3}.~Theorem~\ref{T:1.2} is obtained by applying Theorem~7.4 in~\cite{MN-2015} and by literally repeating the arguments of Section~4 in~\cite{VN-2019}.
\begin{proof}[Proof of uniform irreducibility] Let $P_t(u_0,\cdot)$ be the Markov transition kernel of the family~$(u_t,{\mathbb P}_{u_0})$. The boundedness of
$V$ implies that
\begin{equation}\label{1.5}
P_t^V(u_0,{\textup d} v)\ge e^{-t\|V\|_\infty}P_t(u_0,{\textup d} v) \quad\textup{for } t>0,\, u_0\in H.
\end{equation} According to~\cite{AS-2005, AS-2006}, under Condition~\hyperlink{H}{\rm(H)}, the NS system is approximately controllable in the space~$H$ by controls taking values in the space
$$
{\cal H}_{\cal K}=\mathop{\rm span}\nolimits\{e_l:l\in {\cal K}\}.
$$ This implies that, for any~$u_0, \hat u\in H$ and $r>0$, there is a function $\zeta\in C^\infty([0,1]; {\cal H}_{\cal K})$ such~that
$$
\|u(1,u_0, \zeta)-\hat u\| < r,
$$where $u(t,u_0, \zeta)$ is the solution of the deterministic NS system \eqref{0.3} with the initial condition $u(0)=u_0$ and the (control) force~$\eta= \partial_t \zeta$.~Using the fact that the mapping $(u_0, \zeta)\mapsto u(1,u_0, \zeta)$ is continuous from $H\times C([0,1]; {\cal H}_{\cal K})$ to~$H$, the non-degeneracy of the law of the Wiener process~$ W$ in $C([0,1]; {\cal H}_{\cal K})$ (i.e.,~the~support of the law of~$ W$ coincides with the
entire~space $C([0,1]; {\cal H}_{\cal K})$), a simple compactness argument, and inequality \eqref{1.5}, we~arrive~at~\eqref{1.4}.
\end{proof}
\section{Elements of Malliavin calculus}\label{S:2}
The uniform Feller property is proved by using Malliavin calculus analysis from the papers~\cite{MP-2006, HM-2006, HM-2011}.
In this section, we recall some basic definitions and estimates from there. To match the framework of these papers, we rewrite the NS system \eqref{0.3} in the vorticity formulation:
\begin{equation}
\label{2.1}
\partial_t w -\nu \Delta w+B({\cal K} w,w)=\sum_{l \in {\cal K} } b_l |l|^2 \dot W_l(t) \phi_l,
\end{equation}where $w=\nabla\wedge u$, $B(u,w)=\langle u,\nabla \rangle w$,
and ${\cal K}$ is the Biot--Savart operator
$$
{\cal K} w=\sum_{l\in {\mathbb Z}_*^2} |l|^{-2} l^\bot w_{-l} \phi_l
$$
with $|l|^2=l_1^2+l_2^2$, $l^\bot=(-l_2,l_1)$, $w_l=\langle w,\phi_l\rangle$, and
$$
\phi_l(x)=
\begin{cases}
\sin \langle l,x\rangle& \text{if }l_1>0\text{ or } l_1=0,\, l_2>0, \\ -\cos\langle l,x\rangle & \text{if }l_1<0\text{ or } l_1=0,\, l_2<0, \quad l =(l_1,l_2).
\end{cases}
$$
The operator ${\cal K}$ is continuous from $H^s({\mathbb T}^2;{\mathbb R})$ to $H^{s+1}({\mathbb T}^2;{\mathbb R}^2)$ for any~$s\in {\mathbb R}$; it~allows to recover the velocity field from the vorticity via $u={\cal K} w$.
We consider Eq.~\eqref{2.1} in the space $\tilde H$ of real-valued square-integrable functions on~$\mathbb{T}^2$ with zero mean value (see \eqref{0.6}); it is endowed with the $L^2$ norm $ \|\cdot\|.$ Since~the underlying probability space plays no role,
without loss of generality, we~can assume that $\Omega$ is the Wiener space, $W(t)=\{W_l(t)\}_{l\in{\cal K}}$ is the canonical process, and ${\mathbb P}$ is the Wiener measure.~Furthermore, we denote by $\{\theta_l\}_{l\in{\cal K}}$ the standard basis in ${\mathbb R}^d$ with $d=|{\cal K}|$, and define a
linear map $Q:{\mathbb R}^d\to \tilde H$ by~$Q\theta_l=b_l |l|^2\phi_l.$
Let~$w_t=\Phi(t,w,W_\cdot)$ be~the solution of Eq.~\eqref{2.1} with initial value $w(0)=w\in \tilde{H}$.
For any $0\le s\le t$ and $\xi\in \tilde H$, let~$J_{s,t}\xi$ be the solution of the linearised problem:
\begin{align}
\partial_t J_{s,t}\xi-\nu \Delta J_{s,t}\xi +\tilde{B}(w_t,J_{s,t}\xi)&=0, \label{2.2}
\\ J_{s,s}\xi&=\xi, \nonumber
\end{align}
where $\tilde{B}(w,v)=B({\cal K} w,v)+B({\cal K} v,w)$.
Recall that, for given $T>0$ and $v\in L^2([0,T];{\mathbb R}^d),$
the Malliavin derivative of~$w_t$ in the direction~$v$ is defined by
$$
{\mathcal D}^v w_t=\lim_{\varepsilon \to 0}\frac{1}{\varepsilon}
\left(\Phi(t,w_0,W+\varepsilon \int_0^\cdot v{\textup d} s)-\Phi(t,w_0,W)\right), \quad
$$
where the limit holds almost surely (e.g., see the book~\cite{nualart2006} for finite-dimensional setting or the papers~\cite{MP-2006, HM-2006, HM-2011, FGRT-2015} for Hilbert space~case).~By the Riesz representation theorem, there is a linear operator $ {\mathcal D}:L^2(\Omega, \tilde H)\to L^2(\Omega; L^2([0,T];{\mathbb R}^d)\otimes \tilde H)$ such that
\begin{equation}\label{2.3}
{\mathcal D}^v w =\langle {\mathcal D} w,v \rangle_{L^2([0,T];{\mathbb R}^d)}.
\end{equation}
On the other hand, we have
\begin{equation}\label{2.4}
{\mathcal D}^v w_t={\mathcal A}_{0,t}v,
\end{equation} where ${\mathcal A}_{s,t}:L^2([s,t];{\mathbb R}^{d})\to \tilde H$ is the random operator defined by
\begin{equation}
\label{2.5}
{\mathcal A}_{s,t}v=\int_{s}^{t}J_{r,t}Qv(r){\textup d} r, \quad 0\le s\le t\le T,
\end{equation}i.e., ${\mathcal A}_{s,t}v$ is
the solution of the linearised problem with a source term:
\begin{align*}
\partial_t {\mathcal A}_{s,t}v-\nu \Delta {\mathcal A}_{s,t}v +\tilde{B}(w_t,{\mathcal A}_{s,t}v)&=Qv,
\\ {\mathcal A}_{s,s}v&=0.
\end{align*}
The adjoint ${\mathcal A}_{s,t}^*:\tilde H \to L^2([s,t];{\mathbb R}^{d})$ is given by
$$
({\mathcal A}_{s,t}^*\xi)(r)=Q^{*}J_{r,t}^*\xi, \quad \xi\in \tilde H,\, r\in [s,t],
$$
where $Q^*:\tilde H\rightarrow {\mathbb R}^{d}$ is the adjoint of $Q$.
Let us
denote by $J^{(2)}_{s,t}(\phi,\psi)$ the second derivative of $w_t$ with respect to $w$ in the directions of $\phi$ and $\psi$. It is the solution of the problem
\begin{align*}
\partial_t J^{(2)}_{s,t}(\phi,\psi)- \nu \Delta J^{(2)}_{s,t}(\phi,\psi)+\tilde{B}(J_{s,t}\phi,J_{s,t}\psi)
+\tilde{B}(w_t,J^{(2)}_{s,t}(\phi,\psi))&=0,
\\ J^{(2)}_{s,s}(\phi,\psi)&=0.
\end{align*}The next lemma follows from Lemma 4.10 in \cite{HM-2006}.
\begin{lemma}
\label{L:2.1}
For any $\kappa,p>0$, $0\le \tau< T$, and $w\in \tilde H$, we have
\begin{align}
\label{2.6}
{\mathbb E}_w \sup_{s<t\in [\tau,T]}\|J_{s,t}\|^p_{{\mathcal L}( \tilde H,\tilde{H})}&\le C \exp\{\kappa\|w\|^2\},
\\
\nonumber
{\mathbb E}_w \sup_{s<t\in [\tau,T]} ||| J^{(2)}_{s,t} |||^p &\le C \exp\{\kappa\|w\|^2\} ,
\end{align}
where $ ||| J^{(2)}_{s,t} |||=\sup_{\|\phi\|,\|\psi\|\le1}
\|J^{(2)}_{s,t}(\phi,\psi)\|$, and $C=C(\kappa, p,T-\tau, \mathfrak{B}_0)>0 $ is a constant.
\end{lemma}
For any $0\le s< t,$ the Malliavin operator is defined by
$$
{\mathcal M}_{s,t}={\mathcal A}_{s,t}{\mathcal A}_{s,t}^*:\tilde{H}\rightarrow \tilde{H}.
$$ It is a
non-negative self-adjoint operator, so its regularisation~${\mathcal M}_{s,t}+\beta{\mathbb I}$
is invertible for any $\beta>0$. Here ${\mathbb I}$ is the identity.
The following lemma gathers some estimates from Section~4.8 in~\cite{HM-2006} and Lemma~A.6 in~\cite{FGRT-2015}.
\begin{lemma}
\label{L:2.2}There is a constant $C=C(\mathfrak{B}_0)>0$ such that, for any
$0\le s<t$, $\beta>0$, and~$w\in \tilde H$, we have
\begin{gather}
\label{2.7} \|{\mathcal A}_{s,t}\|^2_{{\mathcal L}(L^2([s,t];{\mathbb R}^d),\tilde{H})} \le C \int_s^t\|J_{r,t}\|_{{\mathcal L}( \tilde H,\tilde{H})}^2{\textup d} r ,
\\ \|{\mathcal A}_{s,t}^*({\mathcal M}_{s,t}+\beta{\mathbb I})^{-1/2}\|_{{\mathcal L}(\tilde{H},L^2([s,t];{\mathbb R}^d))} \le 1,\label{2.8}
\\
\|({\mathcal M}_{s,t}+\beta{\mathbb I})^{-1/2}{\mathcal A}_{s,t}\|_{{\mathcal L}(L^2([s,t];{\mathbb R}^d),\tilde{H})} \le 1,\label{2.9}
\\ \|({\mathcal M}_{s,t}+\beta{\mathbb I})^{-1/2}\|_{{\mathcal L}(\tilde{H},\tilde{H})} \le \beta^{-1/2}.\label{2.10}
\end{gather}
\end{lemma}
We shall use the notation
$$
{\mathcal D}_r F=({\mathcal D} F) (r), \quad {\mathcal D}^j F=({\mathcal D} F)^j,\quad {\mathcal D}_r^j F=({\mathcal D} F)^j(r), \quad j=1,\ldots, d.
$$
From the equalities \eqref{2.3}-\eqref{2.5} it follows that ${\mathcal D}_r^iw_t=J_{r,t}Q\theta_i$, $0\le r\le t$.
From this and \eqref{2.2}, we conclude that, for $0\le s<t$,
$$
\partial_t{\mathcal D}_r^iJ_{s,t}\xi
-\nu \Delta{\mathcal D}_r^iJ_{s,t}\xi+\tilde{B}(w_t, {\mathcal D}_r^iJ_{s,t}\xi)
+\tilde{B}( J_{r,t}Q\theta_i,J_{s,t}\xi)=0.
$$
Furthermore, by the variation of constants formula, we have
$$
{\mathcal D}_r^iJ_{s,t}\xi=
\begin{cases}
J^{(2)}_{r,t}( Q\theta_i,J_{s,r}\xi) & \text{ for } r\geq s, \\ J^{(2)}_{s,t}(J_{r,s}Q\theta_i,\xi) & \text{ for } r\le s.
\end{cases}
$$
This equality and Lemma \ref{L:2.1} imply the following lemma. For further details, we~refer the reader to Section~4.8 in~\cite{HM-2006} and Lemma~A.7 in \cite{FGRT-2015}.
\begin{lemma}\label{L:2.3}
The operators $J_{s,t}$, ${\mathcal A}_{s,t}$, and ${\mathcal A}_{s,t}^*$
are Malliavin differentiable, and for any
$\kappa>0$, $r\in[s,t]$, $p> 0$, and $w\in \tilde H$, the following inequalities~hold
\begin{align}
\label{2.11} {\mathbb E}_w \|{\mathcal D}_r^iJ_{s,t}\|^p_{{\mathcal L}(\tilde{H},\tilde{H})} & \le C \exp\{\kappa \|w\|^2\},
\\ \label {2.12}
{\mathbb E}_w \|{\mathcal D}_r^i {\mathcal A}_{s,t}\|^p_{{\mathcal L}(L^2([s,t];{\mathbb R}^d),\tilde{H})} & \le C \exp\{\kappa \|w\|^2\},
\\ \label{2.13} {\mathbb E}_w \|{\mathcal D}_r^i{\mathcal A}^*_{s,t}\|^p_{{\mathcal L}(\tilde{H},L^2([s,t];{\mathbb R}^d))} & \le C \exp\{\kappa \|w\|^2\},
\end{align}
where $C=C(\kappa,p,t-s,\mathfrak{B}_0)>0$.
\end{lemma}
\section{Proof of uniform Feller property} \label{S:3}
\subsection{Reduction to a gradient estimate} \label{S:3.1}
The aim of this section is to prove the following proposition.
\begin{proposition}\label{P:3.1}
Under Condition~\hyperlink{H}{\rm(H)}, for any $V, \psi\in C_b^1(H)$, there is a number $ R_0=R_0(V)>0$ such that
the family $\left\{\|{\mathfrak P}_t^V{\mathbf 1}\|_R^{-1} {\mathfrak P}_t^V\psi,t\geq 0 \right\}$
is uniformly equicontinuous on $X_R$ for any $R\geq R_0.$
\end{proposition}
\begin{proof} For any $V,\psi\in C_b^1(H),$ let us define functions
$\tilde V, \tilde \psi \in C_b^1(\tilde H)$
by
$\tilde{V}(w)= V({\cal K} w)$
and $ \tilde{\psi}(w)= \psi({\cal K} w)$, $ w\in\tilde H$.
The Feynman--Kac semigroup associated with Eq.~\eqref{2.1} is~given~by
$$\tilde {\mathfrak P}_t^{\tilde V} \tilde{\psi}(w)={\mathbb E}_w \left\{ \exp\left(\int_0^t\tilde V (w_s) {\textup d} s\right) \tilde \psi (w_t)\right\}, \quad \tilde {\mathfrak P}_t^{\tilde V}: C_b^1(\tilde H )\to C_b^1(\tilde H).
$$In what follows,
the number $R_0$ is chosen such that the growth properties~\eqref{1.3} hold.
In the next subsection, we prove the following proposition (cf.~Proposition~4.3 in~\cite{HM-2006}).
\begin{proposition}\label{P:3.2}
Under the conditions of Proposition \ref{P:3.1},
for any numbers $ \kappa>0$ and $ a\in (0,1)$, there is a constant $C=C(\kappa,a,\|\nabla V\|_\infty,\|V\|_\infty )>0$ such that
\begin{equation}\label{3.1}
\|\nabla_\xi \tilde{{\mathfrak P}}_t^{\tilde V}\tilde \psi(w)\|
\le
C \exp\{\kappa \|w\|^2\}\|{\mathfrak P}_t^V{\mathbf 1}\|_{R_0} \left[
\|\nabla \psi\|_\infty a^t+\|\psi\|_\infty \right]\|\xi\|
\end{equation}
for any $w,\xi\in \tilde H$ and $ t\geq 0$. Here
$\nabla_\xi$ is the derivative with respect
to the initial condition in the direction $\xi$.
\end{proposition}This result implies Proposition \ref{P:3.1}. Indeed, let us take any~$u_1,u_2\in X_R$ and set~$w_i=\nabla\wedge u_i$, $i=1,2$.
Using inequality \eqref{3.1} with any~$\kappa>0$ and $a\in (0,1)$
and an interpolation inequality, we see that
\begin{align*}
\big| {\mathfrak P}_t^V\psi(u_1)-{\mathfrak P}_t^V\psi(u_2)\big|
& = \big| \tilde{{\mathfrak P}}_t^{\tilde V} \tilde{\psi}(w_1)-
\tilde{{\mathfrak P}}_t^{\tilde V}\tilde{\psi}(w_2)\big|
\\ & \le C \|{\mathfrak P}_t^V\textbf{1}\|_{R_0} \|w_1-w_2\|
\\ & \le C \|{\mathfrak P}_t^V\textbf{1}\|_{R_0} \|u_1-u_2\|_1
\\ & \le C \|{\mathfrak P}_t^V\textbf{1}\|_{R_0} \|u_1-u_2\|^{1/2}\|u_1-u_2\|_2^{1/2}
\\ &\le C \|{\mathfrak P}_t^V\textbf{1}\|_{R_0} \|u_1-u_2\|^{1/2},
\end{align*}
where $C=C(R,\|V\|_\infty,\|\nabla V\|_\infty,\|\psi\|_\infty,\|\nabla \psi\|_\infty)>0.$
This~completes the proof of Proposition \ref{P:3.1}.
\end{proof}
\subsection{Proof of Proposition \ref{P:3.2}}\label{S:3.2}
Let us take any $\xi\in \tilde{H}$ with $\|\xi\|=1$, denote
$$
\Xi_t = \exp\left(\int_0^t \tilde V(w_s) {\textup d} s\right),
$$and compute the derivative of $\tilde {\mathfrak P}_t^{\tilde V} \tilde{\psi}(w)$ with respect to $w$ in the direction $\xi$:
\begin{equation}
\nabla_\xi \tilde{\mathfrak P}_t^{\tilde{V}}\tilde{\psi}(w) = {\mathbb E}_w\left[\Xi_t \tilde{\psi}( w_t)\int_0^t \nabla \tilde{V}(w_s) J_{0,s} \xi {\textup d} s
+\Xi_t \nabla\tilde{\psi}(w_t) J_{0,t}\xi\right].
\label{3.2}
\end{equation}
Inspired by the papers~\cite{HM-2006, HM-2011}, the idea of the proof of Proposition \ref{P:3.2} is~to approximate the perturbation
$J_{0,t}\xi$ caused by the perturbation $\xi$ of the initial condition with a variation ${\mathcal A}_{0,t}v$ coming from a variation of the noise by an appropriate process $v$. Let us denote by $\rho_t$ the residual error between~$J_{0,t}\xi$ and~${\mathcal A}_{0,t}v$:
$$
\rho_t=J_{0,t}\xi-{\mathcal A}_{0,t}v,
$$ replace the term $J_{0,t}\xi$ in \eqref{3.2} by ${\mathcal A}_{0,t}v+ \rho_t$, and recall that ${\mathcal A}_{0,t}v={\mathcal D}^vw_t$ is the Malliavin derivative of~$w_t$ in the direction~$v$. Then, at least formally, using the Malliavin
chain rule (see Proposition~1.2.3 in~\cite{nualart2006}), we have
\begin{align}
\nonumber \nabla_\xi \tilde{\mathfrak P}_t^{\tilde{V}}\tilde{\psi}(w) & = {\mathbb E}_w\left[\Xi_t \tilde{\psi}(w_t)\int_0^t \nabla \tilde{V}(w_s){\mathcal D}^vw_s {\textup d} s
+\Xi_t \nabla \tilde{\psi}(w_t){\mathcal D}^vw_t \right]
\\ \nonumber &\quad + {\mathbb E}_w\left[\Xi_t\tilde{\psi}(w_t)\int_0^t \nabla \tilde{V}(w_s)\rho_s {\textup d} s
+\Xi_t\nabla \tilde{\psi}(w_t)\rho_t \right]
\\ \nonumber &={\mathbb E}_w \left[{\mathcal D}^v \left(\Xi_t\tilde{\psi}(w_t)\right)\right]
+ {\mathbb E}_w\left[\Xi_t\tilde{\psi}(w_t)\int_0^t \nabla \tilde{V}(w_s)\rho_s {\textup d} s\right]
\\ &\quad +{\mathbb E}_w\left[\Xi_t\nabla \tilde{\psi}(w_t)\rho_t \right]
=I_1+I_2+I_3. \label{3.3}
\end{align}The term $I_1$ is treated using Malliavin integration by parts formula
(see Lemma~1.2.1 in \cite{nualart2006}):
\begin{equation}\label{3.4}
I_1 = {\mathbb E}_w \left[\Xi_t\tilde{\psi}(w_t) \int_0^t v(s){\textup d} W(s) \right],
\end{equation}where the stochastic integral $ \int_0^t v(s){\textup d} W(s) $ is in the Skorokhod sense.~The goal is to choose the process~$v$ in a such way that the terms~$I_i$, $i=1,2,3$ are bounded by the right-hand side of inequality~\eqref{3.1}.~We use the same choice of~$v$ as in the papers~\cite{HM-2006, HM-2011}.~More precisely, for any integer $n\ge0$, the~restriction~$ v_{n,n+1}$ of the process $v$ to the time interval $[n,n+1]$ is defined by
\begin{equation}
\label{3.5}
v_{n,n+1}(t)=\begin{cases} \big({\mathcal A}_{n,n+1/2}^*\left({\mathcal M}_{n,n+1/2}+\beta {\mathbb I}\right)^{-1}
J_{n,n+1/2}\rho_n\big)(t), \,\,\,\,\quad t \in [n,n+1/2], \\ 0,\qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \quad t\in [n+1/2,n+1], \end{cases}
\end{equation}where we set $\rho_0=\xi$ and $\beta>0$ is a small parameter. This choice allows to have an exponential decay for the moments of $\rho_t$ and of the Skorokhod integral as proved in the following lemmas. Inequality~\eqref{3.1} is proved by combining these lemmas (with an appropriate choice of parameters therein) and using a growth property of the Feynman--Kac semigroup.
The following two lemmas are
versions of Propositions~4.13 and~4.14 in~\cite{HM-2006}.
Since their formulations differ from the original ones, we give rather detailed~proofs based on the estimates recalled in Section \ref{S:2}.
\begin{lemma}
\label{L:3.3}
For any $\kappa>0$ and $\alpha>0$, there are constants $\beta=\beta(\kappa,\alpha)>0$ and~$C=C(\kappa,\alpha)>0$ such that
\begin{equation}
\label{3.6}
{\mathbb E}\|\rho_t\|^4 \le C \exp\left\{\kappa \|w\|^2-\alpha t \right\}\quad \text{for any $w\in \tilde H$ and $t\ge0$.}
\end{equation}
\end{lemma}
\begin{proof}
For integer times, this result is established in Proposition~4.13 in~\cite{HM-2006} (this is where Condition~\hyperlink{H}{\rm(H)} is used). Therefore, there are $\beta=\beta(\kappa,\alpha )>0$ and~$C=C(\kappa,\alpha )>0$ such that
\begin{equation}
\label{3.7}
{\mathbb E}\|\rho_n\|^4 \le C \exp\{\kappa \|w\|^2- \alpha n\} \quad \text{for any $w\in \tilde H$ and $n\ge0$}.
\end{equation}From the construction it follows that
$$
\rho_t=\begin{cases} J_{n,t}\rho_n-{\mathcal A}_{n,t}v_{n,t}, & \text{for }t \in [n,n+1/2], \\ J_{n+1/2,t}\rho_{n+1/2}, & \text{for }t\in [n+1/2,n+1] \end{cases}
$$ for any $n\ge0$.
Using \eqref{3.5} and inequalities \eqref{2.8} and \eqref{2.10}, we get
\begin{equation} \label{3.8}
\|v_{n,n+1/2}\|_{L^2([n,n+1/2];{\mathbb R}^d)} \le \beta^{-1/2}\|J_{n,n+1/2} \rho_n\|.
\end{equation}
Hence, for any $t\in [n,n+1/2]$,
\begin{align}
\|\rho_t\|&\le \|J_{n,t}\rho_n\|+\|{\cal A}_{n,t}v_{n,t}\| \nonumber
\\ \nonumber &\le \|J_{n,t}\rho_n\|+\|{\cal A}_{n,t}\|_{{\mathcal L}(L^2([n,t];{\mathbb R}^d),\tilde H)} \|v_{n,t}\|_{L^2([n,n+1/2];{\mathbb R}^d)}
\\ \nonumber &\le \|J_{n,t}\rho_n\|+\|{\cal A}_{n,t}\|_{{\mathcal L}(L^2([n,t];{\mathbb R}^d),\tilde H)} \|v_{n,n+1/2}\|_{L^2([n,n+1/2];{\mathbb R}^d)}
\\ \label{3.9}
&\le C\Big(\|J_{n,t}\rho_n\| +\beta^{-1/2} \|J_{n,n+1/2}\rho_n\| \sup_{s\in [n,t]}\|J_{s,t} \|_{{\mathcal L}( \tilde H,\tilde{H})} \Big),
\end{align}
where we used \eqref{2.7} and \eqref{3.8}. For any
$t\in [n+1/2,n+1],$ it holds that
$$
\|\rho_t\|\le \sup_{s\in [n+1/2,t]}\|J_{s,t}\rho_{n+1/2}\| .
$$
Combining this with inequalities \eqref{2.6}, \eqref{3.7},~\eqref{3.9}, the Cauchy--Schwarz inequality, and the fact that $\kappa>0$ and $\alpha>0$ are arbitrary, we arrive at~\eqref{3.6}.
\end{proof}
\begin{lemma}
\label{L:3.4}The constants $\beta>0$ and $C>0$ in Lemma \ref{L:3.3} can be chosen such that also
\begin{equation}
\label{3.10}
{\mathbb E} \left| \int_{n}^t v(s){\textup d} W(s)\right|^2 \le C\exp\{\kappa \|w\|^2-\alpha n \}
\end{equation}
for any $n\geq 0$, $t\in [n,n+1]$, and $w\in \tilde H$.
\end{lemma}
\begin{proof} In this proof, we consider the endpoint case $t=n+1$; the case $t\in [n,n+1)$ is treated in a similar way.
Using the generalised It$\hat{\text{o}}$ isometry (see Section~1.3 in~\cite{nualart2006}) and the fact that $v(t)=0$ for $t\in [n+1/2,n+1]$ (see \eqref{3.5}), we~obtain
\begin{align}
{\mathbb E} \left| \int_{n}^{n+1} v(s){\textup d} W(s)\right|^2
& = {\mathbb E} \int_{n}^{n+1/2} |v(s)|_{{\mathbb R}^d}^2 {\textup d} s \nonumber\\&\quad+{\mathbb E}\int_n^{n+1/2}\!\!\!\int_{n}^{n+1/2} \text{Tr}({\mathcal D}_sv(r){\mathcal D}_rv(s)){\textup d} s {\textup d} r\nonumber
\\ \nonumber & \le {\mathbb E} \int_{n}^{n+1/2} |v(s)|_{{\mathbb R}^d}^2 {\textup d} s \nonumber\\&\quad+{\mathbb E} \int_{n}^{n+1/2}\!\!\!\int_{n}^{n+1/2} |{\mathcal D}_rv_{n,n+1/2}(s)|_{{\mathbb R}^d\times {\mathbb R}^d}^2{\textup d} s{\textup d} r\nonumber\\ & =L_1+L_2.
\label{3.11}
\end{align}We estimate $L_1$ by using \eqref{2.6},
\eqref{3.7}, and \eqref{3.8}:
\begin{align}
\nonumber {\mathbb E} \int_{n}^{n+1/2} |v(s)|_{{\mathbb R}^d}^2 {\textup d} s &\le \beta^{-1}{\mathbb E} \|J_{n,n+1/2}\rho_{n}\|^2 \\
&\le \nonumber
C \beta^{-1} \exp\{\kappa \|w\|^2/2 \} \left({\mathbb E} \|\rho_n\|^4\right)^{1/2}
\\
\label{3.12}
&\le C\exp\{\kappa \|w\|^2-\alpha n/2 \}.
\end{align}
To estimate $L_2$, we use the explicit form of ${\mathcal D}_r v.$
Notice that, for any $r\in[n,n+1/2]$ and~$i=1,\dots, d$,
\begin{align*}
{\mathcal D}_r^iv_{n,n+1/2}
&= {\mathcal D}_r^i({\mathcal A}_{n,n+1/2}^*) ({\mathcal M}_{n,n+1/2}+\beta {\mathbb I})^{-1}J_{n,n+1/2}\rho_{n}
\\ &\quad + {\mathcal A}_{n,n+1/2}^*({\mathcal M}_{n,n+1/2}+\beta{\mathbb I})^{-1}
\\ & \quad\quad\times \Big( {\mathcal D}_r^i({\mathcal A}_{n,n+1/2}){\mathcal A}_{n,n+1/2}^*
+{\mathcal A}_{n,n+1/2}{\mathcal D}_r^i({\mathcal A}_{n,n+1/2}^*)\Big) \\ & \quad\quad\times
({\mathcal M}_{n,n+1/2}+\beta{\mathbb I})^{-1}
J_{n,n+1/2} \rho_{n}\\ & \quad +{\mathcal A}_{n,n+1/2}^* ({\mathcal M}_{n,n+1/2}+\beta{\mathbb I})^{-1}{\mathcal D}_r^i(J_{n,n+1/2}) \rho_{n}.
\end{align*}
By inequalities \eqref{2.8}-\eqref{2.10}, we have
\begin{align*}
\|{\mathcal D}_r^iv_{n,n+1/2}\|_{L^2([n,n+1/2];{\mathbb R}^d)}
& \le \beta^{-1}\| {\mathcal D}_r^i({\mathcal A}_{n,n+1/2})\|_{{\mathcal L}(L^2([n,n+1/2];{\mathbb R}^d),\tilde{H})} \\&\quad\quad\times
\|J_{n,n+1/2} \rho_{n}\|
\\ &\quad +2\beta^{-1}\| {\mathcal D}_r^i({\mathcal A}_{n,n+1/2}^*)\|_{{\mathcal L}(\tilde{H},L^2([n,n+1/2];{\mathbb R}^d))} \\&\quad\quad\times
\|J_{n,n+1/2} \rho_{n}\| \\ &\quad
+ \beta^{-1/2} \| {\mathcal D}_r^i(J_{n,n+1/2}) \rho_{n}\|.
\end{align*}
Inequalities \eqref{2.6}, \eqref{2.11}-\eqref{2.13}, and \eqref{3.7}, imply that
\begin{align}
{\mathbb E} \int_{n}^{n+1/2}\!\!\!\int_{n}^{n+1/2} |{\mathcal D}_rv_{n,n+1/2}(s)|_{{\mathbb R}^d\times {\mathbb R}^d}^2{\textup d} s{\textup d} r
& \le C \beta^{-2} \exp\{\kappa \|w\|^2/2 \} \left({\mathbb E} \|\rho_n\|^4\right)^{1/2}
\nonumber \\ \label{3.13} &\le C \exp\{\kappa \|w\|^2-n\alpha/2 \}.
\end{align}
Combining estimates \eqref{3.11}-\eqref{3.13} and using the fact that $\kappa>0$ and $\alpha>0$ are~arbitrary,
we obtain the desired result.
\end{proof}
Finally, we will use a growth estimate for the Feynman--Kac semigroup $ \tilde {\mathfrak P}^{\tilde V}_t$. From the first growth esimate in~\eqref{1.3} for the semigroup $ {\mathfrak P}^V_t$ it follows that there are numbers
$R_0>0$, $\gamma\in (0,\gamma_0)$, and $m\ge1$
such that
\begin{equation}\label{3.14}
\tilde {\mathfrak P}^{\tilde V}_t {\mathbf 1} (w)\le C\, {\mathfrak w}_m({\cal K} w) \|{\mathfrak P}^V_t{{\mathbf 1}}\|_{R_0}\quad \text{for any $w\in \tilde H$ and $t\ge0$}.
\end{equation}
Now we are in a position to prove Proposition \ref{P:3.2}.
\begin{proof}[Proof of Proposition \ref{P:3.2}]
Replacing $V$ by $V-\inf_{v\in H} V(v),$ without loss of generality, we can assume that $V\ge0$. Let $v$ be the process defined by \eqref{3.5}, let $\kappa$ and~$\alpha $ be positive numbers (to be chosen later), and let the number $\beta=\beta(\kappa,\alpha )>0$ be such that inequalities~\eqref{3.6} and \eqref{3.10} hold. Furthermore, let the positive numbers $R_0$ and $m$ be such that inequality \eqref{3.14} holds.
Then the computations in \eqref{3.3} are rigorously justified, and we need to estimate the terms $I_1$, $I_2$, and~$I_3$.
{\it Step~1: Estimate for $I_1$}. We write the Skorokhod integral in the term $I_1$ (see~\eqref{3.4})
as follows
$$
\int_0^t v(s){\textup d} W(s) =\sum_{n=1}^{\lfloor t \rfloor}
\int_{n-1}^{n} v(s){\textup d} W(s) + \int_{\lfloor t \rfloor}^{t} v(s){\textup d} W(s),
$$
where $\lfloor t \rfloor $ is the largest number less than or equal to $t$ and the sum in the right-hand side is replaced by zero if $t<1$.
Since $v(s)$ is ${\mathcal F}_{n}$-measurable for~$s\in [n-1,n]$, the Skorokhod integral $ \int_{n-1}^{n} v(s){\textup d} W(s)$ is also ${\mathcal F}_{n}$-measurable. Hence, using the Markov property, we obtain
\begin{align*}
\nonumber I_{1,n} =&\,\, {\mathbb E}_w \left[\Xi_t\tilde{\psi}(w_t) \int_{n-1}^{n} v(s){\textup d} W(s) \right]
\\ \nonumber =& \,\,
{\mathbb E}_w \left[{\mathbb E}_w \left(\Xi_t\tilde{\psi}(w_t) \int_{n-1}^{n} v(s){\textup d} W(s)\Big|{\mathcal F}_{n} \right)\right]
\\ \nonumber =&\,\,
{\mathbb E}_w \left[\Xi_n \int_{n-1}^{n} v(s){\textup d} W(s) \,{\mathbb E}_w \left( \exp\left\{\int_{n}^t\tilde{V}(w_s){\textup d} s \right\}\tilde{\psi}(w_t)\Big|{\mathcal F}_{n}\right) \right]\\ =& \, \,
{\mathbb E}_w \left[\Xi_n \int_{n-1}^{n} v(s){\textup d} W(s) \, \left(\tilde {\mathfrak P}^{\tilde V}_{t-n} \tilde \psi\right) (w_n) \right]
\end{align*}
for any $1\le n\le \lfloor t \rfloor$.~Using inequalities \eqref{1.2}, \eqref{3.10}, \eqref{3.14}, the assumption that~$V\ge0$, and the Cauchy--Schwarz inequality, we see that
\begin{align*}
\nonumber I_{1,n} & \le C \|\psi\|_\infty e^{\|V\|_\infty n}\|{\mathfrak P}_t^V{\mathbf 1}\|_{R_0} {\mathbb E}_w\left[{\mathfrak w}_{m}(u_{n}) \Big| \int_{n-1}^{n} v(s){\textup d} W(s)\Big|\right]
\\ \nonumber & \le C\|\psi\|_\infty e^{\|V\|_\infty n}\|{\mathfrak P}_t^V{\mathbf 1}\|_{R_0} \left({\mathbb E}_u{\mathfrak w}_{m}^2(u_{n})\right)^{1/2} \left( {\mathbb E}_w \Big| \int_{n-1}^{n} v(s){\textup d} W(s)\Big|^2\right)^{1/2}
\\ & \le C \|\psi\|_\infty e^{\|V\|_\infty n}\|{\mathfrak P}_t^V{\mathbf 1}\|_{R_0} {\mathfrak w}_{m}(u) \exp\{(\kappa \|w\|^2 -\alpha n)/2\},
\end{align*}
where $u={\cal K} w$ and $u_s={\cal K} w_s$. Next, using \eqref{3.10} and $V\ge 0$, we get
\begin{align*}
\nonumber I_{1,\lfloor t \rfloor+1}&={\mathbb E}_w \left[\Xi_t\tilde{\psi}(w_t) \int_{\lfloor t \rfloor }^{t} v(s){\textup d} W(s) \right]
\\ &\le \|\psi\|_\infty
e^{\|V\|_\infty t }{\mathbb E}_w \left|\int_{\lfloor t \rfloor }^{t} v(s){\textup d} W(s)\right|\nonumber\\\ & \le C\|\psi\|_\infty e^{\|V\|_\infty t} \|{\mathfrak P}_t^V{\mathbf 1}\|_{R_0} \exp\{(\kappa \|w\|^2-\alpha \lfloor t \rfloor )/2 \}.
\end{align*} Combining the estimates for $I_{1,n}$ and $I_{1,\lfloor t \rfloor+1}$, we arrive at
\begin{align*}
I_1\le C \|\psi\|_\infty \|{\mathfrak P}_t^V{\mathbf 1}\|_{R_0} \exp\{\kappa \|w\|^2\} \sum_{n=1}^{^{\lfloor t \rfloor}} \exp\{(\|V\|_\infty -\alpha /2)n\}.
\end{align*}
{\it Step~2: Estimate for $I_2$}. We first write
\begin{align*}
I_2 &= {\mathbb E}_w\left[\int_0^{\lfloor t \rfloor } \Xi_t\tilde{\psi}(w_t)\nabla \tilde{V}(w_s)\rho_s {\textup d} s\right] +{\mathbb E}_w\left[\int_{\lfloor t \rfloor }^t \Xi_t\tilde{\psi}(w_t)\nabla \tilde{V}(w_s)\rho_s {\textup d} s\right]
\\&= I_{2,1}+I_{2,2}.
\end{align*}Let $\lceil s \rceil$ be the smallest integer greater than or equal to $s$. Then $\rho(s)$ is ${\mathcal F}_{\lceil s\rceil}$-measurable, and using the Markov property, we obtain
\begin{align*}
I_{2,1} &= {\mathbb E}_w\left[ \int_0^{\lfloor t\rfloor} {\mathbb E}_w \left( \Xi_t\tilde{\psi}(w_t) \nabla \tilde{V}(w_s)\rho_s \big| {\mathcal F}_{ \lceil s \rceil } \right) {\textup d} s \right]
\\ &= {\mathbb E}_w \left[ \int_0^{\lfloor t \rfloor }
\Xi_{\lceil s \rceil} \nabla \tilde{V}(w_s)\rho_s
\
{\mathbb E}_w \left(\exp\left\{\int_{\lceil s \rceil}^t\tilde{V}(w_r){\textup d} r\right\}\tilde{\psi}(w_t) \Big| {\mathcal F}_{ \lceil s \rceil } \right) {\textup d} s \right]
\\ &= {\mathbb E}_w \left[ \int_0^{\lfloor t \rfloor }
\Xi_{\lceil s \rceil} \nabla \tilde{V}(w_s)\rho_s
\left(
\tilde {\mathfrak P}^{\tilde V}_{t-\lceil s \rceil} \tilde \psi\right) (w_{\lceil s \rceil}) {\textup d} s \right].
\end{align*}
Then inequalities \eqref{1.2}, \eqref{3.6}, \eqref{3.14}, the assumption that~$V\ge0$, and the Cauchy--Schwarz inequality imply that
\begin{align*}
I_{2,1} & \le
C \|\psi\|_\infty\|\nabla V\|_\infty \|{\mathfrak P}_t^V\textbf{1}\|_{R_0}
\int_0^{\lfloor t\rfloor} e^{\|V\|_\infty \lceil s \rceil } {\mathbb E}_w \left[{\mathfrak w}_m(u_{ \lceil s \rceil }) \|\rho_s\| \right] {\textup d} s
\\ &\le C
\|\psi\|_\infty\|\nabla V\|_\infty \|{\mathfrak P}_t^V\textbf{1}\|_{R_0}
\int_0^{\lfloor t\rfloor} e^{\|V\|_\infty \lceil s \rceil } \left({\mathbb E}_u {\mathfrak w}_m^2(u_{\lceil s \rceil })\right)^{1/2} \left({\mathbb E}_w \|\rho_s\|^2\right)^{1/2} {\textup d} s
\\
& \le
C\|\psi\|_\infty\|\nabla V\|_\infty \|{\mathfrak P}_t^V\textbf{1}\|_{R_0} {\mathfrak w}_{m}(u) \exp\{\kappa \|w\|^2/4\}\\&\quad\quad\quad\quad\times
\int_0^{\lfloor t\rfloor} \exp\{\|V\|_\infty \lceil s \rceil -\alpha s/4 \} {\textup d} s.
\end{align*}
To estimate $I_{2,2}$, we use \eqref{3.6} and $V\ge0$:
\begin{align*}
I_{2,2}&\le \|\psi\|_\infty e^{\|V\|_\infty t }\|\nabla V\|_\infty {\mathbb E}_w \int_{\lfloor t \rfloor }^t \|\rho_s\| {\textup d} s \\&\le
C \|\psi\|_\infty e^{\|V\|_\infty t }\|\nabla V\|_\infty\|{\mathfrak P}_t^V{\mathbf 1}\|_{R_0}
\exp\{(\kappa \|w\|^2-\alpha t)/4\}.
\end{align*}Thus
\begin{align*}
I_2 &\le C\|\psi\|_\infty\|\nabla V\|_\infty \|{\mathfrak P}_t^V\textbf{1}\|_{R_0} \exp\{\kappa \|w\|^2\} \\&\quad\quad\quad\quad\times\left(\exp\{-\alpha t)/4\}+
\int_0^{\lfloor t\rfloor} \exp\{\|V\|_\infty \lceil s \rceil -\alpha s/4 \} {\textup d} s\right).
\end{align*}
{\it Step~3: Estimate for $I_3$}.
By \eqref{3.6}, we have
$$
|I_3|\le C \|\nabla \psi\|_\infty e^{\|V\|_\infty t}{\mathbb E} \|\rho_t\|
\le \|\nabla \psi\|_\infty e^{\|V\|_\infty t} \exp\{(\kappa \|w\|^2-t\alpha )/4\}.
$$ Choosing
$
\alpha \geq 4 \|V\|_\infty-\log a
$ and combining the above estimates of the terms $I_i$, $i=1,2,3$ with \eqref{3.3}, we complete the proof of
Proposition~\ref{P:3.2}.
\end{proof}
\def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
|
\section*{Introduction}\label{sec:fundrepres}
The present paper is devoted to the solution of the following general problem.
\begin{generalproblem}
Let $R$ be an arbitrary commutative associative ring with $1$ and let $\Phi$ be a reduced irreducible root system. $G(\Phi, \blank)$ is a Chevalley--Demazure group scheme and $\rho\colon G(\Phi, \blank) \longrightarrow \GL_{N}(\blank)$ is its arbitrary representation. Describe all overgroups $H$ of the elementary subgroup $\E_{G}(\Phi, R)$ in the representation $\rho$:
$$\E_{G, \rho}(\Phi, R ) \leq H \leq \GL_{N}(R).$$
\end{generalproblem}
The conjectural answer, the \textit{standard overgroup description}, in a general case can be formulated as follows.
For any overgroup $H$ of the elementary group there exists a net of ideals $\mathbb{A}$ of the ring $R$ such that
$$\E_{G,\rho}(\Phi,R)\cdot\E_N(R,\mathbb{A})\leq H \leq N_{\GL_{N}(R)}\big(\E_{G,\rho}(\Phi, R)\cdot\E_N(R,\mathbb{A})\big),$$
where $\E_N(R,\mathbb{A})$ is a \textit{relative elementary subgroup} for the net $\mathbb{A}$.
In the special case of a trivial net, i.\,e., $\mathbb{A} = \{A\}$ consists of one ideal $A$ of the initial ring $R$, overgroups of the group $\E_{G}(\Phi,R)$ can be parametrized by the ideal $A$ of the ring $R$:
\begin{equation}
\E_{G,\rho}(\Phi,R)\cdot\E_N(R,A)\leq H \leq N_{\GL_{N}(R)}\big(\E_{G,\rho}(\Phi, R)\cdot\E_N(R,A)\big),
\label{eq:G_A}
\end{equation}
where $\E_N(R,A)$ equals $\E_N( A)^{\E_N(R)}$ by definition.
\medskip
Based on the classification of finite simple groups in 1984, Michael Aschbacher proved the Subgroup structure theorem~\cite{AschbacherClasses}. It states that every maximal subgroup of a finite classical group either falls into one of the eight explicitly described classes $\mathcal{C}_1$--$\mathcal{C}_8$, or is an ``almost'' simple group in an irreducible representation (class $\mathcal{S}$). In the recent past, many experts studied overgroups of groups from the Aschbacher classes for some special cases of fields. For finite fields and algebraically closed fields maximality of subgroups was obtained by Peter Kleidman and Martin Liebeck, see~\cite{KleidLiebSubStruct,LiebSeitStrucClass}. Oliver King, Roger Dye, and Shang Zhi Li proved maximality of groups from Aschbacher classes for arbitrary fields or described its overgroups in cases where they are not maximal, see~\cite{KingDiag,KingUnit,KingOrthog,DyeSpO,DyeOinSp,DyeSubPolarities,ShangZhiGL,ShangZhiUnit,ShangZhiSUOmega}. We recommend the surveys~\cite{VavSubgroup90,VavSbgs,VavStepSurvey}, which contain necessary preliminaries, complete history, and known results about the initial problem.
In the present paper, we consider the case of the $m$-th fundamental representation of a simply connected group of type $A_{n-1}$, i.\,e., the scheme $G_{\rho}(\Phi,\blank)$ equals a [Zariski] closure of the affine group scheme $\SL_{n}(\blank)$ in the representation with the highest weight $\varpi_{m}$. In our case the extended Chevalley group scheme coincides with the $m$-th fundamental representation of the general linear group scheme $\GL_{n}(\blank)$. This case corresponds to the Aschbacher class ${\mathcal{S}}$ consisting of almost simple groups in certain absolutely irreducible representations. Morally, the paper is a continuation of a series of papers by the St.~Petersburg school on subgroups in classical groups over a commutative ring, see~\cite{BV82,BV84,VavSubgroup90,VP-EOeven,VP-EOodd,VP-Ep,PetOvergr,StepGL,StepNonstandard,LuzE6E7-GL,LuzF4-E6,VavLuzgE6,VavLuzgE7,AnaVavSinI}.
We deal only with the case of a trivial net, i.\,e., a net consists of only one ideal $A$. As shown below (Propositions~\ref{prop:InversInclus} and~\ref{prop:LevelsEqualGeneral}), it imposes a constraint $n\geq 3m$, we proceed with this restriction. In this case the general answer has the following form. Let $N = \binom{n}{m}$ and $H$ be a subgroup in $\GL_{N}(R)$ containing $\bw{m}\E_n(R)$. Then there exists a unique maximal ideal $A\trianglelefteq R$ such that
$$\bw{m}\E_n(R)\cdot\E_N(R,A)\leq H \leq N_{\GL_{N}(R)}\big(\bw{m}\E_n(R)\cdot\E_N(R,A)\big).\eqno{(*)}$$
The present paper is the first part in the serial study of the problem. We construct a level and calculate a normalizer of connected (i.\,e., perfect) intermediate subgroups. Further, it is necessary to construct invariant forms for $\bw{m}\SL_n(R)$ and calculate a normalizer of $\bw{m}\E_n(R)$. Finally, we will extract an elementary transvection from an intermediate subgroup $H$. These steps are enough to solve the problem completely, see~\cite{LubStepSub}.
There are separate results for special cases of the ring $R$ such as a finite field $K$ or an algebraically closed field. For finite fields Bruce Cooperstein proved maximality of the normalizer $N_G\bigl(\bw{2}\E_n(K)\bigr)$ in $\GL_N(K)$~\cite{CoopersteinStructure}. For algebraically closed fields a description of overgroups of $\bw{m}\E_n(K)$ follows from the classical results about maximal subgroups of classical algebraic groups, for instance, see~\cite{SeitzMaxSub}.
\medskip
The present paper is organized as follows. In the next Section we formulate main results of the paper. In Section~\ref{sec:princnot} we set up the notation. Section~\ref{sec:proofs} contains all complete proofs, for instance, in~\S\ref{subsec:square} we have considered an important special case --- a level computation for exterior squares of elementary groups. In~\S\ref{subsec:general}--\S\ref{subsec:levelcomputation} we develop a technique for an arbitrary general exterior power. Finally, a level reduction for exterior powers is proved in~\S\ref{subsec:normgeneral}.
\textbf{Acknowledgment.} We would like to express our sincere gratitude to our scientific adviser Nikolai Vavilov for formulating the problem and for a constant support, without which this paper would never have been written. The authors are grateful to Alexei Stepanov for carefully reading our original manuscript and for numerous remarks and corrections. Also, we would like to thank an anonymous referee for bringing our attention to the paper \cite{GariGura2015}.
\section{Main results}\label{sec:main_results}
Fundamental representations of the general linear group $\GL_{n}$, as well as of the special linear group $\SL_{n}$, are the ones with the highest weights $\varpi_{m} = \underbrace{(1, \dots, 1)}_{m}$ for $m = 1, \dots, n$. The representation with the highest weight $\varpi_{n}$ degenerates for the group $\SL_{n}$. The explicit description of these representations uses exterior powers of the standard representation.
In detail, for a commutative ring $R$ by $\bw{m} R^{n}$ we denote an $m$-th exterior power of the free module $R^n$. We consider the following natural transformation, an \textit{exterior power},
$$\bw{m}: \GL_{n} \rightarrow \GL_{\binom{n}{m}}$$
which extends the action of the group $\GL_{n}(R)$ from $R^n$ to $\bw{m} R^{n}$.
An elementary group $\E_n( R)$ is a subgroup of the group of points $\GL_{n}(R)$, so its exterior power $\bw{m}\E_n(R)$ is a well defined subgroup of the group of points $\bw{m}\GL_{n} (R)$. A more user--friendly description of the elementary group $\bw{m}\E_n(R)$ will be presented in Subsections~\ref{subsec:square} and~\ref{subsec:general}.
Let $H$ be an arbitrary overgroup of an elementary group $\bw{m}\E_n(R)$:
$$\bw{m}\E_n(R) \leq H \leq \GL_{\binom{n}{m}}(R).$$
For any unequal weights $I, J \in \bw{m}[n]$, which are indices for matrix entries of elements from $\GL_{\binom{n}{m}}(R)$, by $A_{I,J}$ we denote the following set
$$A_{I,J}:=\{\xi \in R \;|\; t_{I,J}(\xi) \in H \} \subseteq R.$$
It turns out these sets are ideals that coincide for any pair of unequal weights $I\neq J$.
\begin{proposition}\label{prop:InversInclus}
Sets $A_{I,J}$ coincide for $n \geq 3m$.
\end{proposition}
In the case $\frac{n}{3} \leq m \leq n$ description of overgroups cannot be done by a parametrization only by a single ideal. Moreover, as it could be seen from further calculations we need up to $m$ ideals in some cases for a complete parametrization of overgroups. There are a lot of nontrivial relationships between the ideals. So even the notion of a relative elementary group is far more complicated and depends on a Chevalley group (for instance, see~\cite{AnaVavSinI}), let alone formulations of the Main Theorems. The authors work in this direction and hope this problem would be solved in the near future. In a general case this [partially ordered] set of ideals forms \textit{a net of ideals} (due to Zenon Borevich; for a definition see \cite{BVnets}, for further progress in the direction of subgroup classification see \cite{BVdet, Borevich1984}).
Back to the case $n \geq 3m$, the set $A:= A_{I,J}$ is called a \textit{level of an overgroup} $H$. The description of overgroups goes as follows.
\begin{restatable}[Level computation]{theorem}{LevelForm}
\label{thm:LevelForm}
Let $R$ be a commutative ring and $n$, $m$ be natural numbers with the constraint $n\geq 3m$. For an arbitrary overgroup $H$ of the group $\bw{m}\E_n(R)$ there exists a unique maximal ideal $A$ of the ring $R$ such that
$$\bw{m}\E_n(R)\cdot\E_N(R,A)\leq H.$$
Namely, if a transvection $t_{I,J}(\xi)$ belongs to the group $H$, then $\xi\in A$.
\end{restatable}
The left--hand side subgroup is denoted by $\E\bw{m}\E_n(R,A)$. We note that this group is perfect (Lemma~\ref{lem:PerfectForm}). Motivated by the expected relations $(*)$, we present an alternative description of the normalizer $N_{\GL_{N}(R)}\bigl(\E\bw{m}\E_n(R,A)\bigr)$.
For this we introduce the canonical projection $\rho_A: R\longrightarrow R/A$ mapping $\lambda\in R$ to $\bar{\lambda}=\lambda+A\in R/A$. Applying the projection to all entries of a matrix, we get the reduction homomorphism
$$\begin{array}{rcl}
\rho_{A}:\GL_{n}(R)&\longrightarrow& \GL_{n}(R/A)\\
a &\mapsto& \overline{a}=(\overline{a}_{i,j})
\end{array}$$
Eventually, we have the following explicit \textit{congruence} description.
\begin{restatable}[Level reduction]{theorem}{LevelReduction}
\label{thm:LevelReduction}
Let $n\geq 3m$.
For any ideal $A\trianglelefteq R$, we have
$$N_{\GL_{N}(R)}\bigl(\E\bw{m}\E_n(R,A)\bigr)=\rho_{A}^{-1}\left(\bw{m}\GL_{n}(R/A)\right).$$
\end{restatable}
\section{Principal notation}\label{sec:princnot}
Our notation for the most part is fairly standard in Chevalley group theory. We recall all necessary notion below for the purpose of self-containment.
First, let $G$ be a group. By a commutator of two elements we always mean \textit{the left-normed} commutator $[x,y]=xyx^{-1}y^{-1}$, where $x,y\in G$. Multiple commutators are also left-normed; in particular, $[x,y,z]=[[x,y],z]$. By ${}^xy=xyx^{-1}$ we denote \textit{the left conjugates} of $y$ by $x$. Similarly, by $y^x=x^{-1}yx$ we denote \textit{the right conjugates} of $y$ by $x$. In the sequel, we will use the Hall–Witt identity:
$$[x,y^{-1},z^{-1}]^x\cdot[z,x^{-1},y^{-1}]^z\cdot[y,z^{-1},x^{-1}]^y=e.$$
For a subset $X\subseteq G$, we denote by $\langle X\rangle$ a subgroup it generates. The notation $H\leq G$ means that $H$ is a subgroup in $G$, while the notation $H\trianglelefteq G$ means that $H$ is a normal subgroup in $G$. For $H\leq G$, we denote by $\langle X\rangle^H$ the smallest subgroup in $G$ containing $X$ and normalized by $H$. For two groups $F,H\leq G$, we denote by $[F,H]$ their mutual commutator: $[F,H]=\langle [f,g] \text{ for } f\in F, h\in H\rangle.$
Also we need some elementary ring theory notation. Let $R$ be an associative ring with $1$. By default, it is assumed to be commutative. By an ideal $I$ of a ring $R$ we understand \textit{the two-sided ideal} and this is denoted by $I\trianglelefteq R$. As usual, $R^{*}$ denotes a multiplicative group of a ring $R$. A multiplicative group of matrices over a ring $R$ is called a general linear group and is denoted by $\GL_n(R)=\M_n(R)^{*}$. A special linear group $\SL_{n}(R)$ is a subgroup of $\GL_n(R)$ consisting of matrices of determinant $1$. By $a_{i,j}$ we denote an entry of a matrix $a$ at the position $(i,j)$, where $1\leq i,j\leq n$. Further, $e$ denotes the identity matrix and $e_{i,j}$ denotes the standard matrix unit, i.\,e., the matrix that has $1$ at the position $(i,j)$ and zeros elsewhere. For entries of the inverse matrix we will use the standard notation $a_{i,j}':=(a^{-1})_{i,j}$.
By $t_{i,j}(\xi)$ we denote an elementary transvection, i.\,e., a matrix of the form $t_{i,j}(\xi)=e+\xi e_{i,j}$, $1\leq i\neq j\leq n$, $\xi\in R$. Hereinafter, we use (without any references) standard relations~\cite{StepVavDecomp} among elementary transvections such as
\begin{enumerate}
\item additivity:
$$t_{i,j}(\xi)t_{i,j}(\zeta)=t_{i,j}(\xi+\zeta).$$
\item the Chevalley commutator formula:
$$[t_{i,j}(\xi),t_{h,k}(\zeta)]=
\begin{cases}
e,& \text{ if } j\neq h, i\neq k,\\
t_{i,k}(\xi\zeta),& \text{ if } j=h, i\neq k,\\
t_{h,j}(-\zeta\xi),& \text{ if } j\neq h, i=k.
\end{cases}$$
\end{enumerate}
A subgroup $\E_n(R)\leq \GL_n(R)$ generated by all elementary transvections is called an \textit{$($absolute$)$ elementary group}:
$$\E_n(R)=\langle t_{i,j}(\xi), 1\leq i\neq j\leq n, \xi\in R\rangle.$$
Now define a normal subgroup of $\E_n(R)$, which plays a crucial role in calculating the level of intermediate subgroups. Let $I$ be an ideal in $R$. Consider a subgroup $\E_n(R,I)$ generated by all elementary transvections of level $I$, i.\,e., $\E_n(R,I)$ is a normal closure of $\E_n(I)$ in $\E_n(R)$. This group is called an \textit{$($relative$)$ elementary group of level $I$}:
$$\E_n(R,I)=\langle t_{i,j}(\xi), 1\leq i\neq j\leq n, \xi\in I\rangle^{\E_n(R)}.$$
It is well known (due to Andrei Suslin~\cite{SuslinSerreConj}) that the elementary group is normal in the general linear group $\GL_{n}(R)$ for $n \geq 3$. The normality is crucial for further considerations, so hereafter we suppose that $n \geq 3$. Furthermore, the relative elementary group $\E_n(R,I)$ is normal in $\GL_{n}(R)$ if $n\geq 3$. This fact, first proved in \cite{SuslinSerreConj}, is cited as Suslin’s theorem. Moreover, if $n\geq 3$, then the group $\E_n(R,I)$ is generated by transvections of the form $z_{i,j}(\xi,\zeta)=t_{j,i}(\zeta)t_{i,j}(\xi)t_{j,i}(-\zeta)$, $1\leq i\neq j\leq n$, $\xi\in I$, $\zeta\in R$. This fact was proved by Leonid~Vaserstein and Andrey~Suslin~\cite{VasSusSerre} and, in the context of Chevalley groups, by Jacques~Tits~\cite{TitsCongruence}.
By $[n]$ we denote the set $\{1,2,\dots, n\}$ and by $\bw{m}[n]$ we denote an exterior power of the set $[n]$. Elements of $\bw{m}[n]$ are ordered subsets $I\subseteq [n]$ of cardinality $m$ without repeating entries:
$$\bw{m}[n] = \{ (i_{1}, i_{2}, \dots , i_{m})\; |\; i_{j} \in [n], i_{j} \neq i_{l} \}.$$
We use the lexicographic order on $\bw{m}[n]$ by default: $12\dots (m-1)m < 12\dots (m-1)(m+1) < \dots$
Usually, we write an index $I=\{i_j\}_{j=1}^m$ in the ascending order, $i_1<i_2<\dots<i_m$. Sign $\sign(I)$ of the index $I = (i_1, \dots, i_m)$ equals the sign of the permutation mapping $(i_1, \dots, i_m)$ to the same set in the ascending order. For example, $\sign(1234) = \sign(1342) = +1$, but $\sign(1324) = \sign(4123) = -1$.
Finally, let $n\geq 3$ and $m\leq n$. \textit{By $N$ we denote the binomial coefficient $\binom{n}{m}$}. In the sequel, we denote an elementary transvection in $\E_N(R)$ by $t_{I,J}(\xi)$ for $I,J \in \bw{m}[n]$ and $\xi \in R$. For instance, the transvection $t_{12,13}(\xi)$ equals the matrix with $1$'s on the diagonal and $\xi$ in the position $(12,13)$.
\section{Proofs \& Computations}\label{sec:proofs}
We consider a case of an exterior square of a group scheme $\GL_{n}$ at first. We have two reasons for this way of presentation. Firstly, proofs of statements in a general case belong to the type of technically overloaded statements. At the same time, simpler proofs in the basic case present all ideas necessary for a general case. In particular, for $n=4$ Nikolai Vavilov and Victor Petrov completed the standard description of overgroups\footnote{The restriction of the exterior square map $\bw{2}: \GL_{4}(R) \longrightarrow \GL_{6}(R)$ to the group $\E_4(R)$ is an isomorphism onto the elementary orthogonal group $\EO_6(R)$~\cite{VP-EOeven}.}. Secondly, for exterior squares there are several important results that cannot be obtained for the exterior cube or other powers, see~\cite{LubNek18,LubReverse2,LubReversemArx}. For example, in~\cite{LubReverse2} the author construct a transvection $T\in \bw{2}\E_n(R)$ such that it stabilizes an arbitrary column of a matrix $g$ in $\GL_{\binom{n}{2}}(R)$. And there are no such transvections for other exterior powers.
\subsection{Exterior square of elementary groups}\label{subsec:square}
Let $R$ be a commutative ring with $1$, $n$ be a natural number greater than $3$, and $R^n$ be a right free $R$-module with the standard basis $\{e_1,\dots,e_n\}$. By $\bw{2}R^n$ we denote a universal object in the category of alternating bilinear maps from $R^n$ to $R$-modules. Concretely, take a free module of rank $N=\binom{n}{2}$ with the basis $e_i\wedge e_j$, $1~\leq~i~\neq~j~\leq~n$. The elements $e_i\wedge e_j$ for arbitrary $1\leq i,j\leq n$ are defined by the relation $e_{i} \wedge e_{j} = -e_{j} \wedge e_{i}$.
An action of the group $\GL_{n}(R)$ on the module $\bw{2}R^n$ is diagonal:
$$\bw{2}(g)(e_i\wedge e_j):=(ge_i)\wedge(ge_j) \text{ for any } g\in \GL_{n}(R) \text{ and } 1\leq i\neq j\leq n.$$
In the basis $\{e_I, I\in\bw{2}[n]\}$ of the module $\bw{2}R^n$ a matrix $\bw{2}(g)$ consists of second order minors of the matrix $g$ with lexicographically ordered columns and rows:
$$\left(\bw{2}(g)\right)_{I,J}=\left(\bw{2}(g)\right)_{(i_1,i_2),(j_1,j_2)}=M_{i_1,i_2}^{j_1,j_2}(x) = g_{i_1, j_1}\cdot g_{i_2, j_2} - g_{i_1, j_2}\cdot g_{i_2, j_1}.$$
By the Cauchy--Binet theorem the map $\pi\colon\GL_{n}(R)\longrightarrow \GL_{N}(R)$, $x\mapsto \bw{2}(x)$ is a homomorphism. Thus the map $\pi$ is a representation of the group $\GL_n(R)$, called the \textit{bivector representation} or the \textit{second fundamental representation} (\textit{the representation with the highest weight} $\varpi_2$). The image of the latter action is called the exterior square of the group $\GL_n(R)$. $\E_n(R)$ is a subgroup of $\GL_n(R)$, therefore the exterior square of the elementary group is well defined. The following lemma is a corollary of Suslin's theorem.
\begin{lemma}\label{SuslinFor2}
The image of an elementary group is normal in the image of a general linear group under the exterior square homomorphism:
$$\bw{2}\bigl(\E_n(R)\bigr)\trianglelefteq \bw{2}\bigl(\GL_{n}(R)\bigr).$$
\end{lemma}
Note that $\bw{2}\bigl(\GL_{n}(R)\bigr)$ does not equal $\bw{2}\GL_n(R)$ for arbitrary rings. For detail see the extended description in \S\ref{subsec:general}.
Let us consider a structure of the group $\bw{2}\E_n(R)$ in detail. The following proposition can be extracted from the very definition of $\bw{2}\big(\GL_{n}(R)\big)$.
\begin{proposition}\label{prop:ImageOfTransvFor2}
Let $t_{i,j}(\xi)$ be an elementary transvection. For $n\geq 3$, $\bw{2}t_{i,j}(\xi)$ can be presented as the following product:
\begin{equation}
\bw{2}t_{i,j}(\xi)=\prod\limits_{k=1}^{i-1} t_{ki,kj}(\xi)\,\cdot\prod\limits_{l=i+1}^{j-1}t_{il,lj}(-\xi)\,\cdot\prod\limits_{m=j+1}^n t_{im,jm}(\xi)
\label{eq:def2}
\end{equation}
for any $1\leq i<j\leq n$.
\end{proposition}
\begin{remark}
For $i>j$ a similar equality holds:
$$\bw{2}t_{i,j}(\xi)=\prod\limits_{k=1}^{j-1} t_{ki,kj}(\xi)\,\cdot\prod\limits_{l=j+1}^{i-1}t_{li,jl}(-\xi)\,\cdot\prod\limits_{m=i+1}^n t_{im,jm}(\xi).$$
\end{remark}
Likewise, one can get an explicit form of torus elements $h_{\varpi_2}(\xi)$ of the group $\bw{2}\GL_{n}(R)$.
\begin{proposition}\label{prop:ImageOfDiag2}
Let $d_i(\xi)=e+(\xi-1)e_{i,i}$ be a torus generator, $1\leq i\leq n$. Then the exterior square of $d_i(\xi)$ equals a diagonal matrix, with diagonal entries 1 everywhere except in $n-1$ positions:
\begin{equation}
\bw{2}\bigl(d_i(\xi)\bigr)_{I,I}=
\begin{cases}
\xi,& \text{ if } i\in I,\\
1,& \text{ otherwise}.
\end{cases}
\label{eq:diag2}
\end{equation}
\end{proposition}
It follows from the propositions that $\bw{2}t_{i,j}(\xi)\in \E^{n-2}(N,R)$, where a set $\E^M(N,R)$ consists of products of $M$ or less elementary transvections, e.\,g., $\bw{2}t_{1,3}(\xi)=t_{12,23}(-\xi)t_{14,24}(\xi)t_{15,25}(\xi)\in\bw{2}E_5(R)$.
Let $H$ be an overgroup of the exterior square of the elementary group $\bw{2}\E_n(R)$:
$$\bw{2}\E_n(R)\leq H \leq \GL_N(R).$$
We consider two indices $I, J \in \bw{2}[n]$. By $A_{I,J}$ we denote the set
$$A_{I,J}:=\{\xi \in R \;|\; t_{I,J}(\xi) \in H \} \subseteq R.$$
By definition diagonal sets $A_{I,I}$ equals whole ring $R$ for any index $I$. In the rest of the section, we prove that these sets are ideals, i.\,.e., $A_{I,J}$ form a net of ideals. Moreover, we will get $D$-net in terms of Zenon Borevich~\cite{BVnets} by the latter statement.
Let $t_{I,J}(\xi)$ be an elementary transvection. We define a \textit{height} of $t_{I,J}(\xi)$ (generally, of the pair $(I,J)$) as a cardinality of the intersection $I\cap J$:
$$\height(t_{I,J}(\xi))=\height(I,J)=|I\cap J|.$$
This combinatorial characteristic of transvections is useful in commutator calculations.
The height splits up all sets $A_{I,J}$ into two classes: the one with $\height(I,J)=0$ and the other with $\height(I,J)=1$. In fact, these classes are equal for $n\geq 6$. The set $A := A_{I,J} $ is called a \textit{level} of an overgroup $H$. Note that for $n=4$ the level is unique, that follows from~\cite{VP-EOeven}.
\begin{lemma}\label{lem:IdealFor2}
If $n\geq 6$, then every set $A_{I,J}$ is an ideal of the ring $R$. Moreover, for any $I\neq J$ and $K\neq L$ the ideals $A_{I,J}$ and $A_{K,L}$ coincide.
\end{lemma}
\begin{proof} A complete proof is presented in Section~\ref{subsec:levelcomputation}, Proposition~\ref{prop:LevelsEqualGeneral}. Here we sketch calculations in the case $(n,m) = (4,2)$ exclusively. These calculations present the general idea in a transparent way.
\begin{enumerate}
\item Firstly, take any $\xi\in A_{12,34}$, i.\,e., $t_{12,34}(\xi)\in H$. Then
$$[t_{12,34}(\xi),\bw{2}t_{4,2}(\zeta)]=t_{14,23}(-\xi\zeta^2)t_{14,34}(-\zeta\xi)t_{12,23}(-\xi\zeta)\in H.$$
It remains to provide this calculation with $-\zeta$ and to product two right-hand sides; then we obtain $t_{14,23}(-2\xi\zeta^2)\in H$. By the condition $2 \in R^{*}$, this means that $A_{12, 34} \subseteq A_{14, 23}$. It follows that
$$A_{I,J} \subseteq A_{K,L}\text{ for } I\cup J = K \cup L = \{ 1234 \}.$$
\item Secondly, take any $\xi\in A_{12,34}$, then $[t_{12,34}(\xi),\bw{2}t_{4,5}(\zeta)]=t_{12,35}(\xi\zeta)$. Consequently,
$$A_{I,J} \subseteq A_{K,L}\text{ for } \height(I, J) = \height(K, L) = 0.$$
\item Thirdly, let $\xi\in A_{12,13}$, then $[t_{12,13}(\xi),\bw{2}t_{1,4}(\zeta)]=t_{12,34}(-\xi\zeta)\in H$. Consider two commutators of the latter transvection with $\bw{2}t_{4,1}(\zeta_1)$ and $\bw{2}t_{4,1}(-\zeta_1)$ respectively. We obtain that $t_{24,13}(\zeta_1^2\xi\zeta)\in H$ and also $t_{12,13}(-\xi\zeta\zeta_1)t_{24,34}(\zeta_1\xi\zeta)\in H$. Hence $t_{24,34}(\zeta_1\xi\zeta)\in H$. This means that
$$A_{I,J} \subseteq A_{K,L}\text{ for any } \height(I, J) = \height(K, L) = 1.$$
\item Now, take any $\xi\in A_{12,23}$, then $[t_{12,23}(\xi),\bw{2}t_{4,2}(\zeta)]=t_{14,23}(-\zeta\xi)$. Thus
$$A_{I,J} \subseteq A_{K,L}\text{ for } \height(I, J)= 1, \height(K, L) = 0.$$
\item Finally, let $\xi\in A_{12,34}$. As in (1), consider the commutator $t_{12,34}(\xi)$ with $\bw{2}t_{4,2}(\zeta)$. We obtain $t_{14,23}(-2\xi\zeta^2)\in H$ and $t_{14,34}(-\zeta\xi)t_{12,23}(-\xi\zeta)\in H$. By the same argument we can provide these calculations with the transvection $t_{45,16}(\xi)$ and $\bw{2}t_{6,4}(\zeta_1)$. We get that $t_{56,14}(-2\zeta_1^2\xi)\in H$ and $t_{45,14}(\xi\zeta_1)t_{56,16}(\zeta_1\xi)\in H$. To finish the proof it remains to commutate latter two products. Then $t_{45,34}(-\xi^2\zeta_1\zeta)\in H$, or
$$A_{I,J} \subseteq A_{K,L}\text{ for }\height(I, J)= 0, \height(K, L) = 1.$$
\end{enumerate}
\end{proof}
The following lemma is crucial for the rest. It gives an alternative description of the relative elementary group.
\begin{lemma}\label{lem:AlterRelatFor2}
Let $n\geq 6$. For any ideal $A\trianglelefteq R$, we have
$$\E_N(A)^{\bw{2}\E_n(R)}=\E_N(R,A),$$ where by definition $\E_N(R,A)=\E_N(A)^{\E_N(R)}.$
\end{lemma}
\begin{proof}
The inclusion $\leq$ is trivial. By Vaserstein--Suslin's lemma~\cite{VasSusSerre}, the group $\E_N(R,A)$ is generated by elements of the form
$$z_{ij,hk}(\xi,\zeta) = z_{I,J}(\xi,\zeta) = t_{J,I}(\zeta)\,t_{I,J}(\xi)\,t_{J,I}(-\zeta), \;\xi \in A, \zeta \in R.$$
Hence to prove the reverse inclusion, it sufficient to check the
matrix $z_{ij,hk}(\xi,\zeta)$ to belong to $F:=\E_N(A)^{\bw{2}\E_n(R)}$ for any $\xi \in A$, $\zeta \in R$. Let us consider two cases:
\begin{itemize}
\item Suppose that there exists one pair of the same indices. Without loss of generality, we can assume that $i=k$. Then this inclusion is obvious:
$$z_{ij,hi}(\xi,\zeta)={}^{t_{hi,ij}(\zeta)}t_{ij,hi}(\xi)={}^{\bw{2}t_{h,j}(\zeta)}t_{ij,hi}(\xi)\in F.$$
\item Thus, we are left with the inclusion $z_{ij,hk}(\xi,\zeta)\in F$ with different indices $i,\,j\,,h\,,k$. Firstly, we express
$t_{ij,hk}(\xi)$ as a commutator of elementary transvections:
$$z_{ij,hk}(\xi,\zeta)={}^{t_{hk,ij}(\zeta)}t_{ij,hk}(\xi)={}^{t_{hk,ij}(\zeta)}[t_{ij,jh}(\xi),t_{jh,hk}(1)].$$
Conjugating arguments of the commutator by $t_{hk,ij}(\zeta)$, we get
$$z_{ij,hk}(\xi,\zeta)=[t_{ij,jh}(\xi)t_{hk,jh}(\zeta \xi),t_{jh,ij}(-\zeta)t_{jh,hk}(1)] =:[ab,cd].$$
Next, we decompose the right-hand side with a help of the formula
$$[ab,cd] ={}^a[b,c]\cdot {}^{ac}[b,d]\cdot[a,c]\cdot {}^c[a,d],$$
and observe the exponent $a$ belongs to $\E_N(A)$, so can be ignored. Now a direct calculation, based upon the Chevalley commutator formula, shows that
\begin{align*}
[b,c]&=[t_{hk,jh}(\zeta \xi),t_{jh,ij}(-\zeta)]=t_{hk,ij}(-\zeta^2 \xi) \in \E_N(A);\\
^c[b,d]&={}^{t_{jh,ij}(-\zeta)}[t_{hk,jh}(\zeta \xi),t_{jh,hk}(1)]=\\
&=t_{hk,ik}(-\xi\zeta^2(1+\xi\zeta))t_{jh,ik}(-\xi\zeta^2)\cdot \,{}^{\bw{2}t_{h,i}(\zeta)}[t_{hk,jh}(\xi\zeta),\bw{2}t_{j,k}(-1)];\\
[a,c]&=[t_{ij,jh}(\xi),t_{jh,ij}(-\zeta)]=[t_{ij,jh}(\xi),\bw{2}t_{h,i}(-\zeta)];\\
^c[a,d]&={}^{t_{jh,ij}(-\zeta)}[t_{ij,jh}(\xi),t_{jh,hk}(1)]=\\
&=t_{jh,ik}(\xi\zeta^2)t_{ij,ik}(-\xi\zeta)\cdot\,{}^{\bw{2}t_{h,i}(\zeta)}[t_{ij,jh}(\xi),\bw{2}t_{j,k}(-1)],
\end{align*}
where all factors on the right-hand side belong to $F$.
\end{itemize}
\end{proof}
\begin{remark}
The attentive reader can remark these calculations to be almost completely coincide with the calculations for the orthogonal and symplectic cases~\cite{VP-EOeven,VP-EOodd,VP-Ep}. In the special case $(n,m) = (4,2)$ calculations are the same due to the isomorphism $\bw{2}\E_4(R) \cong \EO_6(R)$. Amazingly this argument proves a similar proposition in the case of general exterior power (see Section~\ref{subsec:levelcomputation}, Lemma~\ref{lem:AlterRelatForm}).
\end{remark}
\begin{corollary}\label{cor:CorolOfL2}
Let $A$ be an arbitrary ideal of $R$. Then
$$\bw{2}\E_n(R)\cdot\E_N(R,A)=\bw{2}\E_n(R)\cdot\E_N(A).$$
\end{corollary}
Summarizing above two lemmas, we get the main result of the paper for bivectors.
\begin{theorem}[Level Computation]\label{thm:LevelFor2}
Let $n\geq 6$ and let $H$ be a subgroup in $\GL_{N}(R)$ containing $\bw{2}\E_n(R)$. Then there exists a unique maximal ideal $A\trianglelefteq R$ such that
$$\bw{2}\E_n(R)\cdot\E_N(R,A)\leq H.$$
Namely, if $t_{I,J}(\xi)\in H$ for some $I$ and $J$, then $\xi\in A$.
\end{theorem}
Lemma~\ref{lem:AlterRelatFor2} asserts precisely $\bw{2}\E_n(R)\cdot\E_N(R,A)$ to be generated as a subgroup by transvections $\bw{2}t_{i,j}(\zeta)$, $\zeta\in R$, and by elementary transvections $t_{ij,hk}(\xi)$, $\xi\in A$ of level $A$. As usual, we assume that $n\geq 6$ and $2\in R^{*}$.
We formulate a perfectness of the lower bound subgroup from the latter Theorem. The proof follows from Lemma \ref{lem:PerfectForm}.
\begin{lemma}\label{lem:PerfectFor2}
Let $n\geq 6$. The group $\bw{2}\E_n(R)\cdot\E_N(R,A)$ is perfect for any ideal $A\trianglelefteq R$.
\end{lemma}
\subsection{Exterior powers of elementary groups}\label{subsec:general}
In this section, we lift the previous statements from the level of an exterior square to the case of an arbitrary exterior power functor.
Let us define an $m$-th exterior power of an $R$-module $R^n$ as follows. A basis of this module consists of exterior products $e_{i_1}\wedge\dots\wedge e_{i_m}$, where $1\leq i_1<\dots<i_m\leq n$. Products $e_{i_1}\wedge\dots\wedge e_{i_m}$ are defined for any set $i_1,\dots,i_m$ as $e_{\sigma(i_{1})}\wedge\ldots\wedge e_{\sigma(i_{m})} = \sign(\sigma)\, e_{i_1} \wedge \ldots \wedge e_{i_{m}}$ for any permutation $\sigma$ in the permutation group $S_{m}$. We denote the $m$-th exterior power of $R^n$ by $\bw{m}R^n$.
For every $m$ the group $\GL_{n}(R)$ acts diagonally on the module $\bw{m}R^n$. Namely, an action of a matrix $g\in\GL_n(R)$ on decomposable $m$-vectors is set according to the rule
$$\bw{m}(g)(e_{i_1}\wedge\dots\wedge e_{i_m}):=(ge_{i_1})\wedge\dots\wedge (ge_{i_m})$$
for every $e_{i_1},\dots,e_{i_m}\in R^n$. In the basis $e_I, I\in\bw{m}[n]$ a matrix $\bw{m}(g)$ consists of $m$-order minors of the matrix $g$ with lexicographically ordered columns and rows:
$$\left(\bw{m}(g)\right)_{I,J}=\left(\bw{m}(g)\right)_{(i_1,\dots,i_m),(j_1,\dots,j_m)}=M_{i_1,\dots,i_m}^{j_1,\dots,j_m}(g).$$
By the Cauchy--Binet theorem the map $\pi\colon\GL_{n}(R)\longrightarrow \GL_{N}(R)$, $x\mapsto \bw{m}(x)$ is homomorphism. Thus, the map $\pi$ is a representation of the group $\GL_n(R)$ called the $m$-th \textit{vector representation} or the $m$-th \textit{fundamental representation} (\textit{the representation with the highest weight} $\varpi_m$). The image of the latter action is called the $m$-th exterior power of the group $\GL_n(R)$. $\E_n(R)$ is a subgroup of $\GL_n(R)$, therefore the exterior power of the elementary group is well defined.
We cannot but emphasize the difference for arbitrary rings between the groups\footnote{The same strict inclusions are still true with changing $\GL$ to $\SL$.}
$$\bw{m} \bigl(\GL_{n}(R)\bigr) < \bw{m}\GL_{n}(R) < \GL_{\binom{n}{m}}(R).$$
The first group is a set-theoretic image of the [abstract] group $\GL_{n}(R)$ under the Cauchy--Binet homomorphism $\bw{m}:\GL_{n}(R)\longrightarrow \GL_{\binom{n}{m}}\left(R\right)$, while the second one is a group of $R$-points of the \textbf{categorical} image of the group scheme $\GL_{n}$ under the natural transformation corresponding to the Cauchy--Binet homomorphism. Since the epimorphism of algebraic groups on points is not surjective in this situation, we see that $\bw{m}\GL_n(R)$ is strictly larger than $\bw{m}\bigl(\GL_{n}(R)\bigr)$. In fact, elements of $\bw{m}\GL_n(R)$ are still images of matrices, but coefficients are not from the ring itself, but from its extensions. This means that for any commutative ring $R$ elements $\widetilde{g}\in\bw{m}\GL_n(R)$ can be represent in the form $\widetilde{g}=\bw{m}g$, $g\in\GL_n(S)$, where $S$ is an extension of the ring $R$. We refer the reader to~\cite{VavPere} for more precise results about the difference between these groups.
As in Section~\ref{subsec:square}, $\bw{m}\E_n(R)$ is a normal subgroup of $\bw{m}(\GL_{n}(R))$ by Suslin's lemma. Moreover, $\bw{m}\E_n(R)$ is normal in $\bw{m}\GL_n(R)$. This fact follows from~\cite[Theorem~1]{PetrovStavrovaIsotropic}.
\begin{theorem}\label{thm:normalityPS}
Let $R$ be a commutative ring, $n\geq 3$, then
$\bw{m}\E_n(R)\trianglelefteq\bw{m}\GL_n(R)$.
\end{theorem}
For further computations we calculate an exterior power of an elementary transvection in the following proposition. The proof is straightforward by the very definition of the [classical] Binet--Cauchy homomorphism.
\begin{proposition}\label{prop:ImageOfTransvForm}
Let $t_{i,j}(\xi)$ be an elementary transvection in $\E_n(R)$, $n\geq 3$. Then $\bw{m}t_{i,j}(\xi)$ equals
\begin{equation}
\bw{m}t_{i,j}(\xi)=\prod\limits_{L\,\in\,\bw{m-1}\,[n\setminus \{i,j\}]} t_{L\cup i,L\cup j}(\sign(L, i)\sign(L, j)\xi)
\label{eq:m}
\end{equation}
for any $1\leq i\neq j\leq n$.
\end{proposition}
Similarly, one can get an explicit form of torus elements $h_{\varpi_m}(\xi)$ of the group $\bw{m}\GL_{n}(R)$.
\begin{proposition}\label{prop:ImageOfDiag}
Let $d_i(\xi)=e+(\xi-1)e_{i,i}$ be a torus generator, $1\leq i\leq n$. Then the exterior power of $d_i(\xi)$ equals a diagonal matrix, with diagonal entries 1 everywhere except in $n-1$ positions:
\begin{equation}
\bw{m}(d_i(\xi))_{I,I}=
\begin{cases}
\xi,& \text{ if } i\in I,\\
1,& \text{ otherwise}.
\end{cases}
\label{eq:diagm}
\end{equation}
\end{proposition}
As an example, consider $\bw{3}t_{1,3}(\xi)=t_{124,234}(-\xi)t_{125,235}(-\xi)t_{145,345}(\xi)\in\bw{3}\E_5(R)$ and $\bw{4}d_2(\xi)=\opn{diag}(\xi,\xi,\xi,1,\xi)\in\bw{4}\E_5(R)$. It follows from the propositions $\bw{m}t_{i,j}(\xi)\in \E^{\binom{n-2}{m-1}}(N,R)$, where by definition every element of the set $\E^M(N,R)$ is a product of $M$ or less elementary transvections. In other words, a residue of a transvection $\opn{res}(\bw{m}t_{i,j}(\xi))$ equals the binomial coefficient $\binom{n-2}{m-1}$. Recall that a residue $\opn{res}(g)$ of a transformation $g$ is called the rank of $g-e$. Finally, there is a simple connection between the determinant of a matrix $g\in\GL_n(R)$ and the determinant of $\bw{m}g\in\bw{m}\GL_n(R)$, see~\cite[Proof of Theorem~4]{WaterhousePGL}:
$$\det\bw{m}g=\bigl(\det(g)\bigr)^{\binom{n}{m}\cdot\frac{m}{n}}=\bigl(\det(g)\bigr)^{\binom{n-1}{m-1}}.$$
\subsection{Elementary calculations technique}\label{subsec:levelgeneral}
For an arbitrary exterior power calculations with elementary transvections are huge. In this section, we organize all possible calculations of a commutator of an elementary transvection with an exterior transvection.
\begin{proposition}\label{prop:TypesOfComm}
Up to the action of the permutation group there exist three types of commutators with a fixed transvection $t_{I,J}(\xi) \in \E_N(R)$:
\begin{enumerate}
\item $[t_{I,J}(\xi), \bw{m}t_{j,i}(\zeta)]=1$ if both $i\not\in I$ and $j \not \in J$ hold;
\item $[t_{I,J}(\xi), \bw{m}t_{j,i}(\zeta)] = t_{\tilde{I}, \tilde{J}}(\pm \zeta\xi)$ if either $i\in I$ or $j \in J$. And then $\tilde{I} = I\backslash i \cup j$ or $\tilde{J} = J\backslash j \cup i$ respectively;
\item If both $i\in I$ and $j \in J$ hold, then we have the equality:
$$[t_{I,J}(\xi), \bw{m}t_{j,i}(\zeta)] = t_{\tilde{I}, J}(\pm \zeta\xi)\cdot t_{I, \tilde{J}}(\pm \zeta\xi)\cdot t_{\tilde{I}, \tilde{J}}(\pm\zeta^2\xi).$$
\end{enumerate}
\end{proposition}
Note that the latter case is true whenever $I\setminus i\neq J\setminus j$, otherwise we obtain $[t_{I,J}(\xi),t_{J,I}(\pm\zeta)]$. This commutator cannot be presented in a simpler form than the very definition.
The rule for commutator calculations from the latter proposition can be translated into the language of \textit{weight diagrams}:\\[7mm]
\textbf{Weight diagrams tutorial.}
\begin{enumerate}
\item Let $G(A_{n-1},\blank)$ be a Chevalley--Demazure group scheme, and let $(I,J)\in\bw{m}[n]^2$ be a pair of different weights for the $m$-th exterior power of $G(A_{n-1},\blank)$. Consider any unipotent $x_\alpha(\xi)$ for a root $\alpha$ of the root system $A_{n-1}$, i.\,e., $x_\alpha(\xi)$ equals an elementary transvection $\bw{m}t_{i,j}(\xi) \in \bw{m}\E_n( R)$;
\item By $\Ar(\alpha)$ denote all paths on the weight diagram\footnote{Recall that we consider the representation with the highest weight $\varpi_{m}$.} of this representation corresponding to the root $\alpha$;
\item Then there exist three different scenarios corresponding to the cases of Proposition~\ref{prop:TypesOfComm}:
\begin{itemize}
\item sets of the initial and the terminal vertices of paths from $\Ar(\alpha)$ do not contain the vertex $(I,J)$;
\item the vertex $(I,J)$ is initial or terminal for one path from $\Ar(\alpha)$;
\item the vertex $(I,J)$ is simultaneously initial and terminal for some path\footnote{From root systems geometry any vertex can be initial or terminal for at most one $\alpha$-path.} from $\Ar(\alpha)$.
\end{itemize}
\item Finally, let us consider a commutator of the transvection $t_{I,J}(\xi)$ and the element $\bw{m}t_{i,j}(\zeta)$. It equals a product of transvections. These transvections correspond to the paths from the previous step. Transvections' arguments are monomials in $\xi$ and $\zeta$. Namely, in the second case the argument equals $\pm\xi\zeta$; in the third case it equals $\pm\xi\zeta^2$.
\end{enumerate}
In Figure~\ref{Fig1}$(a)$ we present all three cases from step $(3)$ for $m=2$ and $\alpha = \alpha_{2}$:
\begin{itemize}
\item $(I,J) = (14,15)$, then $[t_{14, 15}(\xi), \bw{2}t_{2,3}(\zeta)] = 1$;
\item $(I,J) = (13,35)$, then $[t_{13, 35}(\xi), \bw{2}t_{2,3}(\zeta)] = t_{12, 35}(-\xi\zeta)$;
\item $(I,J) = (13,24)$, then $[t_{13, 24}(\xi), \bw{2}t_{2,3}(\zeta)] = t_{12, 24}(-\xi\zeta)t_{12,34}(\xi\zeta^2) t_{13,34}(\zeta\xi)$.
\end{itemize}
Similarly, for the case $m=3$ the elementary calculations can be seen directly from Figure~\ref{Fig1}$(b)$.
\[
\xymatrix @+1.0pc {
{\overset{12}{\bullet}}\ar@{-}[r] &{\overset{13}{\bullet}}\ar@{-}[r]\ar@{-}[d]\textbf{\ar@/_1pc/[l]_2}&{\overset{14}{\bullet}}\ar@{-}[r]\ar@{-}[d]&{\overset{15}{\bullet}}\ar@{-}[d]\\
&{\overset{23}{\bullet}}\ar@{-}[r]&{\overset{24}{\bullet}}\ar@{-}[r]\ar@{-}[d]&{\overset{25}{\bullet}}\ar@{-}[d] \\
&&{\overset{34}{\bullet}}\ar@{-}[r]\textbf{\ar@/_1pc/[u]_2}&{\overset{35}{\bullet}}\ar@{-}[d]\textbf{\ar@/_1pc/[u]_2}\\
&&&{\overset{45}{\bullet}}\\
&&(a)}
\hspace{5mm}
\xymatrix @-2.0pc {
{\overset{123}{\bullet}}\ar@{-}[rrrr] &&&&{\overset{124}{\bullet}}\ar@{-}[rrrr]\ar@{-}[ddrrr]&&&&\textbf{\ar@/_1pc/[llll]_4}{\overset{125}{\bullet}}\ar@{-}[rrrr]\ar@{-}[ddrrr]&&&&{\overset{126}{\bullet}}\ar@{-}[ddrrr]\\
\\
&&&&&&&{\overset{134}{\bullet}}\ar@{-}[rrrr]\ar@{-}[ddd] &&&&{\overset{135}{\bullet}}\ar@{-}[rrrr]\ar@{-}[ddrrr]\ar@{-}[ddd]\textbf{\ar@/_1pc/[llll]_4}&&&&{\overset{136}{\bullet}}\ar@{-}[ddrrr]\ar@{-}[ddd]\\
\\
&&&&&&&&&&&&&&{\overset{145}{\bullet}}\ar@{-}[rrrr]\ar@{-}[ddd] &&&&{\overset{146}{\bullet}}\ar@{-}[ddrrr]\ar@{-}[ddd]\\
&&&&&&&{\overset{234}{\bullet}}\ar@{-}[rrrr] &&&&{\overset{235}{\bullet}}\ar@{-}[rrrr]\ar@{-}[ddrrr]\textbf{\ar@/_1pc/[llll]_4}&&&&{\overset{236}{\bullet}}\ar@{-}[ddrrr]\\
&&&&&&&&&&&&&&&&&&&&&{\overset{156}{\bullet}}\ar@{-}[ddd]\textbf{\ar@/_1pc/[llluu]_4}\\
&&&&&&&&&&&&&&{\overset{245}{\bullet}}\ar@{-}[rrrr]\ar@{-}[ddd]&&&&{\overset{246}{\bullet}}\ar@{-}[ddrrr]\ar@{-}[ddd]\\
\\
&&&&&&&&&&&&&&&&&&&&&{\overset{256}{\bullet}}\ar@{-}[ddd]\textbf{\ar@/_1pc/[llluu]_4}\\
&&&&&&&&&&&&&&{\overset{345}{\bullet}}\ar@{-}[rrrr] &&&&{\overset{346}{\bullet}}\ar@{-}[ddrrr]\\
\\
&&&&&&&&&&&&&&&&&&&&&{\overset{356}{\bullet}}\ar@{-}[ddd]\textbf{\ar@/_1pc/[llluu]_4}\\
\\
\\
&&&&&&&&&&&&&&&&&&&&&{\overset{456}{\bullet}}\\
&&&&&&&&&&&&&&(b)}
\]
\captionof{figure}{Weight diagrams for $(a)$: $(A_4,\varpi_2)$, $\alpha = \alpha_{2}$ and $(b)$: $(A_5,\varpi_3)$, $\alpha = \alpha_{4}$}\label{Fig1}
\subsection{Level computation}\label{subsec:levelcomputation}
We generalize the notion of ideals $A_{I,J}$ to the case of the $m$-th exterior power. Let $H$ be an overgroup of the exterior power of the elementary group $\bw{m}\E_n(R)$:
$$\bw{m}\E_n(R)\leq H \leq \GL_N(R).$$
Let
$$A_{I,J}:=\{\xi \in R \;|\; t_{I,J}(\xi) \in H \}$$
for any indices $I, J \in \bw{m}[n]$. As usual, diagonal sets $A_{I,I}$ equal the whole ring $R$ for any index $I\in\bw{m}[n]$. Thus, we will construct $D$-net of ideals of the ring $R$. Recall that the desired parametrization is given by an explicit juxtaposition for any overgroup $H$ its level, namely an ideal $A$ of the ring $R$. We compute this ideal $A$ in the present section.
We assume that $n \geq 2m$ due to the isomorphism $\bw{m}V^{*} \cong (\bw{\dim(V)-m}V)^{*}$ for an arbitrary free $R$-module $V$. The first step toward the level description is the following observation.
\begin{proposition}\label{prop:LevelsEqualGeneral}
If $|I \cap J|=|K\cap L|$, then sets $A_{I,J}$ and $A_{K,L}$ coincide. In fact, $A_{I,J}$ are ideals of $R$.
\end{proposition}
But first, we prove a weaker statement.
\begin{lemma} \label{lem:LevelsEqual0}
Let $I, J, K, L$ be different elements of the set $\bw{m}[n]$ such that $|I \cap J|=|K\cap L| = 0$. If $n \geq 2m$, then sets $A_{I,J}$ and $A_{K,L}$ coincide.
\end{lemma}
\begin{proof}[Proof of the lemma] The sets $A_{I, J}$ coincide when the set $I \cup J$ is fixed. This fact can be proved by the third type commutation due to Proposition~\ref{prop:TypesOfComm} with $\zeta$ and $-\zeta$. If $\xi \in A_{I,J}$ we get a transvection $t_{I,J}(\xi) \in H$. Then the following two products also belong to $H$:
\begin{align*}
[t_{I,J}(\xi), \bw{m}t_{j,i}(\zeta)] &= t_{\tilde{I}, J}(\pm \zeta\xi)\cdot t_{I, \tilde{J}}(\pm \zeta\xi)\cdot t_{\tilde{I}, \tilde{J}}(\pm\zeta^2\xi)\\
[t_{I,J}(\xi), \bw{m}t_{j,i}(-\zeta)] &= t_{\tilde{I}, J}(\mp \zeta\xi)\cdot t_{I, \tilde{J}}(\mp \zeta\xi)\cdot t_{\tilde{I}, \tilde{J}}(\pm \zeta^2\xi).
\end{align*}
This implies that the product of two factors on the right-hand sides $t_{\tilde{I}, \tilde{J}}(\pm 2\zeta^2\xi)$ belongs to $H$.
It can be easily proved the set $I\cup J$ can be changed by the second type commutations. For example, the set $I_{1}\cup J_{1} = \{1,2,3,4,5,6\}$ can be replaced by the set $I_{2} \cup J_{2} = \{1,2,3,4,5,7\}$ as follows
$$[t_{123,456}(\xi), \bw{3}t_{6,7}(\zeta)] = t_{123, 457}(\xi\zeta).$$
\end{proof}
\begin{proof}[Proof of Proposition~$\ref{prop:LevelsEqualGeneral}$]
Arguing as above, we see that the sets $A_{I,J}$ and $A_{K,L}$ coincide in the case $I\cap J = K\cap L$, where $n_{1} = n - |I\cap J| \geq 2\cdot m - 2 \cdot |I\cap J| = 2 \cdot m_{1}$.
In a general case, we can prove the statement by both the second and the third types commutations. Let us give an example of this calculation with replacing the set $I \cap J = \{1,2\}$ by the set $\{1,5\}$.
Let $t_{123,124}(\xi) \in H$. So we have $[t_{123,124}(\xi), \bw{3}t_{2,5}(\zeta) ] = t_{123,145}(-\xi\zeta) \in H$. We commute this transvection with the element $\bw{3}t_{5,2}(\zeta_1)$. Then the transvection $t_{135,124}(-\zeta_1^2\xi\zeta)$ belongs to $H$ as well as the product $t_{123,124}(\xi\zeta\zeta_1)\cdot t_{135, 145}(-\zeta_1\xi\zeta)\in H$. From the latter inclusion we can see $t_{135, 145}(-\zeta_1\xi\zeta)\in H$ and $I\cap J = \{1,5\}$.
To prove all $A_{I,J}$ are ideals in $R$ it is sufficient to commute any elementary transvection with exterior transvections with $\zeta$ and $1$:
$$t_{I,J}(\xi\zeta)=[t_{I,J}(\xi),\bw{m}t_{j,i}(\zeta),\bw{m}t_{i,j}(\pm 1)]\in H.$$
\end{proof}
Let $t_{I,J}(\xi)$ be an elementary transvection. Let us define a \textit{height} of $t_{I,J}(\xi)$ (more abstractly, of the pair $(I,J)$) as the cardinality of the set $I\cap J$:
$$\height(t_{I,J}(\xi))=\height(I,J)=|I\cap J|.$$
This combinatorial characteristic plays the same role as the distance function $d(\lambda,\mu)$ for roots $\lambda$ and $\mu$ on a weight diagram of a root system. Now Proposition~\ref{prop:LevelsEqualGeneral} can be rephrased as follows.
Sets $A_{I,J}$ and $A_{K,L}$ coincide for the same heights: $A_{I,J} = A_{K,L} =A_{|I\cap J|}$. Suppose that the height of $(I,J)$ is larger than the height of $(K,L)$, then using Proposition~\ref{prop:TypesOfComm}, we get $A_{I,J} \leq A_{K,L}$.
Summarizing the above arguments, we have the height grading:
$$A_{0} \geq A_{1} \geq A_{2} \geq \dots \geq A_{m-2} \geq A_{m-1}.$$
The following result proves a coincidence of the sets $\{A_{k}\}_{k=0\dots m-1}$.
\begin{proposition}\label{prop:IdealEqual}
The ideals $A_{k}$ coincide for $n \geq 3m$. More accurately, the inverse inclusion $A_k\leq A_{k+1}$ takes place if $n\geq 3m-2k$.
\end{proposition}
\begin{proof}
The statement can be proved by the double third type commutation as follows. Let $\xi\in A_k$, i.\,e., a transvection $t_{I,J}(\xi)\in H$ for $\height(t_{I,J}(\xi))=k$. By the third type commutation with a transvection $\bw{m}t_{j,i}(\zeta)$, we have $t_{\tilde{I}, J}(\pm \zeta\xi)\cdot t_{I, \tilde{J}}(\pm \zeta\xi)\in H$. Let us consider an analogous commutator with a specifically chosen transvections $t_{I_1,J_1}(\xi)\in H$ and $\bw{m}t_{j_1,i_1}(\zeta_1)$. We get that $t_{\tilde{I}_1, J_1}(\pm \zeta_1\xi)\cdot t_{I_1, \tilde{J}_1}(\pm \zeta_1\xi)\in H$. The final step is to commute the latter products.
The choice of transvections goes in the way such that the final commutator (initially of the form $[ab,cd]$) equals an elementary transvection. This choice is possible due to the condition $n \geq 3m-2k$.
Let us give a particular example of such calculations for the case $m=4$. This calculation could be easily generalized. The first three steps below correspond to the inclusions $A_{0} \leq A_{1}$, $A_{1} \leq A_{2}$, and $A_{2} \leq A_{3}$ respectively. We emphasize that the ideas of the proof of all three steps are \textbf{completely} identical. The difference has to do only with a choice of the appropriate indices. We replace the numbers $10$, $11$, $12$ with the letters $\alpha$, $\beta$, $\gamma$ respectively.
\begin{enumerate}
\item Let $\xi\in A_0$. Consider the mutual commutator
$$\bigl[[t_{1234,5678}(\xi), \bw{4}t_{8,4}(\zeta)],[t_{49\alpha\beta,123\gamma}(\xi), \bw{4}t_{\gamma,4}(\zeta_1)]\bigr]\in H.$$
It is equal to the commutator
$$[t_{1234,4567}(-\xi\zeta)\cdot t_{1238,5678}(-\zeta\xi),t_{49\alpha\beta,1234}(\xi\zeta_1)\cdot t_{9\alpha\beta\gamma,123\gamma}(\zeta_1\xi)]\in H,$$
which is a transvection $t_{49\alpha\beta,4567}(\xi^2\zeta_1\zeta)\in H$. As the result, $A_0\leq A_1$.
\item For $\xi\in A_1$ consider similar commutator
$$\bigl[[t_{1234,1567}(\xi), \bw{4}t_{7,4}(\zeta)],[t_{1489,123\alpha}(\xi), \bw{4}t_{\alpha,4}(\zeta_1)]\bigr]\in H.$$
Thus,
$$[t_{1234,1456}(\xi\zeta)\cdot t_{1237,1567}(-\zeta\xi),t_{1489,1234}(\xi\zeta_1)\cdot t_{189\alpha,123\alpha}(-\zeta_1\xi)]\in H.$$
Again this commutator is equal to $t_{1489,1456}(-\xi^2\zeta_1\zeta)\in H$, i.\,e., $A_1\leq A_2$.
\item Finally, let $\xi\in A_2$. Consider the commutator
$$\bigl[[t_{1234,1256}(\xi), \bw{4}t_{6,4}(\zeta)],[t_{1248,1237}(\xi), \bw{4}t_{7,4}(\zeta_1)]\bigr]\in H.$$
It is equal to the commutator
$$[t_{1234,1245}(-\xi\zeta)\cdot t_{1236,1256}(-\zeta\xi),t_{1248,1234}(\xi\zeta_1)\cdot t_{1278,1237}(-\zeta_1\xi)]\in H,$$
which is an elementary transvection $t_{1248,1245}(\xi^2\zeta_1\zeta)\in H$. Thus, $A_2\leq A_3$.
\end{enumerate}
\end{proof}
We proved that all ideals $A_i$ coincide for a large enough $n$. However, the following proposition shows relations between the ideals without this restriction. Recall that a residue $\opn{res}$ of an exterior transvection $\bw{m}t_{i,j}(\xi)$ equals the binomial coefficient $\binom{n-2}{m-1}$.
\begin{proposition}\label{prop:Relations}
For ideals $\{A_0,\dots,A_{m-1}\}$ the following relations hold:
\begin{gather*}
A_k\leq A_{k+1}, \text{ for } n\geq 3m-2k;\\
A_{0} \geq A_{1} \geq A_{2} \geq \dots \geq A_{m-2} \geq A_{m-1};\\
\opn{res}\cdot A_{m-2}\leq A_{m-1}.
\end{gather*}
\end{proposition}
Note that we have not included the relations from the notion of $D$-net $A_{I,J}\cdot A_{J,K}\leq A_{I,K}$, since they hold for any net of ideals by the definition.
\begin{proof}
The first two series of relations have already been proved. Therefore, we must only prove that $\opn{res}\cdot A_{m-2}\leq A_{m-1}$. Again we will use the third type commutation.
Let $\xi\in A_{m-2}$,i.\,e., for any indices $I,J$ with $\height(I,J)=m-2$ a transvection $t_{I,J}(\xi)\in H$. Note that if $i\in I, j\in J$, then in the commutator
$$[t_{I,J}(\xi),\bw{m}t_{j,i}(\zeta)]=t_{\tilde{I}, J}(\pm \zeta\xi)\cdot t_{I, \tilde{J}}(\pm \zeta\xi)\cdot t_{\tilde{I}, \tilde{J}}(\pm\zeta^2\xi)$$
the transvection $t_{\tilde{I}, \tilde{J}}(\pm\zeta^2\xi)$ belongs to the group $H$. Indeed, the height of indices $\tilde{I}=I\backslash i \cup j$ и $\tilde{J} = J\backslash j \cup i$ coincide with the height of $I,J$. At the same time the height of $\tilde{I}, J$ and $I, \tilde{J}$ equals $m-1$. Thus $t_{\tilde{I}, J}(\pm \zeta\xi)\cdot t_{I, \tilde{J}}(\pm \zeta\xi)\in H$ for all indices $I,J$ with $\height(I,J)=m-2$ and all different $i\in I, j\in J$.
Consider $\bw{m}t_{1,2}(\xi\zeta)\in H$, where $\zeta\in R$. By the definition of exterior transvections~\eqref{eq:m}, we have $\bw{m}t_{1,2}(\xi\zeta)=\prod\limits_{L} t_{L\cup 1,L\cup 2}(\xi\zeta)$. The proof is to consistently reduce the number of factors in the product by multiplication $\bw{m}t_{1,2}(\xi\zeta)$ on suitable transvections $t_{\tilde{I}, J}(\pm \zeta\xi)\cdot t_{I, \tilde{J}}(\pm \zeta\xi)\in H$. Finally, we get an elementary transvection $t_{P\cup 1,P\cup 2}(c\xi\zeta)$, where the height of indices equals $m-1$ and the coefficient $c$ equals $\binom{n-2}{m-1}$.
Let us give an example such argument for the exterior cube of the elementary group of dimension 5. Take $\xi\in A_1,\zeta\in R$, $\bw{3}t_{1,2}(\xi\zeta)=t_{134,234}(\xi\zeta)t_{135,235}(\xi\zeta)t_{145,245}(\xi\zeta)$.\\
First, consider the commutator $[t_{134,245}(\xi),\bw{3}t_{5,3}(\zeta)]\in H$. As we mentioned above, the matrix $z_1:=t_{134,234}(-\xi\zeta)t_{145,245}(\xi\zeta)\in H$. Thus,
$$\bw{3}t_{1,2}(\xi\zeta)\cdot z_1=t_{135,235}(\xi\zeta)t_{145,245}(2\xi\zeta)\in H.$$
To get an elementary transvection, consider one more commutator\\
$[t_{135,245}(\xi),\bw{3}t_{4,3}(-\zeta)]\in H$. Then the matrix $z_2:=t_{145,245}(\xi\zeta)t_{135,235}(-\xi\zeta)\in H$. It remains to multiply $\bw{3}t_{1,2}(\xi\zeta)$ and $z_1z_2$. We get the transvection $t_{145,245}(3\xi\zeta)\in H$. Therefore, $3\xi\zeta\in A_{2}$.
\end{proof}
As usual, the set $A = A_{I,J}$ is called a \textit{level} of an overgroup $H$. For level computation we need an alternative description of the relative elementary group.
\begin{lemma}\label{lem:AlterRelatForm}
For any ideal $A\trianglelefteq R$, we have
$$\E_N(A)^{\bw{m}\E_n(R)}=\E_N(R,A),$$
where by definition $\E_N(R,A)=\E_N(A)^{\E_N(R)}$.
\end{lemma}
\begin{proof}
Clearly, the left-hand side is contained in the right-hand side. The proof of the inverse inclusion goes by induction on the height of $(I,J)$. By Vaserstein--Suslin's lemma~\cite{VasSusSerre} it is sufficient to check the matrix $z_{I,J}(\xi,\zeta)$ to belong to $F:=\E_N(A)^{\bw{m}\E_n(R)}$ for any $\xi \in A$, $\zeta \in R$.
In the base case $|I\cap J|=m-1$, the inclusion is obvious:
$$z_{I,J}(\xi,\zeta)\cdot t_{I,J}(-\xi)=[t_{J,I}(\zeta),t_{I,J}(\xi)]=\left[\bw{m}t_{j_1,i_1}(\zeta),t_{I,J}(\xi) \right]\in F.$$
Now, let us consider the general case $|I\cap J|=p$, i.\,e., $I=k_1\dots k_{p}i_1\dots i_{q}$ and $J=k_1\dots k_pj_1\dots j_q$. For the following calculations we need two more sets $V:=k_1\dots k_pi_1\dots i_{q-1}j_q$ and $W:=k_1\dots k_pj_1\dots j_{q-1}i_q$.\\
Firstly, we express $t_{I,J}(\xi)$ as a commutator of elementary transvections,
$$z_{I,J}(\xi,\zeta)=^{t_{J,I}(\zeta)}t_{I,J}(\xi)=^{t_{J,I}(\zeta)}[t_{I,V}(\xi),t_{V,J}(1)].$$
Conjugating the arguments of the commutator by $t_{J,I}(\zeta)$, we get
$$[t_{J,V}(\zeta\xi)t_{I,V}(\xi),t_{V,I}(-\zeta)t_{V,J}(1)]=:[ab,cd].$$
Next, we decompose the right-hand side with a help of the formula
$$[ab,cd] ={}^a[b,c]\cdot {}^{ac}[b,d]\cdot[a,c]\cdot {}^c[a,d],$$
and observe the exponent $a$ to belong to $\E_N(A)$, so can be ignored. Now a direct calculation, based upon the Chevalley commutator formula, shows that
\begin{align*}
[b,c]&=[t_{I,V}(\xi),t_{V,I}(-\zeta)]\in F \textrm{ (by the induction step for the height } m-1);\\
^c[b,d]&=^{t_{V,I}(-\zeta)}[t_{I,V}(\xi),t_{V,J}(1)]=t_{V,W}(\xi\zeta^2)t_{I,W}(-\xi\zeta)\cdot\; ^{\bw{m}t_{j_q,i_q}(-\zeta)}t_{I,J}(\xi);\\
[a,c]&=[t_{J,V}(\zeta\xi),t_{V,I}(-\zeta)]=t_{J,I}(-\zeta^2\xi);\\
^c[a,d]&=^{t_{V,I}(-\zeta)}[t_{J,V}(\zeta\xi),t_{V,J}(1)]=\\
=t_{J,W}(-\xi\zeta^2(1+\xi\zeta))&t_{V,W}(-\xi\zeta^2)\cdot\;^{\bw{m}t_{j_q,i_q}(-\zeta)}t_{J,V}(\xi\zeta)\cdot\;^{\bw{m}t_{j_q,i_q}(-\zeta)}z_{J,V}(-\zeta\xi,1)\in F,\\
\hspace{2cm}&\hspace{4cm}\textrm{ (by the induction step for the height }p+1)
\end{align*}
where all factors on the right-hand side belong to $F$.
\end{proof}
\begin{remark}
Since we do not use the coincidental elements of $I$ and $J$, we also can prove this Lemma by induction on $|I\backslash J| = |J\backslash I| = 1/2 \cdot |I \triangle J|$. Then we can assume that $m$ is an arbitrarily large number (mentally, $m =\infty$).
\end{remark}
\begin{corollary}\label{cor:CorolOfL6}
Suppose $A$ be an arbitrary ideal of the ring $R$; then
$$\bw{m}\E_n(R)\cdot\E_N(R,A)=\bw{m}\E_n(R)\cdot\E_N(A).$$
\end{corollary}
Summarizing Proposition~\ref{prop:IdealEqual} and Lemma~\ref{lem:AlterRelatForm}, we get the main result of the paper for the general case.
\LevelForm*
\subsection{Normalizer of \texorpdfstring{$\E\bw{m}\E_n(R,A)$}{TEXT}}\label{subsec:normgeneral}
In this section, we describe a normalizer of the lower bound for a group $H$.
\begin{lemma}\label{lem:PerfectForm}
Let $n\geq 3m$. The group $\E\bw{m}\E_n(R,A) := \bw{m}\E_n(R)\cdot\E_N(R,A)$ is perfect for any ideal $A\trianglelefteq R$.
\end{lemma}
\begin{proof}
It is sufficient to verify all generators of the group $\bw{m}\E_n(R)\cdot\E_N(R,A)$ to lie in its commutator subgroup, which will be denoted by $F$. The proof goes in two steps.
\begin{itemize}
\item For the transvections $\bw{m}t_{i,j}(\zeta)$ this follows from the Cauchy--Binet homomorphism:
$$\bw{m}t_{i,j}(\zeta)=\bw{m}([t_{i,h}(\zeta),t_{h,j}(1)])=\left[\bw{m}t_{i,h}(\zeta),\bw{m}t_{h,j}(1)\right].$$
\item For elementary transvections $t_{I,J}(\xi)$ this can be done as follows. Suppose that $I\cap J=K = k_1\dots k_p$, where $0\leq p\leq m-1$, i.\,e., $I=k_1\dots k_pi_1\dots i_q$ and $J=k_1\dots k_pj_1\dots j_q$. As in Lemma~\ref{lem:AlterRelatForm}, we define a set $V=k_1\dots k_{p}j_1\dots j_{q-1}i_q$. And then
$$t_{I,J}(\xi)=[t_{I,V}(\xi),t_{V,J}(1)]=\left[t_{I,V}(\xi), \bw{m}t_{i_q,j_q}(\pm 1) \right],$$
so we get the required.
\end{itemize}
\end{proof}
Let, as above, $A\trianglelefteq R$, and let $R/A$ be the factor-ring of $R$ modulo $A$. Denote by $\rho_A: R\longrightarrow R/A$ the canonical projection sending $\lambda\in R$ to $\bar{\lambda}=\lambda+A\in R/A$. Applying the projection to all entries of a matrix, we get the reduction homomorphism
$$\begin{array}{rcl}
\rho_{A}:\GL_{n}(R)&\longrightarrow& \GL_{n}(R/A)\\
a &\mapsto& \overline{a}=(\overline{a}_{i,j})
\end{array}$$
The kernel of the homomorphism $\rho_{A}$, $\GL_{n}(R,A)$, is called the \textit{principal congruence subgroup in $\GL_{n}(R)$ of level $A$}. Now, let $\CC_n(R)$ be the center of the group $\GL_{n}(R)$, consisting of the scalar matrices
$\lambda e, \lambda \in R^{*}$. The full preimage of the center of $\GL_{n}(R/A)$, denoted by $\CC_n(R,A)$, is called the \textit{full congruence subgroup of level $A$}. The group $\CC_n(R,A)$ consists of all matrices congruent to a scalar matrix modulo $A$. We further concentrate on a study of the full preimage of the group $\bw{m}\GL_{n}(R/A)$:
$$\CC\bw{m}\GL_{n}(R,A)=\rho_{A}^{-1}\bigl(\bw{m}\GL_{n}(R/A)\bigr).$$
A key point in a reduction modulo an ideal is the following standard commutator formula, proved by Leonid Vaserstein~\cite{VasersteinGLn}, Zenon Borevich, and Nikolai Vavilov~\cite{BV84}.
$$\bigl[\E_n(R),\CC(n,R,A)\bigr]=\E_n(R,A) \textrm{ for a commutative ring }R \textrm{ and } n\geq 3.$$
Finally, we are ready to state the \textit{level reduction} result.
\LevelReduction*
\begin{proof}
In the proof by $N$ we mean $N_{\GL_{N}(R)}$.
Since $\E_N(R,A)$ and $\GL_{N}(R,A)$ are normal subgroups in $\GL_{N}(R)$, we see
$$N\bigl(\underbrace{\E\bw{m}\E_n(R,A)}_{=\bw{m}\E_n(R)\E_N(R,A)}\bigr)\leq N\bigl(\E\bw{m}\E_n(R,A)\GL_{N}(R,A)\bigr)=\CC\bw{m}\GL_{n}(R,A). \eqno(1)$$
Note that the latter equality is due to the normalizer functoriality:
$$N\left(\E\bw{m}\E_n(R,A)\GL_{N}(R,A)\right)=N\left(\rho_A^{-1}\left(\bw{m}\E_n(R/A)\right)\right)=\rho_A^{-1}\left(N \left(\bw{m}\E_n(R/A)\right)\right)=\rho_A^{-1}\left(\bw{m}\GL_{n}(R/A)\right).$$
In particular, using $(1)$, we get
$$\left[\CC\bw{m}\GL_{n}(R,A),\E\bw{m}\E_n(R,A)\right]\leq \E\bw{m}\E_n(R,A)\GL_{N}(R,A).\eqno(2)$$
On the other hand, it is completely clear $\E\bw{m}\E_n(R,A)$ to be normal in the right-hand side subgroup. Indeed, it is easy to prove the following stronger inclusion:
$$\left[\bw{m}\GL_{n}(R)\GL_{N}(R,A),\E\bw{m}\E_n(R,A)\right]\leq \E\bw{m}\E_n(R,A).\eqno(3)$$
To check this, we consider a commutator of the form
$$[xy,hg],\qquad x\in \bw{m}\GL_{n}(R), y\in \GL_{N}(R,A), h\in \bw{m}\E_n(R), g\in \E_N(R,A).$$
Then $[xy, hg]={}^x[y,h] \cdot [x,h] \cdot {}^h[xy,g]$. We need to prove all factors on the right-hand side to belong to $\E\bw{m}\E_n(R,A)$. Right away, the second factor lies in the group $\E\bw{m}\E_n(R,A)$. For the first commutator, we should consider the following inclusions:
$${}^{\bw{m}\GL_{n}(R)}\left[\GL_{N}(R,A),\bw{m}\E_n(R)\right]\leq \Bigl[
\underbrace{{}^{\bw{m}\GL_{n}(R)}\GL_{N}(R,A)}_{=\GL_{N}(R,A)},\underbrace{{}^{\bw{m}\GL_{n}(R)}\bw{m}\E_n(R)}_{=\bw{m}\E_n(R)}\Bigr]\leq \E\bw{m}\E_n(R,A).$$
The element $h\in \bw{m}\E_n(R)$, so we ignore it in conjugation. The third commutator lies in $\E\bw{m}\E_n(R,A)$ due to the following inclusion.
$$\left[\bw{m}\GL_{n}(R)\GL_{N}(R,A),\E_N(R,A)\right]\leq\left[\GL_{N}(R),\E_N(R,A)\right]=\E_N(R,A).$$
Now if we recall $(2)$ and $(3)$, we get
$$\left[\CC\bw{m}\GL_{n}(R,A),\E\bw{m}\E_n(R,A),\E\bw{m}\E_n(R,A)\right]\leq \E\bw{m}\E_n(R,A).\eqno(4)$$
To invoke the Hall–-Witt identity, we need a slightly more precise version of the latter
inclusion:
$$\left[\left[\CC\bw{m}\GL_{n}(R,A),\E\bw{m}\E_n(R,A)\right],\left[\CC\bw{m}\GL_{n}(R,A),\E\bw{m}\E_n(R,A)\right]\right]\leq \E\bw{m}\E_n(R,A).\eqno(5)$$
Observe that by formula $(2)$ we have already checked the left-hand side to be generated by the commutators
of the form
$$[uv,[z,y]], \text{ where } u,y\in \E\bw{m}\E_n(R,A), v\in \GL_N(R,A), z\in \CC\bw{m}\GL_n(R,A).$$
However,
$$[uv,[z,y]]={}^u[v,[z,y]]\cdot[u,[z,y]],$$
By formula $(4)$ the second commutator belongs to $\E\bw{m}\E_n(R,A)$, whereas by $(5)$ the first is an element of $\left[\GL_{N}(R,A),\E_N(R)\right]\leq \E_N(R,A)$.
Now we are ready to finish the proof. By the previous lemma, the group $\E\bw{m}\E_n(R,A)$
is perfect, and thus, it suffices to show $[z,[x,y]]\in \E\bw{m}\E_n(R,A)$ for all $x,y\in \E\bw{m}\E_n(R,A), z\in \CC\bw{m}\GL_{n}(R,A)$. Indeed, the Hall–-Witt identity yields
$$[z,[x,y]]={}^{xz}[[z^{-1},x^{-1}],y]\cdot{}^{xy}[[y^{-1},z],x^{-1}],$$
where the second commutator belongs to $\E\bw{m}\E_n(R,A)$ by $(4)$.
Removing the conjugation by $x\in \E\bw{m}\E_n(R,A)$ in the first commutator and carrying the conjugation by $z$ inside
the commutator, we see that it only remains to prove the relation $[[x^{-1},z],[z,y]y]\in \E\bw{m}\E_n(R,A)$. Indeed,
$$[[x^{-1},z],[z,y]y]=[[x^{-1},z],[z,y]]\cdot{}^{[z,y]}[[x^{-1},z],y],$$
where both commutators on the right--hand side belong to $\E\bw{m}\E_n(R,A)$ by formulas $(4)$ and $(5)$, and moreover,
the conjugating element $[z,y]$ in the second commutator is an element of the group $\E\bw{m}\E_n(R,A)\GL_{N}(R,A)$, and thus by $(3)$, normalizes $\E\bw{m}\E_n(R,A)$.
\end{proof}
\bibliographystyle{unsrt}
|
\section{Introduction}
When speakers of two different languages can effectively communicate each in their own tongue, these languages are considered \textit{mutually intelligible}.
Intelligibility is often assymetric, and considered to be a continuous notion \cite{gooskens}, with some languages exhibiting higher intelligibility (e.g. Bulgarian and Macedonian) while others are only partially understood in the oral form (German and Yiddish) or written media (Russian and Ukrainian).
Does mutual intelligibility, as perceived by humans, make translation an easier task for machines to learn?
We hypothesize that mutually intelligible language pairs require less bitext data to train a neural machine translation model from one language to the other.
To test our hypothesis, we train machine translation models on varying amounts of data for language pairs in the Romance language group (Spanish, Portuguese, Italian, French, and Romanian).
For each language pair, we produce a learning curve that measures the relative BLEU score (with respect to the model trained on 100\% of the data), and compute the area under the curve (AUC), which we interpret as how ``easy'' it is for a machine to learn to translate this particular language pair.
We compare the AUC with two intelligibility scores: an automatically-computed metric of language similarity based on lexical analysis \cite{dinu-ciobanu-2014-romance}, and a human measure of intelligibility derived from a spoken cloze test \cite{gooskens}.
The lexical metric shows high correlation with the AUC (Pearson's $r = 0.539$), while the spoken-language metric exhibits lower correlation ($r = 0.224$).
Further investigation reveals that due to cultural exposure, Romanian speakers find Italo-Western languages intelligible, but not vice versa.
Treating these data points as outliers reveals strong correlation between the spoken-language metric and the AUC ($r = 0.585$),
indicating that translating between languages that humans perceive to be mutually intelligible may indeed be easier.
\section{Measuring Mutual Intelligibility}
\label{sec:background}
There are various methods in applied linguistics for testing to what degree a speaker of one language (source) can understand utterances from a related language (target).
In this work, we focus on five Romance languages (Spanish, Portuguese, Italian, French, and Romanian) and rely on two different approaches for measuring their mutual intelligibility.
\citet{dinu-ciobanu-2014-romance} use automatic lexical analysis to find etymons and cognates (i.e. words that are genetically related) to measure language similarity.
Table~\ref{tab:dinu} shows their intelligibility scores computed over the Europarl corpus \cite{europarl}.
We use this metric as a proxy for written mutual intelligibility, with the significant caveat that it is not based on a study with human participants.
For a more accurate measure of human-perceived mutual intelligibility, we turn to a recent study by \citet{gooskens}.
Here, spoken language intelligibility is measured using a cloze test, in which selected words are removed from a target language utterance and replaced by beeps of uniform length.
The source language speaker must then fill in the missing words; source to target intelligibility is thus how \textit{accurately} the average subject does so.
Since previous exposure to the target language via education, culture, and media can significantly skew the metric, Gooskens et al. select participants with minimal exposure to the target language a priori.
Table~\ref{tab:gooskens} shows their spoken language intelligibility scores among the five Romance languages.
\begin{table}[t]
\small
\centering
\begin{tabular}{lccccc}
\toprule
\textbf{src{\textbackslash}tgt} & \textbf{es} & \textbf{pt} & \textbf{it} & \textbf{fr} & \textbf{ro} \\
\midrule
\textbf{es} & --- & 86.40 & 77.37 & 68.86 & 59.97 \\
\textbf{pt} & 84.56 & --- & 75.33 & 67.11 & 64.54 \\
\textbf{it} & 79.45 & 75.65 & --- & 68.82 & 65.97 \\
\textbf{fr} & 69.98 & 63.15 & 72.93 & --- & 65.12 \\
\textbf{ro} & 70.17 & 65.57 & 69.15 & 62.27 & --- \\
\bottomrule
\end{tabular}
\caption{Written language intelligibility, based on automatic lexical analysis \cite{dinu-ciobanu-2014-romance}.}
\label{tab:dinu}
\end{table}
\begin{table}[t]
\small
\centering
\begin{tabular}{lccccc}
\toprule
\textbf{src{\textbackslash}tgt} & \textbf{es} & \textbf{pt} & \textbf{it} & \textbf{fr} & \textbf{ro} \\
\midrule
\textbf{es} & --- & 35.7 & 38.2 & 28.2 & 13.7 \\
\textbf{pt} & 62.0 & --- & 44.1 & 34.3 & 14.7 \\
\textbf{it} & 56.0 & 23.4 & --- & 18.6 & ~~8.7 \\
\textbf{fr} & 31.5 & 23.5 & 22.9 & --- & 11.0 \\
\textbf{ro} & 46.6 & 20.7 & 47.2 & 47.1 & --- \\
\bottomrule
\end{tabular}
\caption{Spoken language intelligibility, based on human subject spoken cloze tests \cite{gooskens}.}
\label{tab:gooskens}
\end{table}
\section{Experiment}
We hypothesize that mutually intelligible language pairs require less data to train machine translation models than language pairs that exhibit less intelligibility among human speakers.
To test our hypothesis, we sample the training set of each language pair to create training subsets of various sizes, and use these training subsets to create a learning curve for each language pair.
We then compare the area under the learning curves (AUC) to mutual intelligibility scores (Table~\ref{tab:dinu} and Table~\ref{tab:gooskens}); if the hypothesis is correct, we expect to observe correlation between AUC and intelligibility scores.
\subsection{Setup}
\paragraph{Data}
To leverage existing research on mutual intelligibility (Section~\ref{sec:background}), we focus on translation among five Romance languages: Spanish (es), Portuguese (pt), Italian (it), French (fr), and Romanian (ro).
We use the IWSLT'14 shared task \cite{cettolo2014report}, which consists of TED talk subtitles translated from the original spoken English to many other languages, to construct our training data.
Specifically, we pivot through English, enumerating over each sentence in the source language, and matching its translation in the target language by searching for the identical pivot sentence in English.
This process creates equivalent parallel data for all 20 language pairs.
\paragraph{Model and Hyperparameters}
We use fairseq \cite{ott2019fairseq} to train a small transformer model \cite{transformers} with 6 layers of encoder/decoder, 256 model dimensions, 1024 hidden dimensions, and 4 attention heads per layer.
Each model was trained for 50 epochs with dynamic batches of 8000 max tokens on a single 12GB GPU.
We use the Adam optimizer with inverse square root learning rate scheduling, warming up for 4000 steps and peaking at a learning rate of 0.0005, and regularize with 0.3 dropout, smoothed cross entropy loss, weight decay of 0.0001.
The best checkpoint is selected via validation set loss.
During evaluation, we decode with beam search (beam 5) and evaluate with SacreBLEU \cite{post-2018-call}.
\subsection{Measuring Translation Difficulty}
We describe visual and quantitative means for measuring ``learnability'' by plotting the relative performance (BLEU score) as a function of the number of examples provided at training time.
\paragraph{Visual Measure: Learning Curve}
We subsample the original training data to create smaller training subsets, ranging from 20\% of to the original data to 100\%, in increments of 10\%.
For each subset, we train a model and record its BLEU score on the test data (which was not modified).
Since comparing BLEU scores between different language pairs is not informative, we compute the \textit{relative} BLEU score by dividing each value by the score achieved by the model trained on 100\% of the training data.\footnote{For example, if a model trained on half the data achieved 24 BLEU while a model trained on all the data achieved 30 BLEU, their relative performance measure would be 0.8 for the half-data model and 1.0 for the full-data model.}
Using relative performance to plot learning curves allows us to compare learning curves of different language pairs.
Figure~\ref{fig:es_to_all} presents learning curves for translation from Spanish (es) to other Romance languages.
We observe that generally, the Spanish - Portuguese pair learns quickest (i.e., reaches high relative BLEU score with smallest percentage of data used). We can also observe that the initial relative BLEU scores are in the same order as in the scores in both the lexical analysis (Table \ref{tab:dinu}) and spoken analysis (Table \ref{tab:gooskens}). This congruence in ordering is not preserved entirely as percentage of data increases, but is the most prevalent ordering.
\begin{figure}
\centering
\includegraphics[width=\columnwidth,height=\textheight,keepaspectratio]{es_to_all.pdf}
\caption{Machine translation learning curves from Spanish (es) to other Romance languages. The horizontal axis reflects the percentage of training data used. The vertical axis is the BLEU score on the test set, normalized against the BLEU score of the full-data model.}
\label{fig:es_to_all}
\end{figure}
\paragraph{Quantitative Measure: Area Under Curve}
The learning curves visualize how fast the performance saturates when the model is provided with more training data.
For a numerical measure of this phenomenon, we compute the area under the curve (AUC) for each language pair, using the trapezoidal rule.
We interpret AUC as our measure for how ``easy'' it is for a machine to learn to translate a particular language pair.
Table~\ref{tab:auc} shows the AUC score for every language pair.
We observe that the lowest 40\% of the scores involve Romanian in the role of either source or target language.
Languages with high mutual intelligibility on the cloze test, such as Spanish and Portuguese also achieve the highest AUC.
From a manual comparison, these scores seem to align well with the intelligibility scores (Table~\ref{tab:dinu} and Table~\ref{tab:gooskens}), but a more rigorous method is needed to determine how strong this correlation really is.
\begin{table}[t!]
\small
\centering
\begin{tabular}{lccccc}
\toprule
\textbf{src{\textbackslash}tgt} & \textbf{es} & \textbf{pt} & \textbf{it} & \textbf{fr} & \textbf{ro} \\
\midrule
\textbf{es} & --- & 74.71 & 73.74 & 73.63 & 71.77 \\
\textbf{pt} & 75.12 & --- & 72.32 & 72.83 & 70.05 \\
\textbf{it} & 74.34 & 73.32 & --- & 72.69 & 71.71 \\
\textbf{fr} & 72.89 & 72.78 & 72.54 & --- & 69.94 \\
\textbf{ro} & 71.87 & 70.67 & 71.48 & 70.48 & --- \\
\bottomrule
\end{tabular}
\caption{Area under curve (AUC) of machine translation learning curves. More area implies saturation with less data.}
\label{tab:auc}
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{auc_dinu.pdf}
\caption{Correlation between the area under the learning curve (AUC) and written-language mutual intelligibility based on automatic lexical analysis \cite{dinu-ciobanu-2014-romance}. Pearson's $r = 0.539$. Each color and shape of a data point indicate the data point's source language, both in the intelligibility analysis and the translation model.}
\label{fig:auc_dinu}
\end{figure*}
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{auc_gooskens.pdf}
\caption{Correlation between the area under the learning curve (AUC) and spoken-language mutual intelligibility based on cloze tests with human participants \cite{gooskens}. Datapoints with Romanian source (orange pentagons) appear to be outliers to the general trend. Pearson's correlation with full data: $r = 0.224$. Without Romanian as a source language: $r = 0.585$. Each color and shape of a data point indicate the data point's source language, both in the intelligibility analysis and the translation model.}
\label{fig:auc_gooskens}
\end{figure*}
\subsection{Correlation with Mutual Intelligibility}
We plot all 20 language pairs using AUC as the horizontal axis and the intelligibility scores presented in mutual intelligibility studies (Section~\ref{sec:background}) as the vertical axis.
This allows us to visually compare the two metrics, and determine whether a trend exists.
In addition, we compute Pearson's correlation between each pair of metrics to provide a quantitative measure of the correlation.
\paragraph{Written Language Intelligibility}
Figure~\ref{fig:auc_dinu} presents the correlation between AUC scores (Table~\ref{tab:auc}) and written-language intelligibility scores based on lexical analysis (Table~\ref{tab:dinu}).
We observe a strong linear trend with no substantial outliers, which also matches the computed correlation of $r = 0.539$.
While this finding strengthens our hypothesis, it does not completely tease apart mutual intelligibility from language relatedness; we thus compare the AUC scores to an additional intelligibility metric based on human participant data.
\paragraph{Spoken Language Intelligibility}
Figure~\ref{fig:auc_gooskens} shows the correlation between the AUC scores and the spoken-language mutual intelligibility based on cloze tests.
The correlation is modest ($r = 0.224$), possibly due to the different medium (speech versus text).
However, a closer look at the figure brings to attention three datapoints in the top left quadrant, where Romanian is the source language (and the target language is French, Italian, Spanish, from left to right).
These points seem to be outliers to an otherwise strong linear trend.
One possible explanation is that Romanians learn French at school, and it is therefore difficult to find Romanian participants with minimal exposure to other languages in the Romance language group \cite{gooskens}.
When removing Romanian as a source language from the data, we see a much stronger correlation coefficient of $r = 0.585$, similar to the correlation observed with the written-language metric.
\section{Conclusions and Future Work}
This work presents the hypothesis that mutually intelligible language pairs are easier for neural machine translation models to learn, in the sense that they require less training data.
We test the hypothesis by computing the correlation between a proxy measure for learnability (AUC) and pre-existing metrics of mutual intelligibility.
Experiments on five languages in the Romance language group show strong correlation, thus supporting the hypothesis.
To the best of our knowledge, this is the first use of our proposed methodology.
Future work may be able to utilize it to conduct a more extensive investigation of how mutual intelligibility affects machine translation by exploring more language pairs and groups, examine correlations with additional intelligibility metrics, and use true bitext that does not pivot through English.
Our findings might also suggest that there are potential opportunities in the intersection of few-shot learning and cross-lingual transfer between mutually intelligible languages.
\section{Results}
\omer{We will probably remove all of this section}
\subsection{Romance}
Table \ref{table:auc_romance} shows the area under curve (AUC) for each language pair. By looking at the area under curve, we gain a quantifiable insight in to the performance of the machine on a given language pair. The highest AUC score is given to the language pair of highest mutual intelligibility, Italian - Spanish. The correlation between relative BLEU scores and mutual intelligibility score is not significant (Pearson's r = 0.43, p-value = 0.16)
Figure \ref{fig:average_romance} shows the increasing average relative BLEU score of two groups: those who are considered to be of high and low intelligibility. It can be seen that the high intelligibility group achieves higher relative BLEU scores at the start of the process. Table \ref{table:romance1} shows the raw BLEU results on the test set, for the Roman languages, for all data sizes.
Some language pairs are congruent with our hypothesis, for example, Figure \ref{fig:romance_cherry} shows the relative BLEU score in each subset of the data for two pairs of languages- Italian Spanish and Spanish-Romanian. It can be seen that the Spanish - Italian pair reaches a level close to saturation after using only 70\% of the data, which is consistent with their high intelligibility, as opposed to Spanish - Romanian, which saturates at 80\% of the data. We define saturation point as percentage of data which surpasses the 0.95 relative BLEU score
\begin{table*}[t]
\small
\centering
\begin{tabular}{
>{\columncolor[HTML]{EFEFEF}}l l}
{\color[HTML]{000000} ES FR} & 73.632 \\
{\color[HTML]{000000} \textbf{ES IT}} & 73.742 \\
{\color[HTML]{000000} \textbf{RO FR}} & 70.478 \\
{\color[HTML]{000000} \textbf{RO ES}} & 71.871 \\
{\color[HTML]{000000} \textbf{RO IT}} & 71.477 \\
{\color[HTML]{000000} \textbf{IT ES}} & 74.342 \\
{\color[HTML]{000000} \textbf{IT FR}} & 72.688 \\
{\color[HTML]{000000} ES RO} & 71.766 \\
{\color[HTML]{000000} FR IT} & 72.543 \\
{\color[HTML]{000000} FR ES} & 72.891 \\
{\color[HTML]{000000} FR RO} & 69.939 \\
{\color[HTML]{000000} IT RO} & 71.714
\end{tabular}
\caption{AUC score per Romance language pair. In each row, first language is source and last language is destination. Marked in bold are language pairs which are of high intelligibility.}
\label{table:auc_romance}
\end{table*}
\begin{table*}[t]
\small
\centering
\begin{tabular}{
>{\columncolor[HTML]{EFEFEF}}l l}
NL ES & 68.555 \\
ES NL & 68.954 \\
FR DE & 67.309 \\
DE FR & 67.802
\end{tabular}
\caption{AUC score of language pairs between language families (Romance and Germanic). In each row, first language is source and last language is destination.}
\label{table:auc_both}
\end{table*}
\begin{table*}[t]
\small
\centering
\begin{tabular}{lllllllllllll}
\rowcolor[HTML]{EFEFEF}
&
\textbf{ES IT} &
\textbf{ES FR} &
\textbf{ES RO} &
\textbf{IT ES} &
\textbf{IT FR} &
\textbf{IT RO} &
\textbf{FR IT} &
\textbf{FR ES} &
\textbf{FR RO} &
\textbf{RO ES} &
\textbf{RO IT} &
\textbf{RO FR} \\
\cellcolor[HTML]{EFEFEF}\textbf{0.2} & 16.21 & 19.45 & 10.82 & 17.12 & 19.45 & 15.7 & 15.32 & 16.67 & 10.13 & 14.02 & 11.2 & 23.81 \\
\cellcolor[HTML]{EFEFEF}\textbf{0.3} & 19.49 & 23.37 & 14.72 & 20.33 & 23.37 & 17.45 & 19.29 & 21.34 & 14.07 & 18.3 & 15.51 & 26.14 \\
\cellcolor[HTML]{EFEFEF}\textbf{0.4} & 20.82 & 25.33 & 16.07 & 22.74 & 25.33 & 18.73 & 20.72 & 23.12 & 15.96 & 20.3 & 17.4 & 27.92 \\
\cellcolor[HTML]{EFEFEF}\textbf{0.5} & 22.09 & 26.5 & 17.24 & 23.42 & 25.77 & 19.1 & 21.94 & 24.16 & 16.67 & 21.41 & 18.19 & 28.61 \\
\cellcolor[HTML]{EFEFEF}\textbf{0.6} & 22.1 & 27.39 & 17.6 & 23.69 & 26.82 & 19.14 & 22.64 & 24.55 & 17.45 & 22.34 & 19.02 & 29 \\
\cellcolor[HTML]{EFEFEF}\textbf{0.7} & 23.34 & 27.72 & 17.94 & 24.72 & 26.91 & 19.85 & 22.87 & 24.91 & 18 & 22.64 & 19.63 & 29.51 \\
\cellcolor[HTML]{EFEFEF}\textbf{0.8} & 23.48 & 28.56 & 18.72 & 24.88 & 27.77 & 19.67 & 22.92 & 25.59 & 18.59 & 23.51 & 20.18 & 29.89 \\
\cellcolor[HTML]{EFEFEF}\textbf{0.9} & 24.04 & 29.01 & 18.87 & 24.92 & 27.92 & 20.12 & 23.68 & 26.01 & 19.01 & 23.5 & 20.49 & 30.12 \\
\cellcolor[HTML]{EFEFEF}\textbf{1} & 23.92 & 29.33 & 19.31 & 25.16 & 28.37 & 20.09 & 23.87 & 26.25 & 19.06 & 24.35 & 21.14 & 30.01 \\
\cellcolor[HTML]{EFEFEF}\textbf{Total Examples} &
331236 &
312182 &
326634 &
331196 &
310982 &
653202 &
310942 &
312134 &
304960 &
326630 &
326592 &
609974
\end{tabular}
\caption{The relative BLEU score as a function of the dataset size for the Romance languages. }
\label{table:romance1}
\end{table*}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig2.jpeg}
\caption{Average performance of languages considered to be high or low intelligibility, out of the Romance language family. Performance is measured as the BLEU score evaluated on the test set of a model trained on a given subset for 50 epochs, divided by the BLEU score of a model trained on 100\% of the data.}
\label{fig:average_romance}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig3.jpeg}
\caption{Performance of two pairs of languages: Spanish-Romanian and Italian-Spanish. Italian Spanish is highly intelligible while Spanish Romanian is not. Performance is measured as the BLEU score evaluated on the test set of a model trained on a given subset for 50 epochs, divided by the BLEU score of a model trained on 100\% of the data. It can be seen that the intelligible pair scores better for smaller data sizes compared to the unintelligible pair.}
\label{fig:romance_cherry}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig3.jpeg}
\caption{Average relative BLEU results as a function of the data set size, split to the high and low mutual intelligibility in green and red respectively }
\label{fig:german_avg}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig6.jpeg}
\caption{Performance of two pairs of languages: French-Spanish and Spanish-French to explore the a-symmetrical nature of mutual intelligibility. French - Spanish achieve a higher relative BLEU score consistently while also being more mutually intelligible. Performance is measured as the BLEU score evaluated on the test set of a model trained on a given subset for 50 epochs, divided by the BLEU score of a model trained on 100\% of the data. It can be seen that the intelligible pair scores better for smaller data sizes compared to the unintelligible pair.}
\label{fig:symmetry}
\end{figure}
\subsection{Germanic}
Considering the Germanic languages, most of the results must be looked at with a higher degree of skepticism, because of the inherent English exposure and the fact that English is the base for all language pairs, meaning each pair containing English as a source or target language was not pivoted. With this caveat in mind, table \ref{table:auc_germanic} shows the AUC scores for the Germanic language group. The correlation is even weaker (Pearson's r = -0.16, p-value=0.75).
Table \ref{table:germanic} shows the relative BLEU results. For clarity, we will not present the plot of all language pairs. Figure \ref{fig:german_avg} shows the average relative BLEU scores of the high and low intelligibility groups in the Germanic languages. Unlike the Romance, here we do not see a clear advantage to the high intelligibility pairs. A plausible explanation might be that high intelligibility pairs consist mainly of English, because it is measured by people who speak the language, unlike the NMT which does not have this prior knowledge.
\begin{table}[t]
\small
\centering
\begin{tabular}{
>{\columncolor[HTML]{EFEFEF}}l l}
EN NL & 73.223 \\
\textbf{NL DE} & 72.875 \\
\textbf{DE EN} & 71.801 \\
\textbf{NL EN} & 73.578 \\
DE NL & 72.666 \\
EN DE & 72.826
\end{tabular}
\caption{AUC score per Germanic language pair. In each row, first language is source and last language is destination. Marked in bold are language pairs which are of high intelligibility.}
\label{table:auc_germanic}
\end{table}
\begin{table*}[t]
\small %
\centering
\begin{tabular}{
>{\columncolor[HTML]{EFEFEF}}l llllll}
\textbf{} &
\cellcolor[HTML]{EFEFEF}\textbf{EN NL} &
\cellcolor[HTML]{EFEFEF}\textbf{EN DE} &
\cellcolor[HTML]{EFEFEF}\textbf{NL EN} &
\cellcolor[HTML]{EFEFEF}\textbf{NL DE} &
\cellcolor[HTML]{EFEFEF}\textbf{DE EN} &
\cellcolor[HTML]{EFEFEF}\textbf{DE NL} \\
\textbf{0.2} & 22.40 & 18.30 & 26.12 & 14.74 & 22.74 & 20.74 \\
\textbf{0.3} & 26.47 & 22.20 & 30.11 & 18.60 & 26.37 & 22.84 \\
\textbf{0.4} & 28.31 & 24.10 & 32.49 & 20.19 & 29.01 & 23.70 \\
\textbf{0.5} & 29.14 & 25.00 & 34.24 & 21.23 & 30.07 & 24.12 \\
\textbf{0.6} & 29.99 & 25.93 & 34.11 & 21.81 & 31.91 & 24.24 \\
\textbf{0.7} & 30.09 & 26.05 & 34.88 & 21.84 & 32.10 & 24.44 \\
\textbf{0.8} & 31.00 & 26.91 & 35.93 & 22.75 & 32.97 & 24.61 \\
\textbf{0.9} & 31.50 & 27.52 & 36.52 & 23.07 & 33.39 & 24.70 \\
\textbf{100} & 31.91 & 27.55 & 36.65 & 23.11 & 34.01 & 24.73 \\
\textbf{Total Examples} & 307180 & 320478 & 307180 & 291444 & 320478 & 583092
\end{tabular}%
\caption{BLEU results of the Germanic languages}
\label{table:germanic}
\end{table*}
\subsection{Cross Family}
We sampled two languages from each family: French and Spanish from Romance, German and Dutch from Germanic. We then trained a model for each cross family pair, which table \ref{table:auc_both} shows the AUC scores for. AUC scores for cross family translation is much lower, as the highest AUC score for cross family is 68.9 (ES-NL), with the lowest intra group AUC score being 69.9 (FR-RO), shown in table \ref{table:auc_romance}.
\subsection{Asymmetric intelligibility}
An interesting aspect of intelligibility is how a-symmetric it may be- this is very prominent in Table \ref{table:table1}, especially with the Romanian language; Romanian speakers can easily understand other Romance languages but not the other way around.
Figure \ref{fig:symmetry} shows this effect for the Italian and Spanish pair, as Italian-Spanish is more mutually intelligible than Spanish-Italian. It can be seen that Italian Spanish achieves a relative BLEU score of around 92\% on 40\% of the data, while Spanish-Italian requires 50\% to achieve similar results.
|
\section{Introduction}
\label{sec:intro}
Our previous work in~\citet{Sutinjo_AA2021} (hereafter referred to as Paper~I) discussed a formulation for the sensitivity in terms of System Equivalent Flux Density (SEFD) for a polarimetric interferometer that consists of dual polarized antennas. In that work, we provided an example using a dual-polarized MWA dipole element embedded in an array. In this paper, we extend and generalize the SEFD formulation to polarimetric phased arrays interferometers. This is an important generalization as it is directly applicable to low-frequency interferometric phased array telescopes in operation such as the Murchison Widefield Array (MWA)~\citep{2013PASA...30....7T} and LOFAR~\citep{2013A&A...556A...2V}, as well as future Low Frequency Square Kilometre Array (SKA-Low)~\citep{8105424, 7928622}. In particular regarding the SKA-Low, a clear conceptual understanding of array sensitivity, how it varies over telescope pointing angles, and how to calculate it are crucial for validation against SKA sensitivity requirements~\citep{SKA1L1req}.
The reasons for using SEFD as the valid figure of merit (FoM) for a polarimetric radio interferometer as opposed to antenna effective area on system temperature ($A_e/T_{sys}$) was thoroughly reviewed in Paper I. However for convenience, we mention the key ideas here. The primary reason is that antenna effective area, $A_e$, is a number that is defined as matched to the polarization of the incident wave, which is not known in advance for a polarimeter. In contrast, the concept of equivalent flux density is not constrained to the polarization state and it can be readily equated to the system noise. The work in Paper~I allowed us to demonstrate that the often used conversion between $A_e/T_{sys}$ and SEFD is only an approximation that is valid in certain cases in which the Jones matrix is diagonal or anti-diagonal. This is further generalized in our current paper (see Sec.~\ref{sec:RMSunderestimates}), where we show that the SEFD approximation is correct only for row vectors in the Jones matrix (see Eq.~\eqref{eqn:J1}) that are orthogonal. When the row vectors are not orthogonal, the root-mean-square (RMS) SEFD approximation always underestimates the true SEFD.
Furthermore, in our work on this subject, we make an explicit connection to observational radio interferometry which is demonstrated by comparison to radio images. We also calculated and conducted a careful review of second-order noise statistics that form the basis for the SEFD formula. These are the main differences between our work and existing work in radio astronomy phased array sensitivity literature, for example~\citet{Warnick_6018280, Ellingson_5722983, Tokarsky_7987810, 7293140}. However, we acknowledge that there are aspects of our calculations that benefit from the pre-existing collective knowledge in the community, in particular regarding the computation of array response to an incident wave and array noise temperature as discussed in Sec.~\ref{sec:Jones_mat} and Sec.~\ref{sec:Sys_noise}.
The polarimetric phased array under consideration has N dual-polarized elements where each element is connected to a low noise amplifier (LNA), as is typical in practice. The LNA outputs are connected to the phased array weights and subsequently summed to produce the array output as shown in Fig.~\ref{fig:arrays}. We consider the array response to an incident electric field from a target direction, $\mathbf{e}_t$. The voltages that represent the response of Array~1 are
\begin{eqnarray}
\mathbf{v}_1|_t&=&\mathbf{J}\mathbf{e}_t, \nonumber \\
\left[ \begin{array}{c}
V_{X1}|_t \\
V_{Y1}|_t
\end{array} \right]
&=&
\left[ \begin{array}{cc}
l_{X\theta} & l_{X\phi} \\
l_{Y\theta} & l_{Y\phi}
\end{array} \right]
\left[ \begin{array}{c}
E_{t\theta} \\
E_{t\phi}
\end{array} \right],
\label{eqn:J1}
\end{eqnarray}
where $V_{X1}, V_{Y1}$ are the voltages measured by the $X, Y$ arrays, respectively, and $\mathbf{J}$ is the Jones matrix of the array; the elements of the Jones matrix are antenna lengths with units of meter that represent the response of the array to each polarization basis of the electric field, $E_{t\theta},E_{t\phi}$, with units of \si{\volt\per\metre}. There is a similar equation corresponding to Array~ 2. For antenna arrays of an identical design, it is reasonable to assume the same Jones matrix as Array~ 1.
\begin{figure}[t]
\begin{center}
\noindent
\includegraphics[width=3.5in]{figs/Four_elem_array_Et.png}
\caption{Arrays of dual-polarized low-frequency antennas observing the sky. The antennas are at fixed positions above the ground and the $z$-axis is up. The signal from each element is amplified by an LNA, weighted, and summed. The output of each summer is directed into the correlator. The $y$-directed elements (black) follow the same signal flow, but not shown for simplicity. \textcolor{black}{A four-element array is shown as an example.}}
\label{fig:arrays}
\end{center}
\end{figure}
Equation~\eqref{eqn:J1} is applicable to an array of antennas. We consider an array of dual-polarized antennas with outputs that are weighted and summed, then connected to a correlator as depicted in Fig.~\ref{fig:arrays}, thereby forming an interferometer. The antennas receive noise from the sky which includes a partially polarized target source $\bar{E}_t$ and a background with noise temperature distribution given by $T_{sky}$. For brevity, only four elements are shown per array. The extension to N elements is immediately evident. The LNAs, represented by the triangle gain blocks, produce their own noise which is scattered by the array and picked up by all elements in the array. Similarly, the incident sky signal arriving at the array undergoes the same scattering and coupling process. Therefore, the combined noise voltages seen at the LNA inputs are mutually correlated. We consider these noise sources and the interactions thereof to extend the SEFD formula to the interferometric and polarimetric phased array telescope.
This paper is organized as follows. The SEFD derivation and calculation are presented in Sec.~\ref{sec:SEFD_calc}. An example SEFD calculation procedure for the MWA is demonstrated in Sec.~\ref{sec:simulation}.
The SEFD calculation is validated by MWA observations described in Sec.~\ref{sec:obs} with comparison and discussion given in Sec.~\ref{sec:compare}. Concluding remarks are summarized in Sec.~\ref{sec:concl}. The appendix presents a detailed review of fundamental assumptions and statistical calculations that justify the SEFD formula presented in Sec.~\ref{sec:SEFD_calc}.
\section{SEFD of a polarimetric phased array interferometer}
\label{sec:SEFD_calc}
\subsection{Jones matrix of a dual-polarized array}
\label{sec:Jones_mat}
The calculations in Paper~I are reusable for an array but need reinterpretation and adaptation which we now discuss. The first key modification is to take the direction-dependent antenna lengths as that which are seen in the array environment, that is where each antenna element is connected to the input of an LNA. We call this quantity the antenna ``realized length'' \citep{Ung_9040892, Ung_MPhil2020} which is different from the open circuit antenna effective length in Paper~I.
The realized length has the advantage of having a clear and physical interpretation in an array environment depicted in Fig.~\ref{fig:ant_len_cal}. It is obtained by summing the embedded element realized lengths to form an overall equivalent realized length for the array. For example, for the X array,
\begin{eqnarray}
l_{X\theta}=\mathbf{w}_{x}^T\mathbf{l}_{x\theta}, \nonumber \\
l_{X\phi}=\mathbf{w}_{x}^T\mathbf{l}_{x\phi},
\label{eqn:array_lX}
\end{eqnarray}
where $\mathbf{w}_{x}^T=[w_{1x},...,w_{4x}]$ is the vector of weights and $\mathbf{l}_{x\theta}^T=[l_{1x\theta},...,l_{4x\theta}]$ is a vector containing the embedded element realized lengths, and similarly with $\mathbf{l}_{x\phi}$. The quantities $l_{1x\theta}$ and $l_{1x\phi}$ are obtained for a dual-polarized embedded element (number 1) with all other elements in the array terminated with the LNA input impedance, $Z_{LNA}$, as shown in Fig.~\ref{fig:ant_len_cal}. The embedded antenna realized length can be obtained through full-wave electromagnetic simulation~\citep{Ung_9040892} or measurement. Similarly, for the Y array
\begin{eqnarray}
l_{Y\theta}=\mathbf{w}_{y}^T\mathbf{l}_{y\theta}, \nonumber \\
l_{Y\phi}=\mathbf{w}_{y}^T\mathbf{l}_{y\phi},
\label{eqn:array_lY}
\end{eqnarray}
where $\mathbf{w}_{y}^T=[w_{1y},...,w_{4y}]$ is the vector of weights for the Y array, which may differ from $\mathbf{w}_{x}$. We note that Eq.~\eqref{eqn:array_lX} and Eq.~\eqref{eqn:array_lY} apply to the antenna realized length as we emphasized earlier, as opposed to the open-circuit length.
We assume LNAs of an identical design in this paper, and hence the LNA voltage gains are identical and need not be explicitly shown in Eq.~\eqref{eqn:J1}.
\begin{figure}[t]
\begin{center}
\noindent
\includegraphics[width=3in]{figs/Ant_lenght_calc_new.png}
\caption{Antenna length calculation for an array, where $l_{1x\theta}$ is the antenna length associated with $\hat{\theta}$-polarized incident field for $x$-directed antenna in the embedded element number 1 in the array shown, similarly with $l_{ix\theta}, l_{iy\theta}$, where $i$ is the embedded element number. \textcolor{black}{For brevity we consider a four-element array $i=1,\cdots,4$; the extension to $i=1,\cdots,N$ presents no complication.}}
\label{fig:ant_len_cal}
\end{center}
\end{figure}
Given the foregoing discussion regarding the realized length, then for a phased array, the voltages on the left hand side of Eq.~\eqref{eqn:J1} represent those at the outputs of the summer shown in Fig.~\ref{fig:arrays}. Therefore, the Jones matrix, $\mathbf{J}$, translates the incident field to the voltages at the outputs of the summer. This is similarly the case for Array 2 in Fig.~\ref{fig:arrays}. The entries of $\mathbf{J}$ represent the realized lengths of the array (for X or Y component) to an electric field basis ($\hat{\theta}$ or $\hat{\phi}$) such that each entry is a linear combination of the realized lengths of the embedded elements in the array. The Jones matrix then, is
\begin{eqnarray}
\mathbf{J}=\left[ \begin{array}{cc}
l_{X\theta} & l_{X\phi} \\
l_{Y\theta} & l_{Y\phi}
\end{array} \right]=
\left[ \begin{array}{cc}
\mathbf{w}_x^T \mathbf{l}_{x\theta} & \mathbf{w}_x^T \mathbf{l}_{x\phi} \\
\mathbf{w}_y^T \mathbf{l}_{y\theta} & \mathbf{w}_y^T \mathbf{l}_{y\phi}
\end{array} \right]
\label{eqn:J_array}
\end{eqnarray}
We adopt the uppercase ``X'' or ``Y'' to denote the overall array response and lower case ``$x$'' or ``$y$'' for the embedded elements in the array.
\subsection{SEFD expression}
\label{sec:SEFD exp}
The interferometric polarimeter estimates the outer product of the incident electric field in the target direction, which from Paper~I is given by
\begin{eqnarray}
\tilde{\mathbf{e}}\tilde{\mathbf{e}}^H=\mathbf{J}^{-1}\mathbf{v}_1\mathbf{v}_2^H\mathbf{J}^{-H},
\label{eqn:pol}
\end{eqnarray}
where $\mathbf{J}$ is defined by Eq.~\eqref{eqn:J_array} which represents the response of the array. The SEFD is proportional to the standard deviation of the flux estimate
\begin{eqnarray}
\text{SEFD}_I=\frac{\text{SDev}(\tilde{I})}{\eta_0}=\frac{\text{SDev}\left[(\tilde{\mathbf{e}}\tilde{\mathbf{e}}^H)_{1,1}+(\tilde{\mathbf{e}}\tilde{\mathbf{e}}^H)_{2,2}\right]}{\eta_0},
\label{eqn:SEFD_STDev}
\end{eqnarray}
where $\__{1,1}, \__{2,2}$ indicate the diagonal entries of the matrix in Eq.~\eqref{eqn:pol} (hence the subscript $_I$ that refers to Stokes I) and $\eta_0 \approx 120\pi~\Omega$ is the free space impedance. What is needed is the expression for the standard deviations (SDev) for the phased array interferometer, which we develop next.
\subsection{Array system noise}
\label{sec:Sys_noise}
We now compute the standard deviation of the electric field estimates as expressed in Eq.~\eqref{eqn:SEFD_STDev} due to the system noise. As shown in Paper~I, we expect to express the standard deviation in Eq.~\eqref{eqn:SEFD_STDev} in terms of the array realized lengths and the system noise temperatures, $T_{sys}$. Therefore, we need expressions for the mean square voltages for the X and Y arrays due to the system noise.
Fig.~\ref{fig:Tsys} depicts an array illuminated by an equivalent homogeneous black body environment at temperature $T_{sys}$. The value of $T_{sys}$ is that which equals the noise due to the actual diffuse background sky under observation, $T_{sky}$, plus the noise of the LNAs ($T_{rcv})$ in the array environment, and the ohmic loss which is a function of the radiation efficiency ($\eta_{rad}$) of the array. The $T_{sys}$ value for a phased array can be obtained through computation or measurements or a combination thereof using well-documented procedures. For example, computation for $T_{rcv}$ of an array was discussed in the literature~\citep{Warnick_5062509, Belo_7079488, warnick_maaskant_ivashina_davidson_jeffs_2018, Ung_8739942, Ung_9040892}. The computation of $T_{sys}$ for the MWA was demonstrated in~\citet{Ung_9040892, Ung_MPhil2020} and was validated by observation. The $T_{sys}$ value could also be obtained through measurements by carefully calibrating the array gain via known hot and cold sources, for example see~\citet{chippendale_hayman_hay_2014}.
\begin{figure}[t]
\begin{center}
\noindent
\includegraphics[width=3.5in]{figs/Tsys_diagram_new.png}
\caption{Diagrammatic representation of system noise. The array is surrounded by a homogeneous blackbody radiator at temperature $T_{sys}$ that represents the diffuse sky noise and the LNA noise. The $y$ elements are similarly connected to a separate weight and summing circuit that are not shown for brevity.}
\label{fig:Tsys}
\end{center}
\end{figure}
The $T_{sys}$ for the X and Y arrays are generally different; this is similarly noted in~Paper~I for a dual-polarized antenna element. This is expected because for the same sky under observation, the X and Y arrays produce different responses on the sky. Moreover, the $x$ elements and the $y$ elements interact differently depending on the array configuration and weights. Therefore, we anticipate assigning different $T_{sysX}$ and $T_{sysY}$ to the respective arrays, which we discuss next.
An equivalent homogeneous sky at $T_{sys}$ gives rise to partially correlated noise voltages at the antenna ports in the array. A convenient way to quantify this is to start with the case where each antenna port is terminated with an open circuit, that is $Z_{LNA}\rightarrow\infty$. In this case, the mutual coherence is known from the generalized Nyquist theorem~\citep{doi:10.1063/1.1722048, Hillbrand_1084200}
\begin{eqnarray}
\mathbf{C}_o|_{T_{sys}}=\left<\mathbf{v}_o\mathbf{v}_o^H\right>|_{T_{sys}}=4kT_{sys}\Delta f\Re [\mathbf{Z}],
\label{eqn:Re}
\label{eqn:twiss}
\end{eqnarray}
where $\Delta f$ is the noise bandwidth and $\Re [\mathbf{Z}]$ is the real part of the antenna impedance matrix, which is a real and symmetric matrix. We choose the port numbering convention such that
\begin{eqnarray}
\Re [\mathbf{Z}]=\Re
\left[\begin{array}{cc}
\mathbf{Z}_{xx} & \mathbf{Z}_{xy} \\
\mathbf{Z}_{yx} & \mathbf{Z}_{yy}
\end{array} \right],
\label{eqn:ReZ}
\end{eqnarray}
that is the $x_1,\cdots,x_4$ embedded elements ports are numbered 1 to 4 and $y_1,\cdots,y_4$ embedded elements ports are numbered 5 to 8, such that the open circuit voltage vector is $\mathbf{v}_o^T=[V_{ox1},\cdots,V_{ox4},V_{oy1},\cdots,V_{oy4}]$. The open circuit voltage vector is related to the voltage seen at the LNA inputs through a transformation matrix~\citep{Warnick_5062509}, which we call $\mathbf{T}$, such that
\begin{eqnarray}
\mathbf{v}=Z_{LNA}\left[Z_{LNA}\mathbf{I}+\mathbf{Z}\right]^{-1}\mathbf{v}_{o}=\mathbf{T}\mathbf{v}_{o}.
\label{eqn:v_voc}
\end{eqnarray}
Therefore, the coherence matrix of the voltages seen at the inputs of the LNAs loaded with $Z_{LNA}$ is
\begin{eqnarray}
\mathbf{T}\mathbf{C}_o|_{T_{sys}}\mathbf{T}^H&=&\mathbf{T}\left<\mathbf{v}_o\mathbf{v}_o^H\right>|_{T_{sys}}\mathbf{T}^H \nonumber \\
&=&4kT_{sys}\Delta f\mathbf{T}\Re [\mathbf{Z}]\mathbf{T}^H,
\label{eqn:coh_v_LNA}
\end{eqnarray}
which is a Hermitian matrix, since $(\mathbf{T}\Re [\mathbf{Z}]\mathbf{T}^H)^H=\mathbf{T}\Re [\mathbf{Z}]\mathbf{T}^H$.
The sought after quantities are the mean square voltages after weighting and summing
\begin{eqnarray}
\left<|V_{X}|^2\right>&=&[\mathbf{w}_{x}^T,\mathbf{0}_4^T] \mathbf{T}\mathbf{C}_o|_{T_{sysX}}\mathbf{T}^H
\left[\begin{array}{c}
\mathbf{w}_{x}^* \\ \mathbf{0}_4
\end{array} \right] \nonumber \\
&=&4kT_{sysX}\Delta f\mathbf{w}_{x0}^T\mathbf{T}\Re [\mathbf{Z}]\mathbf{T}^H\mathbf{w}_{x0}^*, \nonumber \\
\left<|V_{Y}|^2\right>&=&[\mathbf{0}_4^T,\mathbf{w}_{y}^T] \mathbf{T}\mathbf{C}_o|_{T_{sysY}}\mathbf{T}^H
\left[\begin{array}{c}
\mathbf{0}_4 \\ \mathbf{w}_{y}^*
\end{array} \right] \nonumber \\
&=&4kT_{sysY}\Delta f\mathbf{w}_{y0}^T\mathbf{T}\Re [\mathbf{Z}]\mathbf{T}^H\mathbf{w}_{y0}^*,
\label{eqn:ms_after_weight}
\end{eqnarray}
where $\mathbf{0}_4$ is a 4-long column vector of zeros; $\mathbf{w}_{x0}, \mathbf{w}_{y0}$ reflect the fact that the $x$ and $y$ antennas are summed separately; $T_{sysX}$ and $T_{sysY}$ are distinguished as discussed previously; $k$ is the Boltzmann constant and $\Delta f$ is the noise bandwidth. We know that Eq.~\eqref{eqn:ms_after_weight} is real because the eigenvalues of a Hermitian matrix are real.
For brevity in the SEFD formula, we adopt the following shorthand notations
\begin{eqnarray}
R_X=\mathbf{w}_{x0}^T\mathbf{T}\Re [\mathbf{Z}]\mathbf{T}^H\mathbf{w}_{x0}^*, \nonumber \\
R_Y=\mathbf{w}_{y0}^T\mathbf{T}\Re [\mathbf{Z}]\mathbf{T}^H\mathbf{w}_{y0}^*,
\label{eqn:R_noise}
\end{eqnarray}
which we can think of as the realized array noise resistances, with units of \si{\ohm}, representing the noise voltages at the outputs of the summers. A detailed explanation of this quantity and comparison to current literature can be found in Appendix \ref{apendx:links}. We note that Eq.~\eqref{eqn:ms_after_weight} transforms $T_{sysX}, T_{sysY}$ from a quantity external to the array to the mean square voltages after weighting and summing by the array. Therefore, the output of the summer is the reference plane at which the SEFD is to be calculated, as we show next.
\subsection{$\text{SEFD}_I$ formulation}
\label{sec:SEFD calc}
We assume all antenna arrays in question are of an identical design and the coupling between arrays is negligible due to the inter-array distance of tens of wavelength or larger. The intra-array coupling within each array is accounted for through the process described in Sec.~\ref{sec:Jones_mat} and Sec.~\ref{sec:Sys_noise}. This is the same level of assumptions as~Paper~I. The main difference between Paper~I and this work is that the sensitivity formula is now extended and generalized to the phased array. Following from~Paper~I, the Stokes~I estimate is $\tilde{I}=(\tilde{\mathbf{e}}\tilde{\mathbf{e}}^H)_{1,1}+(\tilde{\mathbf{e}}\tilde{\mathbf{e}}^H)_{2,2}$ from Eq.~\eqref{eqn:pol} and Eq.~\eqref{eqn:SEFD_STDev}. Expanding the matrix equation, we can write
\begin{eqnarray}
|D|^2\tilde{I}&=&X_1X_2^*\left(|l_{Y\phi}|^2+|l_{Y\theta}|^2\right) \nonumber \\
&+&Y_1Y_2^*\left(|l_{X\phi}|^2+|l_{X\theta}|^2\right) \nonumber \\
&-&X_1Y_2^*\left(l_{X\phi}^*l_{Y\phi}+l_{X\theta}^*l_{Y\theta}\right) \nonumber \\
&-&Y_1X_2^*\left(l_{X\phi}l_{Y\phi}^*+l_{X\theta}l_{Y\theta}^*\right),
\label{eqn:Isimp}
\end{eqnarray}
where $D=l_{X\theta}l_{Y\phi}-l_{X\phi}l_{Y\theta}$ is the determinant of the array Jones matrix; the leading $V$'s have been suppressed for brevity, for example $X_1$ is a complex random variable that refers to $V_{X1}$. The components of the antenna realized lengths (which are direction-dependent complex scalars) now refer to that of the array as discussed in Sec.~\ref{sec:Jones_mat}. Following the statistical calculation described in Appendix \ref{apendx:stat},
\begin{eqnarray}
\frac{|D|^4\text{Var}(\tilde{I})}{{(4k\Delta f)^2}
&=&\mathbf{t}_R^T \mathbb{L}\mathbf{t}_R,
\label{eqn:var_I_simp_array}
\end{eqnarray}
where
\begin{eqnarray}
\mathbf{t}_R=\left[ \begin{array}{c}
T_{sysX}R_X \\
T_{sysY}R_Y
\end{array} \right],
\label{eqn:tr}
\end{eqnarray}
is a column vector and the matrix
\begin{eqnarray}
\mathbb{L}=\left[ \begin{array}{cc}
\norm{\mathbf{l}_Y}^4 & |l_{X\phi}^*l_{Y\phi}+l_{X\theta}^*l_{Y\theta}|^2 \\
|l_{X\phi}^*l_{Y\phi}+l_{X\theta}^*l_{Y\theta}|^2 & \norm{\mathbf{l}_X}^4
\end{array} \right],
\label{eqn:Lcal}
\end{eqnarray}
where the vector norms are $\norm{\mathbf{l}_X}^2=|l_{X\theta}|^2+|l_{X\phi}|^2$ and similarly with $\norm{\mathbf{l}_Y}^2$. We note in Eq.~\eqref{eqn:var_I_simp_array} that $R_X$ and $R_Y$ differ and generally cannot be factored out. Again, taking $\sqrt{\text{Var}(\tilde{I})}$ from Eq.~\eqref{eqn:var_I_simp_array} and dividing by the free space impedance, $\eta_0$, produces the desired formula
\begin{eqnarray}
\text{SEFD}_{I}=\frac{\text{SDev}(\tilde{I})}{\eta_0}=\frac{4 k\Delta f}{\eta_0}\frac{\sqrt{\mathbf{t}_R^T \mathbb{L}\mathbf{t}_R}}{|D|^2},
\label{eqn:SEFD_I_res_array}
\end{eqnarray}
with units \si{\watt\per\metre\squared}. If $\mathrm{SEFD}_I$ is stated in \si{\watt\per\metre\squared\per\hertz}, which is often preferred in radio astronomy, then we remove $\Delta f$ from the right hand side.
\subsection{RMS approximation always underestimates SEFD}
\label{sec:RMSunderestimates}
As mentioned in Sec.~\ref{sec:intro}, we now demonstrate the claim that the very commonly used RMS approximation in radio astronomy, for example see~\citep{1999ASPC..180..171W}, always underestimates the true SEFD. The RMS approximation is given by
\begin{eqnarray}
\text{SEFD}_I^{rms}= \frac{1}{2}\sqrt{\text{SEFD}_{XX}^2+\text{SEFD}_{YY}^2},
\label{eqn:SEFD_rms}
\end{eqnarray}
where $\text{SEFD}_{XX}, \text{SEFD}_{YY}$ assume an unpolarized source as discussed in Paper~I. For the notations and approach discussed in this paper, this approximation can be written as
\begin{eqnarray}
\text{SEFD}_I^{rms}&=&\frac{4 k}{\eta_0}\sqrt{\frac{T_{sysX}^2R_X^2}{\norm{\mathbf{l}_X}^4}+\frac{T_{sysY}^2R_Y^2}{\norm{\mathbf{l}_Y}^4}}\nonumber\\
&=&\frac{4 k}{\eta_0}\sqrt{\mathbf{t}_R^T \mathbf{D}_{rms}\mathbf{t}_R},
\label{eqn:SEFD_rms_approx}
\end{eqnarray}
as pointed out in Paper~I. In addition, Appendix~\ref{apendx:RMS2L} shows that Eq.~\eqref{eqn:SEFD_rms_approx} is derivable using $A/T$. The diagonal matrix $\mathbf{D}_{rms}=\text{diag}\left[\norm{\mathbf{l}_{X}}^{-4}, \norm{\mathbf{l}_{Y}}^{-4}\right]$ is different from $\mathbb{L}/|D|^4$ in Eq.~\eqref{eqn:SEFD_I_res_array}. It can be shown that
\begin{eqnarray}
\mathbf{t}_R^T\frac{\mathbb{L}}{|D|^4}\mathbf{t}_R\geq \mathbf{t}_R^T\mathbf{D}_{rms}\mathbf{t}_R.
\label{eqn:to_prove}
\end{eqnarray}
The reason for this is explained as follows. It can be shown through the Gram-Schmidt orthogonalization steps or the related QR factorization~\citep[see Ch.~4]{Strang_ILA2016} of the Jones matrix, that the absolute value of the determinant can be written as
\begin{eqnarray}
|D|=\norm{\mathbf{l}_Y}\norm{\mathbf{e}},
\label{eqn:error_proof}
\end{eqnarray}
where $\mathbf{e}$ is the projection vector of $\mathbf{l}_X$ onto a line which is perpendicular to $\mathbf{l}_Y$, as shown in Fig.~\ref{fig:e_pic}
\begin{eqnarray}
\mathbf{e}=\mathbf{l}_X-\mathbf{p},
\label{eqn:error}
\end{eqnarray}
where $\mathbf{p}= (\mathbf{l}_Y^H\mathbf{l}_X)\mathbf{l}_Y/\norm{\mathbf{l}_Y}^2$,
is the orthogonal projection vector of $\mathbf{l}_X$ onto $\mathbf{l}_Y$~\citep[see Ch.~4]{Strang_ILA2016}, such that $\mathbf{e}$ and $\mathbf{p}$ are orthogonal, $\mathbf{e}\perp \mathbf{p}$. Eq.~\eqref{eqn:error_proof} is easily verified by substitution of $\mathbf{l}_X=[l_{X\theta},l_{X\phi}]^T$ and $\mathbf{l}_Y=[l_{Y\theta},l_{Y\phi}]^T$ into Eq.~\eqref{eqn:error} and simplifying the vector algebra. There is also an insightful geometric interpretation of Eq.~\eqref{eqn:error_proof}: $|D|$ is the area formed by the parallelogram whose parallel sides are represented by vectors $\mathbf{l}_X$ and $\mathbf{l}_Y$ as shown in Fig.~\ref{fig:e_pic}; this is known from the volume property of determinants~\citep[see Ch.~5]{Strang_ILA2016} and is useful for understanding Eq.~\eqref{eqn:ineq_are_C2} discussed next.
\begin{figure}
\centering
\begin{tikzpicture}
\draw [thick, ->] (0,0) -- (3,0);
\node at (3,-0.3) {$\mathbf{l}_Y$};
\draw [thick, ->] (0,0) -- (1,2);
\node at (0.5,1.5) {$\mathbf{l}_X$};
\draw [dashed, thick, ->] (0,-0.05) -- (1,-0.05);
\node at (0.5,-0.3) {$\mathbf{p}$};
\draw [dashed] (1,2) -- (4,2);
\draw [dashed] (3,0) -- (4,2);
\draw [dashed, thick, ->] (1,0) -- (1,1.9);
\node at (1.2,1.0) {$\mathbf{e}$};
\draw [-] (1.2,0.0) -- (1.2,0.2);
\draw [-] (1.2,0.2) -- (1.0,0.2);
\node at (2.5,1.5) {$\text{area}=|D|$};
\%draw [thick, <->] (0,2) -- (0,0) -- (2,0);
\%node at (1,1) {yes};
\end{tikzpicture}
\caption{Depiction of the vectors, their projections, and the area of the parallelogram.}
\label{fig:e_pic}
\end{figure}
Since $\norm{\mathbf{e}}\leq\norm{\mathbf{l}_X}$, then
\begin{eqnarray}
|D|^2\leq\norm{\mathbf{l}_{X}}^2\norm{\mathbf{l}_{Y}}^2.
\label{eqn:ineq_are_C2}
\end{eqnarray}
The equality is fulfilled when the vectors are orthogonal, $\mathbf{l}_{Y} \perp \mathbf{l}_{X}$, which also implies $\mathbf{p}=\mathbf{0}$, $\mathbf{e}=\mathbf{l}_{X}$, and $\mathbf{l}_{X}^H\mathbf{l}_{Y}=0$. This is fully expected given the above-mentioned geometric interpretation and proves Eq.~\eqref{eqn:to_prove}. To see this more clearly, we write
\begin{eqnarray}
\frac{\mathbb{L}}{|D|^4}=\frac{1}{\norm{\mathbf{l}_X-\mathbf{p}}^4\norm{\mathbf{l}_{Y}}^4}\left[ \begin{array}{cc}
\norm{\mathbf{l}_{Y}}^4 & |\mathbf{l}_X^{H}\mathbf{l}_Y|^2 \\
|\mathbf{l}_X^{H}\mathbf{l}_Y|^2 & \norm{\mathbf{l}_{X}}^4
\end{array} \right],
\label{eqn:tLt_C2}
\end{eqnarray}
and
\begin{eqnarray}
\mathbf{D}_{rms}=\frac{1}{\norm{\mathbf{l}_X}^4\norm{\mathbf{l}_{Y}}^4}\left[ \begin{array}{cc}
\norm{\mathbf{l}_{Y}}^4 & 0\\
0 & \norm{\mathbf{l}_{X}}^4
\end{array} \right].
\label{eqn:Drms_mat}
\end{eqnarray}
We note that the denominator on the right hand side of Eq.~\eqref{eqn:tLt_C2}, $\norm{\mathbf{l}_X-\mathbf{p}}^4\norm{\mathbf{l}_{Y}}^4\leq\norm{\mathbf{l}_X}^4\norm{\mathbf{l}_{Y}}^4$, which indeed proves Eq.~\eqref{eqn:to_prove}. Furthermore, it is evident that the matrix in Eq.~\eqref{eqn:tLt_C2} converges to $\mathbf{D}_{rms}$ in Eq.~\eqref{eqn:Drms_mat} when the orthogonality, $\mathbf{l}_{Y} \perp \mathbf{l}_{X}$, is fulfilled. Since in practice orthogonality is approached, but never fulfilled exactly, the statement that the RMS approximation always underestimates SEFD is demonstrably correct.
\textcolor{black}{We note the fact that antenna length is a direction dependent quantity. Therefore orthogonality, $\mathbf{l}_{Y} \perp \mathbf{l}_{X}$, as discussed in this section must be evaluated by taking the inner product, $\mathbf{l}_{X}^H\mathbf{l}_{Y}$, in the every direction of arrival of interest. This is particularly applicable to a wide field-of-view antenna system. For example, a mechanically orthogonal cross dipoles only possess antenna length orthogonality in directions of arrival on the cardinal planes (as discussed in Paper~I). The RMS approximation is only exact in directions of arrivals for which antenna length orthogonality, $\mathbf{l}_{X}^H\mathbf{l}_{Y}=0$, is fulfilled. }
Following the proof, we can then quantify this difference by defining a relative percentage difference of the two results as:
\begin{eqnarray}
\Delta \text{SEFD}_I =\frac{\text{SEFD}_{I}- \text{SEFD}_{I}^{rms}}{\text{SEFD}_{I}} \times 100 \% \geq 0.
\label{eqn:SEFD_I_error}
\end{eqnarray}
We will use Eq.~\eqref{eqn:SEFD_I_error} in Sec.~\ref{sec:compare} to measure the difference between RMS approximation and true SEFD based on simulation and observation, the expected value of which are always positive (for observation data, the ``expected value'' refers to the ensemble mean).
\section{SEFD simulation procedure}
\label{sec:simulation}
\begin{table*}[t]
\centering
\caption{ Calculated parameters required to compute $\text{SEFD}_I$ for the array pointing at Hydra-A source (Az=81\degree, ZA=46\degree) on 2014-12-26 at 16:05:43 UTC. \label{tab:values}}
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{l|ccc||ccc|}
\cline{2-7}
& \multicolumn{3}{c||}{\textbf{X}} & \multicolumn{3}{c|}{\textbf{Y}} \\ \cline{2-7}
& \multicolumn{1}{c|}{110.08 MHz} & \multicolumn{1}{c|}{154.88 MHz} & \multicolumn{1}{c||}{249.6 MHz} & \multicolumn{1}{c|}{110.08 MHz} & \multicolumn{1}{c|}{154.88 MHz} & \multicolumn{1}{c|}{249.6 MHz} \\ \hline
\multicolumn{1}{|l||}{$T_{ant}~(\si{\kelvin})$} & \multicolumn{1}{c|}{603} & \multicolumn{1}{c|}{227} & 66 & \multicolumn{1}{c|}{573} & \multicolumn{1}{c|}{222} & 63 \\ \hline
\multicolumn{1}{|l||}{$T_{\mathrm{rcv}}~(\si{\kelvin})$} & \multicolumn{1}{c|}{180} & \multicolumn{1}{c|}{80} & 59 & \multicolumn{1}{c|}{126} & \multicolumn{1}{c|}{51} & 50 \\ \hline
\multicolumn{1}{|l||}{$\eta_{\mathrm{rad}}$} & \multicolumn{1}{c|}{0.986} & \multicolumn{1}{c|}{0.972} & 0.989 & \multicolumn{1}{c|}{0.981} & \multicolumn{1}{c|}{0.980} & 0.951 \\ \hline
\multicolumn{1}{|l||}{$R~(\Omega)$} & \multicolumn{1}{c|}{15} & \multicolumn{1}{c|}{51} & {71} & \multicolumn{1}{c|}{22} & \multicolumn{1}{c|}{113} & {81} \\ \hline
\multicolumn{1}{|l||}{$l_\theta~(\si{\metre})$} & \multicolumn{1}{c|}{$0.112 +j1.441$} & \multicolumn{1}{c|}{$2.372 - j0.113$} & $-0.095 + j1.931$ & \multicolumn{1}{c|}{$0.022 + j0.237$} & \multicolumn{1}{c|}{$0.441 - j0.106$} & $0.008 + j0.740$ \\ \hline
\multicolumn{1}{|l||}{$l_\phi~(\si{\metre})$} & \multicolumn{1}{c|}{$-0.008 - j0.299$} & \multicolumn{1}{c|}{$-0.494 + j0.018$} & $0.080 - j0.339$ & \multicolumn{1}{c|}{$0.083 + j1.962$} & \multicolumn{1}{c|}{$3.779 - j0.680$} & $-0.880 + j2.045$ \\ \hline
\end{tabular}
\end{table*}
Following the previous section, the steps to computing the SEFD of a polarimetric phased array interferometer are as follows. The first step involves the computation of the array Jones matrix as shown in Eq.~\eqref{eqn:J_array}. The components of the matrix were constructed from embedded antenna realized lengths as shown in Eq.~\eqref{eqn:array_lX} and Eq.~\eqref{eqn:array_lY}, which in our case were obtained from electromagnetic simulation using Altair FEKO\footnote{https://www.altair.com/feko/}. Additional information regarding the simulation setup and results used in this paper were discussed in \citet{sokolowski_colegate_sutinjo_2017}. The embedded antenna realized lengths were found by converting $(E_{\theta}, E_{\phi})$ using \citep{Ung_9040892}
\begin{eqnarray}
l_{n\theta}(\theta,\phi) &=& -j \frac{2\lambda}{\eta_0 V_{t}}Z_{LNA}E_{\theta,n}(\theta,\phi), \nonumber\\
l_{n\phi}(\theta,\phi) &=& -j \frac{2\lambda}{\eta_0 V_{t}}Z_{LNA}E_{\phi,n}(\theta,\phi),
\label{eqn:lr_calc}
\end{eqnarray}
where $\lambda$ is the wavelength, $Z_{LNA}$ is the input impedance of the low-noise amplifier (LNA), and $V_{t}$ is the port excitation voltage used during the simulation which generates the corresponding electric far-field components $E_{\theta,n}$ and $E_{\phi,n}$ of the $n^{th}$ embedded element as a function of direction in the sky.
In the second step, we determine the system temperatures of the array, $T_{sysX}$ and $T_{sysY}$ using
\begin{eqnarray}
T_{sys,p} = T_{ant,p} + T_{rcv,p} + T_0(1-\eta_{rad,p}),
\label{eqn:Tsys_formula}
\end{eqnarray}
where $T_{ant,p}$ is the antenna temperature due to sky for each polarization $p$ ($p$~=~``X'' or ``Y''); $T_{rcv,p}$ is the receiver noise temperature due to the LNA calculated using the methodology discussed in~\citet{Ung_9040892} based on measured LNA noise parameters~\citep{Ung2018}; $\eta_{rad,p}$ is the radiation efficiency of the array and $T_0 = 290$~\SI{}{\kelvin} is the reference temperature (this can be substituted by ambient temperature if known). Radiation efficiency, $\eta_{\mathrm{rad}}$, is the ratio of the total radiated power to the total injected power into the array calculated using methodologies presented in~\citet{5439763} and~\citet{8888935}.
Third, we compute the realized array noise resistances, $R_X$ and $R_Y$, using Eq.~\eqref{eqn:R_noise}. Fourth, we calculate $\mathbf{t}_R^T \mathbb{L}\mathbf{t}_R$ in Eq.~\eqref{eqn:var_I_simp_array}. Finally, we calculate the polarimetric and interferometric array SEFD using Eq.~\eqref{eqn:SEFD_I_res_array}. As an example, we provided several key parameters required to evaluate Eq.~\eqref{eqn:SEFD_I_res_array} at three selected frequency points for an MWA tile pointed at azimuth and zenith angles, Az $= 80.54^\circ$, ZA$= 46.15^\circ$ in Table \ref{tab:values}. A full list of these values (excluding the realized length) can be found in the dataset accompanying this paper. In the next section, we will describe the procedure used to obtain $\text{SEFD}_{I}$ from observational data.
\section{SEFD measurement procedure}
\label{sec:obs}
The data analysis procedure is based on a very similar analysis performed earlier in the Paper~I. The relative difference $\Delta \text{SEFD}_I^{obs}$ defined by Eq.~\eqref{eqn:SEFD_I_error} was measured using MWA observations tracking Hydra-A radio-galaxy recorded in a standard observing mode at the central frequency 154.88\,MHz. On 2014-12-26 between 16:05:42 and 05:13:42 UTC, the MWA recorded 23 112\,s observations centred on the Hydra-A radio-galaxy at a position (Az,ZA) = (80.7\degree, 45.8\degree) and (Az,ZA) = (290.7\degree, 32.4\degree) at the start and end of the observations respectively.
For the purpose of a comparison of $\Delta \text{SEFD}_I$ between the data and simulations, 52\,s of the first 112\,s observation (obsID 1103645160 referred later as ObsA), which started at 2014-12-26 16:05:42 UTC with Hydra-A at (Az,ZA) = (80.7\degree, 45.8\degree), were analysed. This observation was selected at a pointing direction where the simulation predicted that $\Delta \text{SEFD}_I$ would be the most prominent due to the lowest elevation of the pointing direction at gridpoint 114 at the elevation of $\approx~$43.85\degree (ZA = 46.15\degree). We note that MWA observations significantly lower than 45$\degree$ are generally not routinely performed and were not considered for the presented analysis due systematic effects which could affect the data quality.
\subsection{Calibration and imaging}
\label{ref:calibration_and_imaging}
The MWA data were converted into CASA measurement sets \citep{casa} in 40\,kHz frequency (total of 768 channels) and 1\,s time resolution, and downloaded using the MWA All-Sky Virtual Observatory interface \citep{asvo}. In order to avoid aliasing effects at the edges of 1.28\,MHz coarse channels (24 in total), 160\,kHz (4 fine channels) on each end of coarse channels were excluded, which reduced the observing bandwidth to $R_{bw}=$0.75 fraction of the full recorded band (30.72\,MHz).
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figs/wsclean_1103645160_briggs_timeindex037-XX-image_ALL3.png}
\caption{Example 0.5\,s sky and difference images in X polarisation (scale of the colorbar is in Jy) used to calculate SEFD. Left : 0.5\,s sky image. Centre: difference of two consecutive 0.5\,s images. Right: difference of two consecutive 0.5\,s images corrected for the primary beam (divided by the primary beam in the X polarisation). In order to calculate noise in every direction in the sky within the field of view (``noise map''), standard deviation of the noise was calculated in small regions around each pixel in the difference image and divided by $\sqrt{2}$. These noise maps were later converted into SEFD using equation~\ref{eq_rms2sefd}. The same procedure was applied to Y and very similar to Stokes I polarisations (in this case the correction for the primary beam was more complex than the simple division by the corresponding primary beam).}
\label{fig:sky_images}
\end{figure*}
The same observation ObsA was used for calibration. It was sufficient to use a single-source sky model composed of Hydra-A alone, which was the dominant source in the observed field. However, due to the extended structure of Hydra-A, which is resolved by the longest baselines of the MWA Phase 1, it was necessary to use its model by \citet{2014MNRAS.440..327L} converted to the required multi-component format by \citet{gleamx}\footnote{https://github.com/nhurleywalker/GLEAM-X-pipeline}. The total flux density of Hydra-A in this model is within 7\% from the most recent and accurate flux measurements of Hydra-A by \citet{2017ApJS..230....7P} over the entire observing band (139.5 -- 169.0\,MHz). Selecting an appropriate model of Hydra-A was important to ensure correct flux scale of the calibrated visibilities, images and ultimately SEFD values.
The calibration was performed and applied using \textsc{CALIBRATE} and \textsc{APPLYSOLUTIONS} software, which are part of the MWA-reduce software package \citep{offringa-2016} routinely used for the MWA data reduction. The \textsc{calibrate} program uses the 2016 MWA beam model \citep{sokolowski_colegate_sutinjo_2017} to calculate Jones matrices and apparent flux of the sources (just one in this particular case) included in the sky model. Baselines shorter than 58 wavelengths were excluded from the calibration. The phases of the resulting calibration solutions were fitted with a linear function, while the amplitudes were fitted with a $5^{th}$ order polynomial; the resulting fitted calibration solutions were then applied to the un-calibrated visibilities using \textsc{APPLYSOLUTIONS} program.
Sky images with 4096$\times$4096 pixels of angular size $\approx$0.6\,arcmin (image size $\approx$\,42.2\degree$\times$42.2\degree), in 1\,s time resolution, were formed from all correlations products ($XX, YY, XY$ and $YX$)\footnote{We note that the correlation products are in fact $XX^*, YY^*, XY^*$ and $YX^*$, but in the context of this section the conjugate symbols ($^*$) were dropped for brevity.} using the \textsc{WSCLEAN}\footnote{\url{https://sourceforge.net/p/wsclean/wiki/Home/}} program \citep{OffMcK14}. Natural weighting (robust weighting parameter defined by \citet{Briggs_phd_thesis} set to $+2$) was used in order to preserve full sensitivity of the array. We note that natural weighting increases the confusion noise, but it was eliminated by measuring the noise from difference images (see Sec.~\ref{ref:noise2sefd}). The dirty maps were CLEANed with 100000 iterations and threshold of 1.2 standard deviations of the local noise.
In order to reduce contribution from extended emission baselines shorter than 30 wavelengths were excluded, which led to a small reduction of number of baselines used in the imaging to 7466, which is $R_{bs} \approx$0.92 fraction of all the baselines (8128). Corrections for reduced number of baselines and observing bandwidth were taken into account in the later analysis by applying correction factors ($R_{bw}$ and $R_{bs}$) during the conversion from standard deviation of the noise to SEFD.
The resulting $XX$ and $YY$ images (for example left image in Fig.~\ref{fig:sky_images}) were divided by corresponding images of the beam in X and Y polarizations generated with the 2016 beam model \citep{sokolowski_colegate_sutinjo_2017} at the same pointing direction ( (Az,ZA) = (80.5\degree , 46.2\degree) at the MWA gridpoint 114\footnote{MWA gridpoints are pointing directions where delays applied in the analogue beamformers are exact.}).
All the resulting images ($XX, YY, XY$ and $YX$) were converted to Stokes images ($I, Q, U$ and $V$) also using the 2016 beam model. The final products resulting from the above procedure were three sets ($XX, YY$ and Stokes $I$) of $n_t=53$ primary beam corrected images corresponding to 53\,s of the analysed MWA data.
\subsection{Measuring the SEFD from the noise in the sky images}
\label{ref:noise2sefd}
Using the three series (corresponding to $XX$, $YY$ and Stokes $I$) of $n_t=53$ images, difference images (between the subsequent $i$-th and $i-1$ image) were calculated, resulting in $n_d=52$ difference images in each of $XX$, $YY$ and Stokes $I$ polarizations. Difference imaging effectively removes confusion noise resulting in uniform images of thermal noise (entirely due to the system temperature $T_{sys}$). The difference images were visually inspected and verified to have a uniform noise-like structure (examples in the middle and right images in Fig.~\ref{fig:sky_images}). Therefore, the resulting standard deviation is purely due to the instrumental and sky noise ($T_{sys}$).
The standard deviation of the noise calculated in a small region (here a circular region of radius $R_n$=10 pixels corresponding to $\approx$~5 synthesized beams) around pixels in the resulting difference images is distributed around a mean of zero and is not ``contamined'' by the variations of the flux density within the circular regions due to astronomical sources contained inside these regions. The interquartile range divided by 1.35 was used as a robust estimator of standard deviation, which is more robust against outlier data points due to radio-frequency interference (RFI), or residuals of astronomical sources in the difference images. Once this calculation was performed around all the pixels in a difference image, an image of the noise over the field of view (FoV) was created. These images of the noise are later referred to as noise maps or just N (note that N stands for a 2D image and not a number). This procedure was applied to $n_d$ all-sky difference images resulting in $n_d$ noise maps ($N_{I}$, $N_{XX}$ and $N_{YY}$) for each of the polarizations (Stokes $I$, $XX$ and $YY$ respectively).
The corresponding SEFD images (2D maps of SEFD) of the entire FoV were calculated from these noise maps (referred in general as $N$) according to the equation:
\begin{equation}
\text{SEFD}= N \sqrt{ \Delta \nu \Delta t N_b},
\label{eq_rms2sefd}
\end{equation}
where $\Delta t=1$\,s is the integration time, $\Delta \nu = 30.72 \times R_{bw}$\,MHz is the observing bandwidth corrected for reduced number of channels ($R_{bw}$) used in the analysis and $N_b = 7466$ is the number of baselines used in the imaging.
Noise maps in the three polarizations, $N_{I}$, $N_{XX}$ and $N_{YY}$, were converted into corresponding $\text{SEFD}_I$, $\text{SEFD}_{XX}$ and $\text{SEFD}_{YY}$ images using Eq.~(\ref{eq_rms2sefd}), resulting in $n_d$ SEFD images in each polarization (I, XX and YY). Then, median SEFD images $\overline{\text{SEFD}_I}$, $\overline{\text{SEFD}_{XX}}$ and $\overline{\text{SEFD}_{YY}}$ were calculated out of individual $n_d$ SEFD images (a median image of $\text{SEFD}_{I}$ is shown in Figure~\ref{fig:obs_sefd_i}). These median images were used to calculate $\Delta \text{SEFD}_I^{obs}$ as:
\begin{eqnarray}
\Delta \text{SEFD}_I^{\text{obs}} = \frac{( \overline{\text{SEFD}_{I}} - \frac{1}{2}\sqrt{\overline{\text{SEFD}_{XX}}^2 + \overline{\text{SEFD}_{YY}}^2} )}{\overline{\text{SEFD}_{I}}}.
\label{eqn:ratio}
\end{eqnarray}
In the next section, we provide the results derived from simulated and observed data, and discuss their implications.
\subsection{SEFD as a function of frequency}
\label{ref:sefd_vs_frequency}
\begin{figure}[b]
\centering
\includegraphics[width=0.48\textwidth]{figs/sefd_visibilit.png}
\caption{SEFD \textcolor{black}{in the pointing direction of gridpoint 114 with realized area calculated at (Az, ZA) = (80\degree , 46\degree)} as a function of frequency. \textcolor{black}{Observed} SEFD calculated from calibrated visibilities compared with simulated SEFD.}
\label{fig:sefd_visib}
\end{figure}
The method described in Section~\ref{ref:calibration_and_imaging} allows us to determine spatial structure of SEFD by calculating SEFD values in each pixel of 4096$\times$4096 sky images. This method was successfully used in Paper I, where only single dipole in each MWA tile was enabled and the remaining 15 dipoles were terminated. However, it has never been applied to measure SEFD from MWA images resulting from observations in a standard mode with all dipoles in MWA tiles enabled. Therefore, before comparison with the simulations, an initial verification was performed by deriving SEFD values using well-tested method described in Section C of \citep{7293140}. This method uses standard deviation of calibrated visibilities to measure $\text{SEFD}_{XX}$ and $\text{SEFD}_{YY}$ (we highlight that only SEFD of instrumental polarizations can be measured using this method but not $\text{SEFD}_I$) as a function of frequency in the direction of Hydra-A source (no spatial information). These measurements were used as a reference to confirm that the image-based SEFD values were correct. Furthermore, this method allowed us to compare frequency dependence of $\text{SEFD}_{XX}$ and $\text{SEFD}_{YY}$ against the simulations. Fig. \ref{fig:sefd_visib} shows the comparison between the observed and simulated results. \textcolor{black}
At the location in the sky corresponding to (Az, ZA) = (80\degree, 46\degree), the realized area of X-polarized dipole is significantly smaller than Y-polarized dipole and therefore, $\text{SEFD}_{XX}$ is higher than $\text{SEFD}_{YY}$.}
At 154.88~MHz, observed $\text{SEFD}_{XX}=78\pm7.8$~kJy and $\text{SEFD}_{YY}=57.6\pm3.6$~kJy, which is in very good agreement with the values obtained from the imaging method. Overall, there is excellent agreement between simulated and observed $\text{SEFD}_{XX/YY}$ over a wide range of frequencies.
\section{Results and discussion}
\label{sec:compare}
\subsection{Comparison between the data and simulation}
In order to ensure that we are able to draw accurate conclusions regarding the difference in SEFD values calculated using the RMS approximation in Eq.~\eqref{eqn:SEFD_rms} and the true SEFD using Eq.~\eqref{eqn:SEFD_I_res_array}, we compared the simulated SEFD with the measurements. The SEFD of the $X,Y$ and $I$ using difference imaging method probes the SEFD over multiple locations in the sky for a given frequency (154.88 MHz). Similarly, the simulated $\text{SEFD}_{XX/YY}$ images were generated using $A/T$ formulation while the simulated $\text{SEFD}_{I}$ image was generated using Eq.~\eqref{eqn:SEFD_I_res_array}.
Fig. \ref{fig:sefd_sim_meas_comparI} shows the SEFD obtained from simulation (Fig.~\ref{fig:sim_sefd_i}), Stokes I image (Fig.~\ref{fig:obs_sefd_i}) and the difference in percentage between the two (Fig.~\ref{fig:delta_sefd_i}). For reference, contours of the normalized beam pattern for each polarization are displayed in $3$~dB increments down to $-12$~dB level.
\begin{figure}[h!]
\centering
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{figs/MWA_SEFDI_obs1103645160_sim.pdf}
\caption{Simulated $\text{SEFD}_{I}$}
\label{fig:sim_sefd_i}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{figs/MWA_SEFDI_obs1103645160_meas.pdf}
\caption{Observed $\text{SEFD}_{I}$}
\label{fig:obs_sefd_i}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{figs/MWA_SEFDI_obs1103645160_delta3D.pdf}
\caption{$M_{\text{SEFD}_{I}}$, \%, Eq.~(\ref{eqn:delta_s_m})}
\label{fig:delta_sefd_i}
\end{subfigure}
\caption{(a) Simulated $\text{SEFD}_{I}$, (b) observed $\text{SEFD}_{I}$ and (c)~percentage difference between the simulated and observed $\text{SEFD}_{I}$ at 154.88~MHz calculated by Eq.~(\ref{eqn:delta_s_m}).}
\label{fig:sefd_sim_meas_comparI}
\end{figure}
The contours of the beam pattern shown in the $\text{SEFD}_I$ are obtained by normalizing the reciprocal of Eq.~\eqref{eqn:SEFD_I_res_array}. Similar comparison was done for the X and Y polarization, but for brevity, the results are included in Appendix \ref{apendx:XXYY_comp}.
The relative difference in \%, named $M_{SEFD}$, between the simulated and observed SEFD in all images ($XX$, $YY$ and $I$) was computed as:
\begin{eqnarray}
M_{\text{SEFD}}= \frac{\text{SEFD}^{obs} - \text{SEFD}^{sim}}{\text{SEFD}^{sim}} \times 100 \%.
\label{eqn:delta_s_m}
\end{eqnarray}
We analysed the errors using histograms and the observed $\text{SEFD}_{XX}$ and $\text{SEFD}_I$ are on average higher by 9\% and 4\% respectively, while the observed $\text{SEFD}_{YY}$ is lower by 4\% compared to simulated values for all pixels within the $-12$~dB beam width. Overall, the agreement between simulated and observed values is within the $\pm10\%$ range which indicates excellent correspondence of simulated and observed SEFD data.
Following the verification of simulated SEFD with observed data, we can proceed to calculate the error between the RMS and proposed method using Eq.~\eqref{eqn:SEFD_I_error}. Fig.~\ref{fig:relative_ratio_simul} and Fig.~\ref{fig:relative_ratio_data} show this error for an MWA tile for both simulated and observed data, respectively. We can see that the difference between $\text{SEFD}_I^{rms}$ and $\text{SEFD}_I$ is not a uniform offset in the image, but has a noticeable structure. For the observation within the $-3$~dB beam width, the error of SEFD prediction using the RMS approximation is only 7\%. However, it increases as we move further away from the beam center and reaches 23\% in the $-12$~dB beam width. We note that for error in the simulated values, $\Delta \text{SEFD}_{I}^{sim}$, is always positive, meaning the RMS method always underestimates the SEFD as expected.
Fig.~\ref{fig:SEFDI_diff} shows the subtraction of $\Delta \text{SEFD}_I^{obs}$ and $\Delta \text{SEFD}_I^{sim}$ calculated as
\begin{eqnarray}
\Delta_{sim/obs} = \Delta \text{SEFD}_{I}^{obs} - \Delta \text{SEFD}_{I}^{sim}.
\label{eqn:delta_delta}
\end{eqnarray}
The resulting mean difference for all pixels within the $-12$~dB beam width is zero, which allows us to form the following conclusion: (i) the negative values seen in $\Delta \text{SEFD}_I^{obs}$ are due to noise and (ii) we can fully predict the amount of underestimation (\% errors) in $\text{SEFD}_I$ produced by the RMS method. This indicates that Eq.~\eqref{eqn:SEFD_I_res_array} more accurately predicts the array's $\text{SEFD}_{I}$.
\begin{figure}[htb!]
\centering
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{figs/MWA_dSEFDI_obs1103645160_sim.pdf}
\caption{$\Delta \text{SEFD}_{I}^{sim}$, Eq.~(\ref{eqn:SEFD_I_error})}
\label{fig:relative_ratio_simul}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{figs/MWA_dSEFDI_obs1103645160_meas.pdf}
\caption{$\Delta \text{SEFD}_{I}^{obs}$, Eq.~(\ref{eqn:SEFD_I_error})}
\label{fig:relative_ratio_data}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{figs/MWA_dSEFDI_obs1103645160_delta3D.pdf}
\caption{$\Delta_{sim/obs}$, Eq.~(\ref{eqn:delta_delta})}
\label{fig:SEFDI_diff}
\end{subfigure}
\caption{Relative difference between the (a) simulated and (b) observed $\text{SEFD}_I$ and (c) the difference of $\Delta \text{SEFD}_I^{obs}$ and $\Delta \text{SEFD}_I^{sim}$.}
\label{fig:main_fig}
\end{figure}
\subsection{Computation of $\text{SEFD}_I$ for diagonal plane and low elevation pointing angle}
\begin{figure}[h!]
\begin{center}
\noindent
\includegraphics[width=0.5\textwidth]{figs/MWA_delSEFDI_gp150_ch156.pdf}
\caption{$\Delta \text{SEFD}_I^{sim}$ computed for 199.68~MHz at (Az, ZA) = (45\degree, 56.96\degree), showing the impact of the error resulting from RMS expression of $\text{SEFD}_I$.}
\label{fig:worst_case}
\end{center}
\end{figure}
While there was a noticeable structure in the result shown in Fig. \ref{fig:main_fig}, we note that for that observation time, frequency and pointing angle, the difference in SEFD formulation would probably be hard to notice within the 3-dB beam width and indeed, in some cases, only images in the target field down to $-3$~dB are created. Nevertheless, these results show that the improved expression of $\text{SEFD}_I$ provides highly accurate SEFD values, and thus, can be used to precisely calculate the sensitivity of a radio telescope for any pointing angle within the operating frequency range using simulations.
To that end, it is reasonable to continue our analysis based only on simulated results. We completed an iterative search aimed to find a case where $\Delta \text{SEFD}_I$ affects the data within $-3$~dB beam width. Fig.~\ref{fig:worst_case} shows one outcome of this iterative search. It was found that for $\text{Az} = 45\degree$, $ZA=56.96\degree$ ($33.04 \degree $ elevation angle) at 199.68~MHz, the RMS SEFD approximation results in 29\% error within the $-3$~dB beam width and increases up to 36\% within the $-12$~dB beam width. This example further demonstrates the inaccuracies of the RMS approximation, which always predicts lower SEFD than the actual value.
\section{Conclusion} \label{sec:concl}
This work, which is an extension of Paper~I, provides a generalized expression of $\text{SEFD}_I$ for interferometric phased array polarimeters. As in Paper~I, this expression was derived by performing statistical analysis on the standard deviation of the flux density estimate. Our current work further clarifies the fact that the $\text{SEFD}_I$ expression does not depend on assumptions made regarding the background polarization nor orthogonality in the antenna elements. The key array parameters were obtained from the full-wave electromagnetic simulation of the phased array of interest, which is an MWA tile in this example, and used to compute $\text{SEFD}_I$ given by Eq.~\eqref{eqn:SEFD_I_res_array}. The mean percentage difference ($M_{\text{SEFD}_{I}}$) between simulated and observed $\text{SEFD}_I$ is $4\%$, which indicates excellent agreement. Furthermore, we provided proof that the RMS approximation always underestimates the $\text{SEFD}_I$. This proof also demonstrates that the equality between the RMS approximation and $\text{SEFD}_I$ is reached when the row vectors in the Jones matrix are orthogonal, which is a condition that can be approached but never fulfilled in practice.
After verifying the accuracy of our simulation with observed data, we proceeded to compare the relative percentage difference in $\text{SEFD}_I$, as defined in Eq.~(\ref{eqn:SEFD_I_error}), calculated by the often-used RMS approximation and Eq.~(\ref{eqn:SEFD_I_res_array}). A large percentage difference here would indicate that the RMS approximation underestimates the $\text{SEFD}_I$. Indeed, we see that for our chosen observation, the maximum simulated error $\Delta \text{SEFD}_{I}^{sim}=7\%$ within the $-3$~dB beam width and increases to $23\%$ within the $-12$ dB beam width. This result was also verified against the observed data ($\Delta \text{SEFD}_{I}^{obs}$) by taking absolute difference between the two images (Fig.~\ref{fig:relative_ratio_simul} and Fig.~\ref{fig:relative_ratio_data}) defined in Eq.~(\ref{eqn:delta_delta}). The resulting difference between simulated and observed $\Delta \text{SEFD}_I$ has a zero mean value, which indicates a remarkable agreement between the results.
As we had thoroughly validated our simulation against observed data, we then performed an iterative search for pointing angles and frequencies that would yield a higher difference between the RMS approximation and Eq.~\eqref{eqn:SEFD_I_res_array}. We found that for beam pointing (Az, ZA) = (45\degree, 56.96\degree) at 199.68~MHz the RMS approximation produces an error of $29\%$ in $\Delta \text{SEFD}_{I}^{sim}$ within the $-3$~dB beam width, which increases to 36\% within the $-12$~dB beam width. This outcome is in agreement with the prediction made in Paper~I whereby the difference increases at $\text{Az} = 45\degree$ at low elevation angles.
We conclude that the derived $\text{SEFD}_I$ expression improves the fundamental understanding of instrument performance and can be used to accurately calculate sensitivity not only at the principal planes, but also at the diagonal planes (Az$=45\degree$) and low elevation angles. This enables us to more confidently predict sensitivity for detection of pulsars, fast radio bursts (FRBs) or Epoch of Reionisation (EoR) signal. This is particularly important for the cases where the target sources can only be observed at very low elevations (e.g., in effect of response to alerts, about FRBs or any other transient sources).
\section*{Acknowledgement}
\label{sec:Ack}
This scientific work makes use of the Murchison Radio-astronomy Observatory (MRO), operated by CSIRO. We acknowledge the Wajarri Yamatji people as the traditional owners of the Observatory site. Support for the operation of the MWA is provided by the Australian Government (NCRIS), under a contract to Curtin University administered by Astronomy Australia Limited. This work was further supported by resources provided by the Pawsey Supercomputing Centre with funding from the Australian Government and the Government of Western Australia. The authors thank A/Prof. R. B. Wayth and Prof. D. Davidson for discussions on this topic and for reviewing the draft manuscript. The authors thank Dr. Natasha Hurley-Walker for providing the model of the Hydra-A calibrator source.
\bibliographystyle{aa}
|
\section{Introduction} This paper gives new cases of a conjecture made by R. Berger in 1963 \cite{Berger63}. Let $\k$ be an algebraically closed field of characteristic $0$, and let
$(R,\mathfrak{m}_R,\k)$ be an equicharacteristic reduced one-dimensional complete local $\k$-algebra. Berger conjectured that the universally finite module of differentials, $\Omega_R$,
is torsion-free if and only if $R$ is regular. The case in which $R$ is regular is easy, since in that case $\Omega_R$ is free. Hence another formulation of
his conjecture is that $\Omega_R$ is torsion-free if and only if it is a free $R$-module.
There are many approaches to the conjecture which have been partially successful. We refer \cite{Berger63}, \cite{scheja1970differentialmoduln}, \cite{herzog1971wertehalbgruppe}, \cite{Bassein}, \cite{Herzog78}, \cite{Buchweitz}, \cite{Ulrich81}, \cite{Koch}, \cite{HerzogWaldi84}, \cite{HerzogWaldi86}, \cite{yoshino1986torsion}, \cite{Berger88}, \cite{pohl1989torsion1}, \cite{Guttes1990}, \cite{Hubl}, \cite{Isogawa}, \cite{Pohl91}, \cite{ABC1}, \cite{ABC2} and \cite{maitra2020partial} for these approaches. A very nice summary of a majority of these results, along with the main ideas of proofs, can be found in \cite{Berger_article}.
Our generalizations have to do with how the conductor $\mathfrak{C}_R=R:_{\operatorname{Quot}(R)}\overline{R}$ of $R$ sits inside $R$. We first prove that if the conductor is not in the square of the maximal ideal, then Berger's conjecture is true (\Cref{valuation more than conductor}). When the conductor is in the square of the maximal ideal, we construct a certain subring $S$ of $\operatorname{Hom}_R(\mathfrak{m}_R,\mathfrak{m}_R)$. By construction, there always exists torsion in $\Omega_S$. We show that if there are enough torsion elements in $\Omega_S$, we can construct a nonzero torsion element in $\Omega_R$ (\Cref{thm on one more torsion}). One of the first cases we prove is if $S$ is quasi-homogeneous then
Berger's conjecture is true (\Cref{quasihomogeneous}), generalizing a result of Scheja \cite{scheja1970differentialmoduln}.
Let $x$ be a minimal reduction of the maximal ideal $\mathfrak{m}_R$. The next set of results depends on which power of the maximal ideal $\mathfrak{m}_R$ is contained in the conductor $\mathfrak{C}_R$ of $R$. We study the quantity $s(R):=\dim_\k\frac{\(\mathfrak{C}_R,x\)}{(x)}$ which we shall refer to as the \textit{reduced type} of $R$. The terminology reduced type is natural due to the fact that $\frac{(\mathfrak{C}_R,x)}{(x)}\subseteq \frac{(x):\mathfrak{m}}{(x)}$ and the $\k$-dimension of the latter module is precisely the type of $R$.
Our main results extend both those of G\"uttes (\cite{Guttes1990}), who proved that if either $\mathfrak{m}_R^4\subseteq xR$ or $R$ is Gorenstein and $\mathfrak{m}_R^5\subseteq xR$, then Berger's conjecture holds, and of Corti\~{n}as, Geller, and Weibel (\cite{ABC1}), who
proved that if $\mathfrak{m}_R^3\subseteq \mathfrak{C}_R$, then Berger's conjecture holds.
We summarize our extensions below, with $\operatorname{edim}$ denoting the embedding dimension (\Cref{m^4 in C} and \Cref{m^6 in C}):
\begin{theoremx}
Let $(R,\mathfrak{m}_R,\k)$ be an equicharacteristic one-dimensional complete local domain over an algebraically closed field $\k$ of characteristic $0$,
$x$ be a minimal reduction of $\mathfrak{m}_R$ and $\mathfrak{C}_R$ be the conductor ideal of $R$ in its integral closure. Further, let $n=\operatorname{edim} R$ and $s(R)$ be the reduced type of $R$. Then Berger's conjecture is true in the following cases:
\begin{enumerate}
\item $\mathfrak{m}_R^4\subseteq (\mathfrak{C}_R, x)$ and $2\cdot s(R)\leq n(n- 3)$,
\item $\mathfrak{m}_R^6\subseteq (x)$, $n\geq 6$ and $R$ is Gorenstein.
\end{enumerate}
\end{theoremx}
We give other generalizations which relate to the structure of the ring $S$. In particular, when the length of $S/R$ is one, we obtain results similar in spirit to the case in which $R$ is Gorenstein, without having to assume the Gorenstein property.
The structure of the article is as follows: \Cref{conductor not in m^2} takes care of the case when the conductor $\mathfrak{C}_R$ is not contained in the square of the maximal ideal $\mathfrak{m}_R$ (\Cref{valuation more than conductor}) and some related cases. \Cref{basics of S} gives the details of the construction of $S$ and its basic properties (\Cref{conductorbasicprops}, \Cref{canonical module of S} and \Cref{kernelofnew_proof}). And finally, the main results are presented in \Cref{main results} (\Cref{m^4 in C}, \Cref{mainthm} and \Cref{m^6 in C}).
\medskip
\section{Setting and Preliminaries}\label{prelims}
Throughout this paper we assume that $\k$ is an algebraically closed field of characteristic $0$, and $(R,\mathfrak{m}_R,\k)$ is a one-dimensional complete equicharacteristic local $\k$-algebra which is a domain with embedding dimension (denoted by $\operatorname{edim}$) $n$, i.e. $\mu_R(\mathfrak{m}_R)=n$ where $\mu_R(M)$ denotes the minimal number of generators for any $R$-module $M$. We also denote the length of any $R$-module $M$ by $\ell(M)$.
Choosing $t$ to be a uniformizing parameter of the integral closure $\overline{R}$ of $R$, we may assume that
$\overline{R}=\k\ps{t}$. It follows we can write $R=\k\ps{\alpha_1 t^{a_1},\dots,\alpha_nt^{a_n}}$ where $\alpha_i$'s are units in $\overline{R}$ and $a_1\leq a_2\leq \cdots\leq a_n$. Note also that here $\displaystyle \bar{R}$ is finitely generated over $R$ (see for example \cite{SwansonHuneke}[Theorem 4.3.4]).
We define an epimorphism
\begin{align}
\Phi:P=\k\ps{X_1,\dots,X_n}&\twoheadrightarrow R\\
\Phi(X_i)&=\alpha_it^{a_i}\text{ for } 1\leq i\leq n.\nonumber
\end{align}
We denote the kernel of $\Phi$ by $I=(f_1,\dots,f_m)$ and hence have the natural isomorphism $R\cong \k\ps{X_1,\dots,X_n}/I$. Since $\operatorname{edim} R=n$, $I$ is contained in $\mathfrak{m}^2$ where $\mathfrak{m}=(X_1,\dots,X_n)$, the maximal ideal of $\k\ps{X_1,\dots,X_n}$. Such rings are called \textit{analytic $\k$-algebras}. We will interchangeably use $\alpha_i t^{a_i}$ for $x_i$, the images of $X_i$ in the quotient $\k\ps{X_1,\dots,X_n }/I$.
We also define the valuation $\operatorname{ord}(-)$ on $\overline{R}$ given by $\operatorname{ord}(p(t))=a$ if $p(t)=t^a\alpha$ where $\alpha$ is a unit in $\k\ps{t}$ (see for example \cite[Example 6.7.5]{SwansonHuneke}).
\subsection{Universally Finite Module of Differentials}
\begin{definition}\label{defmoduleofdiff}
Let $R$ be an analytic one-dimensional $\k$-algebra as above, which is a domain. Let $I=(f_1,\dots,f_m)$ where $f_j\in P=\k\ps{X_1,\dots,X_n}$. We assume that $I\subseteq \mathfrak{m}_P^2$ where $\mathfrak{m}_P=(X_1,\dots,X_n)$. Then the \textit{universally finite module of differentials over $\k$}, denoted by $\Omega_{R}$, has a (minimal) presentation given as follows:
$$R^{m}\xrightarrow{\left[\frac{\partial f_j}{\partial x_i}\right]} R^n\to \Omega_{ R}\to 0$$ where $\left[\frac{\partial f_j}{\partial x_i}\right]$ is the Jacobian matrix of $I$, with entries in $R$.
\end{definition}
We refer the reader to the excellent resource \cite{Kunzbook} for more information.
Let $\displaystyle \tau(\Omega_{ R})$ denote the torsion submodule of $\Omega_{ R}$. The conjecture of interest in this article is the following:
\begin{conjecturex}[R. W. Berger \cite{Berger63}]
Let $\k$ be a perfect field and let $R$ be a reduced one-dimensional analytic $\k$-algebra. Then $R$ is regular if and only if $\displaystyle \tau(\Omega_{ R})=0$.
\end{conjecturex}
Although the conjecture is for reduced algebras, in this paper we only deal with the case in which $R$ is a domain and $\k$ is algebraically closed of characteristic zero. Our techniques do not immediately seem
to apply otherwise.
\begin{remark}\label{rankomega}
When $\k$ is a perfect field, it is well-known that
$\operatorname{rank}_R(\Omega_{R})=\dim(R)=1$. Hence from \Cref{defmoduleofdiff}, we get that $\operatorname{rank} A=n-1$ where $A=\left[\frac{\partial f_j}{\partial x_i}\right]$ is the Jacobian matrix of $I$, with entries in $R$.
\end{remark}
\begin{remark}
It is clear from \Cref{defmoduleofdiff}, that $\tau(\Omega_{ R})=0$ when $R$ is regular. Thus, from now on we assume that $n\geq 2$.
\end{remark}
\subsection{The Conductor}\label{conductor def}
The conductor ideal $\mathfrak{C}_R$ will be crucial for the purposes of this paper. Recall that the conductor is the largest common ideal of $R$ and its
integral closure, $\overline R $. It follows that $\displaystyle \mathfrak{C}_R=R:_Q\overline{R}$ where $Q=\operatorname{Quot}(R)$, denotes the fraction field of $R$. Since $\displaystyle \overline{R}=k\ps{t}$ and $\mathfrak{C}_R$ is an ideal of $\bar{R}$ as well, we have that $\mathfrak{C}_R=(t^i)_{i\geq c_R}$ where $c_R$ is the least integer such that $t^{c_R-1}\not \in R$, and $t^{c_R+i}\in R$ for all $i\geq 0$. The number $c_R$ is characterized as the least valuation in $\mathfrak{C}_R$. It is clear from this discussion that there cannot be any element $r\in R$, such that $v(r)=c_R-1$. Since $\displaystyle \bar{R}$ is finitely generated over $R$ (\cite[Theorem 4.3.4]{SwansonHuneke}), the conductor ideal is a nonzero ideal of $R$, and it is never all of $R$ unless $R$ is regular.
\subsection{Computing Torsion}
We have the following commutative diagram using the functorial universal properties of the module of differentials and the associated universal derivations.
\[\begin{tikzcd}
\Omega_{R}\ar[r,"f"] & \Omega_{\overline{R}}\\
R\ar[r,hook,"i"] \ar[u,"d"] & \overline{R}\ar[u, swap, "d"]
\end{tikzcd}
\]
We use the same symbols $d$ for both the vertical maps.
Since, $\rk(\Omega_R)=\rk(\Omega_{\overline{R}})=1$ ($R$ and $\bar{R}$ have the same fraction field), and $\Omega_{\overline{R}}$ is free over $\overline{R}$, we get that the $\tau(\Omega_R)=\ker f$.
Also note that by commutativity of the diagram,
$$f(dx_i)=\frac{dx_i}{dt}dt.$$
Since $\Omega_{\overline{R}}$ is isomorphic to $\overline{R}$, we see that $\Omega_{R}$ surjects to a $R$-submodule $\displaystyle \sum_{i=1}^nR\frac{dx_i}{dt}$ of $\displaystyle \overline{R}=\k\ps{t}$. This is a fractional ideal in $\overline{R}$, so multiplying by a suitably high enough power of $t$, it is isomorphic to an ideal of $R$.
The torsion submodule $\tau(\Omega_R)$ is the kernel of the map $\Omega_R\rightarrow\Omega_{\overline{R}}$. Thus, from the above discussion, we get that $\displaystyle \tau(\Omega_R)$ consists of the tuples $\begin{bmatrix}
r_1\\\vdots\\r_n
\end{bmatrix}$ such that $\displaystyle \sum_{i=1}^nr_i\frac{dx_i}{dt}=0$. Evidently, $\tau(\Omega_R)$ is non-zero precisely when the tuples $\begin{bmatrix}
r_1\\\vdots\\r_n
\end{bmatrix}$ are not in the image of the presentation matrix (Jacobian matrix of $I$) of $\Omega_R$, all entries written in terms of the uniformizing parameter $t$. This provides one computational way of computing torsion using Macaulay 2.
\begin{example}
$R=\ps{t^3,t^4,t^5}$ and its defining ideal $I=(y^2-xz,z^2-x^2y,x^3-yz)$ in $\k\ps{x,y,z}$.
Consider the element $$\tau=4ydx-3xdy=\begin{bmatrix}
4y\\-3x\\0
\end{bmatrix}\hspace{-1em}
\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\
\end{array}\hspace{-1em}
\overbrace{
\begin{array}{c}
dx \\ dy \\dz
\end{array}
}^{\text{basis}}$$ in $\Omega_R$. Clearly $4y\frac{dx}{dt}-3x\frac{dy}{dt}=4t^4(3t^2)-3t^3(4t^3)=0$. Now the presentation matrix of $\Omega_{R/\mathfrak{m}_R^2}$ is
\begin{align*}
\begin{bmatrix}
2x & y & z & 0 & 0 &0\\
0 & x & 0 & 2y &z &0\\
0 & 0 & x & 0 &y & 2z
\end{bmatrix}.
\end{align*}
Since the image of $\tau$ in $\Omega_{R/\mathfrak{m}_R^2}$ can never be written as a linear combination of the columns of the above presentation, $\tau$ is nonzero in $\Omega_{R/\mathfrak{m}_R^2}$. Thus $\tau$ is nonzero in $\Omega_R$ as well.
\end{example}
\section{Nonzero torsion when $\mathfrak{C}_R\not\subseteq\mathfrak{m}_R^2$}\label{conductor not in m^2}
Throughout this section, we assume that $\k$ is an algebraically closed field of characteristic $0$, and $(R,\mathfrak{m}_R,\k)$ is a one-dimensional complete equicharacteristic local $\k$-algebra which is a domain with embedding dimension $n$. Choosing $t$ to be a uniformizing parameter of the integral closure $\overline{R}$, we may assume that $\overline{R}=\k\ps{t}$. It follows we can write $R=\k\ps{\alpha_1 t^{a_1},\dots,\alpha_nt^{a_n}}$ where $\alpha_i$'s are units in $\overline{R}$ and $a_1\leq a_2\leq \cdots\leq a_n$. Let $\mathfrak{C}_R$ denote the conductor ideal.
Our primary construction which appears in the sequel will make use of the condition $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$. So prior to our main construction, we settle the case $\mathfrak{C}_R\not\subseteq\mathfrak{m}_R^2$ by showing that this condition always leads to nonzero torsion in $\Omega_R$. In this case at least one of the minimal generators $x_1,\dots,x_n$ of the maximal ideal is in the conductor $\mathfrak{C}_R$. Thus, after a change of variables, the minimal generators in the conductor can be replaced by monomials (i.e., if $x_i=\alpha_it^{a_i}\in \mathfrak{C}_R$, then after a change of variables, $\alpha_i$ can be chosen to be a unit in $R$). We will see in \Cref{two monomials}, the presence of monomials will lead to nonzero torsion in $\Omega_R$.
\begin{theorem}\label{valuation more than conductor} If $\mathfrak{C}_R\not\subseteq\mathfrak{m}_R^2$,
then the torsion $\tau(\Omega_R)$ is nonzero.
\end{theorem}
\begin{proof} Write $R= \k\ps{\alpha_1t^{a_1},\dots,\alpha_nt^{a_n}}$ with conductor $\mathfrak{C}_R=(t^{c_R})\overline{R}$. We first monomialize the $r^{th}$ term as follows:
by multiplying by a nonzero element of $\k$, we may assume that the constant term of the unit $\alpha_r$ is $1$.
By Hensel's lemma \cite[Theorem 7.3]{eisenbud_Commalg}, there exists an element $\beta\in R$ such that $\beta^{a_r} = \alpha_r$. Here we use that the characteristic of $\k$ is $0$. We
write $\beta = 1 + \beta_1t+\cdots$.
Consider the change of variables $s=\beta t$. Under this change of variables, notice that $\k\ps{t}=\k\ps{s}$. Now
\begin{align*}
s=\beta t=t+\beta_1t^{2}+\beta_2t^{3}+\cdots.
\end{align*}
Note that $s^{a_r}=(\beta t)^{a_r}=\alpha_rt^{a_r}\in R$. Furthermore, $\alpha_it^{a_i}=\alpha_i's^{a_i},1\leq i\neq r\leq n$, where $\alpha'_i\in\overline{R}$ are units. Then $R=\k\ps{\alpha_1t^{a_1},\dots,\alpha_nt^{a_n} } = \k\ps{\alpha_1's^{a_1},\dots,\alpha_{r-1}'s^{a_{r-1}},s^{a_r},\alpha_{r+1}'s^{a_{r+1}},\dots ,\alpha_n's^{a_n}}$. We apply this change of variables with $r = 1$ to assume without loss of generality, for the remainder of this proof, that $R = \k\ps{t^{a_1},\alpha_2t^{a_2},\cdots,\alpha_nt^{a_n}}$.
Since $\mathfrak{C}_R$ is not contained in $\mathfrak{m}_R^2$, we must have that $a_n\geq c_R$. Write $\alpha_n = \alpha_{n0} + tb$, where $\alpha_{n0}\ne 0$ is in $\k$ and $b\in \overline{R}$.
Then $\alpha_nt^{a_n} = \alpha_{n0}t^{a_n} + t^{a_n+1}b$, where $b\in \overline{R}$. However, since $a_n\geq c_R$, it follows that $t^{a_n+1}b\in \mathfrak{C}_R\subset R$.
Hence, $t^{a_n} = \alpha_{n0}^{-1}(\alpha_nt^{a_n}-t^{a_n+1}b)\in R$ as well, and then $R = \k\ps{t^{a_1}, \alpha_2t^{a_2},\ldots,\alpha_{n-1}t^{a_{n-1}},t^{a_n}}$.
We now use this particular form for $R$ to find a nonzero torsion element in $\Omega_R$. Namely, $a_nx_ndx_1-a_1x_1dx_n\in\Omega_R$ and the exact sequence
\begin{align*}
0\rightarrow\tau(\Omega_R)\rightarrow\Omega_R\xrightarrow{\phi} R\frac{dx_1}{dt}+\cdots+R\frac{dx_n}{dt}\rightarrow 0
\end{align*}
where the map $\phi$ is the $R$-module map given by $\phi(dx_i)=\frac{dx_i}{dt}, 1\leq i\leq n$. Under this map
\begin{align*}
\phi(a_nx_ndx_1-a_1x_1dx_n)&=a_nt^{a_n}\frac{dt^{a_1}}{dt}-a_1t^{a_1}\frac{dt^{a_n}}{dt}\\
&=(a_1a_nt^{a_1+a_n-1}-a_na_1t^{a_1+a_n-1})dt=0
\end{align*}
Thus $a_nx_ndx_1-a_1x_1dx_n\in \tau(\Omega_R)$. It remains to see that it is nonzero. Consider the image of this element $\overline{a_nx_ndx_1-a_1x_1dx_n}$ in $\Omega_{R/\mathfrak{m}_R^2}$. As $x_1x_n\in \mathfrak{m}_R^2$, it follows that in $\Omega_{R/\mathfrak{m}_R^2}$, $\overline{x_1dx_n+ x_ndx_1} = 0$. Hence $\overline{a_nx_ndx_1-a_1x_1dx_n} = (a_1+a_n)\overline{x_1}d(\overline{x_n})$ in $\Omega_{R/\mathfrak{m}_R^2}$.
Now using \cite[Proposition 2.6, Corollary 2.7]{ABC1}, we have $(a_1+a_n)\overline{x_1}d(\overline{x_n})\neq 0$ in $\Omega_{R/\mathfrak{m}_R^2}$. Thus $a_nx_ndx_1-a_1x_1dx_n\neq 0$ in $\Omega_R$.
\end{proof}
\begin{example}\label{example with a_i is more than c}
Let $R=\k\ps{t^4+t^5,t^7+t^{10},t^8+t^{10},t^9+t^{10}}$ . Macaulay2 computations show that the conductor is $\mathfrak{C}_R=(t^{c_R})\overline{R}=(t^7)\overline{R}$ . Since $a_2\geq c_R$, we have $\tau(\Omega_R)\neq 0$ using the previous result.
\end{example}
\begin{remark}\label{two monomials}
If $\alpha_i,\alpha_j$ are units in $R$ for some $i\neq j$, then we can show that $\tau(\Omega_R)$ is nonzero. Assuming $i=1,j=2$ we can easily see that $R=\k\ps{x_1,\dots,x_n}\cong \k\ps{t^{a_1},t^{a_2},\alpha_3t^{a_3},\dots,\alpha_nt^{a_n}}$. Notice that $a_2x_2dx_1-a_1x_1dx_2\in \tau(\Omega_R)$. This torsion element is nonzero as $\overline{a_2x_2dx_1-a_1x_1dx_2}=(a_2-a_1)\overline{x_2dx_1}$ is nonzero in $\Omega_{R/\mathfrak{m}_R^2}$ (\cite[Corollary 2.7]{ABC1}).
\end{remark}
The above result is also a generalization of \cite[Corollary 3.7]{Isogawa}. The next example illustrates the remark.
\begin{example}
Let $R=\k\ps{t^{8}+t^{9},t^{9}+t^{15},t^{12}+t^{20},t^{14}}$, the conductor $\mathfrak{C}_R=(t^{c_R})\overline{R}=(t^{20})\overline{R}$. Thus $R \cong \k\ps{t^{8}+t^{9},t^{9}+t^{15},t^{12},t^{14}}$ and hence $R$ has at least one torsion element using \Cref{two monomials}. Notice that in this case, none of the $a_i$'s are bigger than the $c_R$.
\end{example}
\section{The transform $R[\frac{\mathfrak{C}_R}{x_1}]$}\label{basics of S}
Throughout this section, we again assume that $(R,\mathfrak{m}_R,\k)$ is a one-dimensional complete equicharacteristic local $\k$-algebra which is a domain with embedding dimension $n$. Further, $\k$ is an algebraically closed field of characteristic $0$. Choosing $t$ to be a uniformizing parameter of the integral closure $\overline{R}$, we may assume that $\overline{R}=\k\ps{t}$. Using the technique as in the first paragraph of the proof of \Cref{valuation more than conductor}, we can also write $R=\k\ps{ t^{a_1},\alpha_2t^{a_2},\dots,\alpha_nt^{a_n}}$ where $\alpha_i$'s are units in $\overline{R}$ and $a_1\leq a_2\leq \cdots\leq a_n$. Let $\mathfrak{C}_R$ denote the conductor ideal.
In this section we study the main construction $S=R[\frac{\mathfrak{C}_R}{x_1}]$ where $x_i=\alpha_it^{a_i}$ with $\alpha_1=1$. Throughout this section, we assume that $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$.
\subsection{Basics of $R[\frac{\mathfrak{C}_R}{x_1}]$}
\qquad \\
We write $\frac{\mathfrak{C}_R}{x_1}$ to denote the set of elements of the form $\frac{c}{x_1}$ where $c\in \mathfrak{C}_R$. We note that the conductor is never inside a proper principal ideal (follows, for instance, from \cite[Corollary 2.6]{maitra2020partial}),
so there are always elements in $\frac{\mathfrak{C}_R}{x_1}$ which are not in $R$ itself. Recall from the introduction that we define the \textit{reduced type} of $R$ to be $$s(R):=\dim_\k\frac{(\mathfrak{C}_R,x_1)}{(x_1)}.$$ We use the notation $s$ whenever the underlying ring is clear.
\smallskip
\begin{lemma}\label{conductorbasicprops} Let $S=R[\frac{\mathfrak{C}_R}{x_1}]$.
The following statements hold.
\begin{enumerate}
\item $ \frac{\mathfrak{C}_R}{x_1}\subseteq \overline{R}$.
\item $\mathfrak{m}_R (\frac{\mathfrak{C}_R}{x_1})\subseteq \mathfrak{C}_R$. In particular, $S\subset \operatorname{Hom}_R(\mathfrak{m},\mathfrak{m})$.
\item Let $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$. Let $\alpha, \beta\in \frac{\mathfrak{C}_R}{x_1}$. Then $\alpha\beta\in \mathfrak{C}_R$.
\item Let $\displaystyle \mathfrak{C}_S$ denote the conductor of $S$ in $\overline{R}$. Then $ \mathfrak{C}_S=\frac{\mathfrak{C}_R}{x_1}$.
\item $S/R$ is a vector space over $\k$ of dimension $s$, where $s$ is the reduced type of $R$.
\end{enumerate}
\end{lemma}
\begin{proof} Since $\displaystyle (x_1)$ is a minimal reduction of $\mathfrak{m}_R$, we have $\frac{\mathfrak{m}_R}{x_1}\subseteq \overline{R}$. Hence $ \frac{\mathfrak{C}_R}{x_1}\subseteq \overline{R}$ proving $(1)$.
For $(2)$ note that $\displaystyle \mathfrak{m}_R\overline{R}=x_1\overline{R}$ and hence $\displaystyle \mathfrak{m}_R\mathfrak{C}_R=x_1\mathfrak{C}_R$. So, $\mathfrak{m}_R\frac{\mathfrak{C}_R}{x_1}\subseteq \mathfrak{C}_R\subseteq \mathfrak{m}_R$.
Write $\alpha=\frac{c}{x_1},\beta=\frac{c'}{x_1}$ where $\displaystyle c,c'\in \mathfrak{C}_R$. For $(3)$, first note that $\displaystyle cc'\in \mathfrak{C}_R^2\subseteq \mathfrak{C}_R\mathfrak{m}_R^2=\mathfrak{C}_R x_1\mathfrak{m}_R =x_1^2\mathfrak{C}_R$ as in the proof of $(2)$. This proves $(3)$.
Finally, note that every valuation more than $c_R-a_1-1$ is present in the valuation semi-group of $S$. But there cannot be any element with valuation $c_R-a_1-1$ in $S$: if possible, let $r$ be such an element. Then by $(2)$, $rx_1\in R$ and has valuation $c_R-1$, a contradiction to the choice of $c_R$. This finishes the proof of $(4)$.
For (5), first note that (2) clearly implies that $S/R$ is a vector space. Now $S/R$ is generated as a $\k$-vector space by elements of the form $c/x_1$ such that $c\in \mathfrak{C}_R$. Choose a basis $\overline{c_1},\dots,\overline{c_s}$ for the $\k$-vector space $(\mathfrak{C}_R,x_1)/(x_1)$. Now construct a $\k$-linear map $\eta:(\mathfrak{C}_R,x_1)/(x_1)\rightarrow S/R$ by mapping $\eta(\overline{c_i})=c_i/x_1$. Suppose $\overline{c}=\sum_{i}k_i\overline{c_i}$ with $k_i\in \k$ be such that $\eta(\overline{c})=0$. Then we get $\sum_ik_i\frac{c_i}{x_1}=0$ in $S/R$. It follows that $(\sum_ik_ic_i)/x_1\in R$ and hence $\sum_{i}k_ic_i\in (x_1)$. This in turn implies that $\overline{c}=0$ and thus, $\eta$ is injective. It is also surjective as any element $c/x_1$ of $S/R$ has a pre-image $\overline{c}\in (\mathfrak{C}_R,x_1)/(x_1)$.
\end{proof}
We shall see that a decrease in the valuation of the conductor can significantly help us in gaining better understanding of torsion elements of $\Omega_R$. If we construct the ring $S=R[\frac{\mathfrak{C}_R}{x_1}]$, then \Cref{conductorbasicprops} guarantees such a drop. We try to explicitly describe the ring $S$ now. First we set up some notation.
\noindent \begin{notation}\label{conventionC/x}
We know that $\displaystyle \mathfrak{C}_R=(t^c,\dots,t^{c+a_1-1})R$ where $c=c_R$ is the valuation (of the conductor ideal) as discussed in \Cref{conductor def}. Hence $\displaystyle \mathfrak{C}_R/x_1$ is generated in $\overline{R}$ (in fact it is the ideal $(t^{c-a_1})\overline{R}$)
by the monomials
\begin{equation}\label{powerlist} t^{c-a_1}, t^{c-a_1+1}, \cdots, t^{c-1}.\end{equation} This is true because if $\alpha_{c-a-1},\ldots,\alpha_{c-1}$ are arbitrary units of the
integral closure $\overline{R}$, the ideal generated by $t^{c-a_1}, t^{c-a_1+1}, \cdots, t^{c-1}$ is the same as the ideal generated by $\alpha_{c-a-1}t^{c-a_1}, \ldots, \\\alpha_{c-1}t^{c-1}$.
The new ring $S$ is constructed by adjoining $\mathfrak{C}_R/x_1$. By \Cref{conductorbasicprops}(5) $S/R$ is a $\k$ vector space of dimension $s$. Using \cite[Proposition 2.9]{herzog1971wertehalbgruppe} we see that this is generated by those powers of $t$ from \Cref{powerlist}, which are not in the valuation semigroup of $R$. We call these powers say $b_1,\dots,b_s$ in ascending order.
\end{notation}
\noindent\begin{remark}\label{canonicalmoduledescription} Since $(\mathfrak{C}_R,x_1)/(x_1)\subseteq ((x_1):\mathfrak{m}_R)/(x_1)$ and the dimension of the latter quantity represents the type of $R$, it is clear that the reduced type $s$ is at most the type of $R$. Since $S$ never equals $R$, it is always at least one. The number $s$ can also be described as $\mu(\omega_R/\omega_S)$ where $\omega_R,\omega_S$ are canonical modules of $R,S$ respectively.
\end{remark}
\begin{proof} For the last statement, dualize the following short exact sequence into the canonical module $\omega_R$ $$0\to R\to S\to \k^s\to 0$$
gives the short exact sequence:
$$0\to \omega_S\to \omega_R\to \operatorname{Ext}^1_R(\k^s, \omega_R)\to 0.$$
Since the number of generators of $\omega_R$ is the type of $R$, and since $\operatorname{Ext}^1_R(\k^s, \omega_R)\cong \k^s$ by duality, the remark follows.
\end{proof}
When $s=1$ we say $R$ is of \textit{reduced type one}. In particular, if $R$ is Gorenstein (type of $R$ equals one), necessarily $s = 1$. The converse is not necessarily true as the next example shows.
\begin{example}
Let $R=\k\ps{t^4,t^{11},t^{17}}$. We can check $R$ is not Gorenstein using \cite{Kunz1970}. M2 computations show that the conductor $\mathfrak{C}_R=(t^{19})R$. It follows that $S = \k\ps{t^4,t^{11},t^{15}, t^{16}, t^{17}, t^{18}} = \k\ps{t^4,t^{11},t^{17}, t^{18}}$ and hence $R$ is of reduced type one.
\end{example}
When $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$, it follows that $\operatorname{edim} S=n+s$. Recall that canonical ideal $\omega_R$ of $R$ exists (\cite[Proposition 3.3.18 ]{bruns_herzog_1998}).
We can also prove that $s=\mu_R\left(\frac{\omega_R}{\overline{\mathfrak{m}_R\omega_R}\cap\omega_R}\right)$,
where $\overline{\mathfrak{m}_R\omega_R}$ is the integral closure of the ideal $\mathfrak{m}_R\omega_R$, thinking of the canonical module $\omega_R$ as an ideal of $R$.
We can show that the canonical module $\omega_S$ of $S$ is in fact $\overline{\mathfrak{m}_R\omega_R}\cap\omega_R$.
\begin{theorem}\label{canonical module of S}
Suppose $S=R[\frac{\mathfrak{C}_R}{x_1}]$, then a canonical module $\omega_S$ of $S$ can be chosen to be $\omega_R\cap \overline{\mathfrak{m}_R\omega_R}$.
\end{theorem}
\begin{proof}
The first part of the proof is essentially due to \cite[Lemma 3]{Brown_Herzog}. But we provide the proof here with more details suitable for our purposes. Let $(y)$ be a minimal reduction of a canonical ideal $\omega_R$ of $R$. Thus, we have $\omega_R\overline{R}=y\overline{R}$ and $\omega_R\mathfrak{C}_R=y\mathfrak{C}_R$. Let $\omega'_R=\frac{\omega_R}{y}$. Clearly, $R\subseteq\omega'_R\subseteq\overline{R}$.
Let $Q=\operatorname{Quot}(R)$. Now recall that $\omega_R:_Q \omega_R=R$ (\cite[Proposition 3.3.11(c)]{bruns_herzog_1998} and \cite[Lemma 2.4.2]{SwansonHuneke}). It is also well-known that $R:_Q\mathfrak{C}_R=\overline{R}$ (see for instance the proof of \cite[Corollary 2.6]{maitra2020partial}). Combining these facts, we obtain that
\begin{align*}
\omega_R':_Q\overline{R}=y^{-1}(\omega_R:_Q\overline{R})&=y^{-1}(\omega_R:_Q(R:_Q\mathfrak{C}_R))\\
&=y^{-1}(\omega_R:_Q((\omega_R:_Q\omega_R):_Q\mathfrak{C}_R))\\
&=y^{-1}(\omega_R:_Q(\omega_R:_Q\omega_R\mathfrak{C}_R))\\
&=y^{-1}\omega_R\mathfrak{C}_R
\end{align*} where the last equalities follow, by applying duality to the maximal Cohen-Macaulay module $\omega_R\mathfrak{C}_R$ (\cite[Theorem 3.3.10(c)]{bruns_herzog_1998}). The equality $\omega_R\mathfrak{C}_R=y\mathfrak{C}_R$ now shows that $\omega_R':_Q\overline{R}=y^{-1}\omega_R\mathfrak{C}_R=\mathfrak{C}_R$. Since $\omega'_R,\mathfrak{C}_R$ are fractional ideals, we have $\operatorname{Hom}(\overline{R},\omega'_R)=\mathfrak{C}_R$ \cite[Lemma 2.4.2]{SwansonHuneke}. This implies $\operatorname{Hom}(\mathfrak{C}_R,\omega'_R)=\overline{R}$ as $\overline{R}$ is a maximal Cohen-Macaulay module. Thus $\omega'_R:\mathfrak{C}_R=\overline{R}$ (again using \cite[Lemma 2.4.2]{SwansonHuneke}).
The canonical module $\omega_S\cong\operatorname{Hom}(S,\omega'_R)\cong \omega'_R:S$. Now let $\alpha\in\operatorname{Quot}(R)$ such that $\alpha\in \omega'_R:S$. Then $\alpha S\subseteq \omega'_R$ or $\alpha R\left[\frac{\mathfrak{C}_R}{x_1}\right]\subseteq \omega'_R$. Thus
\begin{align*}
\alpha R\left[\frac{\mathfrak{C}_R}{x_1}\right]\subseteq \omega'_R&\Leftrightarrow \alpha R\subseteq \omega'_R\text{ and }\alpha(\mathfrak{C}_R/x)\subseteq \omega'_R\\
&\Leftrightarrow\alpha R\subseteq \omega'_R\text{ and }\alpha/x\subseteq \omega'_R:\mathfrak{C}_R=\overline{R}\\
&\Leftrightarrow\alpha R\subseteq \omega'_R\text{ and }\alpha\in x\overline{R}=\mathfrak{m}_R\overline{R}
\end{align*}
Thus $\omega_S\cong \omega'_R\cap \mathfrak{m}_R\overline{R} \cong \omega'_R y\cap \mathfrak{m}_R y\overline{R} =\omega_R \cap \overline{\mathfrak{m}_R y} =\omega_R \cap \overline{\mathfrak{m}_R\omega_R}$.
\end{proof}
Combining the above theorem with \Cref{canonicalmoduledescription}, it is easy to see that $s=\mu_R\left(\frac{\omega_R}{\overline{\mathfrak{m}_R\omega_R}\cap\omega_R}\right)$.
\begin{theorem}\label{kernelofnew_proof}
Let $\displaystyle \mathfrak{C}_R\subseteq \mathfrak{m}_R^2$. Set $P = \k\ps{X_1,...,X_n}$. Choose $S,b_1,\dots,b_s$ as in \Cref{conventionC/x}. Then there exists a presentation of $S$ as follows: $$S=R\Big[\frac{\mathfrak{C}_R}{x_1}\Big]=\frac{\k\ps{X_1,\dots, X_n,T_1,\dots, T_s}}{\ker\Phi + \(X_iT_j-g_{ij}(X_1,\dots,X_n),T_kT_l-h_{kl}(X_1,\dots, X_n)\)_{{1\leq i\leq n, 1\leq j\leq s \atop 1\leq k\leq l\leq s}}}$$ where
$g_{ij}(X_1,\dots,X_n), h_{kl}(X_1,\dots, X_n)\in (X_1,\dots,X_n)^2P$ for all $i,j,k,l$. \\Moreover, $g_{ij}(x_1,\dots,x_n), h_{kl}(x_1,\dots, x_n)\in\mathfrak{C}_R$.
\end{theorem}
\begin{proof}
Define $\Psi : \k\ps{X_1,\dots,X_n,T_1,\dots,T_s}\twoheadrightarrow S$ where $$\Psi(X_i)=\alpha_i t^{a_i},1\leq i\leq n, \qquad\Psi(T_j)=t^{b_j}, 1\leq j\leq s.$$ Note that $c_R-a_1\leq b_1,\dots,b_s\leq c_R-1$. By same arguments as in \Cref{conductorbasicprops}$(3)$ and also by reading off valuations, we see that the images of $\displaystyle X_iT_j$ and $\displaystyle T_kT_l$ are all in $\mathfrak{C}_R$. Hence there exists $g_{ij}(X_1,\dots,X_n)$, $h_{kl}(X_1,\dots,X_n)$ such that $$X_iT_j-g_{ij}(X_1,\dots,X_n),T_kT_l-h_{kl}(X_1,\dots, X_n)\in \ker \Psi.$$ Moreover,
since $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$, we can choose $g_{ij}(X_1,\dots,X_n), h_{kl}(X_1,\dots, X_n)\in (X_1,\dots,X_n)^2P$ for all $i,j,k,l$ as well as $g_{ij}(x_1,\dots,x_n), h_{kl}(x_1,\dots, x_n)\in\mathfrak{C}_R$. Set $$\displaystyle J=\(X_iT_j-g_{ij}(X_1,\dots,X_n),T_kT_l-h_{kl}(X_1,\dots, X_n)\)_{{1\leq i\leq n, 1\leq j\leq s \atop 1\leq k\leq l\leq s}}.$$
By construction, elements in $J$ do not have any purely linear terms in any $X_i$. By the above discussion, $J\subset \ker \Psi$.
Conversely, let $p(X_1,\dots,X_n,T_1,\dots T_s)\in\ker\Psi$. Modulo the ideal $J$, we can write $$p(X_1,\dots,X_n,T_1,\dots, T_s)\equiv p'(X_1,\dots,X_n) +\sum_{i=1}^s\beta_i T_i$$ where $\beta_i \in \k$ and $p'(X_1,\dots,X_n)\in P$. Since $J\subset\ker\Psi$, we have $p'(X_1,\dots,X_n)+\sum_{i=1}^s\beta_i T_i\in\ker\Psi$. Thus, $\displaystyle \sum_{i=1}^s\beta_i t^{b_i}=\Psi(\sum_{i=1}^s\beta_i T_i)=\Psi(- p'(X_1,\dots,X_n))\in R$. By the choice of $b_i$'s, we immediately obtain that $\sum_{i=1}^s\beta_it^{b_i}=0$. Thus $\beta_i=0$ for all $i$ and hence $p'(X_1,\dots,X_n)\in\ker\Phi$. This shows that $\ker\Psi=\ker\Phi+J$.
\end{proof}
\begin{remark}\label{defining matrix of S}
Using the defining ideal of the $S$ in the previous theorem gives the following presentation of $\Omega_S$:
\begin{equation}
\scalemath{0.5}{
\begin{bmatrix}
~&T_1-\partial g_{11}/\partial x_1&\cdots& \partial g_{n1}/\partial x_1&T_2-\partial g_{12}/\partial x_1&\cdots& \partial g_{n2}/\partial x_1 & \cdots & T_n-\partial g_{1n}/\partial x_1&\cdots& \partial g_{nn}/\partial x_1&\partial h_{11}/\partial x_1&\partial h_{12}/\partial x_1&\cdots &\partial h_{1n}/\partial x_1&\cdots &\partial h_{nn}/\partial x_1\\
\partial f_i/\partial x_j &\vdots &\vdots &\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
~&\partial g_{11}/\partial x_n & \cdots &T_1-\partial g_{n1}/\partial x_n&\partial g_{12}/\partial x_n & \cdots &T_2-\partial g_{n2}/\partial x_n & \cdots &\partial g_{1n}/\partial x_n & \cdots &T_n-\partial g_{nn}/\partial x_n&\partial h_{11}/\partial x_1&\partial h_{12}/\partial x_1&\cdots&\partial h_{1n}/\partial x_n&\cdots &\partial h_{nn}/\partial x_n\\
~& x_1 & \cdots & x_n& 0 & \cdots & 0 &\cdots & 0 & \cdots & 0 &2T_1 & T_2 &\cdots &0 &\cdots &0\\
~& 0 & \cdots & 0& x_1 & \cdots & x_n &\cdots & 0 & \cdots & 0 &0 & T_1 &\cdots &0 &\cdots &0\\
~& \vdots & \cdots & \vdots& \vdots & \cdots & \vdots &\cdots & \vdots & \cdots & \vdots &\vdots & \vdots &\cdots &\vdots &\cdots &\vdots\\
~& 0 & \cdots & 0& 0 & \cdots & 0 &\cdots & x_1 & \cdots & x_n &0 & 0 &\cdots &0 &\cdots &2T_n\\
\end{bmatrix}
}
\end{equation}
\end{remark}
By abuse of notation, we denote the images of $T_i$ in $S$, by $T_i$ again. Thus $S$ is also the same as $\k\ps{x_1,\dots,x_n,T_1,\dots,T_s}$. In $S=R\left[\frac{\mathfrak{C}_R}{x_1}\right]=\k\ps{t^{a_1},\dots,\alpha_nt^{a_n},t^{b_1},\dots,t^{b_s} }$, the torsion submodule $\tau(\Omega_S)$ is nonzero (\Cref{two monomials}).
In $\Omega_S$, notice that $\Psi(T_j)=t^{b_j}$ (refer to \Cref{kernelofnew_proof} for definition of $\Psi$). Clearly, $b_iT_idT_j-b_jT_jdT_i\in \tau(\Omega_S), 1\leq i<j\leq s$ as $b_it^{b_j}dt^{b_i}-b_it^{b_i}dt^{b_j}=0$ in $\Omega_{\overline{R}}$. This torsion element is nonzero due to \cite[Proposition 2.6]{ABC1}. Thus $\Omega_S$ always has nonzero torsion elements of the form $\gamma_{ij}=b_iT_idT_j-b_jT_jdT_i, 1\leq i<j\leq s$ (which are ${s\choose 2}$ in number). So the torsion submodule $\tau(\Omega_S)$ has at least ${s\choose 2}$ elements. Moreover, all these elements are $\k$-linearly independent as follows again from \cite[Proposition 2.6]{ABC1}.
\begin{lemma}\label{one more torsion lemma}
Let $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$ and construct $S=R\left[\frac{\mathfrak{C}_R}{x_1}\right]=R[T_1,\dots,T_s]$ as in \Cref{kernelofnew_proof}. Let $\tau(\Omega_R),\tau(\Omega_S)$ represent the torsion submodules of $\Omega_R,\Omega_S$ respectively and $\gamma_{ij}=b_iT_idT_j-b_jT_jdT_i, 1\leq i<j\leq s$. Consider an element $\tau=\sum_i r_idx_i+\sum_j r_j'dT_j$ in $\Omega_S$ where $r_j',1\leq j\leq s$ are not units in $S$. Then there exist $c_{ij}\in \k$ such that $$\tau-\sum_{i,j}c_{ij}\gamma_{ij}=\sum_i r_i''dx_i\in \Omega_S,$$ where $r_i''\in S$. In particular, if $\tau\in \tau(\Omega_{S})$ then $\sum_i r_i''dx_i\in \tau(\Omega_S)$.
\end{lemma}
\begin{proof}
Since $\mathfrak{m}_S^2 = \mathfrak{m}_R^2 + \mathfrak{m}_R(\frac{\mathfrak{C}_R}{x_1}) + (\frac{\mathfrak{C}_R}{x_1})^2 = \mathfrak{m}_R^2$, and since $S/\mathfrak{m}_S = R/\mathfrak{m}_R$, we can represent the
entries $r_j', 1\leq j\leq s$ by elements of $R$ plus linear forms over $\k$ in $T_1,...,T_s$. Let $r_j'=p_j(x_1,\dots,x_n)+\sum_{l}k_{jl}T_{l}, 1\leq j\leq s$ where $k_{jl}\in\k$. Clearly, using the Jacobian matrix (as in \Cref{defining matrix of S}) we can rewrite $\tau$ as $\sum_i u_idx_i+\sum_j u_j'dT_j$ where $u_j'=\sum_{l}k'_{jl}T_l$ where $k_{jl}'\in\k$. We wish to eliminate the variables $T_1,\dots,T_s$ from $u_j'$. Suppose $k_{jl}'\neq 0$.
The column in the Jacobian matrix corresponding to the defining equation $T_lT_j-h_{lj}(X_1,\dots,X_n)$ of $S$ (refer to \Cref{kernelofnew_proof}) is of the form $\theta_{lj}=T_ldT_j+T_jdT_l-\sum_t\frac{\partial h_{lj}}{\partial X_t}dX_t$. Notice that $\gamma_{lj}+b_j\theta_{lj}=(b_l+b_j)T_ldT_j-\sum_tb_l\frac{\partial h_{lj}}{\partial X_t}dX_t$. Since $\theta_{lj}=0$ in $\Omega_S$, $\gamma_{lj}$ can be rewritten as $(b_l+b_j)T_ldT_j-\sum_tb_l\frac{\partial h_{lj}}{\partial x_t}dx_t$ in $\Omega_S$. Now $\tau-\frac{k'_{jl}}{(b_l+b_j)}\gamma_{lj}$ will not have a $T_l$ term in the $(n+j)$-th row (corresponding to $dT_j$). Since $j,l$ were arbitrary, a $\k$-linear combination of $\tau$ and $\gamma_{lj}$ will eliminate all the variables $T_1,\dots,T_s$ from $u_j'$. Since $j$ was arbitrary, we can eliminate the variables $T_1,\dots,T_s$ from $u_1',\dots,u_s'$ as well, to get the result.
\end{proof}
The following theorem is an important technical result of this article. This will serve as the main tool that will help us in pulling back nonzero torsion elements from $\Omega_S$ to $\Omega_{ R}$, as we shall see. We shall write $\ell(M)$ to denote the length of an $R$-module $M$.
\begin{theorem}\label{thm on one more torsion}
Let $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$ and construct $S=R\left[\frac{\mathfrak{C}_R}{x_1}\right]=R[T_1,\dots,T_s]$ as in \Cref{kernelofnew_proof}. Let $\tau(\Omega_R),\tau(\Omega_S)$ represent the torsion submodules of $\Omega_R,\Omega_S$ respectively. If $\ell(\tau(\Omega_S))\geq ns+{s \choose 2}+1$ and all these torsion elements have non-units in the last $s$ rows (corresponding to $dT_1,\dots,dT_s$), then a $\k$-linear combination of these torsion elements can be pulled back to a nonzero torsion element in $\tau(\Omega_R)$. In particular, $\tau(\Omega_R)\neq 0$.
\end{theorem}
\begin{remark} Although the hypothesis of at least $ns+{s\choose 2}+1$ $\k$-linear torsion elements sounds very strong, in fact it is not. We can prove that there are always at least
$ns+{s\choose 2}$ $\k$-linearly independent torsion elements in $\Omega_S$, so this theorem requires only one more new torsion element. Of course, we are only searching for one
nonzero torsion element in $\Omega_R$ in any case, but the point is that the search for one extra torsion element is often better in $S$ than in $R$, since $S$ is usually
a simpler ring.
The number $ns+\binom{s}{2}$ comes from the following observation: suppose some $\alpha_i=1$. Then we can look at the torsion element $a_ix_idT_j-b_jT_jdx_i$ for $1\leq j\leq s$. We can apply the `monomialization technique' as in the proof of \Cref{valuation more than conductor}, to generate $ns$ such torsion elements; we always have $\binom{s}{2}$ torsion elements coming from the variables $T_j$'s. Finally we can use \cite[Proposition 2.6]{ABC1} to prove $\k$-linear independence among these. We defer the technical details to a future article.
\end{remark}
\begin{proof}[Proof of \Cref{thm on one more torsion}]
Since $\ell(\tau(\Omega_S))\geq ns+{s \choose 2}+1$, let $\tau_1,\dots,\tau_{ns+{s\choose 2}+1}$ denote these $\k$-linearly independent torsion elements with non units in the last $s$ rows (corresponding to $dT_1,\dots,dT_s$). By \Cref{one more torsion lemma}, we can use $\gamma_{ij}$ to rewrite $\tau_1,\dots,\tau_{ns+{s\choose 2}+1}$ as $\tau_1',\dots,\tau_{ns+{s\choose 2}+1}'$ which have zeroes in the last $s$ rows. Let $V=\left\{\tau_1',\dots,\tau_{ns+{s\choose 2}+1}'\right\}$ and \\$V'=V\bigcup\left\{\gamma_{ij}~|~1\leq i<j\leq s\right\}$. Thus we have
\begin{align}\label{one more torsion eq1}
\dim_\k\langle V'\rangle\leq \dim_\k \langle V\rangle +{s\choose 2}
\end{align}
where $\langle \cdot\rangle$ denotes the $\k$-linear span. Notice that $\{\tau_1,\dots,\tau_{ns+{s\choose 2}+1}\}\subseteq \langle V'\rangle$ and hence
\begin{align}\label{one more torsion eq2}
ns+{s\choose 2}+1\leq \dim_\k \langle V'\rangle.
\end{align}
Combining \eqref{one more torsion eq1},\eqref{one more torsion eq2}, we have $\dim_\k \langle V\rangle\geq ns+1$. Thus there are at least $ns+1$ $\k$-linearly independent elements $\rho_i,1\leq i\leq ns+1$ all having the last $s$ rows (corresponding to $dT_1,\dots,dT_s$) consisting of zeroes.
Let $B=[\rho_1~\ldots~\rho_{ns+1}]$ be the $(n+s)\times (ns+1)$ matrix obtained by concatenating the column vectors $\rho_i,1\leq i\leq ns+1$.
Since the last $s$ rows of $B$ are zero, it is effectively an $n\times (ns+1)$ matrix.
Our goal is to pullback a $\k-$linear combination of the columns of $B$ to $\Omega_R$. Passing to the vector
space $S/R$ gives an $n\times (ns+1)$ matrix of linear forms in $T_1,...,T_s$. Writing the coefficients of each linear form as its own $s\times 1$ column, we obtain an
$ns\times (ns+1)$ matrix over $\k$. By elementary column operations over $\k$ it follows that we can obtain a column of zeroes. Performing the same operations on the matrix $B$ gives
us a nonzero torsion element whose last $s$ rows are zeroes, and whose entries are in $R$. These necessarily also represent a torsion element in $\Omega_R$, since they
are a syzygy of $\frac{dx_1}{dt},\dots,\frac{dx_n}{dt}$. If this torsion element were zero in $\Omega_R$, it would also be zero in $\Omega_S$, since the presentation of $\Omega_S$ contains the
Jacobian matrix associated to $R$. \end{proof}
As an immediate application, we generalize a result of Scheja (\cite{scheja1970differentialmoduln}), proved by many researchers, who proved Berger's conjecture in the case $R$ is
quasi-homogeneous. If $R$ is quasi-homogeneous, so too is $S$, so the next result is strictly stronger:
\begin{theorem}\label{quasihomogeneous}
Let $R=\k\ps{\alpha_1t^{a_1},\dots,\alpha_nt^{a_n}}$ with conductor $\mathfrak{C}_R$. Construct $S=R\left[\frac{\mathfrak{C}_R}{x_1}\right]=R[T_1,\dots,T_s]$ as in \Cref{kernelofnew_proof}. If $S$ is quasi-homogeneous, then $\tau(\Omega_R)$ is nonzero.
\end{theorem}
\begin{proof} Since $S$ is quasi-homogeneous,
Scheja in \cite{scheja1970differentialmoduln} showed that $\mu(\tau(\Omega_R))\geq {n+s\choose 2}$ where $\operatorname{edim} S=n+s$. After possibly a change of generators, we have that \\ $\Omega_S\twoheadrightarrow \mathfrak{n}=(x_1,\dots,x_n,T_1,\dots,T_s), dx_i\rightarrow x_i,dT_j\rightarrow T_j$. (This map is well-defined because it is induced from the Euler derivation, see \cite[1.5]{Kunzbook}) Let $\tau$ be a torsion element in $\Omega_S$, and write $\tau=\sum_{i=1}^n r_idx_i+\sum_{j=1}^s r'_jdT_j$. Since $\tau\rightarrow 0$ under the above map, we have $\sum_{i=1}^n r_ix_i+\sum_{j=1}^s r'_jT_j=0$. Thus none of the $r'_j$ can be units in $S$, else it would be a contradiction to $x_1,\dots,x_n,T_1,\dots,T_s$ being a minimal generating set for $\mathfrak{n}$. Now since $\Omega_S$ has ${n+s\choose 2}$ $\k$-linearly independent torsion elements and none of the coefficients of $dT_j$ in the description of the torsion elements are units, we can pull back a $\k$-linear combination of these torsion elements to a torsion element in $\Omega_R$ (by \Cref{thm on one more torsion}).
\end{proof}
\begin{example}
Let $R=\k\ps{ t^5,t^8+t^{11},t^9+t^{11},t^{12}+t^{11}}$. Macaulay2 computations show that the conductor $\mathfrak{C}_R=(t^{13})\overline{R}$ and $S=R\left[\frac{\mathfrak{C}_R}{x_1}\right]$ is the monomial curve $\k\ps{t^5,t^8,t^9,t^{11},t^{12}}$. The embedding dimension of $S$ equals $5$ which is one more than that of $R$. Using the previous theorem we see that the torsion $\tau(\Omega_R)\neq 0$. In fact the deviation $d(R)=\mu(I)-\operatorname{edim} R+1=8-4+1=5$. The defining ideal of $R$ is of height three and is not a Gorenstein ideal.
\end{example}
\section{Main Results}\label{main results}
In this section we prove some of the main results of this article. In the rest of this section we assume that $\mathfrak{C}_R\subseteq \mathfrak{m}^2_R$ and $S=R\left[\frac{\mathfrak{C}_R}{x_1}\right]$ as in \Cref{kernelofnew_proof}. Recall that the $\k$-dimension of $S/R$ is called the reduced type of $R$, denoted by $s$. We proved that $s$ is always at most the type of $R$. In particular, if $R$ is Gorenstein, then $s = 1$.
\begin{proposition}\label{last row unit}
Suppose $S=R[\frac{\mathfrak{C}_R}{x_1}]$ and $\tau\in \Omega_S$. If $\tau=\sum r_idx_i+\sum_{j=1}^sr_{n+j}dT_j, r_i\in S$ such that $r_{n+j}$ is a unit in $S$ for some $j$, then for all $1\leq i\leq n$, $x_i\tau\neq 0$.
\end{proposition}
\begin{proof}
Fix $i\leq n$. Let $J=\langle x_i^2,x_iT_j,T_j^2, x_1,\dots,x_{i-1}, x_{i+1},\dots,x_n,T_1,\dots,T_{j-1},T_{j+1},\dots,T_s\rangle$. Writing $\overline{(\,)}$ for images in $\Omega_{R/J}$, $\overline{x_i\tau}=\overline{x_ir_i}dx_i+\overline{x_ir_{n+j}}dT_j$. If $r_i$ is not a unit in $S$, then $x_ir_i\in \mathfrak{m}_S^2\subseteq J$. Thus $\overline{x_ir_i}dx_i=0$. If $r_i$ is a unit then $\overline{x_ir_i}dx_i=0$ as $x_i^2\in J$. Thus we have $\overline{\tau}=\overline{x_ir_{n+j}}dT_j$ which is nonzero in $\Omega_{S/J}$ by \cite[Proposition 2.6]{ABC1}.
\end{proof}
\begin{corollary}
Under the hypothesis of the above theorem, for every $\tau=\sum r_idx_i+\sum_{j=1}^sr_{n+j}dT_j\in 0:_{\Omega_S}x_1$, $r_{n+j}\in S$ cannot be a unit.
\end{corollary}
\begin{proof}
Follows from \Cref{last row unit}.
\end{proof}
\begin{lemma}\label{typecomparisonlemma}
Let $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$ and $S=R[\frac{\mathfrak{C}}{x_1}]$ with $\operatorname{edim}(S)=n+s$ where $s$ is the reduced type of $R$. Then $$\operatorname{type}(S)\leq \operatorname{type}(R)+s(n-1).$$
\end{lemma}
\begin{proof}
Using \Cref{canonicalmoduledescription}, we get the following short exact sequence
$$0\to \omega_S\to \omega_R\to \k^s\to 0$$ and tensoring this sequence with $\k$, we get $$\operatorname{Tor}^R_1(\k,\k^s)\to \omega_S\otimes_R \k\to\omega_R\otimes_R \k\to \k^s\to 0.$$ Comparing the length of the modules appearing this short exact sequence now yields, $\dim_\k\operatorname{Tor}^R_1(\k,\k^s)+\operatorname{type}(R)\geq \operatorname{type}(S)+s$ (\cite[Theorem 3.3.11]{bruns_herzog_1998}), and hence $$\operatorname{type}(S)\leq ns-s+\operatorname{type}(R)=\operatorname{type}(R)+s(n-1).$$
\end{proof}
\begin{proposition}\label{powerofm} Let $R$ and $S$ be as in Lemma \ref{typecomparisonlemma}. Then $\mathfrak{m}_S^k\subseteq x_1S$ if and only if $\mathfrak{m}_R^k\subseteq (x_1,\mathfrak{C}_R)R$.
\end{proposition}
\begin{proof} First we observe that $x_1S = (x_1,\mathfrak{C}_R)R$. Since $\mathfrak{m}_R\subseteq \mathfrak{m}_S$, one implication follows at once. To prove the reverse implication, assume that
$\mathfrak{m}_R^k\subseteq (x_1,\mathfrak{C}_R)R$. It then suffices to prove that $\mathfrak{m}_S^k\subseteq \mathfrak{m}_R^k + \mathfrak{C}_R$. A typical term in the expansion of $\mathfrak{m}_S^k = (\mathfrak{m}_R + \frac{\mathfrak{C}}{x_1})^kS$,
which is not $\mathfrak{m}_R^k$, is of the form $\mathfrak{m}_R^i(\frac{\mathfrak{C}}{x_1})^{k-i}$ for $0\leq i\leq k-1$. Since $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$ and $k-i\geq 1$, this term is contained in
$$\mathfrak{m}_R^i(\mathfrak{m}_R^{2k-2i-2})(\frac{\mathfrak{C}_R}{x_1})\subseteq \mathfrak{C}_R$$ as $\mathfrak{m}_R\mathfrak{C}_R = x_1\mathfrak{C}_R$.
\end{proof}
\begin{lemma}[\cite{Guttes1990}]\label{gutteslemma}
Let $(R,\mathfrak{m}_R,\k)$ be a one dimensional complete local reduced $\k$-algebra with $\operatorname{char}(\k)=0$ and embedding dimension $n\geq 3$. Suppose $y$ is a non-zero divisor such that $\operatorname{edim}(R/R.y)=n-1$
\begin{itemize}
\item[$a)$] If $\mathfrak{m}_R^4\subseteq R.y$, then \[\ell(0:_{\Omega_R} y)\geq \frac{(n-2)(n-1)}{2}.\]
\item[$b)$] If $\mathfrak{m}_R^5\subseteq R.y$, then \[\ell(0:_{\Omega_R} y)\geq \frac{(n-2)(n-1)}{2}-\operatorname{type}(R).\]
\end{itemize}
\end{lemma}
\begin{proof}
For $a)$, we refer the reader to the proof of \cite[Satz 4]{Guttes1990}. For $(b)$, see the proof of \cite[Anmerkung, Page 506-507]{Guttes1990}.
\end{proof}
\begin{theorem}\label{m^4 in C}
Suppose $\mathfrak{m}_R^4\subseteq (\mathfrak{C}_R,x_1)$ and $n(n - 3)\geq 2s$ where $s$ is the reduced type of $R$, then $\tau(\Omega_R)\neq 0$.
\end{theorem}
\begin{proof}
We first construct $S=R\left[\frac{\mathfrak{C}_R}{x_1}\right]$ with $\operatorname{edim} S=n+s$ where $\operatorname{edim} R=n$. By \Cref{powerofm}, we have that $\mathfrak{m}^4_S\subseteq x_1S$. Since $x_1$ is a nonzero divisor on $S$, by definition of torsion submodule we have $0:_{\Omega_S}x_1\subseteq \tau(\Omega_S)$. Now using \Cref{gutteslemma}(a) with $R=S$ and $y=x_1$, we get that
\begin{align*}
\ell(\tau(\Omega_S))\geq \ell(0:_{\Omega_S}x_1)\geq\frac{(n+s-2)(n+s-1)}{2},
\end{align*}
where the latter number is more than $ns+{s\choose 2}+1$ when $n(n-3)\geq 2s$. The result now follows from \Cref{thm on one more torsion}.
\end{proof}
\begin{theorem}\label{mainthm}
Let $S=R\left[\frac{\mathfrak{C}}{x_1}\right]$ with $\operatorname{edim} S=n+s$ and assume that $$\displaystyle\text{type}(R)\leq \frac{n^2-3n-2ns}{2}.$$ If $\mathfrak{m}_R^5\subseteq \mathfrak{C}_R+ x_1R$, then $\tau(\Omega_R)\neq 0$.
\end{theorem}
\begin{proof}
Since $\mathfrak{C}_R\subseteq \mathfrak{m}^2$,
using \Cref{powerofm}, we immediately obtain that $\mathfrak{m}_S^5\subseteq x_1S$.
Since $x_1$ is a nonzero divisor on $S$, by definition of torsion submodule, $0:_{\Omega_S}x_1\subseteq \tau(\Omega_S)$. By \Cref{gutteslemma}(b) with $R=S$ and $y=x_1$, we get that $$\ell(\tau(\Omega_S))\geq \ell(0:_{\Omega_S}x_1)\geq \frac{(n+s-2)(n+s-1)}{2}-\text{type}(S).$$
Here we used that the embedding dimension of $S$ is $n+s$, which follows from the presentation of $S$ given in Remark \ref{defining matrix of S}. All torsion elements in $(0:_{\Omega_S}x_1)$ have non-units in the last row by \Cref{last row unit}. It follows from Theorem \ref{thm on one more torsion} that it suffices to prove
$$\frac{(n+s-2)(n+s-1)}{2}-\text{type}(S)\geq ns+{s\choose 2}+1.$$
Using \Cref{typecomparisonlemma}, it suffices to prove that
$$\frac{(n+s-2)(n+s-1)}{2}-\text{type}(R)-s(n-1)\geq ns+{s\choose 2}+1$$ or
$$\text{type}(R)\leq \frac{(n+s-2)(n+s-1)}{2}-2ns+s-{s\choose 2}-1,$$
which simplifies to our assumption on the type. \end{proof}
\begin{corollary}
Let $R$ be of reduced type one such that $\text{type}(R)\leq {n\choose 2}-2n$. If $\mathfrak{m}_R^5\subseteq \mathfrak{C}_R+ x_1R$, then $\tau(\Omega_R)\neq 0$.
\end{corollary}
\begin{proof} Set $s = 1$ in the above theorem.
\end{proof}
\begin{corollary}\label{m^6 in C}
Let $R$ be Gorenstein and $n=\operatorname{edim} R\geq 6$. If $\mathfrak{m}_R^6\subseteq x_1R$, then $\tau(\Omega_R)\neq 0$.
\end{corollary}
\begin{proof} We may assume that $\mathfrak{C}_R\subseteq \mathfrak{m}_R^2$, because we have already shown that the torsion is nonzero in case the inclusion does not hold (\Cref{valuation more than conductor}).
Since $R$ is Gorenstein, $R$ is also of reduced type one (\Cref{canonicalmoduledescription}). The condition that $\mathfrak{m}_R^6\subseteq x_1R$ implies that
$\mathfrak{m}_R^5\subseteq (x_1R:\mathfrak{m}_R)$. However, it is always true that $\mathfrak{C}_R + x_1R\subseteq x_1:\mathfrak{m}_R$, since $x_1\mathfrak{C}_R = \mathfrak{m}_R\mathfrak{C}_R$. As the conductor can never lie inside a proper principal ideal (follows, for instance, from \cite[Corollary 2.6]{maitra2020partial}) and since $R$ is Gorenstein, we obtain that $\mathfrak{C}_R + x_1R = x_1:\mathfrak{m}_R$. Therefore, $\mathfrak{m}_R^5\subseteq \mathfrak{C}_R+ x_1R$. The inequality $1=\text{type}(R)\leq {n\choose 2}-2n$ is also satisfied as $n\geq 6$. Thus all the conditions of the previous theorem hold, and the result now follows.
\end{proof}
\begin{remark} As mentioned in the above proof, it is always true that $\mathfrak{C}_R + x_1R\subseteq x_1:\mathfrak{m}_R$, since $x_1\mathfrak{C}_R = \mathfrak{m}_R\mathfrak{C}_R$. If equality holds, then $s$ is equal to the type of $R$, and the condition
that $\mathfrak{m}_R^5\subseteq \mathfrak{C}_R + x_1R$ is equivalent to the condition that $\mathfrak{m}_R^6\subseteq x_1R$. This case of ``maximal" reduced type gives a further extension of
the work of G\"uttes. This maximality occurs if $R$ is Gorenstein, but it can also occur in other cases. For one such example, using Macaulay2, we check that
$R = \k[[t^{10},t^{11}+t^{16},t^{12}+t^{16},t^{13}+t^{16}]]=k[[x,y,z,w$]] has conductor $\mathfrak{C}_R = t^{20}\overline{R}$ and $(x):\mathfrak{m}_R=(x,w^2,zw,yw,z^2,yz,y^2)= (x,\mathfrak{C}_R)$.
\end{remark}
\bibliographystyle{siam}
|
\section{Introduction}
\label{sec:intro}
The neutrino oscillation experiments~\cite{Super-Kamiokande:1998kpq,MINOS:2006foh,MINOS:2011amj,PhysRevLett.108.131801,Ling:2013fta,Kim:2013sza} have proved that neutrinos in standard model (SM) have very tiny masses.
Because of the lack of right-handed neutrinos in the SM, the Dirac mass terms cannot be formed as other fermions in the Lagrangian and the SM needs to be extended to explain their non-zero masses.
One important solution is the seesaw mechanism~\cite{FRITZSCH1975256,Minkowski:1977sc,Yanagida:1979as,Sawada:1979dis,Mohapatra:1979ia,Glashow:1979nm,GellMann:1980vs,Keung:1983uu,Foot:1988aq,Mohapatra:1986aw,MAGG198061}, where new gauge singlet right-handed neutrinos $N_{R}$ are introduced and masses of active neutrinos are generated by mixing SM left-handed neutrinos $\nu_L$ with right-handed neutrinos $N_{R}$, resulting in heavy mass eigenstates $N$ that have small SM $\nu_L$ components.
Therefore, searches for heavy neutrinos are crucial to verify the seesaw mechanism and explore the origin of neutrino masses.
At colliders, such heavy neutrinos are usually also called heavy neutral leptons and are extensively searched relying on their effective couplings to SM gauge bosons via their mixing with SM neutrinos.
In theory, the production cross section, decay width, and lifetime of $N$ depend on its mass $m_N$ and the parameter $|V_{\ell N}|^2$ which is related to the matrix element describing the mixing of $N$ with the SM neutrino of flavor $\ell$.
Therefore, limits for such searches are usually given in the plane of the mixing parameter $|V_{\ell N}|^2$ vs. the heavy neutrino mass $m_N$.
Summaries of collider searches of heavy neutrinos
can be found in Refs.~\cite{Atre:2009rg,Deppisch:2015qwa,Das:2015toa,Cai:2017mow,Das:2017rsu,Bolton:2019pcu} and references therein.
Refs.~\cite{CMS:2018iaf, CMS:2018jxx, CMS:2021lzm, ATLAS:2019kpx, LHCb:2020wxx, NA62:2020mcv, Belle:2013ytx, T2K:2019jwa}
are recent experimental studies on heavy neutrino searches.
Among them, the CMS collaboration has analyzed the data with center-of-mass energy $\sqrt{s} =$ 13 TeV and an integrated luminosity of 35.9 $\rm fb^{-1}$~\cite{CMS:2018iaf,CMS:2018jxx} and 137 $\rm fb^{-1}$\cite{CMS:2021lzm}.
Ref.~\cite{CMS:2018iaf} searched for a heavy Majorana neutrino in the trilepton signal process $p p \to W^{(*)} \to \ell\, N \to \ell\, (\ell W^{(*)}) \to \ell\, (\ell \ell \nu)$
and limits are applied on both $|V_{e N}|^2$ and $|V_{\mu N}|^2$ for the heavy neutrino mass range between 1 GeV and 1.2 TeV.
Ref.~\cite{CMS:2018jxx} searched for a heavy Majorana neutrino in the same-sign dilepton channel $ W^{(*)} \to \ell^{\pm}\, N \to \ell^{\pm}\, (\ell^\pm W^{(*)}) \to \ell^\pm\, (\ell^\pm j j)$.
The upper limits are set on $|V_{e N}|^2$, $|V_{\mu N}|^2$, and $|V_{eN }V_{\mu N }^*|^2/(|V_{eN }|^2+|V_{\mu N }|^2)$ for $N$ masses between 20 and 1600 GeV.
Ref.~\cite{CMS:2021lzm} searched for a long-lived heavy neutrino of either Majorana or Dirac type in the signal process $p p \to W \to \ell\, N \to \ell\, (\ell W^* / \nu Z^*) \to \ell \ell \ell \nu$. The final state consists of three leptons among which two leptons form a displaced vertex with respect to the primary proton-proton collision vertex and the third lepton emerges from the primary vertex. Limits are applied on $|V_{e N}|^2$ and $|V_{\mu N}|^2$ in the mass range between 1 and 15 GeV.
In Ref.~\cite{ATLAS:2019kpx}, a heavy Majorana neutrino was investigated by the ATLAS collaboration in the similar signal process to that in Ref.~\cite{CMS:2018iaf}.
Data at $\sqrt{s} =$ 13 TeV and integrated luminosity of 36.1 (32.9) $\rm ab^{-1}$ are analysed for the prompt (displaced) leptonic decay case. Constraints on $|V_{e N}|^2$ and $|V_{\mu N}|^2$ are set for heavy neutrino mass in the range of 4.5 to 50 GeV .
The LHCb collaboration searched for a heavy Majorana neutrino in the signal process $W^+ \to \mu^+\, N \to \mu^+\, (\mu^{\pm} j j)$~\cite{LHCb:2020wxx}. Data corresponding to an integrated luminosity of 3.0 $\rm fb^{-1}$ and center-of-mass energies of 7 and 8 TeV are analysed and upper limits on $|V_{\mu N}|^2$ are set to be $\sim 10^{-3}\,\, (10^{-4})$ in the mass range from 5 to 50 GeV for the lepton number conserving (violating) case.
The NA62 collaboration searched for $N$ produced from $K^+ \to e^+ N$ decays and placed the upper limit on $|V_{e N}|^2 \sim 10^{-9}$ in the mass range 144$-$462 MeV~\cite{NA62:2020mcv}.
Heavy neutrinos from B-meson decays are investigated by the Belle collaboration and upper limits are set on $|V_{e N}|^2$, $|V_{\mu N}|^2$, $|V_{e N}|\,|V_{\mu N}|$ in the mass range 0.5$-$5.0 GeV~\cite{Belle:2013ytx}.
The T2K Collaboration searched for heavy neutrinos from kaon decays and constrained mixing parameters $|V_{\ell N}|^2$ with $\ell = e, \mu, \tau$ for $m_N$ between 140 and 493 MeV~\cite{T2K:2019jwa}.
Moreover, for $N$ produced from meson decays, the SHiP collaboration's search prospect for long-lived neutrinos predicts strong sensitivity
on $|V_{\ell N}|^2$ with $\ell = e, \mu, \tau$ flavors for $m_N$ in the range $0.1-5.8$ GeV~\cite{SHiP:2018xqw}.
The combination of electroweak precision observables and lepton flavor violating decays can also set constraints indirectly on mixing parameters $|V_{\ell N}|^2$, especially when $m_N$ is larger than 80 GeV~\cite{Chrzaszcz:2019inj}.
In this article, we develop the search strategy for a heavy Majorana neutrino via the lepton number violation (LNV) signal process of $p\, e^- \to \mu^+ jjj$ at future electron-proton colliders, the Large Hadron electron Collider (LHeC)~\cite{Klein:2009qt,LHeCStudyGroup:2012zhm,Bruening:2013bga,Klein:2016uwv,LHeC:2020van,Holzer:2021dxw}
and the electron-hadron mode of the Future Circular Collider (FCC-eh)~\cite{Zimmermann:2014qxa,Klein:2016uwv,TOMAS2016149,FCC:2018byv,Holzer:2021dxw}.
We consider the LHeC (FCC-eh) running with an electron beam energy of 60 GeV and a proton beam energy of
7 (50) TeV, which corresponds to $\sqrt{s} =$ 1.3 (3.5) TeV.
The integrated luminosities are assumed to be 1 and 3 $\rm ab^{-1}$ at the LHeC and FCC-eh, respectively.
Compared with the proton colliders, such as the high-luminosity LHC (HL-LHC), the centre-of-mass energy of $ep$ colliders is lower. However, due to lack of gluon-exchange diagrams, the SM QCD backgrounds, which are dominant at $pp$ colliders, are much smaller at $ep$ colliders. Besides, the number of additional interactions in the same event (pileup) is negligible at $ep$ colliders, while it is expected to be very large at the HL-LHC.
Furthermore, heavy neutrinos with mass above 100 GeV can still be produced on-shell from the $t$-channel exchange of $W$-boson at $ep$ colliders, while at the HL-LHC such heavy neutrinos are produced from off-shell $W$- or $Z$-boson processes with limited cross section.
Therefore, future $e p$ colliders could be complementary to the $p p$ collider when searching for beyond standard model (BSM) physics scenarios, particularly for heavy neutrinos.
Ref.~\cite{Azuelos:2021ese} and references therein have reviewed BSM physics searches at the LHeC and FCC-eh, while phenomenology studies on heavy neutrino searches at $e p$ colliders can be found in Refs.~\cite{BUCHMULLER1991465,Buchmuller:1991tu,Buchmuller:1992wm,Ingelman:1993ve,Liang:2010gm,Blaksley:2011ey,Duarte:2014zea,Mondal:2016kof,Antusch:2016ejd,Lindner:2016lxq,Li:2018wut,Das:2018usr,Antusch:2019eiz,Cottin:2021tfo}.
Among them, Ref.~\cite{Duarte:2014zea} investigated heavy Majorana neutrinos produced in an effective Lagrangian approach at the LHeC.
For the LHeC, they considered a 7 TeV proton beam colliding with an electron beam of two energies: $E_e = $ 50 and 150 GeV.
The events are simulated at the parton level and kinematic cuts are applied to reduce the background.
Limits are placed on the neutrino mass and the effective coupling.
Ref.~\cite{Mondal:2016kof} explored heavy neutrinos at the LHeC in the context of an inverse-seesaw model.
The production cross section of various signals of $N j$, $N j W^-$ and $e^- j W^-$ are calculated with and without $80\%$ left-polarized electron beam.
The events are simulated at the parton level and kinematic cuts are applied to reduce the background.
The required integrated luminosities are estimated to achieve a 3-$\sigma$ statistical significance for two different heavy neutrino masses of 150 and 400 GeV.
Ref.~\cite{Li:2018wut} searched for heavy Majorana neutrino via the signal processes $p\,e^- \to e^- \mu^{\pm} \mu^{\pm} + X$ and $p\,e^- \to \nu_e \mu^- \mu^{\pm} + X$.
The events are simulated including detector smearing effects and kinematic cuts are applied to reduce the background.
The required integrated luminosities are estimated for heavy neutrino masses in the range of 100 and 1000 GeV.
Ref.~\cite{Das:2018usr} investigated a heavy neutrino via the signal process $p\,e^- \to j N \to j (e^\pm W^\mp) \to j (e^\pm J)$, where $J$ is a fat-jet from a highly boosted $W-$boson.
The events are simulated at the parton level and passed through selection cuts to reduce the background.
Bounds on $|V_{e N}|^2$ are placed for $m_N$ in the range of 400 to 1000 GeV.
Ref.~\cite{Antusch:2019eiz} probed heavy neutrinos via the lepton flavor violating signal process $p\,e^- \to j N \to j (\mu^- W^+) \to \mu^- + 3 j$ at the LHeC and FCC-eh. Background processes include $j e^- V V$ and $j \nu_e V V$ where $V = Z, W^\pm$ and $V V \to (j j) (\mu^- \mu^+ / \mu^- \bar{\nu}_\mu)$.
Limits on the mixing parameters $|\theta_e \theta_\mu|$ are placed for the heavy neutrino mass in the range from 100 to 1000 GeV.
We note that this work is different from all previous phenomenology studies due to the combination of following aspects:
(i) we consider the LNV signal process $p\,e^- \to \mu^+ + jjj$ assuming $|V_{\ell N}|^2 = |V_{e N}|^2 = |V_{\mu N}|^2$ in the context of a simplified Type-I seesaw model;
(ii) SM background includes four inclusive processes listed in Table~\ref{tab:crsc};
(iii) for both the signal and background event simulation, we utilize the program chain including the event generator, parton shower, hadronization and detector effects;
(iv) for the LHeC (FCC-eh), we consider a 60 GeV electron beam colliding with a proton beam of 7 (50) TeV energy and an integrated luminosity of 1 (3) $\rm ab^{-1}$;
(v) we apply pre-selection criteria and perform multi-variate analysis based on machine-learning to reject the background;
(vi) The 2- and 5-$\sigma$ limits on the mixing parameter $|V_{\ell N}|^2$ are predicted for the heavy neutrino mass in the range 10$-$1000 GeV.
The article is organised as follows.
Sec.~\ref{sec:sig} presents the data simulation and the cross section of the signal process.
Sec.~\ref{sec:bkg} describes the SM background processes and our search strategy.
The analysis details and limits on the mixing parameters $|V_{\ell N}|^2$ at both the LHeC and FCC-eh are shown in Sec.~\ref{sec:result}.
We summarize our study, and comment on the effect of long-lived cases of heavy neutrinos and the positron signal final state in Sec.~\ref{sec:sum}.
\section{The LNV signal}
\label{sec:sig}
To simplify the analyses, we consider the Type-I seesaw model and assume that there is only one generation of heavy neutrinos $N$ which mixes with active neutrinos of electron and muon flavours with the same mixing parameters, i.e. $|V_{\ell N}|^2 = |V_{e N}|^2 = |V_{\mu N}|^2$.
We also assume that $N$ decays promptly in this study.
As shown in Fig.~\ref{fig:Feynman}, the heavy Majorana neutrino $N$ can be produced via the $t$-channel exchange of $W$-boson at $ep$ colliders, and finally decay into $\mu^+$ plus three jets.
The lepton number of this process changes from $+1$ to $-1$, so it violates the conservation of lepton number.
\begin{figure}[h]
\centering
\includegraphics[width=5cm,height=3.5cm]{fig_feynman13.png}
\caption{
The production process of the LNV signal via a Majorana heavy neutrino $N$ at $ep$ colliders.
}
\label{fig:Feynman}
\end{figure}
For the data simulation, we implement the Universal FeynRules Output model file~\cite{Alva:2014gxa,Degrande:2016aje} which extends the SM with additional heavy neutrinos interacting with active neutrinos, into the MadGraph5~\cite{Alwall:2014hca} to generate the signal events.
Similar to our previous work~\cite{Azuelos:2019bwg}, the Pythia6~\cite{Sjostrand:2006za} program is modified to perform the parton showering and hadronization for $ep$ colliders, while the configuration card files~\cite{Delphes_cards} for the LHeC and FCC-eh detectors are implemented to the Delphes program~\cite{deFavereau:2013fsa} to complete the detector simulation.
\begin{figure}[]
\centering
\includegraphics[width=8cm,height=6cm]{fig_crscvsmu10.pdf}
\caption{
The production cross section of the LNV signal $p\, e^- \to \mu^+ jjj$ via the heavy Majorana neutrino $N$ for varying heavy neutrino masses $m_N$ at the LHeC and FCC-eh with $|V_{\ell N}|^2 = 10^{-4}$.
}
\label{fig:crs}
\end{figure}
To maintain consistency throughout this study, the production cross sections calculated by MadGraph5 are used to estimate the number of events for both signal and background processes.
In Fig.~\ref{fig:crs}, we plot the cross sections of the LNV signal $p\, e^- \to \mu^+ jjj$ via the heavy Majorana neutrino $N$ as a function of the heavy neutrino mass $m_N$ at the LHeC and FCC-eh, where the mixing parameter $|V_{\ell N}|^2$ is fixed to be $10^{-4}$.
For large $m_N$, cross sections at the LHeC decrease much faster than those at the FCC-eh.
This behaviour can be understood from the parton distribution function (PDF) of the proton.
In this study, heavy neutrinos are produced from the $t$-channel $W$-boson exchange process $q\,e^- \to j N$.
In order to produce heavy neutrinos with large mass, momenta of incoming quarks need to be large enough, so that the centre of mass energy $\sqrt{s}= 2\sqrt{ E_e E_q}$ is larger than $m_N$.
Considering a quark carries a fraction $x$ of the longitudinal momentum of a proton, only quarks with $x \gtrsim m_N^2 / (4 E_e E_p)$ can contribute to the signal production cross section.
Since the parton density function $f(x)$ of the quark decreases rapidly for large $x$ values~\cite{Placakyte:2011az},
when $m_N$ is very large, the number of quarks which satisfies the condition $x \gtrsim m_N^2 / (4 E_e E_p)$ becomes tiny.
This leads to the rapid decrease in the production cross section for large $m_N$.
\section{background and search strategy}
\label{sec:bkg}
Since the signal process $p\, e^- \to \mu^+ jjj$ violates the lepton number conservation explicitly, it has little SM background in theory.
Considering the signal final state contains one positive charged muon plus multi-jets,
there are mainly four SM background processes, which we label as ``B1 - B4" in this article.
We list their production cross sections in Table~\ref{tab:crsc} .
They can contribute to the background when the final state $e^-$ and/or $\mu ^-$ are undetected. Among these four background processes, $\mu^+ \mu^- e^- j j j$ and $ \mu^+ \nu_\mu e^- j j j$ have large cross sections and are more difficult to eliminate.
We note that the considered four background processes have both QED and QCD interactions.
Processes with pure QED interactions are checked and the sum of their cross sections are found to be only a factor of about 1/90 (1/45) of the considered background at the LHeC (FCC-eh). Because they are much smaller and our computing resources are limited, we do not include the background processes with only QED interactions.
\begin{table}[h]
\centering
\begin{tabular}{cccc}
\hline
\hline
& $\sigma$ [pb] & LHeC & FCC-eh \\
\hline
B1 & $\mu^+ \mu^- e^- j j j$& 0.58 & 2.1 \\
B2 & $\mu^+ \mu^- \nu_e j j j$& $8.6{\mkern-2mu\times\mkern-2mu} 10^{-2}$ & 0.39 \\
B3 & $\mu^+ \nu_\mu e^- j j j$&0.28 & 1.6 \\
B4 & $\mu^+ \nu_\mu \nu_e j j j$& $8.1 {\mkern-2mu\times\mkern-2mu} 10^{-6}$ & $9.3 {\mkern-2mu\times\mkern-2mu} 10^{-5}$ \\
\hline
\hline
\end{tabular}
\caption{
The production cross sections of dominant background processes at the LHeC and FCC-eh.
}
\label{tab:crsc}
\end{table}
The final state muon can also come from tau decays.
However, it will not contribute too much to the background because of the following two reasons: (i) the small tau to muon branching ratio; (ii) the leptonic decay of taus produce neutrinos, resulting in large missing energy in the final state, and cannot pass our missing energy pre-selection cut.
We checked four background processes $\tau^+ \tau^- e^- j j j$, $\tau^+ \tau^- \nu_e j j j$, $\tau^+ \nu_\tau e^- j j j$, and $\tau^+ \nu_\tau \nu_e j j j$ at the LHeC (FCC-eh) and found that after pre-selection the event rate of total background increased only by a factor of $6\%$ ($11\%$).
After performing the full analysis described below, we find that after adding the tau background, the limit on $|V_{\ell N}|^2$ changed from $3.6\,\, (1.10) \times10^{-6}$ to $3.8\,\, (1.12) \times10^{-6}$ for the benchmark $m_N$ = 120 GeV.
Since the effects on the limits are very small, we did not add the tau background in this study.
We apply the following pre-selection cuts to select the signal and reject the background events at the first stage.
(i) Exactly one muon with positive charge, i.e. $N(\mu^+) = 1$ and transverse momentum $p_T(\mu) >$ 5 GeV;
events with final state electron(s) or tau(s) are vetoed.
(ii) All jets are sorted in descending order according to their transverse momenta and we require at least three jets, i.e. $N(j)\geq 3$; for the $p_T$ thresholds of jets, when heavy neutrino mass is below 80 GeV, the $p_T$ of the first three leading jets are required to be greater than 10 GeV, while when masses are above 80 GeV, we require the first two leading jets to have $p_T$ greater than 20 GeV and the third one to have $p_T$ greater than 10 GeV.
(iii) Since both the final state neutrinos and the missing of leptons contribute to the missing energy, the background has much larger missing energy compared with the signal and a pre-selection of missing energy $\slashed{E}_T < 10 $ GeV is applied to reject the background.
For the signal data simulation, we vary the heavy neutrino mass $m_N$ from 10 to 1000 GeV and generate 0.3 million signal events at the LHeC and FCC-eh for each $m_N$.
Due to limited computational resources, we are not able to generate huge number of events for every background processes.
The number of simulated events for each background process is determined according to its importance in reducing the statistical uncertainty on final limits.
For the background, we generate 2.1 (2.0) million $\mu^+ \mu^- e^- j j j$, 10.5 (6.0) million $\mu^+ \mu^- \nu_e j j j$, 27.4 (24.6) million $\mu^+ \nu_\mu e^- j j j$ and 6.0 (6.4) million $\mu^+ \nu_\mu \nu_e j j j$ events at the LHeC (FCC-eh), respectively.
In Table~\ref{tab:cutmn120}, we show the number of events for the signal with benchmark $m_N=120$ GeV and four background processes after applying the pre-selection cuts (i)-(iii) sequentially described above.
\begin{table}[h]
\centering
\begin{tabular}{ccccccc}
\hline
\hline
& & signal & B1 & B2 & B3 & B4 \\
\hline
\multirow{4}{*}{LHeC}
& initial &$1.2\mltp10^{3}$ & $5.8\mltp10^{5}$ & $8.6\mltp10^{4}$ & $2.8\mltp10^{5}$ & 8.1 \\
& (i) & $1.1\mltp10^{3}$ & $2.6\mltp10^{3}$ & $4.0\mltp10^{3}$ & $1.9\mltp10^{4}$ & 6.2 \\
& (ii) & 853 & 799 & $1.3\mltp10^{3}$ & $6.5\mltp10^{3}$ & 4.4 \\
& (iii) & 702 & 699 & 9.3 & 154 & 0.1 \\
\hline
\multirow{4}{*}{FCC-eh}
& initial & $1.0\mltp10^{4}$ & $6.2\mltp10^{6}$ & $ 1.2\mltp10^{6} $ & $4.9\mltp10^{6} $ & 278 \\
& (i) & $8.8\mltp10^{3}$ &$1.3\mltp10^{4}$ & $6.1\mltp10^{4}$ & $2.8\mltp10^{5}$ & 118 \\
& (ii) & $7.2\mltp10^{3}$ & $4.1\mltp10^{3}$ & $2.3\mltp10^{4}$ & $1.2\mltp10^{5}$ & 99 \\
& (iii) & $5.5\mltp10^{3}$ & $2.8\mltp10^{3}$ & 125 & $3.2\mltp10^{3}$ & 1.2 \\
\hline
\hline
\end{tabular}
\caption{
The number of events for the signal with benchmark $m_N$ = 120 GeV and four background processes after applying pre-selection cuts (i)-(iii) sequentially. The numbers correspond to the LHeC and FCC-eh with 1 and 3 $\rm ab^{-1}$ integrated luminosity, respectively.
}
\label{tab:cutmn120}
\end{table}
To further reject the background, we input the following nineteen observables into the TMVA~\cite{Hocker:2007ht} package to perform the multivariate analysis (MVA).
\begin{enumerate}[label*=\Alph*.]
\item The four-momenta of the final state muon:
$E(\mu)$, $p_{x}(\mu)$, $p_{y}(\mu)$, $p_{z}(\mu)$.
\item The number of jets $N(j)$ and the four-momenta of the first three leading jets:
$E(j_{1})$, $p_{x}(j_{1})$, $p_{y}(j_{1})$, $p_{z}(j_{1})$;
$E(j_{2})$, $p_{x}(j_{2})$, $p_{y}(j_{2})$, $p_{z}(j_{2})$;
$E(j_{3})$, $p_{x}(j_{3})$, $p_{y}(j_{3})$, $p_{z}(j_{3})$.
\item The magnitude and the azimuthal angle of the missing transverse momentum:
$\slashed{E}_T$, $\phi(\slashed{E}_T)$;
\end{enumerate}
\begin{figure}[h]
\centering
\includegraphics[width=4cm,height=3cm]{fig_120bdtl1.pdf}
\includegraphics[width=4cm,height=3cm]{fig_120bdtf1.pdf}
\caption{
Distributions of BDT responses for the signal with $m_{N}$ = 120 GeV (black, filled) and total SM background (red) at the LHeC (left) and FCC-eh (right).
}
\label{fig:BDTbench}
\end{figure}
The Boosted Decision Tree (BDT) method in the TMVA package is adopted to separate the background from the signal.
Fig.~\ref{fig:BDTbench} shows the BDT distributions for the total background and the benchmark signal with $m_{N}$ = 120 GeV at the LHeC and FCC-eh. The BDT distributions of the signal and background are well separate, which means that a BDT cut can be applied to reject the background.
Comparing left and right plots, one sees that the signal distributions of LHeC and FCC-eh are similar, but the background distribution of LHeC has an obvious peak, while the background distribution of FCC-eh is flatter. In addition, the distributions of signal and background overlap slightly less at the FCC-eh, indicating that the separation between the signal and background is better than LHeC.
Since the kinematics of the signal varies with $m_N$, the distributions of BDT response also change with $m_N$.
In Figs.~\ref{fig:BDTLHeC} and~\ref{fig:BDTFCC-eh}, we show the BDT distributions of four background processes and the signal corresponding to more representative heavy neutrino masses at the LHeC and FCC-eh.
We observe that as $m_N$ increases, the signal and background B1 ($\mu^+ \mu^- e^- j j j$) become more separate. However, the signal and the other three backgrounds overlap more and more as $m_N$ changes from 20 to 200 GeV, and then separate more and more as $m_N$ changes from 200 GeV to 1 TeV.
When $m_N =$ 1 TeV, BDT distributions of all background processes tend to be similar and are almost completely separate from the signal. However, the limit for $m_N =$ 1 TeV is still restricted by its small signal cross section.
We note that
when the kinematical distributions of background and signal are similar, it is difficult to distinguish between the signal and background,
leading to large overlap between their BDT distributions.
Therefore, the extent of separation between the BDT distributions of signal and background are determined by the degree of deviation of their kinematics.
We note that the above input observables including the four-momenta and angles are very basic and usually called low-level variables for the MVA analysis. One can also construct some complicated observables and input such high-level variables to perform the MVA analysis.
To compare the effects on the final limits by inputting different sets of variables, we construct twenty-nine high-level observables and show the distributions of eight representative ones for the signal with $m_N = 120$ GeV and four background processes at the LHeC and FCC-eh in Fig.~\ref{fig:HLobsLHeC} and Fig.~\ref{fig:HLobsFCCeh}.
The representative observables are ordered according to the separation between the signal and background. At both colliders, the best observables to separate the signal from background are $p_{\rm T}(\mu)$, $E(\mu)$ and $\eta(\mu)$. This is mainly because the two main background processes B1 ($\mu^+ \mu^- e^- j j j$, blue) and B3 ($\mu^+ \nu_\mu e^- j j j$, pink) are well separate from the signal for these observables. One observes that the invariant mass $M(\mu+j_{2}+j_{3})$ can also be a good discriminator. This is because $M(\mu+j_{2}+j_{3})$ has a sharp peak around $m_N$, which means that it can also be used to reconstruct the heavy neutrino mass.
\begin{figure}[h]
\centering
\includegraphics[width=4cm,height=3cm]{fig_120lvl1.pdf}
\includegraphics[width=4cm,height=3cm]{fig_120hvf1.pdf}
\caption{
Distributions of BDT responses for the signal with $m_{N}$ =120 GeV and total background when inputting high-level (HL, dashed) and low-level (LL, solid) observables at the LHeC (left) and FCC-eh (right).
}
\label{fig:BDTcomparehl}
\end{figure}
Fig.~\ref{fig:BDTcomparehl} shows the distributions of BDT responses for the signal with $m_{N}$ =120 GeV and total background when inputting high-level (HL, dashed) and low-level (LL, solid) observables at the LHeC (left) and FCC-eh (right).
One sees that the BDT distributions are similar between the low- and high-level cases.
Although some high-level observables seem to separate the signal better from the background than low-level observables, each high-level observable is not independent of each other and has correlations.
The MVA-BDT analysis combines the information from all input observables with correlations. Since two sets of low- and high-level observables contain similar information, their BDT distributions should be similar as well.
To estimate the effects on final limits, we complete the analyses for the benchmark $m_N =$ 120 GeV using high-level observables at the LHeC (FCC-eh) and find that BDT cut efficiencies change from $7.71\,\,(7.95) \times 10^{-1}$ to $7.27\,\,(7.20) \times 10^{-1}$ for the signal and from $1.03 \times10^{-1}\,\, (8.85 \times10^{-2})$ to $3.95\,\, (3.17) \times10^{-2}$
for the total background. The final 2-$\sigma$ limit decreases slightly from $3.6\times10^{-6} $ to $ 2.4\times10^{-6} $ for LHeC and from $ 1.1\times10^{-6} $ to $ 7.3\times10^{-7} $ for FCC-eh, respectively.
Because the changes in the final results are small and more computing resources are needed to input more observables, for simplicity we input low-level variables to TMVA to obtain the results below.
\section{Results}
\label{sec:result}
In this section, based on our analyses we show the limits on the mixing parameter $|V_{\ell N }|^2$ for the heavy neutrino mass $m_N$ in the range of 10 to 1000 GeV.
After the pre-selection, the BDT cut is optimized according to the signal statistical significance calculated by Eq.~\ref{eqn:statSgf} for each mass case.
\begin{equation}
\sigma_{\rm stat} = \sqrt{2 [(N_s+N_b) {\rm ln}(1+\frac{N_s}{N_b}) - N_s ] },
\label{eqn:statSgf}
\end{equation}
where $N_s$ ($N_b$) is the number of signal (total background) events after all selection cuts.
In Table~\ref{tab:allEfficiencies}, we show selection efficiencies of pre-selection and BDT cuts for both signal and background processes at the LHeC and FCC-eh for representative heavy neutrino masses.
The total selection efficiency is the product of pre-selection and BDT cut efficiencies.
The number of background events after all cuts can be calculated by multiplying the initial number in Table~\ref{tab:cutmn120} by the total selection efficiency, while the number of signal events can be calculated as the product of signal cross section, collider luminosity and total selection efficiency.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,height=6cm]{fig_limit_mu22.pdf}
\caption{
2- and 5-$\sigma$ limits on mixing parameter $|V_{\ell N }|^2$ for the heavy neutrino mass in the range of 10 to 1000 GeV at the LHeC and FCC-eh.
Also shown are the current experimental limits at 95\% confidence level from the trilepton searches (CMS, 35.9 $\rm fb^{-1}$, $3 \ell$)~\cite{CMS:2018iaf}, (CMS, 137 $\rm fb^{-1}$, $3 \ell$)~\cite{CMS:2021lzm}, (ATLAS, 36.1 $\rm fb^{-1}$, $3 \ell$)~\cite{ATLAS:2019kpx} and dilepton searches~(CMS, 35.9 $\rm fb^{-1}$, $2 \ell$)~\cite{CMS:2018jxx}, (LHCb, 3.0 $\rm fb^{-1}$, $2 \ell$)~\cite{LHCb:2020wxx} at the LHC.
}
\label{fig:sensitivity}
\end{figure}
In Fig.~\ref{fig:sensitivity}, we show 2- and 5-$\sigma$ limits on mixing parameter $|V_{\ell N }|^2$ for the heavy neutrino mass in the range of 10 to 1000 GeV at the LHeC (FCC-eh) with an electron beam energy of 60 GeV, a proton beam energy of 7 (50) TeV and an integrated luminosity of 1 (3) $\rm ab^{-1}$.
At the LHeC, as $m_N$ changes from 10 GeV to 100 GeV, the 2-$\sigma$ upper limits on $|V_{\ell N }|^2$ decrease from $7.8 \times 10^{-5}$ to $3.2 \times 10^{-6}$; the limits are relatively flat for $m_N$ between 100 GeV to 500 GeV and increase rapidly to $3.3 \times 10^{-3}$ afterwards.
The 2-$\sigma$ limit at the FCC-eh has similar behavior as that at the LHeC, but its varying range is much smaller, between $\sim 10^{-6}$ and $10^{-5}$.
At both colliders, the 5-$\sigma$ limits are slightly weaker than those for 2-$\sigma$.
The increasing or decreasing behavior of the upper limit is the result of the competition between the signal cross section and the separation extent of the signal and background kinematical distributions.
With the increase of $m_N$, the signal cross section decreases gradually as shown in Fig.~\ref{fig:crs}, while BDT distributions of the signal and the total background become more and more separate (cf. Figs.~\ref{fig:BDTLHeC} and~\ref{fig:BDTFCC-eh}).
Therefore, limits are weaker because of smaller separations between the signal and background kinematics for small masses and because of smaller signal cross sections for heavy masses.
The most stringent limits are achieved at $m_{N} \sim 100$ GeV.
Limits are relatively flat in the middle range of $m_N$ because of the offset between the decrease of the signal cross section and increase of the separation extent of the signal and background kinematics.
At the LHeC, when $m_{N}$ $\gtrsim$ 500 GeV, the signal cross section decreases rapidly with the increase of $m_{N}$, leading to the rapid increase in the upper limit.
Since the signal cross section at the FCC-eh decreases slowly, the limit for heavy masses also increases gently.
The FCC-eh has better limits than the LHeC, mainly because the signal cross section of FCC-eh is larger than that of LHeC for the same $m_{N}$.
To compare with current experiment limits, we also present the recent LHC limits in Fig.~\ref{fig:sensitivity}.
The details of these studies are reviewed in Sec.~\ref{sec:intro}.
The limit curve for (CMS, 35.9 $\rm fb^{-1}$, $2 \ell$) is reinterpreted from the original limit on the parameter $|V_{eN }V_{\mu N }^*|^2/(|V_{eN }|^2+|V_{\mu N }|^2)$ in the CMS same-sign dilepton search~\cite{CMS:2018jxx}. With the same assumption of $|V_{eN }|^2 = |V_{\mu N }|^2$ in this article, the original limit is shown by multiplying a factor of two.
The limit for (CMS, 137 $\rm fb^{-1}$, $3 \ell$) is from the Ref.~\cite{CMS:2021lzm} where the CMS collaboration searched for a long-lived heavy neutrino in the trilepton final state with displaced vertices.
It has excluded some parameter regions when $m_N < 15$ GeV.
Our analyses show that the LHeC limit is slightly weaker than the current CMS and ATLAS trilepton searches when $m_N$ $\lesssim$ 20 GeV, while it is better for heavier masses and about two to three orders of magnitude stronger than the current CMS limits when $m_N$ $\gtrsim$ 100 GeV.
Compared with current LHC limits, the FCC-eh gives more stringent limits when $m_N > 15$ GeV and are much stronger when $m_N$ $\gtrsim$ 100 GeV.
\section{Summary and Discussion}
\label{sec:sum}
In this paper, we utilize the lepton number violation signal process of $p\, e^- \to \mu^+ jjj$ to search for heavy Majorana neutrinos at the future electron-proton colliders.
We consider the LHeC (FCC-eh) running with an electron beam energy of 60 GeV, a proton beam energy of 7 (50) TeV and an integrated luminosity of 1 (3) $\rm ab^{-1}$.
To simplify the analyses, we consider the simplified Type-I model and assume that only one generation of heavy neutrinos $N$ is within the collider access and mixes with active neutrinos of electron and muon flavours with the same mixing parameters, i.e. $|V_{\ell N}|^2 = |V_{e N}|^2 = |V_{\mu N}|^2$ and $|V_{\tau N}|^2 = 0$.
The signal production cross sections are presented in Fig.~\ref{fig:crs} at both LHeC and FCC-eh for the heavy neutrino mass $m_N$ in the range of 10$-$1000 GeV.
We apply detector configurations and simulate signal and four dominant SM background events of $\mu^+ \mu^- e^- j j j$, $\mu^+ \mu^- \nu_e j j j$, $\mu^+ \nu_\mu e^- j j j$ and $\mu^+ \nu_\mu \nu_e j j j$ including detector effects.
We first use the pre-selection cuts to select the final state with exactly one muon with positive charge, at least three jets and small missing energy.
The number of events for the signal with benchmark $m_N =$ 120 GeV and four background processes after applying pre-selection cuts are presented in Table~\ref{tab:cutmn120}.
To reject the background efficiently, nineteen basic observables are input to perform the multi-variate analyses based on machine-learning. The distributions of BDT responses of the signal and the SM background processes at the LHeC and FCC-eh for representative heavy neutrino masses are shown in Appendix~\ref{appendix:BDTdistribution}, while the efficiencies of pre-selection and BDT cuts are shown in Appendix~\ref{appendix:efficiencies}.
To test the effects on the final limits by inputting different sets of observables, we construct and input another set of twenty-nine observables and find that the changes in the final limits are small. The distributions of eight high-level representative kinematic observables are also presented in Appendix~\ref{appendix:HLobs} for the signal with $m_N =$ 120 GeV and four background processes at the LHeC and FCC-eh.
Based on our analyses, we show the 2- and 5-$\sigma$ upper limits on the mixing parameter $|V_{\ell N}|^2$ for the heavy neutrino mass $m_N$ in the range of 10$-$1000 GeV at both LHeC and FCC-eh in Fig.~\ref{fig:sensitivity}.
At the LHeC, the 2-$\sigma$ upper limits on $|V_{\ell N }|^2$ decrease from $7.8 \times 10^{-5}$ to $3.2 \times 10^{-6}$ when $m_N$ changes from 10 GeV to 100 GeV; the limits are relatively flat for $m_N$ in the middle range between 100 GeV and 500 GeV and increase rapidly to $3.3 \times 10^{-3}$ afterwards.
The 2-$\sigma$ limit at the FCC-eh has the similar behavior as that at the LHeC, but its varying range is much smaller, between $\sim 10^{-6}$ and $10^{-5}$.
At both colliders, the 5-$\sigma$ limits are slightly weaker than those for 2-$\sigma$.
The limits are compared with the current LHC experimental limits. Our analyses show that the LHeC limit is slightly weaker than the current CMS and ATLAS trilepton searches when $m_N$ $\lesssim$ 20 GeV, while it is better for heavier masses and about two to three orders of magnitude stronger than the current CMS limits when $m_N$ $\gtrsim$ 100 GeV.
Compared with current LHC limits, the FCC-eh gives more stringent limits when $m_N > 15$ GeV and are much stronger when $m_N$ $\gtrsim$ 100 GeV.
The LNV signal process considered in this study is one typical channel to search for heavy Majorana neutrinos at electron-proton colliders.
If this signal is discovered at colliders, it is also confirmed that the nature of $N$ is of Majorana type.
Therefore, our results are an important complement to the physics goals of electron-proton colliders, and the parameter space probed by our search strategy could explain other fundamental physics problems, such as the leptogenesis~\cite{Drewes:2021nqr}.
We assume that $N$ decays promptly in this study. However,
when $N$'s lifetime is long enough, heavy neutrinos can have sizeable probability to travel through and decay outside the detector.
In this case, such neutrinos will be mis-detected and behave as missing energy, and thus these events cannot contribute to our signal.
The number of detectable signal events needs to be multiplied by the average probability $\bar{P}$ of heavy neutrinos decaying inside the detector's fiducial volume.
Because our signal final state has one $\mu^+$ and two jets from the the decay of $N$, in order to detect these jets, the heavy neutrino needs to decay before the end of the hadronic calorimeter (HCAL).
To estimate this effect on final limits, we consider detector layouts of the LHeC and FCC-eh from Ref.~\cite{AbelleiraFernandez:2012ni} and Ref.~\cite{FCCeh:detector}, respectively.
We generate the signal sample $p e^- \to j N$ at the parton level using MadGraph for different $m_N$ at the LHeC and FCC-eh, and calculate $\bar{P}$ of detectors using a similar method used in Ref.~\cite{Wang:2019orr}.
Results for $\bar{P}$ while varying $|V_{lN}|^2$ when $m_N = 5, 10, 20$ GeV at the LHeC and FCC-eh are derived, and we find that the resulting curves of detectors at LHeC and FCC-eh are very similar.
For $m_N = 5, 10, 20$ GeV, the $\bar{P}$ values begin to reduce from unity when the mixing parameter $|V_{\ell N}|^2 \sim 2 \times 10^{-6}, 3 \times 10^{-8}, 8 \times 10^{-10}$, respectively, and become zero when $|V_{\ell N}|^2 \sim 2 \times 10^{-9}, 5 \times 10^{-11}, 1 \times 10^{-12}$, respectively.
Since Fig.~\ref{fig:sensitivity} shows that the 2-$\sigma$ upper limits on $|V_{\ell N}|^2$ for $m_N = $ 10 GeV at the LHeC and FCC-eh are $7.8 \times 10^{-5}$ and $9.0 \times 10^{-6}$, respectively, which are larger than $3 \times 10^{-8}$, the corresponding $\bar{P}$ values are unity, and thus the final limits are not modified.
For larger $m_N$, $|V_{\ell N}|^2$ is required to be even smaller such that $\bar{P}$ can decrease from unity.
Therefore, the effect of long-lived cases of heavy neutrinos is expected to be negligible for the mass range considered in this study.
Such effect becomes significant for very small masses, and we leave it for future studies.
We note that when $|V_{e N}|^2 = |V_{\mu N}|^2$ the LNV signal $p\, e^- \to e^+ jjj$ also exists with cross section approximately equal to that of signal $p\, e^- \to \mu^+ jjj$.
However, for the signal $p\, e^- \to e^+ jjj$, there exists one additional SM background process of $p\, e^- \to e^- jjj$. When the final state $e^-$ is mis-detected as $e^+$, this process can be one significant background.
To estimate sensitivities on $|V_{\ell N}|^2$ of the $e^+$ final state signal,
we consider its corresponding SM background processes.
The cross sections of background processes of $e^- jjj$, $e^+ e^- e^- j j j$, $e^+ e^- \nu_e j j j$ and $e^+ \nu_e \nu_e j j j$ are found to be $1.4 \times 10^{4}$, $0.53$, $0.33$ and $8.1 \times 10^{-6}$ pb at the LHeC, respectively.
One sees that the production cross section of $e^- jjj$ is much larger than the other background processes. Therefore, it could still dominate even if the detector's charge misidentification rate is small.
With the current detector configuration at the LHeC, after selecting exactly one positron with transverse momentum above 5 GeV, the total background cross section for the $e^+$ final state is $8.62\times10^{-2}$ pb, among which
the $e^- jjj$ process is $6.07 \times 10^{-2}$ pb and the sum of the other three background processes is $2.55 \times 10^{-2}$ pb, so $e^- jjj$ is the main background.
For the $\mu^+$ final state, the total background cross section after selecting exactly one muon with positive charge with transverse momentum above 5 GeV is $2.58 \times 10^{-2}$ pb (cf. cut(i) in Table~\ref{tab:cutmn120}).
Because the total background cross section for the $e^+$ final state is about three times larger than that for the $\mu^+$ final state, the limit on mixing parameter $|V_{\ell N}|^2$ from the signal process $p\, e^- \to e^+ jjj$ is expected to be weaker than that from the signal process $p\, e^- \to \mu^+ jjj$.
However, when $|V_{e N}|^2 \neq 0$ and $|V_{\mu N}|^2 = 0$, the $p\, e^- \to \mu^+ jjj$ signal cannot be produced any more, but the $p\, e^- \to e^+ jjj$ signal can still exist. Because the $p\, e^- \to e^+ jjj$ signal process depends on the mixing parameter $|V_{e N}|^2$ only, it can be a unique channel to probe $|V_{e N}|^2$ independent of other mixing parameters.
In this sense, the detailed analyses of the $e^+$ final state are still meaningful and we leave it for future studies.
|
\section{Introductrion}
Throughout this paper, a curve always means a projective variety in positive characteristic $p \geq 3$ of dimension one. It is well-known that all nonsingular genus-$g$ curves $C$ defined over $\mathbb{F}_q$ with $q = p^n$ satisfy the Hasse-Weil inequality\vspace{-1mm}
\[
1+ q - 2g\sqrt{q} \leq \#C(\mathbb{F}_q) \leq 1 + q + 2g\sqrt{q},\vspace{-1mm}
\]
where $C(\mathbb{F}_q)$ denotes the set of $\mathbb{F}_q$-rational points of $C$. Now, a nonsingular curve $C$ is called \textit{maximal} (resp.\ \textit{minimal}) when the number of $\mathbb{F}_q$-rational points of $C$ attains the upper (resp.\ lower) bound. Maximal curves have been investigated for their applications to coding theory. On the other hand, we call a curve $C$ \textit{superspecial} if ${\rm Jac}(C)$ is isomorphic to the product of supersingular elliptic curves. This is equivalent to saying that the $a$-number of $C$ is equal to $g$. The $a$-number of $C$ is defined to be the dimension of ${\rm Hom}(\alpha_p,{\rm Jac}(C)[p])$ where $\alpha_p$ is the kernel of the Frobenius map on the additive group $\mathbb{G}_a$. And then, it is also known that any maximal or minimal curve over $\mathbb{F}_{p^2}$ is superspecial, though a superspecial curve over $\mathbb{F}_{p^2}$ is not necessarily maximal nor minimal.\par
In this paper, we study the maximality of genus-3 nonhyperelliptic curves of Ciani type, and we will call them Ciani curves briefly. A Ciani curve $C$ is a plane quartic defined by the equation
\[
C: x^4 + y^4 + z^4 + rx^2y^2 + sy^2z^2 + tz^2x^2 = 0,
\]
which was studied by Ciani \cite{Ciani}. Brock \cite[Theorem 3.15]{Brock} studied the superspeciality of Ciani curves and their enumerations, using the result by Hashimoto \cite{Hashimoto} on the computations of the class numbers of quaternion unitary groups. Our main result is as below:\vspace{-1mm}
\begin{Thm}
Assume that a nonsingular curve
\[
C: x^4 + y^4 + z^4 + rx^2y^2 + sy^2z^2 + tz^2x^2 = 0
\]
is superspecial, then $r,s$ and $t$ belong to $\mathbb{F}_{p^2}$. Moreover $C$ is maximal or minimal over $\mathbb{F}_{p^2}$.
\end{Thm}\vspace{-1mm}
\noindent See Corollary 4.6 for a condition determining whether $C$ is maximal or minimal.
The remainder of this paper is structured as follows. In Section 2, we review several properties of Ciani curves. Proposition 2.5 gives us the classification of automorphism groups of a Ciani curve. In Section 3, we look into the structure of a Ciani curve $C$. In particular, we describe explicitly the elliptic curves appearing as quotients of $C$ by involutions. In Section 4, we prove Theorem 1.1.\vspace{-1mm}
\subsection*{Acknowledgments}\vspace{-1mm}
This paper was written while the author is a Ph.D.\,student at Yokohama National University, and I would like to express my special thanks to my supervisor Prof.\ Shushi Harashita for his guidance.
\section{Ciani curve}
\setcounter{equation}{0}
Let $K$ be a perfect field of characteristic $p \geq 3$. In this section, we consider a nonhyperelliptic curve of genus 3 defined over $K$\vspace{-1mm}
\begin{equation}
C: x^4 + y^4 + z^4 + rx^2y^2 + sy^2z^2 + tz^2x^2 = 0,
\end{equation}
which is called a \textit{Ciani curve}. First of all, we discuss the singularity of a Ciani curve.
\begin{Lem}
The curve $C$ is nonsingular if and only if $r,s,t \neq \pm 2$ and $r^2+s^2+t^2-rst-4 \neq 0$.
\begin{proof}
Put $F := x^4 + y^4 + z^4 + rx^2y^2 + sy^2z^2 + tz^2x^2$. Then, we have
\[
\frac{\partial{F}}{\partial{x}} = 2x(2x^2 + ry^2 + tz^2), \quad \frac{\partial{F}}{\partial{y}} = 2y(2y^2 + rx^2 + sz^2), \quad \frac{\partial{F}}{\partial{z}} = 2z(2z^2 + sy^2 + tx^2).\vspace{-1mm}
\]
To show the ``if''-part, assume that $(x:y:z) \in \mathbb{P}^2$ is a singular point of $C$.\par
Firstly, we note that there is no point $(x:y:z)$ on $C$ such that two among $x,y$ and $z$ are zero; e.g., if $x = y = 0$, it follows from $F=0$ that $z=0$, which is a contradiction.\par
Secondly, consider a singular point $(x:y:z)$ on $C$ such that only one among $x,y$ and $z$ is zero, e.g.; the case $x = 0,\,y \neq 0$ and $z \neq 0$. It follows from $2y^2 + sz^2 = 2z^2 + sy^2 = 0$ that $(4-s^2)z^2 = 0$, whence $s = \pm 2$. One can check the case of $y=0$ or $z=0$ in the same way.\par
Lastly, consider a singular point $(x:y:z)$ on $C$ such that all of $x,y$ and $z$ are non-zero. Using the Jacobian criterion, we obtain\vspace{-1mm}
\[
\left( \begin{array}{ccc}
2 & r & t\\
r & 2 & s \\
t & s & 2 \\
\end{array} \right)
\left( \begin{array}{c}
x^2 \\
y^2 \\
z^2 \\
\end{array} \right)
= \left( \begin{array}{c}
0 \\
0 \\
0 \\
\end{array} \right).\vspace{-1mm}
\]
Here, we can calculate\vspace{-2mm}
\[
\det\left( \begin{array}{ccc}
2 & r & t\\
r & 2 & s \\
t & s & 2 \\
\end{array} \right) = -2(r^2+s^2+t^2 - rst - 4),
\]
and thus $r^2+s^2+t^2 - rst - 4 = 0$ since $(x^2,y^2,z^2) \neq (0,0,0)$.\par
Conversely, we show the ``only if''-part. If $r = \pm2$, we can find singular points $(x:y:0)$ on $C$ such that $x^2 \pm y^2 = 0$. One can check the case of $s = \pm 2$ or $t = \pm 2$ in the same way. Moreover, by a tedious calculation, we can find singular points $(x:y:z)$ on $C$ such that
\[
(4-r^2)(rt-2s)x^2 = (4-r^2)(rs-2t)y^2 = (rt-2s)(rs-2t)z^2
\]
if $r^2+s^2+t^2-rst-4 = 0$. The proof is completed.
\end{proof}
\end{Lem}
\begin{Def}
We say that a nonsingular Ciani curve of the form (2.1) is of \textit{$[a,b,c]$-type} when the triple $(r,s,t)$ is a permutation of
\[
(a,b,c),(a,-b,-c),(-a,b,-c),(-a,-b,c).
\]
If both $C$ and $C'$ are two Ciani curves of $[a,b,c]$-type, it is obvious that $C$ is isomorphic to $C'$.
\end{Def}
\newpage
\begin{Rmk}
Suppose that $C$ and $C'$ are two Ciani curves, then the type of $C$ is not necessarily same as that of $C'$ even if $C$ and $C'$ are isomorphic. For example, the Fermat curve $x^4+y^4+z^4 = 0$ is isomorphic to the curve $x^4 + y^4 + 6y^2z^2 + z^4 = 0$. We find a necessary and sufficient condition for two Ciani curves $C$ and $C'$ to be isomorphic in \cite[Proposition 2.2]{Howe}, but do not use it in this paper.
\end{Rmk}
\begin{Exp}
Consider a nonsingular curve
\begin{equation}
C: x^4 + y^4 + z^4 + rx^2yz +sy^2z^2 = 0.
\end{equation}
Note that the automorphism group of $C$ contains ${\rm D}_8$ as shown in \cite[Theorem 3.1]{Ritzenthaler}. Let us confirm that this curve $C$ is of $[a,b,a]$-type with $a=\frac{r}{\sqrt{s+2}}$ and $b = 2-\frac{16}{s+2}$.\par
By replacing $u = (y+z)/2$ and $z = (y-z)/(2\sqrt{-1})$, we obtain $y = u+\sqrt{-1}v$ and $z = u+\sqrt{-1}v$ thus\vspace{-2mm}
\begin{align*}
yz &= (u+\sqrt{-1}v)(u+\sqrt{-1}v) = u^2 + v^2,\\
y^2 + z^2 &= (u+\sqrt{-1}v)^2 + (u+\sqrt{-1}v)^2 = 2(u^2-v^2).
\end{align*}
The left side of (3.2) can be transformed into
\begin{align*}
x^4 + y^4 + z^4 + rx^2yz + sy^2z^2 &= x^4 + (y^2 + z^2)^2 + rx^2yz + (s-2)y^2z^2\\
&= x^4 + (s+2)u^4 + (s+2)v^4 + rx^2u^2 + 2(s-6)u^2v^2 + rv^2x^2,
\end{align*}
and the map
\[
u \mapsto \frac{u}{\sqrt[4]{s+2}} ,\quad v \mapsto \frac{v}{\sqrt[4]{s+2}}
\]
transforms the curve $C$ into
\[
x^4 + u^4 + v^4 + \frac{r}{\sqrt{s+2}}x^2u^2 + \Bigl(2-\frac{16}{s+2}\Bigr)u^2v^2 + \frac{r}{\sqrt{s+2}}v^2x^2.
\]
Hence, the curve $C$ is a Ciani curve of $[a,b,a]$-type.
\end{Exp}
Next result tells us the classification of automorphism groups of Ciani curves. Let ${\rm C}_n$\hspace{1mm}(resp.\ ${\rm D}_n$) be the cyclic (resp.\ dihedral) group of order $n$ and ${\rm S}_n$ the symmetric group of degree $n$.
\begin{Prop}
The automorphism group of a Ciani curve $C$ is either of the following 7 groups:
\begin{enumerate}
\item ${\rm D}_4$. In case $C$ can be written as $[a,b,c]$-type for some $a,b,c \in \mbar{K}$.
\item ${\rm D}_8$. In case $C$ can be written as $[a,b,a]$-type for some $a,b \in \mbar{K}$.
\item ${\rm G}_{16}$. In case $C$ can be written as $[0,b,0]$-type for some $b \in \mbar{K}$.
\item ${\rm S}_4$. In case $C$ can be written as $[a,a,a]$-type for some $a \in \mbar{K}$.
\item ${\rm G}_{48}$. In case $C$ can be written as $[0,2\sqrt{-3},0]$-type.
\item ${\rm G}_{96}$. In case $C$ can be written as $[0,0,0]$-type.\vspace{-1mm}
\item ${\rm G}_{168}$. In case $C$ can be written as $\bigl[\frac{-3+\sqrt{-63}}{2},\frac{-3+\sqrt{-63}}{2},\frac{-3+\sqrt{-63}}{2}\bigr]$-type.
\end{enumerate}
\begin{proof}
It follows from \cite[Theorem 3.1]{Ritzenthaler} about (1),\,(3),\,(4) and (6). It is well-known \cite[Section 6-8]{Meagher} about (5) and (7). It is clear about (2) by Example 2.4.
\end{proof}
\end{Prop}
\section{Elliptic curves associated to a Ciani curve}
\setcounter{equation}{0}
Let $\sigma_1,\sigma_2,\sigma_3$ be automorphisms on a nonsingular Ciani curve $C$ as below:
\begin{align*}
\sigma_1&: (x:y:z) \mapsto (-x:y:z),\\
\sigma_2&: (x:y:z) \mapsto (x:-y:z),\\
\sigma_3&: (x:y:z) \mapsto (x:y:-z).
\end{align*}
Put $E_i := C/\langle\sigma_i\rangle$ for $i \in \{1,2,3\}$, then we obtain the equations
\begin{align*}
E_1 &: X^2 + y^4 + z^4 + rXy^2 + sy^2z^2 + tz^2X = 0,\\
E_2 &: x^4 + Y^2 + z^4 + rx^2Y + sYz^2 + tz^2x^2 = 0,\\
E_3 &: x^4 + y^4 + Z^2 + rx^2y^2 + sy^2Z + tZx^2 = 0
\end{align*}
with $X = x^2,\,Y = y^2$ and $Z = z^2$. One can easily check each $E_i$ is a genus-1 curve, and the next lemma gives the Legendre forms of $E_i$. We choose $\alpha,\beta$ and $\gamma$ such that
\begin{equation}
\alpha^2 = r^2-4, \quad \beta^2 = s^2-4, \quad \gamma^2 = t^2 - 4.
\end{equation}
once and fix them throughout this paper.
\begin{Lem}
We can transform $E_i$ into the Legendre forms $y^2 = x(x-1)(x-\lambda_i)$ with
\begin{align}
\lambda_1 &= \frac{(rt-2s)-\gamma\alpha}{(rt-2s)+\gamma\alpha},\nonumber\\
\lambda_2 &= \frac{(sr-2t)-\alpha\beta}{(sr-2t)+\alpha\beta},\\
\lambda_3 &= \frac{(ts-2r)-\beta\gamma}{(ts-2r)+\beta\gamma}.\nonumber
\end{align}
\begin{proof}
In this proof, we will show about $E_1$. Firstly, we get the equation
\begin{align*}
\Bigl(X + \frac{r}{2}y^2 + \frac{t}{2}z^2\Bigr)^{\!2} &= \Bigl(\frac{r}{2}y^2 + \frac{t}{2}z^2\Bigr)^{\!2} - y^4 - sy^2z^2 - z^4\\
&= \Bigl(\frac{r^2}{4}-1\Bigr)y^4 + \Bigl(\frac{rt}{2} - s\Bigr)y^2z^2 + \Bigl(\frac{t^2}{4}-1\Bigr)z^4.
\end{align*}
By replacing $u = y$ and $v = X + \frac{r}{2}y^2 + \frac{t}{2}z^2$ and multiplying $(r^2-4)/4$ of both sides, we obtain
\[
v^2 = u^4 + \frac{2rt-4s}{r^2-4}u^2z^2 +\frac{t^2-4}{r^2-4}z^4.
\]
Here, the right side is transformed into
\begin{align*}
u^4+ \frac{2rt-4s}{r^2-4}u^2z^2 + \frac{t^2-4}{r^2-4}z^4 &= \Bigl(u^2+\frac{\gamma}{\alpha}z^2\Bigr)^{\!2} \!- 2\frac{\gamma\alpha - (rt-2s)}{r^2-4}u^2z^2\\
&= \Bigl(u^2+ buz + \frac{\gamma}{\alpha}z^2\Bigr)\Bigl(u^2- buz + \frac{\gamma}{\alpha}z^2\Bigr)
\end{align*}
with $b^2 = 2 \cdot \frac{\gamma\alpha - (rt-2s)}{r^2-4}$. Put\vspace{-2mm}
\[
d_1 = \frac{b + \sqrt{b^2 - 4 \cdot \frac{\gamma}{\alpha}}}{2}, \quad d_2 = \frac{b - \sqrt{b^2 - 4 \cdot \frac{\gamma}{\alpha}}}{2},
\]
then we have the factorization
\[
\Bigl(u^2+ buz + \frac{\gamma}{\alpha}z^2\Bigr)\Bigl(u^2- buz + \frac{\gamma}{\alpha}z^2\Bigr) = (u+ d_1z)(u+d_2z)(u-d_1z)(u-d_2z).
\]
The map
\[
u \mapsto \frac{u-d_1z}{u+d_1z} \cdot \frac{d_2+d_1}{d_2-d_1}
\]
transforms this elliptic curve into the Legendre form:
\[
v^2 = u(u-1)\biggl(u-\frac{(d_2+d_1)^2}{(d_2-d_1)^2}\biggr) = u(u-1)\biggl(u-\frac{b^2}{b^2 - 4 \cdot \frac{\gamma}{\alpha}}\biggr) = u(u-1)(u-\lambda_1).
\]
This is the desired conclusion.
\end{proof}
\end{Lem}
Here, we can regard the curve
\[
P: X^2 + Y^2 + Z^2 + rXY + sYZ + tZX = 0
\]
as the quotient $C/\langle\sigma_1,\sigma_2\rangle$. One can confirm that $P$ is nonsingular since $C$ is nonsingular and the genus of $P$ is $0$. Hence, the curve $P$ is isomorphic to the projective line $\mathbb{P}^1$. For the above discussion, we obtain the following diagram:\vspace{-3mm}
\begin{equation}
\vcenter{
\xymatrix{
& C \ar[ld] \ar[d] \ar[rd] &\\
E_1 \ar[rd] & E_3 \ar[d] & E_2 \ar[ld]\\
& P &
}}\vspace{-1mm}
\end{equation}
The diagram (3.3) induces an isogeny ${\rm Jac}(C) \rightarrow E_1 \hspace{-0.3mm}\times\hspace{-0.3mm} E_2 \hspace{-0.3mm}\times\hspace{-0.3mm} E_3$ by \cite[Section 3]{Kani} of degree $2^3$. Since the degree of the isogeny is not divided by $p$, then we obtain ${\rm Jac}(C)[p] \cong (E_1 \hspace{-0.3mm}\times\hspace{-0.3mm} E_2 \hspace{-0.3mm}\times\hspace{-0.3mm} E_3)[p]$. Hence, a Ciani curve $C$ is superspecial if and only if $E_1,E_2$ and $E_3$ are supersingular.
\begin{Prop}
The reverse transformation of that in Lemma 3.1 is given by
\begin{align*}
r &= \frac{\lambda_1\lambda_2 - \lambda_2\lambda_3 - \lambda_3\lambda_1 + 1}{\sqrt{\lambda_1\lambda_2}(1-\lambda_3)},\\
s &=\frac{\lambda_2\lambda_3 - \lambda_3\lambda_1 - \lambda_1\lambda_2 + 1}{\sqrt{\lambda_2\lambda_3}(1-\lambda_1)},\\
t &= \frac{\lambda_3\lambda_1 - \lambda_1\lambda_2 - \lambda_2\lambda_3 + 1}{\sqrt{\lambda_3\lambda_1}(1-\lambda_2)}.
\end{align*}
\begin{proof}
In this proof, we will show the case of $\lambda_1,\lambda_2,\lambda_3 \neq -1$. By linear fractional transformations, one can check that\vspace{-1mm}
\begin{equation}
\frac{1+\lambda_1}{1-\lambda_1} = \frac{rt-2s}{\gamma\alpha},\quad \frac{1+\lambda_2}{1-\lambda_2} = \frac{sr-2t}{\alpha\beta}, \quad \frac{1+\lambda_3}{1-\lambda_3} = \frac{ts-2r}{\beta\gamma}.\nonumber
\end{equation}
Hence we obtain the equation
\[
\frac{(1-\lambda_1)(1-\lambda_2)(1+\lambda_3)}{(1+\lambda_1)(1+\lambda_2)(1-\lambda_3)} = \frac{(r^2-4)(ts-2r)}{(rt-2s)(sr-2t)}.
\]
Since $(r^2-4)(ts-2r) + (rt-2s)(sr-2t) = -2r(r^2+s^2+t^2-rst-4)$, then we have
\begin{equation}
1 + \frac{(1-\lambda_1)(1-\lambda_2)(1+\lambda_3)}{(1+\lambda_1)(1+\lambda_2)(1-\lambda_3)} =-2r \cdot \frac{r^2+s^2+t^2-rst-4}{(rt-2s)(sr-2t)}.
\end{equation}
On the other hand, note that
\begin{align}
(rt-2s)^2 -4(r^2+s^2+t^2-rst-4) &= (r^2-4)(t^2-4),\nonumber\\
(sr-2t)^2 -4(r^2+s^2+t^2-rst-4) &= (s^2-4)(r^2-4),\\
(ts-2r)^2 -4(r^2+s^2+t^2-rst-4) &= (t^2-4)(s^2-4),\nonumber
\end{align}
thus one can check that\vspace{-3mm}
\begin{align*}
\frac{\lambda_1}{(1+\lambda_1)^2} &= \frac{r^2+s^2+t^2-rst-4}{(rt-2s)^2},\\
\frac{\lambda_2}{(1+\lambda_2)^2} &= \frac{r^2+s^2+t^2-rst-4}{(sr-2t)^2},\\
\frac{\lambda_3}{(1+\lambda_3)^2} &= \frac{r^2+s^2+t^2-rst-4}{(ts-2r)^2}.
\end{align*}
Therefore, we have\vspace{-1mm}
\begin{equation}
\frac{\sqrt{\lambda_1\lambda_2}}{(1+\lambda_1)(1+\lambda_2)} = \frac{r^2+s^2+t^2-rst-4}{(rt-2s)(sr-2t)}.
\end{equation}
Using the equations (3.4) and (3.6), then\vspace{-1mm}
\[
\frac{(1+\lambda_1)(1+\lambda_2)(1-\lambda_3) + (1-\lambda_1)(1-\lambda_2)(1+\lambda_3)}{(1+\lambda_1)(1+\lambda_2)(1-\lambda_3)} = -2r \cdot \frac{\sqrt{\lambda_1\lambda_2}}{(1+\lambda_1)(1+\lambda_2)}.
\]
We may solve this with respect to $r$. The other two formulas can be shown in the same way.
\end{proof}
\end{Prop}
\section{Proof of the main theorem}
\setcounter{equation}{0}
In this section, we will show the theorem stated in Introduction.
\setcounter{section}{1}
\begin{Thm}
Assume that a nonsingular curve
\[
C: x^4 + y^4 + z^4 + rx^2y^2 + sy^2z^2 + tz^2x^2 = 0
\]
is superspecial, then $r,s$ and $t$ belong to $\mathbb{F}_{p^2}$. Moreover $C$ is maximal or minimal over $\mathbb{F}_{p^2}$.
\end{Thm}\vspace{-1mm}
\setcounter{section}{4}
\setcounter{Def}{0}
A key to the proof is the following proposition by Auer and Top \cite[Proposition 2.2]{Top}.\vspace{-1mm}
\begin{Prop}
Let $E: y^2 = x(x-1)(x-\lambda)$ be a supersingular elliptic curve, then $\lambda \in (\mathbb{F}_{p^2}\hspace{-0.3mm})^4$. Moreover, the followings are true:\vspace{-1mm}
\begin{itemize}
\item If $p \equiv 3 \pmod{4}$, then an elliptic curve $E$ is maximal over $\mathbb{F}_{p^2}$.\vspace{-1mm}
\item If $p \equiv 1 \pmod{4}$, then an elliptic curve $E$ is minimal over $\mathbb{F}_{p^2}$.
\end{itemize}
In particular, the curve $C$ is maximal or minimal over $\mathbb{F}_{p^2}$.
\end{Prop}
\newpage
Recall the discussions in Section 3. Let $E_1,E_2$ and $E_3$ be the following three elliptic curves:
\begin{align*}
E_1 &: X^2 + y^4 + z^4 + rXy^2 + sy^2z^2 + tz^2X = 0,\\
E_2 &: x^4 + Y^2 + z^4 + rx^2Y + sYz^2 + tz^2x^2 = 0,\\
E_3 &: x^4 + y^4 + Z^2 + rx^2y^2 + sy^2Z + tZx^2 = 0.
\end{align*}
Then, there exist surjectives $C \rightarrow E_i$ defined over $\mathbb{F}_{p^2}$.
\begin{proof}[Proof of the first half of Theorem 1.1]
The elliptic curves $E_i$ are supersingular by the assumption, since the quotient of supersingular curve is supersingular. By Lemma 3.1, each elliptic curve $E_i$ is isomorphic to $y^2 = x(x-1)(x-\lambda_i)$. By using Proposition 4.1, each $\lambda_i$ is a fourth power in $(\mathbb{F}_{p^2}\hspace{-0.3mm})^\times$ and thus $\hspace{-0.3mm}\sqrt{\lambda_i} \in \mathbb{F}_{p^2}$. Hence, it follows from Proposition 3.2 that $r,s,t \in \mathbb{F}_{p^2}$.
\end{proof}
Therefore, the question of whether a Ciani curve $C$ is maximal or minimal over $\mathbb{F}_{p^2}$ makes sense. To prove the second assertion of Theorem 1.1, we need the following four lemmas:
\begin{Lem}
We choose $\varDelta$ such that $\varDelta^2 = r^2+s^2+t^2 - rst- 4$, then the followings are true:
\begin{enumerate}
\item If $C$ is superspecial, then $\alpha\beta, \beta\gamma$ and $\gamma\alpha$ belong to $\mathbb{F}_{p^2}$ where $\alpha,\beta,\gamma$ are chosen in (3.1).\vspace{-1mm}
\item If $C$ is superspecial, then $\varDelta$ belongs to $\mathbb{F}_{p^2}$.
\end{enumerate}
\begin{proof}
The elliptic curve $E_i$ is supersingular by assumption, and so $\lambda_i \in (\mathbb{F}_{p^2}\hspace{-0.3mm})^4$ by Proposition 4.1.\par
(1) This claim holds from (3.2) and the first assertion of Theorem 1.1.\par
(2) We obtain $-\lambda_1 \in \mathbb{F}_{p^2}$ clearly, and so whether $\gamma\alpha+(rt-2s)$ is a square in $\mathbb{F}_{p^2}$ is in accord with whether $\gamma\alpha-(rt-2s)$ is a square in $\mathbb{F}_{p^2}$. This means that\vspace{-1.5mm}
\[
\bigl\{\hspace{-0.3mm}\gamma\alpha+(rt-2s)\bigr\}\bigl\{\hspace{-0.3mm}\gamma\alpha-(rt-2s)\bigr\} = -4(r^2+s^2+t^2-rst-4)\vspace{-1mm}
\]
is a square in $\mathbb{F}_{p^2}$.
\end{proof}
\end{Lem}
\begin{Lem}
If a Ciani curve $C$ is superspecial, then $\frac{r^2-4}{t^2-4},\frac{s^2-4}{r^2-4}$ and $\frac{t^2-4}{s^2-4}$ are fourth powers in $\mathbb{F}_{p^2}$.
\begin{proof}
One can check that
\[
r^2-4 = \frac{D}{\lambda_1\lambda_2(1-\lambda_3)^2}, \quad s^2-4 = \frac{D}{\lambda_2\lambda_3(1-\lambda_1)^2}, \quad t^2-4 = \frac{D}{\lambda_3\lambda_1(1-\lambda_2)^2}
\]
with $D = (\lambda_1\lambda_2 + \lambda_2\lambda_3 + \lambda_3\lambda_1 - 1)^2 - 4\lambda_1\lambda_2\lambda_3(\lambda_1+\lambda_2+\lambda_3 - 2)$. Hence, we have
\[
\frac{r^2-4}{t^2-4} = \frac{\lambda_3(1-\lambda_2)^2}{\lambda_2(1-\lambda_3)^2}, \quad \frac{s^2-4}{r^2-4} = \frac{\lambda_1(1-\lambda_3)^2}{\lambda_3(1-\lambda_1)^2}, \quad \frac{t^2-4}{s^2-4} = \frac{\lambda_2(1-\lambda_1)^2}{\lambda_1(1-\lambda_2)^2}.
\]
Here, we have $\lambda_i, 1-\lambda_i \in (\mathbb{F}_{p^2}\hspace{-0.3mm})^4$ by Proposition 4.1, thus $\frac{r^2-4}{t^2-4},\frac{s^2-4}{r^2-4},\frac{t^2-4}{s^2-4}$ are fourth powers in $\mathbb{F}_{p^2}$. Indeed, the elliptic curve $y^2 = x(x-1)(x-\lambda_i)$ is isomorphic to $y^2 = x(x-1)(x-(1-\lambda_i))$ by the proof of \cite[Proposition\,III.1.7]{Silverman}. Therefore $y^2 = x(x-1)(x-(1-\lambda_i))$ is supersingular.
\end{proof}
\end{Lem}
Next, we construct other elliptic curves ${E'}_{\!i}$ which are 2-isogenous to $E_i$. We define
\begin{align*}
\mu_1 := (rt-2s) + 2\varDelta, \quad \nu_1 := (rt-2s) - 2\varDelta,\\
\mu_2 := (sr-2t) + 2\varDelta, \quad \nu_2 := (sr-2t) - 2\varDelta,\\
\hspace{0.2mm}\mu_3 := (ts-2r) + 2\varDelta, \quad \nu_3 := (ts-2r) - 2\varDelta,
\end{align*}
then we obtain $\mu_i,\nu_i \in \mathbb{F}_{p^2\hspace{-0.3mm}}$ by Lemma 4.2\,(2) if a Ciani curve $C$ is nonsingular and superspecial.
\begin{Lem}
Suppose that a Ciani curve $C$ is superspecial. Let ${E'}_{\!i}$ be the elliptic curve defined by\vspace{-0.5mm}
\[
{E'}_{\!i}: \mu_i\hspace{0.5mm}y^2 = x(x-1)(x-\nu_i/\mu_i).
\]
Then, there exists an isogeny $E_i \rightarrow {E'}_{\!i}$ defined over $\mathbb{F}_{p^2\hspace{-0.3mm}}$ of degree $2$.
\begin{proof}
In this proof, we will show only about ${E'}_{\!1}$. Firstly, the elliptic curve\vspace{-1mm}
\[
E_1: X^2 + y^4 + z^4 + rXy^2 + sy^2z^2 + tz^2X = 0\vspace{-1mm}
\]
is isomorphic to
\begin{equation}
v^2 = (r^2-4)u^4 + 2(rt-2s)u^2 + (t^2-4)
\end{equation}
by replacing $u = y$ and $v = 2X+ru^2 + t$. The elliptic curve of the form (4.1) is 2-isogenous to
\begin{equation}
v^2 = (r^2-4)u^3 + 2(rt-2s)u^2 + (t^2-4)u
\end{equation}
with the morphism $(u,v) \mapsto (u^2,uv)$ defined over $\mathbb{F}_{p^2}$. The right hand side of (4.2) can be factorized as $(r^2-4)u(u-{\mu}'_{1})(u-{\nu}'_{1})$, where
\[
{\mu}'_{1} = \frac{(rt-2s) + 2\varDelta}{r^2-4}, \quad {\nu}'_{1} = \frac{(rt-2s) - 2\varDelta}{r^2-4}.
\]
Here, the elliptic curve $v^2 = (r^2-4)u(u-{\mu}'_{1})(u-{\nu}'_{1})$ is isomorphic to
\[
{E'}_{\!1}: \mu_1y^2 = x(x-1)(x-{\nu}'_{1}/{\mu}'_{1}) = x(x-1)(x-\nu_1/\mu_1).
\]
This is the desired conclusion.
\end{proof}
\end{Lem}
There exists an isogeny ${\rm Jac}(C) \rightarrow {E'}_{\!1} \hspace{-0.3mm}\times\hspace{-0.3mm} {E'}_{\!2} \hspace{-0.3mm}\times\hspace{-0.3mm} {E'}_{\!3}$ defined over $\mathbb{F}_{p^2}$, whose degree is a power of $2$. In particular, a curve $C$ is maximal (resp.\ minimal) if and only if ${E'}_{\!i}$ are maximal (resp.\ minimal).
\begin{Lem}
If $C$ is superspecial, then all $\mu_i$ are squares in $\mathbb{F}_{p^2}$ or none of $\mu_i$ is a square in $\mathbb{F}_{p^2}$.
\begin{proof}
It suffices to show that products $\mu_1\mu_2,\mu_2\mu_3,\mu_3\mu_1$ are squares in $\mathbb{F}_{p^2}$. Recall from (3.5) that
\begin{align*}
\mu_1\nu_1 &= (rt-2s)^2 - 4(r^2+s^2+t^2-rst-4) = (r^2-4)(t^2-4),\\
\mu_2\nu_2 &= (sr-2t)^2 - 4(r^2+s^2+t^2-rst-4) = (s^2-4)(r^2-4),\\
\mu_3\nu_3 &= (ts-2r)^2 - 4(r^2+s^2+t^2-rst-4) = (t^2-4)(s^2-4).
\end{align*}
Since ${E'}_{\!i}$ is supersingular by assumption, thus $\nu_i/\mu_i$ is a fourth power in $(\mathbb{F}_{p^2}\hspace{-0.3mm})^\times$ for all $i \in \{1,2,3\}$ by Proposition 4.1. Hence, the product\vspace{-1mm}
\begin{align*}
\frac{\nu_1}{\mu_1} \cdot \frac{\nu_2}{\mu_2} = \frac{\mu_1\nu_1\mu_2\nu_2}{(\mu_1\mu_2)^2} = \frac{(r^2-4)^2(s^2-4)(t^2-4)}{(\mu_1\mu_2)^2}
\end{align*}
is a fourth power. On the other hand,
\[
(r^2-4)^2(s^2-4)(t^2-4) = (r^2-4)^4 \cdot \frac{s^2-4}{r^2-4} \cdot \frac{t^2-4}{r^2-4}
\]
is a fourth power by Lemma 4.3, so thus $(\mu_1\mu_2)^2 \in (\mathbb{F}_{p^2}\hspace{-0.3mm})^4$. This means that $\mu_1\mu_2$ is a square in $\mathbb{F}_{p^2}$. In the same way, the products $\mu_2\mu_3$ and $\mu_3\mu_1$ are also squares in $\mathbb{F}_{p^2}$.
\end{proof}
\end{Lem}
Now, we show that a nonsingular superspecial Ciani curve $C$ is maximal or minimal over $\mathbb{F}_{p^2}$.
\begin{proof}[Proof of second half of Theorem 1.1]
Suppose that a Ciani curve $C$ is superspecial, then the elliptic curves ${E'}_{\!1},{E'}_{\!2}$ and ${E'}_{\!3}$ are all supersingular.
\begin{itemize}
\item If $p \equiv 3 \pmod{4}$, then the elliptic curve ${E'}_{\!i}$ is maximal if and only if $\mu_i$ is a square in $\mathbb{F}_{p^2}$, and moreover ${E'}_{\!i}$ is minimal if and only if $\mu_i$ is not a square in $\mathbb{F}_{p^2}$ by using Proposition 4.1. It follows from Proposition 4.5 that all ${E'}_{\!i}$ are maximal or minimal, so the proof is done.\vspace{-1mm}
\item If $p \equiv 1 \pmod{4}$, then the elliptic curve ${E'}_{\!i}$ is minimal if and only if $\mu_i$ is a square in $\mathbb{F}_{p^2}$, and moreover ${E'}_{\!i}$ is maximal if and only if $\mu_i$ is not a square in $\mathbb{F}_{p^2}$ by using Proposition 4.1. It follows from Proposition 4.5 that all ${E'}_{\!i}$ are minimal or maximal, so the proof is done.
\end{itemize}\vspace{-7.7mm}
\end{proof}
\begin{Cor}
Suppose that a Ciani curve $C$ is superspecial, then the followings are true:
\begin{itemize}
\item If $p \equiv 3 \pmod{4}$, then $C$ is maximal if and only if $\mu_i$ for an (equivalently, every) $i \in \{1,2,3\}$ is a square in $\mathbb{F}_{p^2}$.\vspace{-1mm}
\item If $p \equiv 1 \pmod{4}$, then $C$ is maximal if and only if $\mu_i$ for an (equivalently, every) $i \in \{1,2,3\}$ is not a square in $\mathbb{F}_{p^2}$.
\end{itemize}
\end{Cor}
\begin{Exp}
Consider the Ciani curve $C$ defined by the equation
\[
C: x^4 + y^4 + z^4 + \frac{3}{\sqrt{-2}}x^2y^2 - \frac{9}{4}y^2z^2 + \frac{3}{\sqrt{-2}}z^2x^2 = 0.
\]
A simple computation says $\lambda_1 = -1$ and $\lambda_2 = \lambda_3 = 2$ by using Lemma 3.1. Hence, the three elliptic curves associated to the Ciani curve $C$ are given as below:\vspace{-1mm}
\begin{align*}
E_1: y^2 = x(x-1)(x+1),\\
E_2: y^2 = x(x-1)(x-2),\\
E_3: y^2 = x(x-1)(x-2).
\end{align*}
It follows from \cite[Example\,V.4.5]{Silverman} that these elliptic curves with $j(E_i) = 1728$ are supersingular if and only if $p \equiv 3 \pmod 4$. Therefore $C$ is superspecial if and only if $p \equiv 3 \pmod 4$. Moreover, the Ciani curve $C$ is maximal over $\mathbb{F}_{p^2}$ if and only if $p \equiv 3 \pmod 4$ since $\mu_1 = \frac{17\sqrt{-1}}{2} \in (\mathbb{F}_{p^2})^2$.
\end{Exp}
Lastly, let us discuss the maximality of a Ciani curve $C$ over $\mathbb{F}_{p^{2e}}$.
\begin{Cor}
Assume that a curve $C: x^4 + y^4 + z^4 + rx^2y^2 + sy^2z^2 + tz^2x^2 = 0$ is superspecial.\vspace{-1mm}
\begin{enumerate}
\item When $e$ is odd, then the followings are true:\vspace{-1mm}
\begin{itemize}
\item If $p \equiv 3 \pmod{4}$, then $C$ is maximal over $\mathbb{F}_{p^{2e}}\!$ if and only if $\mu_i$ for an (equivalently, every) $i \in \{1,2,3\}$ is a square in $\mathbb{F}_{p^2}$.\vspace{-1mm}
\item If $p \equiv 1 \pmod{4}$, then $C$ is maximal over $\mathbb{F}_{p^{2e}}\!$ if and only if $\mu_i$ for an (equivalently, every) $i \in \{1,2,3\}$ is not a square in $\mathbb{F}_{p^2}$.\vspace{-1mm}
\end{itemize}
\item When $e$ is even, then $C$ is not maximal (i.e. minimal) over $\mathbb{F}_{p^{2e}}$.
\end{enumerate}
\begin{proof}
Let $\varGamma$ be a maximal (resp. minimal) curve over $\mathbb{F}_{p^2}$, then $\varGamma$ over $\mathbb{F}_{p^{2e}}\!$ is also maximal (resp. minimal) if $e$ is odd, and is minimal if $e$ is even. This follows immediately from the Weil conjecture (cf. \cite[Appendix\hspace{1mm}C, Exercise 5.7]{Hartshorne}) and the fact that $\varGamma$ over $\mathbb{F}_{p^2}$ is maximal (resp. minimal) if and only if all the eigenvalues of the Frobenius on the first \'{e}tale cohomology group are $-p$ (resp. $p$).
\end{proof}
\end{Cor}
|
\section{Introduction}
Integrated photonic devices have led to game changing new capabilities in applications ranging from high-speed communications \cite{Marpaung2019-jx} to next-generation quantum computing platforms \cite{Arrazola2021-ms} and machine learning hardware accelerators \cite{Wetzstein2020-lz}.
These platforms stand to benefit from photonic components with improved performance, lower losses, larger bandwidths, and more compact footprints.
However, achieving such multi-faceted specifications is extremely challenging through conventional intuition-backed design methodologies.
In contrast, computational inverse design techniques can efficiently explore complex and high-dimensional design landscapes which are inaccessible to human designers.
Enabled by adjoint variable methods \cite{Veronis2004-zk, Molesky2018-oe}, gradient-based optimization techniques have proven to be among the most successful at producing photonic devices with high performance, with examples including wavelength-selective elements (e.g. multiplexers, filters, and resonators), signal routing elements (e.g. waveguide bends and couplers), and active components (e.g. modulators) \cite{Sell2017-hs, Hughes2018-rj, Jin2018-vz, Wang2018-wi, Lin2019-dt, Piggott2020-oa, Yang2020-jd, Sapra2020-zo, Minkov2020-cu, Tseng2021-jh, Hammond2021-pv}.
The approach in gradient-based design strategies is to first formulate an objective function, calculated in terms of a device's simulated performance.
Adjoint methods allow these objective functions and their gradients to be efficiently computed at a cost of only two full-wave simulations \cite{Veronis2004-zk, Molesky2018-oe}, independently of the number of degrees of freedom in a design.
When used in the context of modern automatic differentiation frameworks (e.g. JAX and TensorFlow), these techniques enable flexible end-to-end device optimization frameworks \cite{Minkov2020-cu, Hughes2019-qu}.
However, guaranteeing that an optimized design yields the desired performance, while simultaneously satisfying the numerous design rules of modern semiconductor manufacturing processes \cite{Liebmann2001-ta, Orcutt2010-sl}, still poses a major challenge.
Particularly relevant for integrated photonic devices are the capabilities of the lithographic systems, which define constraints such as the minimum feature length scale that can be reliably printed.
Elements of a design which are smaller than this length scale may print inconsistently or may be completely absent in the fabricated structure.
Moreover, inclusion of a design feature with sub-resolution size can have broad yield implications.
Thus, semiconductor foundries specify design rules which define the minimum width and minimum spacing of design features.
Compliance with these rules, as validated by a design rule check (DRC), is typically a prerequisite for fabrication.
There have been many proposals for DRC-compliant inverse design strategies.
One class of strategies selects a design parameterization with intrinsic geometric length scale guarantees, e.g. by optimizing the placement of geometric primitives which already satisfy the desired minimum feature size constraints \cite{Wang2020-cc}.
A related approach uses a parameterization which enables analytic constraints to be applied directly to the design degrees of freedom, e.g. by optimizing the coordinates of polygon vertices comprising shapes in the design mask \cite{Michaels2018-vd}.
However, the drawback in both of these strategies is that they offer up a limited landscape of potential designs.
For example, when optimizing the vertices of a polygon, an optimizer will be unable to alter the topology of a design by closing or opening holes.
In contrast, topology optimization strategies open up a much larger design landscape by parameterizing a design as a pixelated image, which is typically transformed by a sequence of convolutional filters and pixel-wise nonlinear functions \cite{Frei2007-pl, Jensen2011-br, Wang2012-iv, Qian2013-cc, Zhou2015-hg, Hagg2018-ku, Hammond2021-pv}.
These so-called density or three-field parameterizations allow a gradient-based optimization algorithm (e.g. LBFGS or MMA \cite{Svanberg1987-ya}) to modify any pixel within the design and apply arbitrary changes to the topology.
In practice however, there is often a trade-off between maintaining design feasibility (binarization of the pixels and feature size constraints) and attaining high performance.
This is because a continuous optimizer must traverse \textit{infeasible} regions of the design landscape before reaching \textit{feasible} regions, which may or may not have high performance.
The most common strategy in topology optimization is to begin with a continuous optimization phase before gradually ramping up binarization and applying design rule constraints.
However, if binarization is ramped up too quickly, a continuous optimizer will stall and fail to make progress towards improving performance.
Similarly, constraints or penalties on the feature size must be introduced in a way which will not push the optimizer away from high performance regions of the parameter space.
Often in these strategies, the selection of optimization hyper parameters and the detailed strategy for enforcing design feasibility is an art and, moreover, depends on both the type of device being optimized as well as the performance objective.
In this work we propose an \textit{always feasible} gradient-based inverse design framework with strict guarantees on DRC-compliance.
The key features of our proposal are a \emph{conditional generator} for feasible designs, which is combined with a \emph{straight-through estimator} (STE) to enable gradient backpropagation.
The combination of these two components yields a differentiable transform which can be incorporated into an computational graph, much like a convolutional filter or projection operator in density-based topology optimization.
We apply our framework to several practical photonic design challenges (a waveguide bend, a beam splitter, a waveguide mode converter, and a demultiplexer), and demonstrate that the approach reliably yields designs with high performance.
\section{Methodology}
\subsection{Brush-based minimum length scale constraints}
In this section, we introduce a mathematical framework for measuring whether a binarized pixel array is DRC-compliant, focusing on the minimum width and minimum spacing rules commonly required by semiconductor foundries.
In our framework, we consider binary arrays with values of $-1$ and $+1$, referring to these as \textit{void} and \textit{solid}, respectively.
During the course of a fabrication workflow, such designs may be converted to contours by off-the-shelf algorithms, such as marching squares or dual contouring \cite{Ju2002-jj}.
\begin{figure}
\centering
\includegraphics[width=3.3in]{brushes.pdf}
\caption{(a) Circular brush with diameter 13, and (b) notched square brush with width 13.}
\label{fig:brushes}
\end{figure}
All binary arrays can be realized by sequentially setting individual pixels to void or solid; conceptually, this is equivalent to the repeated placements and translations of single-pixel void or solid brushes. While a single-pixel brush can generate all possible binary arrays, some arrays are infeasible with larger brushes. In the trivial case, a single-pixel solid feature cannot be realized with a $3\times3$ pixel solid brush, since centering the brush on a single pixel also sets adjacent pixels to solid. Throughout this paper, we refer to the pixel on which the brush is centered as a touch, $t$ of the brush.
Our goal is to determine whether a given design, $x$ is feasible with a some brush, $b$, i.e. to determine whether all its solid and void features can be realized with $b$. When this is true, the condition
\begin{equation}
\mathcal{O}(x, b) = \lnot \mathcal{O}(\lnot x, b) = x
\label{eqn:condition}
\end{equation}
holds, where $\mathcal{O}(x, b)$ is the morphological opening of $x$ with brush $b$ and $\lnot$ is a logical \textit{not} operation which changes solid features into void features and void features into solid features, i.e. a value of -1 to 1 and a value of 1 to -1. The morphological opening operation is computed by the application of a binary erosion, $\mathcal{E}$ followed by a dilation \cite{erosion_dilation}, $\mathcal{D}$ with the same brush, i.e. $\mathcal{O}(x, b) = \mathcal{D}(\mathcal{E}(x, b), b)$. Intuitively, the dilation operation causes features in an image to expand while erosion causes features in an image to shrink.
An illustration of these operations is provided in the supporting information.
For a $b$-feasible design $x$, $\mathcal{E}(x, b)$ erodes the solid features and recovers the solid brush touches, while $\mathcal{E}(\lnot x, b)$ recovers the void brush touches. Dilation of the touches then restores the solid and void pixels, respectively. Intuitively, if $x$ contains a small feature that is incompatible with $b$, then erosion will remove it entirely and no solid pixels will be restored by the dilation. Thus, the binary opening operation removes small features, while leaving feasible features unchanged.
Length scale in a topology optimization context is typically associated with the minimum diameter of circular features. Thus, we say that if a design can be realized with an approximately circular brush $b_L$ with a diameter of $L$ pixels, it has a minimum length scale of at least $L \cdot d$, where $d$ is the design pixel pitch. The minimum length scale of the design is the largest value of $L$ for which Eqn. (\ref{eqn:condition}) is satisfied with brush $b_L$. When contours are obtained by dual contouring (i.e. outlining pixels, for binary designs), a design with minimum length scale will also have related minimum width and spacing, equal to the minimum width of the brush itself. An example circular brush with associated measurements is shown in Fig. \ref{fig:brushes}a and an example design which is feasible with the circular brush is shown in Fig. \ref{fig:designs}a.
We turn now to minimum width and spacing constraints which result from foundry design rules. Given a minimum width and spacing constraint \cite{koefferlein, calibre}, there are many brushes for which the associated feasible designs satisfy the constraint. The smallest of these is the notched square, i.e. a square brush which is solid everywhere except at the corners, as shown in Fig. \ref{fig:brushes}b. An example design is shown in Fig. \ref{fig:designs}b. When the constraint requires a physical minimum feature width and spacing of $w$, designs feasible with the notched square brush of width $L = w / d + 2$ will satisfy the constraint. Thus, if we can require that our designs are feasible with a notched square brush of width $L = w / d + 2$, we are also guaranteed to satisfy width and spacing rules of size $w$.
Although reasoning about minimum width and spacing is straightforward for dual-contoured designs feasible with the notched-square brush, we have also extensively validated that such designs are design-rule compliant using industry standard design software such as KLayout \cite{koefferlein} and Siemens Calibre Physical Verification \cite{calibre}.
\begin{figure}
\centering
\includegraphics[width=3.3in]{designs.pdf}
\caption{Random designs feasible for (a) diameter-13 circular brush, and (b) width-13 notched square brush.}
\label{fig:designs}
\end{figure}
\subsection{Generator for feasible designs}
In this section we describe our method of inverse design subject to strict length scale constraints. We have shown that designs which are feasible with a certain brush $b$ ($b$-feasible) will satisfy minimum length scale or width and spacing constraints. Properly, this can be considered a combinatorial optimization problem. However, our approach is to approximately solve this problem, by transforming it into an unconstrained stochastic gradient optimization problem. Critically, we use a conditional generator for feasible designs and a straight-through estimator (STE) for backpropagation; these two components are discussed subsequently.
In general, a binary design $x$ can be produced by dilation with $b$ of the solid and void touches $t_b^s$ and $t_b^v$, i.e.
\begin{equation}
x = \mathcal{D}(t_b^s, b) =\lnot \mathcal{D}(t_b^v, b)
\label{eqn:feasible}
\end{equation}
We consider the generation of $b$-feasible arrays as a process in which $t_b^s$ or $t_b^v$ are updated with a single-element change at each step. In this context, we introduce several touch states and pixel states which may be associated with each location at any step in the process. Defining these states for the solid case: \emph{existing} solid pixels are those which are assigned to solid by some touch in $t_b^s$,
\begin{equation}
p_{\text{existing}}^s = \mathcal{D}(t_b^s, b).
\end{equation}
\emph{Impossible} solid touches are those which would assign an existing void pixel to solid, and consists of all pixels within some distance of existing void pixels,
\begin{equation}
t_{\text{impossible}}^s = \mathcal{D}(p_{\text{existing}}^v, b).
\end{equation}
\emph{Valid} solid touches are those which are not impossible and not already in $t_b^s$,
\begin{equation}
t_{\text{valid}}^s = \lnot t_{\text{impossible}}^s \land \lnot t_b^s,
\end{equation}
where $\land$ is an element-wise logical \textit{and} operation. \emph{Possible} solid pixels are those obtained by dilating all existing or valid touches with $b$, i.e. they are those pixels which are or may be solid in a feasible array,
\begin{equation}
p_{\text{possible}}^s = \mathcal{D}(t_b^s \lor t_{\text{valid}}^s, b),
\end{equation}
where $\lor$ is an element-wise logical \textit{or} operation. \emph{Required} solid pixels are those which are not existing and not possible for void,
\begin{equation}
p_{\text{required}}^s = \lnot p_{\text{existing}}^s \land \lnot p_{\text{possible}}^v.
\end{equation}
\emph{Resolving} solid touches are those which are valid and would assign solid to a required-solid pixel,
\begin{equation}
t_{\text{resolving}}^s = \mathcal{D}(p_{\text{required}}^s, b) \land t_{\text{valid}}^s.
\end{equation}
Finally, \emph{free} touches are valid but only color existing pixels.
\begin{equation}
t_{\text{free}}^s = \lnot \mathcal{D}(p_{\text{possible}}^v \lor p_{\text{existing}}^v, b) \land t_{\text{valid}}^s.
\end{equation}
Corresponding void touches and pixels are be computed using the process. Fig. \ref{fig:traversal_snapshot} shows an example of pixels and touches computed from the steps above and a full example is illustrated in the Supporting Information.
\begin{figure}[t!]
\centering
\includegraphics[width=3.3in]{traversal_snapshot.pdf}
\caption{Design, pixels, and touches for an example traversal using the width-5 notched square brush.}
\label{fig:traversal_snapshot}
\end{figure}
A scheme that generates feasible arrays is outlined in Algorithm \ref{algorithm:generator}. The algorithm begins with empty $t_b^s$ and $t_b^v$ and loops until all pixels are existing for either solid or void. At each step, the pixel and touch states are updated. If free touches exist, all are selected. If resolving touches exist, one of these is selected; otherwise, any of the valid touches are selected.
\begin{figure}[t!]
\begin{algorithm}[H]
\caption{Generator for $b$-feasible designs}
\begin{algorithmic}
\Procedure{Generator}{$b$}
\State initialize empty $t_b^s$ and $t_b^v$
\While{design is incomplete}
\State update pixel and touch states
\If{free touches exist}
\State select all free touches
\ElsIf{resolving touches exist}
\State select a single resolving touch
\Else
\State select a single valid touch
\EndIf
\EndWhile
\EndProcedure
\end{algorithmic}
\label{algorithm:generator}
\end{algorithm}
\end{figure}
The algorithm sets at least one pixel at every step, and thus requires at most $N$ iterations for a reward array with $N$ pixels. A basic implementation of the algorithm involves full-array erosion and dilation operations at each step, which have cost proportional to $O(N)$, and leading to an overall cost $O(N^2)$. However, it is straightforward to implement windowed operations which have fixed cost independent of $N$, and so feasible design generation with $O(N)$ complexity is possible.
If touch selection is random, the algorithm simply yields a random feasible design. In practice, we greedily select the best touch as computed from a pixel reward array $\theta$, where $\theta$ is obtained from a latent design produced in the course of an optimization, and is fixed during each run of the generator. The total reward of a solid touch is the sum of $\theta$ elements that would be set to solid were the touch to be made, i.e. the sum of $\theta$ elements set by the touch. The reward of a void touch is the negative sum of elements set to void by the touch. A simple illustrated example of how rewards for touches are calculated can be found in the supporting information. With this change, the algorithm becomes a conditional generator of feasible designs, where $\theta$ biases the feasible design generation. The designs in Fig. \ref{fig:designs} have been created by running the algorithm on the same random reward array.
To gain some intuition for why this scheme always produces feasible designs, consider that the algorithm will select exclusively solid or exclusively void until reaching a state where all remaining pixels are possible for both solid and void. In this state, no touch can lead to an invalid status. Since the starting state is one where all pixels are possible for both solid and void, we can see that the algorithm will never produce an invalid status.
We emphasize that the full space of feasible designs of brush $b$ is accessible by Algorithm \ref{algorithm:generator}. This can be clearly seen, since e.g. if $\theta = x$ is provided, where $x$ is any feasible design, the output of the algorithm will simply be $x$.
\subsection{Straight through estimator}
Algorithm \ref{algorithm:generator} is not differentiable, and so it cannot be directly used with backpropagation in a gradient-based optimization setting. Drawing inspiration from the field of binary and quantized neural networks, we note that binary optimization can be viewed as topology optimization with a minimum length scale equal to a single pixel. Binary and quantized neural networks often make use of a STE in training, to backpropogate gradients through binary activation functions or to latent weights from which low-precision weights are obtained \cite{bengio2013estimating}. Specifically, in this approach one substitutes the gradient of a non-differentiable function with that of an estimator, i.e.
\begin{equation}
\frac{\partial f}{\partial x} \approx \frac{\partial}{\partial x}f_\text{estimator}
\end{equation}
A typical estimator used for binarization (e.g. via application of the sign function) is the identity operation.
\subsection{Computational graph}
The complete computational graph for the inverse design problem is depicted in Fig. \ref{fig:graph}. A latent design is passed through a transform, yielding the reward array $\theta$. This is passed to the conditional generator, which produces a feasible design. We pass the design to an electromagnetic simulation engine which computes scattering matrix elements; these are the inputs to an objective function, which computes a scalar loss value. Optionally, we symmetrize the transform output to favor symmetric feasible designs. All operations in this graph are differentiable, with the exception of the generator --- for which we use the STE. Thus, we can compute an estimated gradient of the loss with respect to the latent design.
\begin{figure}
\centering
\includegraphics[width=3.3in]{graph.pdf}
\caption{Computational graph of our inverse design scheme. Downward arrows indicate the forward computation, while upward arrows correspond to backpropagation. Gray boxes represent operations and text represents arrays or results.}
\label{fig:graph}
\end{figure}
For the transform, we have found that
\begin{equation}
\label{eqn:estimator}
D(t, b, \beta) = \tanh(\beta (t \circledast b))
\end{equation}
is a good choice, where $\beta$ is a scalar hyperparameter in the range 2-8. The convolutional transform aids in the optimization, as it ensures that the reward array $\theta$ is smooth, restricted in range, and has an associated length scale, determined by the brush size.
The form of our estimator matches that of the transform. As with the transform, outputs of the estimator have a characteristic length scale and are restricted to the range $(-1, 1)$, just as outputs of the generator itself are. Thus, the estimator can be seen as a differentiable approximation of the conditional generator. The success of this estimator is consistent with the finding in binary neural networks that estimators which approximate their forward-pass counterpart outperform simpler functions \cite{yin2019understanding}.
The latent design is randomly initialized with a bias so that the first feasible design is fully solid, and it is updated iteratively using the Adam scheme \cite{kingma2017adam}, which is a common choice for stochastic optimization problems in machine learning. We use a learning rate of 0.01, and $\beta_1=0.667$ and $\beta_2=0.9$ for the decay of the gradient and second moment exponential moving averages, respectively.
Notably, our method does not require scheduling of hyperparameters or changes in parameterization during the course of an optimization, and gives a feasible design at every optimization step. We expect that this simplicity could be advantageous from the perspective of a photonic designer tasked with creating a new component.
\subsection{Objective function}
In a practical design application, one is generally concerned with creating a component that satisfies a performance specification, i.e. a desired level of insertion loss, return loss, and cross talk. We define such specification by defining a cutoff value $|S_{\text{cutoff}}|$ for the magnitude for each scattering parameter element of $S$. Additionally, there is an implicit bound on the allowed values, which is imposed by physical constraints. For example, transmission magnitudes greater than unity are not possible for passive photonic components.
Our scalar objective function has the form
\begin{equation}
\mathcal{L}(S) = \left(\left\vert\left\vert\text{softplus}{\left(g \frac{\vert S\vert^2 -\vert S_{\text{cutoff}}\vert^2}{\min(w_{\text{valid}})}\right)}\right\vert\right\vert_2\right)^2,
\label{eqn:loss_fn}
\end{equation}
where $g$ is a vector of signs (+1 or -1) matching the size of $S$.
The values of $g$ are positive (+1) where the corresponding element of the cutoff array, $|S_{\text{cutoff}}|$ specifies the maximum value, and negative (-1) where $|S_{\text{cutoff}}|$ specifies the minimum value.
$w_{\text{valid}}$ also has the same size as $S$ and defines the range of allowed values for a scattering parameter element, which is defined as the distance between the performance cutoff, specified by $|S_{\text{cutoff}}|$, and the physical bound on the scattering parameter element.
For example, a minimum transmission amplitude cutoff of $0.5$ (-3 dB in power transmission) would have a $w_{\text{valid}}$ of $0.5$, because the maximum transmission possible is 1.0.
All operations other than the min in the softplus of Eqn. (\ref{eqn:loss_fn}) are element-wise.
\section{Results}
\subsection{Nanophotonic optimization problems}
In this section we apply our proposed inverse design framework to several integrated photonic components operating in the O-band, demonstrating designs for a waveguide bend, a mode converter, a beamsplitter, and a wavelength demultiplexer \cite{ceviche_challenges}.
All components are optimized for their characteristics within two 10 nm wavelength bands centered at 1270 nm and 1290 nm.
Performance targets for the components are specified in terms of their scattering parameters, and are summarized in Table \ref{tab:specs}.
Simulations of the components are performed using Ceviche \cite{Hughes2019-qu}, an open source 2D finite difference frequency domain (FDFD) simulator.
Silicon ($\varepsilon_r = 12.25$) and silicon oxide ($\varepsilon_r = 2.25$) are used for solid and void materials, respectively.
All components are coupled to waveguides having 400 nm width, and we deal with the fundamental waveguide mode unless noted otherwise.
\begin{table*}
\begin{tabular}{| r r | c | c | c | c |}
\hline
&
& \begin{tabular}{@{}c@{}}Waveguide \\ bend\end{tabular}
& \begin{tabular}{@{}c@{}}Mode \\ converter\end{tabular}
& \begin{tabular}{@{}c@{}} \ \\ Beamsplitter \end{tabular}
& \begin{tabular}{@{}c@{}}Wavelength \\ demultiplexer\end{tabular} \\
\hline
$S_{11}$& 1270 nm & $\leq$ -20 dB & $\leq$ -20 dB & $\leq$ -20 dB & $\leq$ -20 dB \\
& 1290 nm & $\leq$ -20 dB & $\leq$ -20 dB & $\leq$ -20 dB & $\leq$ -20 dB \\
\hline
$S_{21}$& 1270 nm & $\geq$ -0.5 dB & $\geq$ -0.5 dB & $\geq$ -3.5 dB & $\geq$ -3 dB \\
& 1290 nm & $\geq$ -0.5 dB & $\geq$ -0.5 dB & $\geq$ -3.5 dB & $\leq$ -20 dB \\
\hline
$S_{31}$& 1270 nm & - & - & $\geq$ -3.5 dB & $\leq$ -20 dB \\
& 1290 nm & - & - & $\geq$ -3.5 dB & $\geq$ -3 dB \\
\hline
$S_{41}$& 1270 nm & - & - & $\leq$ -20 dB & - \\
& 1290 nm & - & - & $\leq$ -20 dB & - \\
\hline
\end{tabular}
\caption{Performance specification for component power scattering parameters at the 1270 nm and 1290 nm wavelength bands.}
\label{tab:specs}
\end{table*}
The waveguide bend features a $1.6\times1.6 \ \mu\text{m}^2$ design region, with waveguides connecting to the left (port 1) and bottom (port 2). Given excitation from port 1, we aim to maximize transmission to port 2 while keeping backreflection low. Diagonal reflection symmetry is imposed, so that excitation from port 2 yields identical behavior.
The spatial mode converter features a $1.6\times1.6 \ \mu\text{m}^2$ design region, with waveguides connecting to the left and right. We seek designs that maximally convert the fundamental waveguide mode on the left (port 1) to the second order mode on the right (port 2), with minimal backreflection into port 1.
The beamsplitter design region is $3.2\times2.0 \ \mu\text{m}^2$ in size, with two waveguides connecting to the left (port 1 and port 4) and two waveguides connecting to the right (port 2 and port 3). Given excitation from port 1, the beamsplitter aims to divide power equally into port 2 and port 3 while minimizing backreflection into port 1 or transmission into port 4. We impose reflection symmetry along the horizontal and vertical axes, so that excitation from any other port yields identical behavior.
The wavelength demultiplexer features a $6.4\times6.4 \ \mu\text{m}^2$ design region, with one waveguide connecting to the left (port 1) and two waveguides connecting to the right (port 2 and port 3). Given excitation from port 1, wavelengths in the first band are directed to port 2 while wavelengths in the second band are directed to port 3.
In the optimization context, we consider three wavelengths per band --- the center and extremal wavelengths. The specifications are considered to be fulfilled if the criteria in \ref{tab:specs} are satisfied for all wavelengths. Our design resolution matches the 10 nm simulation resolution, so that e.g. a circular brush with diameter 10 corresponds to a 100 nm length scale.
\subsection{Designs using 100 nm circular brush}
\begin{figure}
\centering
\includegraphics[width=3.3in]{bend.pdf}
\caption{Evolution of a waveguide bend with 100 nm circular brush. Left: field magnitude for 1280 nm excitation with overlaid design. Right: power transmission into port 1 and 2, given excitation from port 1.}
\label{fig:bend}
\end{figure}
Fig. \ref{fig:bend} shows optimization results for the waveguide bend with 100 nm circular brush. Upon initialization, the design is entirely silicon with no enclosed oxide. The initial design is poor, with low transmission to the output port and substantial reflection back to the input port. Subsequent updates significantly modify the topology of the design, adding enclosed oxide features and developing isolated silicon features. We emphasize the key characteristic of our scheme --- the design at every step is fully binary and satisfies the 100 nm length scale constraint.
The target performance specification is achieved in 28 optimization steps, and the design continues to improve with more iterations. The lowest loss in the first 160 steps is found at step 122. Notably, the topology of this design differs from that at step 28, due to the fusing of two void features and the elimination of a third, along with changes in the shape of some solid features. The ability of our scheme to change topology while remaining feasible stands in contrast to some other methods, where the optimizer is locked in to a fixed topology once the optimization reaches an advanced state.
For the mode converter, beamsplitter, and wavelength demultiplexer, the lowest-loss design from the first 160 optimization steps and corresponding field magnitudes are shown in Fig. \ref{fig:other_designs}. The scattering spectra for each is shown in Fig. \ref{fig:other_designs_transmission}, where each satisfies the performance target laid out in Table \ref{tab:specs} and the 100 nm length scale constraint.
\begin{figure}
\centering
\includegraphics{other_designs.pdf}
\caption{Lowest loss designs achieved in the first 160 steps and their respective field magnitudes for (a) mode converter and (b) beamsplitter, both with 1280 nm excitation. Design and fields for the wavelength demultiplexer, with (c) 1270 nm and (d) 1290 nm excitation. All use a 100 nm circular brush.}
\label{fig:other_designs}
\end{figure}
\begin{figure}
\centering
\includegraphics{other_designs_transmission.pdf}
\caption{Scattering spectra for the (a) mode converter, (b) beamsplitter, and (c) wavelength demultiplexer with 100 nm circular brush.}
\label{fig:other_designs_transmission}
\end{figure}
\begin{figure}
\centering
\includegraphics{loss_trajectory_circle.pdf}
\caption{Normalized loss trajectories with 100 nm circular brush. Markers indicate steps where the performance target from Table \ref{tab:specs} is achieved.}
\label{fig:loss_trajectory}
\end{figure}
As we have shown, the topology of a design can change substantially during the course of the optimization. Since the response of a device is generally discontinuous with major changes in the design shape or topology, a noisy loss trajectory is expected. This is illustrated in Fig. \ref{fig:loss_trajectory}, which shows the normalized loss value versus iteration for the waveguide bend of Fig. \ref{fig:bend} and the other three components. Points where the performance target is achieved are also highlighted. For each component, the loss generally decreases with step, but noise in the loss value and occasional regressions to higher loss can be seen.
The waveguide bend and mode converter require fewer than 40 optimization iterations to achieve the target. Meanwhile, the beamsplitter and demultiplexer appear to be relatively more difficult problems, requiring approximately 100 steps to first reach the target. However, this is still well within the range of optimization steps required in an inverse design scheme \cite{Hammond2021-pv}. In general, difficulty is directly related to the target specifications, length scale constraints, the physical size of the component, and the configuration of connecting waveguides. Reducing the length scale or increasing the design size will expand the solution space, and generally allow satisfactory solutions to be found with fewer optimization steps.
\subsection{Designs using 100 nm notched square brush}
Next, we turn to component designs generated using a 100 nm notched square brush. With the design resolution of 10 nm, these designs strictly satisfy an 80 nm minimum width and spacing constraint. The normalized loss trajectories four the components are plotted in Fig. \ref{fig:loss_trajectory_notched_square}. Scattering spectra and designs are found in the Supporting Information.
For all components, designs achieving the performance target from Table \ref{tab:specs} are found. However, the loss trajectories differ in several respects. Specifically, the trajectories are noisier, they generally contain fewer points that achieve the spec, and the spec is achieved later in the optimization. We attribute this to the fact that the notched square is relatively larger in area, limiting the design space accessible to the optimizer and making the problem more challenging. We expect that optimal hyperparameters of the Adam algorithm used to drive our optimization are somewhat problem-dependent, and in future work it would be interesting to explore various configurations. In particular, reducing the learning rate as the optimization progresses, which has been shown to aid in training of quantized neural networks \cite{10.5555/3122009.3242044}, could be beneficial.
\begin{figure}
\centering
\includegraphics{loss_trajectory_notched_square.pdf}
\caption{Normalized loss trajectories with 100 nm notched square brush. Markers indicate steps where the performance target from Table \ref{tab:specs} is achieved.}
\label{fig:loss_trajectory_notched_square}
\end{figure}
\subsection{Reliability and effect of length scale}
Our results indicate that our inverse design framework is capable of finding designs with desired performance for a variety of fabrication-constrained photonic inverse design problems. A useful method will find such solutions for a range of length scale constraints, and do so reliably. To evaluate this, we consider the optimization problems above for circular and notched-square brushes with 60 nm, 80 nm, 100 nm, and 130 nm size. To study reliability of the method, we run 20 independent optimizations with different random initialization, each for 500 steps. Thus, for a single problem configuration (combination of device type, brush shape, and brush size), 10,000 feasible designs were generated.
\begin{figure}
\centering
\includegraphics{fraction_target.pdf}
\caption{Fraction of independent randomly-initialized optimization runs achieving the Table \ref{tab:specs} target performance as a function of the optimization step, for each component. Left column: with circular brush. Right column: with notched square brush.}
\label{fig:fraction_target}
\end{figure}
Our analysis proceeds as follows: for each step $i$ of an optimization run, we identify whether the target performance from Table \ref{tab:specs} is achieved at any step $j \leq i$. If the target performance is achieved at any such step, the run is considered successful at step $i$. Then, for each combination of component, brush shape, and brush size, we compute the number of runs which are successful at each step. From this, the fraction of successful runs is obtained and is shown in Fig. \ref{fig:fraction_target}.
Turning first to the 60 nm brushes, we observe that for all components and both brush shapes, we are able to consistently find designs which achieve the performance target. In most cases, only a few tens of optimization steps (and therefore simulations) are required. Random initialization affects the optimization trajectory, and the 20 different runs produce a distribution in the number of steps required to achieve the target. Comparing the circular and notched square brush, we find that with the circular brush the target performance is achieved slightly earlier, consistent with the hypothesis that the notched square presents a more challenging optimization problem. We observe this trend for other brush sizes discussed below. Across all component types designed with the 60 nm brush, the range of steps required to achieve the target performance is at most tens of steps wide, and in practical application a single optimization run may actually be sufficient to find a design which meets the desired performance specification.
When designing with a 80 nm brush, we again are able to consistently achieve the target performance. Generally, we observe that the number of required steps is larger than for the 60 nm brush. However, in some cases, such as the wavelength demultiplexer with circular brush, the number of steps required follow similar distribution as with the 60 nm brush. This suggests that for some problems, decreasing the length below a threshold value has no benefit. We also see some outlier examples where the target performance specification is not achieved.
As the brush sizes are increased to 100 nm, the general pattern is repeated. The target performance consistently achieved, but more steps are required and a larger number of outlier optimization runs fail to reach the target within 500 steps.
At a brush size of 130 nm, the trend is repeated yet again. With this brush size, designs for the waveguide bend and mode converter which achieve the target continue to be consistently found. For the beamsplitter and wavelength demultiplexer, however, approximately half or fewer optimization runs achieve the target. Therefore, in a practical application of our inverse design method with such large length scales, it may be beneficial to launch multiple optimization runs to ensure that at least one can reach the target performance. Additionally larger design regions may be necessary.
Much of the data supports the conclusion that larger brushes present more challenging optimization problems. We expect this also for theoretical reasons, since the design space with small brushes is strictly larger than that for larger brushes. However, with 130 nm brush we observe that some of the beamsplitter designs achieve the target performance ahead of the best 80 nm and 100 nm brush runs. This warrants further investigation in future work.
\section{Conclusion}
We have introduced an inverse design method which produces designs that are guaranteed to satisfy length scale constraints \textit{throughout} the course of an optimization.
Our always-feasible method is based on a conditional generator for designs and a straight-through gradient estimator, which is commonly used in quantized neural network training \cite{bengio2013estimating}.
Combined, these two components enable the problem of length-scale-constrained topology optimization to be posed as an unconstrained stochastic gradient optimization problem.
While the computational graph of our optimization scheme (Fig. \ref{fig:graph}) resembles that of some established topology optimization methods \cite{Hammond2021-pv, Zhou2015-hg}, where a latent design is transformed into a physical design, the conditional generator used in our scheme avoids the challenges that established topology optimization techniques face in transitioning from performant \textit{infeasible} regions of the design space to performant \textit{feasible} designs.
In such high-dimensional design landscapes there is no guarantee that a high-performing infeasible design will be close to high-performing feasible designs, especially in integrated photonic devices where performance depends strongly on the collective interference of light from multiple scattering interfaces.
Density-based optimization schemes often employ an unconstrained optimization stage \cite{Hammond2021-pv, Vercruysse2019-ej} followed by multiple stages of increasing constraint enforcement, where device performance often drops significantly as design constraints are enabled.
While the optimizer is often able to recover some level of performance, it is not guaranteed to be able to do so.
In some cases, the performance of the final feasible design remains lower than earlier infeasible designs.
A similar challenge in these optimization schemes lies in enforcing design binarization, which is often enabled gradually throughout several optimization stages by tuning a hyperparameter of the latent variable transformation \cite{Zhou2015-hg}.
While these changes to the device binarization reduce device performance, they present another challenge in that increased binarization is fundamentally at odds with the differentiability of the optimization graph. Essentially, the binarization nonlinearity approaches a step function with zero gradients.
The effect of this trade off is that during late stages of the optimization, the algorithm is not able to significantly change the topology of the structure.
In contrast, our always-feasible inverse design scheme always produces binarized designs and, moreover, is able to explore changes to the topology throughout the optimization (Fig. \ref{fig:bend}).
An added benefit of our always-feasible inverse design method is the user's ability to halt an optimization run once a design with sufficiently high performance is generated.
In contrast, obtaining a high performance design in a three field optimization scheme is typically \textit{not} a sufficient condition for halting an optimization because multiple rounds of infeasible optimization are required to achieve high performance.
Moreover, the higher levels of performance that are achieved during the unconstrained three field optimization stages are not guaranteed to be achieved during the stages where fabrication constraints and binarization have been applied.
The results presented in this paper demonstrate that our method can reliably find high-performing designs for a variety of 2D nanophotonic inverse design problems. We believe the good performance of our method, and its simplicity, could make it a useful new scheme in practical applications of inverse design. In the future, it would be important to use a realistic 3D electromagnetic simulator \cite{hughes2021perspective} to design components that can be manufactured in commercial foundries. Another interesting direction would be to apply the method to topology optimization problems in other physical domains.
We also see possible improvements to the underlying method. Specifically, conditional generators for designs satisfying additional constraints would be valuable, such as minimum solid area and void area \cite{Hammond2021-pv}, which are frequently included in foundry design rules. Even for width and spacing constraints addressed by the current scheme, it would be useful to develop new generators which are not conservative, i.e. which do not require a brush larger than the target minimum width. Finally, the estimator and transform functions warrant additional study, and we see potential for learned estimators.
|
\section{Introduction}
Transport in porous media takes place in a vast range of systems, natural as well as man-made. It is thus important to have a deep understanding of the relevant driving forces and their coupling, for instance to describe production of clean water \cite{rauter2021cassie, rauter2021thermo,qasim2019reverse}, CO$_2$ sequestration \cite{li2011pvtxy, ramdin2012state, boot2014carbon}, transport of oxygen, hydrogen, and water in fuel cell catalytic layers \cite{ sauermoser2020flow, wang2015theory}, and transport in lithium-ion battery electrodes and separators \cite{spitthoff2021peltier, gunnarshaug2020reversible,zhao2019review}.
The long-range aim of this work is to obtain a general thermodynamic theory of transport of immiscible fluids in porous media on the macroscale \cite{hassanizadeh1990mechanics, Kjelstrup2018, Kjelstrup2019}. This theory must first describe the thermodynamic state of the fluids in the porous media. To do this we employ a bottom-up approach, a procedure that includes all details of the system on the nanoscale in the construction of a representative elementary volume (REV) \cite{bear1988dynamics, blunt2017multiphase, Bedeaux2020}. The procedure gives a coarse-grained description of the REV on the macroscale, or what is called the Darcy level \cite{bear1988dynamics, hassanizadeh1990mechanics,blunt2017multiphase,Kjelstrup2018,Kjelstrup2019}. A central issue is to find the pressure of the REV. Figure \ref{fig:system} illustrates the problem for the cases studied in this work. A fluid (blue) occupies the pores in a porous material (grey). The pores are so narrow that interactions between fluid and wall become significant, or in other words, that the fluid-wall surface energy becomes significant. But how can we define and determine the properties of the REV, for instance, the pressure? Can we find a representative elementary volume, for which this is possible? In this work, we aim to find answers to these questions.
In bulk fluids, the hydrostatic pressure is well defined, measurable, and well documented as the driving force of the fluid flow. In porous media, where fluids are confined by the pores, however, there is no consensus of neither the thermodynamic nor the mechanical definitions of the pressures. The microscopic mechanical pressure tensor is inherently ambiguous \cite{Irving1950,Schofield1982}. In this context, we make an emphasis on the milestone work of Israelachvili \cite{israelachvili2015intermolecular} who documented short- and long-range forces on fluid particles exerted by the surroundings, and on the thermodynamic analysis that was pioneered by Derjaguin \cite{Derjaguin1934}.
We have previously proposed thermodynamic definitions for the pressure of heterogeneous media, applying Hill's idea of thermodynamics for small systems \cite{Kjelstrup2018, Kjelstrup2019, Galteland2019, Galteland2021, galteland2021legendre}. Hill's definitions have so far only been tested under simple conditions \cite{Rauter2020, Galteland2019, Galteland2021, Erdos2020}. They have not been studied for a variety of shapes and pore sizes.
In Hill's thermodynamics of small systems or nanothermodynamics (See Bedeaux and Kjelstrup \cite{Bedeaux2020}), the REV of the porous medium is considered as an open system, controlled by the temperature, chemical potential, or pressure in the environment. The thermodynamic analysis is thus applied to a grand canonical ensemble of systems and a new variable is introduced, the replica energy, i.e. the energy needed to create one more small system in the ensemble. An adjusted Gibbs-Duhem equation (the Hill-Gibbs-Duhem's equation) appears. This opens up a new route to determine the REV pressure.
Using this as a starting point in section \ref{sec:theory}, we shall find expressions for the so-called integral and differential pressures of the REV, from the system's replica energy. We present a new route to the pressure via the chemical potential of the reservoir in equilibrium with the REV. Molecular dynamics simulations, described in section \ref{sec:methods}, will be used to illustrate and verify the theoretical steps.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{system.png}
\caption{A single-phase and -component fluid in a porous medium. The fluid-solid interfaces can have complex shapes and the volumes can be on the nanoscale.}
\label{fig:system}
\end{figure}
\section{Theory}
\label{sec:theory}
The systematic procedure of Hill consists of the three steps \cite{Bedeaux2020}, which we repeat to give an overview of the procedure. Concepts will be defined in the text that follows.
\begin{enumerate}
\item To start, define the items which make the system small in Hill's sense, by defining the relevant REV. Find the corresponding Hill-Gibbs equation. Next, define the environmental variables that control the system.
\item From an analysis of the ensemble of small systems, find the system's replica energy in terms of its subdivision potential.
\item Derive the corresponding Hill-Gibbs-Duhem equation. This equation can be used to find the REV thermodynamic variables and their interrelations.
\end{enumerate}
\subsection{The general Hill-Gibbs equation of an ensemble of open porous media REVs}
We shall consider the REV of a porous medium, filled with a single-phase, single-component fluid, see figure \ref{fig:system}. The REV volume is $V=V^f+V^s$, where $V^f$ is the volume of the fluid and $V^s$ is the volume of the solid. The porosity is the fraction of the fluid volume to the total volume, $\phi=V^f/V$. These are time-independent properties in the system. The REV is a small system in Hill's sense \cite{Hill1963, Bedeaux2020} because the fluid is confined; it does not have bulk-scale properties. The system is open for the supply of fluid particles and energy. The number of solid particles is fixed, however.
The ensemble of $\mathcal{N}$ REVs has a total entropy $S_t$, total fluid volume $V^f_t$, total number of fluid particles $N^f_{t}$, total number of solid particles $N^s_{t}$ (with total volume $V^s_t$), and total fluid-solid surface area $A_t$. The total differential of the total internal energy of the ensemble of REVs is then
\begin{equation}
\d U_t = T\d S_t-p^f\d V^f_t-p^s\d V^s_t + \mu^f\d N^f_{t} +\mu^s\d N^s_{t} +\gamma\d A_t +\varepsilon \d \mathcal{N}.
\label{eq:total_internal_energy}
\end{equation}
This equation has been called the Hill-Gibbs equation \cite{Bedeaux2020, Galteland2021}. The temperature, fluid pressure, and solid pressures, chemical potentials, and surface tension are defined from the variables involved as
\begin{equation}
\begin{split}
T&\equiv\left(\frac{\partial U_t}{\partial S_t}\right)_{V^f_t,V^s_t,N^f_t,N^s_t,A_t,\mathcal{N}},\qquad
p^f\equiv\left(\frac{\partial U_t}{\partial V^f_t}\right)_{S_t,V^s_t,N^f_t,N^s_t,A_t,\mathcal{N}},\\
p^s&\equiv\left(\frac{\partial U_t}{\partial V^s_t}\right)_{S_t,V^f_t,N^f_t,N^s_t,A_t,\mathcal{N}},\qquad
\mu^f\equiv\left(\frac{\partial U_t}{\partial N^f_t}\right)_{S_t,V^f_t,V^s_t,N^s_t,A_t,\mathcal{N}},\\
\mu^s&\equiv\left(\frac{\partial U_t}{\partial N^s_t}\right)_{S_t,V^f_t,V^s_t,N^f_t,A_t,\mathcal{N}},\qquad
\gamma\equiv\left(\frac{\partial U_t}{\partial A_t}\right)_{S_t,V^f_t,V^s_t,N^f_t,N^s_t,\mathcal{N}}.
\end{split}
\end{equation}
The subdivision potential $\varepsilon$ is the property that deals with system smallness in particular. It is introduced with its conjugate variable, the number of REVs, $\mathcal{N}$. The subdivision potential is the internal energy required to add one more REV to the ensemble under the specified conditions,
\begin{equation}
\varepsilon \equiv \left(\frac{\partial U_t}{\partial \mathcal{N}}\right)_{S_t, V^f_t, V^s_t,N^f_t, N^s_t, A_t}.
\end{equation}
When the subdivision potential differs from zero the system is small. This property will, as we shall see below, adjust the common variables like the pressure, and turn them into effective new variables. In the case of the pressure, the adjustment leads to the integral pressure, central for porous media.
\subsection{The replica energy}
To find the properties of one system, we need to describe the thermodynamic state of a single-phase, single-component fluid in the REV. The fluid is free to move (the system is open), while the solid is not. The system exchanges fluid particles with the surroundings, but not solid particles. The variables that are controlled by contact with the environment are the temperature and the fluid chemical potential. Apart from these, as stated above, the fluid volume, solid volume, surface area, and the number of solid particles are control variables too. The type of ensemble constituted by this set of variables is particularly suited for the transport of fluids through a porous medium.
In order to change the set of variables in the general expression equation \ref{eq:total_internal_energy}, into the above preferred set of control variables, we express the additive variables in the original set of total variables by their controlled value times the number of replicas. This provides also the single system properties that we are after. The controlled fluid and solid volumes, and surface area per REV are $V^f_t = V^f \mathcal{N}$, $V^s_t = V^s\mathcal{N}$, and $A_t = A \mathcal{N}$, respectively. In addition we introduce the controlled number of solid particles per REV, $N_t^s=N^s\mathcal{N}$. By introducing these control variables into the Hill-Gibbs equation for the ensemble, we obtain
\begin{equation}
\d U_t = T\d S_t -p^f\mathcal{N}\d V^f-p^s\mathcal{N}\d V^s + \mu^f\d N_t^f + \mu^s\mathcal{N}\d N^s +\gamma\mathcal{N}\d A+X\d\mathcal{N},
\label{eq:hill_gibbs2}
\end{equation}
where the last term is the replica energy of one small system
\begin{equation}
X = \varepsilon-p^fV^f-p^sV^s+\gamma A+\mu^sN^s.
\label{eq:Replica}
\end{equation}
The replica energy \cite{Hill1963, Hill1964} expresses in a simpler way, the energy required to add one more small system to the ensemble of systems under the conditions controlled. The combination of terms can define the integral pressure minus the integral solid chemical potential times number of solid particles \cite{Galteland2021},
\begin{equation}
-\hat{p}V +\hat{\mu}^sN^s \equiv X,
\end{equation}
however, we need an additional equation to determine both $\hat{p}$ and $\hat{\mu}^s$ separately. Through the introduction of an ensemble of the systems, we have achieved that the internal energy, here $U_t$, is an Euler homogeneous of the first order in the number of REVs. This gave Hill the motivation for the use of an ensemble, and to define the conjugate variables $\varepsilon$ and $\mathcal{N}$.
\subsection{The Hill-Gibbs equation for a single small system}
We are now in a position to integrate the total differential of the total internal energy, see equation \ref{eq:hill_gibbs2}, at constant $T, V^f, V^s, \mu^f, N^s$, and $A$. This gives,
\begin{equation}
U_t = TS_t+\mu^fN^f_t+X\mathcal{N}.
\end{equation}
By introducing the REV average properties, we obtain the internal energy of one REV,
\begin{equation}
U = TS+\mu^fN^f+X.
\label{eq:internal_energy}
\end{equation}
The total differential of the internal energy of one REV is
\begin{equation}
\d U = T\d S -p^f\d V^f-p^s\d V^s+\mu^f\d N^f+\mu^s\d N^s + \gamma \d A.
\end{equation}
By differentiating the internal energy of the REV and using its total differential, we obtain the total differential of the replica energy
\begin{equation}
\d X = \d(-\hat{p}V+\hat{\mu}^sN^s) = -S\d T -p^f\d V^f-p^s\d V^s-N^f\d \mu^f+\mu^s\d N^s +\gamma \d A.
\label{eq:HGD}
\end{equation}
The equation is the outcome of step 2 in Hill's procedure presented in the introduction to this section. The equation can be seen as an extension of Gibbs-Duhem's equation, so we have called it Hill-Gibbs-Duhem's equation \cite{Bedeaux2020}. Applications can now be specified.
The partial derivatives of the replica energy follow
\begin{equation}
\begin{split}
S&=-\left(\frac{\partial X}{\partial T}\right)_{V^f,V^s,\mu^f,N^s, A},\qquad
p^f=-\left(\frac{\partial X}{\partial V^f}\right)_{T,V^s,\mu^f,N^s, A},\\
p^s&=-\left(\frac{\partial X}{\partial V^s}\right)_{T,V^f,\mu^f,N^s, A},\qquad
N^f=-\left(\frac{\partial X}{\partial \mu^f}\right)_{T,V^f,V^s,N^s, A},\\
\mu^s&=\left(\frac{\partial X}{\partial N^s}\right)_{T,V^f,V^s,N^f,A},\qquad
\gamma=\left(\frac{\partial X}{\partial A}\right)_{T,V^f,V^s,N^f,\mu^s}.
\end{split}
\end{equation}
With this set of equations we can calculate all the necessary REV properties of a porous medium. We shall concentrate on the integral fluid pressure and the route to this quantity via the chemical potential.
\subsection{The integral pressure and the chemical potential of the solid in the REV}
We apply here the conditions of constant temperature, fluid volume, solid volume, number of solid particles, and surface area. The Hill-Gibbs-Duhem's equation reduces to
\begin{equation}
\d X = \d(-\hat{p}V+\hat{\mu}^sN^s) = -N^f\d \mu^f
\label{eq:HGD2}
\end{equation}
or, after dividing by $V$, the volume of the REV,
\begin{equation}
\d x = \d(-\hat{p}+ \hat{\mu}^s\rho^s) = -\rho^f\d\mu^f
\label{eq:HGD3}
\end{equation}
Where the density of the fluid in the REV is $\rho^f = N^f/V$ and the density of the solid is $\rho^s = N^s/V$. These densities are of the total REV volume $V$, and not the fluid volume $V^f$ or solid volume $V^s$. The difference of the replica energy density can be calculated from
\begin{equation}
x -x_\infty= -\int_{-\infty}^{\mu^f}\rho^f\d\mu'^f,
\label{eq:replica_energy_density}
\end{equation}
The replica energy density is zero as the fluid chemical potential approaches minus infinity. The replica energy density depends on two unknown variables, $\hat{p}$ and $\hat{\mu}^s$. To proceed we need to know more about these variables.
\subsubsection{Constant integral pressure across boundary}
The integral pressure of the REV can be obtained from equation \ref{eq:HGD3}
\begin{equation}
\d\hat{p} = \rho^f\d\mu^f + \rho^s\d\hat{\mu}^s.
\label{eq:dp_hat}
\end{equation}
The fluid in the porous medium is in equilibrium with its environment, here the bulk phase fluid that surrounds the system (denoted $b$). The environment has the same temperature. The integral pressure was observed to be constant across the phase boundary inside a pore \cite{Rauter2020}. We shall therefore make the assumption that also in this case;
\begin{equation}
\hat{p} = p^b
\label{eq:loce}
\end{equation}
where $p^b$ is the bulk pressure of a fluid in equilibrium with the porous medium. For an ideal gas in a cubic confinement it has been shown that the integral pressure is equal to the bulk pressure \cite{braaten2021small}. If this applies for this system, we also have
\begin{equation}
\hat{p} = \int_{-\infty}^{\mu^f}\rho^b\d\mu'^f,
\end{equation}
where $\rho^b$ is the fluid number density in the bulk phase. The integral pressure is zero as the fluid chemical potential approaches minus infinity. As a consequence, the integral pressure depends only on the fluid chemical potential and temperature. The integral chemical potential of the solid in the REV is then
\begin{equation}
\hat{\mu}^s = \frac{1}{\rho^s}\int_{-\infty}^{\mu^f}(\rho^b-\rho^f)\d \mu'^f.
\label{eq:solid_chemical_potential}
\end{equation}
The solid chemical potential is zero at minus infinite fluid chemical potential. The entropy density of the REV is
\begin{equation}
s = \frac{S}{V} = \frac{1}{T}( u - \mu^f\rho^f-x)
\label{eq:entropy}
\end{equation}
where we used equation \ref{eq:internal_energy} and introduced the replica energy density, $x$.
\subsection{Pressures and surface tension}
There are three pressures; the integral, fluid and solid pressures. The latter two are differential pressures. The integral pressure is a combination of the differential pressures and the (differential) surface tension,
\begin{equation}
\hat{p} = p^f\phi+p^s(1-\phi)-\gamma A/V = p^b.
\label{eq:phat}
\end{equation}
We assume the integral pressure to be equal everywhere in equilibrium and also equal to the bulk pressure of a bulk fluid in equilibrium with the porous medium. If we also assume the distances between solid surfaces to be large and their curvature to be small, we can approximate the fluid pressure to be equal to the bulk pressure \cite{Galteland2021}. In that case, the solid pressure is equal to
\begin{equation}
p^s = p^b + \frac{\gamma A}{V^s}.
\label{eq:solid_pressure}
\end{equation}
\section{Simulation details}
\label{sec:methods}
Systems of fluid and solid particles were investigated with molecular dynamics simulations using LAMMPS \cite{Plimpton1995,thompson2021lammps}. The three different systems have been simulated: A bulk fluid, a single solid particle surrounded by fluid particles, and a face-centered cubic (fcc) lattice of solid particles filled with fluid particles in the pore space. The two latter systems are illustrated in figure \ref{fig:simulation}. The systems were simulated in the grand canonical ensemble and had periodic boundary conditions in all directions. The bulk fluid was simulated to calculate the bulk pressure as a function of the fluid chemical potential. The single solid particle surrounded by fluid particles will be considered as if the lattice constant of the fcc lattice is large. This system was simulated to calculate the fluid-solid surface tension when the solid particles are far apart. The thermodynamic properties of the fluid in the fcc lattice were calculated as a function of the fluid chemical potential.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{simulation.png}
\caption{Visualization of a two of the simulated systems. Fluid particles (red) and solid particles (blue) are visualised with OVITO \cite{Stukowski2009}. Only the particles in a slab centered on the solid particles of thickness $2.5\sigma$ is shown. The fluid chemical potential is $\mu^f=\zeta$ and the radius of the solid particles is $R=5\sigma$. Left: Single solid particle. Right: An fcc lattice with lattice constant $a=17.5\sigma$ and porosity $\phi=0.61$.}
\label{fig:simulation}
\end{figure}
The temperature was controlled by using a Nosé-Hoover thermostat \cite{nose1984unified, hoover1985canonical} adapted by Shinoda \textit{et. al.} \cite{shinoda2004rapid}. The fluid chemical potential was controlled by using grand canonical Monte Carlo insertions and deletions of fluid particles \cite{Frenkel2001}. A particle is inserted at random position in the simulation box with a probability
\begin{equation}
\text{acc}(N^f\rightarrow N^f+1) = \min\left\{1,\quad \frac{V}{\Lambda^3(N^f+1)}\exp[\beta(\mu-E_p(N^f+1)+E_p(N^f)] \right\}.
\end{equation}
A random particle is removed from the simulation box with a probability
\begin{equation}
\text{acc}(N^f\rightarrow N^f-1) = \min\left\{1,\quad \frac{\Lambda^3(N^f+1)}{V}\exp[-\beta(\mu+E_p(N^f-1)-E_p(N^f)] \right\}.
\end{equation}
Where $\Lambda=\sqrt{h^2/(2\pi m k_\text{B}T)}$ is the de Broglie thermal wavelength, $h$ is the Planck constant, $E_p$ is the potential energy and $\beta=1/(k_\text{B}T)$ where $k_\text{B}$ is Boltzmanns constant.
\subsection{Simulation procedure}
In the bulk system, the fluid chemical potential was varied in the range $\mu^f\in[-7,1]$, in which the fluid density varied from a dilute gas to a dense liquid. The system was initialized with an empty cubic simulation box of volume $V=(40\sigma)^3$, and the grand canonical Monte Carlo technique was used to insert particles. The simulation was run until the average number of fluid particles was no longer increasing. The bulk fluid number density $\rho^b=N^f/V$ and pressure $p=(P_{xx}+P_{yy}+P_{zz})/3$ were calculated as a function of the fluid chemical potential for $10^7$ steps. The calculation of the mechanical pressure tensor $P_{\alpha\beta}$ is described below in subsection \ref{sec:mechanical_pressure_tensor}.
The system of the single solid particle was simulated by first placing the solid particle in the center of the cubic simulation box of volume $V=(30\sigma)^3$. The radius of the solid particle was $R=5\sigma$, where $\sigma$ is the diameter of the fluid particle. The fluid chemical potential was varied in the range $\mu\in[-7, 1]\zeta$, where $\zeta$ is the minimum of the pair-wise interaction. The system was initialized by placing the fluid particles in an fcc lattice with a fluid number density approximately equal to the bulk fluid number density around the solid particle at a given fluid chemical potential. This was done to reduce the computational time. The simulation was run until the average number of fluid particles no longer increased, after which it was run for an additional $10^7$ steps to calculate the surface tension as a function of the fluid chemical potential. The calculation of the surface is described below in section \ref{sec:mechanical_pressure_tensor}.
The fcc lattice of solid particles was fixed in space with lattice constant in the range $a\in[15,30]\sigma$, the radius of the solid particles was $R=5\sigma$. The simulation box lengths were $L=2a$, total volume $V=8a^3$, number of solid particles $N^s = 32$, and solid volume $V^s = 128\pi R^3/3$. We have previously found that the smallest REV of such a system is a quarter unit cell \cite{Galteland2019}. The fluid particles were inserted in the pore space in between the solid particles. The porosity of the system was
\begin{equation}
\phi = \frac{V^f}{V}=1-\frac{16\pi R^3}{3a^3},
\end{equation}
where the radius of the solid particles was $R=5\sigma$ given that the lattice constant is $a\geq2\sqrt{2}R\approx14\sigma$. For lower values of $a$ the solid particles overlap, however, we do not look at such systems. The porosity varied from approximately $0.38$ to $0.92$. The fluid chemical potential, number of solid particles, temperature, and volume were controlled.
The system was initialized by placing the solid particles in an fcc lattice with lattice constant $a$ and fluid particles also in an fcc lattice with a number density approximately equal to $\rho^b(\mu^f)$ around the solid particles. The simulation was run until the number of fluid particles was constant, after which the simulation was run for an additional $10^7$ timesteps to calculate thermodynamic properties.
\subsection{The Lennard-Jones/spline potential}
The particles interacted with the Lennard-Jones/spline (LJ/s) potential \cite{Hafskjold2019}, which is equal to the Lennard-Jones potential up to the inflection point $r_s$ shifted by a hard-core diameter $d$. At $r_s$ a third degree polynomial is fitted for the potential to be zero at the cut-off $r_c$ and the force and potential to be continuous at the inflection point $r_s$ and cut-off $r_c$. See the works by Hafskjold \textit{et al.} \cite{Hafskjold2019} and Kristiansen \cite{kristiansen2020transport} for details on the properties of the LJ/s potential. The LJ/s potential is,
\begin{equation}
\nu^\text{LJ/s}(r)=
\begin{cases}
4\zeta\left[\left(\frac{\sigma}{r-d}\right)^{12}-\left(\frac{\sigma}{r-d}\right)^{6}\right] &\text{if } r<r_s,\\
a(r-r_c)^{2}+b(r-r_c)^{3} &\text{if } r_s<r<r_c,\\
0 &\text{else}
\end{cases}
\end{equation}
where $\zeta$ is the minimum of the interaction potential. This symbol is used to avoid confusion with the subdivision potential $\varepsilon$. The soft-core diameter is $\sigma$, $d$ is the hard-core diameter, $r=|\bm{r}_j-\bm{r}_i|$ is the distance between particle $i$ and $j$. The distance $r=\sigma+d$ is the smallest distance where the potential is zero. The Lennard-Jones/spline interaction potential was equal for all particle pairs, but was shifted with hard-core diameter $d$. There were three particle pair interactions, fluid-fluid, fluid-solid, and solid-solid, and only the hard-core diameter varied between them. The solid-solid interaction we set to zero. The hard-core diameter was $d_\text{ff}=0$, $d_\text{fs}=4.5\sigma$, and $d_\text{ss}=9\sigma$ for the fluid-fluid, fluid-solid, and solid-solid interactions, respectively. The solid particle radius was defined as $R\equiv (d_\text{ss}+\sigma)/2 = 5\sigma$. Other definitions of the radii are possible, for example based on Gibbs dividing surface or the surface of tension. The parameters $a$, $b$, $r_s$ and $r_c$ were determined such that the potential and the force were continuous at $r=r_s$ and $r=r_c$. The masses of fluid particles were $m$, and the solid particles were considered to have infinite mass as they were fixed in space.
\subsection{Internal energy density}
The internal energy density was calculated as the sum of the kinetic and potential energy densities,
\begin{equation}
u = \frac{1}{V}\left(\sum_{i=1}^N\sum_{j>i}^N u^\text{LJ/s}(r_{ij}) + \frac{1}{2}\sum_{i=1}^Nm_i(\bm{v}_i\cdot\bm{v}_i)\right)
\end{equation}
where $V$ is the total volume, $N$ is the total number of particles, and $\bm{v}_i$ is the velocity of particle $i$. The solid particles did not contribute to the kinetic energy, as their velocities were controlled to be zero. The internal energy density was used to calculate the entropy density, see equation \ref{eq:entropy}.
\subsection{The mechanical pressure tensor}
\label{sec:mechanical_pressure_tensor}
The mechanical pressure tensor was calculated in spherical and Cartesian coordinates, see Ikeshoji \textit{et.al.} for details \cite{Ikeshoji2003}. It was calculated in spherical coordinates for the single solid particle surrounded by fluid, and in Cartesian coordinates for the bulk fluid and the fcc lattice of solid particles with inserted fluid particles. For the bulk fluid and fcc lattice, the pressure was calculated for the whole simulation box, with sides $L\in[30, 60]\sigma$ and interaction cut-off $r_c\approx 1.74\sigma$. While for the single solid particle it was calculated in spherical shell subvolumes with the origin at the center of the solid particle. This was done to calculate the surface tension.
To avoid confusion between the thermodynamic and mechanical pressures, we will use lower case $p$ for all thermodynamic pressures and upper case $P$ for mechanical pressures. We make this distinction because there is no consensus on how, if possible, to connect the two for heterogeneous media \cite{Long2011, Dijk2020, Long2020, Galteland2021}.
The mechanical pressure tensor can be written as the sum of the ideal gas contribution and a virial contribution,
\begin{equation}
P_{\alpha\beta} = \rho k_\text{B}T\delta_{\alpha\beta}+P_{\alpha\beta}^v,
\end{equation}
where $\delta_{\alpha\beta}$ is the Kronecker delta and the subscripts give the components of the tensor. The virial contribution is due to particle pair interaction, and is calculated as a sum over all particle pairs. For a subvolume $V_k$ the virial contribution is
\begin{equation}
P_{\alpha\beta}^v = - \frac{1}{V_k}\sum_{i=1}^N\sum_{j>i}^Nf_{ij,\alpha}\int_{C_{ij}\in V_k}\d l_\beta.
\end{equation}
Where $N$ is the total number of particles and $f_{ij,\alpha}$ is the $\alpha$-component of force acting on particle $i$ due to particle $j$. The line integral is along the part of the curve $C_{ij}$ that is contained in the subvolume $V_k$. The virial contribution is inherently ambiguous as any continuous curve that starts at the center particle $i$ and ends at the center particle $j$ is permitted \cite{Irving1950, Harasima1958, Schofield1982}. In this work, we have used the Irving-Kirkwood curve, which is the straight line from the center of particle $i$ to the center of particle $j$.
The Cartesian mechanical pressure tensor was calculated for the whole simulation box. The line integral in the virial contribution reduces it to
\begin{equation}
\label{eq:mechanical_pressure}
P_{\alpha\beta}^v = - \frac{1}{V_k}\sum_{i=1}^N\sum_{j>i}^Nf_{ij,\alpha}r_{ij,\beta}.
\end{equation}
where $r_{ij}$ is the line for the center of particle $i$ to the center of particle $j$.
The pressure is calculated as the mean of the diagonal components of the Cartesian mechanical pressure tensor. The bulk pressure is
\begin{equation}
p^b = \frac{1}{3}(P_{xx}+P_{zz}+P_{yy}).
\end{equation}
This was also calculated for the fluid in the fcc lattice of solid spheres. However, we have not equated it to a thermodynamic property.
The fluid-solid surface tension was calculated for the single-sphere simulation case from the spherical mechanical pressure tensor,
\begin{equation}
\gamma = \frac{1}{R^2}\int_R^{r_0}(P_N-P_T) r^2\d r.
\label{eq:surface_tension}
\end{equation}
Where $R=5\sigma$ is the solid particle radius, and $r_0=14\sigma$ is a position in the fluid far away from the fluid-solid surface. The normal component is $P_N = P_{rr}$ and the tangential component is $P_T = (P_{\phi\phi}+P_{\theta\theta})/2$. This is the fluid-solid surface tension of the porous medium given that the surfaces are sufficiently far apart. This will be used to calculate the solid pressure, see equation \ref{eq:solid_pressure}.
\section{Results and discussion}
\subsection{Single solid particle surrounded by fluid}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{profile.png}
\caption{Local fluid number density (left) and normal and tangential pressure tensor components (right) in spherical shells as a function of the distance from the center of the spherical particle in the single-sphere simulation case. The fluid chemical potential is fixed to $\mu^f=1\zeta$ and the particle radius is $R=5\sigma$.}
\label{fig:profile}
\end{figure}
For the single solid sphere, we show the
local fluid number density and mechanical pressure tensor components of the surrounding fluids in figure \ref{fig:profile}. The fluid chemical potential is fixed at the minimum of the pair-wise interaction potential, $\zeta$, giving $\mu^f=1\zeta$. At this fluid chemical potential, the bulk density is $\rho^b = 0.8\sigma^{-3}$. The fluid particles pack in layers close to the surface of the solid particle of radius $R=5\sigma$, as reflected in the density variations in figure \ref{fig:profile}. The tangential component of the mechanical pressure tensor, shown in orange to the right in the figure, has the same variation. The normal and tangential components were used to calculate the surface tension, shown in figure \ref{fig:surface_tension} (see below). In this calculation, we do not equate the solid pressure to the mechanical pressure in any way. Inside the solid particle, the tangential component is zero. This is because the cut-off of the fluid-fluid interaction is relatively short, meaning that no fluid-fluid interaction contributes to the pressure inside the solid sphere. It is only the fluid-solid interactions that contribute to the pressure inside, and these contribute only to the normal component. The normal component is proportional to $r^{-2}$ as a consequence, see Fig.\ref{fig:profile}. This follows from mechanical equilibrium in spherical coordinates,
\begin{equation}
P_T(r) = P_N(r) + \frac{r}{2}\frac{\partial P_N}{\partial r}.
\end{equation}
We used mechanical equilibrium and found that the mechanical pressure calculation in spherical coordinates is consistent with this.
The surface tension was calculated for a single solid sphere surrounded by fluid particles, as a function of the fluid chemical potential. This is presented in figure \ref{fig:surface_tension}. It was calculated from the components of the mechanical pressure tensor, see equation \ref{eq:surface_tension}. Fluid particles have reduced contact with each other and with the solid in the lower chemical potential-regime ($\mu^f\in[-7,-3]\zeta$), which is characterized by low densities. The surface tension is accordingly very small. The surface tension between fluid and solid increases sharply at a chemical potential of around $\mu^f = -2.3\zeta$. The increase indicates increasing interactions between particles (fluid-fluid and fluid-solid). For a chemical potential below $\approx$ 2.3$\zeta$, the fluid is more vapor-like, while above this value, it is more liquid-like.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{surface_tension.png}
\caption{Surface tension for the single-sphere case $\gamma$ as a function of fluid chemical potential $\mu^f$.}
\label{fig:surface_tension}
\end{figure}
In figure \ref{fig:solid_pressure} the corresponding solid pressure is shown as a function of fluid chemical potential for a single-sphere surrounded by fluid particles. It is calculated from the surface tension and bulk pressure, see equation \ref{eq:solid_pressure}. The bulk pressure monotonically increases with the fluid chemical potential.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{solid_pressure.png}
\caption{The solid pressure, $p^s$, in a single sphere, calculated from equation \ref{eq:solid_pressure} as a function of the fluid chemical potential $\mu^f$ of the surrounding fluid.}
\label{fig:solid_pressure}
\end{figure}
So far, the variations are all expected from mechanical equilibrium and standard thermodynamics as given by the Young-Laplace equation. It is reasonable that the pressure of the solid increases monotonously, following the variation in the surface tension.
\subsection{The fluid number density and the replica energy density}
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{fluid_density.png}
\caption{Fluid number density $\rho^f = N^f/V$ as a function of fluid chemical potential $\mu^f$ at varying porosity $\phi=V^f/V$.}
\label{fig:fluid_density}
\end{figure}
Figure \ref{fig:fluid_density} presents the fluid number density $\rho^f=N^f/V$, where $V$ is the total REV volume, as a function of the controlled fluid chemical potential $\mu^f$ for various porosities $\phi$ of the fcc lattice. Results shown as black crosses represent bulk fluid or the limit where the porosity approaches unity. The fluid number density starts at approximately zero density for small fluid chemical potentials (a dilute gas-like phase) and converges to a density of a dense liquid-like phase. The fluid number density decreases with decreasing porosity, which is mainly because of decreased fluid volume compared to total volume. The fluid particles form layers on the solid surface, as was seen in figure \ref{fig:profile}.
The replica energy density is a characteristic property of the small system. It is shown for the first time for a regular fcc-lattice in figure \ref{fig:replica_energy_density}. The property was obtained as the integral over the fluid density, of the fluid chemical potential, see equation \ref{eq:replica_energy_density}. The replica energy density approaches the bulk fluid value (black crosses) in the thermodynamic limit of increasing porosity, as expected. We see a similar development in the replica energy density, as seen in the curve for the pressure of the solid around the single sphere. The replica energy is negative and increases with decreasing porosity. When compared with other thermodynamic properties like the internal energy (see below), the value is sizable. The system is small in the sense that it has non-negligible replica energy.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{replica_energy_density.png}
\caption{The negative replica energy density $-x=-X/V$ as a function of the fluid chemical potential $\mu^f$ at varying porosities $\phi=V^f/V$.}
\label{fig:replica_energy_density}
\end{figure}
The integral solid chemical potential was next calculated from the fluid number density of the REV and the bulk, see equation \ref{eq:solid_chemical_potential}. It is presented in figure \ref{fig:solid_chemical_potential} as a function of the fluid chemical potential. The integral solid chemical potential is much larger than the fluid chemical potential because each solid particle has a radius ten times larger than a fluid particle. This implies that each solid particle interacts with more particles than each fluid particle. The energy required to add one more solid particle is very large compared to adding one more fluid particle. Interestingly, the integral solid chemical potential is independent of the porosity, it is a function of fluid number density and fluid chemical potential. We interpret this to mean that the solid particle are sufficiently far away from each other, such that the integral solid chemical potential is unaffected by the porosity.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{solid_chemical_potential.png}
\caption{Integral solid chemical potential as a function of the fluid chemical potential $\mu^f$ at varying porosities $\phi=V^f/V$.}
\label{fig:solid_chemical_potential}
\end{figure}
Figure \ref{fig:pressure} compares the average of the diagonal components of the mechanical pressure for the bulk fluid (black crosses) and the fcc lattice of spheres filled with fluid particles. The mechanical pressure was calculated in Cartesian coordinates for the whole simulation box, as described by equation \ref{eq:mechanical_pressure}. The bulk value was now equated to the thermodynamic bulk pressure, which was assumed to be equal to the integral pressure. Note that we did not equate the trace of the mechanical pressure tensor in the heterogeneous porous medium to any thermodynamic property. Figure \ref{fig:pressure} serves purely for a comparison. However, even though the geometry inside the porous medium is more complex than what can be captured by the Cartesian mechanical pressure tensor, the values fall almost on the same line. This could suggest that the mean of the diagonal components of the Cartesian mechanical pressure tensor gives the integral pressure. An exception is seen for the data obtained with the smallest porosity, $\phi=0.38$, when fluid chemical potentials vary between approximately $\mu=-2\zeta$ and $\mu=-\zeta$. The given porosity is near the closest packing possible for spheres. While this is interesting, it may only hold for the present case, with a relatively simple structure.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{pressure.png}
\caption{Trace of the mechanical pressure divided by three as a function of the fluid chemical potential $\mu^f$ for varying porosities $\phi=V^f/V$.}
\label{fig:pressure}
\end{figure}
\subsubsection{Internal energy and entropy densities}
The internal energy and entropy densities as a function of the fluid chemical potential are presented in figure \ref{fig:internal_energy_density} and \ref{fig:entropy_density}, respectively. The internal energy and entropy are divided by the REV volume $V$ to give the respective densities. The variation in the internal energy and entropy densities as a function of the fluid chemical potential, is similar to that shown by fluids above the critical point, when they are described by cubic equations of state, for example, the van der Waals equations of state. It is therefore reasonable that the maxima of the internal energy and entropy densities are a consequence of a structural transition that takes place at the chemical potential in question. This will explain why all maxima are located at the same chemical potential (-2.3 $\zeta$). This location is close to the value of the chemical potential, where the fluid-solid surface tension starts to increase and also where the replica energy density begins to deviate stronger from its bulk value (see figures \ref{fig:surface_tension} and \ref{fig:replica_energy_density}). The maxima locations remain at the same chemical potential, as the porosity decreases and are given by the location in the bulk fluid. It may therefore be possible to estimate this position using the equation of state for the bulk fluid. Thus, one can determine for which fluid states it is particularly important to take the system size into account.
The absolute value of the internal energy density decreased with the porosity, reflecting the fact that the fluid number density decreased with the porosity. The entropy density became larger as we approached the bulk value. This observation is typical for phase transitions in small systems, it becomes less clear cut first order, and has less of a discontinuity in the phase variables \cite{Hill1964}.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{internal_energy_density.png}
\caption{Internal energy density $u=U/V$ as a function of the fluid chemical potential $\mu^f$ for varying porosities $\phi=V^f/V$.}
\label{fig:internal_energy_density}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{entropy_density.png}
\caption{The entropy density $s=S/V$ as a function of the fluid chemical potential $\mu^f$ for varying porosities $\phi=V^f/V$. The integral pressure is assumed to be equal to the bulk pressure.}
\label{fig:entropy_density}
\end{figure}
\subsection{Small system effects}
A system is small when we need to take into account the subdivision potential to accurately describe the system. The smallness of a system can also be measured by the replica energy (for the present system, see equation \ref{eq:Replica}). The replica energy depends in general on the set of control variables, and has accordingly an equivalent bulk variable \cite{Bedeaux2020}. It changes proportional to the subdivision potential. The value of the subdivision potential is included in the deviation of the replica energy from the corresponding value of a bulk fluid, see equation \ref{eq:Replica}. This deviation can be seen in figure \ref{fig:replica_energy_density}; it increases with decreasing porosity. While a change in replica energy is expected with additional contributions of the solid, the application of the replica energy as a variable allow us to quantify the smallness of the system, the smallness that originate from the subdivision potential. For evaluation of the state of smallness, the single contributions given in equation \ref{eq:Replica} must be computed independent of one another. We were not able to compute the subdivision potential here, but the replica energy density may still be used to assess whether small system effects can be neglected or not. This can be done by evaluating the total replica energy difference together with the entropy density. Consider for the purpose of such an evaluation the left-hand side of the peak in entropy density. Due to the small replica energy differences between the fluid bulk value and the corresponding values of the nanoporous medium for small chemical potentials, it may be speculated that the subdivision potential is negligible for vapor-like densities. On the right hand side of the peak, the difference is sizeable which is why the subdivision potential cannot be assumed to be negligible a priori. It must therefore be included in the thermodynamic analysis to ensure first-order Euler homogeneity in the total internal energy.
\subsection{The pressure of a REV in a porous medium}
We are now in a position to discuss, if not answer completely, the questions posed upfront; what is the pressure of a representative elementary volume (REV)?
In the present case, the structure of the porous system was regular fcc, and all different microstates are represented by a small REV, a unit cell for all practical purposes. A system with irregular structure has probably a larger REV. All thermodynamic properties defined, refer to the actual REV.
Equations \ref{eq:dp_hat},\ref{eq:loce} and \ref{eq:phat} most central in the description. These relations were derived using Hill's systematic procedure for porous media with nanoscale pores. The effective pressure $\hat{p}$ of the REV was given in equation \ref{eq:dp_hat}. We see that the expression is consistent with Young-Laplace equation. Not only a set of system variables is central in the theory; also the set of control variables need be specified. When this is done, we can arrive, in the present case, at the expression linking $\hat{p}$ to two unknown system variables.
In order to proceed, we have next taken the bold assumption stated in equation \ref{eq:loce}, that the integral pressure $\hat{p}$, which clearly differ from the differential pressure $p$, is constant across the medium at equilibrium. It is then not far from the next possible step; to construct the driving force for mass transport using $\hat{p}$.
The mean of the diagonal components of the Cartesian mechanical pressure tensor, presented in figure \ref{fig:pressure}, gives a good approximation of the integral pressure for the largest porosities. This can give an alternative route to obtain the integral pressure when the porosity is large. However, for small porosities it deviates from the bulk pressure and the integral pressure.
\section{Conclusion}
The thermodynamic method of Hill for nanoscale systems was used to describe the thermodynamic state of a single-phase fluid confined to a porous medium. The size and shape of the porous media were restricted, such that the description can be said to be general for any porous media.
The system was open for fluid to be exchanged with the environment, while the solid was not allowed to be exchanged with the environment. In addition, the temperature was controlled by the environment, and the fluid and solid volumes, surface area, and number of solid particles were controlled.
Nanothermodynamics introduces two new conjugate variables, the subdivision potential and the number of replicas. The subdivision potential was incorporated into the definition of the integral pressure and integral solid chemical potential. A fundamental assumption used in this work is that the integral pressure is constant everywhere in equilibrium. We have shown in previous works that this holds for simple porous media such as the slit pore \cite{Rauter2020,Galteland2021}.
We have used this framework to demonstrate how to compute the fluid number density, replica energy density, integral solid chemical potential, internal energy density, and entropy density of a fluid confined to a face-centered cubic lattice of solid particles. The radius of the solid particles was ten times larger than the fluid particles, and the porosity varied from $\phi=0.38$ to $\phi=0.87$. These porosities range from almost closest packing to rather open structures. We have used this system as a relatively simple model of porous media, computing all its thermodynamic properties, in particular its integral and differential pressures, liquid-solid surface tension, and solid pressure of a single solid particle surrounded by fluid particles.
This way of defining the small system thermodynamic properties, for a given set of control variables, may be useful for the study of transport in non-deformable porous media.
\section*{Conflict of Interest Statement}
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
\section*{Author Contributions}
O.G. contributed to formal analysis, investigation, methodology, software, and visualization. D.B. and S.K contributed to supervision. All authors contributed to conceptualization, writing original drafts, reviewing, and editing. All authors have read and agreed to the published version of the manuscript.
\section*{Funding}
This work was funded by the Research Council of Norway through its Centres of the Excellence funding scheme, project number 262644, PoreLab.
\section*{Acknowledgments}
The simulations were performed on resources provided by UNINETT Sigma2 - the National Infrastructure for High-Performance Computing and Data Storage in Norway. We thank the Research Council of Norway.
\bibliographystyle{ieeetr}
|
\section{Introduction}\label{intro}
Dimensionality reduction (DR) is of central importance when dealing with high-dimensional data \citep{donoho2000high}. It mitigates the curse of dimensionality, allowing for greater statistical flexibility and less computational complexity. DR also enables visualization that can be of great practical interest for understanding and interpreting the structure of large datasets.
Most seminal approaches include Principal Component Analysis (PCA) \cite{pearson1901liii}, multidimensional scaling \cite{kruskal1978multidimensional} and more broadly kernel eigenmaps methods such as Isomap \cite{balasubramanian2002isomap}, Laplacian eigenmaps \citep{belkin2003laplacian} and diffusion maps \citep{coifman2006diffusion}. These methods share the definition of a pairwise similarity kernel that assigns a high value to close neighbors and the resolution of a spectral problem. They are well understood and unified in the kernel PCA framework \citep{ham2004kernel}.
In the past decade, the field has witnessed a major shift with the emergence of a new class of methods. They are also based on pairwise similarities but these are not converted into inner products. Instead, they define pairwise similarity functions in both input and latent spaces and optimize a cost between the two. Among such methods, the Stochastic Neighbor Embedding (SNE) algorithm \cite{NIPS2002SNE}, its heavy-tailed symmetrized version t-SNE \cite{maaten2008tSNE} or more recent approaches like LargeVis \cite{tang2016visualizing} and UMAP \cite{mcinnes2018umap} are arguably the most used in practice. These will be referred to as 'SNE-like' or 'neighbor embedding' methods in what follows. They are increasingly popular and now considered as the state-of-art techniques in many fields \cite{li2017application,kobak2019art,anders2018dissecting}. Their popularity is mainly due to their exceptional ability to preserve local structure, \textit{i.e.}\ close points in the input space have close embeddings, as shown empirically \cite{wang2021understanding}. They also demonstrate impressive performances in identifying clusters \cite{arora2018analysis, linderman2019clustering}. However this is done at the expense of global structure, that these methods struggle in preserving \cite{wattenberg2016use, coenen2019understanding} \textit{i.e.}\ the relative large-scale distances between embedded points do not necessarily correspond to the original ones.
Due to a lack of clear probabilistic foundations, these properties remain mostly empirical. This gap between theory and practice is detrimental as practitioners may rely on strategies that are not optimal for their use case.
While recent software developments are making these methods more scalable \cite{chan2018t,pezzotti2019gpgpu,linderman2019fast} and further expanding their use, the need for a well-established probabilistic framework is becoming more prominent.
In this work we define the generative probabilistic model that encompasses current embedding methods, while establishing new links with the well-established PCA model.
\paragraph{Outline.}
Consider $\bm{X} = (\bm{X}_1,..., \bm{X}_n)^\top \in \mathbb{R}^{n \times p}$, an input dataset that consists of $n$ vectors of dimension $p$. Our task is to embed $\bm{X}$ in a lower dimensional space of dimension $q<p$ (typically $q=2$ for visualization), and we denote by $\bm{Z} = (\bm{Z}_1, ..., \bm{Z}_n)^\top \in \mathbb{R}^{n \times q}$ the unknown embeddings. The rationale of our framework is to suppose that the observations $\bm{X}$ and $\bm{Z}$ are structured by two latent graphs with $\bm{W}_{\scaleto{X}{4pt}}$ and $\bm{W}_{\scaleto{Z}{4pt}}$ standing for their $n$-square weight matrices.
As the goal of DR is to preserve the input's structure in the latent space, we propose to find the best low-dimensional representation $\bm{Z}$ of $\bm{X}$ such that $\bm{W}_{\scaleto{X}{4pt}}$ and $\bm{W}_{\scaleto{Z}{4pt}}$ are close. To build a flexible and robust probabilistic framework, we consider random graphs distributed according to some predefined prior distributions. Our objective is to match the posterior distributions of $\bm{W}_{\scaleto{X}{4pt}}$ and $\bm{W}_{\scaleto{Z}{4pt}}$. Note that as they share the same dimensionality the latter graphs can be easily compared unlike $\bm{X}$ and $\bm{Z}$. The coupling is done with a cross entropy criterion, the minimization of which will be referred to as graph coupling.
In this work, our main contributions are as follows.
\begin{itemize}
\item We show that SNE, t-SNE, LargeVis and UMAP are all instances of graph coupling and characterized by different choices of prior for discrete latent structuring graphs (\cref{sec:GC_unified}). We demonstrate that such graphs essentially capture conditional independencies among rows through a pairwise Markov Random Field (MRF) model which construction can be found in \cref{sec:graph_structure}.
\item We uncover the intrinsic probabilistic property explaining why such methods perform poorly on conserving the large scale structure of the data as a consequence of a degeneracy of the MRF when shift invariant kernels are used (\cref{prop:integrability_pairwise_MRF}). Such degeneracy induces the loss of the relative positions of clusters corresponding to the connected components of the posterior latent graphs which distributions are identified (\cref{prop:posterior_W}). These findings are highlighted by a new initialization of the embeddings that is empirically tested (\cref{sec:towards_large_scale}).
\item We show that for Gaussian MRFs, when adapting graph coupling to precision matrices with suitable priors, PCA appears as a natural extension of the coupling problem in its continuous version (\cref{PCA_graph_coupling}). Such model does not suffer from the aforementioned degeneracy hence preserves the large-scale structure.
\end{itemize}
\section {Shift-Invariant Pairwise MRF to Model Row Dependencies} \label{sec:graph_structure}
We start by defining the distribution of the observations given a graph. The latter takes the form of a pairwise MRF model which as we show is improper (\textit{i.e.}\ not integrable on $\mathbb{R}^{n \times p}$) when shift-invariant kernels are used. We consider a fixed directed graph $\bm{W} \in \mathcal{S}_{\scaleto{W}{4pt}}$ where:
$$\mathcal{S}_{\scaleto{W}{4pt}} = \left\{\bm{W} \in \mathbb{N}^{n \times n} \mid \forall (i,j) \in [n]^2, W_{ii}=0, W_{ij} \leq n \right\}$$
Throughout, $(E, \mathcal{B}(E), \lambda_E)$ denotes a measure space where $\mathcal{B}(E)$ is the Borel $\sigma$-algebra on $E$ and $\lambda_E$ is the Lebesgue measure on $E$.
\subsection{Graph Laplacian Null Space}\label{sec:laplacian_prop}
A central element in our construction is the graph Laplacian linear map, defined as follows, where $\mathcal{S}^n_+(\mathbb{R})$ is the set of positive semidefinite matrices.
\begin{definition}\label{graph_laplacian}
The graph Laplacian operator is the map $L \colon \mathbb{R}_+^{n \times n} \to \mathcal{S}^n_+(\mathbb{R})$ such that
$$\text{for } (i,j) \in [n]^2, \quad L(\bm{W})_{ij} = \left\{
\begin{array}{ll}
- W_{ij} & \text{if } i \neq j \\
\sum_{k \in [n]} W_{ik} & \text{otherwise} \:.
\end{array}
\right. $$
\end{definition}
With an abuse of notation, let $\bm{L} = L(\overline{\bm{W}})$ where $\overline{\bm{W}} = \bm{W} + \bm{W}^\top$. Let $(C_1,...,C_{\scaleto{R}{4pt}})$ be a partition of $[n]$ (\textit{i.e.}\ the set $\{1,2,...,n\}$) corresponding to the connected components (CCs) of $\overline{\bm{W}}$. As well known in spectral graph theory \cite{Chung97}, the null space of $\bm{L}$ is spanned by the orthonormal vectors $\{\bm{U}_{r}\}_{r \in [R]}$ such that for $r \in [R]$,
$\bm{U}_{r} = \left(n_r^{-1/2} 1\!\!1_{i \in C_r}\right)_{i \in [n]}$ with $n_r = \operatorname{Card}(C_r)$. By the spectral theorem, $\bm{U}_{\scaleto{[R]}{5pt}}$ can be completed such that $\bm{L} = \bm{U \Lambda U^\top}$ where $\bm{U} = (\bm{U}_1, ..., \bm{U}_n)$ is orthogonal and $\bm{\Lambda} = \operatorname{diag}((\lambda_i)_{i \in [n]})$ with $0 = \lambda_1 = ... = \lambda_R < \lambda_{R+1} \leq ... \leq \lambda_n$. In the following, the data is split into two parts: $\bm{X}_{\scaleto{M}{4pt}}$, the orthogonal projection of $\bm{X}$ on $\mathcal{S}_{\scaleto{M}{4pt}} = (\ker \bm{L}) \otimes \mathbb{R}^p$, and $\bm{X}_{\scaleto{C}{4pt}}$, the projection on $\mathcal{S}_{\scaleto{C}{4pt}} = (\ker \bm{L})^{\perp} \otimes \mathbb{R}^p$. For $i \in [n]$, $\bm{X}_{\scaleto{M}{4pt},i} = \sum_{r \in [R]} n_r^{-1} 1\!\!1_{i \in C_r}\sum_{\ell \in C_r} \bm{X}_{\ell} $ hence $\bm{X}_{\scaleto{M}{4pt}}$ stands for the empirical means of $\bm{X}$ on CCs, thus modelling the CC positions, while $\bm{X}_{\scaleto{C}{4pt}} = \bm{X} - \bm{X}_{\scaleto{M}{4pt}}$ is CC-wise centered, thus modeling the relative positions of the nodes within CCs. We now introduce the probability distribution of these variables.
\subsection{Pairwise MRF and Shift-Invariances}\label{sec:within_CC}
In this work, the dependency structure among rows of the data is governed by a graph. The strength of the connection between two nodes is given by a symmetric function $k : \mathbb{R}^p \to \mathbb{R}_+$. We consider the following pairwise MRF unnormalized density function:
\begin{align}\label{eq:unnormalized_MRF}
f_{k} \colon (\bm{X},\bm{W}) &\mapsto \prod_{(i,j) \in [n]^2} k(\bm{X}_{i} - \bm{X}_{j})^{W_{ij}} \: .
\end{align}
As we will see shortly, the above is at the heart of DR methods based on pairwise similarities. Note that as $k$ measures the similarity between couples of samples, $f_k$ will take high values if the rows of $\bm{X}$ vary smoothly on the graph $\bm{W}$. Thus we can expect $\bm{X}_i$ and $\bm{X}_j$ to be close if there is an edge between node $i$ and node $j$ in $\bm{W}$. A key remark is that $f_{k}$ is kept invariant by translating $\bm{X}_{\scaleto{M}{4pt}}$. Namely for all $\bm{X} \in \mathbb{R}^{n \times p}$, $f_{k}(\bm{X}, \bm{W}) = f_{k}(\bm{X}_{\scaleto{C}{4pt}}, \bm{W})$. This invariance results in $f_{k}(\cdot, \bm{W})$ being non integrable on $\mathbb{R}^{n \times p}$, as we see with the following example.
\paragraph{Gaussian kernel.} For a positive definite matrix $\bm{\Sigma} \in \mathcal{S}^n_{++}(\mathbb{R})$, consider the Gaussian kernel $k : \bm{x} \mapsto e^{- \frac{1}{2}\| \bm{x} \|_{\bm{\Sigma}}^2}$ where $\bm{\Sigma}$ stands for the covariance among columns. One has:
\begin{align}\label{eq:gaussian_kernel}
\log f_{k}(\bm{X}, \bm{W}) &= -\sum_{(i,j) \in [n]^2} W_{ij} \| \bm{X}_{i}-\bm{X}_{j} \|^2_{\bm{\Sigma}}
= - \operatorname{tr} \left(\bm{\Sigma}^{-1} \bm{X}^{T} \bm{L} \bm{X}\right)
\end{align}
by property of the graph Laplacian (\cref{graph_laplacian}). In this case, it is clear that due to the rank deficiency of $\bm{L}$, $f_{k}(\cdot, \bm{W})$ is only $\lambda_{\mathcal{S}_{\scaleto{C}{3pt}}}$-integrable. In general DR settings one does not want to rely on Gaussian kernels only. A striking example is the use of the Student kernel in t-SNE \cite{maaten2008tSNE}. Heavy-tailed kernels appear useful when the dimension of the embeddings is smaller than the intrinsic dimension of the data \cite{kobak2019heavy}. Our contribution provides flexibility by extending the previous result to a large class of kernels, as stated in the following theorem.
\begin{theorem}\label{prop:integrability_pairwise_MRF}
If $k$ is $\lambda_{\mathbb{R}^p}$-integrable and bounded above $\lambda_{\mathbb{R}^p}$-almost everywhere then $f_{k}(\cdot, \bm{W})$ is $\lambda_{\mathcal{S}_{\scaleto{C}{3pt}}}$-integrable.
\end{theorem}
We can now define a distribution on $(\mathcal{S}_{\scaleto{C}{4pt}}, \mathcal{B}(\mathcal{S}_{\scaleto{C}{4pt}}))$, where $\mathcal{C}_{k}(\bm{W}) = \int f_{k}(\cdot, \bm{W}) d\lambda_{\mathcal{S}_{\scaleto{C}{3pt}}}$:
\begin{align}\label{eq:proba_perp}
\mathbb{P}_{k}(d\bm{X}_{\scaleto{C}{4pt}} | \bm{W}) = \mathcal{C}_{k}(\bm{W})^{-1} f_{k}(\bm{X}_{\scaleto{C}{4pt}}, \bm{W}) \lambda_{\mathcal{S}_{\scaleto{C}{3pt}}}(d\bm{X}_{\scaleto{C}{4pt}}) \: .
\end{align}
\begin{remark}
Kernels may have node-specific bandwidths $\bm{\tau}$, set during a pre-processing step, giving $f_{k}(\bm{X},\bm{W}) = \prod_{(i,j)} k((\bm{X}_{i} - \bm{X}_{j})/\tau_{i})^{W_{ij}}$. Note that such bandwidth does not affect the degeneracy of the distribution and \cref{prop:integrability_pairwise_MRF} still holds.
\end{remark}
\paragraph{Between-Rows Dependency Structure.} By symmetry of $k$, reindexing gives: $f_{k}(\bm{X}, \bm{W}) = \prod_{j \in [n]} \prod_{i \in [j]} k(\bm{X}_{i} - \bm{X}_{j})^{\overline{W}_{ij}}$. Hence distribution \eqref{eq:proba_perp} boils down to a pairwise MRF model \citep{clifford1990markov} with respect to the undirected graph $\overline{\bm{W}}$, $\mathcal{C}_{k}$ playing the role of the partition function. Note that since $f_k$ (Equation \ref{eq:unnormalized_MRF}) trivially factorize according to the cliques of $\overline{\bm{W}}$, the Hammersley-Clifford theorem ensures that the rows of $\bm{X}_{\scaleto{C}{4pt}}$ satisfy the local and global Markov properties with respect to $\overline{\bm{W}}$.
\subsection{Uninformative Model for CC-wise Means}
We showed that the MRF (\ref{eq:unnormalized_MRF}) is only integrable on $\mathcal{S}_{\scaleto{C}{4pt}}$, the definition of which depends on the connectivity structure of $\bm{W}$. As we now demonstrate, the latter MRF can be seen as a limit of proper distributions on $\mathbb{R}^{n \times p}$, see \textit{e.g.}\ \cite{rue2005gaussian} for a similar construction in the Gaussian case.
We introduce the Borel function $f^{\varepsilon}(\cdot, \bm{W}) \colon \mathbb{R}^{n \times p} \to \mathbb{R}_+$ for $\varepsilon > 0$ such that for all $\bm{X} \in \mathbb{R}^{n \times p}$, $f^{\varepsilon}(\bm{X}, \bm{W}) = f^{\varepsilon}(\bm{X}_{\scaleto{M}{4pt}}, \bm{W})$. To allow $f^{\varepsilon}$ to become arbitrarily non-informative, we assume that for all $\bm{W} \in \mathcal{S}_{\scaleto{W}{4pt}}$, $f^\varepsilon(\cdot, \bm{W})$ is $\lambda_{\scaleto{\mathcal{S}_{\scaleto{M}{3pt}}}{6pt}}$-integrable for all $\varepsilon \in \mathbb{R}^*_+$ and $f^{\varepsilon}(\cdot, \bm{W}) \xrightarrow[\varepsilon \to 0]{} 1$ almost everywhere.
We now define the conditional distribution on $(\mathcal{S}_{\scaleto{M}{4pt}}, \mathcal{B}(\mathcal{S}_{\scaleto{M}{4pt}}))$ as follows:
\begin{align}\label{eq:proba_parallel}
\mathbb{P}^{\varepsilon}(d\bm{X}_{\scaleto{M}{4pt}}| \bm{W}) = \mathcal{C}^{\varepsilon}(\bm{W})^{-1} f^{\varepsilon}(\bm{X}_{\scaleto{M}{4pt}}, \bm{W}) \lambda_{\mathcal{S}_{\scaleto{M}{3pt}}}(d\bm{X}_{\scaleto{M}{4pt}})
\end{align}
where $\mathcal{C}^{\varepsilon}(\bm{W}) = \int f^{\varepsilon}(\cdot, \bm{W}) d\lambda_{\mathcal{S}_{\scaleto{M}{3pt}}}$.
With this at hand, the joint conditional is defined as the product measure of (\ref{eq:proba_perp}) and (\ref{eq:proba_parallel}) over the row axis, the integrability of which is ensured by the Fubini-Tonelli theorem. In the following we will use the compact notation $\mathcal{C}^{\varepsilon}_k(\bm{W}) = \mathcal{C}_k(\bm{W})\mathcal{C}^{\varepsilon}(\bm{W})$ for the joint normalizing constant.
\begin{remark}
At the limit $\varepsilon \to 0$ the above construction amounts to setting an infinite variance on the distribution of the empirical means of $\bm{X}$ on CCs, thus loosing the inter-CC structure.
\end{remark}
As an illustration, one can structure the CCs' relative positions according to a Gaussian model with positive definite precision $\varepsilon \bm{\Theta} \in \mathcal{S}_{++}^R(\mathbb{R})$, as it amounts to choosing $f^{\varepsilon} : \bm{X} \to \exp \left(-\frac{\varepsilon}{2} \operatorname{tr}\left(\bm{\Sigma}^{-1}\bm{X}^\top\bm{U}_{\scaleto{[:R]}{5pt}} \bm{\Theta}\bm{U}^\top_{\scaleto{[R]}{5pt}} \bm{X}\right)\right)$ such that: $\bm{X}_{\scaleto{M}{4pt}} | \bm{\Theta} \sim \mathcal{N}\left(\bm{0}, \left(\varepsilon \bm{U}_{\scaleto{[:R]}{5pt}} \bm{\Theta}\bm{U}^\top_{\scaleto{[R]}{5pt}}\right)^{-1} \otimes \bm{\Sigma}\right)$.
\section{Graph Coupling as a Unified Objective for Pairwise Similarity Methods}\label{sec:GC_unified}
In this section, we show that neighbor embedding methods can be recovered in the presented framework. They are obtained, for particular choices of graph priors, at the limit $\varepsilon \to 0$ when $f^{\varepsilon}$ becomes non informative and the CCs' relative positions are lost.
We now turn to the priors for $\bm{W}$. Our methodology is similar to that of constructing conjugate priors for distributions in the exponential family \cite{wainwright2008graphical}, notably we insert the cumulant function $\mathcal{C}_k^{\varepsilon}$ (\textit{i.e.}\ normalizing constant of the conditional) as a multivariate term of the prior.
We consider different forms: binary ($B$), unitary out-degree ($D$) and $n$-edges ($E$), relying on an additional term ($\Omega$) to constraint the topology of the graph. For a matrix $\bm{A}$, $A_{i+}$ denotes $\sum_j A_{ij}$ and $A_{++}$ denotes $\sum_{ij} A_{ij}$. In the following, $\bm{\pi}$ plays the role of the edges prior. The latter can be leveraged to incorporate some additional information about the dependency structure, for instance when a network is observed \cite{li2020high}.
\begin{definition}\label{def:prior_W}
Let $\bm{\pi} \in \mathbb{R}_+^{n \times n}$, $\varepsilon \in \mathbb{R}_+$, $\alpha \in \mathbb{R}$, $k$ satisfies the assumptions of \cref{prop:integrability_pairwise_MRF} and $\mathcal{P} \in \{B,D,E\}$. For $\bm{W} \in \mathcal{S}_{\scaleto{W}{4pt}}$ we introduce:
$$\mathbb{P}_{\scaleto{\mathcal{P},k}{5pt}}^{\varepsilon}(\bm{W}; \bm{\pi}, \alpha) \propto \mathcal{C}^{\varepsilon}_k(\bm{W})^{\alpha} \: \Omega_{\scaleto{\mathcal{P}}{4pt}}(\bm{W}) \prod_{(i,j) \in [n]^2} \pi_{ij}^{W_{ij}}$$
where $\Omega_{\scaleto{B}{4pt}}(\bm{W}) = \prod_{ij} 1\!\!1_{W_{ij} \leq 1}$, $\Omega_{\scaleto{D}{4pt}}(\bm{W}) = \prod_{i} 1\!\!1_{W_{i+} = 1}$ and $\Omega_{\scaleto{E}{4pt}}(\bm{W}) = 1\!\!1_{W_{++} = n}\prod_{ij}(W_{ij}!)^{-1}$.
\end{definition}
When $\alpha = 0$, the above no longer depends on $\varepsilon$ and $k$. We will use the compact notation $\mathbb{P}_{\scaleto{\mathcal{P}}{4pt}}(\bm{W} ; \bm{\pi}) = \mathbb{P}_{\scaleto{\mathcal{P},k}{5pt}}^{\varepsilon}(\bm{W}; \bm{\pi}, 0)$. Note that by $\bm{W} \sim \mathbb{P}_{\scaleto{\mathcal{P}}{4pt}}(\cdot \: ; \bm{\pi})$ we have the following simple Bernoulli $(\mathcal{B})$ and multinomial $(\mathcal{M})$ distributions, where matrix division is element-wise:
\begin{itemize}
\item if $\mathcal{P} = B$, $\forall (i,j) \in [n]^2, \: W_{ij} \stackrel{\perp\!\!\!\!\perp}{\sim} \mathcal{B}\left(\pi_{ij}/(1 + \pi_{ij}) \right)$.
\item if $\mathcal{P} = D$, $\forall i \in [n], \: \bm{W}_{i} \stackrel{\perp\!\!\!\!\perp}{\sim} \mathcal{M}\left(1, \bm{\pi}_{i}/\pi_{i+} \right)$.
\item if $\mathcal{P} = E$, $\bm{W} \sim \mathcal{M}\left(n, \bm{\pi}/\pi_{++} \right)$.
\end{itemize}
We now show that the posterior distribution of the graph given the observations takes a simple form when the distribution of CC empirical means $\bm{X}_{\scaleto{M}{4pt}}$ diffuses \textit{i.e.}\ when $\varepsilon \to 0$.
\begin{proposition}\label{prop:posterior_W}
Let $\bm{\pi} \in \mathbb{R}_+^{n \times n}$, $k$ satisfies the assumptions of \cref{prop:integrability_pairwise_MRF} with $\bm{K}_{\scaleto{X}{4pt}} = (k(\bm{X}_{i} - \bm{X}_{j}))_{(i,j) \in [n]^2}$ and $\mathcal{P}\in \{B, D, E\}$. If $\bm{W}^{\varepsilon} \sim \mathbb{P}_{\scaleto{\mathcal{P},k}{5pt}}^{\varepsilon}(\cdot \: ; \bm{\pi},1)$ then
$$\bm{W}^{\varepsilon} | \bm{X} \xrightarrow[\varepsilon \to 0]{\mathcal{D}} \mathbb{P}_{\scaleto{\mathcal{P}}{4pt}}(\cdot \: ;\bm{\pi} \odot \bm{K}_{\scaleto{X}{4pt}}) \:.$$
\end{proposition}
\begin{remark}
For all $\bm{W} \in \mathcal{S}_{\scaleto{W}{4pt}}$, $\mathcal{C}^{\varepsilon}(\bm{W})$ diverges as $\varepsilon \to 0$, hence the graph prior (\cref{def:prior_W}) is improper at the limit. This compensates for the uninformative diffuse conditional and allows to retrieve a well-defined tractable posterior limit.
\end{remark}
\subsection{Retrieving Well Known DR Methods}\label{sec:retrieving_DR_methods}
We now provide a unified view of neighbor embedding objectives as a coupling between graph posterior distributions. To that extent we derive the cross entropy associated with the various graph priors at hand. In what follows, $k_x$ and $k_z$ satisfy the assumptions of \cref{prop:integrability_pairwise_MRF} and we denote by $\bm{K}_{\scaleto{X}{4pt}}$ and $\bm{K}_{\scaleto{Z}{4pt}}$ the associated kernel matrices on $\bm{X}$ and $\bm{Z}$ respectively. For both graph priors we consider the parameters $\bm{\pi}=\bm{1}$ and $\alpha=1$. For $(\mathcal{P}_{\scaleto{X}{4pt}}, \mathcal{P}_{\scaleto{Z}{4pt}}) \in \{B,D,E\}^2$, we introduce the
cross entropy between the limit posteriors at $\varepsilon \to 0$,
\begin{align*}
\mathcal{H}_{\scaleto{\mathcal{P}_X}{4pt}, \scaleto{\mathcal{P}_Z}{4pt}} = - \mathbb{E}_{\bm{W}_{\scaleto{X}{3pt}} \sim \mathbb{P}_{\scaleto{\mathcal{P}_{\scaleto{X}{2pt}}}{3pt}}(\cdot;\bm{K}_{\scaleto{X}{3pt}})}[\log \mathbb{P}_{\scaleto{\mathcal{P}_{\scaleto{Z}{3pt}}}{4pt}}(\bm{W}_{\scaleto{Z}{4pt}} = \bm{W}_{\scaleto{X}{4pt}}; \bm{K}_{\scaleto{Z}{4pt}})]
\end{align*}
defining a coupling criterion to be optimized with respect to embedding coordinates $\bm{Z}$. We now go through each couple $(\mathcal{P}_{\scaleto{X}{4pt}}, \mathcal{P}_{\scaleto{Z}{4pt}})$ such that $\operatorname{supp}\left(\mathbb{P}_{\scaleto{\mathcal{P}_X}{4pt}}\right) \subset \operatorname{supp}\left(\mathbb{P}_{\scaleto{\mathcal{P}_Z}{4pt}}\right)$ for the cross-entropy to be defined.
\paragraph{SNE.}
When $\mathcal{P}_{\scaleto{X}{4pt}} = \mathcal{P}_{\scaleto{Z}{4pt}} = D$, the probability of the limit posterior graphs factorizes over the nodes and the cross-entropy between limit posteriors takes the form of the objective of SNE \cite{hinton2002stochastic}, where for $i \in [n], \bm{P}^{D}_{i} = \bm{K}_{\scaleto{X}{4pt},i} / K_{\scaleto{X}{4pt},i+}$ and $\bm{Q}^{D}_{i} = \bm{K}_{\scaleto{Z}{4pt},i} / K_{\scaleto{Z}{4pt},i+}$,
$$\mathcal{H}_{D,D}= - \sum_{i \neq j} P^{D}_{ij} \log Q^{D}_{ij} \:.$$
\paragraph{Symmetric-SNE.}
Choosing $\mathcal{P}_{\scaleto{X}{4pt}} = D$ and $\mathcal{P}_{\scaleto{Z}{4pt}} = E$, we define for $(i,j) \in [n]^2$, $\bm{Q}^{E}_{ij} = K_{\scaleto{Z}{4pt},ij} / K_{\scaleto{Z}{4pt},++}$ and $\overline{P}^{D}_{ij} = P^{D}_{ij} + P^{D}_{ji}$. The symmetry of $\bm{Q}^{E}$ yields:
\begin{align*}
\mathcal{H}_{D,E} &= - \sum_{i \neq j} P^{D}_{ij} \log Q^{E}_{ij} = - \sum_{i < j} \overline{P}^{D}_{ij} \log Q^{E}_{ij}
\end{align*}
and the symmetrized objective of t-SNE \cite{maaten2008tSNE} is recovered.
\paragraph{LargeVis.}
Now choosing $\mathcal{P}_{\scaleto{X}{4pt}} = D$ and $\mathcal{P}_{\scaleto{Z}{4pt}} = B$, one can also notice that $\bm{Q}^{B} = \left( K_{\scaleto{Z}{4pt},ij} / (1+K_{\scaleto{Z}{4pt},ij}) \right)_{(i,j) \in [n]^2}$ is symmetric. With this at hand the limit cross-entropy reads
\begin{align*}
\mathcal{H}_{D,B} = - \sum_{i \neq j} P^{D}_{ij} \log Q^{B}_{ij} + \left(1 - P^{D}_{ij} \right) \log\left(1-Q^{B}_{ij} \right)
= - \sum_{i < j} \overline{P}^{D}_{ij} \log Q^{B}_{ij} + \left(2-\overline{P}^{D}_{ij}\right) \log (1- Q^{B}_{ij})
\end{align*}
which is the objective of LargeVis \cite{tang2016visualizing}.
\paragraph{UMAP.}
Let us take $\mathcal{P}_{\scaleto{X}{4pt}} = \mathcal{P}_{\scaleto{Z}{4pt}} = B$ and consider the symmetric thresholded graph $\widetilde{\bm{W}}_{\scaleto{X}{4pt}} = 1\!\!1_{\bm{W}_{\scaleto{X}{4pt}} + \bm{W}_X^\top \geq 1}$. By independence of the edges, $\widetilde{W}_{\scaleto{X}{4pt},ij} \sim \mathcal{B}\left(\widetilde{P}^{B}_{ij} \right) \quad \text{where} \quad \widetilde{P}^{B}_{ij} = P^{B}_{ij} + P^{B}_{ji} - P^{B}_{ij} P^{B}_{ji}$ and $\bm{P}^{B} = \left( K_{\scaleto{X}{4pt},ij} / (1+K_{\scaleto{X}{4pt},ij}) \right)_{(i,j) \in [n]^2}$. Coupling $\widetilde{\bm{W}}_{\scaleto{X}{4pt}}$ and $\bm{W}_{\scaleto{Z}{4pt}}$ gives:
\begin{align*}
\mathcal{H}_{\widetilde{B},B} &= -2 \sum_{i<j} \widetilde{P}_{ij}^{B} \log Q_{ij}^{B} + \left(1 - \widetilde{P}_{ij}^{B} \right) \log \left( 1 - Q_{ij}^{B} \right)
\end{align*}
which is the loss function considered in UMAP \cite{mcinnes2018umap}, the construction of $\widetilde{\bm{W}}_{\scaleto{X}{4pt}}$ being borrowed from section 3.1 of the paper.
\begin{remark}
One can also consider $\mathcal{H}_{E,E}$ but as detailed in \cite{maaten2008tSNE}, this criterion fails at positioning outliers and is therefore not considered.
Interestingly, any other feasible combination of the presented priors relates to an existing method.
\end{remark}
\subsection{Interpretations}\label{sec:interpretations}
\begin{table}[]
\caption{Prior distributions for $\bm{W}_{\scaleto{X}{4pt}}$ and $\bm{W}_{\scaleto{Z}{4pt}}$ associated with the pairwise similarity coupling DR algorithms. Grey-colored boxes are such that the cross-entropy is undefined.}
\begin{center}
\begin{small}
\begin{sc}
\centering
\renewcommand{\arraystretch}{2}
\begin{NiceTabular}{|W{c}{1cm}|W{c}{1.8cm}|W{c}{1.8cm}|W{c}{1.8cm}|}
\hline
\diagbox{{\fontsize{12}{15}\selectfont $\mathcal{P}_{\scaleto{X}{4pt}}$}}{{\fontsize{12}{15}\selectfont $\mathcal{P}_{\scaleto{Z}{4pt}}$}} & $B$ & $D$ & $E$ \\
\hline
$\widetilde{B}$ & UMAP & \cellcolor{black!10} & \cellcolor{black!10} \\
\hline
$D$ & LargeVis & SNE & T-SNE\\
\hline
\end{NiceTabular}
\label{tableau_priors}
\end{sc}
\end{small}
\end{center}
\label{priors_methods}
\end{table}
As we have seen in \cref{sec:retrieving_DR_methods}, SNE-like methods can all be derived from the graph coupling framework. What characterizes each of them is the choice of priors considered for the latent structuring graphs. To the best of our knowledge, the presented framework is the first that manages to unify all these DR algorithms. Such a framework opens many perspectives for improving upon current practices as we discuss in \cref{sec:towards_large_scale} and \cref{Perspectives}.
We now focus on a few insights that our work provides about the empirical performances of these methods.
\paragraph{Repulsion \& Attraction.}
Decomposing $\mathcal{H}_{\scaleto{\mathcal{P}_X}{4pt}, \scaleto{\mathcal{P}_Z}{4pt}}$ with Bayes' rule and simplifying constant terms one has the following optimization problem:
\begin{align}\label{eq:optim_H_Z}
\min_{\bm{Z} \in \mathbb{R}^{n \times q}} -\hspace{-0.2cm}\sum_{(i,j) \in [n]^2} \bm{P}^{\scaleto{\mathcal{P}_X}{4pt}}_{ij}\log k_z(\bm{Z}_i - \bm{Z}_j) + \log \mathbb{P}(\bm{Z}).
\end{align}
The first and second terms in \cref{eq:optim_H_Z} respectively summarize the attractive and repulsive forces of the objective. Recall from \cref{prop:posterior_W}
that $\bm{P}^{\scaleto{\mathcal{P}_X}{4pt}}$ is the posterior expectation of $\bm{W}_{\scaleto{X}{4pt}}$. Hence in SNE-like methods, the attractive forces resume to a pairwise MRF log likelihood with respect to a graph posterior expectation given $\bm{X}$. For instance if $k_z$ is the Gaussian kernel, this attractive term reads $\operatorname{tr} \left(\bm{Z}^\top \bm{L}^\star \bm{Z} \right)$ where $\bm{L}^\star = \mathbb{E}_{\bm{W} \sim \mathbb{P}_{\scaleto{\mathcal{P}_X}{4pt}}(\cdot;\bm{K}_{\scaleto{X}{4pt}})}[L(\bm{W})]$, boiling down to the objective of Laplacian eigenmaps \cite{belkin2003laplacian}. Therefore, for Gaussian MRFs, the attractive forces resume to an unconstrained Laplacian eigenmaps objective. Such link, already noted in \cite{carreira2010elastic}, is easily unveiled in our framework. Moreover, one can notice that only this attractive term depends on $\bm{X}$ as the repulsion is given by the marginal term in (\ref{eq:optim_H_Z}). The latter reads $\mathbb{P}(\bm{Z}) = \sum_{\bm{W} \in \mathcal{S}_{\scaleto{W}{4pt}}} \mathbb{P}(\bm{Z}, \bm{W})$ with $\mathbb{P}(\bm{Z}, \bm{W}) \propto f_k(\bm{Z}, \bm{W})\Omega_{\scaleto{\mathcal{P}_Z}{4pt}}(\bm{W})$. Such penalty notably prevents a trivial solution, as $\bm{0}$, like any constant vector, is a mode of $f_k(\cdot, \bm{W})$ for all $\bm{W}$. Also note that the prior for $\bm{W}_{\scaleto{X}{4pt}}$ only conditions attraction while the prior for $\bm{W}_{\scaleto{Z}{4pt}}$ only affects repulsion. In the present work we focus solely on deciphering the probabilistic model that accounts for neighbor embedding loss functions and refer to \cite{bohm2020unifying} for a quantitative study of attraction and repulsion in these methods.
\paragraph{Global Structure Preservation. }
To gain intuition, consider that $\bm{W}_{\scaleto{X}{4pt}}$ is observed. As we showed in \cref{sec:within_CC}, when one relies on shift invariant kernels, the positions of the CC means are taken from a diffuse distribution. Since the above methods are all derived from the limit posteriors at $\varepsilon \to 0$, $\bm{X}_{\scaleto{M}{4pt}}$ and $\bm{Z}_{\scaleto{M}{4pt}}$ have no influence on the coupling objective. Hence if two nodes belong to different CCs, their low dimensional pairwise distance will likely not be faithful. We can expect this phenomenon to persist when the expectation on $\bm{W}_{\scaleto{X}{4pt}}$ is considered, especially when clusters are well distinguishable in $\bm{X}$. This observation is central to understand the large scale deficiency of these methods. Note that this happens at the benefit of the local structure which is faithfully represented in low dimension, as discussed in \cref{intro}. In the following section we propose to mitigate the global structure deficiency with non-degenerate MRF models.
\section{Towards Capturing Large-Scale Dependencies}\label{sec:towards_large_scale}
In this section, we investigate the ability of graph coupling to faithfully represent global structure in low dimension. To gain intuition on the case where the distribution induced by the graph is not degenerate, we consider a proper Gaussian graph coupling model and show its equivalence with PCA. We then provide a new initialization procedure to alleviate the large scale deficiency of graph coupling when degenerate MRFs are used.
\subsection{PCA as Graph Coupling}
As we argue that the inability of SNE-like methods to reproduce the coarse-grain dependencies of the input in the latent space is due to the degeneracy of the conditional (\ref{eq:proba_perp}), a natural solution would be to consider graphical models that are well defined and integrable on the entire definition spaces of $\bm{X}$ and $\bm{Z}$. For simplicity, we consider the Gaussian model and leave the extension to other kernels for future works. Note that in this case integrability translates into the precision matrix being full-rank. As we see with the following, the natural extension of our framework to such models leads to a well-established PCA algorithm. In the following, for a continuous variable $\bm{\Theta}_{\scaleto{Z}{4pt}}$, $\mathbb{P}(\bm{\Theta}_{\scaleto{Z}{4pt}} = \cdot)$ denotes its density.
\begin{theorem}\label{PCA_graph_coupling}
Let $\nu \geq n$, $\bm{\Theta}_{\scaleto{X}{4pt}} \sim \mathcal{W}(\nu, \bm{I}_n)$ and $\bm{\Theta}_{\scaleto{Z}{4pt}} \sim \mathcal{W}(\nu + p - q, \bm{I}_n)$. Assume that $\bm{\Theta}_{\scaleto{X}{4pt}}$ and $\bm{\Theta}_{\scaleto{Z}{4pt}}$ structure the rows of respectively $\bm{X}$ and $\bm{Z}$ such that:
\begin{align}
\bm{X} | \bm{\Theta}_{\scaleto{X}{4pt}} &\sim \mathcal{N}(\bm{0}, \bm{\Theta}_{\scaleto{X}{4pt}}^{-1} \otimes \bm{I}_p) \label{eq:X_given_thetax}\\
\bm{Z} | \bm{\Theta}_{\scaleto{Z}{4pt}} &\sim \mathcal{N}(\bm{0}, \bm{\Theta}_{\scaleto{Z}{4pt}}^{-1} \otimes \bm{I}_q) \label{eq:Z_given_thetaz}
\end{align}
Then the solution of the precision coupling problem:
\begin{align*}
\min_{\bm{Z} \in \mathbb{R}^{n \times q}} -\mathbb{E}_{\bm{\Theta}_{\scaleto{X}{3pt}} | \bm{X}}\left[\log \mathbb{P}(\bm{\Theta}_{\scaleto{Z}{4pt}}=\bm{\Theta}_{\scaleto{X}{4pt}}|\bm{Z})\right]
\end{align*}
is a PCA embedding of $\bm{X}$ with $q$ components.
\end{theorem}
We now highlight the parallels with the previous construction done for neighbor embedding methods. First note that the multivariate Gaussian with full-rank precision is inherently a pairwise MRF \cite{rue2005gaussian}. When choosing the Gaussian kernel for neighbor embedding methods (equation \ref{eq:gaussian_kernel}), we saw that the graph Laplacian of $\bm{W}$ was playing the role of the among-row precision matrix. Recall that the later always has a null-space which is spanned by the CC indicator vectors (\cref{sec:laplacian_prop}). Here, the key difference is that we impose a full-rank constraint on the precision $\bm{\Theta}$. Concerning the priors, we choose the ones that are conjugate to the conditionals (\ref{eq:X_given_thetax}) and (\ref{eq:Z_given_thetaz}), as previously done when constructing the prior for neigbor embedding methods (definition \ref{def:prior_W}). Hence in the full-rank setting, the prior simply amounts to a Wishart distribution.
The above theorem further highlights the flexibility and generality of the graph coupling framework. Unlike usual constructions of PCA or probabilistic PCA \cite{tipping1999probabilistic}, in the above the linear relation between $\bm{X}$ and $\bm{Z}$ is recovered by solving the graph coupling problem and not explicitly stated beforehand. To the best of our knowledge, it is the first time such a link is uncovered between PCA and SNE-like methods. In contrast with the latter, PCA is well-known for its ability to preserve global structure while being significantly less efficient at identifying clusters \cite{anowar2021conceptual}. Therefore, as suspected in \cref{sec:interpretations}, the degeneracy of the conditional distribution given the graph is key to determine the distance preservation properties of the embeddings. We propose in \cref{sec:hierarchical_modelling} to combine both graph coupling approaches to strike a balance between global and local structure preservation.
\begin{remark}\label{rm:eigenmaps_coupling}
With the above theorem, one can adapt the model on $\bm{X}$ to retrieve an embedding $\bm{Z}$ that corresponds to the projection of $\bm{X}$ onto the principal eigen-subspaces of the expectation of any posterior precision. For instance, when choosing a graph $\bm{W}_{\scaleto{X}{4pt}}$ as constructed as in section \ref{sec:graph_structure} for $\bm{X}$ only, one ends-up diagonalizing the graph Laplacian of a neighborhood graph hence retrieves a Laplacian eigenmaps embedding. This remark places Laplacian eigenmaps as an in-between strategy combining full-rank (\textit{i.e.}\ PCA-like) structure in $\bm{Z}$ and degenerate structure (\textit{i.e.}\ similar to neighbor embedding methods) in $\bm{X}$.
\end{remark}
\subsection{Hierarchical Graph Coupling}\label{sec:hierarchical_modelling}
The goal of this section is to show that global structure in SNE-like embeddings can be improved by structuring the CCs' positions. We consider the following hierarchical model for $\bm{X}$, where $\mathcal{P}_{\scaleto{X}{4pt}} \in \{B,D,E\}$, $k_x$ satisfies the assumptions of \cref{prop:integrability_pairwise_MRF} and $\nu_{\scaleto{X}{4pt}} \geq n$:
\begin{align*}
\bm{W}_{\scaleto{X}{4pt}} \sim \mathbb{P}_{\scaleto{\mathcal{P}_{\scaleto{X}{3pt}},k_x}{5pt}}^{\varepsilon}(\cdot \: ; \bm{1},1), &\quad \bm{\Theta}_{\scaleto{X}{4pt}} | \bm{W}_{\scaleto{X}{4pt}} \sim \mathcal{W}(\nu_{\scaleto{X}{4pt}}, \bm{I}_{\scaleto{R}{4pt}}) \\
\bm{X}_{\scaleto{C}{4pt}} |\bm{W}_{\scaleto{X}{4pt}} \sim \mathbb{P}_{k_x}(\cdot \:| \bm{W}_{\scaleto{X}{4pt}}), &\quad \bm{X}_{\scaleto{M}{4pt}} | \bm{\Theta}_{\scaleto{X}{4pt}} \sim \mathcal{N}\left(\bm{0}, \left(\varepsilon \bm{U}_{\scaleto{[:R]}{5pt}} \bm{\Theta}_{\scaleto{X}{4pt}}\bm{U}^\top_{\scaleto{[R]}{5pt}}\right)^{-1} \otimes \bm{I}_p\right)
\end{align*}
where $\bm{U}_{\scaleto{[R]}{5pt}}$ are the eigenvectors associated to the Laplacian null-space of $\overline{\bm{W}}_{\scaleto{X}{4pt}}$. Given a graph $\bm{W}_{\scaleto{X}{4pt}}$, the idea is to structure the CCs' relative positions with a full-rank Gaussian model.
The same model is considered for $\bm{W}_{\scaleto{Z}{4pt}}$, $\bm{\Theta}_{\scaleto{Z}{4pt}}$ and $\bm{Z}$, choosing $\nu_{\scaleto{Z}{4pt}} = \nu_{\scaleto{X}{4pt}} + p - q$ for the Wishart prior to satisfy the assumption of \cref{PCA_graph_coupling}. With this in place, we aim at providing a complete coupling objective, matching the pairs $(\bm{W}_{\scaleto{X}{4pt}},\bm{\Theta}_{\scaleto{X}{4pt}})$ and $(\bm{W}_{\scaleto{Z}{4pt}},\bm{\Theta}_{\scaleto{Z}{4pt}})$. The joint negative cross-entropy can be decomposed as follows:
\begin{align}
&\mathbb{E}_{(\bm{W}_{\scaleto{X}{3pt}}, \bm{\Theta}_{\scaleto{X}{3pt}})|\bm{X}}\left[\log \mathbb{P}((\bm{W}_{\scaleto{Z}{4pt}},\bm{\Theta}_{\scaleto{Z}{4pt}}) = (\bm{W}_{\scaleto{X}{4pt}},\bm{\Theta}_{\scaleto{X}{4pt}}) | \bm{Z})\right] \nonumber\\
&= \mathbb{E}_{\bm{W}_{\scaleto{X}{3pt}}|\bm{X}}\left[\log \mathbb{P}(\bm{W}_{\scaleto{Z}{4pt}} = \bm{W}_{\scaleto{X}{4pt}} | \bm{Z})\right] + \label{eq:loss_LW} \\
& \mathbb{E}_{(\bm{W}_{\scaleto{X}{3pt}},\bm{\Theta}_{\scaleto{X}{3pt}})|\bm{X}}\left[ \log \mathbb{P}(\bm{\Theta}_{\scaleto{Z}{4pt}} = \bm{\Theta}_{\scaleto{X}{4pt}}| \bm{W}_{\scaleto{Z}{4pt}} = \bm{W}_{\scaleto{X}{4pt}}, \bm{Z}) \right] \label{eq:add_term_Coupling}
\end{align}
where (\ref{eq:loss_LW}) is the usual coupling criterion of $\bm{W}_X$ and $\bm{W}_Z$ capturing intra-CC variability while (\ref{eq:add_term_Coupling}) is a penalty resulting from the Gaussian structure on $\mathcal{S}_{\scaleto{M}{4pt}}$. Constructed as such, the above objective allows a trade-of between local and global structure preservation. Following current trends in DR \cite{kobak2021initialization}, we propose to take care of the global structure first \textit{i.e.}\ focusing on (\ref{eq:add_term_Coupling}) before (\ref{eq:loss_LW}). The difficulty of dealing with (\ref{eq:add_term_Coupling}) lies in the hierarchical construction of the graph and the Gaussian precision (see \cref{fig:graphical_model_hierarchical}). We state the following result.
\begin{wrapfigure}[15]{R}{0.5\textwidth}
\begin{center}
\centerline{\includegraphics[width=0.5\columnwidth]{tSNE_truth.pdf}}
\caption{Left: MNIST t-SNE (perp : 30) embeddings initialized with i.i.d $\mathcal{N}(0,1)$ coordinates. Middle: using these t-SNE embeddings, mean coordinates for each digit are represented. Right: we compute a matrix of mean input coordinates for each of the $10$ digits and embed it using PCA. For t-SNE embeddings, the positions of clusters vary accross different runs and don't visually match the PCA embeddings of input mean vectors (right plot).}
\label{fig:tSNE-clusters-truth}
\end{center}
\end{wrapfigure}
\begin{corollary}\label{corollary_ccPCA}
Let $\bm{W}_{\scaleto{X}{4pt}} \in \mathcal{S}_{\scaleto{W}{4pt}}$, $\bm{L} = L(\overline{\bm{W}}_{\scaleto{X}{4pt}})$ and $\mathcal{S}^q_{\scaleto{M}{4pt}}= (\ker \bm{L}) \otimes \mathbb{R}^q$, then for all $\varepsilon > 0$, given the above hierarchical model, the solution of the problem:
$$\min_{\bm{Z} \in \mathcal{S}^q_{M}} \: -\mathbb{E}_{\bm{\Theta}_{\scaleto{X}{3pt}}| \bm{X}}\left[ \log \mathbb{P}(\bm{\Theta}_{\scaleto{Z}{4pt}} = \bm{\Theta}_{\scaleto{X}{4pt}}| \bm{W}_{\scaleto{Z}{4pt}} = \bm{W}_{\scaleto{X}{4pt}}, \bm{Z}) \right]$$
is a PCA embedding of $\bm{U}_{\scaleto{[:R]}{5pt}}\bm{U}_{\scaleto{[R]}{5pt}}^\top\bm{X}$ where $\bm{U}_{\scaleto{[:R]}{5pt}}$ are the normalized CCs' membership vectors of $\overline{\bm{W}}_{\scaleto{X}{4pt}}$.
\end{corollary}
\begin{remark}
Note that while (\ref{eq:loss_LW}) approximates the objective of SNE-like methods when $\varepsilon \to 0$, the minimizer of (\ref{eq:add_term_Coupling}) given by \cref{corollary_ccPCA} is stable for all $\varepsilon$.
\end{remark}
From this observation, we propose a simple heuristic to minimize (\ref{eq:add_term_Coupling}) that consists in computing a PCA embedding of $\mathbb{E}_{\mathbb{P}_{\scaleto{\mathcal{P}_X}{4pt}}(\cdot;\bm{K}_{\scaleto{X}{3pt}})}\left[ \bm{U}_{\scaleto{[:R]}{5pt}}\bm{U}_{\scaleto{[R]}{5pt}}^\top \right]\bm{X}$. The distribution of the connected components of the posterior of $\bm{W}_{\scaleto{X}{4pt}}$ being intractable, we resort to a Monte-Carlo estimation of the above expectation. The latter procedure called \textit{ccPCA} aims at recovering the inter-CC structure that is filtered by SNE-like methods. \textit{ccPCA} may then be used as initialization for optimizing (\ref{eq:loss_LW}) which is done by running the DR method corresponding to the graph priors at hand (\cref{sec:retrieving_DR_methods}). This second step essentially consists in refining the intra-CC structure.
\subsection{Experiments with \textit{ccPCA}}\label{sec:ccPCA}
\begin{figure*}[t]
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{cluster_positions.pdf}}
\caption{Top: MNIST embeddings produced by PCA, Laplacian eigenmaps, \textit{ccPCA} and finally t-SNE launched after the previous three embeddings to improve the fine-grain structure. Bottom: mean coordinates for each digit using the embeddings of the first row. The color legend is the same as in \cref{fig:tSNE-clusters-truth}. t-SNE was trained during $1000$ iterations using default parameters with the openTSNE implementation \cite{polivcar2019opentsne}.}
\label{fig:methods_embeddings}
\end{center}
\vspace{-0.8cm}
\end{figure*}
\Cref{fig:tSNE-clusters-truth} shows that a t-SNE embedding of a balanced MNIST dataset of 10000 samples \cite{deng2012mnist} with isotropic Gaussian initialization performs poorly in conserving the relative positions of clusters. As each digit cluster contains approximately $1000$ points, with a perplexity of $30$, sampling an edge across digit clusters in the graph posterior $\mathbb{P}_{\scaleto{\mathcal{P}_X}{4pt}}(\cdot;\bm{K}_{\scaleto{X}{4pt}})$ is very unlikely. Recall that the perplexity value \cite{maaten2008tSNE} corresponds to the approximate number of effective neighbors of each point. Hence images of different digits are with very high probability in different CCs of the graph posterior and their CC-wise means are not coupled as discussed in \cref{sec:interpretations}. To remedy this in practice, PCA or Laplacian eigenmaps are usually used as initialization \cite{kobak2021initialization}.
These strategies are tested (\cref{fig:methods_embeddings}) together with \textit{ccPCA}. This shows that
\textit{ccPCA} manages to retrieve the digits that mostly support the large-scale variability as measured by the peripheral positioning of digits $0$ (blue), $2$ (green), $6$ (pink) and $7$ (grey) given by the right side of \cref{fig:tSNE-clusters-truth}. Other perplexity values for \textit{ccPCA} are explored in appendix \ref{sec:other_perp} while the experimental setup is detailed in appendix \ref{sec:setup_exp}. In appendix \ref{sec:quantitative_evaluation}, we perform quantitative evaluations of \textit{ccPCA} for both t-SNE and UMAP on various datasets using K-ary neighborhood criteria. We find that using \textit{ccPCA} as initialization is in general more reliable than PCA and Laplacian eigenmaps for preserving global structure using both t-SNE and UMAP.
Compared to PCA, \textit{ccPCA} manages to aggregate points into clusters, thus filtering the intra-cluster variablity and focusing solely on the inter-cluster structure. Compared to Laplacian eigenmaps which performs well at identifying clusters but suffers from the same deficiency as t-SNE for positioning them, \textit{ccPCA} retains more of the coarse-grain structure. These observations support our unifying probabilistic framework and the theoretical results about the MRF degeneracy which are the leading contributions of this article. Therefore the \textit{ccPCA} initialization appears as a first stepping stone towards more grounded DR methods based on the probabilistic model presented in this article.
\section{Conclusion and Perspectives}\label{Perspectives}
In this work we shed a new light on most popular DR methods by showing that they can be unified within a common probabilistic model in the form of latent Markov Random Fields Graphs coupled by a cross entropy. The definition of such a model constitutes a major step towards the understanding of common dimension reduction methods, in particular their structure preservation properties as discussed in this article.
Our work offers many perspectives, among which the possibility to enrich the probabilistic model with more suited graph priors. Currently considered priors are simply the ones that are conjugate to the MRFs thus they are mostly designed to yield a tractable coupling objective. However they may not be optimal and could be modified to capture targeted features, \textit{e.g.}\ communities, in the input data, and give adapted representations in the latent space.
Finally, the probabilistic model could be leveraged to tackle hyper-parameter calibration, especially kernel bandwidths that have a great influence on the quality of the representations and are currently tuned using heuristics with unclear motivations.
\newpage
|
\section{Introduction} \label{sec:intro}
The prompt emission of gamma-ray bursts (GRBs) is considered to be radiation from relativistic jets \citep[for reviews, see, e.g.,][]{2015PhR...561....1K,2015AdAst2015E..22P}.
The GRB prompt spectra are usually well fitted by the Band function \citep{1993ApJ...413..281B},
which is smoothly-broken power-law with the low and high energy spectral indexes, $\alpha$ (typically $\sim-1$) and $\beta$ ($\sim-2.2$), respectively, and the peak energy of the $\nu F_{\nu}$ spectrum $E_{\mathrm{p}}$ \citep[typically several hundreds $\mathrm{keV}$, see,][]{2000ApJS..126...19P,2006ApJS..166..298K,2011A&A...530A..21N,2014ApJS..211...12G}.
One of the promising models to explain the typical GRB spectrum is synchrotron radiation from electrons injected with a power-law energy distribution with a low-energy cutoff \citep{1996ApJ...466..768T,2000ApJ...543..722L}.
But there exist several problems in this standard synchrotron model.
Representative one of them is the synchrotron fast cooling problem \citep{2000MNRAS.313L...1G}.
Given a magnetic field consistent with $E_{\mathrm{p}}$,
electrons promptly lose energy via synchrotron radiation during the dynamical timescale,
and photons emitted from such cooled electrons contribute to the low-energy spectrum.
As a result, the predicted spectral index becomes $\alpha=-1.5$,
which is significantly softer than the typical index $\alpha\sim-1$.
Several alternative models to reconcile the observed hard spectra have been proposed.
The photosphere emission model is the most promising one \citep{2005ApJ...628..847R,2010MNRAS.407.1033B,2010ApJ...725.1137L,2011ApJ...732...49P}.
\citet{2010MNRAS.407.1033B} and \citet{2010ApJ...725.1137L} reproduce high energy power law component ($\beta\sim-2.2$) by comptonization of thermal photons emitted by subphotospheric dissipation.
Although the simple Planck spectrum implies a too hard spectrum ($\alpha=+1$),
the Doppler-boosted emission components from different angles with respect to the line of sight can soften the spectrum
as $\alpha \sim +0.4$ \citep{2010MNRAS.407.1033B} or $\alpha \sim -1$ \citep{2011ApJ...732...49P}.
\citet{2017MNRAS.468.3202B} proposed the magnetically-dominated jet model with the photospheric emission.
The model reproduces the observed spectrum with $\alpha\sim-1$ by superimposed subdominant synchrotron radiation
on the dominant thermal radiation.
\citet{2001A&A...372.1071D} and \citet{2011A&A...526A.110D} suggested that if the synchrotron self-Compton (SSC) cooling rate in the Klein-Nishina regime exceeds the synchrotron cooling rate, a hard synchrotron spectrum can be reproduced.
However, the SSC dominance in the GeV band is not confirmed in major fraction of GRBs observed with Fermi LAT \citep{2011ApJ...730..141Z,2013ApJS..209...11A}.
If the SSC emission in GeV--TeV range is absorbed in the source via electron--positron pair creation,
the electromagnetic cascade emission makes the low-energy spectrum soft significantly
\citep{2011ApJ...739..103A,2012ApJ...757..115A}.
The SSC emission in 0.1--1 MeV \citep{2004MNRAS.352L..35S,2009A&A...498..677B} is another option
to reproduce the GRB spectrum.
This SSC model predicts strong optical and GeV-TeV emissions \citep{2009MNRAS.393.1107P},
such a signature has not been established as a consensus in the majority of GRBs \citep{2011ApJ...730..141Z,2013ApJS..209...11A}.
The jitter radiation \citep{2000ApJ...540..704M} is an emission process in a small-scale magnetic turbulence.
This process leads to a harder spectrum ($\alpha=0$) than synchrotron emission.
But in the fast cooling regime, the jitter radiation results in $\alpha=-1.5$.
Even in this model, the suppression of the electron cooling is required.
In \citet{2021NatCo..12.4040R}, the observed evolutions of the spectral photon index and the flux density in the steep decay phase of the early afterglow, which is considered as a tail emission of a prompt emission, suggest that adiabatic cooling is the dominant cooling process.
Therefore, the suppression of the radiative cooling for low-energy particles
seems essential for the GRB emission process.
The proton synchrotron model \citep{2020A&A...636A..82G} is one of such models.
Alternatively, a continuous electron acceleration by turbulence
leads to the balance between acceleration and synchrotron cooling,
which produces a hard electron distribution \citep{2009ApJ...705.1714A,2015MNRAS.454.2242A,2017ApJ...846L..28X,2018MNRAS.476.1785B}.
If the decay timescale of the magnetic field is comparable
to the cooling timescale,
the inefficient cooling can produce the required spectral shape
\citep{2006ApJ...653..454P,2011A&A...526A.110D,2013ApJ...769...69B}.
\citet{2014NatPh..10..351U} calculated spectra of instantaneous emissivity after magnetic field decays and obtained hard spectra with $\alpha\sim-1.2$. However, in most of GRB observations, spectra are obtained by integrating photons for significantly longer timescale ($\gtrsim 1\mathrm{s}$) than the variability timescale. In this paper, we focus on the time-integrated spectra.
Recent Particle-in-Cell (PIC) simulations \citep{2018JPlPh..84f7201P,2019ApJ...886..122C} show that magnetic reconnection induced by turbulence in a magnetically dominated plasma produces anisotropic distribution of accelerated particles.
The acceleration process in their simulations are divided into two:
the prompt acceleration along the guiding magnetic field at the reconnection site
and the succeeding turbulence acceleration similar to the 2'nd order Fermi acceleration.
The first prompt acceleration injects relatively low-energy electrons with small pitch angles.
High-energy electrons are accelerated by turbulence, which makes electrons isotropic.
In such cases, the synchrotron emission from low-energy electrons with small pitch angles
is suppressed.
The spectral hardening by such an anisotropic electron distribution has been
pointed out by several authors \citep{2000ApJ...543..722L,2018ApJ...864L..16Y}.
This possibility is discussed in the context of the SSC dominance in blazars \citep{2021MNRAS.506...38S}.
In this paper, we discuss the GRB spectrum with anisotropic distributions as suggested by \citet{2019ApJ...886..122C}.
The small pitch angle for low-energy electrons
suppresses the electron cooling even in a strong magnetic field.
The effect of the energy dependence of the anisotropy is not trivial for the synchrotron spectral shape.
We numerically solve the temporal evolutions of electron energy and pitch angle distributions,
which are essential to compare the time-integrated spectrum with observations.
We take into account adiabatic cooling and decrease of the magnetic field due to the GRB jet expansion.
The structure of this paper is as follows.
In \S \ref{sec:method}, we explain our calculation method of synchrotron spectra taking into account anisotropic electron distributions.
In \S \ref{sec:result}, we show the numerical calculation results of the spectra.
and discuss the required parameter regions.
In \S \ref{sec:sum}, we summarize our results and discuss the validity of the anisotropic model.
\section{Method} \label{sec:method}
We consider a conical jet with a constant bulk Lorentz factor $\Gamma$ and an injection of electrons at a radius $R=R_0$.
We numerically solve the temporal evolutions of the electron energy and pitch angle distributions in the jet comoving frame. The comoving volume expands with its propagation, so we take into account adiabatic cooling and decrease of the magnetic field as well as synchrotron cooling.
Integrating photons emitted from those electrons, the photon spectrum is obtained.
Based on this time-integrated spectrum, we discuss the effects of the anisotropic distribution.
\subsection{\bf{Evolution of the Electron Distribution}}
The injection spectrum is written as follows,
\begin{eqnarray}
\label{eq:initial}
N(\gamma,\mu,t=0)&=&C\gamma^{-p}f(\gamma,\mu)\ \nonumber\\
& &(\gamma_{\rm{min}}<\gamma<\gamma_{\rm{max}}),
\end{eqnarray}
where $C$ is a normalization constant, $N(\gamma,\mu,t)$ is the electron number per $\gamma$ (the Lorentz factor of electrons) per $\mu=\cos\psi$ (the pitch angle cosine) in the comoving frame, and $f(\gamma,\mu)$ expresses the pitch angle distribution, which generally depends on $\gamma$.
We assume a power-law energy distribution with a low energy cutoff $\gamma_{\mathrm{min}}$, which has been conventionally assumed in the interpretation of the GRB emission \citep{1996ApJ...466..768T}.
We fix the electron spectral index $p=2.5$ to reproduce the observed high-energy spectral index $\beta\sim-2.2$ ($\beta \simeq -(p+2)/2$ in the fast cooling synchrotron spectrum), while we discuss cases with different $p$ values in \S \ref{sec:pitch}.
The time evolution of the electron distribution in the comoving frame
is given by the following continuity equation
in the $(\gamma,\mu)$-space,
\begin{eqnarray}
\label{eq:continuity}
\frac{\partial}{\partial t}N(\gamma,\mu,t)
+\frac{\partial}{\partial \gamma}\{(\dot{\gamma}_{\rm syn}+\dot{\gamma}_{\rm adi})N(\gamma,\mu,t)\}\nonumber\\
+\frac{\partial}{\partial \mu}\{(\dot{\mu}_{\rm syn}+\dot{\mu}_{\rm adi})N(\gamma,\mu,t)\}=0,
\end{eqnarray}
where $\dot{\gamma}$ and $\dot{\mu}$ are the energy and pitch angle changing rates, respectively.
The subscripts ``$\mathrm{syn}$'' and ``$\mathrm{adi}$'' mean the synchrotron and adiabatic cooling effects, respectively.
Here, we neglect the SSC cooling, which depends on the electron luminosity or normalization $C$,
and the magnetic field.
Thus, the discussion on the photon spectral shape in this paper does not depend on the normalization $C$.
If we consider a magnetically dominated jet, which is likely for the magnetic reconnection model,
the negligible contribution of the SSC cooling may be adequate.
For simplicity, we also neglect the pitch angle scattering by turbulence.
We will discuss the influence of the pitch angle scattering in \S \ref{sec:sum}.
The synchrotron energy changing rate \citep{1979rpa..book.....R} is
\begin{eqnarray}
\label{eq:synenergy}
\dot{\gamma}_{\rm syn}=-\frac{2e^4B^2(\gamma^2-1)(1-\mu^2)}{3m_{\mathrm{e}}^3c^5},
\end{eqnarray}
where $B$ is the magnetic field strength.
In the frame where $\mu=0$ for an electron,
a circular motion in the same plane is observed, and the radiation is symmetric with respect to the plane
so that the electron is kept in the same plane.
Therefore, even in the laboratory frame,
the radiation reaction does not change the velocity along the magnetic field $\beta \mu$,
where $\beta=\sqrt{1-1/\gamma^2}$.
Combining $\beta\mu=\mathrm{const.}$ and the equation (\ref{eq:synenergy}), we obtain
the pitch angle changing rate by radiation reaction \citep{1985ApJ...299..987P,2016MNRAS.458.2303S} as
\begin{eqnarray}
\label{eq:synpitch}
\dot{\mu}_{\mathrm{syn}}=\frac{2e^4B^2\mu(1-\mu^2)}{3m_{\mathrm{e}}^3c^5\gamma}.
\end{eqnarray}
We assume a toroidal dominant magnetic field decreasing with the jet expansion as $B\propto R^{-1}$,
so we write
\begin{eqnarray}
\label{eq:B}
B=B_{0}\left(\frac{R_{0}}{R}\right)=B_{0}\left(1+\frac{t}{t_\mathrm{d}}\right)^{-1},
\end{eqnarray}
where $B_{0}$ is the magnetic field strength at $R=R_{0}$.
Using the time $t$ in the comoving frame, the radius increases as $R=R_{0}+\Gamma \beta_{\mathrm{j}} ct$, where the jet velocity $\beta_{\mathrm{j}}=\sqrt{1-1/\Gamma^2}$.
We have defined the dynamical timescale as
\begin{eqnarray}
\label{eq:dynamicaltime}
t_\mathrm{d} \equiv \frac{R_{0}}{\Gamma\beta_{\mathrm{j}} c}.
\end{eqnarray}
Even if the field $B$ decreases, an electric field along the magnetic field is not induced
($\dot{\bm{B}}=-c \nabla \times \bm{E}$).
As the force along the field $B$ is zero, the parallel momentum $p_{\parallel}\equiv
m_{\mathrm{e}}c\gamma \beta\mu$ is conserved.
With the adiabatic invariant $p_{\perp}^2/B$ \citep{1963RvGSP...1..283N}, where the perpendicular momentum
$p_{\perp}\equiv m_{\mathrm{e}}c \gamma \beta\sqrt{1-\mu^2}$,
we obtain the adiabatic energy changing rate
\begin{eqnarray}
\label{eq:adienergy}
\dot{\gamma}_{\rm adi}=-\frac{(\gamma^2-1)(1-\mu^2)}{2\gamma}\left(t_{\mathrm{d}}+t\right)^{-1},
\end{eqnarray}
and the adiabatic pitch angle changing rate
\begin{eqnarray}
\label{eq:adipitch}
\dot{\mu}_{\rm adi}=\frac{\mu(1-\mu^2)}{2}\left(t_{\mathrm{d}}+t\right)^{-1}.
\end{eqnarray}
The evolution of the electron distribution is characterized by the ratio $t_{\mathrm{d}}/t_{\mathrm{c}}$,
where $t_{\mathrm{c}}$ is the initial timescale of the synchrotron cooling
for electrons of $\gamma=\gamma_{\mathrm{min}}$,
\begin{eqnarray}
\label{eq:coolingtime}
t_{\mathrm{c}}=\frac{9 m_{\mathrm{e}}^3 c^5}{4 e^4 B_{0}^2 \gamma_{\mathrm{min}}},
\end{eqnarray}
where we adopt the average value in the isotropic distribution for the pitch angle as $\langle\sin^2\psi\rangle=2/3$.
The ratio of the two pitch angle change rates is written as
\begin{eqnarray}
\label{eq:pitchratio}
\frac{\dot{\mu}_{\mathrm{syn}}}{\dot{\mu}_{\mathrm{adi}}}=\frac{3}{\gamma\gamma_{\mathrm{min}}}\frac{t_{\mathrm{d}}}{t_{\mathrm{c}}}\left(1+\frac{t}{t_{\mathrm{d}}}\right)^{-1}.
\end{eqnarray}
For $\gamma^2_{\mathrm{min}} \gg t_{\mathrm{d}}/t_{\mathrm{c}}$,
the pitch angle change due to synchrotron radiation is negligible.
\begin{figure}[htp]
\centering
\includegraphics[width=8cm]{runge.pdf}
\caption{The temporal evolutions of electron trajectories in the $\gamma$-$\mu$ space in the case of $t_{\mathrm{d}}/t_{\mathrm{c}}=1$ with $\gamma_{\mathrm{min}}=10^4$.
Electrons are injected with $\gamma=10^7$ and different values of $\mu$ at $t=0$.}
\label{fig:gammamu}
\end{figure}
Figure \ref{fig:gammamu} shows electron trajectories in the $\gamma$-$\mu$ space, where the initial $\gamma$ is $10^7$ with various pitch angles.
Initially, the cooling timescale of electrons of $\gamma=10^7$ is $10^3$ times shorter
than that for $\gamma=\gamma_{\mathrm{min}}$,
so those electrons promptly lose their energy with negligible changes of $\mu$.
For $t> t_{d}$, the magnetic field decreases significantly and the synchrotron cooling is suppressed.
The electrons mainly lose their energies by adiabatic cooling.
From the equation (\ref{eq:adipitch}), the timescale of the pitch angle change by adiabatic cooling
is longer than $t$ for $t<t_{\mathrm{d}}$.
The pitch angle changes significantly for $t> t_{d}$.
From $\dot{\mu}_{\mathrm{adi}}\propto\mu(1-\mu^2)$, electrons with intermediate values of $\mu$
experience the change of the pitch angles most significantly.
Figure \ref{fig:gammamu} also shows that the synchrotron energy loss is relatively suppressed for smaller pitch angles (larger $\mu$).
\subsection{\bf Initial Pitch Angle Distribution} \label{sec:pitch}
Motivated by the simulations of turbulence magnetic reconnection in
\citet{2019ApJ...886..122C}, we assume anisotropic electron distributions at the injection.
According to \citet{2019ApJ...886..122C}, at the reconnection site
electrons are accelerated along the guiding magnetic field with small pitch angles,
and the succeeding turbulence acceleration gradually accelerates electrons perpendicular to the field.
The pitch angle distribution becomes isotropic at higher energies.
We adopt a Gaussian distribution in the $\mu$-space,
\begin{eqnarray}
\label{eq:pitchdistribution}
f(\gamma,\mu)=f_0(\gamma) \exp\left(-\frac{(\mu-\overline{\mu})^2}{2\Delta\mu^2}\right),
\end{eqnarray}
where $\Delta \mu$ is a function of $\gamma$, and the normalization $f_0(\gamma)$ is adjusted to satisfy
$\int_{-1}^{1} f(\gamma,\mu) d \mu=1$.
Most of low energy electrons have small pitch angles $\mu\simeq1$, so we adopt the peak pitch angle as
\begin{eqnarray}
\label{eq:mubar}
\overline{\mu}=1.
\end{eqnarray}
In the PIC simulations of \citet{2019ApJ...886..122C},
$\overline{\mu}$ becomes smaller at higher energies.
For simplicity, however, we fix $\overline{\mu}=1$ irrespectively of $\gamma$.
Alternatively, we assume a dispersion growing with $\gamma$ as
\begin{eqnarray}
\label{eq:deltamu}
\Delta\mu=\Delta\mu_{\rm{min}}\left(\frac{\gamma}{\gamma_{\rm{min}}}\right)^k,
\end{eqnarray}
where constant parameters $\Delta\mu_{\mathrm{min}}$ and $k$ are introduced.
As fiducial values, we adopt $\Delta\mu_{\mathrm{min}}=0.01$ and $k=1$.
In \citet{2019ApJ...886..122C}, the averaged $\sin\psi$ at minimum Lorentz factor $\gamma_{\mathrm{min}}$ is $\simeq0.14$,
which corresponds to $\Delta\mu_{\mathrm{min}}\simeq0.01$,
for the sigma parameter $\sigma\gtrsim10$ and the initial magnetic turbulent strength $\delta B/B=0.5$.
The pitch angle distribution becomes almost isotropic at $\gamma\simeq10\gamma_{\mathrm{min}}$ in the PIC simulations.
Those detailed results may depend on initial conditions and/or boundary conditions, so the value $k$ is not definite at present. In this paper, we adopt a value of $0.5<k<2$.
\begin{figure}[htp]
\centering
\includegraphics[width=8cm]{delelectron1000.pdf}
\includegraphics[width=8cm]{isoelectron1000.pdf}
\includegraphics[width=8cm]{anisoelectron1000.pdf}
\caption{The evolutions of electron distributions for a single pitch angle
($f(\gamma,\mu)=\delta(\mu-\frac{1}{2})$, upper),
isotopic distribution ($f(\gamma,\mu)=1/2$, middle)
and our anisotropic model ($\Delta\mu_{\mathrm{min}}=0.01,k=1$, lower). The other parameters
are common as $\gamma_{\mathrm{min}}=10^4$ and $t_{\mathrm{d}}/t_{\mathrm{c}}=10^3$.}
\label{fig:cooling}
\end{figure}
In Figure \ref{fig:cooling}, we show the evolutions of the pitch-angle-integrated energy distributions of electrons: $\int_{-1}^{1}d\mu N(\gamma,\mu,t)\equiv N\left(\gamma,t \right)$.
We show three cases with different initial distributions: $f(\gamma,\mu)=\delta(\mu-\frac{1}{2})$ (single pitch angle),
$f(\gamma,\mu)=1/2$ (isotropic), and our anisotropic model with equation (\ref{eq:pitchdistribution}).
For the single pitch angle case, all the electrons have the same $\mu=0.5$, so they all cool down to the same Lorentz factor of $\gamma_{\mathrm{c}}(t)\simeq 9m_{\mathrm{e}}^3c^5/(4e^4B^2t)$.
For the isotropic case, the radiative cooling is suppressed only for electrons with very small pitch angles,
so such electrons remain in higher energy region even at a later stage.
This makes broken power-law energy distributions as shown in the middle panel.
After $t=t_{\mathrm{d}}=10^3t_{\mathrm{c}}$, the magnetic field decreases significantly.
At this stage, the cooling process switches from synchrotron cooling to adiabatic cooling.
For the anisotropic case of $\Delta\mu_{\mathrm{min}}=0.01,k=1$,
the cooling is further suppressed due to the small average of pitch angles compared to the isotropic case.
\subsection{\bf Synchrotron Spectrum}
We calculate time-integrated synchrotron spectra, because the observed spectra are usually provided as time-integrated one with an interval of a few seconds
due to poor photon statistics.
Integrating synchrotron radiation from the time $t=0$ to $t_{\rm end}$,
the spectrum is given by
\begin{eqnarray}
\label{eq:synspectrum}
\nu E_{\nu}=\int_0^{t_{\rm end}} dt \int_1^{\gamma_{\rm{max}}} d\gamma \int_{-1}^{1}d\mu N(\gamma,\mu,t)\nonumber\\
\times\frac{\sqrt{3}\nu e^{3}B\sqrt{1-\mu^2}}{m_{\mathrm{e}} c^2}F_{\mathrm{syn}}\left(\frac{\nu}{\nu_{\mathrm{syn}}}\right),
\end{eqnarray}
where $\nu$ is the photon frequency in the comoving frame, $E_{\nu}$ is the energy emitted per $\nu$, $\nu_{\rm{syn}} \equiv 3\gamma^{2}eB\sqrt{1-\mu^2}/(4\pi m_{e}c)$ is the synchrotron characteristic frequency, $F_{\mathrm{syn}}\left(x\right) \equiv x \int_{x}^{\infty} K_{\frac{5}{3}}(\xi)d\xi$
and $K_{\frac{5}{3}}(\xi)$ is the modified bessel function of the $5/3$ order \citep{1979rpa..book.....R}.
In the following discussion, we focus on only the spectral shape, in particular on the photon spectral index $\alpha$. We define the spectrum $\langle{\nu E_{\nu}}\rangle$ that is normalized by the total energy of electrons,
\begin{eqnarray}
\label{eq:normalizedspectrum}
\langle{\nu E_{\nu}}\rangle\equiv \frac{\nu E_{\nu}}{E_{\mathrm{ini}}},
\end{eqnarray}
where
\begin{eqnarray}
\label{initialenergy}
E_{\mathrm{ini}}=\int_{\gamma_{\rm{min}}}^{\gamma_{\rm{max}}} d\gamma \int_{-1}^{1}d\mu \gamma m_{\mathrm{e}}c^{2}N(\gamma,\mu,t=0).
\end{eqnarray}
We shift the frequency in the jet's comoving frame to the observer frame
with $\nu_{\rm{obs}}=\Gamma \nu$.
\section{Results} \label{sec:result}
\subsection{\bf Spectral Shape}
Figure \ref{fig:spectra} shows the calculated spectra for various $t_{\mathrm{d}}/t_{\mathrm{c}}$ values.
Note again that the cooling time $t_{\mathrm{c}}$ is one for the isotropic distribution so that electrons above $\gamma_{\rm min}$ may not cool even for $t>t_{\rm c}$ depending on the degree of anisotropy.
In Figure \ref{fig:spectra}, we stop the calculation at $t_{\rm end}=10t_{\mathrm{d}}$,
at which the magnetic field decreases enough to suppress the synchrotron radiation.
In spite of the non-trivial anisotropic distributions,
those spectra can be well fitted with the Band function,
but the index $\alpha$ obtained from the fitting depends on the fitting energy range.
In this paper, we estimate the spectral index $\alpha$ at the photon energy of
$1/50$ times the peak energy of the $\nu E_{\nu}$ spectrum.
This energy corresponds to $\sim 10$ keV for the typical peak energy of $E_{\rm p}\sim 500$ keV.
In Figure \ref{fig:spectra},
the hardest spectral index is $\alpha \sim-1.2$ for $t_{d}/t_{c}=45$.
This large value of $t_{d}/t_{c}>1$ is due to suppression of cooling by anisotropic distribution with small pitch angles.
If we can perfectly block the electron cooling to $\gamma<\gamma_{\rm min}$,
the synchrotron emission should provide $\alpha=-0.67$.
but the radiation at $t>t_{\mathrm{d}}$ with the decreasing magnetic field
contributes to the low-energy region, and softens the spectrum as $\alpha \sim-1.2$.
For a much larger value of $t_{\rm{d}}/t_{\rm{c}}$, the index becomes $\alpha\sim-1.5$
(see the case for $t_{\mathrm{d}}/t_{\mathrm{c}}=1000$).
On the other hand, for a smaller value of $t_{d}/t_{c}$,
we again obtain a softer spectrum (see e.g. the case for $t_{\mathrm{d}}/t_{\mathrm{c}}=1$),
where the spectral peak is attributed to the cooling
break $\gamma_{\rm c}$ in the electron spectrum rather than $\gamma_{\rm min}$.
The low-energy spectrum in this case is attributed to the slow cooling electrons below $\gamma_{\rm c}$.
\begin{figure}[htp]
\centering
\includegraphics[width=8cm]{spectrumfidu.pdf}
\caption{The calculated spectra in the comoving frame for $\gamma_{\mathrm{min}}=10^4$ and $B_{0}=2500\mathrm{G}$.
The solid lines are spectra for the anisotropic model with $\Delta \mu_{\mathrm{min}}=0.01$, $k=1$,
and $t_{\rm end}=10t_{\mathrm{d}}$ for different $t_{\mathrm{d}}/t_{\mathrm{c}}$ values.
The dotted black line is an isotropic case with $t_{\mathrm{d}}/t_{\mathrm{c}}=20$ and $t_{\rm end}=10t_{\mathrm{d}}$.}
\label{fig:spectra}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=8cm]{spectrump.pdf}
\caption{The calculated spectra in the comoving frame for $\gamma_{\mathrm{min}}=10^4$ and $B_{0}=2500\mathrm{G}$.
The dotted black line is an isotropic case with $t_{\mathrm{d}}/t_{\mathrm{c}}=20$ and $t_{\rm end}=10t_{\mathrm{d}}$.
The red lines are spectra for the anisotropic model with $\Delta \mu_{\mathrm{min}}=0.01$, changing the parameters as $t_{\rm end}=t_{\mathrm{d}}$ (dashed) and $k=1.5$ (dotted), respectively, from the reference model (solid, $k=1$ and $t_{\rm end}=10t_{\mathrm{d}}$), which is the same as the red line in Figure \ref{fig:spectra}.
The blue, magenta and brown lines are the same anisotropic case with $p=3$ and different $t_{\mathrm{d}}/t_{\mathrm{c}}$ values. The other parameter values are the same as those for the reference case (red solid).}
\label{fig:spectrap}
\end{figure}
A case with the isotropic distribution (the dotted black line)
shows a higher $E_{\rm p}$ than the anisotropic cases.
The lower peak energies in the anisotropic cases are due to the small pitch angles at $\gamma=\gamma_{\rm min}$,
as $E_{\rm p} \propto \sqrt{\langle \sin^2{\psi} \rangle} \sim \sqrt{2 \Delta \mu_{\rm min}}\sim 0.14$.
With decreasing $t_{\rm d}/t_{\rm c}$, the peak values of the $\nu E_{\nu}$
decreases, which shows the suppression of the radiation efficiency.
The peak energy $E_{\rm p}$ also increases with decreasing $t_{\rm d}/t_{\rm c}$
because of the suppression of electron cooling.
In Figure \ref{fig:spectra}, we have assumed $k=1$ for the solid lines,
whose flux above $E_{\rm p}$ is suppressed by the anisotropic distribution
compared to the isotropic case (the dotted black line).
Figure \ref{fig:spectrap} shows spectra for isotropic and anisotropic cases with different $p$, $k$ and $t_{\mathrm{end}}$.
When we stop calculation at $t=t_{\mathrm{d}}$ (dashed red), we obtain hard spectra with $\alpha=-0.90$,
because of the lack of the late radiation.
This optimistic assumption, the inefficient radiation for $t>t_{\mathrm{d}}$,
may be justified if the decay time of the shock-amplified magnetic field is comparable
to the dynamical timescale.
However, a kind of plasma instabilities responsible to the field amplification
implies a highly disturbed field, which may be incompatible with the negligible
pitch-angle diffusion that we have assumed.
Alternatively, the electron escape from the emission region with a strong magnetic field
can implement the termination of radiation at $t\sim t_{\mathrm{d}}$.
Another possible interpretation is that the temporal resolution of gamma-ray spectral observations is comparable to the dynamical timescale $t_{\rm d}/\Gamma$.
The typical time-bin for spectral analysis is a few seconds. So the variability timescale $R_0/(c \Gamma^2)$ should be comparable to a second in this interpretation.
For $k=1.5$ (the dotted red line), the pitch-angle distribution rapidly approaches
an isotropic one with energy, so that the photon spectrum also
approaches to the flux of the isotropic model (the dotted black line)
at a higher energy.
For $p=3$ (blue, magenta and brown lines), the soft electron energy distribution yields radiation dominated by anisotropic electrons at the low energy region.
The contribution of cooled electrons isotropically injected at a higher energy to the low-energy flux is low compared to the cases with $p=2.5$.
The hardest spectral index for $p=3$ is $\alpha=-1.02$ with $t_{\mathrm{d}}/t_{\mathrm{c}}=10$ (magenta line), which is harder than $\alpha=-1.20$ in the $p=2.5$ case (red solid line).
The range of $t_{\mathrm{d}}/t_{\mathrm{c}}$ to make $\alpha$ harder than $-1.25$ also becomes wide as $t_{\mathrm{d}}/t_{\mathrm{c}}=2.2$-130 for $p=3$ case
($t_{\mathrm{d}}/t_{\mathrm{c}}=10$-80 for $p=2.5$).
As demonstrated in \citet{2014NatPh..10..351U}, we also test the continuous injection model, in which electrons with $f(\gamma,\mu)=\delta(\mu-\frac{1}{2})$ are injected continuously until $t=t_{\mathrm{d}}$. Even in this case, $t_{\mathrm{d}}/t_{\mathrm{c}}$
is required to be $\sim1$ to satisfy $\alpha>-1.25$. As shown in \citet{2014NatPh..10..351U}, the instantaneous spectrum of emissivity at $t=t_{d}$ can be as hard as $\alpha=-1.07$ for $t_{\mathrm{d}}/t_{\mathrm{c}}=10$. However, the time-integrated spectrum yields $\alpha=-1.42$.
\subsection{\bf Allowed Parameter Regions}
Free parameters in our model are $\Gamma, \gamma_{\rm{\mathrm{min}}}, \Delta\mu_{\mathrm{min}},k, B_{0}$ and $R_{0}$.
We will show the allowed parameter regions in Figure \ref{fig:regiongamma1000} and \ref{fig:region} for the isotropic and anisotropic cases.
From the following two conditions to be consistent with the observed GRB properties,
we obtain allowed parameter regions of $B_{0}$ and $R_{0}$ for given values of $\Gamma, \gamma_{\rm{\mathrm{min}}}, \Delta\mu_{\mathrm{min}}$,
and $k$.
(I) Peak energy
In the standard synchrotron model,
the peak energy of the observed ${\nu_{\rm{obs}} F_{\nu_{\rm{obs}}}}$ spectrum is
determined by $\gamma_{\rm min}$ as,
\begin{eqnarray}
\label{eq:Ep}
E_{\mathrm{p}}=h\nu_{\mathrm{p}}\simeq\Gamma\frac{3 h \gamma_{\mathrm{min}}^2eB_{0}}{4\pi m_{\mathrm{e}} c}.
\end{eqnarray}
However, as shown in Figure \ref{fig:spectra}, the peak energy
increases with decreasing $t_{\mathrm{d}}/t_{\mathrm{c}}$.
Given $\gamma_{\mathrm{min}}$, $t_{\mathrm{d}}/t_{\mathrm{c}}$ and the other parameters
for the anisotropy,
the peak energy is proportional to $B_0 \Gamma$.
From the results numerically obtained,
we define the allowed region that satisfies $100\mathrm{keV}<E_{\mathrm{p}}<1000\mathrm{keV}$,
(I\hspace{-.1em}I) Low energy spectral index $\alpha$
The key parameter that determines $\alpha$ is $t_{\mathrm{d}}/t_{\mathrm{c}}$.
The maximum $\alpha$ is $\alpha\sim-1.2$ ($-1.0$) for $p=2.5$ ($3.0$) and $B\propto R^{-1}$.
Considering the observation uncertainty,
we require $\alpha>-1.25$ to reproduce the typical spectral index.
\begin{table}[htp]
\caption{The lower and upper limits of $t_{\mathrm{d}}/t_{\mathrm{c}}$ that satisfy $\alpha>-1.25$ in the single $\mu$, isotropic and anisotropic cases with $t_{\rm end}=10t_{\mathrm{d}}$. The radiation efficiency $f_{\mathrm{rad}}$ is also shown in parentheses.}
\label{}
\centering
\begin{tabular}{l|l|l}
case & the lower limit & the upper limit \\ & of $t_{\mathrm{d}}/t_{\mathrm{c}}$ ($f_{\mathrm{rad}}$) & of $t_{\mathrm{d}}/t_{\mathrm{c}}$ ($f_{\mathrm{rad}}$) \\
\hline\hline
single $\mu$ & 0.30 (0.44)& 1.7 (0.75)\\
single $\mu$ steady injection & 0.36 (0.39) & 2.7 (0.71) \\
isotropic & 0.35 (0.43)& 1.8 (0.70)\\
$k=0.5,\Delta\mu_{\rm{min}}=0.01$ & 10.5 (0.48) & 110 (0.80)\\
$k=1,\Delta\mu_{\rm{min}}=0.01$ & 10 (0.55)& 80 (0.80)\\
$k=1.5,\Delta\mu_{\rm{min}}=0.01$ & 12 (0.63)& 50 (0.79)\\
$k=0.5,\Delta\mu_{\rm{min}}=0.1$ & 1.3 (0.47)& 10.5 (0.78)\\
$k=1,\Delta\mu_{\rm{min}}=0.1$ & 1.7 (0.58)& 6 (0.75)\\
$k=1.5,\Delta\mu_{\rm{min}}=0.1$ & 3.0 (0.69)& 3.5 (0.71)\\
$k=1,\Delta\mu_{\rm{min}}=0.01,t_{\mathrm{end}}=t_{\mathrm{d}}$ & 3.5 (0.39) & 310 (0.90) \\
$k=1,\Delta\mu_{\rm{min}}=0.01,p=3$ & 2.2 (0.21) & 130 (0.79) \\
\end{tabular}
\label{tab:tdtc}
\end{table}
In Table \ref{tab:tdtc}, we summarize the lower limit and upper limit of $t_{\mathrm{d}}/t_{\mathrm{c}}$
that satisfy $\alpha>-1.25$ for different cases.
The single-$\mu$ and isotropic cases require $t_{\mathrm{d}}/t_{\mathrm{c}} \sim 1$.
As the anisotropy strengthens, larger values of $t_{\rm d}$ compared to $t_{\rm c}$
are allowed.
Therefore, the fine tuning problem for $t_{\rm d}$ is relaxed in the anisotropic cases.
In Table \ref{tab:tdtc}, we also list the radiation efficiency, $\bm{f_{\mathrm{rad}}}$, defined as the energy ratio of the released photon energy to the initial total electron energy,
\begin{eqnarray}
&&f_{\mathrm{rad}} \nonumber \\ &&=\frac{\int_{0}^{\infty}d\nu E_{\nu}}{\int_{\gamma_{\rm{min}}}^{\gamma_{\rm{max}}} d\gamma \int_{-1}^{1}d\mu \gamma m_{\mathrm{e}}c^{2}N(\gamma,\mu,t=0)}.
\end{eqnarray}
The allowed parameter regions correspond to the marginally fast cooling, so that the radiation efficiency is significantly high as $\sim 0.2$--0.9.
In Figure \ref{fig:regiongamma1000}, we show examples of the allowed parameter regions
in the $R_0$-$B_0$ plane for several cases.
The lines that correspond to a constant value of $E_{\mathrm{p}}$ are curved.
In the fast cooling regime ($t_{\mathrm{d}}/t_{\mathrm{c}} \gg 1$),
the peak energy agrees with equation (\ref{eq:Ep}) so that the lines are horizontal
in Figure \ref{fig:regiongamma1000}.
On the other hand, in weak magnetic cases (slow cooling),
$E_{\mathrm{p}}$ increases with decreasing magnetic field as shown in Figure \ref{fig:spectra}.
This behaviour reflects on the curved feature of the constant $E_{\mathrm{p}}$ lines.
\begin{figure}[htp]
\centering
\includegraphics[width=8cm]{regioniso.pdf}
\includegraphics[width=8cm]{regionmu001k1.pdf}
\includegraphics[width=8cm]{regionmu001k1end.pdf}
\caption{The allowed parameter regions for $\Gamma=1000$ and $\gamma_{\mathrm{min}}=10^4$
are shown in the isotropic (top), anisotropic cases
with $t_{\rm end}=10t_{\mathrm{d}}$ (middle) and $t_{\rm end}=t_{\mathrm{d}}$ (bottom).
In the anisotropic cases, the parameters are $\Delta \mu_{\mathrm{min}}=0.01$, and $k=1$ .
The two red lines enclose the region for $\alpha>-1.25$,
and the two black lines enclose the region for $100\mathrm{keV}<E_{\mathrm{p}}<1000\mathrm{keV}$.
The blue lines show constant magnetic luminosities.
In the allowed regions, $10^{50}\mathrm{erg/s}<L_{B}<10^{51}\mathrm{erg/s}$, $10^{51}\mathrm{erg/s}<L_{B}<10^{52}\mathrm{erg/s}$ and $10^{52}\mathrm{erg/s}<L_{B}<10^{53}\mathrm{erg/s}$ are shown by magenta, orange and cyan areas, respectively.
The regions for $L_{B}<10^{50}\mathrm{erg/s}$ are shown by gray areas.
The green vertical line at $R_0=6\times 10^{13}\mathrm{cm}$ and the right edge $6\times 10^{16}\mathrm{cm}$ correspond to variability timescales $\delta t=R_0/(2 c \Gamma^2)=1\mathrm{ms}$ and $1\mathrm{s}$, respectively.}
\label{fig:regiongamma1000}
\end{figure}
In Figure \ref{fig:regiongamma1000}, we also plot the magnetic luminosity,
\begin{eqnarray}
\label{eq:LB}
L_B=4\pi cR_0^2\Gamma^2\frac{B_{0}^2}{8\pi}.
\end{eqnarray}
If the anisotropic distribution is caused by magnetic reconnection,
the energy source of the non-thermal electrons should be the magnetic energy.
In this case, the magnetic luminosity should be larger than or
comparable to the observed gamma-ray luminosity $L_{\mathrm{\gamma,iso}}$
(note $f_{\mathrm{rad}} \sim 1$ even in our model).
So we present the high magnetic field regions with colored areas in Figure \ref{fig:regiongamma1000}.
As we adopt high values of $\Gamma$ and $\gamma_{\rm min}$ in Figure \ref{fig:regiongamma1000},
the allowed regions are compatible with significantly high $L_B$.
In the case of $t_{\rm end}=t_{\mathrm{d}}$, the allowed region is significantly wider
than the cases with $t_{\rm end}=10 t_{\mathrm{d}}$.
\begin{figure}[htp]
\centering
\includegraphics[width=8cm]{regionmu001k1gm10000all.pdf}
\includegraphics[width=8cm]{regionmu001k1gm1000all.pdf}
\caption{The allowed parameter regions are shown for the anisotropic case of $\Delta \mu_{\mathrm{min}}=0.01$ and $k=1$(red lines) and the isotropic case (black lines)
for different values of $\Gamma$ and $\gamma_{\rm min}$ with $t_{\rm end}=10 t_{\mathrm{d}}$.
Upper panel: $\gamma_{\mathrm{min}}=10^4$ and Lower panel: $\gamma_{\mathrm{min}}=10^3$.
The magnetic luminosity is expressed with the same colors in Figure \ref{fig:regiongamma1000}.}
\label{fig:region}
\end{figure}
In Figure \ref{fig:region}, we show the allowed parameter regions in the $R_0$-$B_0$ plane for different values of $\Gamma$ and $\gamma_{\mathrm{min}}$.
If we require $L_{\rm{B}}>10^{50}\rm{ergs^{-1}}$ for the anisotropic model,
$\Gamma \gtrsim 1000$ and $\gamma_{\rm min} \sim 10^4$ are required.
From equations (\ref{eq:dynamicaltime}), (\ref{eq:coolingtime}), and (\ref{eq:Ep}),
the initial radius is rewritten as
\begin{eqnarray}
\label{eq:R}
R_{0}&=&10^{15}\left(\frac{t_{\rm d}/t_{\rm c}}{80}\right)\left(\frac{\Gamma}{1000}\right)^3\left(\frac{\gamma_{\mathrm{min}}}{10^4}\right)^3\nonumber\\&&\times\left(\frac{3E_{\mathrm{p}}}{300\mathrm{keV}}\right)^{-2}\mathrm{cm}.
\end{eqnarray}
The initial radius $R_{0}$ must be inside the deceleration radius of the afterglow, which is written as
\begin{eqnarray}
\label{eq:Rdec}
R_{\mathrm{dec}}&&=\left(\frac{3E_{\mathrm{iso}}}{4\pi n_{\mathrm{ex}}m_{\mathrm{p}}c^{2}\Gamma^{2}}\right)^{1/3}\nonumber\\
&&\simeq 10^{16}\left(\frac{E_{\mathrm{iso}}}{10^{52}\mathrm{erg}}\right)^{\frac{1}{3}}\left(\frac{n_{\mathrm{ex}}}{1\mathrm{\mathrm{cm^{-3}}}}\right)^{\frac{1}{3}}\nonumber\\&&\times\left(\frac{\Gamma}{1000}\right)^{-\frac{2}{3}}\mathrm{cm}.
\end{eqnarray}
where $m_{\mathrm{p}}$ is proton mass, $E_{\mathrm{iso}}$ is the total isotropic jet energy, $n_{\mathrm{ex}}$ is the ambient medium density.
If we take $\gamma_{\mathrm{min}}>10^5$ in equation (\ref{eq:R}), the allowed $R_0$
is larger than the deceleration radius of the afterglow around $10^{16}\mathrm{cm}$ \citep{2010ApJ...725.2209L}.
Therefore, only $\gamma_{\mathrm{min}}\sim 10^4$ is plausible for the anisotropic model
with magnetic reconnection.
For the isotropic models, a high value of $L_{\rm{B}}$ is not necessarily required.
In the standard internal shock model, in which the isotropic distribution is likely,
the typical value of $\gamma_{\rm min}$ is $\sim m_{\rm p}/m_{\rm e} \sim 10^3$
assuming that all electrons are accelerated.
As shown in the lower panel in Figure \ref{fig:region},
we find allowed parameter regions even for the isotropic distribution with $\gamma_{\rm min}=10^3$.
In this case, the magnetic luminosity is estimated to be lower than
$\sim10^{45}~\mbox{erg}~\mbox{s}^{-1}$ for $\Gamma=300$
or $\sim10^{48}~\mbox{erg}~\mbox{s}^{-1}$ for $\Gamma=1000$.
Such a lower magnetization may be favorable for particle acceleration
by relativistic shocks \citep[see e.g.][]{2020Galax...8...33V},
differently from the reconnection models.
However, the variability timescale $\delta t=R_0/(c \Gamma^2)$ should be shorter than
$\sim 0.03$ ms for $\Gamma=300$ or $0.1$ ms for $\Gamma=1000$, which could conflict
with the lower limit given by the light-crossing time for the Schwarzschild radius
of the central-engine black hole with a mass $M_{\rm BH}$,
$\sim 0.1 (M_{\rm BH}/10 M_\odot)$ ms.
The models with $\gamma_{\rm min}=10^4$ can avoid a too short $\delta t$
(shorter than 30 ms for $\Gamma=300$ or 100 ms for $\Gamma=1000$).
A large $\gamma_{\rm min}$ implies that a small fraction of electrons
are accelerated.
A too short $\delta t$ also leads to absorption of gamma-rays in
the source via electron--positron pair creation.
A simple estimate of the optical depth for $\gamma \gamma$-absorption
for photons of energy $E$
\citep{2003PASJ...55..433A,2009Sci...323.1688A} gives
\begin{eqnarray}
\tau_{\gamma \gamma}\simeq 0.1 \Gamma^{2 \beta-2}
\frac{(-\beta-2)\sigma_{\rm T}L_{\rm iso}}{16 \pi E_{\rm p}c^2 \delta t}
\left( \frac{E E_{\rm p}}{m_{\rm e}^2 c^4} \right)^{-\beta-1},
\end{eqnarray}
where $L_{\rm iso}$ is the isotropically equivalent luminosity.
To avoid absorption of MeV photons, a lower limit for the variability timescale
is given as
\begin{eqnarray}
\delta t > 0.9 \left(\frac{\Gamma}{100}\right)^{-6.4}
\left(\frac{L_{\rm iso}}{10^{52}\mbox{erg}~\mbox{s}^{-1}}\right)^{-1}
\left(\frac{E_{\rm p}}{100\mbox{keV}}\right)^{0.2}
\mbox{ms},\nonumber \\
\end{eqnarray}
where we adopt $\beta=-2.2$.
Only the isotropic model with $\Gamma=100$ and $\gamma_{\rm min}=10^3$
in Figure \ref{fig:region} conflicts the above lower limit.
\section{Conclusions \& Discussion} \label{sec:sum}
In this paper, we have calculated the GRB synchrotron spectra with the anisotropic electron distribution,
which is motivated by the PIC simulations of magnetic reconnection in \citet{2019ApJ...886..122C}.
For that sake, we have followed the evolution of the electron distribution in the energy and pitch angle space
taking into account the synchrotron and adiabatic losses.
Even with the anisotropic distribution,
the obtained spectra can be well fitted with the standard Band function.
Small pitch angles in the anisotropic electron distribution suppress the synchrotron cooling,
which makes synchrotron spectra harder.
Compared to the isotropic distribution, even with a stronger magnetic field,
the inefficient cooling can produce a significantly hard spectrum below the peak energy $E_{\rm p}$.
Given the parameters of $\Gamma$ and $\gamma_{\rm min}$,
allowed parameter regions of $R_0$ and $B_0$ becomes wider in the anisotropic model.
In our conservative assumption for the magnetic field evolution of $B\propto R^{-1}$,
the late radiation at $t>t_{\rm d}$ makes the low-energy spectral index $\alpha\sim-1.2$.
If we seriously consider the difference between the obtained index and the
typical observed index $\alpha\sim-1$,
a rapid decrease of the magnetic field or significant electron escape effect
after the dynamical timescale is required.
If the anisotropic electron injection is due to magnetic reconnection,
the magnetic luminosity should be comparable to or larger than the observed gamma-ray luminosity.
This condition requires the bulk Lorentz factor $\Gamma>1000$, because
the magnetic luminosity is expressed as
\begin{eqnarray}
\label{eq:LBpa}
L_{B}&=&2\times10^{51}\left(\frac{t_{\rm d}/t_{\rm c}}{80}\right)^2\left(\frac{\Gamma}{1000}\right)^6\left(\frac{\gamma_{\mathrm{min}}}{10^4}\right)^2\nonumber\\&&\times\left(\frac{3E_{\mathrm{p}}}{300\mathrm{keV}}\right)^{-2}~\mbox{erg}~\mbox{s}^{-1},
\end{eqnarray}
from equations (\ref{eq:dynamicaltime}), (\ref{eq:coolingtime}), (\ref{eq:Ep}), and (\ref{eq:LB}), where the factor $3$ before $E_{\mathrm{p}}$ is the difference of the peak energy of the anisotropic case from the isotropic fast cooling case (see Figure \ref{fig:spectra}).
This also suggests that $\gamma_{\rm min} \gtrsim 10^4$ is simultaneously required.
There are allowed parameter regions for the isotropic electron distribution as well.
In this case, the standard shock acceleration model seems relevant.
The magnetic luminosity in the allowed region for $\Gamma=300$ and $\gamma_{\mathrm{min}}=10^4$ is around $10^{45}$--$10^{47}~\mbox{erg}~\mbox{s}^{-1}$,
which is consistent with the required low-magnetization for the efficient
particle acceleration by relativistic shocks \citep[e.g.][]{2013ApJ...771...54S,2020Galax...8...33V}.
However, to maintain the variability timescale long enough,
$\gamma_{\mathrm{min}}\gtrsim 10^4$ is required.
In this paper, we have neglected the effect of pitch angle diffusion by turbulence for simplicity.
The pitch angle scattering rate in the Alfv\'{e}n turbulence is given by
\begin{eqnarray}
\label{eq:scatteringrate}
\nu_{\rm{sca}}=\frac{\pi}{4}\left(\frac{k_{\rm res} E(k_{\rm res})}{{B_{0}}^2/{8\pi}}\right)\Omega ,
\end{eqnarray}
where $E(k)$ is the power spectrum of the turbulence,
$k_{\rm res} \equiv \Omega/(c \mu)$ is the resonance wavenumber,
and $\Omega=eB_{0}/(\gamma m_{e}c)$ is the gyro frequency of an electron \citep{1987PhR...154....1B}.
Here, we assume a power spectrum of the Kolmogorov type turbulence as
\begin{eqnarray}
\label{eq:powerspectrum}
kE(k)\equiv\frac{{\delta B_{k}}^2}{8\pi}=\frac{{\delta B}^2}{8\pi}{\left(\frac{k}{k_{\mathrm{inj}}}\right)}^{-\frac{2}{3}},
\end{eqnarray}
where $\delta B$ is the the turbulent component of the magnetic field strength
at the injection wavenumber of the turbulence $k_{\mathrm{inj}}$,
for which we adopt the inverse of the comoving width of the emission region as
$k_{\mathrm{inj}}=\Gamma/R_{0}$,
which is the possible lowest wavenumber.
The condition to maintain the anisotropic distribution
is that the timescale of pitch angle scattering should be longer than $t_{\rm{d}}$;
\begin{eqnarray}
\label{eq:scatteringcondition}
\frac{1}{\nu_{\rm{sca}}}>\frac{R_{0}}{c \Gamma}.
\end{eqnarray}
This condition is written as
\begin{eqnarray}
\label{eq:turbulencestrength}
\frac{\delta B}{B_{0}}&<&
{\left(\frac{4}{\pi}\right)}^{\frac{1}{2}}{\left(\frac{\Gamma r_{\mathrm{L}}}{R_0}\right)}^{\frac{1}{6}}{\left(\frac{1}{\mu}\right)}^{\frac{1}{3}}\nonumber\\
&\simeq&0.06\left(\frac{\Gamma}{1000}\right)^{\frac{1}{3}}\left(\frac{L_{B}}{10^{52} ~\mbox{erg}~\mbox{s}^{-1}}\right)^{-\frac{1}{12}}\nonumber\\&&\times\left(\frac{\gamma}{10^4}\right)^{\frac{1}{6}}{\left(\frac{1}{\mu}\right)}^{\frac{1}{3}},
\end{eqnarray}
where $r_{\mathrm{L}}=\gamma m_{e}c^2/(eB_{0})$ is the Larmor radius of an electron.
The obtained upper bound $\delta B/B_0 \sim 0.1$ seems reasonably large,
though the upper limit becomes smaller for a smaller injection scale,
which is determined by the reconnection layer.
Radiation from an anisotropic electron distribution could have a significant circular polarization degree.
When a circular polarization was detected in the optical afterglow of GRB 121024A, an anisotropic electron distribution was considered \citep{2014Natur.509..201W}.
The circular polarization degree for an anisotropic electron distribution $N(\gamma,\mu)\propto\gamma^{-p}f(\gamma,\mu)$ \citep{1969SvA....13..396S} is given by
\begin{eqnarray}
\label{eq:polarization}
P_{\mathrm{cir}}&=&\frac{1}{\gamma_{\mathrm{min}}}\frac{(2+p)\cot\theta-\sin\theta g(\gamma_{\mathrm{min}},\theta)}{p}\frac{p+1}{p+\frac{7}{3}}\nonumber\\& &\times\frac{\Gamma\left(\frac{3p+8}{12}\right)\Gamma\left(\frac{3p+4}{12}\right)}{\Gamma\left(\frac{3p+7}{12}\right)\Gamma\left(\frac{3p-1}{12}\right)},
\end{eqnarray}
where $\theta$ is the viewing angle from the magnetic field direction in the comoving frame, which is equivalent to the pitch angle of electrons that mostly contributes to radiation
for the observer.
Writing $\mu=\cos\theta$, the function $g$ is written as
\begin{eqnarray}
\label{eq:polarization2}
g(\gamma_{\mathrm{min}},\theta)=\left(\frac{1}{f(\gamma,\mu)}\frac{\partial f(\gamma,\mu)}{\partial \mu}\right)_{\gamma=\gamma_{\mathrm{min}}},
\end{eqnarray}
and $\Gamma$ in equation (\ref{eq:polarization}) is the Gamma function.
From order estimate for the anisotropic pitch angle distribution in our model with equation (\ref{eq:pitchdistribution}) using the typical viewing angle ($\theta \simeq 0.14$), we obtain $P_{\mathrm{cir}}\sim1/\gamma_{\mathrm{min}} \sim 0.01\%$ for $\Delta\mu_{\mathrm{min}}=0.01$ and $\gamma_{\mathrm{min}}=10^4$.
The predicted polarization degree is only a few times larger than that for the isotropic case, which corresponds to $g(\gamma_{\rm min},\theta)\to 0$
in equation (\ref{eq:polarization2}).
\begin{acknowledgements}
First we appreciate the referee for the very helpful suggestions in spite of his/her difficult time.
We thank Kosuke Nishiwaki, Tomohisa Kawashima and Kyohei Kawaguchi for useful comments and discussion. R.G. received helpful advice for numerical calculation from Naotaka Yoshinaga and Chinatsu Watanabe and would like to thank them.
R.G. acknowledges the support by the Forefront
Physics and Mathematics Program to Drive Transformation (FoPM).
This work is supported by the joint research program of the Institute for Cosmic Ray Research (ICRR), the University of Tokyo.
Numerical computations were in part
carried out on FUJITSU Server PRIMERGY CX2550 M5 at ICRR.
\end{acknowledgements}
\section{Abstract}
We examine whether slow synchrotron cooling due to small pitch angles can solve the fast cooling problem in Gamma Ray Burst.
Conventional models have assumed isotropic pitch angle distribution of electrons when calculating cooling of electrons.
A recent PIC simulation showed that low energy electrons accelerated by magnetic reconnection in the magnetically dominated regime have small pitch angles.
We focus on the slow synchrotron cooling of the electrons with the small pitch angles and calculate the cooling of the electrons and the resultant synchrotron spectra.
We show that if accelerated low energy electrons have small pitch angles $\sin \psi\sim0.1$, the case of bulk lorentz factor $\Gamma\sim1000$ and electrons minimum random lorentz factor $\gamma_{\rm{min}}\sim10^4$ can reproduce typical GRBs' spectra with low energy spectral index $\alpha\sim-1$.
\section{Introduction}
Gamma Ray Burst (GRB) is considered to be the radiation from a cosmological relativistic jet.
GRB prompt spectra are usually well fit by the Band function(Band et al. 1993). This function is smoothly broken power law with low energy spectral index $\alpha$(typically $\sim-1$) and high energy spectral index $\beta$(typically $\sim-2.2$) and the peak of the $\nu F_{\nu}$ spectrum $E_{\mathrm{p}}$(typically several hundreds $\mathrm{keV}$) (Preece et al. 2000).
Synchrotron radiation from electrons with a broken power law energy distribution has been considered to be a possible radiative process to explain the GRBs' spectra.
But there exists a crutial problem in this interpretation.
Strong magnetic field is required to reproduce $E_{\mathrm{p}}$ around several hundreds $\mathrm{keV}$.
Radiation from electrons cooled by the strong magnetic field appear in low energy spectra.
The predicted spectra have $\alpha=-1.5$, which is too softer than the typically observed $\alpha\sim-1$.
This problem is known as the synchrotron fast cooling problem in GRBs(Ghisellini et al. 2000).
Alternative models such as the photosphere model (M\'esz\'aros \& Rees 2000) and the synchrotron model which takes into account the Klein-Nishina effect on synchrotron self-Compton process(Derishev et al. 2001) have been proposed.
In this paper, we consider about another solution of the problem, which is slow cooling due to small pitch angles(where pitch angle means the angle between a magnetic field and the velocity of a particle).
In a recent PIC simulation(Comisso et al. 2019), a particle acceleration process by magnetic reconnection in magnetically dominant regime was studied.
They showed that firstly electrons are accelerated along the guiding magnetic field by magnetic reconnection and then accelerated perpendicularly to the field by turbulence.
By this process, accelerated low energy electrons have small pitch angles.
A small pitch angle means a small perpendicular momentum to the magnetic field, so the synchrotron cooling effect is expected to be suppressed.
In this paper, we focus on the possibility that the slow synchrotron cooling due to the small pitch angles solves the fast cooling problem.
The structure of this paper is as follows.
In section 3, we explain our calculation method of synchrotron spectra taking into account the anisotropic electron distribution.
In section 4, we show the numerical calculation results of spectra and explain the results analytically.
And we show the parameter spaces allowed by GRBs' observations.
In section 5, we discuss the possibility of solving the fast cooling problem by the small pitch angles.
\section{Method}
\subsection{the calculation method of the time evolution of electrons distribution}
Our synchrotron model assumes nonthermal power law accelerated electrons as initial condition. Then we calculate the time evolution of the electron distribution by solving the continuity equation.
Considering synchrotron cooling and adiabatic cooling in magnetic field decreasing with jet expansion, the continuity equation is as follows,
\begin{eqnarray}
\frac{\partial}{\partial t}(\frac{dN}{d\gamma d\mu})
+\frac{\partial}{\partial \gamma}\{(\dot{\gamma}_{\rm syn}+\dot{\gamma}_{\rm adi})\frac{dN}{d\gamma d\mu}\}
+\frac{\partial}{\partial \mu}\{(\dot{\mu}_{\rm syn}+\dot{\mu}_{\rm adi})\frac{dN}{d\gamma d\mu}\}=0,
\end{eqnarray}
where $t$ is the time in jet's comoving frame,
$\gamma$ is a electron's random lorentz factor,
$\mu=\cos \psi$ ($\psi$ is a pitch angle of an electron),
$\frac{dN}{d\gamma d\mu}$ is electrons' energy and pitch angle distribution, $\dot{\gamma}_{\rm syn}$ is the synchrotron cooling rate(Eq. (2)), $\dot{\mu}_{\rm syn}$ is the synchrotron pitch angle changing rate(Eq. (3)), $\dot{\gamma}_{\rm adi}$ is the adiabatic cooling rate(Eq. (5)) and $\dot{\mu}_{\rm adi}$ is the adiabatic pitch angle changing rate(Eq. (6)).
The synchrotron cooling rate(Rybicki \& Lightman(1979)) is
\begin{eqnarray}
\dot{\gamma}_{\rm syn}=-\frac{2q^4B^2\gamma^2\beta_{\mathrm{e}}^2(1-\mu^2)}{3m_{\mathrm{e}}^3c^5},
\end{eqnarray}
where $q$ is the electron charge, $m_e$ is the electron mass, $c$ is the velocity of light, $B$ is the magnetic field strength, $\beta_e$ is a velocity of an electron in unit of $c$.
The synchrotron pitch angle changing rate by radiation reaction(Singal(2016)) is
\begin{eqnarray}
\dot{\mu}_{\rm syn}=\frac{2q^4B^2\mu(1-\mu^2)}{3m_{\mathrm{e}}^3c^5\gamma}.
\end{eqnarray}
Here, we assume that magnetic field decreases by a jet's expansion as $B\propto r^{-1}$ (where $r$ is the jet's position) at the radiation region distant from the central engine(Uhm \& Zhang 2014), where toroidal magnetic field $B_{t}\propto r^{-1}$ is more dominant than poloidal magnetic field $B_{p}\propto r^{-2}$,
\begin{eqnarray}
B=B_{0}\left(\frac{R_{0}}{r}\right)=B_{0}\left(\frac{R_{0}}{R_{0}+\Gamma \beta ct}\right).
\end{eqnarray}
Here, after the jet reached the radiation radius $R_{0}$ at a time $t=0\mathrm{s}$, where the magnetic field strength is $B_{0}$, the jet expands as $r=R_{0}+\Gamma \beta ct$, where $\beta=\frac{v}{c}$, $v$ is the jet's velocity in the laboratory frame, $\Gamma=\frac{1}{\sqrt{1-\beta^{2}}}$ is the jet's bulk lorentz factor, which is needed because of the relativistic time dilation effect.
From Equation(4) and the adiabatic invariant $\frac{p_{\perp}^2}{B}$(where $p_{\perp}$ is the electron's perpendicular momentum to the magnetic field)(Northrop 1963), we can obtain the adiabatic cooling rate
\begin{eqnarray}
\dot{\gamma}_{\rm adi}=-\frac{(\gamma^2-1)(1-\mu^2)}{2\gamma}\frac{\Gamma\beta c}{R_0+\Gamma \beta ct}
\end{eqnarray}
and the adiabatic pitch angle changing rate
\begin{eqnarray}
\dot{\mu}_{\rm adi}=\frac{\mu(1-\mu^2)}{2}\frac{\Gamma\beta c}{R_0+\Gamma \beta ct}.
\end{eqnarray}
From equation (4), after the dynamical timescale defined by $t_d=\frac{R_{0}}{\Gamma \beta c}$, magnetic field $B$ is attenuated below the half of the original $B_{0}$.
Thereafter, synchrotron radiation is suppressed and electrons mainly cool by adiabatic cooling and the calculated synchrotron spectra don't change significantly.
Therefore, we stop calculation at the time $t=10t_d$ which is large enough to obtain the low energy photon spectral index $\alpha$.
\subsection{calculation method of synchrotron spectrum}
We calculate time-integrated synchrotron spectra to compare with observed spectra because the spectra observed in GRBs are usually time-integrated due to poor photon statistics.
We integrate synchrotron radiation from the time $t=0\rms$ when the jet reached the radiation region to $t=10t_d$ when the magnetic field decreases enough to suppress the synchrotron radiation(Rybicki \& Lightman(1979)),
\begin{eqnarray}
\nu F_{\nu}=\int_0^{10t_{\rm{d}}} dt \int_1^{\gamma_{\rm{max}}} d\gamma \int_{-1}^{1}d\mu \frac{dN}{d\gamma d\mu}\frac{\sqrt{3}\nu q^{3}B\sin \psi}{m_{\mathrm{e}} c^2}F\left(\frac{\nu}{\nu_{\mathrm{syn}}}\right)
\end{eqnarray}
where $\nu$ is the emitted photons' frequency in the jet's comoving frame, $\nu_{\rm{syn}}=\frac{3\gamma^{2}qB\sin \psi}{4\pi m_{e}c}$ is the synchrotron characteristic frequency, $F\left(\frac{\nu}{\nu_{\rm{syn}}}\right)=\frac{\nu}{\nu_{\rm{syn}}} \int_{\frac{\nu}{\nu_{\rm{syn}}}}^{\infty} K_{\frac{5}{3}}(\xi)d\xi$
and $K_{\frac{5}{3}}(\xi)$ is the modified bessel function of the $\frac{5}{3}$ order.
In the following discussion, we focus on only spectral forms, in particular on the photon spectral index $\alpha$. So, we will use $\widetilde{\nu F_{\nu}}$ that is the $\nu F_{\nu}$ normalized by the total energy which accelerated electrons had,
\begin{eqnarray}
\widetilde{\nu F_{\nu}}=\frac{\nu F_{\nu}}{\int_{\gamma_{\rm{min}}}^{\gamma_{\rm{max}}} d\gamma \int_{-1}^{1}d\mu \frac{dN}{d\gamma d\mu}}_{|t=0}.
\end{eqnarray}
Moreover, to compare the calculated spectra $\widetilde{\nu F_{\nu}}$ with observed spectra, we boost $\widetilde{\nu F_{\nu}}$ in the jet's comoving frame to $\widetilde{\nu_{\rm{obs}} F_{\nu_{\rm{obs}}}}$ in the observer frame by the relation $\nu_{\rm{obs}}=2\Gamma \nu$, where the factor $2\Gamma$ is derived from $\nu_{\rm{obs}}=\frac{1}{\Gamma(1-\beta \cos \theta)}\nu$ and $\theta\simeq0$ from which main radiation comes by the beaming effect, where $\theta$ is the angle from the line of sight.
\subsection{the initial condition in the isotropic electron distribution}
We assume the initial condition of the isotropic electron distribution as follows,
\begin{eqnarray}
\frac{dN}{d\gamma d\mu}_{|t=0}=C\gamma^{-p}\ (\gamma_{\rm{min}}<\gamma<\gamma_{\rm{max}}),
\end{eqnarray}
where $C$ is the normalization constant whose value can be selected arbitrarily because we use normalized $\widetilde{\nu F_{\nu}}$, $\gamma_{\mathrm{min}}$ and $\gamma_{\mathrm{max}}$ are the minimum and maximum electron lorentz factor, respectively.
Through our entire paper, we take the electron spectral index $p=2.5$ to reproduce typically observed high energy spectral index $\beta\sim-2.2$ because $\beta=-\frac{p+2}{2}$ in the fast cooling synchrotron spectrum.
\subsection{the initial condition of the anisotropic electron distribution with small pitch angles}
We assume the initial condition of the anisotropic electron distribution with small pitch angles as Gaussian pitch angle distribution,
\begin{eqnarray}
\frac{dN}{d\gamma d\mu}_{|t=0}=C\gamma^{-p}\exp\left(-\frac{(\mu-\overline{\mu})^2}{2\Delta\mu^2}\right)\ (\gamma_{\rm{min}}<\gamma<\gamma_{\rm{max}}),
\end{eqnarray}
where we assume that the peak pitch angle cosine,
\begin{eqnarray}
\overline{\mu}=1,
\end{eqnarray}
and that the standard deviation of the pitch angle's Gaussian distribution $\Delta\mu$ is energy dependent with constant parameters $\Delta\mu_{min}$ and $k$,
\begin{eqnarray}
\Delta\mu=\Delta\mu_{\rm{min}}\left(\frac{\gamma}{\gamma_{\rm{min}}}\right)^k.
\end{eqnarray}
This pitch angle distribution qualitatively reproduces the distribution of the electrons accelerated by magnetic reconnection in the PIC simulation of Comisso et al.(2019). Namely, most of the low energy electrons have small pitch angles $\cos \psi\sim1$ and high energy electrons are almost isotropic with $\Delta\mu\geq1$.
\subsection{the allowed parameter region by the GRB observations}
Free parameters in our model are $\Gamma, \gamma_{\rm{\mathrm{min}}}, \Delta\mu_{\mathrm{min}},k, B$ and $R$.
In this paper, from the following three conditions which are needed to reproduce the GRB observations, we obtain allowed parameter regions of $B$ and $R$ by fixing the values of $\Gamma, \gamma_{\rm{\mathrm{min}}}, \Delta\mu_{\mathrm{min}},k$.
\\
(1) the typical low energy spectral index $\alpha\sim-1$\\
If low energy electrons don't cool significantly before the magnetic field is attenuated, the typically observed spectral index $\alpha\sim-1$ is reproduced. This condition is that the cooling timescale of the low energy electrons $t_c (\gamma_{min})$ is comparable to the dynamical timescale $t_d$.
Here we define the synchrotron cooling timescale of the lowest energy electrons of the isotropic distribution as
\begin{eqnarray}
t_{\mathrm{c,iso}}=\frac{\gamma_{\mathrm{min}}}{\dot{\gamma}_{\mathrm{syn,iso}}(\gamma_{\mathrm{min}})}=\frac{9 m_{\mathrm{e}}^3 c^5}{4 q^4 B^2 \gamma_{min}},
\end{eqnarray}
and the dynamical timescale is
\begin{eqnarray}
t_{\mathrm{d}}=\frac{R_0}{\Gamma \beta c}.
\end{eqnarray}
We consider $\alpha=-1.2$ as the lower limit of the low energy spectral index $\alpha$ of the typical GRBs and we determine a ratio, $t_d/t_{c,iso}$ such that the calculated spectrum becomes $\alpha=-1.2$ as the upper limit of $t_d/t_{c,iso}$ in the allowed parameter region.
\\
(2) the typical $\nu F_{\nu}$ spectrum peak energy $100\mathrm{keV}<E_{\mathrm{p}}<1000\mathrm{keV}$\\
The peak energy of the observed $\widetilde{\nu_{\rm{obs}} F_{\nu_{\rm{obs}}}}$ spectrum is
\begin{eqnarray}
E_{\mathrm{p}}=2\Gamma\frac{3h\gamma_{\mathrm{min}}^2qB\sin \psi}{4\pi m_{\mathrm{e}} c}.
\end{eqnarray}
So, the allowed region of the magnetic field strength is determined from the observational constraint $100\mathrm{keV}<E_{\mathrm{p}}<1000\mathrm{keV}$.
\\
(3) the magnetic luminosity $10^{52}\mathrm{erg/s}<L_B<10^{54}\mathrm{erg/s}$\\
The magnetic luminosity $L_B$ must be comparable to the observed luminosity $L_{\mathrm{obs}}$ to explain GRB by the radiation from the electrons accelerated by magnetic reconnection.
Here, the magnetic luminosity is
\begin{eqnarray}
L_B=4\pi cR_0^2\Gamma^2\frac{B_0^2}{8\pi}.
\end{eqnarray}
\\
From the observed luminosity $10^{52}\mathrm{erg/s}<L_{\mathrm{obs}}<10^{54}\mathrm{erg/s}$, the allowed region is $10^{52}\mathrm{erg/s}<L_B<10^{54}\mathrm{erg/s}$.
\section{Result}
\section{Conclusion \& Discussion}
\bibliographystyle{plain}
\subsubsection*{#1}}
\pagestyle{headings}
\markright{Reference sheet: \texttt{natbib}}
\usepackage{shortvrb}
\MakeShortVerb{\|}
\begin{document}
\thispagestyle{plain}
\newcommand{\textsc{Bib}\TeX}{\textsc{Bib}\TeX}
\newcommand{\texttt{#1}\def\filedate{#2}\def\fileversion{#3}}}{\texttt{#1}\def\filedate{#2}\def\fileversion{#3}}}
\begin{center}{\bfseries\Large
Reference sheet for \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\ usage}\\
\large(Describing version \fileversion\ from \filedate)
\end{center}
\begin{quote}\slshape
For a more detailed description of the \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\ package, \LaTeX\ the
source file \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\texttt{.dtx}.
\end{quote}
\head{Overview}
The \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\ package is a reimplementation of the \LaTeX\ |\cite| command,
to work with both author--year and numerical citations. It is compatible with
the standard bibliographic style files, such as \texttt{plain.bst}, as well as
with those for \texttt{harvard}, \texttt{apalike}, \texttt{chicago},
\texttt{astron}, \texttt{authordate}, and of course \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}.
\head{Loading}
Load with |\usepackage[|\emph{options}|]{|\texttt{#1}\def\filedate{#2}\def\fileversion{#3}}|}|. See list of
\emph{options} at the end.
\head{Replacement bibliography styles}
I provide three new \texttt{.bst} files to replace the standard \LaTeX\
numerical ones:
\begin{quote}\ttfamily
plainnat.bst \qquad abbrvnat.bst \qquad unsrtnat.bst
\end{quote}
\head{Basic commands}
The \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\ package has two basic citation commands, |\citet| and
|\citep| for \emph{textual} and \emph{parenthetical} citations, respectively.
There also exist the starred versions |\citet*| and |\citep*| that print
the full author list, and not just the abbreviated one.
All of these may take one or two optional arguments to add some text before
and after the citation.
\begin{quote}
\begin{tabular}{l@{\quad$\Rightarrow$\quad}l}
|\citet{jon90}| & Jones et al. (1990)\\
|\citet[chap.~2]{jon90}| & Jones et al. (1990, chap.~2)\\[0.5ex]
|\citep{jon90}| & (Jones et al., 1990)\\
|\citep[chap.~2]{jon90}| & (Jones et al., 1990, chap.~2)\\
|\citep[see][]{jon90}| & (see Jones et al., 1990)\\
|\citep[see][chap.~2]{jon90}| & (see Jones et al., 1990, chap.~2)\\[0.5ex]
|\citet*{jon90}| & Jones, Baker, and Williams (1990)\\
|\citep*{jon90}| & (Jones, Baker, and Williams, 1990)
\end{tabular}
\end{quote}
\head{Multiple citations}
Multiple citations may be made by including more than one
citation key in the |\cite| command argument.
\begin{quote}
\begin{tabular}{l@{\quad$\Rightarrow$\quad}l}
|\citet{jon90,jam91}| & Jones et al. (1990); James et al. (1991)\\
|\citep{jon90,jam91}| & (Jones et al., 1990; James et al. 1991)\\
|\citep{jon90,jon91}| & (Jones et al., 1990, 1991)\\
|\citep{jon90a,jon90b}| & (Jones et al., 1990a,b)
\end{tabular}
\end{quote}
\head{Numerical mode}
These examples are for author--year citation mode. In numerical mode, the
results are different.
\begin{quote}
\begin{tabular}{l@{\quad$\Rightarrow$\quad}l}
|\citet{jon90}| & Jones et al. [21]\\
|\citet[chap.~2]{jon90}| & Jones et al. [21, chap.~2]\\[0.5ex]
|\citep{jon90}| & [21]\\
|\citep[chap.~2]{jon90}| & [21, chap.~2]\\
|\citep[see][]{jon90}| & [see 21]\\
|\citep[see][chap.~2]{jon90}| & [see 21, chap.~2]\\[0.5ex]
|\citep{jon90a,jon90b}| & [21, 32]
\end{tabular}
\end{quote}
\head{Suppressed parentheses}
As an alternative form of citation, |\citealt| is the same as |\citet| but
\emph{without parentheses}. Similarly, |\citealp| is |\citep| without
parentheses. Multiple references, notes, and the starred variants
also exist.
\begin{quote}
\begin{tabular}{l@{\quad$\Rightarrow$\quad}l}
|\citealt{jon90}| & Jones et al.\ 1990\\
|\citealt*{jon90}| & Jones, Baker, and Williams 1990\\
|\citealp{jon90}| & Jones et al., 1990\\
|\citealp*{jon90}| & Jones, Baker, and Williams, 1990\\
|\citealp{jon90,jam91}| & Jones et al., 1990; James et al., 1991\\
|\citealp[pg.~32]{jon90}| & Jones et al., 1990, pg.~32\\
|\citetext{priv.\ comm.}| & (priv.\ comm.)
\end{tabular}
\end{quote}
The |\citetext| command
allows arbitrary text to be placed in the current citation parentheses.
This may be used in combination with |\citealp|.
\head{Partial citations}
In author--year schemes, it is sometimes desirable to be able to refer to
the authors without the year, or vice versa. This is provided with the
extra commands
\begin{quote}
\begin{tabular}{l@{\quad$\Rightarrow$\quad}l}
|\citeauthor{jon90}| & Jones et al.\\
|\citeauthor*{jon90}| & Jones, Baker, and Williams\\
|\citeyear{jon90}| & 1990\\
|\citeyearpar{jon90}| & (1990)
\end{tabular}
\end{quote}
\head{Forcing upper cased names}
If the first author's name contains a \textsl{von} part, such as ``della
Robbia'', then |\citet{dRob98}| produces ``della Robbia (1998)'', even at the
beginning of a sentence. One can force the first letter to be in upper case
with the command |\Citet| instead. Other upper case commands also exist.
\begin{quote}
\begin{tabular}{rl@{\quad$\Rightarrow$\quad}l}
when & |\citet{dRob98}| & della Robbia (1998) \\
then & |\Citet{dRob98}| & Della Robbia (1998) \\
& |\Citep{dRob98}| & (Della Robbia, 1998) \\
& |\Citealt{dRob98}| & Della Robbia 1998 \\
& |\Citealp{dRob98}| & Della Robbia, 1998 \\
& |\Citeauthor{dRob98}| & Della Robbia
\end{tabular}
\end{quote}
These commands also exist in starred versions for full author names.
\head{Citation aliasing}
Sometimes one wants to refer to a reference with a special designation,
rather than by the authors, i.e. as Paper~I, Paper~II. Such aliases can be
defined and used, textual and/or parenthetical with:
\begin{quote}
\begin{tabular}{lcl}
|\defcitealias{jon90}{Paper~I}|\\
|\citetalias{jon90}| & $\Rightarrow$ & Paper~I\\
|\citepalias{jon90}| & $\Rightarrow$ & (Paper~I)
\end{tabular}
\end{quote}
These citation commands function much like |\citet| and |\citep|: they may
take multiple keys in the argument, may contain notes, and are marked as
hyperlinks.
\head{Selecting citation style and punctuation}
Use the command |\bibpunct| with one optional and 6 mandatory arguments:
\begin{enumerate}
\item the opening bracket symbol, default = (
\item the closing bracket symbol, default = )
\item the punctuation between multiple citations, default = ;
\item the letter `n' for numerical style, or `s' for numerical superscript
style, any other letter for
author--year, default = author--year;
\item the punctuation that comes between the author names and the year
\item the punctuation that comes between years or numbers when common author
lists are suppressed (default = ,);
\end{enumerate}
The optional argument is the character preceding a post-note, default is a
comma plus space. In redefining this character, one must include a space if
one is wanted.
Example~1, |\bibpunct{[}{]}{,}{a}{}{;}| changes the output of
\begin{quote}
|\citep{jon90,jon91,jam92}|
\end{quote}
into [Jones et al. 1990; 1991, James et al. 1992].
Example~2, |\bibpunct[; ]{(}{)}{,}{a}{}{;}| changes the output of
\begin{quote}
|\citep[and references therein]{jon90}|
\end{quote}
into (Jones et al. 1990; and references therein).
\head{Other formatting options}
Redefine |\bibsection| to the desired sectioning command for introducing
the list of references. This is normally |\section*| or |\chapter*|.
Define |\bibpreamble| to be any text that is to be printed after the heading but
before the actual list of references.
Define |\bibfont| to be a font declaration, e.g.\ |\small| to apply to
the list of references.
Define |\citenumfont| to be a font declaration or command like |\itshape|
or |\textit|.
Redefine |\bibnumfmt| as a command with an argument to format the numbers in
the list of references. The default definition is |[#1]|.
The indentation after the first line of each reference is given by
|\bibhang|; change this with the |\setlength| command.
The vertical spacing between references is set by |\bibsep|; change this with
the |\setlength| command.
\head{Automatic indexing of citations}
If one wishes to have the citations entered in the \texttt{.idx} indexing
file, it is only necessary to issue |\citeindextrue| at any point in the
document. All following |\cite| commands, of all variations, then insert
the corresponding entry to that file. With |\citeindexfalse|, these
entries will no longer be made.
\head{Use with \texttt{chapterbib} package}
The \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\ package is compatible with the \texttt{chapterbib} package
which makes it possible to have several bibliographies in one document.
The package makes use of the |\include| command, and each |\include|d file
has its own bibliography.
The order in which the \texttt{chapterbib} and \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\ packages are loaded
is unimportant.
The \texttt{chapterbib} package provides an option \texttt{sectionbib}
that puts the bibliography in a |\section*| instead of |\chapter*|,
something that makes sense if there is a bibliography in each chapter.
This option will not work when \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\ is also loaded; instead, add
the option to \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}.
Every |\include|d file must contain its own
|\bibliography| command where the bibliography is to appear. The database
files listed as arguments to this command can be different in each file,
of course. However, what is not so obvious, is that each file must also
contain a |\bibliographystyle| command, \emph{preferably with the same
style argument}.
\head{Sorting and compressing citations}
Do not use the \texttt{cite} package with \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}; rather use one of the
options \texttt{sort} or \texttt{sort\&compress}.
These also work with author--year citations, making multiple citations appear
in their order in the reference list.
\head{Long author list on first citation}
Use option \texttt{longnamesfirst} to have first citation automatically give
the full list of authors.
Suppress this for certain citations with |\shortcites{|\emph{key-list}|}|,
given before the first citation.
\head{Local configuration}
Any local recoding or definitions can be put in \texttt{#1}\def\filedate{#2}\def\fileversion{#3}}\texttt{.cfg} which
is read in after the main package file.
\head{Options that can be added to \texttt{\char`\\ usepackage}}
\begin{description}
\item[\ttfamily round] (default) for round parentheses;
\item[\ttfamily square] for square brackets;
\item[\ttfamily curly] for curly braces;
\item[\ttfamily angle] for angle brackets;
\item[\ttfamily colon] (default) to separate multiple citations with
colons;
\item[\ttfamily comma] to use commas as separaters;
\item[\ttfamily authoryear] (default) for author--year citations;
\item[\ttfamily numbers] for numerical citations;
\item[\ttfamily super] for superscripted numerical citations, as in
\textsl{Nature};
\item[\ttfamily sort] orders multiple citations into the sequence in
which they appear in the list of references;
\item[\ttfamily sort\&compress] as \texttt{sort} but in addition multiple
numerical citations are compressed if possible (as 3--6, 15);
\item[\ttfamily longnamesfirst] makes the first citation of any reference
the equivalent of the starred variant (full author list) and subsequent
citations normal (abbreviated list);
\item[\ttfamily sectionbib] redefines |\thebibliography| to issue
|\section*| instead of |\chapter*|; valid only for classes with a
|\chapter| command; to be used with the \texttt{chapterbib} package;
\item[\ttfamily nonamebreak] keeps all the authors' names in a citation on
one line; causes overfull hboxes but helps with some \texttt{hyperref}
problems.
\end{description}
\end{document}
\section{Introduction} \label{sec:intro}
\latex\ \footnote{\url{http://www.latex-project.org/}} is a document markup
language that is particularly well suited for the publication of
mathematical and scientific articles \citep{lamport94}. \latex\ was written
in 1985 by Leslie Lamport who based it on the \TeX\ typesetting language
which itself was created by Donald E. Knuth in 1978. In 1988 a suite of
\latex\ macros were developed to investigate electronic submission and
publication of AAS Journal articles \citep{1989BAAS...21..780H}. Shortly
afterwards, Chris Biemesdefer merged these macros and more into a \latex\
2.08 style file called \aastex. These early \aastex\ versions introduced
many common commands and practices that authors take for granted today.
Substantial revisions
were made by Lee Brotzman and Pierre Landau when the package was updated to
v4.0. AASTeX v5.0, written in 1995 by Arthur Ogawa, upgraded to \latex\ 2e
which uses the document class in lieu of a style file. Other improvements
to version 5 included hypertext support, landscape deluxetables and
improved figure support to facilitate electronic submission.
\aastex\ v5.2 was released in 2005 and introduced additional graphics
support plus new mark up to identifier astronomical objects, datasets and
facilities.
In 1996 Maxim Markevitch modified the AAS preprint style file, aaspp4.sty,
to closely emulate the very tight, two column style of a typeset
Astrophysical Journal article. The result was emulateapj.sty. A year
later Alexey Vikhlinin took over development and maintenance. In 2001 he
converted emulateapj into a class file in \latex\ 2e and in 2003 Vikhlinin
completely rewrote emulateapj based on the APS Journal's RevTEX class.
During this time emulateapj gained growing acceptance in the astronomical
community as it filled an author need to obtain an approximate number of
manuscript pages prior to submission for cost and length estimates. The
tighter typeset also had the added advantage of saving paper when printing
out hard copies.
Even though author publication charges are no longer based on print pages
\footnote{see Section \ref{sec:pubcharge} in the Appendix for more details
about how current article costs are calculated.} the emulateapj class file
has proven to be extremely popular with AAS Journal authors. An
analysis of submitted \latex\ manuscripts in 2015 revealed that $\sim$65\%
either called emulateapj or have a commented emulateapj classfile call
indicating it was used at some stage of the manuscript construction.
Clearly authors want to have access to a tightly typeset version of the
article when corresponding with co-authors and for preprint submissions.
When planning the next \aastex\ release the popularity of emulateapj played
an important roll in the decision to drop the old base code and adopt and
modify emulateapj for \aastex\ v6.+ instead. The change brings \aastex\
inline with what the majority of authors are already using while still
delivering new and improved features. \aastex\ v6.0 through v6.31 were
written by Amy Hendrickson. The release dates were January 2016 (v6.0),
October 2016 (v6.1), January 2018 (v6.2), June 2019 (v6.3), and March 2010
(v6.31) respectively.
The new features in the recent releases includes:
\begin{itemize}
\item{v6.0}
\begin{enumerate}
\item line numbering and watermarking,
\item improved citations for third party data repositories and software,
\item easier construction of matrix figures consisting of multiple
encapsulated postscript (EPS) or portable document format (PDF) files,
\item figure set mark up for large collections of similar figures,
\item color mark up to easily enable/disable revised text highlighting,
\item improved url support, and
\item numerous table options such as the ability to hide columns, column decimal alignment, automatic column math mode and numbering, plus splitting of wide tables (see Section \ref{subsec:tables} for details).
\end{enumerate}
\item{v6.1}
\begin{enumerate}
\item ORCID support for preprints,
\item improved author, affiliation and collaboration mark up,
\item reintroduced the old AASTeX v5.2 {\tt\string\received}, {\tt\string\revised}, {\tt\string\accepted}, and {\tt\string\published} commands plus added the new {\tt\string\submitjournal} command to document which AAS Journal the manuscript was submitted to, plus
\item new typeset style options including {\tt\string modern}.
\end{enumerate}
\item{v6.2}
\begin{enumerate}
\item A new RNAAS style option for Research Note manuscripts,
\item Titles no longer put in all caps,
\item No page skip between the title page and article body,
\item re-introduce RevTeX's widetext environment for long lines in two column style formats, and
\item upgrade to the {\tt\string\doi} command.
\end{enumerate}
\item{v6.3}
\begin{enumerate}
\item New {\tt\string interactive} environment to highlight interactive figures (see Section \ref{animation}),
\item Improved collaboration commands,
\item New {\tt\string anonymous} style to keep the authors, affiliations and acknowledgments from showing in the compiled pdf for dual anonymous review, and
\item Adoptions of IAU approved syntax for nominal units, see Section \ref{nominal}.
\end{enumerate}
\item{v6.31}
\begin{enumerate}
\item Fixes a bug in the {\tt\string anonymous} style for dual anonymous review.
\item Improves line numbering with the {\tt\string linenumbers} style around equations due to the amsmath and lineno package compatibility issues.
\item Depreciate the {\tt\string \\acknowledgment} command in favor of the {\tt\string acknowledgment} environment.
\end{enumerate}
\end{itemize}
The rest of this article provides information and examples on how to create
your own AAS Journal manuscript with v6.31. Special emphasis is placed on
how to use the full potential of \aastex\ v6+. The next section describes
the different manuscript styles available and how they differ from past
releases. Section \ref{sec:floats} describes table and figure placement.
Specific examples of tables, Section
\ref{subsec:tables}, and figures, Section \ref{subsec:figures}, are also
provided. A special emphasis is placed on interactive figures.
Section \ref{sec:displaymath} discusses how to display math and
incorporate equations in a manuscript while Section \ref{sec:highlight}
discuss how to use different ways to highlight revisions. The last section,
\ref{sec:cite}, shows how recognize software and external data as first
class references in the manuscript bibliography. An appendix is included
for additional information readers might find useful.
More documentation is embedded in the comments of this \latex\ file and in the online documentation at
\url{http://journals.aas.org/authors/aastex.html}.
\section{Manuscript styles} \label{sec:style}
The default style in \aastex\ v6.31 is a tight single column style, e.g. 10
point font, single spaced. The single column style is very useful for
article with wide equations. It is also the easiest to style to work with
since figures and tables, see Section \ref{sec:floats}, will span the
entire page, reducing the need for address float sizing.
To invoke a two column style similar to the what is produced in
the published PDF copy use \\
\noindent {\tt\string\documentclass[twocolumn]\{aastex631\}}. \\
\noindent Note that in the two column style figures and tables will only
span one column unless specifically ordered across both with the ``*'' flag,
e.g. \\
\noindent{\tt\string\begin\{figure*\}} ... {\tt\string\end\{figure*\}}, \\
\noindent{\tt\string\begin\{table*\}} ... {\tt\string\end\{table*\}}, and \\
\noindent{\tt\string\begin\{deluxetable*\}} ... {\tt\string\end\{deluxetable*\}}. \\
\noindent This option is ignored in the onecolumn style.
Some other style options are outlined in the commented sections of this
article. Any combination of style options can be used.
Two style options that are needed to fully use the new revision tracking
feature, see Section \ref{sec:highlight}, are {\tt\string linenumbers} which
uses the lineno style file to number each article line in the left margin and
{\tt\string trackchanges} which controls the revision and commenting highlight
output.
There is also a new {\tt\string modern} option that uses a Daniel
Foreman-Mackey and David Hogg design to produce stylish, single column
output that has wider left and right margins. It is designed to have fewer
words per line to improve reader retention. It also looks better on devices
with smaller displays such as smart phones.
The {\tt\string anonymous} option will prevent the author and affiliations
from being shown in the compiled pdf copy. This option allows the author
to keep this critical information in the latex file but prevent the reviewer
from seeing it during peer review if dual anonymous review (DAR) is requested.
Likewise, acknowledgments can also be hidden if placed in the new
{\tt\string\begin\{acknowledgments\}} ... {\tt\string\end\{acknowledgments\}}
environment. The use of this option is highly recommended for PSJ submissions.
Advice for anonymizing your manuscript for DAR is provided at
\url{https://journals.aas.org/manuscript-preparation/#dar}.
\section{Floats} \label{sec:floats}
Floats are non-text items that generally can not be split over a page.
They also have captions and can be numbered for reference. Primarily these
are figures and tables but authors can define their own. \latex\ tries to
place a float where indicated in the manuscript but will move it later if
there is not enough room at that location, hence the term ``float''.
Authors are encouraged to embed their tables and figures within the text as
they are mentioned. Please do not place the figures and text at the end of
the article as was the old practice. Editors and the vast majority of
referees find it much easier to read a manuscript with embedded figures and
tables.
Depending on the number of floats and the particular amount of text and
equations present in a manuscript the ultimate location of any specific
float can be hard to predict prior to compilation. It is recommended that
authors textbf{not} spend significant time trying to get float placement
perfect for peer review. The AAS Journal's publisher has sophisticated
typesetting software that will produce the optimal layout during
production.
Note that authors of Research Notes are only allowed one float, either one
table or one figure.
For authors that do want to take the time to optimize the locations of
their floats there are some techniques that can be used. The simplest
solution is to placing a float earlier in the text to get the position
right but this option will break down if the manuscript is altered.
A better method is to force \latex\ to place a
float in a general area with the use of the optional {\tt\string [placement
specifier]} parameter for figures and tables. This parameter goes after
{\tt\string \begin\{figure\}}, {\tt\string \begin\{table\}}, and
{\tt\string \begin\{deluxetable\}}. The main arguments the specifier takes
are ``h'', ``t'', ``b'', and ``!''. These tell \latex\ to place the float
\underline{h}ere (or as close as possible to this location as possible), at
the \underline{t}op of the page, and at the \underline{b}ottom of the page.
The last argument, ``!'', tells \latex\ to override its internal method of
calculating the float position. A sequence of rules can be created by
using multiple arguments. For example, {\tt\string \begin\{figure\}[htb!]}
tells \latex\ to try the current location first, then the top of the page
and finally the bottom of the page without regard to what it thinks the
proper position should be. Many of the tables and figures in this article
use a placement specifier to set their positions.
Note that the \latex\ {\tt\string tabular} environment is not a float. Only
when a {\tt\string tabular} is surrounded by {\tt\string\begin\{table\}} ...
{\tt\string\end\{table\}} is it a true float and the rules and suggestions
above apply.
In AASTeX v6.31 all deluxetables are float tables and thus if they are
longer than a page will spill off the bottom. Long deluxetables should
begin with the {\tt\string\startlongtable} command. This initiates a
longtable environment. Authors might have to use {\tt\string\clearpage} to
isolate a long table or optimally place it within the surrounding text.
\subsection{Tables} \label{subsec:tables}
Tables can be constructed with \latex's standard table environment or the
\aastex's deluxetable environment. The deluxetable construct handles long
tables better but has a larger overhead due to the greater amount of
defined mark up used set up and manipulate the table structure. The choice
of which to use is up to the author. Examples of both environments are
used in this manuscript.
Tables longer than 200 data lines and complex tables should only have a
short example table with the full data set available in the machine
readable format. The machine readable table will be available in the HTML
version of the article with just a short example in the PDF. Authors are
required to indicate in the table comments that the data in machine
readable format in the full article.
Authors are encouraged to create their own machine
readable tables using the online tool at
\url{http://authortools.aas.org/MRT/upload.html}.
\aastex\ v6 introduced five new table features that were designed to make
table construction easier and the resulting display better for AAS Journal
authors. The items are:
\begin{enumerate}
\item Declaring math mode in specific columns,
\item Column decimal alignment,
\item Automatic column header numbering,
\item Hiding columns, and
\item Splitting wide tables into two or three parts.
\end{enumerate}
Full details on how to create each type are given in the following
sections. Additional details are available in the AASTeX
guidelines at \url{http://journals.aas.org/authors/aastex.html}
\subsubsection{Column math mode}
Both the \latex\ tabular and \aastex\ deluxetable require an argument to
define the alignment and number of columns. The most common values are
``c'', ``l'' and ``r'' for \underline{c}enter, \underline{l}eft, and
\underline{r}ight justification. If these values are capitalized, e.g.
``C'', ``L'', or ``R'', then that specific column will automatically be in math
mode meaning that \$s are not required. Note that having embedded dollar
signs in the table does not affect the output.
\subsubsection{Decimal alignment}
Aligning a column by the decimal point can be difficult with only center,
left, and right justification options. It is possible to use phantom calls
in the data, e.g. {\tt\string\phn}, to align columns by hand but this can
be tedious in long or complex tables. To address this \aastex\ introduces
the {\tt\string\decimals} command and a new column justification option,
``D'', to align data in that column on the decimal. In deluxetable the
{\tt\string\decimals} command is invoked before the {\tt\string\startdata}
call but can be anywhere in \latex's tabular environment.
Two other important thing to note when using decimal alignment is that each
decimal column \textit{must end with a space before the ampersand}, e.g.
``\&\&'' is not allowed. Empty decimal columns are indicated with a decimal,
e.g. ``.''. Do not use deluxetable's {\tt\string\nodata} command.
The ``D'' alignment token works by splitting the column into two parts on the
decimal. While this is invisible to the user one must be aware of how it
works so that the headers are accounted for correctly. All decimal column
headers need to span two columns to get the alignment correct. This can be
done with a multicolumn call, e.g {\tt\string\multicolumn2c\{\}} or
{\tt\string\multicolumn\{2\}\{c\}\{\}}, or use the new
{\tt\string\twocolhead\{\}} command in deluxetable. Since \latex\ is
splitting these columns into two it is important to get the table width
right so that they appear joined on the page. You may have to run the
\latex\ compiler twice to get it right.
\subsubsection{Automatic column header numbering} \label{subsubsec:autonumber}
The command {\tt\string\colnumbers} can be included to automatically number
each column as the last row in the header. Per the AAS Journal table format
standards, each column index numbers will be surrounded by parentheses. In
a \latex\ tabular environment the {\tt\string\colnumbers} should be invoked
at the location where the author wants the numbers to appear, e.g. after
the last line of specified table header rows. In deluxetable this command
has to come before {\tt\string\startdata}. {\tt\string\colnumbers} will
not increment for columns hidden by the ``h'' command, see Section
\ref{subsubsec:hide}.
Note that when using decimal alignment in a table the command
{\tt\string\decimalcolnumbers} must be used instead of
{\tt\string\colnumbers} and {\tt\string\decimals}.
\subsubsection{Hiding columns} \label{subsubsec:hide}
Entire columns can be \underline{h}idden from display simply by changing
the specified column identifier to ``h''. In the \latex\ tabular environment
this column identifier conceals the entire column including the header
columns. In \aastex's deluxetables the header row is specifically
declared with the {\tt\string\tablehead} call and each header column is
marked with {\tt\string\colhead} call. In order to make a specific header
disappear with the ``h'' column identifier in deluxetable use
{\tt\string\nocolhead} instead to suppress that particular column header.
Authors can use this option in many different ways. Since column data can
be easily suppressed authors can include extra information and hid it
based on the comments of co-authors or referees. For wide tables that will
have a machine readable version, authors could put all the information in
the \latex\ table but use this option to hid as many columns as needed until
it fits on a page. This concealed column table would serve as the
example table for the full machine readable version. Regardless of how
columns are obscured, authors are responsible for removing any unneeded
column data or alerting the editorial office about how to treat these
columns during production for the final typeset article.
Table \ref{tab:messier} provides some basic information about the first ten
Messier Objects and illustrates how many of these new features can be used
together. It has automatic column numbering, decimal alignment of the
distances, and one concealed column. The Common name column
is the third in the \latex\ deluxetable but does not appear when the article
is compiled. This hidden column can be shown simply by changing the ``h'' in
the column identifier preamble to another valid value. This table also
uses {\tt\string\tablenum} to renumber the table because a \latex\ tabular
table was inserted before it.
\begin{deluxetable*}{cchlDlc}
\tablenum{1}
\tablecaption{Fun facts about the first 10 messier objects\label{tab:messier}}
\tablewidth{0pt}
\tablehead{
\colhead{Messier} & \colhead{NGC/IC} & \nocolhead{Common} & \colhead{Object} &
\multicolumn2c{Distance} & \colhead{} & \colhead{V} \\
\colhead{Number} & \colhead{Number} & \nocolhead{Name} & \colhead{Type} &
\multicolumn2c{(kpc)} & \colhead{Constellation} & \colhead{(mag)}
}
\decimalcolnumbers
\startdata
M1 & NGC 1952 & Crab Nebula & Supernova remnant & 2 & Taurus & 8.4 \\
M2 & NGC 7089 & Messier 2 & Cluster, globular & 11.5 & Aquarius & 6.3 \\
M3 & NGC 5272 & Messier 3 & Cluster, globular & 10.4 & Canes Venatici & 6.2 \\
M4 & NGC 6121 & Messier 4 & Cluster, globular & 2.2 & Scorpius & 5.9 \\
M5 & NGC 5904 & Messier 5 & Cluster, globular & 24.5 & Serpens & 5.9 \\
M6 & NGC 6405 & Butterfly Cluster & Cluster, open & 0.31 & Scorpius & 4.2 \\
M7 & NGC 6475 & Ptolemy Cluster & Cluster, open & 0.3 & Scorpius & 3.3 \\
M8 & NGC 6523 & Lagoon Nebula & Nebula with cluster & 1.25 & Sagittarius & 6.0 \\
M9 & NGC 6333 & Messier 9 & Cluster, globular & 7.91 & Ophiuchus & 8.4 \\
M10 & NGC 6254 & Messier 10 & Cluster, globular & 4.42 & Ophiuchus & 6.4 \\
\enddata
\tablecomments{This table ``hides'' the third column in the \latex\ when compiled.
The Distance is also centered on the decimals. Note that when using decimal
alignment you need to include the {\tt\string\decimals} command before
{\tt\string\startdata} and all of the values in that column have to have a
space before the next ampersand.}
\end{deluxetable*}
\subsubsection{Splitting a table into multiple horizontal components}
Since the AAS Journals are now all electronic with no print version there is
no reason why tables can not be as wide as authors need them to be.
However, there are some artificial limitations based on the width of a
print page. The old way around this limitation was to rotate into
landscape mode and use the smallest available table font
sizes, e.g. {\tt\string\tablewidth}, to get the table to fit.
Unfortunately, this was not always enough but now along with the hide column
option outlined in Section \ref{subsubsec:hide} there is a new way to break
a table into two or three components so that it flows down a page by
invoking a new table type, splittabular or splitdeluxetable. Within these
tables a new ``B'' column separator is introduced. Much like the vertical
bar option, ``$\vert$'', that produces a vertical table lines
the new ``B'' separator indicates where to \underline{B}reak
a table. Up to two ``B''s may be included.
Table 2
shows how to split a wide deluxetable into three parts with
the {\tt\string\splitdeluxetable} command. The {\tt\string\colnumbers}
option is on to show how the automatic column numbering carries through the
second table component, see Section \ref{subsubsec:autonumber}.
\begin{splitdeluxetable*}{lccccBcccccBcccc}
\tabletypesize{\scriptsize}
\tablewidth{0pt}
\tablenum{5}
\tablecaption{Measurements of Emission Lines: two breaks \label{tab:deluxesplit}}
\tablehead{
\colhead{Model} & \colhead{Component}& \colhead{Shift} & \colhead{FWHM} &
\multicolumn{10}{c}{Flux} \\
\colhead{} & \colhead{} & \colhead{($\rm
km~s^{-1}$)}& \colhead{($\rm km~s^{-1}$)} & \multicolumn{10}{c}{($\rm
10^{-17}~erg~s^{-1}~cm^{-2}$)} \\
\cline{5-14}
\colhead{} & \colhead{} &
\colhead{} & \colhead{} & \colhead{Ly$\alpha$} & \colhead{N\,{\footnotesize
V}} & \colhead{Si\,{\footnotesize IV}} & \colhead{C\,{\footnotesize IV}} &
\colhead{Mg\,{\footnotesize II}} & \colhead{H$\gamma$} & \colhead{H$\beta$}
& \colhead{H$\alpha$} & \colhead{He\,{\footnotesize I}} &
\colhead{Pa$\gamma$}
}
\colnumbers
\startdata
{ }& BELs& -97.13 & 9117$\pm 38$& 1033$\pm 33$&$< 35$&$< 166$& 637$\pm 31$& 1951$\pm 26$& 991$\pm 30$& 3502$\pm 42$& 20285$\pm 80$& 2025$\pm 116$& 1289$\pm 107$\\
{Model 1}& IELs& -4049.123 & 1974$\pm 22$& 2495$\pm 30$&$< 42$&$< 109$& 995$\pm 186$& 83$\pm 30$& 75$\pm 23$& 130$\pm 25$& 357$\pm 94$& 194$\pm 64$& 36$\pm 23$\\
{ }& NELs& \nodata & 641$\pm 4$& 449$\pm 23$&$< 6$&$< 9$& -- & 275$\pm 18$& 150$\pm 11$& 313$\pm 12$& 958$\pm 43$& 318$\pm 34$& 151$\pm 17$\\
\hline
{ }& BELs& -85 & 8991$\pm 41$& 988$\pm 29$&$< 24$&$< 173$& 623$\pm 28$& 1945$\pm 29$& 989$\pm 27$& 3498$\pm 37$& 20288$\pm 73$& 2047$\pm 143$& 1376$\pm 167$\\
{Model 2}& IELs& -51000 & 2025$\pm 26$& 2494$\pm 32$&$< 37$&$< 124$& 1005$\pm 190$& 72$\pm 28$& 72$\pm 21$& 113$\pm 18$& 271$\pm 85$& 205$\pm 72$& 34$\pm 21$\\
{ }& NELs& 52 & 637$\pm 10$& 477$\pm 17$&$< 4$&$< 8$& -- & 278$\pm 17$& 153$\pm 10$& 317$\pm 15$& 969$\pm 40$& 325$\pm 37$&
147$\pm 22$\\
\enddata
\tablecomments{This is an example of how to split a deluxetable. You can
split any table with this command into two or three parts. The location of
the split is given by the author based on the placement of the ``B''
indicators in the column identifier preamble. For more information please
look at the new \aastex\ instructions.}
\end{splitdeluxetable*}
\subsection{Figures\label{subsec:figures}}
\begin{figure}[ht!]
\plotone{cost.pdf}
\caption{The subscription (squares) and author publication (asterisks)
costs from 1991 to 2013. Subscription cost are on the left Y axis while
the author costs are on the right Y axis. All numbers in US dollars and
adjusted for inflation. The author charges also account for the change
from page charges to digital quanta in April 2011. \label{fig:general}}
\end{figure}
Authors can include a wide number of different graphics with their articles
but encapsulated postscript (EPS) or portable document format (PDF) are
encouraged. These range from general figures all authors are familiar with
to new enhanced graphics that can only be fully experienced in HTML. The
later include figure sets, animations and interactive figures. All
enhanced graphics require a static two dimensional representation in the
manuscript to serve as the example for the reader. All figures should
include detailed and descriptive captions. These captions are absolutely
critical for readers for whom the enhanced figure is inaccessible either
due to a disability or offline access. This portion of the article
provides examples for setting up all these types in with the latest version
of \aastex.
\subsection{General figures\label{subsec:general}}
\aastex\ has a {\tt\string\plotone} command to display a figure consisting
of one EPS/PDF file. Figure \ref{fig:general} is an example which shows
the approximate changes in the subscription costs and author publication
charges from 1991 to 2013 in the AAS Journals. For a general figure
consisting of two EPS/PDF files the {\tt\string\plottwo} command can be
used to position the two image files side by side.
Both {\tt\string\plotone} and {\tt\string\plottwo} take a
{\tt\string\caption} and an optional {\tt\string\figurenum} command to
specify the figure number\footnote{It is better to not use
{\tt\string\figurenum} and let \latex\ auto-increment all the figures. If you
do use this command you need to mark all of them accordingly.}. Each is
based on the {\tt\string graphicx} package command,
{\tt\string\includegraphics}. Authors are welcome to use
{\tt\string\includegraphics} along with its optional arguments that control
the height, width, scale, and position angle of a file within the figure.
More information on the full usage of {\tt\string\includegraphics} can be
found at \break
\url{https://en.wikibooks.org/wiki/LaTeX/Importing\_Graphics\#Including\_graphics}.
\subsection{Grid figures}
Including more than two EPS/PDF files in a single figure call can be tricky to
easily format. To make the process easier for authors \aastex\ v6 offers
{\tt\string\gridline} which allows any number of individual EPS/PDF file
calls within a single figure. Each file cited in a {\tt\string\gridline}
will be displayed in a row. By adding more {\tt\string\gridline} calls an
author can easily construct a matrix X by Y individual files as a
single general figure.
For each {\tt\string\gridline} command a EPS/PDF file is called by one of
four different commands. These are {\tt\string\fig},
{\tt\string\rightfig}, {\tt\string\leftfig}, and {\tt\string\boxedfig}.
The first file call specifies no image position justification while the
next two will right and left justify the image, respectively. The
{\tt\string\boxedfig} is similar to {\tt\string\fig} except that a box is
drawn around the figure file when displayed. Each of these commands takes
three arguments. The first is the file name. The second is the width that
file should be displayed at. While any natural \latex\ unit is allowed, it
is recommended that author use fractional units with the
{\tt\string\textwidth}. The last argument is text for a subcaption.
Figure \ref{fig:pyramid} shows an inverted pyramid of individual
figure constructed with six individual EPS files using the
{\tt\string\gridline} option.
\begin{figure*}
\gridline{\fig{V2491_Cyg.pdf}{0.3\textwidth}{(a)}
\fig{HV_Cet.pdf}{0.3\textwidth}{(b)}
\fig{LMC_2009.pdf}{0.3\textwidth}{(c)}
}
\gridline{\fig{RS_Oph.pdf}{0.3\textwidth}{(d)}
\fig{U_Sco.pdf}{0.3\textwidth}{(e)}
}
\gridline{\fig{KT_Eri.pdf}{0.3\textwidth}{(f)}}
\caption{Inverted pyramid figure of six individual files. The nova are
(a) V2491 Cyg, (b) HV Cet, (c) LMC 2009, (d) RS Oph, (e) U Sco, and (f)
KT Eri. These individual figures are taken from \citet{2011ApJS..197...31S}.
\label{fig:pyramid}}
\end{figure*}
\subsection{Enhanced graphics}
Enhanced graphics have an example figure to serve as an example for the
reader and the full graphical item available in the published HTML article.
This includes Figure sets, animations, and interactive figures. The
Astronomy Image Explorer (\url{http://www.astroexplorer.org/}) provides
access to all the figures published in the AAS Journals since they offered
an electronic version which was in the mid 1990s. You can filter image
searches by specific terms, year, journal, or type. The type filter is
particularly useful for finding all published enhanced graphics. As of
March 2021 there are over 4000 videos, 1300 figure sets, and 100 interactive
figures. The next sections describe how to include these types of graphics
in your own manuscripts.
\subsubsection{Figure sets}
The grid commands given above works great for a limited set of individual
figure files but what do you do if you have many 10s or 100s or even 1000s of
individual figure files? Figure sets represents a virtual flip book of a
large group of similar style figures. The derived PDF article will only
shows an example figure while the enhanced content is available in the
figure set in the HTML edition. The advantage of a figure set gives the
reader the ability to easily sort through a large collection to find
individual component figures. The advantage to the author is that grouping
similar figures into a figure set can result in significant cost savings in
terms of reduced publication charges, see Appendix B. All of the figure set
components, along with their html framework, are also available to the reader
for download in a single .tar.gz package.
Special \latex\ mark up is required to create a figure set. Prior to
\aastex\ v6 the underlying mark up commands had to be inserted by hand
but is now included. Note that when an article with figure set is compiled
in \latex\ none of the component figures are shown and a floating Figure
Set caption will appear in the resulting PDF.
\figsetstart
\figsetnum{4}
\figsettitle{Swift X-ray light curves}
\figsetgrpstart
\figsetgrpnum{1.1}
\figsetgrptitle{KT Eri}
\figsetplot{KT_Eri.pdf}
\figsetgrpnote{The Swift/XRT X-ray light curve for the first year after
outburst of KT Eri.}
\figsetgrpend
\figsetgrpstart
\figsetgrpnum{1.2}
\figsetgrptitle{RS Oph}
\figsetplot{RS_Oph.pdf}
\figsetgrpnote{The Swift/XRT X-ray light curve for the first year after
outburst of RS Oph.}
\figsetgrpend
\figsetgrpstart
\figsetgrpnum{1.3}
\figsetgrptitle{U Sco}
\figsetplot{U_Sco.pdf}
\figsetgrpnote{The Swift/XRT X-ray light curve for the first year after
outburst of U Sco.}
\figsetgrpend
\figsetgrpstart
\figsetgrpnum{1.4}
\figsetgrptitle{V2491 Cyg}
\figsetplot{V2491_Cyg.pdf}
\figsetgrpnote{The Swift/XRT X-ray light curve for the first year after
outburst of V2491 Cyg.}
\figsetgrpend
\figsetgrpstart
\figsetgrpnum{1.5}
\figsetgrptitle{Nova LMC 2009}
\figsetplot{LMC_2009.pdf}
\figsetgrpnote{The Swift/XRT X-ray light curve for the first year after
outburst of nova LMC 2009.}
\figsetgrpend
\figsetgrpstart
\figsetgrpnum{1.6}
\figsetgrptitle{HV Cet}
\figsetplot{HV_Cet.pdf}
\figsetgrpnote{The Swift/XRT X-ray light curve for the first year after
outburst of HV Cet.}
\figsetgrpend
\figsetend
\begin{figure}
\plotone{KT_Eri.pdf}
\caption{The Swift/XRT X-ray light curve for the first year after
outburst of the suspected recurrent nova KT Eri. At a maximum count rate of
328 ct/s, KT Eri was the brightest nova in X-rays observed to date. All
the component figures (6) are available in the Figure Set. Note that
these components that are {\bf not} shown in the compiled pdf. The figure
set consists of the same figures as shown in Figure \ref{fig:pyramid}.
The example figure shown for figure sets can be one component or many.
\label{fig:fig4}}
\end{figure}
Authors are encouraged to use an online tool at
\url{http://authortools.aas.org/FIGSETS/make-figset.html} to generate their
own specific figure set mark up to incorporate into their \latex\ articles.
\subsubsection{Animations \label{animation}}
Authors may, and are in fact encouraged, to include animations in their
manuscripts. The video will stream inline with the published article and
also be available for download. When writing the manuscript, a stand alone
figure is necessary to serve as an example for the reader. Ideally, this
is a single still frame from the animation but in some case the animation
may only represent a small portion of the example figure, say one many
panels as shown in Figure \ref{fig:video}. Regardless, it is very
important that the author provide descriptive text in the figure caption
including start and stop times and the video duration. Authors should
review the AAS animation guidelines in the graphics guide at
\url{https://journals.aas.org/graphics-guide/#animations}.
\begin{figure}
\begin{interactive}{animation}{movie.mp4}
\plotone{f4.pdf}
\end{interactive}
\caption{Figure 1 from \citet{2018ApJ...868L..33L}. AIA 171\AA (a,b),
AIA 131\AA (c), and AIA 304\AA images are shown. The red rectangle
in (a) shows the field of view of the other panels. An animation of
panels (b-d) is available. It covers 8 hours of observing beginning
at 01:00 UT on 2012 January 19. The video duration is 20 seconds.
\label{fig:video}}
\end{figure}
Animations and interactive figures (Section \ref{sec:interactive}) should
use the {\tt\string\begin{interactive}} environment in the figure call. This
environment
places a blue border around the figure to indicate that the figure is
enhanced in the published HTML article. The
command also serves to alert the publisher what files are used to generate
the dynamic HTML content. {\tt\string\interactive} takes two arguments. The
first details the type and currently only three are allowed. The types are
{\tt\string js} for generic javascript interactive figures,
{\tt\string animation} for inline videos, and
{\tt\string timeseries} for interactive light curves produced
by astropy \citet{2013A&A...558A..33A}\footnote{To be release in the
summer of 2019}. If these types are not provide the compiler will issue an
error and quit. The second argument is the file that produces the enhanced
feature in the HTML article.
\subsubsection{Interactive figures \label{sec:interactive}}
Interactive figures give the reader the ability to manipulate the
information contained in an image which can add clarity or help further the
author's narrative. These figures consist of two parts, a static
representative figure for the manuscript and the dynamic javascript plus
HTML framework that allows for interactive control.
An example of an interactive figure is a 3D model.
The underlying figure is a X3D file while x3dom.js is the javascript driver
that displays it. An author created interface is added via a html wrapper.
The first 3D model published by the AAS Journals using this technique was
\citet{2014ApJ...793..127V}.
Figure \ref{fig:interactive} provides an interactive example which can be
run locally to demonstrate how a simple javascript plus html interface
allows a reader to switch between figures. The necessary files for this
particular interactive figure are in the {\tt\string interactive.tar.gz}
file included with this package. Unpack the file and point the browser to
the local html file. In this case, the javascript that runs the interactive
buttons is embedded in the html file but it could just as easily be calls
to external javascript libraries. Ideally, the javascript should be
included with the submitted package of interactive files to minimize
external dependencies within the published article.
\begin{figure}
\begin{interactive}{js}{interactive.tar.gz}
\plotone{f5.pdf}
\end{interactive}
\caption{Figure 4 from \citet{2018AJ....156...82C}. \emph{Upper panel}: the
cumulative median observing time to measure the $3\sigma$ RV masses of TESS
planets as a function of host star spectral type and up to $10^3$ hours.
The \emph{dashed blue curves} represent the results from the optical
spectrograph whereas the \emph{solid red curves} represent the near-IR
spectrograph. \emph{Lower panel}: the time derivative of the cumulative
observing time curves used to indicate the RV planet detection efficiency.
The \emph{horizontal dashed line} highlights the value of the detection
efficiency at 20 hours per detection. Note that unlike the lower panels,
the upper panels do not share a common ordinate due to the differing number
of planet detections around stars in each spectral type bin. The
interactive version has two buttons that allows one to turn the optical and
NIR layers. \label{fig:interactive}}
\end{figure}
Authors should consult the online tutorials at
\url{https://journals.aas.org/graphics-guide/#interactive_figures}
for more information on what is currently supported and links to
tutorials and examples.
\section{Displaying mathematics} \label{sec:displaymath}
The most common mathematical symbols and formulas are in the amsmath
package. \aastex\ requires this package so there is no need to
specifically call for it in the document preamble. Most modern \latex\
distributions already contain this package. If you do not have this
package or the other required packages, revtex4-1, latexsym, graphicx,
amssymb, longtable, and epsf, they can be obtained from
\url{http://www.ctan.org}
Mathematics can be displayed either within the text, e.g. $E = mc^2$, or
separate from in an equation. In order to be properly rendered, all inline
math text has to be declared by surrounding the math by dollar signs (\$).
A complex equation example with inline math as part of the explanation
follows.
\begin{equation}
\bar v(p_2,\sigma_2)P_{-\tau}\hat a_1\hat a_2\cdots
\hat a_nu(p_1,\sigma_1) ,
\end{equation}
where $p$ and $\sigma$ label the initial $e^{\pm}$ four-momenta
and helicities $(\sigma = \pm 1)$, $\hat a_i=a^\mu_i\gamma_\nu$
and $P_\tau=\frac{1}{2}(1+\tau\gamma_5)$ is a chirality projection
operator $(\tau = \pm1)$. This produces a single line formula. \latex\ will
auto-number this and any subsequent equations. If no number is desired then
the {\tt\string equation} call should be replaced with {\tt\string displaymath}.
\latex\ can also handle a a multi-line equation. Use {\tt\string eqnarray}
for more than one line and end each line with a
\textbackslash\textbackslash. Each line will be numbered unless the
\textbackslash\textbackslash\ is preceded by a {\tt\string\nonumber}
command. Alignment points can be added with ampersands (\&). There should be
two ampersands per line. In the examples they are centered on the equal
symbol.
\begin{eqnarray}
\gamma^\mu & = &
\left(
\begin{array}{cc}
0 & \sigma^\mu_+ \\
\sigma^\mu_- & 0
\end{array} \right) ,
\gamma^5= \left(
\begin{array}{cc}
-1 & 0\\
0 & 1
\end{array} \right) , \\
\sigma^\mu_{\pm} & = & ({\bf 1} ,\pm \sigma) ,
\end{eqnarray}
\begin{eqnarray}
\hat a & = & \left(
\begin{array}{cc}
0 & (\hat a)_+\\
(\hat a)_- & 0
\end{array}\right), \nonumber \\
(\hat a)_\pm & = & a_\mu\sigma^\mu_\pm
\end{eqnarray}
\section{Revision tracking and color highlighting} \label{sec:highlight}
Authors sometimes use color to highlight changes to their manuscript in
response to editor and referee comments. In \aastex\ new commands
have been introduced to make this easier and formalize the process.
The first method is through a new set of editing mark up commands that
specifically identify what has been changed. These commands are
{\tt\string\added\{<text>\}}, {\tt\string\deleted\{<text>\}}, and
{\tt\string\replaced\{<old text>\}\{<replaced text>\}}. To activate these
commands the {\tt\string trackchanges} option must be used in the
{\tt\string\documentclass} call. When compiled this will produce the
marked text in red. The {\tt\string\explain\{<text>\}} can be used to add
text to provide information to the reader describing the change. Its
output is purple italic font. To see how {\tt\string\added\{<important
added info>\}}, {\tt\string\deleted\{<this can be deleted text>\}},
{\tt\string\replaced\{<old data>\}\{<replaced data>\}}, and \break
{\tt\string\explain\{<text explaining the change>\}} commands will produce
\added{important added information}\deleted{, deleted text, and }
\replaced{old data}{and replaced data,} toggle between versions compiled with
and without the {\tt\string trackchanges} option.\explain{text explaining
the change}
A summary list of all these tracking commands can be produced at the end of
the article by adding the {\tt\string\listofchanges} just before the
{\tt\string\end\{document\}} call. The page number for each change will be
provided. If the {\tt\string linenumbers} option is also included in the
documentclass call then not only will all the lines in the article be
numbered for handy reference but the summary list will also include the
line number for each change.
The second method does not have the ability to highlight the specific
nature of the changes but does allow the author to document changes over
multiple revisions. The commands are {\tt\string\edit1\{<text>\}},
{\tt\string\edit2\{<text>\}} and {\tt\string\edit3\{<text>\}} and they
produce {\tt\string<text>} that is highlighted in bold, bold+italic and
bold+underline, respectively. Authors should use the first command to
\edit1{indicated which text has been changed from the first revision.} The
second command is to highlight \edit2{new or modified text from a second
revision}. If a third revision is needed then the last command should be used
\edit3{to show this changed text}. Since over 90\% of all manuscripts are
accepted after the 3rd revision these commands make it easy to identify
what text has been added and when. Once the article is accepted all the
highlight color can be turned off simply by adding the
{\tt\string\turnoffediting} command in the preamble. Likewise, the new commands
{\tt\string\turnoffeditone}, {\tt\string\turnoffedittwo}, and
{\tt\string\turnoffeditthree} can be used to only turn off the
{\tt\string\edit1\{<text>\}}, {\tt\string\edit2\{<text>\}} and
{\tt\string\edit3\{<text>\}}, respectively.
Similar to marking editing changes with the {\tt\string\edit} options there
are also the {\tt\string\authorcomments1\{<text>\}},
{\tt\string\authorcomments2\{<text>\}} and
{\tt\string\authorcomments3\{<text>\}} commands. These produce the same
bold red, italic blue and underlined purple text but when the
{\tt\string\turnoffediting} command is present the {\tt\string<text>}
material does not appear in the manuscript. Authors can use these commands
to mark up text that they are not sure should appear in the final
manuscript or as a way to communicate comments between co-authors when
writing the article.
\section{Software and third party data repository citations} \label{sec:cite}
The AAS Journals would like to encourage authors to change software and
third party data repository references from the current standard of a
footnote to a first class citation in the bibliography. As a bibliographic
citation these important references will be more easily captured and credit
will be given to the appropriate people.
The first step to making this happen is to have the data or software in
a long term repository that has made these items available via a persistent
identifier like a Digital Object Identifier (DOI). A list of repositories
that satisfy this criteria plus each one's pros and cons are given at \break
\url{https://github.com/AASJournals/Tutorials/tree/master/Repositories}.
In the bibliography the format for data or code follows this format: \\
\noindent author year, title, version, publisher, prefix:identifier\\
\citet{2015ApJ...805...23C} provides a example of how the citation in the
article references the external code at
\doi{10.5281/zenodo.15991}. Unfortunately, bibtex does
not have specific bibtex entries for these types of references so the
``@misc'' type should be used. The Repository tutorial explains how to
code the ``@misc'' type correctly. The most recent aasjournal.bst file,
available with \aastex\ v6, will output bibtex ``@misc'' type properly.
\begin{acknowledgments}
We thank all the people that have made this AASTeX what it is today. This
includes but not limited to Bob Hanisch, Chris Biemesderfer, Lee Brotzman,
Pierre Landau, Arthur Ogawa, Maxim Markevitch, Alexey Vikhlinin and Amy
Hendrickson. Also special thanks to David Hogg and Daniel Foreman-Mackey
for the new "modern" style design. Considerable help was provided via bug
reports and hacks from numerous people including Patricio Cubillos, Alex
Drlica-Wagner, Sean Lake, Michele Bannister, Peter Williams, and Jonathan
Gagne.
\end{acknowledgments}
\vspace{5mm}
\facilities{HST(STIS), Swift(XRT and UVOT), AAVSO, CTIO:1.3m,
CTIO:1.5m,CXO}
\software{astropy \citep{2013A&A...558A..33A,2018AJ....156..123A},
Cloudy \citep{2013RMxAA..49..137F},
Source Extractor \citep{1996A&AS..117..393B}
}
|
\subsection*{Acknowledgements} {#1}
\addtocontents{toc}{\protect\setcounter{tocdepth}{1}} }
\vfuzz2pt
\hfuzz2pt
\begin{document}
\bibliographystyle{plain}
\title{Well-posedness and singularity formation for Vlasov--Riesz system}
\author{Young-Pil Choi\thanks{Department of Mathematics, Yonsei University, Seoul 03722, Republic of Korea. E-mail: <EMAIL>} \and In-Jee Jeong\thanks{Department of Mathematics and RIM, Seoul National University, Seoul 08826. E-mail: injee\[email protected]} }
\date\today
\maketitle
\renewcommand{\thefootnote}{\fnsymbol{footnote}}
\footnotetext{\emph{Key words: Vlasov--Riesz system, well-posedness, averaging lemma, singularity formation, Fokker--Planck} \\
\emph{2010 AMS Mathematics Subject Classification:} 76B47, 35Q35 }
\renewcommand{\thefootnote}{\arabic{footnote}}
\begin{abstract}
We investigate the Cauchy problem for the Vlasov--Riesz system, which is a Vlasov equation featuring an interaction potential generalizing previously studied cases, including the Coulomb $\Phi = (-\Delta)^{-1}\rho$, Manev $(-\Delta)^{-1} + (-\Delta)^{-\frac12}$, and pure Manev $(-\Delta)^{-\frac12}$ potentials. For the first time, we extend the local theory of classical solutions to potentials more singular than that for the Manev. Then, we obtain finite-time singularity formation for solutions with various attractive interaction potentials, extending the well-known blow-up result for attractive Vlasov--Poisson for $d\ge4$. Our local well-posedness and singularity formation results extend to cases when linear diffusion and damping in velocity are present.
\end{abstract}
\section{Introduction}
In this paper, we study the initial value problem for the \textit{Vlasov--Riesz} system and its generalizations, which are Vlasov equations with general interaction forces, featuring the Coulomb and Manev potential as special cases. To be more specific, let $f = f(t,x,v)$ be the number density of particles at position $x \in \mathbb{R}^d$ with velocity $v \in \mathbb{R}^d$ at time $t$. Then, our main system reads as
\begin{equation} \tag{VR}\label{eq:VR}
\left\{
\begin{aligned}
&\partial_t f + v\cdot\nabla_x f + \nabla \Phi \cdot \nabla_v f = \sigma \nabla_v \cdot (\nabla_v f + vf),\quad (x,v) \in \mathbb{R}^d \times \mathbb{R}^d, \quad t > 0, \\
&\Phi = \kappa\Lambda^{-\beta} \rho, \quad \rho = \int_{\R^d} f \, dv,
\end{aligned}
\right.
\end{equation}
Here, $\Lambda^{-\beta}$ is defined by the Fourier multiplier with symbol $|\xi|^{-\beta}$, where $\xi$ is the dual variable for $x$. The cases $\kappa > 0$ and $\kappa < 0$ correspond to the attractive and repulsive interactions, respectively. The right hand side of \eqref{eq:VR} is the linear Fokker--Planck operator, where the coefficient $\sigma$ is nonnegative.
The system \eqref{eq:VR} covers not only the classical Coulomb interaction $\beta = 2$, i.e. $\Phi = (-\Delta)^{-1}\rho$ but also the so-called {\it pure Manev} or {\it Manev correction} \cite{BDIV97} given by $\Phi = \Lambda^{-1}\rho$. We shall also consider potentials given by linear combination of $\Lambda^{-\beta}$, the principal example being the {\it Manev} potential $\Phi = \Lambda^{-1}\rho + (-\Delta)^{-1}\rho$ proposed by Manev in \cite{Man1, Man2, Man4, Man3} as a modification of the Newtonian gravitation law. Apart from the Einstein's theory of general relativity, the additional term $\Lambda^{-1}\rho$ allows to explain various phenomena observed in the solar system, such as the anomalous secular precession of Mercury's perihelion, gravitational redshift, free gravitational collapse, and so on; see \cite{Hall94,H07, Man1, Man2, Man4, Man3, Ure98, BDIV97, Ure99, Kirk}. It is interesting to note that Sir Isaac Newton himself used a Manev-type gravitational potential to explain the dynamics of the Moon (\cite[Book I,
Section IX, Proposition XLIV, Theorem XIV]{Newton}). Moreover, Hall in \cite{Hall94} suggests an interaction potential which is \textit{even more singular} than that of Manev, to deal with Mercury's precession. We refer to \cite{DMMS95,H07, Ill00,IP05, SSM99, Diacu} and references therein for a detailed discussion of the history and applications of the modified Newtonian's law including the Manev's nonrelativistic gravitational law. Then, the Riesz interaction $\Lambda^{-\beta}\rho$ can be considered as a generalization, and such interaction has been indeed studied in the physics literature, e.g. \cite{BBDR05, Maz11, Tor16}. It also appears in the study of equilibrium properties of a system of point particles interacting via Coulomb or Riesz interactions and confined by an external potential \cite{LS17, PR18, RS16}.
\subsection{Main Results}
The main purpose of this work is to develop a local well-posedness theory of strong solutions for the system \eqref{eq:VR} which is applicable even in the singular regime $\beta<1$ and establish finite-time singularity formation within our local well-posedness framework. Our main results cover both $\sigma=0$ and $\sigma>0$, and it seems that there have been no results on singularity formation in the latter case, namely when the Fokker--Planck term is present.
\subsubsection{Local well-posedness}
To put our well-posedness result in context, let us briefly review the classical results on the initial value problem for \eqref{eq:VR}. To begin with, the case $\beta=2$ which corresponds to the famous Vlasov--Poisson ($\sigma=0$) or Vlasov--Poisson--Fokker--Planck system ($\sigma>0$) has been studied extensively. The global existence of weak solutions are discussed in \cite{CS95, LP91, Pal12, Vic91}, and classical solutions (which are global when $d\le3$) are obtained in \cite{Bou93,BD85, Degond86,Pfa92, UO78, VO90}. The asymptotic behavior of solutions is studied in \cite{BD95, CSV95, Her07, HJ13} and \cite{Gla96, Pert04, Rein07} are a few standard references for general theory on the Vlasov--Poisson system and related kinetic equations. In the case $\beta=1$ and $\sigma=0$, i.e. Vlasov--Manev system where the interaction potential is more singular by order 1, the local-in-time existence and uniqueness of classical solutions are obtained in \cite{IVDB98}, using the method of characteristics.
Although it is not clearly stated in \cite{IVDB98}, the arguments therein can be easily adapted to produce local-in-time unique classical solution to the system \eqref{eq:VR} when $\beta \in [1,2]$, namely the cases ``interpolating'' between Coulomb and Manev type potentials.
In the case $\beta \ge 1$, a standard energy estimate in Sobolev spaces $H^s$ can be employed to prove local well-posedness, for large $s$. This seems to be sharp, since when $\beta=1$, the advecting velocity $\nabla\Phi \sim \nabla\Lambda^{-1}\rho$ has the same Sobolev regularity with $f$ in the $x$ variable for each $t$. While one may expect ill-posedness in the more singular regime $\beta<1$, our main result shows that local well-posedness for smooth solutions still persists as long as $\beta>\frac34$. The key observation is that the \textit{velocity averaging} effect, coming from the kinetic transport $\partial_t + v\cdot\nabla_x$, improves regularity of $\rho$ compared with $f$ when suitably \textit{averaged in time}. While such a kinetic averaging lemma is widely used in the study of global-in-time regularity of solutions to kinetic transport equations, see e.g. \cite{GLPS88, GPS85, DLM91, Vil02, Jab09}, we are not aware of any previous applications of the averaging lemma in the construction of local smooth solutions for systems with singular advection.
For the averaging lemma to be applicable, we shall need to introduce the following weighted Sobolev norm: with $\brk{v}^2 = 1 + |v|^2$ and an integer $N\ge0$, we define $H^{s,2N}_{x,v}(\mathbb R^d\times\mathbb R^d)$--norm by \begin{equation*}
\begin{split}
\nrm{f}_{H^{s,2N}_{x,v}}^2 & := \iint_{\R^d \times \R^d} |\brk{v}^{2N}f|^2 + | \brk{v}^{2N} \Lambda^s_x f |^2 + | \brk{v}^{2N} \Lambda^s_v f |^2 \, dvdx .
\end{split}
\end{equation*} When $N = 0$, we recover the usual Sobolev space $H^s_{x,v} = H^{s,0}_{x,v}$. We are now ready to state our main result.
\begin{theorem}\label{thm:lwp}
Assume that $\frac{3}{4}< \beta \le 1$ and $\kappa \in \mathbb R$. For any $\sigma\ge0$, the Cauchy problem for \eqref{eq:VR} is locally well-posed in $H^{s,2N}_{x,v}(\mathbb R^d\times \mathbb R^d)$ for any real $s > \frac{d}{2}+1$ and integer $N>\frac{d}{4}$. That is, for any initial data $f_0 \in H^{s,2N}_{x,v}(\mathbb R^d\times \mathbb R^d)$, there exist $T>0$ and a unique solution $f$ to \eqref{eq:VR} with $f(t=0) = f_0$ belonging to $L^\infty([0,T]; H^{s,2N}_{x,v}(\mathbb R^d\times \mathbb R^d))$.
When $\sigma = 0$ and $f_0 \in H^{s}_{x,v}(\mathbb R^d\times \mathbb R^d)$ is compactly supported in $v$, there exist $T>0$ and a unique corresponding solution in $L^\infty([0,T]; H^{s}_{x,v}(\mathbb R^d\times \mathbb R^d))$ which is compactly supported in $v$ for each $t \in [0,T]$.
\end{theorem}
\begin{remark} Note that the sign of $\kappa$ is irrelevant in the above. Moreover, Theorem \ref{thm:lwp} readily extends to systems where the operator $\Lambda^{-\beta}$ in \eqref{eq:VR} is replaced with convolution-type operators which have the same order of singularity in the kernel. For instance, one may treat $\Lambda^{-\beta} + \sum_{i} c_i \Lambda^{-\beta_i}$ with any $c_i\in\mathbb R$ and $\beta_i>\beta$.
\end{remark}
\subsubsection{Finite-time singularity formation}
Our second main result is the finite-time loss of smoothness of solutions to the equation \eqref{eq:VR}. It is a well-known result of Horst \cite{Horst} (also see \cite[Section 4.6]{Gla96}) that when $d\ge4$, any smooth solution to the Vlasov--Poisson system, i.e. \eqref{eq:VR} with $\beta = 2$ and $\sigma=0$, in the gravitational case can exist only on a finite interval of time. For the Vlasov--(pure) Manev system, i.e. \eqref{eq:VR} with $\beta = 1$ and $\sigma=0$, singularity formation is shown in \cite{BDIV97}. We extend these classical results to the case with attractive singular power-law potential, even when the Fokker--Planck term is present. In particular, our result shows the finite time loss of smoothness of solutions to the Vlasov--Poisson--Fokker--Planck equations in the attractive case under suitable assumptions on the initial data, for high dimensions. Later, we shall also prove singularity formation in certain cases when both the attractive and repulsive potentials are present.
Although our well-posedness theory mainly focuses on the case $\beta \in (\frac34 , 1]$, our second result covers more general interaction potentials. Thus for the analysis of the finite-time singularity formation, it will be convenient to use the notation $K\star\rho$ in place of $\Phi=\Lambda^{-\beta}\rho$ and consider the following general form of $K$:
\[
K(x) = \sum_{i=1}^N c_i K_i(x), \quad K_i(x) = \frac1{|x|^{\alpha_i}},
\]
where $c_i > 0$ and $\alpha_i \in (0,d)$ for all $i=1,\dots,N$. Note that for $\beta \in (0,2)$, $\Lambda^{-\beta}\rho$ equals $K \star\rho$ up to an absolute constant, where
\[
K(x) = \frac{1}{|x|^{d-\beta}}.
\]
When $d\ge3$, the Coulomb case corresponds to $\beta=2$. For $d\ge2$, the case $\beta \in (0,2)$ corresponds to the Riesz potential, with the case $\beta=1$ corresponding to the pure Manev potential.
Since our strategy for singularity formation relies on the energy-type estimates on several physical quantities, including the total energy, radial-weighted momentum, and momentum of inertia, the proof requires a little bit of integrability of solutions. (However, the initial data does not need to have compact support or contain vacuum in any finite regions.) For this purpose, let us introduce a solution space $X$ as follows.
\begin{definition}\label{def_sol} For a given $T>0$, we call $f \in X(T)$ if $f$ is a classical solution to the Cauchy problem \eqref{eq:VR} on the time interval $[0,T]$ satisfying the following conditions of decay at far fields: \begin{equation}\label{eq:decay-assumptions}
\begin{split}
f (|v| |x|^2 + (|v| + |x|)|v|^2 + |\ln f| + |\nabla K \star \rho| ) \to 0
\end{split}
\end{equation}
as $|x|, |v| \to +\infty$ for all $t \in [0,T]$.
\end{definition}
The total energy $E=E(t)$ and momentum of inertia $I=I(t)$ are defined as follows:
\[
E:= \frac12\iint_{\R^d \times \R^d} |v|^2 f\,dxdv + \iint_{\R^d \times \R^d} f \ln f\,dxdv - \frac12 \int_{\R^d} \rho K\star\rho\,dx
\]
and
\[
I := \frac12\iint_{\R^d \times \R^d} |x|^2 f\,dxdv.
\]
The decay conditions appeared in Definition \ref{def_sol} enable us to estimate the time derivative of the above functions. It is not difficult to show that, under the framework of Theorem \ref{thm:lwp}, \eqref{eq:decay-assumptions} is satisfied for the unique local-in-time solution once we assume enough decay in $x,v$ on the initial data. Let us now state our second main result.
\begin{theorem}\label{thm:blow} Let $T>0$ and $d \geq 3$, and let $f$ be a solution to the system \eqref{eq:VR} satisfying $f \in X(T)$. Suppose $\max_{i=1,\dots,N}\alpha_i > 2$.
\begin{itemize}
\item (Vlasov equation: $\sigma =0$) Suppose the initial kinetic energy is small enough compared to the initial interaction energy:
\[
\iint_{\R^d \times \R^d} |v|^2 f_0 \,dxdv + \left( \sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\right)^{1 - \frac{2}{\alpha_M - 2}} \left(c_1 \left(\frac{\alpha_M}2 - 1\right)\right)^{\frac2{\alpha_M-2}} < \int_{\R^d} \rho_0 K\star\rho_0 \,dx,
\]
where $\alpha_M:=\max_{i=1,\dots,N}\alpha_i$.
\item (Vlasov--Fokker--Planck equation: $\sigma>0$) Suppose that the initial total energy and radial-weighted momentum is sufficiently small compared to the initial momentum of inertia so that
\[
2(1+\delta)E(0) + \beta I'(0) < -\beta(\sigma + \beta) I(0) - C_\delta - C_0
\]
for some $\delta > 0$ satisfying $2(1+\delta) < \max_{i=1,\dots, N} \alpha_i$, where $I'(0) := I'(t)|_{t=0}$, i.e.
\[
I'(0) = \iint_{\R^d \times \R^d} (x\cdot v) f_0\,dxdv,
\]
and the constants $C_\delta$, $\beta$, and $C_0$ are explicitly given as
\[
C_\delta := 4(1+\delta)(1 + \delta^{-1})^{\frac d{2+d}} (e^{-1} 2^{3d} \pi^{2d})^{\frac1{2+d}}, \qquad \beta := \frac{-\sigma + \sqrt{\sigma^2 + 4C_\delta}}{2},
\]
and
\[
C_0 := \left( \sum_{\alpha_i < 2(1+\delta)}c_i \left(1 + \delta - \frac{\alpha_i}2\right)\right)^{1 - \frac{2}{\alpha_M - 2}} \left(c_1 \left(\frac{\alpha_M}2 - 1 - \delta\right)\right)^{\frac2{\alpha_M-2}},
\]
respectively.
\end{itemize}
Then the life-span of the strong solution is finite.
\end{theorem}
\begin{figure}[h
\centering
\includegraphics[scale=1]{table}\label{fig:table}
\caption{Well-posedness and singularity formation for strong solutions of \eqref{eq:VR} in the case $\sigma=0$ and $\kappa>0$}
\end{figure}
See Figure 1 for a simple illustration of the current status of well-posedness and singularity formation of \eqref{eq:VR} in the case $\sigma=0$ and $\kappa>0$. To the best of our knowledge, it is not known whether smooth solutions exist globally in time as soon as the interaction potential becomes more singular than Coulomb in $d=2$. The situation seems to be the same for $1<\beta<2$ when $d=3$.
\begin{remark}
We give several remarks on the statement of Theorem \ref{thm:blow}.
\begin{itemize}
\item It is not difficult to find a class of initial data which satisfies the above sufficient conditions for singularity formation: one can consider ``concentrated'' initial data \begin{equation*}
\begin{split}
f_{0}(x,v) = \frac{1}{(\epsilon')^d} \varphi\left( \frac{x}{\epsilon'} \right) \frac{1}{\epsilon^d} \psi\left( \frac{v}{\epsilon} \right)
\end{split}
\end{equation*} with small $\epsilon',\epsilon>0$ and compactly supported smooth bump functions $\varphi,\psi\ge0$. Then, in the limit $\epsilon',\epsilon\to0$, we have $I, I' \to 0$ while $E\to-\infty$.
\item In the biharmoic case, i.e., $K$ satisfying $(-\Delta)^2 K = \delta_0$ in the sense of distributions, we obtain finite-time singularity formation under the assumptions of Theorem \ref{thm:blow} when $d>6$.
\item When $\sigma = 0$, we also have the finite-time singularity formation when $\min_{i=1,\dots,N} \alpha_i \geq 2$ and
\begin{equation}\label{manev}
\iint_{\R^d \times \R^d} |v|^2 f_0 \,dxdv < \int_{\R^d} \rho_0 K\star\rho_0 \,dx.
\end{equation}
In particular, if $K$ is given as the pure Manev potential \cite{BDIV97} in $d=3$:
\[
K(x) = \frac{c_1}{|x|^2}, \quad c_1 > 0,
\]
then classical solutions $f \in X(T)$ with initial data satisfying \eqref{manev} cannot exist globally in time. Note that our well-posedness theory covers the regime $\alpha_i \in [2, \frac94)$ in the three-dimensional case.
\item Later in Section \ref{subsec:mixed}, we prove singularity formation with potentials having both repulsive and attractive terms.
\end{itemize}
\end{remark}
\section{Local well-posedness}\label{sec_well}
This section is devoted to the proof of Theorem \ref{thm:lwp}. First, we review the classical kinetic averaging lemma in Section \ref{subsec:avg}. Then, we establish a priori estimates in Sobolev spaces in Sections \ref{subsec:apriori-cpt} and \ref{subsec:apriori-moment}. We conclude Theorem \ref{thm:lwp} in Section \ref{subsec:eandu} by proving existence and uniqueness of solutions. In what follows, we shall take $\kappa = \pm 1$ for simplicity and write $u(t,x) = \nabla\Phi (t,x) = \kappa \nabla \Lambda^{-\beta}\rho (t,x)$.
\subsection{Averaging lemma} \label{subsec:avg}
\begin{lemma}[{{\cite[Theorem 7.2]{Gla96}}}] \label{lem:avg}
Let $h \in L^2_{t,x,v}$ be a solution of
\begin{equation*
\begin{split}
\partial_t h + v\cdot \nabla_{x} h = \sum_{|\alpha|\le m} D_{v}^{\alpha} g_{\alpha},
\end{split}
\end{equation*} and assume that $h$ and $g_{\alpha}$ are supported in $v \in B(0,R)$. Here $m \in \mathbb{N}$. Then, for any $\psi \in C^{\infty}_{c}(\mathbb R^{d})$ compactly supported in $B(0,1)$, we have for any $R\ge 1$
\begin{equation*
\begin{split}
\int_{B(0,R)} h(\cdot,\cdot,v) \psi(R^{-1}v) \, dv \in H^{s}(\mathbb R\times\mathbb R^{d}), \quad s = \frac{1}{2(1+m)}
\end{split}
\end{equation*} with \begin{equation}\label{eq:avg-est}
\begin{split}
\left\| \int_{B(0,R)} h(\cdot,\cdot,v) \psi(R^{-1}v) \, dv \right\|_{H^{s}_{t,x}} \le C R^{\frac{d}{2}} \left( \nrm{h}_{L^2_{t,x,v}} + \sum_{|\alpha|\le m}\nrm{g_{\alpha}}_{L^2_{t,x,v}} \right).
\end{split}
\end{equation} Here, $C>0$ is a constant depending on $m$ and $\psi$ but not on $R$.
\end{lemma}
\begin{remark}
Strictly speaking, \cite[Theorem 7.2]{Gla96} proves a version of \eqref{eq:avg-est} in which dependency in $R$ is not given explicitly. However, following their proof gives $C R^{\frac{d}{2}} $ with $C$ independent of $R$, assuming $R\ge1$.
\end{remark}
\subsection{A priori estimates: compactly supported case} \label{subsec:apriori-cpt}
As a warm-up, we shall first prove the Sobolev a priori estimates in the compactly supported case, assuming $\sigma = 0$. In this case, we can simply use the standard Sobolev spaces $H^s(\mathbb R^d\times\mathbb R^d)$. To be precise, we will prove the following result.
\begin{lemma}\label{lem:apriori-cptspt}
Assume that $\sigma=0$, $\frac34 < \beta \le 1$, $\kappa \in \{ \pm 1 \}$, and $s>\frac{d}{2}+1$. Let $f$ be a sufficiently smooth and fast decaying solution to \eqref{eq:VR} on a time interval $[0,T]$. Furthermore, assume that for some $R_0>0$, ${\mathrm{supp}}\,_{v}( f(t=0) ) \subset B_{0}(R_{0})$, where $B_{0}(R_{0}) \subset \mathbb R^d_v$ is the ball of radius $R_0$ centered at the origin. Then, by taking $T>0$ smaller if necessary, but in a way depending only on $\nrm{f(t=0)}_{H^{s}(\mathbb R^d\times\mathbb R^d)}$ and $R_{0}$, we have for all $0<t\le T$ \begin{equation*}
\begin{split}
\nrm{f(t)}_{H^{s}(\mathbb R^d\times\mathbb R^d)} \le 4 \nrm{f(t=0)}_{H^{s}(\mathbb R^d\times\mathbb R^d)}, \qquad {\mathrm{supp}}\,_{v}( f(t) ) \subset B_{0}(4R_{0}).
\end{split}
\end{equation*}
\end{lemma}
\begin{proof} We proceed in several steps.
\medskip
\noindent \textbf{Step I: $H^{s}$--estimate}. Taking the Fourier transform in $x$ and $v$ and denoting the dual variables by $\xi$ and $\mu$ respectively, we obtain \begin{equation*}
\begin{split}
\partial_t \widehat{f}(t,\xi,\mu) + \nabla_\mu \cdot \xi \widehat{f}(t,\xi,\mu) - \int \widehat{u}(t, \xi-\eta) \cdot i\mu \widehat{f}(t,\eta,\mu) \,d\eta = 0 .
\end{split}
\end{equation*} With $\brk{\mu}^2 = 1 + |\mu|^2$ and $\brk{\xi}^2 = 1+|\xi|^2$, we have that \begin{equation*}
\begin{split}
\frac{1}{2} \frac{d}{dt} \iint (\brk{\mu}^{2s}+\brk{\xi}^{2s})\left| \widehat{f}(t,\xi,\mu)\right|^2 \, d\xi d\mu = I + II,
\end{split}
\end{equation*} where \begin{equation*}
\begin{split}
I = -\Re \iint (\brk{\mu}^{2s}+\brk{\xi}^{2s}) \overline{\widehat{f}(t,\xi,\mu)} \xi \cdot \nabla_\mu \widehat{f} (t,\xi,\mu) \, d\xi d\mu
\end{split}
\end{equation*} and \begin{equation*}
\begin{split}
II = \Re \iint (\brk{\mu}^{2s}+\brk{\xi}^{2s})\overline{\widehat{f} (t,\xi,\mu)} \int \widehat{u}(t, \xi-\eta) \cdot i\mu \widehat{f}(t,\eta,\mu) \, d\eta d\mu d\xi.
\end{split}
\end{equation*}
After integrating by parts, \begin{equation*}
\begin{split}
|I| =\left| \iint \xi \cdot \nabla_\mu (\brk{\mu}^{2s}+\brk{\xi}^{2s}) |\widehat{f}|^2 \, d\xi d\mu \right| \le C \iint (\brk{\mu}^{2s}+\brk{\xi}^{2s}) | \widehat{f} |^2 \, d\xi d\mu \le C \nrm{f}_{H^s_{x,v}}^2.
\end{split}
\end{equation*} Next, we observe that from (anti-)symmetry, \begin{equation*}
\begin{split}
\Re \iint \brk{\mu}^{2s} \overline{\widehat{f}(t,\xi,\mu)} \int \widehat{u}(t, \xi-\eta) \cdot i\mu \widehat{f}(t,\eta,\mu) d\eta d\mu d\xi = 0.
\end{split}
\end{equation*} Indeed, this follows from changing variables as $(\eta,\xi) \mapsto (\xi,\eta)$ and taking the complex conjugate. Similarly, if we consider the expression \begin{equation*}
\begin{split}
II' := \Re \iiint \brk{\xi}^{ s} \overline{\widehat{f}(t,\xi,\mu)} \widehat{u}(t, \xi-\eta) \cdot i\mu \brk{\eta}^{s} \widehat{f}(t,\eta,\mu) \, d\eta d\mu d\xi,
\end{split}
\end{equation*} we have that $II' = 0$ by symmetry. Therefore, we may rewrite $II$ as \begin{equation*}
\begin{split}
II = \Re \iiint \brk{\xi}^{ s} \overline{\widehat{f}(t,\xi,\mu)} (\brk{\xi}^s -\brk{\eta}^s ) \widehat{u}(t, \xi-\eta) \cdot i\mu \widehat{f}(t,\eta,\mu) \, d\eta d\mu d\xi.
\end{split}
\end{equation*}
Using that \begin{equation*}
\begin{split}
\left| \brk{\xi}^s -\brk{\eta}^s \right|\le C\brk{\xi-\eta}(\brk{\xi-\eta}^{s-1} + \brk{\eta}^{s-1}),
\end{split}
\end{equation*} we have \begin{equation*}
\begin{split}
\left| II \right| \le C(II_1 + II_2 ),
\end{split}
\end{equation*} where \begin{equation*}
\begin{split}
II_1 = \iiint \brk{\xi}^{ s} |\overline{\widehat{f}(t,\xi,\mu)}| \brk{\xi-\eta}^{s}|\widehat{u}(\xi-\eta)| |\mu \widehat{f}(t,\eta,\mu)| \, d\eta d\mu d\xi
\end{split}
\end{equation*} and \begin{equation*}
\begin{split}
II_2 = \iiint \brk{\xi}^{ s} |\overline{\widehat{f}(t,\xi,\mu)}| \brk{\xi-\eta} |\widehat{u}(\xi-\eta)| \brk{\eta}^{s-1}|\mu \widehat{f}(t,\eta,\mu)| \, d\eta d\mu d\xi.
\end{split}
\end{equation*} Term $II_2$ is easy to handle: first using Cauchy-Schwartz in $\xi,\mu$, Young's convolution inequality in $\eta$, and then Sobolev embedding, \begin{equation*}
\begin{split}
\left| II_2 \right| & \le \nrm{ \brk{\xi}^s \widehat{f} }_{L^2_{\xi,\mu}} \left\Vert \int \brk{\xi-\eta} |\widehat{u}(\xi-\eta)| \brk{\eta}^{s-1}|\mu \widehat{f}(t,\eta,\mu)|\, d\eta\right\Vert_{L^2_{\xi,\mu}}\\
&\le C\nrm{ \brk{\xi}^s \widehat{f} }_{L^2_{\xi,\mu}} \nrm{\brk{\xi}\widehat{u}(\xi)}_{L^1_\xi} \nrm{ \brk{\xi}^{s-1}\brk{\mu} \widehat{f} }_{L^2_{\xi,\mu}} \\
&\le C \nrm{u}_{H^{\frac{d}{2} + 1 + \epsilon}_x } \nrm{f}_{H^s_{x,v}}^2
\end{split}
\end{equation*} for any $\epsilon > 0$.
We proceed similarly for $II_1$: \begin{equation*}
\begin{split}
\left| II_1 \right| & \le \nrm{ \brk{\xi}^s \widehat{f}(t) }_{L^2_{\xi,\mu}} \left\Vert \int |\xi-\eta|^{s} |\widehat{u}(\xi-\eta)| |\mu \widehat{f}(t,\eta,\mu)|\, d\eta\right\Vert_{L^2_{\xi,\mu}}\\
&\le C\nrm{ \brk{\xi}^s \widehat{f}(t) }_{L^2_{\xi,\mu}} \nrm{\brk{\xi}^{s}\widehat{u}(\xi)}_{L^2_\xi} \nrm{ \brk{\mu} \widehat{f}(t) }_{L^2_{\mu}L^1_{\xi}} \\
&\le C \nrm{u}_{H^{s}} \nrm{f}_{H^s_{x,v}} \nrm{f}_{H^{\frac{d}{2}+1+\epsilon}_{x,v}}.
\end{split}
\end{equation*} {Therefore}, we conclude for any $s> \frac{d}{2}+1$ that \begin{equation*}
\begin{split}
\frac{d}{dt} \nrm{f}_{H^s_{x,v}}^{2}\le C\nrm{u}_{H^{s}} \nrm{f }^{2}_{H^s_{x,v}}.
\end{split}
\end{equation*}
\medskip
\noindent \textbf{Step II: {Averaging Lemma}}. To apply the averaging lemma, we consider the equation for $\Lambda_x^sf$: \begin{equation*}
\begin{split}
\partial_t (\Lambda_x^sf) + v\cdot\nabla_x (\Lambda_x^sf) = - \nabla_v \cdot \Lambda_x^s ( u f ).
\end{split}
\end{equation*} We shall apply Lemma \ref{lem:avg} with $m=1$, $h = \Lambda^{s}_{x}f$ and $g = \Lambda^{s}_{x}(uf)$. We assume that the support of $f$ in $v$ is compactly contained in the open ball $B(0,R)$ for all $0 \le t \le T$, and take $\psi$ to be a $C^\infty_c$--function which is identically 1 on the support of $\Lambda^{s}_{x}f$ for all $0\le t \le T$. We now estimate
\begin{equation*}
\begin{split}
\nrm{\Lambda_x^sf}_{L^2([0,T];L^2_{x,v})} \le CT^{\frac{1}{2}} \nrm{f}_{L^\infty([0,T]; H^s_{x,v})}
\end{split}
\end{equation*} and \begin{equation*}
\begin{split}
\nrm{\Lambda_x^s ( u f )}_{L^2([0,T];L^2_{x,v})} & \le C \nrm{u f }_{L^2([0,T]; L^2_vH^s_{x})} \\
& \le C \nrm{u}_{L^2([0,T];H^s_x) } \nrm{f}_{L^\infty([0,T]; L^{2}_vH^s_x )} .
\end{split}
\end{equation*} We have used that $s>\frac{d}{2}$. We now recall that $u = \pm \Lambda^{-\beta}\nabla_{x}\rho$ holds, which gives for each $t$ the estimate \begin{equation*}
\begin{split}
\nrm{u(t,\cdot)}_{H^{s} }\le C \nrm{\rho(t,\cdot)}_{H^{s+1-\beta}} .
\end{split}
\end{equation*} This gives \begin{equation*}
\begin{split}
\nrm{\Lambda_x^s ( u f )}_{L^2([0,T];L^2_{x,v})} & \le C \nrm{\rho}_{L^{2}_{t}H^{s+1-\beta}} \nrm{f}_{L^\infty([0,T]; L^{2}_vH^s_x )} .
\end{split}
\end{equation*}
We now apply the averaging lemma: since \begin{equation*}
\begin{split}
\int \Lambda_{x}^{s}f \psi(v)\, dv = \Lambda_{x}^{s}\rho,
\end{split}
\end{equation*} we obtain
\begin{equation*}
\begin{split}
\nrm{\Lambda^s\rho}_{L^2_{t}H^{\frac{1}{4}}_{x}} \le \nrm{\Lambda^s\rho}_{H^{\frac{1}{4}}_{t,x}} \le C(1+R^{\frac{d}{2}})\left(T^{\frac{1}{2}} \nrm{f}_{L^\infty_t L^{2}_{v} H^{s}_{x}} + \nrm{\rho}_{L^{2}_{t}H^{s+1-\beta}} \nrm{f}_{L^\infty_{t} L^{2}_vH^s_x} \right).
\end{split}
\end{equation*} On the other hand, we can directly estimate \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_{t,x}} \le CR^{\frac{d}{2}} \nrm{f}_{L^2_{t,x,v}} \le CR^{\frac{d}{2}}T^{\frac12} \nrm{f}_{L^\infty_{t}L^2_{x,v}}.
\end{split}
\end{equation*} Combining the previous estimates and using interpolation \begin{equation*}
\begin{split}
\nrm{\rho}_{H^{s+1-\beta}} & \le \nrm{\rho}_{H^{s+\frac14}}^{4(1-\beta)}\nrm{\rho}_{H^{s}}^{1-4(1-\beta)} \\
&\le \varepsilon \nrm{\rho}_{H^{s+\frac14}} + C_{\beta} \varepsilon^{-\frac{4(1-\beta)}{1-4(1-\beta)}} \nrm{\rho}_{H^{s}},
\end{split}
\end{equation*} (here it is used that $\frac{3}{4}<\beta<1$) we deduce that \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_{t}H^{\frac14 + s}_{x}} & \le C(1 + R^{\frac{d}{2}}) \nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}}\left( T^{\frac12} + \nrm{\rho}_{L^{2}_{t}H^{s+1-\beta}} \right) \\
&\le C(1 + R^{\frac{d}{2}})\nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}}\left( T^{\frac12} + \varepsilon \nrm{\rho}_{L^2_tH^{s+\frac14}} + C_{\beta} \varepsilon^{-\frac{4(1-\beta)}{1-4(1-\beta)}} \nrm{\rho}_{L^2_tH^{s}} \right).
\end{split}
\end{equation*}
Therefore, we can pick \begin{equation*}
\begin{split}
\varepsilon := \frac{1}{10C} \frac{1}{ (1 + R^{\frac{d}{2}})\nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}} }
\end{split}
\end{equation*} and absorb the $\varepsilon \nrm{\rho}_{L^2_tH^{s+\frac14}}$--term on the right hand side to derive \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_{t}H^{\frac14 + s}_{x}} & \le C(1 + R^{\frac{d}{2}}) \nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}}\left( T^{\frac12} +((1 + R^{\frac{d}{2}})\nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}})^{\frac{4(1-\beta)}{1-4(1-\beta)}} \nrm{\rho}_{L^2_tH^{s}} \right).
\end{split}
\end{equation*} Finally, using \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_tH^{s}} \le C T^{\frac12}R^{\frac{d}{2}} \nrm{f}_{L^\infty_t L^2_vH^s_x},
\end{split}
\end{equation*} we conclude the estimate \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_{t}H^{\frac14 + s}_{x}} & \le C(1 + R^{\frac{d}{2}})T^{\frac12} \nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}} \left( 1 + (1 + R^{\frac{d}{2}}) \nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}}\right)^{\frac{1}{1-4(1-\beta)}} .
\end{split}
\end{equation*}
\medskip
\noindent \textbf{Step III: Closing the a priori estimate}. Returning to the $H^s$ estimate for $f$, we have for $t\le T$ that
\begin{equation*}
\begin{split}
\nrm{f(t) }_{H^s_{x,v}} & \le \nrm{f_0}_{H^s_{x,v}}\exp\left( C\int_0^t \nrm{u(\tau)}_{H^{s}} d\tau \right) \\
& \le \nrm{f_0}_{H^s_{x,v}}\exp\left( CT^{\frac12}\nrm{\rho}_{L^2_{t}H^{s+1-\beta}} \right) \\
& \le \nrm{f_0}_{H^s_{x,v}}\exp\left( C(1 + R^{\frac{d}{2}})T \nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}} \left( 1 + (1 + R^{\frac{d}{2}}) \nrm{f}_{L^\infty_{t}L^2_{v}H^{s}_{x}}\right)^{\frac{1}{1-4(1-\beta)}} \right).
\end{split}
\end{equation*} We assume that $r(t)>0$ is the smallest satisfying \begin{equation*}
\begin{split}
{\mathrm{supp}}\,_{v}( f(t,x,\cdot) ) \subset B(0,r(t)), \quad \forall x \in \mathbb R^{d}
\end{split}
\end{equation*} and set \begin{equation}\label{eq:R-def}
\begin{split}
R(T) = 1 + \sup_{t\in[0,T]} r(t).
\end{split}
\end{equation} Furthermore, writing \begin{equation}\label{eq:F-def}
\begin{split}
F(T) = \sup_{t\in[0,T]} \nrm{f(t)}_{H^s_{x,v}},
\end{split}
\end{equation} we have that \begin{equation}\label{eq:F-estimate}
\begin{split}
F(T) \le F(0)\exp\left( CT (R^{\frac{d}{2}}(T)F(T))^{1+ \frac{1}{1-4(1-\beta)}} \right).
\end{split}
\end{equation} Moreover, \begin{equation*}
\begin{split}
\left|\frac{d}{dt} R(t)\right| \le C\nrm{u(t)}_{L^\infty} \le C\nrm{ u(t)}_{H^{\frac{d}{2}+\epsilon}} \le C \nrm{\rho(t)}_{H^{s}} \le CR^{\frac{d}{2}}(t) \nrm{f(t)}_{H^{s}_{x,v}}.
\end{split}
\end{equation*} This gives \begin{equation}\label{eq:R-estimate}
\begin{split}
R(T) \le R(0)\exp\left( CT R^{\frac{d}{2}-1}(T)F(T) \right).
\end{split}
\end{equation} One can prove that $F(T)$ and $R(T)$ are continuous in $T$, see Subsection \ref{subsec:eandu} below. With a continuity argument in time, we finally conclude the a priori estimate \begin{equation}\label{eq:apriori-compact}
\begin{split}
\sup_{t\in[0,T]} \nrm{f(t)}_{H^s_{x,v}} \le 4\nrm{f_0}_{H^s_{x,v}}, \qquad R(T) \le 4R(0)
\end{split}
\end{equation} for $T>0$ sufficiently small. This finishes the proof.
\end{proof}
\subsection{Finite moment case} \label{subsec:apriori-moment}
In this section, we shall take $\sigma>0$ and obtain a priori estimates using Sobolev spaces with finite $v$-moments.
\begin{lemma}\label{lem:apriori-moment}
Assume that $\sigma>0$, $\frac34 < \beta \le 1$, $\kappa \in \{ \pm 1 \}$, $s>\frac{d}{2}+1$, and $N>\frac{d}{4}$. Let $f$ be a sufficiently smooth and fast decaying solution to \eqref{eq:VR} on a time interval $[0,T]$. Then, by taking $T>0$ smaller if necessary, but in a way depending only on $\nrm{f(t=0)}_{H^{s,2N}(\mathbb R^d\times\mathbb R^d)}$ and $\sigma$, we have for all $0<t\le T$ \begin{equation*}
\begin{split}
\sup_{t\in[0,T]}\nrm{f(t, \cdot ) }_{H^{s,2N}_{x,v}} + 2\sigma \int_0^{T} \nrm{ \nabla_v ( \brk{v}^{2N} \Lambda^s_{x} f ) }_{L^{2}}^{2} \, dt & \le C(\nrm{f(t=0)}_{ {H^{s,2N}_{x,v}} }, \sigma).
\end{split}
\end{equation*}
\end{lemma}
\begin{proof}
While the arguments are largely parallel to the compactly supported case, we now need to understand how the moments change in time and incorporate the Fokker--Planck term into the above computations.
\medskip
\noindent \textbf{Step I: $H^{s,2N}_{x,v}$--estimate}. Since the Fourier transform of the operator $\brk{v}^2$ is given by $1 - \Delta_\mu$, we have from \begin{equation*}
\begin{split}
\partial_t \widehat{f}(t,\xi,\mu) + \nabla_\mu \cdot \xi \widehat{f}(t,\xi,\mu) - \int \widehat{u}(\xi-\eta) \cdot i\mu \widehat{f}(t,\eta,\mu) d\eta = \sigma i\mu \cdot \left( i\mu \widehat{f}(t,\xi,\mu) + \frac{1}{i}\partial_{\mu} \widehat{f}(t,\xi,\mu) \right)
\end{split}
\end{equation*} that \begin{equation*}
\begin{split}
& \frac{1}{2} \frac{d}{dt} \left( \nrm{ \brk{v}^{2N} {(1 - \Delta_{x})^{\frac{s}{2}} } f }_{L^{2}}^{2} + \nrm{ { \brk{v}^{2N} (1 - \Delta_{v})^{\frac{s}{2}} } f }_{L^{2}}^{2} \right) \\
& \qquad = \frac{1}{2} \frac{d}{dt} \iint \left| (1 - \Delta_\mu)^{N} \left[ \brk{\xi}^{s} \widehat{f}(t,\xi,\mu) \right] \right|^{2} + \left| (1 - \Delta_\mu)^{N} \left[ \brk{\mu}^{s} \widehat{f}(t,\xi,\mu) \right] \right|^{2}\, d\xi d\mu \\
& \qquad = I + II + III,
\end{split}
\end{equation*} where \begin{equation*}
\begin{split}
I & = -\Re \iint (1 - \Delta_\mu)^{ N} \left[ \brk{\xi}^{s} \overline{\widehat{f}} \right] (1-\Delta_\mu)^{N} \left[ \brk{\xi}^{s} \xi \cdot \nabla_\mu \widehat{f} \right] d\xi d\mu \\
&\qquad -\Re \iint (1 - \Delta_\mu)^{ N} \left[ \brk{\mu}^{s} \overline{\widehat{f}} \right] (1-\Delta_\mu)^{N} \left[ \brk{\mu}^{s} \xi \cdot \nabla_\mu \widehat{f} \right] d\xi d\mu ,
\end{split}
\end{equation*} \begin{equation*}
\begin{split}
II &= \Re \iint (1 - \Delta_\mu)^{ N} \left[ \brk{\xi}^{ s} \overline{\widehat{f}(t,\xi,\mu)} \right] (1-\Delta_\mu)^{N} \left[ \brk{\xi}^{s} \int \widehat{u}(\xi-\eta) \cdot i\mu \widehat{f}(t,\eta,\mu) d\eta\right] d\mu d\xi \\
&\qquad + \Re \iint (1 - \Delta_\mu)^{ N} \left[ \brk{\mu}^{ s} \overline{\widehat{f}(t,\xi,\mu)} \right] (1-\Delta_\mu)^{N} \left[ \brk{\mu}^{s} \int \widehat{u}(\xi-\eta) \cdot i\mu \widehat{f}(t,\eta,\mu) d\eta\right] d\mu d\xi
\\
& =: II_\xi + II_\mu,
\end{split}
\end{equation*} and \begin{equation*}
\begin{split}
III & = \sigma \Re \iint (1 - \Delta_\mu)^{ N} \left[ \brk{\xi}^{ s} \overline{\widehat{f}} \right] (1-\Delta_\mu)^{N} \brk{\xi}^{s} \left[ -|\mu|^{2} \widehat{f} + \mu\cdot\nabla_\mu \widehat{f} \right] d\mu d\xi \\
&\qquad + \sigma \Re \iint (1 - \Delta_\mu)^{ N} \left[ \brk{\mu}^{ s} \overline{\widehat{f}} \right] (1-\Delta_\mu)^{N} \brk{\mu}^{s} \left[ -|\mu|^{2} \widehat{f} + \mu\cdot\nabla_\mu \widehat{f} \right] d\mu d\xi
\\
& =: III_\xi + III_\mu.
\end{split}
\end{equation*}
The term $I$ can be handled similarly as before; since $(1-\Delta_\mu)^N$ commutes with $\xi\cdot\nabla_\mu$, we obtain \begin{equation*}
\begin{split}
|I| \le C \nrm{f}_{H^{s,2N}_{x,v}}^2.
\end{split}
\end{equation*} Next, {to treat $II_\xi$, we first note that} \begin{equation*}
\begin{split}
(1-\Delta_\mu)^N [\mu \widehat{f}] = \mu (1-\Delta_\mu)^{N}[\widehat{f}] - 2N (1-\Delta_\mu)^{N-1} [ \nabla_\mu\widehat{f} ]
\end{split}
\end{equation*} {holds, which gives rise to two terms. The first term can be treated similarly as the corresponding term from the compact support case (replacing $\brk{\xi}^s$ by $\brk{\xi}^{s}- \brk{\eta}^s$), with bound \begin{equation*}
\begin{split}
C \left( \nrm{ \brk{\xi}\widehat{u}(\xi) }_{L^1} \nrm{ (1 - \Delta_\mu)^{N} \brk{\xi}^s \widehat{f}}_{L^{2}}^{2} + \nrm{ \brk{\xi}^{s}\widehat{u}(\xi) }_{L^{2}} \nrm{ (1 - \Delta_\mu)^{N} \widehat{f}}_{L^{1}}^{2} \right).
\end{split}
\end{equation*} On the other hand, the other term can be bounded directly, with bound \begin{equation*}
\begin{split}
C \nrm{ \widehat{u} (\xi)}_{L^1} \nrm{ (1 - \Delta_\mu)^{N} \brk{\xi}^s \widehat{f}}_{L^{2}} \nrm{ (1 - \Delta_\mu)^{N-1} \nabla_\mu \brk{\xi}^s \widehat{f}}_{L^{2}}.
\end{split}
\end{equation*} Here, we note that \begin{equation*}
\begin{split}
\nrm{ (1 - \Delta_\mu)^{N-1} \nabla_\mu \brk{\xi}^s \widehat{f}}_{L^{2}}\le C \nrm{ (1 - \Delta_\mu)^{N} \brk{\xi}^s \widehat{f}}_{L^{2}}
\end{split}
\end{equation*} holds, which can be seen by taking the inverse Fourier transform in the $\mu$-variable. Together, we obtain that} \begin{equation*}
\begin{split}
\left|II_\xi\right| \le C\nrm{u}_{H^{s}} \nrm{ f }_{H^{s,2N}_{x,v}}^2.
\end{split}
\end{equation*} Moving on to the estimate of $II_\mu$, observing the cancellation
\begin{equation*}
\begin{split}
\Re \iint (1-\Delta_\mu)^{N} \left[\brk{\mu}^{s} \overline{\widehat{f}}(t,\xi,\mu) \right] \int \widehat{u}(\xi-\eta) \cdot i\mu (1-\Delta_\mu)^{N}\left[ {\brk{\mu}^{s} } \widehat{f}(t,\eta,\mu) \right] d\eta d\mu d\xi = 0,
\end{split}
\end{equation*} we have that $II_\mu$ can be rewritten as \begin{equation*}
\begin{split}
II_\mu = \Re \iint (1 - \Delta_\mu)^{ N} \left[ \brk{\mu}^{ s} \overline{\widehat{f}} (t,\xi,\mu) \right] \int \widehat{u}(\xi-\eta) \cdot i [(1-\Delta_\mu)^{N}, \mu] \brk{\mu}^{s} \widehat{f}(t,\eta,\mu) \,d\eta d\mu d\xi,
\end{split}
\end{equation*} where $[(1-\Delta_\mu)^{N}, \mu ]$ is the commutator between $(1-\Delta_\mu)^{N}$ and $\mu$. {It is not difficult to obtain the estimate \begin{equation}\label{eq:comm}
\begin{split}
\nrm{ [(1-\Delta_\mu)^{N}, \mu] \brk{\mu}^{s} \widehat{f} }_{L^2_{\xi,\mu}} \le C \nrm{ f }_{H^{s,2N}_{x,v}},
\end{split}
\end{equation} which implies} \begin{equation*}
\begin{split}
\left| II_\mu\right| \le C \nrm{u}_{H^{s}} \nrm{ f }_{H^{s,2N}_{x,v}}^2.
\end{split}
\end{equation*} {To prove a general commutator estimate from which \eqref{eq:comm} follows as a special case, we first note an inductive formula: for $k\ge1$ and a smooth function $G$ of $\mu$, we have \begin{equation}\label{eq:comm-i}
\begin{split}
[(1-\Delta_\mu)^{k+1}, G(\mu) ] = (1-\Delta_{\mu}) [(1-\Delta_\mu)^{k}, G(\mu) ] + [(1-\Delta_\mu) , G(\mu) ] (1-\Delta_{\mu})^{k},
\end{split}
\end{equation} which follows from a direct computation. Next, when $N=1$ and $G(\mu)=\mu$, we have \begin{equation*}
\begin{split}
[(1-\Delta_\mu) , \mu ] = - \nabla_\mu.
\end{split}
\end{equation*} Using the above inductive formula, it follows that \begin{equation}\label{eq:comm1}
\begin{split}
[(1-\Delta_\mu)^{N} , \mu ] = -N (1-\Delta_\mu)^{N-1}\nabla_\mu ,
\end{split}
\end{equation} from which \eqref{eq:comm} follows.
}
We now estimate $III$; in the case of $III_\xi$, introducing $H:= \brk{\xi}^s \widehat{f}$ for simplicity, we have that \begin{equation*}
\begin{split}
III_\xi & = \sigma \Re \iint (1 - \Delta_\mu)^{ N} H (1-\Delta_\mu)^{N} \left[ -|\mu|^{2} H + \mu\cdot\nabla_\mu H \right] d\mu d\xi \\
& = -\sigma \iint \left| |\mu| (1-\Delta_\mu)^{N} H \right|^2 - \frac{\mu}{2}\cdot \nabla_\mu \left| (1-\Delta_\mu)^{N} H \right|^{2} d\mu d\xi \\
& \qquad + \sigma \Re \iint (1 - \Delta_\mu)^{ N} H \, \left[ (1-\Delta_\mu)^{N} , ( -|\mu|^{2} + \mu\cdot\nabla_\mu) \right] H d\mu d\xi,
\end{split}
\end{equation*} where $ \left[ (1-\Delta_\mu)^{N} , ( -|\mu|^{2} + \mu\cdot\nabla_\mu) \right]$ is the commutator between the operators $(1-\Delta_\mu)^{N}$ and $( -|\mu|^{2} + \mu\cdot\nabla_\mu)$. {To begin with, note that \begin{equation}\label{eq:comm2}
\begin{split}
\left[ (1-\Delta_\mu)^{N} , \mu\cdot\nabla_\mu \right] =\left[ (1-\Delta_\mu)^{N} , \mu \right] \cdot\nabla_\mu = -N (1-\Delta_\mu)^{N-1} \Delta_{\mu} ,
\end{split}
\end{equation} using \eqref{eq:comm1}. Next, we claim that \begin{equation*}
\begin{split}
\left[ (1-\Delta_\mu)^{N} , |\mu|^{2} \right] = C_{N} \mu\cdot \nabla_\mu (1-\Delta_\mu)^{N-1} + \sum_{j=0}^{N-1} C_{j,N} \Delta_\mu^{j}
\end{split}
\end{equation*} holds for some combinatorial constants $C_{j,N}$ and $C_{N}$. This assertion can be easily proved with an induction in $N$, using \eqref{eq:comm-i} and \eqref{eq:comm2}.
} Therefore, we can bound \begin{equation}\label{eq:III}
\begin{split}
\left| III_\xi + \sigma \iint \left|\brk{\mu}(1-\Delta_\mu)^{N} H \right|^2 + d\left| (1-\Delta_\mu)^{N} H \right|^{2} d\mu d\xi \right| \le C\sigma \nrm{ (1-\Delta_\mu)^{N} H }_{L^{2}_{\xi,\mu}}^{2}.
\end{split}
\end{equation} {In the above, we have simply used that $\nrm{ \Delta_\mu^{j} H}_{L^2} \le \nrm{ (1-\Delta_\mu)^N H}_{L^2}$ for any $j\le N$, which can be seen by taking the inverse Fourier transform.} The estimate of $III_\mu$ is similar; we can obtain \eqref{eq:III} with $III_\xi$ and $H$ replaced with $III_\mu$ and $\brk{\mu}^s \widehat{f}$, respectively.
Collecting the estimates for $I$, $II$, and $III$, we conclude for $s> \frac{d}{2}+1$ that \begin{equation}\label{eq:Hs-moment}
\begin{split}
\frac12 \frac{d}{dt} \nrm{f}_{H^{s,2N}_{x,v}}^{2} + \sigma \left( \nrm{ \nabla_v ( \brk{v}^{2N} \Lambda^s_{x} f ) }_{L^{2}}^{2} + \nrm{ \nabla_v ( \brk{v}^{2N} \Lambda^s_{v} f ) }_{L^{2}}^{2} \right) \le C(\nrm{u}_{H^{s}} +\sigma )\nrm{f }_{H^{s,2N}_{x,v}}^{2}.
\end{split}
\end{equation}
\medskip
\noindent \textbf{Step II: Averaging lemma via moments}. We now fix a smooth and radial bump function $\psi(v)\ge 0$ supported on $\{ |v| \le 2 \}$ and equals 1 on $\{ |v| \le 1 \}$. Then, we set $\phi_j(v) := \psi(2^{-j}v) - \psi(2^{-j+1}v)$ and $\phi_0 = \psi$ so that $1 = \sum_{j\ge 0} \phi_j$. From \begin{equation*}
\begin{split}
\partial_t f + v\cdot\nabla_x f = - \nabla_v\cdot(uf) + \sigma \nabla_v \cdot ( \nabla_v f + vf ),
\end{split}
\end{equation*} we have \begin{equation*}
\begin{split}
\partial_t( \phi_j(v) \Lambda^s_x f) + v\cdot\nabla_x (\phi_j(v) \Lambda^s_x f) &= - \nabla_v \cdot (\phi_j(v) \Lambda^s_x (uf)) + \nabla_v \phi_j (v) \cdot \Lambda_x^s(uf) \\
&\qquad + \sigma \nabla_v \cdot ( \phi_j(v)\nabla_v (\Lambda^s_xf) + v \phi_j(v) \Lambda^s_xf ) \\
&\qquad + \sigma \nabla_v\phi_j(v) \cdot ( \nabla_v (\Lambda_x^sf) + v\Lambda_x^sf ).
\end{split}
\end{equation*} Using that \begin{equation*}
\begin{split}
\Lambda^s\rho(x) = \sum_{j\ge 0} \int \Lambda^s_x f(x,v) \phi_j(v) \, dv ,
\end{split}
\end{equation*} we shall apply the Averaging Lemma for each $j$ to $\phi_j(v) \Lambda^s_x f$: \begin{equation*}
\begin{split}
&\left\Vert \int \Lambda^s_x f(x,v) \phi_j(v) \, dv \right\Vert_{L^2_{t}H^{\frac14}_{x}} \\
& \le C2^{\frac{(j+1)d}{2}}\left( \nrm{ \phi_j(v) \Lambda^s_x f }_{L^2_{t,x,v}} + \nrm{ \phi_j(v) \Lambda^s_x (uf)}_{L^2_{t,x,v}} + \nrm{ \nabla \phi_j (v) \cdot \Lambda_x^s(uf) }_{L^2_{t,x,v}} \right. \\
&\qquad + \left. \sigma \nrm{ \phi_j(v)\nabla_v (\Lambda^s_xf) + v \phi_j(v) \Lambda^s_xf }_{L^2_{t,x,v}} + \sigma \nrm{\nabla_v\phi_j(v) \cdot ( \nabla_v (\Lambda_x^sf) + v\Lambda_x^sf ) }_{L^{2}_{t,x,v}} \right) \\
& \le C2^{\frac{(j+1)d}{2}}( 1 + \sigma 2^{j}) 2^{-2N(j-1)} \left( ( T^{\frac12} + \nrm{\rho}_{L^2_t H^{s+1-\beta}}) \nrm{f}_{L^\infty_t H^{s,2N}_{x,v}} + \sigma\nrm{ \nabla_v ( \brk{v}^{2N} \Lambda_x^s f ) }_{L^2_{t,x,v}} \right).
\end{split}
\end{equation*} Therefore, for $N> \frac{d}{4}$ ( {this implies that $N \ge \frac{d}{4}+1$ since $N$ was assumed to be an integer}), we obtain from triangle inequality that \begin{equation*}
\begin{split}
\nrm{ \Lambda^s\rho}_{L^2_t H^{\frac14}_x} &\le \sum_{j\ge 0} \left\Vert \int \Lambda^s_x f(x,v) \phi_j(v) \, dv \right\Vert_{L^2_{t}H^{\frac14}_{x}} \\
& \le C \left( ( T^{\frac12} + \nrm{\rho}_{L^2_t H^{s+1-\beta}}) \nrm{f}_{L^\infty_t H^{s,2N}_{x,v}} + \sigma\nrm{ \nabla_v ( \brk{v}^{2N} \Lambda_x^s f ) }_{L^2_{t,x,v}} \right).
\end{split}
\end{equation*} Interpolating similarly as in the compactly supported case, \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_{t} \dot{H}^{\frac14 + s}_{x}} & \le C \nrm{f}_{L^\infty_{t} H^{s,2N}_{x,v}}\left( T^{\frac12} +(1 + \nrm{f}_{L^\infty_{t} H^{s,2N}_{x,v}})^{\frac{4(1-\beta)}{1-4(1-\beta)}} \nrm{\rho}_{L^2_tH^{s}} \right) + C\sigma\nrm{ \nabla_v ( \brk{v}^{2N} \Lambda_x^s f ) }_{L^2_{t,x,v}} .
\end{split}
\end{equation*} On the other hand, from the Cauchy--Schwartz inequality, we obtain that \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_tH^{s}} \le C T^{\frac12} \nrm{f}_{L^\infty_t H^{s,2N}_{x,v}},
\end{split}
\end{equation*} for $N>\frac{d}{4}$. Therefore, we conclude the estimate \begin{equation*}
\begin{split}
\nrm{\rho}_{L^2_{t}H^{\frac14 + s}_{x}} & \le C T^{\frac12} \nrm{f}_{L^\infty_{t} H^{s,2N}_{x,v}} \left( 1 + \nrm{f}_{L^\infty_{t} H^{s,2N}_{x,v}}\right)^{\frac{1}{1-4(1-\beta)}} + C\sigma\nrm{ \nabla_v ( \brk{v}^{2N} \Lambda_x^s f ) }_{L^2_{t,x,v}} .
\end{split}
\end{equation*}
\medskip
\noindent \textbf{Step III: Closing the a priori estimate}. Returning to \eqref{eq:Hs-moment} and integrating in time from $t = 0$ to $t = T$, we obtain that \begin{equation*}
\begin{split}
\nrm{f(T,\cdot)}_{H^{s,2N}_{x,v}}^{2} + 2\sigma \int_0^{T} \nrm{ \nabla_v ( \brk{v}^{2N} \Lambda^s_{x} f ) }_{L^{2}}^{2} \, dt \le \nrm{f_0}_{H^{s,2N}_{x,v}}^{2} + \int_0^T C(\nrm{u}_{H^{s}} +\sigma )\nrm{f }_{H^{s,2N}_{x,v}}^{2} \, dt .
\end{split}
\end{equation*}
In the above inequality, the same estimate holds for $\nrm{f(T,\cdot)}_{H^{s,2N}_{x,v}}^{2}$ replaced with any $\nrm{f(\tau,\cdot)}_{H^{s,2N}_{x,v}}^{2}$ as long as $0\le \tau<T$. To ease notation, let us introduce \begin{equation*}
\begin{split}
X := \sup_{t\in[0,T]} \nrm{f(t,\cdot)}_{H^{s,2N}_{x,v}}^{2}, \qquad A(t) := \nrm{ \nabla_v ( \brk{v}^{2N} \Lambda^s_{x} f(t, \cdot) ) }_{L^{2}_{x,v}} .
\end{split}
\end{equation*} Then, we simply have \begin{equation*}
\begin{split}
X + 2\sigma\int_0^T A^{2}(t) \, dt \le C\left( 1 + \sigma TX + X \int_0^T \nrm{u}_{H^{s}} \,dt \right).
\end{split}
\end{equation*} Note that at this point we can absorb the $\sigma TX$ term on the right hand side to the left hand side by taking $T \le 1/(2\sigma)$. Then, we recall the averaging lemma: \begin{equation*}
\begin{split}
\int_0^T \nrm{u}_{H^{s}} \,dt \le CT^{\frac12}\left(T^{\frac12} X \left( 1 + X \right)^{\frac{1}{1-4(1-\beta)}} + \sigma \nrm{A}_{L^2_t([0,T])} \right).
\end{split}
\end{equation*} This gives the estimate \begin{equation}\label{eq:prelim}
\begin{split}
X + 4\sigma\int_0^T A^{2}(t) \, dt \le C(1+X)T^{\frac12}\left(T^{\frac12} X \left( 1 + X \right)^{\frac{1}{1-4(1-\beta)}} + \sigma \nrm{A}_{L^2_t([0,T])} \right).
\end{split}
\end{equation} On the right hand side, the term involving $A$ can be handled as follows: \begin{equation*}
\begin{split}
C(1+X) T^{\frac12} \sigma \nrm{A}_{L^2_t([0,T])} \le C_\sigma(1+X)^2T + 2 \sigma \nrm{A}_{L^2_t([0,T])}^{2}.
\end{split}
\end{equation*} Then, the last term on the right hand side can be absorbed into the left hand side of \eqref{eq:prelim}. This gives \begin{equation*
\begin{split}
X + 2\sigma\int_0^T A^{2}(t) \, dt \le C(1+X)T^{\frac12}\left(T^{\frac12} X \left( 1 + X \right)^{\frac{1}{1-4(1-\beta)}} + C_\sigma(1+X)^2T \right) .
\end{split}
\end{equation*} The right hand side is small when $T\ll1$. Therefore, with a standard continuity argument in time, we conclude that \begin{equation*}
\begin{split}
\sup_{t\in[0,T]}\nrm{f(t, \cdot ) }_{H^{s,2N}_{x,v}} + 2\sigma \int_0^{T} \nrm{ \nabla_v ( \brk{v}^{2N} \Lambda^s_{x} f ) }_{L^{2}}^{2} \, dt & \le C(\nrm{f_0}_{ {H^{s,2N}_{x,v}} }, \sigma)
\end{split}
\end{equation*} for some $T = T(\nrm{f_0}_{ {H^{s,2N}_{x,v}} }, \sigma )>0$. This finishes the proof of the desired a priori estimate. \end{proof}
\subsection{Existence and uniqueness} \label{subsec:eandu}
In the above, we have obtained a priori estimates both in the compactly supported and the finite moment cases. Given these estimates, to conclude Theorem \ref{thm:lwp} it only remains to prove existence and uniqueness of solutions, which is rather standard.
\begin{proof}[Proof of Theorem \ref{thm:lwp}]
We shall sketch a proof in the case $\sigma = 0$, assuming compact support in $v$. The case $\sigma>0$ does not require any significant change in the arguments. Let us fix some $\frac34 < \beta \le 1$ and initial data $f_0 \in H^{s}_{x,v}(\mathbb R^d\times\mathbb R^d)$ compactly supported in $v$. To obtain existence of an $L^\infty_tH^s$--solution to \eqref{eq:VR} with initial data $f_0$, we may consider the regularized system
\begin{equation} \label{eq:VR-reg}
\left\{
\begin{aligned}
&\partial_t f^\epsilon + v\cdot\nabla_x f^\epsilon \pm \Lambda^{-\beta}( \varphi_\epsilon * \nabla_x \rho^\epsilon ) \cdot \nabla_v f^\epsilon = 0, \\
& \rho^\epsilon = \int_{\R^d} f^\epsilon \, dv, \\
& f^\epsilon(t=0) = f_0.
\end{aligned}
\right.
\end{equation} Here $\varphi_\epsilon(x) := \epsilon^{-d}\varphi(\epsilon^{-1}x)$ is the standard mollifier scaled to $\epsilon>0$. Then, for each $\epsilon>0$, it is not difficult to establish existence (e.g. following the same argument as in the Vlasov--Poisson system \cite[Chapter 4]{Gla96}) of a solution $f^\epsilon$ belonging to $L^{\infty}([0,T_\epsilon); H^{s}_{x,v})$ and compactly supported in $v$. While in principle $T_\epsilon>0$ could decrease as $\epsilon\to0$, the time interval of existence can be extended as long as the support radius in $v$ and the quantity $\nrm{f(t,\cdot)}_{H^{s}_{x,v}}$ remains bounded. By defining $F^\epsilon$ and $R^\epsilon$ by \eqref{eq:F-def} and \eqref{eq:R-def} with $f$ replaced with $f^\epsilon$, one can repeat the proof of a priori estimates for \eqref{eq:VR-reg} to derive \eqref{eq:F-estimate} and \eqref{eq:R-estimate} with $(F,R)$ replaced with $(F^\epsilon, R^\epsilon)$. Furthermore, it is clear that $R^\epsilon$ is continuous in time. Therefore, applying a continuity argument in time is justified, and we conclude \eqref{eq:apriori-compact} for $(F^\epsilon, R^\epsilon)$, which in particular guarantees uniform-in-$\epsilon$ existence time interval $[0,T]$ with $T>0$ depending only on $f_0$. The sequence $\{ f^\epsilon \}_{\epsilon>0}$ has a weakly convergent subsequence in $L^\infty([0,T];H^{s}_{x,v})$, and we denote the limit by $f$. Then, \begin{itemize}
\item $f$ is compactly supported in $v$ and belongs to $L^\infty([0,T];H^{s}_{x,v})$.
\item $f(t=0)=f_0$.
\item $f$ is a solution to \eqref{eq:VR}.
\end{itemize}
This gives existence. To prove uniqueness, assume that there are two compactly supported solutions $f, \bar{f}$ with the same initial data $f_0$. We may assume that the solutions belong to $L^\infty([0,T];H^s)$ with $T>0$ and supported in $v \in B(0,R)$ for some $R\ge 1$ for all $t \in [0,T]$. We denote the corresponding velocities by $u = \pm \Lambda^{-\beta} \nabla\rho $ and $\bar{u} = \pm \Lambda^{-\beta}\nabla\bar{\rho}$ where $\rho = \int f dv$ and $\bar{\rho} = \int \bar{f} dv$. Then, the equation for $f- \bar{f}$ reads \begin{equation*}
\begin{split}
\frac{d}{dt} (f - \bar{f}) + v \cdot \nabla_x (f - \bar{f}) + u \cdot \nabla_v (f - \bar{f}) + (u - \bar{u}) \cdot \nabla \bar{f} = 0.
\end{split}
\end{equation*} We estimate for $t \in [0,T]$ \begin{equation}\label{eq:uniq}
\begin{split}
\frac{d}{dt} \nrm{f - \bar{f}}_{L^{2}_{x,v}}^2 \le C_R\nrm{ \nabla \bar{f}}_{L^\infty} \nrm{u - \bar{u}}_{L^{2}_{x}}\nrm{f - \bar{f}}_{L^{2}_{x,v}},
\end{split}
\end{equation} with $C_R>0$ depending on $R$. Then, for some small $0<\tau\le T$ to be determined, the averaging lemma applied to $f - \bar{f}$ gives \begin{equation*}
\begin{split}
\nrm{\rho-\bar{\rho}}_{L^2([0,\tau];H^{\frac14}_x)} \le C_{R,f,\bar{f}} \tau^{\frac12} \nrm{f - \bar{f}}_{L^\infty([0,\tau];L^{2}_{x,v})}.
\end{split}
\end{equation*} Here and in the following, $C_{R,f,\bar{f}}$ is a constant which depends only on $R$ and $\nrm{f}_{L^\infty([0,T];H^s)}$, $\nrm{\bar{f}}_{L^\infty([0,T];H^s)}$. This gives \begin{equation*}
\begin{split}
\nrm{u-\bar{u}}_{L^2([0,\tau];L^{2}_x)} \le C_{R,f,\bar{f}} \tau^{\frac12} \nrm{f-\bar{f}}_{L^\infty([0,\tau];L^{2}_x)}.
\end{split}
\end{equation*} Returning to \eqref{eq:uniq} and integrating in time, we have from $f-\bar{f} = 0$ at $t = 0$ that \begin{equation*}
\begin{split}
\nrm{f-\bar{f}}_{L^2_{x,v}}(\tau) & \le C_{R,f,\bar{f}}\int_0^\tau \nrm{u-\bar{u}}_{L^2} (s) \,ds \\
& \le C_{R,f,\bar{f}}\tau \nrm{f - \bar{f}}_{L^\infty([0,\tau];L^{2}_x)}.
\end{split}
\end{equation*} We could have taken $\tau>0$ small in a way that $C_{R,f,\bar{f}}\tau<\frac12$. This forces that \begin{equation*}
\begin{split}
\nrm{f - \bar{f}}_{L^\infty([0,\tau];L^{2}_x)} = 0.
\end{split}
\end{equation*} Repeating the same argument starting from $t =\tau$ gives $f = \bar{f}$ on $[0,T]$. This finishes the proof of uniqueness. \end{proof}
\begin{comment}
\section{Global well-posedness}
In this section, we prove the global well-posedness in the case $d\le2 $. To begin with, we formally obtain a uniform in time bound on the kinetic energy, assuming that the solution exists and smooth. This argument is originally due to Horst \cite{Horst} in the three-dimensional Vlasov--Poisson system.
\subsection{Kinetic energy bound}
Recall that
\begin{equation*}
\begin{split}
E(t) = \frac12 \nrm{ |v|^2 f}_{L^1} - \frac12 \nrm{\Lambda^{\frac{\alpha-d}{2}} \rho}_{L^{2}}^{2} + \iint f \ln f \, dxdv
\end{split}
\end{equation*}
We have that $E(t) \le E(0)$. This gives \begin{equation}\label{eq:energy-prelim}
\begin{split}
\nrm{ |v|^2 f(t,\cdot)}_{L^1} \le 2E(0) + \nrm{\Lambda^{\frac{\alpha-d}{2}} \rho(t,\cdot)}_{L^2}^{2} -2 \iint f \ln f \, dxdv
\end{split}
\end{equation}
\medskip
\noindent \textbf{Estimate of the interaction energy}. We begin with \begin{equation*}
\begin{split}
\nrm{\Lambda^{\frac{\alpha-d}{2}} \rho }_{L^{2}} \le C \nrm{\rho}_{L^{q}}, \quad q := \frac{d}{d-\alpha/2}.
\end{split}
\end{equation*} On the other hand, from kinetic interpolation we have \begin{equation*}
\begin{split}
\nrm{\rho}_{L^{\frac{2+d}{1+d}}} \le C \nrm{f}_{L^{\infty}}^{\frac{2}{2+d}} \nrm{ |v|^2 f }_{L^1}^{\frac{d}{2+d}}.
\end{split}
\end{equation*} Whenever \begin{equation*}
\begin{split}
1 \le q \le \frac{2+d}{1+d},
\end{split}
\end{equation*} we can estimate \begin{equation*}
\begin{split}
\nrm{\rho}_{L^q} \le \nrm{\rho}_{L^{\frac{2+d}{1+d}}}^\theta \nrm{\rho}_{L^{1}}^{1-\theta}
\end{split}
\end{equation*} for some $0 \le \theta=\theta(q,d) \le 1$. Hence \begin{equation*}
\begin{split}
\nrm{\Lambda^{\frac{\alpha-d}{2}} \rho }_{L^{2}}^{2}\le C \nrm{\rho}_{L^{1}}^{2(1-\theta)} (\nrm{f}_{L^{\infty}}^{\frac{2}{2+d}} \nrm{ |v|^2 f }_{L^1}^{\frac{d}{2+d}})^{2\theta}
\end{split}
\end{equation*} When $d = 1$, this gives for any $0\le\theta\le 1$ \begin{equation}\label{eq:interaction-bound}
\begin{split}
\nrm{\Lambda^{\frac{\alpha-d}{2}} \rho(t,\cdot) }_{L^{2}}^{2}\le C(\epsilon,f_0) \exp(C\sigma t) + \epsilon \nrm{|v|^2 f}_{L^1}
\end{split}
\end{equation} for any $\epsilon>0$ using $\epsilon$-Young inequality. In the case $ d = 2$, we need $\theta < 1$ to obtain the previous inequality, unless the initial data is small. This gives the range $0<\alpha<1$ in 2D.
Assuming \eqref{eq:interaction-bound}, we may return to \eqref{eq:energy-prelim} to obtain the estimate \begin{equation*}
\begin{split}
\nrm{ |v|^2 f(t,\cdot)}_{L^1} \le 2E(0) + C \exp(C\sigma t) + \frac14 \nrm{|v|^2 f}_{L^1}-2 \iint f \ln f \, dxdv
\end{split}
\end{equation*} which gives \begin{equation}\label{eq:energy-prelim2}
\begin{split}
\nrm{ |v|^2 f(t,\cdot)}_{L^1} \le C\left(E(0) + \exp(C\sigma t)\right) + 2 \iint f \ln \frac{1}{f} \chi_{0\le f \le 1} \, dxdv
\end{split}
\end{equation}
\medskip
\noindent \textbf{Estimate of the entropy and $x$ moment}. Recall that \begin{equation*}
\begin{split}
2 \iint f \ln \frac{1}{f} \chi_{0\le f \le 1} \, dx dv \le \frac14 \nrm{|v|^2 f}_{L^{1}} + C\nrm{|x|^2 f}_{L^{1}} + C.
\end{split}
\end{equation*} Inserting this estimate in \eqref{eq:energy-prelim2} gives the bound \begin{equation}\label{eq:energy-prelim3}
\begin{split}
\nrm{ |v|^2 f(t,\cdot)}_{L^1} \le C\left(1 + \exp(C\sigma t)\right) + C\nrm{|x|^2 f(t,\cdot)}_{L^{1}} .
\end{split}
\end{equation} On the other hand, we have \begin{equation*}
\begin{split}
\frac{d}{dt} \nrm{|x|^2 f(t,\cdot)}_{L^1} & = \iint x\cdot v f \, dx dv \le \left( \nrm{|x|^2 f(t,\cdot)}_{L^1} \nrm{|v|^2 f(t,\cdot)}_{L^1} \right)^{\frac12} \\
& \le C\nrm{|x|^2 f(t,\cdot)}_{L^1} + C\left(1 + \exp(C\sigma t)\right) + C\nrm{|x|^2 f(t,\cdot)}_{L^{1}} .
\end{split}
\end{equation*} We have used \eqref{eq:energy-prelim3} in the last inequality. This results in the estimate \begin{equation*}
\begin{split}
\nrm{|x|^2 f(t,\cdot)}_{L^{1}} \le C \exp( C_\sigma t ) .
\end{split}
\end{equation*} Returning to \eqref{eq:energy-prelim3}, we conclude the following kinetic energy bound \begin{equation*}
\begin{split}
\nrm{ |v|^2 f(t,\cdot)}_{L^1} \le C \exp( C_\sigma t ) .
\end{split}
\end{equation*}
\subsection{Moment estimate}
\end{comment}
%
%
%
%
%
%
\section{Singularity formation}\label{sec_blow}
\subsection{Proof of Theorem \ref{thm:blow}}
In this section, we give the details on the proof for Theorem \ref{thm:blow}. For this, we first start with the total energy estimate for the system \eqref{eq:VR}.
\begin{lemma}\label{lem_energy} Let $f$ be a strong solution to the system \eqref{eq:VR} satisfying $f \in X(T)$. Then we have for $t<T$
\[
E(t) + \sigma\int_0^t \iint_{\R^d \times \R^d} \frac1f |\nabla_v f + vf|^2\,dxdvds = E(0).
\]
\end{lemma}
\begin{proof} Straightforward computations give
\begin{align*}
\frac12\frac{d}{dt}\left(\iint_{\R^d \times \R^d} |v|^2 f \,dxdv\right)
&= \iint_{\R^d \times \R^d}( \nabla K\star\rho)\cdot v f \,dxdv - \sigma\iint_{\R^d \times \R^d} v \cdot (\nabla_v f + vf) \,dxdv.
\end{align*}
On the other hand, we get
\[
\frac{d}{dt}\left(\frac{1}{2}\int_{\R^d} \rho K\star\rho \,dx\right) = \int_{\R^d} \partial_t\rho K\star\rho \,dx = \int_{\R^d} \rho u \cdot \nabla K\star\rho \,dx = \iint_{\R^d \times \R^d}( \nabla K\star\rho)\cdot v f \,dxdv.
\]
This yields
\[
\frac{d}{dt}\left(\frac12\iint_{\R^d \times \R^d} |v|^2 f \,dxdv - \frac{1}{2}\int_{\R^d} \rho K\star\rho \,dx\right)= - \sigma\iint_{\R^d \times \R^d} v \cdot (\nabla_v f + vf) \,dxdv.
\]
We then combine the previous estimates with the following entropy estimate
\begin{align*}
\frac{d}{dt}\left(\iint_{\R^d \times \R^d} f \log f \,dxdv\right)
&= \iint_{\R^d \times \R^d} (\partial_t f) \log f\,dxdv\\
&= - \iint_{\R^d \times \R^d} ( \nabla K\star\rho)\cdot \nabla_v f \,dxdv -\sigma \iint_{\R^d \times \R^d} \frac{\nabla_v f}{f} \cdot (\nabla_v f + vf)\,dxdv
\end{align*}
to conclude the desired result.
\end{proof}
\begin{remark}\label{rmk_s0} In the absence of diffusion, i.e. $\sigma = 0$, we easily find from Lemma \ref{lem_energy} that
\[
\frac{d}{dt}\left(\frac12\iint_{\R^d \times \R^d} |v|^2 f \,dxdv - \frac{1}{2}\int_{\R^d} \rho K\star\rho \,dx\right) = 0.
\]
Thus if we define $\tilde E = \tilde E(t)$ as
\[
\tilde E = \frac12\iint_{\R^d \times \R^d} |v|^2 f \,dxdv - \frac{1}{2}\int_{\R^d} \rho K\star\rho \,dx,
\]
then $\tilde E(t) = \tilde E(0)$ for all $t\geq 0$.
\end{remark}
Then in the lemma below, we show the estimate on the second-order time-derivative of $I$.
\begin{lemma}\label{lem_mom} Let $f$ be a solution to the system \eqref{eq:VR} satisfying $f \in X(T)$. Then we have
\[
I''(t) = \iint_{\R^d \times \R^d} |v|^2 f\,dxdv -\frac12\sum_{i=1}^Nc_i \alpha_i\int_{\R^d} \rho K_i \star \rho\,dx - \sigma I'(t).
\]
\end{lemma}
\begin{proof} Straightforward computations gives
\begin{align*}
\frac12\frac{d^2}{dt^2}\iint_{\R^d \times \R^d} |x|^2f\,dxdv &= \frac{d}{dt}\iint_{\R^d \times \R^d} (x \cdot v) f \,dxdv \cr
&= - \iint_{\R^d \times \R^d} (x\cdot v) \left(\nabla_x \cdot (vf) + \nabla_v \cdot ((\nabla K\star\rho) f) - \sigma\nabla_v \cdot(\nabla_vf + vf) \right) dxdv\cr
&= \iint_{\R^d \times \R^d} |v|^2 f\,dxdv + \int_{\R^d} x \cdot (\nabla K\star\rho)\rho\,dx - \sigma \iint_{\R^d \times \R^d} (x \cdot v) f \,dxdv \cr
&= \iint_{\R^d \times \R^d} |v|^2 f\,dxdv -\frac12\sum_{i=1}^Nc_i \alpha_i\int_{\R^d} \rho K_i \star \rho\,dx - \sigma I'(t),
\end{align*}
where we used
\[
\iint_{\R^d \times \R^d} (x\cdot v) \Delta _v f\,dxdv = 0
\]
and
\begin{align*}
\int \rho x \cdot \nabla K\star\rho\,dx &= -\sum_{i=1}^Nc_i \alpha_i \iint_{\R^d \times \R^d} \rho(x) x \cdot \frac{x-y}{|x-y|^{\alpha_i+2}} \rho(y)\,dxdy\cr
&=\sum_{i=1}^Nc_i \alpha_i \iint_{\R^d \times \R^d} \rho(y) y \cdot \frac{x-y}{|x-y|^{\alpha_i+2}} \rho(x)\,dxdy\cr
&=-\frac12\sum_{i=1}^Nc_i \alpha_i \iint_{\R^d \times \R^d} \rho(x) \frac{1}{|x-y|^{\alpha_i}} \rho(y)\,dxdy\cr
&=-\frac12\sum_{i=1}^Nc_i \alpha_i\int_{\R^d} \rho K_i \star \rho\,dx.
\end{align*}
This implies
\[
I''(t) = \iint_{\R^d \times \R^d} |v|^2 f\,dxdv -\frac12\sum_{i=1}^Nc_i \alpha_i\int_{\R^d} \rho K_i \star \rho\,dx - \sigma I'(t) \qedhere
\]
\end{proof}
Using the above estimates on the momentum of inertia, we obtain Theorem \ref{thm:blow} in the case $\sigma = 0$.
\begin{proof}[Proof of Theorem \ref{thm:blow} in the case $\sigma = 0$] It follows from Lemma \ref{lem_mom} that
\[
\frac12\frac{d^2}{dt^2}\iint_{\R^d \times \R^d} |x|^2f\,dxdv = \iint_{\R^d \times \R^d} |v|^2 f\,dxdv - \frac12\sum_{i=1}^Nc_i \alpha_i\int_{\R^d} \rho K_i \star \rho\,dx.
\]
This together with Remark \ref{rmk_s0} yields
\begin{equation}\label{est_mom}
\frac12\frac{d^2}{dt^2}\iint_{\R^d \times \R^d} |x|^2f\,dxdv = 2\tilde E(0) + \sum_{i=1}^Nc_i \left(1 - \frac{\alpha_i}2\right)\int_{\R^d} \rho K_i\star\rho\,dx.
\end{equation}
Since $\max_{i=1,\dots,N}\alpha_i > 2$, without loss of generality, we may assume $\alpha_1 = \max_{i=1,\dots,N}\alpha_i > 2$. Note that if $\alpha_i > \alpha_j$, then for any $\varepsilon >0$,
\begin{align*}
\int_{\R^d} \rho K_j\star\rho\,dx &= \left(\int_{|x-y|\leq \varepsilon} + \int_{|x-y|\geq \varepsilon}\right) \rho(x) \frac{1}{|x-y|^{\alpha_j}} \rho(y)\,dxdy\cr
&\leq \varepsilon^{\alpha_i - \alpha_j} \int_{|x-y|\leq \varepsilon} \rho(x) \frac{1}{|x-y|^{\alpha_i}} \rho(y)\,dxdy + \frac1{\varepsilon^{\alpha_j}}\|\rho\|_{L^1}^2\cr
&\leq \frac1{\varepsilon^{\alpha_j}}\left( \varepsilon^{\alpha_i} \int_{\R^d} \rho K_i\star\rho\,dx + 1\right).
\end{align*}
Using the above observation, we estimate
\begin{align*}
&\sum_{i=1}^N c_i \left(1 - \frac{\alpha_i}2\right)\int_{\R^d} \rho K_i\star\rho\,dx \cr
&\quad = \left(\sum_{\alpha_i > 2} + \sum_{\alpha_i < 2}\right) c_i \left(1 - \frac{\alpha_i}2\right)\int_{\R^d} \rho K_i\star\rho\,dx\cr
&\quad \leq c_1 \left(1 - \frac{\alpha_1}2\right)\int_{\R^d} \rho K_1\star\rho\,dx + \sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\int_{\R^d} \rho K_i\star\rho\,dx\cr
&\quad \leq c_1 \left(1 - \frac{\alpha_1}2\right)\int_{\R^d} \rho K_1\star\rho\,dx + \sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\frac1{\varepsilon^{\alpha_i}}\left(\varepsilon^{\alpha_1}\int_{\R^d} \rho K_1\star\rho\,dx + 1\right)\cr
&\quad \leq c_1 \left(1 - \frac{\alpha_1}2\right)\int_{\R^d} \rho K_1\star\rho\,dx + \sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\frac1{\varepsilon^2}\left(\varepsilon^{\alpha_1}\int_{\R^d} \rho K_1\star\rho\,dx + 1\right).
\end{align*}
We now choose $\varepsilon>0$ such that
\[
\varepsilon = \left(c_1 \left(\frac{\alpha_1}2 - 1\right) \left(\sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\right)^{-1}\right)^{\frac1{\alpha_1-2}},
\]
then this gives
\[
\sum_{i=1}^N c_i \left(1 - \frac{\alpha_i}2\right)\int_{\R^d} \rho K_i\star\rho\,dx \leq \left( \sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\right)^{1 - \frac{2}{\alpha_1 - 2}} \left(c_1 \left(\frac{\alpha_1}2 - 1\right)\right)^{\frac2{\alpha_1-2}}.
\]
We then combine the above and \eqref{est_mom} to get
\[
\frac12\frac{d^2}{dt^2}\iint_{\R^d \times \R^d} |x|^2f\,dxdv \leq 2\tilde E(0) + \left( \sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\right)^{1 - \frac{2}{\alpha_1 - 2}} \left(c_1 \left(\frac{\alpha_1}2 - 1\right)\right)^{\frac2{\alpha_1-2}}.
\]
This shows that if the right hand side of the above is negative, i.e.
\[
\iint_{\R^d \times \R^d} |v|^2 f_0 \,dxdv + \left( \sum_{\alpha_i < 2}c_i \left(1 - \frac{\alpha_i}2\right)\right)^{1 - \frac{2}{\alpha_1 - 2}} \left(c_1 \left(\frac{\alpha_1}2 - 1\right)\right)^{\frac2{\alpha_1-2}} < \int_{\R^d} \rho_0 K\star\rho_0 \,dx,
\]
then the momentum of inertia should be negative in a finite time. Hence the life-span $T$ of the solution should be finite.
\end{proof}
We next focus on the case with the linear Fokker--Planck operator. In this case, the total energy includes the entropy term.
\begin{lemma}\label{lem_mom2} Let $f$ be a solution to the system \eqref{eq:VR} satisfying $f \in X(T)$. Then we have
\[
I''(t) \leq 2(1+\delta)E(0) + C_0 - 2(1+\delta)\iint_{\R^d \times \R^d} f \ln f\chi_{0 \leq f \leq 1}\,dxdv - \sigma I'(t) - \delta \iint_{\R^d \times \R^d} |v|^2 f\,dxdv
\]
for some $\delta > 0$ satisfying $2(1+\delta) < \max_{i=1,\dots, N} \alpha_i$, where $C_0 > 0$ is given by
\[
C_0 = \left( \sum_{\alpha_i < 2(1+\delta)}c_i \left(1 + \delta - \frac{\alpha_i}2\right)\right)^{1 - \frac{2}{\alpha_1 - 2}} \left(c_1 \left(\frac{\alpha_1}2 - 1 - \delta\right)\right)^{\frac2{\alpha_1-2}}.
\]
\end{lemma}
\begin{proof}
By Lemmas \ref{lem_energy} and \ref{lem_mom}, we deduce
\begin{align*}
I''(t) &= (1+\delta) \iint_{\R^d \times \R^d} |v|^2 f\,dxdv -\frac12\sum_{i=1}^Nc_i \alpha_i\int_{\mathbb{R}^d} \rho K_i \star \rho\,dx - \sigma I'(t) - \delta \iint_{\R^d \times \R^d} |v|^2 f\,dxdv\cr
&=2(1+\delta)E(t) + \sum_{i=1}^Nc_i \left(1+\delta - \frac{\alpha_i}2\right)\int_{\R^d} \rho K_i\star\rho\,dx - 2(1+\delta)\iint_{\R^d \times \R^d} f \ln f\,dxdv\cr
&\quad - \sigma I'(t) - \delta \iint_{\R^d \times \R^d} |v|^2 f\,dxdv,
\end{align*}
where $\delta >0$ is chosen such that $2(1+\delta) < \max_{i=1,\dots, N} \alpha_i$. Then similarly as in the proof of Theorem \ref{thm:blow} in the case $\sigma = 0$, we estimate
\begin{align*}
\sum_{i=1}^N c_i \left(1 +\delta - \frac{\alpha_i}2\right)\int_{\R^d} \rho K_i\star\rho\,dx &\leq \left( \sum_{\alpha_i < 2(1+\delta)}c_i \left(1 + \delta - \frac{\alpha_i}2\right)\right)^{1 - \frac{2}{\alpha_1 - 2}} \left(c_1 \left(\frac{\alpha_1}2 - 1 - \delta\right)\right)^{\frac2{\alpha_1-2}} \cr
&= C_0.
\end{align*}
Here we again set $\alpha_1 = \max_{i=1,\dots, N} \alpha_i$ without loss of generality. Thus we obtain
\begin{align*}
I''(t) &\leq 2(1+\delta)E(0) + C_0 - 2(1+\delta)\iint_{\R^d \times \R^d} f (\ln f) \chi_{0 \leq f \leq 1}\,dxdv - \sigma I'(t) - \delta \iint_{\R^d \times \R^d} |v|^2 f\,dxdv.
\end{align*}
\end{proof}
In order to handle the entropy term appeared in Lemma \ref{lem_mom2}, we modify the classical lemma \cite{CIP94} to suit our methodology.
\begin{lemma}\label{lem_log} For given $\delta > 0$, there exists $C_\delta >0$, independent of $f \geq 0$, such that
\[
- 2(1+\delta)\iint_{\R^d \times \R^d} f (\ln f) \chi_{0 \leq f \leq 1}\,dxdv \leq C_\delta(I+1) + \delta \iint_{\R^d \times \R^d} |v|^2 f\,dxdv,
\]
where $I = \frac12\iint_{\R^d \times \R^d} |x|^2 f\,dxdv$ and the constant $C_\delta$ is given by
\[
C_\delta = 4(1+\delta)(1 + \delta^{-1})^{\frac d{2+d}} (e^{-1} 2^{3d} \pi^{2d})^{\frac1{2+d}}.
\]
\end{lemma}
\begin{proof}Note that the following holds for any $s, \sigma \geq 0$:
\begin{align*}
-s\ln(s) \chi_{0 \leq s \leq 1} &= -s\ln(s) \chi_{e^{-\sigma} \leq s \leq 1} - s\ln(s) \chi_{e^{-\sigma} \geq s}\cr
&\leq s\sigma + 2e^{-2}\sqrt s \chi_{e^{-\sigma} \geq s}\cr
&\leq s\sigma + 2e^{-2} e^{-\sigma/2},
\end{align*}
where we used
\[
\sup_{s \in (0,1)} -s\ln(s) \leq 2e^{-2}.
\]
We then take
\[
s = f \quad \mbox{and} \quad \sigma = |x|^2 \epsilon_1 + |v|^2\epsilon_2,
\]
where $\epsilon_i > 0$, $i=1,2$ will be determined later. Thus, we have
\begin{align*}
& - 2(1+\delta)\iint_{\R^d \times \R^d} f (\ln f) \chi_{0 \leq f \leq 1}\,dxdv \cr
&\quad \leq 2(1+\delta)\iint_{\R^d \times \R^d} \left(|x|^2 \epsilon_1 + |v|^2\epsilon_2 \right)f\,dxdv + 4e^{-2}(1 + \delta)\iint_{\R^d \times \R^d} \exp\left(- \frac{|x|^2 \epsilon_1 + |v|^2\epsilon_2}2 \right)dxdv\cr
&\quad = 4(1+\delta)\epsilon_1 I(t) + 2(1+\delta)\epsilon_2\iint_{\R^d \times \R^d} |v|^2 f\,dxdv + 4e^{-2}(1 + \delta)\frac{(2\pi)^d}{(\epsilon_1 \epsilon_2)^{d/2}}.
\end{align*}
We finally choose $\epsilon_i>0$, $i=1,2$ so that
\[
2(1+\delta)\epsilon_2 = \delta \quad \mbox{and} \quad 4(1+\delta)\epsilon_1 = 4e^{-2}(1 + \delta)\frac{(2\pi)^d}{(\epsilon_1 \epsilon_2)^{d/2}}
\]
to conclude the desired result.
\end{proof}
By combining Lemmas \ref{lem_mom} and \ref{lem_log}, we estimate
\[
I''(t) \leq 2(1+\delta)E(0) + C_0 + C_\delta + C_\delta I(t) - \sigma I'(t).
\]
Our main strategy is to show that $I(t)$ can be negative in a finite time under certain assumption on the initial data, and thus it leads to a contradiction. For this, we need to have the following Gr\"onwall-type lemma.
\begin{lemma}\label{lem_diff} Let $h = h(t)$ be a nonnegative $\mathcal C^2$-function satisfying the following differential inequality:
\begin{equation}\label{eq_h}
h''(t) + c_1 h'(t) \leq c_2 h(t) + c_3, \quad h(0) = h_0, \quad h'(0) = h'_0
\end{equation}
for some $c_i >0$, $i=1,2$ and $c_3 \in \mathbb{R}$. Then we have
\begin{align*}
h(t) &\leq \left( h_0 + \frac{c_3}{\beta(\beta + c_1)} + \frac{1}{c_1 + 2\beta}\left(h'_0 - \beta h_0 - \frac{c_3}{\beta + c_1}\right) \right)e^{\beta t}\cr
&\quad - \frac{1}{c_1 + 2\beta}\left(h'_0 - \beta h_0 - \frac{c_3}{\beta + c_1} \right) e^{-(c_1 + \beta)t} - \frac{c_3}{\beta (c_1 + \beta)},
\end{align*}
where $\beta > 0$ is given by
\begin{equation}\label{beta}
\beta := \frac{-c_1 + \sqrt{c_1^2 + 4c_2}}{2}.
\end{equation}
\end{lemma}
\begin{proof} Set
\[
\tilde h(t) := h'(t) - \beta h(t) = e^{\beta t} \left(h(t) e^{-\beta t}\right)',
\]
then it follows from \eqref{eq_h} that
\[
\tilde h' = h'' - \beta h' \leq -(c_1 + \beta)\left(h' - \frac{c_2}{c_1 + \beta}h\right) + c_3.
\]
Since $\beta$ given by \eqref{beta} satisfies
\[
\beta = \frac{c_2}{c_1 + \beta},
\]
we obtain
\[
\tilde h' \leq -(c_1 + \beta)\tilde h + c_3.
\]
Solving the above implies
\[
e^{\beta t} \left(h(t) e^{-\beta t}\right)' = \tilde h(t) \leq \tilde h(0) e^{-(c_1 + \beta)t} + \frac{c_3}{c_1 + \beta}\left(1 - e^{-(c_1 + \beta)t} \right),
\]
and again solving the resulting differential inequality yields the desired result.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:blow} in the case $\sigma > 0$] We use Lemma \ref{lem_diff} with $h = I$, $c_1 = \sigma > 0$, $c_2 = C_\delta > 0$, and $c_3 = 2(1+\delta)E(0) + C_0 + C_\delta$ to obtain
\begin{align*}
I(t) &\leq \left( I(0) + \frac{2(1+\delta)E(0) + C_0 +C_\delta}{\beta(\beta + \sigma)} + \frac{1}{\sigma + 2\beta}\left(I'(0) - \beta I(0) - \frac{2(1+\delta)E(0)+ C_0 + C_\delta}{\beta + \sigma}\right) \right)e^{\beta t}\cr
&\quad - \frac{1}{\sigma + 2\beta}\left(I'(0) - \beta I(0) - \frac{2(1+\delta)E(0)+ C_0+C_\delta}{\beta + \sigma} \right) e^{-(\sigma + \beta)t} - \frac{2(1+\delta)E(0) + C_0 + C_\delta}{\beta (\sigma + \beta)},
\end{align*}
where
\[
\beta = \frac{-\sigma + \sqrt{\sigma^2 + 4C_\delta}}{2}.
\]
Since $\beta>0$, this implies that if
\[
I(0) + \frac{2(1+\delta)E(0) + C_0 +C_\delta}{\beta(\beta + \sigma)} + \frac{1}{\sigma + 2\beta}\left(I'(0) - \beta I(0) - \frac{2(1+\delta)E(0) + C_0 + C_\delta}{\beta + \sigma}\right) < 0,
\]
which is equivalent to
\[
2(1+\delta)E(0) + \beta I'(0) < -\beta(\sigma + \beta) I(0) - C_\delta - C_0,
\]
then the life-span $T$ of solutions should be finite.
\end{proof}
\subsection{Further discussion: repulsive and attractive interactions}\label{subsec:mixed}
In this part, we show the singularity formation for the system \eqref{eq:VR} with the repulsive-attractive singular interaction potentials. More precisely, let us consider the potential $K$ given by
\begin{equation}\label{pot_K}
K(x) = \frac{1}{|x|^{\alpha_1}} - \frac{1}{|x|^{\alpha_2}} =: K_a(x) + K_r(x),
\end{equation}
where $0< \alpha_1 \neq \alpha_2 < d$.
In this case, we estimate
\[
\int_{\R^d} \rho x \cdot \nabla K\star\rho\,dx=-\frac{\alpha_1}2 \int_{\R^d} \rho K_a \star \rho\,dx + \frac{\alpha_2}2 \int_{\R^d} \rho K_r \star \rho\,dx.
\]
Thus,
\begin{align*}
I''(t) &=2(1+\delta)E(t) + \left(1+\delta - \frac{\alpha_1}2\right)\int_{\R^d} \rho K_a\star\rho\,dx - \left(1+\delta - \frac{\alpha_2}2\right)\int_{\R^d} \rho K_r\star\rho\,dx \cr
&\quad - 2(1+\delta)\iint_{\R^d \times \R^d} f \ln f\,dxdv - \sigma I'(t) - \delta \iint_{\R^d \times \R^d} |v|^2 f\,dxdv.
\end{align*}
Let us denote the sum of the second and third terms on the right hand side by $I_K$.
Analogous to that of Theorem \ref{thm:blow}, we assume that the exponent $\alpha_1$ in the attractive potential satisfies $\alpha_1 > 2$. We then consider two cases: $\alpha_2 < 2$ and $\alpha_2 \geq 2$.
In the first case, we can choose $\delta > 0$ such that $\frac{\alpha_1}2 - 1 \geq \delta > 0 > \frac{\alpha_2}2 - 1$. This implies $1+\delta - \frac{\alpha_2}2 \geq 0$ and $1+\delta - \frac{\alpha_1}2 \leq 0$. Thus, $I_K \leq 0$.
On the other hand, if $\alpha_2 \geq 2$, we further assume $\alpha_1 > \alpha_2$, then
\begin{align*}
\int_{\R^d} \rho K_r\star\rho\,dx &= \left(\int_{|x-y|\leq 1} + \int_{|x-y|\geq 1}\right) \rho(x) \frac{1}{|x-y|^{\alpha_2}} \rho(y)\,dxdy\cr
&\leq \int_{|x-y|\leq 1} \rho(x) \frac{1}{|x-y|^{\alpha_1}} \rho(y)\,dxdy + \|\rho\|_{L^1}^2\cr
&\leq \int_{\R^d} \rho K_a\star\rho\,dx + 1.
\end{align*}
Thus,
\begin{align*}
I_K &\leq \left(1+\delta - \frac{\alpha_1}2\right)\int_{\R^d} \rho K_a\star\rho\,dx - \left(1+\delta - \frac{\alpha_2}2\right)\left(\int_{\R^d} \rho K_a\star\rho\,dx + 1\right)\cr
&\leq -\frac{\alpha_1 - \alpha_2}{2} \int_{\R^d} \rho K_a\star\rho\,dx +\frac{\alpha_2}2 -1 -\delta\cr
&\leq \frac{\alpha_2}2 -1 -\delta.
\end{align*}
In summary, one can deduce the following result.
\begin{theorem}\label{thm:blow2} Let $T>0$ and $d \geq 3$, and let $f$ be a solution to the system \eqref{eq:VR} satisfying $f \in X(T)$. Consider the interaction potential $K$ given as \eqref{pot_K}.
\begin{itemize}
\item (Vlasov equation: $\sigma =0$) Suppose $\alpha_1 \geq \max\{2, \alpha_2\}$ and
\[
\iint_{\R^d \times \R^d} |v|^2 f_0 \,dxdv < \int_{\R^d} \rho_0 K\star\rho_0 \,dx + \left( \frac{\alpha_2}2 - 1\right){\bf 1}_{\{\alpha_2 \geq 2\}}.
\]
\item (Vlasov--Fokker--Planck equation: $\sigma>0$) Suppose $\alpha_1 > \max\{2, \alpha_2\}$ and the initial total energy and radial-weighted momentum is sufficiently small compared to the initial momentum of inertia so that
\[
2(1+\delta)E(0) + \beta I'(0) < -\beta(\sigma + \beta) I(0) - \left(\frac{\alpha_2}2 -1 -\delta\right){\bf 1}_{\{\alpha_2 \geq 2\}}.
\]
\end{itemize}
Then the life-span $T$ of the solution should be finite.
\end{theorem}
\begin{remark}One may extend Theorem \ref{thm:blow2} to the case where the interaction potential $K$ is given by
\[
K(x) = \sum_{i=1}^N \frac{c_i}{|x|^{\alpha_i}}, \quad \alpha_i \in (0,d), \ \ c_i \in \mathbb{R}, \quad i=1,\dots, N.
\]
\end{remark}
\subsection*{Acknowledgments}
YPC has been supported by NRF grant (No. 2017R1C1B2012918) and Yonsei University Research Fund of 2021-22-0301. IJJ has been supported by the New Faculty Startup Fund from Seoul National University and the National Research Foundation of Korea grant (No. 2019R1F1A1058486).
|
\section{Introduction}
The systematic study of one-loop Feynman integrals in perturbative quantum field theories dates back to the end of the 1970s when 't~Hooft and Veltman \cite{tHooft:1978jhc} calculated the generic one-, two-, three- and four-point scalar integrals in dimensional regularization (DREG) up to order $\epsilon^0$ , where $\epsilon=(4-d)/2$ with spacetime dimension $d$. Passarino and Veltman \cite{Passarino:1978jh} then demonstrated that tensor integrals up to four points can be systematically reduced to scalar ones, and later it was shown \cite{Bern:1992em, Bern:1993kr} that integrals with more than four external legs in $4-2\epsilon$ dimensions can be expressed as lower-point ones up to order $\epsilon^0$. These developments in principle solved the problem of next-to-leading order (NLO) calculations for tree-induced scattering processes.
The improvements of experimental precision and the progress of theoretical studies require the understanding of scattering amplitudes and cross sections at higher orders in perturbation theory. In this respect, we need to compute the one-loop integrals to higher orders in $\epsilon$. These allow us to predict the infrared divergences appearing in the two-loop amplitudes \cite{Catani:1998bh, Aybat:2006wq, Aybat:2006mz, Sterman:2002qn, Becher:2009cu, Becher:2009qa, Becher:2009kw, Ferroglia:2009ep, Ferroglia:2009ii}, and are also necessary for computing the one-loop squared amplitudes which are essential ingredients of next-to-next-to-leading order (NNLO) cross sections.
Unlike the terms up to order $\epsilon^0$, generic results for the higher order terms are not available yet. Part of the reason is that integrals with more than four external legs are in general not reducible to lower-point ones when considering higher orders in $\epsilon$. These require further calculations which are often rather complicated due to the increasing number of physical scales involved.
It is known \cite{Bourjaily:2019exo, Herrmann:2019upk, Chen:2020uyk} that one-loop integrals in a given family admit a uniform transcendentality (UT) basis satisfying canonical differential equations of the form \cite{Henn:2013pwa}
\begin{equation}
d\vec{f}(\vec{x},\epsilon) = \epsilon \, d\bm{A}(\vec{x}) \, \vec{f}(\vec{x},\epsilon) \, ,
\label{eq:can_de}
\end{equation}
where $\vec{x}$ is the set of independent kinematic variables, and the matrix $d\bm{A}$ takes the $d\log$-form:
\begin{equation}
d\bm{A}(\vec{x}) = \sum_i \bm{C}_i \, d\log(W_i(\vec{x})) \, .
\end{equation}
In the above expression, $\bm{C}_i$ are matrices consisting of rational numbers, and $W_i(\vec{x})$ are algebraic functions of the variables. The functions $W_i$ are called the ``letters'' for this integral family, and the set of all independent letters is called the ``alphabet''.
At one-loop, a canonical basis can be generically constructed by looking for $d\log$-form integrands \cite{ArkaniHamed:2010gh, Gehrmann:2011xn, Drummond:2013nda, Henn:2013pwa, Arkani-Hamed:2014via, Bern:2014kca, Herrmann:2019upk, Bourjaily:2019exo, Chen:2020uyk, Chen:2022lzr}. On the other hand, obtaining the $d\log$ matrix $d\bm{A}(\vec{x})$ is not always a trivial task when the number of variables is large. We note that the $d\log$ matrix can be easily reconstructed if we have the knowledge of the alphabet $\{W_i(\vec{x})\}$ in advance, since the coefficient matrices $\bm{C}_i$ can then be obtained by bootstrapping.
Having the alphabet (and hence the matrix $d\bm{A}(\vec{x})$) in a good form also helps to solve the differential equations \eqref{eq:can_de} order-by-order in the dimensional regulator $\epsilon$. The (suitably normalized) solution can be written as a Taylor series:
\begin{equation}
\vec{f}(\vec{x},\epsilon) = \sum_{n=0}^{\infty} \epsilon^n \, \vec{f}^{(n)}(\vec{x}) \, ,
\end{equation}
where the $n$th order coefficient function can be written as a Chen iterated integral \cite{Chen:1977oja}
\begin{align}
\vec{f}^{(n)}(\vec{x}) =& \int_{\vec{x}_0}^{\vec{x}} d\bm{A}(\vec{x}_n) \cdots \int_{\vec{x}_0}^{\vec{x}_2} d\bm{A}(\vec{x}_1)
+ \vec{f}^{(n)}(\vec{x}_0) \, .
\end{align}
Such iterated integrals can be analyzed using the language of ``symbols'' \cite{Brown:2009qja, Goncharov:2010jf, Duhr:2011zq} that encodes the algebraic properties of the resulting functions. In certain cases, these iterated integrals can be solved analytically (either by direct integration or by bootstrapping). The results can often be written in terms of generalized polylogarithms (GPLs) \cite{Goncharov:1998kja} which allow efficient numeric evaluation~\cite{Vollinga:2004sn, Naterop:2019xaf, Wang:2021imw}. When an analytic solution is not available, it is straightforward to evaluate them numerically either by numerical integration or by a series expansion~\cite{Moriello:2019yhu, Hidding:2020ytt}.
In this paper, we describe a generic method to construct the letters systematically from cut integrals in the Baikov representation \cite{Baikov:1996iu, Frellesvig:2017aai}. The letters can be generically written in terms of various Gram determinants. The letters and symbols of one-loop integrals have been considered in \cite{Arkani-Hamed:2017ahv, Abreu:2017ptx, Abreu:2017enx, Abreu:2017mtm}, and our method is similar to that in \cite{Abreu:2017ptx, Abreu:2017enx, Abreu:2017mtm}. Nevertheless, we evaluate the cut integrals differently and obtain equivalent but simpler expressions in certain cases utilizing properties of Gram determinants. Furthermore, we consider the cases of divergent cut integrals which were ignored in earlier studies. Using our results, it is easy to write down all letters for a given integral family even before constructing the differential equations. These letters will also appear in the corresponding two-loop integrals.
\section{The canonical basis of one-loop integrals}
We use the method of \cite{Chen:2020uyk, Chen:2022lzr} to construct the canonical basis in the Baikov representation. In this section, we give a brief review of the construction procedure since it will also be relevant for obtaining the alphabet in the matrices $d\bm{A}(\vec{x})$.
Consider a generic one-loop integral topology with $N = E + 1$ external legs, where $E$ is the number of independent external momenta. Integrals in this topology can be written as
\begin{align}
\label{eq:intI}
I_{a_1, \cdots, a_N} = \int \frac{d^dl}{i \pi^{d/2}} \frac{1}{z_1^{a_1} z_2^{a_2}\cdots z_N^{a_N}} \,,
\end{align}
where $z_i$ are the propagator denominators given by
\begin{gather}
\label{eq:propagator_general}
z_1=l^2-m_1^2 \,, \quad z_2= (l+p_1)^2-m_2^2 \,, \quad \cdots \,, \nn \\
\quad z_N= (l+p_1+\cdots+p_{E})^2-m_N^2 \,.
\end{gather}
Here $p_1,\ldots,p_E$ are external momenta which we assume to span a space-like subspace of the $d$-dimensional Minkowski spacetime. This corresponds to the so-called (unphysical) Euclidean kinematics. Results in the physical phase-space region can be defined by analytic continuation.
The idea of the Baikov representation is to change the integration variables from loop momenta $l^\mu$ to the Baikov variables $z_i$, and the result is given by
\begin{multline}
I_{a_1,\ldots,a_N} = \frac{1}{(4 \pi)^{E/2} \, \Gamma\big((d-E)/2\big)} \\
\times \int_{\mathcal{C}} \frac{ \left| G_N(\bm{z}) \right|^{(d-E-2)/2}}{\left| \mathcal{K}_N \right|^{(d-E-1)/2} } \prod_{i=1}^{N} \frac{dz_i}{z_i^{a_i}} \, ,
\label{eq:1loopbaikov}
\end{multline}
where $\bm{z}=\{z_1,\ldots,z_N\}$ is the collection of the Baikov variables. The function $G_N(\bm{z})$ is a polynomial of the $N$ variables, while $\mathcal{K}_N$ is independent of $\bm{z}$. They are given by
\begin{equation}
G_N(\bm{z}) \equiv G(l,p_1,\ldots,p_E) \, , \quad \mathcal{K}_N = G(p_1, \cdots, p_{E}) \, ,
\label{eq:GN}
\end{equation}
where the Gram determinant is defined as
\begin{equation}
\label{eq:gram1}
G(q_1,\ldots,q_n) \equiv \det
\begin{pmatrix}
q_1 \cdot q_1 & q_1 \cdot q_2 & \cdots & q_1 \cdot q_n
\\
q_2 \cdot q_1 & q_2 \cdot q_2 & & \vdots
\\
\vdots & & \ddots & \vdots
\\
q_n \cdot q_1 & \cdots & \cdots & q_n \cdot q_n
\end{pmatrix}
\, .
\end{equation}
Note that in Eq.~\eqref{eq:GN}, the scalar products involving the loop momentum $l$ should be re-expressed in terms of $\bm{z}$:
\begin{align}
l^2 &= z_1 + m_0^2 \, , \nonumber
\\
l \cdot p_i &= \frac{z_{i+1} + m_{i+1}^2 - p_i^2 - z_i - m_i^2}{2} - \sum_{j=1}^{i-1} p_i \cdot p_j \, .
\end{align}
The integration domain $\mathcal{C}$ in Eq.~\eqref{eq:1loopbaikov} is determined by the condition $G_N(\bm{z})/\mathcal{K}_N \leq 0$ with Euclidean kinematics.
We are now ready to write down the UT integrals $g_N$ for any $N$ according to \cite{Chen:2020uyk}. We need to distinguish between the cases of odd $N$ and even $N$:
\begin{align}
g_{N} \big|_{\text{$N$ odd}} &= \frac{\epsilon^{(N+1)/2}}{(4\pi)^{(N-1)/2} \, \Gamma(1-\epsilon)} \nn \\
&\times \int \left( -\frac{\mK_N}{G_N(\bm{z})} \right)^\epsilon \prod_{i=1}^N \frac{dz_i}{z_i} \, , \nn
\\
g_{N} \big|_{\text{$N$ even}} &= \frac{\epsilon^{N/2}}{(4\pi)^{(N-1)/2} \, \Gamma(1/2-\epsilon)} \nn \\
&\times \int \frac{\sqrt{G_N(\bm{0})}}{\sqrt{G_N(\bm{z})}} \left( -\frac{\mK_N}{G_N(\bm{z})} \right)^\epsilon \prod_{i=1}^N \frac{dz_i}{z_i} \, ,
\label{eq:UTint}
\end{align}
where we set $\mK_1 = 1$, and $\bm{0}$ means that all $z_i$'s are zero.
Note that $g_{2n-1}$ and $g_{2n}$ can be naturally identified as Feynman integrals in $2n-2\epsilon$ dimensions:
\begin{align}
g_N \big|_{N=2n-1} &= \epsilon^n \sqrt{\mK_N} \, I_{1 \times N}^{(2n-2\epsilon)} \,, \nn
\\
g_N \big|_{N=2n} &= \epsilon^n \sqrt{G_N(\bm{0})} \, I_{1 \times N}^{(2n-2\epsilon)} \,,
\label{eq:UTMIs}
\end{align}
where $I^{(d)}_{1 \times N}$ denotes the $d$-dimensional $N$-point Feynman integral with all powers $a_i = 1$:
\begin{align}
I^{(d)}_{1 \times N} \equiv \int \frac{d^dl}{i \pi^{d/2}} \frac{1}{z_1 z_2 \cdots z_N} \,.
\end{align}
They can be related to Feynman integrals in $4-2\epsilon$ dimensions using the dimensional recurrence relations \cite{Tarasov:1996br, Lee:2009dh}. Applying the above to all sectors of a family, we build a complete canonical basis satisfying $\epsilon$-form differential equations.
\section{Letters in differential equations: convergent cases}
\label{sec:letter}
Given a basis of Feynman integrals, it is straightforward to calculate the derivatives with respect to some kinematic variable $x_i$. For a UT basis $\vec{f}(\vec{x},\epsilon)$ we write
\begin{equation}
\frac{\partial}{\partial x_i} \vec{f}(\vec{x},\epsilon) = \epsilon \, \bm{A}_i(\vec{x}) \, \vec{f}(\vec{x},\epsilon) \, ,
\end{equation}
where the elements in the matrix $\bm{A}_i(\vec{x})$ have the property that they only contain simple poles. In principle, one may already attempt to solve these differential equations by direct integration. However, this is usually rather difficult when $\bm{A}_i(\vec{x})$ contains many irrational functions (square roots). Therefore it is often very useful to combine the partial derivatives into a total derivative, and rewrite the differential equations in the form of Eq.~\eqref{eq:can_de}. To do that it is important to know the alphabet (i.e., the set of independent letters $W_i(\vec{x})$) in the matrix $d\bm{A}(\vec{x})$. With the knowledge of the alphabet, it is straightforward to reconstruct the whole matrix $d\bm{A}(\vec{x})$ by comparing the coefficients in the partial derivatives.
In principle, one may obtain the letters by directly integrating the matrices $\bm{A}_i(\vec{x})$ over the variables $x_i$, and manipulating the resulting expressions. However, in the presence of many square roots (containing high-degree polynomials) in multi-scale problems, these integrations are not easy to perform, and the results are often extremely complicated. One may find examples in various one-loop and multi-loop calculations, e.g., Refs.~\cite{Heller:2019gkq, Bonciani:2019jyb, Chen:2022nxt}.
With such kind of expressions, it is highly non-trivial to decide whether a set of letters are independent or not.\footnote{There is a package \texttt{SymBuild} \cite{Mitev:2018kie} which can carry out such a task, but the computational burden is rather heavy when there are many square roots.} Furthermore, from experience we know that letters involving square roots can often be written in the form
\begin{equation}
\frac{P(\vec{x}) - \sqrt{Q(\vec{x})}}{P(\vec{x}) + \sqrt{Q(\vec{x})}} \, ,
\end{equation}
where $P$ and $Q$ are polynomials. Such letters have nice properties under analytic continuation: they are real when $Q(\vec{x}) > 0$, and become pure phases when $Q(\vec{x}) < 0$. But it is not easy to recover this structure from direct integration.
Given the above considerations, we now describe a novel method to obtain the letters, especially those with square roots and multiple scales. Our method is based on the $d\log$-form integrals in the Baikov representation under various cuts. We will work with the generic propagator denominators in Eq.~\eqref{eq:propagator_general} and the Baikov representation \eqref{eq:1loopbaikov}. Without loss of generality, we define the Baikov cut on the first $r$ variable $z_1,\ldots,z_r$ by \cite{Frellesvig:2017aai}
\begin{multline}
\label{eq:baikov_cut}
I_{a_1,\ldots,a_N}\big|_{\text{$r$-cut}} = \frac{1}{(4 \pi)^{E/2} \, \Gamma((d-E)/2)} \\
\times \int \prod_{j=r+1}^N \frac{dz_j}{z_j^{a_j}} \prod_{i=1}^r \oint_{z_i=0} \frac{dz_i}{z_i^{a_i}} \frac{| G_N(\bm{z}) |^{(d-E-2)/2}}{|\mathcal{K}_N|^{(d-E-1)/2} } \, .
\end{multline}
An important property of the Baikov cut is that if one of the powers $a_i$ $(1 \leq i \leq r)$ is non-positive, the cut integral vanishes according to the residue theorem. The coefficient matrices in the differential equations are invariant under the cuts, and we will utilize this fact to obtain the letters by imposing various cuts.
To begin with, we write the differential equation satisfied by an $N$-point one-loop UT integral $g_N$ (see Eqs.~\eqref{eq:UTint} and \eqref{eq:UTMIs}) as
\begin{align}
\label{eq:DE-letter}
dg_N(\vec{x},\epsilon) &= \epsilon \, dM_N(\vec{x}) \, g_N(\vec{x},\epsilon) \nn \\
&+ \epsilon \sum_{m < N} \sum_i dM_{N,m}^{(i)}(\vec{x}) \, g_m^{(i)}(\vec{x},\epsilon) \, ,
\end{align}
where $g_N(\vec{x},\epsilon)$ and $g_m^{(i)}(\vec{x},\epsilon)$ are components of the canonical basis $\vec{f}(\vec{x},\epsilon)$, while $dM_N(\vec{x})$ and $dM_{N,m}^{(i)}(\vec{x})$ are entries in the matrix $d\bm{A}(\vec{x})$. It should be clear from the above equation that the derivative of $g_N$ cannot depend on higher-point integrals, and cannot depend on other $N$-point integrals as well. It may depend on several $m$-point integrals for each $m < N$, and we use a superscript like in $g_m^{(i)}$ and $dM_{N,m}^{(i)}$ to distinguish them. These $m$-point integrals can be obtained by ``squeezing'' some of the propagators in the $N$-point diagram.
From Eq.~\eqref{eq:DE-letter}, one sees that it's possible to focus on a particular entry of the $d\bm{A}$ matrix by imposing some cuts. We elaborate on this in the following. In this Section we will assume that the master integrals (after imposing cuts) have no divergences, such that the integrands can be expanded as Taylor series in $\epsilon$ before integration. It can be shown that in this situation only $g_N$, $g_{N-1}^{(i)}$ and $g_{N-2}^{(i)}$ appear in the right side of Eq.~\eqref{eq:DE-letter}. It turns out that the most complicated letters are given by these cases. Occasionally we encounter divergences in the cut integrals, and one has to expand the integrands as Laurent series in terms of distributions. We will discuss these cases in the next Section.
\subsection{The self-dependence $dM_{N}$}
\label{sec:self}
The self-dependent term in Eq.~\eqref{eq:DE-letter} is easy to extract by imposing the ``maximal-cut'', i.e., cut on all variables $\bm{z}$. All the lower-point integrals vanish under this cut, and the differential equation becomes
\begin{equation}
d\tilde{g}_{N}(\vec{x},\epsilon) = \epsilon \, dM_N(\vec{x}) \, \tilde{g}_{N}(\vec{x},\epsilon) \, ,
\end{equation}
where $\tilde{g}_N$ denotes the cut integral. Using the generic form of UT integrals in Eq.~\eqref{eq:UTint}, it is easy to see that
\begin{equation}
dM_N(\vec{x}) = d\log \left(-\frac{\mK_N(\vec{x})}{\widetilde{G}_N(\vec{x})}\right) ,
\label{eq:dM_N}
\end{equation}
where
\begin{equation}
\widetilde{G}_N(\vec{x}) \equiv G_N(\bm{0}) \, .
\end{equation}
Hence we see that the corresponding letter can be chosen as
\begin{equation}
W_N(\vec{x}) = \frac{\widetilde{G}_N(\vec{x})}{\mK_N(\vec{x})} \, .
\end{equation}
We note that two letters are equivalent if they only differ by a constant factor or a constant power, i.e.,
\begin{equation}
W(\vec{x}) \sim c \, W(\vec{x}) \sim \left[ W(\vec{x}) \right]^n \, .
\label{eq:letter_equiv}
\end{equation}
Therefore in practice, we may choose a form that is convenient for the particular case at hand.
It is possible that $G_N(\bm{0}) = 0$ such that $W_N(\vec{x}) = 0$ and cannot be a letter. In this case, the integral $\tilde{g}_{N}$ itself vanishes under the maximal cut. This means that the integral is reducible to integrals in sub-sectors, and we don't need to consider it as a master integral.
\subsection{Dependence on sub-sectors with one fewer propagator}
We now consider the dependence of the derivative of $g_N$ on sub-sectors with $N-1$ propagators. There can be $N$ such sub-sectors, corresponding to ``squeezing'' one of the $N$ propagators. Focusing on one of the sub-sector integral $g_{N-1}^{(i)}$, we can always reorganize the propagators (by shifting the loop momentum and relabel the external momenta) such that the squeezed one is $z_N$. We can then impose cut on the first $N-1$ variables, and write the differential equation as
\begin{align}
d\tilde{g}_{N}(\vec{x},\epsilon) &= \epsilon \, dM_{N}(\vec{x}) \, \tilde{g}_{N}(\vec{x},\epsilon) \nn
\\
&+ \epsilon \, dM_{N,N-1}(\vec{x}) \, \tilde{g}_{N-1}(\vec{x},\epsilon) \,,
\end{align}
where we have suppressed the superscript since only one sub-sector survives the cut. The letter in $dM_{N}(\vec{x})$ has been obtained in the previous step, and we now need to calculate the letter in $dM_{N,N-1}(\vec{x})$.
\subsubsection{Odd number of propagators}
We first consider the case where $N$ is an odd number. Using the generic form of one-loop UT integrals Eq.~\eqref{eq:UTint}, we can write
\begin{multline}
d\int_{r_-}^{r_+} \left(-\frac{\mK_N}{G_N(\bm{0}',z_N)}\right)^\epsilon \frac{dz_N}{z_N} \\
= \epsilon \, dM_N \int_{r_-}^{r_+} \left(-\frac{\mK_N}{G_N(\bm{0}',z_N)}\right)^\epsilon \frac{dz_N}{z_N}
\\
+ dM_{N,N-1} \, \frac{2^{1-2\epsilon} \, \Gamma^2(1-\epsilon)}{ \Gamma(1-2\epsilon)} \left(-\frac{\mK_{N-1}}{\widetilde{G}_{N-1}}\right)^\epsilon \, ,
\end{multline}
where the integration boundary is determined by the two roots $r_{\pm}$ of the polynomial $G_N(\bm{0}',z_N)$, and $\bm{0}'$ means that the vector $\bm{z}'\equiv\{z_1,\ldots,z_{N-1}\}$ is zero.
If both $r_+$ and $r_-$ are non-zero, the integration over $z_N$ is convergent for $\epsilon \to 0$. We can then set $\epsilon = 0$ in the equation and get
\begin{equation}
dM_{N,N-1} = \frac{1}{2} \, d\int_{r_-}^{r_+} \frac{dz_N}{z_N} = \frac{1}{2} \, d\log\frac{r_+}{r_-} \, .
\end{equation}
We may already set the letter to $r_+/r_-$ and stop at this point. However, it will be useful to write $r_\pm$ in terms of certain Gram determinants. This not only simplifies the procedure to compute the letter, but also tells us about the physics in the divergent situations $r_+ = 0$ or $r_- = 0$.
Given the propagator denominators \eqref{eq:propagator_general} and the definition of the Gram determinant \eqref{eq:gram1}, it is easy to see that $z_N$ only appears in the top-right and bottom-left corners of the Gram matrix. Using the expansion of the determinant in terms of cofactors, we can write
\begin{align}
G_N(\bm{0}',z_N)=-\frac{1}{4}\mK_{N-1} z_N^2 - \widetilde{B}_N z_N +\widetilde{G}_N \, ,
\end{align}
where $\widetilde{B}_N \equiv B_N(\bm{0})$ with (recall that $E=N-1$)
\begin{align}
B_N(\bm{z}) \equiv G(l,p_1,\ldots,p_{E-1};\, p_{E},p_1,\ldots,p_{E-1})\, ,
\end{align}
Here we have defined an extended Gram determinant
\begin{multline}
G(q_1,\ldots,q_n;\, k_1,\ldots,k_n) \\
= \det
\begin{pmatrix}
q_1 \cdot k_1 & q_1 \cdot k_2 & \cdots & q_1 \cdot k_n
\\
q_2 \cdot k_1 & q_2 \cdot k_2 & & \vdots
\\
\vdots & & \ddots & \vdots
\\
q_n \cdot k_1 & \cdots & \cdots & q_n \cdot k_n
\end{pmatrix}
.
\end{multline}
We may further use the geometric picture of Gram determinants to simplify the two roots. The Gram determinants can be expressed as
\begin{align}
G(q_1,\ldots,q_n) &= \det \left( q_i^\mu q_j^\nu g_{\mu\nu} \right) \nn
\\
&= \det(g_{\mu\nu}) \left[ V(q_1,\ldots,q_n) \right]^2 \, ,
\end{align}
where $q_i^\mu$ is the $\mu$th component of $q_i$ in the subspace spanned by $\{q_1,\ldots,q_n\}$ (with an arbitrary coordinate system), and $g_{\mu\nu}$ is the metric tensor of this subspace. $V(q_1,\ldots,q_n)$ is the volume of the parallelotope formed by the vectors $q_1,\ldots,q_n$ (in the Euclidean sense).
Let $l^\star$ denote a solution to the equation $\bm{z} = 0$ (recall that $z_i$ contain scalar products involving the loop momentum $l$), we can write
\begin{align}
\widetilde{G}_{N-1} &= G(l^\star,p_1,\ldots,p_{E-1}) \, , \nonumber
\\
\widetilde{G}_N &= G(l^\star,p_1,\ldots,p_E) \, , \nonumber
\\
\widetilde{B}_N &= G(l^\star,p_1,\ldots,p_{E-1};\, p_{E},p_1,\ldots,p_{E-1}) \, .
\end{align}
We let $l^\star_\perp$ and $p_{E\perp}$ to denote the components of $l^\star$ and $p_E$ perpendicular to the subspace spanned by $p_1,\ldots,p_{E-1}$, respectively. We are working in the region that the subspace of external momenta is space-like, and $l^\star_\perp$ must be time-like (since $l^\star$ is either time-like or light-like due to $(l^\star)^2 - m_1^2 = 0$). We can write the components of $l^\star_\perp$ perpendicular and parallel to $p_{E\perp}$ as $|l^\star_\perp|\cosh(\eta)$ and $|l^\star_\perp|\sinh(\eta)$, respectively, where $|l^\star_\perp| \equiv \sqrt{(l^\star_\perp)^2}$. We also denote $|p_{E\perp}| \equiv \sqrt{-p_{E\perp}^2}$. These allow us to write
\begin{align}
\frac{\widetilde{B}_N}{\mK_{N-1}} &= - |l_{\perp}^\star| |p_{E\perp}| \sinh(\eta) \,, \quad \frac{\mK_{N}}{\mK_{N-1}} = -|p_{E\perp}|^2 \,, \nn
\\
\frac{\widetilde{G}_N}{\mK_{N-1}} &= -|l_{\perp}^\star|^2 |p_{E\perp}|^2 \cosh^2(\eta) \,, \quad \frac{\widetilde{G}_{N-1}}{\mK_{N-1}} = |l_{\perp}^\star|^2 \,.
\label{eq:triangle1}
\end{align}
It then follows that
\begin{equation}
\widetilde{B}_N^2 + \mK_{N-1} \widetilde{G}_N = -\mK_{N-1}^2 |l_{\perp}^\star|^2 |p_{E\perp}|^2 = \mK_{N} \widetilde{G}_{N-1} \, .
\end{equation}
Note that the above relation can also be obtained from the Sylvester's determinant identity applied to Gram determinants (for other applications of this relation, see, e.g., \cite{Chen:2020uyk, Dlapa:2021qsl, Chen:2022lzr}). We will encounter further instances of this relation later in this work.
Expressing $r_\pm$ in terms of the Gram determinants, we can finally write the letter in $dM_{N,N-1}$ (for odd $N$) as
\begin{equation}
W_{N,N-1}(\vec{x}) = \frac{\widetilde{B}_{N}-\sqrt{\widetilde{G}_{N-1}\mathcal{K}_N}}{\widetilde{B}_{N}+\sqrt{\widetilde{G}_{N-1}\mathcal{K}_N}} \, .
\label{eq:W_n1_odd}
\end{equation}
We emphasize that the ingredients $\widetilde{B}_N$, $\widetilde{G}_{N-1}$ and $\mK_N$ can be very complicated functions of the kinematic variables $\vec{x}$ when $N$ and the length of $\vec{x}$ are large, and it is not easy to obtain the letter through direct integration in multi-scale problems.
If one of $r_\pm$ is zero, the integration over $z_N$ is divergent when $\epsilon \to 0$, and we cannot expand the integrand as a Taylor series. Actually, one can see that $W_{N,N-1}(\vec{x})$ in Eq.~\eqref{eq:W_n1_odd} becomes zero in this situation. On the other hand, this requires $\widetilde{G}_N = 0$, which means that $g_N$ vanishes under maximal cut, and hence is not a master integral. It is also possible that $\widetilde{G}_{N-1} = 0$ and $g_{N-1}$ is not a master. In this case $\log W_{N,N-1} = \log(1) = 0$ drops out of the differential equations. We therefore do not need to consider these cases here. Similar considerations apply to the $N$-even case coming next.
\subsubsection{Even number of propagators}
We now turn to the situation where $N$ is an even number. We proceed similarly as the odd case, and arrive at the cut differential equation
\begin{align}
&d \int_{r_-}^{r_+} \frac{dz_N}{z_N} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_N)}} \left[-\frac{\mK_N}{G_N(\bm{0}',z_N)}\right]^\epsilon \nonumber
\\
&= \epsilon \, dM_{N} \int_{r_-}^{r_+} \frac{dz_N}{z_N} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_N)}} \left[-\frac{\mK_N}{G_N(\bm{0}',z_N)}\right]^\epsilon \nonumber
\\
&+ 2\pi \, \epsilon \, \frac{2^{2\epsilon} \, \Gamma(1-2\epsilon)}{\Gamma^2(1-\epsilon)} \, dM_{N,N-1} \left(-\frac{\mK_{N-1}}{\widetilde{G}_{N-1}}\right)^\epsilon \, .
\end{align}
We again assume that the integration over $z_N$ is convergent for $\epsilon \to 0$. We can then expand the integrands on both sides of the above equation. At order $\epsilon^0$, the integral on the left side is
\begin{equation}
\int_{r_-}^{r_+} \frac{dz_N}{z_N} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_N)}} = i \pi \, .
\label{eq:pi}
\end{equation}
Hence its derivative is zero. Comparing the order $\epsilon^1$ coefficients, and plugging in the form of $dM_N$ obtained earlier in Eq.~\eqref{eq:dM_N}, we get
\begin{multline}
dM_{N,N-1} = -\frac{1}{2\pi} \, d \int_{r_-}^{r_+} \frac{dz_N}{z_N} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_N)}}
\\
\times \log \frac{G_N(\bm{0}',z_N)}{\widetilde{G}_N} \, .
\end{multline}
The above integrand involves multi-valued functions such as square roots and logarithms. To define the integral, we need to choose a convention including branch cuts for these functions and also the path from $r_-$ to $r_+$. Different conventions will lead to results differing by some constants or an overall minus sign, but these do not affect the letter up to the equivalence mentioned in Eq.~\eqref{eq:letter_equiv}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.4\textwidth]{Contour_M43}
\caption{\label{fig:contour_M43}The branch cuts and integration paths for $M_{N,N-1}$ with even $N$.}
\end{figure}
We denote $G_N(\bm{0}',z_N)$ as $(r_+-z_N)(z_N-r_-) \mK_{N-1}/4$ with $\mK_{N-1} > 0$, and write the integral as
\begin{multline}
M_{N,N-1} = -\frac{1}{2\pi} \, \int_{r_-}^{r_+} \frac{dz_N}{z_N} \sqrt{\frac{r_+ r_-}{(z_N-r_+)(z_N-r_-)}}
\\
\times \log \frac{(z_N-r_+)(z_N-r_-)}{r_+ r_-} \, .
\end{multline}
The branch cuts involve the points $r_\pm$ and $\infty$ on the complex $z_N$ plane. To represent the cuts more clearly, we perform the change of variable:
\begin{align}
z_N = \frac{1}{t} \,, \quad t_\pm = \frac{1}{r_\mp} \, .
\end{align}
The branch points then become $t_\pm$ and $0$, and we write the integral as
\begin{equation}
M_{N,N-1} = -\frac{1}{2\pi} \int_{t_-}^{t_+} I(t) \, dt \, ,
\label{eq:M43_dt}
\end{equation}
with the integrand
\begin{align}
I(t) &= \frac{1}{\sqrt{(t-t_+)(t-t_-)}} \left[ \log\frac{t-t_+}{t} + \log\frac{t-t_-}{t} \right] .
\label{eq:M43_It}
\end{align}
With this form of the integrand, we choose the branch cut for the square root to be the line segment between $t_+$ and $t_-$, and the branch cuts for the two logarithms to be the line segments between $0$ and $t_\pm$, respectively. These branch cuts are depicted as the wiggly lines in Fig.~\ref{fig:contour_M43}, together with several paths $C_{i\pm}$ which lie infinitesimally close to the cuts. We define the square root following the convention that $\sqrt{(t-t_+)(t-t_-)} \to t$ when $t \to \infty$.
We choose the integration path in Eq.~\eqref{eq:M43_dt} to along the line segment $\mathcal{C}_{1+}$, and write the integral as
\begin{equation}
M_{N,N-1} = -\frac{1}{4\pi} \left[ \int_{\mathcal{C}_{1+}} I(t) \, dt - \int_{\mathcal{C}_{1-}} I(t) \, dt \right] ,
\end{equation}
where we have used the fact that the values of $I(t)$ on $\mathcal{C}_{1\pm}$ differ by a sign. Since there are no other singularities in the complex $t$ plane (including $\infty$), we may deform the paths as long as we don't go across the branch cuts. Hence we know that
\begin{align}
M_{N,N-1} &= \frac{1}{4\pi} \left[ \int_{\mathcal{C}_{2+}} I(t) \, dt - \int_{\mathcal{C}_{2-}} I(t) \, dt \right] \nn \\
&+ \frac{1}{4\pi} \left[ \int_{\mathcal{C}_{3+}} I(t) \, dt - \int_{\mathcal{C}_{3-}} I(t) \, dt \right].
\end{align}
On the paths $\mathcal{C}_{2+}$ and $\mathcal{C}_{2-}$, there is a $2\pi i$ difference coming from the first logarithm in Eq.~\eqref{eq:M43_It}. A similar difference of $-2\pi i$ arising from the second logarithm is there between $\mathcal{C}_{3+}$ and $\mathcal{C}_{3-}$. Therefore we have
\begin{align}
dM_{N,N-1} &= -\frac{i}{2} \, d\int_0^{t_+} \frac{dt}{\sqrt{(t-t_+)(t-t_-)}} \nn \\
&\quad - \frac{i}{2} \, d\int_0^{t_-} \frac{dt}{\sqrt{(t-t_+)(t-t_-)}} \nonumber
\\
&= -i \, d\log \frac{\sqrt{r_+} - \sqrt{r_-}}{\sqrt{r_+} + \sqrt{r_-}} \, .
\end{align}
Note that with the above convention, we have
\begin{equation}
\int_{r_-}^{r_+} \frac{dz_N}{z_N} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_N)}} = \int_{t_-}^{t_+} \frac{dt}{\sqrt{(t-t_+)(t-t_-)}} = i \pi \, .
\end{equation}
We can now express the roots $r_\pm$ in terms of Gram determinants. The result can be written as
\begin{align}
dM_{N,N-1} = \frac{i}{2} \, d\log \frac{\widetilde{B}_{N}-\sqrt{-\widetilde{G}_{N} \mK_{N-1}}}{\widetilde{B}_{N}+\sqrt{-\widetilde{G}_{N} \mK_{N-1}}} \, ,
\end{align}
where the definition of $\widetilde{B}_{N}$, $\widetilde{G}_{N}$ and $\mK_{N-1}$ is similar as before.
Hence we can write the letter in $dM_{N,N-1}$ (for even $N$) as
\begin{equation}
W_{N,N-1}(\vec{x}) = \frac{\widetilde{B}_{N}-\sqrt{-\widetilde{G}_{N} \mK_{N-1}}}{\widetilde{B}_{N}+\sqrt{-\widetilde{G}_{N} \mK_{N-1}}} \, .
\label{eq:W_n1_even}
\end{equation}
As mentioned earlier, we don't need to consider the divergent case $\widetilde{G}_{N-1}=0$ or the trivial case $\widetilde{G}_N=0$ here.
\subsection{Dependence on sub-sectors with two fewer propagators}
As in the previous subsection, we consider the dependence of the derivative of $g_N$ on sub-sectors with $N-2$ propagators. Without loss of generality, we cut on the variables $\bm{z}'=\{z_1,\ldots,z_{N-2}\}$. Now we are left with two sub-sectors with $N-1$ propagators: one with $\bm{z}',z_{N-1}$ and the other with $\bm{z}',z_{N}$. We use a superscript to distinguish these two, and the differential equation then reads
\begin{align}
d\tilde{g}_{N} &= \epsilon \Big( dM_{N} \, \tilde{g}_{N} + dM_{N,N-1}^{(1)} \, \tilde{g}_{N-1}^{(1)} \nn \\
&\quad + dM_{N,N-1}^{(2)} \, \tilde{g}_{N-1}^{(2)} + dM_{N,N-2} \, \tilde{g}_{N-2} \Big) \, ,
\end{align}
where we have suppressed the arguments of the functions for simplicity.
\subsubsection{Odd number of propagators}\label{sec:W53}
If $N$ is an odd number, assuming convergence and expanding the integrands, we find
\begin{align}
&d\int_\mathcal{C} \frac{dz_{N-1}}{z_{N-1}}\frac{dz_N}{z_N} = 4\pi \, dM_{N,N-2} \nn \\
&+ 2 \, dM_{N,N-1}^{(1)} \int_{r_{-}^{(1)}}^{r_{+}^{(1)}} \frac{dz_{N-1}}{z_{N-1}} \frac{\sqrt{\widetilde{G}_{N-1}^{(1)}}}{\sqrt{G_{N-1}^{(1)}(\bm{0}',z_{N-1})}} \nonumber\\
&+ 2 \, dM_{N,N-1}^{(2)} \int_{r_{-}^{(2)}}^{r_{+}^{(2)}} \frac{dz_N}{z_N} \frac{\sqrt{\widetilde{G}_{N-1}^{(2)}}}{\sqrt{G_{N-1}^{(2)}(\bm{0}',z_N)}} \,,
\end{align}
where the domain $\mathcal{C}$ is determined by $G_N(\bm{0}',z_{N-1},z_N)\geq 0$, and $r_{\pm}^{(i)}$ are the two roots of the polynomial $G_{N-1}^{(i)}(\bm{0}',z)$.
The two integrals on the right-hand side can be easily performed using Eq.~\eqref{eq:pi}, and we have
\begin{align}
dM_{N,N-2} = dI_{N,N-2} - \frac{i}{2} \left( dM_{N,N-1}^{(1)} + dM_{N,N-1}^{(2)} \right)\, ,
\end{align}
where $I_{N,N-2}$ is the double integral
\begin{align}
I_{N,N-2} = \frac{1}{4\pi} \int_\mathcal{C} \frac{dz_{N-1}}{z_{N-1}}\frac{dz_N}{z_N} \, .
\end{align}
The integration domain $\mathcal{C}$ is controlled by the positivity of the polynomial
\begin{align}
G_N(\bm{0}',z_{N-1}, z_N) &= -\frac{1}{4}\mK_{N-1} z_N^2 - B_N(\bm{0}',z_{N-1},0) \, z_N \nn \\
&\quad + G_N(\bm{0}',z_{N-1},0) \, .
\end{align}
The integration over $z_N$ can be easily performed to arrive at
\begin{align}
I_{N,N-2} &= \frac{1}{4\pi} \int_{r_{N-1,-}}^{r_{N-1,+}} I(z_{N-1}) \, dz_{N-1} \nonumber
\\
&\equiv \frac{1}{4\pi} \int_{r_{N-1,-}}^{r_{N-1,+}} \frac{dz_{N-1}}{z_{N-1}} \nn \\
&\quad \times \log \frac{B_N(\bm{0}',z_{N-1},0) - \sqrt{\Delta(z_{N-1})}}{B_N(\bm{0}',z_{N-1},0) + \sqrt{\Delta(z_{N-1})}} \,,
\label{eq:I53}
\end{align}
where $r_{N-1,\pm}$ are the two roots of the polynomial
\begin{equation}
G_{N-1}^{(1)}(\bm{z}',z_{N-1}) = G(l,p_1,\ldots,p_{E-1}) \, ,
\end{equation}
and
\begin{align}
\Delta(z_{N-1}) &= \left[ B_N(\bm{0}',z_{N-1},0) \right]^2 + \mK_{N-1} G_N(\bm{0}',z_{N-1},0) \nn \\
&= \mK_N G_{N-1}^{(1)}(\bm{0}',z_{N-1}) \,.
\end{align}
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{Contour_M53}
\caption{\label{fig:contour_M53}The branch cuts and integration paths for $M_{N,N-2}$ with odd $N$.}
\end{figure}
We are now concerned with the singularities of the integrand $I(z_{N-1})$ in Eq.~\eqref{eq:I53}. There are two poles at $0$ and $\infty$, respectively. There is a branch cut between $r_{N-1,-}$ and $r_{N-1,+}$ for the square root. There is also a branch cut between $R_{N-1,-}$ and $R_{N-1,+}$ for the logarithm, where $R_{N-1,\pm}$ are the two roots of the polynomial $G_N(\bm{0}',z_{N-1},0)$. These singularities are depicted in Fig.~\ref{fig:contour_M53}. We define the integral path of Eq.~\eqref{eq:I53} to be the upper half of the contour $\mathcal{C}_1$. Hence we have
\begin{align}
I_{N,N-2} &= \frac{1}{8\pi} \int_{\mathcal{C}_1} I(z_{N-1}) \, dz_{N-1} \nn \\
&= -\frac{1}{8\pi} \int_{\mathcal{C}_2+\mathcal{C}_3+\mathcal{C}_4} I(z_{N-1}) \, dz_{N-1} \,.
\end{align}
The integration around $\mathcal{C}_3$ is just $(-2\pi i)$ multiplying the residue at $z_{N-1} = 0$, i.e.
\begin{align}
-\frac{1}{8\pi} \, d\int_{\mathcal{C}_3} I(z_{N-1}) \, dz_{N-1} &= \frac{i}{4} \, d\log\frac{\widetilde{B}_N - \sqrt{\mK_N \widetilde{G}_{N-1}^{(1)}}}{\widetilde{B}_N + \sqrt{\mK_N \widetilde{G}_{N-1}^{(1)}}} \nn \\
&= \frac{i}{2} dM_{N,N-1}^{(1)} \,.
\end{align}
On the two sides of $\mathcal{C}_2$, the logarithm differs by $2\pi i$, and
\begin{align}
&-\frac{1}{8\pi} \, d\int_{\mathcal{C}_2} I(z_{N-1}) \, dz_{N-1} = \frac{i}{4} \, d\log\frac{R_{N-1,+}}{R_{N-1,-}} \nn \\
&= \frac{i}{4} \, d\log\frac{\widetilde{B}_N - \sqrt{\mK_N \widetilde{G}_{N-1}^{(2)}}}{\widetilde{B}_N + \sqrt{\mK_N \widetilde{G}_{N-1}^{(2)}}} = \frac{i}{2} dM_{N,N-1}^{(2)} \,.
\end{align}
From the above we see that the genuine contribution to $dM_{N,N-2}$ only comes from the integration along $\mathcal{C}_4$. For that we need to investigate the behavior of the logarithm in Eq.~\eqref{eq:I53} in the limit $z_{N-1} \to \infty$. We first note that $G_{N-1}^{(1)}(\bm{0}',z_{N-1}) \sim -\mK_{N-2} z_{N-1}^2 / 4$ in that limit. As for $B_N(\bm{0}',z_{N-1},0)$, it is a linear function of $z_{N-1}$ and the coefficient can be extracted as
\begin{align}
&\frac{\partial B_N(\bm{0}',z_{N-1},0)}{\partial z_{N-1}} \nn \\
&= \frac{\partial G(l,p_1,\ldots,p_{E-1};\, p_E,p_1,\ldots,p_{E-1})}{\partial z_{N-1}} \nn
\\
&= \frac{\partial l \cdot p_E}{\partial z_{N-1}}\frac{\partial G(l,p_1,\ldots,p_{E-1};\, p_E,p_1,\ldots,p_{E-1})}{\partial l \cdot p_E} \nn
\\
&\quad + \frac{\partial l \cdot p_{E-1}}{\partial z_{N-1}}\frac{\partial G(l,p_1,\ldots,p_{E-1};\, p_E,p_1,\ldots,p_{E-1})}{\partial l\cdot p_{E-1}} \nn
\\
&= \frac{1}{2} G(p_1,\ldots,p_{E-1};\, p_1,\ldots,p_{E-1}) \nn
\\
&\quad + \frac{1}{2} G(p_1,\ldots,p_{E-2},p_{E-1};\, p_1,\ldots,p_{E-2},p_{E}) \nn
\\
&= \frac{1}{2}G(p_1,\ldots,p_{E-2},p_{E-1};\, p_1,\ldots,p_{E-2},p_{E-1}+p_E) \, .
\end{align}
We hence have
\begin{align}
dM_{N,N-2} &= -\frac{1}{8\pi} \, d\int_{\mathcal{C}_4} I(z_{N-1}) \, dz_{N-1} \nn \\
&= \frac{i}{4} d\log \frac{C_N - \sqrt{-\mK_N \mK_{N-2}}}{C_N + \sqrt{-\mK_N \mK_{N-2}}} \, ,
\label{eq:M53}
\end{align}
where
\begin{equation}
C_N = G(p_1,\ldots,p_{E-2},p_{E-1};\, p_1,\ldots,p_{E-2},p_{E-1}+p_E) \, .
\end{equation}
The letter $W_{N,N-2}$ can be readily read off. Note that the Gram determinants in this letter only involve external momenta. The letter hence has a well-defined limit when $\widetilde{G}_{N-2} = 0$ and $g_{N-2}$ is not a master. We will see what this means later.
\subsubsection{Even number of propagators}
\label{sec:W42}
If $N$ is an even number, assuming no divergence, we have the differential equation
\begin{align}
d\int_\mathcal{C} \frac{dz_{N-1}}{z_{N-1}} \frac{dz_N}{z_N}\frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},z_N)}} = 4\pi \, dM_{N,N-2} \,,
\end{align}
where the domain $\mathcal{C}$ is determined by $G_N(\bm{0}',z_{N-1},z_N) \geq 0$. Note that the dependence on $g_{N-1}^{(i)}$ drops out in this case. We choose to integrate over $z_N$ first, and have
\begin{align}
dM_{N,N-2} &=\frac{1}{4\pi} \, d\int_{r_{N-1,-}}^{r_{N-1,+}} \frac{dz_{N-1}}{z_{N-1}} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},0)}} \nn \\
&\quad \times \int_{r_{N,-}}^{r_{N,+}} \frac{dz_N}{z_N}\frac{\sqrt{G_N(\bm{0}',z_{N-1},0)}}{\sqrt{G_N(\bm{0}',z_{N-1},z_N)}} \,,
\end{align}
where $r_{N,\pm}$ are the two roots of the polynomial $G_N(\bm{0}',z_{N-1},z_N)$ with respect to $z_N$ (treating $z_{N-1}$ as a constant). The integration range of $z_{N-1}$ is in turn determined by the discriminant $\Delta$ of $G_N(\bm{0}',z_{N-1},z_N)$ (with respect to the variable $z_N$). Writing $\Delta = \mK_4 G_{N-1}^{(1)}(\bm{0}',z_{N-1})$, we know the the bounds $r_{N-1,\pm}$ are just the two roots of the polynomial $G_{N-1}^{(1)}(\bm{0}',z_{N-1})$. Here we define
\begin{align}
G_{N-1}^{(1)}(\bm{z}',z_{N-1}) &= G(l,p_1,\ldots,p_{E-1}) \, , \nn \\
G_{N-1}^{(2)}(\bm{z}',z_{N}) &= G(l,p_1,\ldots,p_{E-1}+p_E) \, .
\end{align}
The integration over $z_N$ can be carried out using Eq.~\eqref{eq:pi}. We then arrive at
\begin{align}
dM_{N,N-2} &= \frac{i}{4} \, dI_{N,N-2} \,,
\end{align}
where
\begin{align}
I_{N,N-2} &= \int_{r_{N-1,-}}^{r_{N-1,+}} \frac{dz_{N-1}}{z_{N-1}} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},0)}} \, ,
\end{align}
where $r_{N-1,\pm}$ are the two roots of $G_{N-1}^{(1)}(\bm{0}',z_{N-1})$. We denote the two roots of $G_N(\bm{0}',z_{N-1},0)$ as $R_{N-1,\pm}$. We can then write
\begin{align}
&G_N(\bm{0}',z_{N-1},0) \nn \\
&= -\frac{1}{4} \mK_{N-1}^{(2)} (z_{N-1}-R_{N-1,+})(z_{N-1}-R_{N-1,-}) \, ,
\end{align}
where
\begin{equation}
\mK_{N-1}^{(2)} = G(p_1,\ldots,p_{E-2},p_{E-1}+p_E) \, .
\end{equation}
We define
\begin{equation}
t = \frac{1}{z_{N-1}} \, , \quad t_{\pm} = \frac{1}{r_{N-1,\mp}} \, , \quad T_{\pm} = \frac{1}{R_{N-1,\mp}} \, .
\end{equation}
The integral can then be written as
\begin{align}
I_{N,N-2} &= \int_{t_-}^{t_+} \frac{dt}{\sqrt{(t-T_+)(t-T_-)}} \nn \\
&= 2 \log \frac{\sqrt{t_+-T_+} + \sqrt{t_+-T_-}}{\sqrt{t_--T_+} + \sqrt{t_--T_-}} \, .
\end{align}
We now want to rewrite the above expression in terms of Gram determinants. To do that we first write
\begin{align}
G_N(\bm{0}',z_{N-1},0) &= -\frac{1}{4} \mK_{N-1}^{(2)} z_{N-1}^2 - \widetilde{B}_{N} z_{N-1} + \widetilde{G}_N \, , \nn
\\
G_{N-1}^{(1)}(\bm{0}',z_{N-1}) &= -\frac{1}{4}\mK_{N-2} z_{N-1}^2 - \widetilde{B}_{N-1}^{(1)} z_{N-1} + \widetilde{G}_{N-1}^{(1)} \, ,
\end{align}
where
\begin{align}
\mK_{N-2} &= G(p_1,\ldots,p_{E-2}) \, , \nn
\\
B_{N-1}^{(1)}(\bm{z}) &= G(l,p_1,\ldots,p_{E-2};\, p_{E-1},p_1,\ldots,p_{E-2}) \, .
\end{align}
The roots are given by
\begin{align}
t_{\pm} &= \frac{\widetilde{B}_{N-1}^{(1)} \pm \sqrt{\mK_{N-1}^{(1)} \widetilde{G}_{N-2}}}{2\widetilde{G}_{N-1}^{(1)}} \, , \nn
\\
T_{\pm} &= \frac{\widetilde{B}_{N} \pm \sqrt{\mK_{N} \widetilde{G}_{N-1}^{(2)}}}{2\widetilde{G}_{N}} \, ,
\end{align}
where
\begin{equation}
G_{N-1}^{(2)} = G(l,p_1,\ldots,p_{E-2},p_{E-1}+p_E) \, ,
\end{equation}
and we have used the relations
\begin{align}
B_N^2 + \mK_{N-1}^{(2)} G_N &= \mK_N G_{N-1}^{(2)} \, , \nn \\
\left( B_{N-1}^{(1)} \right)^2 + \mK_N G_{N-1}^{(1)} &= \mK_{N-1}^{(1)} G_{N-2} \, .
\end{align}
We can now employ the geometric representations of the Gram determinants in Eq.~\eqref{eq:triangle1} to simplify the expressions. Let $l^\star$ be the solution to $\bm{z}=0$, we will be concerned with the components of $l^\star$, $p_{E-1}$ and $p_{E-1}+p_E$ orthogonal to the subspace spanned by $\{p_1,\ldots,p_{E-2}\}$. For convenience we denote these components as $k^\mu$ (for $l^\star$), $p^\mu$ (for $p_{E-1}$) and $q^\mu$ (for $p_{E-1}+p_E$). We note that $k^\mu$ is time-like while $p^\mu$ and $q^\mu$ are space-like. Hence we can define the norms $|k| = \sqrt{k^2}$, $|p| = \sqrt{-p^2}$ and $|q| = \sqrt{-q^2}$. We further denote the components of $k^\mu$ and $p^\mu$ perpendicular to $q$ as $k_\perp^\mu$ and $p_\perp^\mu$, and define the corresponding norms as $|k_\perp|$ and $|p_\perp|$. We can finally write
\begin{align}
t_{\pm} = \frac{\sinh(\eta_1) \pm i}{2 |k| |p| \cosh^2(\eta_1)} \, , \quad T_{\pm} = \frac{\sinh(\eta_2) \pm i}{2 |k_{\perp}| |p_\perp| \cosh^2(\eta_2)} \, ,
\end{align}
where $\eta_1$ is the hyperbolic angle between $k$ and $p$, and $\eta_2$ is the hyperbolic angle between $k_\perp$ and $p_\perp$. It will be convenient to define the imaginary angle $\theta_{kp} \equiv \pi/2 - i\eta_1$, such that $\cosh(\eta_1) = \sin\theta_{kp}$ and $i \sinh(\eta_1) = \cos\theta_{kp}$; and similarly $\theta_{kp,\perp q} \equiv \pi/2 - i\eta_2$.
We use $\theta_{pq}$ to denote the angle between $p$ and $q$, and define $\xi$ as the hyperbolic angle between $k$ and $q$ (with the corresponding imaginary angle $\theta_{kq} \equiv \pi/2 - i\xi$). We then have the relations
\begin{align}
|p_\perp| &= |p| \sin\theta_{pq} \, , \quad |k_\perp| = |k| \sin\theta_{kq} \, , \nonumber
\\
\cos\theta_{kp} &= \cos\theta_{kq} \cos\theta_{pq} + \cos\theta_{kp,\perp q} \sin\theta_{kq} \sin\theta_{pq} \, .
\label{eq:tri2}
\end{align}
It then follows that
\begin{align}
t_{\pm} - T_{\pm} &\equiv \frac{P_{\pm\pm}}{2|k_\perp| |p_\perp| \sin^2\theta_{kp} \sin^2\theta_{kp,\perp q}} \, ,
\end{align}
where
\begin{align}
P_{\pm\pm} &= (-i\cos\theta_{kp} \pm i) \sin^2\theta_{kp,\perp q} \sin\theta_{pq} \sin\theta_{kq} \nn \\
&\quad - (-i\cos\theta_{kp,\perp q} \pm i) \sin^2\theta_{kp} \, .
\end{align}
Plugging in the relation \eqref{eq:tri2}, we may write the functions $P_{\pm\pm}$ as
\begin{align}
P_{++} &= -8 i \, \sin^2\left( \frac{\theta_{kp}}{2} \right) \cos^2\left( \frac{\theta_{kq}+\theta_{pq}}{2} \right) \sin^2\left( \frac{\theta_{kp,\perp q}}{2} \right) , \nonumber
\\
P_{+-} &= 8 i \, \sin^2\left( \frac{\theta_{kp}}{2} \right) \cos^2\left( \frac{\theta_{kq}-\theta_{pq}}{2} \right) \cos^2\left( \frac{\theta_{kp,\perp q}}{2} \right) , \nonumber
\\
P_{-+} &= -8 i \, \cos^2\left( \frac{\theta_{kp}}{2} \right) \sin^2\left( \frac{\theta_{kq}+\theta_{pq}}{2} \right) \sin^2\left( \frac{\theta_{kp,\perp q}}{2} \right) , \nonumber
\\
P_{--} &= 8 i \, \cos^2\left( \frac{\theta_{kp}}{2} \right) \sin^2\left( \frac{\theta_{kq}-\theta_{pq}}{2} \right) \cos^2\left( \frac{\theta_{kp,\perp q}}{2} \right) .
\end{align}
Using trigonometry identities together with the relations
\begin{align}
\cos\theta_{pq} &= \cos\theta_{kp} \cos\theta_{kq} + \cos\theta_{pq,\perp k} \sin\theta_{kp} \sin\theta_{kq} \, , \nonumber
\\
\sin\theta_{pq} &= \sin\theta_{pq,\perp k} \, \frac{\sin\theta_{kp}}{\sin\theta_{kp,\perp q}} \, ,
\end{align}
we can arrive at a surprisingly simple result
\begin{equation}
I_{N,N-2} = 2 \log e^{-i \theta_{pq,\perp k}} = \log \frac{\cos\theta_{pq,\perp k} - i \sin\theta_{pq,\perp k}}{\cos\theta_{pq,\perp k} + i \sin\theta_{pq,\perp k}} \, ,
\end{equation}
where $\theta_{pq,\perp k}$ is the angle between $p_{\perp k}$ and $q_{\perp k}$. It is straightforward to rewrite the above expression in terms of Gram determinants, and we finally obtain
\begin{align}
dM_{N,N-2} = \frac{i}{4} d\log \frac{\widetilde{D}_N - \sqrt{-\widetilde{G}_N\widetilde{G}_{N-2}}}{\widetilde{D}_N + \sqrt{-\widetilde{G}_N\widetilde{G}_{N-2}}} \,,
\label{eq:W42}
\end{align}
where $\widetilde{D}_{N} = D_N(\bm{0})$ and
\begin{equation}
D_N(\bm{z}) = G(l,p_1,\ldots,p_{E-1};l,p_1,\ldots,p_{E-1}+p_E) \, .
\end{equation}
\subsection{Dependence on further lower sub-sectors}
In the convergent case, $dg_{N}$ cannot depend on $g_{N-3}$ or integrals with even fewer propagators. For odd $N$, this can be easily seen from the powers of $\epsilon$ in Eq.~\eqref{eq:UTint}. For even $N$, however, $dg_{N}$ and $g_{N-3}$ are multiplied by the same power of $\epsilon$ in the differential equations. We then need to examine the three-fold integrals appearing in the differential equations under the $(N-3)$-cut. The first two folds can be performed following the calculations in Section~\ref{sec:W42}, and the last fold can be studied similar to Section~\ref{sec:W53}. Finally we can arrive at the conclusion that $dM_{N,N-3}=0$ in the convergent case. Note however, such dependence can be present in the divergence cases to be discussed in the next Section.
\section{Letters in differential equations: divergent cases}
We now consider the situation when some cut integrals become divergent and one cannot perform a Taylor expansion for the integrands. As discussed earlier, this happens when certain Gram determinants vanish under maximal cut, and the corresponding integrals are reducible to lower sectors. A classical example is the massless 3-point integral that can be reduced to 2-point integrals. Reducible higher-point integrals can occur with specific configurations of external momenta, which appear, e.g., at boundaries of differential equations or in some effective field theories. Divergent cut integrals can have two kinds of consequences, which we will discuss in the following.
\subsection{$N,N-2$ dependence with a reducible $(N-1)$-point integral}
We consider the dependence of $dg_{N}$ on $g_{N-2}$ when $g_{N-1}^{(1)}$ is reducible, where $N$ is even. Following the derivation in Section~\ref{sec:W42}, we see that now one of $r_{N-1,\pm}$ is zero and $G_{N-1}^{(1)}(\bm{0}',0) = 0$. The integration over $z_{N-1}$ is hence divergent and one cannot Taylor expand the integrand in $\epsilon$. One can also find that the entry $dM_{N,N-2}$ obtained in Section~\ref{sec:W42} is divergent. To proceed, we can keep the regulator in the differential equation:
\begin{align}
&d\int_\mathcal{C} \frac{dz_{N-1}}{z_{N-1}} \frac{dz_N}{z_N} \nn
\\
&\quad \times \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},z_N)}} \left[-\frac{\mK_N}{G_N(\bm{0}',z_{N-1},z_N)}\right]^\epsilon \nonumber
\\
&= \epsilon \, dM_{N} \int_\mathcal{C} \frac{dz_{N-1}}{z_{N-1}} \frac{dz_N}{z_N} \nn \\
&\quad \times \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},z_N)}} \left[-\frac{\mK_N}{G_N(\bm{0}',z_{N-1},z_N)}\right]^\epsilon \nonumber
\\
&+ 4\pi \, dM_{N,N-2}^\star \left(-\frac{\mK_{N-2}}{\widetilde{G}_{N-2}}\right)^\epsilon + \mathcal{O}(\epsilon) \,,
\end{align}
where $dM_{N,N-2}^\star$ denotes the entry in the divergent case. Note that $g_{N-1}^{(1)}$ is not a master integral and does not contribute to the right-hand side, while the last $\mathcal{O}(\epsilon)$ denotes a suppressed contribution from another $(N-1)$-point integral $g_{N-1}^{(2)}$. Here we assume that $G_{N-1}^{(2)}(\bm{0}',0)$ is non-zero and the integration over $z_N$ is convergent for $\epsilon \to 0$.
We now need to perform Laurent expansions of the integrands in terms of distributions. We write
\begin{align}
G_{N-1}^{(1)}(\bm{0}',z_{N-1}) &= \frac{1}{4}\mK_{N-2} z_{N-1} \, (t - z_{N-1}) \, , \nn
\\
t &= - \frac{4\widetilde{B}_{N-1}^{(1)}}{\mK_{N-2}} \, .
\end{align}
We can then use
\begin{equation}
\int_0^t \frac{dz}{z^{1+\epsilon}} \, f(z) = -\frac{t^{-\epsilon}}{\epsilon} \, f(0) + \int_0^t \frac{dz}{z^{1+\epsilon}} \left[ f(z) - f(0) \right] ,
\end{equation}
to perform the series expansion. In particular we have
\begin{align}
&\int_\mathcal{C} \frac{dz_{N-1}}{z_{N-1}} \frac{dz_N}{z_N} \nn
\\
&\quad \times \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},z_N)}} \left[-\frac{\mK_N}{G_N(\bm{0}',z_{N-1},z_N)}\right]^\epsilon \nonumber
\\
&= i\pi \int_0^t \frac{dz_{N-1}}{z_{N-1}^{1+\epsilon}} \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},0)}} \nn
\\
&\quad \times \left[ 1 + \epsilon \, h(z_{N-1}) + \mathcal{O}(\epsilon^2) \right] \nonumber
\\
&= i\pi \Bigg[ -\frac{1}{\epsilon} + \log(t) - h(0) \nn \\
&\quad + \int_0^t \frac{dz_{N-1}}{z_{N-1}} \left( \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},0)}} - 1 \right) \Bigg] + \mathcal{O}(\epsilon) \, ,
\label{eq:M42_div}
\end{align}
where the function $h(z_{N-1})$ arises from the expansion in $\epsilon$ after integrating over $z_N$. When $z_{N-1} \to 0$, it reduces to
\begin{equation}
h(0) = \log \left( \frac{4\mK_{N-1}^{(1)}}{\widetilde{B}_{N-1}^{(1)}} \right) + 4\log(2) \, .
\end{equation}
The last integral in Eq.~\eqref{eq:M42_div} can be obtained by taking the limit $\widetilde{G}_{N-1}^{(1)} \to 0$ in the difference between Eq.~\eqref{eq:W42} and a simple integral of $1/z_{N-1}$:
\begin{align}
&\int_0^t \frac{dz_{N-1}}{z_{N-1}} \left( \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},0)}} - 1 \right) \nn\\
&= \lim_{\widetilde{G}_{N-1}^{(1)} \to 0} \Bigg( \log\frac{\widetilde{D}_{N} - \sqrt{-\widetilde{G}_{N} \widetilde{G}_{N-2}}}{\widetilde{D}_{N} + \sqrt{-\widetilde{G}_{N} \widetilde{G}_{N-2}}} \nn \\
&\hspace{4em} - \log\frac{\widetilde{B}_{N-1}^{(1)} +\sqrt{\widetilde{G}_{N-2} \mK_{N-1}^{(1)}}}{\widetilde{B}_{N-1}^{(1)} -\sqrt{\widetilde{G}_{N-2} \mK_{N-1}^{(1)}}} \Bigg) \,.
\end{align}
Using the relations
\begin{gather}
-G_{N} G_{N-2}=D_{N}^2-G_{N-1}^{(1)}G_{N-1}^{(2)} \, , \nn \\
G_{N-2} \mK_{N-1}^{(1)}=\left(B_{N-1}^{(1)}\right)^2+G_{N-1}^{(1)}\mK_{N-2} \, ,
\label{eq:W42_div_G1}
\end{gather}
we can simplify the expression and arrive at
\begin{multline}
\int_0^t \frac{dz_{N-1}}{z_{N-1}} \left( \frac{\sqrt{\widetilde{G}_N}}{\sqrt{G_N(\bm{0}',z_{N-1},0)}} - 1 \right)
\\
= \log \frac{\widetilde{G}_N \mK_{N-2}}{\mK_{N-1}^{(1)}\widetilde{G}_{N-1}^{(2)} } \, .
\end{multline}
Now we can combine everything and find in the divergent case (for even $N$) that
\begin{equation}
W_{N,N-2}^\star = \frac{\widetilde{G}_{N-2} \, \mK_N}{\mK_{N-1}^{(1)} \, \widetilde{G}_{N-1}^{(2)}} \, .
\label{eq:W42}
\end{equation}
Comparing to Eq.~\eqref{eq:W42}, we note that the letter in the divergent case is simpler (without square roots) than that in the convergent case. Interestingly, this simple letter can be obtained without going through the tedious calculation in the above. We observe that in the divergent case $\widetilde{G}_{N-1}^{(1)} \to 0$, we have the relation
\begin{equation}
\tilde{g}_{N-1}^{(1)} = -\frac{1}{2} \tilde{g}_{N-2} \, .
\end{equation}
This hints that we should combine $dM_{N,N-1}^{(1)}$ and $dM_{N,N-2}$ to arrive at $dM_{N,N-2}^\star$:
\begin{align}
dM_{N,N-2}^{\star} &= \lim_{\widetilde{G}_{N-1}^{(1)}\to 0} \left( -\frac{1}{2} dM_{N,N-1}^{(1)} + dM_{N,N-2} \right) \nn
\\
&= \frac{i}{4} \lim_{\widetilde{G}_{N-1}^{(1)} \to 0} \Bigg( \log\frac{\widetilde{D}_{N} - \sqrt{-\widetilde{G}_{N} \widetilde{G}_{N-2}}}{\widetilde{D}_{N} + \sqrt{-\widetilde{G}_{N} \widetilde{G}_{N-2}}} \nn
\\
&\hspace{4em} - \log\frac{\widetilde{B}_{N}^{(1)} -\sqrt{-\widetilde{G}_{N} \mK_{N-1}^{(1)}}}{\widetilde{B}_{N}^{(1)} +\sqrt{-\widetilde{G}_{N} \mK_{3}^{(N)}}} \Bigg) \,.
\end{align}
Using the relations in Eq.~\eqref{eq:W42_div_G1} as well as
\begin{align}
-G_{N} \mK_{N-1}^{(1)}&=\left(B_{N}^{(1)}\right)^2+G_{N-1}^{(1)}\mK_{N} \, ,
\label{eq:gramdetforM42IR}
\end{align}
we can easily arrive at Eq.~\eqref{eq:W42}.
Further divergences may arise if $\widetilde{G}_{N-1}^{(2)} = 0$ in Eq.~\eqref{eq:W42}. In this case both $g_{N-1}^{(1)}$ and $g_{N-1}^{(2)}$ are reducible to lower-point integrals. It can be shown that the corresponding letter can be obtained by including $dM_{N,N-1}^{(2)}$, and we do not elaborate on the calculation here. We finally note that the above considerations can also be applied to the $N$-odd cases, although here $g_{N-1}^{(i)}$ can only be reducible for specific configurations of external momenta. We will encounter similar situations in the next subsection.
\subsection{$N,N-3$ dependence with a reducible $(N-2)$-point integral}
In the convergent case, we have seen that $dg_{N}$ can only depends on $g_N$, $g_{N-1}^{(i)}$ and $g_{N-2}^{(i)}$. This picture changes in the divergent case when one of $g_{N-2}^{(i)}$ is reducible, and $dg_{N}$ may develop dependence on some $(N-3)$-point integrals. As a practical example, we consider the dependence of 5-point integrals on 2-point ones. According to Eq.~\eqref{eq:DE-letter}, we have
\begin{align}
d\tilde{g}_5 &= \epsilon \, dM_5 \, \tilde{g}_5 + \epsilon \sum_i dM_{5,4}^{(i)} \, \tilde{g}_4^{(i)} + \epsilon \sum_i dM_{5,3}^{(i)} \, \tilde{g}_3^{(i)} \nn
\\
&+ \epsilon \, dM_{5,2} \, \tilde{g}_2 \, ,
\end{align}
where the cut on $z_1$ and $z_2$ is imposed. Using Eq.~\eqref{eq:UTint} we arrive at
\begin{align}
dM_{5,2} + \mathcal{O}(\epsilon) &= \frac{\epsilon}{8\pi} \, dI_{5,2}(\epsilon) - \frac{\epsilon}{4\pi} \sum_{i=3}^5 dM_{5,4}^{(i)} \, I_{4,2}^{(i)}(\epsilon) \nn \\
&- \frac{\epsilon}{2} \sum_{i=3}^5 dM_{5,3}^{(i)} \, I_{3,2}^{(i)}(\epsilon) \, ,
\label{eq:M52eq}
\end{align}
where
\begin{align}
I_{5,2}(\epsilon) &= \int \frac{dz_3}{z_3} \frac{dz_4}{z_4} \frac{dz_5}{z_5} \left( -\frac{\mK_5}{G_5(0,0,z_3,z_4,z_5)} \right)^\epsilon , \nonumber
\\
I_{4,2}^{(i)}(\epsilon) &= \int \frac{dz_j}{z_j} \frac{dz_k}{z_k} \frac{\sqrt{G_4^{(i)}(0,0,0,0)}}{\sqrt{G_4^{(i)}(0,0,z_j,z_k)}} \nn
\\
&\hspace{4em} \times \left( -\frac{\mK_4^{(i)}}{G_4^{(i)}(0,0,z_j,z_k)} \right)^\epsilon , \nonumber
\\
I_{3,2}^{(i)}(\epsilon) &= \int \frac{dz_i}{z_i} \left( -\frac{\mK_3^{(i)}}{G_3^{(i)}(0,0,z_i)} \right)^\epsilon ,
\end{align}
where $j<k$ and $j,k \neq i$. We note that there is a factor of $\epsilon$ in each term on the right-hand side of Eq.~\eqref{eq:M52eq}. Therefore the term can only contribute if the integral is divergent in the limit $\epsilon \to 0$. For that to happen, at least one of $G_3^{(i)}(0,0,0)$ needs to vanish. For simplicity we assume $G_3^{(3)}(0,0,0) = 0$, while the other two $G_3^{(i)}(0,0,0)$'s are non-zero. In any case, it is clear that the $I_{3,2}^{(i)}(\epsilon)$ terms do not contribute, since they are either zero or non-divergent. The integrals $I_{4,2}^{(4)}(\epsilon)$ and $I_{4,2}^{(5)}(\epsilon)$ are similar to Eq.~\eqref{eq:M42_div} with the result $-i\pi/\epsilon + \mathcal{O}(\epsilon^0)$. Therefore we only need to deal with the divergent part of $I_{5,2}(\epsilon)$:
\begin{align}
I_{5,2}(\epsilon) &= \int \frac{dz_3}{z_3} \frac{dz_4}{z_4} \left[ \Delta(z_3,z_4) \right]^{-\epsilon} \nn \\
&\log\frac{B_5^{(5)}(0,0,z_3,z_4,0) - \sqrt{\Delta(z_3,z_4)}}{B_5^{(5)}(0,0,z_3,z_4,0) + \sqrt{\Delta(z_3,z_4)}} + \mathcal{O}(\epsilon^0) \, ,
\end{align}
where
\begin{equation}
\Delta(z_3,z_4) = \mK_5 G_{4}^{(5)}(0,0,z_3,z_4) \, .
\end{equation}
The integration over $z_4$ is similar to Eq.~\eqref{eq:I53}, except the additional factor $\Delta^{-\epsilon}$, which regularizes the divergence as $z_3 \to 0$. Since we are only interested in the leading term in $\epsilon$, it is equivalent to replace this factor by $z_3^{-\epsilon}$. We can then expand $z_3^{-1-\epsilon}$ in terms of distributions. Keeping only the $1/\epsilon$ terms, we have
\begin{align}
&dI_{5,2}(\epsilon) + \mathcal{O}(\epsilon^0) \nn \\
&= -\frac{1}{\epsilon} d\int \frac{dz_4}{z_4} \log\frac{B_5^{(5)}(0,0,0,z_4,0) - \sqrt{\Delta(0,z_4)}}{B_5^{(5)}(0,0,0,z_4,0) + \sqrt{\Delta(0,z_4)}} \nn\\
\nn \\
&= -\frac{1}{\epsilon}\left( 2\pi i \, dM_{5,4}^{(4)} + 2\pi i\, dM_{5,4}^{(5)} + 4\pi \, dM_{5,3}^{(3)} \right)\, ,
\end{align}
where the second line follows from the calculation of Eq.~\eqref{eq:I53}. We finally arrive at
\begin{equation}
dM_{5,2} = -\frac{1}{2} dM_{5,3}^{(3)} = -\frac{i}{8} d\log \frac{C_5 - \sqrt{-\mK_5 \mK_{3}}}{C_5 + \sqrt{-\mK_5 \mK_{3}}}
\label{eq:W52}
\end{equation}
where
\begin{equation}
C_5 = G(p_1,p_2,p_3,p_4;\, p_1,p_2,p_3,p_4+p_5) \, .
\end{equation}
The result in Eq.~\eqref{eq:W52} is unsurprising given the relation $g_3^{(3)}=-g_2/2$. Similar behaviors are observed when more than one $\widetilde{G}_3$ vanish. The corresponding $dM_{5,2}$ is then a linear combination of several $dM_{5,3}$'s.
We hence conclude that letters in these cases can also be obtained straightforwardly without tedious calculations.
The above discussion relates the appearance of $dM_{N,N-3}$ to the reducibility of one or more $g_{N-2}^{(i)}$'s. One may imagine that, if in addition, one or more $g_{N-3}^{(i)}$'s becomes reducible, there can be $dM_{N,N-4}$ appearing in the differential equations. This is impossible for integrals with generic external momenta (i.e., the $E$ external momenta are indeed independent). However, such cases may arise at certain boundaries of kinematic configurations. When this happens, the corresponding letters can be easily obtained following the reduction rules among the integrals, as was done in the previous paragraph.
\section{Summary and outlook}
\label{sec:summary}
In summary, we have studied the alphabet for one-loop Feynman integrals. The alphabet governs the form of the canonical differential equations, and provides important information on the analytic solution of these equations. We find that the letters in the alphabet can be generically constructed utilizing the UT integrals in the Baikov representation under various cuts. We first considered cases where all the cut integrals are convergent in the limit $\epsilon \to 0$. The corresponding letters coincide with the results in \cite{Abreu:2017ptx, Abreu:2017enx, Abreu:2017mtm}, while our expressions are simpler in certain cases. We have also thoroughly studied the cases of divergent cut integrals. We find that letters in the divergent cases can be easily obtained from the convergent cases by taking certain limits. The letters admit universal expressions in terms of various Gram determinants. We have checked our general results in several known examples, and found agreements. We have also applied our results to the complicated case of a $2 \to 3$ amplitude with 7 physical scales. The details about that is presented in Ref.~\cite{Chen:2022nxt}.
We expect that our results will be useful in many calculations of $2 \to 3$ and $2 \to 4$ amplitudes which are theoretically and/or phenomenologically interesting. It is also interesting to see whether similar universal structures can be obtained at higher loop orders, using the UT integrals in the Baikov representation of \cite{Chen:2020uyk, Chen:2022lzr}.
\begin{acknowledgments}
This work was supported in part by the National Natural Science Foundation of China under Grant No. 11975030, 11635001 and 11925506.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
\IEEEPARstart{I}{nternet} of things (IoT) applications and services have become popular in recent years due to major technological advancements in sensing, communications, and computation \cite{zhang2021survey}. In many IoT applications, devices operate with extremely low power due to the size, cost, and technological limitations. Traditionally, the devices upload their data to a central node (CN), where all data gathered will be analyzed together. However, such an approach is not feasible in many IoT settings, due to privacy issues, limited power, and usually insufficient communication bandwidth \cite{imteaj2021survey}. An alternative solution is to use federated learning (FL) \cite{wahab2021federated}, where each user is responsible for computing the updates to the current global model based on its own local training data and transmitting them to the CN. Using the updates from the devices, CN improves the global model and redistributes the updated global model to the users. This process is repeated until the convergence is achieved \cite{jin2022communication}.
In FL, communication between devices and CN can be significantly slower than local computing \cite{9272666,zhao2021federated}. Since devices need to communicate with CN repeatedly until the convergence is acquired, there exists a trade-off between local computational power and communication overhead. As we increase the number of local iterations, the required communication between users and CN reduces, and vice versa. The effect of large communication overhead on the scalability of FL and trade-off between local updates and global aggregation was investigated in \cite{wang2018giant,8664630}. Another challenge of FL is the system heterogeneity in which different devices have different computational capabilities, power levels, and storage capacities. In \cite{9187874}, wight-based federated averaging was proposed to tackle this issue.
The majority of works on FL examined a simplified model for the communication channel, where transmissions are error-free but rate-limited \cite{amiri2021convergence,ren2020accelerating,yang2019scheduling}. However, since many IoT devices are operating with low power, have limited computational capabilities, and send short packets, the wireless channel between IoT devices and CN is usually erroneous. Recently, a few studies considered FL over a fading wireless channel \cite{zhang2020federated}. In \cite{amiri2020federated}, to enhance the learning accuracy over a fading channel, a new scheme was proposed, in which at each iteration, based on the channel state, only one device is selected for transmission using a capacity-achieving channel code. However, there is still a lack of full understanding of the effect of communication error on the accuracy and convergence of FL approaches. This is particularly important for IoT applications, since the channel between devices and CN is usually weak. Due to the large number of devices, the local updates from some of the devices may not reach the CN, which deteriorates the FL performance. In this paper, we shed some light on the effect of communication errors on the convergence and accuracy of FL algorithms.
We consider a distributed learning problem, where the links between the devices and the CN is modeled by packet erasure channels. In particular, we assume the local updates sent by devices face communication errors; that is, the update is erased or received successfully with a certain probability. We consider two scenarios to calculate the global parameter, where the CN only uses the received fresh local updates and discard missing updates or reuse past updates for devices with missing updates. We further analyse the convergence of these approaches, and prove that by using old local updates in case of errors, the FL algorithm employing gradient descent (GD) converges, and the global parameter will converge to the optimal global parameter. This means that the CN does not necessarily needs fresh updates in every communication round of FL to calculate the global parameter. Instead it can reuse past updates in case of error and continue the FL without jeopardizing the global accuracy. We provide simulation results to verify our analysis and further discuss the convergence behaviour for various datasets with different statistical properties.
The rest of the paper is organized as follows. In Section II, we explain the system model and describe different FL approaches in the presence of communication errors. In Section III, we analyze the convergence and accuracy of the FL approaches in the presence of communication error. Simulation results are provided in Section IV. Finally, Section V concludes the paper.
\section{System Model}
We consider a general federated learning problem, where $N$ device with local datasets, $\mathcal{D}_1$, $\mathcal{D}_2$, $\cdots$, $\mathcal{D}_N$, communicate with the central node (CN). The channel between the $i^{th}$ device and CN is modeled by a packet erasure channel with erasure probability $\epsilon_i$; that is a packet sent from the device to the CN is erased with probability $\epsilon_i$ and received correctly with probability $1-\epsilon_i$. The erasure events for all channels are independent of each other. This model is particularly important for massive IoT scenarios, where many IoT devices communicate with a CN using short packets and due to limited power at devices, the communication channel is mostly erroneous. In particular, let us assume that the local updates from each device is a packet of length $k$ bits, which is encoded using a channel code of rate $R=k/n$, where $n$ is the codeword length. By using the normal approximation bound \cite{Polyanskiy2010}, the packet error rate at the receiver, when the signal-to-noise ratio is $\gamma$, is given by:
\begin{align}
\epsilon\approx Q\left(\frac{n\log_2(1+\gamma)-k+\log_2(n)}{\sqrt{nV(\gamma)}}\right),
\end{align}
where $Q(.)$ is the standard $Q$-function and $V(\gamma)=\left(1-(1+\gamma)^{-2}\right)\log_2^2(e)$ is the channel dispersion \cite{Polyanskiy2010}. The short packet communication can then be modeled by the packet erasure channel with erasure probability $\epsilon$, when $k$, $n$, and $\gamma$ is known. Using this simplified model, the channel quality can be characterized by a single parameter $\epsilon$.
We also assume that the downlink channel, i.e., the channel from the CN to devices, is error-free since the CN can transmit with high power; therefore, the error in the downlink channel can be appropriately mitigated.
\vspace{-4ex}
\subsection{Federated Learning in the error-free scenario}
In FL, each device $i$ calculates the local update and sends the parameters to the CN. Then, CN aggregates all the parameters received from all nodes and calculates the general parameters. In particular, let $w^{(t)}$ denote the local parameter calculated at device $i$ at time instant $t$. By using the gradient descent (GD) method with learning rate $\eta$ and local loss function $F_i:\mathbb{R}^d\to\mathbb{R}$, $w^{(t)}_i$ can be calculated as follows:
\begin{equation}
w^{(t)}_i = w^{(t-1)} - \eta \nabla F_i(w^{(t-1)}).
\end{equation}
Upon receiving all local parameters (assuming that $\epsilon_i=0$ for all devices) from all devices, the CN calculates the global parameter as follows:
\begin{equation}
w^{(t)} = \frac{1}{D}\sum_{i=1}^{N}D_iw^{(t)}_i,
\end{equation}
where $D_i=|\mathcal{D}_i|$ is the size of dataset $\mathcal{D}_i$ and $D=\sum_{i=1}^ND_i$.
\vspace{-3ex}
\subsection{Federated Learning in the presence of communication error}
Here, we consider that the channel between the nodes and the CN is erroneous and that the local updates calculated from some of the nodes may not reach the CN.
\subsubsection{FL with erroneous communication and no memory at CN}
In this scenario, the CN calculates the global parameter using the received local updates only. The global update, in this case, is calculated as follows:
\begin{align}
w^{(t)} = \frac{1}{\sum_{i\in\mathcal{S}(t)}D_i}\sum_{i\in\mathcal{S}(t)}D_iw^{(t)}_i,
\label{eq:globalnomemory}
\end{align}
where $\mathcal{S}(t)$ is the set of all nodes that their updates have been successfully received at CN at time instant $t$.
\subsubsection{FL with erroneous communication and reuse of old local updates}
We consider another scenario, in which the CN uses the previous updates received from the devices in case their new updates are not received. In this case, the global update is calculated as follows:
\begin{align}
w^{(t)} = \frac{1}{D}\left(\sum_{i\in\mathcal{S}(t)} D_iw^{(t)}_i+\sum_{j\in\mathcal{F}(t)}D_jw^{(t-1)}_j\right),
\label{eq:globalreuse}
\end{align}
where $\mathcal{F}(t)$ is the set of all nodes that their updates have not been received at CN at time instant $t$. Here, we assumed that the previous local updates for missing nodes are always available at the CN. This assumption is valid as long as the erasure probability is low.
\begin{figure}[!t]
\centering
\begin{subfigure}[t]{0.69\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{Comp1.eps}
\caption{MSE vs. Comm. round.}
\label{fig:comp1mse}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.29\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{Dataset.eps}
\caption{Datasets.}
\label{fig:dataset1}
\end{subfigure}
\caption{A comparison between error-free and erroneous FL, when the number of devices is $N=3$, $|D_i|=1000$, and GD with learning rate $\eta=0.005$ and maximum 10 iterations at devices is used.}
\label{fig:comp1}
\end{figure}
Fig. \ref{fig:comp1} shows a comparison between the FL schemes with and without communication errors. As can be seen in this figure, when CN does not have a memory, the overall mean squared error (MSE) fluctuates and the FL does not converge. However, when CN can store previous updates from the devices, it can reuse them if the fresh updates are missing (see \eqref{eq:globalreuse}). In this case, the FL algorithm converges to the same MSE as that for FL without any communication error.
\vspace{-2ex}
\section{Performance Analysis of the FL Algorithm in the Presence of Communication Error}
For the simplicity of the analysis, we assume that the size of all local datasets are the same, i.e., $D_i=D/N$, for $i=1,\cdots,N$. When the CN does not have memory, all updates which have not been received will be discarded from the global aggregation step. Let us, for a moment, assume that the number of local iterations at the nodes is sufficiently large. That is, when each node performs a local update, the GD at that device converges. In this case, we define $w^{(t)}_i=w_i$ for $i=1,\cdots,N$. By using \eqref{eq:globalnomemory}, the global update at time instant $t$ is given by:
\begin{align}
w^{(t)} = \frac{1}{|\mathcal{S}(t)|}\sum_{i\in\mathcal{S}(t)}w_i.
\label{eq:convnomemory}
\end{align}
Since the link between the $i^{th}$ device and CN is an erasure channel with erasure probability $\epsilon_i$, the probability mass function $(\mathrm{pmf})$ of $w^{(t)}$ can be calculated as follows:
\begin{align}
\mathrm{Prob}\left\{w^{(t)}=\frac{\sum_{i=1}^NI_iw_i}{\sum_{i=1}^NI_i}\right\} =\prod_{i=1}^N\epsilon_i^{1-I_i}(1-\epsilon_i)^{I_i},
\label{eq:pmfnomemory}
\end{align}
where $I_i\in\{0,1\}$ denote the erasure event for the link between the $i^{th}$ device and CN, i.e., $\mathrm{Prob}\{I_i=1\}=1-\epsilon_i$ and $\mathrm{Prob}\{I_i=0\}=\epsilon_i$. From \eqref{eq:pmfnomemory} it can be easily observed that the global parameter always fluctuates due to the random nature of the erasure events. This can be clearly seen in Fig. \ref{fig:comp1mse}.
\vspace{-2ex}
\subsection{Performance Analysis of the FL algorithm with erroneous communication and reuse of old local updates}
Here, we assume that $F_i(.)$, for all $i\in\{1,\cdots,N\}$, is convex and $L$-smooth. For function $f:\mathbb{R}^d\to\mathbb{R}$ that is convex and $L$-smooth, the following relations hold, $\forall x,y\in\mathbb{R}^{d}$ \cite{MAL-050}:
\begin{align}
\label{convexity}
f(y)&\le f(x)+\nabla f(x)(y-x)'+\frac{L}{2}||y-x||_2^2,\\
\label{smoothness}
f(x^*)&-f(x)\le -\frac{1}{2L}||\nabla f(x)||_2^2, ~x^*=\arg\min_x f(x),\\
\label{smoothconvex}
||\nabla f(&x)- \nabla f(y)||_2\le L||x-y||_2,\\
\left(\nabla f(\right.&x)-\left.\nabla f(y)\right)(x-y)'\ge\frac{1}{L}||\nabla f(x)-\nabla f(y)||_2^2,
\label{smoothc2}
\end{align}
where $(.)'$ denote the matrix transpose operand.
\begin{lemma}
Let $F_i:\mathbb{R}^d\to \mathbb{R}$, for all $i\in\{1,\cdots,N\}$, is convex and $L$-smooth. $F_{\mathcal{G}}(x)=(1/N)\sum_{i\in\mathcal{G}}F_i(x)$ is also convex and $\frac{|\mathcal{G}|L}{N}$-smooth.
\end{lemma}
\begin{IEEEproof}
This can be easily proved from the sub-additivity of the norm and additivity of the gradient \cite{MAL-050}.
\end{IEEEproof}
From Lemma 1, it can be easily proved that $F(x)=(1/N)\sum_{i=1}^N F_i(x)$ is also convex and $L$-smooth. In the following theorem, we show that the FL algorithm in the presence of communication error, where the CN uses the past local updates in case of communication error, converges to the global minima of the global loss function and accordingly the optimal global parameter.
\begin{theorem}
Let us consider a FL problem with $N$ devices, where the channel from each device to the CN is modeled by an erasure channel with erasure probability $\epsilon$. We assume that the local loss function $F_i(x)$ at device $i$ is convex and $L$-smooth. We further assume that $||\nabla F(x)- \nabla F(y)||_2\ge\mu||x-y||_2$, for all $x,y\in \mathbb{R}^d$, where $F(x)=\frac{1}{N}\sum_{i=1}^NF_i(x)$. Let $\delta_t=||w^{(t)}-w^*||_2^2$, where $w^*=\arg \min_w F(w)$, and $\bar{\delta}_{t+1}=\frac{1}{t+1}\sum_{i=0}^{t}\delta_i$. For the FL algorithm \eqref{eq:globalreuse}, when $\epsilon\le\frac{\mu}{2L}$ and $\eta=\frac{1}{L}$, $\bar{\delta}_{k}$ is upper bounded by:
\begin{align}
\bar{\delta}_{t}\le\frac{F(w^{(0)})-F(w^*)}{t\beta^2},~~\text{for}~t>0,
\end{align}
where $\beta^2=\frac{\mu^2}{2L}-2L\epsilon^2$.
\end{theorem}
\begin{IEEEproof}
The proof is provided in Appendix A.
\end{IEEEproof}
Theorem 1 states that the gap to the global minima ($w^*$) decreases with the iteration number $t$ and converges to zero, when $t$ is arbitrary large, i.e., $\lim_{t\to\infty} \bar{\delta}_{t}=0$. it is also easy to show that $\lim_{t\to\infty} \delta_{t}=0$, since otherwise, if $\delta_{t}\ge\epsilon$, where $\epsilon>0$, $\bar{\delta}_{t}$ will be always bounded above $\epsilon$. It is important to note that Theorem 1 does not state that $\delta_t$ is a decreasing function of $t$. Instead it shows that in the limit that $t$ is sufficiently large, the global parameter, $w^{(t)}$, converges to the optimal global parameter $w^*$, even in the presence of communication error.
\section{Results and Discussion}
In this section, we focus on the FL algorithm, where the CN uses old local updates received from devices when their fresh updates are not available due to communication error. Fig. \ref{fig:comp3mse} shows the overall MSE of the FL algorithm at various erasure probabilities, when $N=3$. We use the same dataset as in Fig. \ref{fig:dataset1}. As can be seen in Fig. \ref{fig:comp3mse}, when the erasure rate is small, i.e., $\epsilon=0.1$, the performance of the FL with reuse of old updates closely approaches that of the FL without communication errors. However, when the erasure probability increases, the MSE in the early iterations has a significant gap to the ideal FL case with no error. In all scenarios, even when the erasure rate is large, i.e., $\epsilon=0.5$, the FL algorithm with reuse of old updates in the presence of error converges to that of the ideal FL without error.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{3users.eps}
\caption{A comparison between error-free and erroneous FL, when the number of devices is $N=3$, $|D_i|=1000$, and GD with learning rate $\eta=0.005$ and maximum 1 iterations at devices is used.}
\label{fig:comp3mse}
\vspace{-1ex}
\end{figure}
Fig. \ref{fig:comp10mse} shows the performance of FL algorithm with reuse of old updates at various erasure probabilities, when the number of devices is $N=10$. The datasets are shown in Fig. \ref{fig:dataset10}. The data is created by using a non-linear model $y=x^2+z$, where $z\sim\mathcal{N}(0,\sigma^2)$, is additive white Gaussian noise. As can be seen in Fig. \ref{fig:dataset10}, the linear regression curve for each local dataset has a different slope; therefore, losing any dataset due to error will result in a significant change in the global parameter. This model is of particular importance for IoT scenarios, where devices are distributed in a field at various locations; therefore, their measurements are location/time dependent and accordingly their datasets are non-iid. As can be seen in Fig. \ref{fig:comp10mse}, when the CN reuses past local updates in case of error, the MSE performance of the FL algorithm converges to that of the ideal FL algorithm without error.
\begin{figure}[!t]
\centering
\begin{subfigure}[t]{0.69\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{10users.eps}
\caption{MSE vs. Comm. round.}
\label{fig:comp10mse}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.29\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{10usersdataset.eps}
\caption{Datasets.}
\label{fig:dataset10}
\end{subfigure}
\vspace{-1ex}
\caption{A comparison between error-free and erroneous FL, when the number of devices is $N=10$, $|D_i|=100$, and GD with learning rate $\eta=0.005$ and maximum 1 iterations at devices is used.}
\label{fig:comp10}
\vspace{-0.5em}
\end{figure}
It is important to note that when the dataset is uniformly distributed among devices and the local parameters are not significantly different, even without reusing old local updates, the FL algorithm converges. An example is provided in Fig. \ref{fig:compunif}, where the dataset is uniformly distributed between 3 devices. In this case, the local parameter from all devices are relatively close to each other. Therefore, missing some devices' updates does not affect the overall performance. As can be seen in Fig. \ref{fig:comp1mseunif}, the FL algorithm without memory at the CN outperforms the FL algorithm with reusing local updates, when the dataset is uniformly distributed among devices. This can be explained for an extreme case, when $D$ is relatively large, no error occurs up until the $i$th communication round, and all local updates are the same in each round. In this case, by using \eqref{eq:globalnomemory}, we will have $w^{(t)}=w^{(t)}_i=w^{(t-1)}-\eta\nabla F_i(w^{(t-1)})$. Since the datasets are uniformly distributed, $w^{(t)}=w^{(t-1)}-\eta\nabla F(w^{(t-1)})$, even if an error occurs. However, if in the case of error we reuse old updates, by using \eqref{eq:globalreuse} the global parameter will be $w^{(t)}=w^{(t)}_i-\epsilon\left(w^{(t)}_i-w^{(t-1)}_i\right)$. This is equivalent to $w^{(t)}=w^{(t-1)}-\eta\nabla F(w^{(t-1)})-\epsilon\left(w^{(t)}_i-w^{(t-1)}_i\right)$, which means that the global parameter will have a gap, proportional to the erasure rate, to that of the error-free case. Therefore, for uniformly distributed datasets, by reusing old updates in case of communication error, the FL algorithm takes longer to converge.
\begin{figure}[!t]
\centering
\begin{subfigure}[t]{0.69\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{CompUnif.eps}
\caption{MSE vs. Comm. round.}
\label{fig:comp1mseunif}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.29\columnwidth}
\centering
\includegraphics[width=0.95\columnwidth]{unifDataset.eps}
\caption{Datasets.}
\label{fig:datasetunif}
\end{subfigure}
\vspace{-1ex}
\caption{A comparison between error-free and erroneous FL, when the number of devices is $N=3$, $|D_i|=100$, and GD with learning rate $\eta=0.01$ and maximum 1 iterations at devices is used.}
\label{fig:compunif}
\vspace{-0.5em}
\end{figure}
We further consider an image classification task using a real dataset from MNIST \cite{mcmahan2017communication}, which consists of 4000 handwritten images of the numbers $0$ to $3$. The example runs in parallel using 4 workers (i.e., devices), each processing images of a single digit. In particular, worker $i$ has 700 handwritten images of number $i-1$ as its training set. The validation and test set at the CN each has 150 images of each number $0$ to $3$. Similar to \cite{mcmahan2017communication}, at each worker, we use a CNN with two $5\times 5$ convolution layers (the first with 32 channels, the second with 64, each followed with $2\times 2$ max pooling), a fully connected layer with 512 units and ReLu activation, and a final softmax output layer. We also use the stochastic GD optimizer with learning rate $0.001$. Fig. \ref{fig:cifar10} shows the accuracy of the FL algorithm after each communication round in the presence of communication error. As can be seen in this figure, when the CN uses the old local updates if the fresh updates are erased due to communication error, the FL algorithm converges to the same level of accuracy as that for the FL algorithm without any communication error. However, when the CN does not store past local updates and only uses fresh updates and discard missing updates in the aggregation stage, the accuracy changes significantly with the communication round.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{4u_1epoch_accuracy_CIFAR10.eps}
\vspace{-1ex}
\caption{Test accuracy of FL using the MNIST dataset and $N=4$ devices.}
\label{fig:cifar10}
\vspace{-1ex}
\end{figure}
\section{Conclusions}
In this paper, we studied the federated learning algorithm in the presence of communication errors. We modelled the channel between the devices and the central node (CN) by packet erasure channels. We presented two approaches to deal with error events. That is when an error occurs, and local updates are not received at CN, the CN can either calculate the global parameter by using only the fresh local updates received correctly or reuse the old updates for missing local updates. We proved that the FL algorithm with reusing local updates in case of error converges to the same global parameters as that with FL without any communication error. This means that the CN does not need to wait for the retransmission of missing updates. Instead, it can reuse the past local updates to calculate the global parameters and continue the FL algorithm. This approach converges to the same result as FL achieves with no communication error. This is of practical importance as IoT devices can significantly save energy by relaxing the communication reliability requirements without jeopardizing the overall learning accuracy. We further provided simulation results to verify our theoretical results. We also highlighted that when the dataset is uniformly distributed among devices, the FL algorithm without memory at the CN converges faster than that with reusing local updates.
\appendices
\section{Proof of Theroem 1}
Assuming that $D_i=D/N$, \eqref{eq:globalreuse} can be written as follows:
\begin{align}
w^{(t+1)}&= \frac{1}{N}\sum_{i\in\mathcal{S}(t)}w^{(t)}_i+\frac{1}{N}\sum_{j\in\mathcal{F}(t)}w^{(t-1)}_j\\
\nonumber&=\frac{1}{N}\sum_{i\in\mathcal{S}(t)}\left(w^{(t)}-\eta\nabla F_i(w^{(t)})\right)\\
\nonumber&+\frac{1}{N}\sum_{j\in\mathcal{F}(t)}\left(w^{(t-1)}-\eta\nabla F_j(w^{(t-1)})\right)\\
\nonumber&=w^{(t)}-\eta\nabla F(w^{(t)})+\frac{|\mathcal{F}(t)|}{N}\left(w^{(t-1)}-w^{(t)}\right)\\
&+\eta\nabla F_{\mathcal{F}}(w^{(t)})-\eta\nabla F_{\mathcal{F}}(w^{(t-1)}),
\label{eq:iteraton}
\end{align}
where $F_{\mathcal{F}}(x)=\frac{|\mathcal{F}(x)|}{N}\sum_{j\in\mathcal{F}(t)}F_j(x)$. Since $F(x)$ is convex and $L$-smooth, by using \eqref{convexity}, we have:
\begin{align}
\nonumber F(&w^{(t+1)})\le F(w^{(t)})+\nabla F(w^{(t)})\left(w^{(t+1)}-w^{(t)}\right)'\\
\nonumber &+\frac{L}{2}||w^{(t+1)}-w^{(t)}||_2^2\\
\nonumber &\overset{\eqref{eq:iteraton}}{=}F(w^{(t)})-\eta||\nabla F(w^{(t)})||_2^2\\
\nonumber&+\frac{|\mathcal{F}(t)|}{N}\nabla F(w^{(t)})(w^{(t-1)}-w^{(t)})'\\
\nonumber &+\eta\nabla F(w^{(t)})(\nabla F_{\mathcal{F}}(w^{(t)})-\nabla F_{\mathcal{F}}(w^{(t-1)}))'\\
\nonumber &+\frac{L}{2}\eta^2||\nabla F(w^{(t)})||_2^2+\frac{L|\mathcal{F}(t)|^2}{2N^2}||w^{(t-1)}-w^{(t)}||_2^2\\
\nonumber &+\frac{L}{2}\eta^2||\nabla F_{\mathcal{F}}(w^{(t)})-\nabla F_{\mathcal{F}}(w^{(t-1)})||_2^2\\
\nonumber &-\frac{L|\mathcal{F}(t)|}{N}\eta\nabla F(w^{(t)})(w^{(t-1)}-w^{(t)})'\\
\nonumber &-L\eta^2\nabla F(w^{(t)})(\nabla F_{\mathcal{F}}(w^{(t)})-\nabla F_{\mathcal{F}}(w^{(t-1)}))'\\
\nonumber &-\frac{\eta L|\mathcal{F}(t)|}{N}(\nabla F_{\mathcal{F}}(w^{(t)})-\nabla F_{\mathcal{F}}(w^{(t-1)}))(w^{(t)}-w^{(t-1)})'
\end{align}
Now, assuming that $\eta=\frac{1}{L}$ and due to the fact that $|\mathcal{F}_F|\approx \epsilon N$, when $N$ is sufficiently large, this can be simplified to:
\begin{align}
\nonumber &F(w^{(t+1)})\le F(w^{(t)})-\frac{1}{2L}||\nabla F(w^{(t)})||_2^2\\
\nonumber &+\frac{L\epsilon^2||w^{(t)}-w^{(t-1)}||_2^2}{2}
+\frac{||\nabla F_{\mathcal{F}}(w^{(t)})-\nabla F_{\mathcal{F}}(w^{(t-1)})||_2^2}{2L}\\
\nonumber &-\epsilon(\nabla F_{\mathcal{F}}(w^{(t)})-\nabla F_{\mathcal{F}}(w^{(t-1)}))(w^{(t)}-w^{(t-1)})'\\
\nonumber &\overset{(a)}{\le} F(w^{(t)})-\frac{1}{2L}||\nabla F(w^{(t)})||_2^2+\frac{L\epsilon^2}{2}||w^{(t)}-w^{(t-1)}||_2^2\\
\nonumber &-\frac{1}{2L}||\nabla F_{\mathcal{F}}(w^{(t)})-\nabla F_{\mathcal{F}}(w^{(t-1)})||_2^2\\
&\le F(w^{(t)})-\frac{||\nabla F(w^{(t)})||_2^2}{2L} +\frac{L}{2}\epsilon^2||w^{(t)}-w^{(t-1)}||_2^2,
\end{align}
where step $(a)$ follows from \eqref{smoothc2} and Lemma 1, which indicates that $F_{\mathcal{F}}(.)$ is convex and $Le$-smooth. Since we assumed that $||\nabla F(x)- \nabla F(y)||_2\ge\mu||x-y||_2$, for all $x,y\in\mathbb{R}^d$, we have:
\begin{align}
\nonumber F(w^{(t+1)})&\le F(w^{(t)})-\frac{\mu^2}{2L}||w^{(t)}-w^*||_2^2\\
&+\frac{L}{2}\epsilon^2||w^{(t-1)}-w^{(t)}||_2^2.
\label{2nditeration}
\end{align}
It is easy to show that $||w^{(t-1)}-w^{(t)}||_2^2\le 2\left(\delta_t+\delta_{t-1}\right)$. We can further simplify \eqref{2nditeration} as follows:
\begin{align}
F(w^{(t+1)})\le F(w^{(t)})+\left(L\epsilon^2-\frac{\mu^2}{2L}\right)\delta_t+L\epsilon^2\delta_{t-1}.
\end{align}
Summing up both sides over $t=1,\cdots, k$, and using telescopic cancellation, we have:
\begin{align}
\nonumber F(w^{(k+1)})&\le F(w^{(0)})+(2L\epsilon^2-\frac{\mu^2}{2L})\sum_{i=1}^{k-1}\delta_i\\
&+(L\epsilon^2-\frac{\mu^2}{2L})(\delta_k+\delta_0),
\label{eqtelescope}
\end{align}
where we assumed that the first global update ($t=1$), is calculated without any communications error. That is $F(w^{(1)})\le F(w^0)-\frac{\mu}{2L}\delta_0$. Assuming that $\epsilon\le\frac{\mu}{2L}$, we have $\frac{\mu^2}{2L}-2L\epsilon^2<\frac{\mu^2}{2L}-L\epsilon^2$. Therefore, \eqref{eqtelescope} is simplified to:
\begin{align}
F(w^{(k+1)})&\le F(w^{(0)})-\beta^2(k+1)\bar{\delta}_{k+1},
\end{align}
where $\beta^2=\frac{\mu^2}{2L}-2L\epsilon^2$. By rearranging the above inequality, we have:
\begin{align}
\bar{\delta}_{k+1}\le\frac{F(w^{(0)})-F(w^{(k+1)})}{(k+1)\beta^2}.
\end{align}
Since $F(w^*)\le F(w^{(k+1)})$, We will have:
\begin{align}
\bar{\delta}_{k+1}\le\frac{F(w^{(0)})-F(w^*)}{(k+1)\beta^2}.
\end{align}
\bibliographystyle{IEEEtran}
\footnotesize
|
\section{Introduction}
Networks are a useful abstraction for many real-world systems, representing interactions between objects within a system. Network analysis examines relationships among entities, such as persons, organizations, or documents \cite{wright_2015} and has been extensively adopted for modeling systems in various domains with a long history of applications \cite{doi:10.1137/S003614450342480} \cite{Costa_2011}, including neuroscience \cite{şimşek2021geometry}, biology \cite{Leifer_2020}, and social networks \cite{doi:10.1177/016555150202800601}. Networks exist in various forms: weighted or unweighted, directed or undirected. However, many complex systems are more accurately modeled by directed weighted networks, where the relation between two different entities in the system is asymmetric and exists in a range of intensities.
For example, citation networks, immigration networks, and trade networks are all accurately modeled by directed weighted networks.
Directed networks are mathematically represented as a directed graph where vertices represent the objects or entities in the system, and edges encode the interaction between individual objects. Previous studies have implemented a wide range of measures for studying networks \cite{rider_2005}, including node centrality \cite{ficara2021correlation}, clustering coefficients \cite{Masuda_2018} and path lengths between nodes \cite{melnik2016simple}. In this paper, we focus on analyzing node centrality in directed weighted networks.
Node centrality measures the influence of a node in the entire network by assigning rankings and numbers to nodes within the network based on their network position. Centrality measures enable us to detect various real-world phenomena, including the identification of banks that are too-connected-to-fail \cite{GOFMAN2017113} and decisions regarding human capital or education \cite{RePEc:cpr:ceprdp:10631}. Different centrality measures have been developed to capture different behaviors. For example, betweenness centrality has been widely used in social networks \cite{Lee_2021}, eigenvector centrality has proved to be useful in temporal networks \cite{taylor2016eigenvectorbased}, optimal percolation centrality has been widely used in large networks \cite{delima2020estimating}, and K-core centrality has been widely used in dynamic networks \cite{Liu_2020}.
In this paper, we are interested in studying the propagating properties of a network since influences of individuals in many real-world complex systems have the potential to propagate over the entire system. This phenomena can be meaningfully illustrated on the global trade network, where globalization of trade and intertwined economies around the world can cause economic perturbations originated in a single country propagate elsewhere.
Moreover, since world economies are exhibiting increasing levels of local heterogeneity and global interdependency \cite{Serrano_2007}, it is crucial to take a topological standpoint on defining the centrality measure for understanding propagating effects in trade networks. Thus, we define a new centrality measure using persistent homology, namely the \emph{quasi-centrality} (Definition \ref{def: quasi-central nonno}) to determine the ranks of the nodes in directed weighted networks based on this property.
Persistent homology is a central tool used in topological data analysis (TDA), which is a burgeoning field in math and data analysis where concepts from algebraic topology are used to simplify, summarize, and compare complex data sets.
TDA has found successful applications in neuroscience \cite{DBLP:journals/ficn/Dabaghian20}, biological models \cite{Topaz_2015} machine learning \cite{DBLP:journals/corr/ChepushtanovaEH15}, and other related fields in statistics, math, physics, and biology. However, while many applications of TDA have focused on simplifying and identifying the intrinsic shapes of complex data sets, it has not been extensively applied to analyze networks. Hence the significance of our work is shown by utilizing TDA concepts in examining network properties.
The quasi-centrality measure (Definition \ref{def: quasi-central nonno}) is defined based on the idea that if a node plays a crucial role in the connectivity of the network, we would expect perturbations originating from this node to propagate significantly through the network. Roughly, we measure a node's role in the overall connectivity of a directed network by computing the difference between the connectivity of the network before and after deleting this node. In particular, we use the ``size'' of homology groups as a proxy to determine the connectivity of a directed network (Remark \ref{def: rem}).
We show that the quasi-centrality is optimal for measuring the influence of perturbations originated from individual nodes by (1) computing the quasi-centralities on the star-shape network in Example \ref{ex: example} and comparing with existing centrality measures, and (2) analyzing the importance of country-industry pairs in the Asia machinery production network (Section 4) by computing the quasi-centralities and comparing with existing centrality measures.
Moreover, given a directed weighted network $G$, we introduce a method (Definition \ref{def: hierarc}) that extracts a hierarchical representation of nodes in $G$ based on their topological impacts. This method combines the bottleneck distance (Definition \ref{defn:bottleneck}), a tool used in TDA, with hierarchical clustering, a method used in cluster analysis that determines similarity between objects.
We present our findings as follows: in Section 2, we provide background knowledge in networks and existing centrality measures, algebraic topology, and TDA (specifically persistent homology).
In Section 3, we define quasi-centrality (Definition \ref{def: quasi-central nonno}).
In Section 4, we illustrate the practicality of quasi-centrality by computing it on the Asia machinery production network.
In Section 5, we present our method that determines the hierarchy of nodes in a given directed weighted network, and analyze the hierarchical dendrograms extracted from the Asia machinery production networks. And finally in Section 6, we summarize our results and discuss future directions.
\paragraph{Data and implementations:} Our data sets and software are available on \url{https://github.com/lindahe8989/A-topological-centrality-measure-for-directed-networks}.
For plotting persistence diagrams and hierarchical dendrograms, we partially used \url{https://github.com/fmemoli/PersNet} for reference.
\section{Preliminary}
In this section, we provide background knowledge in algebraic topology and persistent homology, as well as related work in network analysis. We begin by introducing networks.
\subsection{Networks}
\begin{definition}
A \emph{network} is a pair $(X, w_X)$ where $X$ is a finite set and $w_X:X \times X \rightarrow \mathbb{R}$ is called the weight function. An \emph{undirected network} is a network such that $w_X(x_1,x_2)=w_X(x_2,x_1)$ for all $x_1, x_2 \in X$, and a \emph{directed network} is a network such that $w_X(x_1,x_2) \not = w_X(x_2,x_1)$ for some $x_1, x_2 \in X$.
\end{definition}
Our definition of quasi-centrality (Definition \ref{def: quasi-central nonno}) is restricted to dissimilarity networks, introduced below:
\begin{definition}
A \emph{dissimilarity network} is a network $G=(X, w_X)$ where $w_X: X \times X \rightarrow \mathbb{R}$ is called a \emph{dissimilarity function}, i.e. a map such that $w_X(x, x')=0$ if and only if $x=x'$ for all $x, x' \in X$.
\end{definition}
\begin{remark}
A dissimilarity network can be mathematically represented as a directed graph without self-loops.
\end{remark}
Given a directed network $G=(X,m_X)$, the dissimilarity function $m_X(x_1, x_2)$ describes the intensity of interaction between the two nodes. We can associate a distance metric $w_X$ to $G$ such that if two nodes interact frequently, i.e., $m_X(x_1, x_2)>>0$, then they are close to each other in this distance metric, i.e., $w_X(x_1, x_2) \approx 0$. We adapt the definition of \emph{effective distance} \cite{brockmann_helbing_2013}\footnote{Successful applications include \cite{Iannelli_2017}, \cite{brockmann_helbing_2013}, etc.}, which has been proved to have successfully interpreted the relationship between distance and the volume of interaction between two nodes into our distance metric $m_X$ as follows:
\begin{definition}
Let $G=(X, w_X)$ be a (dissimilarity) network. Define $\gamma(G)$ to be $(X, m_X)$ where $m_X: X \times X \rightarrow \mathbb{R}$ is given by:
\[m(x_i, x_j)=\begin{cases}
1-\log \frac{w(x_i, x_j)}{\sum_{k \not = j}w(x_k, x_j)} \geq 1 \text{ if } i \not = j\\
0 \text{ if } i=j.
\end{cases} \]
\label{def: gamma weight}
\end{definition}
Hence in $\gamma(G)$, two vertices $x_i,x_j$ with large volume of interaction $w(x_i, x_j)$ will correspond to a closer distance $m(x_i, x_j)$.
\begin{remark}
The notion of effective distance will be used heavily in our definition of quasi-centrality (Definition \ref{def: quasi-central nonno}).
\end{remark}
\begin{remark}
For simplicity in our calculation, when $w(x_i,x_j)=0$, we choose a large number for $m(x_i,x_j)$, i.e., we let the distance between nodes $x_i, x_j$ in the new metric $m$ be very far apart so that it is almost indistinguishable when no relation exists between the two nodes. In Example \ref{ex: example}, we let $m(x_i,x_j)=24.026$.
\end{remark}
\subsection{Existing centrality measures}
Given a network $G=(X, w_X)$, we can study the qualitative features of the network. One such interesting feature is to detect the importance of individual nodes in the network. In particular, \emph{node centrality} measures have been extensively applied on empirical networks to detect ranks and influence of individual nodes.
Here we summarize existing measures of node centrality for directed networks. In Section 5 we will use these as a comparison with quasi-centrality (Definition \ref{def: quasi-central nonno}).
In the following definitions, we let $G=(X, w_X)$ be a directed network and $A_{ij}$ be its adjacency matrix.
\begin{itemize}
\item \emph{Degree centrality}: for a directed network $G$, the \emph{in-degree centrality} and \emph{out-degree centrality} for a node $i \in X$ are defined to be the number of edges going in or out of $i$.
\item \emph{Katz centrality}: for a directed network $G$, the the Katz centrality $x_i$ for node $i \in X$ is defined to be $x_i=\alpha \sum_j A_{ij}x_j+\beta$, where $\alpha$ and $\beta$ are constants and $A_{ij}$ is an element of the adjacency matrix.
\item \emph{Pagerank centrality}: for a directed network $G$, the Pagerank centrality $x_i$ for node $i \in X$ is defined to be $x_i=\alpha \sum_j A_{ij}\frac{x_j}{k_j^{out}}+\beta$, where $\alpha$ and $\beta$ are constants, $A_{ij}$ is an element of the adjacency matrix, and $k_j^{out}$ is the out-degree of the node $i$.
\item \emph{HITS hubs and authorities centrality}: The HITS hubs and the HITS authorities centrality measures are defined such that one measure depends on the centrality measure determined by the other measure.
Let $\alpha$ and $\beta$ be non-negative constants. For a directed network $G$, the HITS-authority centrality $x_i$ for node $i \in X$ is defined to be the sum of the hub centralities $y_j$ which point to the node $i$: $x_i=\alpha \sum_jA_{ij}y_j$, and the HITS-hubs centrality $y_i$ for node $i \in X$ is defined to be the sum of the authority centrality $x_j$ which are pointed by the node $i$: $y_i=\beta \sum_jA_{ji}x_j$.
\end{itemize}
We note that all of the above centrality measures are based on the adjacency matrix $A_{ij}$, showing that the centrality measure for a node $i$ is completely dependent on the direct neighbors of the node, while not taking the propagating influence of the node on the entire network into account. The goal of our research is to provide a centrality measure that takes propagating influences into account, and we do so by using algebraic topology and persistent homology, which we will provide background knowledge in the following section.
\subsection{Algebraic topology and homology}
We begin by introducing simplicial complexes, which are the foundational objects of study in algebraic topology and homology.
\begin{definition}
Given a finite set $ T $, an \emph{(abstract) simplicial complex} $ X $ is a subset of $Pow(T)$, the power set on $ T $, such that
\begin{itemize}
\item singletons $\{x\} \subseteq T$ belong to $ X$, and
\item whenever $\sigma \in Pow(T)$ belongs to $X$, then any subset $\tau \subseteq \sigma$ belongs to $ X $ as well.
We refer to a proper subset $\tau$ of $\sigma$ is as a \emph{face} of $\sigma$.
\end{itemize}
\end{definition}
The elements of $T$ are called the \emph{vertices} of $ X $, the two-element subsets of $ T $ that belong to $ X $ are called the \emph{edges}, and in general the $k$-element subsets of $T$ are called the ($k$-1)-simplices in $ X $.
Given a simplex $ \sigma = \{x_0, \ldots , x_n\} $, an \emph{orientation} of $ \sigma$ is a choice of an equivalence class of the orderings of its vertices. We say that two orderings are \emph{equivalent} if they differ by an even permutation. For $n > 0$, there are thus two equivalence classes of orderings.
A choice of one of the classes will be called an orientation of $\sigma$. If $\sigma = \{v_0, ..., v_n\}$, then $\sigma$ with a choice of an orientation will be denoted by $\sigma = [v_0: ...: v_n]$.
We now define homology for simplicial complexes. Let $X$ be a simplicial complex, and let $X_n$ denote the set of oriented $n$-simplices in $X$, then for any dimension $n \in \mathbb{Z}_{+}$, the \emph{vector space of $n$-chains} of $X$ over the field $\mathbb{K}$, denoted by $C_n(X, \mathbb{K})$, is the free $\mathbb{K}$-vector space with basis the set of oriented $n$-simplices:
\[C_n(X,\mathbb{K})=\left \{ \sum \alpha_i\sigma_i: \sigma_i \in X_n, \alpha_i \in \mathbb{K}\right \}. \]
\begin{definition}
Given any $n$-chain vector space $C_n(X,\mathbb{K})$, where $n \in \mathbb{Z}_{+}$, we define the boundary map $\partial_n: C_n(X,\mathbb{K}) \rightarrow C_{n-1}(X,\mathbb{K})$ as the the linear transformation specified on the generators as
\[\partial_n([v_0,...,v_n])=\sum_{i=1}^{n} (-1)^i[v_0,...,v_{i-1},v_{i+1},...,v_n].\]
\end{definition}
Informally, $\partial_n$ maps each $(n+1)$-simplex to its boundary comprised of its faces. It is easy to check that the boundary of a boundary is empty, and hence the linear maps $\partial_n$ satisfy the property that composing any two consecutive boundary maps yields the zero map:
\begin{prop}\cite{Munkers84}
\label{prop: chain linear transform}
For all $n \in \mathbb{Z}_{+}$, $\partial_{n-1} \circ \partial_{n}=0$.
\end{prop}
We axiomatize the properties of $C_n(X; \mathbb{K})$ in the definition of a chain complex. Namely, a \emph{chain complex of vector spaces} $A_{\bullet}$ is
a collection of vector spaces $\{ A_n\}$ such that
the linear transformations $\partial_n: A_n \rightarrow A_{n-1}$ satisfy $\partial_{n-1} \circ \partial_{n}=0$ for all $n \in \mathbb{Z}_+$.
From the chain complex $A_{\bullet}$, one may define the following two subspaces:
\[\emph{k-cycles}: Z_k(A_{\bullet}):= \ker(\partial_k)=\{a \in A_k: \partial_ka=0 \}, \]
\[\emph{k-boundaries}: B_k(A_{\bullet}):= \im(\partial_k)=\{a \in A_k: a= \partial_{k+1}b\}. \]
\begin{remark}
By Proposition \ref{prop: chain linear transform}, the image of $\partial_{n+1}$ is contained in the kernel of $\partial_{n}$, so we can take the quotient of kernel($\partial_{n}$) by the image($\partial_{n+1}$).
\end{remark}
\begin{definition}
For a chain complex $A_{\bullet}$, the $n$th \emph{homology group} $H_n(A_{\bullet})$ is defined as the quotient
\[H_n(A_{\bullet})=\text{ker}(\partial_n)/\text{im}(\partial_{n+1}).\]
\end{definition}
Roughly, we can think of homology groups as the subspace of homotopy classes of cycles in $Z_k(A_{\bullet})$ that are not the boundaries of elements of $B_{k+1}(A_{\bullet})$.
The quotient vector space $H_n(A_{\bullet})$ is called the \emph{k-th homology of the chain complex} $A_{\bullet}$.
\begin{definition}
The dimension of $H_n(A_{\bullet})$ over $ \mathbb{K}$ is called the \emph{$n$-th Betti number of} $A_{\bullet}$, denoted by $\beta_n(A_{\bullet})$.
\end{definition}
\begin{example}
Suppose $X$ is the simplicial complex homeomorphic to the boundary of a triangle with vertices $v_0, v_1$, and $v_2$. Then $C_0(X,\mathbb{K}) = \mathbb{K}^3$, generated by the vertices $[v_1], [v_2], [v_3]$, and $C_1(X, \mathbb{K}) = \mathbb{K}^3$, generated by the edges $[v_0, v_1], [v_1, v_2], [v_0, v_2]$, and $C_n(X, \mathbb{K}) = 0$ for all $n > 1$. Then
\[H_0(X) = \frac{Z_0(X)}{B_0(K)}= \frac{\langle [v_0], [v_1], [v_2] \rangle}{\langle [v_0]-[v_1], [v_0]-[v_2] \rangle}=\mathbb{K}, \]
\[H_1(X) = \frac{Z_1(X)}{B_1(K)}= \frac{\langle [v_0, v_1] - [v_0, v_2] + [v_0, v_2] \rangle}{\langle 0 \rangle}=\mathbb{K}. \]
All other homology groups are trivial because $ C_n(X; \mathbb{K}) = 0 $ for $ n \geq 2$. The \emph{geometric realization}\footnote{See Section 1.8 in \cite{rabadan.raul.blumberg.20}: vaguely, the geometric realization $|X|$ of a simplicial complex $X$ is the operation that interprets each algebraic $n$-simplex as a standard topological $n$-simplex gluding together in a ``nice'' way. } $|X|$ of this simplicial complex $X$ is homotopic\footnote{See Section 1.8 in \cite{rabadan.raul.blumberg.20}: vaguely, two topological spaces $X$ and $Y$ are homotopic if one can be continuously deformed into the other.} to $S^1$, so this example computes the homology of $S^1$.
Note that the dimension $ H_0 (S^1)$ corresponds to the number of path components of $S^1$, and the dimension of $H_1 (S^1) $ corresponds to the number of 1-dimensional cycles in $S^1$.
In general, if we have a $n$-dimensional sphere $S^n$, then $H_n (S^n)=\mathbb{K} $ and $H_0 (S^n)=\mathbb{K} $ and all other homology groups are trivial.
\end{example}
In fact, the homology groups have a very nice interpretation: for all simplicial complexes $ X $, $ H_0(X) $ is the free $\mathbb{K}$-vector space on the set of \emph{0-dimensional cycles} of $X$ modulo boundaries. i.e., two vertices are equivalent if there exists a sequence of edges between them, which can be visualized as \emph{path components} in $|X|$; $H_1(X)$ is the free $\mathbb{K}$-vector space on the set of \emph{1-dimensional cycles} of $X$ modulo boundaries, which can be visualized as \emph{1-dimensional holes} in $|X|$; more generally, $H_k(X)$ is the free $\mathbb{K}$-vector space on the set of \emph{k-dimensional cycles} of $X$ modulo boundaries, and can be visualized as \emph{k-dimensional holes} in $|X|$.
\begin{fact}
Let $X$ and $Y$ be topological spaces\footnote{In fact, we can define homology for ``sufficiently nice" topological spaces, not just abstract simplicial complexes.}, then a continuous map $ f: X \to Y $ induces a $ \mathbb{K}$-linear map $ H(f): H_*(X; \mathbb{K}) \to H_*(Y; \mathbb{K}) $.
\end{fact}
\begin{remark}
Let $X$ and $Y$ be topological spaces.
If $ f, g: X \to Y $ are homotopic, i.e., they can be continuously deformed into each other, then they induce the same $\mathbb{K}$-linear homomorphisms on the respective homology groups.
\end{remark}
\begin{remark}
If two topological spaces $X,Y$ are homotopy equivalent\footnote{Two topological spaces $X$ and $Y$ are homotopy equivalent if there exists a pair of continuous maps $f: X \rightarrow Y$ and $g: Y \rightarrow X$ such that $f \circ g$ is homotopic to $\text{id}_Y$ and $g \circ f$ is homotopic to $\text{id}_X$.}, then $X$ and $Y$ have isomorphic homology groups.
\end{remark}
\subsection{Persistent homology}
Computing the homology of finite simplicial complexes can be simplified to computing linear algebra. However, most empirical data sets do not exist in the form of simplicial complexes. Here we introduce the method in which we build simplicial complexes from directed networks.
Given a directed network $G$, one can induce a sequence of simplicial complexes $\mathcal{D}_{\delta, G}$ for a sequence of values of parameter $\delta \in \mathbb{R}_+$ such that $\mathcal{D}_{\delta=m, G} \subseteq \mathcal{D}_{\delta=n, G}$ for all $m \leq n$, where $\mathcal{D}_{\delta, G}$ is called the Dowker sink complex (Definition \ref{def: dowker}), introduced first in \cite{chowdhury_mémoli_2018}:
\begin{definition}
Given a network $G=(X, w_X)$ and $\delta \in \mathbb{R}$, define $R_{\delta, G} \subseteq X \times X$ as:
\[R_{\delta, G} := \{(x, x'): w_X(x, x') \leq \delta \}. \]
Using $R_{\delta, G}$, we build a simplicial complex $\mathcal{D}_{\delta, G}$ called the \emph{Dowker sink complex} as:
\[\mathcal{D}_{\delta, G}:= \{ \sigma \in Pow(X): \text{ there exists } p \in X \text{ s.t. } (x_i,p) \in R_{\delta, G} \text{ for each } x_i\in \sigma \}. \]
The node $p$ is called the \emph{sink} for the simplex $ \sigma$, and one can check that $\mathcal{D}_{\delta, G} \subseteq \mathcal{D}_{\delta', G}$ for all $\delta \leq \delta'$.
\label{def: dowker}
\end{definition}
\begin{remark}
Let $G=(X, w_X)$ be a dissimilarity network, then $R_{\delta=0, G}$ is precisely the set of $|X|$ vertices, i.e., 0-simplices.
\end{remark}
We note that the Dowker sink complex and the \emph{Vietoris-Rips complex} \cite{otter_porter_tillmann_grindrod_harrington_2017}, one of the most prevalent tools used in TDA, are analogous. While the Vietoris-Rips complex induces a filtration of simplicial complexes from a point cloud data set, the Dowker sink complex induces a filtration of simplicial complexes from a directed network.
\begin{definition}
A \emph{filtration of simplicial complexes} is a set of simplicial complexes $X_1, X_2, ..., X_n$ such that each simplicial complex is contained in its successive simplical complex:
\[ X=X_1 \subseteq X_2 \subseteq ... \subseteq X_n.\]
$X$ is called a \emph{filtered simplicial complex}.
\end{definition}
\begin{example}
Let $G=(X, w_X)$ be a directed network in a finite metric space, and let the simplicial complexes $\mathcal{D}_{\delta_1, G}$, $\mathcal{D}_{\delta_2, G}$, ..., $\mathcal{D}_{\delta_n, G}$ be the Dowker sink complexes at $\delta_1 \leq \delta_2 \leq ... \leq \delta_n$ respectively, then the sequence
\[ \mathcal{D}_{\delta_1, G} \subseteq \mathcal{D}_{\delta_2, G} \subseteq \mathcal{D}_{\delta_3, G} \subseteq ... \subseteq \mathcal{D}_{\delta_n, G}=\mathcal{D}_{G}\]
is a filtration of simplicial complexes, and $\mathcal{D}_{G}$ is a filtered simplicial complex.
\end{example}
One can compute the homology groups of the simplicial complex $\mathcal{D}_{\delta, G}$ for a given value of $\delta$.
As the value of $\delta$ increases, one can record the birth (the value of $\delta$ at which a homological feature appears) and death (the value of $\delta$ at which a homological feature disappears) values for all homological features (path components, 1-cycles, etc.) that appear in $\mathcal{D}_{G}$.
\begin{remark}
The definitions of birth, death values, and a persistence barcode of a homological feature in a filtered simplicial complex are presented in Section 4 of \cite{otter_porter_tillmann_grindrod_harrington_2017}. We will omit these definitions since they are not necessary for the work presented in this paper.
\end{remark}
Thus, given a filtered simplicial complex $X$, we can associate $X$ a union of \emph{persistence barcodes}, each representing the birth and death values of a homological feature of $X$. Persistence barcodes record changes in the $n$-dimensional homology group $H_n$ for some $n \in \mathbb{Z}_{\geq 0}$ when one simplicial complex $X_i$ is included in its successor $X_{i+1}$ in the filtration. In particular, each persistence barcode $[b,d)$ represents the lifetime of a generator of a homology group $H_n$, where $\delta=b$ is the birth value of the feature and $\delta=d$ is the death value of the feature in the filtered simplicial complex $X$. The value $d-b$ is called the \emph{persistence} of the feature.
\begin{definition}
A \emph{persistence diagram} $\textbf{D}$ is a union of a finite set of points above the diagonal $D = \{(x, y) \in \mathbb{R}^2 | x = y\}$ in ${\mathbb{R}_{\geq 0}}^2$ and the entire diagonal $D$.
\end{definition}
\begin{remark}
Each point $(x,y) \in \textbf{D}$ represents the birth and death $\delta$ values of a persistent barcode $[b,d)$ with $x=b$ and $y=d$.
\end{remark}
\begin{remark}
Roughly, a persistence diagram inherits the same data as the persistence barcodes. However, the persistence diagram $\textbf{D}$ contains the diagonal $D$, which will be useful for comparing the similarity of data sets. Specifically, a matching between two persistence diagrams map each point of a persistence diagram to another point of the other persistence diagram or to the diagonal, For more concrete details, see Section 5.1.
\end{remark}
Generally, given any dissimilarity network $G(X, w_X)$, we denote $\Dgm_{n}(G)$ as the $n$-dimensional persistence diagram for the filtered simplicial complex $\mathcal{D}_{G}$. Moreover, we denote \[\textbf{P}_n(G):= \{[b_1, d_1),[b_2, d_2),...,[b_k, d_k) \}\] to be the set of n-dimensional persistence barcodes for the filtered simplicial complex $\mathcal{D}_{G}$.
\section{Quasi-centrality measure}
In this section, we present our definition of \emph{quasi-centrality} for a directed dissimilarity network based on persistent homology and the Dowker sink complex, which is given in Definition \ref{def: quasi-central nonno}.
\begin{definition}
\label{def: function f gamma}
Let $G=(X, w_X)$ be a (dissimilarity) network and let $x \in X$, define ${ f(G, x)=(X\setminus\{x\}, w_X)} $ to be the sub-network induced by deleting $x$ and all edges incident to $x$ in $G$.
\end{definition}
\begin{remark}
We note that for a trade network $G=(X,w_X)$, where $w_X(x_1,x_2)$ represents the trade volume between individual countries or industries $x_1, x_2 \in X$, the induced network ${f(G,x)}$ models an embargo trade barrier\footnote{For our purposes, an embargo on $ x $ refers to the situation in which all other countries/nodes refuse to trade with $x$.} on $x$.
\end{remark}
\begin{definition}
Let $G=(X, w_X)$ be a (dissimilarity) network and $\gamma(G)=(X, m_X)$ be defined as in Definition \ref{def: gamma weight}. Given, $x \in X$, define $\mu(x)$ to be the minimum distance between $x$ and any other node $x' \in X$ with respect to the metric $m_X$.
\label{def: muu}
\end{definition}
\begin{definition}
Let $G=(X, w_X)$ be a (dissimilarity) network and let $x \in X$. We define the \emph{quasi-centrality} $C(x)$ as follows:
\[C(x)=\sum_{b \in \textbf{P}_0(f(\gamma(G),x))} l(b) -\sum_{b \in \textbf{P}_0(\gamma(G))} l(b) + \mu(x) \]
where $b$ represents a persistence interval in $\textbf{P}_0$ and $l(b)$ is the persistence of the interval, i.e., the length of the interval, and $\gamma(G)$ is given in Definition \ref{def: gamma weight} and $\mu(x)$ is given in Definition \ref{def: muu}.
\label{def: quasi-central nonno}
\end{definition}
\begin{remark}
We first give an intuitive explanation of what $C(x)$ measures before proving that $C(x)$ is nonnegative for all $x \in X$. We note that $\gamma(G)$ not only can be treated as a topological space where $\mathcal{D}_{\gamma(G)}$ encompasses the data of its homological features, but it also can be treated as a metric space since it comprises the data of the pairwise distances between nodes in the network.
Roughly speaking, we think of the sum $\sum_{b \in \textbf{P}_0(\gamma(G))} l(b)$ as a measure of how disconnected $\gamma(G)$ is, i.e., if $\sum$ is large, then $\gamma(G)$ is considered to be highly disconnected and vice versa. In particular, since the dimension of the 0-th homology group of a simplicial complex $X$ measures the number of path components (disconnected components) in $|X|$, we consider $\sum_{b \in \textbf{P}_0(\gamma(G))} l(b)$ to be the ``size" of the 0-th homology group of $\gamma(G)$,
and $\gamma(G)$ is considered to be highly disconnected if the sum $\sum$ is large and vice versa.
Similarly, the sum $\sum_{b \in \textbf{P}_0(f(\gamma(G),x))} l(b)$ as a measure of how disconnected $f(\gamma(G),x)$ is. We think of the difference as measuring of how much node $x$ contributes to the overall connectivity of the network.
Intuitively, a network should become `more disconnected' after deleting a node, which is made precise by the following theorem.
\label{def: rem}
\end{remark}
\begin{thm}
For a (dissimilarity) network $G=(X, w_X)$, $C(x)$ is nonnegative for all $x \in X$.
\end{thm}
\begin{proof}
Recall that each $[0,d_j) \in \textbf{P}_0(\gamma(G))$ represents the lifetime of a path component in $\gamma(G)$, hence by definition one connected component $\alpha$ joins another connected component $\beta$ at $\delta=d_j$.
Therefore, at least two $0-$simplices $[x_1] \in \alpha$ and $[x_2] \in \beta$ belong to a 1-simplex $\pm [x_1:x_2]$ at $\delta=d_j$.
By definition of the Dowker sink complex, there exists $p \in X$ such that $m(x_1, p) \leq d_j$ and $m(x_2, p) \leq d_j$.
If $x \in \{x_1,x_2,p\} $, then the 0-th persistence barcode in $\textbf{P}_0(f(\gamma(G), x)))$ representing the lifetime of path component $\alpha$ or $\beta$ will have a death resolution $\delta$ greater than or equal to $d_j$.
If $ {x \not \in \{x_1,x_2,p\}} $, then the 0-th persistence barcode in $\textbf{P}_0(f(\gamma(G), x)))$ representing the lifetime of path component $\alpha$ or $\beta$ is still equal to $d_j$.
Hence all persistence barcodes representing the lifetime of connected components have death values in $\textbf{P}_0(\gamma(G))$ greater than or equal to the death values of corresponding components in $\textbf{P}_0(f(\gamma(G), x)))$. Therefore
\[\sum_{b \in \textbf{P}_0(f(\gamma(G),x))} l(b)+\mu(x) \geq \sum_{b \in \textbf{P}_0(\gamma(G))} l(b),\]
and our proof is complete.
\end{proof}
\begin{remark}
As the value of $\delta$ increases, $C(x)$ measures how often node $x$ bridges between two disconnected components.
We think of the node $x$ as serving as a $\delta$-sink not only for a pair of nodes in a neighborhood of $x$, but rather the pair of connected components the nodes belong to.
In this way, the quasi-centrality $C(x)$ takes into account not only the local neighborhood of node $x$, but also propagating effects in said neighborhood.
\end{remark}
\begin{figure}
\centering
\hspace{-1cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\begin{tikzcd}
& x_1 \\
x_2 & x_3 \arrow[u, "2"] \arrow[ld, "1"] \arrow[rd, "4"] \arrow[l, "1"] \arrow[r, "3"] & x_4 \arrow[d, "6"] \\
x_5 & & x_6
\end{tikzcd}
\end{minipage}
\hspace{1cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\begin{tikzcd}
& x_1 \\
x_2 & x_3 \arrow[u, "2.70"] \arrow[ld, "3.40"] \arrow[rd, "2.01"] \arrow[l, "3.40"] \arrow[r, "2.30"] & x_4 \arrow[d, "1.00"] \\
x_5 & & x_6
\end{tikzcd}
\end{minipage}
\hfill
\caption{Left: $G=(X,w_X)$. Right: $\gamma(G)=(X,m_X)$. }
\label{fig: example 1}
\end{figure}
\begin{figure}
\centering
\hspace{-1cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\begin{tikzcd}
& x_1 \\
x_2 & & x_4 \arrow[d, shift left, "1.00"] \\
x_5 & & x_6
\end{tikzcd}
\end{minipage}
\hspace{1cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\begin{tikzcd}
& x_1 \\
x_2 & x_3 \arrow[u, "2.70"] \arrow[ld, "3.40"] \arrow[l, "3.40"] \arrow[r, "2.30"] & x_4 \\
x_5 &
\end{tikzcd}
\end{minipage}
\hfill
\caption{Left: $f(\gamma(G), x_3))$. Right: $f(\gamma(G), x_6))$. }
\label{fig: example 2}
\end{figure}
\begin{example}
\label{ex: example}
In this example, we compute the quasi-centrality measure on nodes in an example directed weighted network $G$ as shown in Figure \ref{fig: example 1} (left), and compare the result with existing centrality measures.
The illustrations of $G$ and $\gamma(G)$ are shown in Figure 1 (left and right respectively). In the illustration of $\gamma(G)$, we leave out the edge $e_{ij}$ when $w(x_i, x_j)=0$. However, in our calculation of $C(x)$, we let the $m(x_i,x_j)$ to be a very big number when $w(x_i, x_j)=0$ so that nodes $x_i,x_j$ in $m_X$ are very far apart that it is almost indistinguishable when no relation exists between the two nodes. In this example, we let $m(x_i,x_j) = 24.026$ to make computations more convenient.
Figure \ref{fig: example 2} illustrates the networks for $f(\gamma(G), x_3))$ and $f(\gamma(G), x_6))$, and Figure \ref{fig: example 4} illustrates the Dowker persistence barcodes for $f(\gamma(G), x_3)))$, $f(\gamma(G), x_6)))$, and $\gamma(G)$.
\begin{figure*}[ht!]
\subfloat[ $\Dgm_0(\gamma(G)))$ \label{fig:PKR}]{%
\includegraphics[ width=0.31\textwidth]{pers-5star.png}}
\hspace{\fill}
\subfloat[ $\Dgm_0(f(\gamma(G), x_3)))$ \label{fig:PKR}]{%
\includegraphics[ width=0.31\textwidth]{pers-5star-0.png}}
\hspace{\fill}
\subfloat[$\Dgm_0(f(\gamma(G), x_6)))$ \label{fig:tie5}]{%
\includegraphics[ width=0.31\textwidth]{pers-5star-5.png}}\\
\caption{}
\label{fig: example 4}
\end{figure*}
We illustrate the computation of $C(x_3)$ as an example. Since $\mu(x_3)=2.01$,
\begin{align*}
C(x_3) &= \sum_{j=1}^{n-1} d_{j}' -\sum_{j=1}^{n} d_j +2.01 \\
&=(1+24.026\times 4)-(1+2.01+2.70+3.40+3.40+24.026)+2.01 \\
&=65.978.
\end{align*}
Similarly, we have $C(x_1)=C(x_2)=C(x_4)=C(x_5)=0$ and $C(x_6)=0.29$. This shows that in $G$, node $x_3$ has the most propagating influence in the entire network, followed by node $x_6$.
\begin{figure}
\centering
\begin{tabular}{ | m{0.7cm} | m{1.8cm} | m{1.8cm} | m{1.8cm} | m{1.8cm} | m{1.8cm} | m{1.8cm} |}
\hline
Node & Quasi-centrality & Katz & Pagerank 1 & Pagerank 2 & HITS-hubs & HITS-authorities \\
\hline
$x_1$ & 0 & 0.38 & 0.15 & 0.50 & 0.00 & 0.11 \\
\hline
$x_2$ & 0 & 0.38 & 0.13 & 0.09 & 0.00 & 0.06 \\
\hline
$x_3$ & 65.978 & 0.23 & 0.12 & 0.09 & 0.47 & 0.00\\
\hline
$x_4$ & 0 & 0.38 & 0.16 & 0.09 & 0.53 & 0.17 \\
\hline
$x_5$ & 0 & 0.38 & 0.13 & 0.14 & 0.00 & 0.06\\
\hline
$x_6$ & 0.29 & 0.61 & 0.31 & 0.09 & 0.00 & 0.60\\
\hline
\end{tabular}
\caption{Centrality values for $G$ obtained from the quasi-centrality, Katz centrality, HITS-hubs and authorities centralities, and Pagerank centrality.}
\label{fig: table of quasi-cen ex}
\end{figure}
Figure \ref{fig: table of quasi-cen ex} illustrates the values obtained by other existing centrality measures. We have included two measures for the Pagerank centrality: Pagerank 1 outputs the centrality values for nodes in the original network $G$ and Pagerank 2 outputs the centrality measures for nodes in the network obtained by switching the direction of edges from $G$.
We included Pagerank 2 for the following reason: since all existing centrality measures are based on the term $\sum A_{ij}x_j$, i.e., the centrality value $x_i$ for node $i$ only sums up the contribution values $x_j$ of nodes $j$ that points toward node $i$, by computing Pagerank 2 and comparing to the results obtained by the quasi-centrality, we essentially eliminates the possibility that the differences in the centrality values is solely due to the fact of the direction of the edges.
We note that $x_3$ has low centrality values for all existing centrality measures, including both Pagerank centrality measures. On the contrary, we note that $x_3$ has the high quasi-centrality value, which is what we would expect since node $x_3$ holds the center position in the directed dissimilarity network $G$. Therefore in this case, the quasi-centrality better measures the propagating influences of nodes in this example.
We expect the reason for all existing centrality measures having resulted in a low centrality value for $x_3$ for two reasons: firstly, all existing centrality measures depend on the adjacency matrix $A_{ij}$ and contain the term $\sum A_{ij}x_j$, which sums up the contribution values of nodes $j$ that links with node $i$. However, it is possible that the nodes $j$ are not central in the network and hence does not contribute to the centrality value of node $i$, which can be shown in this example. However, it is still possible that node $i$ holds the center position in the network even when its neighbors are not important in the network (such as in this example), in which all existing centrality measure fail to accurately measure its importance. Secondly, the adjacency matrix term $\sum A_{ij}x_j$ only sums up the contribution values of nodes $j$ that has an ingoing edge to node $i$ while ignoring the nodes that have an outgoing edge from node $i$, meaning that the centrality values are only taking into account of propagating influences on one side while ignoring the other.
Generally, when measuring the propagating influence of a node, both incoming edges and outgoing edges should positively impact the centrality value. For example, when measuring the influence of perturbations originated from a country or industry in the empirical trade production network, if a particular country or industry is both an important exporter and importer, it will have a greater influence on the entire economy compared to a country or industry that is only an important importer. The quasi-centrality in this case would give a more thorough measure of a node's propagating influence as it takes into account of the nodes' impact in both directions, whereas other centrality measures might only focus on ingoing edges or outgoing edges.
\end{example}
\section{Quasi-centrality applications}
In this section, we apply the quasi-centrality measure to an empirical network describing trade among countries and industries--the Asia machinery production network in 2007 and 2011.
We show that the quasi-centrality measure gives a more refined evaluation of the propagating influences of nodes in the trade network than existing centrality measures.
The trade network exhibits many properties common to complex networks, including high clustering coefficient and degree-degree correlation between different vertices \cite{PhysRevE.68.015101}.
Moreover, research suggests that random microeconomic shocks have the ability to propagate through a network with high inter-connectivity and cause “cascade effects” in the entire network and economy \cite{econometrica_2012}--a prime example is the Asiatic crisis.
Therefore, it is important to take into account of the propagating effects of the trade network when evaluating node centrality. We later show that the quasi-centrality accurately predicts the influences of nodes in the trade network as it takes into account of such propagating effects.
We examine the machinery production network in Asia during 2007 and 2011 (network illustrations are shown in Figure \ref{fig: 2007 and 2011 Asia Network Visualization}), before and after the 2008-2009 financial crisis. Our network consists of both different intermediate products and different countries as nodes, as either counterpart plays a crucial role in supply and production chains\footnote{\label{footnote:supply}A supply/production chain is a system of intermediate transactions, suppliers, distributors, consisting of different countries and intermediate products.}, which are central for our understanding of the entire network.
\begin{figure*}[ht!]
\subfloat[ 2007 Asia machinery production network \label{fig:PKR}]{%
\includegraphics[ width=0.5\textwidth]{2007_asia_network_visualization.png}}
\hspace{\fill}
\subfloat[2011 Asia machinery production network \label{fig:tie5}]{%
\includegraphics[ width=0.5\textwidth]{2011_asia_network_visualization.png}} \\
\caption{}
\label{fig: 2007 and 2011 Asia Network Visualization}
\end{figure*}
Our data consists of $8 \times 4$ pairs of countries and industries, obtained from the OECD Inter-Country Input-Output (ICIO) Tables \cite{OECD2021}, which provides transport values between country-industry pairs. The eight countries and four industries are labeled in the following table:
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
Country Label & Country Name \\
\hline
KOR & North Korea \\
\hline
JPN & Japan \\
\hline
MYS & Malaysia \\
\hline
SGP & Singapore \\
\hline
TWN & Chinese Taipei \\
\hline
CN & China \\
\hline
VNM & Vietnam \\
\hline
THA & Thailand \\
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
Industry Label & Industry Name \\
\hline
C29 & Machinery and equipment \\
\hline
C30 & Computer, electronic and optical equipment \\
\hline
C31 & Electrical machinery and apparatus \\
\hline
C34 & Motor vehicles, trailers and semi-trailers \\
\hline
\end{tabular}
\end{center}
For brevity, we abbreviate the 32 nodes by [label of industry]-[label of the country].
For example, the machinery and equipment industry in Korea is referred to as KOR-C29.
\begin{remark}
In our data obtained from the OECD Inter-Country Input-Output (ICIO) Tables, imports are valued at basic prices of the country of origin, i.e., the domestic and international distribution included in goods imports in c.i.f. purchasers' prices are re-allocated to trade, transport and insurance sectors of foreign and domestic industries.
\end{remark}
\begin{remark}
We exclude taxes paid and subsidies received in foreign countries in our network.
\end{remark}
\begin{remark}
For the rest of Section 4, we will use ``node'' and ``country-industry pair'' interchangeably.
\end{remark}
\subsection{2007 Asia machinery production network}
We first examine nodes' propagating influences in the 2007 Asia machinery production network. We compute the quasi-centralities (Figure \ref{fig: 2007-asia-quasi-centrality}) for nodes in the network and compare with existing centrality measures (Figure \ref{fig: 2007-asia-other-centrality}) and later demonstrate that quasi-centrality successfully captures propagating effects and more accurately evaluates nodes' propagating influences in the network.
\begin{figure}
\centering
\includegraphics[width=13cm]{2007_Quasi-horizontal.png}
\caption{Quasi-centralities for the 2007 Asia machinery production network.}
\label{fig: 2007-asia-quasi-centrality}
\end{figure}
\begin{figure}
\centering
\hspace{-2cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2007_pagerank.png}
\end{minipage}
\hspace{3cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2007_katz_horizontal.png}
\end{minipage}
\vfill
\hspace{-2cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2007_HIT_hubs_horizontal.png}
\end{minipage}
\hspace{3cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2007_HITS_Authorities_horizontal.png}
\end{minipage}
\hfill
\caption{Existing centralities (Katz, Pagerank, HITS-hubs, HITS-authority) for the 2007 Asia machinery production network.}
\label{fig: 2007-asia-other-centrality}
\end{figure}
According to Figure \ref{fig: 2007-asia-quasi-centrality}, JPN-C34 (marked in orange) has the highest quasi-centrality value, implying that it has the highest propagating influence over the network, followed by CN-C30 (marked in pink), JPN-C29 (marked in grey), THA-C34 (marked in green), etc.
These values reflect that China and Japan were the two largest contributors to the machinery production industry in Asia in 2007 \cite{doi:10.1080/10168737.2016.1148398}.
We note that THA-C34 also has a relatively high quasi-centrality value. Although Thailand only accounted for 1.12 $\%$ of global exports of vehicle parts in 2007, while Japan, Korea, and China accounted for 10.2 $\%$, 4.05 $\%$, and 3.98 $\%$ (\cite{OEC.2021}) respectively, Thailand holds an important network position locally in Asia due to its high inter-correlation with China, the leading player in the network.
According to \cite{/content/publication/9789264237087-en}, between 2000 to 2010, Thailand's exports in motor vehicles parts, electronics and electrical appliances increased by $>40$ percent, and growth in trade with China (both imports and exports) also grew substantially.
By 2010, China became Thailand's lead export destination.
Since quasi-centrality is designed to capture how much a node contributes to the overall connectivity of the network, meaning that higher intensity of links between nodes will lead to a higher connectivity, we may also conclude the following: if a node $a$ is highly influential in the network, and if another node $b$ exhibits strong linkage with node $a$, then node $b$ also plays a relatively important role in the network because perturbations originated from $b$ will have a high influence on $a$, and will then be highly influential to the entire network. Since Thailand is highly interconnected with China, one of the main players in the network, we may infer that perturbations originating from Thailand have a high potential to propagate over the entire network.
\paragraph{Comparison with existing centrality measures.}
We observe that the relative importance of THA-C34 according to the quasi-centrality is not reflected in other existing centrality measures. We hypothesize that this is due the fact that the existing centrality measures are dependent on the adjacency matrix $A_{ij}$, implying that the centrality value for THA-C34 is based on its local neighborhood, while the quasi-centrality evaluates the THA-C34's contribution to the connectivity of the entire network, emphasizing its interconnectedness with CN-C30.
According to Figure \ref{fig: 2007-asia-other-centrality}, the Katz centrality returns similar centrality values for all nodes in the network, failing to capture subtleties in the influences of nodes.
Since the trade network exhibits high clustering properties, i.e., every country-industry pair is linked with almost every other country-industry pair, the Katz centrality is unable to capture the relative differences in the influence of nodes.
We may conclude that this is due to the fact that Katz centrality ignores the importance of the intensity of interaction between edge links, i.e. it does not take edge weights into account.
Moreover, both HITS-hubs and HITS-authorities centralities are not refined since their output values fail to evaluate the relative scale of the the influences of nodes in the network:
they are only able detect a single node that is highly influential, while ignoring the influences of other nodes.
This is not helpful for detecting potential souces of shocks in supply chains because we are only able to detect one single chain instead of many chains.
Of the existing centrality measures, the Pagerank centrality gives similar relative information of the influence of nodes compared to the gradation given by the quasi-centrality.
However, we note that the Pagerank centrality outputs a high centrality value for KOR-C34 (marked in yellow), while the quasi-centrality measure for KOR-C34 is zero.
According to \cite{doi:10.1080/10168737.2016.1148398}, even though Korea has high import value for machinery final products, its export value is relatively low--in particular, KOR-C34 has a many more incoming edges than outgoing edges.
Since the Pagerank centrality $x_i$ for a node $i$ is $\alpha \sum_{j} A_{ij}x_j+\beta$, which only sums up the contribution value of nodes $j$ that has an edge that points toward node $i$, KOR-C34 has a high Pagerank centrality value as it has a large number of incoming edges, i.e., imports.
However, a country-industry pair's propagating influence in the entire trade network should depend on both its export and import values, as a country-industry pair with both high export and import should have a larger propagating influence than a country-industry pair with only high import values.
Thus in this case, quasi-centrality offers a better assessment of a node's influence than Pagerank does.
While the Pagerank centrality only remembers one direction at a time, meaning that it captures some propagating influence but not all, the quasi-centrality captures both directions and gives a more comprehensive evaluation of the propagating influence of a node.
\subsection{2011 Asia machinery production network}
Now we proceed to analyze the propagating influences of nodes in the 2011 Asia machinery production network, following the 2007-2008 financial crises. The objective is to identify changes in power dynamics of nodes in the network compared to 2007.
\begin{figure}
\centering
\includegraphics[width=13cm]{2011_quasi-horizontal.png}
\caption{Quasi-centralities for the 2011 Asia machinery production network.}
\label{fig: 2011-asia-quasi-centrality}
\end{figure}
\begin{figure}
\centering
\hspace{-2cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2011_Pagerank_horizontal.png}
\end{minipage}
\hspace{3cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2011_Katz_Horizontal.png}
\end{minipage}
\vfill
\hspace{-2cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2011_HITS_Hubs_horizontal.png}
\end{minipage}
\hspace{3cm}
\begin{minipage}[b]{0.3\textwidth}
\centering
\includegraphics[width=7.3cm]{2011_HITS_authorities_horizontal.png}
\end{minipage}
\hfill
\caption{Existing centralities (Katz, Pagerank, HITS-hubs, HITS-authority) for the 2011 Asia machinery production network.}
\label{fig: 2011-asia-other-centrality}
\end{figure}
According to Figure \ref{fig: 2011-asia-quasi-centrality}, CN-C30 (marked in pink) has the highest quasi-centrality value, followed by JPN-C34 (marked in orange) and SGP-C30 (marked in red).
It is worth noting the change in the quasi-centrality values for JPN-C34, CN-C30, and SGP-C30 from 2007 to 2011. While the quasi-centrality value for CN-C30 and SGP-C30 have significantly increased, the value for JPN-C34 has significantly decreased.
According to \cite{doi:10.1080/10168737.2016.1148398}, China became a dominant player in the global machinery production network in terms of \emph{both} export value and the diversity of industry-destination pairs following the 2008-2009 financial crisis. Since the quasi-centrality measures how a node contributes to the overall connectivity of the network, accounting for \emph{all} factors of import, export values, and industry-destination pairs, China's growth of impact in the network is substantially illustrated by its increase in the quasi-centrality value.
According to \cite{doi:10.1080/10168737.2016.1148398}, Japan's machinery production growth, on the other hand, stagnated because of two main reasons after the financial crisis of 2008-2009. First, transportation links between Japan and multiple countries weakened, specifically with Korea.
Second, there was a noticeable decrease in the number of product–destination pairs in Japan’s exports of machinery. Both of these factors have led to Japan's stagnation in economic growth, and the stagnation is distinctively illustrated by the change of Japan's quasi-centrality value relative to others as computed in Figure \ref{fig: 2011-asia-quasi-centrality}.
In addition to the significant power dynamic changes of Japan and China in the network, we also note that the quasi-centrality value for SGP-C30 has significantly increased compared to 2007. According to \cite{hoon_2022}, the Singapore electrical machinery and apparatus industry became more \emph{central} in numerous supply chains (see footnote \ref{footnote:supply}) after the financial crisis, thereby playing an important role as an intermediary between an abundance of production, supply, distribution, and post-sales activities of goods and services.
\paragraph{Comparison with existing centrality measures.}
We note that China's dominance in the network and Japan's stagnation of economic growth are not reflected in the existing centrality measures. Figure \ref{fig: 2011-asia-other-centrality} shows that the HITS-hubs and authorities centrality measures are not refined in that they continue to only detect the node with the highest influence in the network (KOR-C34 and JPN-C34 respectively), and that the Katz centrality continues to output the same centrality measure for almost all the nodes.
Moreover, the Pagerank centrality continues to measure JPN-C34 (marked in yellow) as being the most influential node in the network. We hypothesize that this is due to the fact that since the Pagerank centrality only sums up the contribution values of nodes having an incoming to JPN-C34, the decrease in product–destination pairs for JPN-C34 of \emph{exports} would not affect its Pagerank centrality.
Furthermore, CN-C30 (marked in pink) also has a relatively low Pagerank centrality value even though it is the dominant player in the network in 2011. We hypothesize that this is again due to the fact that the Pagerank centrality only sums up the contribution values of nodes having an incoming edge to CN-C30, but CN-C30 exports far more than imports. Specifically, China's raw export value of machines globally was $ \$ 894 $ billion dollars in 2011, while its raw import value of machines globally was only $ \$ 328 $ billion dollars.
Moreover, the increase of Singapore electrical machinery industry's influence in the network is substantially reflected by its quasi-centrality value, while not reflected in the existing centrality measures, again showing how the quasi-centrality is effective in measuring a node's propagating influence in the network.
\subsection{Summary}
We computed the quasi-centralities of nodes in the 2007 and 2011 Asia machinery production network and compared the result with the existing centrality measures. Based on empirical trade statistics, we showed that quasi-centrality is optimal for detecting propagating influences of a node in the network as it takes into account both the import and export values of a country and industry pair, as well as the node's number of industry-destination pairs and its role in intermediary transactions.
Specifically, we showed that empirical facts support the information obtained by computing the quasi-centrality measure: a set of countries and industries such that perturbations originating from this set of nodes can propagate significantly through the network.
\paragraph{Potential applications}
Trade is often adversely affected by perturbations in a single country, including natural disasters such as a hurricane, earthquake or flood, and political turmoil or an armed conflict \cite{korniyenko_pinat_dew_2017}. Understanding which collections of nodes in the network might potentially cause significant propagating impacts can aid in identifying bottlenecks, thereby making the trade network more resilient.
We propose that using the quasi-centrality measure can ultimately serve to protect world economies.
\section{Clustering nodes by topological influences}
In section 5.1, we introduce a method (Definition \ref{def: hierarc}) that combines hierarchical clustering and persistent homology to construct a hierarchy of nodes based on their impact in the overall topology of a directed weighted network.
In sections 5.2 and 5.3, we apply our method to the 2007 and 2011 Asia machinery production network.
\subsection{Node hierarchy using TDA}
Before introducing our method in Definition \ref{def: hierarc}, we recall how hierarchical clustering works.
Given a set of objects and a specified weight function \cite{kulish_2002} that computes the similarity between two objects in the set, hierarchical clustering is a method for detecting community structures within this set of objects by arranging the objects into a hierarchy of groups according to the weight function.
For a given value of parameter $t$, hierarchical clustering groups objects that are ``at most $t$ far apart from each other'' according to the weight function.
\begin{definition}
Given a finite set $X$, a \emph{partition} $ P $ of $X$ is a collection $ P=\{B_1, ..., B_k\}$ of subsets of $X$ such that:
\begin{enumerate}[nolistsep]
\item For all $i \not = j$, $B_i \cap B_j = \emptyset $, and
\item The union $\cup_{i=1}^{k}B_i =X$.
\end{enumerate}
Each $ B_i $ is referred to as a \emph{block} of $P$. We denote the set of all partitions of $X$ by $ \text{Part}(X) $.
\end{definition}
Because the partition depends on t, we would like a family of t-partitions, hence a hierarchical dendrogram is defined to be a function that outputs a partition given an input value of t:
\begin{definition}
Let $X$ be a finite set. A \emph{dendrogram over $X$} is a weight function \\
$ {D_X: \mathbb{R}_{+} \rightarrow \text{Part}(X)} $ such that:
\begin{enumerate}[nolistsep]
\item For $t' \geq t$, every block of $D_X(t)$ is contained in a block of $D_X(t')$.
\item There exists $t_F\in \mathbb{R}_{+}$ such that for all $t \geq t_F$ , $D_X(t)=\{X\}$.
\item The weight function $D_X(0)$ consists of singletons $\{x\}$ for all $x \in X$ .
\item For all $t\in \mathbb{R}_{+}$, there exists $\epsilon > 0$ such that $D_X(t)=D_X(t')$ for all $t' \in [t, t+\epsilon ]$.
\end{enumerate}
\end{definition}
As the value of parameter $t$ increases, the partition $D_X(t)$ gets ``coarser'', i.e., objects that are within $t$ similar to each other belong to the same block in $D_X(t)$.
A dendrogram can be thought of as a (nested) family of partitions of $X$ into \emph{similarity classes}, with the variable $ t $ specifying the desired fineness and granularity of the partition $ D_X(t) $. Elements that belong to the same block in $D_X(t)$ for small values of $t \in \mathbb{R}_{+}$ are more similar according to the weight function $D_X$, and vice versa.
\begin{remark}
We will not be going into detail of the computer algorithm that we use to construct our hierarchical dendrogram since the exact computer algorithm will not be necessary for explaining our work. We will use \emph{single-linkage clustering}\footnote{\label{footnote:single}You can read more concrete details about the algorithm in Section II of \cite{8862232}.} as our clustering algorithm, which is also used in Definition \ref{def: hierarc}.
\end{remark}
Now we explain how we can extract a hierarchical dendrogram from a dissimilarity directed network $G=(X,w_X)$ based on nodes' impact in the topology of the network (Definition \ref{def: hierarc}).
We first introduce the set of objects included in our hierarchy:
\begin{definition}
Given a dissimarity network $G=(X,w_X)$, let $\gamma(G)$ be as defined in Definition \ref{def: gamma weight} and assume that $|X|=n$, the number of nodes in $G$. For each node $x \in X$, let $f(\gamma(G),x)$ be as defined in Definition \ref{def: function f gamma}. Define the set of objects denoted by $S_G$ in the hierarchical dendrogram associated to $G$ as:
\[S_G :=\left\{\Dgm(f(\gamma(G),x)) \mid x \in \gamma(G) \right\} \cup \{ \Dgm(\gamma(G))) \}.\]
\end{definition}
We are interested in using dendrograms to express the hierarchy of nodes in a given directed network based on nodes' topological influences in the network.
The main idea is that the bottleneck distance between the Dowker persistence diagram associated to $G$ and the diagram associated to the induced network $G_x=(X\setminus\{x\},w_X)$ is small if $ x $ does not play an important role in the topology of $G$, and vice versa. We first introduce the set of objects $S_G$ in the hierarchical dendrogram associated to $G$:
Recall that given a directed weighted network, the Dowker persistence diagram summarizes the homological features in the Dowker complex associated to the network.
We may then conclude that given two directed weighted networks, if the two Dowker persistence diagrams associated to the networks are similar, then the networks have similar topology.
We measure the similarity between two persistence diagrams using the bottleneck distance (Definition \ref{defn:bottleneck}), which describes the cost of the optimal \emph{matching}\footnote{A matching $\eta$ between persistence diagrams $D_1$ and $D_2$ pairs each point in $D_1$ with a point in $D_2$ or a point on the diagonal line, and pairs each point in $D_2$ with a point in $D_1$ or a point on the diagonal.} between points of the two diagrams \cite{agami2020comparison}, introduced below:
\begin{definition}\label{defn:bottleneck}
Given two persistence diagrams $D_1$ and $D_2$, the \emph{bottleneck distance} of $D_1$ and $D_2$ denoted by $d_{B_\infty} (D_1,D_2)$ is defined as
\[d_{B_\infty} (D_1,D_2)=\inf_{\eta: D_1 \rightarrow D_2} \sup_{x\in D_1} ||x-\eta(x) ||_\infty \]
where $\eta$ ranges over all embeddings $\eta: D_1 \hookrightarrow D_2 \cup \Delta $, where $ \Delta $ denotes the diagonal and $||(x,y)||_\infty$ is the usual $L_\infty$ norm.
\end{definition}
If the bottleneck distance between two Dowker persistence diagrams is close to zero, i.e., if $d_{B_\infty} (\Dgm_1,\Dgm_2) \approx 0$, then we may conclude that two corresponding directed weighted networks have similar topological structure, and vice versa.
Using $S_G$, we may obtain the hierarchical dendrogram associated to the dissimilarity directed network $G$ as follows:
\begin{definition}
Given a dissimilarity network $G=(X,w_X)$ with the bottleneck distance, the \emph{hierarchical dendrogram $\mathcal{H}_G$ associated to $G$} is the function
\begin{align*}
\mathcal{H}_G:& \mathbb{R}_+ \rightarrow \text{Part}(S_G)\\
&t \mapsto \text{partition of } S_G \hspace{1mm} (S_G, \text{bottleneck distance}).
\end{align*}
And we implement the single-linkage clustering (see footnote \ref{footnote:single}) algorithm to extract the hierarchical dendrogram from the function $\mathcal{H}_G$.
\label{def: hierarc}
\end{definition}
For a given object $ x \in X $ and parameter value $t=t'$, we will denote the block containing $ x $ in $\mathcal{H}_G(t')$ by $\mathcal{B}_{\mathcal{H}_G}(x,t')$.
\subsection{Hierarchical clustering of the 2007 Asia machinery production network}
In this section, we examine the 2007 Asia machinery production network by analyzing the hierarchical dendrogram associated to the network obtained by the method we introduced in the previous section (Definition \ref{def: hierarc}).
Figure \ref{fig: 2007 asia gamma dendrogram} shows the hierarchical dendrogram associated to the 2007 Asia machinery production network, where the $y$-axis shows the 33 \emph{nodes}\footnote{For brevity, we will refer to the objects in the hierarchy to be the nodes in the dendrogram.} in the dendrogram, corresponding to the elements in $S_G$, where each of $\Dgm(f(\gamma(G),x))$ is labeled by the node $x$ deleted, and $\Dgm(\gamma(G))$ is labeled by STANDARD, and the values along the $x$-axis corresponds to the values of the parameter $t$ in the weight function $\mathcal{H}_G$.
The vertical line at $t=t'$ corresponds to the blocks of the partition $\mathcal{H}_G(t')$.
We consider the smallest value of $t$ for which $ \Dgm(f(\gamma(G), x)) $ belongs to the cluster containing STANDARD to be its topological impact on the network.
In other words, we expect nodes in the dendrogram belonging to the block containing the STANDARD node at smaller values of $t$ in the partition $\mathcal{H}_G(t)$ to have a smaller impact in the topology of the network, i.e., if given two nodes $a,b$ and $t < t'$, and that $a \in \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t)$ and $b \in \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t')$ then $b$ has a greater topological influence in the network topology.
According to Figure \ref{fig: 2007 asia gamma dendrogram}, we see that $\text{THA-C30}, \text{VNM-C34} \in \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t=0)$, implying that the country-industry pairs THA-C30 and VNM-C34 are insignificant in the topological structure of the network. We consider these nodes as ``peripheral,'' i.e. deleting them doesn't really change the network topology.
\begin{figure}
\centering
\includegraphics[width=14cm]{2007_asia_gamma_final.png}
\caption{Hierarchichal dendrogram associated to the 2007 Asia machinery production network.}
\label{fig: 2007 asia gamma dendrogram}
\end{figure}
On the contrary, when $t < 0.48$, $\text{JPN-C34} \not \in \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t)$ while \\
$ {\mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},0.42) = \{S_G \setminus \{\text{JPN-C34}\} \} } $, thus we conclude that deleting JPN-C34 has the most significant impact in the topological structure of the network.
Indeed, we see that the 1-dimensional barcodes of $\Dgm(f(\gamma(G),\text{JPN-C34}))$ (Figure \ref{fig: 2007 dendrogram 3}(b)) differs significantly from the 1-dimensional barcodes of $\Dgm(\gamma(G))$ (Figure \ref{fig: 2007 dendrogram 3}(a)): while all 1-dimensional barcodes in $\Dgm(\gamma(G))$ have death parameters $\delta \leq 4.38$, there exists a 1-dimensional barcode in $\Dgm(f(\gamma(G),\text{JPN-C34}))$ with death parameter $\delta = 5.64 $, indicating that JPN-C34 has a significant impact in the one-dimensional homological features of $\gamma(G)$, reflecting its importance in the network topology.
\begin{figure*}[ht!]
\subfloat[ $\Dgm(\gamma(G))$. \label{fig:PKR}]{%
\includegraphics[ width=0.5\textwidth]{pers-STANDARD.png}}
\hspace{\fill}
\subfloat[ $\Dgm(f(\gamma(G),\text{JPN-C34}))$. \label{fig:PKR}]{%
\includegraphics[ width=0.5\textwidth]{pers-JPN-C34.png}}
\hspace{\fill}
\subfloat[$\Dgm(f(\gamma(G),\text{VNM-C31}))$. \label{fig:tie5}]{%
\includegraphics[ width=0.5\textwidth]{pers-VNM-C31.png}}
\hspace{\fill}
\subfloat[$\Dgm(f(\gamma(G),\text{CN-C31}))$. \label{fig:tie5}]{%
\includegraphics[ width=0.5\textwidth]{pers-CN-C31.png}} \\
\caption{Dowker persistence diagrams for sub-networks obtained from the 2007 Asia machinery production network.}
\label{fig: 2007 dendrogram 3}
\end{figure*}
We note that from the result illustrated in Section 4.1, JPN-C34 also has the highest quasi-centrality value in the 2007 Asia machinery production network. Since the quasi-centrality measures the propagating influence of JPN-C34 and the network hierarchy reflects JPN-C34's impact in the overall topological structure of the network, we conclude that perturbations originating from JPN-C34 not only effectively propagate throughout the entire network, but also changes the network topology significantly.
Similarly, when $t < 0.33$, $\text{CN-C31} \not \in \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t)$, implying that CN-C31 also has a significant impact in the network topology. Indeed, the 1-dimensional barcodes of $\Dgm(f(\gamma(G),\text{CN-C31}))$ (Figure \ref{fig: 2007 dendrogram 3}(d)) also differ significantly from the 1-dimensional barcodes of $\Dgm(\gamma(G))$ (Figure \ref{fig: 2007 dendrogram 3}(a)).
Hence similar to JPN-C34, we conclude that perturbations originating from CN-C31 not only effectively propagate throughout the entire network, but also changes the network topology significantly.
On the contrary, although VNM-C31 has zero quasi-centrality value from the result in Section 4.1, when $t < 0.43$, $\text{VNM-C31} \not \in \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t)$, implying that VNM-C31 has a significant impact in the network topology. Indeed, the 1-dimensional barcodes of $\Dgm(f(\gamma(G),\text{VNM-C31}))$ (Figure \ref{fig: 2007 dendrogram 3}(c)) exhibit considerable differences compared to the one-dimensional barcodes of $\Dgm(\gamma(G))$ (Figure \ref{fig: 2007 dendrogram 3}(a)). However, the 0-dimensional persistence diagrams do not reflect significant differences. Since the quasi-centrality is measured based on how much VNM-C31 contributes to the overall connectivity of the network and is measured in terms of connected components (0-dimensional homological features), we conclude that although perturbations originating from VNM-C31 cannot propagate effectively throughout the entire network, it can potentially cause changes in the network topology.
\begin{figure*}[ht!]
\subfloat[ $\Dgm(f(\gamma(G),\text{THA-C31}))$. \label{fig:PKR}]{%
\includegraphics[ width=0.5\textwidth]{pers-THA-C31.png}}
\hspace{\fill}
\subfloat[$\Dgm(f(\gamma(G),\text{MYS-C29}))$. \label{fig:tie5}]{%
\includegraphics[ width=0.5\textwidth]{pers-MYS-C29.png}}
\hspace{\fill}
\subfloat[$\Dgm(f(\gamma(G),\text{VNM-C30}))$. \label{fig:tie5}]{%
\includegraphics[ width=0.5\textwidth]{pers-VNM-C30.png}}
\hspace{\fill}
\subfloat[$\Dgm(f(\gamma(G),\text{VNM-C29}))$. \label{fig:tie5}]{%
\includegraphics[ width=0.5\textwidth]{pers-VNM-C29.png}} \\
\caption{Dowker persistence diagrams for sub-networks obtained from the 2007 Asia machinery production network.}
\label{fig: 2007 dendrogram 4}
\end{figure*}
In fact, the hierarchical dendrogram produced using the method in Definition \ref{def: hierarc} not only illustrates the hierarchy of the nodes based on their topological impact in the network, but also allows us to identify which collections of nodes have \emph{similar} impacts in the network. In particular, the nodes that first belong to $\mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t)$ at approximately the same value of $t$ are considered to have similar topological influence in the network.
For example, it is clear from Figure \ref{fig: 2007 asia gamma dendrogram} that the nodes THA-C31, MYS-C29, VNM-C30, VNM-C29 all first join $\mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t)$ at \emph{close} values of $t$ ($0.26 <t< 0.28$), implying that this set of nodes have similar topological influence in the network.
Indeed, a close resemblance between the Dowker persistence diagrams corresponding to this set of nodes can be observed in Figure \ref{fig: 2007 dendrogram 4}.
\subsection{Hierarchical clustering for the 2011 Asia machinery production network}
\begin{figure}
\centering
\includegraphics[width=14cm]{2011_asia_dendrogram_final.png}
\caption{Hierarchichal dendrogram associated to the 2011 Asia machinery production network.}
\label{fig: 2011 asia dendrogram}
\end{figure}
Now we analyze the hierarchical dendrogram associated to the 2011 Asia machinery production network obtained from Definition \ref{def: hierarc}.
We hope that comparing the 2011 network hierarchical dendrogram (Figure \ref{fig: 2011 asia dendrogram}) with the 2007 network hierarchical dendrogram (Figure \ref{fig: 2007 asia gamma dendrogram}) will afford insights into how the country-industry pairs' influence in the topological structure of the network changed in the wake of the financial crisis.
In contrast to the 2007 hierarchical dendrogram where $\text{JPN-C34} \not \in \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},t)$ when $t<0.48$ while $ \mathcal{B}_{\mathcal{H}_G}(\text{STANDARD},0.42) = \{S_G \setminus \{\text{JPN-C34}\} \} $, the 2011 hierarchical dendrogram (Figure \ref{fig: 2011 asia dendrogram}) reflects that the absolute power of the node JPN-C34 in the network has significantly weakened; it now shares the influential role with another node: VNM-C31. The position change of JPN-C34 in the hierarchy further reflects that over the course of 2007 to 2011, the Japanese industry has stagnated \cite{doi:10.1080/10168737.2016.1148398} both in terms of its topological impacts in the structure of the network (reflected by the hierarchy) and its propagating influence through the entire network (reflected by the quasi-centrality).
The Dowker persistence diagram of the original network $\gamma(G)$ is illustrated in Figure \ref{fig: 2011 dendrogram 2}(a) and the Dowker persistence diagrams obtained by deleting the node JPN-C34 (VNM-C31, resp.) from $\gamma(G)$ are illustrated in Figure \ref{fig: 2011 dendrogram 2}(b) ((c), resp.) respectively. We see that both JPN-C34 and VNM-C31 had cause big changes in the network topology in terms of the 1-cycles. Indeed, because JPN-C34 and VNM-C31 are "well-connected" to other nodes, deleting these nodes have a high potential disrupt the 1-dimensional homological features.
\begin{figure*}[ht!]
\subfloat[$\Dgm(\gamma(G))$. \label{fig:PKR}]{
\includegraphics[ width=0.5 \textwidth]{pers-STANDARD-2011.png}}
\hspace{\fill}
\subfloat[ $\Dgm(f(\gamma(G),\text{JPN-C34}))$. \label{fig:PKR}]{
\includegraphics[ width=0.5\textwidth]{pers-JPN-C34-2011.png}}
\hspace{\fill}
\subfloat[$\Dgm(f(\gamma(G),\text{VNM-C31}))$. \label{fig:tie5}]{
\includegraphics[ width=0.5\textwidth]{pers-VNM-C31-2011.png}}
\hspace{\fill}
\subfloat[$\Dgm(f(\gamma(G),\text{CN-C30}))$. \label{fig:tie5}]{
\includegraphics[ width=0.5\textwidth]{pers-CN-C30-2011.png}} \\
\caption{Dowker persistence diagrams for sub-networks obtained from the 2011 Asia machinery production network.}
\label{fig: 2011 dendrogram 2}
\end{figure*}
Although CN-C30 had the largest quasi-centrality value as computed in Section 4.2, and that the Dowker persistence diagram of $f(\gamma(G),\text{CN-C30})$ illustrated in \ref{fig: 2011 dendrogram 2}(c) also reflects significant changes in the 0-dimensional barcodes compared to the Dowker persistence diagram of $\gamma(G)$, its impact on the 1-dimensional homological features is notably less than JPN-C34 and VNM-C31.
We note that the hierarchical clustering method incorporates higher dimensional topology in the network while the quasi-centrality gives a direct measure of the extent of propagating influence of a node in terms of changes in the 0-dimensional homological features.
\subsection{Summary}
We used the bottleneck distance between persistence diagrams obtained from the Dowker sink complex, to cluster nodes by their impact on the topology of the network, with a high emphasis on higher order homological features.
We showed that our hierarchical clustering dendrogram reflects both the extent to which each node changes the topology of the network and which collections of persistence diagrams have similar topological effects.
We note the key distinction between computing quasi-centrality on nodes and hierarchical clustering: while the quasi-centrality gives a direct measure of the node's propagating influence by recording how frequently the node serves as a bridge between path components, i.e., 0-dimensional homological features, the hierarchical clustering method directly compares the Dowker persistence diagrams, which includes higher dimensional homological features.
While the deletion of a node can potentially have a smaller impact on the 0-dimensional persistence barcodes, it might disrupt 1-cycles or higher dimensional cycles, illustrating its impact in the overall network topology.
\section{Discussion}
\paragraph{Summary}In this paper, we apply persistent homology to develop a network centrality measure \emph{quasi-centrality} that captures propagating influences in directed networks.
We computed the quasi-centrality on the Asia machinery production network and demonstrated that our centrality measure accurately reflects the propagating influences of perturbations originating from individual nodes.
Furthermore, we introduced a method that allows us to express the hierarchy of nodes based on their impact in the topology of a directed network. By incorporating hierarchical clustering and the bottleneck distance, we are able to not only detect which nodes play a significant role in the topology of the network, but also determine which nodes change the topology of the network in a similar fashion.
\paragraph{Future directions} We expect that many other properties in networks (directed or undirected) admit characterizations using tools in TDA: one direction for future research is to define other measures in network analysis using TDA, such as density, robustness, efficiency, connectivity, etc. We expect that using TDA to investigate network properties allows us to bring a different perspective to examine networks.
Another open question is how to relate higher dimensional homological features in the simplicial complexes that arises from networks to real-world phenomena, such as trade flow, embargo, value and supply chains, patterns of trade, biological cycles, etc.
Another direction for future research is to apply the quasi-centrality measure on other directed networks. We expect that the quasi-centrality can not only be useful in assessing the influence of individual components within industries or trade networks, but also in other highly clustered complex directed networks, such as biological networks, air flight networks, etc.
We believe that many network structures such as trade flows, supply chains, and biological cycles can be better understood from a topological point of view by employing tools in TDA. Just as TDA has been successfully applied to detect intrinsic shapes in complex data sets, we believe that TDA can also find interesting structures in network topology and
detect structural properties in networks, allowing us to better understand these networks, and gain more insights into the complex systems they model.
\paragraph{Limitations} We note that although the quasi-centrality measure and the hierarchical clustering function we defined in this paper can be useful for detecting nodes' propagating influences and topological impacts in a given directed network, computations and dendrograms generated are highly computationally costly as the number of nodes in the network exceeds 70-80 nodes.
\section{Acknowledgements}
I would like to sincerely thank my mentor Lucy Yang of Harvard University for providing thorough guidance and invaluable advice. I would like to thank the MIT-PRIMES program for providing me this invaluable opportunity and resources to work on this project. I would like to thank Professor Memoli of the Ohio State University and Dr. Chowdury of Stanford University for their generous help and discussions. I would also like to thank Dr. Tanya Khovanova and Dr. Kent Vashaw of MIT for carefully proofreading the paper and offering helpful suggestions.
\printbibliography
\end{document}
|
\section{Introduction}
Technological advancement has made high-dimensional data readily available. For example, in bioinformatics, the researchers seek to understand the gene expression level with microarray or next-generation sequencing techniques where each point consists of over 50,000 measurements \cite{Pease5022,shalon1996dna,metzker2010sequencing,reuter2015high}.
The abundance of features demands the development of feature selection algorithms to improve a Machine Learning task, e.g., classification. Another important aspect of feature selection is knowledge discovery from data. Which biomarkers are important to characterize a biological process, e.g., the immune response to infection by respiratory viruses such as influenza \cite{o2013iterative}? Additional benefits of feature selection include improved visualization and understanding of data, reducing storage requirements, and faster algorithm training times.
Feature selection can be accomplished in various ways that can be broadly categorized into the filter, wrapper, and embedded methods. In a filter method, each variable is ordered based on a score. After that, a threshold is used to select the relevant features \cite{lazar2012survey}. Variables are usually ranked using correlation \cite{guyon2003introduction,yu2003feature}, and mutual information \cite{vergara2014review,fleuret2004fast}. In contrast, a wrapper method uses a model and determines the importance of a feature or a group of features by the generalization performance of the predetermined model \cite{el2016review, hsu2002annigma}. Since evaluating every possible combination of features becomes an NP-hard problem, heuristics are used to find a subset of features. Wrapper methods are computationally intensive for larger data sets, in which case search techniques like Genetic Algorithm (GA) \cite{goldberg1988genetic} or Particle Swarm Optimization (PSO) \cite{kennedy1995particle} are used. In embedded methods, feature selection criteria are incorporated within the model, i.e., the variables are picked during the training process \cite{lal2006embedded}. Iterative Feature Removal (IFR) uses the absolute weight of a Sparse SVM model as a criterion to extract features from the high dimensional biological data set \cite{o2013iterative}.
Mathematically feature selection problem can be posed as an optimization problem on $\ell_0$-norm, i.e., how many predictors are required for a machine learning task. As the minimization of $\ell_0$ is intractable (non-convex and non-differentiable), $\ell_1$-norm is used instead, which is a convex proxy of $\ell_0$ \cite{tibshirani1996regression}. Note $\ell_1$ is not differentiable at 0, but the problem can be tackled by leveraging on sub-gradient methods \cite{boyd2003subgradient,duchi2011adaptive}. Since the introduction of these seminal papers \cite{candes2006stable,candes2005decoding}, the use of 1-norm has surged significantly in research work. Despite some difficulties \cite{zou2006adaptive,zou2005regularization}, the $\ell_1$ has been used for the feature selection task in linear \cite{fonti2017feature,muthukrishnan2016lasso,kim2004gradient,o2013iterative,chepushtanova2014band} as well as in nonlinear regime \cite{li2016deep,scardapane2017group,li2020efficacy}.
This paper proposes a new embedded variable selection approach called Sparse Centroid-Encoder (SCE) to extract features when class labels are available. Our method extends the Centroid-Encoder model \cite{GHOSH201826,GhKi2020}, where we applied a $l_1$ penalty to a sparsity promoting layer between the input and the first hidden layer. We evaluate this Sparse Centroid-Encoder on diverse data sets and show that the selected features produce better generalization than other state-of-the-art techniques. Our results showed that SCE picked fewer features to obtain high classification accuracy. As a feature selection tool, SCE uses a single model for the multi-class problem without the need to create multiple one-against-one binary models typical of linear methods, e.g., Lasso \cite{tibshirani1996regression}, or Sparse SVM \cite{chepushtanova2014band}. Although SCE can be used both in binary and multi-class problems, we focused on the multi-class feature selection problem in this paper. The work of \cite{li2016deep} also uses a similar sparse layer between the input and the first hidden with an Elastic net penalty while minimizing the classification error with a softmax layer.
The authors used Theano's symbolic differentiation \cite{bergstra2010theano} to impose sparsity.
In contrast, our approach minimizes the centroid-encoder loss with an explicit differentiation of the $l_1$ function using the sub-gradient. Unlike DFS, our model can capture the intra-class variability by using multiple centroids per class. This property is beneficial for multi-modal data sets.
The article is organized as follows:
In Section \ref{SCE} we present the Sparse Centroid-Encoder algorithm.
In Section \ref{experiments} we apply SCE to a range of bench-marking
data sets taken from the literature.
In Section \ref{lit}, we review related work, for both linear and
non-linear feature selection.
In Section \ref{dis_cons}, we present our discussion and conclusion.
\section{Sparse Centroid-Encoder }
\label{SCE}
Centroid-encoder (CE) neural networks are the starting point of our approach~\cite{GhKi2020,GHOSH201826,aminian2021early}. We present a brief overview of CEs and demonstrate
how they can be extended to perform non-linear feature selection.
\subsection{Centroid-encoder}
The CE neural network is a variation of an autoencoder and can be used for both visualization and classification tasks. Consider a data set with $N$ samples and $M$ classes. The classes denoted $C_j, j = 1, \dots, M$ where the indices of the data associated with class $C_j$ are denoted $I_j$. We define centroid of each class as
$c_j=\frac{1}{|C_j|}\sum_{i \in I_j} x^i$ where $|C_j|$ is the cardinality of class $C_j$. Unlike autoencoder, which maps each point $x^i$ to itself, the CE maps each point $x^i$ to its class centroid $c_j$
by minimizing the following cost function over the parameter set $\theta$:
\begin{equation}
\begin{aligned}
\mathcal{L}_{ce}(\theta)=\frac{1}{2N}\sum^M_{j=1} \sum_{i \in I_j}\|c_j-f(x^i; \theta))\|^2_2
\label{equation:CECostFunction}
\end{aligned}
\end{equation}
The mapping $f$ is composed of a dimension reducing mapping
$g$ (encoder) followed by a dimension increasing reconstruction
mapping $h$ (decoder). The output of the encoder is used as a supervised visualization tool \cite{GhKi2020,GHOSH201826}, and attaching another layer to map to the one-hot encoded labels performs robust classification \cite{aminian2021early}.
\subsection{Sparse Centroid-encoder for feature selection}
The Sparse Centroid-encoder (SCE) is a modification to the centroid-encoder architecture as shown in Figure \ref{fig:SCE_arch}. Unlike centroid-encoder, we haven't used a bottleneck architecture as visualization is not our aim here. The input layer is connected to the first hidden layer via the sparsity promoting layer (SPL). Each node of the input layer has a weighted one-to-one connection to each node of the SPL.
The number of nodes in these two layer are the same. The nodes in SPL don't have any bias or non-linearity. The SPL is fully connected to the first hidden layer, therefore the weighted input from the SPL will be passed to the hidden layer in the same way that of a standard feed forward network. During training, a $l_{1}$ penalty will be applied to the weights connecting the input layer and SPL layer. The sparsity promoting $l_{1}$ penalty will drive most of the weights to near zero and the corresponding input nodes/features can be discarded. Therefore, the purpose of the SPL is to select important features from the original input. Note we only apply the $l_{1}$ penalty to the parameters of the SPL.
\begin{figure}[t!]
\centering
\includegraphics[width=10.0cm,height=16.cm]{CE_SCE_Diagram.png}
\caption{The architecture of Centroid-encoder and Sparse Centroid-encoder. Notice the Centroid-encoder uses a bottleneck architecture which is helpful for visualization. In contrast, the Sparse Centroid-encoder doesn't use any bottleneck architecture; instead, it employs a sparse layer between the input and the first hidden layer to promote feature sparsity.}
\label{fig:SCE_arch}
\end{figure}
Denote $\theta_{spl}$ to be the parameters (weights) of the SPL and $\theta$ to be the parameters of the rest of the network. The cost function of sparse centroid-encoder is given by
\begin{equation}
\begin{aligned}
\mathcal{L}_{sce}(\theta)=\frac{1}{2N}\sum^M_{j=1} \sum_{i \in I_j}\|c_j-f(x^i; \theta))\|^2_2 + \lambda \|\theta_{spl}\|_{1}
\label{equation:SCECostFunction}
\end{aligned}
\end{equation}
where $\lambda$ is the hyper-parameter which controls the sparsity. A larger value of $\lambda$ will promote higher sparsity resulting more near-zero weights in SPL. In other words, $\lambda$ is a knob that controls the number of features selected from
the input data.
Like centroid-encoder, we trained sparse centroid-encoder using error backpropagation, which requires the gradient of the cost function of Equation \ref{equation:SCECostFunction}. As $\ell_1$ function is not differentiable at $0$, we implement this term using the sub-gradient\cite{boyd2003subgradient}. We trained SCE using Scaled Conjugate Gradient Descent \cite{moller1993scaled} on the full training set. Like any neural network-based model, the hyperparameters of SCE need to be tuned for optimum performance. Table \ref{table:SCEHyperparameters} contains the list with the range of values we used in this paper. We used validation set to choose the optimal value. For a small sample size data set ( high-dimensional biological data), we ran a five-fold cross validation on the training set to pick the optimum value. The Python implementation of Sparse Centroid-Encoder will be provided in Supplementary Material.
\begin{table}[!ht]
\vspace{1mm}
\centering
\begin{tabular} {|c|c|}
\hline
Hyper parameter & Range of Values \\
\hline
\#SCG Iteration & \{25,50,75,100\} \\
\hline
\# Hidden Layers (L) & \{1, 2\} \\
\hline
\# Hidden Nodes (H) & \{50,100,200,250,500\} \\
\hline
Activation Function & Hyperbolic tangent (tanh) \\
\hline
$\lambda$ & \{0.01, 0.001, 0.0001, 0.0002, 0.0004, 0.0006, 0.0008\} \\
\hline
\# Center/Class & \{1, 2, 3, 4, 5\} \\
\hline
\end{tabular}
\caption{Hyperparameters for Sparse Centroid-Encoder.}
\label{table:SCEHyperparameters}
\end{table}
\subsubsection{Feature Cut-off}
\label{feature_cut-off}
The 1-norm of the sparse layer (SPL) drives a lot of weight to near zero. Often hard thresholding or a ratio of two consecutive weights is used to pick the nonzero weight \cite{o2013iterative}. We take a different approach. After training SCE, we arrange the absolute value of the weights of the SPL in descending order. And then find the elbow of the curve. We measure the distance of each point of the curve to the straight line formed by joining the first and last points of the curve. The point with the largest distance is the position (P) of the elbow. We pick all the features whose absolute weight is greater than that of P; see panel (c), (d) of Figure\ref{fig:SCE_analysis} and panel (b), (c) of Figure \ref{fig:Feature_Stability}.
\subsection{Empirical Analysis of SCE}
\label{sparsity_analysis}
\begin{figure}[b!]
\centering
\includegraphics[width=12.5cm, height=10.5cm]{SCE_Sparsity_Analysis.png}
\label{fig:GSE73072_Sparsity}
\caption{Analysis of Sparse Centroid-encoder. (a) Change of the two costs over $\lambda$. (b) Change of validation accuracy over $\lambda$. (c) Sparsity plot of the weight of $W_{SPL}$ for $\lambda = 0.001$. (d) Same as (c) but $\lambda = 0.1$.}
\label{fig:SCE_analysis}
\end{figure}
In this section we present an empirical analysis of our model. The results of feature selection for the digits 5 and 6 from the MNIST set are displayed in Figure \ref{fig:SCE_analysis}. In panel (a), we compare the two terms
that contribute to Equation \ref{equation:SCECostFunction}, i.e., the
centroid-encoder and $l_{1}$ costs,
weighted with different values of $\lambda$. As expected, we observe that the CE cost monotonically decreases with
$\lambda$, while the $l_{1}$ cost increases as
$\lambda$ decreases.
For larger values of $\lambda$, the model focuses more on minimizing the $l_{1}$-norm of the sparse layer, which results in smaller values. In contrast, the model pays more attention to minimizing the CE cost for small $\lambda$s; hence we notice smaller CE cost and higher $l_{1}$ cost.
Panel (b) of Figure \ref{fig:SCE_analysis} shows the accuracy on a validation set as a function nine different values of $\lambda$; the validation accuracy reached its peak for $\lambda = 0.001$.
In panels (c) and (d), we plotted the magnitude of the feature weights of the sparse layer in descending order.
The sharp decrease in the magnitude of the weights
demonstrates the promotion of sparsity by SCE. The model
effectively ignores features by setting their weight to approximately zero. Notice the model produced a sparser solution for $\lambda = 0.1$, selecting only 32 features compared to 122 chosen variables for $\lambda = 0.001$. Figure \ref {fig:SCE_Sparsity} shows the position of the selected features, i.e., pixels, on the digits 5 and 6. The intensity of the color represents the feature's importance. Dark blue signifies a higher absolute value of the weight, whereas light blue means a smaller absolute weight.
\begin{figure}[h!]
\centering
\includegraphics[width=13.25cm,height=12.5cm]{SCE_Sparsity_Colors.png}
\caption{Demonstration of the sparsity of Sparse Centroid-encoder on MNIST digits 5 and 6. The digits are shown in white, and the selected pixels are marked using blue—the darkness of blue indicates the relative importance of the pixel to distinguish the two digits. We showed the selected pixels for two choices of $\lambda$. Notice that for $\lambda = 0.1$, the model chose the lesser number of features, whereas it picked more pixels for $\lambda = 0.001$. $\lambda$ is the nob which controls the sparsity of the model.}
\label{fig:SCE_Sparsity}
\end{figure}
Our next analysis shows how SCE extracts informative features from a multi-modal dataset, i.e., data sets whose classes appear to have multiple clusters. In this case, one center per class may not be optimal, e.g., ISOLET data. To this end, we trained SCE using a different number of centers per class where the centers were determined using standard $k$-Means algorithm \cite{lloyd1982least,macqueen1967some}. After the feature selection, we calculated the validation accuracy and plotted it against the number of centers per class in Figure \ref {fig:Isolet_Analysis}. The validation accuracy jumped significantly from one center to two centers per class. The increased accuracy indicates that the speech classes are multi-modal, further validated by the two-dimensional PCA plot of the three classes shown in panel (b)-(d).
\begin{figure}[h!]
\centering
\includegraphics[width=14.5cm,height=13.5cm]{Isolet_Accuracy-vs-nCenter_Class1-3_2D_PCA.png}
\caption{Sparse Centroid-encoder for multi-modal data set. Panel (a) shows the increase in validation accuracy over the number of centroids per class. Panel (b)-(d) shows the two-dimensional PCA plot of the three speech classes.}
\label{fig:Isolet_Analysis}
\end{figure}
Our last analysis sheds light on the feature selection stability of the SCE as shown in Figure \ref{fig:Feature_Stability}. In panel (a), we present the position of the selected pixels over two runs (194 and 198 respectively with 167 overlapping ones). Most of the selected pixels reside in the middle of the image, making sense as the MNIST digits lie in the center of a 28 x 28 grid. Notice that the non-overlapping pixels of the two runs are neighbors, making sense as the neighboring pixels perhaps contain similar information about the digits. On the SMK\_CAN dataset, which has over 19,000 features, the 1-norm of the sparse layer makes most of the variables to zero/near zero, selecting only 570 and 594 biomarkers. We didn't use hard thresholding to induce sparsity. Our feature cut-off technique, mentioned in Section \ref{feature_cut-off}, picks the non-zero biomarkers. Also, notice the absolute weight of the selected biomarkers, which suggests that the $\ell_1$ didn't shrink all parameters of the sparse layers.
\begin{figure}[ht!]
\centering
\includegraphics[width=13.5cm,height=9.0cm]{./figures/FeatureSparsity_MNIST_SMK_CAN.png}
\caption{Sparse Centroid-Encoder on MNIST (all ten classes) and high dimensional ($\#$features 19993) SMK\_CAN data. Panel (a) shows position of the selected pixels over two run ($\lambda = 0.0002$). Panel (b)-(c) shows the sparsity plot ($\lambda = 0.0002$) of the SPL layer over two run on SMK\_CAN training data.}
\label{fig:Feature_Stability}
\end{figure}
\newpage
\iffalse
\subsection{ Feature Selection Workflow Using Sparse Centroid-Encoder}
By design, sparse methods identify a small number
of features that accomplish a classification task. If one is interested in {\it all} the discriminatory features that can be used to separate multiple classes, then one can repeat the process of removing good features.
This section describes how sparse centroid-encoder (SCE) can be used iteratively to extract all discriminatory features from a data set; see \cite{o2013iterative} for an application of this approach to sparse support vector machines.
SCE is a model based on neural network architecture; hence, it's a non-convex optimization. As a result, multiple runs will produce different solutions, i.e., different feature sets on the same training set. These features may not be optimal given an unseen test set. To find out the robust features from a training set, we resort to frequency-based feature pruning. In this strategy, first, we divide the entire training set into $k$ folds. On each of these folds, we ran the SCE and picked the top $N$ (user select) number of features. We repeat the process $T$ times to get $k\times T$ feature sets. Then we count the number of occurrences of each feature and call this number the frequency of a feature. We ordered the features based on the frequency and picked the optimum number from a validation set. We present the feature selection work flow in Figure \ref{IFS_SCE}. We trained SCE using Scaled Conjugate Gradient Descent \cite{moller1993scaled}. The architectural details are kept in Appendices.
\begin{figure*}[ht!]
\centering
\includegraphics[width=15.0cm,height=9.0cm]{FeatureSelectionWorkFlowFinal.png}
\caption{Feature selection workflow using Sparse Centroid-encoder. \textbf{a} First, the data set has been partitioned into training and validation. \textbf{b} We further partitioned the training set into $\textit{n}$ splits. \textbf{c} On each of the training splits, we ran Sparse Centroid-encoder to get $\textit{n}$ feature sets. \textbf{d} We calculated the occurrence of each feature among the $\textit{n}$ sets and called it the frequency of the feature. We ranked features from high to a low frequency to get an ordered set. \textbf{e} At last, we picked the optimum number of features using a validation set.}
\label{IFS_SCE}
\end{figure*}
\fi
\section{Experimental Results}
\label{experiments}
We present the comparative evaluation of our model on various data sets using several feature selection techniques.
\subsection{Experimental Details}
We used twelve data sets from a variety of domains (image, biology, speech, and sensor; see Table\ref{table:dataDescription}) and five neural network-based models to run three benchmarking experiments. To this end, we picked the published results from four papers \cite{lemhadri2021lassonet, singh2020fsnet, li2016deep, yamada2020feature} for benchmarking. We followed the same experimental methodology described in those papers for an apples-to-apples comparison. This approach permitted a direct comparison of LassoNet, FsNet, Supervised CAE, DFS, and Stochastic Gates using the authors' best results. All three experiments follow the standard workflow.
\begin{itemize}
\item Split each data sets into training and test partition.
\item Run SCE on the training set to extract top $K\in\{10,16,50\}$ features.
\item Using the top $K$ features train a one hidden layer ANN classifier with $H$ ReLU units to predict the test samples. The $H$ is picked using a validation set.
\item Repeat the classification 20 times and report average accuracy.
\end{itemize}
Now we describe the details of the three experiments.
\textbf{Experiment 1:} The first bench-marking experiment is conducted on five real-world high dimensional biological data sets: ALLAML, GLIOMA, SMK\_CAN, Prostate\_GE, and GLI\_85\footnote{Available at http://featureselection.asu.edu/datasets.php} to compare SCE with FsNet and Supervised CAE (SCAE). Following the experimental protocol of Singh et al. \cite{singh2020fsnet}, we randomly partitioned each data into a 50:50 ratio of train and test and ran SCE on the training set. After that, we calculated the test accuracy using the top $K=\{10,50\}$ SCE features. We repeated the experiment 20 times and reported the mean accuracy. We ran a 5-fold cross-validation on the training set to tune the hyperparameters.
\textbf{Experiment 2:} In the second bench-marking experiment, we compared SCE with LassoNet\cite{lemhadri2021lassonet} and Stochastic Gate\cite{yamada2020feature} on six data sets: Mice Protein\footnote{There are some missing entries that are imputed by mean feature values.}, COIL20, Isolet, Human Activity, MNIST, and FMNIST\footnote{Available at UCI Machine Learning repository}. Following the experimental set of Lemhadri et al., we split each data set into 70:10:20 ratio of training, validation, and test sets. We ran SCE on the training set to pick the top $K=50$ features to predict the class labels of the sequester test set. We extensively used the validation set to tune the hyperparameters.
\textbf{Experiment 3:} In the last benchmark, we used the single cell GM12878 data \footnote{The authors of \cite{li2016deep} shared the data with us} which has separate training, validation, and test sets. The SCE is run to select the top $K=16$ features to compare the prediction performance with Deep/Shallow DFS \cite{li2016deep}, and Lasso. Again we used the validation set for hyperparameters tuning.
\begin{table}[!ht]
\vspace{1mm}
\centering
\begin{tabular} {|c|c|c|c|c|c|}
\hline
Dataset & No. Features & No. of Classes & No. of Samples & Domain \\
\hline
ALLAML & 7129 & 2 & 72 & Biology \\
GLIOMA & 4434 & 4 & 50 & Biology \\
SMK\_CAN & 19993 & 2 & 187 & Biology \\
Prostate\_GE & 5966 & 2 & 102 & Biology \\
GLI\_85 & 22283 & 2 & 85 & Biology \\
GM12878 & 93 & 3 & 6468 & Biology \\
Mice Protein & 77 & 8 & 975 & Biology \\
COIL20 & 1024 & 20 & 1440 & Image \\
Isolet & 617 & 26 & 7797 & Speech \\
Human Activity & 561 & 6 & 5744 & Accelerometer Sensor \\
MNIST & 784 & 10 & 70000 & Image \\
FMNIST & 784 & 10 & 70000 & Image \\
\hline
\end{tabular}
\caption{Descriptions of the data sets used for benchmarking experiments.}
\label{table:dataDescription}
\end{table}
\subsection{Results}
Now we discuss the results of the three benchmarking experiments. In Table \ref{table:exp1_results} we present the results of the first experiment where we compare SCE, SCAE, and FsNet on five high-dimensional biological data sets. Apart from the results using a subset (10 and 50) of features, we also provide the prediction using all the features. In most cases, feature selection helps improve classification performance. Generally, SCE features perform better than SCAE and FsNet; out of the ten classification tasks, SCE produces the best result on eight. Notice that the top fifty SCE features give a better prediction rate than the top ten in all the cases. Interestingly, the accuracy of SCAE and FsNet drop significantly on SMK\_CAN and GLI\_85 using the top fifty features.
\begin{table}[h!]
\vspace{-0.25cm}
\centering
\begin{tabular} {|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{Data set} & \multicolumn{3}{c|} {Top 10 features} & \multicolumn{3}{c|} {Top 50 features} & \multirow{1}{*}{All features} \\
\cline{2-7}
& \multicolumn{1}{c|} {FsNet} & \multicolumn{1}{c|} {SCAE} & \multicolumn{1}{c|} {SCE} & \multicolumn{1}{c|} {FsNet} & \multicolumn{1}{c|} {SCAE} & \multicolumn{1}{c|} {SCE} & \multirow{1}{*}{ANN} \\
\hline
ALLAML & $91.1$ & $83.3$ & $\textbf{92.5}$ & $92.2$ & $93.6$ & $\textbf{95.9}$ & $89.9$\\
\hline
Prostate\_GE & $87.1$ & $83.5$ & $\textbf{89.5}$ & $87.8$ & $88.4$ & $\textbf{89.9}$ & $75.9$\\
\hline
GLIOMA & $62.4$ & $58.4$ & $\textbf{63.2}$ & $62.4$ & $60.4$ & $\textbf{69.0}$ & $70.3$\\
\hline
SMK\_CAN & $\textbf{69.5}$ & $68.0$ & $68.6$ & $64.1$ & $66.7$ & $\textbf{69.4}$ & $65.7$\\
\hline
GLI\_85 & $87.4$ & $\textbf{88.4}$ & $84.1$ & $79.5$ & $82.2$ & $\textbf{85.5}$ & $79.5$\\
\hline
\end{tabular}
\caption{Comparison of mean classification accuracy of FsNet, SCAE, and SCE features on five real-world high-dimensional biological data sets. The prediction rates are averaged over twenty runs on the test set. Numbers for FsNet and SCAE are being reported from \cite{singh2020fsnet}.}
\label{table:exp1_results}
\end{table}
Now we turn our attention to the results of the second experiment, as shown in Table\ref{table:exp2_results}. The features of the Sparse Centroid-Encoder produce better classification accuracy than LassoNet and STG in all the cases. Especially for Mice Protein, Activity, Isolet, FMNIST, and MNIST, our model has better accuracy by $2.5\% - 4.5\%$. The results for Stochastic Gates (STG) in \cite{yamada2020feature} are not in a table form, but our eyeball comparison of classification accuracy with the top 50 features on ISOLET, COIL20, and MNIST suggests that stochastic gate is not more accurate than SCE. For example, using the top 50 features, STG obtains approximately 85\% accuracy on ISOLET while SCE obtains 91.1\%; STG obtains about 97\% on COIL20 while SCE obtains 99.3\%; on the dataset, MNIST STG achieves approximately 91\% while SCE 93.8\%. In this experiment, we ran SCE with multiple centroids per class and observed an improved prediction rate than one center per class on Isolet, Activity, MNIST, and FMNIST. The observation suggests that the classes are multi-modal, providing a piece of valuable information. The optimum number of centers was picked using the validation set, see Figure \ref{fig:Isolet_Analysis}.
\begin{table}[ht!]
\vspace{1mm}
\centering
\begin{tabular} {|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{Data set} & \multicolumn{3}{c|} {Top 50 features} & \multirow{1}{*}{\#Centers} & \multirow{1}{*}{All features} \\
\cline{2-4}
& \multicolumn{1}{c|} {LassoNet} & \multicolumn{1}{c|} {STG} & \multicolumn{1}{c|} {SCE} & \multicolumn{1}{c|}{for SCE} & {ANN} \\
\hline
Mice Protein &$95.8$ & NA &$\textbf{98.4}$ & 1 & $100.00$\\
\hline
MNIST & $87.3$ & $91.0$ & $\textbf{93.8}$ & 3 & $97.60$\\
\hline
FMNIST & $80.0$ & NA &$\textbf{84.7}$ & 3 & $90.16$\\
\hline
ISOLET & $88.5$ & $85.0$ &$\textbf{91.1}$ & 5 & $96.96$\\
\hline
COIL-20 & $99.1$ & $97.0$ &$\textbf{99.3}$ & 1 & $98.87$\\
\hline
Activity & $84.9$ & NA &$\textbf{89.4}$ & 4 & $92.81$\\
\hline
\end{tabular}
\caption{Classification results using LassoNet, STG, and SCE features on six publicly available data sets. The column '\#Centers for SCE' denotes how many centroids per class are used to train SCE. Numbers for LassoNet and STG are reported from \cite{lemhadri2021lassonet} and \cite{yamada2020feature} respectively. All the reported accuracies are measured on the test set. NA means the result has not been reported.}
\label{table:exp2_results}
\end{table}
In Table\ref{table:exp3_results}, we present the results of our last experiment on the single cell data GM12878. We use the published results for deep feature selection (DFS), shallow feature selection, and Lasso from the work of Li et al. to evaluate SCE. To compare with Li et al., we used the top 16 features to report the mean accuracy of the test samples. We see that the SCE features outperform all the other models. Among all the models, Lasso exhibits the worst performance with an accuracy of $81.86\%$. This relatively low accuracy is not surprising, given Lasso is a linear model.
\begin{table}[ht!]
\centering
\begin{tabular} {|c|c|c|c|c|}
\hline
\multirow{2}{*}{Data set} & \multicolumn{4}{c|} {Top 16 features} \\
\cline{2-5}
& \multicolumn{1}{c|} {SCE} & \multicolumn{1}{c|} {Deep DFS} & \multicolumn{1}{c|} {Shallow DFS} & \multicolumn{1}{c|}{Lasso} \\
\hline
GM12878 & \textbf{87.51} & 85.67 & 85.34 & 81.86 \\
\hline
\end{tabular}
\caption{Classification accuracies using the top 16 features by various techniques. Results of Deep DFS, Shallow DFS, Lasso, and Random Forest are reported from \cite{li2016deep}.}
\label{table:exp3_results}
\end{table}
The classification performance gives a quantitative measure that doesn't reveal the biological significance of the selected genes. We did a literature survey of the top genes selected by sparse centroid-encoder and provided a detailed description in the appendix. Some of these genes play an essential role in transcriptional activation, e.g., H4K20ME1 \cite{barski2007high}, TAF1 \cite{wang2014crystal}, H3K27ME3 \cite{cai2021h3k27me3}, etc. Gene H3K27AC \cite{creyghton2010histone} plays a vital role in separating active enhances from inactive ones. Besides that, many of these genes are related to the proliferation of the lymphoblastoid cancer cells, e.g., POL2 \cite{yamada2013nuclear}, NRSF/REST \cite{kreisler2010regulation}, GCN5 \cite{yin2015histone}, PML \cite{salomoni2002role}, etc. This survey confirms the biological significance of the selected genes.
\iffalse
\subsection{Quantitative and Qualitative Analysis}
\label{quant_qual_analysis}
Now we present the results from a comprehensive analysis across nine data sets.
\subsubsection{Comparison with LassoNet on UCI data sets}
In our first set of experiments, we compared SCE with LassoNet \cite{lemhadri2021lassonet} on six publicly data sets taken from different domains, including image (MNIST, Fashion MNIST, COIL-20), speech (ISOLET), activity recognition (Human Activity Recognition Using Smartphones), and a biological data set (Mice Protein data). These data sets have been used in the literature for benchmarking \cite{lemhadri2021lassonet,balin2019concrete}. Following the experimental protocol of Lemhadri et al., we randomly partitioned each data set into a 70:10:20 split of training, validation, and test set. We normalized the training partition by subtracting the mean and dividing each feature by its corresponding standard deviation. We used the mean and standard deviation of the training to standardize the test samples. We used the training set for feature selection and the validation set for hyperparameter tuning. After the feature selection, we used a one hidden layer neural network classifier to predict the class label on the test set. We ran the classifier ten times ($K=10$) and presented the mean accuracy with standard deviation in Table \ref{table:LassoNet-vs-SCE}. Apart from showing the accuracy with top-50 features, we also give classification results with an optimum number of features selected using the validation set.
As we can see from Table \ref{table:LassoNet-vs-SCE}, the features of the Sparse Centroid-encoder produce better classification accuracy than LassoNet in most of the cases. Especially for Mice Protein, Activity, and MNIST, our model has better accuracy by a margin of $2.5\% - 3.5\%$. On COIL-20, Lemhadri et al. preprocessed the data by resizing them to 20 x 20 images. We ran our experiment on the original images. Apart from comparing with LassoNet, we also presented accuracies using all features and the optimum number of features. In most cases, the optimum number of features gave a reasonably good performance compared to all features. It's noteworthy that MNIST and COIL-20 classification using the optimum number of variables is slightly better than all; we used only $48\%$ and $8.7\%$ of total features, respectively. The superior performance of our model on a wide range of data sets suggests that Sparse Centroid-encoder may be applied as a feature detector in many domains.
\begin{table*}[h!]
\vspace{1mm}
\centering
\begin{tabular} {|c|c|c|c|c|c|c|}
\hline\hline
\multirow{2}{*}{Data set} & \multicolumn{2}{c|} {Top 50 features} & \multicolumn{2}{c|} {Optimum no. of features} & \multirow{1}{*}{Centers / Class} & \multirow{2}{*}{All features} \\
\cline{2-5}
& \multicolumn{1}{c|} {LassoNet} & \multicolumn{1}{c|} {SCE} & \multicolumn{1}{c|} {No. of features} & \multicolumn{1}{c|} {SCE} & \multicolumn{1}{c|}{for SCE} & \\
\hline
Mice Protein & $95.8$ & $\textbf{99.10} \pm \textbf{0.51}$ & 22 & $99.46 \pm 0.44$ & 1 & $100.00$\\
\hline
MNIST & $87.3$ & $\textbf{90.78} \pm \textbf{0.09}$ & 377 & $97.93 \pm 0.03$ & 1 & $97.60 \pm 0.08$\\
\hline
FMNIST & $80.0$ & $\textbf{80.27} \pm \textbf{0.11}$ & 355 & $88.61 \pm 0.10$ & 3 & $90.16 \pm 0.17$\\
\hline
ISOLET & $88.5$ & $\textbf{89.06} \pm \textbf{0.27}$ & 245 & $95.51 \pm 0.21$ & 5 & $96.96 \pm 0.16$\\
\hline
COIL-20 & $\textbf{99.1}$ & $98.03 \pm 0.28$ & 90 & $99.83 \pm 0.17 $ & 1 & $98.87 \pm 0.52$\\
\hline
Activity & $84.9$ & $\textbf{87.53} \pm \textbf{0.49}$ & 90 & $ 89.42 \pm 0.67$ & 4 & $92.81 \pm 0.29$\\
\hline \hline
\end{tabular}
\caption{Classification results using LassoNet and SCE features on six publicly available data sets. We compared LassoNet with SCE using the top 50 variables, and we also reported accuracy using the optimum number and all features for each data set. Numbers for LassoNet are reported from \cite{lemhadri2021lassonet}. All the reported accuracies were measured on the test set.}
\label{table:LassoNet-vs-SCE}
\end{table*}
\subsubsection{Single Cell Data}
GM12878 is a single cell data set that has been previously used to test multiclass feature selection algorithms \cite{li2016deep}. The samples were collected from the annotated DNA region of lymphoblastoid cell line. Each sample is represented by a 93 dimensional features sampled from three classes: active enhancer (AE), active promoter (AP) and background (BG) where each class contains $2,156$ number of samples. The data set is split equally into a separate training, validation and test sets.
We used the validation set to tune hyper-parameters and to pick the optimal number of features. After the feature selection step, we merged the training and validation set and trained $K=10$ centroid-encoder classifiers with the selected features, and reported classification accuracy on the test set.
We use the published results
for deep feature selection (DFS), shallow feature selection, LASSO, and Random Forest (RF) from the work of Li et al.
to evaluate SCE as shown in Table \ref{table:GM12878_Result}.
To compare with Li et al., we used the top 16 features to report the mean accuracy of the test samples. We also report the test accuracy using the top 48 features picked from the validation set, this was the best result in our experiment.
When restricted to the top 16, we see that the SCE features still outperform all the other models.
\begin{table}[ht!]
\vspace{1mm}
\centering
\begin{tabular} {|c|c|c|}
\hline\hline
{Models} & {No. of Features} & {Accuracy}\\
\hline
SCE & $\textbf{48}$ & $\textbf{89.40} \pm \textbf{0.24}$ \\
SCE & $\textbf{16}$ & $\textbf{88.33} \pm \textbf{0.13}$ \\
Deep DFS & $16$ & $85.67$\\
Shallow DFS & $16$ & $85.34$\\
LASSO & $16$ & $81.86$\\
Random Forest & $16$ & $88.21$\\
\hline \hline
\end{tabular}
\caption{Classification accuracies using the top 16 features by various techniques. Results of Deep DFS, Shallow DFS, LASSO, and Random Forest are reported from \cite{li2016deep}. We present accuracy with the top 48 features which were selected from a validation set.}
\label{table:GM12878_Result}
\end{table}
Among all the models, LASSO features exhibit the worst performance with an accuracy of $81.86\%$. This relatively low accuracy is not surprising, given LASSO is a linear model.
The classification performance gives a quantitative measure that doesn't reveal the biological significance of the selected genes. We did a literature survey of the top genes selected by sparse centroid-encoder and provided the detailed description in the appendix. Some of these genes play an essential role in transcriptional activation, e.g., H4K20ME1 \cite{barski2007high}, TAF1 \cite{wang2014crystal}, H3K27ME3 \cite{cai2021h3k27me3}, etc. Gene H3K27AC \cite{creyghton2010histone} plays a vital role in separating active enhances from inactive ones. Besides that, many of these genes are related to the proliferation of the lymphoblastoid cancer cells, e.g., POL2 \cite{yamada2013nuclear}, NRSF/REST \cite{kreisler2010regulation}, GCN5 \cite{yin2015histone}, PML \cite{salomoni2002role}, etc. This survey confirms the biological significance of the selected genes.
\subsubsection{Indian Pines Hyperspectral Imagery}
Here we compare SCE with sparse support vector machines
on the well-known Indian Pines hyperspectral imagery of a variety of crops following the experiments in
(\cite{chepushtanova2014band}). The task is to identify the frequency bands which are essential to assign a test sample correctly to one of the sixteen classes. We took all the 220 bands in the feature selection step, including the twenty water absorption bands. Note that in literature these noisy water absorption bands are often excluded before the experiments \cite{reshma2016dimensionality,cao2017hyperspectral}. We wanted to check whether our model was able to reject them. In fact, our model did discard them as no water absorption bands were in the top $100$ features. It appeared that SSVM included some of these noisy bands as described in\cite{chepushtanova2014band}.
We followed the experimental protocol in \cite{chepushtanova2014band}. The entire data set is split in half into a training and test set. Because of the small size of the training set, we did a 5-fold cross-validation on the training samples to tune hyper-parameters. After the feature selection on the training set, we took top $n=1,2,3,4,5,10,20,40,60,80$ features to build a CE classifier on the training set to predict the class labels of the test samples. For each $n$ we repeat the classification task $K=10$ times.
Note that we compared the performance of SCE and SSVM features without spatial smoothing for a more direct comparison of the classification rates.
Figure \ref{fig:IndianPine_Result} presents the accuracy on the test data using the top $n$ bands ($n=1, 2, 3, 4, 5, 10, 20, 40, 60, 80$) which were calculated on the training set.
Classification using SCE features generally produces better accuracy. Notice that SCE features yield better classification performance using fewer bands. In particular, the accuracy of the top SCE feature (band 13) is at least $15\%$ higher than the top SSVM feature (band 1). See the Appendix for additional details on the feature sets selected.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{SCE_vs_SSVM_IP.png}
\caption{Comparison of classification accuracy using SCE and SSVM features.}
\label{fig:IndianPine_Result}
\end{figure}
\subsubsection{Respiratory Infections in Humans}
\textbf{GSE73072} This microarray data set is a collection of gene expressions taken from human blood samples as part of multiple clinical challenge studies \cite{liu2016individualized} where individuals were infected with the following respiratory viruses HRV, RSV, H1N1, and H3N2. In our experiment we excluded the RSV study. Blood samples were taken from the individuals before and after the inoculation. RMA normalization \cite{irizarry2003exploration} is applied to the entire data set, and the LIMMA \cite{ritchie2015limma} is used to remove the subject-specific batch effect. Each sample is represented by 22,277 probes associated with gene expression. The data is publically available on the NCBI GeneExpression Omnibus (GEO) with identifier GSE73072.
We conducted our last experiment on the GSE73072 human respiratory infection data where the goal is to predict the classes control, shedders, and non-shedders
at the very early phase of the infection, i.e., at time bin spanning hours 1-8.
Controls are the pre-infection samples, whereas shedders and non-shedders are post-infection samples picked from the time bin 1-8 hr.
Shedders actually disseminate virus while non-shedders do not.
We considered six studies, including two H1N1 (DEE3, DEE4), two H3N2 (DEE2, DEE5), and two HRV (Duke, UVA) studies. We used $10\%$ training samples as a validation set—the training set comprised all the studies except for the DEE5, which was kept out for testing. We did a leave-one-subject-out (LOSO) cross-validation on the test set using the selected features from the training set. In this experiment we compared SCE with Random Forest (RF).
The results on this data set in shown in Table \ref{table:GSE73072_Result}.
The top 35 features of SCE produce the best Balanced Success Rate (BSR) of $90.61\%$ on the test study DEE5. For the Random Forest model, the best result is achieved with 30 features. We also included the results with 35 biomarkers, but the BSR didn't improve. Note both the models picked a relatively small number of features, 30 and 35 out of the 22,277 genes, but SCE features outperform RF by a margin of $7\%$. Although RF selects features with multiple classes using a single model, it weighs a single feature by measuring the decrease of out-of-bag error. In contrast, SCE looks for a group of features while minimizing its cost. We think the multivariate approach of SCE makes it a better feature detector than RF.
\begin{table}[ht!]
\vspace{1mm}
\centering
\begin{tabular} {|c|c|c|c|}
\hline\hline
{Time Bin} & {Model} & {No. of Features} & {BSR}\\
\hline
\multirow{2}{*}{$1-8$} & SCE & $35$ & $90.61 \pm 2.38$ \\
\cline{2-4}
& RF & 30 & $83.05 \pm 2.42$ \\
\cline{2-4}
& RF & 35 & $82.65 \pm 2.51$ \\
\hline \hline
\end{tabular}
\caption{Balanced success rate (BSR) of LOSO cross-validation on the DEE5 test set. The selected features from training set is used to predict the classes of control, shedder, and non-shedder.}
\label{table:GSE73072_Result}
\end{table}
\fi
\section{Related Work}
\label{lit}
Feature selection has a long history spread across many fields, including bioinformatics, document classification, data mining, hyperspectral band selection, computer vision, etc. It's an active research area, and numerous techniques exist to accomplish the task. We describe the literature related to the embedded methods where the selection criteria are part of a model. The model can be either linear or non-linear.
\subsection{Feature Selection using Linear Models}
Adding an $l_1$ penalty to classification and regression methods naturally produce feature selectors.
For example, least absolute shrinkage and selection operator or Lasso \cite{tibshirani1996regression}
has been used extensively for feature selection on various data sets \cite{fonti2017feature,muthukrishnan2016lasso,kim2004gradient}.
Elastic net, proposed by Zou et al. \cite{zou2005regularization}, combined the Lasso penalty with the Ridge Regression penalty \cite{hoerl1970ridge} to overcome some limitations of Lasso.
Elastic net has been widely applied, e.g., \cite{marafino2015efficient,shen2011identifying,sokolov2016pathway}. Note both Lasso and Elastic net are convex in the parameter space. Also see the following works using $\ell_1$ based features selection on linear models see \cite{lindenbaum2021randomly,candes2008enhancing,daubechies2010iteratively,bertsimas2017trimmed,xie2009scad}.
Support Vector Machines (SVM) \cite{cortes1995support} is a state-of-the-art model for classification, regression and feature selection. SVM-RFE is a linear feature selection model which iteratively removes the least discriminative features until a parsimonious set of predictive features are selected \cite{guyon2002gene}. IFR \cite{o2013iterative}, on the other hand, selects a group of discriminatory features at each iteration and eliminates them from the data set. The process repeats until the accuracy of the model starts to drop significantly. Note IFR uses Sparse SVM (SSVM), which minimizes the $l_1$ norm of the model parameters. Lasso, Elastic Net, and SVM-based techniques are mainly applied to binary problems. These models are extended to the multi-class problem by combining multiple binary one-against-one (OAO) or one-against-all (OAA) models. \cite{chepushtanova2014band} used 120 Sparse SVM models to select discriminative bands from the Indian Pine data set, which has 16 classes.
On the other hand, Random forest \cite{breiman2001random}, a decision tree-based technique, finds features from multi-class data using a single model. The model doesn't use Lasso or Elastic net penalty for feature selection. Instead, the model weighs the importance of each feature by measuring the out-of-bag error.
\subsection{Feature Selection using Deep Neural Networks}
While the linear models are fast and convex, they don't capture the non-linear relationship among the input features (unless a kernel trick is applied). Because of the shallow architecture, these models don't learn a high-level representation of input features. Moreover, there is no natural way to incorporate multi-class data in a single model. Non-linear models based on deep neural networks overcome these limitations. In this section, we will briefly discuss a handful of such models.
\cite{scardapane2017group} used group Lasso \cite{tibshirani1996regression} to impose the sparsity on a group of variables instead of a single variable. They applied the group sparsity simultaneously on the input and the hidden layers to remove features from the input data and the hidden activation.
Li et al. proposed deep feature selection (DFS), which is a multilayer neural network-based feature selection technique \cite{li2016deep}. DFS uses a one-to-one linear layer between the input and the first hidden layer. As a sparse regularization, the authors used elastic-net \cite{zou2005regularization} on the variables of the one-to-one layer to induce sparsity. The standard soft-max function is used in the output layer for classification. With this setup, the network is trained in an end-to-end fashion by error backpropagation. Despite the deep architecture, its accuracy is not competitive, and experimental results have shown that the method did not outperform the random forest (RF) method. \cite{kim2016opening} proposed a heuristics based technique to assign importance to each feature. Using the ReLU activation, \cite{roy2015feature} provided a way to measure the contribution of an input feature towards hidden activation of next layer.
\cite{han2018autoencoder} developed an unsupervised feature selection technique based on the autoencoder architecture. Using a $l_{2,1}$-norm to the weights emanating from each input node, they measure the contribution of each feature while reconstructing the input. The model excavates the input features, which have a minimum contribution.
\cite{taherkhani2018deep} proposed a RBM \cite{Hinton:2006:FLA:1161603.1161605,hinton2006reducing} based feature selection model.
This algorithm runs the risk of combinatorial explosion for data set with $50K-60K$ features (e.g., microarray gene expression data set).
\section{Discussion and Conclusion}
\label{dis_cons}
\iffalse
In this paper, we presented Sparse Centroid-Encoder as an effective feature selection tool for multiclass classification problems. The benchmarking results span nine diverse data sets and six methods providing evidence
that the features of SCE produce better generalization performance compared to other models.
We compared SCE with deep feature selection (deep DFS), a deep neural network-based model, and found that the features of SCE significantly improved the classification rate on test data. The survey of the known functionality of these genes indicates plausible biological significance.
In addition to extracting the most robust features, the model shows the ability to discard irrelevant features. E.g., on the Indian Pines hyperspectral image data set, our model didn't pick any of the water absorption bands considered to be noisy.
We also compare our model
to multi-class extensions
sparse binary classifiers. Chepushtanova et al. used 120
binary class SSVM models on the Indian Pine data set. Similarly, Lasso needed three models for the GM12878 data. These models will suffer a combinatorial explosion when the number of classes increases. In contrast, SCE uses a single model to extract features from a multi-class data set.
SCE compares favorably to the neural network-based models deep DFS, LassoNet, where samples are mapped to class labels.
SCE employs multiple centroids to capture the variability within a class, improving the prediction rate of unknown test samples. In particular, the prediction rate on the ISOLET improved significantly from one centroid to multiple centroids suggesting the speech classes are multi-modal. The two-dimensional PCA of ISOLET classes further confirms the multi-modality of the classes. We also observed an enhanced classification rate on FMNIST and Activity data with multiple centroids. In contrast, single-center per class performed better for other data sets (e.g., COIL-20, Mice Protein, GM12878, etc.). Hence, apart from producing an improved prediction rate using features that capture intra-class variance, our model can provide extra information about whether the data is unimodal or multi-modal. This aspect of sparse centroid-encoder distinguishes it from the LassoNet and DFS, which do not model the multi-modal nature of the data.
We also presented a feature selection workflow to determine the optimal number of robust features. Our experimental results showed that the prediction rate using the optimal number of features produces the best result. These results are often very competitive compared to using all features. For example, on Mice Protein data, we got an accuracy of $99.46\%$ using only 22 variables. On MNIST and COIL-20, our approach used only $48\%$ and $9\%$, respectively, of the total feature sets and actually outperform the prediction rate of all features. On the GSE73072 data set, we got more than $90\%$ accuracy on the holdout test set using only 35 out of 22,277 features.
\fi
In this paper, we presented Sparse Centroid-Encoder as an efficient feature selection tool for binary and multiclass problems. The benchmarking results span twelve diverse data sets and six methods providing evidence that the features of SCE produce better generalization performance than other state-of-the-art models. We compared SCE with FsNet, mainly designed for high-dimensional biological data, and found that our proposed method outperformed it in most cases. The comparison also includes Supervised CAE, which is not more accurate than SCE. On the data sets, where the no. of observations are more than the no. of variables, SCE produces the best classification results than LassoNet, Stochastic Gate, and DFS in each case. These experiments involve image, speech, and accelerometer sensor data. Moreover, the survey of the sixteen SCE genes of GM12878 indicates plausible biological significance. The empirical evaluation using an array of diverse data sets establishes the value of the Sparse Centroid-Encoder as a nonlinear feature detector.
Our analysis of the Sparse Centroid-Encoder in Section \ref{sparsity_analysis} demonstrates that the 1-norm induces good feature sparsity. We chose the $\lambda$ from the validation set from a wide range of values and saw that smaller values work better for classification. The visualization of the MNIST pixels (panel (a) of Figure \ref{fig:Feature_Stability}) provides a qualitative justification for a high prediction rate. SCE selected most of the pixels from the central part of the image, ignoring the border, making sense as the digits lie in the center of a 28 x 28 grid. The $\ell_1$ penalty on the SPL layer induces sharp sparsity on the SMK\_CAN data (panel (b) and (c) of Figure \ref{fig:Feature_Stability}) without shrinking all the variables. Our feature cut-off technique correctly demarcates the important features from the rest.
SCE compares favorably to the neural network-based model FsNet, SCAE DFS, LassoNet, and STG, where samples are mapped to class labels. SCE employs multiple centroids to capture the variability within a class, improving the prediction rate of unknown test samples. In particular, the prediction rate on the ISOLET improved significantly from one centroid to multiple centroids suggesting the speech classes are multi-modal. The two-dimensional PCA of ISOLET classes further confirms the multi-modality of data. We also observed an enhanced classification rate on MNIST, FMNIST, and Activity data with multiple centroids.
In contrast, single-center per class performed better for other data sets (e.g., COIL-20, Mice Protein, GM12878, etc.). Hence, apart from producing an improved prediction rate using features that capture intra-class variance, our model can provide extra information about whether the data is unimodal or multi-modal. This aspect of sparse centroid-encoder distinguishes it from the techniques which do not model the multi-modal nature of the data.
\bibliographystyle{unsrt}
|
\section{Introduction}
The quantum graphs - differential operators on metric graphs - is a rapidly growing branch of mathematical physics lying on the border between differential equations, spectral geometry and operator theory. During the last decade, particular attention was paid to study the impact on the spectrum of several differential operators such as Laplacian and Schr\"odinger operators on the metric graphs under the perturbations of the topology and geometry, i.e., graph surgery. These surgery principles proved to be a useful tool for the
spectral analysis of quantum graphs, in particular, for eigenvalue estimates. The usual goal is to derive spectral estimates that depend on the simple geometric properties of the graph, such as total length, the Betti number, the diameter or the number of vertices or edges of the underlying metric graph, see \cite{R,ZS,KNN,RL,A,BKKM1,KKK,F,KN,KKMM} and references therein. \\
The focus of the current work is to
\begin{itemize}
\item extend the existing surgical tool box to the graph Laplacian with $\delta'$ and anti-standard vertex interactions and
\item exploit this broader set of surgery principles to obtain new spectral bounds for graph Laplacians equipped with $\delta$ and $\delta'$ vertex interactions including standard, anti-standard, Neumann and Dirichlet vertex conditions.
\end{itemize}
Theorem 4.3 of \cite{BKKM} provides a very useful tool to derive interlacing inequalities between the eigenvalues of Laplacians with different sets of vertex conditions for a fixed underlying metric graph. Our derivation of some new eigenvalue bounds make intensive use of this result. Wherever possible we derive estimates on the general eigenvalues and not just the lowest non-zero eigenvalue.\\
Some of the surgical transformations for Laplacian with $\delta'$ vertex conditions appeared in \cite{RS}.
We will consider three additional (or a more general form of perturbations that appeared in \cite{RS}) types of perturbations of the graphs for $\delta'$-type vertex conditions.
\begin{enumerate}[(i)]
\item Increasing or decreasing vertex coupling parameter,
\item a pendant graph is added to a graph, or
\item a graph is inserted at some vertex of other graph.
\end{enumerate}
It is well known that under these transformations, eigenvalues of Laplacian with different set of vertex interactions behave differently.
For the Kirchoff (Standard) or $\delta$ vertex conditions, increasing the $\delta$-coupling parameter at a certain vertex of the graph increases the eigenvalues. That is, a general monotonicity principle is valid. Moreover, under the perturbation (i), the \textcolor{blue}{Theorem (3.4)} of \cite{BKKM} and \textcolor{blue}{Theorem (3.1.8)} of \cite{BK1} establishes the interlacing of the eigenvalues of the two graphs. Furthermore, the Dirichelet eigenvalues of the graph are sandwiched between two consecutive eigenvalues of same underlying $\delta$-graph. For the anti-standard as well as $\delta'$ type conditions, we show that the effect on the eigenvalues under the transformation (i), in contrast to standard or $\delta$-conditions, depends on the sign of the coupling parameters. It turns out that the perturbation (i) divides the permutation invariant conditions into two classes: those for which increasing the coupling parameter from a negative (positive) number to a negative (positive) number leads to the non-increasing eigenvalues, and those for which increasing the coupling parameter from a negative number to a positive number lead to the non-decreasing eigenvalues. Moreover, the eigenvalues of graph equipped with anti-standard conditions are squeezed between two consecutive eigenvalues of same underlying $\delta'$-graph, see, \textcolor{blue}{Theorem \eqref{strength}}. \\
The later part of section $3$ is devoted to the surgery transformation which increases the total length of the metric graph. In \textcolor{blue}{Theorem \eqref{length1}} it is shown that for permutation invariant condition such as anti-standard and $\delta'$ conditions the eigenvalues behave non-increasingly if the length of an edge is scaled by a factor greater than one. Moreover, if graph and the coupling parameters at each vertex are scaled up by some factor $t>1$ then then eigenvalues are pushed down by the factor $\frac{1}{t^2}$, see \textcolor{blue}{Theorem \eqref{length2}}.
In \cite{R,BK,KMN,KN} eigenvalue estimates were derived for certain basic surgical operations of quantum graphs, namely, gluing vertices and attaching edges. The paper \cite{RS} deals with these operations for more general (permutation invarient vertex conditions) self-adjoint vertex conditions.
More sophisticated surgery operations where the set of edges is changed were investigated in\cite{BKKM1,KKMM, RL}, using the symmetrisation technique
first applied to quantum graphs by L. Friedlander \cite{F} (see also \cite{KKK}). For the $\delta$-type vertex conditions it is known that the eigenvalues behave non-increasing under the surgery transformations (i) and (ii), see \cite{BKKM}.
In \textcolor{blue}{Theorem \eqref{pendant}} of the this paper, we show that for anti-standard and $\delta'$-type vertex conditions, the effect on the eigenvalues under the transformation (ii) also depend on the sign of the coupling parameters. This perturbation divides the vertex conditions into two parts: transformation (ii) that decreases the eigenvalues and those for which the perturbation (ii) increases the eigenvalues. Moreover, in \textcolor{blue}{Proposition \eqref{insert}} it is shown that for $\delta'$-type conditions the perturbation (iii) decreases the eigenvalues. In Section $4$, we provide some lower and upper bounds on the eigenvalues of the graphs equipped with (among others) $\delta$ and $\delta'$-type conditions. First, we provide some upper bounds on the lowest eigenvalue of $\delta'$-graph using some trial functions from the domain of the quadratic form. Later we use the surgical transformation to obtain a few bounds on the eigenvalues. \textcolor{blue}{Theorem \eqref{ariturk1}} provides an upper bound in terms of the betti number and the number of edges for the Neumann-Kirchoff eigenvalues and \textcolor{blue}{Theorem \eqref{delta1}} give bounds on the eigenvalues of $\delta$-graph. A similar but a better estimate can be found in \cite[\textcolor{blue}{Lemma 1.5}]{A}. However, for any flower graph the estimate in \textcolor{blue}{Theorem \eqref{ariturk1}} coincides with the estimate presented in \cite[\textcolor{blue}{Lemma 1.5}]{A}. In \textcolor{blue}{Theorem \eqref{ariturk2}} we provide the upper bounds on the eigenvalues of a metric graph with standard conditions at internal vertices and Dirichlet conditions at the vertices of degree one. For the lower and upper estimates on the eigenvalues of graph with anti-standard conditions, see \textcolor{blue}{Theorem \eqref{ec} and \eqref{anti2}}, similar estimates for standard conditions can be found in \cite{BKS}. \textcolor{blue}{Theorem \eqref{delta'} and \eqref{deltaprime}} provide lower bounds on the eigenvalues of $\delta'$-graph with negative and positive coupling parameters, respectively. The later part of this section includes a few more bounds on the eigenvalues of $\delta$ and $\delta'$-graph, these bounds are obtained using the well-known inequalities and especially the estimates on the eigenvalues of metric tree presented in \cite{R} and \cite{ZS} for standard and anti-standard vertex conditions, respectively. We also include the equivalent form of these estimates expressed in terms of other topological parameters of the underlying metric graph. In the last part, we have used a simple relation between the eigenvalues of some bipartite graphs equipped with standard and anti-standard conditions to first obtain a relation between the eigenvalues of a bipartite graph and then for general graphs equipped with $\delta$ and $\delta'$ conditions.\\
We would like to mention that all results depend
only on the vertex conditions at those vertices which are changed by the graph transformation. At all other vertices, general self-adjoint conditions are allowed. The proofs of our results are all variational comparing Rayleigh quotients, which is a standard method in obtaining eigenvalue estimates. In fact, estimates on the quadratic form associated to the Laplace operator on suitable finite-dimensional subspaces together with an application of the min-max principle yields the desired estimates for the eigenvalues.
\section{Preliminaries}
This section provides some basics of the quantum graph. It also provides the set of interlacing inequalities between the eigenvalues of the same underlying graph equipped with different conditions such as Dirichlet, Neumann, standard, anti-standard, $\delta$ and $\delta'$. These inequalities are obtained using Theorem \cite[\textcolor{red}{Theorem 4.3}]{BKKM} and help to prove some bounds on the eigenvalues. For more detail on this subject, we refer to \cite{BK1}.\\
Let $\Gamma$ be a finite and connected graph having a finite set of edges $E$ and a finite set of vertices $V$. In addition, we identify each edge $e \in E$ with an interval $[0,\ell_e]$, where $\ell_e >0$ denote the length of the edge and let $x_e$ be coordinate along this edge which induces a metric on $\Gamma$. The sum of the lengths of each edge denotes the total length of a metric graph $\Gamma$. That is, $$\mathcal{L}(\Gamma)=\sum\limits_{e \in E} \ell_e.$$
Let $|E|$ and $|V|$ denote the number of edges and the number of vertices in a metric graph $\Gamma$. Each edge $e$ connects a pair of vertices $v_1$ and $v_2$, where a vertex is viewed as a subset of endpoints of edges. The vertices are disjoint such that their union coincides with the set containing each end point of all edges. An edge of $\Gamma$ can be a loop with both endpoints connected to a single vertex, and we also allow multiple edges to run between a pair of vertices. We assume that the graph $\Gamma$ is compact, which means there are a finite number of edges, each having a finite length. An edge $e$ is said to be an incident to a vertex $v$ if at least one of the endpoints belongs to that vertex and denote by $e \sim v$, and let $E_v$ denote the set of all edges incident to a vertex $v$. The number of edges incidents to a vertex $v$, denoted by $d_v$, is called the degree of a vertex. A vertex of degree one is the boundary vertex, and an edge incident to such a vertex is a boundary edge. The first Betti number, the number of independent cycles, of a metric graph $\Gamma$ is given by,
$$\beta=|E|-|V|+1.$$
\\
Our main objective is to study the spectrum of the Laplacian $\frac{-d^2}{dx^2}$ acting on the functions living on the edges of $\Gamma$ with vertices equipped with one of the below-mentioned vertex conditions. The operator is self-adjoint in the Hilbert space
$$L^2(\Gamma)=\bigoplus\limits_{e \in E} L^2(0,\ell_e).$$
Which consists of functions $\varphi: \Gamma \mapsto \mathbb{C}$, that are square-integrable along each edge. Let $\varphi_e$ be the restriction of $\varphi$ to some edge $e \in E$. We define this operator on a subspace of the Sobolev space $W^2_2(\Gamma \backslash V) $, consisting of functions $\varphi \in L^2(\Gamma)$, whose weak derivatives up to order two exist on each edge and are also square-integrable. We denote these Sobolev spaces by
$$H^2(\Gamma)=\bigoplus\limits_{e \in E} H^2(0, \ell_e).$$
equipped with the standard Sobolev norm and inner product. To describe the vertex conditions of our interest, let us denote the limiting value of a function $\varphi$ and its derivative at any endpoint of an edge by
$$\varphi(x_i)= \lim\limits_{x \mapsto x_i} \varphi(x) $$ and
\begin{equation*}
\partial \varphi(x_i)
=\begin{cases} \varphi'(x_i), \quad \quad \text{if $x_i$ is the left endpoint}, \\
- \varphi'(x_i), \quad \ \text{if $x_i$ is the right endpoint.}
\end{cases}
\end{equation*}
The vertex conditions define the linear relation between values of function and their derivatives at a vertex. The operator $L=-\frac{d^2}{dx^2}$ is a self-adjoint defined on the functions from the Sobolev space $W^2_2(\Gamma \backslash V)$ satisfying the following $\delta$-type vertex conditions.
\begin{equation} \label{delta}
\begin{cases} \varphi(x_i)= \varphi(x_j) \equiv \varphi(v_m),\quad x_i,x_j\in v_m, \quad (\text{continuity condition})\\
\sum\limits_{x_i \in v_m}\partial \varphi(x_i)=\alpha_m \varphi(v_m).
\end{cases}
\end{equation}
Here $\alpha_m$ are the fixed real numbers associated with each vertex $v_m$ and are commonly known as interaction strength of $\delta$-conditions. The corresponding quadratic form and its domain is given by,
\begin{equation} \label{qdelta}
h[\varphi]:=\int_\Gamma |\varphi'(x)|^2 \,dx+\sum_{m=1}^{|V|}\alpha_m|\varphi(v_m)|^2
\end{equation}
with domain,
$$D(h)= \{ \varphi \in H^1(\Gamma): \varphi(x_i)=\varphi(x_j), \quad x_i, x_j \in v_m \}.$$
The standard conditions are defined by setting $\alpha_m=0$ in \eqref{delta} at all vertices of $\Gamma$, and the quadratic form is given by $h[\varphi]:=\displaystyle{\int_\Gamma |\varphi'(x)|^2 dx}$ , defined on continuous functions in $H^1(\Gamma)$. We remark that the vertices of degree two equipped with standard vertex conditions do not affect the spectrum of the graph. Thus, any interior point can be regarded as a vertex of degree two equipped with standard vertex conditions and similarly, any vertex of degree two can be suppressed. \\
The Dirichlet vertex conditions are those in which the value of a function $\varphi$ is zero at each end point of an interval, that is, $$\varphi(x_i)=0, \quad x_i \in v_m. $$
Equivalently, this means the function vanishes at a vertex $v_m$ of a metric graph $\Gamma$. The Dirichlet conditions corresponds to $\delta$-type conditions if we let $\alpha_m=\infty$ in \eqref{delta}.
When Dirichlet conditions are imposed at some vertex, the functions along the incident edges do not interact (no communication between edges). Suppose all vertices are equipped with Dirichlet condition. In that case, the metric graph is the collection of independent intervals. The differential operator decouples into a direct sum of the operators on separate intervals with Dirichlet condition imposed at endpoints. The spectrum of the metric graph coincides with the disjoint union of the spectrum of each interval. That is the operator is not changed if two Dirichlet vertices are identified to form a single Dirichlet vertex. The quadratic form is given by the expression
$h[\varphi]:=\int_\Gamma |\varphi'(x)|^2 dx$, with domain $D(h)=\{\varphi \in H^1(\Gamma): \varphi(v)=0 \}.$\\
In the $\delta'$-type condition, the role of the functions and their derivative are switched. In this case, the Laplacian is still self-adjoint on the functions from the Sobolev space $W^2_2(\Gamma \backslash V)$ satisfying the following $\delta'$-type vertex conditions.
\begin{equation} \label{delta'}
\begin{cases}
\partial \varphi(x_i)=\partial \varphi (x_j), \quad x_i,x_j\in v_m, \quad (\text{continuity condition}) \\
\sum \limits_{x_i \in v_m}\varphi(x_i)=\alpha'_m \dfrac{d \varphi(v_m)}{dx}, \quad \alpha'_m \neq 0.
\end{cases}
\end{equation}
The corresponding quadratic form is defined on the functions in $H^1(\Gamma)$, and is given by,
\begin{equation} \label{qdelta'}
h[\varphi]:=\int_\Gamma |\varphi'(x)|^2 \,dx+\sum\limits_{m=1}^{|V|} \frac{1}{\alpha'_m} \left|\sum \limits_{x_i \in v_m}\varphi(x_i)\right|^2.
\end{equation}
Consider the case when $\alpha'_m=0$ in \eqref{delta'} at some vertex $v_m$ of a metric graph $\Gamma$. The quadratic form in \eqref{qdelta'} only makes sense if we force that the sum of the vertex values of functions living on the incidents edges of $v_m$ is zero. The following vertex conditions are known as the anti-standard vertex conditions,
\begin{equation} \label{anti}
\begin{cases}
\partial \varphi(x_i)=\partial \varphi (x_j), \quad x_i,x_j\in v_m \quad (\text{continuity condition}), \\
\sum \limits_{x_i \in v_m}\varphi(x_i)=0 \hspace{1.0in} \quad (\text{balance condition}).
\end{cases}
\end{equation}
And the quadratic form of a metric graph equipped with anti-standard vertex conditions is defined by the expression $h[\varphi]:=\int_\Gamma |\varphi'(x)|^2 \,dx$ defined on the domain.
$$D(h)=\left \{ \varphi \in H^1(\Gamma): \sum \limits_{x_i \in v_m}\varphi(x_i)=0\right \}.$$
Another type of vertex condition that decouples the graph into disjoints interval is the \textbf{Neumann conditions}. For these conditions, the derivatives of functions living on incident edges are zero at each vertex, and no conditions are assumed on the values of functions. That is, $$\varphi_e'(v)=0$$
Equivalently, this means that the derivatives of functions vanish at each endpoint of intervals in a metric graph $\Gamma$. The Neumann conditions corresponds to $\delta'$-type conditions if we let $\alpha'_m=\infty$ in \eqref{delta'}. The quadratic form of a metric graph equipped with such conditions is defined on the function from $H^1(\Gamma)$, given by the following expression.
\begin{equation} \label{qneumann}
h[\varphi]:=\int_\Gamma |\varphi'(x)|^2 \,dx.
\end{equation}
We remark that at the vertex of degree one, the anti-standard vertex conditions coincide with Dirichlet conditions, and the standard condition simplifies to Neumann conditions. Provided the length of an edge is preserved, the eigenvalues of operators corresponding to the above mentioned vertex conditions are independent of the choice of parametrization of edges in $\Gamma$. We further remark that these operators are self-adjoint and semi-bounded, and the underlying metric graph is compact, which guarantees that the spectrum of Laplacians consists of a sequence of real eigenvalues of finite multiplicity and can be listed as follows.
$$\lambda_1 \leq \lambda_2 \leq \lambda_3 \leq \cdots,$$
and each eigenvalue is repeated according to its multiplicity, and the corresponding eigenfunctions form an orthonormal basis of the Hilbert space $L^2(\Gamma)$. The eigenvalues and the corresponding eigenfunctions depend on the length, vertex conditions and the topological structure of the underlying metric graph \cite{BKKM,RS,BK1, KN}. For this reason, we will write $\lambda_k=\lambda_k(\Gamma)$ to reflect the spectrum's dependence on the quantum graph- a metric graph, differential operator and the vertex conditions. \\
Generally, the explicit computation of spectrum is challenging to calculate, except in the case of a few simple
graphs, impossible to calculate explicitly. Because the relevant secular equation is transcendental to a very high degree. The zeros of these secular equations are the eigenvalues of a quantum graph, and it is not straightforward to find reliable analytical and numerical solutions to such equations. Therefore, the interesting question is whether we can find universal bounds on the eigenvalues depending on a few simple parameters of the underlying metric graph. To obtain these bounds and the corresponding extremal graphs, a better understanding of how the geometry of the graph influences the eigenvalues is required. Therefore, we attempt to understand how the eigenvalues change subject to specific changes in the graph structure. For example, one may be interested to know if a particular graph minimizes or maximizes a given eigenvalue among all graphs with specific fixed geometric quantities such as length, number of edges, diameter, Betti number. To answer such questions, one needs to be able to make comparisons between two or more graphs. For this, we need to make some fundamental changes in the geometry of the underlying graph, such as lengthening (or shortening) an edge, gluing together vertices to obtain a new graph etc. These changes are made so that the impact on the eigenvalues are predictable; such alterations in the graph were referred to as surgery in \cite{BKKM}.\\
A standard method in obtaining eigenvalue estimates is based on variational arguments. The variational argument is based on comparing the Rayleigh quotients on suitable finite-dimensional subspaces. The Rayleigh quotient $R(\varphi)$ associated with Laplacian $L=-\dfrac{d^2}{dx^2}$ is defined as
\begin{equation}
R(\varphi)=\frac{h[\varphi]}{||\varphi||^2}
\end{equation}
on the space $L^2(\Gamma)$, the inner product is given by
\begin{equation}
<\varphi,\psi>=\int_{\Gamma} \varphi(x)\overline{\psi(x)}dx.
\end{equation}
Here, $h[\varphi]=< L \varphi, \varphi >$ is the quadratic form of the operator $L$. It is evidenced that computing the Rayleigh quotient for certain trial functions from the domain of the quadratic form leads to estimates for eigenvalues, and by minimizing the Rayleigh quotient, the eigenvalues and eigenfunctions of an operator can be obtained. Thus, the Rayleigh quotient is related to eigenvalues via the following min-max principle,
\begin{align} \label{minmax}
\lambda_k(\Gamma)= \min_{\underset{\dim(X) =k }{X \subset D\left(h\right) }}{\max_{0 \neq \varphi \in X}{\frac{h[\varphi]}{||\varphi||^2}}}.
\end{align}
If a $k$-dimensional subspace realises the minimum $\lambda_k(\Gamma)$ in \eqref{minmax}, we call it a minimising subspace for $\lambda_k(\Gamma)$.
Let $H$ and $\tilde{H}$ be two self-adjoint operators with discrete spectrum and $h$ and $\tilde{h}$ be their corresponding semi-bounded from below quadratic forms. We say that $\tilde{h}$ is a \textit{positive rank-n perturbation} of $h$ if $\tilde{h}=h$ on some $Y\subset_n \mathrm{dom}(h)$ and either $Y=\mathrm{dom}(\tilde{h})\subset_n \mathrm{dom(h)}$ or $\tilde{h}\geq h$ with $\mathrm{dom}(\tilde{h})=\mathrm{dom}(h)$. Here, the symbol $\subset_n$ denotes the subspace such that the quotient space $\mathrm{dom}(h)/Y$ is $n$-dimensional. \\
The following theorem provides the interlacing with equality characterisation between the eigenvalues of the two forms where one form is the rank $1$-perturbation of the other.
\begin{theorem}\cite[\textcolor{red}{Theorem 4.3}]{BKKM}\label{berk1}. Let $H$ and $\tilde{H}$ be two self-adjoint operators with discrete spectrum and $h$ and $\tilde{h}$ be their corresponding semi-bounded from below quadratic forms. If form $\tilde{h}$ is a positive rank-1 perturbation of the form $h$ then, eigenvalues of $H$ and $\tilde{H}$ satisfy
$$
\lambda_k(\Gamma)\leq \lambda_k(\tilde{\Gamma})\leq \lambda_{k+1}(\Gamma)\leq \lambda_{k+1}(\tilde{\Gamma}), \quad k\geq 1.
$$
Moreover, if $\lambda$ is a common eigenvalue of $H$ and $\tilde{H}$, then their respective multiplicities $m$ and $\tilde{m}$ differ by at most $1$ and the dimension of the intersection of their $\lambda$-eigenspaces is $\mathrm{min}(m,\tilde m)$.
\end{theorem}
We will use the following notation to represent the metric graph equipped with respective vertex conditions.
\begin{center}
\begin{tabular}{ |c|c| }
\hline
Metric Graph & Vertex Conditions \\
\hline
$\Gamma^s$ & Standard \\
$\Gamma^a$ & Anti-Standard \\
$\Gamma^d$ & Dirichlet \\
$\Gamma^n$ & Neumann \\
$\Gamma^\delta$ & $\delta$ \\
$\Gamma^{\delta'}$ & $\delta'$ \\
\hline
\end{tabular}
\end{center}
Furthermore, if arbitrary vertex conditions are assumed at all other vertices of $\Gamma$ except $v_1$ where Dirichlet conditions are imposed, we will use the notation $\lambda_k(\Gamma;v_1)$ to represent the eigenvalues for this graph. Moreover, if more than one vertex is equipped with Dirichlet condition, we will denote the eigenvalues by $\lambda_k(\Gamma;D)$, where $D$ is the subset of vertex set $V$ containing all Dirichlet vertices. We will only consider the case when the degree one vertex are endowed with Dirichlet conditions. \\
Let $\Gamma^n$ be the metric graph obtained from $\Gamma^a$ by imposing Neumann conditions at a vertex $v_0$. Then the corresponding quadratic form for $\Gamma^a$ and $\Gamma^n$ are defined by the same expression \eqref{qneumann}. However, the domain $D(h^a)$ is a co-dimension one subspace of $D(h^n)$. Thus by \textcolor{blue}{Theorem} \eqref{berk1},
\begin{equation*}
\lambda_k(\Gamma^n) \leq \lambda_k(\Gamma^a) \leq \lambda_{k+1}(\Gamma^n) \leq \lambda_{k+1}(\Gamma^a).
\end{equation*}
Moverover, for any Neumann graph, graph equipped with Neumann conditions, obtained from $\Gamma^a$ the domain $D(h^a)$ is a co-dimension $|V|$ subspace of $D(h^n)$, then applying the \textcolor{blue}{Theorem} \eqref{berk1} we obatined the following inequality.
\begin{equation*}
\lambda_k(\Gamma^n) \leq \lambda_k(\Gamma^a) \leq \lambda_{k+|V|}(\Gamma^n).
\end{equation*}
Let $\Gamma^{\delta'}$ be the graph obtained from the Neumann graph $\Gamma^n$ by imposing the $\delta'$-type conditions at vertex $v_0$ in $\Gamma^n$ with strength $\alpha'_0 >0 $, then the domains of the quadratic forms are same, $D(h^n)=D(h^{\delta'})$, and
$$h^n[\varphi]=\int_\Gamma |\varphi'(x)|^2 \,dx \leq \int_\Gamma |\varphi'(x)|^2 \,dx+ \frac{1}{\alpha'_0} \left|\sum \limits_{x_i \in v_0}\varphi(x_i)\right|^2=h^{\delta'}[\varphi].$$
Since $D(h^a) \subset_1 D(h^n)$ and on this domain $h^n=h^{\delta'}$. Thus, $h^{\delta'}$ is a positive rank-$1$ perturbation of the form $h^n$. Therefore,
$$\lambda_k(\Gamma^n) \leq \lambda_k(\Gamma^{\delta'}) \leq \lambda_{k+1}(\Gamma^n).$$
Similarly if $\alpha'_0 <0 $, then
$$h^{\delta'}[\varphi]=\int_\Gamma |\varphi'(x)|^2 \,dx+ \frac{1}{\alpha'_0} \left|\sum \limits_{x_i \in v_0}\varphi(x_i)\right|^2 \leq \int_\Gamma |\varphi'(x)|^2 \,dx=h^{n}[\varphi],$$
and $D(h^a) \subset_1 D(h^{\delta'})$ and on this domain $h^n=h^{\delta'}$. Thus, $h^{n}$ is a positive rank-$1$ perturbation of the form $h^{\delta'}$. Therefore,
$$\lambda_k(\Gamma^{\delta'}) \leq \lambda_k(\Gamma^{n}) \leq \lambda_{k+1}(\Gamma^{\delta'}).$$
An equality between a Neumann and a $\delta'$- eigenvalue is possible only if the eigenspace of
the Neumann eigenvalue contains a function which satisfy the balance condition at $v_0$, or, equivalently, the eigenspace of the $\delta'$-eigenvalue contains a function whose derivative vanishes at $v_0$.\\
Let $\Gamma^n$ be the Neumann graph and obtain the graph $\Gamma^d$ by prescribing the Dirichlet conditions at a vertex $v_0$ in $\Gamma^n$. Since the quadratic forms of $\Gamma^n$ and $\Gamma^d$ are defined by the same expression and agree on the domain $D(h^d) \subset_1 D(h^n)$. Thus,
$$\lambda_k(\Gamma^n) \leq \lambda_k(\Gamma^d;v_0) \leq \lambda_{k+1}(\Gamma^n).$$
For the Dirichlet graph $\Gamma^d$ and the graph $\Gamma^a$ obtained from $\Gamma^d$ by imposing the anti-standard vertex condition at $v_0$ in $\Gamma^d$, the respective forms $h^d$ and $h^a$ coincides on $D(h^d) \subset_1 D(h^a)$. Hence,
$$\lambda_k(\Gamma^a) \leq \lambda_k(\Gamma^d) \leq \lambda_{k+1}(\Gamma^a).$$
The equality is attained when the eigenspace of the Dirichlet eigenvalue contains a function such that its derivatives are continuous at $v_0$, or, equivalently, the eigenspace of the eigenvalue of $\Gamma^a$ contains a function that vanishes at $v_0$. \\
Moreover, for the Dirichlet graph $\Gamma^d$ and the graph $\Gamma^{\delta'}$ obtained from $\Gamma^d$ by imposing the $\delta'$-type condition at $v_0$ in $\Gamma^d$, the respective forms $h^d$ and $h^{\delta'}$ coincides on $D(h^d) \subset_1 D(h^{\delta'})$. Hence,
$$\lambda_k(\Gamma^{\delta'}) \leq \lambda_k(\Gamma^d) \leq \lambda_{k+1}(\Gamma^{\delta'}).$$
Let $\Gamma^{\delta'}$ be finite compact metric graph with arbitrary self-adjoint vertex conditions at each vertex, except $v_0$ where $\delta'$-condition are imposed with strength $\alpha'_{0}$, and obtain a graph $\Gamma^{\delta}$ from $\Gamma^{\delta'}$ by imposing $\delta$-condition at $v_0$ with strength $\alpha_{0}$. Let $h^{\delta'}$ and $h^{\delta}$ denote the quadratic forms with domain $D(h^{\delta'})$ and $D(h^{\delta})$, respectively. Then the domain $D(h^{\delta})$ is a subspace of $D(h^{\delta'})$, and on $D(h^{\delta})$, we have
\begin{align*}
h^{\delta}-h^{\delta'}&=\alpha_{0}|\varphi(v_0)|^2- \frac{1}{\alpha'_{0}}\left| \sum\limits_{x_i \in v_0} \varphi(x_i) \right|^2 \\
&= \alpha_{0}|\varphi(v_0)|^2- \frac{1}{\alpha'_{0}}\left| d_{v_0} \varphi(v_0) \right|^2 \\
&= |\varphi(v_0)|^2 \left(\alpha_{0} - \frac{d_{v_0}^2}{\alpha'_{0}}\right)
\end{align*}
Now, if $\alpha_{0} - \frac{d_{v_0}^2}{\alpha'_{0}} \geq 0$, then $h^{\delta} \geq h^{\delta'}$.\\
Moreover, if $\alpha_{0} = \frac{d_{v_0}^2}{\alpha'_{0}}$, then $h^{\delta}=h^{\delta'}$ on $D(h^{\delta}) \subset_1 D(h^{\delta'})$, and by rank one perturbation the following interlacing inequalities holds
$$\lambda_k(\Gamma^{\delta'}) \leq \lambda_k(\Gamma^{\delta}) \leq \lambda_{k+1}(\Gamma^{\delta'}). $$
For our later purpose, we provide the explicit eigenvalues of an interval $I$ and the path graph $P$.
\begin{example}
Let $I$ be a metric graph consist of single edge of length $\ell$, the eigenvalues for the operator $-\frac{d}{dx^2}$ on an interval $[0,\ell]$ are given by,
\begin{align*}
\lambda_k(I,v_1,v_2)&=\frac{k^2\pi^2}{\ell^2}, \quad k \geq 1, \quad \text{Dirichlet Conditions}\\
\lambda_{k+1}(I^s)&=\frac{k^2\pi^2}{\ell^2}, \quad k \geq 0, \quad \text{Standard Conditions} \\
\lambda_{k+1}(I^s)&=\frac{k^2\pi^2}{\ell^2}, \quad k \geq 0, \quad \text{Neumann Conditions} \\
\lambda_k(I^a)&=\frac{k^2\pi^2}{\ell^2}, \quad k \geq 1, \quad \text{Anti-standard Conditions}
\end{align*}
\end{example}
\begin{example}
Let $P$ be the path graph with standard vertex condition prescribed at each vertex. Due to standard vertex condition the eigenvalue problem for Laplacian $-\frac{d^2}{dx_e^2}$ on the path $P$ can be identified with eigenvalue problem on an interval of length $L(P)$ for the operator $-\frac{d^2}{dx^2}$ subject to Neumann boundary condition prescribed at both endpoints. Thus the eigenvalues of path graph $P$ are given by
$$\lambda_{k+1}(P^s)=\frac{k^2 \pi^2}{L(P)^2}, \quad k \geq 0.$$
\end{example}
\begin{example}
The eigenvalues for $-\frac{d^2}{dx_e^2}$ on the path graph $P$ of length $L(P)$ with anti standard vertex conditions imposed at each vertex coincides with the eigenvalues of $-\frac{d^2}{dx^2}$ on an interval of length $L(P)$ with Dirichlet conditions at both endpoints. Thus the eigenvalues of $P$ are given by
$$\lambda_k(P^a)=\frac{k^2\pi^2}{L(P)^2}, \quad k \geq 1.$$
\end{example}
\section{Surgical Transformations}
In this section, we present the tools that make some fundamental changes in the geometry of the underlying metric graph, such as lengthening (or shortening) an edge, gluing together vertices; these tools will alter the given metric graph into some other graph for which the spectrum is known. We discuss the effects of surgery principles on the spectrum of a Laplacian on a finite compact quantum graph with arbitrary self-adjoint vertex conditions ($\delta$-type, $\delta'$-type). For the metric graph equipped with $\delta$-conditions, these surgical transformations and their effect on the spectrum were studied in \cite{BKKM}. Here, we would like to investigate how some surgical operations affect the spectrum of a given metric graph $\Gamma$ and whether we can use a set of surgical operations to transform the graph into a simpler one by precisely predicting the corresponding change in the spectrum.\\
\subsection{Operation changing vertex conditions}
In this part of a section, we study the eigenvalue's dependence on the vertex conditions. The operation changing the vertex conditions does not affect the total length of a metric graph. However, it changes its connectivity in case of gluing of vertices. First, we study the dependence of eigenvalues of Laplacian on the $\delta'$-coupling parameter $\alpha'_m$. This operation of changing the $\delta$-potential at a vertex $v$ is well understood. Let $\Gamma^\delta_{\tilde{\alpha}}$ be obtained from $\Gamma^\delta_{\alpha}$ by changing the interaction strength at a vertex $v$ from $\alpha$ to $\tilde{\alpha}$. If $\infty < \alpha \leq \tilde{\alpha} \leq \infty$, then the following interlacing inequalities were established in \cite[\textcolor{blue}{ Theorem 3.4}]{BKKM} and \cite[\textcolor{blue}{Theorem 3.1.8}]{BK1}.
\begin{equation} \label{interlacing1}
\lambda_k(\Gamma^\delta_\alpha) \leq \lambda_k(\Gamma^\delta_{\tilde{ \alpha}}) \leq \lambda_k(\Gamma^\delta;v) \leq \lambda_{k+1}(\Gamma^\delta_\alpha)
\end{equation}
The repeated application of \eqref{interlacing1} leads to the following inequalities.
\begin{equation} \label{interlacing2}
\lambda_k(\Gamma^\delta_\alpha) \leq \lambda_k(\Gamma^d) \leq \lambda_{k+|V|}(\Gamma^\delta_\alpha)
\end{equation}
Let $\Gamma$ and $\tilde{ \Gamma}$ be two quantum graphs with the same underlying metric graph but with different strengths $\alpha'_m$ in the vertex conditions. Then the domains of the quadratic forms for $\Gamma$ and $\tilde{\Gamma}$ are the same. That is, $D(h)=D(\tilde{h})$. Thus, any $\varphi \in D(h)$ which minimizes the Rayleigh quotient for $\lambda_k(\Gamma)$ will also be in $D(\tilde{h})$ and will also minimize the Rayleigh quotient for $\lambda_k(\tilde{\Gamma})$.
The following theorem shows the spectrum's dependence on the vertex parameter $\alpha'_m$.
\begin{theorem} \label{strength}
Let $\Gamma$ be finite, compact and connected metric graph with local, self-adjoint vertex condition at all vertices of $\Gamma$ except $v_0$, where $\delta'$-condition is imposed with strength $\alpha_0'$. Let $\tilde{\Gamma}$ be a graph obtained from $\Gamma$ by changing the strength from $\alpha_0'$ to $\tilde{\alpha}'$, and let $\Gamma_0$ be obtained from $\Gamma$ by imposing anti-standard vertex conditions at $v_0$.
\begin{enumerate}
\item If $0 < \alpha_0' < \tilde{\alpha}_0'$, or $ \alpha'_0 < \tilde{\alpha}_0' < 0$, then
\begin{equation} \label{interlacing3}
\lambda_k(\tilde{ \Gamma}) \leq \lambda_k(\Gamma) \leq \lambda_k(\Gamma_0) \leq \lambda_{k+1}(\tilde{ \Gamma}).
\end{equation}
\item If $\alpha_0' < 0$ and $\tilde{\alpha}_0' > 0$, then
\begin{equation} \label{interlacing4}
\lambda_k( \Gamma) \leq \lambda_k(\tilde{\Gamma}) \leq \lambda_k(\Gamma_0) \leq \lambda_{k+1}( \Gamma).
\end{equation}
\end{enumerate}
\end{theorem}
\begin{proof}
Let $h, \tilde{h} \ \text{and} \ h_0$ be the quadratic forms of corresponding Laplacian associated to graphs $\Gamma,\tilde{\Gamma} \ \text{and} \ \Gamma_0 $ respectively, then we have
\begin{equation*}
h[\varphi]=h_0[\varphi]+\frac{1}{\alpha_0'} \left|\sum \limits_{x_i \in v_0}\varphi(x_i)\right|^2 \text{and} \ \ \tilde{h}[\varphi]=h_0[\varphi]+\frac{1}{\tilde{\alpha}_0'} \left|\sum \limits_{x_i \in v_0}\varphi(x_i)\right|^2,
\end{equation*}
on the appropriate sub-spaces of $H^1(\Gamma)=\bigoplus\limits_{e \in E} H^1(e)$. Since the underlying metric graph is the same and the domain of the quadratic form does not depend on strength, therefore
$$D(h)=D(\tilde{h})\ \text{and} \ D(h_0)=\left \{\varphi \in D(h): \sum \limits_{x_i \in v_0}\varphi(x_i)=0 \right \},$$
and
\begin{equation*}
h[\varphi]-\tilde{h}[\varphi]=\left|\sum\limits_{e_{i} \in E_v} \varphi_i(v)\right|^2\left(\frac{1}{\alpha_0'}-\frac{1}{\tilde{\alpha}_0'}\right).
\end{equation*}
Since $0 < \alpha_0' < \tilde{\alpha}_0'$, or $ \alpha'_0 < \tilde{\alpha}_0' < 0 \implies \frac{1}{\alpha_0'}-\frac{1}{\tilde{\alpha}_0'} > 0$. Thus $h \geq \tilde{h}$ for all $\varphi$, and the inequality follows from the min-max description of eigenvalues. Furthermore, we have
$$h[\varphi] = h_0[\varphi] \ \text{and} \ \tilde{h}[\varphi] = h_0[\varphi] \quad \text{for all} \quad \varphi \in D(h_0).$$
The domain $D(h)$ is larger than the domain $D(h_0)$ and the corresponding quadratic forms $h$ and $h_0$ agree on $D(h_0)$. Therefore, minimizing the Rayleigh quotient over smaller space results in larger eigenvalues. The last inequality follows from the fact that $D(h_0)$ is a subspace of $D(h)$ of co-dimension one.\\
If $\alpha_0' < 0$ and $\tilde{\alpha}_0' > 0$, then $\frac{1}{\alpha_0'} <0 $ and $\frac{1}{\tilde{ \alpha}_0'} >0 \implies \left(\frac{1}{\alpha_0'}-\frac{1}{\tilde{\alpha}_0'}\right)<0 $. Thus, $h \leq \tilde{h}$ and hence by min-max principle, $\lambda_k(\Gamma) \leq \lambda_k(\tilde{\Gamma}).$
\end{proof}
Equality between an eigenvalue of $\Gamma$ and $\Gamma_0$ is only possible if the eigenspace of $\lambda_k(\Gamma)$ contains an eigenfunction $\varphi $ such that at vertex $v_0$ the sum of components of $\varphi$ living on the incident edges to $v_0$ is zero.
\begin{definition} {\textbf{(Gluing).}}
The gluing of vertices is a surgical transformation in which a graph $ \tilde{\Gamma}$ is obtained by identifying some vertices $v_1,v_2,\cdots,v_n$ of a graph $\Gamma$. If the strengths $\alpha_m$ for $m=1,2,\cdots,n$ were associated to the vertices $v_1,v_2,\cdots,v_n$, then the glued vertex $v_0$ will be specified with strength $\alpha_0=\sum\limits_{m=1}^{n} \alpha_m$.
\end{definition}
The reverse operation to gluing of vertices is named as splitting, in which a vertex $v_0$ of $\tilde{\Gamma}$ is cut into $n$ descendent vertices $v_1,v_2,\cdots,v_n$ to obtain a new graph $\Gamma$. The obtained graph $\Gamma$, using this surgery transformation, is not unique in general, as the incident edges to $v_0$ may be assigned to descendent vertices in several different ways. Furthermore, if $\varphi$ is a function on $\tilde{\Gamma}$ satisfying $\delta$-type conditions at vertex $v_0$, one can lift this function on the graph $\Gamma$ by assigning the strengths $$\alpha_m=-\dfrac{\sum\limits_{e \sim v_m}\partial \varphi_e(v_0)}{\varphi(v_0)},$$ to the descendent vertices $v_m$ for $m=1,2,\cdots, n$ in $\Gamma$. If the function $\varphi$ vanishes at $v_0$, then the descendent vertices will be endowed with Dirichlet conditions \cite{BKKM}.
Similarly, if the graph $\tilde{\Gamma}$ is equipped with $\delta'$-type conditions then any function on $\tilde{\Gamma}$ can be lifted to $\Gamma$ by specifying the strengths $\alpha'_m$
$$\alpha'_m=-\dfrac{\sum\limits_{e \sim v_m} \varphi_e(v_0)}{\partial \varphi(v_0)},$$ to the descendent vertices $v_m$ for $m=1,2,\cdots, n$ in $\Gamma$.\\
Let $\tilde{\Gamma}^{\delta}$ be the graph obtained by gluing two vertices of $\Gamma^{\delta}$, then the following interlacing inequalities can be easily proved using variational principles. (c.f., e.g., \cite[\textcolor{blue}{Theorem 3.4}]{BKKM}, \cite[ \textcolor{blue}{Theorem 3.1.10}]{BK1}).
\begin{equation}\label{dinterlac}
\lambda_k(\Gamma^\delta) \leq \lambda_k(\tilde{ \Gamma}^\delta) \leq \lambda_{k+1}(\Gamma^\delta) \leq \lambda_{k+1}(\tilde{ \Gamma}^\delta).
\end{equation}
For each metric graph there exist a unique graph obtained by identifying all the vertices to a single vertex, such a graph is known as flower graph. Note that, there is a one-to-one correspondence between a unique flower graph and the class of graphs with fix total number of edges, total length of metric graph and length of each edge, $(|E|,\mathcal{L},\ell_e)$. Let $\Gamma_f^{\delta}$ be the flower graph corresponding to metric graph $\Gamma^{\delta}$, then the repeated applications of \eqref{dinterlac} leads to the following inequalities.
\begin{equation}\label{dflower}
\lambda_{k-|V|+1}(\Gamma_{f}^\delta) \leq \lambda_k(\Gamma^\delta) \leq \lambda_k( \Gamma_{f}^\delta) \leq \lambda_{k+|V|-1}(\Gamma^\delta).
\end{equation}
The gluing operation does not depend on the sign of the strengths $\alpha_m$ for the $\delta$-graph. However, for the $\delta'$-graph we have the following theorem, which shows that the effect of gluing on the spectrum depends on the sign of the strengths $\alpha'_m$.
\begin{theorem}\cite[\textcolor{blue}{Theorem 4.2}]{RS} \label{roh}.
Let $\Gamma^{\delta'}$ be the metric graph equipped with local, arbitrary self-adjoint vertex conditions at each vertex, except $v_1$ and $v_2$, where $\delta'$-type conditions are imposed with strengths $\alpha'_1$ and $\alpha'_2$. Let $\tilde{\Gamma}^{\delta'}$ be graph obtained by gluing vertices $v_1$ and $v_2$ of $\Gamma^{\delta'}$, producing new vertex $v_0$ with strength $\alpha'_0=\alpha'_1+\alpha'_2$. Then for all $k \geq 1$ the following assertions hold.
\begin{enumerate}
\item If $\alpha'_1, \alpha'_2 >0$ then $\lambda_k(\tilde{\Gamma}^{\delta'}) \leq \lambda_k(\Gamma^{\delta'}).$
\item If $\alpha'_1, \alpha'_2 < 0$ then $\lambda_k(\Gamma^{\delta'}) \leq \lambda_k(\tilde{\Gamma}^{\delta'}).$
\item If $\alpha'_1 \cdot \alpha'_2 < 0$ and $\alpha_0 >0 $ then $\lambda_k(\Gamma^{\delta'}) \leq \lambda_k(\tilde{\Gamma}^{\delta'}).$
\item If $\alpha'_1 \cdot \alpha'_2 <0 $ and $\alpha'_0 <0 $ then $\lambda_k(\tilde{\Gamma}^{\delta'}) \leq \lambda_k(\Gamma^{\delta'}).$
\item If $\alpha'_1 \cdot \alpha'_2 <0 $ and $\alpha'_0 =0 $ then $\lambda_k(\Gamma^{\delta'}) \leq \lambda_k(\tilde{\Gamma}^{\delta'}).$
\item If $\alpha'_1 \cdot \alpha'_2=0$ then $\lambda_k(\tilde{\Gamma}^{\delta'}) \leq \lambda_k(\Gamma^{\delta'}).$
\end{enumerate}
\end{theorem}
\begin{remark} \label{d'flower}
\noindent Let $\Gamma^{\delta'}$ be the metric graph and $\Gamma^{\delta'}_f$ be its corresponding flower graph. Then the following statements hold.
\begin{enumerate}
\item If $\alpha'_m > 0 $ for all $m=1,2,\cdots, |V|$ in $\Gamma^{\delta'}$, then
$$\lambda_k(\Gamma^{\delta'}_f)\leq \lambda_k(\Gamma^{\delta'}).$$
\item If $\alpha'_m < 0$ for all $m=1,2,\cdots, |V|$ in $\Gamma^{\delta'}$, then
$$\lambda_k(\Gamma^{\delta'})\leq \lambda_k(\Gamma^{\delta'}_f).$$
\end{enumerate}
\end{remark}
\subsection{Operations Increasing the total length}
One of the most important quantities of a metric graph is its total length, which contains information about the eigenvalues. Naturally, any change in the length of a metric graph will shift the spectrum of the respective metric graph in some direction. A more profound study on how the change in the length of the metric graph effect the spectrum can be found in \cite{EJ, RS,BKKM}. For the standard vertex condition, the scaling of a metric graph by a factor of $t>1$ reduce the eigenvalues by $t^2$. If $\mathcal{L}(\tilde{\Gamma}^s)=t \mathcal{L}(\Gamma^s)$, then
$$\lambda_k(\tilde{ \Gamma}^s)=\frac{\lambda_k(\Gamma^s)}{t^2}.$$
Moreover, for general $\delta$-conditions, if the metric graph $\tilde{\Gamma}^{\delta}$ is obtained by scaling each edge of the graph $\Gamma^{\delta}$ by a factor of $t >0$ and coupling parameters $\alpha_m$ at each vertex are scaled with the factor $\frac{1}{t}$, then
$$\lambda_k(\tilde{\Gamma}^{\delta})= \frac{\lambda_k(\Gamma^{\delta})}{t^2}.$$
Since the quadratic form and its domain depend on the graph's length, the following proposition show that increasing any edge's length of $\Gamma^{\delta'}$ lowers all eigenvalues.
\begin{proposition} \label{length1}
Let $\tilde{\Gamma}^{\delta'} $ be the metric graph obtained from $\Gamma^{\delta'}$ by scaling one of the edge say $e_0$ of $\Gamma^{\delta'}$ by a factor $t > 1$, then
\begin{equation}
\lambda_k(\tilde{\Gamma}^{\delta'}) \leq \lambda_k(\Gamma^{\delta'}).
\end{equation}
\end{proposition}
\begin{proof}
We parametrize each edge $e_j$ by $[0,\ell_j]$,
let the function $\psi(x)$ minimizes the expression \eqref{qdelta'} for the $\Gamma^{\delta'}$. Let $\varphi(x) \in D(\tilde{h}^{\delta'})$ be defined by
\begin{equation}
\varphi(x)=\begin{cases}
\psi \left(\frac{x}{t}\right) \quad \text{for} \ x \in e'_0=[0,t \ell_0] \\
\psi(x) \quad \text{ for } \ x \in \{e\in E:e \neq e_0 \}
\end{cases}
\end{equation}
\begin{align*}
\int_{e'_0}|\varphi(x)|^2dx&=\int^{tl_0}_{0}|\varphi(x)|^2dx=\int^{tl_0}_{0}\left|\psi \left(\frac{x}{t}\right)\right|^2dx=t\int_{e_0}|\psi(x)|^2dx,
\\ &\implies \int_{e'_0}|\varphi(x)|^2dx > \int_{e_0}|\psi(x)|^2dx,\\
\int_{e'_0}|\varphi'(x)|^2dx&=\int^{tl_0}_{0}|\varphi'(x)|^2dx=\frac{1}{t^2}\int^{tl_0}_{0} \left|\psi'\left(\frac{x}{t}\right)\right|^2dx=\frac{1}{t}\int_{e_0}|\psi'(x)|^2dx,\\
& \implies \int_{e'_0}|\varphi'(x)|^2dx < \int_{e_0}|\psi'(x)|^2dx.
\end{align*}
We have constructed $\varphi(x)$ in such a way that the values at the vertices are unchanged $\psi(x_i)=\varphi(x_i)$, and hence
$$\sum\limits_{m=1}^{|V|}\frac{1}{\alpha'_m} \left|\sum\limits_{x_i \in v_m} \psi(x_i)\right|^2=\sum\limits_{m=1}^{|V|}\frac{1}{\alpha'_m} \left|\sum\limits_{x_i \in v_m} \varphi(x_i)\right|^2$$
\begin{align*}
R(\psi)&=\frac{\sum\limits_{e \in E}\int_{e}|\psi'(x)|^2dx+\sum\limits_{m=1}^{|V|}\frac{1}{\alpha'_m} \left|\sum\limits_{x_i \in v_m} \psi(x_i)\right|^2}{\sum\limits_{e \in E}\int_{e}|\psi(x)|^2dx} \\
&=\frac{\sum\limits_{e \in E:e \neq e_0}\int_{e}|\psi'(x)|^2dx+\int_{e_0}|\psi'(x)|^2dx+\sum\limits_{m=1}^{|V|}\frac{1}{\alpha'_m} \left|\sum\limits_{x_i \in v_m} \psi(x_i)\right|^2}{\sum\limits_{e \in E:e \neq e_0}\int_{e}|\psi(x)|^2dx+\int_{e_0}|\psi(x)|^2dx} \\
&\geq \frac{\sum\limits_{e \in E:e \neq e_0}\int_{e}|\psi'(x)|^2dx+\frac{1}{t}\int_{e_0}|\psi'(x)|^2dx+\sum\limits_{m=1}^{|V|}\frac{1}{\alpha'_m} \left|\sum\limits_{x_i \in v_m} \psi(x_i)\right|^2}{\sum\limits_{e \in E:e \neq e_0}\int_{e}|\psi(x)|^2dx+t \int_{e_0}|\psi(x)|^2dx} \\
&= \frac{\sum\limits_{e \in E:e \neq e_0}\int_{e}|\varphi'(x)|^2dx+\int_{e'_0}|\varphi'(x)|^2dx+\sum\limits_{m=1}^{|V|}\frac{1}{\alpha'_m} \left|\sum\limits_{x_i \in v_m} \varphi(x_i)\right|^2}{\sum\limits_{e \in E:e \neq e_0}\int_{e}|\varphi(x)|^2dx+ \int_{e'_0}|\varphi(x)|^2dx}=R(\varphi)
\end{align*}
Since $R(\psi) \geq R(\varphi)$, so by max-min description of eigenvalues we obtain the required result.
Note that $\varphi(x) \not\in D(L^{\delta'}),$ but $\varphi(x)\in D(h^{\delta'})$.
\end{proof}
The result of the above theorem still holds when each edge $e$ of the graph $\Gamma^{\delta'}$ is scaled by a factor $t_e > 1$. This show that the spectrum of a quantum graph depends on its total length. Increasing the graph's total length also decreases the eigenvalues, so one can shift the eigenvalues in some direction by scaling the graph's total length. When $0<t<1$, the above result is reversed. The following theorem describes the relationship between the Rayleigh quotients of the graphs with total lengths $\mathcal{L}$ and $t\mathcal{L}$.
\begin{theorem} \label{length2}
Let $\Gamma^{\delta'} $ be the metric graph with total length $\mathcal{L}(\Gamma^{\delta'})$. Obtain a graph $ \tilde{\Gamma}^{\delta'}$ with length $t \mathcal{L}(\Gamma^{\delta'})$ by scaling each edge of $\Gamma^{\delta'}$ by $t > 0$ and scaling each strength $\alpha'_m$ by same factor, then
\begin{equation}
\lambda_k(\tilde{\Gamma}^{\delta'})=\frac{1}{t^2} \lambda_k(\Gamma^{\delta'}).
\end{equation}
\end{theorem}
\begin{proof}
Let $\psi(x)$ with $||\psi||_{L^2(\Gamma^{\delta'})}=1$ minimizes the Rayleigh quotient for graph $\Gamma^{\delta'}$. Then the function defined by $\varphi(x)=\frac{1}{\sqrt{t}}\psi(\frac{x}{t})$ with $||\varphi||_{L^2(\tilde{\Gamma}^{\delta'})}=1$ minimizes the Rayleigh quotient for the graph $\tilde{\Gamma}^{\delta'}$.
Let $e_j=[0,\ell_j]$ and $\tilde{e_j}=[0,t\ell_j]$ be the parametrization of edges in $\Gamma^{\delta'}$ and $\tilde{\Gamma}^{\delta'}$, respectively.
\begin{align*}
\tilde{R}(\varphi(\tilde{x}))&=\int_{\tilde{\Gamma}^{\delta'}}|\varphi'(\tilde{x})|^2d\tilde{x}+\sum\limits_{m=1}^{|V|}\frac{1}{\tilde{\alpha}'_m} \left|\sum\limits_{\tilde{x}_i \in \tilde{v}_m} \varphi(\tilde{x}_i)\right|^2 \\
&=\sum\limits_{j=1}^{|\tilde{E}|}\int_{\tilde{e_j}}|\varphi'(\tilde{ x})|^2d\tilde{x}+\sum\limits_{m=1}^{|V|}\frac{1}{\tilde{\alpha}'_m} \left|\sum\limits_{\tilde{x}_i \in \tilde{v}_m} \varphi(\tilde{x}_i)\right|^2 \\
&=\sum\limits_{j=1}^{|\tilde{E}|}\int^{t \ell_j}_{0}\left|\frac{1}{\sqrt{t}}\psi'\left(\frac{\tilde{x}}{t}\right)\frac{1}{t}\right|^2d \tilde{x}+\sum\limits_{m=1}^{|V|}\frac{1}{\tilde{\alpha}'_m} \left|\sum\limits_{\tilde{x}_i \in \tilde{v}_m} \frac{1}{\sqrt t}\psi \left(\frac{\tilde{x}_i}{t} \right)\right|^2 \\
\end{align*}
Let $x=\frac{\tilde{x}}{t} \implies t dx=d \tilde{x}$, and the limits of integration become $x=0 , x=\ell_e.$
\begin{align*}
&=\frac{1}{t^2}\left[\sum\limits_{j=1}^{|E|} \int_{e_j}|\psi'(x)|^2dx+\sum\limits_{m=1}^{|V|}\frac{1}{ {\alpha}'_m} \left|\sum\limits_{ x_i \in { v}_m}\psi(x_i)\right|^2\right] \\
\end{align*}
The expression inside the bracket is the Rayleigh quotient for the graph $\Gamma^{\delta'}$. Therefore, by min-max description of eigenvalues.
\begin{align*}
\lambda_k(\tilde{ \Gamma}^{\delta'})&=\min_{\underset{\dim(X) =k}{X \subset D\left(\tilde{h}\right) }}{\max_{0 \neq \varphi \in X}} \tilde{R}(\varphi) \\
&=\min_{\underset{\dim(X) =k }{X \subset D\left(\tilde{h}\right) }}{\max_{\varphi \in X: ||\varphi||=1}}\frac{1}{t^2}\left[\sum\limits_{j=1}^{|E|} \int_{e_j}|\psi'(x)|^2dx+\sum\limits_{m=1}^{|V|}\frac{1}{ {\alpha}'_m} \left|\sum\limits_{ x_i \in { v}_m}\psi(x_i)\right|^2\right] \\
&=\frac{1}{t^2}\left(\min_{\underset{\dim(X) =k }{X \subset D\left(h\right) }}{\max_{\psi \in X: ||\psi||=1}}\left[\sum\limits_{j=1}^{|E|} \int_{e_j}|\psi'(x)|^2dx+\sum\limits_{m=1}^{|V|}\frac{1}{{\alpha}'_m} \left|\sum\limits_{ x_i \in { v}_m}\psi(x_i)\right|^2\right]\right)\\
&=\frac{1}{t^2} \lambda_k(\Gamma^{\delta'})
\end{align*}
If we let $t$ approaches infinity, the values of a Rayleigh quotient, and thus the eigenvalues are also pushed towards zero. And letting $t$ go to zero will force the Rayleigh quotient approaches to $\pm \infty$. Furthermore, the result is still valid, if both the graphs are equipped anti-standard vertex conditions.
\end{proof}
We now consider the surgical transformation that will increase the volume of a given metric graph by attaching a new subgraph to it.
\begin{definition}
{\textbf{ (Attaching a pendant graph).}}
Given two metric graphs $\Gamma$ and $\hat{ \Gamma}$ and let $v_1 \in \Gamma$ and $w_1 \in \hat{\Gamma}$. If the metric graph $\tilde{\Gamma}$ is obtained by gluing vertices $v_1$ and $w_1$, then we speak of attaching a pendant graph $\hat{\Gamma}$ to $\Gamma$.
\end{definition}
Let $\tilde{\Gamma}^{\delta}$ be obtained by attaching a pendant graph $\hat{\Gamma}^{\delta}$ to $\Gamma^{\delta}$, then it was proved in \cite[\textcolor{blue}{Theorem 3.10}]{BKKM} that if for some $r$ and $k$, $\lambda_r(\hat{\Gamma}^{\delta}) \leq \lambda_k(\Gamma^{\delta})$ then $$\lambda_{k+r-1}(\tilde{\Gamma}^{\delta}) \leq \lambda_k(\Gamma^{\delta}).$$
For the graphs equipped with $\delta'$-type condition it was established in \cite[\textcolor{blue}{Theorem 3.5}]{RS} that attaching a pendant edge to some vertex of $\Gamma^{\delta'}$ to obtained a graph $\tilde{\Gamma}^{\delta'}$ leads to the following inequality,
$$\lambda_{k}(\tilde{\Gamma}^{\delta}) \leq \lambda_k(\Gamma^{\delta}).$$
This inequality also hold when the graphs are equipped with anti-standard vertex conditions. Moreover, the repeated application of this result also shows that the same result is true when the graph $\tilde{\Gamma}^{\delta'}$ is formed by attaching a graph $\hat{\Gamma}^{\delta}$ to $\Gamma^{\delta'}$.
\begin{theorem} \label{pendant}
Let $\Gamma$ be a finite, compact and connected metric graph, let $v_1$ be a vertex of $\Gamma$ and let $L=-\frac{d^2}{dx^2}$ be the Laplacian in $L^2(\Gamma)$ subject to arbitrary local, self-adjoint vertex conditions at each vertex $v_m \in V \setminus\{v_1\}$, with $\delta'$ vertex condition at $v_1$ , with strength $\alpha'_1$. Let $\hat{\Gamma}$ be a finite, compact and connected metric graph, with arbitrary local, self-adjoint vertex conditions at each vertex $w_m \in \hat{V} \setminus \{w_1\}$, with $\delta'$ vertex condition at $w_1$, with strength $\hat{ \alpha}'_2$. Let $\tilde{\Gamma}$ be graph formed by attaching the pendant graph $\hat{\Gamma}$ to $\Gamma$, if for some $r,k_0$, we have, $$\lambda_r(\hat{\Gamma})\leq \lambda_{k_0}(\Gamma),$$ then the following assertions hold.
\begin{enumerate}
\item If $\alpha'_1, \hat{\alpha}'_2 > 0$, then
$${\lambda_{k+r}(\tilde{\Gamma}) \leq \lambda_k(\Gamma), \quad k \geq k_0}.$$
\item If $\alpha'_1 \cdot \hat{\alpha}'_2 = 0$ then, $${\lambda_{k+r}(\tilde{\Gamma}) \leq \lambda_k(\Gamma), \quad k \geq k_0}.$$
\item If $\alpha'_1 \cdot \hat{\alpha}'_2 < 0 \ \text{and} \ \alpha'_1+ \hat{\alpha}'_2 < 0$ then,
$${\lambda_{k+r}(\tilde{\Gamma}) \leq \lambda_k(\Gamma), \quad k \geq k_0}.$$
\item If $\alpha'_1, \hat{\alpha}'_2 < 0$, then
$${\lambda_{k+r}(\tilde{\Gamma}) \geq \lambda_k(\Gamma), \quad k \geq k_0}.$$
\item If $\alpha'_1 \cdot \hat{\alpha}'_2 < 0 \ \text{and} \ \alpha'_1+ \hat{\alpha}'_2 > 0$ then
$${\lambda_{k+r}(\tilde{\Gamma}) \geq \lambda_k(\Gamma), \quad k \geq k_0}.$$
\item If $\alpha'_1 \cdot \hat{\alpha}'_2 < 0 \ \text{and}\ \alpha_1+ \hat{\alpha}'_2 = 0$ then
$${\lambda_{k+r}(\tilde{\Gamma}) \geq \lambda_k(\Gamma), \quad k \geq k_0}.$$
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item
Let $\Gamma'$ be the disconnected graph whose components are $\Gamma$ and $\hat{\Gamma}$, that is, $\Gamma'=\Gamma \cup \hat{\Gamma}$. The assumption $\lambda_r(\hat{\Gamma}) \leq \lambda_{k_0}(\Gamma)$, and the fact that the spectrum of $\Gamma'$ is equal to the union of the spectra of $\Gamma$ and $\hat{\Gamma}$ implies that every $k$-th eigenvalue of $\Gamma$ is some $m$-th eigenvalue of $\Gamma'$,
$$\lambda_{k_0}(\Gamma)=\lambda_m(\Gamma')$$ for some $m=m(r,k_0) \geq r+k_0$. Now attach the vertex $w_1$ of $\hat{\Gamma}$ with a vertex $v_1$ of $\Gamma$ to form $\tilde{\Gamma}$, then by \textcolor{blue}{Theorem} \eqref{roh}, when $\alpha'_1,\hat{\alpha}'_2 > 0$, we obtain $$\lambda_m(\tilde{\Gamma}) \leq \lambda_m(\Gamma '),$$
combining this inequality with the estimate $m \geq r+k_0$, we obtain
$$\lambda_{r+k}(\tilde{\Gamma}) \leq \lambda_k(\Gamma), \quad k \geq k_0.$$
(4) Let $\Gamma'$ be the disconnected graph whose components are $\Gamma$ and $\hat{\Gamma}$, that is, $\Gamma'=\Gamma \cup \hat{\Gamma}$. The assumption $\lambda_r(\hat{\Gamma}) \leq \lambda_{k_0}(\Gamma)$, and the fact that the spectrum of $\Gamma'$ is equal to the union of the spectra of $\Gamma$ and $\hat{\Gamma}$ implies that
$$\lambda_{k_0}(\Gamma)=\lambda_m(\Gamma')$$ for some $m=m(r,k_0) \geq r+k_0$. Now attach the vertex $w_1$ of $\hat{\Gamma}$ with a vertex $v_1$ of $\Gamma$ to form $\tilde{\Gamma}$, then by \textcolor{blue}{Theorem} \eqref{roh}, when $\alpha'_1, \hat{\alpha}'_2 < 0$, we obtain $$\lambda_m(\tilde{\Gamma}) \geq \lambda_m(\Gamma '),$$
combining this inequality with the estimate $m \geq r+k_0$, we obtain
$$\lambda_{r+k}(\tilde{\Gamma}) \geq \lambda_k(\Gamma), \quad k \geq k_0.$$
\end{enumerate}
The other parts of the theorem can be proved correspondingly.
\end{proof}
\begin{definition}
The inserting a graph at vertex is a surgical operation in which a metric graph $\tilde{\Gamma}$ is obtained by removing a vertex $v_0$ of $\Gamma$ and attaching its incident edges to some vertices of $\hat{\Gamma}$.
Let $e_1,e_2,\cdots,e_n$ be the incident edges to a vertex $v_0$ of a metric graph $\Gamma$ and $w_1,w_2,\cdots,w_m, m \leq n$ be the vertices in $\hat{\Gamma}$ to which an edge has been so attached. If $\alpha_0$ and $\hat{\alpha}_1,\hat{\alpha}_2, \cdots, \hat{\alpha}_m$ are the coupling parameters associated to the vertices $v_0$ and $w_1,w_2,\cdots,w_m$, respectively. Then the coupling parameters should be placed at the vertices of $\tilde{w}_1,\tilde{w}_2,\cdots,\tilde{w}_m,$ of $\tilde{\Gamma}$ in such a way that their sum is equal to $\alpha_0 +\hat{\alpha}_1+\hat{\alpha}_2+ \cdots+ \hat{\alpha}_m$.
\end{definition}
A similar result to the following theorem was proved in \cite[\textcolor{blue}{Theorem $3.10$}]{BKKM}, , in which it was assumed that if the vertices of the graphs $\Gamma$ and $\hat{\Gamma}$ are equipped with $\delta$-type and standard vertex conditions, respectively. Then, for all $k$ such that $\lambda_k(\Gamma) \geq 0$, $$\lambda_k(\tilde{\Gamma}) \leq \lambda_k(\Gamma).$$
The reason for imposing standard conditions on $\hat{\Gamma}$ is that we want that the sum of the strengths assigned to the new vertices of $ \tilde{\Gamma}$ must equal to the strength assigned at $v_0$ of $\Gamma$. The following result omits this restriction. Therefore, the eigenvalues of $\tilde{\Gamma}$ are bounded above by the eigenvalues of the graphs with different strengths at $v_0$ each strength equal to the sum of the strengths of vertices in $\tilde{\Gamma}$ to which the incident edges of $v_0$ are attached. Moreover, we believe that the following theorem gives an improved result when $r \geq m$.
\begin{theorem}
Let $\tilde{\Gamma}^{\delta}$ be obtained by inserting vertices $w_1,w_2,\cdots,w_m$ of $\hat{\Gamma}^{\delta}$ at vertex $v_0$ of $\Gamma^{\delta}$. If for some $r\geq m$ and $k$, we have, $ \lambda_r(\hat{\Gamma}^{\delta}) \leq \lambda_{k}(\Gamma^{\delta}).$ Then,
\begin{equation}
\lambda_{k+r-m}(\tilde{\Gamma}^{\delta})\leq \lambda_k(\Gamma^{\delta}).
\end{equation}
\end{theorem}
\begin{proof}
Let $\check{\Gamma}^{\delta}$ be a graph obtained by gluing the vertices $w_1,w_2,\cdots,w_m$ of $\hat{\Gamma}^{\delta}$ to form a single vertex $w^*$, then
$$\lambda_{r-m+1}(\check{\Gamma}^{\delta}) \leq \lambda_r(\hat{\Gamma}^{\delta}). $$
The assumption $ \lambda_r(\hat{\Gamma}^{\delta}) \leq \lambda_{k}(\Gamma^{\delta}),$ implies that $\lambda_{r-m+1}(\check{\Gamma}^{\delta}) \leq \lambda_{k}(\Gamma^{\delta})$.
Obtain the graph $\check{\check{\Gamma}}^{\delta}$ by attaching the vertex $w^*$ of pendant graph $\check{\Gamma}^{\delta}$ at vertex $v_0$ of $\Gamma^{\delta}$, this implies,
$$\lambda_{k+r-m}(\check{\check{\Gamma}}^{\delta}) \leq \lambda_k(\Gamma^{\delta}).$$
Now, cut through the vertex $v_0$ of $\check{\check{\Gamma}}^{\delta}$ to restore the vertices $\tilde{w}_1,\tilde{w}_2,\cdots,\tilde{w}_m$ to form a graph $\tilde{\Gamma}^{\delta}$, thus
$$\lambda_{k+r-m}(\tilde{\Gamma}^{\delta})\leq \lambda_{k+r-m}(\check{\check{\Gamma}}^{\delta}).$$
\end{proof}
\begin{corollary}
Let $\Gamma$ be a finite, compact and connected metric graph with arbitrary self adjoint vertex conditions at each vertex. Let $\tilde \Gamma$ be a metric graph obtained by lengthening an edge of $\Gamma$.
If for some $k_0$, $\lambda_{k_0}(\Gamma)\geq 0$, then
\begin{equation}
\lambda_k(\tilde \Gamma) \leq \lambda_k(\Gamma), \quad k \geq k_0.
\end{equation}
\end{corollary}
\begin{proof}
Let $\Gamma^s$ be an edge of finite length with endpoints equipped with standard vertex conditions, and $\mathcal{C}^s$ be a loop obtained by joining the boundary points of $\Gamma^s$. Consider any interior point $x^*$ on the edge of $\Gamma$ as a vertex of degree two equipped with standard condition.
Since
\begin{equation*}
0=\lambda_1(\mathcal{C}^s) \leq \lambda_{k_0}(\Gamma)
\end{equation*}
Obtain a graph $\hat{\Gamma}$ by attaching $\mathcal{C}^s$ as a pendant graph to $\Gamma$ at an interior point $x^*$ then.
\begin{equation*}
\lambda_k (\hat{\Gamma}) \leq \lambda_k(\Gamma), \quad k \geq k_0.
\end{equation*}
Now restore the endpoints of $\Gamma^s$ in accordance with the definition of splitting of the vertex to obtained a graph $\tilde \Gamma$, having two interior points as a vertex of degree two each equipped with standard conditions. Which can easily be removed and considered as interior points of an edge of $\Gamma$. Thus,
\begin{equation*}
\lambda_k (\tilde{\Gamma} ) \leq \lambda_k(\hat{\Gamma}), \quad k \geq k_0.
\end{equation*}
\end{proof}
\begin{proposition} \label{insert}
Suppose $\tilde{\Gamma}$ is formed by inserting a graph $\Gamma^a$ at a vertex $v_0$ of $\Gamma$ in a way that incident edges of $v_0$ are attached to only two vertices , say $w_1$ and $w_2$, of $\Gamma^a$. Assume that arbitrary local self-adjoint vertex conditions are imposed on each vertex $v_m$ of $\Gamma$ except $v_0$, where $\delta'$-condition is imposed with strength $\alpha'_0$, and anti-standard conditions are prescribed at each vertex of $\Gamma^a$ prior to insertion. If for some $k_0$, $\lambda_{k_0}(\Gamma)\geq 0$, then
\begin{equation*}
\lambda_{k+1}(\tilde{\Gamma}) \leq \lambda_k(\Gamma), \quad k\geq k_0.
\end{equation*}
\end{proposition}
\begin{proof}
Let $\hat{\Gamma}^a$ be a graph obtained by gluing the vertices $w_1$ and $w_2$ of $\Gamma^a$ to form a single vertex $w^*$ with total strength zero. The assumption $ \lambda_{k}(\Gamma) \geq 0$, for all $k \geq k_0$, along with the fact that the spectrum of $\hat{\Gamma}^a$ is non-negative, implies that $\lambda_1(\hat{\Gamma}^a)=0 \leq \lambda_k(\Gamma)$, for all $k \geq k_0$.
Let $\check{\Gamma}$ be obtained by attaching the vertex $w^*$ of pendant graph $\hat{\Gamma}^a$ at vertex $v_0$ of $\Gamma$, then by \textcolor{blue}{Theorem \eqref{pendant}(2)},
$$\lambda_{k+1}(\check{\Gamma}) \leq \lambda_k(\Gamma), \quad k \geq k_0.$$
Now, cut through the vertex $v_0$ of $\check{\Gamma}$ to restore the vertices $w_1$ and $w_2$ to obtain the graph $\tilde{\Gamma}$, the restoration of vertices is performed in such a way that the sum of strengths $\tilde{\alpha}'_1$, $\tilde{\alpha}'_2$, assigned to $\tilde{w}_1$ and $\tilde{w}_2$, is $\alpha'_0$. Moreover, the strengths are assigned such that the the operation of splitting the vertex $v_0$ lowers all eigenvalues of $\check{\Gamma}$. Thus,
$$\lambda_{k+1}(\tilde{\Gamma})\leq \lambda_{k+1}(\check{\Gamma}).$$
\end{proof}
\section{Eigenvalue Estimates}
This section provides upper and lower estimates for Laplacian on a metric graph equipped with self-adjoint vertex conditions ($\delta$ and $\delta'$-type conditions). This section is divided into two subsections.
\subsection{Estimates on the eigenvalues of \texorpdfstring{$\delta$}{Lg}-graph} In this part of the section we provide some upper and lower estimates on the eigenvalues of the $\delta$-graph.
\begin{theorem} \label{ariturk1}
Let $\Gamma^s$ be finite, compact connected metric graph with standard conditions, then
\begin{equation}\label{est1}
\lambda_k(\Gamma^s) \leq \left( \frac{(k-1+\beta+|E|)\pi}{\mathcal{L}(\Gamma^s)}\right)^2.
\end{equation}
\end{theorem}
\begin{proof}
Consider an interval $I^s$ of length $\mathcal{L}(\Gamma^s)$ equipped with standard vertex conditions. Obtain a path graph $P^s$ by creating vertices of degree two at interior points of $I^s$ such that the number of edges and their lengths are same as in $\Gamma^s$. Cut through the degree two vertices of $P^s$ to form a new graph $\hat{\Gamma}^s$ which consists of $|E|$ disconnected edges. Create a metric tree $T^s$ by gluing the vertices of $\hat{\Gamma}^s$ in such a way that the graph $\Gamma^s$ can be obtained from $\Gamma^s$ by pairwise gluing of $\beta$ pair of vertices of $T^s$. Thus,
\begin{equation}
\left(\frac{k \pi}{\mathcal{L}(\Gamma^s)}\right)^2=\lambda_k(I^s)=\lambda_k(P^s) \geq \lambda_k(\hat{\Gamma}^s) \geq \lambda_{k-|E|+1}(T^s) \geq \lambda_{k-|E|+1-\beta}(\Gamma^s).
\end{equation}
Equivalently, this estimates can be written as,
$$\lambda_k(\Gamma^s) \leq \left( \frac{(k-2+2\beta+|V|)\pi}{\mathcal{L}(\Gamma^s)}\right)^2.$$
\end{proof}
We can also obtain a lower bound on $\lambda_k(\Gamma^s)$ as follows. However, this bound is not a better estimate as it only makes sense when $k\geq |E|-1$.
$$\left(\frac{(k-|E|+1) \pi}{\mathcal{L}(\Gamma^s)} \right)^2=\lambda_{k-|E|+1}(I)=\lambda_{k-|E|+1}(P) \leq \lambda_k(\hat{\Gamma}^s) \leq \lambda_k(T^s) \leq \lambda_k(\Gamma^s).$$
We use this estimate to obtain a few bounds on the eigenvalues of $\Gamma^{\delta}$. Let $\alpha=\sum\limits_{m=1}^{|V|} \alpha_m$ denote the total strength of $\Gamma^{\delta}$ and $\Gamma^{\delta}_f$ be the corresponding flower graph. Let $\Gamma^s_f$ be the same underlying flower graph equipped with standard conditions. If $\alpha \leq 0$, then the above result implies that $$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^{\delta}_f) \leq \lambda_k(\Gamma^s_f) \leq \left( \frac{(k-1+2|E|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.$$
Similarly, if $\alpha > 0$, then
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^{\delta}_f) \leq \lambda_{k+1}(\Gamma^s_f) \leq \left( \frac{(k+2|E|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.$$
\begin{theorem} \label{delta1}
Let $\Gamma^{\delta}$ be a metric graph with $|E|$ number of edges and $|V|$ number of vertices. Then the following assertions hold.
\begin{enumerate}
\item If $\alpha=0$, then
\begin{equation*}
\left( \frac{(k-|V|+1)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2 \leq \lambda_k(\Gamma^{\delta}) \leq \left( \frac{(k+|E|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.
\end{equation*}
\item If $\alpha > 0 $, then
\begin{equation}
\left( \frac{(k-|V|+1)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2 \leq \lambda_k(\Gamma^{\delta}) \leq \left( \frac{(k+|E|+1)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.
\end{equation}
\item If $\alpha < 0$, then
\begin{equation}
\left( \frac{(k-|V|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2 \leq \lambda_k(\Gamma^{\delta}) \leq \left( \frac{(k+|E|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.
\end{equation}
\end{enumerate}
\end{theorem}
\begin{proof}
Let $I^s$ be an interval of length $\mathcal{L}(\Gamma^{\delta})$ equipped with standard vertex conditions and let $P^s$ be a path graph obtained by creating vertices of degree two at interior points of $I^s$ such that the number of edges and their lengths are same as in $\Gamma^{\delta}$. Then,
$$\lambda_k(I^s)=\lambda_k(P^s).$$
Obtain a flower graph $P^s_f$ from $P^s$, then \eqref{dinterlac} implies that
$$\lambda_{k-|E|}(P^s_f) \leq \lambda_{k}(P^s) \leq \lambda_k(P^s_f) \leq \lambda_{k+|E|}(P^s).$$
Thus,
\begin{equation*}
\left( \frac{k\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2=\lambda_k(I^s)=\lambda_k(P^s)\leq \lambda_k(P^s_f) \leq \lambda_{k+|E|}(P^s)=\lambda_{k+|E|}(I^s)= \left(\frac{(k+|E|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.
\end{equation*}
\begin{enumerate}
\item If $\alpha=0$, then
$$ \left( \frac{(k-|V|+1)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2 \leq \lambda_{k-|V|+1}(P^s_f) \leq \lambda_k(\Gamma^{\delta})\leq \lambda_k(P^s_f) \leq \left( \frac{(k+|E|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.$$
\item If $\alpha > 0$, then
$$ \left( \frac{(k-|V|+1)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2 \leq \lambda_{k-|V|+1}(P^s_f) \leq \lambda_k(\Gamma^{\delta})\leq \lambda_{k+1}(P^s_f) \leq \left( \frac{(k+|E|+1)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.$$
\item If $\alpha < 0$, then
$$ \left( \frac{(k-|V|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2 \leq \lambda_{k-|V|}(P^s_f) \leq \lambda_k(\Gamma^{\delta})\leq \lambda_k(P^s_f) \leq \left( \frac{(k+|E|)\pi}{\mathcal{L}(\Gamma^{\delta})}\right)^2.$$
\end{enumerate}
\end{proof}
\begin{theorem} \label{ariturk2}
Let $\Gamma$ be finite compact connected metric graph with pendant vertices equipped with Dirichlet conditions and all other internal vertices the standard conditions are imposed. Let $|D|$ and $|S|$ denote the number of Dirichlet and number of standard vertices in $\Gamma$. Then \begin{equation}
\lambda_k(\Gamma) \leq \left( \frac{(k-2+2\beta+2|D|+|S|)\pi}{\mathcal{L}(\Gamma)}\right)^2.
\end{equation}
\end{theorem}
\begin{proof}
Let $\Gamma^s$ be same underlying metric graph as $\Gamma$, but equipped with standard vertex conditions. Then by \eqref{est1}
$$\lambda_k(\Gamma^s) \leq \left( \frac{(k-1+\beta+|E|)\pi}{\mathcal{L}(\Gamma^s)}\right)^2.$$
Obtain the graph $\Gamma$ by imposing Dirichlet conditions at required vertices of $\Gamma^s$. Then by \eqref{interlacing1}
$$\lambda_{k-|D|}(\Gamma) \leq \lambda_k(\Gamma^s) \leq \lambda_k(\Gamma)\leq \lambda_{k+|D|}(\Gamma^s).$$
$$\lambda_k(\Gamma) \leq \left( \frac{(k-2+2\beta+|D|+|V|)\pi}{\mathcal{L}(\Gamma^s)}\right)^2. $$
Equivalently, this estimates can be written as,
$$\lambda_k(\Gamma) \leq \left( \frac{(k-2+2\beta+2|D|+|S|)\pi}{\mathcal{L}(\Gamma)}\right)^2.$$
\end{proof}
For the metric graphs in \textcolor{blue}{Theorem} \eqref{ariturk1} and \textcolor{blue}{Theorem} \eqref{ariturk2}, similar but improved estimates on the eigenvalues can be found in \cite[\textcolor{blue}{Lemma 1.5}]{A}. However, for any flower graph equipped with standard conditions the estimate \eqref{est1} coincides with the estimate presented in \cite[\textcolor{blue}{Lemma 1.5}]{A}.
\subsection{Estimates on the eigenvalues of \texorpdfstring{$\delta'$}{Lg}-graph}
We establish a few estimates on the lowest eigenvalue of the $\delta'$-graph using some trials functions from the domain of the quadratic form and also on general eigenvalues using some established inequalities. \\
Since the domain of the quadratic form of the $\delta'$-graph $\Gamma^{\delta'}$ contain a constant function, in particular it contains $\hat \varphi(x)=\frac{1}{\sqrt{\mathcal{L}(\Gamma^{\delta'})}}$. Therefore,
\begin{align*}
\lambda_1(\Gamma^{\delta'})&= \underset{\varphi \in D(h^{\delta'})}{\min} h^{\delta'}[\varphi] \\
&=\underset{\varphi \in D(h^{\delta'})}{\min} \left( \int_{\Gamma^{\delta'}}|\varphi'(x)|^2 dx +\sum\limits_{m=1}^{|V|} \frac{1}{\alpha'_m} \left|\sum \limits_{x_i \in v_m}\varphi(x_i)\right|^2 \right) \\
&\leq \int_{\Gamma^{\delta'}}|\hat \varphi'(x)|^2 dx +\sum\limits_{m=1}^{|V|} \frac{1}{\alpha'_m} \left|\sum \limits_{x_i \in v_m}\hat \varphi(x_i)\right|^2 \\
&= \sum\limits_{m=1}^{|V|} \frac{1}{\alpha'_m} \left|\sum \limits_{x_i \in v_m} \frac{1}{\sqrt{\mathcal{L}(\Gamma^{\delta'})}} \right|^2 \\
&=\frac{1}{\mathcal{L}(\Gamma^{\delta'})} \sum\limits_{m=1}^{|V|} \frac{1}{\alpha'_m} \left|\sum \limits_{x_i \in v_m} 1 \right|^2 \\
&=\frac{1}{\mathcal{L}(\Gamma^{\delta'})} \sum\limits_{m=1}^{|V|} \frac{1}{\alpha'_m} |d_{v_m}|^2 \\
&=\frac{1}{\mathcal{L}(\Gamma^{\delta'})} \sum\limits_{m=1}^{|V|} \frac{1}{\alpha'_m} |d_{v_m}|^2
\end{align*}
\begin{remark}
Even if there are some vertices where $\alpha'_m=0$, we can still use this upper bound and in this case the summation is taken over the vertices where strengths $\alpha'_m$ are non zero.
\begin{equation*}
\lambda_1(\Gamma^{\delta'}) \leq \frac{1}{\mathcal{L}(\Gamma^{\delta'})} \left( \sum\limits_{\{v \in V| \alpha'_v \neq 0\}} \frac{d^2_{v}}{\alpha'_v} \right)
\end{equation*}
\end{remark}
\begin{enumerate}
\item If the degree of each vertex is $n$ then, $$\lambda_1(\Gamma^{\delta'}) \leq \frac{n^2} { \mathcal{L}(\Gamma^{\delta'})} \sum\limits^{|V|}_{m=1} \frac{1}{\alpha'_m}.$$
\item If the strength at each vertex is same, say $\alpha'$, then
$$\lambda_1(\Gamma^{\delta'}) \leq \frac{1}{\alpha' \mathcal{L}(\Gamma^{\delta'})} \sum\limits^{|V|}_{m=1} |d_{v_m}|^2.$$
\item In addition, if the total number of vertices are $|V|$ and degree of each vertex is $n$ then, $$\lambda_1(\Gamma^{\delta'}) \leq \frac{n^2 \ |V|}{\alpha' \mathcal{L}(\Gamma^{\delta'})}.$$
\item If the strengths are evenly distributed, then
$$\lambda_1(\Gamma^{\delta'}) \leq \frac{n^2 \ |V|^2 }{\alpha' \mathcal{L}(\Gamma^{\delta'})}.$$
\end{enumerate}
The question now is how close one can push the lowest eigenvalue towards this upper bound. The quadratic form does not converge very close to lowest eigenvalue when the strengths are evenly distributed. So, evenly distribution does not increase the eigenvalue. How can we get better estimation? \\
Total degree of a graph is defined as the sum of degree of each vertex. Handshaking theorem states that the degree of a graph $\Gamma$ equals twice the number of edges in $\Gamma$ and the total degree of a graph is even.
\begin{equation*}
degree \ of \ a \ graph = \sum_{v \in V} deg(v)=2 \ (number \ of \ edges \ in \ \Gamma)
\end{equation*}
\begin{remark}
We express the upper bound on the lowest eigenvalue in terms of total degree of a graph and total number of edges. Although, the following is not a good approximation as compare to earlier one.
Let $|V|$ and $|E|$ denote the total number of vertices and total number of edges in a graph $\Gamma$, respectively. Since
\begin{align*}
|d_{v_1}|^2+|d_{v_2}|^2+\cdots+|d_{v_{|V|}}|^2 &\leq (|d_{v_1}|+|d_{v_2}|+\cdots+|d_{v_{|V|}}|)^2 \\
&=[deg(\Gamma)]^2 \\
&=4 \ |E|^2.
\end{align*}
If the strengths are evenly distributed, then
\begin{equation*}
\lambda_1(\Gamma^{\delta'}) \leq \frac{|V|}{\beta \mathcal{L}(\Gamma^{\delta'})} \ [deg(\Gamma^{\delta'})]^2,
\end{equation*}
and in terms of number of edges,
\begin{equation*}
\lambda_1(\Gamma^{\delta'}) \leq \frac{4 \ |V| \ |E|^2}{\beta \mathcal{L}(\Gamma^{\delta'})}.
\end{equation*}
If we let $\alpha'_m=\infty$ at each vertex then, $\lambda_1(\Gamma)=0$. Maybe we can increase the eigenvalue and push it towards the upper bound by adding the maximum number of edges with a large length and minimum vertices.
\end{remark}
Since there is no restriction on the domain of the quadratic form, so we could have used other functions that belongs to $H^1(\Gamma^{\delta'})$ as a test functions i.e. $\cos(x), \sin(x)$, but the bounds that we get using these functions are not so explicitly nice. Let us use the functions $\varphi(x)=\sin(\frac{\pi x}{\ell_j}), \quad x \in e_j=[0,\ell_j]$ and $\varphi(x)=\cos(\frac{\pi x}{\ell_j}), \quad x \in e_j=[\frac{-\ell_j}{2},\frac{\ell_j}{2}]$. Then,
\begin{equation*}
\lambda_1(\Gamma^{\delta'}) \leq \frac{\pi^2}{L(\Gamma^{\delta'})} \sum\limits_{j=1}^{|E|}\frac{1}{\ell_j}.
\end{equation*}
If the length of each edge is equal, say, $\ell$, then
\begin{equation*}
\lambda_1(\Gamma^{\delta'}) \leq \left(\frac{\pi}{\ell} \right)^2.
\end{equation*}
We can also use the function $\varphi(x)=\sin(\frac{2 \pi x}{\ell_j}), \quad x \in e_i=[0,\ell_j]$, as a test function and we get an upper bound
\begin{equation*}
\lambda_1(\Gamma^{\delta'} ) \leq \frac{4 \pi^2}{L(\Gamma^{\delta'})} \sum\limits_{j=1}^{|E|} \frac{1}{\ell_j}.
\end{equation*}
Let us consider the function $\varphi(x)=\cos(\frac{2 \pi x}{\ell_j}), \quad x \in e_j=[0,\ell_j]$. Clearly, this function is in the domain of the quadratic form and can be used as a test function to get an upper bound on first eigenvalue and $\varphi(0)=\varphi(\ell_j)=1$ implies that $\sum \limits_{x_i \in v_m}\varphi(x_i)=d_{v_m}$, for each $m=1,2,\cdots,|V|.$
\begin{align*}
h^{\delta'}[\varphi]&=\int_{\Gamma^{\delta'}}|\varphi'(x)|^2 dx+\sum\limits_{m=1}^{|V|}\frac{1}{\alpha'_m} \left|\sum\limits_{x_i \in v_m} \varphi(x_i)\right|^2,\\
&= 2\pi^2 \sum\limits_{j=1}^{|E|} \frac{1}{\ell_j}+\sum\limits_{m=1}^{|V|}\frac{d^2_{v_m}}{\alpha'_m}.
\end{align*}
\begin{equation*}
<\varphi,\varphi>=\int_{\Gamma^{\delta'}}|\varphi(x)|^2 dx=\sum\limits_{j=1}^{|E|} \frac{\ell_j}{2}=\frac{L(\Gamma^{\delta'})}{2}.
\end{equation*}
\begin{equation*}
\lambda_1(\Gamma^{\delta'}) \leq \frac{2}{L(\Gamma^{\delta'})} \left( 2\pi^2 \sum\limits_{j=1}^{|E|} \frac{1}{\ell_j}+\sum\limits_{m=1}^{|V|}\frac{d^2_{v_m}}{\alpha'_m} \right).
\end{equation*}
For any fix real number $\lambda$, the eigenvalue counting function $N_{\Gamma}(\lambda)$ is defined as the number of eigenvalues of a graph $\Gamma$ smaller than $\lambda$. Since the quantum graph $\Gamma$ is finite compact, and the operator is self-adjoint with a discrete spectrum bounded from below. Therefore the value of function $N_{\Gamma}(\lambda)$ is finite.
$$ N_{\Gamma}(\lambda)= \# \{\lambda_i \in \sigma(\Gamma) : \lambda_i \leq \lambda\}. $$
\begin{theorem} \label{ec}
Let $\Gamma^a$ be finite compact graph of total length $\mathcal{L}(\Gamma^a)$ with $|V|$ the number of vertices and $|E|$ the number of edges. Let each vertex of $\Gamma^a$ be equipped with anti-standard vertex conditions. Then
\begin{equation} \label{ec1}
\left(\frac{\pi}{\mathcal{L}(\Gamma^a)} \right)^2 (k-|V|)^2 \leq \lambda_k(\Gamma^a) \leq \left(\frac{\pi}{\mathcal{L}(\Gamma^a)} \right)^2 (k+|E|-1)^2.
\end{equation}
\end{theorem}
\begin{proof}
Since the interaction strengths are zero at each vertex, therefore the quadratic form is non-negative, and thus the eigenvalues $\lambda \geq 0$.
Hence the lower estimate in \eqref{ec1}
is interesting only if $k > |V|$.\\
Let $\Gamma^d$ be the graph obtained from $\Gamma^a$ by imposing the Dirichlet Conditions at each vertex of $\Gamma^a$, and let $h^a$ and $h^d$ denotes the quadratic form of the graphs $\Gamma^a$ and $\Gamma^d$ with domains $D(h^a)$ and $D(h^d)$, respectively. Since the expression of the quadratic forms $h^a$ and $h^d$ are same; moreover, if a function $\varphi$ satisfy Dirichlet conditions at some vertex $v$, then it also satisfies the anti-standard vertex condition that vertex. Therefore, the domain $D(h^d)$ is a subspace of $D(h^a)$ and the quadratic forms $h^a$ and $h^d$ agree on $D(h^d)$, thus minimizing over a smaller domain results in large eigenvalues. Furthermore, the domain $D(h^d)$ is a co-dimension one subspace of $D(h^a)$. Thus, by the rank one nature perturbation, the following interlacing inequalities holds.
$$\lambda_k(\Gamma^a) \leq \lambda_k(\Gamma^d) \leq \lambda_{k+|V|}(\Gamma^a)$$
Let for some $\lambda \in \mathbb{R}$, $N_{\Gamma^a}(\lambda)$ and $N_{\Gamma^d}(\lambda)$ denotes the eigenvalue counting functions for the graphs $\Gamma^a$ and $\Gamma^d$, respectively. Then the above interlacing inequalities implies the following relation between eigenvalue counting functions.
\begin{equation} \label{cfrelation}
N_{\Gamma^d}(\lambda) \leq N_{\Gamma^a}(\lambda) \leq N_{\Gamma^d}(\lambda) +|V|.
\end{equation}
Consider the Laplace operator acting on a interval of length $\ell$ , and assume that the end points are equipped with Dirichlet Conditions. The eigenvalues are $\lambda_k=\left( \frac{k \pi}{\ell}\right)^2$, for $\lambda \geq 0$ the value of eigenvalue counting function is $N_{[0,\ell]}(\lambda)=\left[\frac{\sqrt{\lambda }}{\pi} \ell \right].$ Where square brackets mean to take the integer part of the argument. Since Dirichlet conditions imposed at a vertex of degree two or more does not connect the individual function living on the incident edge in any way. The Dirichlet conditions have the effect of disconnecting the vertex of degree $d_v$ into $d_v$ vertices of degree one, so the graph $\Gamma^d$ is now decoupled into a set of intervals, and the set of eigenvalues is just the union of eigenvalues of each interval (counting multiplicities). Moreover,
The counting function is given by
\begin{equation} \label{c1}
N_{\Gamma^{d}}(\lambda)= \sum\limits_{j=1}^{|E|}N_{[0,\ell_j]}(\lambda) =\left[\frac{\sqrt {\lambda }}{\pi} \ell_1 \right] +\left[\frac{\sqrt {\lambda }}{\pi} \ell_2 \right] + \cdots+ \left[\frac{\sqrt {\lambda }}{\pi} \ell_{|E|} \right] \leq \left[\frac{\sqrt {\lambda }}{\pi} \mathcal{L}(\Gamma^a) \right].
\end{equation}
Since for any $a$ and $b$, $ [a]+[b] \leq [a+b]$, therefore taking integer part of the sum of the terms is increased at most by the number of terms minus one as compared to adding integer parts only. As the number of terms in \eqref{c1} are equal to number of edges $|E|$, thus
$$\left[\frac{\sqrt {\lambda }}{\pi} \mathcal{L}(\Gamma^a) \right] =\left[\frac{\sqrt {\lambda }}{\pi} (\ell_1 +\ell_2+\cdots+\ell_{|E|})\right] \leq \left[\frac{\sqrt {\lambda }}{\pi} \ell_1 \right] +\left[\frac{\sqrt {\lambda }}{\pi} \ell_2 \right] + \cdots+ \left[\frac{\sqrt {\lambda }}{\pi} \ell_{|E|} \right] +|E|-1, $$
and we have
\begin{equation} \label{c2}
\left[\frac{\sqrt {\lambda }}{\pi} \mathcal{L}(\Gamma^a) \right] -|E|+1 \leq N_{\Gamma^{d}}(\lambda).
\end{equation}
\end{proof}
Thus, from \eqref{cfrelation},\eqref{c1} and \eqref{c2}, we obtain the following inequalities.
\begin{equation*}
\left[\frac{\sqrt {\lambda }}{\pi} \mathcal{L}(\Gamma^a) \right] -|E|+1 \leq N_{\Gamma^{d}}(\lambda) \leq N_{\Gamma^{a}}(\lambda) \leq N_{\Gamma^d}(\lambda) +|V| \leq \left[\frac{\sqrt {\lambda }}{\pi} \mathcal{L}(\Gamma^a) \right]+|V|.
\end{equation*}
Setting $\lambda=\left (\dfrac{k \pi } {\mathcal{L}(\Gamma^a)} \right)^2$ we get
$$k-|E|+1 \leq N_{\Gamma^{a}} \left (\frac{k \pi } {\mathcal{L}(\Gamma^a)} \right)^2 \leq k+|V|,$$ so
$$\lambda_{k - |E|+1} \leq \left ( \frac{k \pi } {\mathcal{L}(\Gamma^a)} \right)^2 \leq \lambda_{k+|V|}.$$
This estimate implies that the multiplicity of the eigenvalues is uniformly bounded by $|E|+|V|$.
setting $\tilde{k}=k-|E|+1$ we get $\lambda_{\tilde{k}} \leq \left(\dfrac{\pi}{\mathcal{L}(\Gamma^a)}\right)^2 (\tilde{k}+|E|-1)^2$ and similarly setting $\tilde{k}=k+|V|$ we get $\left(\dfrac{\pi}{\mathcal{L}(\Gamma^a)}\right)^2 (\tilde{k}-|V|)^2 \leq \lambda_{\tilde{k}}$. Since $\tilde{k}$ is a dummy subscript so we can replace it with $k$ and finally, we obtain,
$$\left(\frac{\pi}{\mathcal{L}(\Gamma^a)} \right)^2 (k-|V|)^2 \leq \lambda_k(\Gamma^a) \leq \left(\frac{\pi}{\mathcal{L}(\Gamma^a)} \right)^2 (k+|E|-1)^2.$$
Another type of vertex condition that decouples the graph into disjoints interval is the Neumann conditions. For these conditions, the derivatives of functions living on incident edges are zero at each vertex, and no conditions are assumed on the values of functions. That is, $$\varphi_e'(v)=0$$
Let $\Gamma^n$ be the finite compact metric graph with Neumann vertex conditions, then the quadratic form is defined by,
$$h^n(\varphi)=\int_{\Gamma^{n}} |\varphi'|^2dx$$
with domain $D(h^n)=H^1(\Gamma^{n} \backslash V)$. The domain $D(h^a)$ is a co-dimension one subspace of $D(h^n)$, and the quadratic forms $h^a$ and $h^n$ agree on $D(h^a)$. Thus, by rank one perturbation, the following interlacing inequalities holds.
$$\lambda_k(\Gamma^n) \leq \lambda_k(\Gamma^a) \leq \lambda_{k+|V|}(\Gamma^n)$$
Let for some $\lambda \in \mathbb{R}$, $N_{\Gamma^{a}}(\lambda)$ and $N_{\Gamma^{n}}(\lambda)$ denotes the eigenvalue counting functions for the graphs $\Gamma^a$ and $\Gamma^n$, respectively. Then the above interlacing inequalities implies the following relation between eigenvalue counting functions.
\begin{equation}
N_{\Gamma^a}(\lambda) \leq N_{\Gamma^n}(\lambda) \leq N_{\Gamma^a}(\lambda) +|V|.
\end{equation}
In the previous theorem, if we impose Neumann conditions instead of Dirichlet conditions to decouple the graph $\Gamma^a$, then one can prove the following theorem correspondingly.
\begin{theorem} \label{anti2}
Let $\Gamma^a$ be finite compact graph of total length $\mathcal{L}(\Gamma^a)$ with $|V|$ the number of vertices and $|E|$ the number of edges. Let each vertex of $\Gamma^a$ be equipped with anti-standard vertex conditions. Then
\begin{equation}
\left(\frac{k \pi}{\mathcal{L}(\Gamma^a)} \right)^2 \leq \lambda_k(\Gamma^a) \leq \left(\frac{\pi}{\mathcal{L}(\Gamma^a)} \right)^2 (k+|E|+|V|-1)^2.
\end{equation}
\end{theorem}
\begin{theorem} \label{deltap}
Let $\Gamma^{\delta'}$ be metric graph with negative strengths $\alpha'_m <0$ at each vertex. Then,
\begin{equation}
\left(\frac{(k-2|V|)\pi}{2 \mathcal{L}(\Gamma^{\delta'})}\right)^2 \leq \lambda_k(\Gamma^{\delta'}), \quad k \geq 2|V|.
\end{equation}
\end{theorem}
\begin{proof}
Let $\hat{\Gamma}^{\delta'}$ be a double cover obtained by doubling each edge in $\Gamma^{\delta'}$ and assigning strengths $2\alpha'_m$ at each vertex. Then each eigenvalue of $\Gamma^{\delta'}$ is also an eigenvalue of $\hat{\Gamma}^{\delta'}$. However, the converse is not true. Let $\lambda(\Gamma^{\delta'})$ be an eigenvalue corresponding to the eigenfunction $\varphi$. The eigenfunction $\varphi$ can be lifted on $\hat{\Gamma}^{\delta'}$ by letting it assume the same values on the new edges as on edges of $\Gamma^{\delta'}$. Then newly constructed function $\hat{\varphi}$ is an eigenfunction for $\hat{\Gamma}^{\delta'}$ corresponding to eigenvalue $\lambda(\Gamma^{\delta'})$, because for $\lambda(\Gamma^{\delta'})$ this function satisfies the eigenvalue equation and vertex conditions on $\hat{\Gamma}^{\delta'}$. Thus,
$$\lambda_k(\hat{\Gamma}^{\delta'}) \leq \lambda_k(\Gamma^{\delta'}).$$
Let $\mathcal{C}^{\delta'}$ be the graph obtained from $\hat{\Gamma}^{\delta'}$ by splitting each of the vertices of $\hat{\Gamma}^{\delta'}$ to form a cycle of length $2 \mathcal{L}(\mathcal{C}^{\delta'})$, since each vertex has strength $2\alpha'_m<0 $, therefore [\textcolor{blue}{Theorem} \eqref{roh}] implies,
\begin{equation*}
\lambda_k(\mathcal{C}^{\delta'}) \leq \lambda_k(\hat{\Gamma}^{\delta'}).
\end{equation*}
Obtain a graph $\mathcal{C}^{a}$ by assigning the zero at all vertices of $\mathcal{C}^{\delta'}$. Then by \textcolor{blue}{Theorem} \eqref{strength},
$$\lambda_{k-2|V|}(\mathcal{C}^a)\leq \lambda_k(\mathcal{C}^{\delta'}).$$
The graph $\mathcal{C}^a$ is a bipartite metric graph, therefore \cite[\textcolor{blue}{Theorem 3.6}]{PR} implies that
$$\lambda_{k-2|V|}(\mathcal{C}^s)=\lambda_{k-2|V|}(\mathcal{C}^a)$$
The graph $\mathcal{C}$ is a loop of length $2\mathcal{L}(\Gamma^{\delta'})$, and now
we cut the graph $\mathcal{C}^s$ to form an edge $I$ of length $2\mathcal{L}(\Gamma^{\delta'}).$ Thus,
$$\left(\frac{(k-2|V|)\pi}{2 \mathcal{L}(\Gamma^{\delta'})}\right)^2=\lambda_{k-2|V|}(I)\leq \lambda_{K-2|V|}(\mathcal{C}^s) .$$
\end{proof}
A similar lower bound on the eigenvalues of the Laplacian acting on the edges of a star graph with anti-standard vertex condition was proved in \cite[\textcolor{blue}{Theorem 1} ]{ZS}. We will use it to get a lower estimate on the metric graph $\Gamma$ with $\delta'$-conditions.
\begin{theorem} \label{deltaprime}
Let $\Gamma^{\delta'}$ be finite compact connected metric graph with $|E| \geq 2$, and $\delta'$-conditions equipped at each vertex with total strength $\alpha'=\sum\limits^{|V|}_{m=1} \alpha'_m$. Let $\ell_1 \geq \ell_2 \geq \cdots \geq\ell_{|E|}$ denotes the length of each edge $e_j$, then for $k,|E| \in \mathbb{N}$, and $1 \leq j \leq |E|$, we have
\begin{equation}
min \left \{\frac{\left(2\left( k- \left( 1+\frac{|V|+1}{|E|}\right) \right)+3\right)^2 \pi^2}{4 \ell_1^2}, \frac{\left(2\left( k- \left( 1+\frac{|V|+1}{|E|}\right) \right)+1\right)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k|E|+j}(\Gamma^{\delta'})
\end{equation}
\end{theorem}
\begin{proof}
\begin{enumerate}
\iffalse \item
Let $S^a$ denote the star graph constructed from the edges of $\Gamma^{\delta'
}$ and let each vertex of $S^a$ be equipped with anti-standard vertex condition. Then by \cite[\textcolor{blue}{Theorem 1} ]{ZS}
\begin{equation}
min \left \{ \frac{(2k+3)^2 \pi^2}{4 \ell_1^2}, \frac{(2k+1)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k|E|+j}(S^a).
\end{equation}
Construct a new star graph $S^{\delta'}$ by assigning the strength $\frac{-1}{\alpha'}$ at boundary vertices of $S^a$, and assign $\alpha'+\frac{|E|}{\alpha'}$ at the center vertex of $S^{a}$. Then by \textcolor{blue}{Theorem} \eqref{strength},
$$\lambda_{k|E|+j}(S^a) \leq \lambda_{k|E|+j+|E|+1}(S^{\delta'}).$$
Now, create a new flower graph $\Gamma^{\delta'}_f$ from $S^{\delta'}$ in accordance to \textcolor{blue}{Theorem} \eqref{roh}(3). Thus,
$$\lambda_{k|E|+j+|E|+1}(S^{\delta'}) \leq \lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}_f).$$
Finally, we restore the vertices of $\Gamma^{\delta'}$ by cutting, step by step, the vertex of flower graph, then by \textcolor{blue}{Theorem}\eqref{roh}(1)
$$\lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}_f) \leq \lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}).$$
\begin{align*}
k|E|+j+|E|+1&=k|E|+|E|+1+j \\
&=k|E|+\frac{(|E|+1)|E|}{|E|}+j \\
&= \left(k+ \frac{(|E|+1)}{|E|}\right)|E|+j
\end{align*}
Let $k'=\left(k+ \frac{(|E|+1)}{|E|}\right)$ $\implies$ $k=k'-\left(1+\frac{1}{|E|} \right).$
\begin{equation}
min \left \{ \frac{(2k+3)^2 \pi^2}{4 \ell_{1}^2}, \frac{(2k+1)^2 \pi^2}{4 \ell_{j}^2} \right \} \leq \lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}).
\end{equation}
\begin{equation*}
min \left \{ \frac{\left(2\left( k'- \left( 1+\frac{1}{|E|}\right) \right)+3\right)^2 \pi^2}{4 \ell_1^2}, \frac{\left(2\left( k'- \left( 1+\frac{1}{|E|}\right) \right)+1\right)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k'|E|+j}(\Gamma^{\delta'})
\end{equation*}
\item Let $S^a$ denote the star graph constructed from the edges of $\Gamma^{\delta'}$ and let each vertex of $S^a$ be equipped with anti-standard vertex condition. Then by \cite[\textcolor{blue}{Theorem 1} ]{ZS}
\begin{equation}
min \left \{ \frac{(2k+3)^2 \pi^2}{4 \ell_1^2}, \frac{(2k+1)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k|E|+j}(S^a).
\end{equation}
Construct a new star graph $S^{\delta'}$ by assigning the strength $\frac{-1}{\alpha'}$ at boundary vertices of $S^a$, and assign $\alpha'+\frac{|E|}{\alpha'}$ at the center vertex of $S^a$. Then by \textcolor{blue}{Theorem} \eqref{strength},
$$\lambda_{k|E|+j}(S^a) \leq \lambda_{k|E|+j+|E|+1}(S^{\delta'}).$$
Now, create a new flower graph $\Gamma^{\delta'}_f$ from $S^{\delta'}$ in accordance to \textcolor{blue}{Theorem}\eqref{roh}(3). Thus,
$$\lambda_{k|E|+j+|E|+1}(S^{\delta'}) \leq \lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}_f).$$
We restore the vertices of $\Gamma^{\delta'}$ by cutting, step by step, the vertex of flower graph, then by \textcolor{blue}{Theorem}\eqref{roh}(6), the obtained graph is same as $\Gamma^{\delta'}$ but one vertex of it is equipped with $\delta'$-condition with strength $\alpha'$, and at all other $|V|-1$ vertices the anti-standard conditions are imposed. Call this new graph $\Gamma_{\alpha'}$, Thus
$$\lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}_f) \leq \lambda_{k|E|+j+|E|+1}(\Gamma_\alpha').$$
Now, replace the strengths at all vertices of $\Gamma_\alpha'$ to obtain the graph $\Gamma^{\delta'}$. Thus $$\lambda_{k|E|+j+|E|+1}(\Gamma_\alpha) \leq \lambda_{k|E|+j+|E|+|V|}(\Gamma^{\delta'}).$$
\begin{align*}
k|E|+j+|E|+|V|&=k|E|+|E|+|V|+j \\
&=k|E|+\frac{(|E|+|V|)|E|}{|E|}+j \\
&= \left(k+ \frac{(|E|+|V|)}{|E|}\right)|E|+j
\end{align*}
Let $k'=\left(k+ \frac{(|E|+|V|)}{|E|}\right)$ $\implies$ $k=k'-\left(1+\frac{|V|}{|E|} \right).$
\begin{equation}
min \left \{ \frac{(2k+3)^2 \pi^2}{4 \ell_1^2}, \frac{(2k+1)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k|E|+j+|E|+|V|}(\Gamma^{\delta'}).
\end{equation}
\begin{equation*}
min \left \{\frac{\left(2\left( k'- \left( 1+\frac{|V|}{|E|}\right) \right)+3\right)^2 \pi^2}{4 \ell_1^2}, \frac{\left(2\left( k'- \left( 1+\frac{|V|}{|E|}\right) \right)+1\right)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k'|E|+j}(\Gamma^{\delta'})
\end{equation*}
\item
\fi
Let $S^a$ denote the star graph constructed from the edges of $\Gamma^{\delta'
}$ and let each vertex of $S^a$ be equipped with anti-standard vertex condition. Then by \cite[\textcolor{blue}{Theorem 1} ]{ZS}
\begin{equation}
min \left \{ \frac{(2k+3)^2 \pi^2}{4 \ell_1^2}, \frac{(2k+1)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k|E|+j}(S^a).
\end{equation}
Construct a new star graph $S^{\delta'}$ by assigning the negative strengths at vertices of $S^a$ . Then by \textcolor{blue}{Theorem} \eqref{strength},
$$\lambda_{k|E|+j}(S^a) \leq \lambda_{k|E|+j+|E|+1}(S^{\delta'}).$$
Now, first create a flower graph $\Gamma^{\delta'}_f$ from $S^{\delta'}$ in accordance to \textcolor{blue}{Theorem}\eqref{roh}(2). Thus,
$$\lambda_{k|E|+j+|E|+1}(S^{\delta'}) \leq \lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}_f).$$
and then obtain a graph $\Gamma^a_f$ by prescribing anti-standard conditions to $\Gamma^{\delta'}_f$.\\
We restore the vertices of $\Gamma^{\delta'}$ by cutting, step by step, the vertex of flower graph $\Gamma^a_f$, then by \textcolor{blue}{Theorem}\eqref{roh}(6), the obtained graph is same as $\Gamma^{\delta'}$ but equipped with anti-standard conditions. Call this new graph $\Gamma^a$, Thus
$$\lambda_{k|E|+j+|E|+1}(\Gamma^{\delta'}_f) \leq \lambda_{k|E|+j+|E|+1}(\Gamma^a_f)\leq \lambda_{k|E|+j+|E|+1}(\Gamma^a).$$
Now, assign the strengths at all vertices of $\Gamma^a$ to obtain the graph $\Gamma^{\delta'}$. Thus $$\lambda_{k|E|+j+|E|+1}(\Gamma^a) \leq \lambda_{k|E|+j+|E|+|V|+1}(\Gamma^{\delta'}).$$
\begin{align*}
k|E|+j+|E|+|V|+1&=k|E|+|E|+|V|+1+j \\
&=k|E|+\frac{(|E|+|V|+1)|E|}{|E|}+j \\
&= \left(k+ \frac{(|E|+|V|+1)}{|E|}\right)|E|+j
\end{align*}
Let $k'=\left(k+ \frac{(|E|+|V|+1)}{|E|}\right)$ $\implies$ $k=k'-\left(1+\frac{|V|+1}{|E|} \right).$
\begin{equation}
min \left \{ \frac{(2k+3)^2 \pi^2}{4 \ell_1^2}, \frac{(2k+1)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k|E|+j+|E|+|V|+1}(\Gamma^{\delta'}).
\end{equation}
\begin{equation*}
min \left \{\frac{\left(2\left( k'- \left( 1+\frac{|V|+1}{|E|}\right) \right)+3\right)^2 \pi^2}{4 \ell_1^2}, \frac{\left(2\left( k'- \left( 1+\frac{|V|+1}{|E|}\right) \right)+1\right)^2 \pi^2}{4 \ell_j^2} \right \} \leq \lambda_{k'|E|+j}(\Gamma^{\delta'})
\end{equation*}
\end{enumerate}
\end{proof}
\begin{definition}
A {\bf spanning tree }$T$ of a graph $\Gamma$ is an undirected subgraph of $\Gamma$ containing maximal edges of $\Gamma$ with no cycles, or it contains all the vertices of $\Gamma$ connected by a minimal set of edges of $\Gamma$.
\end{definition}
A graph $\Gamma$ can have more than one spanning tree, each having an equal number of edges, let $|E_T|$ denotes the cardinality of the edge set of spanning tree $T$. Each spanning tree has it's own length so we can compare their lengths, we define a new term for our later use.
\begin{definition}
A {\bf maximal spanning tree} $T_m$ is a spanning tree for a graph $\Gamma$ having the largest length among all the spanning trees of $\Gamma$.
\end{definition}
For a finite, compact connected metric tree $T^s$ with standard conditions, let $P^s$ be the largest path inside $T^s$. Since the spectrum of Laplacian on the path graph $P^s$ with standard conditions consists of discrete points $$\lambda_{k+1}(P^s)=\left(\frac{k \pi}{\mathcal{L}(P^s)}\right)^2, \quad k=0,1,2,\cdots$$
The tree graph $T^s$ can be obtained by attaching pendant edges with standard conditions to the vertices of $P^s$ in accordance to \cite[\textcolor{blue}{Theorem 3.10}]{BKKM}. Thus,
\begin{equation}
\lambda_{k+1}(T^s) \leq \lambda_{k+1}(P^s)=\left(\frac{k \pi}{\mathcal{L}(P^s)}\right)^2, \quad k=0,1,2,\cdots
\end{equation}
Now consider a finite compact connected metric tree $T^a$ with anti-standard conditions and
let $P^a$ be the longest path inside $T^a$. Since the spectrum of Laplacian on the path graph $P^a$ with anti-standard conditions consists of discrete points $$\lambda_{k}(P^a)=\left(\frac{k \pi}{\mathcal{L}(P^a)}\right)^2, \quad k=1,2,\cdots$$
The tree graph $T^a$ can be obtained by attaching pendant edges with anti-standard conditions to the vertices of $P^a$ in accordance to \cite[\textcolor{blue}{Theorem 3.5}]{RS}. Thus $$\lambda_{k}(T^a) \leq \lambda_{k}(P^a)=\left(\frac{k \pi}{\mathcal{L}(P^a)}\right)^2, \quad k=1,2,\cdots$$
Or using \textcolor{blue}{Theorem}\eqref{pendant}
$$\lambda_{k+1}(T^a) \leq \lambda_{k}(P^a)=\left(\frac{k \pi}{\mathcal{L}(P^a)}\right)^2, \quad k=1,2,\cdots$$
One can obtain these estimates with any path graph inside the metric tree, even an interval would give an estimate. However, we choose the longest path to get better estimates. The weak version of these bounds were established in \cite[\textcolor{blue}{Theorem 3.2}]{R} for standard conditions and in \cite[\textcolor{blue}{Theorem 3}]{ZS} for anti-standard conditions, in which a particular path graph $P$ was chosen which starts with the longest edge and ends at second largest edge. These estimates were further expressed in terms of average length of metric tree.
For standard conditions,
\begin{equation} \label{tree1}
\lambda_{k+1}(T^s) \leq \lambda_{k+1}(P^s)=\left(\frac{k \pi}{\mathcal{L}(P^s)}\right)^2 \leq \left(\frac{k \pi |E|}{2 \mathcal{L}(T^s)}\right)^2, \quad k=0,1,2,\cdots
\end{equation}
Similarly, for anti-standard conditions,
\begin{equation} \label{tree2}
\lambda_{k}(T^a) \leq \lambda_{k}(P^a)=\left(\frac{k \pi}{\mathcal{L}(P^a)}\right)^2 \leq \left(\frac{k \pi |E|}{2 \mathcal{L}(T^a)}\right)^2, \quad k=1,2,\cdots
\end{equation}
The following simple but useful relation was proved in \cite[\textcolor{blue}{Corollary 3.8} ]{PR}, which connects the eigenvalues of metric tree with standard and anti-standard vertex conditions.
\begin{equation}
\lambda_k(T^a)=\lambda_{k+1}(T^s), \quad k\ \in \mathbb{N}.
\end{equation}
We use these estimates to obtain some bounds on the spectrum of the Laplacian acting on a finite compact connected metric graph $\Gamma$ with standard and anti-standard vertex conditions, and then we will generalize these estimates for $\delta$ and $\delta'$- type vertex conditions. Most of the following estimates are equal in magnitude but are expressed in terms of different geometrical parameters of the metric graph $\Gamma$.
\\ For the anti-standard vertex conditions, we know that either attaching a pendant graph (particularly an edge), or attaching an edge between existing vertices lowers all eigenvalues \cite[\textcolor{blue}{Theorem 3.2 , 3.5} ]{RS}. Therefore,
\begin{equation*}
\lambda_{k}(\Gamma^a) \leq \lambda_k(T^a) \leq \lambda_k(P^a)=\left(\frac{k \pi \gamma}{\mathcal{L}(\Gamma^a)}\right)^2, \quad k=1,2,\cdots
\end{equation*}
Where $\gamma=\frac{\mathcal{L}(\Gamma^a)}{\mathcal{L}(P^a)}$.
If we let $\gamma=\mathcal{L}(\Gamma^a)-\mathcal{L}(P^a)$, then the same estimate obtained by the path graph can be rewritten as
\begin{equation*}
\lambda_{k}(\Gamma^a) \leq \lambda_k(P^a)=\left(\frac{k \pi }{\mathcal{L}(\Gamma^a)-\gamma}\right)^2, \quad k=1,2,\cdots
\end{equation*}
Note that the above bounds are equal to $\lambda_{k}(P^a)=\left(\frac{k \pi}{\mathcal{L}(P^a)}\right)^2$ in magnitude, but are expressed in terms of total length of the graph.
\\ Let $T^a$ be the tree graph inside $\Gamma^a$ with the longest length among all the tree inside $\Gamma^a$. Then we use the estimate \eqref{tree2} on $T^a$ to obtain an estimate on $\Gamma^a$.
$$\lambda_k(\Gamma^a) \leq \lambda_k(T^a) \leq \left(\frac{k \pi |E_T|}{2 \mathcal{L}(T^a)}\right)^2, \quad k=1,2,\cdots$$
Moreover, if the graph $\Gamma^a$ with $|E|$ number of edges is obtained by from metric tree $T^a$ by either attaching edges or loops of lengths $\ell_1, \ell_2, \cdots, \ell_m$, then
$$\lambda_k(\Gamma^a) \leq \left(\frac{k \pi (|E|-m)}{2 \left(\mathcal{L}(\Gamma^a)-\sum\limits_{i=1}^m \ell_i\right)}\right)^2, \quad k=1,2,\cdots$$
If $T^a_m$ is the maximal spanning tree inside $\Gamma^a$, then
$$\lambda_k(\Gamma^a) \leq \lambda_k(T^a_m) \leq \left(\frac{k \pi |E_{T_m}|}{2 \mathcal{L}(T^a_m)}\right)^2, \quad k=1,2,\cdots$$
The number of vertices in $T^a_m$ are equal to the number of vertices in $\Gamma^a$, and the number of edges in $T^a_m$ are one less than the number of vertices in $\Gamma^a$. Let $|V|$ denotes the number of vertices in $\Gamma^a$, and the graph $\Gamma^a$ can be obtained from $T^a_m$ by attaching pendant loops or multiple edges between existing vertices. Thus
\begin{equation}
\lambda_k(\Gamma^a) \leq \lambda_k(T^a_m) \leq \left(\frac{k \pi \gamma (|V|-1)}{2 \mathcal{L}(\Gamma^a)}\right)^2, \quad k=1,2,\cdots
\end{equation}
The above bounds can be expressed in terms of the first Betti number of the graph $\Gamma^a$. Since the first Betti number of maximal spanning tree is zero, and while constructing $\Gamma^a$ from $T^a_m$ each attachment of either loop or an edge between existing vertices increases the Betti number by one. Let $\beta$ be number of cycles in graph $\Gamma^a$ and let $\ell_1, \ell_2, \cdots, \ell_\beta$ be the number of required edges to obtain $\Gamma^a$ from $T^a_m$. Then,
$$\lambda_k(\Gamma^a) \leq \left(\frac{k \pi (|E|-\beta)}{2 \left(\mathcal{L}(\Gamma^a)-\sum\limits_{i=1}^m \ell_i\right)}\right)^2, \quad k=1,2,\cdots$$
In addition, if $\ell_1= \ell_2= \cdots= \ell_\beta$, then
$$\lambda_k(\Gamma^a) \leq \left(\frac{k \pi (|E|-\beta)}{2 \left(\mathcal{L}(\Gamma^a)- \beta \ell \right)}\right)^2, \quad k=1,2,\cdots$$
For different types of metric tree inside the graph $\Gamma^a$, we obtain different expression for bounds, some of these were equal in magnitude but expressed in different topological parameters. Now we chose a particular type of a metric tree inside $\Gamma^a$, and will use the estimate \eqref{tree2}. Let $T^a$ be the tree inside $\Gamma^a$ such that the number of edges and the length of each edge in $T^a$ and $\Gamma^a$ are equal. Then $\Gamma^a$ can be obtained by gluing some vertices of $T^a$ and each gluing produces a cycle. Thus
\begin{equation} \label{anti1}
\lambda_k(\Gamma^a) \leq \lambda_k(T^a) \leq \left(\frac{k \pi |E|}{2 \mathcal{L}(\Gamma^a)}\right)^2, \quad k=1,2,\cdots
\end{equation}
Similarly, for the graph $\Gamma^s$ with standard vertex condition the following bounds hold.
Let $P^s$ be the longest path inside $\Gamma^s$, then we attach the pendant edges, step by step, with standard condition to the vertices of $P^s$ unless we obtain a tree $T^s$ with number of edges equal to the number of edges $|E|$ of graph $\Gamma^s$. Now $\Gamma^s$ can be obtained by pairwise gluing of $\beta$ pair of vertices.
\begin{equation*}
\lambda_{k+1-\beta}(\Gamma^s) \leq \lambda_{k+1} (T^s) \leq \lambda_{k+1} (P^s)=\left(\frac{k \pi }{\mathcal{L}(P^s)}\right)^2, \quad k=1,2,\cdots,
\end{equation*}
$$\lambda_{k+1}(\Gamma^s) \leq \left(\frac{(k+\beta) \pi \gamma}{\mathcal{L}(\Gamma^s)}\right)^2, \quad k=1,2,\cdots$$
where $\gamma=\frac{\mathcal{L}(\Gamma^s)}{L(P^s)}$.
If we consider the longest metric tree $T^s$, then we can use the estimate \eqref{tree1},
$$\lambda_{k+1}(\Gamma^s) \leq \lambda_{k+1}(T^s) \leq \left(\frac{(k+\beta) \pi |E_T|}{2\mathcal{L}(T^s)}\right)^2=\left(\frac{(k+\beta) \pi |E_T|\gamma}{2\mathcal{L}(\Gamma^s)}\right)^2, \quad k=1,2,\cdots$$
If we choose a tree $T^s$ such that each vertex of $T^s$ is equipped with standard vertex conditions and the number of edges and the length of each edge is same as of $\Gamma^s$. However, the only difference between $T^s$ and $\Gamma^s$ is that $\Gamma^s$ can be obtained from $T^s$ by pairwise gluing of $\beta$ pair of vertices, then
\begin{equation} \label{standard2}
\lambda_{k+1}(\Gamma^s) \leq \left(\frac{(k+\beta) \pi |E|}{2 \mathcal{L}(\Gamma^s)}\right)^2
\end{equation}
If $T^s_m$ is the maximal spanning tree inside $\Gamma^s$, then
$$\lambda_{k+1}(\Gamma^s) \leq \left(\frac{(k+\beta) (|V|-1)\pi }{2 \left(\mathcal{L}(\Gamma^s)-\sum\limits_1^{m} \ell_i\right)}\right)^2.$$
Or
$$\lambda_{k+1}(\Gamma^s) \leq \left(\frac{(k+\beta) (|E|-\beta)\pi }{2 \left(\mathcal{L}(\Gamma^s)-\sum\limits_1^{m} \ell_i\right)}\right)^2.$$
Where $|E|$, $|V|$ and $\beta$ denotes the number of edges, number of vertices and the first Betti number of the graph $\Gamma^s$.\\
\begin{definition}
Let $\Gamma$ be a finite compact connected metric graph. Then diameter of $\Gamma$ is defined as,
$$d(\Gamma):=\max \{ d(x,y) : x,y \in \Gamma \}.$$
\end{definition}
Let $T^s$ be a finite, connected metric tree and let $d(T^s)$ denote the diameter of $T^s$, then the following inequality was proved in \cite[\textcolor{blue}{Theorem 3.4}]{R}
\begin{equation} \label{diameter1}
\lambda_{k+1}(T^s) \leq \left(\frac{k \pi}{d(T^s)}\right)^2.
\end{equation}
A similar inequality for $T^a$ was proved in \cite[\textcolor{blue}{Theorem 3}]{ZS},
\begin{equation} \label{diameter2}
\lambda_{k}(T^a) \leq \left(\frac{k \pi}{d(T^a)}\right)^2.
\end{equation}
In order to establish better estimates on the eigenvalues of metric trees $T^s$ and $T^a$, we require the following lemma which is also contained in \cite[\textcolor{blue}{Lemma 4.2}]{RL}. However, we provide a brief proof for the completeness of the presentation.
\begin{lemma}
Let $T$ be a metric tree of total length $\mathcal{L}(T)$ with $|E_p|\geq 2$ the pendant edges, then
\begin{equation}
\frac{2 \mathcal{L}(T)}{|E_p|} \leq d(T).
\end{equation}
\end{lemma}
\begin{proof}
We will show that the diameter of a metric tree is the distance between two pendant points.
Suppose that $d(\Gamma)=d(x_1,x_2)$ for some $x_1,x_2 \in T$. To prove by contradiction we assume that $x_1$ is not a pendant point of $T$, then $x_1$ is some interior point and the graph $T \backslash x_1$ has at least two connected components. Let $T_1$ denote of the component such that $x_2 \notin T_1$ and $x_1\neq y \in T_1$. Then $$d(y,x_2) > d(x_1,x_2)=d(T).$$
This holds because $T$ is a tree and any path inside $T$ that between $x_2$ and $y$ contains the point $x_1$. This contradicts our assumption of $d(x_1,x_2)=d(T)$.
Let $x_0$ be a middle point of a shortest path $P$ between $x_1$ and $x_2$. Thus,
$$d(x_1,x_0)=\frac{d(T)}{2}=d(x_0,x_2).$$
Since the tree $T$ has $E_p$ pendant edges, therefore there are exactly $E_p$ path of length at most $d(x_1,x_0)$, that connect the point $x_0$ and a pendant vertex of $T$. Moreover, these paths covers the metric tree $T$, thus
$$\mathcal{L}(T) \leq \sum\limits_{v \text{ is a pendant}} d(x_0,v) \leq \sum\limits_{v \text{ is a pendant}} d(x_0,x_1)= \frac{E_p d(T)}{2}.$$
\end{proof}
The bounds \eqref{diameter1} and \eqref{diameter2} implies that
\begin{equation} \label{newtree1}
\lambda_{k+1}(T^s) \leq \left( \frac{k \pi |E_p|}{2 \mathcal{L}(T^s)} \right)^2
\end{equation}
and,
\begin{equation} \label{newtree2}
\lambda_{k}(T^a) \leq \left( \frac{k \pi |E_p|}{2 \mathcal{L}(T^a)} \right)^2.
\end{equation}
Thus, we get improved estimates as compared to the estimates \eqref{tree1} and \eqref{tree2} on the eigenvalues of a metric trees $T^s$ and $T^a$. Therefore, the above listed bounds can be improved when the estimates \eqref{tree1} and \eqref{tree2} are replaced by \eqref{newtree1} and \eqref{newtree2}. \\
We now use previously established estimates on metric graph with standard conditions to obtain some estimates on the metric graph $\Gamma^{\delta}$. For two same underlying metric graph $\Gamma^{\delta}_\alpha$ and $\Gamma^{\delta}_{\alpha'}$ with strengths $\alpha$ and $\alpha'$ at some vertex $v_0$, respectively, the following interlacing inequality holds.
$$\lambda_k(\Gamma^{\delta}_\alpha) \leq \lambda_k(\Gamma^{\delta}_{\alpha'}) \leq \lambda_{k+1}(\Gamma^{\delta}_\alpha).$$
Let $\Gamma^{\delta}$ be finite compact connected metric graph, and let $|V^p|$ and $|V^n|$ denotes the number of vertices of $\Gamma^{\delta}$ with positive and negative strengths, respectively, and let $\Gamma^s$ be the same underlying metric graph with standard conditions. If $\Gamma^p$ is a graph obtained from $\Gamma^s$ by replacing the interaction strengths from zero to positive at those vertices where positive strengths were specified in $\Gamma^{\delta}$. Now $\Gamma^{\delta}$ can be obtained from this new graph $\Gamma^p$ by assigning negative strengths at some vertices, where negative strengths were assigned in $\Gamma^{\delta}$. Then following inequalities holds.
$$\lambda_k(\Gamma^s) \leq \lambda_k(\Gamma^p) \leq \lambda_{k+|V^p|}(\Gamma^s),$$
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^p) \leq \lambda_{k+|V^n|}(\Gamma^{\delta})$$
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^p) \leq \lambda_{k+|V^p|}(\Gamma^s) \leq \left(\frac{(k+\beta+|V^p|-1) \pi |E|}{2 \mathcal{L}(\Gamma^{\delta})}\right)^2 $$
Where last inequality is due to \eqref{standard2}. If we replace the zero strengths with negative strengths first and then with the positive strengths even then we obtain the same bound.
$$ \lambda_k(\Gamma^n) \leq \lambda_k(\Gamma^s) \leq \lambda_{k+|V^n|}(\Gamma^n)$$
$$\lambda_k(\Gamma^n) \leq \lambda_k(\Gamma^{\delta}) \leq \lambda_{k+|V^p|}(\Gamma^n) $$
\begin{equation}
\lambda_k(\Gamma^{\delta}) \leq \lambda_{k+|V^p|}(\Gamma^n) \leq \lambda_{k+|V^p|}(\Gamma^s) \leq \left(\frac{(k+\beta+|V^p|-1) \pi |E|}{2 \mathcal{L}(\Gamma^{\delta})}\right)^2
\end{equation}
The other way to establish an upper estimate for eigenvalues of the metric graph $\Gamma^{\delta}$ is using the application of the corresponding flower graph. Let $\Gamma^{\delta}$ be the metric graph with total interaction strength $\alpha=\sum\limits_{m=1}^{|V|} \alpha_m$, and $\Gamma^s$ be same graph with standard conditions at each vertex. Let $\Gamma^s_f$ and $\Gamma^{\alpha}_f$ be the corresponding flower graph with strengths zero and $\alpha$.\\
If $\alpha >0$, then
$$\lambda_k(\Gamma^s) \leq \lambda_k(\Gamma^s_f) \leq \lambda_{k+|V|-1}(\Gamma^s)$$
$$\lambda_k(\Gamma^s_f) \leq \lambda_k(\Gamma^\alpha_f) \leq
\lambda_{k+1}(\Gamma^s_f)$$
Now restore the vertices of $\Gamma^{\delta}$ from $\Gamma^\alpha_f$, then
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^\alpha_f) \leq \lambda_{k+|V|-1}(\Gamma^{\delta})$$
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^\alpha_f) \leq \lambda_{k+1}(\Gamma^s_f) \leq \lambda_{k+|V|}(\Gamma^s) \leq \left(\frac{(k-1+\beta+|V|) |E| \pi}{2 L(\Gamma^{\delta})}\right)^2. $$
If $\alpha \leq0$,
$$\lambda_k(\Gamma^s) \leq \lambda_k(\Gamma^s_f) \leq \lambda_{k+|V|-1}(\Gamma^s)$$
$$\lambda_k(\Gamma^\alpha_f) \leq \lambda_k(\Gamma^s_f) \leq
\lambda_{k+1}(\Gamma^\alpha_f)$$
Now restore the vertices of $\Gamma^{\delta}$ from $\Gamma^\alpha_f$, then
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^\alpha_f) \leq \lambda_{k+|V|-1}(\Gamma^{\delta})$$
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^\alpha_f) \leq \lambda_{k}(\Gamma^s_f) \leq \lambda_{k+|V|-1}(\Gamma^s) \leq \left(\frac{(k-2+\beta+|V|) |E| \pi}{2 L(\Gamma^{\delta})}\right)^2. $$
Let $\Gamma_a^{\delta'}$ be the finite compact connected metric graph equipped with $\delta'$-type conditions and possibly anti-standard conditions at some vertices. Let $\Gamma^a$ be the same underlying metric graph as $\Gamma_a^{\delta'}$ with anti-standard vertex conditions at each vertex. Then, with the help of \textcolor{blue}{Theorem} \eqref{strength}, and the estimate \eqref{anti1}, we obtain
$$\lambda_k(\Gamma_a^{\delta'}) \leq \lambda_k(\Gamma^a) \leq \left(\frac{k |E|\pi}{2 L(\Gamma_a^{\delta'})}\right)^2. $$
Let $B$ denote a finite compact connected bipartite metric graph, and let $B^a$ and $B^s$ be the bipartite metric graphs equipped with anti-standard and standard vertex conditions at each vertex of both graphs, respectively. The following relation was established in \cite[\textcolor{blue}{Theorem 3.6}]{PR}.
\begin{equation}
\lambda_{k+\beta}(B^a) = \lambda_{k+1}(B^s).
\end{equation}
We use this relation and the \textcolor{blue}{Theorem} \eqref{strength} to obtain a relation between the eigenvalues of a graph with $\delta'$-condition and the same graph equipped with $\delta$-conditions.
Let $B^{\delta'}$ be the bipartite metric graph and $B^s$ be same underlying bipartite metric graph equipped with standard vertex conditions. Then from \textcolor{blue}{Theorem} \eqref{strength}, we know that either increasing or decreasing the strength at vertex from zero, decreases the eigenvalues. Thus,
$$\lambda_{k+\beta}(B^{\delta'}) \leq \lambda_{k+\beta}(B^a)=\lambda_{k+1}(B^s) \leq \lambda_{k+\beta+|V|}(B^{\delta'}).$$
Let $B^{\delta}$ be a bipartite graph with non-negative strengths at each vertex, then
$$\lambda_{k+\beta}(B^{\delta'}) \leq \lambda_{k+1}(B^{\delta}).$$
Let $|V^p|$ and $|V^n|$ denote the number of vertices in $B^{\delta}$ with positive and negative strengths, respectively, and $B^s$ be the same underlying metric graph. Then the graph $B^{\delta}$ can be obtained from $B^s$ in two steps. First, obtain a graph $B^p$ from $B^s$ by assigning positive strengths to those vertices where positive strengths are assigned in $B^{\delta}$, and then obtain $B^{\delta}$ by assigning negative strengths to some vertices of $B^p$ wherever is required.
Thus
$$\lambda_{k+1}(B^{s}) \leq \lambda_{k+1}(B^p)\leq \lambda_{k+1+|V^p|}(B^s) $$
$$\lambda_{k+1}(B^{\delta}) \leq \lambda_{k+1}( B^p) \leq \lambda_{k+1+|V^n|}(B^{\delta})$$
$$\lambda_{k+\beta}(B^{\delta'}) \leq \lambda_{k+1}(B^s) \leq \lambda_{k+1}(B^p) \leq \lambda_{k+1+|V^n|}(B^{\delta}).$$
Therefore,
\begin{equation}
\lambda_{k+\beta}(B^{\delta'}) \leq \lambda_{k+1+|V^n|}(B^{\delta}) \quad \text{and} \quad \lambda_{k+1}(B^{\delta}) \leq \lambda_{k+\beta+|V^p|}(B^a).
\end{equation}
Moreover, for any metric graph $\Gamma$, not necessarily bipartite, one can obtain several bipartite metric graphs $B$ by creating vertices of degree two equipped with standard conditions at some interior points of $\Gamma$. Therefore, for any metric graph $\Gamma^{\delta}$ with non-negative strengths at each vertex, the following relation hold.
\begin{equation}
\lambda_{k+\beta}(B^{\delta'}) \leq \lambda_{k+\beta}(B^a)=\lambda_{k+1}(B^s)=\lambda_{k+1}(\Gamma^s) \leq \lambda_{k+1}(\Gamma^{\delta}).
\end{equation}
Note that in the above relation the total length of the graphs $B^{\delta'}$ and $\Gamma^{\delta}$ are equal. That is, $\mathcal{L}(B^{\delta'})=\mathcal{L}(\Gamma^{\delta}).$
Let $\Gamma^{\delta}$ and $\Gamma^{\delta'}$ be same underlying metric graphs with strengths $\alpha_m$ and $\alpha'_m$, respectively.
Let $\Gamma^{\delta}_f$ be the corresponding flower graph obtained from $\Gamma^{\delta}$, then $\Gamma^{\delta}_f$ can be turned into a $\mathfrak{B}^s$ bipartite graph by introducing the degree two vertices at its leave. Therefore,
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^{\delta}_f) \leq \lambda_{k+1} (\Gamma^s_f)=\lambda_{k+1}(\mathfrak{B}^s)=\lambda_{k+\beta}(\mathfrak{B}^a)$$
Let us impose the Dirichlet conditions at degree two vertices of $\mathfrak{B}^a$ to obtain a graph $\mathfrak{B}^a_d$. Now again impose the standard conditions at the Dirichlet vertices in $\mathfrak{B}^a_d$. Thus
\begin{equation}
\lambda_{k+\beta}(\mathfrak{B}^a) \leq \lambda_{k+\beta} (\mathfrak{B}^a_d) \leq \lambda_{k+2\beta}(\mathfrak{B}^a_s)= \lambda_{k+2|E|}(\Gamma^a_f) \leq \lambda_{k+2|E|+1}(\Gamma^{\delta'}_f).
\end{equation}
Where $\Gamma^{\delta'}_f$ has strength $\alpha=\sum\limits_{m=1}^{|V|} \alpha'_m$, and
if for all $m=1,2,\cdots,|V|$, the strengths $\alpha'_m \geq 0$ in $\Gamma^{\delta'}$, then restoring the vertices of $\Gamma^{\delta'}$ from $\Gamma_f^{\delta'}$ implies that
$$\lambda_k(\Gamma^{\delta}) \leq \lambda_{k+2|E|+1}(\Gamma^{\delta'}_f) \leq \lambda_{k+2|E|+1}(\Gamma^{\delta'}).$$
If $|V^{p'}|$ and $|V^{n'}|$ denote the number vertices in $\Gamma^{\delta'}$ with positive and negative strengths, respectively. Then,
\begin{align*}
\lambda_k(\Gamma^{\delta}) \leq \lambda_k(\Gamma^p) &\leq \lambda_{k+|V^p|}(\Gamma^s)\\
&= \lambda_{k+\beta}(B^s)\\
&= \lambda_{k+\beta+|V^p|-1} (B^a)\\
&\leq \lambda_{k+\beta+|V^p|+|V^{p'}|+|V^{n'}|-1}(B^{\delta'})
\end{align*}
Since for the graph endowed with $\delta'$-type conditions, attaching a pendant edge, attaching an edge between existing vertices, or gluing two vertices with at least one equipped with anti-standard vertex conditions lowers all eigenvalues \cite[\textcolor{blue}{Theorem 3.2, 3.5 and 4.2}]{RS}. Therefore, for any metric graph $\Gamma^{\delta'}$ with length $\mathcal{L}(\Gamma^{\delta'}) \geq \mathcal{L}(\tilde{\Gamma}^{\delta})$, we obtain,
$$\lambda_{k+\beta}(\Gamma^{\delta'}) \leq \lambda_{k+\beta}(B^{\delta'}) \leq \lambda_{k+1+|V^n|}(\tilde{\Gamma}^{\delta}).$$
|
\section{Introduction}
Magnetic materials with high spin polarization have always been the centre of attraction for spintronics based research. Heusler alloys are indeed one of the most promising class of such materials. They have shown immense potential for spintronics application by virtue of their tunable and diverse electronic/magnetic properties. Half-metallic ferromagnets (HMF)\cite{PhysRevLett.50.2024} and spin gapless semiconductors (SGS)\cite{WANG20201,doi:10.1063/5.0028918} hold a special place in this class of materials due to their unique band structure and high spin polarization. Due to this, intensive efforts have been put in to study the magnetic and electronic properties of these materials. First-principles calculations predict half-metallicity with 100\% spin-polarization for several pure Heusler alloys. It is, however, important to note that most of the ab-intio investigations are based on defect free or stoichiometric composition of the concerned material. On the contrary, Heusler alloys are know to be prone to disorder and when synthesized, they show either partial B2-type or A2-type structural disorders, which may destroy the half-metallicity of the concerned alloy. This fundamental discrepancy is mainly because of a lack of microscopic understanding of the structural and electronic/magnetic properties of such alloys. Thus, it is quite important to resolve the above mentioned discrepancy between theoretical and experimental findings in order to design new plausible Heusler-based materials for potential applications.
In this paper, we report a detailed study of structure, electronic, magnetic and transport properties of a equiatomic quaternary alloy NbVTiAl. Experimental findings are corroborated with the results obtained from first principles calculations. NbVTiAl is a Heusler-type alloy whose magnetic properties deviate from the Slater Pauling (SP) rule.\cite{doi:10.1063/1.4805063} Magnetic measurements yield an almost vanishing magnetization ($\sim 10^{-3} \mu_B/f.u.$), while an ordered NbVTiAl is expected to yield a net moment of 1.0 $\mu_B$/f.u. as per SP rule. Our ab-initio band structure calculation also confirms a net moment of 0.8 $\mu_B$/f.u. with a high spin polarization of 90\% (nearly half metallic) for a completely ordered NbVTiAl structure. This is a clear disagreement between the theoretical predictions and experimental results in terms of magnetic properties. To gain insight into the origin of this discrepancy, we have investigated the effect of disorder on the magnetic and electronic properties of NbVTiAl using state of art ab-intio calculations. To the best of our knowledge, the present study is one out of few, which correctly highlights the significance of disorder in explaining the correct magnetic and electronic state of a quaternary Heusler alloy within the DFT framework.
The room temperature X-ray diffraction (XRD) data of NbVTiAl confirms an A2 type structural disorder, as confirmed by the absence of (111) and (200) superlattice reflection peaks. The quenching of magnetic moment was successfully explained by simulating the experimentally observed A2 disordered phase. The disorder also causes a sharp decrease in the spin polarization. Magnetization measurements also reveal a high Curie temperature ($>$ 800 K). The transport measurements confirm a metallic-like behaviour, as also supported by our first principles calculation for A2 phase of NbVTiAl. Hall resistivity shows a linear dependence on the field with a positive slope. The calculated value of total density of states simulated for experimentally observed A2 structure was found to be comparable with that obtained from the heat capacity measurements. The detailed explanation in the present work sets up a stage for quaternary Heusler alloys in general, where disorder can play a decisive role in determining the correct magnetic and electronic properties and hence the reduced/quenched moment.
\section{Experimental details}
\label{exp}
Arc melting technique was used to prepare polycrystalline sample of NbVTiAl, using the stoichiometric amounts of high pure (at least 99.9\% purity) constituent elements in argon atmosphere. To reduce any contamination in the chamber, a Ti ingot was used as an oxygen getter. The ingots formed were flipped and melted several times for better homogeneity. Room temperature X-ray diffraction patterns were taken using Cu K$_\alpha$ radiation with the help of Paralytic X-pert diffractometer. Phase purity and crystal structure analysis were done using Fullprof suite\cite{RR}. Magnetization measurements were done using a vibrating sample magnetometer (VSM) attached to the physical property measurement system (PPMS) (Quantum design) for fields up to 40 kOe. Thermo-magnetic curves in the higher temperature range were obtained using a VSM attached with high temperature oven, under a field of 500 Oe.
For resistivity and Hall effect measurements, polycrystalline alloys were cut into rectangular shape using a diamond wheel saw to dimensions nearly 1 mm thickness, 4.87 mm wide and length 8.62 mm. These rectangular pieces were then polished in order to further reduce the thickness (up to 0.52 mm). Electrical resistivity measurements (using dc-linear four probe method by applying a 5 mA current) have been carried out using PPMS. In this method, four contacts were made on the sample along a line. A constant current was allowed to flow through the two outermost contacts. The voltage was measured between the two inner contacts and the gap between them is taken as the sample length. In a four-probe geometry (see Fig. S2(a) in supplementary \cite{RR3}), the resistivity can be calculated as, $\rho$=RxA/l, where, R is the resistance, A is area of cross section through which the current passes (A =width x thickness), and l is the voltage lead separation ($V_+V_-$ = 5.76 mm). Hall measurements were carried out using five probe measurement, an additional module attached to the PPMS. In five-probe geometry, a fifth voltage lead is attached in parallel to one of the other voltage leads (see Fig. S2(b) in supplementary \cite{RR3}) to get an accurate value for the Hall potential. While the magnetic field is turned off, a potentiometer between the two leads is used to null the offset that is due to sample resistance. Then, once a field is applied, the measured potential drop gives only the Hall potential, plus components due to instrumental non-idealities that can be removed with AC filtering technique. 150 mA current was applied along the direction of length while the developed transverse voltage was measured using the leads connected on sides of the sample. The Hall Effect measurements were carried out in an applied magnetic field of 5 T in five probe contact geometry at different temperatures. The electrical contacts were made using silver epoxy. Specific heat measurements (by relaxation calorimetry) have been carried out using PPMS.
\section{Computational details}
\label {comp}
First principles calculations of NbVTiAl were performed using spin polarized density functional theory (DFT)\cite{hohenberg1964inhomogeneous} as implemented within the Vienna ab initio simulation package (VASP),\cite{kresse1996efficient,kresse1996efficiency,kresse1993ab} with a projected augmented-wave (PAW) basis.\cite{kresse1999ultrasoft} Pseudopotential formalism with Perdew, Burke, and Ernzerhof (PBE) exchange-correlation functional\cite{perdew1996generalized} was used for all the calculations. A 24$\times$24$\times$24 $\Gamma$-centered K-mesh was used to perform the Brillouin zone(BZ) integration. A plane wave energy cut-off of 500 eV was used for all the calculations. All the structures are fully relaxed with total energies (forces) converged to values less than 10$^{-6}$ eV (0.01 eV/\AA). A 64 atom special quasirandom structure (SQS)\cite{zunger1990special} is generated to simulate the A2 disordered phase of NbVTiAl. SQS is a periodic structure known to mimic the random correlations accurately for a given alloy concentration. Alloy Theoretic Automated Toolkit (ATAT) \cite{van2013efficient} was used to generate the SQS structures. BZ integration for SQS case was performed using $6\times6\times6$ k-mesh.
\section{Results and Discussion}
\label{result}
\subsection{Structural Analysis}
\label{result1}
For any quaternary Heusler alloy (QHA) XX$'$YZ, there are 4 inequivalent Wyckoff positions, namely 4a, 4b, 4c and 4d. If one fix the position of Z-atom to be at 4a site, then there exists three possible non degenerate crystallographic configurations,\cite{RANI2019165662}
\begin{enumerate}
\item Type I $\rightarrow$ X at 4d, X$'$ at 4c and Y at 4b site,
\item Type II $\rightarrow$ X at 4b , X$'$ at 4c and Y at 4d site,
\item Type III $\rightarrow$ X at 4c , X$'$ at 4b and Y at 4d site
\end{enumerate}
For the configuration considering X at 4b, X$'$ at 4c, Y at 4d and Z at 4a Wyckoff positions, the structure factors for the superlattice $(111)$ and $(200)$ reflections can be written as,\cite{PhysRevB.99.104429}:\\
\begin{equation}
\mathrm{F_{111}} = 4[\mathrm{(f_Y-f_Z)-i(f_X-f_{X'})}]
\end{equation}
\begin{equation}
\mathrm{F_{200}} = 4[\mathrm{(f_Y+f_Z)-(f_X+f_{X'})}]
\end{equation}
These equations imply that, the intensity of the (111) peak should reduce or disappear in presence of B2 disorder (X and X$^\prime$ \& Y and Z atoms are randomly distributed). For a completely disordered structure i.e., A2-type (all the four atoms are randomly distributed at four Wyckoff positions), both the superlattice peaks $(111)$ and $(200)$ should be absent.
Figure \ref{XRD} shows the rietveld refined room temperature powder XRD pattern of NbVTiAl. It is clear that the alloy exhibits a cubic crystal structure. The absence of superlattice reflection peaks (111) and (200) indicates complete A2 disorder. The lattice parameter as deduced from the refinement considering the A2-type structure (Im-3m, with all the atoms sharing 2a site according to the stoichiometry) was found to be 3.186 $\mathrm{\AA}$. The $\chi ^2$ value was found to be 2.06. The chemical nature of the constituent elements greatly affects their ordering. In Heusler alloys, the least electronegative element occupies octahedral site.\cite{Graf20111} However, if the constituent elements have comparable electronegativities, they compete to occupy the same site, and hence cause the disorder. A possible reason of A2-type disorder in NbVTiAl could be the nearly similar electronegativities of the constituent elements (the electronegativity of Nb, V, Ti, and Al are 1.60, 1.63, 1.54, and 1.61 respectively) which make all the four sites equally probable to be occupied by the four atoms.
\begin{figure}[hbt]
\centering
\includegraphics[width=0.7\linewidth]{Fig1}
\caption{Rietveld refined XRD pattern of NbVTiAl at room temperature. }
\label{XRD}
\end{figure}
\begin{figure}[hbt]
\centering
\includegraphics[width=\linewidth]{Fig2}
\caption{For NbVTiAl (a) Isothermal magnetization curve at 5 K, and (b) Temperature dependence of magnetization at H=500 Oe field.}
\label{VMHT}
\end{figure}
\subsection{Magnetic Properties}
For an ordered quaternary Heusler alloy XX$'$YZ, when X$'$ and Y have less than half-filled d-orbitals and Z = Al or Si, the magnetization obeys the so-called Slater-Pauling (SP) rule,\cite{doi:10.1063/1.4805063} according to which the net magnetization of the compound is given by,
\begin{equation}
M = (N_v - 18)\ \mu_B/f.u.
\end{equation}
where N$_v$ is the total number of valence electrons in the alloy.
NbVTiAl has 17 valence electrons and hence expected to have a magnetic moment of 1 $\mu_B/f.u.$, as per the SP rule. Figure \ref{VMHT}(a) shows the M-H curve for NbVTiAl measured at 5 K. It clearly indicate a very small net magnetization ($\sim 10^{-3} \mu_B/f.u.$). However, the initial slope of the M-H loop clearly indicates ferro/ferri magnetic nature of the alloy. These observations along with the M-T data (discussed below) are indicative of nearly compensated ferrimagnetism in this alloy. It should be noted that in the case of an antiferromagnet, the compensated magnetic moment is due to the same kind of ions present at different sites, whereas, FCF materials usually contain three or more magnetic ions with their moments aligned in such a way that the total magnetization is zero. This claim, however, can only be established using atom projected magnetic states determined using advanced experimental probes such as X-ray magnetic circular dichroism (XMCD) or neutron diffraction, which is beyond the scope of the present work. However, in order to get a better insight, we have carried out theoretical simulations for various magnetic configurations in Y-type ordered structures as well as for random disordered structures. We showed that this deviation in moment possibly arise due to the random disorder (see Section \ref{theo} for detailed analysis of theoretical results).
Figure \ref{VMHT}(b) shows the thermo-magnetic (M-T) curve (measured in ZFC mode) in the temperature range 5-1000 K recorded under an applied field of H = 500 Oe. This curve clearly indicates a high magnetic ordering temperature ($>$ 800 K).
\subsection{Transport Properties}
\subsubsection{Longitudinal Resistivity}
{\color{black}{Figure \ref{RHC}(a) shows the temperature (T) dependence of the longitudinal electrical resistivity of NbVTiAl at 0 and 50 kOe. In the moderately high T range, resistivity increases with temperature revealing metallic nature with a very small negative MR. In case of a metal, the resistivity is generally dominated by electron-phonon scattering which accounts for the linear dependence at high temperatures and saturation near residual resistivity at low temperatures. In the lower temperature region, T $<$ 25 K, the resistivity is almost constant and thus is independent of temperature. A similar behavior has also been observed in other half-metallic Heusler alloys such as CoRhMnGe,\cite{PhysRevB.96.184404} CoRuFeSi,\cite{BAINSLA2015631} Co$_2$Fe$_{1-x}$Cr$_x$Si\cite{PhysRevApplied.10.054022}, Co$_2$FeSi\cite{PhysRevLett.110.066601} etc. and was attributed to the absence of spin-flip scattering. It should be noted that, due to the presence of gap in the minority sub-band, spin-flip scattering is not possible in half-metals. For T $>$ 50 K, the resistivity increases almost linearly with T, which arises from the electron-phonon scattering.}}
\begin{figure}[hbt]
\centering
\includegraphics[width=0.8\linewidth]{Fig3}
\caption{(a) Temperature dependence of longitudinal electrical resistivity($\rho$) in 0 and 50 kOe fields, (b) Hall resistivity ($\rho_{xy}$) vs. magnetic field (H), measured at 10 K, along with the linear fit (c) Temperature dependence of carrier concentration (n), and (d) Heat capacity vs. T curve for high T range, up to 250 K. Inset of (d) shows the (C/T) vs.$\mathrm{T^2}$ plot, in low temperature range for NbVTiAl.}
\label{RHC}
\end{figure}
\subsubsection{Hall Effect}
Figure \ref{RHC}(b) shows the Hall resistivity $\rho_{xy}$ vs. field (H) at 10 K for NbVTiAl. Notably, $\rho_{xy}$ increases almost linearly with the field. The positive slope of the $\rho_{xy}$ versus H curve indicates that holes are the majority charge carriers. The carrier concentration 'n' was calculated from the slope ($R_H$) {of $\rho_{xy}$ versus H curves measured at different temperatures using $R_H = 1/(ne)$. Figure \ref{RHC}(c) shows the variation of n with temperature. Clearly, the order of magnitude of carrier concentration is $\sim 10^{22}$ $cm^{-3}$.
\subsubsection{Heat capacity measurements}
In the low temperature range, the heat capacity of a ferromagnetic material with high ${T_C}$ can be described by the Sommerfeld model\cite{PhysRevB.83.235211}(magnetic excitations have insignificant contribution), according to which:
\begin{equation}
C(T)=\gamma T + \beta T^3
\label{CT}
\end{equation}
where, the first and second terms corresponds to the electronic and lattice contributions to the total heat capacity respectively. Here, $\gamma$ is the Sommerfeld coefficient representing the electronic part and $\beta$ is the lattice coefficient. Figure \ref{RHC}(d) shows the C vs. T plot in zero field, while the inset shows the $T^2$ dependence of measured heat capacity (C/T). Evidently, C/T vs. $\mathrm{T^2}$ shows a linear behavior, and the slope and the intercept of the curve correspond to the value of $\gamma$ and $\beta$ of Eq. (\ref{CT}) respectively. In the free electron model, the value of $\gamma$ corresponds to the density of states (DoS) at the Fermi level ($E_F$) according to the relation \cite{met7100414}
\begin{equation}
N(E_F)=\frac{3\gamma}{\pi^2 N_A {k_B}^2}
\end{equation}
where, $N_A$ is the Avogadro number and $k_B$ is the Boltzmann constant.
Also, from the value of $\mathrm{\beta}$, the Debye temperature can be calculated using the following relation\cite{met7100414}
\begin{equation}
\theta_D = \frac{234ZR}{\beta}
\end{equation}
where, $R$ is the universal gas constant and $Z$ is the number of atoms per formula unit.\\
A fitting of C/T vs. $T^2$ curve with Eq. (\ref{CT}) gives $\gamma$ = 16.54 $\mathrm{m J\ mole^{-1} K^{-2}}$ and $\beta$ = 0.23 $\mathrm{m J\ mole^{-1} K^{-4}}$. The calculated value of density of states $N(E_F)$ from the extracted Sommerfeld constant ($\gamma$) is found to be 7.0 states $\mathrm{eV^{-1} f.u.^{-1}}$, while $\theta_D$ is found to be 323 K. It should be noted that, these estimates of $N(E_F)$ and $\theta_D$ are completely based on the free electron model and they only guide us to get a qualitative trend of these numbers as compared to other similar HMFs.
\subsection{Theoretical Results}
\label{theo}
As already mentioned, quaternary Heusler alloys generally crystallise in Y-type structure and there are three possible non-degenerate configurations (namely, Type I, Type II and Type III) depending on the atomic positions of constituent atoms. Total energy ab-initio calculations revealed that Type II configuration in the ferrimagnetic state is energetically the most stable configuration with a net magnetic moment of 0.80 $\mathrm{\mu_B}$/f.u. (The structural optimisation details of the ordered Y-type structure of NbVTiAl can be found in the supplementary material\cite{RR3}.) Also, the calculated spin-polarized band structure and DoS for the energetically most stable, Type II configuration of NbVTiAl alloy confirms a nearly half metallic state with high spin polarization.(See Fig. S1 of supplementary material \cite{RR3}.)
\begin{figure}[hbt]
\centering
\includegraphics[width=0.8\linewidth]{Fig4}
\caption{Spin polarized density of states for (a) completely disordered A2 structure, and two swap disordered configurations namely (a) 1V-1Al and (b) 1Nb-1V \& 1V-1Al of NbVTiAl. }
\label{dos_dis}
\end{figure}
\begin{table*}[hbt]
\centering
\caption{Theoretically optimized lattice constants (in \AA), total energy (E) and total magnetic moment of various swap disordered and A2 disordered (SQS) configurations of NbVTiAl . Results for the purely ordered structure is also given for comparison. }
\begin{tabular}{c| c| c| c}
\hline \hline
System & Lattice constants (a/b/c) & E (eV/atom) &$\mathrm{m_{total}}$($\mu_B$/f.u.) \\
\hline
Completely Ordered & 6.35\ /6.35\ /6.35 & -7.874 & 0.80\\
\hline
1Nb-1Al & 6.35\ /6.35\ /6.35 & -7.869 & 0.50 \\
2Nb-2Al& 6.37\ /6.37\ /6.37 &-7.852 &0.28\\
\hline
1Ti-1Al & 6.35\ /6.35\ /6.35 & -7.846 & 0.86\\
2Ti-2Al& 6.38\ /6.35\ /6.35 &-7.817 &0.90\\
\hline
1V-1Al & 6.34\ /6.34\ /6.34 & 7.849& 0.02\\
2V-2Al& 6.33\ /6.33\ /6.37 & 7.816 &-0.10\\
\hline
1Nb-1Ti & 6.36\ /6.36\ /6.36 & -7.853 & 0.99\\
2Nb-2Ti& 6.36\ /6.36\ /6.36 & -7.831&0.88\\
\hline
1Nb-1V & 6.35\ /6.35\ /6.35 &-7.863 & 0.64\\
2Nb-2V& 6.35\ /6.35\ /6.37& -7.855&0.50\\
\hline
1V-1Ti & 6.35\ /6.35\ /6.35 &-7.871 & 0.51\\
2V-2Ti& 6.35\ /6.35\ /6.35& -7.865 &0.43\\
\hline
1Nb-1Al \& 1Ti-1Al & 6.34\ /6.36\ /6.36 &-7.838 & 0.61\\
\hline
1Nb-1V \& 1V-1Al & 6.36\ /6.36\ /6.35 & -7.841 & 0.01\\
\hline
1Nb-1Ti \& 1V-1Al & 6.36\ /6.36\ /6.35 & -7.833& 0.03\\
\hline
A2 disorder & 6.36\ /6.36\ /6.35 & -7.837& 0.05\\
\hline \hline
\end{tabular}
\label{MM}
\end{table*}
The simulated net magnetic moment (0.80 $\mathrm{\mu_B}$/f.u.) for ordered NbVTiAl agrees fairly well with the Slater-Pauling rule (1.0 $\mathrm{\mu_B}$/f.u.) but goes in complete contrast with the experimental observation where the alloy showed almost negligible ($\sim 10^{-3} \mu_B/f.u.$) net moment.
Such a discrepancy may arise due to several reasons such as long range magnetic ordering, structural disorder \cite{RANI2019165662, PhysRevB.60.13237, doi:10.1063/1.4998308, FENG20157, PhysRevB.66.104429} etc. We have indeed checked several magnetic orders including antiferromagnetic (AFM), 2$^{nd}$ neighbour ferrimagnetic etc., but all of them are either much higher in energy as compared to the ferrimagnetic Type II configuration, or of competing energy but with finite net moment or non-metallic electronic structure, in complete disagreement with the experimental findings. As such, these calculations could not reconcile the discrepancy. The other option left was to explore the effect of substitutional disorder on the electronic and magnetic properties. As demonstrated in Sec. \ref{result1}, our XRD study revealed a A2 type disorder in NbVTiAl i.e., all the four Wyckoff positions are equally probable to be randomly occupied by four atoms Nb, V, Ti and Al.
To understand the experimental observations better, we dig deep into the local atomic arrangement of the system and discuss its relevance to the observed net magnetic moment. To do this, we generated a completely disordered A2 structure of NbVTiAl and simulated the same to analyze the repercussions associated with it, on the system's magnetic properties. Interestingly, this structure yields an almost negligible ($\sim$0.05 $\mu_B$/f.u.) net moment for NbVTiAl.
This A2 phase remains energetically very close to the ordered Type II configuration (within 37 meV/atom) (see Table \ref{MM}). It is expected that due to disorder, new states can emerge near the Fermi level, which can further reduces the spin polarization. Interestingly, the A2 disordered phase of NbVTiAl indeed follows this expectation and correctly confirms a metallic phase for NbVTiAl, as observed experimentally. This is clear from the spin-polarized density of states, as shown in Fig. \ref{dos_dis}(a). Also, the calculated value of total density of states $N(E_F)$ for the SQS structure was found to be $\sim$ 4.90 states $\mathrm{eV^{-1} f.u.^{-1}}$ which is comparable to the value ($\sim$7.0 states $\mathrm{eV^{-1} f.u.^{-1}}$) obtained from the extracted Sommerfeld constant (using heat capacity measurements).
To get a better insight into the actual reason for negligible moment for A2 disordered case, we first look at the moments of the individual atoms. We found that there are a few V atoms with moment as high as $\sim$0.4 $\mu_B$, but there are also nearby atomic sites involving V, Ti and Nb which gives small compensating magnetic moments. Overall, it can be inferred that a completely disordered environment can effectively induce randomly oriented magnetic moments which, when averaged out, collectively yields zero net moment. Though it can be concluded that a complete random disorder statistically explains the vanishing net moment, we try to provide more material specific discussion in the following paragraph, as the percentage of disorder might be different in the actual, synthesized sample. In other words, we try to understand the effect of neighboring environment on the local moment of each individual atom.
NbVTiAl possesses a crystal structure with four interpenetrating fcc lattices. In a completely ordered structure, nearest neighbour environment of V involves 4 Al and 4 Nb atoms. Similarly, Nb has 4 V and 4 Ti atoms in its immediate environment, and so on. In a completely disordered structure, this nearest neighboring environment changes completely, which can make a huge impact on the atom projected local moments and hence the net moment. To get a clear picture of the effect of small change of atomic environment on the local magnetic arrangement in NbVTiAl, we have simulated a few simple point defects introduced in an ordered environment. A $2\times2\times2$ supercell of the primitive cell of the most stable ordered configuration (Type II) of NbVTiAl was constructed. This supercell contains a total of 32 atoms, including 8-atoms of each kind. Several swap-disordered (where one atom was exchanged by another atom) configurations have been simulated to check if such a disorder is responsible for the almost negligibly small measured moment in this alloy. In a $2\times2\times2$ supercell, exchanging one of the eight 'P' atom positions with one of the eight 'Q' atom positions leads to a 12.5\% swap disorder between 'P' and 'Q' atoms. All possible configurations for replacement of 'P' by 'Q' and vice versa were simulated, and the results of energetically most stable configurations are chosen to present here. When the position of one Nb atom was interchanged with 1 Al atom, the disorder configuration was labelled as "1Nb-1Al". Similarly, "2Nb-2Al" type disorder indicates that the positions of two Nb atoms were interchanged with two Al atoms. Various disordered structures have been considered and labeled in a similar manner as described above. Each of these structures was fully relaxed. Table \ref{MM} summarises the optimized lattice parameters, total energies and the net magnetic moments for various simulated disordered structures. It can be seen that all the disordered states, except those by involving swap disorder between Nb and Ti atoms, results in reduced moment as compared to that for a purely ordered structure. In fact, in the case of Nb-Ti swap disordered structures, the total magnetic moment increases as compared to the ordered structure. In some cases, the reduced moment was accompanied by a cubic to pseudo cubic distortion (lattice constants a, b, c differ by a small amount). Almost zero magnetic moment was observed in three cases, 1V-1Al, 1Nb-1V \& 1V-1Al, and 1Nb-1Ti \& 1V-1Al. As compared to the completely ordered structure, the atom projected local moments in the disordered case changes considerably (some even get quenched and/or antiferromagnetically aligned) due to the change in the local atomic environment. Thus, the reduced moment observed in NbVTiAl possibly originates from some specific swap disorder configurations between the constituent atoms, as discussed above.
To further understand the effect of disorder on the electronic structure of a few swapped disordered configurations having negligibly small total magnetic moments, we have simulated the spin-polarized density of 1V-1Al, 1Nb-1V \& 1V-1Al cases. These are shown in Fig. \ref{dos_dis}(b) and \ref{dos_dis}(c). One can notice that due to disorder, new states emerge near the Fermi level, which make the system metallic, and reducing the spin polarization. Overall, this study provides a clear understanding on how the degree of disorder and hence the altered local atomic environment can affect the net moment and electronic structure of quaternary Heusler alloys.
\section{Conclusion}
\label{conc}
In summary, we report an interesting quaternary alloy NbVTiAl which is observed to crystallize in A2-type (Im-3m) structure. Disorder mainly arises because of the similar electronegativity of the constituent elements. Magnetization measurements confirm a high Curie temperature ($>$ 800 K) and a negligibly small magnetic moment ($\sim 10^{-3} \mu_B/f.u.$) at 5 K. Temperature dependence of longitudinal resistivity reveals a metallic nature. Holes act as majority carries with a carrier concentration of $\sim$ $10^{22}$ cm$^{-3}$, as confirmed by the Hall measurements. Band structure calculation of purely ordered NbVTiAl confirms a nearly half-metallic behavior with a high spin polarization ($\sim$ 90\%). The simulated net magnetization of the purely ordered phase is 0.8 $\mu_B/f.u.$, which is in fair agreement with the SP rule, but deviates significantly from the experimentally measured magnetization. The detailed study confirms that disorder changes the magnetic moment remarkably, leading to a very small value as observed experimentally. Such a disorder is also responsible for the emergence of states at the Fermi level which impacts the half metallicity and hence the spin polarization of NbVTiAl. Thus, it has been observed that even if theoretical simulation reveals high spin polarization for the ordered structure, the magnetic and electronic properties obtained from experiment may differ completely. NbVTiAl is one of a kind in the quaternary alloys family where disorder plays a decisive role in controlling the magnetic/electronic properties of the alloy.
\section*{Acknowledgment}
DR acknowledges Prof. K. Hono for allowing to use the high-temperature magnetic measurements facility at NIMS, Japan. JK acknowledge the financial support provided by IIT Bombay. AA acknowledges DST- SERB (Grant No. CRG/2019/002050) for funding to support this research.
\section*{References}
|
\section{Introduction}
Noisy Intermediate-Scale Quantum (NISQ) computers, which have tens to hundreds of quantum bits (qubits) are anticipated to speed up the computational power of classical computers for certain classes of problems. However, the noisy nature of their qubits and operations degrades the reliability of NISQ systems.
Thus, errors should be mitigated to improve the fidelity of NISQ algorithms, which necessitates error characterization and circuit design approaches to minimize the impact of errors on the quantum circuit output. Randomized benchmarking is typically used to characterize different sources of noise in the quantum hardware including gate, measurement, and decoherence errors~\cite{PhysRevA.85.042311}. The computed error rates are used by noise-aware quantum compilers to generate physical quantum circuits with high fidelity.
While quantum noise is very complex, simplified error models based on error rates collected during the calibration process are typically used for quantum circuit compilation~\cite{{DBLP:journals/corr/abs-1903-10963}}.
Depending on the permissible access to the quantum computer, more advanced error characterization can be unavailable. A predictive analysis capable of delivering a projection with a reasonable accuracy about the circuit error rates, despite the lack of complete and accurate error model is also required. A predictive technique, which is computationally fast and applicable to many quantum computing technologies, and takes into account the circuit structure can enhance the compilation process~\cite{10.1145/3408039}.
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{1\textwidth}
\centering
\includegraphics[width=5.8in]{Figures/cnot_n.eps}
\caption{}
\label{fig:(a)}
\end{subfigure}
\begin{subfigure}[b]{1\textwidth}
\centering
\includegraphics[width=5.8in]{Figures/rz_n.eps}
\caption{}
\label{fig:(b)}
\end{subfigure}
\caption{Commutation rules of (a) CNOT, and (b) $R_z$ gates~\cite{Optimiz_Q}.}
\label{fig:three graphs}
\end{figure*}
In this paper, we propose error propagation-based predictive techniques to project the quantum circuit success rate primarily based on gate errors. We take into account the propagation path of different gate errors in the circuit. {\color{black} While the existing estimated success probability typically used by noise-aware quantum compilers to generate physical quantum circuit incorporates the gate error rates into its model, it ignores the order of the gates that don't affect the circuit depth or the gate count.} Our proposed predictive techniques can differentiate between different physical implementations of the same quantum circuit, which use the same physical qubits, share the same set of quantum gates, and have the same depth, but vary in the order of quantum gates in the circuit and their corresponding error rates. {\color{black} We experimentally show the advantage of considering the order of the physical quantum circuit gates in the presence of variable gate errors.} Accordingly, we also propose gate rescheduling algorithms at different quantum circuit abstractions to maximize the fidelity of the quantum circuit. {\color{black} Our proposed approaches can be integrated with other gate rescheduling algorithms that reduce the gate count and the circuit depth.} To the best of our knowledge, no such analysis has been proposed to
incorporate the impact of the error propagation paths to the quantum gate scheduling procedure.
The main contributions of our paper are as follows:
\begin{itemize}
\item We propose a Weighted Estimated Success Probability (WESP) metric based on the quantum circuit structure.
\item We propose new quantum circuit gate rescheduling algorithms at different design levels including complex and elementary gate-level based on our proposed WESP metric to boost the fidelity of the quantum circuit output.
\item We experimentally validate the effectiveness of our proposed metrics and gate rescheduling algorithms using a variety of quantum circuits executed on different IBM quantum computers.
\end{itemize}
The remainder of paper is organized as follows. Section II provides a background on the quantum circuit compilation, quantum hardware errors, an application of NISQ computers, and different success criteria for quantum circuit evaluation. Section III discusses related works on quantum circuit mapping and gate reordering approaches. Section IV shows the implication of the gate error propagation path on the output state fidelity of the quantum circuit. Section V provides our proposed reliability metric and rescheduling algorithms. Section VI validates the effectiveness of our approaches through several experiments. We conclude the paper in Section VII.
\iffalse
Future is all about quantum computing due to the increase in computation speed when compared to classical computers. Currently, several companies like IBM~\cite{8322045}, Google~\cite{7934217}, etc., have developed their own quantum devices, but all these devices have been built using noisy qubits, which is a major concern when it comes to the reliability of the device. These noisy devices are known as Noisy Intermediate-Scale Quantum computers. The researchers started showing interest in developing algorithms to reduce the impact of noise by mitigating the errors in these quantum devices. There are several factors that influence the noise in the quantum device, which is discussed in Section II.B in detail. Quantum circuits are executed on quantum devices to measure the amount of error in the device\todo{Not quite right}. The quantum circuits are executed on a software-based compiler to study the impact of the error based on the output success probability\todo{incorrect, you don't use a compiler to study the impact of error}. The compilation stack considers\todo{consider is not the right word, uses could be meaningful} the error rate of the quantum devices during the execution process. To date, we do not have any accurate failure model to analyze the failure of quantum devices, so we use a simple model to estimate how good is the quantum circuit based on the success rate.
In this paper, we propose to explore the structure of the quantum circuit when we try to model the error rates. For a given quantum circuit with the same set of gates and the same set of physical qubits, the success probability is affected while reordering the gates without altering the depth. We reorder the gates that satisfy the commutative property. To the best of our knowledge, existing works do not consider the structure of the quantum circuit to model the error rates to enhance the output probability of the circuit.The main contributions of our paper are the following:
\begin{itemize}
\item We propose a metric that takes into account the structure of the quantum circuit when we predict the output success probability.
\item Based on the metric, we propose another level of quantum circuit scheduling that tries to enhance the success probability of the circuit after post mapping.
\item We experimentally validate the impact of the proposed work based on the success probability of different quantum circuits using IBM Q16 Melbourne, IBM Q7 Casablanca, and IBM Q Santiago quantum computers..
\end{itemize}
\fi
\section{Background}
\subsection{Quantum circuit compilation}
Quantum algorithms are described using quantum circuits, which are executed on the quantum computer. A quantum circuit comprises of quantum gates, which change the state of the qubits. Single-qubit gates operate on a single qubit such as the Hadamard (H) gate, which creates a superposition state, and the $R_z$ gate, which rotates the qubit around the z-axis. A multi-qubit gate such as Controlled Not (CNOT) entangles two qubits~\cite{nielsen_chuang_2019}. To enable a quantum circuit execution on the targeted NISQ computer, the circuit has to go through several compilation steps. Complex gates should be constructed using elementary gates supported by the NISQ architecture. To reduce the depth and the gate count of the circuit, several gate-level optimization techniques are applied including template matching~\cite{1432873} and gate reordering~\cite{10.1145/3434318} based techniques. In the former one, a cascade of quantum gates is substituted with a sub-circuit with a lower gate count, while in the later one, consecutive gates cancel each other based on the commutation rules of the quantum gates~\cite{Optimiz_Q}, which are defined as follows:
\begin{definition}
Let $U_1$ and $U_2$ be two unitary matrices. $U_1$ and $U_2$ are said to be commutative if $U_1U_2 = U_2U_1$ for any input state. Commutation rules of different quantum gates are shown in Figure~\ref{fig:three graphs}.
\end{definition}
Next, the physical qubits of the quantum circuit are allocated and their gates are scheduled to meet the constraints of the quantum architecture. This process is referred to as quantum circuit mapping. A main challenge in the mapping process is the restricted qubits connectivity in superconducting quantum architectures. Two-qubit gates can be applied to certain pairs of physical qubits described using the coupling graph of the quantum architecture, in which nodes represent qubits and edges show the connectivity between different qubits. To enable arbitrary multi-qubit gates in the presence of the coupling constraint, SWAP operations consisting of three CNOT gates are used for gate scheduling, resulting in an excessive gate count and a large circuit depth. Efficient mapping approaches are used to generate physical quantum circuits with as a minimum gate count as possible to be executed on the quantum computer. These approaches often represent the quantum circuit {\color{black}as a directed acyclic graph, referred to as a} gate dependency graph, where each node represents a gate and each edge represents a direct dependency between two gates. {\color{black} The dependency graph is divided into levels, where each level consists of quantum gates applied simultaneously in the same circuit layer. All the nodes reachable from a given node $i$ in the graph are considered as the gates which are dependent on the $g_i$ gate ($g_i$ reachable gates).}
{\color{black}
\begin{example}
Figure~\ref{fig:gate_dep_general} shows an example of a quantum circuit and its corresponding gate dependency graph in which the gates in each circuit layer are mapped to the graph nodes applied at the same level. For each quantum gate, the list reachable gates is extracted from the gate dependency graph. For example, there are 3 reachable gates from $g_3$ gate, which are $g_5$, $g_6$, $g_7$, and $g_8$ gates.
\end{example}
}
\begin{figure}[t]
\vspace{-0.05in}
\centering
\includegraphics[width=3.3in]{Figures/gate_dep_wo.eps}
\caption{{\color{black}(a) A quantum circuit and (b) its corresponding gate dependency graph.}}
\label{fig:gate_dep_general}
\end{figure}
\subsection{Reliability of quantum devices}
Different sources of noise contribute to the noise complexity of NISQ systems. The elevated noise levels endanger the reliability of quantum circuits. Gate errors are a major source of quantum computing noise. They are modeled as Pauli gates that operate on a single qubit. They are represented as $X$ (a $\pi$ rotation around the x-axis), $Y$ (a $\pi$ rotation around the y-axis), $Z$ (a $\pi$ rotation around the z-axis), and $I$ (identity matrix) single-qubit gates~\cite{janardan2016analytical}. The propagation of some of the Pauli errors through the CNOT gate is shown in Figure~\ref{fig:pauli}. The measurement/read-out error can occur while measuring the output state of the quantum circuit. A qubit can lose its state after a period of time, resulting in decoherence errors. Quantum gates that act on different qubits simultaneously can cause crosstalk errors.
\begin{figure}[b]
\centering
\vspace{-0.1in}
\begin{subfigure}{.35\columnwidth}
\centering
\includegraphics[width=0.8in]{Figures/p1.eps}
\caption{}
\label{fig:(a)}
\end{subfigure}
\begin{subfigure}{.35\columnwidth}
\centering
\includegraphics[width=0.8in]{Figures/p2.eps}
\caption{}
\label{fig:(b)}
\end{subfigure}
\vspace{-0.05in}
\caption{The propagation path of (a) X and (b) Z Pauli gate errors through a CNOT gate.}
\label{fig:pauli}
\end{figure}
As NISQ computers do not support quantum error correction, they often rely on noise-aware quantum compilers to mitigate errors. The quantum compiler should maximize the success probability of the physical quantum circuit. An Estimated Success Probability (ESP) can be used by a noise-aware quantum compiler to
select the physical qubits of the quantum circuit and schedule the quantum gates. Given the gate ($e_{g_i}$) and measurement ($e_{m_i}$) errors for all the quantum circuit gates, and its qubits, ESP is computed as $\textstyle \prod^{G-1}_{i=0}(1 - e_{g_i}) \times \prod^{Q-1}_{i=0}(1- e_{m_i})$~\cite{DBLP:journals/corr/abs-1903-10963}. {\color{black} Randomized Benchmarking sequences are executed on the quantum hardware to compute single- and two-qubit gate errors based on the average sequence errors. The qubit measurement/readout error is computed as the average measurement error 0 and 1, in which the qubit is initially assigned to state 1 and 0, respectively.
}
\subsection{Quantum approximation optimization algorithm}
Quantum approximation optimization algorithm (QAOA) is a hybrid quantum-classical algorithm for solving optimization problems. It relies on not only the output of the quantum circuit but also the classical optimizer that updates the circuit parameters to improve the solution of the optimization problem~\cite{farhi2014quantum}. The objective of the algorithm is to minimize/maximize a cost function (C(X)) of a given problem described using the following Hamiltonian: $H= \sum_{x\in{0,1}^n}{C(x)\ket{x} \bra{x}}$. A QAOA quantum circuit contains two main components, namely, phase separation and mixing operations, which are applied $p$ times repeatedly with ($\gamma_1$,$\beta_1$), ..., ($\gamma_p$,$\beta_p$) parameters.
The phase separation can be represented using ZZ gates, while mixing operations are represented using single-qubit gate rotation around the x-axis (R$_x$). An example of a problem to be solved using the QAOA algorithm is the Maximum Cut (Max-Cut) problem. For a graph with n vertices, the solution of the Max-Cut problem divides the graph into two subsets such that the total weight of the edges between those subsets is as maximum as possible. The graph vertices are described as a string of n qubits in which each qubit can be measured as 0 or 1. The QAOA circuit output provides a candidate solution of the Max-Cut problem. Using an iterative procedure, we can find the n-qubit string that maximizes the total weight of the edges between the two subsets, referred to as the cost function.
\iffalse
\hl{Quantum approximation optimization algorithm (QAOA) is a hybrid (classical part) algorithm that helps in solving optimization problems of NISQ devices~\cite{farhi2014quantum}. The objective of the algorithm is to minimize the error for a given problem. The Hamiltonian is defined as $H= \sum_{x\in{0,1}^n}{C(x)\ket{x} \bra{x}}$, where the objective function (C(x)) acts on the computational basis states $\ket{x} $. For a given n bit string and m clauses, the cost function $C(x)$ has eigenvectors $\ket{x}$ and eigenvalues for each $\ket{x}$. The parameters $\alpha$ and $\beta$ helps in finding the maximize the eigenvalue to $C_{MAX}$ for a given $C(x)$. One of the examples of QAOA is to solve the Max-Cut problem. For a graph ($G$) with n-nodes, the goal of Max-Cut is to achieve a maximum number of edges that are cut by a given partition of vertices into two sub-graphs. The nodes of $G$ can be described as a string of n bits in which each bit of the string can hold a value of 0 or 1. In QAOA for the Max-Cut problem, the cost function represented by Hamiltonian is used to evaluate the maximum number of cut edges for $G$. QAOA aims to find the maximum eigenvalue for the Hamiltonian function, which provides an approximate solution for the given function. }
\fi
\subsection{Quantum circuit success metrics}
Different success criteria/metrics have been proposed to show the impact of the quantum compilation approaches on the quantum circuit output. For a single-output quantum circuit, the Probability of Successful Trails (PST) measures the probability of the correct output of the quantum circuit as $\frac{\text{Number\:of\:Successful\:Trials}}{\text{Total\:Number\:of\:Trials}}$~\cite{Tannu:2019:QCE:3297858.3304007}.
For hybrid quantum algorithms, application-specific success criteria are used. For example for QAOA, which maximizes or minimizes a cost function, the Approximation Ratio (AR) of the cost function quantifies the success of the quantum circuit~\cite{9251960, PhysRevX.10.021067}. It is defined as $\frac{\text{Mean\:of\:Cost\:Function\:Over\:all\:Sampled\:Output}}{\text{Maximum\:Cost\:Function\:Value}}$.
Given AR, we can compute the Approximation Ratio Gap (ARG) to quantify how close is the output of the actual execution of QAOA circuit to the simulation result~\cite{9251960}. It is computed as $\frac{\text{AR\:of\:Simulation\:-\:AR\:of\:Execution}}{\text{AR\:of\:Simulation}}\times100$. The smaller the gap, the more reliable the QAOA circuit.
\section{Related work}
\label{related_work}
\subsection{Quantum circuit mapping}
Since quantum circuit mapping is an NP-Complete problem~\cite{siraichi2018qubit}, traversing the search space is speculative for quantum architectures with a large number of qubits. Various heuristics-based techniques have been proposed to address this problem~\cite{9384317}. Some of these techniques target minimizing the number of circuit SWAP operations~\cite{itoko2019optimization,8702439,wille2019mapping,zulehner2018efficient,10.1145/3297858.3304023}, while others focus on minimizing the circuit error rates~\cite{Tannu:2019:QCE:3297858.3304007,Ash-Saki:2019:QQR:3316781.3317888}.
For example, quantum circuit mapping can be achieved by using BRIDGE and SWAP operations as transformation rules inserted based on a dynamic programming or a heuristic-based look-ahead schemes~\cite{itoko2019optimization}. The BRIDGE gates are composed of a sequence of four CNOT gates.
SAT solver minimizes the number of SWAP operations in the physical quantum circuit at the cost of extensive computation overhead (e.g.~\cite{wille2019mapping,8702439}). The proposed technique in~\cite{8702439} divides the quantum circuit into sub-circuits that require no SWAP operations within each sub-circuit.
Next, the SAT solver is used to construct the quantum circuit by inserting a minimum number of SWAP operations between the sub-circuits that share qubits.
In~\cite{wille2019mapping}, a logical quantum circuit is mapped to a physical quantum circuit using a SAT solver too.
Constraints are added to the SAT solver based on the coupling graph of the quantum architecture to reduce its complexity at the expense of the gate count.
An A* algorithm explores the search space in the coupling graph to find the shortest path between qubits for each two-qubit gate, and thus, reduce the number of SWAP operations in the quantum circuit~\cite{zulehner2018efficient}.
To address the difference in the error rates of various qubits, a mapping technique has been proposed based on Dijkstra algorithm, which schedules quantum gates and allocates physical qubits with low error rates~\cite{Tannu:2019:QCE:3297858.3304007}.
The success probability of the physical quantum circuit can be further improved by searching for isomorphic sub-graphs in the quantum architecture's coupling constraint with the highest ESP~\cite{Ash-Saki:2019:QQR:3316781.3317888}. Other mapping approaches address decoherence, correlated, and unexpected errors (e.g~\cite{8824907,Tannu:2019:EDM:3352460.3358257,10.1145/3400302.3415684}).
To mitigate crosstalk errors, they should be characterized first, followed by an efficient mapping approach that selectively serializes circuit gates, while satisfying the required coherence time~\cite{10.1145/3373376.3378477}.
\subsection{Gate reordering of quantum circuits}
A quantum circuit consists of layers of gates that can be applied in parallel. Consecutive commuting gates can be reordered while preserving the state of the quantum system~\cite{Optimiz_Q, Guerreschi_2018}. The commutation rules of the gates have been utilized to reduce the quantum circuit depth, gate count~\cite{9218558,10.1145/3400302.3415620,Optimiz_Q, Guerreschi_2018}, and optimize the control pulses~\cite{Shi_2019}. A heuristics-based approach has been developed to perform a gate-level optimization by reordering quantum gates to reduce the circuit gate count~\cite{Optimiz_Q}.
A two-step approach has been proposed to reduce the depth of the quantum circuit by constructing a dependency graph based on the circuit gates and then swapping the gates that satisfy the commutation rules to reduce the circuit depth~\cite{Guerreschi_2018}.
A compilation approach has been developed to enhance the performance of QAOA circuits by reordering the complex gates~\cite{9218558,9251960}. Due to the commutation relation, most of the dependencies are ignored.
Instead, a two-level search process is adopted to minimize the circuit depth using a breadth-first search algorithm, which finds a possible two-layer interchanges in each iteration until the desired depth is achieved.
A layout synthesis algorithm has been proposed to minimize the circuit depth by reordering
complex gates of QAOA circuits using a Satisfiability Modulo Theories (SMT) solver~\cite{10.1145/3400302.3415620}.
An optimized pulse level compilation approach has been proposed to address the inefficiency of the standard gate compilation process that directly translates the logical instructions to control pulses~\cite{Shi_2019}. In this approach, the gates are reordered based on commutation rules, and then a small set of gates are aggregated to a larger operation by finding the optimal control pulse based on the gradient descent method.
Gate reordering has also been applied to minimize the impact of decoherence errors~\cite{9296804}. Faults in the form of additional gates due to decoherence errors are injected into the quantum circuit to analyze their impact on the circuit output. Accordingly, gates are reordered at the circuit locations, in which the decoherence errors significantly affect the circuit output.
{\color{black}
In this work, we propose additional compilation layers applied to the physical quantum circuit post-mapping. We restrict the search space of our gate rescheduling algorithms to maintain the gate count and the circuit depth of the optimized quantum circuit using the previously proposed compilation approaches. Thus, our work can be incorporated with the other compilation/optimization layers to further improve the output state fidelity of the quantum circuit.
}
\begin{figure}[t]
\centering
\includegraphics[width=3.4in]{Figures/fig__3.eps}
\caption{(a) A quantum sub-circuit and its gate dependency graph, and (b) its corresponding equivalent sub-circuit and its gate dependency graph.}
\label{fig:pro_example}
\end{figure}
\iffalse
\begin{figure*}
\centering
\begin{subfigure}{0.05\textwidth}
\centering
\includegraphics[width=0.6in]{Figures/rome_n.eps}
\caption{}
\label{fig:(a)}
\end{subfigure}\hspace{1em}
\begin{subfigure}{0.35\textwidth}
\centering
\includegraphics[width=2.0in]{Figures/m1_n.eps}
\caption{}
\label{fig:(b)}
\end{subfigure}
\begin{subfigure}{0.15\textwidth}
\centering
\includegraphics[width=2.0in]{Figures/fig4_nn.eps}
\caption{}
\label{fig:(c)}
\end{subfigure}\hfill
\begin{subfigure}{0.415\textwidth}
\centering
\includegraphics[width=1.11in]{Figures/m1m2_n.eps}
\caption{}
\label{fig:(d)}
\end{subfigure}
\caption{(a) IBM Q Santiago architecture, 4-qubit BV with different mappings (b) and (c), and (d) PST for M1 and M2.}
\label{fig:mapping}
\end{figure*}
\fi
\iffalse
\begin{figure*}
\hspace{-11.1cm}
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=0.6in]{Figures/rome_n.eps}
\caption{}
\label{}
\end{subfigure}
\hspace{-2cm}
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=5.7in]{Figures/n_all.eps}
\label{}
\end{subfigure}
\caption{(a) IBM Q Santiago architecture, 4-qubit BV with different mappings (b) and (c), and (d) PST for M1 and M2.}
\label{fig:mapping}
\end{figure*}
Figure~\ref{fig:mapping1}
\fi
\begin{figure*}
\centering
\includegraphics[width=6.9in]{Figures/update_fig4.eps}
\caption{(a) IBM Q Santiago architecture, 4-qubit BV implemented using (a) M1 and (b) M2 mappings, and (d) their corresponding PST value.}
\label{fig:mapping}
\end{figure*}
\section{Implication of the gate error propagation path on the output state fidelity}
Stochastic Pauli noise is a primary factor to estimate the circuit error rates~\cite{{DBLP:journals/corr/abs-1903-10963}}. While tracking the precise impact of the error propagation on the circuit can be computationally expensive
in the presence of universal
gates~\cite{janardan2016analytical},
two-qubit gates enable error propagation across different qubits. Thus, the location of the quantum gates especially the two-qubit gates can affect the circuit error rates. {\color{black} For the sake of simplicity}, we assume that the two-quit gate (CNOT) error impacts the control and the target qubits, and thus, propagates to the gates applied next to the control and the target qubits.
\begin{example}
To illustrate the gate error propagation through the quantum circuit, {\color{black}let's consider} the example provided in Figure~\ref{fig:pro_example}. Two equivalent sub-circuits using the commutation rules of quantum gates and their corresponding gate dependency graphs are provided. {\color{black} We use a dotted line to show the output qubits affected by the gates in the dependency graph.}
Every gate in the sub-circuit can produce Pauli errors that propagate to the qubit output through the gate dependency graph~\cite{janardan2016analytical}. In Figure~\ref{fig:pro_example}(a),
$q_0$, $q_1$, and $q_2$ output qubits are affected by $\{g_0\}$, $\{g_0, g_1, g_2, g_3\}$, and $\{g_0, g_1, g_2\}$ gate errors, respectively.
On the other hand, in Figure~\ref{fig:pro_example}(b), $q_0$, $q_1$, and $q_2$ output qubits are affected by $\{g_0, g_1, g_2, g_3\}$, $\{g_0, g_1, g_2, g_3\}$, and $\{g_1, g_2\}$ gate errors, respectively. Thus, while the two quantum sub-circuits in Figure~\ref{fig:pro_example} are equivalent, their output qubits $q_0$, and $q_2$ are susceptible to different sets of gate errors.
\end{example}
Due to the spatial variation in error rates of the quantum hardware, gates applied to different qubits result in different error rates. Although all gate errors affect the fidelity of the quantum circuit, the two-qubit gate errors has lower fidelity than single-qubit gate errors.
Thus, reordering quantum gates while maintaining the circuit functionality changes not only the error propagation paths but also the gate error impact on the output state of the circuit. In the previous example, if $g_0$ gate has the highest error rate in the circuit, which significantly deviates from other gate errors in the circuit, applying $g_0$ gate at the first layer of the sub-circuit as in Figure~\ref{fig:pro_example}(a) will spread $g_0$ error to all the output qubits unlike the equivalent sub-circuit in Figure~\ref{fig:pro_example}(b) in which $g_0$ error affects only $q_0$ and $q_1$ output qubits.
\begin{example}
Figure~\ref{fig:mapping} shows the impact of gate reordering on the quantum circuit output fidelity. A 4-qubit Bernstein–Vazirani (BV) circuit is mapped to satisfy the coupling constraint of IBM Q Santiago architecture. Figure~\ref{fig:mapping}(a) represents the coupling graph and the error rates of IBM Q Santiago architecture, in which the values in the upper and lower part of each node represent the read-out and single-qubit errors, respectively, while the edge label is the two-qubit gate error applied to the corresponding pair of qubits. All error rates are multiplied by $10^{-3}$. Two physical implementations of the BV quantum circuit, namely M1 and M2, are provided in Figure~\ref{fig:mapping}(b) and~\ref{fig:mapping}(c), respectively. The difference between the two physical implementations of the circuit is the location of the first two CNOT gates of the quantum circuit, which can be reordered based on the commutation rules. The CNOT gate highlighted in red color has the highest two-qubit gate error rate. According to the order of the quantum gates, the error propagation path also changes as shown in Figure~\ref{fig:mapping}(b) and~\ref{fig:mapping}(c). We demonstrate the impact of the gate reordering on the quantum circuit output by executing both M1 and M2 quantum circuits on IBM Q Santiago architecture, and reporting their PST as shown in Figure~\ref{fig:mapping}(d). To eliminate the impact of gate reordering on decoherence errors, we insert barriers before and after applying all the two-qubit gates in the two circuit implementations in addition to a barrier after the second CNOT gate of M2 quantum circuit in Figure~\ref{fig:mapping}(c), which ensure that the qubits in the two implementations are used for the same period of time. The PST values are $0.784$ and $0.861$ for the BV quantum circuit generated using the M1 and M2 mapping, respectively, and executed for 8192 shots/trials. Our observation indicates that reordering gates with higher error rates to earlier circuit layers will result in error propagation to larger number of circuit gates, and thus, degrade the output state fidelity. While PST is different for both mapping, the ESP of both M1 and M2 mapping is $0.7757$, which implies the need to consider the gate error propagation path for estimating the success rate of the circuit, and thus, refining the quantum circuit mapping policies.
\iffalse
The gates of the quantum circuit after mapping to the quantum device can be reordered without changing the depth. An example of gate reordering for a 4-qubit Bernstein–Vazirani (BV) circuit executed on IBM Q5 Rome architecture is shown in Figure~\ref{fig:mapping}. Figure~\ref{fig:mapping}(c) represents the error rates of IBM Q5 Rome architecture, in which the values inside the circle represent the read-out error and single-qubit error, and the values between the qubits are CNOT gate errors. Though all types of errors affect the fidelity of the circuit, the CNOT gate error is more crucial than other errors. So for the given quantum architecture, the initial mapping (M1) of the 4-qubit BV circuit is obtained as shown in Figure~\ref{fig:mapping}(a) in which the CNOT gate with a high error rate is highlighted in red according to Figure~\ref{fig:mapping}(c). Then we move the CNOT gate with a high error rate to the right based on the commutative property for the mapping M2 as shown in Figure~\ref{fig:mapping}(b). Later we execute both M1 and M2 quantum circuits on IBM Q5 Rome architecture, and the PST is computed as shown in Figure~\ref{fig:mapping}(b). The PST is $54\%$ and $71\%$ for the mapping M1 and M2, respectively. From the output probability, it is clear that rescheduling the gate based on the commutative property helps to improve the fidelity of the quantum circuit. This motivates us to consider the structure of the quantum circuit to model error rates.
The fidelity of the quantum circuit is affected based on the location of the gates on the noisy qubits, which will propagate the error through the circuit. To illustrate the impact of propagation of the error, let us consider the same example M1 from Figure~\ref{fig:mapping}(b), and the error propagation path for the respective circuit is highlighted in red. According to the error rates shown in Figure~\ref{fig:mapping}(a), the single-qubit and two-qubit gate errors from high to low for the qubits are $q2\rightarrow q1\rightarrow q4\rightarrow q3\rightarrow q0$ and $q1\leftrightarrow q2\rightarrow q2\leftrightarrow q3\rightarrow q0\leftrightarrow q1\rightarrow q3\leftrightarrow q4$, respectively. The error path highlighted in red will cause a huge impact on the circuit since the qubit $q2$ and $q1$ has the CNOT gate with highest error value compared to other qubits, and the CNOT gates in those qubits are executed first, which will later affect the gates in the qubits $q1, q2, q3$ and $q4$. It is clear from the discussion that if the qubits with high error rates are executed first, it will immensely affect the fidelity of the quantum circuit.
\fi
\end{example}
\iffalse
\todo{
First you need to emphasize that with a given quantum circuit mapping and the same depth there is a room to reoder gates while maintaining the circuit depth. You take a simple BV example with 4 qubits and you show that you can reorder the CNOT gates in different ways and show the corresponding PST for each reordering. Make sure you add a barrier after the single qubit gate at layer one and before the single qubit gates at the last layer to insure that the circuit depth is the same. This motivates the need of taking the circuit structure into account.
Then, you need to explain that the location of the gate in the circuit impacts the error propagation path of that gate errors. THus, applying noisy qubit gates at the beginning of the circuit will increase the impact of the errors in the circuit. Here you also take the same BV example and show for a very noisy CNOT gate applies at the beginning its error propagation path as arrows through the circuit (in red color). You can give also red color to the CNOT gate with the highest error (make sure that there is a significant difference in the different CNOT error gates). Use the same BV example as before. You conclude this section with the need to take location of very noisy gates in the circuit as they may spread errors aggressively thought the circuit.
Next section you can start talking about your reliability model WESP using the same example as before to demonstrate how to compute it.
}
\fi
\section{Error propagation-based gate rescheduling}
\subsection{Weighted estimated success probability}
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[width=3.4 in]{Figures/w_m1.eps}
\caption{}
\label{fig:m1wesp}
\end{subfigure} \vspace{1mm}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=3.4 in]{Figures/fig5b_n.eps}
\caption{}
\label{fig:m2wesp}
\end{subfigure
\caption{{(a) The number of reachable gates ($S_{i}$}) for each quantum gate in M1, and (b) M2 mappings.}
\label{fig:wesp}
\end{figure}
ESP enables quick and easy estimation of the circuit success probability. Yet, it overlooks the circuit structure since different orders of the quantum gates yield the same ESP.
To account for the penalty of rescheduling the same set of physical quantum gates with variable error rates on the output state fidelity of the quantum circuit, we modify the ESP metric and propose a new metric referred to as Weighted Estimated Success Probability (WESP) defined as
\begin{equation*}
\textstyle
WESP =\prod^{G-1}_{i=0}(1 - (\lambda_{i} + e_{g_i})) \times \prod^{Q-1}_{i=0}(1- e_{m_i})\:\:\end{equation*}
$\lambda_{i}$ tunes the error rate of the corresponding $g_i$ gate to account for the gate position as
\begin{equation*}
\lambda_{i} = w_i(e_{g_i}- {\color{black} \min(E_{G})}),
\end{equation*} where $w_i$ and {\color{black} $\min(E_{G})$} are the weight of the $i^{th}$ gate and the minimum gate error rate in set of gate errors {\color{black}($E_{G}$)} of the quantum circuit, respectively. Instead of multiplying the weight of each gate by its error rate, which can significantly reduce the gate error of very noisy gates applied in later layers of the circuit, we multiply the weight with the difference between the current gate error and the minimum gate error in the circuit {\color{black} to account for the variation in the gate errors}. The resulting number is added to the gate error. To compute $w_i$, the gate dependency graph of the circuit is constructed first. For each gate ($g_i$) (node) in the circuit (graph), we identify the number of all subsequent dependent gates (reachable gates) that are applied to the same qubit or other qubits that are directly or indirectly connected to the current gate through two-qubit gates ($S_i$). For a quantum circuit with $G$ gates, the weight of the $i^{th}$ gate ($w_i$) is computed as $w_{i}=\frac{S_i}{G}$.
We exclude the $R_z$ gate since the error rate is zero. Thus, the term $\lambda$ is used to approximate the impact of noisy quantum gates on the output state fidelity of the quantum circuit given the number of subsequent dependent (reachable) gates. To simplify our analysis, we omit how Pauli errors propagate through different quantum gates in the circuit. Instead, we assume that for each gate, Pauli errors will propagate to its corresponding qubits.
\begin{example}
Figure~\ref{fig:wesp}(a) and (b) show the physical quantum circuits implemented using elementary gates supported by IBM quantum computers for M1 and M2 quantum circuits in Figure~\ref{fig:mapping}, respectively. The $S_i$ values for each single- and two-qubit gates are shown in red color. For example, in the physical circuit provided in Figure~\ref{fig:wesp}(a), the $\sqrt X$ gate applied to qubit $q_4$ in layer 2 is connected to the next $\sqrt X$ gate applied to the same qubit at a later layer in the gate dependency graph. Hence, the value of $S_i$ for the first $\sqrt X$ gate is 1. Similarly, the value of $S_i$ for all the other gates except $R_z$ gates are computed. The number of erroneous gates ($G$) is 16. Based on the device error rates as shown in Figure~\ref{fig:mapping}(a), the value of WESP for M1 mapping is $0.7748$. On the other hand, WESP for M2 mapping is $0.7764$, which indicates that M2 circuit provides a higher PST. This observation is aligned with the PST value of M1 and M2 circuits provided in Figure~\ref{fig:mapping}(d).
\iffalse
Lets consider the quantum circuits with M1 mapping as shown in Figure~\ref{fig:m1wesp}. The ${w_{g_i}}$ values for each single and two-qubit gates are highlighted in red, and the value of $N$ is 16. For example, the output of $\sqrt X$ gate in qubit $q_0$ is feed as input to the next $\sqrt X$ gate in the same qubit, hence the value of ${w_{g_i}}$ for the first gate is 1. Similarly, the value of ${w_{g_i}}$ for all the gates except $R_z$ gate is computed. Based on the device error is shown in Figure~\ref{fig:mapping}(c), the value of WESP for $M_1$ mapping is 0.817531.
\fi
\end{example}
\subsection{Elementary gate rescheduling based on WESP}
We exploit the gate commutation rules to reduce the impact of Pauli errors while maintaining the depth of the circuit. We propose a gate
rescheduling algorithm post-quantum circuit mapping guided by the proposed WESP metric as a new optimization layer to reduce the circuit gate errors. Since the gates of the physical quantum circuit satisfy the coupling constraint of the quantum hardware, no additional SWAP operation is required. We develop a greedy approach for gate rescheduling based on the circuit dependency graph ($D(V,E)$) as described in Algorithm~\ref{alg:alg2}. The objective of our rescheduling algorithm is to minimize WESP while maintaining the circuit depth ($Depth_C$) to avoid any additional decoherence errors. An important condition is added to ensure that the updated gate scheduling does not increase the depth of the circuit ($No\_Inc\_Depth$). For $g_i$ gate operating on $q_x$ and $q_y$ qubits at layer $t$ and $g_j$ gate operating on $q_x$ and $q_z$ qubits at layer $t+k$, $g_i$ and $g_j$ can be reordered if they are commuting and $q_y$ and $q_z$ are {\color{black}idle} at layers $t+k$ and $t$, respectively. Our evaluation is repeated for every pair of {\color{black} immediate} dependent gates in the gate dependency graph. For a quantum circuit with $D(V,E)$ gate dependency graph, in which $V$ (nodes) represents the {\color{black}set} of quantum circuit gates and $|E|$ (edges) shows the dependency between different gates, the time complexity of calculating WESP is {\color{black}$O(|V|+|E|)$}. Thus, the time complexity of our greedy heuristic, which is provided in Algorithm~\ref{alg:alg2} is {\color{black}$O(|V|^{2} \cdot (|V|+|E|))$}. Our proposed gate rescheduling approach is computationally fast compared to the exhaustive approach that requires checking the entire search space, which can be very large for quantum circuits with large number of commuting gates.
\begin{example}
To motivate the need and the effectiveness of our greedy approach, we exhaustively generate all possible ways of gate scheduling for our BV circuit provided in Figure~\ref{fig:mapping} and~\ref{fig:wesp}. We pick the circuit with the highest WESP and compare it with the generated circuit using our greedy approach in terms of WESP and PST. Specifically, we label each CNOT gate in Figure~\ref{fig:wesp} to list all possible combinations of gate scheduling. The total number of all possible ways of gate scheduling, which maintain the depth of the circuit is 14. Among all the gate scheduling combinations, B-C-D-E-F-A has the highest WESP of 0.7768, while the resulting circuit based on our greedy algorithm has a CNOT combination of B-A-C-D-E-F with a WESP of 0.7764. The PST value of the quantum circuit based on the exhaustive and our greedy approach are 0.863 and 0.861, respectively. Accordingly, the gap between the exhaustive and the greedy solutions is very small.
\end{example}
\begin{algorithm
\SetAlgoLined
\begin{small}
{\color{black}
\caption{Gate rescheduling}
\label{alg:alg2}
\KwIn{$C$ = The physical quantum circuit}
\KwOut{$C^{'}$ = The updated physical quantum circuit}
Construct the $D$ graph of $C$; \\Initialize $C^{'}$= $C$; \\
Initialize $L(i)$ = {gates applied at layer $i$ $\mid$ $0 < i <$ $Depth_C$; }\\
\For{each each layer $i$}{
\For{each gate $g_{j}$ in $L(i)$ }{
\For{ each immediate dependent gate $g_{k}$}{
\If{Are\_commutative($D$, $g_{j}$, $g_{k}$)}{
\If{No\_Inc\_Depth($D$, $g_{j}$, $g_{k}$)}{
Compute WESP of $C^{'}$ under rescheduled gates;
}
}
}
Select rescheduling of $g_{j}$ with max. WESP;\\
Update $C^{'}$;
}
}
Return $C^{'}$;
}
\end{small}
\end{algorithm}
\subsection{Complex gate rescheduling based on WESP for QAOA}
QAOA quantum circuits consist of single-qubit rotations and two-qubit phase gates implemented using ZZ complex gate. Since ZZ gates are commutative~\cite{crooks2018performance,Venturelli_2018}, we exploit the freedom of the ZZ gate placement at higher quantum circuit design level to further reduce the quantum circuit errors. ZZ gates are reordered without violating the circuit depth and the gate count.
\begin{example}
An example of ZZ gate reordering is provided in {Figure}~\ref{fig:ZZ}, in which both quantum sub-circuits share the same gate count and circuit depth but vary in the ZZ gate scheduling.
\end{example}
\begin{figure}[t]
\centering
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.1in]{Figures/ZZ1.eps}
\caption{}
\label{fig:(a)}
\end{subfigure}
\begin{subfigure}{.4\columnwidth}
\centering
\includegraphics[width=1.1in]{Figures/ZZ2.eps}
\caption{}
\label{fig:(b)}
\end{subfigure}
\vspace{-0.05in}
\caption{Two quantum sub-circuits with different ZZ gate schedulings.}
\vspace{-0.05in}
\label{fig:ZZ}
\end{figure}
We propose a look-ahead approach that approximates the error rate of each complex gate in the circuit. The approximation entitles the generation of physical quantum circuit first, using an efficient quantum mapping approach that allocates physical qubits with minimum error rates and minimizes the number of SWAP operations. We scan the physical quantum circuit next to identify the corresponding two-qubit gates used for building each ZZ complex gate. Given the decomposed sub-circuit of each complex gate, we compute the error rate of each complex gate.
We adjust the WESP metric to operate on the intermediate representation of the quantum circuit prior to complex gate decomposition. In the updated WESP, $G$ is the number of ZZ gates, $e_{g_i}$ is the ZZ gate error rate computed as the product of the error rates of its elementary gates, $w_i$ is the ratio of dependent (reachable) ZZ gates, and {\color{black} $\min(E_{G})$ }is the minimum error rate of the ZZ complex gate in the circuit. The ZZ gates are reordered based on the updated WESP using a similar rescheduling algorithm as the one provided in Algorithm~\ref{alg:alg2}, in which $C$ is updated to be the intermediate representation of the quantum circuit and every pair of quantum gates ($g_i$, $g_j$) corresponds to a pair of ZZ gates in the quantum circuit.
The look-ahead approach requires scanning the entire physical circuit including all its gates. The time complexity of the look-ahead approach is linear in the number of gates ({\color{black}$O(|V|)$}). The time complexity of our ZZ gate rescheduling algorithm is {\color{black}$O({|V^{'}}|^{2} \cdot (|V^{'}|+|E^{'}|))$} for $D^{'}(V^{'},E^{'})$ complex gate dependency graph, in which $V^{'}$ represents the {\color{black} set} of ZZ complex gates and ${E^{'}}$ is the {\color{black} set} of edges that show the dependency between different ZZ gates.
{\bf We emphasize that our rescheduling approaches, which are applied post-mapping can be integrated with other quantum compilation approaches that exploit the gate commutation rules to reduce the circuit depth and the gate count since our algorithms do not alter the circuit depth nor the gate count. Furthermore, our proposed approaches can be easily applied to quantum circuits executed on different quantum hardware in the presence of significantly variable hardware gate errors. }
\iffalse
The look-ahead approach requires scanning the entire physical circuit including all its gates. The time complexity is linear linear that is the same as scanning the whole dependency graph; the complexity is $O(V+E)$ and the complexity of our greedy heuristic appro
ach is $O(V^{2} \cdot (V+E))$. The complexity of the Algorithm is dominant to the complexity of the look-ahead approach, thus neglecting the complexity of the look-ahead approach, the total time complexity of our approach for complex gate-based rescheduling is $O(V^{2} \cdot (V+E))$.
\fi
\section{Experimental Evaluation}
\subsection{Experimental setup}
We show the effectiveness of our proposed rescheduling algorithms based on WESP in improving the output state fidelity of different quantum circuits executed on NISQ computers. We conduct two experiments. In the first experiment, we use PST to evaluate our proposed elementary gate rescheduling approach applied to different quantum circuits, while in the second experiment, we show the impact of our proposed complex ZZ gate rescheduling on the AR of QAOA quantum circuits.
We use a total of 12 benchmark circuits with different number of qubits and gate count. The \textbf{Bernstein-Vazirani (BV)} algorithm identifies a hidden string, which is encoded in the circuit~\cite{10.1137/S0097539796300921}. The discrete fourier transform is represented using \textbf{Quantum Fourier Transform (QFT)} circuit~\cite{nielsen_chuang_2019}. A shift operation is carried out for a given input Boolean function using \textbf{Hidden Shift (HS)} algorithm~\cite{10.5555/644108.644189}. The \textbf{Grover Search (Grover)} algorithm provides a solution ($x$) of a function $f(x)$ equal to 1. Given a unitary operator, the eigenvalue of the eigenvector is estimated using \textbf{Quantum Phase Estimation (QPE)} algorithm~\cite{nielsen_chuang_2019}. \textbf{Quantum Approximate Optimization Algorithm (QAOA)} solves optimization problems~\cite{farhi2014quantum}. Other reversible quantum circuits such as \textbf{Adder, Toffoli gate, and Decoder} are also used. We obtained QFT and QPE circuits from Qiskit~\cite{Qiskit}, HS and QAOA circuits from Cirq~\cite{quantum_ai_team_and_collaborators_2020_4062499}, adder from~\cite{2004quant.ph.10184C}, decoder from RevLib~\cite{4539430}, and BV, Toffoli, and Grover Search by manual construction. The properties of all the benchmark circuits are provided in Table~\ref{tab:prop} prior to the circuit mapping, which includes the number of qubits (\# Qubits), single-qubit gates (U), CNOT gates (CNOT), and measurement operations (M), the circuit depth (Depth), and the expected output. We use QAOA to solve the Max-Cut problem of three different graphs, which consist of 15 nodes and different number of edges of random weight per node. The QAOA quantum circuits are \textbf{QAOA\_1}, \textbf{QAOA\_2}, and \textbf{QAOA\_3}, which target graphs of 15 nodes with 5, 6, and 7 edges per node, respectively. To construct and test the QAOA circuits, we set $p$ to 1 with default $\gamma$ and $\beta$ parameters provided by Cirq~\cite{quantum_ai_team_and_collaborators_2020_4062499}. Unlike the other quantum circuits listed above, we consider the entire output distribution of QAOA quantum circuits to evaluate their cost function.
\begin{table}[t]
\centering
\caption{Properties of quantum circuit benchmarks.}
\label{tab:prop}
{%
\begin{tabular}{|c|c|c|c|c|c|c|}\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Benc-\\hmark\end{tabular}}} & \multicolumn{1}{|c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\#\\Qubits\end{tabular}}} & \multicolumn{3}{c|}{\#} & \multirow{2}{*}{Depth}&\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}Expected\\ Output\end{tabular}} \\ \cline{3-5}
& & U & CNOT & M & & \\ \hline
BV\_3 & 4 & 9 & 2 & 3 & 6 & 110 \\
Adder & 3 & 30 & 17 & 2 & 28&0100 \\
Grover & 3 & 24 & 7 & 2 & 22&10 \\
Toffoli & 3 & 11 & 6 & 3 & 12&111 \\
QFT\_3 & 3 & 10 & 6 & 3 & 14&000 \\
QFT\_4 & 4 & 26 & 18 & 4 & 28&0000 \\
Decoder & 5 & 28 & 21 & 4 & 29&00100 \\
QPE\_3 & 3 & 19 & 7 & 2 & 17&01 \\
QPE\_4 & 4 & 24 & 14 & 3 & 31&001 \\
QAOA\_1 & 15 & 437 & 111 & 15 & 176 & NA \\
QAOA\_2 & 15 & 525 & 135 & 15 &214 & NA\\
QAOA\_3 & 15 & 602 & 156 & 15 &224 & NA \\ \hline
\end{tabular}
}
\end{table}
\begin{table}[b]
\centering
\caption{Properties of different IBM quantum computers.}
\label{tab:Res1}
{%
\begin{tabular}{|c|c|c|c|c|c|}\hline
\begin{tabular}[c]{@{}l@{}}Quantum Computer\end{tabular} &
\begin{tabular}[c]{@{}l@{}}\# Qubits\end{tabular} &
\begin{tabular}[c]{@{}l@{}}QV\end{tabular} \\\hline
IBM Q16 Melbourne & 15 & 8 \\
IBM Q Casablanca & 7 & 32 \\
IBM Q Santiago & 5 & 32 \\
IBM Q Rome & 5 & 32 \\ \hline
\end{tabular}%
}
\end{table}
\begin{figure*}[t]
\centering
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=3.5in]{Figures/m.eps}
\caption{}
\label{fig:(a)}
\end{subfigure}\hfill
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=3.5in]{Figures/c.eps}
\caption{}
\label{fig:(b)}
\end{subfigure}\hfill
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=3.5in]{Figures/s.eps}
\caption{}
\label{fig:(c)}
\end{subfigure}\hfill
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=3.5in]{Figures/r.eps}
\caption{}
\label{fig:(d)}
\end{subfigure}
\caption{{Probability of Successful Trails (PST) for different benchmarks executed on various IBMQ architectures}}
\label{fig:PST1}
\end{figure*}
\begin{table*}[h]
\centering
\caption{Properties of physical quantum circuits generated using Prop. I.}
\label{tab:acc}
{%
\begin{tabular}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|c|c|c@{\hspace*{0.1cm}}|c@{\hspace*{0.1cm}}|c|c|c@{\hspace*{0.1cm}}|c@{\hspace*{0.1cm}}|c|c|c@{\hspace*{0.1cm}}|c@{\hspace*{0.1cm}}|c|c|c@{\hspace*{0.1cm}}|c@{\hspace*{0.1cm}}|}
\hline
\multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}c@{}}Benc-\\hmark\end{tabular}}} & \multicolumn{4}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{Melbourne} & \multicolumn{4}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{Casablanca} & \multicolumn{4}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}@{\hspace*{0.1cm}}|}{Santiago} & \multicolumn{4}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{Rome}\\ \cline{2-17}
\multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{} & \multicolumn{2}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{WESP} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{R}} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{D}} & \multicolumn{2}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{WESP} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{R}} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{D}} & \multicolumn{2}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{WESP} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{R}} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{D}} & \multicolumn{2}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{WESP} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{R}} & \multicolumn{1}{@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{\multirow{2}{*}{D}} \\ \cline{2-3} \cline{6-7} \cline{10-11} \cline{14-15}
\multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{0.1cm}}|}{Base} & \multicolumn{1}{l@{\hspace*{0.1cm}}|}{Prop. I} & \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{}& \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{0.1cm}}|}{Base} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{0.1cm}}|}{Prop. I} & \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{}& \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{0.1cm}}|}{Base} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{0.1cm}}|}{Prop. I} & \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{} & \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{0.1cm}}|}{Base} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{0.1cm}}|}{Prop. I} & \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{}& \multicolumn{1}{|@{\hspace*{0.1cm}}c@{\hspace*{0.1cm}}|}{}\\ \hline
BV\_3 & 6.666E-1 & 6.672E-1 & 1 & 34 & 7.064E-1 & 7.071E-1 & 1 & 16 & 8.590E-1 & 8.591E-1 & 1 & 15 & 9.078E-1 & 9.079E-1 & 1 & 15 \\
Adder & 4.243E-1 & 4.945E-1 & 7 & 74 & 5.992E-1 & 6.442E-1 & 4 & 102 & 4.662E-1 & 5.339E-1 & 5 & 69 & 7.330E-1 & 7.743E-1 & 3 & 74 \\
Grover & 3.909E-1 & 3.940E-1 & 3 & 51 & 8.312E-1 & 8.432E-1 & 2 & 56 & 8.052E-1 & 8.218E-1 & 10 & 73 & 8.756E-1 & 8.831E-1 & 1 & 65 \\
Toffoli & 4.246E-1 & 4.292E-1 & 2 & 16 & 7.516E-1 & 7.870E-1 & 2 & 51 & 7.820E-1 & 7.998E-1 & 3 & 45 & 9.006E-1 & 9.107E-1 & 2 & 44 \\
QFT\_3 & 5.940E-1 & 6.278E-1 & 3 & 65 & 7.942E-1 & 8.127E-1 & 2 & 68 & 7.146E-1 & 7.284E-1 & 7 & 71 & 8.681E-1 & 8.833E-1 & 4 & 65 \\
QFT\_4 & 4.784E-1 & 4.789E-1 & 3 & 94 & 6.978E-1 & 7.235E-1 & 2 & 137 & 4.271E-1 & 4.572E-1 & 4 & 150 & 7.604E-1 & 7.800E-1 & 2 & 76 \\
Decoder & 2.886E-2 & 4.869E-2 & 5 & 141 & 2.648E-1 & 3.059E-1 & 3 & 200 & 2.117E-1 & 2.824E-1 & 6 & 215 & 5.079E-1 & 5.705E-1 & 6 & 229 \\
QPE\_3 & 3.430E-1 & 3.438E-1 & 1 & 77 & 7.664E-1 & 7.911E-1 & 2 & 71 & 7.422E-1 & 7.635E-1 & 9 & 92 & 8.374E-1 & 8.377E-1 & 1 & 92 \\
QPE\_4 & 1.528E-1 & 1.705E-1 & 3 & 127 & 6.691E-1 & 7.096E-1 & 3 & 134 & 3.467E-1 & 4.034E-1 & 5 & 117 & 6.842E-1 & 7.232E-1 & 8 & 102 \\
\hline
\end{tabular}
}
\end{table*}
We run our circuits on a variety of quantum architectures with different number of qubits and Quantum Volume (QV), as shown in Table~\ref{tab:Res1}. The QV determines the largest size of random quantum circuits in terms of the number of qubits and the circuit depth, which can be successfully executed on the quantum computer with high fidelit
~\cite{Cross_2019}. Each execution of the quantum circuit consists of 8192 shots/trials.
All the benchmark circuits are mapped to the quantum architecture using Qiskit software development kit based on the error rates of the quantum hardware~\cite{Qiskit}. We implement our rescheduling algorithms using Python programming language for easy integration with Qiskit. We run our algorithms on a 2.10 GHz Intel Xeon(R) CPU E5-2620 processor with 62.8 GB memory.
We execute 150, 153, 121, and 117 pairs of base and reordered circuits on IBM Q Melbourne, IBM Q Santiago, IBM Q Rome, and IBM Q Casablanca, respectively, based on the availability of these devices. The results of all these circuits show that the reordered quantum circuits according to WESP provided in Section V yield a better solution than the base circuits generated by Qiskit. We select a subset of these circuits to show the effectiveness of our approach.
\iffalse
We use a total of 9 benchmark circuits with different qubits based on the quantum architecture. We use a variety of quantum architectures with different qubit sizes and quantum volumes, as shown in Table~\ref{tab:Res1}. The quantum volume measures the error rates and potential of a quantum computer~\cite{Cross_2019}\todo{please verify. My understanding is that it has to do with the depth and the number of qubits of quantum circuits that can be executed on the hardware. Check the reference and write it nicely}. The \textbf{Bernstein-Vazirani (BV)} algorithm finds a string for a given input function~\cite{10.1137/S0097539796300921}. The Discrete Fourier Transform representation of the quantum circuit is \textbf{Quantum Fourier Transform (QFT)}~\cite{nielsen_chuang_2019}. A shift operation is carried out for a given input Boolean function by \textbf{Hidden Shift (HS)} algorithm~\cite{10.5555/644108.644189}. The value of $x$ is found using \textbf{Grover Search} algorithm that sets the function $f(x)$ to one. Given a unitary operator, the eigenvalue of the eigenvector is estimated using \textbf{Quantum Phase Estimation (QPE)} algorithm~\cite{nielsen_chuang_2019}.
\textbf{Quantum Approximate Optimization Algorithm (QAOA)}~\cite{farhi2014quantum} helps in solving optimization problems such as Max-cut problem. \textbf{Other reversible quantum circuits} such as \textbf{Adder, Toffoli gate, Graycode, and Decoder} are also used. We obtained QFT and QPE circuits from Qiskit~\cite{Qiskit}, HS and QAOA circuits from Cirq~\cite{quantum_ai_team_and_collaborators_2020_4062499}, and adder and decoder circuits from RevLib~\cite{4539430}. The structural properties of all the benchmark circuits are described in Table~\ref{tab:prop}. To extract the QAOA benchmark circuit, we used $p=1$ for default alpha and beta parameters provided by Cirq.
We perform the rescheduling algorithm on Ubuntu, having a 2.10 GHz Intel Xeon(R) CPU E5-2620 processor with 62.8 GB memory and IBMQ quantum computers. The execution time for commutative property-based reordering for each benchmark circuit on various quantum architectures is shown in Table~\ref{tab:exec}. All the benchmark circuits are mapped to the quantum architecture using Qiskit SDK~\cite{Qiskit}. We execute the post mapped quantum circuits on different IBMQ computers by using 8192 shots for each run. TThe algorithms are implemented using python for easy integration with qiskit, which slows down the execution time. C/C++ can be used to implement the same algorithm to improve the execution time.
\fi
\subsection{Elementary gate rescheduling results}
In the first experiment, we study the impact of our proposed elementary gate rescheduling on the PST of different benchmark circuits. For each quantum circuit to be executed on a given quantum computer, we first generate the corresponding physical quantum circuit using the highest optimization level of Qiskit based on the calibration data of the quantum hardware. We refer to this approach as a \textit{Base}. Next, we apply our proposed elementary gate rescheduling algorithm based on WESP to the physical quantum circuit to generate a new optimized physical quantum circuit. We refer to this process as \textit{Prop. I}. We run the two physical quantum circuits on the quantum computer and evaluate their PST.
Figure~\ref{fig:PST1} provides a comparison of the PST of quantum circuits generated using Base and Prop. I and executed on different quantum computers.
We also report the WESP of quantum circuits generated using Base and Prop. I, the number of gate reordering using Prop. I (R), and the depth of each physical quantum circuit (D) executed on different quantum computers in Table~\ref{tab:acc}. We emphasize that the depth of the physical quantum circuit prior to and post applying Prop. I is always the same to avoid additional decoherence errors.
Figure~\ref{fig:PST1} shows that our proposed rescheduling algorithm improves the PST, and thus, the fidelity of the quantum circuit output. We observe significant improvement of PST for some quantum circuits more than others such as Adder, Toffoli, and Grover Search quantum circuits for most of the quantum computers, which implies that the circuit structure can impact the effectiveness of our approach. We also observe that our proposed approach can also perform well for quantum computers with a larger quantum volume, and thus, a better fidelity such as IBM Q Santiago, IBM Q Casablanca, and IBM Q Rome.
Furthermore, as our approach exploits the variation in gate error rates, which can result in different error rates propagating throughout different circuit paths, it is expected to perform better in the presence of higher gate error standard deviation of the quantum hardware. This is further illustrated in Figure~\ref{fig:PST1}, for which the standard deviation of the two-qubit gate error of IBM Q16 Melbourne, IBM Q Casablanca, IBM Q Santiago, and IBM Q Rome on average are 1.31E-02, 1.86E-02, 1.33E-02, 1.20E-02, respectively, which are considered relatively high.
As shown in~\cite{9296804}, the output state fidelity of the quantum circuit can be improved by rescheduling quantum gates as late as possible in the circuit to reduce the usage time of the qubit, and thus, reduce the qubit decoherence errors.
To ensure that the improvement in PST in Figure~\ref{fig:PST1} is not due to a reduction in the qubit life time, and thus, decoherence errors, we identify the first layer in which each physical qubit is being used in each pair of quantum circuits generated using the Base and the Prop. I approaches. For non-measured qubits we also identify the last layer being used in both physical quantum circuits. For measured qubits, all measurement operations will be at the end of the circuit according to IBM quantum computers. We observe that none of the circuits used in Figure~\ref{fig:PST1} exhibits any reduction in their physical qubit lifetime after running the rescheduling algorithm, which confirms the effectiveness of our approach in reducing the impact of Pauli errors on the output state fidelity of the quantum circuit.
The run time of our proposed elementary gate rescheduling algorithm for each benchmark circuit to be executed on various quantum architectures is shown in Table~\ref{tab:exec}. Each
quantum circuit requires only a small fraction of a second for gate rescheduling time.
\begin{table}
\centering
\caption{Run time of Prop. I algorithm applied to different quantum circuits mapped to different IBM Q computers}
\label{tab:exec}
\begin{tabular}{|l|r|r|r|r|}
\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}c@{}}Benc-\end{tabular}}} & \multicolumn{4}{c|}{Prop. I Execution Time (msec)} \\ \cline{2-5}
\multicolumn{1}{|c|}{hmark} & \multicolumn{1}{l|}{\begin{tabular}[c]{@{}c@{}}Melb-\\ourne\end{tabular}} & \multicolumn{1}{l|}{\begin{tabular}[c]{@{}c@{}}Casa-\\blanca\end{tabular}} & \multicolumn{1}{l|}{\begin{tabular}[c]{@{}c@{}}San-\\tiago\end{tabular}}& \multicolumn{1}{l|}{Rome} \\
\hline
BV\_3& 0.54 & 0.17 & 0.17 & 0.23 \\
Adder& 1.25 & 2.17 & 1.37 & 1.52 \\
Grover & 1.14 & 1.27 & 1.52 & 1.34 \\
Toffoli & 0.20 & 0.53 & 0.85 & 0.90 \\
QFT\_3 & 1.12 & 1.38 & 1.49 & 1.26 \\
QFT\_4 & 1.47 & 2.59 & 2.92 & 1.57 \\
Decoder & 2.52 & 3.27 & 3.11 & 3.28 \\
QPE\_3 & 1.57 & 1.51 & 1.93 & 1.72 \\
QPE\_4 & 2.51 & 1.86 & 2.32 & 1.89 \\
\hline
\end{tabular}
\end{table}
\iffalse
\begin{table*}[t]
\centering
\caption{WESP.}
\label{tab:acc}
{%
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multicolumn{1}{|c|}{\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}c@{}}Benc-\\hmark\end{tabular}}} & \multicolumn{4}{c|}{Melbourne} & \multicolumn{4}{c|}{Casablanca} & \multicolumn{4}{c|}{Santiago} & \multicolumn{4}{c|}{Rome}\\ \cline{2-17}
\multicolumn{1}{|c|}{} & \multicolumn{2}{c|}{WESP} & \multicolumn{1}{c|}{\multirow{2}{*}{R}} & \multicolumn{1}{c|}{\multirow{2}{*}{D}} & \multicolumn{2}{c|}{WESP} & \multicolumn{1}{c|}{\multirow{2}{*}{R}} & \multicolumn{1}{c|}{\multirow{2}{*}{D}} & \multicolumn{2}{c|}{WESP} & \multicolumn{1}{c|}{\multirow{2}{*}{R}} & \multicolumn{1}{c|}{\multirow{2}{*}{D}} & \multicolumn{2}{c|}{WESP} & \multicolumn{1}{c|}{\multirow{2}{*}{R}} & \multicolumn{1}{c|}{\multirow{2}{*}{D}} \\ \cline{2-3} \cline{6-7} \cline{10-11} \cline{14-15}
\multicolumn{1}{|c|}{} & \multicolumn{1}{l|}{Base} & \multicolumn{1}{l|}{\emph{Prop. I}} & \multicolumn{1}{|c|}{}& \multicolumn{1}{|c|}{} & \multicolumn{1}{l|}{Base} & \multicolumn{1}{l|}{\emph{Prop. I}} & \multicolumn{1}{|c|}{}& \multicolumn{1}{|c|}{} & \multicolumn{1}{l|}{Base} & \multicolumn{1}{l|}{\emph{Prop. I}} & \multicolumn{1}{|c|}{} & \multicolumn{1}{|c|}{} & \multicolumn{1}{l|}{Base} & \multicolumn{1}{l|}{\emph{Prop. I}} & \multicolumn{1}{|c|}{}& \multicolumn{1}{|c|}{}\\ \hline
BV\_3 & 6.666E-1 & 6.672E-1 & 1 & 34 & 7.678E-1 & 7.683E-1 & 1 & 15 & 8.590E-1 & 8.591E-1 & 1 & 15 & 9.078E-1 & 9.079E-1 & 1 & 15 \\
Adder & 4.243E-1 & 4.945E-1 & 7 & 74 & 3.226E-1 & 3.895E-1 & 4 & 102 & 4.662E-1 & 5.339E-1 & 5 & 69 & 7.330E-1 & 7.743E-1 & 3 & 74 \\
Grover & 3.909E-1 & 3.940E-1 & 3 & 51 & 7.418E-1 & 7.587E-1 & 4 & 58 & 8.052E-1 & 8.218E-1 & 10 & 73 & 8.756E-1 & 8.831E-1 & 1 & 65 \\
Toffoli & 4.246E-1 & 4.292E-1 & 2 & 16 & 8.339E-1 & 8.462E-1 & 2 & 36 & 7.820E-1 & 7.998E-1 & 3 & 45 & 9.006E-1 & 9.107E-1 & 2 & 44 \\
QFT\_3 & 5.940E-1 & 6.278E-1 & 3 & 65 & 8.428E-1 & 8.603E-1 & 3 & 65 & 7.146E-1 & 7.284E-1 & 7 & 71 & 8.681E-1 & 8.833E-1 & 4 & 65 \\
QFT\_4 & 4.784E-1 & 4.789E-1 & 3 & 94 & 7.116E-1 & 7.421E-1 & 5 & 140 & 4.271E-1 & 4.572E-1 & 4 & 150 & 7.604E-1 & 7.800E-1 & 2 & 76 \\
Decoder & 2.886E-2 & 4.869E-2 & 5 & 141 & 4.267E-2 & 8.362E-2 & 2 & 239 & 2.117E-1 & 2.824E-1 & 6 & 215 & 5.079E-1 & 5.705E-1 & 6 & 229 \\
QPE\_3 & 3.430E-1 & 3.438E-1 & 1 & 77 & 6.369E-1 & 6.713E-1 & 5 & 74 & 7.422E-1 & 7.635E-1 & 9 & 92 & 8.374E-1 & 8.377E-1 & 1 & 92 \\
QPE\_4 & 1.528E-1 & 1.705E-1 & 3 & 127 & 6.928E-1 & 7.396E-1 & 6 & 103 & 3.467E-1 & 4.034E-1 & 5 & 117 & 6.842E-1 & 7.232E-1 & 8 & 102 \\
\hline
\end{tabular}
}
\end{table*}
\fi
\subsection{Complex gate rescheduling results}
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\includegraphics[width=3.3 in]{Figures/ar_.eps}
\caption{}
\label{fig:a1}
\end{subfigure} \vspace{1mm}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=3.3 in]{Figures/arg_.eps}
\caption{}
\label{fig:a2}
\end{subfigure
\caption{(a) The approximation ratio, and (b) the corresponding approximation ratio gap for different QAOA circuits.}
\label{fig:approxqaoa}
\end{figure}
In the second experiment, we show the effectiveness of our proposed ZZ complex gate rescheduling algorithm with respect to AR, and the corresponding ARG of different QAOA quantum circuits, namely QAOA\_1, QAOA\_2, and QAOA\_3, executed on IBM Q16 Melbourne quantum computer. We refer to the complex gate rescheduling algorithm as \textit{Prop. II}. We also study the impact of applying Prop. II followed by Prop. I approach to generate the physical QAOA quantum circuit, which implies the use of ZZ complex gate rescheduling algorithm first followed by elementary gate rescheduling algorithm after complex gate decomposition based on the commutation rules in Figure~\ref{fig:three graphs}. This process is referred to as \textit{Prop. I+II}.
Figure~\ref{fig:approxqaoa}(a) and (b) provide a comparison of the AR, and the corresponding ARG, respectively, of different QAOA circuits generated using Base, Prop. I, Prop. II, and Prop. I+II.
The depth of physical QAOA\_1, QAOA\_2, and QAOA\_3 quantum circuits are 315, 375, and 412, respectively, for all the proposed rescheduling approaches.
The range of the number of reordering for QAOA\_1, QAOA\_2, and QAOA\_3 quantum circuits are 3-35, 5-39, and 7-53, respectively.
The WESP computed at the complex and the elementary levels of the base and reordered QAOA quantum circuits and their corresponding ESP are provided in Table~\ref{tab:wesp_qaoa}.
Our results show that while ZZ gates rescheduling algorithm improves the AR of the QAOA circuits, and thus, reduces the ARG, applying elementary gate rescheduling can provide a better output state fidelity. Our results also show that using both rescheduling algorithms at the complex and the elementary gate-level can deliver the best reduction in the ARG, and therefore, the best output state fidelity.
Thus, as both decoherence and gate errors contribute to the quantum circuit noise, by incorporating our gate rescheduling approaches that push very noisy gates to as later circuit layers as possible we can improve the output state fidelity even for QAOA circuits with large depth.
\begin{table}[]
\centering
{
\caption{{WESP of QAOA quantum circuits generated using the Base, Prop.I, Prop. II, and Prop. I+II and their corresponding ESP value.}}
\label{tab:wesp_qaoa}
\begin{tabular}{|@{\hspace*{0.08cm}}l@{\hspace*{0.1cm}}|@{\hspace*{0.008cm}}c@{\hspace*{0.08cm}}|@{\hspace*{0.008cm}}c@{\hspace*{0.08cm}}|@{\hspace*{0.01cm}}c@{\hspace*{0.08cm}}|@{\hspace*{0.008cm}}c@{\hspace*{0.08cm}}|@{\hspace*{0.009cm}}c@{\hspace*{0.08cm}}|@{\hspace*{0.01cm}}c@{\hspace*{0.08cm}}|}
\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}c@{}}Benc-\\hmark\end{tabular}}} & \multicolumn{1}{c|}{\multirow{2}{*}{ESP}} & \multicolumn{2}{l|}{WESP (complex)} & \multicolumn{3}{c|}{WESP (elementary)} \\
\cline{3-7}
\multicolumn{1}{|c|}{hmark} &\multicolumn{1}{c|}{} & \multicolumn{1}{l|}{Base} & \multicolumn{1}{@{\hspace*{0.1cm}}l@{\hspace*{-0.8cm}}|}{Prop. II}& \multicolumn{1}{l|}{Base} & \multicolumn{1}{@{\hspace*{0.02cm}}l@{\hspace*{0.2cm}}|}{Prop. I} & \multicolumn{1}{@{\hspace*{0.02cm}}l@{\hspace*{0.02cm}}|}{Prop. I+II} \\
\hline
QAOA\_1 & 1.03E-04 & 5.01E-05 & 6.32E-05 & 1.69E-05 & 5.09E-05 & 7.19E-05 \\
QAOA\_2 & 6.81E-07 & 2.49E-07 & 4.58E-07 & 5.61E-08 & 2.31E-07 & 5.36E-07 \\
QAOA\_3 & 9.05E-10 & 3.78E-10 & 5.63E-10 & 1.66E-11 & 4.98E-10 & 7.92E-10 \\
\hline
\end{tabular}
}
\end{table}
The run time of our proposed rescheduling approaches for QAOA\_1, QAOA\_2, and QAOA\_3 quantum circuits to be executed on IBM Q16 Melbourne quantum computer is shown in Table~\ref{tab:qaoa}. Our proposed approaches consume only a small fraction of a second despite the large circuit depth and the number of qubits.
\begin{table}
\centering
\caption{Run time of Prop. I, Prop. II, and Prop. I+II algorithms for different QAOA circuits mapped to IBM Q16 Melbourne.}
\label{tab:qaoa}
\begin{tabular}{|l|r|r|r|}
\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}c@{}}Benc-\\hmark\end{tabular}}} & \multicolumn{3}{c|}{Execution Time (msec)} \\
\cline{2-4}
\multicolumn{1}{|c|}{hmark} & \multicolumn{1}{l|}{Prop. I} & \multicolumn{1}{l|}{Prop. II} & \multicolumn{1}{l|}{Prop. I+II} \\
\hline
QAOA\_1 & 3.15 & 3.01 & 6.17 \\
QAOA\_2 & 3.22 & 3.11 & 6.35 \\
QAOA\_3 & 3.53 & 3.37 & 7.26 \\
\hline
\end{tabular}
\end{table}
\section{Conclusion}
In this paper, we propose gate rescheduling algorithms based on the gate error propagation paths in the quantum circuits. Given the variation in the error rates of NISQ computers, we show that the location of the quantum gate can significantly affect the output state fidelity of the quantum circuit. Our proposed approaches can be easily integrated with other quantum compilation approaches, which ensure that the quantum circuit mapping process maximizes ESP. We also anticipate further improvements in the output state fidelity when applying our approaches with other compilation approaches that target decoherence and correlated errors. Our future work will further investigate the quantum circuit structures that benefit the most from our proposed gate rescheduling algorithms.
\iffalse
In this paper, we propose gate rescheduling algorithms based on the impact of the gate error propagation paths on the output state fidelity of quantum circuits executed on quantum computers with variable error rates. Our proposed approaches can be easily integrated with other quantum compilation approaches that carefully map quantum circuits to the target hardware to maximize ESP. We anticipate further improvements in the output state fidelity when applying our approaches with other compilation approaches that target decoherence errors and correlated errors. Our future work will further investigate the quantum circuit structures that benefit the most from our proposed gate rescheduling algorithms.
\fi
\section{Acknowledgement}
We thank IBM Quantum for the access to the IBM processors through the IBM Quantum Researchers Program.
\bibliographystyle{IEEEtran}
|
\subsection{Few-Shot Learning}
MAML~\cite{finn2017model} is a representative of gradient-based meta-learning approach, its goal is to obtain optimal initial model parameters for \textit{unseen} tasks. Suppose there are a set of meta-training tasks sampled from a task distribution $p(\mathcal{T})$. Each task $\mathcal{T}_i$ is split into support (training) set and query (validation) set $\{\mathcal{S}_i, \mathcal{Q}_i\}$. As a bi-level optimization problem, in the \textit{inner loop}, MAML adapts the task-specific model parameters $\phi_i$ from initialization parameters $\theta$ for task $\mathcal{T}_i$ based on its support set: $\phi_i = \operatornamewithlimits{arg\,min}_{\theta} \left[L({\theta}; \mathcal{S}_i)\right]$ ($L$ is the loss of model parameterized by $\theta$ on data $\mathcal{S}_i$). The loss of adapted model $\phi_i$ on the corresponding query set $L(\phi_i; \mathcal{Q}_i)$ is used to evaluate the performance.
In the \textit{outer loop}, the averaged query set loss is minimized to obtain the optimal initial parameters.
Therefore, the objective function could be formulated as follows:
\begin{equation}
\label{meta-objective}
\theta^*= \operatornamewithlimits{arg\,min}_{\theta \in {\Theta}}{\mathbb{E}_{\mathcal{T}_i\sim p(\mathcal{T})}}\left[L(\text{Alg}(\theta; \mathcal{S}_{i}); \mathcal{Q}_i)\right]
\end{equation}
where $\text{Alg}(\theta; \mathcal{S}_{i})$ corresponds to single or multiple gradient descent steps in the inner-level task-specific adaptation. In case of single-step gradient update, $\text{Alg}(\theta; \mathcal{S}_{i})$ can be specified as following:
\begin{equation}
\label{param-adaptation}
\phi_i = \text{Alg}(\theta; \mathcal{S}_{i}) \approx \theta - \alpha \nabla_{\theta}{L}(\theta; \mathcal{S}_{i})
\end{equation}
where $\alpha $ is a learning rate. The learned meta-parameters $\theta^{*}$ from outer-level will be leveraged as model initialization for the \textit{unseen} tasks during meta-test stage. A table of notations with corresponding explanations is given in Appendix~\ref{app:notations}.
\section{Notation}
\label{app:notations}
For clear interpretation, we list the notations used in this paper and their corresponding explanation, as shown in Table~\ref{tab:notation}.
{\renewcommand{\arraystretch}{1.2}
\begin{table}[ht!]
\caption{Important Notations and Descriptions}
\label{tab:notation}
\centering
\vskip 0.15in
\begin{center}
\begin{small}
\begin{tabular}{ll}
\toprule
\bfseries{Notation} & \bfseries{Description} \\
\midrule
$p(\mathcal{T})$ & Probability distribution of meta-training tasks \\
$N$ & The number of meta-training tasks\\
$M$-way, $K$-shot & The number of classes in one task $M$, and each class contains $K$ examples in the support set \\
$\mathcal{T}_i$ & The $i$-th meta-training task \\
$\{\mathcal{S}_i$, $\mathcal{Q}_i$, $\mathcal{U}_i$\} & Support set, query set, and unlabeled set of meta-training task $\mathcal{T}_i$ \\
\{$\mathcal{S}^{new}$, $\mathcal{Q}^{new}$, $\mathcal{U}^{new}$\} & Support set, query set, and unlabeled set for task $\mathcal{T}'_i$ in meta-test \\
$\mathcal{A}^{s}_{i}$ & Selected unlabeled examples from Unlabeled set for task $\mathcal{T}_i$ in the inner loop in meta-training \\
$\mathcal{A}^{q}_{i}$ & Selected unlabeled examples from Unlabeled set for task $\mathcal{T}_i$ in the outer loop in meta-training \\
$\theta$ & Initial parameters of base learner \\
$\phi_i$ & Task-specific parameters for task $\mathcal{T}_i$ \\
${L}(\phi; \mathcal{D})$ & Loss function on dataset $\mathcal{D}$ characterized by model parameter $\phi$ \\
$L_{l}$, $L_{u}$ & Cross entropy loss on labeled data ( or hypothesized labeled data) \\
$\text{Alg}(\theta; \mathcal{D})$ & One or multiple steps of gradient descent initialized at $\theta$ on dataset $\mathcal{D}$\\
$\alpha, \beta$ & Learning rate in the inner loop and outer loop \\
$\tau_{\text{in}}$, $\tau_{\text{out}}$ & Temperature annealing coefficient in the inner (or outer) loop \\
$B_{\text{in}}$, $B_{\text{out}}$ & Budget in the inner (or outer) loop selection among all classes in the task\\
$T_{\text{in}}$, $T_{\text{out}}$ & Total number of steps in the inner loop; The number of epochs in the outer loop; \\
$T_{\text{warm}}$ & Warm start epoch in the outer loop \\
$f$ & A submodular function\\
$S_{ij}$ & similarity between sample $i$ and $j$ \\
$I_f$ & A submodular mutual information function instantiated using a submodular function $f$\\
$\Xcal$ & Pairwise similarity matrix used to instantiate a submodular function $f$\\
\bottomrule
\end{tabular}
\end{small}
\end{center}
\end{table}
}
\section{Details of Inner and Outer SMI Subset Selection} \label{app:inner_outer}
\subsection{Inner loop}
Task specific model parameters for task $\mathcal{T}_i$:
\begin{equation}
\begin{aligned}
\phi_{i}=& \operatorname{argmin}_{\theta} L_{l}\left(\theta ; \mathcal{S}_{i}\right)+\tau_{\text{in}} L_{u}\left(\theta ; \mathcal{A}^{s}_{i}\right) \\
=&\theta-\alpha \nabla_{\theta} L_{l}\left(\theta ; \mathcal{S}_{i}\right)-\alpha \tau_{\text{in}} \nabla_{\theta} L_{u}\left(\theta ; \mathcal{A}^{s}_{i}\right) \text{(one step gradient update example)}
\end{aligned}
\end{equation}
In which, $\alpha$ is learning rate. $\tau$ is the coefficient from the pseudo labeling approach. Since it is an increasing temperature variable, let $\tau^{(t)}$ denote the $\tau$ in step $t$.
Since there are several gradient update steps in inner loop.
\begin{equation}
\phi_i^{(t+1)} = \phi_i^{(t)} - \nabla L\left(\theta ; \mathcal{S}_{i}\cup\mathcal{A}^{s}_{i}\right)
\end{equation}
Let $\phi_{i}^{(t)}$ denote the model parameters for $t$-th step for task $\mathcal{T}_i$.
\begin{itemize}
\item initialization: $\phi_{i}^{(0)} = \theta$, $\mathcal{A}^s_i=\emptyset$
\item inner step 1:
$$
\phi_i^{(1)}=\theta-\alpha \nabla_{\theta} L_{l}(\theta ; \mathcal{S}_i)-\alpha \tau_{\text{in}}^{(1)} \nabla_{\theta} L_{u}\left(\theta ; \mathcal{A}^{s}_{i}\right)
$$
Select subset for this step:
$\mathcal{A}^{s}_{i1} \leftarrow \operatorname{argmax}_{\mathcal{A}_{i1}^s \subseteq \mathcal{U}_i,|\mathcal{A}^s_{i1}| \leq B_{\text{in}}} I_{f}(\mathcal{A}^s_{i1}; \mathcal{S}_i \cup \mathcal{Q}_i)$
Set of selected examples: $\mathcal{A}^s_i$ = $\mathcal{A}^s_i \cup \mathcal{A}^s_{i1}$
In this step, the CNN model used to calculate the class probabilities for SMI is parameterized by {\color{blue}$\theta$}.
\item inner step 2:
$$
\phi_i^{(2)}=\theta-\alpha \nabla_{\phi^{(1)}} L_{l}(\phi^{(1)} ; \mathcal{S}_i)-\alpha \tau_{\text{in}}^{(2)} \nabla_{\phi^{(1)}} L_{u}\left(\phi^{(1)} ; \mathcal{A}^{s}_{i}\right)
$$
Select subset for this step:
$\mathcal{A}^{s}_{i2} \leftarrow \operatorname{argmax}_{\mathcal{A}_{i2}^s \subseteq \mathcal{U}_i,|\mathcal{A}^s_{i2}| \leq B_{\text{in}}} I_{f}(\mathcal{A}^s_{i2}; \mathcal{S}_i \cup \mathcal{Q}_i)$
Set of selected examples: $\mathcal{A}^s_i$ = $\mathcal{A}^s_i \cup \mathcal{A}^s_{i2}$
In this step, the CNN model used to calculate the class probabilities for SMI is parameterized by {\color{blue}$\phi^{(1)}$}.
\item inner step 3:
$$
\phi_i^{(3)}=\theta-\alpha \nabla_{\phi^{(2)}} L_{l}(\phi^{(2)} ; \mathcal{S}_i)-\alpha \tau_{\text{in}}^{(3)} \nabla_{\phi^{(2)}} L_{u}\left(\phi^{(2)}; \mathcal{A}^{s}_{i}\right)
$$
Select subset for this step:
$\mathcal{A}^{s}_{i3} \leftarrow \operatorname{argmax}_{\mathcal{A}_{i3}^s \subseteq \mathcal{U}_i,|\mathcal{A}^s_{i3}| \leq B_{\text{in}}} I_{f}(\mathcal{A}^s_{i3}; \mathcal{S}_i \cup \mathcal{Q}_i)$
Set of selected examples: $\mathcal{A}^s_i$ = $\mathcal{A}^s_i \cup \mathcal{A}^s_{i3}$
In this step,the CNN model used to calculate the class probabilities for SMI is parameterized by {\color{blue}$\phi^{(2)}$}.
\item continue repeat until the end of inner loop: step $T_{\text{in}}-1$.
Model parameters for task $T_i$ udpate process:
$$\phi^{(0)} (:=\theta) \rightarrow \phi^{(1)}\rightarrow \phi^{(2)}\rightarrow \phi^{(3)} ...\rightarrow \phi^{(T_{\text{in}}-1)} $$
\end{itemize}
\subsection{Outer Loop}
Meta-parameter update according to:
\begin{equation}
\theta = \operatornamewithlimits{arg\,min}_{\theta} \mathcal{J}(\theta)
\end{equation}
The final loss function is:
\begin{equation}
\begin{aligned}
\mathcal{J}(\theta)&=\frac{1}{M}\sum_{i=1}^{M} L_{l}\left(\phi_{i} ; \mathcal{Q}_{i}\right)+ \tau_{\text{out}} L_{u}\left(\phi_{i} ; \mathcal{A}^{q}_{i}\right) \\
&=\frac{1}{M}\sum_{i=1}^{M} L_{l}\left(\operatorname{argmin}_{\theta} L_{l}\left(\theta ; \mathcal{S}_{i}\right)+\tau_{\text{in}} L_{u}\left(\theta ; \mathcal{A}^{s}_{i}\right) ; \mathcal{Q}_{i}\right)+ \tau_{\text{out}} L_{u}\left(\operatorname{argmin}_{\theta} L_{l}\left(\theta ; \mathcal{S}_{i}\right)+\tau_{\text{in}} L_{u}\left(\theta ; \mathcal{A}^{s}_{i}\right) ; \mathcal{A}^{q}_{i}\right)
\end{aligned}
\end{equation}
in which, $T$ is the set of meta-training tasks. $\tau_{\text{out}}$ is still a coefficient borrowed from the pseudo label formulation.
The second equal in the above equation is according to the inner loop update:
$$
\phi_i = \operatorname{argmin}_{\theta} L_{l}\left(\theta ; \mathcal{S}_{i}\right)+\tau L_{u}\left(\theta ; \mathcal{A}^{s}_{i}\right)
$$
Since there is only one step in the outer loop, subset selection only happens one time.
\begin{equation}
\mathcal{A}^{q}_{i} \leftarrow \operatornamewithlimits{arg\,max}_{\mathcal{A}\subseteq \mathcal{U}_i\backslash\mathcal{A}^s_i, |\mathcal{A}|\leq B_{\text{out}}}I_{f}(\mathcal{A}; \mathcal{S}_i\cup \mathcal{Q}_i)
\end{equation}
The CNN model used here should be the final model parameter after the inner loop: {\color{blue}$\phi_i^{(T_{\text{in}}-1)}$}.
The motivation here to use $\phi_i^{(T_{\text{in}}-1)}$ instead of meta-parameter $\theta$ is that SMI needs a CNN model which has powerful representation. For task $\mathcal{T}_i$, $\phi_i^{(T_{\text{in}}-1)}$ is more powerful than $\theta$.
\section{Additional Experimental Detail}
\label{app:exp_details}
As aforementioned, our backbone follows the same architecture as the embedding function used by \citep{finn2017model}. Concretely, the backbone structure is made of 4 modules, each of which contains a 3$\times$3 convolutions and 64 filters, followed by batch normalization, a ReLU, and a 2$\times$2 max-pooling with stride 2. To reduce overfitting, 32 filters per layer are considered. Cross entropy loss function is used for all experiment including the loss of selected unlabeled set between the predicted labels and the hypothesized labels.
\subsection{Hyparameters tuning}
All baseline approaches including three extended prototypical networks~\cite{ren2018meta} and TPN-semi~\cite{liu2018learning} are reimplemented via their official code following the original implementation including hyper-parameters. For our \mbox{\textsc{Platinum}}{} algorithm, all step sizes ($\alpha, \beta$) are chosen from $\{$0.0001, 0.001, 0.01, 0.1$\}$. The Batch size (number of tasks per iteration) is chosen from $\{$1, 2, 4$\}$. The number of iterations are chosen from $\{$10,000,\enspace 20,000,\enspace 30,000,\enspace 40,000,\enspace 60,000$\}$. The selected best ones are: learning rate in the inner loop $\alpha=0.01$, meta parameters step size (outer learning rate) $\beta=0.0001$; the number of iterations for all experiments is set to be 60,000 (600 epochs, each epoch has 100 iterations). We monitor the accuracy and loss from meta-validation stage and save the model which has the best validation accuracy and use that to evaluate the performance on unseen novel tasks in meta-test stage.
\subsection{Additional Results}
\begin{table*}[!htbp]
\caption{1-shot classification accuracies ($\%$) of different labeled ratios on \textit{mini}ImageNet.}
\label{app:tab:classification_main_ratio_1_shot}
\small
\vspace{-3mm}
\centering
\begin{small}
\begin{tabular}{l|ccccc}
\toprule
& \multicolumn{5}{c}{\textit{1-shot}} \\
Methods & 1\% & 5\% & 10\% & 20\% & 30\% \\
\midrule
Soft k-Means~\cite{ren2018meta} & 24.61{\tiny $\pm$0.64} & 38.45{\tiny $\pm$0.81} & 40.65{\tiny $\pm$0.92} & 42.55{\tiny $\pm$0.99} & 44.09{\tiny $\pm$1.08} \\
Soft k-Means+Cluster~\cite{ren2018meta} & 15.76{\tiny $\pm$0.59} & 38.34{\tiny $\pm$0.82} & 41.15{\tiny $\pm$0.99} & 45.17{\tiny $\pm$0.95} & 47.05{\tiny $\pm$1.08} \\
Masked Soft k-Means~\cite{ren2018meta} & 25.48{\tiny $\pm$0.67} & 39.03{\tiny $\pm$0.89} & 42.91{\tiny $\pm$0.93} & 45.31{\tiny $\pm$1.01} & 47.17{\tiny $\pm$1.07} \\
TPN-semi~\cite{liu2018learning} & 40.25{\tiny $\pm$0.92} & 42.40{\tiny $\pm$0.77} & 45.78{\tiny $\pm$0.80} & 48.02{\tiny $\pm$0.82} & 47.52{\tiny $\pm$0.83} \\
MAML~\cite{finn2017model} & 35.26{\tiny $\pm$0.85} & 42.51{\tiny $\pm$0.78} & 44.29{\tiny $\pm$0.78} & 45.10{\tiny $\pm$0.75} & 45.26{\tiny $\pm$0.78} \\
PL & 37.71{\tiny $\pm$0.94} & 44.04{\tiny $\pm$0.82} & 46.58{\tiny $\pm$0.72} & 45.13{\tiny $\pm$0.78} & 44.37{\tiny $\pm$0.81} \\
\hline
\textsc{Gcmi} (ours) & 41.94{\tiny $\pm$0.96} & 44.98{\tiny $\pm$0.80} & 46.85{\tiny $\pm$0.74} & 47.72{\tiny $\pm$0.76} & 48.93{\tiny $\pm$0.70} \\
\textsc{Flmi} (ours) & \textbf{42.27}{\tiny $\pm$0.95} & \textbf{45.01}{\tiny $\pm$0.83} & \textbf{47.84}{\tiny $\pm$0.86} & \textbf{47.82}{\tiny $\pm$0.78} & \textbf{48.98}{\tiny $\pm$0.72} \\
\bottomrule
\end{tabular}
\end{small}
\vspace{-3mm}
\end{table*}
\begin{table*}[!htbp]
\caption{5-shot classification accuracies ($\%$) of different labeled ratios on \textit{mini}ImageNet.}
\label{app:tab:classification_main_ratio_5_shot}
\small
\vspace{-3mm}
\centering
\begin{small}
\begin{tabular}{l|ccccc}
\toprule
& \multicolumn{4}{c}{\textit{5-shot}} \\
Methods & 1\% & 10\% & 20\% & 30\% \\
\midrule
Soft k-Means~\cite{ren2018meta} & 38.20{\tiny $\pm$1.64} & 56.27{\tiny $\pm$1.75} & 60.13{\tiny $\pm$1.79} & 62.47{\tiny $\pm$1.65}\\
Soft k-Means+Cluster~\cite{ren2018meta} & 33.65{\tiny $\pm$1.53} & 56.87{\tiny $\pm$1.77} & 60.33{\tiny $\pm$1.81} & 62.43{\tiny $\pm$1.79}\\
Masked Soft k-Means~\cite{ren2018meta} & 39.33{\tiny $\pm$1.55} & 57.20{\tiny $\pm$1.64} & 62.50{\tiny $\pm$1.78} & 63.00{\tiny $\pm$1.77} \\
TPN-semi~\cite{liu2018learning} & 46.27{\tiny $\pm$1.67} & 60.55{\tiny $\pm$0.72} & 62.43{\tiny $\pm$0.69} & 63.10{\tiny $\pm$0.69}\\
MAML~\cite{finn2017model} & 60.22{\tiny $\pm$0.83} & 61.06{\tiny $\pm$0.81} & 63.18{\tiny $\pm$0.76} & 65.60{\tiny $\pm$0.82} \\
PL~\cite{lee2013pseudo} & 60.64{\tiny $\pm$0.92} & 61.17{\tiny $\pm$0.85} & 63.06{\tiny $\pm$0.79} & 65.14{\tiny $\pm$0.74} \\
\hline
\textsc{Gcmi} (ours) & 63.62{\tiny $\pm$0.95} & \textbf{62.72}{\tiny $\pm$0.88} & 64.78{\tiny $\pm$0.76} & 65.96{\tiny $\pm$0.74} \\
\textsc{Flmi} (ours) & \textbf{63.80}{\tiny $\pm$0.92} & 62.60{\tiny $\pm$0.86} & \textbf{65.16}{\tiny $\pm$0.74} & \textbf{66.10}{\tiny $\pm$0.79} \\
\bottomrule
\end{tabular}
\end{small}
\vspace{-3mm}
\end{table*}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=0.6\columnwidth]{figures/num_OOD_classes_1_shot.pdf}}
\caption{Comparison under different number of OOD classes in the unlabeled set for 1-shot case on \textit{mini}ImageNet. }
\label{fig:num_OOD_classes_1shot}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=0.6\columnwidth]{figures/select_accu_pl.pdf}}
\caption{Selection accuracy in the unlabeled set for 1-shot case on \textit{mini}ImageNet during meta-test for PL. }
\label{fig:select_pl}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=0.6\columnwidth]{figures/select_accu_gcmi.pdf}}
\caption{Selection accuracy in the unlabeled set for 1-shot case on \textit{mini}ImageNet during meta-test for \textsc{Gcmi}. }
\label{fig:select_gcmi}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=0.6\columnwidth]{figures/select_accu_flmi.pdf}}
\caption{Selection accuracy in the unlabeled set for 1-shot case on \textit{mini}ImageNet during meta-test for \textsc{Flmi}. }
\label{fig:select_flmi}
\end{center}
\vskip -0.2in
\end{figure}
\textbf{Selection accuracy.}
\figref{fig:select_pl}, \figref{fig:select_gcmi}, and \figref{fig:select_flmi} show the selection accuracy of three SSL algorithms: PL, \textsc{Gcmi} and \textsc{Flmi} in the inner loop during meta-test. Although \textsc{Gcmi} and \textsc{Flmi} has slightly low accuracy than PL, this verifies that our proposed \mbox{\textsc{Platinum}}{} is able to select more balanced and diverse data which are more important for model training.
\subsection{Datasets and Implementation details} \label{sec:implementation_details}
\textbf{Datasets.} We conduct experiments on three datasets: \textit{mini}ImageNet \cite{vinyals2016matching}, \textit{tiered}ImageNet \cite{ren2018meta}, and CIFAR-FS \cite{bertinetto2018meta}.
Both \textit{mini}ImageNet and \textit{tiered}ImageNet are modified subsets of the ILSVRC-12 dataset~\cite{russakovsky2015imagenet}. \textbf{\textit{mini}ImageNet} consists of 100 classes and each class has 600 images. Following the disjoint class split from ~\cite{ravi2016optimization}, we split it into 64 classes for training, 16 for validation, and 20 for test. Similarly, \textbf{\textit{tiered}ImageNet} is a larger dataset, consisting of 608 classes and each class has 768$\sim$1300 images. Classes are split into 351 for training, 97 for validation, and 160 for test~\cite{ren2018meta}. All images in these two datasets are of resolution 84$\times$84$\times$3.
\textbf{CIFAR-FS} contains 60,000 images of size 32$\times$32$\times$3 from 100 classes. We use the same class split as \textit{mini}ImageNet.
\textbf{Implementation details.} We follow the ``$K$-shot, $M$-way'' episode training setting in \cite{finn2017model} and~\cite{ren2018meta} to do semi-supervised few-shot classification experiments to evaluate \mbox{\textsc{Platinum}}{}. In all our experiments, we use first-order approximation of the MAML algorithm \cite{finn2017model}.
We implement image classification experiments in 5-way, 1-shot (5-shot) settings. Concretely, at first, all examples of each class will be randomly divided into labeled portion (where $\Scal$ and $\Qcal$ are sampled from) and unlabeled portion (where $\Ucal$ is sampled from) based on a predefined labeled ratio $\rho$, where $\rho$ is the ratio of the number of data points in the labeled portion to the total number of data points in the current class. Then, we sample each task to contain 1 (5) data points in the support set $\Scal$, and 15 (15) data points in the query set $\Qcal$ per class. For the unlabeled set $\Ucal$, we sample 50 (50) data points for each class. To select a subset for semi-supervision using SMI functions, we use a budget $B_{in}=$ 25 (25) for the inner loop, and a budget $B_{out}=$ 50 (50) for the outer loop. Note that we perform a per-class selection to assign pseduo-labels using the SMI functions, which leads to a budget of 5 and 10 data points for the inner and outer loop respectively. For our experiments in \tabref{tab:classification_main_1}, \tabref{tab:classification_main_2} and \tabref{tab:classification_main_3}, we use a labeled set ratio $\rho=0.01$. However, we also compare with a number of other $\rho$ values (see \tabref{tab:classification_main_ratio} and \tabref{tab:add_orig_40}). For our experiments with OOD classes in the unlabeled set (\tabref{tab:classification_main_2}), we use 5 distractor classes with 50 data points for each class. To make a fair comparison, we apply the same 4-layer CONV backbone architecture given in~\citep{vinyals2016matching, finn2017model} for our model and all baselines. We provide detailed hyperparameters for our experiments in Appendix \ref{app:exp_details}. We use an NVIDIA RTX A6000 GPU for our experiments. The PyTorch implementation is available at \url{https://github.com/Hugo101/PLATINUM}.
\textbf{Baselines.} We consider meta-learning based semi-supervised few-shot classification approaches as baselines, including the extended prototypical network~\cite{ren2018meta} (including Soft k-Means, Soft k-Means+Cluster, Masked Soft k-Means), TPN-semi~\cite{liu2018learning}, and reimplemented LST~\cite{li2019learning} based on the 4-layer CONV (\textit{small}/\textit{large}:small/large inner loop steps, {\em i.e.} 5/10 (meta-train/test) for \textit{small}, 20/100 for \textit{large}.). We also compare with MAML which serves as the supervised classification baseline without the additional unlabeled set. In addition, we compare with two well known approaches from the semi-supervised learning literature and implement them in the inner and outer loop on the top of MAML. The first one is Pseudo-labeling (PL)~\cite{lee2013pseudo} and the second one is a consistency regularization method, VAT~\cite{miyato2018virtual}.
\begin{table*}[!htbp]
\caption{Few-shot classification accuracy (\%) of different labeled ratios ($\rho=$1\%, 5\%, 10\%, 20\%, 30\%) on \textit{mini}ImageNet. (Due to the space limitation, we show results with 95$\%$ confidence interval in Appendix~\ref{app:exp_details}. {\textdagger}: only supervised setting is considered.)}
\label{tab:classification_main_ratio}
\small
\vspace{-2mm}
\centering
\begin{small}
\begin{tabular}{l|ccccc|cccc}
\toprule
& \multicolumn{5}{c}{\textit{1-shot}} & \multicolumn{4}{c}{\textit{5-shot}} \\
Methods & 1\% & 5\% & 10\% & 20\% & 30\% & 1\% & 10\% & 20\% & 30\% \\
\midrule
Soft k-Means~\cite{ren2018meta} & 24.61 & 38.45 & 40.65 & 42.55 & 44.09 & 38.20 & 56.27 & 60.13 & 62.47 \\
Soft k-Means+Cluster~\cite{ren2018meta} & 15.76 & 38.34 & 41.15 & 45.17 & 47.05 & 33.65 & 56.87 & 60.33 & 62.43 \\
Masked Soft k-Means~\cite{ren2018meta} & 25.48 & 39.03 & 42.91 & 45.31 & 47.17 & 39.33 & 57.20 & 62.50 & 63.00 \\
TPN-semi~\cite{liu2018learning} & 40.25 & 42.40 & 45.78 & 48.02 & 47.52 & 46.27 & 60.55 & 62.43 & 63.10 \\
MAML{}\textsuperscript{\textdagger}~\cite{finn2017model} & 35.26 & 42.51 & 44.29 & 45.10 & 45.26 & 60.22 & 61.06 & 63.18 & 65.60 \\
PL~\cite{lee2013pseudo} & 37.71 & 44.04 & 46.58 & 45.13 & 44.37 & 60.64 & 61.17 & 63.06 & 65.14 \\
\hline
\textsc{Gcmi} (ours) & 41.94 & 44.98 & 46.85 & 47.72 & 48.93 & 63.62 & \textbf{62.72} & 64.78 & 65.96 \\
\textsc{Flmi} (ours) & \textbf{42.27} & \textbf{45.01} & \textbf{47.84} & \textbf{47.82} & \textbf{48.98} & \textbf{63.80} & 62.60 & \textbf{65.16} & \textbf{66.10} \\
\bottomrule
\end{tabular}
\end{small}
\end{table*}
\begin{table*}[!ht]
\caption{\small Few-shot classification accuracy (\%) with $\rho=$40\% labeled data per class for \textit{mini}ImageNet.}
\label{tab:add_orig_40}
\small
\centering
\begin{small}
\begin{tabular}{l | c c |c c }
\toprule
& \multicolumn{2}{c|}{\textit{1-shot}} & \multicolumn{2}{c}{\textit{5-shot}}\\
Methods & \textit{w/o OOD} & \textit{w/ OOD} & \textit{w/o OOD} & \textit{w/ OOD} \\
\midrule
Soft k-Means~\cite{ren2018meta} & 50.09{\tiny $\pm$0.45} & 48.70{\tiny $\pm$0.32} & 64.59{\tiny $\pm$0.28} & 63.55{\tiny $\pm$0.28} \\
Soft k-Means Cluster~\cite{ren2018meta} & 49.03{\tiny $\pm$0.24} & 48.86{\tiny $\pm$0.32} & 63.08{\tiny $\pm$0.18} & 61.27{\tiny $\pm$0.24} \\
Masked Soft k-Means~\cite{ren2018meta} & 50.41{\tiny $\pm$0.31} & 49.04{\tiny $\pm$0.31} & 64.39{\tiny $\pm$0.24} & 62.96{\tiny $\pm$0.14} \\
TPN-semi~\cite{liu2018learning} & \textbf{52.78}{\tiny $\pm$0.27} & \textbf{50.43}{\tiny $\pm$0.84} & 66.42{\tiny $\pm$0.21} & 64.95{\tiny $\pm$0.73} \\
\hline
\textsc{Gcmi} (\textit{large}, ours) & \textbf{51.35}{\tiny $\pm$0.93} & \textbf{50.85}{\tiny $\pm$0.89} & \textbf{66.65}{\tiny $\pm$0.75} & \textbf{66.66}{\tiny $\pm$0.74} \\
\textsc{Flmi} (\textit{large}, ours) & 51.06{\tiny $\pm$0.96} & 49.83{\tiny $\pm$0.91} & \textbf{67.34}{\tiny $\pm$0.72} & \textbf{66.20}{\tiny $\pm$0.73} \\
\bottomrule
\end{tabular}
\vspace{-3mm}
\end{small}
\end{table*}
\subsection{Results} \label{sec:main_results}
In this section, we present extensive experiments that compare the performance of \mbox{\textsc{Platinum}}\ with other methods. We provide the results for 1-shot (5-shot), 5-way experiments for \textit{mini}ImageNet in \tabref{tab:classification_main_1}, \textit{tiered}ImageNet in \tabref{tab:classification_main_2}, and CIFAR-FS in \tabref{tab:classification_main_3}. On all datasets, we conduct experiments with (\textit{w/}) and without (\textit{w/o}) OOD classes in the unlabeled set. Since these experiments use $\rho=0.01$, we conduct experiments for $\rho=$0.1, 0.2 and 0.3 on \textit{mini}ImageNet, and present the results in \tabref{tab:classification_main_ratio}. \tabref{tab:add_orig_40} shows the results of $\rho=$0.4 on \textit{mini}ImageNet which is the default setting in recent papers.
\textbf{Analysis across multiple datasets.} We observe that \textsc{Flmi} outperforms other methods for the 1-shot setting on all datasets. The performance of \textsc{Flmi} is slightly better than \textsc{Gcmi} due to the additional diversity that the \textsc{Flmi} function models (see \tabref{tab:submod_inst}). When compared to other methods, the SMI functions (\textsc{Flmi} and \textsc{Gcmi}) improve the accuracy by $\approx 2 - 4\%$ over existing methods. Interestingly, in \tabref{tab:classification_main_3}, we observe that \textsc{Gcmi} outperforms \textsc{Flmi} and other baselines in the presence of OOD classes in the unlabeled set. This is expected since \textsc{Gcmi} \emph{only} models query-relevance \cite{kothawade2021prism} as opposed to \textsc{Flmi} which also models diversity.
\textbf{Varying the labeled set ratio $\rho$. } In \tabref{tab:classification_main_ratio}, we analyze different values of $\rho$ for 1-shot and 5-shot on the \textit{mini}ImageNet dataset. We observe that the gain using \mbox{\textsc{Platinum}}\ is higher when the number of labeled data points is lower than the number of unlabeled data points, \ie, $\rho$ is small. This further reinforces the need for a framework like \mbox{\textsc{Platinum}}\ which performs well in the low labeled data regime. It should be noted that, in our experiments, we use a much lower labeled ratio $\rho=$0.01, since that is a more realistic scenario. This could explain why the results of many baselines look low. We also conduct an experiment with the exact setup as~\cite{ren2018meta} where they assume labels for 40\% of examples per class for \textit{mini}ImageNet. We present the results in \tabref{tab:add_orig_40} and observe that our results still outperform~\citet{ren2018meta} and~\citet{liu2018learning}'s for 5-shot case at least. As mentioned by~\citet{ren2018meta}, their method \textit{overfits} when the labeled data is as low as $\rho=$0.1 (\textit{mini}ImageNet), whereas our method works still well in the low data regime as well.
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=0.9\columnwidth]{figures/num_OOD_classes_5_shot.pdf}}
\caption{Comparison under different number of OOD classes in the Unlabeled Set for 5-shot case on \textit{mini}ImageNet. TPN-semi is much worse than MAML by 20$\%$, so we do not put it in this figure. (1-shot case is shown in Appendix~\ref{app:exp_details}.)}
\label{fig:num_OOD_classes_5shot}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{figures/inner_outer_select.pdf}}
\caption{\textit{w/} outer selection vs. \textit{w/o} outer selection. \textbf{Left}: 1-shot, \textbf{Right}: 5-shot. Both of them are on \textit{mini}ImageNet.}
\label{fig:inner_outer_select}
\end{center}
\vskip -0.2in
\end{figure}
\subsection{Ablation Study} \label{sec:ablation_results}
\textbf{Varying the number of distractor classes.} To explore the effect of the number of OOD classes in the unlabeled set, we increased the number of OOD classes, while keeping the number of in-distribution classes to be 5. We keep using the same number of unlabeled images per class as previous experiment. In \figref{fig:num_OOD_classes_5shot}, we present the result for this ablation study on the 5-way 5-shot setting for the \textit{mini}ImageNet dataset. As expected, we observe that the accuracy during meta-testing decreases as the complexity of in-distribution subset selection increases due to larger number of OOD classes. We observe that the semi-supervision provided by the SMI based methods (\textsc{Flmi} and \textsc{Gcmi}) consistently aids MAML and outperform other methods as the number of OOD classes increase, while PL suffers and eventually performs slightly worse than MAML.
\textbf{Inner and outer loop selection.}
One of the key components of \mbox{\textsc{Platinum}}\ is embedding semi-supervision in the outer-loop. We conduct an ablation study using the 5-way 1-shot and 5-shot setting on the \textit{mini}ImageNet dataset to analyze the effect of outer-loop semi-supervision and present the results in \figref{fig:inner_outer_select}. Particularly, we evaluate the meta-test accuracy of few-shot classification with (\textit{w/}) and without (\textit{w/o}) the outer-loop selection for three methods: PL, \textsc{Gcmi} and \textsc{Flmi}. MAML is also included for comparison. We observe that providing semi-supervision in the outer loop consistently improves the performance across all experiments. Especially for 1-shot of PL, we observe an improvement of $\approx4\%$. Interestingly, PL performs worse than MAML without the outer-loop semi-supervision, and outperforms MAML with it.
\begin{table}[t!]
\caption{\small The accuracy (\%) of 5-way 5-shot experiment for $\rho=$0.4 on \textit{mini}ImageNet based on a pretrained ResNet-12 (\textit{w/o OOD}).}
\label{tab:add_resnet}
\small
\vspace{-1mm}
\centering
\begin{center}
\begin{tabular}{c c c }
\toprule
MAML & LST~\cite{li2019learning} & \textsc{Gcmi} (\textit{large}, ours)\\
\midrule
75.21{\tiny $\pm$0.65} & 78.70{\tiny $\pm$0.80} & 79.44{\tiny $\pm$0.76} \\
\bottomrule
\end{tabular}
\vspace{-3mm}
\end{center}
\vspace{-2mm}
\end{table}
\textbf{Other backbones.} We primarily use the 4-layer CONV architecture so that we could do a fair comparison with existing methods~\cite{ren2018meta,liu2018learning} which have a similar setting as ours. The more recent semi-supervised few-shot learning and meta-learning methods~\cite{li2019learning, tian2020rethinking} use ResNet-12 as backbone to boost the performance. Note that they usually require a feature pretraining procedure, which may not be available in various real-world scenarios, especially in the case of low labeled ratio in our paper. To compare, we conduct experiment based on a pretrained ResNet-12 from~\citet{tian2020rethinking} to replace 4-layer CONV, and only update the final classifier. We report results in \tabref{tab:add_resnet} and observe that SMI could improve further. It is reasonable that using a more powerful feature extractor model would still result in gains over current methods.
\section{Introduction}
Deep neural networks (DNNs) have proven to be successful in a variety of domains. However, they require large amounts of data, which might not be available for all desired tasks. In such low data regimes, they struggle to perform well. A well known approach to mitigate this problem is meta-learning, which aims to learn from multiple smaller tasks that are related to the target task. The most promising meta-learning techniques that improve the performance of DNNs are gradient based meta-learning methods \cite{finn2017model, rusu2018meta, sun2019meta}. Typically, these methods are designed to operate for few-shot learning. A natural way to improve the performance of meta-learning is by using additional unlabeled data. Semi-supervised techniques are known to use unlabeled data to improve the performance on tasks with relatively small number of labeled data \cite{oliver2018realistic, chapelle2009semi}.
\begin{figure}[t]
\centering
\begin{subfigure}[]{0.5\textwidth}
\includegraphics[width = \textwidth]{figures/intro_fig_2.pdf}
\end{subfigure}
\caption{ Semi-supervised few-shot learning setup. During meta-training, the goal is to iterate overs tasks $\Tcal_1 \cdots \Tcal_N$ and meta-learn a parametrization using the support set $\Scal$, query set $\Qcal$, and the unlabeled set $\Ucal$. During meta-testing, the learned parametrization is used as an initialization and a model is trained using the $\Scal$ and $\Ucal$ to perform well on $\Qcal$. In any task, $\Ucal$ may contain data points that are out-of-distribution, \ie not pertinent to the classes of data points in $\Scal$.
\vspace{-3ex}}
\label{fig:Intro}
\end{figure}
In this paper, we focus on few-shot classification using Model Agnostic Meta-Learning (MAML) and improve it via semi-supervised learning (see \figref{fig:Intro}). In many realistic classification tasks, although the labeled data is scarce, there is plenty of unlabeled data available for training a classifier. Towards this goal, we propose \mbox{\textsc{Platinum}}, a novel framework that embeds semi-supervision in the MAML framework by using \textbf{submodular mutual information (SMI)} \cite{smi} functions as per-class acquisition functions. We observe that embedding semi-supervision in the MAML framework is non-trivial, since simply using a pseudo-labeling approach in the inner loop does not improve the performance. This lack of improvement occurs due to either noisy pseudo-labels or class imbalance caused due to pseudo-labels being confident only for certain classes. To overcome these issues, \mbox{\textsc{Platinum}}\ uses a class-wise unique instantiations of SMI functions to provide per-class semi-supervision. Furthermore, these per-class acquired subsets are diverse, leading to a richer meta-learned parameterization (see Fig.~\ref{fig:overview}).
\subsection{Related Work}
\textbf{Few-shot learning.} There has been an extensive amount of work in few-shot learning, which has mainly revolved around supervised learning. Although our framework is embedding semi-supervision into MAML which belongs to the gradient descent family of methods, the few-shot learning literature can be broadly divided into the following categories: 1) Metric learning methods \cite{vinyals2016matching, snell2017prototypical} which deal with learning a similarity space where the task can be efficiently done with a few labeled data points. 2) Memory networks \cite{munkhdalai2017meta, santoro2016meta, oreshkin2018tadam, mishra2017simple}, which focus on learning to store ``experience" from previously observed tasks in the interest of generalizing to newer tasks. 3) Gradient based meta-learning methods \cite{finn2017model, finn2018probabilistic, antoniou2018train, ravi2016optimization, grant2018recasting, sun2019meta, killamsetty2020nested, zhao2022adaptive} which aim to meta-learn a model in the outer loop that is used as a starting point in the inner loop for a new few-shot task. In addition, some researches use pre-training strategy to boost the performance~\cite{chen2019closer, tian2020rethinking, wang2021role}. The \mbox{\textsc{Platinum}}\ framework embeds semi-supervision for gradient descent based methods that use an inner-outer bi-level optimization.\\
\textbf{Semi-supervised learning (SSL).} The goal of SSL methods is to leverage unlabeled data alongside the labeled data to obtain a better representation of the dataset than supervised learning \cite{oliver2018realistic}. The most basic SSL method, Pseudo-labeling \cite{lee2013pseudo} uses model predictions as target labels as a regularizer, and a standard supervised loss function for the unlabeled dataset. Some SSL methods like $\Pi$-Model \cite{laine2016temporal, sajjadi2016regularization} and Mean Teacher \cite{tarvainen2017mean} use consistency regularization, by using data augmentation and dropout techniques. Mean Teacher obtains a more stable target output by using an exponential moving average of parameters across previous epochs. Virtual Adversarial Training (VAT) \cite{miyato2018virtual} uses an effective regularization technique that uses slight perturbations such that the prediction of the unlabeled samples is affected the most. More recent techniques like FixMatch \cite{sohn2020fixmatch}, MixMatch \cite{berthelot2019mixmatch} and UDA \cite{xie2019unsupervised} use data augmentations like flip, rotation, and crops to predict pseudo-labels. In this paper, we propose a new SSL technique that uses class-wise instantiations of SMI functions that mitigates the issue of class-imbalance in selected subsets and is comparatively robust to OOD classes in the unlabeled set.
\textbf{Semi-supervised few-shot learning.}
There are two categories for semi-supervised few-shot learning: 1) \textit{meta-learning based}: \citet{ren2018meta} propose new extensions of Prototypical Networks \cite{snell2017prototypical} by viewing each prototype as a cluster center and tuning the cluster locations using soft K-means such that the data points in support and unlabeled sets fit better. \citet{liu2018learning} learn a graph construction module to propagate labels from labeled examples to unlabeled examples. In addition, \citet{li2019learning} propose learning to self-train (LST) which aims to meta-learn how to cherry-pick and label data points from the unlabeled set and optimizes weights of these pseudo-labels. However, their method is on the top of a pretrained meta-transfer learning \cite{sun2019meta} model which requires the labels across all training tasks to be known beforehand. Unfortunately, such meta-data about the dataset may not be available in most realistic scenarios. On the other hand, our \mbox{\textsc{Platinum}}\ framework does not require a pre-trained network or any meta-data for embedding semi-supervision in gradient descent based methods. 2) \textit{Transfer learning based} \cite{yu2020transmatch,Wang_2020_CVPR,Huang_2021_ICCV, huang2021ptn,lazarou2021iterative}: this is the main focus in more recent works. Similar to LST~\cite{li2019learning}, they assume all examples of base classes are labeled so that a feature extractor could be pretrained based on them. In contrast, we assume there are few examples per class are labeled during both meta-training and meta-test, which is more realistic than that in transfer learning based approaches. In addition, transfer learning based approaches do not leverage episodes training strategy, which is different from ours.
\textbf{Data subset selection (DSS).} DSS using submodular functions has been studied in the context of various applications like video summarization \cite{kaushal2020realistic, kaushal2019framework}, image-collection summarization \cite{tschiatschek2014learning, kothawade2020deep}, efficient learning \cite{kaushal2019learning,killamsetty2020glister,killamsetty2021grad,liu2017svitchboard}, targeted learning \cite{kothawade2021prism, kothawade2021talisman}, \etc\ Recently, \cite{kothawade2021prism} used the SMI functions for improving the performance of rare classes in the context of image classification, and \cite{kothawade2021talisman} used them for mining rare objects and slices for improving object detectors. \cite{kothawade2021similar} used the submodular information measures as acquisition functions for active learning in scenarios with class imbalance, redundancy and OOD data. \cite{killamsetty2020glister,killamsetty2021grad} studied the role of submodular functions and coresets for compute-efficient training of deep models.
\subsection{Our Contributions}
The following are our main contributions: 1) Given the limitations of existing approaches, we propose \mbox{\textsc{Platinum}}\ (see \secref{sec:method}) that uses per-class semi-supervision using SMI functions, thereby preventing class-imbalance in the selected subset for semi-supervision. 2) Our framework learns richer parameterizations by embedding semi-supervision in the inner and outer loop of MAML. 3) We conduct extensive experiments on \textit{mini}ImageNet \cite{vinyals2016matching}, \textit{tiered}ImageNet \cite{ren2018meta}, and CIFAR-FS \cite{bertinetto2018meta} datasets where the unlabeled set has in-distribution and out-of-distribution (OOD) classes. 4) We conduct various ablation experiments where we study the effect of varying the: i) ratio of labeled and unlabeled data points, ii) number of OOD classes, and iii) inner and outer loop selection for semi-supervision.
\section{Preliminaries} \label{sec:preliminaries}
\subsection{Model Agnostic Meta Learning (MAML)}
\input{MAML}
\subsection{Submodular Mutual Information}
\textbf{Submodular functions.} Submodular functions \cite{tohidi2020submodularity, bach2011learning, bach2019submodular} have been widely used for data subset selection as they naturally model properties like coverage, representation, diversity, \etc. Given a ground-set of $n$ data points $\Vcal = \{1,2,3, \cdots, n\}$, and a set function $f: 2^{\Vcal} \xrightarrow{} \mathbb{R}$. The set function $f$ is known to be submodular \cite{fujishige2005submodular} if for $x \in \Vcal$, $f(\Acal \cup x) - f(\Acal )\geq f(\Bcal \cup x) - f(\Bcal)$, $\forall \Acal \subseteq \Bcal \subseteq \Vcal$ and $x \notin \Bcal$. We use two well known submodular functions in this work, facility location (\textsc{Fl}) and graph-cut (\textsc{Gc}) (see \tabref{tab:submod_inst}(a)) that can be instantiated using a similarity kernel containing pairwise similarities between all data points. In general, submodular functions admit a $1-\frac{1}{e}$ constant factor approximation \cite{nemhauser1978analysis} for cardinality constraint maximization. Furthermore, they can be optimized in \emph{near-linear} time using greedy algorithms \cite{mirzasoleiman2015lazier}.
\\
\textbf{Submodular Mutual Information (SMI).}
In this work, we use the SMI instantiations of the above submodular functions to provide semi-supervision. Particularly, we use \textsc{Flmi} and \textsc{Gcmi} where the underlying submodular function is \textsc{Fl} and \textsc{Gc} respectively (see \tabref{tab:submod_inst}(b)). The SMI functions can be used to select data points that are semantically \emph{similar} to the data points in a given query set \cite{kothawade2021similar, kothawade2021prism, kothawade2021talisman}. To obtain pseudo-labels, we use exemplars from a particular class in the query set used to instantiate an SMI function. The subset obtained by optimizing this SMI function is then assigned labels of the class of the data points used in the query set. Formally, the submodular mutual information (SMI) is defined as $I_f(\Acal; \Rcal) = f(\Acal) + f(\Rcal) - f(\Acal \cup \Rcal)$, where $\Rcal$ is a query set. Note that \cite{smi, kothawade2021prism} propose a few other SMI functions. However, we use only the \textsc{Flmi} and \textsc{Gcmi} variants in the interest of scalability (see \secref{sec:smi_scalability}).
\begin{table}[!htb]
\caption{Instantiations of different submodular functions.
}
\label{tab:submod_inst}
\begin{subtable}{.40\linewidth}
\centering
\caption{Instantiations of submodular functions.}
\begin{tabular}{|l|l|}
\hline
\textbf{SF} & \textbf{$f(\Acal)$} \\ \hline
\scriptsize{\textsc{Fl}} & \scriptsize{$\sum\limits_{i \in \Ucal} \max\limits_{j \in \Acal} S_{ij}$} \\
\scriptsize{\textsc{Gc}} & \scriptsize{$\sum \limits_{i \in \Acal, j \in \Ucal} S_{ij} -$} \\ & \scriptsize{$ \sum\limits_{i, j \in \Acal} S_{ij}$} \\
\hline
\end{tabular}
\vspace{0.5ex}
\end{subtable}
\begin{subtable}{.40\linewidth}
\centering
\caption{Instantiations of SMI functions.}
\begin{tabular}{|c|c|c|}
\hline
\textbf{SMI} & \textbf{$I_f(\Acal;\Rcal)$} \\ \hline
\scriptsize{FLMI} & \scriptsize{$\sum\limits_{i \in \Rcal} \max\limits_{j \in \Acal} S_{ij} + $ $ \sum\limits_{i \in \Acal} \max\limits_{j \in \Rcal} S_{ij}$} \\
\scriptsize{GCMI} & \scriptsize{$2 \sum\limits_{i \in \Acal} \sum\limits_{j \in \Rcal} S_{ij}$} \\
\hline
\end{tabular}
\end{subtable}%
\end{table}
\begin{figure*}[!htbp]
\centering
\includegraphics[width = \textwidth]{figures/overview.pdf}
\caption{Overview of our \mbox{\textsc{Platinum}}{} framework that solves a semi-supervised few-shot learning problem. For a specific task $\mathcal{T}_i$, in each inner loop and outer loop gradient update step, we select a subset from the unlabeled set by maximizing the per-class SMI function (see \AlgRef{alg-ssl-fsl}). In each \textit{inner loop} step, the selected subset $\mathcal{A}^{s}_i$ and support set $\Scal_i$ will be used to update model parameters $\phi_i$ using \equref{param-adaptation-ssl}. In the \textit{outer-loop} of the meta-training stage, another subset $\mathcal{A}^{q}_{i}$ will be selected after inner loop selection according to the updated model parameters $\phi_i$. Meta-parameters $\theta$ would be updated based on $\mathcal{A}^{q}_{i}$ and the query set $\Qcal_i$ using \equref{meta-objective-ssl}. It should be noted that, as temperature annealing coefficients, $\tau_{\text{in}}$ is a function of inner step $t_{\text{in}}$, and $\tau_{\text{out}}$ is a function of the global iteration index.}
\label{fig:overview}
\vspace{-4mm}
\end{figure*}
\section{\mbox{\textsc{Platinum}}: Our Semi-Supervised Meta-Learning Framework} \label{sec:method}
\input{formulation}
\subsection{Leveraging Full Potential of SSL during Meta-Training} \label{sec:meta_training}
In this section, we discuss the meta-training pipeline of \mbox{\textsc{Platinum}}. Particularly, we emphasis on \emph{inner loop} and \emph{outer loop} semi-supervision embeded using \emph{class-wise} SMI instantiations. We detail our pipeline in \AlgRef{alg-ssl-fsl}.
For any task $\Tcal_i \sim p(\mathcal{T})$, we first initialize the model parameters $\phi_i \leftarrow \theta$, where $\theta$ is meta-learned by optimizing the outer loop on the previous tasks. Using parameters $\phi_i$, we compute an embedding containing class probabilities for each data point belonging to the unlabeled set $\Ucal_i$. Emprically, we found out that using the class probabilities based on the classes belonging to $\Tcal_i$ was as effective as using a larger and computationally expensive embedding like last layer features or gradients. Since the support and query set have labels during meta-training, we use a one-hot vector to represent the data points in $\Scal_i$ and $\Ucal_i$, where the probabilty of the class corresponding to the label is set to one. Next, we compute a pairwise similarities $\Xcal_{pq}$, where $p \in \{\Scal_i \cup \Qcal_i\}, q \in \Ucal_i$. For each class $c$, we instantiate an SMI function $I_f^c$ (\tabref{tab:submod_inst}) by using a sub-matrix $\Xcal^c$ with pairwise similarities $\Xcal^c_{pq}$ such that $p$ belongs to class $c$, $\forall p \in \{\Scal_i \cup \Qcal_i\}$. We then maximize $I_f^c$ with a budget of $B/C$ as follows:
\begin{equation}
\label{eq:per_class_smi}
\mathcal{A}^c \leftarrow \operatornamewithlimits{arg\,max}_{\mathcal{A}^c\subseteq \mathcal{U}_i, |\mathcal{A}^c|\leq B/C}I_{f}^c(\mathcal{A}^c; \mathcal{S}_i\cup \mathcal{Q}_i)
\end{equation}
where, $C$ is the number of classes in the current task $\Tcal$. Since $\Acal^c$ is obtained by optimizing the SMI function $I_f^c(\Acal^c;\Rcal)$, where $\Rcal$ contains data points \emph{only} from class $c$, we assigned the hypothesized label $c$ to all data points in $\Acal^c$. We obtain $\Acal^s = \{\Acal^1 \cup \Acal^2 \cdots \cup \Acal^C\}$ by repeating the selection for each class. Finally, we update $\phi_i$ in the inner loop by using gradient descent on $\Scal_i$ and $\Acal^s_i$. Note that $\phi_i$ is updated for $T_{\text{in}}$ steps in the inner loop. Similarly, we obtain $\Acal^q_i$ by using the SMI functions to embed semi-supervision in the outer loop and update $\theta$ by gradient descent on $\Qcal_i$ and $\Acal^q_i$. We summarize the inner loop and outer loop optimization problems in \equref{eq:inner_sel} and \equref{eq:outer_sel} respectively, and discuss them in more detail:
\begin{equation}
\mathcal{A}^{s}_{i} \leftarrow \operatornamewithlimits{arg\,max}_{\mathcal{A}^{s}_{i}\subseteq \mathcal{U}_i, |\mathcal{A}^{s}_{i}|\leq B_{\text{in}}}I_{f}(\mathcal{A}^{s}_{i}; \mathcal{S}_i\cup \mathcal{Q}_i)
\label{eq:inner_sel}
\end{equation}
\begin{equation}
\mathcal{A}^{q}_{i} \leftarrow \operatornamewithlimits{arg\,max}_{\mathcal{A}^{q}_{i}\subseteq \mathcal{U}_i\backslash\mathcal{A}^s_i, |\mathcal{A}^{q}_{i}|\leq B_{\text{out}}}I_{f}(\mathcal{A}^{q}_{i}; \mathcal{S}_i\cup \mathcal{Q}_i)
\label{eq:outer_sel}
\end{equation}
where $B_{\text{in}}$ and $B_{\text{out}}$ are selection budget per class.
\textbf{Inner loop.}
Although MAML could achieve single step gradient update in the inner loop, it is not common to have good adaptation in practice especially considering the involving of additional unlabeled set. To illustrate this clearly, we assume there are $T_{\text{in}}$ steps during the model adaptation in the inner loop. Inspired from ~\cite{lee2013pseudo}, we add some unlabeled examples to update a task-specific model $\phi_i$. Different from ~\cite{lee2013pseudo}, we do not use all examples in the unlabeled set because efficiency matters in meta-learning training procedure. Therefore, the loss function in the inner loop is formulated as below:
\begin{equation}
L\left(\theta ; \mathcal{S}_{i}\cup\mathcal{A}^{s}_{i}\right) = L_{l}\left(\theta ; \mathcal{S}_{i}\right)+\tau_{\text{in}} L_{u}\left(\theta ; \mathcal{A}^{s}_{i}\right)
\end{equation}
where $\mathcal{A}^{s}_{i}$ is the selected examples with pseudo labels.We define $L_{l}$ as the loss based on examples with true labels, and $L_{u}$ as the loss function based on hypothesized labeled examples. Similar formulations have been used in conventional semi-supervised learning approaches, such as Pseudo-Label~\cite{lee2013pseudo} and VAT~\cite{miyato2018virtual}. $\tau_{\text{in}}$ is a temperature annealing coefficient:
\begin{equation}
\small
\tau_{\text{in}}(t)= \begin{cases}0 & t<2 \\ \exp{(-5(1 - \frac{t}{T_{\text{in}}})^2)} & 2 < t\leq T_{\text{in}}\end{cases}
\end{equation}
Note that we consider multi inner step updates, and SMI subset selection happens in each step (line 6-13 in Algorithm~\ref{alg-ssl-fsl}).
\textbf{Outer loop.}
Considering the meta-parameters are updated in the outer loop based on the labeled query set, and there are few labeled examples per class, it is beneficial to augment the query set aiming to generalize well for novel class in the meta-test stage. Considering the unlabeled examples, the loss function in the outer loop could be:
\begin{equation}
\begin{aligned}
{\mathcal{J}(\theta)} &= {L}({\phi_i}; \mathcal{Q}_{i}\cup \mathcal{A}^{q}_{i}) \\
&=L_{l}\left(\phi_{i} ; \mathcal{Q}_{i}\right)+ \tau_{\text{out}} L_{u}\left(\phi_{i} ; \mathcal{A}^{q}_{i}\right) \\
\end{aligned}
\end{equation}
where $\mathcal{A}^{q}_{i}$ is the selected examples with pseudo labels, and $\tau_{\text{out}}$ is a temperature annealing coefficient:
\begin{equation}
\small
\tau_{\text{out}}(j)= \begin{cases}\exp{(-5(1 - \frac{t}{T_{\text{warm}}})^2)} & 0 < j\leq T_{\text{warm}} \\ 1 & T_{\text{warm}}<j\leq T_{\text{out}} \end{cases}
\end{equation}
$T_{\text{out}}$ is the total number of epochs during meta-training procedure. $T_{\text{warm}}$ is a warm starting epoch index. More detailed selection process explanation is given in Appendix~\ref{app:inner_outer}.
\begin{table}[!htbp]
\caption{Running time (training time of 100 tasks) comparison on \textit{mini}ImageNet domains for 1-shot (5-shot) 5-way experiments without OOD classes in unlabeled set.}
\label{tab: running_time}
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lcc}
\toprule
Methods & 1-shot (s) & 5-shot (s) \\
\midrule
MAML & 23.92 & 49.83\\
\textsc{Gcmi} & 27.33 & 58.91\\
\textsc{Flmi} & 28.94 & 56.04\\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vspace{-5mm}
\end{table}
\begin{table*}[!htbp]
\caption{Few-shot classification accuracy ($\%$) on \textit{mini}ImageNet. ({\textdagger}: only supervised setting is considered.)}
\label{tab:classification_main_1}
\small
\centering
\begin{small}
\begin{tabular}{l|cc|cc}
\toprule
& \multicolumn{2}{c|}{\textit{1-shot}} & \multicolumn{2}{c}{\textit{5-shot}}\\
Methods & \textit{w/o OOD} & \textit{w/ OOD} & \textit{w/o OOD} & \textit{w/ OOD}\\
\midrule
Soft k-Means~\cite{ren2018meta} & 24.61{\tiny $\pm$0.64} & 23.57{\tiny $\pm$0.63} & 38.20{\tiny $\pm$1.64} & 38.07{\tiny $\pm$1.53} \\
Soft k-Means+Cluster~\cite{ren2018meta} & 15.76{\tiny $\pm$0.59} & 9.77{\tiny $\pm$0.51} & 33.65{\tiny $\pm$1.53} & 30.47{\tiny $\pm$1.42} \\
Masked Soft k-Means~\cite{ren2018meta} & 25.48{\tiny $\pm$0.67} & 25.03{\tiny $\pm$0.68} & 39.33{\tiny $\pm$1.55} & 38.48{\tiny $\pm$1.74} \\
TPN-semi~\cite{liu2018learning} & 40.25{\tiny $\pm$0.92} & 26.70{\tiny $\pm$0.98} & 46.27{\tiny $\pm$1.67} & 36.81{\tiny $\pm$0.87} \\
LST(\textit{small})~\cite{li2019learning} & 37.65{\tiny $\pm$0.78} & 37.82{\tiny $\pm$0.91} & 61.50{\tiny $\pm$0.92} & 57.67{\tiny $\pm$0.85} \\
LST(\textit{large})~\cite{li2019learning} & 41.36{\tiny $\pm$0.98} & 39.32{\tiny $\pm$0.95} & 61.51{\tiny $\pm$0.98} & 59.24{\tiny $\pm$0.95} \\
{MAML}{}\textsuperscript{\textdagger} \cite{finn2017model} & 35.26{\tiny $\pm$0.85} & 35.26{\tiny $\pm$0.85} & 60.22{\tiny $\pm$0.83} & 60.20{\tiny $\pm$0.83} \\
VAT \cite{miyato2018virtual} & 36.55{\tiny $\pm$0.86} & 34.03{\tiny $\pm$0.84} & 61.60{\tiny $\pm$0.83} & 61.24{\tiny $\pm$0.88} \\
PL \cite{lee2013pseudo} & 37.71{\tiny $\pm$0.94} & 35.16{\tiny $\pm$0.85} & 60.64{\tiny $\pm$0.92} & 60.31{\tiny $\pm$0.87} \\
\hline
\textsc{Gcmi} (ours) & 41.94{\tiny $\pm$0.96} & \textbf{42.57}{\tiny $\pm$0.93} & 63.62{\tiny $\pm$0.95} & \textbf{63.54}{\tiny $\pm$0.94} \\
\textsc{Flmi} (ours) & \textbf{42.27}{\tiny $\pm$0.95} & 41.53{\tiny $\pm$0.97} & \textbf{63.80}{\tiny $\pm$0.92} & 63.44{\tiny $\pm$0.99} \\
\bottomrule
\end{tabular}
\end{small}
\end{table*}
\begin{table*}[!ht]
\caption{Few-shot classification accuracy ($\%$) on \textit{tiered}ImageNet. ({\textdagger}: only supervised setting is considered.)}
\label{tab:classification_main_2}
\small
\centering
\begin{small}
\begin{tabular}{l|cc|cc}
\toprule
& \multicolumn{2}{c|}{\textit{1-shot}} & \multicolumn{2}{c}{\textit{5-shot}}\\
Methods & \textit{w/o OOD} & \textit{w/ OOD} & \textit{w/o OOD} & \textit{w/ OOD}\\
\midrule
Soft k-Means~\cite{ren2018meta} & 27.53{\tiny $\pm$0.74} & 27.04{\tiny $\pm$0.76} & 44.63{\tiny $\pm$1.19} & 44.78{\tiny $\pm$1.05} \\
Soft k-Means+Cluster~\cite{ren2018meta} & 30.48{\tiny $\pm$0.84} & 31.30{\tiny $\pm$0.86} & 46.93{\tiny $\pm$1.18} & 49.33{\tiny $\pm$1.17} \\
Masked Soft k-Means~\cite{ren2018meta} & 33.85{\tiny $\pm$0.84} & 32.99{\tiny $\pm$0.87} & 47.63{\tiny $\pm$1.12} & 47.35{\tiny $\pm$1.08} \\
TPN-semi~\cite{liu2018learning} & 44.13{\tiny $\pm$1.04} & 31.83{\tiny $\pm$1.09} & 58.53{\tiny $\pm$1.57} & 56.92{\tiny $\pm$1.67} \\
LST(\textit{small})~\cite{li2019learning} & 42.86{\tiny $\pm$0.86} & 42.33{\tiny $\pm$0.95} & 59.55{\tiny $\pm$0.92} & 58.82{\tiny $\pm$0.93} \\
LST(\textit{large})~\cite{li2019learning} & 44.34{\tiny $\pm$0.97} & 44.59{\tiny $\pm$0.99} & 61.45{\tiny $\pm$0.90} & 60.75{\tiny $\pm$0.93} \\
{MAML}{}\textsuperscript{\textdagger} \cite{finn2017model} & 41.96{\tiny $\pm$0.84} & 41.96{\tiny $\pm$0.84} & 61.30{\tiny $\pm$0.85} & 61.30{\tiny $\pm$0.85} \\
VAT \cite{miyato2018virtual} & 41.52{\tiny $\pm$0.82} & 41.51{\tiny $\pm$0.79} & 59.98{\tiny $\pm$0.83} & 60.01{\tiny $\pm$0.87} \\
PL \cite{lee2013pseudo} & 41.22{\tiny $\pm$0.89} & 40.87{\tiny $\pm$0.83} & 61.70{\tiny $\pm$0.77}& 60.57{\tiny $\pm$0.87} \\
\hline
\textsc{Gcmi} (ours) & 45.49{\tiny $\pm$0.91} & 45.55{\tiny $\pm$0.90} & 63.67{\tiny $\pm$0.83} & \textbf{62.59}{\tiny $\pm$0.85} \\
\textsc{Flmi} (ours) & \textbf{45.63}{\tiny $\pm$0.86} & \textbf{46.19}{\tiny $\pm$0.94} & \textbf{63.75}{\tiny $\pm$0.87} & 62.19{\tiny $\pm$0.91} \\
\bottomrule
\end{tabular}
\end{small}
\vspace{-1mm}
\end{table*}
\begin{table*}[!ht]
\caption{Few-shot classification accuracy ($\%$) on CIFAR-FS. ({\textdagger}: only supervised setting is considered.)}
\label{tab:classification_main_3}
\small
\centering
\begin{small}
\begin{tabular}{l|cc|cc}
\toprule
& \multicolumn{2}{c|}{\textit{1-shot}} & \multicolumn{2}{c}{\textit{5-shot}}\\
Methods & \textit{w/o OOD} & \textit{w/ OOD} & \textit{w/o OOD} & \textit{w/ OOD}\\
\midrule
LST(\textit{small})~\cite{li2019learning} & 38.60{\tiny $\pm$0.94} & 38.61{\tiny $\pm$0.89} & 53.43{\tiny $\pm$0.94} & 51.83{\tiny $\pm$0.98} \\
LST(\textit{large})~\cite{li2019learning} & 37.65{\tiny $\pm$0.91} & 37.74{\tiny $\pm$0.97} & \textbf{55.77}{\tiny $\pm$0.93} & 52.23{\tiny $\pm$0.95} \\
{MAML}{}\textsuperscript{\textdagger} \cite{finn2017model} & 37.90{\tiny $\pm$0.91} & 37.90{\tiny $\pm$0.91} & 52.60{\tiny $\pm$0.89} & 52.60{\tiny $\pm$0.89} \\
VAT \cite{miyato2018virtual} & 39.48{\tiny $\pm$0.83} & 38.91{\tiny $\pm$0.88} & 53.20{\tiny $\pm$0.80} & 52.44{\tiny $\pm$0.83} \\
PL \cite{lee2013pseudo} & 38.11{\tiny $\pm$0.87} & 37.29{\tiny $\pm$0.92} & 52.83{\tiny $\pm$0.82} & 52.42{\tiny $\pm$0.91} \\
\hline
\textsc{Gcmi} (ours) & 40.47{\tiny $\pm$0.88} & 40.10{\tiny $\pm$0.89} & {55.01}{\tiny $\pm$0.84} & \textbf{54.42}{\tiny $\pm$0.92} \\
\textsc{Flmi} (ours) & \textbf{40.96}{\tiny $\pm$0.86} & \textbf{40.48}{\tiny $\pm$0.87} & 54.94{\tiny $\pm$0.80} & 54.16{\tiny $\pm$0.91} \\
\bottomrule
\end{tabular}
\end{small}
\vspace{-3mm}
\end{table*}
\subsection{Scalability of SMI Optimization} \label{sec:smi_scalability}
We chose to embed semi-supervision using \textsc{Flmi} and \textsc{Gcmi} in our framework due to their scalability benefits \cite{kothawade2021similar, kothawade2021talisman}. Asymptotically, the time and space complexity of computing a similarity matrix $\Xcal$ for \textsc{Flmi} and \textsc{Gcmi} is only $|\Rcal| \times |\Ucal|$. Since in the few-shot learning setting, we set $\Rcal \leftarrow {\Scal \cup \Qcal}$ which is comparatively much smaller than $\Ucal$, the complexity of these SMI functions is only $|\Ucal|$. Moreover, the SMI functions that we use are monotone and submodular which allows a $1 - \frac{1}{e}$ constant factor approximation \cite{nemhauser1978analysis}. Hence, for optimizing the SMI functions, we use a greedy algorithm \cite{nemhauser1978analysis} using memoization \cite{iyer2019memoization}. This leads to an amortized cost of $|\Ucal|\log|\Ucal|$ which can be further reduced to $|\Ucal|$ using the lazier than lazy greedy algorithm \cite{mirzasoleiman2015lazier}. Hence, \textsc{Flmi} and \textsc{Gcmi} can be optimized in linear time, making it applicable to few-shot learning datasets with a large number of tasks and large unlabeled sets.
\textbf{Time complexity of \textsc{Platinum}.} Since \textsc{Flmi} and \textsc{Gcmi} can be optimized in linear time, asymptotically it does not change in terms of the worst case in MAML. Therefore, the iteration complexity is still $\mathcal{O}(1/\epsilon^2)$~\cite{fallah2020convergence}. \tabref{tab: running_time} shows the empirical running time per epoch (100 iterations, one task per iteration) for MAML and our proposed \textsc{Gcmi} and \textsc{Flmi}.
Therefore, it is safe to say that our proposed framework \mbox{\textsc{Platinum}}{} would not slow down the original meta-learning framework (such as MAML).
\section{Experiments} \label{sec:experiments}
\input{experiment}
\section{Conclusion}
In this paper, we propose a novel semi-supervised model-agnostic meta-learning framework \mbox{\textsc{Platinum}}{}. It leverages submodular mutual information functions as per-class acquisition functions to select more balanced and diverse data from unlabeled data in the inner and outer loop of meta-learning. Meta-learning based semi-supervised few-shot learning experiments validates the effectiveness of embedding semi-supervision in the MAML by \mbox{\textsc{Platinum}}{}, especially for small ratio of labeled to unlabeled samples. We also notice that it might be useful to involve some diversity measurements for the selected subset to do quantitative analysis, we leave this as future work.
\section*{Acknowledgements}
We gratefully thank anonymous reviewers for their valuable comments. This work is supported by the National Science Foundation under Grant Numbers IIS-2106937, IIS-1954376, IIS-1815696, a gift from Google Research, and the Adobe Data Science Research award. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation, Google or Adobe.
\bibliographystyle{icml2022}
\section{How to Select: Submodular Information Measure (SIM) based approach}
This section is to introduce how to select samples from the unlabeled set $\mathcal{U}_i$ via the submodular information measures from Prof.Iyer's recent paper.
The algorithm is shown in Algorithm~\ref{alg-ssl-fsl}.Some details could be found from Prof.Iyer's paper. i.e., the instantiations of SIM, definition of similarity matrix, etc.
Notes:
1. For SMI, or SCG, we just need to do some modification according to the Table.1 from Prof.Iyer's paper.
2. The loss $\mathcal{H}$ shown in the algorithm should be the same to $L$, could it be cross entropy loss?
3. If we only consider the imbalance, and OOD scenario, which means we might only consider SMI through the setting: $\mathcal{Q} \leftarrow \mathcal{Q}, \mathcal{P} \leftarrow \emptyset $. \textcolor{blue}{So in this case, the query set in meta-learning will be viewed as the $\mathcal{Q}_i$ in SIM-based approach.}
\begin{algorithm}[t]
\small
\caption{SIM-based SSL for few shot learning}
\begin{algorithmic}[1]
\label{alg-ssl-fsl}
\REQUIRE Training tasks: $\mathcal{T}_{tr}=\{\mathcal{S}_i,\mathcal{Q}_i, \mathcal{U}_i\}_{i=1}^{T}$, Loss function $\mathcal{H}$ for learning model $\mathcal{M}$, Batch size of tasks: $b$, Budge of selected samples from unlabeled set: B
\STATE Randomly initialize $\boldsymbol{\theta}$
\WHILE{not done}
\STATE Sample mini-batch of tasks $\{\mathcal{S}_i,\mathcal{Q}_i, \mathcal{U}_i\}_{i=1}^{b} \sim \mathcal{T}_{tr}$
\FOR{each task $\mathcal{T}_i=\{\mathcal{S}_i,\mathcal{Q}_i, \mathcal{U}_i\},\forall i\in [1,b]$ }
\STATE Compute adapted parameters $\phi_i$ with gradient descent by Eq.~(\ref{param-adaptation}) [namely Eq.(\ref{param-adaptation-ssl})]
\STATE Using model parameters $\phi_i$, compute gradients using hypothesized labels $
\left\{\nabla_{\theta} \mathcal{H}\left(\tilde{x}_{j}, \hat{y}_{j}, \theta\right), \forall j \in \mathcal{U}\right\}
$, and obtain a similarity matrix $X$.
\STATE Instantiate a submodular function $f$ based on $X$.
\STATE $
\mathcal{A}_{i} \leftarrow \operatorname{argmax}_{\mathcal{A}_i \subseteq \mathcal{U}_i,|\mathcal{A}_i| \leq B} I_{f}(\mathcal{A}_i; \mathcal{Q}_i \mid \mathcal{P}_i)
$ (Optimize SCMI with an appropriate choice of $\mathcal{Q}_i$ and $\mathcal{P}_i$, see Table.1 in Prof.Iyer's paper)
\STATE Update model parameters from $\phi_i$ to $\phi_i'$ via Eq.(\ref{st1})
\ENDFOR
//\textit{meta update}
\STATE Update $\boldsymbol{\theta}^{(t+1)}$ by Eq.~(\ref{meta-update}) using $\{\mathcal{Q}_i\}_{i=1}^{b}$
\ENDWHILE
\STATE Return the model initialization $\theta$.
\end{algorithmic}
\end{algorithm}
\subsection{Notation}
Some notations are listed here:
\begin{itemize}
\item $\theta$: meta-parameters
\item $\phi$: task specific model parameters
\item Task $i$ in meta-training: $\mathcal{T}_i = \{\mathcal{S}_i, \mathcal{Q}_i, \mathcal{U}_i\}$
\item Task $i$ in meta-test: $\mathcal{T}'_i = \{\mathcal{S}'_i, \mathcal{Q}'_i, \mathcal{U}'_i\}$
\item $\mathcal{S}_i$, $\mathcal{Q}_i$, $\mathcal{U}_i$: Support set, Query set, and Unlabeled set for task $\mathcal{T}_i$ in meta-training
\item $\mathcal{S}'_i$, $\mathcal{Q}'_i$, $\mathcal{U}'_i$: Support set, Query set, and Unlabeled set for task $\mathcal{T}'_i$ in meta-test
\item $\mathcal{S}_{iu}$: Selected unlabeled examples from Unlabeled set for task $\mathcal{T}_i$ in the inner loop in meta-training
\item $\mathcal{Q}_{iu}$: Selected unlabeled examples from Unlabeled set for task $\mathcal{T}_i$ in the outer loop in meta-training
\item $L(\theta;\cdot)$: Cross entropy loss based on model with parameters $\theta$
\end{itemize}
For now, we consider subset selection in the inner loop based on Support set and subset selection in the outer loop based on Query set. Namely:
\begin{equation}
\mathcal{S}_{iu} = \operatornamewithlimits{arg\,max}_{A\subseteq \mathcal{U}_i, |A|\leq B}I_{f}(A; \mathcal{S}_i)
\end{equation}
\begin{equation}
\mathcal{Q}_{iu} = \operatornamewithlimits{arg\,max}_{A\subseteq \mathcal{U}_i, |A|\leq B}I_{f}(A; \mathcal{Q}_i)
\end{equation}
$f$ here is the SMI function, which is FLQMI. Details could be found from Table 2.(a) in the paper~\cite{kothawade2021similar}.
\subsection{Formulation}
Meta-parameter update according to:
\begin{equation}
\theta = \operatornamewithlimits{arg\,min}_{\theta} L_{maml+SMI} = \operatornamewithlimits{arg\,min}_{\theta} \mathcal{F}(\boldsymbol{\theta})
\end{equation}
The final loss function is:
\begin{equation}
\begin{aligned}
\label{meta-objective-ssl}
\overbrace{\boldsymbol{\theta}^*= \operatornamewithlimits{arg\,min}_{\boldsymbol{\theta} \in \boldsymbol{\Theta}}{\mathcal{F}(\boldsymbol{\theta})}}^{outer-level}\text{\hspace{1.7cm}}\\
\text{where, }\mathcal{F}(\boldsymbol{\theta}) = \frac{1}{M}\sum_{i=1}^{M}L(\overbrace{\color{blue}\phi_i}^{inner-level}, (\mathcal{Q}_{i}, \mathcal{U}_{i}))\text{\hspace{0.5cm}}
\end{aligned}
\end{equation}
$$
\phi_i = \mathcal{A}lg(\boldsymbol{\theta}, (\mathcal{S}_{i},\mathcal{U}_i)) = \operatorname{argmin}_{\theta} L\left(\theta ; S_{i}\right)+\tau L_{ssl}\left(\theta ; S_{iu}\right)
$$
\begin{equation}
\begin{aligned}
&{\mathcal{F}(\boldsymbol{\theta})} = \frac{1}{M}\sum_{i=1}^{M}L({\phi_i}, (\mathcal{Q}_{i}, \mathcal{U}_{i})) \\
&=\frac{1}{M}\sum_{i=1}^{M} L\left(\phi_{i} ; Q_{i}\right)+ \tau' L_{ssl}\left(\phi_{i} ; Q_{iu}\right) \\
&=\frac{1}{M}\sum_{i=1}^{M} L\left(\operatorname{argmin}_{\theta} L\left(\theta ; S_{i}\right)+\tau L_{ssl}\left(\theta ; S_{iu}\right) ; Q_{i}\right)+ \tau' L_{ssl}\left(\operatorname{argmin}_{\theta} L\left(\theta ; S_{i}\right)+\tau L_{ssl}\left(\theta ; S_{iu}\right) ; Q_{iu}\right)
\end{aligned}
\end{equation}
in which:
\begin{equation}
\mathcal{S}_{iu} = \operatornamewithlimits{arg\,max}_{A\subseteq \mathcal{U}_i, |A|\leq B}I_{f}(A; \mathcal{S}_i)
\end{equation}
\begin{equation}
\mathcal{Q}_{iu} = \operatornamewithlimits{arg\,max}_{A\subseteq \mathcal{U}_i, |A|\leq B}I_{f}(A; \mathcal{Q}_i)
\end{equation}
\subsection{Inner loop}
Task specific model parameters for task $\mathcal{T}_i$:
\begin{equation}
\begin{aligned}
\phi_{i}=& \operatorname{argmin}_{\theta} L\left(\theta ; S_{i}\right)+\tau L\left(\theta ; S_{iu}\right) \\
=&\theta-\alpha \nabla_{\theta} L\left(\theta ; S_{i}\right)-\alpha \tau \nabla_{\theta} L\left(\theta ; S_{iu}\right) \text{(one step gradient update example)}
\end{aligned}
\end{equation}
In which, $\alpha$ is learning rate. $\tau$ is the coefficient from the pseudo labeling approach. Since it is an increasing temperature variable, let $\tau^{(j)}$ denote the $\tau$ in step $j$.
Since there are several gradient update steps in inner loop. Take three steps for example.
Let $\phi_{i}^{(j)}$ denote the model parameters for $j$-th step for task $T_i$.
\begin{itemize}
\item initialization: $\phi_{i}^{(0)} = \theta$
\item inner step 1:
$$
\phi_i^{(1)}=\theta-\alpha \nabla_{\theta} L(\theta ; S_i)-\alpha \tau^{(1)} \nabla_{\theta} L\left(\theta ; S_{iu}\right)
$$
$\mathcal{S}_{iu} = \operatornamewithlimits{arg\,min}_{A\subseteq \mathcal{U}_i}I_{f}(A; \mathcal{S}_i)$. In this step, the CNN model used to calculate the gradient in SMI is parameterized by {\color{blue}$\theta$}.
\item inner step 2:
$$
\phi_i^{(2)}=\theta-\alpha \nabla_{\phi^{(1)}} L(\phi^{(1)} ; S_i)-\alpha \tau^{(2)} \nabla_{\phi^{(1)}} L\left(\phi^{(1)} ; S_{iu}\right)
$$
$\mathcal{S}_{iu} = \operatornamewithlimits{arg\,min}_{A\subseteq \mathcal{U}_i}I_{f}(A; \mathcal{S}_i)$. In this step, the CNN model used to calculate the gradient in SMI is parameterized by {\color{blue}$\phi^{(1)}$}.
\item inner step 3:
$$
\phi_i^{(3)}=\theta-\alpha \nabla_{\phi^{(2)}} L(\phi^{(2)} ; S_i)-\alpha \tau^{(3)} \nabla_{\phi^{(2)}} L\left(\phi^{(2)}; S_{iu}\right)
$$
$\mathcal{S}_{iu} = \operatornamewithlimits{arg\,min}_{A\subseteq \mathcal{U}_i}I_{f}(A; \mathcal{S}_i)$. In this step, the CNN model used to calculate the gradient in SMI is parameterized by {\color{blue}$\phi^{(2)}$}.
\item inner loop end.
Model parameters for task $T_i$ udpate process:
$$\phi^{(0)} (:=\theta) \rightarrow \phi^{(1)}\rightarrow \phi^{(2)}\rightarrow \phi^{(3)}$$
\end{itemize}
\subsection{Outer Loop}
Meta-parameter update according to:
\begin{equation}
\theta = \operatornamewithlimits{arg\,min}_{\theta} L_{maml+SMI}
\end{equation}
The final loss function is:
\begin{equation}
\begin{aligned}
&L_{maml+SMI}=\frac{1}{M}\sum_{i=1}^{M} L\left(\phi_{i} ; Q_{i}\right)+ \tau' L\left(\phi_{i} ; Q_{iu}\right) \\
&=\frac{1}{M}\sum_{i=1}^{M} L\left(\operatorname{argmin}_{\theta} L\left(\theta ; S_{i}\right)+\tau L\left(\theta ; S_{iu}\right) ; Q_{i}\right)+ \tau' L\left(\operatorname{argmin}_{\theta} L\left(\theta ; S_{i}\right)+\tau L\left(\theta ; S_{iu}\right) ; Q_{iu}\right)
\end{aligned}
\end{equation}
in which, $T$ is the set of meta-training tasks. $\tau'$ is still a coefficient borrowed from the pseudo label formulation.
The second equal in the above equation is according to the inner loop update:
$$
\phi_i = \operatorname{argmin}_{\theta} L\left(\theta ; S_{i}\right)+\tau L\left(\theta ; S_{iu}\right)
$$
Since there is only one step in the outer loop, subset selection only happens one time.
\begin{equation}
\mathcal{Q}_{iu} = \operatornamewithlimits{arg\,min}_{A\subseteq \mathcal{U}_i}I_{f}(A; \mathcal{Q}_i)
\end{equation}
The CNN model used here should be the final model parameter after the inner loop: {\color{blue}$\phi_i^{(3)}$}.
The motivation here to use $\phi_i^{(3)}$ instead of meta-parameter $\theta$ is that SMI needs a CNN model which has powerful representation. For task $T_i$, $\phi_i^{(3)}$ is more powerful than $\theta$.
|
\section*{TBD}
Compiled on \today\ at \currenttime
\section{Introduction}
\label{sect:intro}
Low frequency radio astronomy is under-going a rebirth, with the arrival of the next-generation instruments that will provide two orders of magnitude improvement in sensitivity and an order of magnitude improvement in frequency coverage. In particular the Low frequency Square Kilometre Array (SKA-Low) will provide an instantaneous frequency coverage of 50 to 350 MHz, a collecting area of 0.4 sq. kilometres, spread over 65km baselines in the Phase-1 and the full 1.0 sq. kilometres in Phase-2, and exciting prospects of new science.
Thus it is vital that we update the radio-astronomical methods and strategies to match the potential of the radio-astronomical instruments.
In this paper we focus on the challenges at the foundation of accurate direction-dependent ionospheric calibration, as these are
among the most difficult obstacles to achieving the nominal SKA-Low performance.
Directional dependent (DD) calibration is vital at low frequencies, because the field of view (FoV) is intrinsically large and the
propagation effects caused by inhomogeneties in the distribution of the ionospheric plasma ($\Delta{\rm I}$, measured in TECU or 10$^{16}$ electrons per m$^2$) imposes temporal and spatial phase disturbances on the incoming wavefront whose magnitude scales inversely proportional to the observing frequency ($\phi_{\rm ion}({\rm deg})=480\,\Delta{\rm I}\, \nu_{\rm GHz}^{-1}$, with $\nu_{\rm GHz}$ the observing frequency in GHz \cite{sovers_98}).
Traditionally, with observations of smaller fields of view and at higher frequencies, these effects could largely be handled with a single station-based correction (or direction independent (DI) correction), valid for all of the field of view.
Lonsdale (2005, Figure 1)\cite{lonsdale_05} classified four operating regimes based on the relative sizes of the FoV (V), the ionospheric phase fluctuation disturbances (S) and the projected size of the array (A).
Traditionally, with observations of smaller fields of view and at higher frequencies (V$\ll$S), these effects could largely be handled with a single correction per station valid within the FoV. This applies for both compact (A$\ll$S) and extended (A$\gg$S) arrays, which are cases 1 and 2, respectively, for which conventional self-calibration techniques are suitable.
For modern low-frequency arrays this is no longer sufficient, as the FoV is greater than the scale size of the disturbances (V$\ge$S).
Case 3 is for sufficiently compact arrays such that the DD wavefront disturbances over the array can be treated as planar (i.e. linear approx.) and yield coherent apparent source position shifts varying in time and direction. When the linear approximation is no longer valid (i.e. extended array) the presence of non-planar (i.e. curvature) disturbances additionally lead to source shape deformations; this is case 4, which requires a more complex calibration algorithm, where an independent solution for each direction and station must be applied.
We show here that this is an important issue even for the pathfinders such as Murchison Widefield Phase-2 Array (MWA-2), which has baselines less than 5 kilometres long.
A prime challenge with DD corrections is in the imaging of the data. The Van-Citter relationship that is the basis of all radio-interferometric imaging requires that the data in the Fourier domain has a uniform calibration applied before transforming it to the image domain.
Several research groups are developing solutions for this issue, for example DDFacet\cite{ddf_tasse}, which takes a facetting approach, and is widely used for LOFAR LBA imaging and WSClean\cite{offringa_14} with Image Domain Gridding\cite{tol_18}, which uses gridding in the image domain to correct for direction-dependent effects, and is widely used for MWA imaging.
We have been looking at efficient methods to generate accurate DD-calibrations, for application in these next generation imaging tools.
We have been developing a method called LEAP (Low-frequency Excision of Atmosphere in Parallel), which is discriminated from other methods by the fact that every direction can be treated as independent, and thus analysed in parallel. This compares favourably to sequential (`peeling') methods (e.g. SPAM\cite{intema_09} or RTS\cite{mitch_cal}) that correct for each source in order of strength, and simultaneous methods (e.g. SAGECal\cite{sagecal}) that solve for all directions with a very large solution matrix, which do not scale well to large numbers of sources or stations.
LEAP of course has its limitations; the fact that it
does not perform source subtraction can elevate the noise floor, and the flux density cutoff for calibrator sources.
Nevertheless, as it does not require a complete sky model (other than a catalogue of LEAP calibrator positions, which are each treated independently) and is very
simple and robust, it can play an important role in providing a preliminary solution to the more sophisticated and complicated methods, and a unique role
in providing the real-time direction-dependent calibration for the SKA tied-array beamforming. The latter is used to form multiple simultaneous beams within
the field of view, as required in VLBI\cite{skavlbi_wp10} and equally in pulsar studies.
Traditional MWA DD-calibration strategies, as used by GLEAM\cite{nhw_gleam} for example,
assume that observations fall in ``Case 3'', and ignore the potential defocusing effect arising from non-linear disturbances in the wavefront above the array. The ionospheric calibration is carried out in the image domain, as a rubber sheet-type correction to the individual 2-min long scan images before mosaicing.
This approach worked perfectly well for GLEAM (with a maximum MWA Phase-1 baseline of 2\,km), and has had some success with GLEAM-X (with a maximum MWA Phase-2 baseline of 5\,km).
However, as we will show, for the increased sensitivity
of SKA-Low this is going to be completely inadequate, even for this range of short baseline lengths.
The ionospheric calibration data itself provides a means to gain information on the physical nature, and spatial and temporal structure scales of the propagation media
\cite{mevius_16}.
Previous MWA large-scale studies have been performed using image-based analysis \cite{loi_15,jordon_17,helmboldt_20}, however these cannot probe the fine scale.
In this paper we use the LEAP station-based calibration data, in the visibility domain, to quantify the presence of higher order spatial and temporal phase structure over the MWA Phase-2 array imprinted in the incoming wavefront as it propagates through the ionospheric plasma irregularities, for directions within the FoV.
As LEAP measures the phase variations within the baselines formed by the array, we are able to study the small scale spatial and temporal structures; that is at scales from 30m to 5km. Our particular interest is to discover at what level the planar distortions approximations breakdown, entering the calibration regime 4, as that is crucial information for the planning of the SKA-Low calibration schemes.
\subsection{A LEAP primer}
We quickly recap the discussions in Rioja et.\,al., 2018 \cite{rioja_18} (hereafter LEAP-I)
to summarise the operations of LEAP.
The core insight of the LEAP algorithm is that the extremely wide fractional frequency bandwidth of current and further low-frequency arrays allows the use of frequency smearing to suppress all sources away from the phase centre.
LEAP processing only requires as input the directions for calibrator sources that are sufficiently strong to dominate the phase observable
after the other directions are suppressed by frequency averaging. For this we have used the GLEAM-I catalogue\cite{nhw_gleam}, so the positions are accurate for the frequencies being used. This simple approach is remarkably effective, as the contributions from other directions are suppressed by a factor $\propto {{\Delta\nu}\over{\nu}} \theta_0 $, where $\Delta\nu$ is the bandwidth and $\nu$ is the central frequency, and $\theta_0$ is the source distance from the centre in beamwidth units.
The fractional bandwidth is very high for modern low-frequency instruments. SKA-Low, for example, has a frequency span ratio close to 1; MWA has one as high as 1:3 and MWA Phase-3 (MWA-3) will be even higher.
We require the input data to be, at a minimum, bandpass corrected so that the frequency channels can be averaged. For example,
if using {\it a priori} conventional direction independent (DI) self-calibration (i.e. a flux-weighted average solution across the FoV), LEAP will correct
for the DD ionospheric residuals with respect to the flux-weighted DI calibration direction; these DD residuals being small we can then determine the DD dispersive delay unambiguously from the phase term only.
An example of the phase screens ({\it hereafter ionospheric screen}) over the array observed with the MWA-2, for a given direction, is presented in Figure \ref{fig:example_phase}.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{./example_phase_surf.pdf}
\caption
An example of a LEAP ionospheric phase screen at 88MHz from a 2-min scan considered in this paper; the span in the sky is 30$^\prime$-60$^\prime$, depending on the height of the ionosphere.
Shown as a mesh are the station-based LEAP calibration phases above the MWA Phase-2 extended configuration array for the direction of one LEAP calibrator along a line of sight with an elevation of 65$^o$, color-coded in degrees, with the X and Y axis in meters for the 128 stations (at the nodes of the mesh).
The wireframe shows the second order fit to the data, underlining the high curvature of the surface.
\label{fig:example_phase}}
\end{figure}
The interpolation and smoothing of these solutions over the array, and over different pointings, provides an input into, for example, WSClean for direction dependent calibration and imaging. An example is shown in Figure \ref{fig:WSC_corrections}, where the interpolated TEC surface for all directions on the sky is projected onto a regular grid, for each station.
\begin{figure}
\centering
\includegraphics[width=0.85\textwidth]{./fig2-TEC-screen.pdf}
\caption
Example of the ionospheric phase screens as would be provided to WSClean for DD wide-FoV imaging, from a 2min scan at 88\,MHz in our dataset. Each plot represents the ionospheric phase solutions for a given station across the wide MWA FoV, for four stations in the outer region of the extended configuration (i.e. long baselines).
The TEC values have been interpolated from the LEAP station-based phase solutions towards the $\sim$200 calibrator directions, which are shown with white diamonds. The location of calibrator shown in Figure \ref{fig:example_phase} is marked with a black square (labelled Fig. 1) and the location of the few fast changing sources (labelled Fast) are marked with a star.
\label{fig:WSC_corrections}}
\end{figure}
\section{Observations and Analysis}\label{sec:obs_ana}
We used two datasets of GLEAM\cite{nhw_gleam,gleam-x} observations with the MWA Phase-2 long baseline configuration\cite{mwa2} at two of the lowest frequency bands, with an instantaneous bandwidth of 30 MHz for the explorations presented for this paper.
One dataset consisted of a series of interleaving 2-min long scans, centred at 88MHz and 154MHz, on December
18th, 2017, under moderate-weather conditions.
The interval between observations at the same frequency was 10 minutes, spanning an hour and a half.
The other dataset includes a small number of 2-min long scans carried out on June 2018, with the strong source
3C444 (J221425-170140) near the phase centre, under a variety of weather conditions.
That is, good weather on June 3$^{\rm rd}$ and poor weather on June 12$^{\rm th}$\cite{nhw_pc}.
The latter was flagged to be rejected and reobserved based on severe coherent artefacts (residual sidelobes from several strong sources caused by ionospheric distortions)
that contaminated the whole of the image, following the standard GLEAM analysis.
All the observations were calibrated using LEAP, following the calibration and imaging procedure described in Section 1.1, as part of a wider-scope project to advance the effective imaging implementation (e.g. a single gridding and imaging inversion) of LEAP DD calibration
along with a comparative study of `refocusing' calibration in wide FoV imaging using MWA
observations (Dodson et\ al., 2021, in prep).
However the aim of this paper is to provide answers to some of the questions about the local spatial and temporal scales of the ionospheric disturbances pertinent for the SKA, and identify the requirements for an optimum strategy for the removal of systematic errors
depending on the observing frequency and under different ionospheric conditions.
Of special interest for this work are the LEAP calibration by-products, namely the measured station-based phase solutions which define the so-called ionospheric phase screens above the array, along the directions of the stronger LEAP calibrators
(i.e. those with reduced thermal noise error contribution). They are a measure of the wavefront disturbances introduced by its propagation through the ionosphere.
Hence, each screen conveys information of the spatial (and temporal) structure of the DD ionospheric disturbances, at scales smaller than the array size
(i.e. less than 5km and potentially down to 30m). They are the main observable for the ionospheric analysis in the visibility domain followed here.
The station-based phase observables are obtained with the script discussed in LEAP-I, where {\sc casa} was used to: call WSClean for phase rotation to calibrator source directions and imaging for analysis and validation, and to solve for gain solutions.
As part of ICRAR bridging contributions to the SKA science processing a GPU version has been written\footnote{https://jira.skatelescope.org/browse/SP-994} that can efficiently process multiple directions in parallel. This addresses the issue of parallel reading of the data, however the processing here predated the release of that tool.
We have followed two methods of analysis of the individual ionospheric phase screens.
Firstly, a regression analysis fitting, using linear and second order surfaces.
As a result we measured the planar `slope' and orientation from the 1$^{\rm st}$ order 2D linear fits, along with the `curvature'; obtained as
the quadratic mean of the three 2$^{\rm nd}$ order fits coefficients.
Also we measured the Root Mean Squared residual errors (RMSE) to the first and second-order fit (${\rm RMSE}_1$ and ${\rm RMSE}_2$ respectively), and computed the RMSE fractional improvement, defined, for a given screen, as:
\[
\Delta {\rm RMSE}=\frac{{\rm RMSE}_1 - {\rm RMSE}_2}{ {\rm RMSE}_1}
\]
Secondly, we compute the 2nd-moment structure function ({\it hereafter} Full Structure Function or SF) of the ionospheric screens ($\phi_{ion}$) over the range of (projected) baselines in the MWA array, $r$, following the approach in, for example, Van de Tol, 2009\cite{vantol_phd}:
\[
D(r)=\langle (\phi(r^\prime)-\phi(r^\prime+r))^2\rangle
\]
were $\langle ....\rangle$ stands for a statistical average.
The SF analysis is traditionally used to measure the scaling features of the electron density fluctuations in the ionosphere in order to understand the nature of the physical processes (e.g. turbulences) at the origin of the measured signal.
Here we use the ability of the SF as a tool to characterize complex surfaces, particularly with non-uniformly sampled and non-periodic data, using the measured phase values that form the ionospheric screens.
\cite{mevius_16,nature_sf_swarm,sf_toys_neuroscience}. Of particular interest here is the capacity to detect deviations from linearity.
We calculated the SFs using the baseline phase values, deduced from the station solutions.
In addition we formed the Detrended Structure Function (DSF) from the residual phase values, after subtracting the corresponding planar fits from the station solutions.
This eliminates the overall dominant influence of a planar slope in the data set.
We expected D(r), in a noise-free environment, to follow:
\[
D(r) = C^2 r^\beta
\]
where $C$ would be the slope of the ionospheric phase screen (i.e. change of phase per unit baseline distance), and
where $\beta$ is the so-called scaling exponent, which allows us to characterize the scaling nature of the signal under investigation (in our case, the ionospheric disturbances); this value is equal to 5/3 for pure thin-screen Kolmogorov turbulence, or 2 for perfectly planar surfaces. This behaviour holds over the inertial region, which is the approximately linear regime in log-log space between the outer scale at which power is injected and the inner scale at which power is dissipated in the Kolmogorov theory of turbulence.
We also generated spatial structure functions using simulated datasets of synthetic phase surfaces base on our MWA observations, i.e ionospheric toy-models with known properties, to test the applicability of SF analysis to the characterization of ionospheric phase screens
and to help with interpretation of our empirical findings in terms of ionospheric behaviour.
The toy models consisted of a family of surfaces defined by 2D linear and quadratic polynomial functions with and without added noise,
for a range of values for the polynomial coefficients and noise parameters, compatible with our findings from the regression analysis.
Dodson et\ al.{} (2021, in prep.) will present a complementary analysis of the ionospheric disturbances based on the comparison of pre- and post-calibration images.
The latter has the capacity to additionally correct for defocusing artifacts in the image, such as changes in the source shape and residual sidelobe patterns, which arise from the small scale ionospheric disturbances.
\section{Results}
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{./time_var_1212870680.pdf}
\caption{
A highly time variable phase screen, with 30s sampling, observed on a nominally good weather day (2018/06/12) on 3C444 at 154MHz. A compact ($<$1km) knot of plasma can be seen passing through the line of sight of the array.
About 1\% of the sources with SNR$>$6 have high temporal variability.
\label{fig:example_phase_temporal}}
\end{figure}
Figure \ref{fig:example_phase} showed an example of LEAP ionospheric phase screens above the MWA array in the direction of a strong LEAP calibrator, for a 2-min scan solution.
For weaker sources, it is expected that the ionospheric signature will be diluted, and ultimately buried under the increasing thermal noise. The results presented here comprise the analysis of the ionospheric screens in the directions of all ($\sim$200) LEAP calibrators (with SNR$>$2, per station) within the MWA FoV in each 2-min scan, for the set of scans and the analysis described in Section \ref{sec:obs_ana}.
{\it Regression Analysis}
Figures \ref{fig:Corrections}a) and b) show the outcomes from the linear 2D regression analysis of the unwrapped station-based phase solutions in an ionospheric screen, for 8 consecutive 2-min scans, at 88 and 154 MHz, respectively.
Unwrapping, that is taking into account full wraps of phase by enforcing continuity, is essential to accurately fit the phase slopes on the longer baselines.
Shown are the magnitude of the `slope' (circle size, mTECU/km) and the orientation angle (circle colour, degrees) of the fitted plane.
The sky location of the symbols correspond to that of the LEAP calibrators across the FoV, for each scan.
The FoV at 154\,MHz is smaller than for 88\,MHz, and each subplot is taken at a different moment in time (although all those shown have a common phase centre). Thus the station gains vary between each subplot and the selected sources vary between scans.
Figure \ref{fig:Curved_corrections} corresponds to the second order 2D regression analysis,
showing the magnitude of the fitted curvature (circle colour, mTECU/km$^{2}$) and slope (circle size with same scale as Figure \ref{fig:Corrections}, in mTECU/km), at 88\,MHz, with SNR$>$6.
Because higher SNR is required to measure the curvature, there are significantly fewer data points. Furthermore, we do not show the corresponding figure for 154\,MHz; there were too few data points per scan to show the spatial correlation.
The median value for the ionospheric slope is 5$\pm$3\,mTECU/km, across the $\sim$5km span of the array, for all LEAP calibrator directions, at both 88 and 154 MHz.
The median value for the curvature is 3$\pm$2\,mTECU/km$^2$, at both frequencies for SNR$>$6.
For each screen, the linear and second order RMSE errors from the regression analysis are combined to form the fractional $\Delta$RMSE quantity, that would show
a significant improvement of one fit above the other, as expected if deviations from linearity are significant.
That is, cases when the curvature signature is large enough so that its magnitude is comparable to or greater than the thermal noise contribution and results in a large fractional reduction of the RMSE.
Figure \ref{fig:Hist} shows a histogram of the fractional $\Delta$RMSE values corresponding to the screens shown in Figure \ref{fig:Corrections}, binned by SNR. This shows
that the ionospheric curvature signature, if present, appears more significant towards the directions of strong LEAP calibrators, where the thermal noise is reduced.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{./PhaseSurf_2D_80_hsv.pdf}
\includegraphics[width=0.7\textwidth]{./PhaseSurf_2D_150_hsv.pdf}
\caption
Outcomes of 2D linear regression analysis of LEAP ionospheric screens: fitted values of slope (circle size, see last frame) and orientation angle (East through North, cyclic colour scale}) across the array,
for the RA and DEC directions of LEAP calibrators with SNR$>$2 per station. {\it Upper} panel for 88 MHz, and {lower} panel for 154 MHz. Each subplot is for a 2-min scan at the time indicated in the title. The results for any snapshot at either frequency are spatially correlated, even though the ionospheric screens in each direction are measured independently.
\label{fig:Corrections}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{./PhaseSurf_altcurve_2D_80.pdf}
\caption
Outcomes of 2D second-order regression analysis of LEAP ionospheric screens: fitted values of slope (circle size, see last frame and identical to Figure \ref{fig:Corrections}) and curvature (colour) across the array,
for the RA and DEC directions of LEAP calibrators with SNR$>$6 per station, at 88 MHz. Each subplot is for a 2-min scan at the time indicated in the title. The curvature shows significantly less spatial correlation than the slope.}
\label{fig:Curved_corrections}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{./frac_of_rms_0-10_snr.pdf}
\caption
The fraction of LEAP ionospheric screens showing significant curvature increases with the SNR per station (derived from the RMSE of the 2D fit residuals). The metric for significant curvature is that the fractional difference of the RMSE values from the linear and second order 2D fits
is greater than the expectation values from the low signal to noise cases.
Here we show the normalised counts of directions with ionospheric screen curvatures, for SNR ranges indicated with color codes, at 88 MHz. We conclude that, for SNR$>$6 (or RMSE$<10^o$) most screens are poorly described by a planar surface, even with the small size of MWA-2 array.
Therefore, the incidence of curvature is expected to grow larger, in more directions, with increased sensitivity, such as MWA-3, and much more for SKA-Low, which includes much longer baselines, in addition to a giant leap in sensitivity.
}
\label{fig:Hist}
\end{figure}
Some limited insights in the temporal behaviour also come from the regression analysis.
We searched the solutions from all the stronger sources for indications of significant variation on shorter time spans than the full scan length of 2min.
Three of the more than 300 source directions with a SNR$>$6 per station, for all scans, at each frequency, showed significant temporal variability.
Their location is marked on Figure \ref{fig:WSC_corrections}.
Figure \ref{fig:example_phase_temporal} shows a similar plot to Figure \ref{fig:example_phase} for the example with the best SNR, with solutions every 30s. Clear changes in their spatial and temporal structure are easily seen, even by visual inspection.
{\it Structure Functions:}
Structure function analysis is the traditional method to infer the nature of a propagation media.
We have formed the spatial structure functions of the ionospheric screens corresponding to the strongest sources in our data across the limited MWA baseline range. For the ``full'' SF we use the station-based phase values measured with LEAP in the direction of the source, and the values after subtracting the linear plane from the linear regression analysis, for the ``detrended'' SF, respectively.
The clearest example, for the direction towards the strongest source 3C444 with a point source flux of 44\,Jy at 154MHz, is shown in
Figure \ref{fig:3C444}, for the full (solid lines) and detrended (dotted lines) SF analysis, for observations on two days with ``good'' (red for 154MHz and green for 88MHz) and ``poor'' (blue for 154MHz) weather conditions.
In this case the SFs increase with baseline length with gradients of
$1.8^{+0.2}_{-0.2},1.7^{+0.2}_{-0.2}, 1.82^{+0.04}_{-0.04}$, respectively. The noise floor is not reached until very short baselines, below a few hundred meters.
The detrended SF analysis shows a rise of the residual signal over the noise at baselines greater than 0.6km, which is more clear for the ``poor'' weather observations and lower frequencies.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{./3c44_struct_func_detrend_scaled_phi_three.pdf}
\caption{Full (solid lines) and detrended (dotted lines) spatial structure function D(r) of the LEAP ionospheric screens measured in the direction of the
strong source 3C444 (J221425-170140), from MWA-2 observations carried out on June 2018 under good (red crosses) and poor (blue circles) weather conditions at 154MHz, and under good weather conditions (green pluses) at 88MHz.
The full structure functions follow a near-constant gradient across the baseline range, here shown in units of radians$^2$ for comparison to Mevius et\ al.{} 2016.
The detrended structure functions have been calculated after removal of the 2D linear fit from the ionospheric screens.
In this case, the transition from noise dominated domain to baseline length dependent domain occurs at about 0.6km, which would be our recommended scale for the introduction of higher order DD-calibrations. To scale the y-axis to units of mTECU$^2$, for comparison to Figure \ref{fig:SF}, multiply the 88/154 MHz data by 110/338, respectively.}
\label{fig:3C444}
\end{figure}
Figure \ref{fig:toy} shows the behaviour of the spatial structure functions ({\it left} full; {\it right} detrended) for simulated 2D ionospheric screens (i.e. ``toy models'')
defined by ionospheric slope and curvature values comparable with the measurements from our regression analysis in the 2017 datasets
(i.e. a slope of 5.1mTECU/km and curvature of 6.8mTECU/km$^2$), to
four times greater, for moderate (blue dots) and poor (yellow cross) weather, respectively; for comparison, the case for a linear screen (i.e. no curvature, and a slope double of the minimum) is shown with green squares.
In addition, we tested the effect of measurement noise in the SF by adding a random noise signal with an RMSE of up to 10$^o${} per
station (i.e. $\sim$2mTECU at 88MHz) to these toy models
(red circles and purple crosses for moderate and poor weather, respectively).
Figure \ref{fig:toy} ({\it lower}) shows the behaviour of the gradient of the structure functions, across the limited range of
MWA baseline lengths, as a function of the added noise signals to the toy models.
\begin{figure}\hspace{-1cm}
\centering
\includegraphics[width=0.95\textwidth]{fig8.pdf}
\caption{Spatial structure functions calculated for simulated toy ionospheric models at 88MHz, based on the ionospheric screens measured with MWA-2.
{\it Upper Left} shows the traditional full structure function (SF), for a 2D-polynomial model with parameters for curvature and slope typical for our moderate-weather observations (circles), noise-free and with added-noise equivalent to a per station SNR of 20. Plotted with crosses are the SF for poor-weather with parameters approximately four times greater, with and without noise. Finally, with dotted lines and squares are the SF calculated for a planar model, with and without noise.
For a given weather condition, the effect of added noise limit the SF measurements at short baselines and shifts the SF upwards, compared to noise-free case;
however, the values and gradients of the SFs are preserved at longer baselines.
{\it Upper Right} shows the detrended SF, calculated after a 2D linear model has been subtracted from the data describe above, showing the signature of the residual non-linear ionospheric structures.
We consider any excess signal above the noise-floor region at the shortest baselines, as the indicator for the presence of non-linear screen structure.
The turn-over (or potentially turn up with even higher order terms) towards the longest baselines indicate the end of the inertial region.
{\it Lower} shows the gradients measured from the SF plots verses added noise, for the full SF ({\it blue circles and cyan squares} fitted over the range of 0.5 to 5km and for the detrended SF {\it red crosses} fitted over 0.2 to 2km baseline lengths. The blue line is for a second order screen model and the cyan line is for a planar model. For the latter the noise-free gradient should be exactly 2.
This shows how the gradient is a sensitive function of the noise-level, particularly with the narrow range of MWA-2 baselines.
\label{fig:toy}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\textwidth]{fig9.pdf}
\caption
{\it Upper plots} show the empirical spatial structure functions calculated from the MWA-2 LEAP ionospheric screens in the directions of sources with SNR $>$20 (per station) versus projected baseline bins at 88 ({\it left}) and 154 MHz ({\it right}). The intermittent lines are for measurements of an individual screen, colour coded to identify the time of the 2-min scan and source direction. The bold thick black line comprises the measurements from the ensemble of all screens, that is multiple scans and directions, to calculate the structure function.
The bold coloured lines, labelled with 3C444, are repeats of the data from Figure \ref{fig:3C444}.
The structure functions are shown in units of mTECU$^2$ to allow for direct comparison between 88 and 154 MHz.
For the case of lowest measurement noise individual screens (towards 3C444, with SNR$>$50, shown with the same bold colours as in Fig. \ref{fig:3C444}) the SF extends in a quasi linear fashion to baselines of $\sim$100m.
b) A histogram of the SF gradients in a) measured between 0.5 and 5\,km, plotted in blue solid for 88MHz and red outlines for 154MHz. The median gradients is 1.72$\pm$0.15, which is consistent within errors to that of Mevius et\ al.{} (2016\cite{mevius_16}) and the expectations for a 2D Kolmogorov spectrum (1.89$\pm$0.1 and 1.67, respectively).
c) A histogram of the empirical MWA-2 LEAP ionospheric screen slopes $C$ for all the LEAP calibrators from the regression analysis, in mTECU/km, at
88MHz and 154MHz. The median slope is 5$\pm$3mTECU/km at both frequencies.
%
\label{fig:SF}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\textwidth]{./all_struct_func_phase_detrend_4.pdf}
\caption
\label{fig:SF-detrend}
Empirical detrended spatial structure functions calculated from the MWA-2 LEAP ionospheric screens in the directions of sources with SNR $>$20 (per station) after subtracting a planar fit versus projected baseline bins at 88 ({\it left}) and 154 MHz ({\it right}).
Any excess signal above the noise unambiguously exposes the presence of the `non-linear' components in the ionospheric screens.
The DSF for individual screens in the direction of 3C444 are shown in bold (green, blue, red) matching the colours in Fig. \ref{fig:3C444}, and for the ensemble over time and directions, shown in bold and black.
The bold coloured lines, labelled with 3C444, are repeats of the data from Figure \ref{fig:3C444}.
The ensemble DSF at 154MHz does not show significant non-linear terms greater than the noise, although individual datasets do show such behaviour. On the other hand, both the ensemble and individual detrended structure functions at 88MHz indicate the presence of significant non-linear structure at small scales, i.e. baseline lengths greater than 0.6km.}
\end{figure}
Figure \ref{fig:SF} shows a compilation of the full structure functions of the ionospheric screens towards strong (SNR$>$20) sources, for each individual 2-min measurement set and source direction
({\it dashed lines}).
Also using multiple scans and directions to form the combined SF (black solid line, in bold), at 88 MHz ({\it upper left}) and 154 MHz ({\it upper right}).
A histogram of the structure function gradients over the range of 0.5 to 5 km baselines is shown ({\it lower left}),
along with the histogram of the ionospheric slopes (measured on the input ionospheric screens with the linear regression analysis ({\it lower right}).
Figure \ref{fig:SF-detrend} shows the detrended structure functions for the same datasets as Figure \ref{fig:SF},
at 88 MHz ({\it left}) and 154 MHz ({\it right});
here, after the contribution from a nominal planar surface is subtracted out, the detrended SF is dominated by the presence of non-linear structures in the ionospheric screens.
\section{Discussions and Conclusions}
\subsection{Regression Analysis}
We have used the LEAP measurements of the ionospheric phase surfaces over the MWA-2 array to make a
detailed analysis of the spatial structure of the ionospheric distortions imposed on the incoming wavefronts towards $\sim$200 simultaneous directions across the FoV, for 20 datasets spanning a range of different weathers, at 154 and 88 MHz.
LEAP, being an station-based phase calibration, is sensitive to higher order ionospheric phase structural changes at small scales, less than
the size of the array.
Alternative calibration strategies that use image-domain apparent source position shifts to measure the ionospheric phases only measure a simple gradient\cite{loi_15,jordon_17,helmboldt_20}.
However, LEAP analysis requires observations of stronger sources, as we now are determining antenna-based corrections\cite{rioja_18}, rather than corrections from the array-averaged data in the image.
Nevertheless, ignoring higher order effects results in
severe artefacts in the images and introduces an intractable bias in studies such as EoR\cite{vedantham_15,trott_18}.
Here we discuss the results of two approaches to measure spatial small scale phase deviations
which are undetectable to analysis based on the image-domain position shifts.
We are confident that the LEAP calibration method is providing measurements of only the ionospheric surface. Standard practise in MWA analysis is to assume that the beam response is the same for all stations. If there were deviations from this assumption random distortions would appear in the phase surfaces. Such behaviour would be expected to occur mainly on the short baselines, where the stations can interfere with each other. Given that we don’t see such behaviour, i.e. the surfaces are smooth, we can take that the assumption is valid, particularly for this extended baseline configuration.
For the regression analysis
we fitted 1$^{\rm st}$ and 2$^{\rm nd}$ order polynomial to the surfaces to estimate the ionospheric screen parameters: slope, direction and curvature.
These display large scale coherent behaviour in the slopes (Figure \ref{fig:Corrections}), over the FoV, as previously reported \cite{rioja_18,jordon_17,helmboldt_20}.
The RMSE values for the 2$^{\rm nd}$ order fit correlates well with the inverse of the LEAP calibrator flux densities and the expected thermal SNR; that is, the stronger sources have smaller fitted parameter errors,
as expected from high signal to noise ratio measurements.
The coherent behaviour in the curvatures, derived in the 2$^{\rm nd}$ order fit, over the sky is less clear because of the sparsity of strong calibrator signals, but does indicate spatial coherence (Figure \ref{fig:Curved_corrections}) at the lower frequency.
The fractional RMSE change between the 1$^{\rm st}$ and 2$^{\rm nd}$ order fits provides a useful metric for the significant detection of curvature in the presence of measurement noise.
When considering all observations, the RMSE values from the first and second order fits are not
significantly different for the majority of the cases. This is to be expected, as the small collecting area
of the MWA stations and the low sensitivity result in a thermal noise contribution larger than the ionospheric phase signature
for the majority of LEAP calibrator sources.
Figure \ref{fig:Hist} shows that, when introducing LEAP calibrator flux density cutoffs, the impact of the curvature is increasingly visible
for higher SNRs
with 20\% and 40\% of cases showing significant ($>\sim 15$\%) fractional RMSE improvements,
for 6$<$SNR$<$12 and SNR$>$12, respectively.
The presence of curvature in the ionospheric screens over the array, if uncorrected, result in increased residual sidelobes after deconvolution, particularly for stronger sources.
This underlines the importance of station-based calibration to reduce the residual sidelobes in the cleaned images.
This will be fully discussed in Dodson et\ al.{} (2021 in prep.), where we show the
improvement in the recovered source peak fluxes in the images.
We have limited information for the discussion of high temporal variability of the phase screens, other than noting that three out of the about three hundred strong (SNR$>$6) sources in all the datasets showed significant variation at short timescales. One example is plotted in Figure \ref{fig:example_phase_temporal}. Because of the few lines of sight towards strong sources it is impossible to track the nature of this behaviour, which will affect the more sensitive SKA observations.
Significant small scale deviations from linearity in the spatial ionospheric phase distributions are detected from the regression analysis,
even with the limited sensitivity and size of MWA Phase-2.
This indicates that MWA Phase-2 observations reside in the Lonsdale regime 4.
Under these conditions, the performance of image-based apparent source position shift ionospheric calibration degrades,
thus there will be benefits from using higher order calibration with MWA Phase-2.
These effects are expected to become more significant in observations with MWA Phase-3, and even more with SKA, as discussed below.
\subsection{Structure Function Analysis}
Because of the excellent ground coverage of the MWA stations, our observations are highly applicable for SF analysis, to characterise
the fine scale structure of the ionospheric wavefront distortions above the array towards multiple viewing directions, and probe the underlying physical nature of the distorting media.
Because of the denser station coverage we are able to reconstruct the fine structure much better than the similar study on LOFAR\cite{mevius_16}, whilst the latter explores a much large range of baseline lengths. The observational differences lead to slightly different methods; for example the tracking LOFAR observations allow for the subtraction of the temporal mean phase, where as for the snapshot MWA observations we have subtracted a DI calibration.
Because of this,
in cases where one very strong source dominates the DI calibration, the residual ionospheric screen slope in the direction of that source (e.g. 3C444) can be significantly lower than the median slope over all directions.
An example of this can be seen in Figure \ref{fig:3C444} where the intercept at 1\,km (that corresponds to the slope of the ionospheric screen) of the SF at 88\,MHz is less than the intercept of the 154\,MHz SF, where prima-facie one would expect the reverse. This occurs because of the greater dominance of this source in the DI calibration at the lower frequency. However, the log-log SF gradients are preserved.
The benefits from high signal-to-noise ratio measurements are the potential to probe smaller scales, increased precision of ionospheric calibration, and tighter constraint of the upper limit of the inner scale of turbulence, in the absence of other systematic errors.
MWA is also free of potential clock errors, which were one of the contributions to the systematic noise floor limits in the LOFAR analysis;
for 3C444 we reach a noise floor of 0.1, 0.3 and 0.05 mTECU
for our three datasets in poor and good weather at 154MHz and good weather at 88MHz, as measured in the detrended data over the baseline range of 50 to 100m.
Nevertheless the two approaches reach similar results as to the SF gradients being most compatible with Kolmogorov turbulence. Mevius et\ al.{} use the `refractive scale' $R_{\rm diff}$ as a physically relevant measurement (and a proxy of ionospheric weather quality), whereas we use the screen slope (mTECU/km) which is independent of the SF gradient. These two quantities are trivially convertible and comparable between the two studies.
Our median slopes of 5$\pm$3mTECU/km is equivalent to $R_{\rm diff}$ being 2 to 12km, and $R_{\rm diff}$ values of more than 5\,km are considered suitable for EoR observations\cite{mevius_16}; this of course assumes that the impact of severe weather manifested as a high ionospheric slope is also turbulent, and ignores the potential for a high quality DD-calibration to turn a nominally `bad weather day' into a good weather day.
In our comparison of the range of $R_{\rm diff}$ we note the differences in the methods; LOFAR is a pointed tracking observation with a constant ionospheric wedge over the array subtracted, whereas this analysis is for a wide-FoV experiment with many lines of sight where a global DI model has been subtracted. These differences will affect the deduced $R_{\rm diff}$, making exact comparison difficult.
We have used toy surface models to improve our interpretations of the
SF method.
These show that in the presence of noise, interpreting a gradient of $\sim$5/3 as proof of Kolmogorov turbulence is risky, particularly with the short baseline range of the MWA, as noise flattens the gradient in the full structure function.
We have investigated the impact of joint fitting of the noise floor plus the noise-free structure function (i.e. $C^2r^\beta + \sigma^2$), but found that for the MWA the terms $\beta$ and $\sigma$ are highly correlated.
We found our best results by limiting the measurement of the gradient to the range of baselines least affected by the noise (that is 0.5 to 5km).
Finally, the signature of the toy model's curvature is hard to see in the SF, as it is dominated by the linear component.
However, the detrended structure function does allow for the robust detection of non-linear distortions, the spatial scales for the onset of these effects and for the end of the inertial-like regime.
For the strong source 3C444 (Figure \ref{fig:3C444}) we can track the structure functions down to the shortest baselines without being overcome by the thermal noise; the gradient in this case is 1.8 at 154\,MHz in both good and poor weather days, and 1.7 at 88\,MHz, between 0.5 and 5km.
Thus at both frequencies the best fit interpretation would agree with being Kolmogorov turbulence, but we don't consider this proven, given the difficulty of separating the gradient from the noise.
Using DSF (i.e. after subtracting a linear surface) the detectable divergence from the linear fit occurs at less than 1km, which we can confidently interpret as the first indications of non-linear behaviour above the noise.
At the longest baselines there is a change in the detrended structure function gradient, which is characteristic of curvature.
In summary, we would recommend that non-linear corrections are applied for baselines greater than 1km.
Similar conclusions come from the study of the ensemble of directions and times to strongest sources in all the datasets;
the gradients are consistent with Kolmogorov turbulence, being 1.6$\pm$0.2 and 1.8$\pm$0.1 for the 88MHz and 154MHz datasets (Figure \ref{fig:SF}). But we stress that these measured gradients could be underestimated because of the impact of the thermal noise.
The detrended SF (Figure \ref{fig:SF-detrend}) allow us to better detect deviations from a planar surface.
These show that in general at 88MHz, even in moderate-weather, higher order ionospheric corrections should be used for baselines greater than 0.6km and must be used for baselines greater than 2km.
The combined SF at 154MHz are more consistent with planar solutions, except for strong sources (i.e. as shown in Figure \ref{fig:3C444}) and/or poor weather, when higher order corrections are required over the same baseline ranges.
This implies that MWA-2 falls in Lonsdale's regime 4.
\subsection {Implications for SKA}
The SKA-Low stations will be approximately ten times larger, thus have close to a hundred times larger collecting area (per station). Furthermore the instantaneous bandwidth will be ten times larger.
Therefore the sensitivity will be much higher than for MWA, and the importance of precise ionospheric calibration (and other systematic errors), will be even more significant.
The expected continuum baseline sensitivity is 3mJy in 1 minute\cite{ska_perf}, from which we predict more than 100 sources with SNR$>$6 in the FoV, using the TREC models\cite{TREC}.
We have found that for MWA at 88MHz, even in moderate-weather, the influence of higher order terms are detectable for baselines greater than 0.6km and are significant for baselines greater than 2km.
On the other hand, for most sources at 154MHz in moderate-weather planar solutions should be sufficient, but for strong sources (i.e. as shown in Figure \ref{fig:3C444}) and/or poor weather (or both) higher order corrections are required.
These conclusions will be most applicable for the SKA, where the increase in sensitivity mean that many sources will be at similar or higher SNR that the few limited examples with the MWA.
Thus we conclude that station based directional dependent calibration solutions are required for both the final imaging,
and also for the tied-array beam forming, even for $\sim$kilometre-long baselines.
The VLBI beamforming, for example, is assuming that all stations with in a radius of 20\,km (or as a fall-back position 4\,km) will be included in the formed tied-array beams\cite{skavlbi_wp10}.
In this case real-time station-based direction-dependent calibration will be required; this can be provided by LEAP, as we are currently working on demonstrating.
In summary the opportunities opened up by the amazing potential of SKA-Low to revolutionise low-frequency astronomy require that the calibration of the data is equally accurate. Thus station-based directional-dependent corrections will be needed for all baselines and operations.
|
\section{Introduction}
Excitons are the bound states formed by an electron-hole pair in a semiconductor crystal \cite{schaefer-book,haugkoch}, and as such, they are close analogues of the hydrogen atom. Excitons manifest themselves as optical absorption or emission lines within the band gap of the material. In the theoretical treatment of exciton formation, light-matter interaction is most often described within the dipole approximation. This approximation disregards the spatial structure of the light field, which is justified by the tiny length scale on which the de Broglie waves vary, as compared to the wavelength of the light \cite{cohentannoudji}.
The dipole approximation gives rise to optical selection rules related to the conservation of angular momentum during an optical transition. Within the dipole approximation, light may carry only one quantum of angular momentum per photon, realized through the circular polarization of the light; therefore, optical transitions can change angular momentum quantum numbers of the matter only by one unit. In atoms, this rule selects the $s$-to-$p$ or $p$-to-$d$ transitions; in semiconductors, the selection rule affects the orbitals of the bands in an analogous way. As a consequence of the dipole approximation, excitons created from such a dipole transition do not carry angular momentum; that is, the respective quantum number $j$ is zero. Moreover, as established by the Elliott formula \cite{elliott57}, the transition amplitude quickly decays with the principal quantum $n$ (as $(n+1/2)^{-3}$ in 2D), such that the exciton spectrum is strongly dominated by transitions into the 1$s$ state, i.e. the state corresponding to the hydrogenic ground state. A common technique to allow optical access to the $p$ exciton series in semiconductors is nonlinear, two-photon spectroscopy \cite{berkelbach15,wang15}.
It is clear that effects beyond the dipole approximation can modify the selection rule of the excitons. In particular, the dipole approximation disregards the possible spatial structure of the light beam, which in the case of twisted light results in a well-defined orbital angular moment (OAM) per photon \cite{allen92,yao11,torres2011twisted}. The photon OAM essentially adds another tunable degree of freedom for tailoring light-matter interaction. It has been proposed to use this new degree of freedom for generating a topological band structure by breaking time-reversal symmetry \cite{bhattacharya21}, for pumping electrons in a magnetic field through the Landau level \cite{grass18,fujita19,cao21}, or for producing topological defects such as vortices or skyrmions \cite{fujita17,fujita172,cian20,kim21}.
A striking demonstration of how optical selection rules are modified by photon OAM has been achieved in an experiment with trapped ions \cite{schmiegelow16}, showing a dipole-forbidden atomic $s$-to-$d$ transition in the presence of a twisted light field. Given the analogy between an exciton and a hydrogen atom, it might be expected that twisted light can generate transitions into dark excitonic levels, where the photonic OAM is absorbed in the internal degree of freedom. This indeed has theoretically been suggested for the case of Rydberg excitons \cite{konzelmann19}. On the other hand, there have also been experiments with atomic and polaritonic condensates in which twisted light has led to the formation of vortices \cite{andersen2006,kwon19}. This indeed would suggest that the orbital angular momentum of the photon is absorbed by the center-of-mass (COM) degree of freedom of the exciton, rather than by the relative motion of electron and hole. Also, in the strong drive limit and absence of Coulomb binding, OAM of light can lead to Floquet vortex creation, but it is an open question how strong drive and exciton formation compete with each other.
To better understand the fate of the photonic OAM in excitonic transitions, the present paper studies the case of a single exciton in a two-band semiconductor model in 2D in the presence of a twisted light source. We extend theoretical studies of band-to-band transitions in semiconductors or graphene with twisted light, presented in Refs. \onlinecite{quinteiro09,quinteiro10,farias13}, to the case where Coulomb interactions give rise to exciton formation. Our analysis demonstrates that, under the assumption that the spatial variation of the light occurs on a length scale much larger than the size of the unit cell of the semiconductor crystal, transitions into excitonic levels $j\neq 0$ remain completely forbidden even in the presence of twisted light. Instead, the structure of the light field selects the COM degree of freedom of the excitons. Since it is the relative motion that essentially determines the energy of an exciton, it follows that the twist of the light source does not modify the excitonic spectrum. In this context, we also note that small shifts of the spectrum are possible if the COM dispersion of the exciton is taken into account. This indeed has been observed in a recent experiment with excitons in a Dirac material, which found a blueshift of the exciton lines for sufficiently large values of photon OAM \cite{simbulan20}.
Our paper is organized in the following way: In Sec. II, we develop the general analytical formalism to describe exciton transitions in structured light beam. In Sec. III, we specifically address the case of a Bessel beam. To evaluate this case, we make use of the rotational symmetry of the beam which makes an explicit numerical treatment feasible. With this we are able to show, for a finite system size, that the $s$ states are optically bright, in quantitative good agreement with the 2D Elliot formula, independent from the choice of the photon OAM. Our numerical calculation also confirms that the COM momentum of the exciton is peaked at the linear momentum of the photon.
\section{General Analytical Model}
\subsection{Light-matter coupling}
We consider a 2D semiconductor with Bloch bands $\lambda$ and wave vector ${\bf k}$, described by Bloch functions $\varphi_{\lambda,{\bf k}}({\bf r}) = \frac{1}{\sqrt{S}} e^{i {\bf k}\cdot {\bf r}} u_{\lambda,{\bf k}}({\bf r})$, where $u_{\lambda,{\bf k}}({\bf r}+{\bf R_i}) = u_{\lambda,{\bf k}}({\bf r})$ with ${\bf R}_i$ a lattice vector.
In this basis, the crystal Hamiltonian reads $H_0 = \sum_{\lambda,{\bf k}} \epsilon_{\lambda,\bf k} c^{\dagger}_{\lambda,{\bf k}} c_{\lambda,{\bf k}}$,
with $ c_{\lambda,{\bf k}}$ ($c^\dagger_{\lambda,{\bf k}}$) being the annihilation (creation) operators, and $\epsilon_{\lambda,\bf k}$ the dispersion. We assume a light field given by a vector potential ${\bf A}({\bf r})=A({\bf r}) \cdot {\bf e}$ in the Coulomb gauge, such that the light-matter Hamiltonian is given by:
\begin{align}
H_{\rm LM} = \sum_{\lambda',\lambda} \sum_{{\bf k}',{\bf k}} \frac{i e \hbar}{M} \langle \lambda', {\bf k}' | {\bf A}({\bf r}) \cdot \nabla_{\bf r} | \lambda, {\bf k} \rangle c^\dagger_{\lambda',{\bf k}'} c_{\lambda, {\bf k}}.
\end{align}
We are only interested in the matrix element $h^{\lambda',\lambda}_{{\bf k}',{\bf k}} = \langle \lambda', {\bf k}' | {\bf A}({\bf r}) \cdot \nabla_{\bf r} | \lambda, {\bf k} \rangle = \int d^2r \varphi^*_{\lambda',{\bf k}'}({\bf r}) A({\bf r}) {\bf e}\cdot\nabla_{\bf r} \varphi_{\lambda,{\bf k}}({\bf r})$ with $\lambda={\rm c}$ and $\lambda'={\rm v}$, i.e. transitions amplitudes between conduction and valence band. Taking into account the orthonormality of the bands, the derivative operator has to act onto the lattice-periodic function $u_{\lambda,{\bf k}}$ to yield non-zero contributions. Explicitly, we have
\begin{align}
h^{\rm v,c}_{{\bf k}',{\bf k}} = \frac{1}{S} \int d^2r A({\bf r}) e^{i({\bf k}-{\bf k}')\cdot {\bf r}} u^*_{{\rm v}, {\bf k}'}({\bf r}) {\bf e}\cdot\nabla_{\bf r} u_{{\rm c},{\bf k}}({\bf r}).
\label{hvc}
\end{align}
Here, $S$ is the size of the system.
At this stage, we make the approximation {\bf (A1)}: the vector potential and the exponential do not vary within a unit cell. By keeping variations beyond the scale of a unit cell, this approximation is less restrictive than the dipole approximation which would fully ignore the spatial structure of the light. Yet without considering a particular choice of vector potential, the spatial variations of the beam is generally limited by a length scale on the order of the wavelength of the light. The same length scale also determines the variation of the exponential $e^{i({\bf k}-{\bf k}')\cdot{\bf r}}$, since it will turn out \textit{a posteriori} that ${\bf k}-{\bf k}'$ is determined through the photon momentum. Since the optical wavelength is usually several orders of magnitude larger than the size of unit cell, the approximation is totally valid in the usual cases, but might not hold in some special cases, e.g. of Moir\'{e} lattices with enlarged unit cells \cite{seyler19,tran19,jin19,alexeev19}. Applying (A1) to Eq.~(\ref{hvc}), we write:
\begin{align}
h^{\rm v,c}_{{\bf k}',{\bf k}} &= \frac{1}{N}_{\rm sites} \sum_{{\bf R}_i} A({\bf R}_i) e^{i({\bf k}-{\bf k}')\cdot {\bf R}_i} \times \nonumber \\ &
\times \frac{1}{S_{\rm cell}} {\bf e}\cdot \left[ \int_{\rm cell} d^2r u^*_{{\rm v},{\bf k}'}({\bf r}) \nabla_{\bf r} u_{{\rm c},{\bf k}}({\bf r}) \right]
\equiv \nonumber \\ & \equiv
A_{\bm \kappa} \times {\bf e}\cdot {\bf p}_{{\bf k}',{\bf k}}^{\rm vc},
\label{eqn:hvc}
\end{align}
where ${\bm \kappa}={\bf k}-{\bf k}'$ and $A_{\bm \kappa}= \frac{1}{N}_{\rm sites} \sum_{{\bf R}_i} A({\bf R}_i) e^{i {\bm \kappa} \cdot {\bf R}_i}$ the Fourier transform of the vector potential. The dipole moment between the ${\bf k}'$ state in the valence band and the ${\bf k}$ state in the conduction band is denoted by ${\bf p}_{{\bf k}',{\bf k}}^{\rm vc}$.
We proceed by making a second approximation {\bf(A2)}: the dipole moment depends only weakly on the wave vectors ${\bf k}$ and ${\bf k}'$. Indeed, the most radical implementation of this approximation in which the dipole moment is set to a constant ${\bf p}^{\rm vc}_0$ is commonly used in the literature, cf. Ref. \onlinecite{haugkoch}. To be less restrictive, we argue that ${\bf p}_{{\bf k}',{\bf k}}^{\rm vc}$ may depend on ${\bf k}+{\bf k}'$ (which can take relatively large values), whereas the dependence on ${\bf k}-{\bf k}'$ (which remains small since it is equivalent to the photon momentum) is neglible. To this end we introduce the quantity ${\bf K}= \frac{1}{2}({\bf k}+{\bf k}')$ and assume a linear (or linearized) dependence on ${\bf K}$:
\begin{align}
{\bf p}^{\rm vc}_{\bf K} = {\bf p}^{\rm vc}_0 + ({\bm \alpha}\cdot {\bf K}){\bf p}^{\rm vc}_1.
\end{align}
For notational convenience, we write ${\bf e}\cdot{\bf p}^{\rm vc}_{\bf K} = p^{\rm vc}_{\bf K} = p^{\rm vc}_0 + ({\bm \alpha}\cdot {\bf K})p^{\rm vc}_1$. The light-matter matrix element is finally written as:
\begin{align}
h^{\rm v,c}_{{\bf k}',{\bf k}} = h^{\rm vc}_{{\bm \kappa},{\bf K}} = A_{\bm \kappa} p_{\bf K}^{\rm vc}.
\end{align}
This expression makes it immediately clear that the wave vector $\bm \kappa$ is exclusively selected by properties of the light field, whereas the wave vector $\bf K$ is exclusively determined by material properties. In the following, we will find that, in the case of exciton transitions, $\bm \kappa$ ($\bf K$) is related to the COM (relative) momentum of the exciton.
\subsection{Exciton transitions}
We are now interested in the transition amplitude for exciton formation ${\cal T}_X \equiv \langle X | H_{\rm LM} | {\rm vac} \rangle$. Here, $| X \rangle$ denotes an excitonic state, which in 2D is characterized through four quantum numbers for relative and COM motion. We choose $|X\rangle = |{\bf k}_{\rm com},n,j\rangle$, i.e. we describe the excitonic state by its linear COM momentum ${\bf k}_{\rm com}$, and its hydrogenic quantum numbers $n$ and $j$, representing the relative degrees of freedom. The vacuum state $|{\rm vac}\rangle$ corresponds to a filled valence band and an empty conduction band.
The excitonic wave function can be written as
\begin{align}
\langle {\bf R}, {\bf r} | X \rangle \equiv
\Phi_{n,j}^{{\bf k}_{\rm com}}({\bf R},{\bf r}) \equiv
\Phi_{{\bf k}_{\rm com}}^{\rm (com)}({\bf R}) \times
\Phi_{n,j}^{\rm (rel)}({\bf r}),
\end{align}
where ${\bf r}={\bf r}_{\rm e}-{\bf r}_h$ are relative coordinates of an electron-hole pair, and ${\bf R}=\frac{1}{2}({\bf r}_{\rm e}+ {\bf r}_{\rm h})$ are the COM coordinates. The relative motion of electron and hole is described by the solutions to the 2D hydrogen atom, which are given by \cite{haugkoch}
\begin{align}
\label{phirel}
\Phi_{n,j}^{\rm (rel)}({\bf r}) = \tilde{\cal N}_{n,j} f_{nj}(r) e^{i j \phi_r} = \tilde{\cal N}_{n,j} \rho^{|j|} e^{-\frac{|\rho|}{2}} L_{n-|j|}^{2|j|}(\rho) e^{i j \phi_r},
\end{align}
where $\rho=r\rho_n$, with the inverse length scale given by $\rho_n=\frac{2}{(n+1/2)a_0}$. The material-specific length scale $a_0=\hbar^2 \epsilon/(e^2M)$ is the effective Bohr radius depending on effective mass $M$ and dielectric constant $\epsilon$. The normalization of the relative wave function is given by
\begin{align}
\tilde{\cal N}_{n,j} = \sqrt{\frac{(n-|j|)!}{(n+|j|)!} \left(\frac{\rho_n}{2}\right)^2 \frac{1}{\pi} \frac{1}{n+1/2}}.
\end{align}
For the COM part, we simply assume plane waves, $\Phi_{{\bf k}_{\rm com}}^{({\rm com})}({\bf R}) = e^{i {\bf k}_{\rm com}\cdot {\bf R}}$.
Without making use of the explicit solution for the excitonic wave functions, we write for the transition amplitude:
\begin{align}
\label{AX}
{\cal T}_X &= \int d^2{\bf R} \int d^2{\bf r} \sum_{{\bm \kappa},{\bf K}} \langle X|{\bf R},{\bf r} \rangle \langle {\bf R},{\bf r} | {\bm \kappa},{\bf K} \rangle
\times \nonumber \\ & \times \langle {\bm \kappa},{\bf K} | H_{\rm LM} | {\rm vac} \rangle.
\end{align}
The last term corresponds to the band-to-band transition amplitude evaluated above,
\begin{align}
\langle {\bm \kappa},{\bf K} | H_{\rm LM} | {\rm vac} \rangle &= \langle {\bf K}-\frac{\bm \kappa}{2} | H_{\rm LM} | {\bf K} + \frac{\bm \kappa}{2} \rangle = \nonumber \\ &=
h^{\rm v,c}_{{\bf K} - \frac{\bm \kappa}{2},{\bf K} + \frac{\bm \kappa}{2}} = A_{\bm \kappa} p_{\bf K}^{\rm vc},
\end{align}
which is Fourier transformed to spatial coordinates by the second term,
\begin{align}
\langle {\bf R},{\bf r} | {\bm \kappa},{\bf K} \rangle = \frac{1}{S^2} e^{i ({\bf K}+\frac{\bm \kappa}{2}) \cdot{\bf r}_{\rm e} - i ({\bf K}-\frac{\bm \kappa}{2}) \cdot{\bf r}_{\rm h} } = \frac{1}{S^2} e^{i ({\bf K}\cdot{\bf r} + i{\bm \kappa}\cdot{\bf R})}.
\end{align}
This expression explicitly shows that the wave vector ${\bm \kappa}$ ($\bf K$) is conjugate to the COM (relative) variable.
Plugging all expressions into Eq.~(\ref{AX}), the ${\bf R}$ integral is immediately evaluated into a Kronecker-Delta $\delta_{{\bm \kappa},{\bf k}_{\rm com}}$, so photon momentum ${\bm \kappa}$ and COM momentum ${\bf k}_{\rm com}$ must match.
We obtain:
\begin{align}
\label{AX2}
{\cal T}_X &= \frac{(2\pi)^2 \tilde {\cal N}_{nj}}{S^2} A_{{\bf k}_{\rm com}} \int d^2{\bf r} f_{nj}(r) e^{ij \phi_r} \times \nonumber \\ & \times \sum_{\bf K} e^{i {\bf K}\cdot{\bf r}} (p_0^{\rm vc} + {\bm \alpha}\cdot{\bf K} p_1^{\rm vc}).
\end{align}
The ${\bf K}$-sum is a Fourier transform into the relative variable ${\bf r}$, and we can write $\sum_{\bf K} e^{i {\bf K}\cdot{\bf r}} (p_0^{\rm vc} + {\bm \alpha}\cdot{\bf K} p_1^{\rm vc}) = (2\pi)^2 [p^{\rm vc}_0 \delta({\bf r}) - i p_1^{\rm vc} {\bm \alpha}\cdot \nabla_{\bm r} \delta({\bf r})]$. From this expression it can immediately be seen that a constant dipole moment leads to non-zero transition amplitudes only if the relative exciton wavefunction $f_{nj}(r)$ is non-zero at $r=0$. This is the case only for $s$-excitons. The linear dependence of the dipole momentum on ${\bf K}$, expressed by the second term, gives rise to non-vanishing transition amplitudes if the first derivative of $f_{nj}(r)$ is non-zero at $r=0$.
This second term enables the formation of excitons in higher momentum states than the $s$ series. However, we emphasize that this term is independent from the light source, and with respect to the relative degrees of freedom ($n,j$), we get the same transitions, no matter what the spatial structure of the light might be (as long as approximation (A1) holds). Our analysis shows that the Elliott formula is unchanged by spatial structure of light beyond the scale of the unit cell.
Our result agrees with a recent experiment in 2D transition metal dichalcogenides (TMDs) where twisted light has been used to reveal light-like exciton dispersion \cite{simbulan20}. Using non-resonant Laguerre-Gaussian beams, they observed a blue shift of the exciton energy that increased with $\ell$; this indicates that the OAM was transferred preferentially to the COM of the exciton during its creation. However, we stress that our results apply much more generally since no particular 2D semiconductor or spatial light profile was specified during the analysis. This implies that the dispersion of all 2D excitons could be probed in a similar manner, presenting an alternative method to the traditional angle-resolved photoluminescence measurements \cite{kwon19}.
\section{Example: Excitons from twisted light}
\begin{figure*}
\centering
\includegraphics[scale=0.6]{fig1.eps}
\caption{
(a) A Bessel light beam ${\bf A}$ with photon momentum/angular momentum $(q_{\|},\ell)$ creates an electron-hole pair in a 2D electron gas (2DEG). The electron [hole] is characterized by quantum numbers $(m,\nu)$ [$(m',\nu')$] for angular momentum/ momentum. The pair can form a bound state, and the degrees of freedom of such an exciton are the center-of-mass motion, characterized by angular momentum/ momentum quantum numbers $(J,N)$, and the relative motion, characterized by quantum numbers $(j,n)$ for angular momentum and energy. (b) The selection rules for optical transitions and exciton formation reflect (i) conservation of angular momentum, and (ii) conservation of linear momentum, reflected by the illustrated triangle conditions.}
\label{Tnu}
\end{figure*}
Thus far, we have been very general in our treatment with respect to the profile of the optical excitation. In the following, we are going to treat the specific case of a Bessel beam and, besides the analytical treatment along the lines presented in the previous sections, we will also present the result of numerical evaluations. With this choice of the vector potential, our system exhibits a cylindrical symmetry, since the light field has an azimuthal phase dependence $\exp(i \ell \phi)$, where $\ell$ defines the OAM per photon (in units $\hbar$). To match this symmetry, we consider a cylindrical sample, noting that the sample geometry becomes irrelevant in the thermodynamic limit. Accordingly, we adopt our theoretical description to this symmetry, and express the light-matter coupling in terms of a cylindrical wave functions, see also Refs. \onlinecite{quinteiro10,farias13}. We illustrate this case in Fig.~\ref{Tnu}, where we also sketch the resulting selection rules in terms of the good quantum numbers for the cylindrical symmetry.
\subsection{Band-to-band transitions in cylindrical basis}
Instead of plane waves with linear momentum quantum number, the electronic basis in a cylindrical sample is best described by wave functions $ \varphi_{m,\nu}({\bf r}) = {\cal N}_{m,\nu} J_m(k_{m,\nu} r) \exp(i m \phi)$, which solve the Schr\"odinger equation for free electrons with cylindrical boundary conditions. Here, $J_m(x)$ denotes the $m$th Bessel function, and the momenta $k_{m,\nu}$ must be chosen such that the wave function vanishes at the system boundary (i.e. for $|{\bf r}|=R_0$). Therefore, we have $k_{m,\nu}=x_{m,\nu}/R_0$, with $x_{m,\nu}$ the $\nu$th zero of the $m$th Bessel function. The normalization is given by ${\cal N}_{m,\nu} = (R_0\sqrt{\pi} |J_{|m-1|}(x_{m,\nu})|)^{-1}$. As in the previous section, the crystal lattice is taken into account by multiplying the wave functions $\varphi_{m,\nu}$ with lattice-periodic Bloch functions $u_\lambda({\bf r})$ for the bands $\lambda$. For simplicity, the Bloch functions are assumed to be independent from the quantum numbers $m$ and $\nu$. With this, the electronic basis is given
\begin{align}
\varphi_{\lambda;m,\nu}({\bf r}) = {\cal N}_{m,\nu} J_m(k_{m,\nu}r) \exp(im\phi) u_{\lambda}({\bf r}).
\end{align}
In this basis, the light-matter transition amplitudes in the Coulomb gauge (within the weak field limit) are given by
\begin{align}
\label{sumcell}
h^{\rm v,c}_{m,\nu;m',\nu'} = \frac{-i\hbar e}{SM} \int d{\bf r} \bar \varphi_{+, m',\nu'}({\bf r}) \left[ {\bf A}({\bf r})\cdot \nabla_{\bf r} \right] \varphi_{-, m,\nu}({\bf r}),
\end{align}
with $\bar \varphi$ denoting the complex conjugate of $\varphi$.
Before we proceed, let us first fix the vector potential. We consider a vector potential which in the sample plane reads ${\bf A}({\bf R}) = A_0 a(R) e^{i\ell \phi} {\bf e}_{\sigma}$, where ${\bf e}_{\sigma}$ is the polarization in the plane. For a circularly polarized Bessel beam with OAM $\ell$, we have $a(R) = J_\ell(q_{\|} R)$, with $q_{\|}$ the in-plane photon momentum. Note that a vertical contribution to the vector potential, needed to fulfill Maxwell's equation, is neglected here since it is not relevant for light-matter interaction with a two-dimensional medium. However, it is important to keep in mind that the frequency $\omega$ of the photon depends also on the perpendicular momentum component $q_z$, $\omega = \frac{c}{\hbar}\sqrt{q_{\|}^2 +q_z^2}$.
Invoking the approximations (A1) and (A2), both ${\bf A}({\bf r})$ and $J_m(k_{m,\nu}r) \exp(im\phi)$ can be considered constant on the scale of the lattice constant $a$. Thus, the evaluation of $h^{\rm v,c}_{m,\nu;m',\nu'}$ can be split into an integral $\cal I$ restricted to the unit cell and a sum over units cells ${\cal S}_{m,\nu;m',\nu'}$, that is, we can write $h^{\rm v,c}_{m,\nu;m',\nu'}= \frac{-i\hbar }{ S M} {\cal S}_{m,\nu;m',\nu'} \times {\cal I}$. As before, the sum over the whole system takes into account the variation of the light field, occurring on larger scales, whereas the unit cell integral determines the material's dipole moment taken to be a constant.
Explicitly, the two contributions are given by
\begin{align}
{\cal I} = e \int_{\rm c} {\rm d}{\bf r}\bar u_+({\bf r}) ( {\bf e}_\sigma \cdot \nabla_{\bf r} ) u_-({\bf r}),
\end{align}
and
\begin{align}
{\cal S}_{m,\nu;m',\nu'} =& {\cal N}_{m,\nu} {\cal N}_{m',\nu'} A_0 \sum_i J_m(k_{m,\nu}R_i) J_{m'}(k_{m',\nu'}R_i) \nonumber \\ &
\times J_\ell(q_{\|} R_i) \exp[i(m-m'+\ell)\phi_i],
\end{align}
where $(R_i,\phi_i)$ denote the lattice vectors. In analogy to Eq.~(\ref{eqn:hvc}), we can read off the results of the cell integral ${\cal I}$ as the interband dipole moment $p_{\rm vc} \equiv {\bf e}_{\sigma} \cdot {\bf d} = {\cal I}$, which depends only on the material. Since we take it to be constant here, it will enter the transition amplitudes only as a prefactor. To evaluate ${\cal S}_{m,\nu;m',\nu'}$, we replace the summation over cells by an integral. With this, we immediately arrive at a first selection rule from the angular part of the integral:
\begin{align}
\label{kronS}
{\cal S}_{m,\nu;m',\nu'} \propto \delta_{\ell+m-m'}.
\end{align}
The radial integral in ${\cal S}_{m,\nu;m',\nu'}$ is over a product of three Bessel functions, ${\cal S}_{m,\nu;m',\nu'} \propto \int_0^{R_0} dr \ r J_m(k_{m,\nu}r) J_m'(k_{m',\nu'}r) J_\ell(q_{\|} r)$. Its analytic solution (in the limit $R_0 \rightarrow \infty$) has been derived in Ref. \onlinecite{jackson72}, and can also be found in the Supplemental Material (SM). Here, we only consider that, from this solution, the integral takes non-zero values only if a triangle condition is fulfilled: The three scalars $k_{m,\nu}$, $k_{m',\nu'}$, and $q_{\|}$ must be such that they can form a triangle (including the limit in which the triangle is squeezed to a line). Therefore, this condition yields a second selection rule: the change of electron momentum upon a band-to-band transition is bounded by the in-plane momentum of the photon.
\subsection{Exciton transitions}
The amplitudes $h^{\rm v,c}_{m,\nu;m',\nu'}$ quantify the band-to-band transition which generates an electron-hole pair characterized by $m-m'=\ell$ and $|k_{m,\nu}-k_{m',\nu'}| \leq q_{\|}$. Next, we have to ask which excitonic states can be formed from these pairs. In accordance with the presumed cylindrical symmetry of the system, we now also describe the excitonic states in terms of cylindrical-symmetric quantum numbers, $|X\rangle = |N,J,n,j\rangle$, where $n,j$ account for the state of relative motion (as before), and $N,J$ for the COM degrees of freedom (instead of ${\bf k}_{\rm com}$ used in the previous section).
Again, the excitonic wave function is a product of the relative ($\bf{r}$) and COM ($\bf{R}$) contributions: $\langle {\bf R},{\bf r}|X \rangle \equiv \Phi_{n,j}^{J,N}({\bf R},{\bf r}) \equiv \Phi_{N,J}^{\rm (com)}({\bf R}) \times \Phi_{n,j}^{\rm (rel)}({\bf r})$. The relative part is unchanged, given by Eq.~(\ref{phirel}). Since the COM of the exciton is subject to the same boundary conditions as electron and hole individually, its wave function is given by:
\begin{align}
\Phi_{J,N}^{\rm (com)}({\bf R}) = {\cal N}_{J,N} J_J(k_{J,N} R) \exp(i J \phi_{\rm com}),
\end{align}
where the quantum number $J$ denotes the angular momentum of the COM, and both $J$ and $N$ together define the total COM momentum $Q_{\rm com} = x_{J,N}/R_0$.
Projecting the excitonic wave function onto the rotationally symmetric basis for electron and hole wave functions is equivalent to a Hankel transform. This projection yields a quantity
$ {\cal B}_{m,\nu;m',\nu'}^{n,j;J,N}$:
\begin{align}
\label{B}
{\cal B}_{m,\nu;m',\nu'}^{n,j;J,N} =& {\cal N}_{m,\nu} {\cal N}_{m',\nu'} \int {\rm d}{\bf r}_{\rm e} \int {\rm d}{\bf r}_{\rm h}
\bar \Phi_{n,j}^{J,N}({\bf R},{\bf r}) \nonumber \\ &
J_m(k_{m,\nu}r_{\rm e}) J_{m'}(k_{m',\nu'}r_{\rm h}) \exp[i(m \phi_{\rm e} - m' \phi_{\rm h})].
\end{align}
An explicit analytic expression which solves this integral is provided in the SM. As before for ${\cal S}_{m,\nu;m',\nu'}$, we also encounter a triangle condition in the evaluation of $ {\cal B}_{m,\nu;m',\nu'}^{n,j;J,N}$: it is non-zero, only if the lengths $k_{m,\nu}$, $k_{m',\nu'}$, and $k_{J,N}$ form a triangle, i.e. $ |k_{m,\nu}-k_{m',\nu'}|\leq k_{J,N}$. More importantly, as shown in the SM, one of the integrals in Eq. (\ref{B}) yields a Kronecker-$\delta$:
\begin{align}
\label{kronB}
{\cal B}_{m,\nu;m'\nu'}^{n,j;J,N} \sim \delta_{j+J,m-m'}.
\end{align}
Together with the selection rule for band-to-band transitions, Eq.~(\ref{kronS}), Eq.~(\ref{kronB}) reflects conservation of angular momentum.
We are now in the position to calculate the exciton transition amplitude ${\cal T}_{n,j}^{J,N} \equiv \langle X_{n,l}^{J,N} | H_{\rm LM} | {\rm vac} \rangle$:
\begin{align}
{\cal T}_{n,j}^{J,N} = \sum_{m,\nu;m',\nu'} h^{\rm v,c}_{m,\nu;m',\nu'} {\cal B}_{m,\nu;m'\nu'}^{n,j;J,N}.
\label{Tnj}
\end{align}
These sums should go over all occupied (empty) levels $m',\nu'$ ($m,\nu$), but a more practical limitation of these sums is due to the fact that ${\cal B}_{m,\nu;m'\nu'}^{n,j;J,N} \approx 0$ when either $k_{m,\nu}$ or $k_{m',\nu'}$ become much larger than the inverse of the Bohr radius, $a_0^{-1}$.
\subsection{Numerical evaluation}
The last observation allows us to introduce a cutoff momentum $k_{\rm cut} \gg a_0^{-1}$ at which the sums can be truncated. With this, the numerical evaluation of Eq.~(\ref{Tnj}) become feasible. For concreteness, by comparison of ${\cal T}_{n,j}^{J,N}$ obtained from different cutoff momenta $k_{\rm cut}$, we estimate that the relative error remains below 0.1 for $k_{\rm cut} a_0 \geq 3.75$. In the numerical evaluation of ${\cal T}_{n,j}^{J,N}$ presented below, we have included 47,100 Bessel functions. With that, $k_{\rm cut} a_0>3.75$ for system sizes up to $R_0/a_0=12,500$. We note that $k_{\rm cut}$ also restricts the sums in $m$ and $\nu$ in the following way: $|m|<\pi k_{\rm cut} R$ and/or $\nu<k_{\rm cut} R$.
The numerical evaluation confirms the analytical result from Sec. II that the Elliott formula remains unchanged by the spatial structure of the light source. To this end, we obtained the height of the spectral lines, $\tilde {\cal T}_{n,j}$, by summing the contributions from all COM momentum modes at a given $n$ and $j$:
\begin{align}
\tilde {\cal T}_{n,j} = \frac{1}{{\cal N}_{\cal T}} \sum_{N,J} {\cal T}_{n,j}^{J,N}.
\end{align}
To make this quantity independent from the intensity of the light, we normalize by ${\cal N}_{\cal T} = \sqrt{ \sum_{n,j,N} \left| {\cal T}_{n,j}^{J,N} \right|^2 }$.
The results are shown in Table~\ref{tab:1} for a system of size $R_0=10^4 a_0$ in a Bessel beam with OAM $\ell=0$ and $\ell=1$ and in-plane photon momentum $q_{\|}=10^{-3}a_0^{-1}$. For comparison, we also provide the results from the 2D Elliott formula for an infinite system. All values agree very well with each other.
\begin{table}[t]
\centering
\begin{tabular}{|c||c|c|c|}
\hline
& 1s & 2s & 3s \\
\hline
\hline
OAM 0 & 0.998 & 0.037 & 0.0079 \\
OAM 1 & 0.995 & 0.037 & 0.0079 \\
2D Elliott & 0.9993 & 0.037 & 0.0080\\
\hline
\end{tabular}
\caption{Relative transition strength, $|\tilde {\cal T}_{n,j}|^2$, for a system of size $R_0=10^4 a_0$ in a Bessel beam with OAM $\ell=0$ and $\ell=1$ and in-plane photon momentum $q_{\|}=10^{-3}a_0^{-1}$. For comparison, we also provide the results from 2D Elliott formula for an infinite system in a Gaussian beam.}
\label{tab:1}
\end{table}
We note that the numerical evaluation also yields small but finite values for transitions into $p$-states. However, in contrast to the values for the transitions into $s$-states, these values show a strong and non-monotonic dependence on the system size and/or photon momentum. This suggests that, in accordance with our general arguments presented in Sec. II, the finite transition amplitudes into $p$-states are numerical artifacts, and the only bright transitions occur into the $s$ states.
Our numerical evaluation also confirms the selection rule that the COM momentum of the exciton is determined by the linear in-plane momentum of the photon. To this end, we focus on the 1$s$ transition and evaluate the transition strengths ${\cal T}_{0,0}^{\ell,N}$ into the different COM modes $k_{\ell,N}$. The results, normalized by the peak value ${\rm max}_N ( {\cal T}_{0,0}^{\ell,N} )$, are shown in Fig. \ref{fig:results} for different values of photon OAM $\ell$ and photon momentum. The transition strength is clearly peaked for the COM momenta which match with the momentum of the photon, but barely depends on the OAM.
\begin{figure}[t]
\centering
\includegraphics[scale=0.25]{fig2.eps}
\caption{For the 1$s$ transition, we plot the transition strength ${\cal T}_{0,0}^{\ell,N}$ into the different COM modes $k_{\ell,N}$, normalized by the peak value ${\rm max}_N ( {\cal T}_{0,0}^{\ell,N} )$, for illumination with $\ell=0$ and $\ell=1$ Bessel beams. The peak is obtained for the best match between COM momentum $k_{\ell,N}$ and in-plane photon momentum, $q_{\|}=10^{-3}a_0^{-1}$.
\label{fig:results}}
\end{figure}
Let us finally discuss the different length scales which appear in the calculation, that is, the effective Bohr radius $a_0$, the sample size $R_0$, and the inverse of the photon momentum $q_{\|}^{-1}$. In the calculation, we have taken the effective Bohr radius $a_0$ as the unit of length.
With typical values of the dielectric constant being much greater than 1 (e.g. $\approx 13$ in GaAs\cite{Strazalkowski76} and 7 in semiconducting TMDs\cite{Laturia18}), and the effective mass being much smaller than the electron mass (e.g. $0.067$ and $0.39$ electron masses for the conduction band in GaAs\cite{Cardona61} and model TMDs\cite{Kormanyos15}, respectively), the effective Bohr radius can significantly exceed the Bohr radius of the hydrogen atom ($\approx 0.05$nm). Typical values range between 0.1 to 1 nm. Taking the numerical constraints into account (i.e. truncation errors), our study examines sample sizes $R_0$ on the order of $10^4$ effective Bohr radii which corresponds to sample sizes on the order of 1-10 microns. Importantly, this size is significantly larger than the optical vortex. Regarding the in-plane photon momentum $q_{\|}$, an upper limit is given by
the inverse of the wavelength, $2\pi/\lambda_0$, assuming vertical incidence on the sample. The wave length $\lambda_0$ is determined by the band gap of the material. As an estimate for this limit, we obtain $100{\rm nm}^{-1}$. Thus, our choice of $q_{\|}=10^{-3}a_0^{-1}$ corresponds to the upper limit if $a_0=0.1$ nm, while this choice remains below that limit if $a_0$ is larger.
\section{Summary and Conclusions}
We have shown that the vector potential selects the COM quantum numbers (absolute value of COM momentum + COM angular momentum), but has no effect on the transition amplitudes into states with different relative quantum numbers $n$ and $j$. This implies that Elliott's formula is unchanged by the structure in the light field. The approximation which gives rise to these conclusion is the separation of length scales: unit cell vs. wavelength. This assumption implies that $A(x)$ and $e^{iqx}$ are constant on the the level of a unit cell, and for the dipole moment, $p_{{\bf k},{\bf k}+{\bf q}}^{\rm vc} \approx p_{{\bf k},{\bf k}}^{\rm vc}$. We have confirmed our general analytical result by performing numerical evaluations for the concrete case of Bessel beams in a circularly symmetric sample. Qualitatively, we have shown that, for a transition to be optically bright, the sum of both relative and COM angular momenta, $j+J$, must be equal to the OAM value $\ell$ of the light. Quantitatively, we have evaluated that the transition amplitudes are given by the Elliott formula.
While our results rule out twisted light for the generation of dark excitons, the predicted transfer of OAM to the COM degree of freedom can be useful from the perspective of quantum simulation, and especially from the point of view of artificial gauge fields. In Ref.~\cite{joerg20}, it has been shown that artificial flux is generated in a photonic system when OAM light is injected into a waveguide lattice. Excitons in tunable lattices have recently be shown to form strongly correlated many-body phases, such as Mott insulating phases \cite{lagoin21} or checkerboard phases \cite{lagoin22}. If, in the future, twisted light provided excitonic lattices with artificial magnetic fluxes, this could give rise to chiral Mott insulators \cite{dhar12} or extended supersolid regimes \cite{suthar20}.
\acknowledgments{
We thank Bin Cao, Valentin Kasper, Maciej Lewenstein, Andrey Grankin, Sunil Mittal, Jay Sau, Deric Session, Glenn Solomon, Daniel Suarez, and Jon Vanucci for thoughtful comments and fruitful discussions.
T.G. and U.B. acknowledge funding from “la Caixa” Foundation (ID 100010434, fellowship code LCF/BQ/PI19/11690013), the European Research Council for ERC Advanced Grant NOQIA; Agencia Estatal de
Investigación (R\&D project CEX2019-000910-S, funded by
MCIN/AEI/10.13039/501100011033, Plan National FIDEUA PID2019-106901GB-I00, FPI, QUANTERA MAQS PCI2019-111828-2, Proyectos de I+D+I “Retos Colaboración”
RTC2019-007196-7); Fundació Cellex; Fundació Mir-Puig; Generalitat de Catalunya through the CERCA program, AGAUR Grant No. 2017 SGR 134, QuantumCAT U16-011424, co-funded by ERDF Operational Program of Catalonia 2014-2020; EU Horizon 2020 FET-OPEN OPTOLogic (Grant No 899794); National Science Centre, Poland (Symfonia Grant No. 2016/20/W/ST4/00314); Marie Sk\l odowska-Curie grant STREDCH No 101029393; “La Caixa” Junior Leaders fellowships (ID100010434) and EU Horizon 2020 under Marie Sk\l odowska-Curie grant agreement No 847648 (LCF/BQ/PI20/11760031, LCF/BQ/PR20/11770012, LCF/BQ/PR21/11840013).
J.S. acknowledges support from the NSF Graduate Research Fellowship Program (GRFP) and the ARCS Foundation. J.S. and M.H. were supported by AFOSR FA95502010223, NSF PHY1820938, and NSF DMR-2019444, ARL W911NF1920181, ARO W911NF2010232, Simons and Minta Martin Foundations.
}
\onecolumngrid
|
\section{Introduction}
Experiments all over the world are attempting to observe neutrinoless
double-beta ($0\nu\beta\beta$) decay, which occurs only if neutrinos are
Majorana particles, at a rate that is related to neutrino masses. To learn
anything quantitative from an observed decay rate, one must know a nuclear
matrix element that cannot be measured independently and so must be computed
\cite{RevModPhys.80.481,0034-4885-80-4-046301,EJIRI20191,universe6120225}. Such
computations, which must handle the exchange of a virtual neutrino among
nucleons and mesons, are difficult, and so the matrix elements for isotopes used
in experiments are not known with high precision. A related process,
two-neutrino double-beta ($2\nu\beta\beta$) decay, has been observed, however,
and its rates are often a part of attempts to reduce the uncertainty in
$0\nu\beta\beta$ nuclear matrix elements.
The $2\nu\beta\beta$ nuclear matrix elements have been extracted from measured
half-lives in 11 nuclei at present \cite{universe.6.159}. To believe the
results of many-body computations of $0\nu\beta\beta$ decay, one would like to
see similar computations that reproduce these $2\nu\beta\beta$ matrix elements.
Because the closure approximation---replacing the energies of states in the
decay's intermediate nucleus with an average---is accurate for
$0\nu\beta\beta$, some approaches rely on it and have a harder time with
$2\nu\beta\beta$ decay, for which the approximation is poor. Nuclear density
functional theory (DFT) in combination with the proton-neutron
quasiparticle-random phase approximation (pnQRPA) is not one of these
approaches, however; $0\nu\beta\beta$ and $2\nu\beta\beta$ matrix elements can
be computed in similar ways. In many applications of the pnQRPA, in fact, the
$2\nu\beta\beta$ matrix element is used to constrain the strength of the piece
of the density functional associated with isoscalar proton-neutron pairing,
which suppresses both the $0\nu\beta\beta$ and $2\nu\beta\beta$ matrix elements
\cite{PhysRevLett.57.3148,PhysRevC.37.731}.
The pnQRPA can be used in conjunction with a phenomenological Hamiltonian as
well as in nuclear DFT. The advantage of DFT is its large single-particle model
space and universality; a single energy-density functional (EDF) is taken to
describe all the isotopes in the nuclear chart. The authors of Ref.\
\cite{PhysRevC.87.064302} used a deformed-nucleus pnQRPA with a Skyrme EDF,
computing $2\nu\beta\beta$ and $0\nu\beta\beta$ matrix elements in several
experimentally important isotopes. They fit the strength of the isoscalar
pairing interaction, on which rates depend sensitively, and it obtains the correct
$2\nu\beta\beta$ matrix elements before computing the $0\nu\beta\beta$ matrix
elements.
Nuclear EDFs are commonly optimized to reproduce a number of experimental
observables from a wide range of nuclei
\cite{Bogner20132235,PhysRevC.82.024313,PhysRevC.85.024304,
PhysRevC.89.054314,0954-3899-42-3-034024,PhysRevC.79.034310}. The optimization
is easiest when the observables are ground-state expectation values in even-even
isotopes. The time-odd part of the EDF
and the proton-neutron pairing strength have no effect on even-even ground states, however, and so
cannot be fixed in the same way. Instead, they are usually optimized globally,
through the use of single-$\beta$-decay rates and Gamow-Teller and spin-dipole
giant-resonance energies in nuclei all over the table of isotopes
\cite{PhysRevC.93.014304}.
We would like to assess the ability of these globally determined EDFs to
reproduce 2$\nu\beta\beta$ matrix elements so that we can further optimize them if necessary and then confidently apply them to $0\nu\beta\beta$ decay.
Standard pnQRPA calculations, however, require the construction of a QRPA matrix
Hamiltonian, which within large single-particle spaces can consume too much
computational time and memory. We can turn instead to the finite amplitude
method (FAM) within time-dependent DFT. The FAM, which is formally equivalent
to the QRPA \cite{nakatsukasa:024318,PhysRevC.84.014314}, computes the linear
response induced by an external field with a complex frequency. One-body
induced fields and the response of quasiparticle states are calculated by
iteration, without the need to compute the two-body QRPA matrix elements. The
proton-neutron version of the FAM (pnFAM) was developed and implemented in
Ref.~\cite{PhysRevC.90.024308} in order to calculate $\beta$-decay rates and
Gamow-Teller strength distributions
\cite{PhysRevC.94.055802,PhysRevC.102.034326}. Because of its efficiency, it
was used in Ref.\ \cite{PhysRevC.93.014304} in an attempt to optimize the
neutron-proton part of a particular nuclear EDF.
In this paper we show how to use the pnFAM to efficiently compute
$2\nu\beta\beta$ nuclear matrix elements. Our procedure, a preliminary version
of which was reported on in Ref.~\cite{doi:10.1063/1.5130971}, employs a
complex-plane integration technique \cite{PhysRevC.87.064309,PhysRevC.91.044323}
to perform the summation over intermediate states. We compare our
$2\nu\beta\beta$ nuclear matrix elements in $^{76}$Ge, $^{130}$Te, $^{136}$Xe,
and $^{150}$Nd to those obtained from matrix diagonalization with the same EDF
in Ref.\ \cite{PhysRevC.87.064302}. Then we use the EDFs with time-odd terms
fit in Ref.\ \cite{PhysRevC.93.014304} to compute the matrix elements for all 11
nuclei in which the $2\nu\beta\beta$ decay rate has been measured, and for 27 nuclei in which it has not.
The rest of this paper is organized as follows: Section~\ref{sec:qrpa} briefly
presents the definition of the $2\nu\beta\beta$ matrix element and describes the
pnQRPA. Section~\ref{sec:fam} formulates our scheme for computing
$2\nu\beta\beta$ nuclear matrix elements in the pnFAM. Section~\ref{sec:unc}
compares the pnFAM $2\nu\beta\beta$ matrix elements with those obtained by
matrix diagonalization in the pnQRPA, and Sec.\ \ref{sec:global} assesses the
performance of globally fit functionals and offers predictions for unmeasured
rates. Section~\ref{sec:conclusion} is a conclusion.
\section{\texorpdfstring{$2\nu\beta\beta$}{2nbb} matrix element and the QRPA\label{sec:qrpa}}
\subsection{\texorpdfstring{$2\nu\beta\beta$}{2nbb} matrix element}
The nuclear matrix element governing the $2\nu\beta\beta$ decay of the nucleus
$(N,Z)$ to the ground state of the nucleus $(N-2,Z+2)$ contributes to the
half-life $T^{2\nu}_{1/2}$ as follows:
\begin{equation}
\label{eq:T2nu}
[T^{2\nu}_{1/2}]^{-1} = G_{2\nu}(Q_{\beta\beta},Z) |M^{2\nu}|^2 \,,
\end{equation}
where $G_{2\nu}$ is a phase space factor, and the 2$\nu\beta\beta$ matrix
element is a sum of Fermi and Gamow-Teller parts \cite{RevModPhys.80.481},
\begin{align}
M^{2\nu} &= M^{2\nu}_{\rm GT} - \frac{g_V^2}{g_A^2} M^{2\nu}_{\rm F}, \\
M^{2\nu}_{\rm F}&= \sum_n \frac{ \displaystyle\langle 0^+_f|\sum_a \tau^-_a |n\rangle
\langle n|\sum_b \tau^-_b |0^+_i\rangle}
{\displaystyle E_n - \frac{M_i+M_f}{2}}, \\
M^{2\nu}_{\rm GT}&= \sum_n \frac{ \displaystyle\langle 0^+_f|\sum_a \bm{\sigma}_a \tau^-_a |n\rangle
\cdot \langle n|\sum_b \bm{\sigma}_b \tau^-_b |0^+_i\rangle}
{\displaystyle E_n - \frac{M_i+M_f}{2}}.
\label{eq:2nGT}
\end{align}
Here $\tau^-_a$ is the isospin-lowering operator for nucleon $a$,
$\bm{\sigma}_a$ is the corresponding spin operator, $M_i$ and $M_f$ are the
ground-state energies of the initial and final states of the decay, and $|n\rangle$, with energy $E_n$ is one of a
complete set of intermediate states in the nucleus $(N-1,Z+1)$. The Fermi part
of the $2\nu\beta\beta$ matrix element is very small because isospin is nearly
conserved \cite{PhysRevC.87.045501}, and we neglect it here.
\subsection{The pnQRPA}
The proton-neutron QRPA evaluates the transition matrix elements between the
initial or final state and the intermediate states that appear in the numerator
of Eq.~(\ref{eq:2nGT}), taking into account the effect of the proton-neutron
residual interaction beyond the mean-field approximation. In the pnQRPA, both
the initial and final states $|0_{i/f, {\rm QRPA}}^+\rangle$ are based on
Hartree-Fock-Bogoliubov (HFB) quasiparticle vacua, which incorporate
axially-symmetric deformation in our work. The intermediate states are related
to the initial or final state by a QRPA phonon operator
\begin{equation}
\begin{aligned}
|\lambda,K\rangle &=
\hat{\cal Q}^{\lambda\dag}_{K}
|0^+_{{\rm QRPA}}\rangle \\
\hat{\cal Q}^{\lambda\dag}_{K}
&= \!\!\!\!\!\! \sum_{\substack{pn \\ j_{z,p}+j_{z,n}=K}}
\!\!\!\! \!\! X_{pn,K}^{\lambda} \hat{a}^{\dag}_{p} \hat{a}^{\dag}_{n}
- Y_{pn,K}^{\lambda}
\hat{a}_{\bar{n}}
\hat{a}_{\bar{p}} \,,
\end{aligned}
\end{equation}
where $\hat{a}_{\tau=n,p}$ is a neutron or proton quasiparticle operator,
defined so that $\hat{a}_{\tau}|0_{\rm HFB}^+\rangle = 0$. Here, the indices
$p$ and $n$ label proton and neutron quasiparticles. $j_{z,\tau}$ and $K$ are
the projections along the symmetry axis of the quasiparticle and phonon angular
momentum, and the index $\bar{\tau}$ labels the time-reversal partner of the
state $\tau$ ($j_{z,\bar{\tau}} = -j_{z,\tau}$). From now on, for the sake of
simplicity we omit the restriction $j_{z,p}+j_{z,n}=K$ when summing over the
proton and neutron quasiparticle states.
The QRPA amplitudes $X_{pn,K}^\lambda$ and $Y_{pn,K}^\lambda$ are solutions of
the QRPA equations,
\begin{align}
\sum_{p'n'}
\begin{pmatrix} A_{pn,p'n'} & B_{pn,p'n'} \\ B^{\ast}_{pn,p'n'} & A^{\ast}_{pn,p'n'} \end{pmatrix}
\begin{pmatrix} X^{\lambda}_{p'n',K} \\ Y^{\lambda}_{p'n',K} \end{pmatrix}
=
\Omega^{\lambda}_{K}
\begin{pmatrix} X^{\lambda}_{pn,K} \\ -Y^{\lambda}_{pn,K} \end{pmatrix},
\end{align}
where $\Omega^{\lambda}_{K}$ is an excitation energy, measured from the QRPA
ground state of the initial/final state. The $A$ and $B$ matrices contain
residual interactions, computed from the second functional derivative of the
EDF. The $2\nu\beta\beta$ matrix element can be calculated by combining the
pnQRPA transition matrix elements from the initial and final states of the decay
to the intermediate states. Because the procedure introduces two sets of the
intermediate states, an additional approximation for matching them is necessary.
We thus approximate the Gamow-Teller matrix element in Eq.~(\ref{eq:2nGT}) by
\begin{widetext}
\begin{align}
M^{2\nu}_{\rm GT} &= \sum_{K=-1}^1 (-1)^K \sum_{\substack{\lambda_i>0 \\ \lambda_f>0}} \frac{\displaystyle
\langle 0_{f,{\rm QRPA}}^+| \hat{F}^{{\rm GT}-}_{-K} |\lambda_f,K\rangle
\langle \lambda_f,K|\lambda_i,K\rangle \langle \lambda_i,K|\hat{F}^{{\rm GT}-}_{K} |0_{i,{\rm QRPA}}^+\rangle}
{\displaystyle \frac{\Omega^{\lambda_i}_{K} + \Omega^{\lambda_f}_{K}}{2}}.
\label{eq:2nGTQRPA}
\end{align}
\end{widetext}
In the summation, the expression $\lambda>0$ denotes the states with
$\Omega_K^\lambda>0$.
The Gamow-Teller operator in the quasiparticle basis is
\begin{align}
\hat{F}^{{\rm GT}\pm}_{K} &= \sum_a (\sigma_K)_a \tau^\pm_a \nonumber\\
&= \sum_{pn} \left[F_{20,K}^{{\rm GT}\pm}(pn) \hat{a}^\dag_{p} \hat{a}^\dag_{n}
+ F_{02,K}^{{\rm GT}\pm}(pn) \hat{a}_{\bar{n}} \hat{a}_{\bar{p}}\right] \nonumber \\
&\quad + ( \hat{a}^\dag \hat{a}{\rm -terms}),
\end{align}
and its transition amplitudes in Eq. (\ref{eq:2nGTQRPA}) are given by
\begin{equation}
\begin{aligned}
\langle \lambda_i, K|
\hat{F}^{{\rm GT-}}_{K}|0^+_{i,{\rm QRPA}}\rangle=
& \\
\sum_{pn} \left[F^{{\rm GT-}}_{20,K}(pn) X^{\lambda_i\ast}_{pn,K}
\right.&\left.+ F^{{\rm GT-}}_{02,K}(pn) Y^{\lambda_i\ast}_{pn,K} \right] \\
\langle 0^+_{f,{\rm QRPA}}|
\hat{F}^{{\rm GT-}}_{-K}|\lambda_f,K\rangle=
& \\
\sum_{pn}\left[ F^{{\rm GT-}}_{02,K}(pn) X^{\lambda_f}_{pn,K}
\right.&\left.+ F^{{\rm GT-}}_{20,-K}(pn) Y^{\lambda_f}_{pn,-K}\right] \,.
\end{aligned}
\end{equation}
To compute the overlap of the two intermediate states $\langle
\lambda_f,K|\lambda_i,K\rangle$ we adapt expressions based on the QRPA
\cite{Simkovic2004321} and the quasiparticle Tamm-Dancoff approximation (QTDA)
\cite{PhysRevC.87.064302}. The result is
\begin{align}
\langle \lambda_f,K|\lambda_i,K\rangle
&= \sum_{pnp'n'}
\left( X_{p'n',K}^{\lambda_f\ast} X_{pn,K}^{\lambda_i}
- \alpha Y_{p'n',K}^{\lambda_f\ast} Y_{pn,K}^{\lambda_i}
\right) \nonumber \\
& \quad \times{\cal O}_{pp'}(\alpha) {\cal O}_{nn'}(\alpha)
\nonumber\\
&= \sum_{pn}
\left( \bar{X}_{pn,K}^{\lambda_f\ast} \bar{X}_{pn,K}^{\lambda_i}
- \alpha \bar{Y}_{pn,K}^{\lambda_f\ast} \bar{Y}_{pn,K}^{\lambda_i}
\right), \label{eq:overlap1}
\end{align}
where $\alpha$ is a parameter that is 0 for the QTDA overlap and 1 for the QRPA
overlap, and the ${\cal O}_{\tau\tau'}(\alpha)$ are elements of the matrix that
connect the quasiparticles associated with the initial and final states of the
decay. Explicit expressions for these elements, together with the derivation of
Eq.~(\ref{eq:overlap1}), are in Appendix~\ref{sec:overlap}. $\bar{X}$ and
$\bar{Y}$ are defined by
\begin{subequations}\label{eq:XbarYbar}
\begin{align}
\bar{X}_{pn,K}^{\lambda_i} &=\sum_{p'} {\cal O}^{T}_{pp'}(\alpha)
X_{p'n,K}^{\lambda_i},\\
\bar{Y}_{pn,K}^{\lambda_i} &= \sum_{p'} {\cal O}^{T}_{pp'}(\alpha) Y_{p'n,K}^{\lambda_i},\\
\bar{X}_{pn,K}^{\lambda_f} &= \sum_{n'} X_{pn',K}^{\lambda_f}{\cal O}^{T}_{n'n}(\alpha), \\
\bar{Y}_{pn,K}^{\lambda_f} &= \sum_{n'} Y_{pn',K}^{\lambda_f}{\cal O}^{T}_{n'n}(\alpha).
\end{align}
\end{subequations}
\section{The FAM \label{sec:fam}}
\subsection{pnFAM}
The FAM is formally equivalent to the QRPA and enables us to compute DFT
response functions efficiently. A detailed formulation of the like-particle FAM
and the pnFAM in the presence of the pairing correlations appear, respectively,
in Refs.~\cite{PhysRevC.84.014314} and \cite{PhysRevC.90.024308}.
In the pnFAM, one applies a time-dependent external field of the form
\begin{align}
\hat{F}_{K}^T(t) = \eta( \hat{F}_{K}^T e^{i\omega t} + \hat{F}^{T\dag}_{K}
e^{-i\omega t}) \,,
\end{align}
with $\hat{F}^T_K$ a one-body proton-neutron excitation operator and $\omega$ a
complex frequency. The excitation operator induces oscillations of
quasiparticle annihilation operators (e.g., for neutrons) of the form
\begin{align}
\delta \hat{a}_n(t) &= \eta \sum_p \hat{a}^\dag_p \left[
X_{pn}(\omega,\hat{F}^T_K) e^{-i\omega t} + Y^\ast_{pn}(\omega,\hat{F}^T_K)
e^{i\omega t} \right].
\end{align}
Solving the time-dependent DFT equations results in the FAM amplitudes
$X_{pn}(\omega,\hat{F}^T_K)$ and $Y_{pn}(\omega,\hat{F}^T_K)$, which are related
to the QRPA amplitudes $X^\lambda_{pn,K}$ and $Y^\lambda_{pn,K}$ through
\cite{PhysRevC.87.064309}
\begin{align}
X_{pn}(\omega,\hat{F}_{K}^T)&= - \sum_{\lambda>0} \left\{
\frac{X_{pn,K}^{\lambda} \langle \lambda,K|\hat{F}_{K}^T|0^+\rangle}
{\Omega^\lambda_K- \omega} \right. \nonumber \\
&\quad + \left. \frac{ Y_{pn,K}^{\lambda\ast}\langle
0^+|\hat{F}_{K}^T|\lambda,-K\rangle} {\Omega^\lambda_K+\omega}
\right\}, \label{eq:FAMX}\\
Y_{pn}(\omega,\hat{F}_{K}^T)&= - \sum_{\lambda>0} \left\{
\frac{Y_{pn,K}^{\lambda} \langle \lambda,K|\hat{F}_{K}^T|0^+\rangle}
{\Omega^\lambda_K- \omega} \right.\nonumber \\
&\quad + \left. \frac{ X_{pn,K}^{\lambda\ast}\langle
0^+|\hat{F}_{K}^T|\lambda,-K\rangle} {\Omega^\lambda_K+\omega}
\right\}. \label{eq:FAMY}
\end{align}
\subsection{\texorpdfstring{$2\nu\beta\beta$}{2nbb} matrix elements in the pnFAM}
To calculate the QRPA $2\nu\beta\beta$ nuclear matrix element in Eq.\
\eqref{eq:2nGTQRPA}, we separately solve the pnFAM computations in the initial
and final nuclei, distinguishing quantities from the two nuclei with the
superscripts $(i)$ and $(f)$. We then compute a quantity that is a combination
of the two sets of pnFAM amplitudes
\begin{align}
{\cal T}(\alpha;\omega_i, \hat{F}_{K_i}^{T_i}; \omega_f, \hat{F}_{K_f}^{T_f})
& \equiv \sum_{pn} \left[
\bar{Y}^{(f)}_{pn}(\omega_f,\hat{F}_{K_f}^{T_f})
\bar{X}^{(i)}_{pn}(\omega_i,\hat{F}_{K_i}^{T_i})
\right. \nonumber \\
& \quad\left. -\alpha \bar{X}^{(f)}_{pn}(\omega_f,\hat{F}_{K_f}^{T_f})
\bar{Y}^{(i)}_{pn}(\omega_i,\hat{F}_{K_i}^{T_i}) \right],
\label{eq:Tfunc}
\end{align}
\begin{figure}[t]
\includegraphics[width=\columnwidth]{ocontour.pdf}
\caption{Contours $C_i$ and $C_f$. \label{fig:contour}}
\end{figure}
where $\bar{X}^{(i/f)}$ and $\bar{Y}^{(i/f)}$ are the amplitudes in
Eqs.~(\ref{eq:FAMX}) and (\ref{eq:FAMY}), labeled in the same way as the QRPA
amplitudes in Eq.~(\ref{eq:XbarYbar}). By substituting Eqs.~(\ref{eq:FAMX}) and
(\ref{eq:FAMY}) into Eq.~(\ref{eq:Tfunc}), we obtain an expression for
$\mathcal{T}$ in terms of the QRPA amplitudes:
\begin{widetext}
\begin{align}
{\cal T}(\alpha;\omega_i,\hat{F}_{K_i}^{T_i}; \omega_f, \hat{F}_{K_f}^{T_f})
&=
\sum_{pn}
\sum_{\substack{\lambda_i>0 \\ \lambda_f>0}} \left[
\frac{ (
\bar{X}^{\lambda_f\ast}_{pn,K_f} \bar{X}^{\lambda_i}_{pn,K_i} - \alpha \bar{Y}^{\lambda_f\ast}_{pn,K_f} \bar{Y}^{\lambda_i}_{pn,K_i})
\langle 0_f^+|\hat{F}_{K_f}^{T_f}|\lambda_f,-K_f\rangle \langle \lambda_i,K_i|\hat{F}_{K_i}^{T_i}|0_i^+\rangle}
{ (\Omega^{\lambda_f}_{K_f} + \omega_f)(\Omega^{\lambda_i}_{K_i} - \omega_i)} \right. \nonumber \\
&\quad +
\frac{ (
\bar{Y}^{\lambda_f\ast}_{pn,K_f}
\bar{X}^{\lambda_i}_{pn,K_i} - \alpha
\bar{X}^{\lambda_f\ast}_{pn,K_f}
\bar{Y}^{\lambda_i}_{pn,K_i})
\langle \lambda_f,K_f|\hat{F}_{K_f}^{T_f}|0_f^+\rangle \langle \lambda_i,K_i|\hat{F}_{K_i}^{T_i}|0_i^+\rangle}
{ (\Omega^{\lambda_f}_{K_f} - \omega_f)(\Omega^{\lambda_i}_{K_i} - \omega_i)} \nonumber \\
&\quad +
\frac{ (
\bar{X}^{\lambda_f\ast}_{pn,K_f}
\bar{Y}^{\lambda_i}_{pn,K_i} - \alpha \bar{Y}^{\lambda_f\ast}_{pn,K_f}
\bar{X}^{\lambda_i}_{pn,K_i})
\langle 0_f^+|\hat{F}_{K_f}^{T_f}|\lambda_f,-K_f\rangle \langle 0_i^+|\hat{F}_{K_i}^{T_i}|\lambda_i,-K_i\rangle}
{ (\Omega^{\lambda_f}_{K_f} + \omega_f)(\Omega^{\lambda_i}_{K_i} + \omega_i)} \nonumber \\
&\quad + \left.
\frac{ (
\bar{Y}^{\lambda_f\ast}_{pn,K_f}
\bar{Y}^{\lambda_i}_{pn,K_i} - \alpha \bar{X}^{\lambda_f\ast}_{pn,K_f}
\bar{X}^{\lambda_i}_{pn,K_i})
\langle \lambda_f,K_f|\hat{F}_{K_f}^{T_f}|0_f^+\rangle \langle 0_i^+|\hat{F}_{K_i}^{T_i}|\lambda_i,-K_i\rangle}
{ (\Omega^{\lambda_f}_{K_f} - \omega_f)(\Omega^{\lambda_i}_{K_i} + \omega_i)} \right]. \label{eq:T}
\end{align}
\end{widetext}
${\cal T}$ has first-order poles at $\omega_i = \pm\Omega^{\lambda_i}_{K_i}$ and
$\omega_f=\pm\Omega^{\lambda_f}_{K_f}$. We choose a counterclockwise contour
$C_i$ ($C_f$) for $\omega_i$ ($\omega_f$) that includes positive-energy
(negative-energy) poles and excludes all the negative-energy (positive-energy)
poles, as in Fig.~\ref{fig:contour}, to extract the residues from the first term
on the right side of Eq.~(\ref{eq:T}).
Integrating $\mathcal{T}$ together with an arbitrary regular complex function of
$\omega_i$ and $\omega_f$ over those frequencies, we obtain an expression that
can be used for two-body matrix elements:
\begin{align}
&{\cal M}\bm{\bigl(}\alpha; \hat{F}_{K_i}^{T_i}, \hat{F}_{K_f}^{T_f}; f(\omega_i,\omega_f)\bm{\bigr)} \nonumber \\
&=-\frac{1}{4\pi^2}
\oint_{C_i} d\omega_i \oint_{C_f} d\omega_f {\cal T}(\alpha;\omega_i,\hat{F}_{K_i}^{T_i};\omega_f,\hat{F}_{K_f}^{T_f})
f(\omega_i,\omega_f) \nonumber \\
&= \sum_{\substack{\lambda_i>0\\ \lambda_f>0}}
\sum_{pn}
\left(\bar{X}^{\lambda_f\ast}_{pn,K_f} \bar{X}^{\lambda_i}_{pn,K_i} - \alpha \bar{Y}^{\lambda_f\ast}_{pn,K_f} \bar{Y}^{\lambda_i}_{pn,K_i} \right) \nonumber \\
&\quad \times
f(\Omega^{\lambda_i}_{K_i},-\Omega^{\lambda_f}_{-K_f})
\langle 0_f^+|\hat{F}_{K_f}^{T_f}|\lambda_f,-K_f\rangle \langle \lambda_i,K_i|\hat{F}_{K_i}^{T_i}|0_i^+\rangle.
\label{eq:M}
\end{align}
The Fermi and Gamow-Teller $2\nu\beta\beta$ decay nuclear matrix elements are
then given by
\begin{align}
M^{2\nu}_{{\rm F}} &= {\cal M}\left(\alpha; \hat{F}^{{\rm F-}},\hat{F}^{{\rm F-}}; f=\frac{2}{\omega_i - \omega_f}\right), \\
M^{2\nu}_{{\rm GT}} &= \sum_{K=-1}^1 (-1)^K {\cal M}\left(\alpha; \hat{F}_K^{{\rm GT-}},\hat{F}_{-K}^{{\rm GT-}}; f=\frac{2}{\omega_i - \omega_f}\right),
\end{align}
under the assumptions that $X^{\lambda_f}_{-K}=X^{\lambda_f}_K$,
$Y^{\lambda_f}_{-K}=Y^{\lambda_f}_{K}$, and $\Omega^{\lambda_f}_{-K_f} =
\Omega^{\lambda_f}_{K_f}$. Even when starting from the final state we use the
external operator $\sigma_{-K}\tau^-$ that changes neutrons into protons to
properly include the backward amplitudes in Eqs.~(\ref{eq:FAMX}) and
(\ref{eq:FAMY}).
By setting $f=1$, $\alpha=1$, and taking the same HFB vacuum for the initial and
final states in Eq.~(\ref{eq:M}), we can use that equation to compute the
unweighted summed strengths:
\begin{align}
{\cal M}^{i=f}(1;\hat{F}^{\rm F\mp}, \hat{F}^{\rm F\pm}; 1)
&= \sum_{\lambda>0} |\langle \lambda,0|\hat{F}^{\rm F\mp}|0^+\rangle|^2, \\
{\cal M}^{i=f}(1; \hat{F}^{\rm GT\mp}_{K}, \hat{F}^{\rm GT\pm}_{-K};1)
&= (-1)^K \nonumber \\
&\quad \times \sum_{\lambda>0} |\langle \lambda,K|\hat{F}^{{\rm GT\mp}}_{K}|0^+\rangle|^2.
\end{align}
Sum rules can be used to check the routines that compute matrix elements.
\section{Results with SkM* and Comparison with Prior Work \label{sec:unc}}
Our calculation of $2\nu\beta\beta$ nuclear matrix elements uses an extension of
the pnFAM code developed in Ref.~\cite{PhysRevC.90.024308}, which is in turn
based on the nuclear DFT solver {\sc hfbtho}
\cite{PEREZ2017363,Stoitsov20131592,Stoitsov200543}. That last code uses the
harmonic oscillator basis in a cylindrical coordinate system and allows axial
deformation. In this section we provide details of our calculations with the
SkM* functional and compare our $2\nu\beta\beta$ matrix elements for $^{76}$Ge,
$^{130}$Te, $^{136}$Xe, and $^{150}$Nd with those obtained in
Ref.~\cite{PhysRevC.87.064302} by diagonalizing the pnQRPA matrix.
\subsection{Parameter values}
\begin{table}[t]
\caption{Experimental values of $\tilde{\Delta}_n^{(3)}$ and
$\tilde{\Delta}_p^{(3)}$ (in MeV) and the volume-pairing strengths $V_n$ and
$V_p$ fit to those values (in MeV fm$^3$). The averages of the strengths in
the initial and final nuclei used in the pairing EDF. Experimental binding
energies are taken from Ref.~\cite{ame2020}. Data in parentheses are not used
to fit the pairing strengths.
\label{table:gap}}
\begin{ruledtabular}
\begin{tabular}{ccccc}
& $\tilde{\Delta}_n^{(3)}$ & $\tilde{\Delta}_p^{(3)}$ & $V_n$ & $V_p$
\\ \hline
$^{76}$Ge & 1.393 & 1.114 & $-$182.70 & $-$194.49 \\
$^{76}$Se & 1.551 & 1.392 & $-$185.40 & $-$202.22 \\
average && & $-$184.05 & $-$198.36 \\ \hline
$^{130}$Te& 1.114 & (0.801)& $-$166.21 & N/A \\
$^{130}$Xe& 1.170 & 1.016 &$-$173.80 & $-$194.00 \\
average & & & $-$170.01 & $-$194.00 \\ \hline
$^{136}$Xe& (0.841) & 0.751 & N/A & $-$148.66 \\
$^{136}$Ba& 0.960 & 1.005 &$-$184.16 & $-$172.54 \\
average & & & $-$184.16 & $-$160.60 \\ \hline
$^{150}$Nd& 1.070 & 0.918 & $-$181.64 & $-$202.31 \\
$^{150}$Sm& 1.194 & 1.196 & $-$184.84 & $-$195.24 \\
average & & & $-$183.24 & $-$198.78
\end{tabular}
\end{ruledtabular}
\end{table}
To integrate in cylindrical coordinates, we use Gauss-Hermite quadrature with
$N_{\rm GH}=40$ points for the $z$ direction and Gauss-Laguerre quadrature with
$N_{\rm GL}=40$ points for the $r$ direction. To compute the direct Coulomb
mean field, we use the prescription described in Ref.~\cite{Stoitsov20131592}
with length parameter $L=50$ fm and $N_{\rm Leg}=80$ Gauss-Legendre points.
We include $N_{\rm sh}=20$ harmonic-oscillator major shells to describe the HFB
wave functions. This corresponds to 1771 single-particle states for neutrons
and protons (with axial and time-reversal symmetry taken into account), and, in
the pnFAM, to 257 686 $K=0$ two-quasiparticle states and 256 025 $K=\pm 1$
two-quasiparticle states. We include all such states, with no additional
model-space truncation, in the pnFAM calculations. The dimension of the pnQRPA
matrix corresponding to our pnFAM calculations is about 500 000 for each $K$
quantum number.
We employ the same Skyrme SkM* functional \cite{Bartel198279} and volume-type
pairing with 60-MeV energy cutoff (with $\hbar^2/2m=20.73$ MeV fm$^2$ and the
one-body center-of-mass correction included in the kinetic term) as that in
Ref.\ \cite{PhysRevC.87.064302}. The HFB solver cited in that paper, however,
works in a cylindrical box with $r_{\rm max}=z_{\rm max}=20$ fm and a coordinate
spacing of 0.7 fm, and is thus different from ours.
\begin{table}[b]
\caption{Properties of HFB ground states with the SkM* + volume pairing (with
average pairing strengths) EDF. The table shows pairing gaps (in MeV),
quadrupole deformation, and total HFB energies (in MeV) and compares the
quadrupole deformation to the value in Ref.~\cite{PhysRevC.87.064302}.
\label{table:dft}}
\begin{ruledtabular}
\begin{tabular}{cccccc}
& \textrm{$\Delta_n$} & \textrm{$\Delta_p$} & \textrm{$\beta$} &
\textrm{$E_{\rm HFB}$} & \textrm{$\beta$ (Ref \cite{PhysRevC.87.064302})}
\\
\colrule $^{76}$Ge & 1.609 & 1.473 & $-$0.021 & $-$661.804 & $-$0.025 \\
& 1.435 & 1.205 & 0.185 & $-$662.274 & 0.184 \\
& 1.612 & 1.475 & 0. & $-$661.802 & \\ \hline
$^{76}$Se & 1.589 & 1.648 & 0. & $-$659.315 & $-$0.018 \\
& 1.508 & 1.257 & $-$0.194 & $-$659.594 & \\ \hline
$^{130}$Te
& 1.178 & 1.028 & 0. & $-$1096.839 & 0.01 \\ \hline
$^{130}$Xe & 1.078 & 1.009 & 0.141 & $-$1093.423 & 0.13 \\
& 1.107 & 1.113 & $-$0.124 & $-$1093.152 & \\
& 1.359 & 1.351 & 0. & $-$1092.393 & \\ \hline
$^{136}$Xe
& 0. & 0.878 & 0. & $-$1143.253 & 0.004 \\ \hline
$^{136}$Ba & 1.025 & 0.931 & $-$0.047 & $-$1139.268 & $-$0.021\\
& 0.928 & 0.735 & 0.094 & $-$1139.538 & \\
& 1.057 & 0.985 & 0. & $-$1139.231 & \\ \hline
$^{150}$Nd & 1.129 & 0.764 & 0.292 & $-$1235.794 & 0.27 \\
& 1.375 & 1.358 & $-$0.177 & $-$1232.563 & \\
& 1.422 & 1.688 & 0. & $-$1231.080 & \\ \hline
$^{150}$Sm & 1.131 & 1.307 & 0.223 & $-$1234.675 & 0.22 \\
& 1.294 & 1.707 & 0. & $-$1232.436 & \\
& 1.305 & 1.534 & $-$0.137 & $-$1233.068 &
\end{tabular}
\end{ruledtabular}
\end{table}
We adjust the volume pairing strengths to reproduce experimental odd-even
staggering (OES) with the density-averaged pairing gap. To reduce fluctuations
\cite{PhysRevC.82.024313}, we take as the experimental data an average of the
results of the three-point formula evaluated at the two even-odd or odd-even
systems:
\begin{equation}
\begin{aligned}
\tilde{\Delta}^{(3)}_n(N,Z) &= \frac{ \Delta^{(3)}_n(N-1,Z) +
\Delta^{(3)}_n(N+1,Z)}{2} \\
\tilde{\Delta}^{(3)}_p(N,Z) &= \frac{\Delta^{(3)}_p(N,Z-1) +
\Delta^{(3)}_p(N,Z+1)}{2} \,,
\end{aligned}
\end{equation}
where $\Delta^{(3)}_{n/p}$ is the result of the three-point formula \cite{Bender2000}.
Table \ref{table:gap} lists the experimental values for this quantity and the
neutron and proton volume pairing strengths that best reproduce them. In order
to use the same EDF for both nuclei in the decay, we take the average of the
pairing strengths fit in the initial and final nuclei. We note that the
experimental $\Delta^{(3)}_{n/p}$ values do not provide useful information if
the series of isotopes used to calculate them includes closed-shell nuclei.
$\Delta^{(3)}_{p}$ in $^{130}$Te ($Z=50$ included) and $\Delta^{(3)}_{n}$ in
$^{136}$Xe and $^{136}$Ba ($N=82$ included) are such cases if the average of the
results of two odd-even mass formulas is used. We avoid using the pairing gap
$\tilde{\Delta}_p^{(3)}$ of $^{130}$Te to fit the proton pairing strength,
fitting the pairing strength instead to the proton gap in $^{130}$Xe. We do
adopt the neutron $\tilde{\Delta}^{(3)}$ of $^{136}$Ba, however, to determine
the neutron pairing strength because the strengths fit to $\Delta_n^{(3)}$ and
$\tilde{\Delta}_n^{(3)}$ are quite similar in that nucleus. The globally fit
EDFs described in Sec.~\ref{sec:global} are free from these problems.
Table \ref{table:dft} shows the results of the DFT calculations for the initial
and final nuclei. The quadrupole deformations of the HFB states are quite close
to those in Ref.~\cite{PhysRevC.87.064302}. We choose the HFB solution in the
top line for each nucleus in which several HFB solutions coexist.
\begin{table}[t]
\caption{$Q$ values for each double-beta decay in units of MeV. Experimental
$Q$ values are obtained from atomic masses \cite{ame2020}.
\label{table:qvalue}}
\begin{ruledtabular}
\begin{tabular}{ccccc}
& This paper & SkM* (Ref.~\cite{PhysRevC.87.064302}) & Exp. \\ \colrule
$^{76}$Ge $\to$ $^{76}$Se & 4.05 & 4.84 & 2.04 \\
$^{130}$Te $\to$ $^{130}$Xe & 4.98 & 4.22 & 2.53 \\
$^{136}$Xe $\to$ $^{136}$Ba & 5.55 & 5.60 & 2.46 \\
$^{150}$Nd $\to$ $^{150}$Sm & 2.68 & 2.35 & 3.37 \\
\end{tabular}
\end{ruledtabular}
\end{table}
Table~\ref{table:qvalue} shows $\beta\beta$ $Q$ values. Our calculation does
not perfectly reproduce the values in Ref.~\cite{PhysRevC.87.064302}, which were obtained
from the same SkM* EDF but a different HFB code. We suspect that the
differences are due to the different representations of the oscillator basis
states and treatments of pairing.
\begin{table}[b]
\caption{Neutron and proton parts of the HFB overlap $\langle 0^+_{f,{\rm
HFB}}|0^+_{i,{\rm HFB}}\rangle$ between the initial and the final states
compared with values from previous QRPA calculations.
\label{table:overlap}}
\begin{ruledtabular}
\begin{tabular}{cccccccccc}
& neutron & proton & total & Ref.~\cite{PhysRevC.87.064302} &
Ref.~\cite{PhysRevC.83.034320} & Ref.~\cite{PhysRevC.97.045503} \\ \colrule
$^{76}$Ge & 0.907 & 0.886 & 0.803 & & 0.81 & 0.72, 0.73\\
$^{130}$Te & 0.329 & 0.403 & 0.133 & & & 0.73, 0.73\\
$^{136}$Xe & 0.480 & 0.787 & 0.378 & 0.47 & & 0.43, 0.39\\
$^{150}$Nd & 0.679 & 0.589 & 0.400 & & 0.52 & 0.51, 0.52
\end{tabular}
\end{ruledtabular}
\end{table}
Table~\ref{table:overlap} lists the values of the HFB overlap included in the
matrix ${\cal O}$. Our values agree with those of previous calculations with
similar nuclear deformation. The overlap becomes small when the deformation of
the initial and final states are different. That situation arises in $^{130}$Te
and $^{136}$Xe, where the initial states are spherical while the final states
are prolate and oblate, respectively. The overlap also becomes small when the
initial or final state has no pairing gap. That is the case for neutrons in
$^{136}$Xe.
\subsection{Contour}
To use the expression in Eq.\ \eqref{eq:M}, we must choose the contours $C_i$
and $C_f$. We take each to be centered on the real axis and circular, with the
circle specified by the two energies $\omega_{\rm L}$ and $ \omega_{\rm R}$ at
they cross the real axis. The radius $r$ and the center of the contour
$\omega_0$ are then given by $r = (\omega_{\rm R} - \omega_{\rm L})/2$ and
$\omega_0 =(\omega_{\rm L} + \omega_{\rm R})/2$, and every point on the contour
can be written in the form $\omega = \omega_0 + r e^{i\theta}$. We use
$\omega_{\rm L} = 0.1$ MeV and $\omega_{\rm R}=120$ MeV for $C_i$ and
$\omega_{\rm L} = -120$ MeV and $\omega_{\rm R} = -0.1$ MeV for $C_f$ to include
all the unperturbed two-quasiparticle states within the quasiparticle-energy
cutoff.
\begin{figure}[b]
\includegraphics[width=\columnwidth]{76Geintegrand.pdf}
\caption{Integrand in computation of summed Gamow-Teller strength and
$2\nu\beta\beta$ matrix element for $^{76}$Ge as a function of the angle
$\theta$ for the contour $C_i$. The integrand is normalized to 1 at
$\theta=\pi$.}
\label{fig:integrand}
\end{figure}
Figure~\ref{fig:integrand} shows the integrands for the summed GT strength and
the $2\nu\beta\beta$ matrix element in $^{76}$Ge as a function of the angle
$\theta$ for the contour $C_i$. The main contribution to each comes from the
peak at $\theta=\pi$, the point at $\omega=\omega_{\rm L}$ where the two
contours are closest. While the integrand of the sum is distributed broadly
along the whole contour $C_i$, the energy denominator $2/(\omega_i-\omega_f)$
concentrates the $2\nu\beta\beta$ contribution at $\theta=\pi$. To take this
contribution into account precisely and efficiently, we introduce a parameter
$\gamma$ to control the distribution of the discretized points near the origin,
i.e., we discretize the angle $\theta$ as follows:
\begin{align}
\theta_k &= (m + x_k^\gamma) \pi,\\
x_k &= -1 + 2 \frac{k-1}{n_r - 1} \quad
(k = 1, 2, \cdots, n_r) \,,
\end{align}
where $\gamma$ is an odd number, and $m$ is 1 for $C_i$ and $0$ for $C_f$. The
parameter $\theta_k$ runs from 0 to $2\pi$ for $C_i$ and $-\pi$ to $\pi$ for
$C_f$. We use $n_r=202$ and omit the contribution from $(\omega_i,\omega_f) =
(\omega_{\rm R}, -\omega_{\rm L})$, because those points are on the real axis and can be very
close to the QRPA poles, although their contribution to the nuclear matrix
element should be small because of the factor $2/(\omega_i-\omega_f)$.
Table~\ref{table:gamma} shows the dependence of the summed strengths and the
$2\nu\beta\beta$ matrix element in $^{76}$Ge on the parameter $\gamma$. The
matrix element converges by $\gamma = 5$, which is the value we use.
\subsection{Summed Strengths}
Table~\ref{table:sumrule} shows the unweighted summed Fermi and Gamow-Teller
strengths obtained from the double contour integration for selected nuclei of
interest to experimentalists. Integration up to 120 MeV reproduces more than
99.9\% of the Ikeda sum rule in all these nuclei.
\begin{table}[t]
\caption{ Dependence on the discretization parameter $\gamma$ in $^{76}$Ge of
summed Fermi and Gamow-Teller strengths $m({\rm
F\pm})=\sum_{\lambda>0}|\langle\lambda,0|\hat{F}^{\rm F\pm}|0^+\rangle|^2$ and
$m({\rm GT\pm})=\sum_K(-1)^K\sum_{\lambda>0}|\langle\lambda,K|\hat{F}^{\rm
GT\pm}_K|0^+\rangle|^2$ and of the dimensionless Gamow-Teller $2\nu\beta\beta$
nuclear matrix element. We use volume like-particle pairing and no isoscalar
pairing.
\label{table:gamma}}
\begin{ruledtabular}
\begin{tabular}{ccccc}
$\gamma$ & \textrm{1} & \textrm{3} & \textrm{5} & \textrm{7}\\ \colrule
$m({\rm F-})$ & 12.0213 & 12.0209 & 12.0201 & 12.0189 \\
$m({\rm F+})$ & 0.0252 & 0.0255 & 0.0260 & 0.0269 \\
$m({\rm F-})-m({\rm F+})$ & 11.9961 & 11.9954 & 11.9940 & 11.9920 \\
\colrule $m({\rm GT-})$ & 37.5860 & 37.5837 & 37.5811 & 37.5774 \\
$m({\rm GT+})$ & 1.6065 & 1.6070 & 1.6085 & 1.6109 \\
$m({\rm GT-})-m({\rm GT+})$ & 35.9795 &35.9767 & 35.9726 & 35.9664 \\
\colrule $M^{2\nu} m_ec^2$ & 0.1802 & 0.1574 & 0.1574 & 0.1575
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{table}[b]
\caption{Summed Fermi and Gamow-Teller transitions, from double contour
integration ($\omega_{\rm L}=0.1$ MeV and $\omega_{\rm R}=120$ MeV,
$n_r=202$, and $\gamma=5$), as percentages of the corresponding sum rules.
\label{table:sumrule}}
\begin{ruledtabular}
\begin{tabular}{rcc}
& $\displaystyle\frac{m(\text{F}-)-m(\text{F}+)}{N-Z}$ & $\displaystyle\frac{m(\text{GT}-)-m(\text{GT}+)}{3(N-Z)}$\\ \colrule
$^{76}$Ge & 0.9995 & 0.9992 \\
$^{76}$Se & 0.9994 & 0.9992 \\
$^{130}$Te & 0.9996 & 0.9993 \\
$^{130}$Xe & 0.9996 & 0.9993 \\
$^{136}$Xe & 0.9998 & 0.9997 \\
$^{136}$Ba & 0.9996 & 0.9994 \\
$^{150}$Nd & 0.9996 & 0.9994 \\
$^{150}$Sm & 0.9996 & 0.9995
\end{tabular}
\end{ruledtabular}
\end{table}
\subsection{\texorpdfstring{$2\nu\beta\beta$}{2nbb} matrix element}
We calculate the $2\nu\beta\beta$ matrix elements for $^{76}$Ge, $^{130}$Te,
$^{136}$Xe, and $^{150}$Nd, setting the neutron-proton isovector pairing
strength to the average of the neutron and proton like-particle pairing
strengths [$V_1=(V_n+V_p)/2$] and varying the isoscalar pairing strength $V_0$
from 0 to $-300$ MeV fm$^3$. We use the QTDA ($\alpha=0$) to compute the
overlap among intermediate states. Figure~\ref{fig:2nbb_benchmark_aved3_r2sdmat}
displays the dependence of the $2\nu\beta\beta$ Gamow-Teller nuclear matrix
elements on the isoscalar pairing strength.
Like the authors of that
paper, we use two values of $g_A$: one ``unquenched'' ($g_A=1.25$, though the
currently accepted value is greater than 1.27) and one quenched ($g_A=1.0$), and
compare results for the EDF SkM* with and without a modified proton-neutron
piece [$C_1^s=100$ MeV fm$^3$, $C_1^T=C_1^{\nabla s} = 0$, see
Eq.~(\ref{eq:time-odd})]. Our matrix elements agree reasonably well with those
of Ref.~\cite{PhysRevC.87.064302} in $^{130}$Te, $^{136}$Xe, and $^{150}$Nd,
while they are about twice as large in $^{76}$Ge.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{unc_skms_2nbb.pdf}
\caption{ Dependence on the isoscalar pairing strength of the 2$\nu\beta\beta$
Gamow-Teller nuclear matrix element (in units of MeV$^{-1}$) for $^{76}$Ge,
$^{130}$Te, $^{136}$Xe, and $^{150}$Nd, with the SkM* + volume pairing EDF.
Red curves indicate the results with the time-odd functional derived from the
SkM* interaction and blue curves the results with the modified time-odd
functional. The thick solid and dotted curves correspond to $g_A=1.25$, and
the dashed and thin solid curves to the quenched value $g_A=1.0$.
\label{fig:2nbb_benchmark_aved3_r2sdmat}}
\end{figure}
\section{global edf \label{sec:global}}
\subsection{Performance of global EDFs}
As Fig.~\ref{fig:2nbb_benchmark_aved3_r2sdmat} shows, the $2\nu\beta\beta$ and
$0\nu\beta\beta$ nuclear matrix elements are suppressed by isoscalar
proton-neutron pairing correlations that cannot be constrained from the
ground-state properties of even-even nuclei. The usual QRPA approach uses
$2\nu\beta\beta$ decay rates to determine the strength of isoscalar pairing
separately in each decaying nucleus. The philosophy of nuclear DFT, however, is
that one EDF parameter set should, if possible, describe all the
$\beta\beta$-decaying nuclei in the nuclear chart. In this section we
assess the ability of globally-fit EDFs to describe $2\nu\beta\beta$ decay,
without using that observable at all in the fitting. We use Skyrme-type EDFs, with
the isovector time-odd and isoscalar pairing parts globally fit to
single-$\beta$ decay rates and to Gamow-Teller and spin-dipole resonances.
Reference~\cite{PhysRevC.93.014304}, which deals with single-$\beta$ decay in many
isotopes, undertakes the global fitting and proposes ten parameter sets, called
1A, 1B, 1C, 1D, 1E, 2, 3A, 3B, 4, and 5, each corresponding to a different EDF.
The time-even parts for all the sets except set 2 are taken from the functional
SkO$'$ \cite{PhysRevC.60.014316}; that of parameter set 2 comes from the
functional SV-min \cite{PhysRevC.79.034310} (though tensor-density terms are
neglected, and the neutron and proton have different masses). In both cases,
center-of-mass corrections to the mean field are neglected. The paper uses
mixed volume-surface like-particle isovector pairing terms, fit to reproduce
odd-even staggering in ten isotopes with $50\le A \le 250$; the strengths are
$V_n = -253.75$ MeV fm$^3$, $V_p = -274.68$ MeV fm$^3$ for SkO$'$
and $V_n = -244.06$ MeV fm$^3$, $V_p = -257.90$ MeV fm$^3$ for
SV-min \cite{mustonen-priv}.
\begin{table}[t]
\caption{ The neutron and proton pairing gaps (in MeV) and quadrupole
deformation for the lowest-energy HFB solutions in the initial and final nuclei
of the decay, computed with the SkO$'$ and SV-min EDFs, together with a mixed
pairing EDF. Solutions with parentheses are not the lowest-energy ones, but we
use them in addition when calculating matrix elements.
\label{table:hfb-global1}}
\begin{ruledtabular}
\begin{tabular}{ccccccc}
&& \textrm{SkO$'$} & & & \textrm{SV-min} & \\
& $\Delta_n$ & $\Delta_p$ & $\beta$ & $\Delta_n$ & $\Delta_p$ & $\beta$ \\
\colrule $^{48}$Ca & 0.771& 0.000& 0.000& 0.793& 0.000& 0.000 \\
$^{48}$Ti & 1.270& 1.386& 0.000& 1.275& 1.309& 0.000 \\ \colrule
$^{76}$Ge & 1.063& 1.189& 0.136& 1.123& 1.094& 0.131 \\
$^{76}$Se & 1.134& 1.532& 0.000& 1.165& 1.352& 0.000 \\ \colrule
$^{82}$Se & 0.619& 1.106& 0.152& 0.689& 1.124& 0.134 \\
$^{82}$Kr & 1.014& 1.353& 0.112& 1.041& 1.230& 0.101 \\ \colrule
$^{96}$Zr & 1.153& 1.133& $-$0.173& 1.041& 0.986& 0.000 \\
& (1.354 & 1.129 &0.000)& \\
$^{96}$Mo & 1.202& 1.174& 0.000& 0.991& 1.090& 0.000 \\ \colrule
$^{100}$Mo & 1.200& 1.089& $-$0.192& 1.299& 1.078& 0.000 \\
& (1.123& 1.246& 0.214)& \\
$^{100}$Ru & 0.994& 1.092& 0.186& 1.189& 1.137& 0.000 \\ \colrule
$^{116}$Cd & 1.430& 0.854& 0.000& 1.463& 0.492& 0.120 \\
$^{116}$Sn & 1.406& 0.000& 0.000& 1.553& 0.000& 0.000 \\ \colrule
$^{128}$Te & 1.139& 0.970& 0.000& 1.209& 0.907& 0.000 \\
$^{128}$Xe & 1.136& 0.912& 0.142& 1.152& 0.841& 0.156 \\
&(1.147& 1.064&$-$0.112)&(1.179& 0.986&$-$0.122) \\ \colrule
$^{130}$Te & 1.013& 0.971& 0.000& 1.043& 0.902& 0.000 \\
$^{130}$Xe & 1.051& 1.001& 0.111& 1.077& 0.947& 0.118 \\ \colrule
$^{136}$Xe & 0.000& 1.180& 0.000& 0.000& 1.143& 0.000 \\
$^{136}$Ba & 0.767& 1.349& 0.000& 0.775& 1.296& 0.000 \\ \colrule
$^{150}$Nd & 0.962& 0.686& 0.311& 0.886& 0.830& 0.266 \\
$^{150}$Sm & 0.901& 1.074& 0.238& 0.823& 1.101& 0.203 \\ \colrule
$^{238}$U & 0.863& 0.735& 0.265& 0.763& 0.596& 0.269 \\
$^{238}$Pu & 0.828& 0.640& 0.269& 0.745& 0.572& 0.272
\end{tabular}
\end{ruledtabular}
\end{table}
The isovector time-odd part of any Skyrme-type EDF is given by
\begin{align}
\chi^{\rm odd}_1(\bm{r}) &=
C_1^s[\rho_0]\bm{s}_1^2 + C_1^{\Delta s} \bm{s}_1\cdot \Delta \bm{s}_1 + C_1^j \bm{j}_1^2
\nonumber \\
&\quad + C_1^T \bm{s}_1\cdot \bm{T}_1
+ C_1^{\nabla j} \bm{s}_1\cdot \bm{\nabla}\times \bm{j}_1
\nonumber \\ &\quad
+ C_1^F \bm{s}_1\cdot \bm{F}_1
+ C_1^{\nabla s} ( \bm{\nabla}\cdot\bm{s}_1)^2 \,,
\label{eq:time-odd}
\end{align}
where $\bm{s}_1, \bm{j}_1, \bm{T}_1$, and $\bm{F}_1$ are the isovector spin,
current, spin-kinetic, and tensor-kinetic densities, respectively. The
isoscalar pairing functional in all these parametrizations has the mixed
density dependence
\begin{align}
\tilde{\chi}_0(\bm{r}) = \frac{V_0}{4}\left[
1 - \frac{1}{2} \frac{\rho_0(\bm{r})}{\rho_c}
\right] | \tilde{\bm{s}}_0(\bm{r})|^2 \,,
\end{align}
where $\tilde{\bm{s}}_0$ is the isoscalar pair density, $\rho_c=0.16$ fm$^{-3}$,
and $\rho_0$ is the usual isoscalar density. In the parameter sets 1A, 1B, 1C,
1D, 1E, only $C_1^s$ (with no density dependence) and $V_0$ are fit. In sets 3A
and 3B $C_1^T$ and $C_1^F$ are fit as well. In the parameter set 4, $C_1^j,
C_1^{\nabla j}$, and $C_1^{\nabla s}$ are adjusted, while other parameters are
the same as in set 3A. In set 5, $V_0$, $C_1^s$, and $C_1^j$ are fit.
\begin{table}[t]
\caption{The HFB overlap $\langle 0^+_{f,{\rm HFB}}|0^+_{{\rm HFB},i}\rangle$
between the initial and the final states of the double-beta decay, computed with
SkO$'$ and SV-min EDFs.
The numbers in parentheses denote powers of 10. \label{table:overlap-global1}}
\begin{ruledtabular}
\begin{tabular}{cccccccccc}
& & SkO$'$ & & & SV-min & & \\
& neutron & proton & total & neutron & proton & total & \\ \hline
$^{48}$Ca & 0.764 & 0.513 & 0.392 & 0.776 & 0.512 & 0.398 \\
$^{76}$Ge & 0.577 & 0.559 & 0.323 & 0.586 & 0.587 & 0.344 \\
$^{82}$Se & 0.729 & 0.829 & 0.604 & 0.772 & 0.862 & 0.665 \\
$^{96}$Zr & 0.283 & 0.306 & 0.087 & 0.882 & 0.877 & 0.774 \\
(sph.$\to$sph.) & 0.915 & 0.893 & 0.818 & \\
$^{100}$Mo & $1.8(-3)$& $1.4(-2)$ & $2.6(-5)$ &0.914 & 0.905 & 0.828 \\
(pro.$\to$pro.) & 0.864 & 0.875 & 0.755 \\
$^{116}$Cd & 0.932 & 0.521 & 0.485 & 0.507 & 0.293 & 0.148 \\
$^{128}$Te & 0.342 & 0.388 & 0.133 & 0.294 & 0.343 & 0.101 \\
(obl.$^{128}$Xe) & 0.440 & 0.533 & 0.235 & 0.403 & 0.487 & 0.197 \\
$^{130}$Te & 0.489 & 0.523 & 0.256 & 0.464 & 0.509 & 0.236 \\
$^{136}$Xe & 0.517 & 0.921 & 0.476 & 0.522 & 0.931 & 0.486 \\
$^{150}$Nd & 0.624 & 0.601 & 0.375 & 0.711 & 0.683 & 0.485 \\
$^{238}$U & 0.912 & 0.882 & 0.805 & 0.902 & 0.873 & 0.787
\end{tabular}
\end{ruledtabular}
\end{table}
Table~\ref{table:hfb-global1} lists the pairing gaps and quadrupole deformation
of the HFB states used to compute $2\nu\beta\beta$ nuclear matrix elements.
Neutron pairing collapses only in $^{136}$Xe and proton pairing collapses in
$^{48}$Ca and $^{116}$Sn.
SkO$'$ and SV-min cause different
amounts of deformation. $^{96}$Zr, $^{100}$Mo, and $^{100}$Ru are oblate,
oblate, and prolate (respectively) with SkO$'$, while they are all spherical
with SV-min. $^{116}$Cd is spherical with SkO$'$, but is
prolate with SV-min.
Table \ref{table:overlap-global1} contains the overlaps of the initial and final
HFB vacua. Significant differences in deformation and pairing between the two
HFB states lead to small overlaps, and because the two EDFs can produce
different levels of deformation and pairing in any nucleus, the overlaps depend
significantly on the EDF. In $^{96}$Zr and $^{100}$Mo, the HFB overlaps with
SkO$'$ are extremely small because the initial state is oblate and the final state
spherical or prolate. In $^{116}$Cd, the HFB overlap with
SV-min is smaller for a similar reason. The QRPA may not be adequate when the
overlaps, like those with SkO$'$ in $^{100}$Mo, are very small. Our treatment
omits both projection onto states with good angular momentum, which involves the mixing of states with different orientations, and the fluctuation in shape and pairing
captured, e.g., by the generator coordinate method \cite{LopezVaquero2011520,PhysRevC.90.031301}.
The effects of the physics we have neglected can be significant when the matrix
elements are small at the HFB or QRPA levels.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{2nbb_benchmark.pdf}
\caption{Comparison of dimensionless $2\nu\beta\beta$ nuclear matrix elements obtained from
global EDFs with experimental values. The matrix elements computed with the
lowest-energy HFB solutions are marked with crosses, while those elements
computed with the other HFB solutions are marked with triangles. Orange symbols
come from computations with the QTDA overlap and blue symbols from computations
with the QRPA overlap. The EDFs that give rise to each particular point appear
in Table \ref{table:2nbb-global1} in Appendix \ref{sec:appen-num}.
\label{fig:2nbb}}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=\textwidth]{2nbb_prediction.pdf}
\caption{Dimensionless $2\nu\beta\beta$ nuclear matrix elements calculated with the global
EDFs. The matrix elements from parameter set 4 are excluded.
\label{fig:2nbb_prediction}}
\end{figure*}
In Fig.~\ref{fig:2nbb} we compare the Gamow-Teller $2\nu\beta\beta$ nuclear
matrix element, scaled by $g_A^2 m_e c^2$ to be dimensionless, that results from
calculations with the ten different SkO$'$- and SV-min-based EDFs discussed just above. We
also show the experimental matrix elements, extracted from the half-lives in
Ref.~\cite{universe.6.159}. We use a quenched axial-vector coupling constant
$g_A=1.0$ to match the value from Ref.~\cite{PhysRevC.93.014304}, which
determines the EDF parameters. Despite the differences among the EDFs in the
pieces of the functional that were fit and in the data chosen to fit them, the
$2\nu\beta\beta$ matrix elements that they produce are quite close to one
another in some of the heavier nuclei. This fact means that the parts of the
EDF that affect the $2\nu\beta\beta$ matrix element are determined almost fully
by the $\beta$ decay rates and giant resonance energies used in fitting them.
In some lighter isotopes such as $^{48}$Ca, $^{96}$Zr, and $^{100}$Mo, on the
other hand, the values of the nuclear matrix element, like those of the overlap,
depend significantly on the EDF. Although such matrix elements may provide an
additional constraint on the pnEDF, the disagreement also suggests, as we noted
earlier, that correlations that escape the QRPA are important
\cite{Rodriguez2011436,0954-3899-44-3-034002}. Table~\ref{table:2nbb-global1}
in Appendix \ref{sec:appen-num} contains more detail than Fig.\ \ref{fig:2nbb},
in particular the values for each individual EDF of all the matrix elements.
In some nuclei, such as $^{76}$Ge, the EDFs all produce comparable values for
the matrix element, but those values are quite different from the experimental
one. The reason for the discrepancy, again, is the quite different degrees of
deformation in the initial and final nuclei, a difference that in reality is
probably made less significant by shape fluctuations. In other nuclei,
$^{96}$Zr, $^{100}$Mo, and $^{128}$Te (and especially the first two), the values
span a wide range. The reason is that two local minima appear in the initial
isotopes, and the value of the $2\nu\beta\beta$ matrix element depends strongly
on which minimum is used.
In these two cases, the HFB overlaps associated with the lowest minima for SkO$'$ are very small (0.087 and 2.6$\times$10$^{-5}$ in $^{96}$Zr and $^{100}$Mo), and the $2\nu\beta\beta$ matrix elements from the lowest minimum are consequently smaller than the experimental values,
while the matrix elements associated with the other HFB solutions are larger than or comparable to the experimental values.
Correlations that admix states near those other minima, if they were taken into account, would probably increase
the $2\nu\beta\beta$ matrix elements produced by the lowest minima. Such admixtures are beyond what the QRPA includes, however, and seeing their effects would require an approximation such as the generator-coordinate method.
Figure \ref{fig:2nbb} also shows that overlaps computed with the
QTDA prescription result in larger matrix elements than those computed with the
QRPA prescription.
\subsection{Predictions}
Using the same global EDFs as in the previous section, we compute the
$2\nu\beta\beta$ matrix elements for all the nuclei in which that decay might
conceivably be observed: $^{46}$Ca, $^{70}$Zn, $^{80}$Se, $^{86}$Kr, $^{94}$Zr,
$^{98}$Mo, $^{104}$Ru, $^{110}$Pd, $^{114}$Cd, $^{122}$Sn, $^{124}$Sn, $^{134}$Xe, $^{142}$Ce,
$^{146}$Nd, $^{148}$Nd, $^{154}$Sm, $^{160}$Gd, $^{170}$Er, $^{176}$Yb,
$^{186}$W, $^{192}$Os, $^{198}$Pt, $^{204}$Hg, $^{226}$Ra, $^{232}$Th,
$^{244}$Pu, and $^{248}$Cm. Figure \ref{fig:2nbb_prediction}
summarizes the results, while Table \ref{table:2nbb-prediction} in Appendix
\ref{sec:appen-num} indicates the individual EDFs responsible for each symbol in
the figure.
We emphasize that we are able to make these predictions only because we use EDFs
that are fit globally and without considering $2\nu\beta\beta$ half-lives. In
typical QRPA calculations, by contrast, the strength of isoscalar pairing is
adjusted in each nucleus individually to reproduce the $2\nu\beta\beta$ half-life.
As we mentioned in the previous section, the QRPA nuclear matrix elements may
not be reliable if the deformations of the initial and final states of the decay
are different. The deformation parameters differ by more than 0.1 for the following
decays: $^{70}$Zn $\to$ $^{70}$Ge, $^{80}$Se $\to$ $^{80}$Kr, $^{134}$Xe $\to$
$^{134}$Ba, and $^{146}$Nd $\to$ $^{146}$Sm with SkO$'$ and $^{80}$Se $\to$
$^{80}$Kr, $^{114}$Cd $\to$ $^{114}$Sn, $^{122}$Sn $\to$ $^{122}$Te, $^{134}$Xe
$\to$ $^{134}$Ba, and $^{146}$Nd $\to$ $^{146}$Sm with SV-min. We also saw
earlier that the QRPA can go awry if shape mixing is important. A full
treatment of shape mixing requires something like the generator coordinate
method \cite{Ring-Schuck,RevModPhys.75.121}, but we can get a good idea of when
it will be significant by examining potential energy curves. These turn out to
be broad near the minimum for the nuclei $^{46}$Ti, $^{70}$Ge, $^{94}$Zr (only
SkO$'$), $^{98}$Mo, $^{104}$Ru, $^{110}$Pd, $^{114}$Cd, $^{122}$Te,
$^{124}$Te, $^{134}$Ba,
$^{142}$Ce, $^{198}$Pt, and $^{198}$Hg. Unfortunately, the generator coordinate
method, while it has been applied to $0\nu\beta\beta$ decay
\cite{PhysRevC.90.031301,PhysRevC.93.014305,PhysRevLett.105.252503,Rodriguez2011436,Rodriguez2013174,PhysRevLett.124.232501,PhysRevC.98.054311,PhysRevC.95.024305,PhysRevC.91.024316,PhysRevC.90.054309,PhysRevC.96.054310,
PhysRevC.98.064324, PhysRevC.100.031303} is difficult to apply to
$2\nu\beta\beta$ decay because the closure approximation is poor there and a
complete set of intermediate states is required.
With the parameter set 4, the pnFAM converges more slowly than with the other
parameter sets, and the resulting matrix elements are often quite different from
those produced by the other sets. Thus we exclude set 4 from the distribution
of the nuclear matrix elements shown in Fig.~\ref{fig:2nbb_prediction}. We see
better agreement among the other EDFs in heavier isotopes as a general rule, and
the QTDA prescription for the overlap again leads to larger numbers than does
QRPA prescription.
\section{Conclusions \label{sec:conclusion}}
We have presented a computationally efficient framework for calculating the
matrix elements for two-neutrino double-beta decay within nuclear
density functional theory. We employ the finite amplitude method to compute the
QRPA approximation to the matrix elements. Our approach allows large
single-particle model spaces and the use of a single nuclear EDF for all nuclei.
It also eliminates the need to truncate two-quasiparticle spaces.
We first used harmonic-oscillator-based HFB and FAM codes together with familiar
EDFs to compute the $2\nu\beta\beta$ matrix elements in a few important nuclei,
comparing the results with those obtained previously by diagonalizing the QRPA
matrix. Using EDFs that had been fit globally to single-$\beta$ decay rates and
giant-resonance energies, we then computed the $2\nu\beta\beta$ matrix elements
in all nuclei in which double-beta decay has or could be observed. Agreement
with the matrix elements extracted from already measured half-lives is good in
general, and we offered predictions for those nuclei that have unmeasured
half-lives.
Although we focus on $2\nu\beta\beta$ decay in this paper, we can also compute
double-electron capture matrix elements in the same way. The most interesting
extension of our work is to neutrinoless double-beta decay. The presence of a
neutrino propagator in that matrix element, however, will make that process more
challenging to treat than $2\nu\beta\beta$ decay.
\section*{Acknowledgments}
We are grateful to Mika T. Mustonen for the valuable discussions. This work is
supported by JSPS KAKENHI Grants No.~17H05194, No.~19KK0343, No.~20H05242,
and No.~20K03964, and by the U.S. Department of Energy, Office of Science, Office
of Nuclear Physics, under grant No. DE-FG02-97ER41019.
This research was conducted in part during the INT program INT-17-2a ``Neutrinoless Double-Beta Decay," at the Institute for Nuclear Theory, University of Washington.
Numerical calculations were performed at the Oakforst-PACS Systems through the Multidisciplinary Cooperative Research Program of the Center for Computational Sciences,
University of Tsukuba.
|
\section{Algebraic preliminaries} \label{sec:algebra}
\subsection{Hypercubes and hyperboxes}
In this section, we recall the hypercube formalism of Manolescu and Ozsv\'{a}th \cite{MOIntegerSurgery}. We write $\mathbb{E}_n$ for the set of points $\{0,1\}^n\subset \mathbb{Z}^n$. Similarly if $\ve{d}=(d_1,\dots, d_n)$ is a tuple of $n$ positive integers, we write $\mathbb{E}(\ve{d})=\{0,\dots, d_1\}\times \cdots \times \{0,\dots, d_n\}$. If $\varepsilon,\varepsilon'\in \mathbb{R}^n$, we say that $\varepsilon\le\varepsilon'$ if the inequality holds for each component of $\varepsilon$ and $\varepsilon'$.
\begin{define}
An \emph{$n$-dimensional hypercube of chain complexes} $(C^\varepsilon, D^{\varepsilon,\varepsilon'})_{\varepsilon\in \mathbb{E}_n}$ is a collection of groups $C^{\varepsilon}$, ranging over $\varepsilon\in \mathbb{E}_n$, together with a collection of maps $D^{\varepsilon,\varepsilon'}\colon C^\varepsilon\to C^{\varepsilon'}$, whenever $\varepsilon\le \varepsilon'$. Furthermore, we assume the following compatibility condition is satisfied whenever $\varepsilon\le \varepsilon''$:
\begin{equation}
\sum_{\varepsilon': \varepsilon\le \varepsilon'\le\varepsilon''} D^{\varepsilon',\varepsilon''}\circ D^{\varepsilon,\varepsilon'}=0. \label{eq:hypercube-structure-relations}
\end{equation}
\end{define}
We usually call $D^{\varepsilon,\varepsilon}$ the internal differential of $C^\varepsilon$. The hypercube structure relation in equation~\eqref{eq:hypercube-structure-relations} is equivalent to $(\bigoplus_{\varepsilon\in \mathbb{E}_n} C^\varepsilon,\sum_{\varepsilon\le \varepsilon'} D^{\varepsilon,\varepsilon'})$ being a chain complex.
\begin{define}
A \emph{hyperbox of chain complexes of size~$\ve{d}$} consists of a collection of groups $(C^\varepsilon)_{\varepsilon\in \mathbb{E}(\ve{d})}$, together with a choice of map $D^{\varepsilon,\varepsilon'}\colon C^\varepsilon\to C^{\varepsilon'}$ whenever $|\varepsilon'-\varepsilon|_{L^\infty}\le 1$. We assume equation~\eqref{eq:hypercube-structure-relations} holds whenever $|\varepsilon''-\varepsilon|_{L^\infty}\le 1$.
\end{define}
An important operation involving hyperboxes is \emph{compression} \cite{MOIntegerSurgery}*{Section~5}. This operation takes a hyperbox of size $\ve{d}=(d_1,\dots, d_n)$ and returns an $n$-dimensional hypercube of chain complexes. We illustrate with an example. Consider a 1-dimensional hyperbox
\[
\begin{tikzcd}
C_0
\mathrm{a.r.}[r, "f_{0,1}"]
&
C_1
\mathrm{a.r.}[r, "f_{1,2}"]
&
\cdots
\mathrm{a.r.}[r, "f_{n-1,n}"]
&
C_n.
\end{tikzcd}
\]
The compression of the above hyperbox is the hypercube
\[
\begin{tikzcd}[column sep=2.5cm]
C_0
\mathrm{a.r.}[r, "f_{n-1,n}\circ \cdots \circ f_{0,1}"]
&
C_n.
\end{tikzcd}
\]
\noindent More generally, the same description works for an $n$-dimensional hyperbox $\mathcal{C}$ of size $(1,\dots, 1,d)$. That is, we may view $\mathcal{C}$ as a 1-dimensional hyperbox of size $(d)$, which we call $\mathcal{C}'$, where the complex at each point $\varepsilon_0$ of $\mathbb{E}(d)$ is the $(n-1)$-dimensional hypercube $\mathcal{C}'_{\varepsilon_0}=(\oplus_{\varepsilon\in \mathbb{E}_{n-1}} C^{\varepsilon\varepsilon_0},\sum_{\varepsilon\leq \varepsilon'}D^{\varepsilon \varepsilon_0,\varepsilon'\varepsilon_0})$, viewed as a chain complex. Schematically, $\mathcal{C}'$ is represented as a diagram
\[
\begin{tikzcd}
\mathcal{C}'_0
\mathrm{a.r.}[r, "f_{0,1}"]
&
\mathcal{C}'_1
\mathrm{a.r.}[r, "f_{1,2}"]
&
\cdots
\mathrm{a.r.}[r, "f_{d-1,d}"]
&
\mathcal{C}'_d.
\end{tikzcd}
\]
where the maps $f_{i,i+1}$ are constructed as a sum of the maps $D^{\varepsilon i,\varepsilon' i+1}$ for $\varepsilon\leq \varepsilon'\in \mathbb{E}_{n-1}$.
The definition given above for compression of a $1$-dimensional hyperbox of chain complexes then applies, to construct a $1$-dimensional hypercube:
\[
\begin{tikzcd}[column sep=2.5cm]
\mathcal{C}'_0
\mathrm{a.r.}[r, "f_{d-1,d}\circ \cdots \circ f_{0,1}"]
&
\mathcal{C}'_n.
\end{tikzcd}
\]
For $\varepsilon\leq \varepsilon'\in \mathbb{E}_n$ with $\varepsilon_n\neq \varepsilon_n'$, the map $D^{\varepsilon,\varepsilon'}$ of the compression of $\mathcal{C}$ is the component of $f_{d-1,d}\circ\dots \circ f_{0,1}$ from $\mathcal{C}^\varepsilon$ to $\mathcal{C}^{\varepsilon'}$.
For a hyperbox $C$ of size $(d_1,\dots, d_n)$, the above description may be iterated to give the compression, as follows. This description depends on a choice of ordering of the axes of the cube. (The motivated reader may verify that re-ordering the axes results in a homotopic hypercube.) Firstly, we may view $C$ as a collection of $d_1\cdots d_{n-1}$ hyperboxes of size $(1,\dots, 1,d_n)$. We compress each of these hyperboxes using the function-composition description above. Stacking these hyperboxes results in a hyperbox of size $(d_1,\dots, d_{n-1},1)$. We may view this as a collection of $d_1\cdots d_{n-2}$ hypercubes of size $(1,\dots, 1,d_{n-1}, 1)$, which we compress using the function-composition description. Stacking the resulting hyperboxes gives a hyperbox of size $(d_1,\dots, d_{n-2},1,1)$. Repeating this procedure gives the compression of $C$. It is possible to relate this description to the description in terms of the algebra of songs in \cite{MOIntegerSurgery}, cf. \cite{Liu2Bridge}*{Section~4.1.2}.
\subsection{Hypercubes of attaching curves}
In this section, we describe the notion of a hypercube in the Fukaya category. Such objects are described by Manolescu and Ozsv\'{a}th \cite{MOIntegerSurgery}, and are referred to as \emph{hyperboxes of Heegaard diagrams}. The reader may also compare this construction to Lipshitz, Ozsv\'{a}th and Thurston's notion of a \emph{chain complex of attaching circles} \cite{LOTDoubleBranchedII}. The construction is also described in \cite{HHSZExact}*{Section~5.5}.
We begin with a preliminary definition about $\Spin^c$ structures. Firstly, we say $\mathcal{L}_{\b}=(\ve{\beta}^\varepsilon)_{\varepsilon\in \mathbb{E}_n}$ is an \emph{empty hypercube of beta (or alpha)-attaching curves} if it is a collection of attaching curves on a surface $\Sigma$, indexed by points of $\mathbb{E}_n$.
\begin{define}
\begin{enumerate}
\item Suppose that $\mathcal{L}_{\b}=(\ve{\beta}^\varepsilon)_{\varepsilon\in \mathbb{E}_n}$ is an empty hypercube of beta-curves. A \emph{hypercube of $\Spin^c$-structures} for $\mathcal{L}_{\b}$ consists of a collection of $\Spin^c$ structures
\[
\mathfrak{S}_{\varepsilon_1,\dots, \varepsilon_m}\subset \Spin^c(X_{\varepsilon_1,\dots, \varepsilon_m})
\]
for each sequence $\varepsilon_1<\cdots< \varepsilon_m$ in $\mathbb{E}_n$, satisfying $m>1$ and the following compatibility relations. Firstly, if $1\le i<j\le m$, then $\mathfrak{S}_{\varepsilon_1,\dots, \varepsilon_m}$ is closed under the action of $\delta^1 H^1(Y_{\varepsilon_i,\varepsilon_j})$. Secondly if $\varepsilon_1<\cdots< \varepsilon_m$ is a sequence, and $\varepsilon_{i_1}<\dots< \varepsilon_{i_j}$ is a subsequence, where $1\le i_1<\dots< i_j\le m$, then $\mathfrak{S}_{\varepsilon_{i_1},\dots, \varepsilon_{i_j}}$ is the image of $\mathfrak{S}_{\varepsilon_1,\dots, \varepsilon_m}$ under the natural restriction map
\[
\Spin^c(X_{\varepsilon_1,\dots, \varepsilon_m})\to \Spin^c(X_{\varepsilon_{i_1},\dots, \varepsilon_{i_j}}).
\]
\item If $\mathcal{L}_{\a}=(\ve{\alpha}^\varepsilon)_{\varepsilon\in \mathbb{E}_n}$ and $\mathcal{L}_{\b}=(\ve{\beta}^\varepsilon)_{\varepsilon\in \mathbb{E}_m}$ are two empty hypercubes of alpha and beta attaching curves, then a hypercube of $\Spin^c$ structures for the pair $(\mathcal{L}_{\a},\mathcal{L}_{\b})$ consists of the following. For every pair of sequences $(\nu_1<\dots<\nu_k)$ in $\mathbb{E}_n$ and $(\varepsilon_1<\dots< \varepsilon_\ell)$ in $\mathbb{E}_m$ such that $k+\ell>1$, a set of $\Spin^c$ structures $\mathfrak{S}_{\nu_k,\dots, \nu_1,\varepsilon_1,\dots, \varepsilon_\ell}\subset\Spin^c(X_{\nu_k,\dots, \nu_1,\varepsilon_1,\dots, \varepsilon_{\ell}})$ that are closed under $\delta^1$-orbits and compatible with restriction, similar to above. The case that one of $(\nu_1<\dots<\nu_k)$ and $(\varepsilon_1<\dots <\varepsilon_\ell)$ is the empty list is allowed.
\end{enumerate}
\end{define}
Note that a hypercube of $\Spin^c$ structures for the pair $(\mathcal{L}_{\a}, \mathcal{L}_{\b})$ induces a hypercube of $\Spin^c$ structures for each of $\mathcal{L}_{\a}$ and $\mathcal{L}_{\b}$.
\begin{define}
An $n$-dimensional hypercube of beta attaching curves $\mathcal{L}_{\b}$ on a pointed surface $(\Sigma,w)$ consists of an empty hypercube of beta attaching curves $(\ve{\beta}^\varepsilon)_{\varepsilon\in \mathbb{E}_n}$, a hypercube of $\Spin^c$ structures on $\mathcal{L}_{\b}$, together with a choice of Floer chain $\Theta_{\varepsilon,\varepsilon'}\in \mathit{CF}^-(\Sigma, \ve{\beta}^\varepsilon,\ve{\beta}^{\varepsilon'}, w)$ whenever $\varepsilon<\varepsilon'$. Furthermore, the chains are required to satisfy the following compatibility condition, whenever $\varepsilon<\varepsilon'$:
\begin{equation}
\sum_{\varepsilon=\varepsilon_1<\cdots<\varepsilon_n=\varepsilon'} f_{\b^{\varepsilon_1},\b^{\varepsilon_2},\dots, \b^{\varepsilon_n}} (\Theta_{\varepsilon_1,\varepsilon_2},\dots, \Theta_{\varepsilon_{n-1},\varepsilon_n})=0.
\label{eq:hypercube-Lagrangians}
\end{equation}
In the above, the sum is taken over $\Spin^c$ structures according to the hypercube of $\Spin^c$ structures.
\end{define}
We will also write the Floer chain $\Theta_{\varepsilon,\varepsilon'}$ as $\Theta_{\b^{\varepsilon}, \b^{\varepsilon'}}$ when it seems clearest to refer to the sets of curves involved. In the above, we assume that each Heegaard multi-diagram is weakly admissible, and the appropriate finiteness of counts so that the sum makes sense. Assuming weak admissibility, one may always obtain a sensible expression by working over the power series ring $\mathbb{F}[[U]]$. Hypercubes of \emph{alpha} attaching curves are defined by a notational modification of the above definition.
We may \emph{pair} hypercubes of attaching curves, as follows. If $\mathcal{L}_{\a}=(\ve{\alpha}^\nu,\Theta_{\nu',\nu})_{\nu\in \mathbb{E}_m}$ and $\mathcal{L}_{\b}=(\ve{\beta}^\varepsilon,\Theta_{\varepsilon,\varepsilon'})_{\varepsilon\in \mathbb{E}_n}$ are hypercubes of attaching curves for some hypercubes of $\Spin^c$ structures, and we have a hypercube of $\Spin^c$ structures $\mathfrak{S}$ for the pair $(\mathcal{L}_{\a},\mathcal{L}_{\b})$ which extends the hypercubes for $\mathcal{L}_{\a}$ and $\mathcal{L}_{\b}$, then we may form an $(n+m)$-dimensional hypercube of chain complexes denoted $\mathit{CF}^-(\mathcal{L}_{\a},\mathcal{L}_{\b}, \mathfrak{S})$. The group at index $(\nu,\varepsilon)$ is the ordinary Floer complex $\mathit{CF}^-(\Sigma, \ve{\alpha}^{\nu},\ve{\beta}^\varepsilon,\mathfrak{S}_{\nu,\varepsilon})$. The morphism from $(\nu,\varepsilon)$ to $(\nu',\varepsilon')$ is the map
\[
D_{(\nu,\varepsilon),(\nu',\varepsilon')}(\ve{x})=\sum_{\substack{\nu=\nu_1<\cdots<\nu_k=\nu'\\ \varepsilon=\varepsilon_1<\cdots<\varepsilon_j=\varepsilon'}}f_{\nu_k,\dots, \nu_1,\varepsilon_1,\dots, \varepsilon_j,\mathfrak{S}_{\nu_k,\dots, \nu_1,\varepsilon_1,\dots, \varepsilon_j}}(\Theta_{\nu_k,\nu_{k-1}},\dots, \Theta_{\nu_2,\nu_1}, \ve{x},\Theta_{\varepsilon_1,\varepsilon_2},\dots, \Theta_{\varepsilon_{j-1},\varepsilon_{j}}).
\]
It is straightforward to use the compatibility condition in~\eqref{eq:hypercube-Lagrangians} to see that $\mathit{CF}^-(\mathcal{L}_{\a},\mathcal{L}_{\b},\mathfrak{S})$ is a hypercube of chain complexes. In specific examples, we will also use the abbreviated notation
\[
f_{\a^{\nu_1}\to \dots\to \a^{\nu_k}}^{\b^{\varepsilon_1}\to \cdots \to \b^{\varepsilon_j}}(\ve{x})=f_{\nu_k,\dots, \nu_1,\varepsilon_1,\dots, \varepsilon_j,\mathfrak{S}_{\nu_k,\dots, \nu_1,\varepsilon_1,\dots, \varepsilon_j}}\left(\Theta_{\nu_k,\nu_{k-1}},\dots, \Theta_{\nu_2,\nu_1}, \ve{x},\Theta_{\varepsilon_1,\varepsilon_2},\dots, \Theta_{\varepsilon_{j-1},\varepsilon_{j}}\right).
\]
\noindent If $\nu=\nu'$, then we will usually write $f_{\a^{\nu}}^{\b^{\varepsilon}\to \b^{\varepsilon'}}$, and similarly if $\varepsilon=\varepsilon'$. With respect to this notation the hypercube map becomes
\[
D_{(\nu,\varepsilon),(\nu',\varepsilon')}(\ve{x})=\sum_{\substack{\nu=\nu_1<\dots<\nu_k=\nu' \\ \varepsilon=\varepsilon_1<\dots< \varepsilon_j=\varepsilon'}}f_{\a^{\nu_1}\to \dots\to \a^{\nu_k}}^{ \b^{\varepsilon_1}\to \cdots \to \b^{\varepsilon_j}}.\]
We will also frequently substitute the letter $h$ for $f$ in the specific case of a map counting pseudoholomorphic rectangles.
\subsection{$\iota$-complexes}
We recall the following definition from \cite{HMInvolutive}.
\begin{define}
\label{def:iota-complex}
An \emph{$\iota$-complex} is a chain complex $(C,\d)$ which is free and finitely generated over $\mathbb{F}[U]$ and equipped with an endomorphism $\iota$. Furthermore, the following hold:
\begin{enumerate}
\item\label{iota-1} $C$ is equipped with a $\mathbb{Z}$-grading, such that $U$ has grading $-2$. We call this grading the \emph{Maslov} or \emph{homological} grading.
\item\label{iota-2} There is a grading preserving isomorphism $U^{-1}H_*(C)\cong \mathbb{F}[U,U^{-1}]$.
\item \label{iota-3} $\iota$ is a grading preserving chain map and $\iota^2\simeq \id$.
\end{enumerate}
\end{define}
\begin{define}\label{def:enhanced-iota-homomorphisms}
If $(C,\iota)$ and $(C',\iota')$ are $\iota$-complexes, we define the group of \emph{enhanced $\iota$-morphisms} to be
\[
\underline{\Mor}((C,\iota),(C',\iota')):=\Hom_{\mathbb{F}[U]}(C,C')\oplus \Hom_{\mathbb{F}[U]}(C,C')[1],
\]
where $[1]$ denotes a grading shift. We define
\[
\d_{\underline{\Mor}}(F,h)=(\d' F+F\d, F\iota+\iota'F+\d' h+h\d),
\]
which makes the category of $\iota$-complexes into a dg-category.
An enhanced $\iota$-\emph{homo}morphism is an enhanced $\iota$-morphism $(F,h)$ satisfying $\d_{\underline{\Mor}}(F,h)=0$. An \emph{enhanced $\iota$-homotopy equivalence} is an enhanced $\iota$-homomorphism $(F,h)$ such that there exists an enhanced $\iota$-homomorphism $(G,k)$ with the property that $(G,k)\circ (F,h)$ and $(F,h)\circ (G,k)$ both differ from $(\id,0)$, the identity enhanced $\iota$-morphism, by a boundary in $\underline{\Mor}$. Composition is given by $(F,h)\circ (F',h')=(F\circ F',F'h+h'F)$.
\end{define}
\subsection{Hypercubes and $\iota$-complexes} \label{subsec:hypercube-equivalence}
In this section, we describe a relation between enhanced $\iota$-morphisms and hypercubes. An enhanced $\iota$-morphism $(F,h)\colon (C,\iota)\to (C',\iota')$ may be encoded into a diagram of the following form
\begin{equation}
\begin{tikzcd}[labels=description,row sep=1.5cm, column sep=2cm]
C\mathrm{a.r.}[d, "\iota"] \mathrm{a.r.}[r, "F"]\mathrm{a.r.}[dr,dashed, "h"] & C\mathrm{a.r.}[d, "\iota' "]\\
C\mathrm{a.r.}[r,"F"]& C'
\end{tikzcd}
\label{eq:hypercube=morphism-iota-complexes}
\end{equation}
The pair $(F,h)$ is an enhanced $\iota$-homomorphism if and only if the above diagram is a hypercube of chain complexes. Composition of $\iota$-morphisms is encoded by stacking and compressing hypercubes.
\begin{rem}
The diagram in equation~\eqref{eq:hypercube=morphism-iota-complexes} is algebraically equivalent to the following diagram:
\begin{equation}
\begin{tikzcd}[labels=description,row sep=1.5cm, column sep=2cm]
C\mathrm{a.r.}[d, "Q(\iota+\id)"] \mathrm{a.r.}[r, "F"]\mathrm{a.r.}[dr,dashed, "Q\cdot h"] & C\mathrm{a.r.}[d, "Q(\iota' +\id)"]\\
Q\cdot C\mathrm{a.r.}[r,"F"]& Q\cdot C'
\end{tikzcd}
\end{equation}
In particular, we may view enhanced iota morphisms between iota complexes as containing equivalent information to an $\mathbb{F}[U,Q]/Q^2$-equivariant map between the associated complexes over $\mathbb{F}[U,Q]/Q^2$. We will use the two perspectives interchangeably.
\label{rem:q-to-iota}
\end{rem}
Relations between compositions of $\iota$-morphisms are also naturally encoded into hypercubes:
\begin{lem}\label{lem:iota-maps-from-hyperboxes}
Suppose that $(A,\iota_A)$, $(B,\iota_B)$, $(C,\iota_C)$ and $(D,\iota_D)$ are $\iota$-complexes, and $(F,f)$, $(G,g)$, $(H,h)$ $(I,i)$ and $(J,j)$ are enhanced $\iota$-morphisms which fit into the following diagram
\[
\begin{tikzcd}[column sep={1.5cm,between origins},row sep=.8cm,labels=description]
A
\mathrm{a.r.}[dr, "G"]
\mathrm{a.r.}[ddd, "\iota_A"]
\mathrm{a.r.}[dddrr,dashed, "i"]
\mathrm{a.r.}[rr, "I"]
\mathrm{a.r.}[ddddrrr,dotted, "j"]
&[.7 cm]\,
&B
\mathrm{a.r.}[rd, "H"]
\mathrm{a.r.}[ddd,"\iota_B"]
\mathrm{a.r.}[ddddr,dashed, "h"]
&[.7 cm]\,
\\
&[.7 cm] C
\mathrm{a.r.}[rr, crossing over, "F"]
\mathrm{a.r.}[dddrr,dashed, crossing over, "f"]
&\,
&[.7 cm] D
\mathrm{a.r.}[from=ulll, dashed, crossing over, "J"]
\mathrm{a.r.}[ddd, "\iota_D"]
\\
\\
A
\mathrm{a.r.}[dr, "G"]
\mathrm{a.r.}[rr,"I"]
\mathrm{a.r.}[drrr,dashed, "J"]
&[.7 cm]\,&
B
\mathrm{a.r.}[dr, "H"]&[.9 cm]\,\\
& [.7 cm] C
\mathrm{a.r.}[from=uuuul, dashed, crossing over, "g"]
\mathrm{a.r.}[from=uuu,crossing over, "\iota_C"]
\mathrm{a.r.}[rr, "F"]&\,
&[.9 cm] D
\end{tikzcd}
\]
Then the hypercube relations for the above diagram are equivalent to each of $(F,f)$, $(G,g)$, $(H,h)$ and $(I,i)$ being enhanced $\iota$-homormophisms, as well as the relation
\[
(F,f)\circ (G,g)+(H,h)\circ (I,i)=\d_{\underline{\Mor}}(J,j).
\]
\end{lem}
\begin{proof}The length 1 relations imply that $F$, $G$, $H$ and $I$ are chain maps. The length 2 relations along the left, right, front and back faces are easily seen to be equivalent to $(F,f)$, $(G,g)$, $(H,h)$ and $(I,i)$ being enhanced $\iota$-homormorphisms. The length 2 relation along the top and bottom face is equivalent to
\[
FG+H I=[\d, J].
\]
The length 3 relation is equivalent to
\[
F g+f G+H i+h I=\iota_D J+J\iota_A+[\d, j].
\]
On the other hand, by definition,
\[
(F,f)\circ (G,g)+(H,h)\circ (I,i)=(FG+HI,Fg+fG+Hi+hI)
\]
while
\[
\d_{\underline{\Mor}}(J,j)=([\d,J], \iota_D J+J\iota_A+[\d, j]).
\]
The main claim follows immediately.
\end{proof}
\subsection{$\iota_K$-complexes and $\iota_L$-complexes} \label{sec:knot-and-link-complexes} We now review the refinement of the analog of an $\iota$-complex for knots, called an $\iota_K$-complex, and its extension to links. First we recall some algebraic background. Let $(C_K,\d)$ be a free, finitely generated complex over the ring $\mathbb{F}[\mathscr{U},\mathscr{V}]$. There are two naturally associated maps
\[
\Phi,\Psi\colon C_K\to C_K,
\]
constructed as follows. Write $\d$ as a matrix with respect to a free $\mathbb{F}[\mathscr{U},\mathscr{V}]$-basis of $C_K$. We then define $\Phi$ to be the endomorphism obtained by differentiating each entry of this matrix with respect to $\mathscr{U}$. We define $\Psi$ to be the endomorphism obtained by differentiating each entry with respect to $\mathscr{V}$. These maps appear naturally in the study of knot Floer homology, see \cites{SarkarMaslov, ZemQuasi, ZemCFLTQFT}. The maps $\Phi$ and $\Psi$ are independent of the choice of basis up to $\mathbb{F}[\mathscr{U},\mathscr{V}]$-equivariant chain homotopy \cite{ZemConnectedSums}*{Corollary~2.9}.
An $\mathbb{F}$-linear map $F\colon C_K\to C_K'$ is \emph{skew-$\mathbb{F}[\mathscr{U},\mathscr{V}]$-equivariant} if
\[
F\circ \mathscr{V}=\mathscr{U}\circ F\quad \text{and} \quad F\circ \mathscr{U}=\mathscr{V}\circ F.
\]
We recall the following definition from \cite{HHSZExact}. (Compare also \cite{HMInvolutive}*{Definition~6.2} and \cite{ZemConnectedSums}*{Definition~2.2}.)
\begin{define}\label{def:iota-K-complex}
An \emph{$\iota_K$-complex} $(C_K,\d, \iota_K)$ is a finitely generated, free chain complex $(C_K,\d)$ over $\mathbb{F}[\mathscr{U},\mathscr{V}]$, equipped with a skew-equivariant endomorphism $\iota_K$ satisfying
\[
\iota_K^2\simeq \id+\Phi\Psi.
\]
\end{define}
\begin{rem}\label{rem:iotaKU}
If $(C_K,\d,\iota_K)$ is an $\iota_K$-complex, then $\iota_K$ commutes with $U=\mathscr{U}\mathscr{V}$, and hence we can view $C_K$ as an (infinitely generated) complex over $\mathbb{F}[U]$ with an $\mathbb{F}[U]$-equivariant endomorphism $\iota_K$.
\end{rem}
We have the following notions of morphism of $\iota_K$-complexes:
\begin{define} Suppose that $\mathscr{C}=(C_K,\d, \iota_K)$ and $\mathscr{C}'=(C'_K,\d',\iota_K')$ are $\iota_K$-complexes.
\begin{enumerate}
\item An \emph{$\iota_K$-homomorphism} from $\mathscr{C}$ to $\mathscr{C}'$ consists of an $\mathbb{F}[\mathscr{U},\mathscr{V}]$-equivariant chain map $F\colon C_K\to C_K'$, which satisfies $\iota_K'F+F\iota_K\eqsim 0$ (where $\eqsim$ denotes skew-equivariantly chain homotopy equivalence).
\item The group of \emph{enhanced $\iota_K$-morphisms} is
\[
\underline{\Mor}(\mathscr{C},\mathscr{C}'):=\Hom_{\mathbb{F}[\mathscr{U},\mathscr{V}]}(C_K,C_K')\oplus \bar{\Hom}_{\mathbb{F}[\mathscr{U},\mathscr{V}]}(C_K,C_K')[1,1],
\]
where $\bar{\Hom}_{\mathbb{F}[\mathscr{U},\mathscr{V}]}(C_K,C_K')$ denotes the group of $\mathbb{F}[\mathscr{U},\mathscr{V}]$-skew-equivariant maps. The differential on $\underline{\Mor}(\mathscr{C}_K,\mathscr{C}'_K)$ is given by
\[
\d_{\underline{\Mor}}(F,g)=(F\d+\d'F, F\iota_K+\iota_K' F+\d' g+g\d).
\]
We say $(F,g)$ is an enhanced $\iota_K$-\emph{homo}morphism if $\d_{\underline{\Mor}}(F,g)=0$. Two enhanced $\iota_K$-morphisms are $\iota_K$-\emph{homotopic} if their sum is a boundary in $\underline{\Mor}(\mathscr{C}_K,\mathscr{C}_K')$.
\end{enumerate}
\end{define}
We will also be interested in the case of free, finitely-generated complexes $(C_L,\d)$ over the ring $\mathbb{F}[\mathscr{U}_1,\mathscr{V}_1, \dots \mathscr{U}_{\ell}, \mathscr{V}_{\ell}]$, which arise when we study links $L$ with multiple components. In this case, there are maps $\Phi_i$ and $\Psi_i$ for $1 \leq i \leq {\ell}$ arising from differentiating with respect to each variable. An $\mathbb{F}$-linear map $F \colon C_L \rightarrow C_L'$ is said to be skew-equivariant if it exchanges $\mathscr{U}_i$ and $\mathscr{V}_i$ for each $i$. We may then consider $\iota_L$-complexes $(C_L, \d, \iota_L)$, where the map $\iota_L$ satisfies the formula
\[
\iota_L^2 \simeq \left(\id+\Phi_{\ell}\Psi_{\ell}\right)\circ \dots \circ \left(\id + \Phi_1\Psi_1\right).
\]
This is the diffeomorphism map for performing a Dehn twist on each link component. See \cite{ZemQuasi}*{Theorem~B} for further discussion. The definitions of $\iota_K$-morphisms and their variations now extend straightforwardly to $\iota_L$-complexes.
As in Section \ref{subsec:hypercube-equivalence}, an enhanced $\iota_K$-homomorphism is equivalent to a hypercubes of chain complexes, compositions of $\iota_K$-morphisms are encoded as hypercubes in the same manner as in Lemma~\ref{lem:iota-maps-from-hyperboxes}.
\section{Introduction}
Heegaard Floer homology, defined by Ozsv{\'a}th and Szab{\'o} in the early 2000s \cites{OSDisks, OSProperties}, is a powerful suite of invariants of 3-manifolds, knots and links inside of them, and 4-dimensional cobordisms between them. Given a basepointed 3-manifold $(Y,w)$ together with a choice of $\Spin^c$ structure $\mathfrak{s}$ on $Y$, the initial construction of the invariant goes by choosing a Heegaard diagram $\mathcal{H}$ for $(Y,w)$ and associating to it a free finitely-generated chain complex $\mathit{CF}^{-}(\mathcal{H},\mathfrak{s})$ over the ring $\mathbb F[U]$, where $U$ is a variable of degree $-2$. If $\mathcal{H}$ and $\mathcal{H}'$ are two Heegaard diagrams representing $(Y,w)$, the Reidemeister-Singer theorem implies that they may be connected by a sequence of Heegaard moves. Ozsv{\'a}th and Szab{\'o} \cite[Section 9.2]{OSDisks} proved that sequences of moves induce transition maps, that is, chain homotopy equivalences
\[
\Psi_{\mathcal{H} \to \mathcal{H}'}\colon \mathit{CF}^{-}(\mathcal{H},\mathfrak{s}) \rightarrow \mathit{CF}^{-}(\mathcal{H}',\mathfrak{s}),
\]
\noindent showing that the isomorphism class of the homology group $\mathit{HF}^-(\mathcal{H}) = \oplus_{\mathfrak{s} \in \Spin^c(Y)}\mathit{HF}^-(\mathcal{H}, \mathfrak{s})$ is an invariant of $(Y,w)$.
Juh{\'a}sz, Thurston, and the fourth author \cite{JTNaturality} proved that this data is \emph{first-order natural} in the sense that the maps $\Psi_{\mathcal{H} \to \mathcal{H}'}$ are independent of the choice of Heegaard moves up to chain homotopy. (Here, ``first-order'' means that the chain homotopies relating the transition maps associated to two possible sequences of Heegaard moves are not yet known to be independent of the choices involved in their definition.) It follows that Heegaard Floer homology associates to $(Y,w)$ a specific module $\mathit{HF}^{-}(Y,w)$ rather than an isomorphism class of $\mathbb F[U]$-modules.
In 2015 the first author and Manolescu \cite{HMInvolutive} put additional structure on the Heegaard Floer package in the form of a homotopy involution $\iota$, leading to involutive Heegaard Floer homology. The construction of the invariant goes as follows: given a basepointed 3-manifold $(Y,w)$ together with a conjugation-invariant $\Spin^c$ structure $\mathfrak{s}$ on $Y$, there is a chain isomorphism
\[ \eta \colon \mathit{CF}^-(\mathcal{H}, \mathfrak{s}) \to \mathit{CF}^-(\overline{\mathcal{H}},\mathfrak{s})\]
\noindent where $\overline{\mathcal{H}}$ denotes the conjugate Heegaard diagram. One may then compose with the transition map
\[ \Psi_{\overline{\mathcal{H}} \to \mathcal{H}} \colon \mathit{CF}^-(\overline{\mathcal{H}}, \mathfrak{s}) \to \mathit{CF}^{-}(\mathcal{H}, \mathfrak{s})\]
\noindent obtaining the map $\iota$. This data is variously packaged either as the pair $(\mathit{CF}^{-}(Y), \iota)$, called an \emph{$\iota$-complex}, or as the mapping cone
\[
\mathit{CFI}^{-}(\mathcal{H}, \mathfrak{s}) = \Cone(\mathit{CF}^{-}(\mathcal{H},\mathfrak{s}) \xrightarrow{Q(1+\iota)} Q\cdot\mathit{CF}^-(\mathcal{H}, \mathfrak{s})[-1]).
\]
\noindent Here, $Q$ is a formal variable of degree $-1$, and we view the cone as a complex over $\mathbb F[U,Q]/Q^2$. Juh{\'a}sz-Thurston-Zemke naturality implies that the equivariant chain homotopy equivalence class of the pair $(\mathit{CF}^{-}(\mathcal{H},\mathfrak{s}), \iota)$, or equivalently the $\mathbb F[U,Q]/Q^2$ chain homotopy equivalence class of $\mathit{CFI}^{-}(\mathcal{H}, \mathfrak{s})$, is an invariant of $Y$ \cite[Section 2.1]{HMInvolutive}. However, this does not suffice to show that involutive Heegaard Floer homology is itself natural.
In this paper we prove that involutive Heegaard Floer homology is first-order natural. Our proof relies on equipping a Heegaard diagram $\mathcal{H}$ for a basepointed 3-manifold $(Y,w)$ with a set of \emph{doubling data} which is used to give a tractable model for the involution. We write $\mathfrak{D}$ for a Heegaard diagram $\mathcal{H}$ equipped with such a collection of data, and we refer to such a $\mathfrak{D}$ as a \emph{doubling enhanced Heegaard diagram}. We refer to the involutive Heegaard Floer homology as defined using the Heegaard diagram $\mathcal{H}$ with this model for the involution as $\mathit{CFI}^{-}(\mathfrak{D})$. (For more information on the doubling model for the involution, see Section \ref{subsec:doubling-models}.)
We also describe a set of moves which relate any two choices of doubling enhanced Heegaard diagrams. Given a sequence of such moves relating $\mathfrak{D}$ and $\mathfrak{D}'$, we define a transition map
\[ \Psi_{\mathfrak{D} \to \mathfrak{D}'}: \mathit{CFI}^{-}(\mathfrak{D}) \rightarrow \mathit{CFI}^{-}(\mathfrak{D}')\]
\noindent and show they are independent of the sequence chosen, as follows.
\begin{thm}\label{thm:final-naturality}
The transition map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ is independent up to $\mathbb{F}[U,Q]/Q^2$-equivariant chain homotopy from the sequence of moves between doubling enhanced Heegaard diagrams. Furthermore, $\Psi_{\mathfrak{D}\to \mathfrak{D}}\simeq \id$, and $\Psi_{\mathfrak{D}'\to \mathfrak{D}''}\circ \Psi_{\mathfrak{D}\to \mathfrak{D}'}\simeq \Psi_{\mathfrak{D}\to \mathfrak{D}''}$.
\end{thm}
\begin{rem}
Our involutive transition map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ contains data equivalent to a pair $(\Psi_{\mathcal{H}\to \mathcal{H}'},H_{\mathfrak{D}\to \mathfrak{D}'})$ where $\Psi_{\mathcal{H}\to \mathcal{H}'}$ is the non-involutive transition map from $\mathit{CF}^-(\mathcal{H})$ to $\mathit{CF}^-(\mathcal{H}')$, and $H_{\mathfrak{D}\to \mathfrak{D}'}$ is a distinguished chain-homotopy between $\Psi_{\mathcal{H}\to \mathcal{H}'}\iota_{\mathfrak{D}}$ and $\iota_{\mathfrak{D}'}\Psi_{\mathcal{H}\to \mathcal{H}'}$. In this manner, Theorem~\ref{thm:final-naturality} implies that the chain homotopy $H_{\mathfrak{D}\to \mathfrak{D}'}$ is also well-defined, up to a suitable notion of further chain-homotopy.
\end{rem}
Given a 4-dimensional cobordism $W$ with $\partial W = -Y_1 \coprod Y_2$ and a $\Spin^c$ structure $\mathfrak{s}$ on $W$, Ozsv{\'a}th and Szab{\'o} \cite{OSDisks} also constructed cobordism maps
\[
\mathit{CF}(W,\mathfrak{s}) : \mathit{CF}^{-}(Y_1, w_1,\mathfrak{s}|_{Y_1}) \rightarrow \mathit{CF}^{-}(Y_2, w_2, \mathfrak{s}|_{Y_2}).\]
\noindent These maps depend on a choice of path $\gamma$ connecting the basepoints in $Y_1$ and $Y_2$ but are otherwise independent of choices made in their definition \cite[Sections 8 and 9]{OSDisks}, cf. also \cite[Corollary F]{ZemGraphTQFT}.
In \cite{HMInvolutive}, the first author and Manolescu gave a construction of cobordism maps on involutive Heegaard Floer homology, but did not show their construction was invariant of a choice of handle decomposition of the cobordism. In this paper we refine the construction to give maps
\[
\mathit{CFI}(W,\mathfrak{s}): \mathit{CFI}^-(Y_1, w_1, \mathfrak{s}|_{Y_1}) \rightarrow \mathit{CFI}^-(Y_2,w_2,\mathfrak{s}|_{Y_2})
\]
\noindent associated to a cobordism $W$ from $Y_1$ to $Y_2$ and a conjugation-invariant $\Spin^c$ structure on $W$, and prove they are well-defined in the following sense.
\begin{thm}\label{thm:well-defined-cobordism-map}
The cobordism map $\mathit{CFI}(W,\mathfrak{s})$ is well-defined; that is, it depends only on the choice of $W$, $\mathfrak{s}$ and $\gamma$.
\end{thm}
Heegaard Floer homology also has a counterpart for knots and links in 3-manifolds, introduced by Ozsv{\'a}th and Szab{\'o} \cite{OSKnots} and independently in the case of knots by J.~Rasmussen \cite{RasmussenKnots}. In its modern form, this theory associates to a Heegaard diagram $\mathcal{H}$ for a link $L$ in a 3-manifold $Y$, with two basepoints $w_i$ and $z_i$ on each component, a chain complex $\mathcal{C\!F\!L}(\mathcal{H})$ over the ring $\mathbb F[\mathscr{U}_1, \mathscr{V}_1, \dots, \mathscr{U}_{\ell}, \mathscr{V}_{\ell}]$. This again decomposes over $\Spin^c$ structures on $Y$, such that $\mathcal{C\!F\!L}(\mathcal{H}) = \oplus_{\mathfrak{s} \in \Spin^c(Y)}\mathcal{C\!F\!L}(\mathcal{H},\mathfrak{s})$. As previously there are maps associated to sequence of Heegaard moves between Heegaard diagrams $\mathcal{H}$ and $\mathcal{H}'$ for $(Y, L, \ve{w},\ve{z})$
\[ \Psi_{\mathcal{H} \to \mathcal{H}'}: \mathcal{C\!F\!L}(\mathcal{H}, \mathfrak{s}) \rightarrow \mathcal{C\!F\!L}(\mathcal{H}', \mathfrak{s}).\]
These maps are again well-defined up to chain homotopy \cite[Theorem 1.8]{JTNaturality}.
The first author and Manolescu \cite{HMInvolutive} defined an endomorphism $\iota_L$ on $\mathcal{C\!F\!L}(\mathcal{H})$, called the link involution, in analogy with the procedure for defining the 3-manifold involution $\iota$, and showed that the equivariant chain homotopy equivalence class of the pair $\mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(\mathcal{H}) = (\mathcal{C\!F\!L}(\mathcal{H}),\iota_L)$ is an invariant of $(Y,L)$. (We note that \cite{HMInvolutive} focused on the case of knots, rather than links, but the same construction works for links). The doubling model for the involution $\iota$ on 3-manifolds may be adapted to a model for the endomorphism $\iota_L$. We use this model to prove naturality and functoriality of the link variant of Floer homology in the following sense. Once again, we let $\mathfrak{D}$ denote an appropriate set of doubling data for a Heegaard diagram $\mathcal{H}$ for $(Y, L, \ve{w}, \ve{z})$, and let $\mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(\mathfrak{D})$ refer to $\iota_L$-complex defined using this data. Given a sequence of Heegaard moves relating two doubling-enhanced Heegaard link diagrams, we define a transition map
\[
\Psi_{\mathfrak{D} \rightarrow \mathfrak{D}'} \colon \mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(\mathfrak{D}) \rightarrow \mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(\mathfrak{D}')
\]
\noindent which are enhanced $\iota_L$-homomorphisms. (Types of morphisms between $\iota_L$ complexes are reviewed in Section \ref{sec:knot-and-link-complexes}.) We then prove the following.
\begin{thm} \label{thm:link-naturality}
The transition maps $\Psi_{\mathfrak{D} \rightarrow \mathfrak{D}'}$ are well-defined up to homotopies of morphisms of $\iota_L$-complexes.
\end{thm}
The fourth author \cite{ZemCFLTQFT}, following work of Juhasz \cite{JCob}, defined maps associated to decorated link cobordisms $(W, \Sigma, \mathfrak{s})$ between $(Y_1,L_1, \ve{w}_1,\ve{z}_2,\mathfrak{s}|_{Y_1})$ and $(Y_2, L_2, \ve{w}_2,\ve{z}_2, \mathfrak{s}|_{Y_2})$
\[\mathcal{C\!F\!L}(W,\Sigma,\mathfrak{s}): \mathcal{C\!F\!L}(Y_1,L_1, \ve{w}_1,\ve{z}_1,\mathfrak{s}|_{Y_1}) \to \mathcal{C\!F\!L}(Y_2,L_2,\ve{w}_2,\ve{z}_2, \mathfrak{s}|_{Y_2})\]
\noindent which are again invariant of the choices involved in their definitions. Here the pair $(W,\Sigma)$ consists of a compact 4-manifold $W$ with embedded surface $\Sigma$ such that $W$ is a cobordism from $Y_1$ to $Y_2$ and $\partial \Sigma = -L_1 \coprod L_2$, together with some decorations on $\Sigma$. We will be interested in the case that $\Sigma$ is a collection of annuli, each with one boundary component in $Y_1$ and one in $Y_2$, and each decorated with two parallel longitudinal arcs. Maps associated to knot concordances or knot cobordisms have also appeared in \cite{JMConcordance, TNConcordance, AECobordism}.
Given a link cobordism $(W, \Sigma)$ from $(Y_1, L_1, \ve{w}, \ve{z})$ to $(Y_2, L_2, \ve{w}, \ve{z})$ consisting of a set of annuli each with two parallel longitudinal arcs as described above, and a $\Spin^c$ structure $\mathfrak{s}$ on $W$ with the property that $\bar{\mathfrak{s}} = \mathfrak{s} + PD([\Sigma])$, we construct cobordism maps
\[
\mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(W, \Sigma, \mathfrak{s}) \colon \mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(Y_1, L_1, \ve{w}, \ve{z}, \mathfrak{s}|_{Y_1}) \rightarrow \mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(Y_2, L_2, \ve{w}, \ve{z}, \mathfrak{s}|_{Y_2}).
\]
\noindent which are again enhanced homomorphisms of $\iota_L$-complexes. We prove the following functoriality theorem.
\begin{thm} \label{thm:link-functoriality}
The cobordism maps $\mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(W, \Sigma, \mathfrak{s})$ are well-defined up to homotopies of $\iota_L$-complexes.
\end{thm}
\subsection{Computations and examples}
In our paper, we also perform several example computations. The first example we present is for the cobordism $W=S^2\times S^2$, with its unique self-conjugate $\Spin^c$ structure $\mathfrak{s}$. If we remove two 4-balls from $W$, we obtain a cobordism from $S^3$ to $S^3$, and hence a map
\[
\mathit{CFI}(W,\mathfrak{s})\colon \mathbb{F}[U,Q]/Q^2\to \mathbb{F}[U,Q]/Q^2.
\]
In Section~\ref{sec:example}, we show via a direct diagrammatic computation that $\mathit{CFI}(W,\mathfrak{s})$ is multiplication by $Q$. This is agrees with the prediction from $\Pin(2)$-equivariant monopole Floer homology; see the proof of \cite{LinExact}*{Theorem~5}.
As a second example, we compute certain 2-handle cobordisms using the knot surgery formula from \cite{HHSZExact}. We recall that to a knot $K\subset S^3$, Ozsv\'ath and Szab\'{o} \cite{OSIntegerSurgeries}*{Theorem~1.1} described a complex $\mathbb{X}_n(K)$ which computes $\mathit{CF}^-(S^3_n(K))$. Therein they also described a way of computing the non-involutive cobordism map $\mathit{CF}(W_n(K),\mathfrak{s})$ in terms of $\mathbb{X}_n(K)$, where $W_n(K)$ denotes the natural 2-handle cobordism from $S^3$ to $S^3_n(K)$, for each $\mathfrak{s}\in \Spin^c(W_n(K))$. In \cite{HHSZExact}, we described an involutive refinement of Ozsv\'{a}th and Szab\'{o}'s mapping cone formula, which we denoted by $\mathbb{XI}_n(K)$, and which computes $\mathit{CFI}(S^3_n(K))$. In Section~\ref{sec:knot-surgery}, we describe a refinement of our work from \cite{HHSZExact} which computes certain cobordism maps. From the description in \cite{HHSZExact}, it turns out to be most convenient to consider the cobordism $W_n'(K)$ from $S^3_n(K)$ to $S^3$ obtained by reversing the orientation of $W_n(K)$. This manifold possesses a self-conjugate $\Spin^c$ structure if and only if $n$ is even. In Theorem~\ref{thm:knot-surgery-cobordism-map}, we describe an algebraic formula for the involutive cobordism map $\mathit{CFI}(W_{2n}'(K),\mathfrak{s})$ when $\mathfrak{s}$ is the unique self-conjugate $\Spin^c$ structure on $W_{2n}'(K)$.
\subsection{The doubling model of the involution and several variations} \label{subsec:doubling-models} The constructions of this paper rely on a model for the involution $\iota$, used in our previous paper \cite{HHSZExact}, based on the procedure of \emph{doubling} a Heegaard diagram. We now briefly recall this model. If $\Sigma$ is a Heegaard splitting of $Y$, containing a basepoint $w$, we construct another Heegaard splitting $D(\Sigma)$ of $Y$, with $D(\Sigma)\cong \Sigma\# \bar{\Sigma}$, which is embedded as the boundary of a regular neighborhood of $\Sigma\setminus N(w)$. A schematic appears in Figure \ref{fig:double}. The attaching curves for doubled diagrams are described in Section~\ref{sec:doubling-def}. An important property of the doubling operation is that if $\mathcal{H}$ is a diagram for $Y$, and $D(\mathcal{H})$ is a double, then the maps relating $\mathit{CF}^-(\mathcal{H})$ and $\mathit{CF}^-(D(\mathcal{H}))$ have a conceptually simple form, and similarly for the maps from $\mathit{CF}^-(D(\mathcal{H}))$ to $\mathit{CF}^-(\bar{\mathcal{H}})$; see Section \ref{sec:doubling-model-manifolds}. (Doubled Heegaard diagrams had previously been considered in \cite{ZemDuality}, \cite{JZContactHandles}, and \cite{JZStabilizationDistance}.)
\begin{figure}[H]
\centering
\input{double.pdf_tex}
\caption{Realizing the involution on $S^3$ by doubling.
}\label{fig:double}
\end{figure}
The proofs in the present paper require additional variations of this model for the involution, based on adding extra basepoints or extra tubes between the two copies of $\Sigma$ and $\bar \Sigma$. In the 3-manifold case, we begin by describing transition maps between the involutive complexes associated to Heegaard diagrams for a 3-manifold which are related by \emph{elementary equivalences} in terms of maps on the complexes associated to the doubled diagrams, and define maps for general handleslide equivalences as compositions of the maps for elementary equivalences. However, it is somewhat difficult to see that this construction fails to depend on the sequence of elementary equivalences chosen. To show this, we consider an expanded version of this model, in which an additional basepoint and pair of curves are added in the doubling region. We introduce this model in detail in Section~\ref{sec:expanded-def}. Using this model we define transition maps for handleslide equivalences which are more obviously invariant of the choices involved, and then use this to prove invariance of the transition maps for the standard doubling model. In the case of knots and links, our situation is somewhat more technically complex, so for simplicity we in fact work solely with the analog of the expanded model for the involution.
\subsection{Organization} This paper is organized as follows. In Section \ref{sec:algebra} we discuss some algebraic preliminaries and their Floer-theoretic counterparts; in particular, we discuss hypercubes of chain complexes and hypercubes of attaching curves for a Heegaard surface. We additionally review the algebraic formalism of $\iota$-complexes and their relationship to hypercubes of chain complexes. In Section \ref{sec:involutive} we review some background on involutive Heegaard Floer homology, and in particular recall the operation of doubling a Heegaard diagram and the resulting model for the involution on Heegaard Floer homology from \cite{HHSZExact}. In Section \ref{sec:elementary-equivalences} we define transition maps for Heegaard diagrams related by elementary equivalences, and define general transition maps for Heegaard diagrams related by handleslide equivalences as compositions of maps for elementary equivalences. We prove in Proposition \ref{prop:continuity} that these maps satisfy the continuity axiom, in analogy with \cite[Proposition 9.27]{JTNaturality}. In Section \ref{sec:polygons} we prove some necessary technical results concerning stabilizations and holomorphic polygons, generalizing the results from our previous paper \cite{HHSZExact}. In Section \ref{sec:expanded-def} we introduce the basepoint exanded doubling diagrams and the basepoint expanded model of the involution, and construct transition maps for handleslide equivalences using this model. In Section \ref{sec:relate-expanded} we define a chain homotopy equivalence between involutive chain complexes associated to the ordinary and basepoint expanded doubles of a Heegaard diagram, and prove that these maps commute with the transitions maps associated to handleslide equivalences. We further show that the transition maps defined in Section \ref{sec:expanded-def} for generalized handleslide equivalences are unique up to chain homotopy equivalence, which proves the result for the non-expanded model. We then turn our attention to constructing the maps between involutive complexes associated to cobordism; in Section \ref{sec:1-and-3-handles} we construct the maps associated to one- and three-handles and in Section \ref{sec:2-handles} we construct the maps associated to two-handles. In Section \ref{sec:stable} we define construct naturality maps for stabilizations as compositions of cobordism maps associated to cancelling handles and prove they commute with the transition maps associated to handleslide equivalences and with each other. In Section \ref{sec:handleswap} we prove invariance of our transition maps under handleswaps, in analogy with \cite[Section 9.3]{JTNaturality}. Finally, in Section~\ref{sec:final-overview} we complete the proofs of Theorem~\ref{thm:final-naturality}, following the structure of the proof in \cite[Section 9.4]{JTNaturality}, and of Theorem~\ref{thm:well-defined-cobordism-map}. As an example, in Section \ref{sec:example} we present the calculation of the cobordism map for a twice punctured copy of $S^2\times S^2$, decomposed as two handle $2$-handle cobordisms $S^3\to S^1\times S^2$ and $S^1\times S^2\to S^3$. In Section \ref{sec:knots} we describe how to adapt the naturality and functoriality results of the previous sections to the case of knots and links, and prove Theorems \ref{thm:link-naturality} and \ref{thm:link-functoriality}. We then conclude in Section \ref{sec:knot-surgery} with a computation of the involutive cobordism map associated to the cobordism $W'_n(K)$ from $S^3_n(K)$ to $S^3$ for $n$ even in terms of the involutive mapping cone formula.
\subsection*{Acknowledgments} The authors are grateful to Andr{\'a}s Juh{\'a}sz, Ciprian Manolescu, and Dylan Thurston for helpful conversations (over many years).
\section{Involutive Heegaard Floer homology} \label{sec:involutive}
In this section, we review Hendricks and Manolescu's construction of involutive Heegaard Floer homology \cite{HMInvolutive}, and also define the doubling models which feature in our statements of naturality and functoriality.
\subsection{The involutive Floer complexes}
We presently recall Hendricks and Manolescu's original construction \cite{HMInvolutive}. Suppose that $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},w)$ is a weakly admissible Heegaard diagram for $Y$. Suppose that $\mathfrak{s}$ is a self-conjugate $\Spin^c$ structure. Write $\bar \mathcal{H}=(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\alpha}, w)$ for the diagram obtained from $\mathcal{H}$ by reversing the orientation of $\Sigma$ and reversing the roles of $\ve{\alpha}$ and $\ve{\beta}$. There is a canonical chain isomorphism
\[
\eta\colon \mathit{CF}^-(\bar{\mathcal{H}},\mathfrak{s})\to \mathit{CF}^-(\mathcal{H},\mathfrak{s}).
\]
Hendricks and Manolescu consider the map
\[
\iota \colon \mathit{CF}^-(\mathcal{H},\mathfrak{s})\to \mathit{CF}^-(\mathcal{H},\mathfrak{s})
\]
given by the formula
\[
\iota := \eta \circ \Psi_{\mathcal{H}\to \bar {\mathcal{H}}},
\]
where $\Psi_{\mathcal{H}\to \bar \mathcal{H}}$ is the map from naturality as in \cite{JTNaturality}. They define the involutive Heegaard Floer complex $\mathit{CFI}(Y,\mathfrak{s})$ to be the $\mathbb{F}[U,Q]/(Q^2)$-chain complex whose underlying $\mathbb{F}[U,Q]/(Q^2)$-module is
\[
\mathit{CF}^-(Y,\mathfrak{s})\otimes_{\mathbb{F}[U]} \mathbb{F}[U,Q]/(Q^2)=\mathit{CF}^-(Y,\mathfrak{s})\oplus Q\mathit{CF}^-(Y,\mathfrak{s})
\] with differential $\partial_{\mathit{CFI}(Y,\mathfrak{s})}=\partial_{\mathit{CF}^-(Y,\mathfrak{s})}\otimes \id +(\id+\iota)\otimes Q$. That is, the involutive Heegaard Floer complex is the mapping cone
\[
\Cone\left(Q(\id+\iota)\colon \mathit{CF}^-(Y,\mathfrak{s})\to Q \mathit{CF}^-(Y,\mathfrak{s})\right).
\]
with the evident action of $\mathbb{F}[U,Q]/(Q^2)$. Hendricks and Manolescu prove that $\mathit{CFI}^-(Y,\mathfrak{s})$ is well-defined up to chain homotopy equivalence.
Hendricks and Manolescu also define a refinement for knots, which we recall presently along with its extension to links. Beginning with the knot case, suppose that $K$ is a null-homologous, oriented knot in a 3-manifold $Y$, and $\mathfrak{s}\in \Spin^c(Y)$ is self-conjugate. We recall the following standard definition.
\begin{define} A Heegaard diagram for a pair $(Y,K)$ consists of a tuple $(\Sigma,\ve{\alpha},\ve{\beta},w,z)$, such that $(\Sigma,\ve{\alpha},\ve{\beta})$ is a Heegaard diagram for $Y$, such that the knot $K$ intersects $\Sigma$ in the points $w$ and $z$. Furthermore, if $U_{\a}$ and $U_{\b}$ are the closures of the two components of $Y\setminus \Sigma$, then $K$ intersects each of $U_{\a}$ and $U_{\b}$ in a boundary-parallel arc. The standard convention is that $K$ intersects $\Sigma$ positively at $z$, and negatively at $w$.
\end{define}
We will work over the version of knot Floer homology \cite{OSKnots} \cite{RasmussenKnots} which is freely generated over a 2-variable polynomial ring $\mathbb{F}[\mathscr{U},\mathscr{V}]$. We denote this chain complex by $\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(K)$. See, e.g. \cite{ZemConnectedSums}*{Section~3} for background on this version of knot Floer homology.
Let $\phi^+ \colon (Y,K,w,z)\to (Y,K,z,w)$ be the diffeomorphism of tuples supported in a neighborhood of $K$ corresponding to a positive half-twist along $K$, and similarly for $\phi^{-}$ and the negative half-twist. There is a canonical chain isomorphism
\[
\eta_K\colon\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\bar\Sigma,\bar \ve{\beta}, \bar \ve{\alpha},z,w)\to \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\Sigma, \ve{\alpha}, \ve{\beta},w,z)
\]
which satisfies $\eta_K(\mathscr{U}\cdot x)=\mathscr{V} \cdot \eta_K(x)$ and $\eta_K(\mathscr{V} \cdot x)=\mathscr{U}\cdot \eta_K(x)$. We define
\[
\iota_{K,+}:=\eta_K\circ \Psi_{\phi^+(\mathcal{H})\to \bar \mathcal{H}} \circ \phi_*^+,
\]
and we define $\iota_{K,-}$ similarly. Here, $\phi_*^+$ denotes the tautological diffeomorphism map, and $\Psi_{\phi^+(\mathcal{H})\to \bar \mathcal{H}}$ denotes the map from naturality obtained by picking a sequence of Heegaard moves relating $\phi^+(\mathcal{H})$ and $\bar \mathcal{H}$. There is a map $\iota_K^-$ defined similarly. Hendricks and Manolescu define the involutive knot Floer complex to be the data consisting of $\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(K)$ equipped with the endomorphism $\iota_{K,\pm}$, which gives an $\iota_K$-complex.
\begin{rem}
In \cite{HMInvolutive}, Hendricks and Manolescu considered only the involution we call $\iota_{K,+}$, and wrote simply $\iota_K$. It is unknown whether there are any knots $K$ where $(\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(K), \iota_{K,+})$ is inequivalent to $(\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(K),\iota_{K,-})$. The distinction between $\iota_{K,+}$ and $\iota_{K,-}$ has previously been observed in \cite{BKST:genus}.
\end{rem}
The case of links is similar. For a link $L$ of $\ell=|L|$ components, we consider Heegaard diagrams $(\Sigma, \ve{\alpha}, \ve{\beta}, \ve{w}, \ve{z})$ such that $\ve{w} = \{w_1,\dots, w_n\}$ and $\ve{z} = \{z_1, \dots, z_{\ell}\}$ and the $i$-th component $K_i$ of $L$ intersects $\Sigma$ positively at $z_i$ and negatively at $w_i$. The resulting chain complex $\mathcal{C\!F\!L}(L)$ is freely-generated over the polynomial ring $\mathbb{F}[\mathscr{U}_1, \mathscr{V}_1, \dots, \mathscr{U}_{\ell}, \mathscr{V}_{\ell}]$. In principle there are $2^{\ell}$ choices of orientation which may be used to define the involution; however for simplicity we assume that an orientation is fixed, and we consider only the two orientations which are either coherent, or opposite to our preferred orientation. It is now straightforward to adapt the construction of $\iota_{K,\pm}$ given above to models for $\iota_{L,\pm}$ for these versions of the link involution.
\subsection{Doubled Heegaard diagrams}
\label{sec:doubling-def}
Suppose that $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},w)$ is a pointed Heegaard diagram for $Y$. We view a regular neighborhood of $\Sigma$ as $\Sigma\times [0,1]$. Pick a small disk $D\subset \Sigma$, which contains $w$ along its boundary. The submanifold $U_0=(\Sigma\setminus D)\times [0,1]$ is a handlebody of genus $2g$, where $g$ denotes $g(\Sigma)$. The boundary of this submanifold is canonically identified with $\Sigma\# \bar \Sigma$, where the connected sum occurs at $w$. Write $U_1$ for the closure of the complement of $U_0$. Clearly $U_1$ is also a handlebody of genus $2g$. We write $D(\Sigma)=\Sigma\# \bar \Sigma$ for this Heegaard surface.
We may naturally equip $D(\Sigma)$ with attaching curves, as follows. For $U_1$, we may use the curves $\ve{\alpha} \cup \bar \ve{\beta}$, where $\ve{\alpha}\subset \Sigma$ and $\bar\ve{\beta} \subset \bar \Sigma$. For $U_0$, we pick compressing disks by picking a collection $\delta_1,\dots, \delta_{2g}$ of properly embedded arcs on $\Sigma\setminus N(w)$. We add a basepoint to $\d N(w)$, for which we also write $w$. We assume that the arcs $\delta_1,\dots, \delta_{2g}$ are disjoint from $w$, and also form a basis of $H_1(\Sigma\setminus D, \d D\setminus\{w\})$. We form attaching curves $\ve{\Delta}\subset \Sigma\# \bar \Sigma$ by doubling the curves $\delta_1,\dots, \delta_{2g}$ onto $\Sigma\# \bar \Sigma.$ We write $D(\mathcal{H})=(\Sigma \# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\Delta},w)$ for the resulting diagram.
\begin{rem}
The diagram $D(\mathcal{H})$ may also be described by gluing two bordered Heegaard diagrams for genus $g$ handlebodies \cite{LOTBordered}.
\end{rem}
\subsection{Doubling and the involution} \label{sec:doubling-model-manifolds}
We use doubled Heegaard diagrams to give a natural model of the involution. This model appears in \cite{ZemDuality} and \cite{HHSZExact}. In this section, we describe the construction.
Firstly, we define the following map as a composition of 1-handle maps
\[
F_1^{\bar \b,\bar\b}\colon \mathit{CF}^-(\Sigma,\ve{\alpha},\ve{\beta},w)\to \mathit{CF}^-(\Sigma\# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\beta}\bar \ve{\beta}, w).
\]
Next, we note that the diagram $(\Sigma\# \bar \Sigma, \ve{\beta}\bar \ve{\beta},\ve{\Delta},w)$ is a doubled diagram for $(S^1\times S^2)^{\# g}$, and hence we may pick a cycle $\Theta_{\b\bar \b,\Delta}$ generating the top degree of homology. The cycle $\Theta_{\b\bar \b, \Delta}$ is not unique, but any two choices are related by a boundary. We define the holomorphic triangle map
\[
f_{\a\bar \b}^{\b \bar \b\to \Delta}\colon \mathit{CF}^-(\Sigma\# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta}, w)\to \mathit{CF}^-(\Sigma\# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\Delta},w)
\]
by the formula $f_{\a \bar \b}^{\b \bar \b\to \Delta}(x)=f_{\a \bar \b, \b \bar \b,\Delta}(x,\Theta_{\b \bar \b,\Delta}).$
Analogously, there is another holomorphic triangle map
\[
f_{\a \bar \b}^{\Delta\to \a \bar \a}\colon \mathit{CF}^-(\Sigma \# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\Delta}, w)\to \mathit{CF}^-(\Sigma\# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha},w)
\]
as well as a 3-handle map
\[
F_3^{\a,\a}\colon \mathit{CF}^-(\Sigma \# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}, w)\to \mathit{CF}^-(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\alpha},w).
\]
\begin{prop}[\cite{ZemDuality}*{Propositions~7.2, 7.8}]
\label{prop:doubling-involution}
The composition $f_{\a \bar \b}^{\b \bar \b\to \Delta} \circ F_1^{\bar \b, \bar\b}$ is chain homotopic to the map $\Psi_{\mathcal{H} \to D(\mathcal{H})}$ from naturality. Dually, the composition $F_3^{\a,\a}\circ f_{\a \bar \b}^{\Delta\to \a \bar \a}$ is chain homotopic to the map $\Psi_{D(\mathcal{H})\to \bar \mathcal{H}}$.
\end{prop}
The idea of the proof of the above result is that the composition is topologically equivalent to a composition of the maps for canceling 1-handles and 2-handles, or canceling 2-handles and 3-handles.
\begin{cor} \label{cor:doubling-model}
The involution satisfies
\[
\iota \simeq \eta\circ F_3^{\a,\a} \circ f_{\a \bar \b}^{\Delta \to \a \bar \a}\circ f_{\a \bar \b}^{\b \bar \b \to \Delta} \circ F_1^{\bar \b, \bar \b},
\]
where $\eta$ is the canonical chain isomorphism
\[
\eta\colon \mathit{CF}^-(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\alpha},w)\to \mathit{CF}^-(\Sigma,\ve{\alpha},\ve{\beta},w).
\]
\end{cor}
\begin{define}
A \emph{doubling enhanced Heegaard diagram} $\mathfrak{D}$ consists of the following data:
\begin{enumerate}
\item A Heegaard diagram $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},w)$.
\item A set of attaching curves $\ve{\Delta}$ on $\Sigma\# \bar \Sigma$, constructed by doubling a basis of arcs $\delta_1,\dots, \delta_{2g}$.
\item Choices of almost complex structures to compute the four Floer complexes $\mathit{CF}(\Sigma,\ve{\alpha},\ve{\beta})$, $\mathit{CF}(\Sigma \# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})$, $\mathit{CF}(\Sigma\# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\Delta})$ and $\mathit{CF}(\Sigma\# \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar\ve{\alpha})$, as well as almost complex structures to compute the triangle maps $f_{\a \bar \b}^{\Delta \to \a \bar \a}$ and $f_{\a \bar \b}^{\b \bar \b \to \Delta}$.
\end{enumerate}
We say that $\mathfrak{D}$ is a \emph{doubling enhancement} of $\mathcal{H}$. We say that $\mathfrak{D}$ is \emph{admissible} if each of the Heegaard diagrams and triples are weakly admissible.
\end{define}
\subsection{Doubling and the knot and link involution}
\label{sec:doubling-knots}
The knot involution of \cite{HMInvolutive} also admits a convenient doubling model, which has appeared in \cite{HHSZExact} and \cite{JZStabilizationDistance}*{Section~9.4}.
The doubling model for knots has a similar description to Corollary~\ref{cor:doubling-model}, but more care will be needed to define the maps. There is furthermore one additional choice to be made, which will correspond to a choice of which direction along $K$ to perform a half-twist.
We begin by describing doubled knot diagrams. If $(\Sigma,\ve{\alpha},\ve{\beta},w,z)$ is a diagram for $(Y,K)$, we consider the following two diagrams:
\[
D^+(\mathcal{H})=(\Sigma\#_z \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\Delta}, w ,\bar w)\quad \text{and} \quad D^-(\mathcal{H})=(\Sigma\#_w \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\Delta},\bar z, z).
\]
We start by analyzing the diagram $\mathcal{D}^+(\mathcal{H}).$ There is a map
\[
F_{1,+}^{\bar \b, \bar \b}\colon \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\Sigma,\ve{\alpha},\ve{\beta},w,z)\to \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\Sigma\#_z \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta}, w, \bar w)
\]
obtained as the composition of a 1-handle map, as well as a diffeomorphism map for moving $z$ to $\bar w$. As in the setting of closed 3-manifolds, we have holomorphic triangle maps
$f_{\a \bar \b}^{\b \bar \b\to \Delta}$ and $f_{\a \bar \b}^{\Delta\to \a \bar \a}$. Finally, there is a map
\[
F_{3,+}^{\a,\a}\colon \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\Sigma\#_z \bar \Sigma, \ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}, w, \bar w)\to \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\bar\Sigma, \bar \ve{\beta}, \bar \ve{\alpha}, \bar z, \bar w),
\]
obtained by moving $w$ to the position of $\bar z$, and then applying the 3-handle maps to remove the $\ve{\alpha}$ curves. Similar logic to Proposition~\ref{prop:doubling-involution} then implies that
\[
\iota_{K,+}\simeq \eta_K\circ F_{3,+}^{\a,\a}\circ f_{\a \bar \b}^{\Delta \to \a \bar \a}\circ f_{\a \bar \b}^{\b \bar \b\to \Delta}\circ F_{1,+}^{\bar \b, \bar \b}.
\]
\noindent There are analogous maps $F_{1,-}^{\bar \b, \bar \b}$ and $F_{3,-}^{\a, \a}$ which instead involve the diagram $D^-(\mathcal{H})$, such that
\[
\iota_{K,-}\simeq \eta_K\circ F_{3,-}^{\a,\a}\circ f_{\a \bar \b}^{\Delta \to \a \bar \a}\circ f_{\a \bar \b}^{\b \bar \b\to \Delta}\circ F_{1,-}^{\bar \b, \bar \b}.
\]
The same procedure with only notational changes may be used to give doubling models for the link involutions $\iota_{L^,\pm}$ for links with more than one component. See Section~\ref{sec:knots} for more details.
\subsection{Alpha versus beta doubling}
The reader may note that a somewhat arbitrary choice has been made in the definition of $\iota$. Indeed, there is another diagram $\bar{D}(\mathcal{H})=(\Sigma\# \bar \Sigma, \ve{\Delta}, \ve{\beta} \bar \ve{\alpha},w)$, conjugate to the diagram $D(\mathcal{H})$ considered above. Similar to Proposition~\ref{prop:doubling-involution} and using the same notation from that proposition, $f^{\b\bar\a}_{\a\bar\a\to \Delta}\circ F_1^{\bar\a,\bar\a}$ is also chain homotopic to $\Psi_{\mathcal{H}\to \bar D(\mathcal{H})}$. Dually, $F_3^{\b,\b}\circ f^{\b\bar\a}_{\Delta\to\b\bar\b}$ is homotopic to $\Psi_{\bar D(\mathcal{H})\to \bar{\mathcal{H}}}$. As a consequence, parallel to Corollary \ref{cor:doubling-model},
\begin{equation}
\iota \simeq \eta\circ F_3^{\b,\b} \circ f_{\Delta\to \b \bar \b}^{\b\bar\a}\circ f_{\a \bar \a\to \Delta}^{\b \bar \a} \circ F_1^{\bar \a, \bar \a}.\label{eq:alpha-doubling}
\end{equation}
We will refer to the composite on the right side of equation~\eqref{eq:alpha-doubling} as the \emph{alpha-doubling model} for $\iota$, and we refer to the model $\eta\circ F_3^{\a,\a} \circ f_{\a \bar \b}^{\Delta \to \a \bar \a}\circ f_{\a \bar \b}^{\b \bar \b \to \Delta} \circ F_1^{\bar \b, \bar \b}$ from Corollary~\ref{cor:doubling-model} as the \emph{beta-doubling} model. In spite of the fact that the alpha-doubling and beta-doubling models are homotopic by naturality, we do not claim there is a canonical homotopy between them. In this paper, we arbitrarily work over the beta-doubling model exclusively.
\section{Maps for elementary handleslide equivalences}
\label{sec:elementary-equivalences}
In this section, we define the transition maps for changes of the doubling data which fix the Heegaard surface $\Sigma$. These correspond to changes of the $\ve{\alpha}$, $\ve{\beta}$ and $\ve{\Delta}$ curves by handleslides and isotopies. Furthermore, in this section, we consider only a restricted set of alpha and beta equivalences, as follows.
\begin{define}\,
\begin{enumerate}
\item Suppose that $(\Sigma,\ve{w})$ is a Heegaard surface with $n\ge 1$ basepoints and $\ve{\gamma}$ and $\ve{\gamma}'$ are attaching curves on $\Sigma$. We say that $\ve{\gamma}$ and $\ve{\gamma}'$ are \emph{handleslide equivalent} if they may be related by a sequence of handleslides and isotopies.
\item Suppose $\Sigma$ is a Heegaard surface and $\ve{\gamma}$ and $\ve{\gamma}'$ are handleslide equivalent attaching curves on $\Sigma$. We say that $\ve{\gamma}$ and $\ve{\gamma}'$ are related by an \emph{elementary equivalence} if $(\Sigma,\ve{\gamma}',\ve{\gamma},\ve{w})$ is a weakly admissible diagram, and $\mathbb{T}_{\gamma'}\cap \mathbb{T}_{\gamma}$ consists of exactly $2^{|\gamma|-1}$ generators.
\end{enumerate}
\end{define}
Suppose $(\Sigma,\ve{\alpha},\ve{\beta})$ is a Heegaard diagram, and $\ve{\alpha}'$ and $\ve{\beta}'$ are handleslide equivalent to $\ve{\alpha}$ and $\ve{\beta}$, respectively. Suppose further that $\ve{\Delta}$ and $\ve{\Delta}'$ are choices of doubling curves. Let $\mathfrak{D}$ denote the data consisting of $(\Sigma,\ve{\alpha},\ve{\beta})$, the curves $\ve{\Delta}$, and appropriate choices of almost complex structures to compute the involution. Let $\mathfrak{D}'$ denote the analogous data with $\ve{\alpha}'$, $\ve{\beta}'$ and $\ve{\Delta}'$. We say that $\mathfrak{D}'$ is obtained from $\mathfrak{D}$ by an \emph{elementary equivalence} if $\ve{\alpha}'$ and $\ve{\beta}'$ are obtained from $\ve{\alpha}$ and $\ve{\beta}$ by elementary equivalences, and also if the tuple
\[
(\Sigma\# \bar \Sigma,\ve{\alpha}'\bar \ve{\alpha}',\ve{\alpha}\bar \ve{\alpha},\ve{\beta}\bar \ve{\beta},\ve{\beta}'\bar \ve{\beta}',\ve{\Delta},\ve{\Delta}',w)
\]
is weakly admissible.
In the case that $\mathfrak{D}'$ is obtained from $\mathfrak{D}$ by an elementary equivalence, we define the transition map
\[
\Psi_{\mathfrak{D}\to \mathfrak{D}'}\colon \mathit{CFI}(\mathfrak{D})\to \mathit{CFI}(\mathfrak{D}')
\]
as the compression of the hyperbox of chain complexes shown in Figure~\ref{def:transition-map-elementary-handleslide}. Therein, the rows with the 1-handle maps are constructed similarly to \cite{HHSZExact}*{Section~14}, where they are called \emph{hypercubes of stabilization}. The construction here requires that the $\ve{\alpha}'$ be obtained from $\ve{\alpha}$ by an elementary equivalence, and similar for $\ve{\beta}$ and $\ve{\beta}'$, so that the degeneration argument from \cite{HHSZExact}*{Section~10} applies. The rows with 3-handles are similar. The remaining subcubes are obtained by pairing hypercubes of attaching curves.
Note that in Figure~\ref{def:transition-map-elementary-handleslide} we have omitted a final level involving the tautological map $\eta\colon \mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha})\to \mathit{CF}(\ve{\alpha},\ve{\beta})$. Throughout the paper, we will usually omit this row.
\begin{figure}[h]
\[
\Psi_{\mathfrak{D}\to \mathfrak{D}'}:=\quad
\begin{tikzcd}[labels=description, row sep=1cm]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{\bar{\b},\bar{\b}}"]
& \mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{\bar{\b}',\bar{\b}}"]
& \mathit{CF}(\ve{\alpha}',\ve{\beta}')
\mathrm{a.r.}[d,"F_{1}^{\bar{\b}',\bar{\b}'}"]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}'\bar{\ve{\beta}}', \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}' \bar{\ve{\beta}}', \ve{\beta}' \bar{\ve{\beta}}')
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}' \bar \ve{\beta}',\ve{\Delta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}'\bar \ve{\beta}',\ve{\Delta}')
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\alpha} \bar{\ve{\alpha}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a,\a }"]
&\mathit{CF}(\ve{\alpha}' \bar{\ve{\beta}}', \ve{\alpha} \bar{\ve{\alpha}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a',\a }"]
& \mathit{CF}(\ve{\alpha}' \bar{\ve{\beta}}', \ve{\alpha}' \bar{\ve{\alpha}}')
\mathrm{a.r.}[d, "F_3^{\a',\a'}"]
\\
\mathit{CF}(\bar{\ve{\beta}}, \bar{\ve{\alpha}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d,"\id"]
\mathrm{a.r.}[drr,dashed]
& \mathit{CF}(\bar{\ve{\beta}}',\bar{\ve{\alpha}})
\mathrm{a.r.}[r]
& \mathit{CF}(\bar{\ve{\beta}}',\bar{\ve{\alpha}}')
\mathrm{a.r.}[d,"\id"]
\\
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha})
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta}', \bar \ve{\alpha}')
\end{tikzcd}
\]
\caption{The transition map for an elementary equivalence. An additional row involving the canonical map $\eta$ is omitted.}
\label{def:transition-map-elementary-handleslide}
\end{figure}
If $\mathfrak{D}$ and $\mathfrak{D}'$ are handleslide equivalent (but perhaps not related by an elementary equivalence), we pick an arbitrary sequence
\[
\mathfrak{D}=\mathfrak{D}_1,\dots, \mathfrak{D}_n=\mathfrak{D}'
\]
such that $\mathfrak{D}_{i+1}$ and $\mathfrak{D}_i$ are related by an elementary equivalence for all $i$. We define
\begin{equation}
\Psi_{\mathfrak{D}\to \mathfrak{D}'}:=\Psi_{\mathfrak{D}_{n-1}\to \mathfrak{D}_n}\circ \cdots \circ \Psi_{\mathfrak{D}_1\to \mathfrak{D}_2}.
\label{eq:definition-transition-map-compose-elementary}
\end{equation}
A key step in our proof of naturality will be proving the following theorem:
\begin{thm}
\label{thm:naturality-fixed-Sigma}
Suppose that $\mathfrak{D}$, $\mathfrak{D}'$ and $\mathfrak{D}''$ are doubling enhanced Heegaard diagrams which all have the same underlying Heegaard surface $\Sigma$ and are handleslide equivalent.
\begin{enumerate}
\item\label{thm:naturality-fixed-Sigma-1} The map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ defined in~\eqref{eq:definition-transition-map-compose-elementary} is independent from the intermediate sequence of doubling enhanced Heegaard diagrams, up to $\mathbb{F}[U,Q]/Q^2$-equivariant chain homotopy.
\item\label{thm:naturality-fixed-Sigma-2} $\Psi_{\mathfrak{D}\to \mathfrak{D}}\simeq \id_{\mathit{CFI}(\mathfrak{D})}.$
\item\label{thm:naturality-fixed-Sigma-3} $\Psi_{\mathfrak{D}'\to \mathfrak{D}''}\circ \Psi_{\mathfrak{D}\to \mathfrak{D}'}\simeq \Psi_{\mathfrak{D}\to \mathfrak{D}''}$.
\end{enumerate}
\end{thm}
Part~\eqref{thm:naturality-fixed-Sigma-2} of Theorem~\ref{thm:naturality-fixed-Sigma} is proven in the subsequent Section~\ref{sec:continuity}. Parts~\eqref{thm:naturality-fixed-Sigma-1} and \eqref{thm:naturality-fixed-Sigma-3} are more technical and are proven in Sections \ref{sec:expanded-def} and ~\ref{sec:relate-expanded}; in particular, they follow from Proposition~\ref{prop:expansion}.
\begin{rem}
We are only able to define the map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ using Figure~\ref{def:transition-map-elementary-handleslide} when $\mathfrak{D}$ and $\mathfrak{D}'$ are related by an elementary equivalence because we are not able to construct the 1-handle hypercubes if $\mathfrak{D}$ and $\mathfrak{D}'$ are related by an arbitrary handleslide equivalence. To show the map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ is independent of the choice of intermediate diagrams, we will define a more complicated transition map $\tilde{\Psi}_{\mathfrak{D}\to \mathfrak{D}'}$ in Section~\ref{sec:maps-expanded-model}.
\end{rem}
\subsection{Continuity}
\label{sec:continuity}
In this section, we prove the continuity axiom from \cite{JTNaturality}*{Definition~2.32} for our maps associated to a handleslide equivalence. In our present context, the continuity axiom amounts to the following:
\begin{prop}
\label{prop:continuity}
Let $\mathfrak{D}$ denote an admissible doubling enhanced Heegaard diagram for $(Y,w)$. Let $g\colon (\Sigma,w)\to (\Sigma,w)$ be a pointed diffeomorphism which is smoothly isotopic to the identity. Let $g(\mathfrak{D})$ be the push-forward of the diagram $\mathfrak{D}$ under $g$, and let
\[
\mathit{CFI}(g)\colon \mathit{CFI}(\mathfrak{D})\to \mathit{CFI}(g(\mathfrak{D}))
\]
denote the tautological map. Then
\[
\Psi_{\mathfrak{D}\to g(\mathfrak{D})}\simeq \mathit{CFI}(g).
\]
\end{prop}
\begin{proof}
We adapt the argument of \cite{JTNaturality}*{Proposition~9.27} to the involutive case by also using the small translate theorems for triangles and quadrilaterals from our previous paper \cite{HHSZExact}. Write $(\Sigma,\ve{\alpha},\ve{\beta},w)$ for the underlying Heegaard diagram of $\mathfrak{D}$, and let $\ve{\Delta}$ be the doubling curves. As a first step, pick a $C^\infty$ small diffeomorphism $g_0$ of $(\Sigma,w)$ and write $\ve{\alpha}'=g_0(\ve{\alpha})$, and define $\ve{\beta}'$ and $\ve{\Delta}'$ similarly. Assume that $|\a'_i\cap \a_j|=2 \delta_{ij}$, where $\delta_{ij}$ denotes the Kronecker delta, and assume the analogous statement holds for $\ve{\beta}'$, $\ve{\beta}$, $\ve{\Delta}'$ and $\ve{\Delta}$. It follows from the small translate theorems \cite[Propositions 11.1 and 11.5]{HHSZExact} that if $g_0$ is chosen suitably small, then we may pick an almost complex structure on $\Sigma\times \Delta$ which interpolates between a fixed $J$ on the $(\Sigma,\ve{\alpha},\ve{\beta})$ cylindrical end and $(g_0)_*(J)$ on the $(\Sigma,\ve{\alpha}',\ve{\beta})$ cylindrical end such that the map $f_{\a',\a,\b}(\Theta_{\a',\a},\ve{x})$ counts only small triangles, so that $f_{\a',\a,\b}(\Theta_{\a',\a},\ve{x})=\ve{x}_{np}$ for all $\ve{x}\in \mathbb{T}_{\a}\cap \mathbb{T}_{\b}$. Here $\ve{x}_{np}$ denotes the nearest point of $\mathbb{T}_{\a'}\cap \mathbb{T}_{\b}$ to $\ve{x}$. In a similar manner, the small translate theorem for quadrilaterals implies that all of the length 2 maps counted by $\Psi_{\mathfrak{D}\to g_0(\mathfrak{D})}$ vanish (as no holomorphic curves are counted). In particular, $\Psi_{\mathfrak{D}\to g_0(\mathfrak{D})}$ coincides with the nearest point map.
Similarly, if $g_0'$ is $C^\infty$ close to $g_0$, then the same argument as the small translate theorems also imply that $\Psi_{\mathfrak{D}\to g_0'(\mathfrak{D})}$ and $\Psi_{g_0'(\mathfrak{D})\to \mathfrak{D}}$ coincide with the nearest point maps.
More generally, if $g$ is $C^\infty$ small relative to $g_0$, then we can decompose
\[
\Psi_{\mathfrak{D}\to g(\mathfrak{D})}\simeq \Psi_{g_0(\mathfrak{D})\to g(\mathfrak{D})}\circ \Psi_{\mathfrak{D}\to g_0(\mathfrak{D})}
\]
and apply the nearest point results to the latter two maps to identify $\Psi_{\mathfrak{D}\to g(\mathfrak{D})}$ with $\mathit{CFI}(g)$. This establishes the claim when $g$ is $C^\infty$-small. Since both sides of the desired equality are functorial under composition, and any diffeomorphism $g\colon (\Sigma,w)\to (\Sigma,w)$ which is smoothly isotopic to the identity may be decomposed into a composition of $C^\infty$-small diffeomorphisms, this proves the claim in general.
\end{proof}
\section{Completing the proofs of naturality and functoriality} \label{sec:final-overview}
In this section, we complete our proofs of naturality and functoriality.
\subsection{Naturality}
If $\mathfrak{D}$ and $\mathfrak{D}'$ are two arbitrary doubling-enhanced Heegaard diagrams, we define our transition map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ by composing the maps for an arbitrary sequence stabilizations, elementary equivalences of attaching curves, and pointed isotopies of the Heegaard diagram in $(Y,w)$.
We now prove Theorem \ref{thm:final-naturality}, showing that this transition map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ defined above is independent up to chain homotopy from the sequence of Heegaard moves.
\begin{proof}[Proof of Theorem \ref{thm:final-naturality}] By \cite{JTNaturality}*{Theorem~2.38}, it suffices to show that the transition maps we have defined satisfy the axioms of a strong Heegaard invariant \cite{JTNaturality}*{Definition~2.32}. We verify these axioms presently.
The \emph{functoriality} axiom asserts two claims. Firstly, it asserts that we have well defined morphisms associated to handleslide equivalences of alpha curves and morphisms associated to handleslide equivalences of beta curves (for a fixed Heegaard surface). This part of the axiom follows from Theorem~\ref{thm:naturality-fixed-Sigma}. (We remind the reader that unlike in the setting of ordinary Heegaard Floer homology, it is not particularly natural to define an involutive transition map which changes just the alpha curves, or just the beta curves, since the transition map is defined by compressing hyperboxes as in~\eqref{def:transition-map-elementary-handleslide}. Such hyperboxes necessarily require changing both the alpha and beta curves simultaneously).
The second part of the functoriality axiom asserts that the morphisms for stabilizations are inverse to the morphisms for destabilizations. This is proven in Proposition~\ref{prop:stabilization-naturality}.
The next axiom from \cite{JTNaturality}*{Definition~3.32} is the \emph{commutativity} axiom, concerning commutativity of the distinguished rectangles from \cite{JTNaturality}*{Definition~2.29}. We now verify that there is no monodromy around the distinguished rectangles from \cite{JTNaturality}*{Definition~2.29}. In our present setting, we may rephrase the distinguished rectangles from \cite{JTNaturality} as diagrams of embedded, doubling enhanced Heegaard diagrams with the following shape, satisfying one of five configurations:
\[
\begin{tikzcd}
\mathfrak{D}_1
\mathrm{a.r.}[r, "e"]
\mathrm{a.r.}[d, "f"]
&
\mathfrak{D}_2
\mathrm{a.r.}[d, "g"]
\\
\mathfrak{D}_3
\mathrm{a.r.}[r, "h"] & \mathfrak{D}_4
\end{tikzcd}
\]
We presently enumerate the five configurations, and also prove that there is no monodromy:
\begin{enumerate}
\item $e$, $f$, $g$ and $h$ are handleslide equivalences. In our setting, we must consider changes of both of the attaching curves, as well as changes of the doubling datum. Note these are also referred to as \emph{strong equivalences} in the literature, e.g.~in \cite{MOIntegerSurgery}. The diagram commutes up to chain homotopy by Theorem~\ref{thm:naturality-fixed-Sigma}. We note that in \cite{JTNaturality}, the stated rectangle has the property that $e$ and $h$ are $\alpha$-equivalences, while $f$ and $g$ are $\beta$-equivalences. However the rectangle therein is also a rectangle of \emph{isotopy diagrams}, so we must consider changes of both the alpha and beta curves along each edge.
\item $e$ and $h$ are handleslide equivalences of attaching curves and doubling curves, while $f$ and $g$ are both stabilizations. Commutativity follows from Proposition~\ref{prop:stabilization-naturality}.
\item $e$ and $h$ are handleslide equivalences, while $f$ and $g$ are diffeomorphisms. Commutativity around such rectangles is tautological.
\item $e$, $f$, $g$ and $h$ are all stabilizations. Furthermore, there are disjoint 3-balls $B_1$ and $B_2$, such that $e$ and $h$ correspond to a stabilization in $B_1$, while $f$ and $g$ correspond to a stabilization in $B_2$. Commutativity around such rectangles follows from Proposition~\ref{prop:stabilization-naturality}.
\item The maps $e$ and $h$ are stabilizations, while $f$ and $g$ are diffeomorphisms. Furthermore, the stabilization ball for $h$ is the image of the stabilization ball for $e$ under the diffeomorphism $g$. Commutativity around such rectangles is tautological.
\end{enumerate}
The remaining axioms of \cite{JTNaturality}*{Definition~3.32}, continuity and handleswap invariance, were verified in Proposition~\ref{prop:continuity} and Theorem~\ref{thm:handleswap}.
\end{proof}
\subsection{Functoriality}
We now finish defining our cobordism maps and sketch the proof of Theorem \ref{thm:well-defined-cobordism-map}, showing that these maps are well-defined. Our proof is modeled on \cite{OSTriangles}. Since our approach is standard, we only provide an overview and sketch the points at which our construction deviates from \cite{OSTriangles}.
We begin with the construction, and will shortly sketch a proof of invariance. Suppose $W$ is a cobordism from $Y_1$ to $Y_2$, and assume basepoints in $Y_1$ and $Y_2$, as well as a path $\gamma$ connecting them, have been chosen. We assume that $W$, $Y_1$, $Y_2$ are connected. We pick a Morse function $f\colon W\to [0,1]$. We may assume that the indices of the critical points of $f$ are non-decreasing, and we assume all critical values are distinct and that $f$ has critical points only of index 1, 2 and 3. Next, we pick a gradient like vector field $v$ such that our path $\gamma$ is a flow-line. A choice of gradient like vector field gives a diffeomorphism
\[
W\cong W(\mathbb{S}_2)\cup W(\mathbb{S}_1)\cup W(\mathbb{S}_0),
\]
where $\mathbb{S}_i$ are framed $i$-dimensional links and $\mathbb{S}_0\subset Y_1$, $\mathbb{S}_1\subset Y_1(\mathbb{S}_0)$, and $\mathbb{S}_2\subset Y_1(\mathbb{S}_0)(\mathbb{S}_1)$. If $\mathfrak{s}\in \Spin^c(W)$ is self-conjugate, we define our cobordism map as the composition
\[
\mathit{CFI}(W,\mathfrak{s})=\mathit{CFI}(W(\mathbb{S}_2))\circ \mathit{CFI}(W(\mathbb{S}_1), \mathfrak{s}|_{W(\mathbb{S}_1)})\circ \mathit{CFI}(W(\mathbb{S}_0)).
\]
We now prove Theorem \ref{thm:well-defined-cobordism-map}. As a first step, we show that the 2-handle map is well-defined:
\begin{lem}\label{lem:well-defined-2handle}
Our 2-handle map $\mathit{CFI}(W(\mathbb{S}_1),\mathfrak{s}|_{W(\mathbb{S}_1)})$ is independent of the choice of diagram subordinate to a particular bouquet. Furthermore, it is also invariant from the choice of bouquet, as well as handleslides amongst the components of $\mathbb{S}_1$.
\end{lem}
\begin{proof}
Ozsv\'{a}th and Szab\'{o} \cite{OSTriangles}*{Lemma~4.5} described five moves which relate any two bouquets of a fixed link. Invariance under the first four of these moves follows immediately from invariance of our map from handleslides amongst the curves in our Heegaard triple, Proposition~\ref{prop:2-handles-handle-slides}. The final move described by Ozsv\'{a}th and Szab\'{o} is stabilization of the Heegaard triple. Invariance from this move is a formal consequence of our definition of the stabilization map. Indeed we defined the naturality map for stabilizations as the composition of the 1-handle map, followed by a canceling 2-handle map. Both the 1-handle map and map for a canceling 2-handle map commute with the 2-handle map $\mathit{CFI}(W(\mathbb{S}_1), \mathfrak{s}|_{W(\mathbb{S}_1)})$ by Propositions~\ref{prop:composition-law} and ~\ref{prop:commute-1handles/2-handles}. Hence our 2-handle map is independent of the choice of diagram subordinate to a fixed bouquet.
Next, we note that independence from the choice of bouquet, as well as independence from handleslides amongst the components of $\mathbb{S}_1$ both follow from independence of the map from handleslides of the attaching curves, which we proved in Proposition~\ref{prop:2-handles-handle-slides}. Compare \cite{OSTriangles}*{Lemma~4.8}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:well-defined-cobordism-map}]
Firstly, we verify that the construction is independent of the choice of gradient-like vector field $v$ for $f$. Subsequently we sketch that the cobordism map is invariant from $f$.
The space of gradient like vector fields for a fixed Morse function is connected. Since we have already decomposed $W$ along level sets which separate critical points of different indices, the codimension 1 singularities of a path of gradient like vector fields are handleslides amongst link components of the same indices. For critical points of index 1 or 3, invariance under these moves can instead be proven by reordering the attachment of the handles using the composition law for 1-handles and 3-handles, Proposition~\ref{lem:commute-1/3-handles}, since a handleslide is the same as an isotopy after attaching one of the handles. For 2-handles, we cannot add additional level sets because of the $\Spin^c$ decomposition, so instead handleslide invariance is proven directly in Lemma~\ref{lem:well-defined-2handle}.
We now consider invariance under $f$. By standard Cerf theory, \cite{CerfStratification} \cite{KirbyCalculus} any two Morse functions $f_0$ and $f_1$ with critical points ordered monotonically may be connected with a 1-parameter family $(f_{t})_{t\in [0,1]}$ whose critical points are also ordered monotonically, except at finitely many points where a birth-death singularity occurs. Furthermore, if $f_0$ and $f_1$ have no index 0 or 4 critical points, then each $f_t$ may also be chosen to have no index 0 or 4 critical points. We now show that the maps are invariant under index 1/2 handle cancellations, as well as index 2/3 handle cancellations. Invariance under such handle cancellation is essentially automatic from the definition of our naturality map for stabilizations. Indeed, consider the case that $W$ has a Morse function with only index 2 critical points, whose descending manifolds intersect $Y_0$ in a framed link $\mathbb{S}_1$. Let $\mathbb{S}_0\subset Y_0$ be a 0-sphere, and let $\mathbb{K}$ be a framed knot in $Y_0(\mathbb{S}_0)$ which intersects the co-core of $\mathbb{S}_0$ in exactly 1-point. Assume $\mathbb{S}_1$ is disjoint from $\mathbb{S}_0$ and $\mathbb{K}$. Using the composition law of Propositions~\ref{prop:composition-law}, we obtain
\[
\mathit{CFI}(W(\mathbb{S}_1\cup \mathbb{K}))\circ \mathit{CFI}(W(\mathbb{S}_0))\simeq \mathit{CFI}(W(\mathbb{S}_1))\circ \mathit{CFI}(W(\mathbb{K}))\circ \mathit{CFI}(W(\mathbb{S}_0)):=\mathit{CFI}(W(\mathbb{S}_1))\circ \sigma,
\]
where $\sigma$ is the naturality map for stabilization (and we are omitting $\Spin^c$ structures from the notation). Invariance under index 2/3 handle cancelations follows entirely analogously. The proof is complete.
\end{proof}
\section{An expanded model of the involution}
\label{sec:expanded-def}
In this section, we describe a \emph{basepoint-expanded} doubling model of the involution, which will allow us to understand the transition maps for non-elementary handleslide equivalences. An outline of this section is as follows. In Section~\ref{sec:expanded-model-def-details}, we define the notion of a \emph{basepoint expanded, doubling enhanced Heegaard diagram} $\tilde{\mathfrak{D}}$, which can be used to construct an $\mathbb{F}[U,Q]/Q^2$-complex $\mathit{CFI}(\tilde{\mathfrak{D}})$. In Section~\ref{sec:maps-expanded-model}, we construct transition maps
\[
\tilde{\Psi}_{\tilde{\mathfrak{D}}_1\to \tilde{\mathfrak{D}}_2}\colon \mathit{CFI}(\tilde{\mathfrak{D}}_1)\to \mathit{CFI}(\tilde{\mathfrak{D}}_2)
\]
in the case that $\tilde\mathfrak{D}_1$ and $\tilde\mathfrak{D}_2$ are two basepoint expanded, doubling enhanced Heegaard diagrams which have the same underlying Heegaard surface, and which satisfy a weak admissibility condition. Unlike in Section~\ref{sec:elementary-equivalences}, we do not require $\tilde{\mathfrak{D}}_1$ and $\tilde{\mathfrak{D}}_2$ to be related by an elementary handleslide. In the subsequent Section~\ref{sec:relate-expanded}, we relate the complexes $\mathit{CFI}(\tilde{\mathfrak{D}})$ and the transition maps $\tilde{\Psi}_{\tilde{\mathfrak{D}}_1\to \tilde{\mathfrak{D}}_2}$ to the non-expanded models described earlier. The key motivation for considering the maps $\tilde{\Psi}_{\tilde\mathfrak{D}_1\to \tilde \mathfrak{D}_2}$ is that they may be defined for non-elementary handleslide equivalences in a manner which clearly is independent of additional choices.
\subsection{Doubling with extra basepoints}
\label{sec:expanded-model-def-details}
We now describe our expanded model of the involution, which uses an extra basepoint. The presence of the extra basepoint simplifies some of the gluing arguments.
\begin{define}
A \emph{basepoint expanded, doubling enhanced Heegaard diagram} $\tilde{\mathfrak{D}}$ consists of a singly pointed Heegaard diagram $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},w)$, a small disk $D\subset \Sigma$, containing $w$ along its boundary, a choice of point $w'$, as well as a collection of $2g+2$ attaching curves $\ve{\Delta}\subset \Sigma\# \bar \Sigma$, as follows. The curves $\ve{\Delta}$ are constructed by doubling a basis of pairwise disjoint, properly embedded arcs $\delta_1,\dots, \delta_{2g+2}$ on $\Sigma\setminus D$, which avoid $w$ and $w'$ and which form a basis of $H_1(\Sigma\setminus D, \d D\setminus \{w,w'\})$. Additionally, $\tilde{\mathfrak{D}}$ contains the choice of almost complex structures used to compute the holomorphic triangle maps.
\end{define}
Using a basepoint expanded diagram $\tilde{\mathfrak{D}}$, as above, we obtain a model of the involution by modifying the formula in Corollary~\ref{cor:doubling-model}, as follows. We let $c$ and $c'$ be small perturbations of the circle $\d D$, as in Figure~\ref{fig:5}.
We define a 1-handle map
\[
F_1^{c\bar \a, c\bar \a}\colon \mathit{CF}(\Sigma,\ve{\alpha},\ve{\beta},w)\to \mathit{CF}(\Sigma\# \bar \Sigma, \ve{\alpha} c \bar \ve{\beta}, \ve{\beta} c \bar \ve{\beta}, w,w')
\]
by tensoring with the top degree generator. This map adds the basepoint $w'$. Similarly, we define a map
\[
F_3^{\b c',\b c'}\colon \mathit{CF}(\Sigma \# \bar \Sigma, \ve{\alpha} c' \bar \ve{\beta}, \ve{\alpha} c'\bar \ve{\alpha},w,w')\to \mathit{CF}(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\alpha}, w)
\]
by using the same formula as the standard 3-handle map.
We define the basepoint-expanded model for the involution via the formula
\[
\iota:=\eta\circ F_3^{\b c',\b c'}\circ f_{\b c' \bar \b}^{\a c \bar \b\to \a c'\bar \b}\circ A_{\lambda}\circ f_{\Delta\to \b c'\bar{\b} }^{\a c\bar{\b}} \circ f_{\a c\bar{\a}\to \Delta}^{\b c\bar{\a}} \circ F_1^{c\bar{\a},c\bar{\a}}.
\]
In the above equation, $A_{\lambda}$ denotes the relative homology map for an arc $\lambda$ connecting $w$ and $w'$ in the connected sum region. This map is discussed further in the subsequent Section~\ref{sec:hypercubes-rel-homology}.
\begin{figure}[H]
\centering
\input{fig5.pdf_tex}
\caption{The basepoint expanded doubling model of the involution. (Frames read in alphabetical order).}\label{fig:5}
\end{figure}
\subsection{Homology actions on hypercubes}
\label{sec:hypercubes-rel-homology}
In this section, we describe how to construct a homology action on a hypercube of Floer chain complexes. We can construct both a version for closed curves on $\Sigma$, as well as a relative one for arcs on $\Sigma$ with boundary in $\ve{w}$. We focus on the relative case.
We recall the construction from \cite{ZemGraphTQFT}*{Section~5}. If $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},\ve{w})$ is a multi-pointed Heegaard diagram, and $\lambda$ is a path on $\Sigma$ connecting $w_1,w_2\in \ve{w}$, there is an endomorphism
\begin{equation}
A_\lambda\colon \mathit{CF}(\mathcal{H},\mathfrak{s})\to \mathit{CF}(\mathcal{H},\mathfrak{s}),\label{eq:A-lambda-original}
\end{equation}
which satisfies
\begin{equation}
[\d, A_\lambda]=U_{w_1}+U_{w_2}.\label{eq:[d,A-lambda]}
\end{equation}
In our present paper, we work with just one $U$-variable, so $A_{\lambda}$ becomes a chain map. The map $A_\lambda$ is defined as follows. If $\phi\in \pi_2(\ve{x},\ve{y})$ is a class of disks, we may define a quantity $a(\lambda,\phi)\in \mathbb{Z}$ by summing the changes of the multiplicities of $\phi$ across the alpha curves as one travels along $\lambda$. (Compare \cite{Ni-homological}). Sometimes it is helpful to write $a(\lambda,\phi)=(\d_{\a} D(\phi))\cdot \lambda$, where $\d_{\a} D(\phi)$ is the alpha boundary of the domain of $\phi$. The map $A_\lambda$ is defined via the formula
\[
A_{\lambda}(\ve{x}):=\sum_{\substack{\phi\in \pi_2(\ve{x},\ve{y})\\ \mu(\phi)=1}} a(\lambda,\phi) \# (\mathcal{M}(\phi)/\mathbb{R})U^{n_{w_1}(\phi)+\cdots+n_{w_n}(\phi)} \cdot \ve{y}.
\]
The map $A_{\lambda}$ is sometimes a helpful tool when adding and removing basepoints. We now describe a version of $A_\lambda$ for hypercubes. Suppose we are given two hypercubes of attaching curves $\mathcal{L}_\a$ and $\mathcal{L}_{\b}$ on a multi-pointed Heegaard diagram, each consisting of handleslide equivalent attaching curves, such that the length 1 morphisms are all the top degree generators of $\mathit{CF}^-(\#^n (S^1\times S^2))$. We now suppose that $\lambda$ is an arc which connects two basepoints, $w_1$ and $w_2$. Consider the pairing $\mathit{CF}(\mathcal{L}_\a,\mathcal{L}_\b)$, where we identify the variables of the basepoints $w_1$ and $w_2$ to a single $U$.
We now construct a morphism of hypercubes
\[
\mathscr{A}_{\lambda}\colon \mathit{CF}(\mathcal{L}_{\a}, \mathcal{L}_{\b})\to \mathit{CF}(\mathcal{L}_{\a}, \mathcal{L}_{\b}).
\]
Recall that a morphism of hypercubes can itself be thought of as a hypercube of dimension 1 larger than $\dim \mathit{CF}(\mathcal{L}_{\a}, \mathcal{L}_{\b})$. Our morphism of hypercubes $\mathscr{A}_{\lambda}$ will have the property that the length one morphisms coincide with the chain maps $A_\lambda$ from~\eqref{eq:A-lambda-original}. We will show that $\mathscr{A}_{\lambda}$ is well-defined up to chain homotopy of hypercube morphisms, and is natural with respect to restriction to sub-cubes of $\mathit{CF}(\mathcal{L}_{\a},\mathcal{L}_{\b})$.
We note that we may formally view $\lambda$ as an input for a holomorphic polygon counting map by setting
\[
\begin{split}
&f_{\gamma_1,\dots,\gamma_j,\gamma_j,\dots, \gamma_m}(\ve{x}_{1,2},\dots,\ve{x}_{j-1,j}, \lambda, \ve{x}_{j,j+1},\dots, \ve{x}_{m-1,m})
\\:=&\sum_{\substack{\psi\in \pi_2(\ve{x}_{0,1},\dots, \ve{x}_{m-1,m},\ve{z})\\ \mu(\psi)=3-m}} (\d_{\gamma_j}D(\psi) \cdot \lambda)\# \mathcal{M}(\psi)U^{n_{w_1}(\psi)+\cdots+n_{w_n}(\psi)}\cdot \ve{z},
\end{split}
\]
where $\mathcal{M}(\psi)$ denotes the moduli space of holomorphic $m$-gons (in the usual sense).
\noindent As particular examples,
\[
f_{\a,\a,\b}(\lambda,\ve{x})=A_{\lambda}(\ve{x})\quad \text{and} \quad f_{\a,\b,\b}(\ve{x},\lambda)=B_{\lambda}(\ve{x}),
\]
where $A_{\lambda}$ counts holomorphic disks weighted by $\#(\d_{\a}(\phi)\cap \lambda)$, while $B_{\lambda}$ counts disks weighted by $\#(\d_{\b}(\phi)\cap \lambda)$.
\begin{lem}\label{lem:associativity-extra-lambda}
The holomorphic polygon counting maps, defined with an extra input $\lambda$, satisfy the standard associativity rule as long as one uses the convention that $\d \lambda=U_{w_1}+U_{w_2}$, together with conventions that the polygon maps are $U_{w_i}$-equivariant and strictly unital (i.e. they vanish if they have $1$ as an input when there are more than 2 total inputs). In particular, if we work over a single $U$ variable, they satisfy the associativity relations with the convention that $\d \lambda=0$.
\end{lem}
\begin{proof}The proof differs based on whether $\ell=2$ or $\ell>2$. If $\ell=2$, then the result follows from Equation~\eqref{eq:[d,A-lambda]}. For the case that $\ell>2$, one counts the ends of index $(3-\ell)$ moduli spaces weighted by $\d_{\gamma_i} D(\psi)\cdot \lambda$ and quickly obtains the result.
\end{proof}
The construction of $\mathscr{A}_{\lambda}$ is as follows. We will formally construct a cube-shaped diagram $\mathcal{L}_{\a}^\lambda$ of dimension $\dim \mathcal{L}_{\a}+1$. To begin, we take two copies of $\mathcal{L}_{\a}$, for which we write $\mathcal{L}_{\a}\times \{0\}$ and $\mathcal{L}_{\a}\times \{1\}$, and formally adjoin additional length 1 arrows from $\mathcal{L}_{\a}\times \{0\}$ to $\mathcal{L}_{\a}\times \{1\}$, which we label by the character $\lambda$.
We begin by constructing the length 2 chains of $\mathcal{L}_{\a}^\lambda$, via the following argument. We will show that the 2-dimensional faces can be constructed to have the following form:
\begin{equation}
\begin{tikzcd}[column sep=2cm, row sep=2cm, labels=description]
\ve{\alpha}
\mathrm{a.r.}[r, "\Theta_{\a',\a}"]\mathrm{a.r.}[d, "\lambda"]
\mathrm{a.r.}[dr,dashed, "U\eta_{\a',\a}"]& \ve{\alpha}'\mathrm{a.r.}[d, "\lambda"]
\\
\ve{\alpha}\mathrm{a.r.}[r,"\Theta_{\a',\a}"]& \ve{\alpha}'
\end{tikzcd}
\label{eq:length-2-A-lambda}
\end{equation}
(We have not yet defined $\eta_{\a',\a}$.) The desired relation is that
\begin{equation}
A_\lambda(\Theta_{\a',\a})+A'_{\lambda}(\Theta_{\a',\a})=\d U \eta_{\a',\a}.\label{eq:length-2-relation-A-lambda}
\end{equation}
Here $A_{\lambda}$ counts changes of a disk class across $\ve{\alpha}$, while $A'_{\lambda}$ counts changes across $\ve{\alpha}'$. Note that by \cite{HMZConnectedSum}*{Section~3.5}, we have that
\begin{equation}
A_{\lambda}(\Theta_{\a',\a})+A'_{\lambda}(\Theta_{\a',\a})=U(\Phi_{w_1}+\Phi_{w_2})(\Theta_{\a',\a}), \label{eq:telescope-A-lambda}
\end{equation}
where $\Phi_{w_i}$ denotes the map
\[
\Phi_{w_i}(\ve{x})=U^{-1}\sum_{\substack{\phi\in \pi_2(\ve{x}, \ve{y}) \\ \mu(\phi)=1}} n_{w_i}(\phi) \# (\mathcal{M}(\phi)/\mathbb{R}) U^{(n_{w_1}+\cdots+n_{w_n})(\phi)},
\]
extended $\mathbb{F}[U]$-equivariantly.
We note that $\Phi_{w_i}(\Theta_{\a',\a})$ is a cycle of grading 1 higher than $\Theta_{\a',\a}$, so it is null-homologous, since $\Theta_{\a',\a}$ is the top degree generator. Let $\eta_{\a',\a}$ be any chain of homogeneous grading one higher that $\Theta_{\a',\a}$ so that
\[
\d \eta_{\a',\a}=(\Phi_{w_1}+\Phi_{w_2})(\Theta_{\a',\a}).
\]
We perform this construction to build all of the length 2 chains in $\mathcal{L}_{\a}^{\lambda}$. (If we did not quotient by $U_{w_1}+U_{w_2}$, then we would instead obtain a length 2 morphism of the form $U_1\eta_{\a',\a}^1+U_2\eta_{\a',\a}^2$).
We now construct the higher length chains in the cube $\mathcal{L}^\lambda_{\a}$. Their construction is similar to the construction of the length 2 chains. We assume, by induction, that each length $m>1$ chain is of the form $U \eta_{\a',\a}$. (This is only necessary when $m\in \{2,3\}$, but is satisfiable for all $m>1$).
Let $\varepsilon<\varepsilon'$ be points $\mathbb{E}_{d}$, where $d=\dim \mathcal{L}_\a$, and suppose that all chains of length less than $|\varepsilon'-\varepsilon|_{L^1}$ have already been defined. Set
\begin{equation}
\begin{split}
C_{\varepsilon',\varepsilon}:=&\sum_{\varepsilon=\varepsilon_1<\cdots <\varepsilon_j=\varepsilon'} \sum_{i=0}^j f_{\a^{\varepsilon_j},\dots, \a^{\varepsilon_1}}(\Theta_{\a^{\varepsilon_j},\a^{\varepsilon_{j-1}}},\dots,\Theta_{\a^{\varepsilon_{i+1}},\a^{{\varepsilon}_{i}}},\lambda,\Theta_{\a^{\varepsilon_i},\a^{\varepsilon_{i-1}}}\dots, \Theta_{\a^{\varepsilon_2},\a^{\varepsilon_1}})\\
+&\sum_{\varepsilon=\varepsilon_1<\cdots <\varepsilon_j=\varepsilon'} \sum_{i=0}^j f_{\a^{\varepsilon_j},\dots, \a^{\varepsilon_1}}(\Theta_{\a^{\varepsilon_j},\a^{\varepsilon_{j-1}}},\dots,\Theta_{\a^{\varepsilon_{i+1}},\a^{\varepsilon_i}},U\eta_{\a^{\varepsilon_i},\a^{\varepsilon_{i-1}}},\Theta_{\a^{{\varepsilon}_{i-1}},\a^{{\varepsilon}_{i-2}}}\dots, \Theta_{\a^{\varepsilon_2},\a^{\varepsilon_1}})
\end{split}
\label{eq:A-lambda-1st-equation}
\end{equation}
However, by an entirely analogous argument to ~\eqref{eq:telescope-A-lambda}, we have
\begin{equation}
\begin{split}
C_{\varepsilon',\varepsilon}=&\sum_{\varepsilon=\varepsilon_1<\cdots <\varepsilon_j=\varepsilon'} f_{\a^{\varepsilon_j},\dots, \a^{\varepsilon_1}}^{n_{w_1}+n_{w_2}}(\Theta_{\a^{\varepsilon_j},\a^{{\varepsilon}_{j-1}}},\dots,\Theta_{\a^{\varepsilon_2},\a^{\varepsilon_1}})\\
+&\sum_{\varepsilon=\varepsilon_1<\cdots <\varepsilon_j=\varepsilon'} \sum_{i=0}^j f_{\a^{\varepsilon_j},\dots, \a^{\varepsilon_1}}(\Theta_{\a^{\varepsilon_j},\a^{{\varepsilon}_{j-1}}},\dots,\Theta_{\a^{\varepsilon_{i+1}},\a^{\varepsilon_i}},U\eta_{\a^{\varepsilon_i},\a^{\varepsilon_{i-1}}},\Theta_{\a^{{\varepsilon}_{i-1}},\a^{\varepsilon_{i-2}}}\dots, \Theta_{\a^{\varepsilon_2},\a^{\varepsilon_1}})
\end{split}
\label{eq:A-lambda-2nd-equation}
\end{equation}
where $f_{\a^{\varepsilon_n},\dots, \a^{\varepsilon_1}}^{n_{w_1}+n_{w_2}}$ counts holomorphic polygons with a factor of $n_{w_1}(\psi)+n_{w_2}(\psi)$. In particular, $C_{\varepsilon',\varepsilon}=U C'_{\varepsilon',\varepsilon}$, for some chain $C'_{\varepsilon',\varepsilon}$. Since the hypercube relations are satisfied for all proper faces, the element $C_{\varepsilon',\varepsilon}$ is a cycle. Since the complexes are free $\mathbb{F}[U]$-modules, this implies that $C'_{\varepsilon',\varepsilon}$ is also a cycle. However $C_{\varepsilon',\varepsilon}$ has grading $m-2$ higher than the top degree generator, where $m=|\varepsilon'-\varepsilon|_{L^1}$. Hence $C'_{\varepsilon',\varepsilon}$ has grading $m$ higher than the top degree generator. As long as $m\ge 1$, we conclude that $C'_{\varepsilon',\varepsilon}$ is a boundary. We set $\eta_{\a^{\varepsilon'}, \a^{\varepsilon}}$ to be any chain of homogeneous grading $m+1$ such that $\d \eta_{\a^{\varepsilon'},\a^{\varepsilon}}=C'_{\varepsilon',\varepsilon}$.
Finally, we define the hypercube corresponding to the morphism $\mathscr{A}_\lambda$ to be the pairing of $\mathit{CF}(\mathcal{L}^\lambda_{\a}, \mathcal{L}_\b)$, where the hypercube maps are defined by using the normal formulas for the maps in a pairing, while allowing the chains $\lambda$ to be inputs.
\begin{lem}\,
\begin{enumerate}
\item The diagram $\mathcal{L}_{\a}^{\lambda}$ satisfies the hypercube relations (interpreted in the way described above).
\item The map $\mathscr{A}_{\lambda}$ satisfies $\d_{\Mor}(\mathscr{A}_{\lambda})=0$ (i.e. $\mathscr{A}_{\lambda}$ is a cycle).
\item The map $\mathscr{A}_{\lambda}$ is independent, up to chain-homotopy of hypercube morphisms, from the choices of chains used in its construction.
\end{enumerate}
\end{lem}
\begin{proof}The first claim follows immediately from the construction.
The second claim follows quickly from Gromov compactness interpreted via Lemma~\ref{lem:associativity-extra-lambda}.
We consider the final claim. This is proven as follows. Suppose we have two models of $\mathcal{L}_{\a}^{\lambda}$, which are both constructed using the above procedure. We consider a partially constructed hypercube of dimension $d+2$, given by the following diagram:
\[
\mathcal{L}_{\a}^{\lambda_1\to \lambda_2}=
\begin{tikzcd}[column sep=2cm, row sep=2cm, labels=description]
\mathcal{L}_{\a}\mathrm{a.r.}[d, "\id"] \mathrm{a.r.}[dr,dashed, "*"]\mathrm{a.r.}[r,"\lambda_1"] &\mathcal{L}_{\a} \mathrm{a.r.}[d,"\id"]\\
\mathcal{L}_{\a}\mathrm{a.r.}[r, "\lambda_2"]& \mathcal{L}_{\a}.
\end{tikzcd}
\]
Here $\lambda_1$ denotes the chains in one model of $\mathcal{L}_{\a}^{\lambda}$, while $\lambda_2$ denotes the chains constructed in the other model, and the $*$ arrows are yet to be defined. The hypercube relations are satisfied on the four $d$-dimensional faces of the above diagram. We wish to define chains which increment both of the displayed coordinates, such that the hypercube relations are satisfied. The construction of such arrows (corresponding to the $*$ arrow above) follows nearly verbatim from the construction of the cube $\mathcal{L}^\lambda_\a$ itself. Pairing $\mathcal{L}_{\a}^{\lambda_1\to \lambda_2}$ with $\mathcal{L}_{\b}$ gives the homotopy between the two models of $\mathscr{A}_{\lambda}$, completing the proof.
\end{proof}
\begin{rem}
\begin{enumerate}
\item If one works over the ring $\mathbb{F}[U_{w_1},\dots, U_{w_n}]$, one obtains instead that $\d_{\Mor}(\mathscr{A}_{\lambda})=(U_{w_1}+U_{w_2})\cdot \id$, where $\d \lambda=U_{w_1}+U_{w_2}$.
\item The simplest case of the map $\mathscr{A}_{\lambda}$ is when $\mathcal{L}_{\a}$ is 0-dimensional. Then $\mathcal{L}_\a^\lambda$ is just the formal hypercube
\[
\begin{tikzcd} \ve{\alpha} \mathrm{a.r.}[r, "\lambda"] & \ve{\alpha} \end{tikzcd}
\]
If $\mathcal{L}_{\b}$ is a beta hypercube, then the map
\[
\mathscr{A}_{\lambda}\colon \mathit{CF}(\ve{\alpha}, \mathcal{L}_{\b})\to \mathit{CF}(\ve{\alpha}, \mathcal{L}_{\b})
\]
requires no additional choices.
\item Naturally, one may also build a hypercube $\mathcal{L}_{\b}^{\lambda}$, similar to above, and define
\[
\mathscr{B}_{\lambda}\colon \mathit{CF}(\mathcal{L}_\a, \mathcal{L}_{\b})\to \mathit{CF}(\mathcal{L}_{\a},\mathcal{L}_{\b})
\]
as the hypercube $\mathit{CF}(\mathcal{L}_{\a}, \mathcal{L}_{\b}^{\lambda})$.
\end{enumerate}
\end{rem}
\subsection{Maps for general equivalences on the expanded model}
\label{sec:maps-expanded-model}
We now define the naturality maps for a general alpha or beta equivalence in terms of the expanded model of the involution. Suppose that $\tilde\mathfrak{D}$ and $\tilde\mathfrak{D}'$ are two basepoint expanded doubling enhanced Heegaard diagrams, which share the same underlying Heegaard surface $\Sigma$. Write $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},w)$ and $\mathcal{H}'=(\Sigma,\ve{\alpha}',\ve{\beta}',w)$ for the underlying non-involutive Heegaard diagrams of $\tilde\mathfrak{D}$ and $\tilde\mathfrak{D}'$, respectively. Write $\ve{\Delta}$ and $\ve{\Delta}'$ for the choices of doubling curves. Furthermore, we assume that the diagram
\[
(\Sigma\# \bar{\Sigma},\ve{\alpha}',\bar \ve{\alpha}',\ve{\alpha},\bar \ve{\alpha},\ve{\beta},\bar \ve{\beta},\ve{\beta}', \bar \ve{\beta}',c,c',\ve{\Delta},\ve{\Delta}',w,w')
\]
is weakly admissible. (This may be achieved by winding $\ve{\alpha}$, $\ve{\alpha}'$, $\ve{\beta}$ and $\ve{\beta}'$ sufficiently).
\begin{rem}
\label{rem:admissibility-unconventional}
Note that the above is not a valid Heegaard diagram since none of the attaching curve sets have $2g(\Sigma)+1$ curves, however it still makes sense to require weak-admissibility. Indeed the definition of weak-admissibility is that if $P$ is a non-zero integral 2-chain with boundary equal to a linear combination of curves on the diagram and $n_{w}(P)=n_{w'}(P)=0$, then $P$ has both positive and negative multiplicities.
\end{rem}
\begin{figure}[ht]
\[
\begin{tikzcd}[labels=description, row sep=1cm]
\mathit{CF}(\ve{\alpha},\ve{\beta},w)
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{c\bar{\b},c\bar{\b}}"]
& \mathit{CF}(\ve{\alpha}',\ve{\beta},w)
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{c\bar{\b}',c\bar{\b}}"]
& \mathit{CF}(\ve{\alpha}',\ve{\beta}',w)
\mathrm{a.r.}[d,"F_{1}^{c\bar{\b}',c\bar{\b}'}"]
\\
\mathit{CF}(\ve{\alpha} c \bar{\ve{\beta}}, \ve{\beta} c \bar{\ve{\beta}},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}'c \bar{\ve{\beta}}', \ve{\beta} c \bar{\ve{\beta}},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}' c \bar{\ve{\beta}}', \ve{\beta}' c \bar{\ve{\beta}}',w,w')
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha} c \bar \ve{\beta}, \ve{\Delta},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "A_{\lambda}"]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}'c\bar \ve{\beta}',\ve{\Delta},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "A_{\lambda}"]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}' c \bar \ve{\beta}', \ve{\Delta}',w,w')
\mathrm{a.r.}[d, "A_{\lambda}"]
\\
\mathit{CF}(\ve{\alpha} c \bar \ve{\beta}, \ve{\Delta},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}'c\bar \ve{\beta}',\ve{\Delta},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}' c \bar \ve{\beta}', \ve{\Delta}',w,w')
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha} c\bar{\ve{\beta}}, \ve{\alpha} c' \bar{\ve{\alpha}},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
&\mathit{CF}(\ve{\alpha}'c \bar{\ve{\beta}}', \ve{\alpha} c' \bar{\ve{\alpha}},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
&\mathit{CF}(\ve{\alpha}' c \bar{\ve{\beta}}', \ve{\alpha}' c \bar{\ve{\alpha}}',w,w')
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha} c' \bar{\ve{\beta}}, \ve{\alpha} c' \bar{\ve{\alpha}},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a c',\a c'}"]
&\mathit{CF}(\ve{\alpha}' c' \bar{\ve{\beta}}', \ve{\alpha} c' \bar{\ve{\alpha}},w,w')
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a'c',\a c'}"]
& \mathit{CF}(\ve{\alpha}' c' \bar{\ve{\beta}}', \ve{\alpha}' c' \bar{\ve{\alpha}}',w,w')
\mathrm{a.r.}[d, "F_3^{\a'c',\a'c'}"]
\\
\mathit{CF}(\bar{\ve{\beta}}, \bar{\ve{\alpha}},w)
\mathrm{a.r.}[r]
\mathrm{a.r.}[d,"\id"]
\mathrm{a.r.}[drr,dashed, "h_{\bar \b\to \bar \b'}^{\bar \a \to \bar \a'}"]
& \mathit{CF}(\bar{\ve{\beta}}',\bar{\ve{\alpha}},w)
\mathrm{a.r.}[r]
& \mathit{CF}(\bar{\ve{\beta}}',\bar{\ve{\alpha}}',w)
\mathrm{a.r.}[d,"\id"]
\\
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha},w)
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}',w)
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta}', \bar \ve{\alpha}',w)
\end{tikzcd}
\]
\caption{The hyperbox whose compression is the basepoint expanded transition map $\tilde\Psi_{\tilde\mathfrak{D}\to \tilde\mathfrak{D}'}$.}
\label{fig:transition-map-1}
\end{figure}
Our transition map $\tilde \Psi_{\tilde \mathfrak{D}\to \tilde \mathfrak{D}'}$ is defined by the hyperbox in Figure~\ref{fig:transition-map-1}. We now describe aspects of the construction in more detail. The second and third levels are obtained by pairing hyperboxes of attaching curves in the obvious manner. The fourth level is obtained from the argument in Section~\ref{sec:hypercubes-rel-homology}, where we take $\lambda$ to be a path in the neck region which connects the two basepoints $w$ and $w'$.
We now consider the levels involving the 1-handle and 3-handle maps.
The map $F_{1}^{c\bar{\a}',c \bar{\a}}$ is different than the other 1-handle maps we have seen so far, since we do not require the diagram $(\Sigma, \bar{\ve{\alpha}}', \bar{\ve{\alpha}})$ to be a standard diagram for $(S^1\times S^2)^{\# g}$. Instead, we only require the diagram to be admissible. The construction of the map is as follows. We pick a cycle $\Theta_{\a',\a}$ representing the top degree generator of $\mathit{CF}^-(\Sigma,\ve{\alpha}',\ve{\alpha},w)$, and define
\[
F_1^{c\bar{\a}',c\bar{\a}}(\ve{x})=\ve{x}\times \theta^+_{c,c}\times \Theta_{\bar{\a}',\bar{\a}},
\]
extended $\mathbb{F}[U]$-equivariantly. Here $\theta^+_{c,c}$ is the top degree intersection point of $c$ with a small Hamiltonian translate of itself. This map is a chain map by \cite{ZemGraphTQFT}*{Proposition~8.5}.
\section{Relating the expanded and ordinary models of the involution}
\label{sec:relate-expanded}
In this section, we relate the expanded model of the involution from Section~\ref{sec:expanded-def} with the ordinary model from Section~\ref{sec:doubling-def}. If $\mathfrak{D}$ is a doubling enhanced Heegaard diagram, and $\tilde{\mathfrak{D}}$ is a basepoint expanded doubling diagram, such that the underlying (non-involutive) Heegaard diagrams for $\mathfrak{D}$ and $\tilde{\mathfrak{D}}$ coincide, we will construct a chain homotopy equivalence
\[
F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}\colon \mathit{CFI}(\mathfrak{D})\to \mathit{CFI}(\tilde{\mathfrak{D}}).
\]
\begin{prop} \label{prop:expansion}\,
\begin{enumerate}
\item\label{prop:expansion-part-2}
If $\mathfrak{D}$ and $\mathfrak{D}'$ are ordinary doubling diagrams which differ by an elementary equivalence, then
\[
\tilde\Psi_{\tilde \mathfrak{D}\to \tilde \mathfrak{D}'}\circ F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}+F_{\mathfrak{D}'\to \tilde \mathfrak{D}'}\circ \Psi_{\mathfrak{D}\to \mathfrak{D}'}\simeq 0.
\]
\item \label{prop:expansion-part-3} If $\tilde \mathfrak{D}, $ $\tilde \mathfrak{D}'$ and $\tilde \mathfrak{D}''$ are three basepoint expanded doubling enhancements of three handleslide equivalent diagrams, then
\[
\tilde{\Psi}_{\tilde \mathfrak{D}'\to \tilde \mathfrak{D}''}\circ \tilde{\Psi}_{\tilde \mathfrak{D}\to \tilde \mathfrak{D}'}\simeq \tilde{\Psi}_{\tilde \mathfrak{D}\to \tilde \mathfrak{D}''}.
\]
\end{enumerate}
\end{prop}
The definition of the map $F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}$ is in Section~\ref{subsec:expanding-the-involution}. The proof of Part~\eqref{prop:expansion-part-2} of Proposition~\ref{prop:expansion} is in Section~\ref{sec:expansion-relation-transition-maps}. Finally we prove Part~\eqref{prop:expansion-part-3} in Section~\ref{sec:composition-law-expanded-proof}.
\subsection{Preliminaries}
In this section, we review a few maps that will appear in the construction of $F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}$, and also construct an important hypercube.
Firstly, we recall the \emph{free-stabilization} maps from \cite{ZemGraphTQFT}*{Section~6}. If $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},\ve{w})$ is a Heegaard diagram, and $w'$ is a point in $\Sigma\setminus(\ve{\alpha}\cup \ve{\beta} \cup \ve{w})$, we may form a new diagram $\mathcal{H}^+=(\Sigma,\ve{\alpha}\cup \a_0,\ve{\beta}\cup \b_0, \ve{w}\cup \{w'\})$ by adding attaching curves $\a_0$ and $\b_0$ which are contained in a small ball centered at $w'$. We assume that $\a_0$ and $\b_0$ intersect in two points, and furthermore are disjoint from $\ve{\alpha}\cup \ve{\beta}\cup \ve{w}$.
In the above setting, there are chain maps
\[
S_{w'}^+\colon \mathit{CF}^-(\mathcal{H},\mathfrak{s})\to \mathit{CF}^-(\mathcal{H}^+, \mathfrak{s})\quad \text{and} \quad S_{w'}^-\colon \mathit{CF}^-(\mathcal{H}^+,\mathfrak{s})\to \mathit{CF}^-(\mathcal{H},\mathfrak{s})
\]
called the free-stabilization maps. They are given by the formulas
\[
S_{w'}^+(\ve{x})=\ve{x}\times \theta^+,\qquad S_{w'}^-(\ve{x}\times \theta^+)=0,\quad \text{and} \quad S_{w'}^-(\ve{x}\times \theta^-)=\ve{x},
\]
extended equivariantly over $\mathbb{F}[U]$.
If $\lambda$ is an arc which connects $w'$ to another basepoint in $\ve{w}$, then
\begin{equation}
S_{w'}^- A_{\lambda} S_{w'}^+=\id,
\label{eq:free-stabilization}
\end{equation}
for appropriately chosen almost complex structure. This is proven in \cite{ZemGraphTQFT}*{Lemma~7.10} by a model computation.
We now generalize Equation~\eqref{eq:free-stabilization} to the setting of hypercubes. Suppose that $\mathscr{H}=\mathit{CF}(\mathcal{L}_{\a},\mathcal{L}_{\b})$ is a hypercube formed by pairing hypercubes of attaching curves $\mathcal{L}_{\a}$ and $\mathcal{L}_{\b}$.
We form new hypercubes, $\mathcal{L}_{\a}^+$ and $\mathcal{L}_{\b}^+$, by adding small translates of $\a_0$ and $\b_0$ to each attaching curve of $\mathcal{L}_{\a}$ and $\mathcal{L}_{\b}$. We form the morphisms of $\mathcal{L}_{\a}^+$ by tensoring each morphism of $\mathcal{L}_{\a}$ with $\theta_{\a_0,\a_0}^+$. We construct $\mathcal{L}_{\b}^+$ similarly. The hypercube relations for $\mathcal{L}_{\a}^+$ and $\mathcal{L}_{\b}^+$ follow from Proposition~\ref{prop:multi-stabilization-counts}. Finally, we define a new hypercube of chain complexes, $\mathscr{H}^+=\mathit{CF}(\mathcal{L}_{\a}^+,\mathcal{L}_{\b}^+)$.
We assume, additionally, that $\mathcal{L}_{\a}$ is a hypercube of handleslide equivalent alpha attaching curves, and that the length 1 morphisms are all cycles representing the top degree generator of homology. We construct the morphism
\[
\mathscr{A}_{\lambda}\colon \mathit{CF}(\mathcal{L}_{\a}^+,\mathcal{L}_{\b}^+)\to \mathit{CF}(\mathcal{L}_{\a}^+,\mathcal{L}_{\b}^+)
\]
as in Section~\ref{sec:hypercubes-rel-homology}.
We consider the following diagram:
\begin{equation}
\begin{tikzcd}[labels=description, row sep=1.2cm, column sep=1.2cm]
\mathscr{H}
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[ddd,"\id"]
& \mathscr{H}
\mathrm{a.r.}[d,"\mathscr{S}_{w'}^+"]
\\
&
\mathscr{H}^+
\mathrm{a.r.}[d, "\mathscr{A}_{\lambda}"]
\\
&
\mathscr{H}^+
\mathrm{a.r.}[d, "\mathscr{S}_{w'}^-"]
\\
\mathscr{H}
\mathrm{a.r.}[r,"\id"]
&
\mathscr{H}
\end{tikzcd}
\label{eq:add-basepoint-hypercube}
\end{equation}
The maps $\mathscr{S}_{w'}^+$ and $\mathscr{S}_{w'}^-$ denote hypercube versions of the free-stabilization maps; that is, they are hypercube morphisms with only length 1 maps, which are the ordinary free-stabilization maps on the Floer complexes. In the language of \cite{HHSZExact}, these are \emph{hypercubes of stabilization}.
If we compress the right-hand side of~\eqref{eq:add-basepoint-hypercube} we obtain a cube-shaped diagram. In the following, we verify that the resulting diagram is a hypercube of chain complexes:
\begin{prop}\label{prop:simple-expansion-hypercube}
Let $\mathscr{H}$ and $\mathscr{H}^+$ be as above. Then $\mathscr{A}_{\lambda}$ may be chosen so that the diagram in ~\eqref{eq:add-basepoint-hypercube} becomes a hypercube of chain complexes once we compress the right hand-side. In the hypercubes along the right-hand side, we use almost complex structures which are nodal near the free-stabilization region.
\end{prop}
We begin with a lemma:
\begin{lem}
\label{lem:A_lambda-cube-stabilized}
Let $\mathscr{H}$ and $\mathscr{H}^+$ be as above. The hypercube $(\mathcal{L}_\a^+)^{\lambda}$ may be chosen so that each chain of length at least 2 in the $\lambda$ direction is of the form $U\eta_{\a',\a}\otimes \theta^+_{\a_0,\a_0}$.
\end{lem}
\begin{proof} The proof is to follow the steps of the construction in Section~\ref{sec:hypercubes-rel-homology} and verify that each of the chains in $\mathcal{L}_{\a}^\lambda$ of length 2 or more may be taken to be of the stated form. Suppose $\varepsilon<\varepsilon'$ and $|\varepsilon'-\varepsilon|_{L^1}=1$. The chain of $\mathcal{L}_{\a}^+$ from $\ve{\alpha}_{\varepsilon}$ to $\ve{\alpha}_{\varepsilon'}$ is of the form $\Theta_{\a_{\varepsilon'},\a_{\varepsilon}}\otimes \theta^+_{\a_0,\a_0}$ by definition. As in Equation~\eqref{eq:length-2-relation-A-lambda}, the desired length 2 hypercube relation is
\begin{equation}
(A_\lambda+A_{\lambda}')(\Theta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+)=\d (U\eta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+). \label{eq:A_lambda-length-2-stabilized}
\end{equation}
From Equation~\eqref{eq:telescope-A-lambda}, we know
\[
(A_\lambda+A_{\lambda}')(\Theta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+)=U(\Phi_{w}+\Phi_{w'})(\Theta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+).
\]
Using the differential computation from \cite{ZemGraphTQFT}*{Proposition~6.5}, we see that $\Phi_{w'}(\Theta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+)=0$ and
\[
\Phi_{w}(\Theta_{\a',\a}\otimes \theta^+_{\a_0,\a_0})=\Phi_{w}^-(\Theta_{\a',\a})\otimes \theta_{\a_0,\a_0}^+
\]
where $\Phi_{w}^-$ denotes the basepoint action on the unstabilized complexes. In particular $\Phi_{w}^-(\Theta_{\a',\a})=\d \eta_{\a',\a}$, for some $\eta_{\a',\a}$, since $\Phi_w^-$ increases grading by 1, and $\Theta_{\a',\a}$ is in the maximal grading of homology. Using the fact that $S_{w'}^+$ is a chain map \cite{ZemGraphTQFT}*{Proposition~6.5}, we also have
\[
\d(\eta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+)=\d(\eta_{\a',\a})\otimes \theta_{\a_0,\a_0}^+.
\]
In particular, the chain $\eta_{\a',\a}$, constructed as above, satisfies Equation~\eqref{eq:A_lambda-length-2-stabilized}.
The higher length chains of $(\mathcal{L}_{\a}^+)^\lambda$ are analyzed similarly. In our present context, by using the stablization result for holomorphic polygons in Proposition~\ref{prop:multi-stabilization-counts}, we see that the right-hand side of Equation~\eqref{eq:A-lambda-1st-equation} can be written as $C_{\varepsilon',\varepsilon}\otimes \theta_{\a_0,\a_0}^+$, where $C_{\varepsilon',\varepsilon}$ is a chain on the unstabilized complex. By the exact same reasoning as before, $C_{\varepsilon',\varepsilon}$ is a multiple of $U$, and is a cycle, so we may pick a primitive of the form $U\eta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+$, completing the proof.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:simple-expansion-hypercube}] The hypercube relations for the right-hand side follow from the fact that each of the maps $\mathscr{S}_{w'}^+$, $\mathscr{A}_{\lambda}$ and $\mathscr{S}_{w'}^-$ is a homomorphism of hypercubes (i.e. a chain map).
The hypercube relations for the entire cube, after compressing the right-hand side of the diagram in Equation~\eqref{eq:add-basepoint-hypercube} are equivalent to the relation
\[
\mathscr{S}_{w'}^- \mathscr{A}_{\lambda} \mathscr{S}_{w'}^+=\id,
\]
as hypercube morphisms. First, we use Lemma~\ref{lem:A_lambda-cube-stabilized} to construct the hypercube $(\mathcal{L}_{\a}^+)^{\lambda}$ so that each chain of length at least 2 which has non-trivial $\lambda$ direction is of the form $U \eta_{\a',\a}\otimes \theta_{\a_0,\a_0}^+$. In the composition $\mathscr{A}_{\lambda} \mathscr{S}_{w'}^+$, each arrow with length greater than 1 will be obtained by a holomorphic polygon count which only has inputs of the form $\ve{x}\otimes \theta^+_{\a_0,\a_0}$. By Proposition~\ref{prop:multi-stabilization-counts}, the output of such a polygon count will also have a tensor factor of $\theta^+_{\a_0,\a_0}$. Such elements lie in the kernel of $S_{w'}^-$. Hence the composition $\mathscr{S}_{w'}^- \mathscr{A}_{\lambda} \mathscr{S}_{w'}^+$ has no arrows of length greater than 1. The arrows of length 1 correspond to the composition $S_{w'}^- A_{\lambda} S_{w'}^+$, which is the identity, by \cite{ZemGraphTQFT}*{Lemma~7.10}.
\end{proof}
\subsection{The map \texorpdfstring{$F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}$}{F:D->D-tilde}}\label{subsec:expanding-the-involution}
In this section, we define the map $F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}$. The map will correspond to a hypercube relating the expanded model of the involution from Section~\ref{sec:expanded-def} and the non-expanded model from Section~\ref{sec:doubling-def}.
The construction proceeds in two steps. We first relate the singly-pointed involution to a model which adds a basepoint in a very simple way. We call this the \emph{trivially expanded} involution. As a second step, we relate the trivially expanded model to the expanded model defined in Section~\ref{sec:expanded-def}.
Let $(\Sigma,\ve{\alpha},\ve{\beta},w)$ denote a Heegaard diagram, and let $w'$ be a new basepoint, chosen near $w$. Let $\a_0$ and $\b_0$ be two new alpha and beta circles, centered at $w'$. We will let $s$ denote either of the curves $\a_0$ or $\b_0$ (or small translates thereof).
\begin{figure}[ht]
\[
\begin{tikzcd}[labels=description, row sep=.8cm, column sep=1.5cm]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[d,"F_1^{\bar \b, \bar \b}"]
\mathrm{a.r.}[dr,phantom, "\fbox{A}"]
&[-1cm]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d,"F_1^{\bar \b, \bar \b}"]
\mathrm{a.r.}[r, "\id"]
\mathrm{a.r.}[dr,phantom, "\fbox{B}"]
&
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[r, "\id"]
\mathrm{a.r.}[d,"F_{1}^{s\bar \b , s\bar \b}"]
\mathrm{a.r.}[dr,phantom, "\fbox{C}"]
&[-1cm]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d,"F_{1}^{s\bar \b, s\bar \b}"]
\\[1cm]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[d, "f_{\a \bar \b}^{\b\bar \b\to \Delta}"]
\mathrm{a.r.}[dr,phantom, "\fbox{D}"]
&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[d, "f_{\a \bar \b}^{\b\bar \b\to \Delta}"]
\mathrm{a.r.}[r, "S_{w'}^+"]
\mathrm{a.r.}[dr,phantom, "\fbox{E}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\beta} s \bar \ve{\beta})
\mathrm{a.r.}[d, "f_{\a s \bar \b}^{\b s\bar \b\to s\Delta}"]
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[dr,phantom, "\fbox{F}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta} , \ve{\beta} s \bar \ve{\beta})
\mathrm{a.r.}[d, "f_{\a s \bar \b}^{\b s\bar \b\to s\Delta}"]
\\[1cm]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[ddd, "\id"]
\mathrm{a.r.}[dddr,phantom, "\fbox{G}"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[d, "S_{w'}^+"]
\mathrm{a.r.}[r, "S_{w'}^+"]
\mathrm{a.r.}[dr,phantom, "\fbox{H}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[r, "\id"]
\mathrm{a.r.}[d," A_{\lambda}"]
\mathrm{a.r.}[dr,phantom, "\fbox{I}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s\ve{\Delta} )
\mathrm{a.r.}[d,"A_\lambda"]
\\
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d, "A_\lambda"]
\mathrm{a.r.}[r, "A_\lambda"]
\mathrm{a.r.}[dr,phantom, "\fbox{J}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[r, "\id"]
\mathrm{a.r.}[d,"f_{\a s \bar \b}^{ s\Delta\to \a s \bar \a}"]
\mathrm{a.r.}[dr,phantom, "\fbox{K}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d,"f_{\a s \bar \b}^{ s\Delta\to \a s \bar \a}"]
\\
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d, "S_{w'}^-"]
\mathrm{a.r.}[r, "f_{\a s \bar \b}^{ s\Delta\to \a s \bar \a}"]
\mathrm{a.r.}[dr,phantom, "\fbox{L}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\alpha}, \ve{\alpha} s \bar \ve{\alpha})
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[d,"S_{w'}^-"]
\mathrm{a.r.}[dr,phantom, "\fbox{M}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\alpha}, \ve{\alpha} s \bar \ve{\alpha})
\mathrm{a.r.}[d,"F_3^{\a s,\a s}"]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[d, "f_{\a \bar \b}^{\Delta\to \a \bar \a}"]
\mathrm{a.r.}[dr,phantom, "\fbox{N}"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[d,"f_{\a \bar \b}^{\Delta\to \a \bar \a}"]
\mathrm{a.r.}[r," f_{\Delta\to \b\bar \b}^{\b\bar{\a}}"]
\mathrm{a.r.}[dr,phantom, "\fbox{O}"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[r,"F_3^{\a,\a}"]
\mathrm{a.r.}[d,"F_3^{\a,\a}"]
\mathrm{a.r.}[dr,phantom, "\fbox{P}"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[d,"\id "]
\\[1cm]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[d, "F_3^{\a,\a}"]
\mathrm{a.r.}[dr,phantom, "\fbox{Q}"]
&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}\bar \ve{\alpha})
\mathrm{a.r.}[d, "F_3^{\a,\a}"]
\mathrm{a.r.}[r,"F_3^{\a,\a}"]
\mathrm{a.r.}[dr,phantom, "\fbox{R}"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r, "\id"]
\mathrm{a.r.}[d,"\id"]
\mathrm{a.r.}[dr,phantom, "\fbox{S}"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[d,"\id"]
\\[1cm]
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r,"\id"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r,"\id"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r,"\id"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\end{tikzcd}
\]
\caption{The first hyperbox used to construct the map $F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}$. None of the faces have length 2 maps. The boxed letter will be referred to in the proof of part \eqref{prop:expansion-part-2} of Proposition~\ref{prop:expansion}.}
\label{eq:transition-standard-to-trivial}
\end{figure}
In Figure~\ref{eq:transition-standard-to-trivial} each square commutes, and no length 2 maps are necessary. As our definition, we take the compression of the right-hand side of Figure~\ref{eq:transition-standard-to-trivial} as the trivially expanded model of the involution.
We now describe how to relate the trivially expanded model in Figure~\ref{eq:transition-standard-to-trivial} to the expanded model described in Section~\ref{sec:expanded-def}. To do this, we use the hyperbox shown in Figure~\ref{eq:trivial-model-to-neck}.
\begin{figure}[ht]
\[
\begin{tikzcd}[labels=description, row sep=1.2cm, column sep=2.2cm]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d,"F_{1}^{s\bar \b, s\bar \b}"]
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[dr,phantom, "\fbox{A'}"]
&
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d,"F_{1}^{s\bar \b,c\bar \b}"]
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[dr,phantom, "\fbox{B'}"]
&
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d,"F_{1}^{c\bar \b,c\bar \b}"]
\\
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\beta} s \bar \ve{\beta})
\mathrm{a.r.}[d, "f_{\a s \bar \b}^{\b s \bar \b\to s \Delta}"]
\mathrm{a.r.}[r,"f_{\a s \bar \b}^{\b s \bar \b\to \b c \bar \b}"]
\mathrm{a.r.}[dr,dashed]
\mathrm{a.r.}[dr,phantom, shift left=6,pos=.6, "\fbox{C'}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\beta} c \bar \ve{\beta})
\mathrm{a.r.}[d,"f_{\a s \bar \b}^{\b c \bar \b\to s \Delta}"]
\mathrm{a.r.}[r,"f_{\a s \bar \b\to \a c \bar \b}^{\b c \bar \b}"]
\mathrm{a.r.}[dr,dashed]
\mathrm{a.r.}[dr,phantom, shift left=6,pos=.6, "\fbox{D'}"]
&
\mathit{CF}(\ve{\alpha} c \bar \ve{\beta}, \ve{\beta} c \bar \ve{\beta})
\mathrm{a.r.}[d,"f_{\a c \bar \b}^{\b c\bar \b \to s\Delta}"]
\\
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d,"A_\lambda"]
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[dr,phantom, "\fbox{E'}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d,"A_\lambda"]
\mathrm{a.r.}[r,"f_{\a s \bar \b\to \a c \bar \b}^{s \Delta}"]
\mathrm{a.r.}[dr,dashed]
\mathrm{a.r.}[dr,phantom, shift left=6,pos=.6,"\fbox{F'}"]
&
\mathit{CF}( \ve{\alpha} c \bar \ve{\beta}, s\ve{\Delta})
\mathrm{a.r.}[d,"A_\lambda"]
\\
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d,"f_{\a s \bar \b}^{s \Delta \to \a s \bar \a}"]
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[dr,dashed]
\mathrm{a.r.}[dr,phantom, shift left=6,pos=.6,"\fbox{G'}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d,"f_{\a s \bar \b}^{s \Delta\to \a c'\bar \a}"]
\mathrm{a.r.}[r,"f_{\a s \bar \b\to \a c \bar \b}^{s \Delta}"]
\mathrm{a.r.}[dr,dashed]
\mathrm{a.r.}[dr,phantom, shift left=6,pos=.6,"\fbox{H'}"]
&
\mathit{CF}(\ve{\alpha} c \bar \ve{\beta}, s \ve{\Delta})
\mathrm{a.r.}[d,"f_{\a c \bar \b}^{s \Delta\to \a c'\bar \a}"]
\\
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\alpha} s \bar \ve{\alpha})
\mathrm{a.r.}[d,"\id"]
\mathrm{a.r.}[r,"f_{\a s \bar \b}^{\a s \bar \a\to \a c'\bar \a}"]
\mathrm{a.r.}[dr,phantom, "\fbox{I'}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\alpha} c'\bar \ve{\alpha})
\mathrm{a.r.}[r,"f_{\a s \bar \b\to \a c \bar \b}^{\a c'\bar \a}"]
\mathrm{a.r.}[d, "\id"]
\mathrm{a.r.}[dr,dashed]
\mathrm{a.r.}[dr,phantom, shift left=6,pos=.6,"\fbox{J'}"]
&
\mathit{CF}(\ve{\alpha} c \bar \ve{\beta}, \ve{\alpha} c' \bar \ve{\alpha})
\mathrm{a.r.}[d, "f_{\a c \bar \b\to \a c'\bar \b}^{\a c'\bar \a}"]
\\
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\alpha} s \bar \ve{\alpha})
\mathrm{a.r.}[d,"F_3^{\a s,\a s}"]
\mathrm{a.r.}[r, "f_{\a s \bar \b}^{\a s \bar \a\to \a c'\bar \a}"]
\mathrm{a.r.}[dr,phantom,"\fbox{K'}"]
&
\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\alpha} c'\bar \ve{\alpha})
\mathrm{a.r.}[d,"F_3^{\a s, \a c'}"]
\mathrm{a.r.}[r,"f_{\a s \bar \b\to \a c'\bar \b}^{\a c'\bar \a}"]
\mathrm{a.r.}[dr,phantom,"\fbox{L'}"]
&
\mathit{CF}(\ve{\alpha} c'\bar \ve{\beta}, \ve{\alpha} c'\bar \ve{\alpha})
\mathrm{a.r.}[d,"F_3^{\a c',\a c'}"]
\\
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r,"\id"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r,"\id"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\end{tikzcd}
\]
\caption{The second hyperbox used to construct the map $F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}$. The boxed letters label the faces (in particular, they do not label maps in the faces).}
\label{eq:trivial-model-to-neck}
\end{figure}
We now explain the maps appearing in equation Figure~\ref{eq:trivial-model-to-neck}. Except for the top and bottom-most cubes, all of the faces either commute on the nose, or are obtained by pairing hypercubes of attaching curves. The hypercubes are all of handleslide equivalent attaching curves, or handleslide equivalent attaching curves with an extra direction corresponding to $\lambda$. Hence the construction follows by the technique of filling empty hypercubes of handleslide equivalent attaching curves \cite{MOIntegerSurgery}*{Lemma~8.6}, or Section~\ref{sec:hypercubes-rel-homology}.
We now consider the top-most and bottom-most hypercubes. We claim that the hypercube relations may be verified by appealing to Proposition~\ref{prop:multi-stabilization-counts}. To this end, we state the following lemma:
\begin{lem}\,
\label{lem:s->c-multi-stabilize-verify}
\begin{enumerate}
\item With respect to the unique $\Spin^c$ structures which have torsion restriction to the boundaries of the corresponding 4-manifolds, the Heegaard triples
\[
(\bar \Sigma, s \bar \ve{\beta}, s \bar \ve{\beta}, c \bar \ve{\beta},w,w')\quad \text{and}\quad (\bar \Sigma, c \bar \ve{\beta}, s \bar \ve{\beta}, c \bar \ve{\beta},w,w')
\]
are weakly admissible multi-stabilizing Heegaard triples. Furthermore
\[
\widehat{f}_{s\bar \b, s \bar \b, c \bar \b}(\Theta^+_{s\bar \b, s \bar \b},\Theta^+_{s\bar \b, c \bar \b})=\Theta^+_{s\bar \b, c \bar \b}\quad \text{and} \quad \widehat{f}_{c \bar \b, s \bar \b, c \bar \b}(\Theta^+_{c \bar \b,s \bar \b},\Theta^+_{s \bar \b, c \bar \b})=\Theta^+_{c \bar \b, c \bar \b}.
\]
Here, $\widehat{f}_{s\bar \b, s \bar \b, c \bar \b}$ denotes the map on $\widehat{\mathit{CF}}$, and similarly for the other holomorphic triangle map.
\item Similarly, with respect to the unique $\Spin^c$ structures which have torsion restriction to the boundary,
\[
(\Sigma, \ve{\alpha} s, \ve{\alpha} s, \ve{\alpha} c',w,w')\quad \text{and} \quad (\Sigma, \ve{\alpha} c', \ve{\alpha} s, \ve{\alpha} c',w,w')
\]
are weakly admissible multi-stabilizing diagrams, and
\[
\widehat{f}_{\a s, \a s, \a c'} (\Theta^-_{\a s, \a s}, \Theta^+_{\a s, \a c'})=\Theta^-_{\a s, \a c'}\quad \text{and}\quad \widehat{f}_{\a c', \a s, \a c'}(\Theta^+_{\a c', \a s}, \Theta^-_{\a s, \a c'})=\Theta^-_{\a c',\a c'}.
\]
\end{enumerate}
\end{lem}
\begin{proof} The claims about admissibility are proven by reducing to the genus zero triples $(S^2, s,s,c,w,w')$, and so forth, by removing pairwise isotopic triples of curves from $\bar \ve{\beta}$ or $\ve{\alpha}$. See Figure~\ref{fig:7} for the model case. It is straightforward to verify admissibility for the genus 0 diagrams. Clearly the other conditions of a multi-stabilizing diagram are satisfied.
We now prove the claim about the holomorphic triangle maps. Both triangle maps may be interpreted as naturality maps for a sequence of handleslides. Since each of the Floer complexes appearing in the statement have a unique top degree intersection point, and a unique bottom degree intersection point, and also since the claims only involve the maps on $\widehat{\mathit{CF}}$, the claims follow from grading considerations.
\end{proof}
\begin{figure}[H]
\centering
\input{fig7.pdf_tex}
\caption{The curves $s$, $c$ and $c'$ in the connected sum neck region. The $w$ basepoint is indicated $\bullet$ and the $w'$ is indicated $\bullet'$.}\label{fig:7}
\end{figure}
\subsection{Proof of Part \eqref{prop:expansion-part-2} of Proposition~\ref{prop:expansion}}
\label{sec:expansion-relation-transition-maps}
We now show that the map $F_{\mathfrak{D}\to \tilde \mathfrak{D}}$, defined in the previous section, commutes up to chain homotopy for the transition maps for elementary equivalences of the alpha and beta curves. This is part~\eqref{prop:expansion-part-2} of Proposition~\ref{prop:expansion}.
Our proof is to build a 3-dimensional hypercube which realizes the homotopy commutation
\[
F_{\mathfrak{D}'\to \tilde{\mathfrak{D}}'}\circ \Psi_{\mathfrak{D}\to \mathfrak{D}'}\simeq \tilde{\Psi}_{\tilde \mathfrak{D}\to \tilde \mathfrak{D}'}\circ F_{\mathfrak{D}\to \tilde \mathfrak{D}},
\]
whenever $\mathfrak{D}$ and $\mathfrak{D}'$ differ by an elementary equivalence. We build this hypercube as follows. For each of the faces labeled (A)--(S) in Figure~\ref{eq:transition-standard-to-trivial}, and each of the faces labeled (A')--(L') in Figure~\ref{eq:trivial-model-to-neck}, we build a hyperbox of size $(1,1,2)$, which extends the corresponding sub-cube of the map $F_{\mathfrak{D}\to \tilde{\mathfrak{D}}}$ in the direction of the maps $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$ and $\tilde{\Psi}_{\tilde \mathfrak{D}\to \tilde \mathfrak{D}'}$. We can view each of these hyperboxes as a sequence of two hypercubes. The third coordinate of the first hypercube is an alpha equivalence, and the third coordinate of the second hypercube is a beta equivalence. (This pattern is parallel to the construction of the map $\Psi_{\mathfrak{D}\to \mathfrak{D}'}$; see Figure~\ref{def:transition-map-elementary-handleslide}). Write $\ve{\alpha}',$ $\ve{\beta}'$ and $\ve{\Delta}'$ for the attaching curves of $\mathfrak{D}'$.
The construction of the size $(1,1,2)$ hyperbox over the facet labeled B in Figure~\ref{eq:transition-standard-to-trivial} is shown below.
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep={2cm,between origins},
column sep={3cm,between origins},
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}
]
|[alias=A]|\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr,"f_{\a\to a'}^{\b}"]
\mathrm{a.r.}[dd,"F_1^{\bar \b, \bar \b}"]
\mathrm{a.r.}[dr,"\id"]
\mathrm{a.r.}[dddr,phantom, "\fbox{B}"]
&[-1cm]
&
\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[dd,"F_1^{\bar \b', \bar \b}"]
\mathrm{a.r.}[rr, "f_{\a'}^{\b\to\b'}"]
\mathrm{a.r.}[dr,"\id"]
&[-1cm]
&
\mathit{CF}(\ve{\alpha}',\ve{\beta}')
\mathrm{a.r.}[dd,"F_{1}^{\bar \b',\bar \b'}",]
\mathrm{a.r.}[dr,"\id"]
&[-1cm]
\,
\\[-.6cm]
&
|[alias=B]|\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr,"f_{\a\to \a'}^{\b}",crossing over]
&&
\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[rr, "f_{\a'}^{\b\to \b'}",crossing over]
&&
\mathit{CF}(\ve{\alpha}',\ve{\beta}')
\mathrm{a.r.}[dd,"F_1^{s \bar \b', s \bar \b'}"]
\\[.7cm]
|[alias=D]|
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a \bar \b\to \a'\bar \b'}^{\b \bar \b} ", pos=.7]
\mathrm{a.r.}[dr,"S_{w'}^+"]
&&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}', \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a' \bar \b'}^{\b \bar \b \to \b'\bar \b'}", pos=.7]
\mathrm{a.r.}[dr,"S_{w'}^+"]
&&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}', \ve{\beta}' \bar \ve{\beta}')
\mathrm{a.r.}[dr,"S_{w'}^+"]
\\[-.6cm]
&
|[alias=C]|\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\beta} s \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a s \bar \b \to \a' s \bar \b'}^{\b s \bar \b}"]
\mathrm{a.r.}[from=uu,crossing over,"F_1^{s\bar \b',s \bar \b}"]
&&
\mathit{CF}(\ve{\alpha}'s \bar \ve{\beta}', \ve{\beta} s \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a' s \bar \b'}^{\b s \bar \b\to \b' s \bar \b'} "]
\mathrm{a.r.}[from=uu, crossing over,"F_1^{s\bar \b,s \bar \b}"]
&&
\mathit{CF}(\ve{\alpha}' s \bar \ve{\beta}', \ve{\beta}' s \bar \ve{\beta}')
\end{tikzcd}
\label{eq:hyperbox-over-B}
\end{equation}
The hypercube relations for the diagram in~\eqref{eq:hyperbox-over-B} follow by using Proposition~\ref{prop:multi-stabilization-counts} to destabilize the holomorphic triangle counts.
The hyperbox over the face labeled (G) in Figure~\ref{eq:transition-standard-to-trivial} follows from the construction from Proposition~\ref{prop:simple-expansion-hypercube}. The size $(1,1,2)$ hyperboxes over the remaining faces (A)--(S) are similar to the construction in equation~\eqref{eq:hyperbox-over-B}. We leave the details to the reader.
Next, we move on to extending the hyperboxes from Figure~\ref{eq:trivial-model-to-neck}. The extensions of the faces labeled (C')--(J') are obtained by pairing hypercubes of handleslide equivalent attaching curves (with an extra dimension corresponding to $\lambda$ in the case of E' and F'). In particular, the extensions are constructed using techniques which are now standard, so we leave them to the reader.
We investigate the hypercubes labeled A', B', K' and L'. All four are constructed similarly, so we focus our attention on the face labeled A'. The extension is shown below:
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep={2cm,between origins},
column sep={3cm,between origins},fill opacity=.8,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}
]
|[alias=A]|\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr,"f_{\a\to a'}^{\b}"]
\mathrm{a.r.}[dd,"F_1^{s\bar \b, s\bar \b}"]
\mathrm{a.r.}[dr,"\id"]
\mathrm{a.r.}[dddr,phantom, "\fbox{A'}"]
&[-1cm]
&
\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[dd,"F_1^{\bar \b', \bar \b}"]
\mathrm{a.r.}[rr, "f_{\a'}^{\b\to\b'}"]
\mathrm{a.r.}[dr,"\id"]
&[-1cm]
&
\mathit{CF}(\ve{\alpha}',\ve{\beta}')
\mathrm{a.r.}[dd,"F_{1}^{\bar \b',\bar \b'}",]
\mathrm{a.r.}[dr,"\id"]
&[-1cm]
\,
\\[-.6cm]
&
|[alias=B]|\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr,"f_{\a\to \a'}^{\b}",crossing over]
&&
\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[rr, "f_{\a'}^{\b\to \b'}",crossing over]
&&
\mathit{CF}(\ve{\alpha}',\ve{\beta}')
\mathrm{a.r.}[dd,"F_1^{s \bar \b', s \bar \b'}"]
\\[.7cm]
|[alias=D]|
\mathit{CF}(\ve{\alpha} s\bar \ve{\beta}, \ve{\beta} s \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a s \bar \b\to \a' s\bar \b'}^{\b s \bar \b} ", pos=.7]
\mathrm{a.r.}[dr,"f_{\a s \bar \b}^{\b s \bar \b \to \b c \bar \b}"]
&&
\mathit{CF}(\ve{\alpha}' s \bar \ve{\beta}', \ve{\beta} s \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a' s \bar \b'}^{\b s \bar \b \to \b' s\bar \b'}", pos=.7]
\mathrm{a.r.}[dr,"f_{\a' s \bar \b'}^{\b s \bar \b \to \b c \bar \b}"]
&&
\mathit{CF}(\ve{\alpha}'s \bar \ve{\beta}', \ve{\beta}'s \bar \ve{\beta}')
\mathrm{a.r.}[dr,"f_{\a' s \bar \b'}^{\b' s \bar \b'\to \b' c \bar \b'}"]
\\[-.6cm]
&
|[alias=C]|\mathit{CF}(\ve{\alpha} s \bar \ve{\beta}, \ve{\beta} c \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a s \bar \b \to \a' s \bar \b'}^{\b c \bar \b}"]
\mathrm{a.r.}[from=uu,crossing over,"F_1^{s\bar \b,c \bar \b}"]
&&
\mathit{CF}(\ve{\alpha}'s \bar \ve{\beta}', \ve{\beta} c \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a' s \bar \b'}^{\b c \bar \b\to \b' c \bar \b'} "]
\mathrm{a.r.}[from=uu, crossing over,"F_1^{s\bar \b',s \bar \b}"]
&&
\mathit{CF}(\ve{\alpha}' s \bar \ve{\beta}', \ve{\beta}' c \bar \ve{\beta}')
\end{tikzcd}
\label{eq:hyperbox-over-A'}
\end{equation}
The hypercube relations follow from the stabilization results of Proposition~\ref{prop:multi-stabilization-counts}, together with Lemma~\ref{lem:s->c-multi-stabilize-verify}.
\subsection{The composition law for expanded transition maps}
\label{sec:composition-law-expanded-proof}
In this section we prove part \eqref{prop:expansion-part-3} of Proposition \ref{prop:expansion}. As in the hypotheses of the proposition, suppose that we have three handleslide equivalent Heegaard diagrams $(\Sigma,\ve{\alpha},\ve{\beta},w)$, $(\Sigma,\ve{\alpha}',\ve{\beta}',w)$ and $(\Sigma,\ve{\alpha}'',\ve{\beta}'',w)$, all with underlying surface $\Sigma$. We also pick three choices of basepoint expanded doubling curves $\ve{\Delta}$, $\ve{\Delta}'$ and $\ve{\Delta}''$, and let $c$ and $c'$ denote the curves in the connected sum region. We further assume that the tuple
\[
(\Sigma \# \bar \Sigma, \ve{\alpha}'',\bar \ve{\alpha}'', \ve{\alpha}',\bar \ve{\alpha}', \ve{\alpha}, \bar \ve{\alpha}, \ve{\beta}, \bar \ve{\beta}, \ve{\beta}',\bar \ve{\beta}', \ve{\beta}'',\bar \ve{\beta}'',\ve{\Delta},\ve{\Delta}',\ve{\Delta}'',c,c',w)
\]
is weakly admissible (cf. Remark~\ref{rem:admissibility-unconventional}). Write $\tilde \mathfrak{D}$, $\tilde\mathfrak{D}'$ and $\tilde\mathfrak{D}''$ for the associated basepoint expanded, doubling enhanced Heegaard diagrams. We will show that
\[
\tilde{\Psi}_{\tilde\mathfrak{D}\to \tilde\mathfrak{D}''}\simeq \tilde{\Psi}_{\tilde\mathfrak{D}'\to \tilde\mathfrak{D}''}\circ \tilde{\Psi}_{\tilde\mathfrak{D}\to \tilde\mathfrak{D}'}.
\]
Our proof goes by constructing a hyperbox which realizes the stated relation.
For each vertical level in the construction of $\tilde{\Psi}_{\tilde \mathfrak{D}\to \tilde\mathfrak{D}'}$, the hyperbox we construct will have a level which takes the form of a hyperbox of size $(2,2,1)$.
We describe the 1-handle hyperbox first. See Figure~\ref{fig:naturality-Sigma-fixed-1-handles} for the overall shape of the 1-handle hyperbox.
As a first step, we construct the hypercube shown in Equation~\eqref{eq:top-degree-generator-hypercube}.
\begin{equation}
\begin{tikzcd}[
column sep={2.3cm,between origins},
row sep=1.4cm,
labels=description,
fill opacity=.7,
text opacity=1,
]
\mathbb{F}[U] \mathrm{a.r.}[dr, "\id", pos=.6]
\mathrm{a.r.}[ddd, "\Theta_{\bar \b', \bar \b}"]
\mathrm{a.r.}[rr, "\id"]
\mathrm{a.r.}[ddddrrr,dotted, "\omega_{\bar \b'', \bar \b'}"]
&[.7 cm]
&\mathbb{F}[U]
\mathrm{a.r.}[rd, "\id"]
\mathrm{a.r.}[ddd, "\Theta_{\bar \b', \bar \b'}"]
&[.7 cm]
\,
\\
&[.7 cm]
\mathbb{F}[U]
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "\id"]
&\,
&[.7 cm]
\mathbb{F}[U]
\mathrm{a.r.}[ddd, "\Theta_{\bar \b'', \bar \b'}"]
\\
\\
\mathit{CF}(\bar \ve{\beta}', \bar \ve{\beta})
\mathrm{a.r.}[dr,"f_{\bar \b'\to \bar \b''}^{\bar \b}"]
\mathrm{a.r.}[rr, "f_{\bar \b'}^{\bar \b\to\bar \b'}"]
\mathrm{a.r.}[drrr,dashed, "h_{\bar \b'\to \bar \b''}^{\bar \b \to \bar \b'}"]
&[.7 cm]\,&
\mathit{CF}(\bar \ve{\beta}', \bar \ve{\beta}')
\mathrm{a.r.}[dr, "f_{\bar \b'\to \bar \b''}^{\bar \b'}"]
&[.7 cm]
\\
& [.7 cm]
\mathit{CF}(\bar \ve{\beta}'', \bar \ve{\beta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu, "\Theta_{\bar \b'', \bar \b}"]
\mathrm{a.r.}[rr, "f_{\bar \b''}^{\bar \b\to \bar \b'}"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul, dashed, "\lambda_{\bar \b'', \bar \b}"]
&
&[.7 cm]
\mathit{CF}(\bar \ve{\beta}'', \bar \ve{\beta}')
\mathrm{a.r.}[from=uuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuull,dashed,"\lambda_{\bar \b'', \bar \b'}"]
\end{tikzcd}
\label{eq:top-degree-generator-hypercube}
\end{equation}
The hypercube in Equation~\eqref{eq:top-degree-generator-hypercube} is constructed similarly to the procedure for \emph{filling} a hypercube of handleslide equivalent attaching curves, described by Manolescu and Ozsv\'{a}th \cite{MOIntegerSurgery}*{Lemma~8.6}. In Equation~\eqref{eq:top-degree-generator-hypercube}, an arrow labeled by an element $\ve{x}\in \mathbb{T}_{\a}\cap \mathbb{T}_{\b}$ means the map from $\mathbb{F}[U]$ to $\mathit{CF}(\ve{\alpha},\ve{\beta})$ given by
\[
1\mapsto \ve{x},
\]
extended equivariantly over $U$. The existence of $\lambda_{\bar \b'',\bar \b'}$, $\lambda_{\bar \b'', \bar \b'}$ and $\omega_{\bar \b'', \bar \b'}$ so that the diagram in~\eqref{eq:top-degree-generator-hypercube} satisfies the hypercube relations follows from the fact that $\bar \ve{\beta}$, $\bar\ve{\beta}'$ and $\bar\ve{\beta}''$ are handleslide equivalent.
It is helpful to concretely understand the hypercube relations in~\eqref{eq:top-degree-generator-hypercube}. The length 1 relations amount to each chain marked with $\Theta$ being a cycle. The length 2 relations are as follows:
\begin{equation}
\begin{split}
f_{\bar \b'}^{ \bar \b\to \bar \b'}(\Theta_{\bar \b', \bar \b})&=\Theta_{\bar \b', \bar \b'}\\
f_{\bar \b'\to \bar \b''}^{\bar \b'}(\Theta_{\bar \b', \bar \b'})&=\Theta_{\bar \b'', \bar \b'}\\
\Theta_{\bar \b'', \bar \b}+f_{\bar \b'\to \bar \b''}^{\bar \b}(\Theta_{\bar \b', \bar \b})&=\d \lambda_{\bar \b'', \bar \b}\\
f_{\bar \b''}^{\bar \b\to \bar \b'}(\Theta_{\bar \b'', \bar \b})+ \Theta_{\bar \b'', \bar \b'}&= \d \lambda_{\bar \b'', \bar \b'}\\
f_{\bar \b'\to \bar \b''}^{\bar \b'}\circ f_{\bar \b'}^{\bar \b\to \bar \b'}+f_{\bar \b''}^{\bar \b\to \bar \b'}\circ f_{\bar \b'\to \bar \b''}^{\bar \b}&=\left[\d, h_{\bar \b'\to \bar \b''}^{\bar \b\to \bar \b'}\right].
\end{split}
\label{eq:hypercube-relations-diagonal-product}
\end{equation}
Note that the first relation is forced by grading considerations, and the second is implied by a nearest point map computation. The third and fourth relations are clearly satisfiable, and the final relation is the associativity relations for quadrilaterals.
Finally, the length 3 relation is
\[
f_{\bar \b''}^{\bar \b \to \bar \b'}(\lambda_{\bar \b'', \bar \b})+\lambda_{\bar \b'', \bar \b'}+h_{\bar \b'\to \bar \b''}^{\bar \b \to \bar \b'}(\Theta_{\bar \b', \bar \b})=\d \omega_{\bar \b'', \bar \b'}.
\]
Note that the hypercube in~\eqref{eq:top-degree-generator-hypercube} is a new type of hypercube. It is not naturally realized by pairing alpha and beta hypercubes together. We can think of the left side as determining the following hypercube of \emph{alpha} attaching curves
\begin{equation}
\begin{tikzcd}[labels=description, column sep=1.5cm, row sep=1.5cm]
\bar{\ve{\beta}}
\mathrm{a.r.}[r,"1"]
\mathrm{a.r.}[dr,dashed,"\lambda_{\bar\b'',\bar\b}"]
\mathrm{a.r.}[d,"\Theta_{\bar \b', \bar \b}"]
&\bar{\ve{\beta}}
\mathrm{a.r.}[d, "\Theta_{\bar\b'',\bar\b}"]
\\
\bar{\ve{\beta}}'
\mathrm{a.r.}[r, "\Theta_{\bar\b'',\bar\b'}"]
&\bar{\ve{\beta}}''
\end{tikzcd}
\label{eq:subface-left-pure-1-handle}
\end{equation}
On the other hand, the front side is more naturally viewed as corresponding to the following hypercube of \emph{beta} attaching curves
\begin{equation}
\begin{tikzcd}[labels=description, column sep=1.5cm, row sep=1.5cm]
\bar{\ve{\beta}}''
\mathrm{a.r.}[r,"1"]
\mathrm{a.r.}[dr,dashed,"\lambda_{\bar\b'',\bar\b'}"]
\mathrm{a.r.}[d,"\Theta_{\bar \b'', \bar \b}"]
&\bar{\ve{\beta}}''
\mathrm{a.r.}[d, "\Theta_{\bar\b'',\bar\b'}"]
\\
\bar{\ve{\beta}}
\mathrm{a.r.}[r, "\Theta_{\bar\b,\bar\b'}"]
&\bar{\ve{\beta}}'
\end{tikzcd}
\label{eq:subface-front-pure-1-handle}
\end{equation}
The bottom face of~\eqref{eq:top-degree-generator-hypercube} is the pairing of hypercubes of attaching curves.
Next, we pick hypercubes of alpha attaching curves and beta attaching curves (respectively):
\begin{equation}
\begin{tikzcd}[labels=description, column sep=1.5cm, row sep=1.5cm]
\ve{\alpha}
\mathrm{a.r.}[r,"\Theta_{\a',\a}"]
\mathrm{a.r.}[dr,dashed,"\lambda_{\a'',\a}"]
\mathrm{a.r.}[d,"1"]
&\ve{\alpha}'
\mathrm{a.r.}[d, "\Theta_{\a'',\a'}"]
\\
\ve{\alpha}
\mathrm{a.r.}[r, "\Theta_{\a'',\a}"]
&\ve{\alpha}''
\end{tikzcd}
\quad \text{and} \quad
\begin{tikzcd}[labels=description, column sep=1.5cm, row sep=1.5cm]
\ve{\beta}
\mathrm{a.r.}[r,"\Theta_{\b,\b'}"]
\mathrm{a.r.}[dr,dashed,"\lambda_{\b,\b''}"]
\mathrm{a.r.}[d,"1"]
&\ve{\beta}'
\mathrm{a.r.}[d, "\Theta_{\b',\b''}"]
\\
\ve{\beta}
\mathrm{a.r.}[r, "\Theta_{\b,\b''}"]
&\ve{\beta}''
\end{tikzcd}
\label{eq:alpha-hypercubes}
\end{equation}
As usual, we assume that the length 1 morphisms represent top degree cycles. Note that we do not assume that the chains in~\eqref{eq:alpha-hypercubes} have any relation with the chains in~\eqref{eq:subface-left-pure-1-handle} or~\eqref{eq:subface-front-pure-1-handle}.
\begin{figure}[hp]
\[
\begin{tikzcd}[
column sep={2.1cm,between origins},
row sep=1.5cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}]
(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[dr, "\id"]
\mathrm{a.r.}[ddd, "F_1^{c \bar{\b}, c \bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a\to \a'}^{\b}"]
\mathrm{a.r.}[ddddrrr,dotted]
&[.7 cm]
&(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[rd, "f_{\a'\to \a''}^{\b}"]
\mathrm{a.r.}[ddd, "F_1^{c\bar{\b}', c \bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a'}^{\b\to \b'}"]
\mathrm{a.r.}[ddddrrr,dotted]
\mathrm{a.r.}[ddddr,dashed]
&[.7 cm]
&
(\ve{\alpha}',\ve{\beta}')
\mathrm{a.r.}[dr, "f_{\a'\to \a''}^{\b'}"]
\mathrm{a.r.}[ddd, "F_{1}^{c \bar{\b}', c \bar{\b}'}"]
&[.7 cm]
\\
&[.7 cm]
|[alias=A]|(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "f_{\a\to \a''}^{\b}"]
&\,
&[.7 cm]
(\ve{\alpha}'',\ve{\beta})
\mathrm{a.r.}[ddd, "F_1^{c\bar{\b}'', c\bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a''}^{\b\to \b'}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll, dashed]
&&
|[alias=B]|(\ve{\alpha}'',\ve{\beta}')
\mathrm{a.r.}[ddd, "F_{1}^{c\bar{\b}'', c\bar{\b}'}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll, dashed]
\\
\\
(\ve{\alpha} c \bar{\ve{\beta}}, \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[dr,"\id"]
\mathrm{a.r.}[rr, "f_{\a c \bar{\b}\to \a' c \bar{\b}'}^{\b c \bar{\b}}"]
\mathrm{a.r.}[drrr,dashed]
&[.7 cm]\,&
(\ve{\alpha}' c \bar{\ve{\beta}}', \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[dr, "f_{\a' c \bar{\b}'\to \a'' c \bar{\b}''}^{\b c \bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a'c \bar{\b}'}^{\b c \bar{\b}\to \b' c \bar{\b}'}"]
\mathrm{a.r.}[drrr,dashed]
&[.7 cm]\,
&
(\ve{\alpha}'c\bar{\ve{\beta}}',\ve{\beta}' c \bar{\ve{\beta}}')
\mathrm{a.r.}[dr, "f_{\a'c \bar{\b}'\to \a'' c \bar{\b}''}^{\b' c \bar{\b}'}"]
\\
& [.7 cm]
|[alias=D]|(\ve{\alpha} c \bar{\ve{\beta}}, \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu, "F_1^{c\bar{\b}, c\bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a c \bar{\b}\to \a'' c \bar{\b}''}^{\b c \bar{\b}}"]
&
&[.7 cm]
(\ve{\alpha}'' c \bar{\ve{\beta}}'', \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[rr, "f_{\a''c \bar{\b}''}^{\b c \bar{\b}\to \b' c \bar{\b}'}" ]
&&
|[alias=C]|(\ve{\alpha}'' c \bar{\ve{\beta}}'', \ve{\beta}' c \bar{\ve{\beta}}')
\mathrm{a.r.}[from=uuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuull,dashed]
\end{tikzcd}
\]
\vspace{1cm}
\[
\begin{tikzcd}[
column sep={2.1cm,between origins},
row sep=1.5cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}]
|[alias=A]|(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[dr, "\id"]
\mathrm{a.r.}[ddd, "F_1^{c \bar{\b}, c \bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a\to \a''}^{\b}"]
&[.7 cm]
&(\ve{\alpha}'',\ve{\beta})
\mathrm{a.r.}[rd, "\id"]
\mathrm{a.r.}[ddd, "F_1^{c\bar{\b}'', c \bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a''}^{\b\to \b'}"]
\mathrm{a.r.}[ddddrrr,dotted]
\mathrm{a.r.}[dddrr,dashed]
&[.7 cm]
&
|[alias=B]|(\ve{\alpha}'',\ve{\beta}')
\mathrm{a.r.}[dr, "f_{\a''}^{\b'\to \b''}"]
\mathrm{a.r.}[ddd, "F_{1}^{c \bar{\b}'', c \bar{\b}'}"]
&[.7 cm]
\\
&[.7 cm]
(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "f_{\a\to \a''}^{\b}"]
&\,
&[.7 cm]
(\ve{\alpha}'',\ve{\beta})
\mathrm{a.r.}[ddd, "F_1^{c\bar{\b}'', c\bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a''}^{\b\to \b''}"]
&&
(\ve{\alpha}'',\ve{\beta}'')
\mathrm{a.r.}[ddd, "F_{1}^{c\bar{\b}'', c\bar{\b}''}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll, dashed]
\\
\\
|[alias=D]|(\ve{\alpha} c \bar{\ve{\beta}}, \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[dr,"\id"]
\mathrm{a.r.}[rr, "f_{\a c \bar{\b}\to \a'' c \bar{\b}''}^{\b c \bar{\b}}"]
&[.7 cm]\,&
(\ve{\alpha}'' c \bar{\ve{\beta}}'', \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[dr, "\id"]
\mathrm{a.r.}[rr, "f_{\a''c \bar{\b}''}^{\b c \bar{\b}\to \b' c \bar{\b}'}"]
\mathrm{a.r.}[drrr,dashed]
&[.7 cm]\,
&
|[alias=C]|(\ve{\alpha}''c\bar{\ve{\beta}}'',\ve{\beta}' c \bar{\ve{\beta}}')
\mathrm{a.r.}[dr, "f_{\a''c \bar{\b}''}^{\b' c \bar{\b}'\to \b'' c \bar{\b}''}"]
\\
& [.7 cm]
(\ve{\alpha} c \bar{\ve{\beta}}, \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu, "F_1^{c \bar{\b}, c \bar{\b}}"]
\mathrm{a.r.}[rr, "f_{\a c \bar{\b}\to \a'' c \bar{\b}''}^{\b c \bar{\b}}"]
&
&[.7 cm]
(\ve{\alpha}'' c \bar{\ve{\beta}} '', \ve{\beta} c \bar{\ve{\beta}})
\mathrm{a.r.}[rr, "f_{\a''c \bar{\b}''}^{\b c \bar{\b}\to \b'' c \bar{\b}''}" ]
&&
(\ve{\alpha}'' c \bar{\ve{\beta}}'', \ve{\beta}'' c \bar{\ve{\beta}}'')
\end{tikzcd}
\]
\caption{Four hypercubes. The length 2 maps are sums of quadrilateral and triangle maps (obtained by pairing hypercubes). The two boxes are stacked along the gray faces.}
\label{fig:naturality-Sigma-fixed-1-handles}
\end{figure}
We now consider the diagram in Figure~\ref{fig:naturality-Sigma-fixed-1-handles}. We will presently define the maps appearing therein, and then subsequently prove the hypercube relations. All the maps are built from the hypercubes in~\eqref{eq:top-degree-generator-hypercube} and~\eqref{eq:alpha-hypercubes}. We note that the hypercube relations are slightly subtle, since they do not arise from simply pairing hypercubes of attaching curves together.
We focus first on the top left cube of Figure~\ref{fig:naturality-Sigma-fixed-1-handles}. The maps appearing there are as follows:
\begin{enumerate}
\item The length 2 map along the top is the one associated to pairing the 0-dimensional beta hypercube $\ve{\beta}$ with the alpha hypercube
\[
\begin{tikzcd}[labels=description, column sep=1.5cm, row sep=1.5cm]
\ve{\alpha}
\mathrm{a.r.}[r,"\Theta_{\a',\a}"]
\mathrm{a.r.}[dr,dashed,"\lambda_{\a'',\a}"]
\mathrm{a.r.}[d,"1"]
&\ve{\alpha}'
\mathrm{a.r.}[d, "\Theta_{\a'',\a'}"]
\\
\ve{\alpha}
\mathrm{a.r.}[r, "\Theta_{\a'',\a}"]
&\ve{\alpha}''
\end{tikzcd}
\]
\item The length 2 map along the bottom is the one obtained by pairing the 0-dimensional beta hypercube $\ve{\beta} c \bar \ve{\beta}$ with the hypercube
\[
\begin{tikzcd}[labels=description, column sep=3cm, row sep=1.5cm]
\ve{\alpha} c \bar \ve{\beta}
\mathrm{a.r.}[r,"\Theta_{\a',\a}|\theta^+|\Theta_{\bar \b', \bar\b}"]
\mathrm{a.r.}[d,"1"]
\mathrm{a.r.}[dr,dashed, "\lambda"]
&\ve{\alpha}' c \bar \ve{\beta}'
\mathrm{a.r.}[d, "\Theta_{\a'',\a'}|\theta^+|\Theta_{\bar \b '', \bar \b '}"]
\\
\ve{\alpha} c \bar \ve{\beta}
\mathrm{a.r.}[r, "\Theta_{\a'',\a}|\theta^+|\Theta_{\bar \b'', \bar \b}"]
&\ve{\alpha}'' c \bar \ve{\beta}''
\end{tikzcd}
\]
where
\[
\lambda=\lambda_{\a'',\a}|\theta^+|f_{\bar{\b}'', \bar{\b}', \bar{\b}}(\Theta_{\bar{\b}'', \bar{\b}'}, \Theta_{\bar{\b}', \bar{\b}}) +\Theta_{\a'',\a}|\theta^+|\lambda_{\bar{\b}'', \bar{\b}}.
\]
\item The length 2 map along the right side is
\[
\ve{x}\mapsto f_{\a'\to \a''}^{\b}(\ve{x})| \theta^+| \lambda_{\bar{\b}'',\bar{\b}}.
\]
\item The length 3 map is
\[
\ve{x} \mapsto h_{\a \to \a'\to \a''}^{\b}(\ve{x})| \theta^+| \lambda_{\bar{\b}'', \bar \b}+f_{\a'', \a,\b}(\lambda_{\a'',\a},\ve{x})|\theta^+|\lambda_{\bar{\b}'',\bar{\b}}.
\]
\end{enumerate}
\begin{lem}\label{lem:top-left-1-handle-subcube}
The maps described above for the top left subcube of the diagram in Figure~\ref{fig:naturality-Sigma-fixed-1-handles} satisfy the hypercube relations.
\end{lem}
\begin{proof} The length 1 relations are immediate, as are the length 2 relations for all but the right face of the cube. For the right face of the cube, using the stabilization result from Proposition~\ref{prop:multi-stabilization-counts}, applied to holomorphic triangles, we see that the hypercube relations amount to showing that if $\ve{x}\in \mathbb{T}_{\a''}\cap \mathbb{T}_{\b}$, then
\[
f_{\a'\to \a''}^{\b}(\ve{x})| \theta^+|\left(\Theta_{\bar \b'', \bar \b}+f_{\bar \b'', \bar \b', \bar \b}(\Theta_{\bar \b'', \bar \b'}, \Theta_{\bar \b', \bar \b})\right)=\d \left(f_{\a' \to \a''}^{\b}(\ve{x})| \theta^+ | \lambda_{\bar{\b}'', \bar \b}\right)+f_{\a'\to \a''}^{\b}(\d \ve{x})| \theta^+|\lambda_{\bar{\b}'', \bar{\b}},
\]
which follows from the form of the differential on the Heegaard diagram obtained by attaching a 1-handle. The key point is that the right-hand side of the above equation simplifies to just $f_{\a'\to \a''}^{\b}(\ve{x})| \theta^+|\d \lambda_{\bar{\b}'', \bar \b}$, which coincides with the left-hand side by Equation~\eqref{eq:hypercube-relations-diagonal-product}.
Next, we consider the length 3 relation of the cube. The desired relation is that if $\ve{x}\in \mathbb{T}_{\a}\cap \mathbb{T}_{\b},$ then
\begin{equation}
\begin{split}
&h_{\a\to \a'\to \a''}^{\b}(\ve{x})|\theta^+|\Theta_{\bar \b'', \bar \b} \\
&+f_{\a'',\a,\b}(\lambda_{\a'',\a},\ve{x})|\theta^+|\Theta_{\bar \b'', \bar \b}\\
&+h_{\a'' c \bar \b'', \a' c \bar \b', \a c\bar \b, \b c \bar \b}(\Theta_{\a'',\a'}|\theta^+|\Theta_{\bar \b'' \bar \b'}, \Theta_{\a',\a}|\theta^+|\Theta_{\bar \b' \bar \b},\ve{x}|\theta^+|\Theta_{\bar \b, \bar \b})\\
&+ (f_{\a'\to \a''}^{\b}\circ f_{\a \to \a'}^{\b})(\ve{x})|\theta^+|\lambda_{\bar \b'', \bar \b}\\
&+f_{\a'' c \bar \b'', \a c \bar \b, \b c \bar \b}(\lambda_{\a'',\a}|\theta^+|f_{\bar \b'', \bar \b', \bar \b}(\Theta_{\bar \b'', \bar \b'}, \Theta_{\bar \b', \bar \b}), \ve{x} |\theta^+|\Theta^+_{\bar \b, \bar \b})\\
&+f_{\a'' c \bar \b'', \a c \bar \b, \b c \bar \b}(\Theta_{\a'', \a}|\theta^+|\lambda_{\bar \b'', \bar \b},\ve{x} |\theta^+|\Theta_{\bar \b, \bar \b})\\
=&\left[ h_{\a\to \a'\to \a''}^{\b}(-)|\theta^+|\lambda_{\bar{\b}'', \bar{\b}}+f_{\a'',\a,\b}(\lambda_{\a'',\a},-)|\theta^+|\lambda_{\bar \b'', \bar \b},\d\right](\ve{x}).
\end{split}
\label{eq:hypercube-relations-diagonal-1-handle}
\end{equation}
We note that
\begin{equation}
\begin{split}
&\left[ h_{\a\to \a'\to \a''}^{\b}(-)|\theta^+|\lambda_{\bar{\b}'', \bar{\b}},\d\right](\ve{x})
\\
=&f_{\a'',\a,\b}(f_{\a'',\a',\a}(\Theta_{\a'',\a'},\Theta_{\a',\a}),\ve{x})|\theta^+|\lambda_{\bar{\b}'',\bar \b}\\
&+\left(f_{\a'\to \a''}^{\b}\circ f_{\a\to \a'}^{\b}\right)(\ve{x})|\theta^+|\lambda_{\bar \b'', \bar \b}\\
&+h_{\a\to \a'\to \a''}^{\b}(\ve{x})|\theta^+| \Theta_{\bar \b'', \bar \b}\\
&+h_{\a\to \a'\to \a''}^{\b}(\ve{x})|\theta^+| f_{\bar \b'', \bar \b', \bar \b}(\Theta_{\bar \b'', \bar \b'}, \Theta_{\bar \b', \bar \b}).
\end{split}
\label{eq:1-handle-hypercube-simpl-1}
\end{equation}
Also
\begin{equation}
\begin{split}
&\left[f_{\a'',\a,\b}(\lambda_{\a'',\a},-)|\theta^+|\lambda_{\bar \b'', \bar \b},\d\right](\ve{x})\\
=&f_{\a'', \a, \b}(\Theta_{\a'',\a}+f_{\a'',\a',\a}(\Theta_{\a'',\a'}, \Theta_{\a',\a}),\ve{x})|\theta^+|\lambda_{\bar \b'', \bar \b}
\\
+&
f_{\a'',\a,\b}(\lambda_{\a'',\a},\ve{x})|\theta^+|(\Theta_{\bar \b'', \bar \b}+f_{\bar \b'', \bar \b', \bar \b}(\Theta_{\bar \b'', \bar \b'}, \Theta_{\bar \b', \bar \b})).
\end{split}
\label{eq:1-handle-hypercube-simpl-2}
\end{equation}
Next, we note that the stabilization result from Proposition~\ref{prop:multi-stabilization-counts}
implies that
\begin{equation}
\begin{split}f_{\a'' c \bar \b'', \a c \bar \b, \b c \bar \b}(\lambda_{\a'',\a}|\theta^+|f_{\bar \b'', \bar \b', \bar \b}(\Theta_{\bar{\b}'', \bar{\b}'}, \Theta_{\bar \b', \bar \b}), \ve{x} |\theta^+|\Theta_{\bar \b, \bar \b})
&=f_{\a'', \a, \b}(\lambda_{\a'', \a}, \ve{x})|\theta^+|f_{\bar \b'', \bar \b', \bar \b}(\Theta_{\bar{\b}'', \bar{\b}'}, \Theta_{\bar \b', \bar \b}),
\\
f_{\a'' c \bar \b'', \a c \bar \b, \b c \bar \b}(\Theta_{\a'', \a}|\theta^+|\lambda_{\bar \b'', \bar \b},\ve{x} |\theta^+|\Theta_{\bar \b, \bar \b})
&=f_{\a'', \a, \b}(\Theta_{\a'',\a}, \ve{x})|\theta^+|\lambda_{\bar \b'', \bar \b}.
\end{split}
\label{eq:1-handle-hypercube-simpl-3}
\end{equation}
Using~\eqref{eq:1-handle-hypercube-simpl-1}, ~\eqref{eq:1-handle-hypercube-simpl-2} and~\eqref{eq:1-handle-hypercube-simpl-3} to simplify~\eqref{eq:hypercube-relations-diagonal-1-handle}, we see that the length 3 hypercube relation becomes equivalent to
\begin{equation}
\begin{split}
&h_{\a'' c \bar \b'', \a' c \bar \b', \a c\bar \b, \b c \bar \b}(\Theta_{\a'',\a'}|\theta^+|\Theta_{\bar \b'' \bar \b'}, \Theta_{\a',\a}|\theta^+|\Theta_{\bar \b',\bar \b},\ve{x}|\theta^+|\Theta_{\bar \b, \bar \b})\\
=&h_{\a\to \a'\to \a''}^{\b}(\ve{x})|\theta^+| f_{\bar \b'', \bar \b', \bar \b}(\Theta_{\bar \b'', \bar \b'}, \Theta_{\bar \b', \bar \b}).
\end{split}
\label{eq:hypercube-reformulated-diagonal-1handle}
\end{equation}
We now prove ~\eqref{eq:hypercube-reformulated-diagonal-1handle}. Suppose that $\psi$ and $\psi_0$ are classes of rectangles on $(\Sigma,\ve{\alpha}'',\ve{\alpha}',\ve{\alpha},\ve{\beta})$ and $(\bar{\Sigma}, \bar{\ve{\beta}}'', \bar{\ve{\beta}}',\bar{\ve{\beta}}, \bar{\ve{\beta}})$ such that $\mu(\psi)+\mu(\psi_0)=-1$. Let $C(\psi,\psi_0)$ denote the set of all homology classes which coincide with $\psi$ and $\psi_0$ outside the 1-handle region, and have $\theta^+$ as their input and output in the 1-handle region. (Such rectangles necessarily have index $-1$). Applying Proposition~\ref{prop:multi-stabilization-counts} to the 1-handle region allows us to remove the 1-handle region connecting $\Sigma$ and $\bar{\Sigma}$. We obtain
\begin{equation}
\sum_{\phi\in C(\psi,\psi_0)} \# \mathcal{M}(\phi)\equiv \#( \mathcal{M}(\psi)\times_{\ev_{\Box}} \mathcal{M}(\psi_0)) \pmod 2 \label{eq:rectangles-and-1-handles}.
\end{equation}
Here, we are also using the nodal almost complex structures around the 1-handle region on the left-hand side. Furthermore any class of rectangles which is not in some $C(\psi,\psi_0)$ for a pair $(\psi,\psi_0)$ satisfying $\mu(\psi)+\mu(\psi_0)=-1$ has no holomorphic representatives. Also, $\times_{\ev_{\Box}}$ denotes the fibered product over the evaluation map to the parameter space $\mathbb{R}\cong \mathcal{M}(\Box)$.
We write $h^0$ for the map from $\mathit{CF}^-(\ve{\alpha},\ve{\beta})$ to $\mathit{CF}^-(\ve{\alpha}''c\bar{\ve{\beta}}'',\ve{\beta} c \bar{\ve{\beta}})$ which is the composition
\[
h^0=h_{\a c \bar{\b} \to \a' c \bar{\b}'\to \a'' c \bar{\b}''}^{\b c \bar{\b}}\circ F_1^{c \bar{\b}, c \bar{\b} }.
\]
By Equation~\eqref{eq:rectangles-and-1-handles}, we may identify $h^0$ as the map which counts holomorphic rectangles on the disjoint union of two diagrams, viewed as a fibered product over $\ev_{\Box}$. We may consider a map $h^t$ for any generic $t\in \mathbb{R}$, which counts curve pairs $(u,u_0)$ on the disjoint union consisting of holomorphic quadrilaterals such that $\ev_{\Box}(u)=\ev_{\Box}(u_0)+t$.
We claim that $h^t=h^{t'}$ for generic $t$ and $t'$ in $\mathbb{R}$. Firstly, we note that since $\bar{\ve{\beta}}$ appears twice (once as a small Hamiltonian isotopy), the nearest point argument of \cite{HHSZExact}*{Proposition~11.5} (cf. \cite{LOTDoubleBranchedII}*{Section~3.4}) implies that for suitably small choices of translations, the only rectangles on $\bar{\Sigma}$ with representatives have Maslov index at least 0. In particular all holomorphic rectangles counted by $h^t$ for generic $t$ have $\mu(\psi)=-1$ and $\mu(\psi_0)=0$.
We consider the ends of the moduli spaces counted by $h^t$, ranging over $t\in [0,\infty)$. No degenerations are possible on the $\Sigma$-side, since all rectangles there have index $-1$. On the $\bar{\Sigma}$-side, rectangles have index $0$, so the possible degenerations are a holomorphic disk splitting off at some finite $t$, as well as the ends which appear as $t\to \infty$. The ends appearing at $t\to \infty$ correspond exactly to the map
\[
h_{\a\to \a'\to \a''}^{\b}(\ve{x})|\theta^+|f_{\bar{\b}'', \bar{\b}',\bar{\b}}(\Theta_{\bar{\b}'', \bar{\b}'}, f_{\bar{\b}', \bar{\b}, \bar{\b}}(\Theta_{\bar{\b}', \bar{\b}}, \Theta_{\bar{\b},\bar{\b}})).
\]
The ends at finite $t$ are slightly more subtle. First, one end which can appear is an end where a disk breaks off into the $\bar{\ve{\beta}}$-$\bar{\ve{\beta}}$ end. These cancel modulo two since $\Theta_{\bar{\b}, \bar{\b}}$ is a cycle. Holomorphic disks breaking off into the other ends are prohibited, since they would leave an index $-1$ holomorphic rectangle, with $\Theta_{\bar{\b},\bar{\b}}$ as input. (The most important end where this is prohibited is the outgoing $\bar{\ve{\beta}}''$-$\bar{\ve{\beta}}$ end, where a more general diagram might have degenerations.) In particular, we obtain~\eqref{eq:hypercube-reformulated-diagonal-1handle}, completing the proof.
\end{proof}
The hypercube in the bottom right of Figure~\ref{fig:naturality-Sigma-fixed-1-handles} is constructed by simply interchanging the roles of the alpha and beta curves in the above procedure.
We now construct the maps in the top right hypercube of Figure~\ref{fig:naturality-Sigma-fixed-1-handles}. As a first step, we pick chains to give hypercubes of attaching curves, as follows:
\begin{enumerate}
\item The top face is obtained by pairing the two hypercubes
\[
\begin{tikzcd}
\ve{\alpha}'\mathrm{a.r.}[r, "\Theta_{\a'',\a'}"]
&
\ve{\alpha}''
\end{tikzcd}
\quad \text{and} \quad
\begin{tikzcd}
\ve{\beta}\mathrm{a.r.}[r, "\Theta_{\b,\b'}"]
&
\ve{\beta}'
\end{tikzcd}
\]
\item The bottom face is obtained by pairing the two hypercubes
\[
\begin{tikzcd}[column sep =2.5cm]
\ve{\alpha}'c \bar{\ve{\beta}}'\mathrm{a.r.}[r, "\Theta_{\a'',\a'}|\theta^+|\Theta_{\bar \b'', \bar \b'}"]
&
\ve{\alpha}''c \bar{\ve{\beta}}''
\end{tikzcd}
\quad \text{and} \quad
\begin{tikzcd}[column sep =2.5cm]
\ve{\beta} c \bar \ve{\beta}\mathrm{a.r.}[r, "\Theta_{\b,\b'}|c|\Theta_{\bar \b, \bar \b'}"]
&
\ve{\beta}' c \bar \ve{\beta}'
\end{tikzcd}
\]
\item The length 2 map along the left face is
\[
\ve{x}\mapsto f_{\a'\to \a''}^{\b}(\ve{x})| \theta^+| \lambda_{\bar{\b}'',\bar{\b}}.
\]
\item The length 2 map along the front face is
\[
\ve{x}\mapsto f_{\a''}^{ \b\to \b'}(\ve{x})| \theta^+| \lambda_{\bar \b'', \bar \b'}.
\]
\item The length 3 map is the sum of the following two maps
\[
\begin{split}
\ve{x}&\mapsto \left(f_{\a''}^{\b\to \b'}\circ f_{\a'\to \a''}^{\b}\right)(\ve{x})|\theta^+|\omega_{\bar \b'', \bar \b'}\\
\ve{x}&\mapsto h^{t<t_0}(\ve{x})
\end{split}
\]
\end{enumerate}
The map $h^{t<t_0}$ is a new map, as follows. It counts pairs $(u,u_0)$, where $u$ and $u_0$ are holomorphic rectangles on $(\Sigma,\ve{\alpha}'',\ve{\alpha}',\ve{\beta},\ve{\beta}')$ and $(\bar{\Sigma}, \bar \ve{\beta}'',\bar \ve{\beta}',\bar\ve{\beta},\bar \ve{\beta}')$, respectively, both of index $-1$, such that
\[
\ev_{\Box}(u)<\ev_{\Box}(u_0).
\]
If $\ve{y}$ and $\ve{z}$ are the outputs of $u$ and $u_0$, respectively, then the $h^{t<t_0}(\ve{x})$ has a summand of $\ve{y}\times \theta^+\times \ve{z}$ (weighted by the appropriate $U$-power).
\begin{lem}
With the maps described above, the top right subcube of Figure~\ref{fig:naturality-Sigma-fixed-1-handles} is a hypercube of chain complexes.
\end{lem}
\begin{proof} The proof is similar to the proof of Lemma~\ref{lem:top-left-1-handle-subcube}. The length 1 and 2 hypercube relations are similar to the previous case, so we focus on the length 3 relation. If $\ve{x}\in \mathbb{T}_{\a'}\cap \mathbb{T}_{\b}$, the desired relation is
\begin{equation}
\begin{split}
&h_{\a'\to \a''}^{\b\to \b'}(\ve{x})|\theta^+|\Theta_{\bar \b'', \bar\b'}\\
&+f_{\a'' c \bar \b'', \b c \bar \b,\b' c \bar \b'}\left(f_{\a'\to \a''}^{\b}(\ve{x})|\theta^+|\lambda_{\bar \b'', \bar \b}, \Theta_{\b, \b'}|\theta^+|\Theta_{\bar \b, \bar \b'}\right)\\
&+\left(f_{\a''}^{\b\to \b'}\circ f_{\a'\to \a''}^{\b}\right) (\ve{x})|\theta^+|\lambda_{\bar \b'', \bar \b'}\\
&+h_{\a' c \bar \b'\to \a'' c \bar \b''}^{\b c \bar \b\to \b' c \bar \b'}(\ve{x}|\theta^+|\Theta_{\bar \b', \bar \b})\\
=&\left[\left(f_{\a''}^{\b\to \b'}\circ f_{\a'\to \a''}^{\b}\right)(-)|\theta^+|\omega_{\bar \b'', \bar \b'}+h^{t<t_0}, \d \right](\ve{x})
\end{split}
\label{eq:length-3-relation-well-defined-ab}
\end{equation}
By excising the special curves in the connected sum region using the results about holomorphic triangles and 1-handles from \cite{ZemGraphTQFT}*{Theorem~8.8} (compare Proposition~\ref{prop:multi-stabilization-counts}, above), we obtain
\begin{equation}
\begin{split}
&f_{\a'' c \bar \b'', \b c \bar \b,\b' c \bar \b'}\left(f_{\a'\to \a''}^{\b}(\ve{x})|\theta^+|\lambda_{\bar \b'', \bar \b}, \Theta_{\b, \b'}|\theta^+|\Theta_{\bar \b, \bar \b'}\right)\\
=&\left(f^{\b\to \b'}_{\a''}\circ f_{\a' \to \a''}^{\b}\right)(\ve{x})|\theta^+|f_{\bar \b''}^{\bar \b \to \bar \b'}(\lambda_{\bar \b'', \bar \b}).
\end{split}
\label{eq:hyperbox-holomorphic-triangles-ab-well-def}
\end{equation}
We now analyze the quadrilateral map $h_{\a' c \bar \b'\to \a'' c \bar \b''}^{\b c \bar \b\to \b' c \bar \b'}(\ve{x}|\theta^+|\Theta_{\bar \b', \bar \b})$. As with the previous case, we can destabilize the 1-handle region and identify the expression with the count of index $-1$ holomorphic quadrilaterals on $(\Sigma\sqcup \bar \Sigma)\times \Box$. We can view this moduli space as the fibered product over the evaluation map to $\mathcal{M}(\Box)\cong \mathbb{R}$. If $(\psi,\psi_0)$ is a homology class on $(\Sigma \sqcup \bar \Sigma)\times \Box$ which admits a holomorphic representative, then there are generically two possibilities:
\begin{enumerate}
\item\label{case:mu-1} $\mu(\psi)=0$ and $\mu(\psi_0)=-1$;
\item\label{case:mu-2} $\mu(\psi)=-1$ and $\mu(\psi_0)=0$.
\end{enumerate}
As before we will write $h^0$ for the map $h_{\a' c \bar \b'\to \a'' c \bar \b''}^{\b c \bar \b\to \b' c \bar \b'}(\ve{x}|\theta^+|\Theta_{\bar \b', \bar \b})$. For $t\in [0,\infty)$, we write $h^t$ for a deformation, which counts curves such that
\[
\ev_{\Box}(u_0)-\ev_{\Box}(u)=t.
\]
We count the ends of the moduli spaces defining the maps $h^t$ (ranging over $t\in [0,\infty)$). We obtain
\begin{equation}
\begin{split}
[\d, h^{t<t_0}](\ve{x})=&h_{\a' c \bar \b'\to \a'' c \bar \b''}^{\b c \bar \b\to \b' c \bar\b'}(\ve{x}|\theta^+|\Theta_{\bar \b', \bar \b})\\
&+(f^{\b\to \b'}_{\a''}\circ f_{\a'\to \a''}^{\b})(\ve{x})|\theta^+|h_{\bar \b'\to \bar \b''}^{\bar \b\to \bar \b'}(\Theta_{\bar \b', \bar \b})\\
&+h_{\a'\to \a''}^{\b\to \b'}(\ve{x})|\theta^+|\Theta_{ \bar \b'', \bar \b'}.
\end{split}
\label{eq:deform-diagonals-quadrilaterals}
\end{equation}
Note that instead of $\Theta_{\bar \b'', \bar \b'}$ in the final factor, what more naturally appears is
\[
f_{\bar \b'', \bar \b', \bar \b'}(\Theta_{\bar \b'', \bar \b'}, f_{\bar \b', \bar \b, \bar \b'}(\Theta_{\bar \b', \bar \b}, \Theta_{\bar \b, \bar \b'})).
\]
However this coincides with $\Theta_{\bar \b'', \bar \b'}$, by a nearest point computation, cf. \cite{HHSZExact}*{Proposition~11.1}.
Finally,
\begin{equation}
\begin{split}
&\left[\left(f_{\a''}^{\b\to \b'}\circ f_{\a'\to \a''}^{\b}(-)\right)|\theta^+|\omega_{\bar \b'', \bar \b'}, \d \right](\ve{x})\\
=&\left(f_{\a''}^{\b\to \b'}\circ f_{\a'\to \a''}^{\b}(\ve{x})\right)|\theta^+|\left(f_{\bar \b''}^{\bar \b \to \bar \b'}(\lambda_{\bar \b'', \bar \b})+\lambda_{\bar \b'', \bar \b'}+h_{\bar \b'\to \bar \b''}^{\bar \b \to \bar \b'}(\Theta_{\bar \b', \bar \b})\right)
\end{split}
\label{eq:def-omega-b''b'}
\end{equation}
By combining ~\eqref{eq:hyperbox-holomorphic-triangles-ab-well-def}, ~\eqref{eq:deform-diagonals-quadrilaterals} and ~\eqref{eq:def-omega-b''b'}, we quickly obtain ~\eqref{eq:length-3-relation-well-defined-ab}, completing the proof.
\end{proof}
\section{Holomorphic polygons and stabilizations} \label{sec:polygons}
In this section, we describe some technical results concerning stabilizations of Heegaard diagrams and holomorphic polygons. The results we cover are a generalization of those considered in \cite{HHSZExact}*{Section~9}.
\begin{define} Suppose that $\mathcal{D}_0=(\Sigma_0,\ve{\gamma}_1,\dots, \ve{\gamma}_n,\ve{w})$ is a weakly admissible Heegaard tuple and $\mathfrak{S}$ is a set of $\Spin^c$ structures on $X_{\gamma_1,\dots, \gamma_n}$, which is closed under the action of $\delta H^1(Y_{\gamma_i,\gamma_j})$ for $i<j$.
\begin{enumerate}
\item We say that $(\mathcal{D}_0,\mathfrak{S})$ is a \emph{multi-stabilizing diagram} if the following are satisfied. All elements of $\mathfrak{S}$ restrict to a single element of $\Spin^c(\d X_{\gamma_1,\dots, \gamma_n})$ which is furthermore torsion. Additionally, we require that if $\mathfrak{s}_1,\mathfrak{s}_2\in \mathfrak{S}$, then $d(\mathfrak{s}_1)=d(\mathfrak{s}_2)$, where $d(\mathfrak{s})=(c_1^2(\mathfrak{s})-2\chi-3\sigma)/4$.
\item We say that $(\mathcal{D}_0,\mathfrak{S})$ is \emph{algebraically rigid} if $\d=0$ on $\widehat{\mathit{CF}}(\Sigma_0, \ve{\gamma}_i, \ve{\gamma}_{i+1},\mathfrak{s}_{i,i+1},\ve{w})$, where $\mathfrak{s}_{i,i+1}$ is the restriction to $Y_{\gamma_i,\gamma_{i+1}}$ of the elements of $\mathfrak{S}$.
\end{enumerate}
\end{define}
\begin{example}
Let $\mathcal{D}_0$ be a Heegaard triple where all $\ve{\gamma}_i$ are pairwise related by small Hamiltonian translates. Let $\mathfrak{S}=\{\mathfrak{s}_0\}$, where $\mathfrak{s}_0$ is the unique $\Spin^c$ structure which has torsion restriction to the boundary. Then $(\mathcal{D}_0,\mathfrak{S})$ is a multi-stabilizing diagram. For suitable choices of translates, $\mathcal{D}_0$ is algebraically rigid.
\end{example}
\begin{rem} When the set $\mathfrak{S}$ is understood from context (such as the case when there is a unique $\Spin^c$ structure on $X_{\gamma_1,\dots, \gamma_n}$ which has torsion restriction to $\d X_{\gamma_1,\dots, \gamma_n}$), we will usually refer to the diagram $\mathcal{D}_0$ as being a multi-stabilizing diagram.
\end{rem}
In this section, we state a helpful result about stabilizations and holomorphic curves. Before we state our result, we introduce some notation.
As a first step, we recall some notation and basic facts about Stasheff's associahedron, $K_n$. We view $K_n$ as a convex polytope which models the compactification of $n+1$ marked points on the boundary of a disk. It is well known that $K_n$ admits an embedding into a Euclidean space and has the homology of a point. There is a well known cell structure on $K_n$, giving a convenient model for its homology, which we denote $C_*^{\mathrm{cell}}(K_n)$. The cells of this decomposition are in bijective correspondence with planar trees with $n$ inputs, one output, and no internal vertices of valence less than 3. If $T$ is such a tree, then the degree of the corresponding cell is
\[
n-1-\#(\text{internal vertices}).
\]
The differential of a tree $T$ is the sum of all ways of breaking an internal vertex into two vertices, both with valence at least 3. The degree 0 cells correspond to trees with only valence 3 internal vertices. There is a codimension $0$ cell of degree $n-2$, which has just one internal vertex.
Suppose that $\mathcal{D}=(\Sigma,\ve{\gamma}_1,\dots, \ve{\gamma}_n,w)$ is a weakly admissible Heegaard tuple and $J=(J_x)_{x\in K_{n-1}}$ is a stratified family of almost complex structures on $\Sigma\times D_n$ for computing holomorphic $n$-gons. Suppose also that $\mathfrak{S}\subset \Spin^c(X_{\gamma_1,\dots, \gamma_n})$ is a set which is closed under the actions of $\delta^1 H^1(Y_{\gamma_i,\gamma_j})$ for all $i<j$, and assume for simplicity that all elements of $\mathfrak{S}$ restrict to a single element of $\Spin^c(\d X_{\gamma_1,\dots, \gamma_n})$. Given a tree $T\in C_*^{\mathrm{cell}}(K_{n-1})$, there is a map
\[
f_{\mathcal{D}, \mathfrak{S}; J,T}\colon \widehat{\mathit{CF}}(\ve{\gamma}_1,\ve{\gamma}_2, \mathfrak{s}_{1,2})\otimes \cdots \otimes \widehat{\mathit{CF}}(\ve{\gamma}_{n-1},\ve{\gamma}_n,\mathfrak{s}_{n-1,n})\to \widehat{\mathit{CF}}(\ve{\gamma}_1,\ve{\gamma}_n, \mathfrak{s}_{1,n}),
\]
obtained by composing the polygon maps corresponding to each internal vertex of $T$.
If $T$ has only valence 3 internal vertices, then there is a unique point of $K_{n-1}$ corresponding to $T$, which is in the strata of maximal codimension. We write $J_T$ for the corresponding almost complex structure. The map $F_{\gamma_1,\dots, \gamma_n, \mathfrak{S};J_T}$ is obtained by successively composing holomorphic triangle maps according to the tree $T$.
More generally, we can view the holomorphic polygon maps as fitting together compatibly to give a chain map
\[
f_{\mathcal{D},\mathfrak{S}, J}\colon C_*^{\mathrm{cell}}(K_{n-1})\otimes \widehat{\mathit{CF}}(\ve{\gamma}_1,\ve{\gamma}_2,\mathfrak{s}_{1,2})\otimes \cdots \otimes \widehat{\mathit{CF}}(\ve{\gamma}_{n-1}, \ve{\gamma}_n,\mathfrak{s}_{n-1,n})\to \widehat{\mathit{CF}}(\ve{\gamma}_{1},\ve{\gamma}_n, \mathfrak{s}_{1,n}).
\]
The same statement holds for $\mathit{CF}^-$ as long as we are more careful about finiteness of curve counts, either by restricting $\Spin^c$ structures or working over $\mathbb{F}[[U]]$.
\begin{lem}
Suppose that $\mathcal{D}_0=(\Sigma_0,\ve{\gamma}_1,\dots, \ve{\gamma}_n,\ve{w})$ is a diagram and $\mathfrak{S}$ is a set of $\Spin^c$ structures such that $(\mathcal{D}_0,\mathfrak{S})$ is algebraically rigid and multi-stabilizing. Let $T_1$ and $T_2$ be two degree 0 trees. Then,
\[
\widehat{f}_{\mathcal{D}_0,\mathfrak{S};J_{T_1}}(\ve{x}_1,\dots, \ve{x}_{n-1})=\widehat{f}_{\mathcal{D}_0,\mathfrak{S};J_{T_2}}(\ve{x}_1,\dots, \ve{x}_{n-1}),
\]
for any $\ve{x}_i\in \mathbb{T}_{\gamma_i}\cap \mathbb{T}_{\gamma_{i+1}}.$
\end{lem}
\begin{proof}
The trees $T_1$ and $T_2$ are homologous as elements of $C_*^{\mathrm{cell}}(K_{n-1})$, since $H_*^{\mathrm{cell}}(K_{n-1})\cong \mathbb{Z}$. Hence,
\[
\widehat{f}_{\mathcal{D}_0,\mathfrak{S},J_{T_1}}, \widehat{f}_{\mathcal{D}_0,\mathfrak{S},J_{T_2}}\widehat{\mathit{CF}}(\ve{\gamma}_1,\ve{\gamma}_2, \mathfrak{s}_{1,2})\otimes \cdots \otimes \widehat{\mathit{CF}}(\ve{\gamma}_{n-1},\ve{\gamma}_n, \mathfrak{s}_{n-1,n})\to \widehat{\mathit{CF}}(\ve{\gamma}_1,\ve{\gamma}_n,\mathfrak{s}_{1,n})
\]
and chain homotopic.
The differentials on domain and codomain of the two maps vanish, so any two chain homotopic maps between them are equal.
\end{proof}
\begin{prop}\label{prop:multi-stabilization-counts}
Let $\mathcal{D}=(\Sigma,\ve{\delta}_1,\dots, \ve{\delta}_{n},\ve{w})$ and $\mathcal{D}_0=(\Sigma_0,\ve{\gamma}_1,\dots, \ve{\gamma}_{n},\ve{w}_0)$ be Heegaard $n$-diagrams, where $n>2$. Assume $\mathfrak{s}\in \Spin^c(X_{\delta_1,\dots, \delta_{n}})$ and $\mathfrak{S}\subset \Spin^c(X_{\gamma_1,\dots, \gamma_{n}})$ and assume that $(\mathcal{D}_0,\mathfrak{S})$ is an algebraically rigid multi-stabilizing Heegaard diagram. Form the connected sum $\mathcal{D}\# \mathcal{D}_0$ by adding a tube from $\mathcal{D}_0$ at some $w_0\in \ve{w}_0$ to a point on $\mathcal{D}$. Let $T$ be a degree 0 tree representing the generator of $C_*^{\mathrm{cell}}(K_{n-1})$. Suppose $J=(J_x)_{x\in K_{n-1}}$ and $J^0=(J^0_x)_{x\in K_{n-1}}$ are stratified families of almost complex structures on $\Sigma\times D_n$ and $\Sigma_0\times D_n$, respectively, for counting holomorphic $n$-gons. Let $\theta_1,\dots, \theta_{n-1}$ be homogeneously graded elements (necessarily cycles) of $\widehat{\mathit{CF}}(\Sigma_0,\ve{\gamma}_1,\ve{\gamma}_2,\mathfrak{s}_{1,2}),\dots, \widehat{\mathit{CF}}(\Sigma_0,\ve{\gamma}_{n-1}, \ve{\gamma}_{n},\mathfrak{s}_{n-1,n})$, respectively, and suppose that
\[
\ve{y}=\widehat{f}_{\mathcal{D}_0,\mathfrak{S};J^0_T}(\theta_1,\dots, \theta_{n-1})
\]
is non-zero. Then
\[
f_{\mathcal{D}\# \mathcal{D}_0, \mathfrak{s}\# \mathfrak{S}; J \wedge J^0}(\ve{x}_1\times \theta_1,\dots, \ve{x}_{n-1}\times \theta_{n-1})=f_{\mathcal{D}, \mathfrak{s}; J}(\ve{x}_1,\dots, \ve{x}_{n-1})\otimes \ve{y}+\sum_{\substack{
\ve{z}\in \mathbb{T}_{\gamma_1}\cap \mathbb{T}_{\gamma_n}\\
\gr(\ve{z})> \gr(\ve{y})}} \ve{q}_{\ve{z}}\otimes \ve{z},
\]
where $\ve{q}_{\ve{z}}$ are elements of $\ve{\mathit{CF}}^-(\Sigma,\ve{\delta}_1,\ve{\delta}_n)$.
\end{prop}
\begin{proof} The proof follows from similar reasoning to \cite{HHSZExact}*{Propositions~10.2 and 10.3}, as we now sketch. See also \cite{OSLinks}*{Section~6.1}. Since we are using almost complex structures which are singular at the connected sum point, the relevant moduli spaces are fibered products over $D_{n}^k\times K_{n-1}$, where $k$ is the multiplicity of a class at the connected sum point. Consider a class $\psi\wedge \psi_0$, where $\psi_0\in \pi_2(\theta_1,\dots, \theta_{n-1}, \ve{y}_0)$, for some intersection point $\ve{y}_0$, and such that $\mathfrak{s}_{\ve{w}_0}(\psi_0)\in \mathfrak{S}$. The assumption that $\widehat{f}_{\gamma_1,\dots, \gamma_n,\mathfrak{S};J_T^0}(\theta_1,\dots, \theta_{n-1})=\ve{y}$ is non-zero implies that there exists a class of triangles $\psi'\in \pi_2(\theta_1,\dots, \theta_{n-1},\ve{y})$ which has $\mathfrak{s}_{\ve{w}_0}(\psi')\in \mathfrak{S}$, $n_{\ve{w}_0}(\psi')=0$ and $\mu(\psi')=0$.
Since all elements of $\mathfrak{S}$ have the same degree,
we conclude that
\[
\gr(\ve{y}_0)-2 n_{\ve{w}_0}(\psi_0)+\mu(\psi_0)=\gr(\ve{y}).
\]
Using Sarkar's formula for the Maslov index, we see that
\[
\mu(\psi\wedge \psi_0)=\mu(\psi)+\mu(\psi_0)-2n_{w_0}(\psi_0).
\]
(The only difference between the formulas for $\mu(\psi\wedge \psi_0)$ and $\mu(\psi)+\mu(\psi_0)$ is in the Euler measure, which is corrected on the right hand side by $2n_{w_0}(\psi_0)$).
Hence, combining the above two formulas, we obtain
\[
\mu(\psi\wedge \psi_0)=\mu(\psi)+\gr(\ve{y})-\gr(\ve{y}_0)+n_{\ve{w}_0\setminus \{w_0\}}(\psi_0).
\]
The main claim concerns only classes where $\gr(\ve{y})-\gr(\ve{y}_0)\ge 0$. Since the map $f_{\mathcal{D}\# \mathcal{D}_0, \mathfrak{s}\# \mathfrak{s}_0; J\wedge J^0}$ counts index $3-n$ curves, we assume
\[
\mu(\psi)+\gr(\ve{y})-\gr(\ve{y}_0)+n_{\ve{w}_0\setminus \{w_0\}}(\psi_0)=3-n.
\]
By transversality for curves representing $\psi$, we may assume that $\mu(\psi)\ge 3-n$. Hence, if $\gr(\ve{y})-\gr(\ve{y}_0)\ge 0$, then we must have $\mu(\psi)=3-n$, $\gr(\ve{y})=\gr(\ve{y}_0)=0$ and $n_{\ve{w}_0\setminus \{w_0\}}(\psi_0)=0$. In particular, the unconstrained moduli space $\mathcal{M}(\psi)$ is 0-dimensional. It suffices to understand the moduli spaces on the $\Sigma_0$ side.
Let $k\ge 0$ be a fixed integer, and assume that $\gr(\ve{y})=\gr(\ve{y}_0)$. Write $\mathcal{M}_k(\theta_1,\dots, \theta_{n-1},\ve{y}_0)$ for the moduli space of curves representing a class in $\pi_2(\theta_1,\dots, \theta_{n-1},\ve{y}_0)$ with multiplicity $k$ at $w_0$. We now claim that
\[
\ev\colon \mathcal{M}_k(\theta_1,\dots, \theta_{n-1},\ve{y}_0)\to D_{n}^k\times K_{n-1}
\]
is odd degree if $\ve{y}_0$ is a summand of $\ve{y}$ in $\widehat{\mathit{CF}}(\ve{\gamma}_1,\ve{\gamma}_n)$, and is even degree if $\ve{y}_0$ is not a summand of $\ve{y}$. This clearly implies the claim.
Let $(\ve{d}, x)\in D_{n}^k\times K_{n-1}$ be generic, and pick a path $\gamma\colon [0,\infty)\to D_{n}^k\times K_{n-1}$, such that $\gamma(0)=(\ve{d},x)$ and $\gamma(t)$ has the following behavior as $t\to \infty$. As $t\to \infty$, we assume that $x(t)$ approaches the tree $T$, viewed as a point in $\d K_{n-1}$, while the $k$-tuple of points $\ve{d}(t)$ travel towards one of the boundary punctures of $D_{n}^k$. Furthermore, we assume that under the identification of this end as a half cylinder $[0,1]\times[0,\infty),$ they approach some fixed $\ve{d}'\subset ((0,1)\times \mathbb{R})^k$, modulo the $\mathbb{R}$-action on $[0,\infty)$. We consider the 1-dimensional moduli space of curves which have $\ev(u)\in \im \gamma$. The generic degenerations of this moduli space at finite $t$ correspond exactly to index 1 disks breaking off of the cylindrical ends. These cancel modulo 2, since the differentials vanish on each $\widehat{\mathit{CF}}(\ve{\gamma}_i,\ve{\gamma}_{i+1}, \mathfrak{s}_{i,i+1})$. The ends appearing as $t\to \infty$ correspond to points in the following set:
\[
\bigg(\coprod_{\substack{
\phi\in \pi_2(\theta_1,\theta_1)\\ \mu(\phi)=2k} } \mathcal{M}(\phi, \ve{d}')\bigg)\times \bigg(\coprod_{\substack{\psi_0\in \pi_2(\theta_1,\dots, \theta_{n-1},\ve{y}_0)\\ \mu(\psi_0)=0 \\ n_{\ve{w}_0}(\psi_0)=0}} \mathcal{M}_{J_T}(\psi_0)\bigg).
\]
The left-hand factor has odd cardinality by \cite{ZemDuality}*{Equation~(31)}. The count of the right-hand side is exactly the $\ve{y}_0$ coefficient of $\widehat{f}_{\mathcal{D}_0,\mathfrak{S}; J_T^0}(\theta_1,\dots, \theta_{n-1})$, completing the proof.
\end{proof}
\subsection{Remarks and special cases}
Proposition~\ref{prop:multi-stabilization-counts} generalizes \cite{HHSZExact}*{Propositions~10.2 and 10.3}, which concern triangles and quadrilaterals. In this section we describe some special cases.
\begin{rem} The results of \cite{HHSZExact}*{Propositions~10.2 and 10.3} are stated in terms of an \emph{index tuple}, which does not appear in our Proposition~\ref{prop:multi-stabilization-counts}. The stabilization formulas from \cite{HHSZExact} are stated only in the case that one of the index tuple coordinates was 0. The entry of the index tuple being zero implied a specific formula for $\ve{y}=\widehat{f}_{\mathcal{D}_0,\mathfrak{S};T}(\theta_1,\dots, \theta_n)$. See \cite{HHSZExact}*{Lemma~10.4}. In our present paper, we also need the stabilization formulas in cases when the entry of the index tuple is non-zero, but where we know the value of $\widehat{f}_{\mathcal{D}_0,\mathfrak{S};T}(\theta_1,\dots, \theta_n)$.
\end{rem}
\begin{example} Suppose that $\mathcal{D}$ and $\mathcal{D}_0$ are as in the statement of Proposition~\ref{prop:multi-stabilization-counts}, and suppose further that $\mathcal{D}_0$ has the property that each $\ve{\gamma}_i$ is obtained from $\ve{\gamma}_1$ by a small Hamiltonian translation of each of the curves. Then Proposition~\ref{prop:multi-stabilization-counts} implies that
\[
f_{\mathcal{D}\# \mathcal{D}_0, \mathfrak{s} \# \mathfrak{s}_0; J\wedge J^0}(\ve{x}_1\times \theta_{1,2}^+,\dots, \ve{x}_n\times \theta_{n-1,n}^+)=f_{\mathcal{D}, \mathfrak{s};J}(\ve{x}_1,\dots, \ve{x}_n)\otimes \Theta^+_{1,n}.
\]
Similarly,
\[
\begin{split}
&f_{\mathcal{D}\# \mathcal{D}_0, \mathfrak{s} \# \mathfrak{s}_0; J\wedge J^0}(\ve{x}_1\times \theta_{1,2}^-,\ve{x}_2\times \theta_{2,3}^+,\dots, \ve{x}_n\times \theta_{n-1,n}^+)\\
=&f_{\mathcal{D}, \mathfrak{s};J}(\ve{x}_1,\dots, \ve{x}_n)\otimes \theta^-_{1,n}+\sum_{\substack{\Theta\in \mathbb{T}_{\gamma_1}\cap \mathbb{T}_{\gamma_n} \\ \gr(\Theta)>\gr(\theta^-_{1,n})}} \ve{q}_{\ve{z}}\otimes \Theta.
\end{split}
\]
\end{example}
\section{1- and 3-handles} \label{sec:1-and-3-handles}
In this section, we define our cobordism maps for 1-handle and 3-handles.
\subsection{The construction}
We now construct the map for a 4-dimensional 1-handle. Suppose that $\mathbb{S}_0=\{p,p'\}$ is a (framed) 0-sphere in $Y$. We pick a Heegaard surface $(\Sigma,\ve{\alpha},\ve{\beta})$ so that $\mathbb{S}_0\subset \Sigma\setminus (\ve{\alpha}\cup \ve{\beta})$. We form a Heegaard surface for the surgered manifold $Y(\mathbb{S}_0)$ by cutting out small disks centered at $p$ and $p'$, and gluing in an annular diagram $(A,\gamma_0,\gamma_0)$, where $\gamma_0$ and $\gamma_0$ are meridians of the annulus which intersect in two points. (Here and onwards, we abuse notation and view one copy of $\gamma_0$ as a Hamiltonian translate of the other). Write $(\Sigma', \ve{\alpha} \gamma_0, \ve{\beta} \gamma_0,w)$ for the diagram obtained by this procedure.
We now define the 4-dimensional 1-handle map. To do this, we first pick paths $\lambda$ and $\lambda'$ from $p$ and $p'$ to $w$. We pick doubling arcs $\ve{\delta}$ for $\Sigma$ which are disjoint from $\lambda$ and $\lambda'$. Furthermore, we choose two more doubling arcs $\ve{\delta}_0$ on $\Sigma'$, which are contained in the union of a neighborhood of $\lambda$ and $\lambda'$, as well as the annular 1-handle region. We assume $\ve{\delta}\cup \ve{\delta}_0$ is a valid collection of doubling arcs for $\Sigma'$. Write $\ve{\Delta}$ and $\ve{\Delta}_0$ for the curves obtained by doubling $\ve{\delta}$ and $\ve{\delta}_0$, respectively.
Additionally, the involutive 1-handle map requires a choice of closed curve $\tau\subset \Sigma'\# \bar \Sigma'$, such that the path $\tau$ is obtained by doubling a properly embedded arc on $\Sigma'\setminus N(w)$ which intersects $\gamma_0$ transversely in a single point.
We define the 1-handle map to be the diagram obtained by compressing the following hyperbox:
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=2.4cm,
fill opacity=.7,
text opacity=1
]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d, "F_{1}^{\bar{\b},\bar{\b}}"]
\mathrm{a.r.}[rrr, "F_1^{\gamma_0,\gamma_0}"]
&[-1.5cm]
&[-1.5cm]
&[-.2cm]
\mathit{CF}(\ve{\alpha}\gamma_0,\ve{\beta}\gamma_0)
\mathrm{a.r.}[d,"F_{1}^{\bar{\b}\bar \gamma_0,\bar{\b} \bar \gamma_0}"]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[d,"f^{\b \bar{\b}\to \Delta}_{\a \bar{\b}}"]
&
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[r,equal]
\mathrm{a.r.}[d, "f^{\b \bar \b\gamma_0 \bar \gamma_0\to \Delta \gamma_0 \bar \gamma_0}_{\a \bar \b \gamma_0 \bar \gamma_0}"]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[d,"f^{\b \bar \b \gamma_0 \bar \gamma_0\to \Delta \Delta_0}_{\a \bar \b \gamma_0 \bar \gamma_0}"]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[d, "f^{\Delta\to \a \bar{\a}}_{\a \bar{\b}}"]
&
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[r, "f_{\a \bar \b \gamma_0 \bar \gamma_0}^{\Delta \gamma_0 \bar \gamma_0 \to \Delta \Delta_0}"]
\mathrm{a.r.}[d, "f_{\a \bar \b \gamma_0 \bar \gamma_0}^{\Delta \gamma_0 \bar \gamma_0\to \a \bar \a \gamma_0 \bar \gamma_0}"]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}\bar{\ve{\beta}} \gamma_0 \bar \gamma_0,\ve{\Delta} \ve{\Delta}_0)
\mathrm{a.r.}[d, "f_{\a \bar \b \gamma_0 \bar \gamma_0}^{\Delta \Delta_0\to \a \bar \a \gamma_0 \bar \gamma_0}"]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\alpha} \bar{\ve{\alpha}})
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[d, "F_{3}^{\a,\a}"]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[r, "B_\tau"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[d, "F_3^{\a\gamma_0,\a\gamma_0}"]
\\
\mathit{CF}(\bar{\ve{\beta}}, \bar{\ve{\alpha}})
\mathrm{a.r.}[rrr, "F_1^{\bar \gamma_0, \bar \gamma_0}"]
&&&
\mathit{CF}(\bar{\ve{\beta}} \bar{\gamma}_0,\bar{\ve{\alpha}} \bar{\gamma}_0)
\end{tikzcd}
\label{eq:1-handles}
\end{equation}
It will be helpful to observe that the bottom-most level of~\eqref{eq:1-handles} has the following slightly expanded form
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=1.5cm,
fill opacity=.7,
text opacity=1
]
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\alpha} \bar{\ve{\alpha}})
\mathrm{a.r.}[r, "F_1^{\bar \gamma_0, \bar \gamma_0}"]
\mathrm{a.r.}[d, equal]
&[-.5cm]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}\bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}\bar \gamma_0)
\mathrm{a.r.}[r,"F_1^{\gamma_0,\gamma_0}"]
\mathrm{a.r.}[d,equal]
&[-.5cm]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[r, "B_\tau"]
&[-1cm]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[d, "F_3^{\gamma_0,\gamma_0}"]
\\
\mathit{CF}(\ve{\alpha}\bar{\ve{\beta}}, \ve{\alpha}\bar{\ve{\alpha}})
\mathrm{a.r.}[r, "F_1^{\bar \gamma_0, \bar \gamma_0}"]
\mathrm{a.r.}[d, "F_3^{\a,\a}"]
&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \bar \gamma_0,\ve{\alpha} \bar \ve{\alpha} \bar \gamma_0)
\mathrm{a.r.}[rr,equal]
&&
\mathit{CF}(\ve{\alpha}\bar{\ve{\beta}} \bar{\gamma}_0,\ve{\alpha}\bar{\ve{\alpha}} \bar{\gamma}_0)
\mathrm{a.r.}[d,"F_3^{\a,\a}"]
\\
\mathit{CF}(\bar{\ve{\beta}}, \bar{\ve{\alpha}})
\mathrm{a.r.}[rrr, "F_1^{\bar \gamma_0, \bar \gamma_0}"]
&&&
\mathit{CF}(\bar{\ve{\beta}} \bar{\gamma}_0,\bar{\ve{\alpha}} \bar{\gamma}_0)
\end{tikzcd}
\label{eq:1-handles-expanded-bottom-portion}
\end{equation}
We now describe the construction of~\eqref{eq:1-handles}. The top-most face which has a diagonal map is obtained by pairing a hypercube of beta attaching curves with the hypercube of alpha attaching curves consisting of $\ve{\alpha}\bar \ve{\beta} \gamma_0\bar \gamma_0$.
We now consider the lowest face of~\eqref{eq:1-handles} which has a diagonal map. This map is obtained by pairing the 0-dimensional alpha hypercube $\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0$ with the diagram below. The map $B_\tau$ is the $H_1$-action for $\tau$, defined by counting changes across the beta curves, $\ve{\alpha}\bar \ve{\alpha} \gamma_0\bar \gamma_0$.
\begin{equation}
\begin{tikzcd}[labels=description, row sep =2cm, column sep =2.5cm]
\ve{\Delta} \gamma_0\bar \gamma_0
\mathrm{a.r.}[r,"\Theta_{\Delta \gamma_0\bar \gamma_0, \Delta \Delta_0}"]
\mathrm{a.r.}[d,"\Theta_{\Delta \gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[dr, dashed, "\eta"]
&
\ve{\Delta} \ve{\Delta}_0
\mathrm{a.r.}[d, "\Theta_{\Delta \Delta_0, \a \bar \a \gamma_0 \bar \gamma_0}"]
\\
\ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0
\mathrm{a.r.}[r ,"\tau"]
&
\ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0
\end{tikzcd}
\label{eq:1-handle-hypercube-involving-H1}
\end{equation}
In equation~\eqref{eq:1-handle-hypercube-involving-H1}, the hypercube relations are to be interpreted using the formalism of Section~\ref{sec:hypercubes-rel-homology}. In particular, the hypercube relations amount to the equation
\begin{equation}
\d \eta=B_\tau(\Theta_{\Delta\gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0})+f_{\Delta \gamma_0\bar \gamma_0, \Delta \Delta_0, \a \bar \a\gamma_0\bar \gamma_0}(\Theta_{\Delta \gamma_0\bar \gamma_0, \Delta\Dt_0}, \Theta_{\Delta\Dt_0,\a \bar \a \gamma_0\bar \gamma_0}).
\label{eq:1-handle-hypercube-involving-H1-unpacked}
\end{equation}
\begin{rem}
\label{rem:1-handle-many-choices}
There are many choices of morphisms in the diagram in equation~\eqref{eq:1-handle-hypercube-involving-H1}. Any two choices of top-degree generators $\Theta_{\Delta \gamma_0 \bar \gamma_0, \Delta \Delta_0}$ are homologous. Similarly any two choices of $\Theta_{\Delta \gamma_0 \bar \gamma_0, \a \bar\a \gamma_0 \bar \gamma_0}$ and $\Theta_{\Delta \Delta_0, \a \bar \a \gamma_0 \bar \gamma_0}$ are homologous. Unlike these morphisms, it is not the case that any two choices of diagonal morphism $\eta=\eta_{\Delta \gamma_0\bar \gamma_0, \a \bar \a \gamma_0 \bar \gamma_0}$ are homologous. Indeed this chain lies in the top degree of $\mathit{CF}^-(\#^{g+1} S^1\times S^2)$, and hence there two choices which are not homologous. Similarly, it is not the case that any two choices for the curve $\tau$ are homologous. Nonetheless, we will show in Lemma~\ref{lem:1-handles-and-handleslides} that resulting involutive 1-handle map obtained by compressing the diagram in equation~\eqref{eq:1-handles-expanded-bottom-portion} is independent from these choices, up to chain homotopy.
\end{rem}
We now show that this is satisfiable:
\begin{lem}
\label{lem:1-handle-bottom-hypercube}There is a homogeneously graded $\eta\in \mathit{CF}(\ve{\Delta}\gamma_0\bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0\bar \gamma_0)$ satisfying~\eqref{eq:1-handle-hypercube-involving-H1-unpacked}.
\end{lem}
\begin{proof}Any two choices of $\ve{\Delta}\Ds_0$ are related by a sequence of handleslides and isotopies. Via an associativity argument, it is sufficient to show the claim for any convenient choice of $\ve{\Delta}$ and $\ve{\Delta}_0$. In particular, we may assume that the curves of $\ve{\Delta}_0$ are disjoint from $\ve{\alpha}$ and $\bar \ve{\alpha}$. In particular, the relevant triple diagram $(\Sigma' \# \bar \Sigma', \ve{\Delta} \gamma_0 \bar \gamma_0,\ve{\Delta} \ve{\Delta}_0,\ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)$ may be decomposed as a connected sum
\[
(\Sigma\# \bar \Sigma,\ve{\Delta},\ve{\Delta},\ve{\alpha} \bar \ve{\alpha})\#(\mathbb{T}^2\# \mathbb{T}^2,\gamma_0 \bar \gamma_0, \ve{\Delta}_0, \gamma_0 \bar \gamma_0).
\]
By the stabilization result in Proposition~\ref{prop:multi-stabilization-counts}, it suffices to show the claim for the genus 2 triple $(\mathbb{T}^2\# \mathbb{T}^2,\gamma_0 \bar \gamma_0, \ve{\Delta}_0, \gamma_0 \bar \gamma_0)$.
There are two ways to verify the claim in this situation. The first strategy is to argue topologically via the following reasoning:
\begin{enumerate}
\item Topologically the triple $(\mathbb{T}^2\# \mathbb{T}^2,\gamma_0 \bar \gamma_0, \ve{\Delta}_0, \gamma_0 \bar \gamma_0)$ is for 0-surgery on an unknot in $S^1\times S^2$.
\item By picking $\ve{\Delta}_0$ appropriately, we may assume that $\tau$ runs parallel to a curve $\Delta\in \ve{\Delta}_0$. Furthermore, $\Delta$ is the meridian of the knot on which we are performing surgery, and in particular defines the dual of the surgery knot.
\item In general, if $U$ is a 0-framed unknot in $Y$, and $\mu$ is a meridian, then $\mathit{CF}(W(U))=A_\mu\circ F_{1}$, where $F_1$ is the 1-handle map. This is easily verified using a genus 1 diagram and a stabilization result for triangles.
\item By combining the ideas above, the proof is complete.
\end{enumerate}
Alternately, one may explicitly perform the computation by counting holomorphic triangles. Write $\theta^{\pm}$ (resp. $\bar{\theta}{}^{\pm}$) for the intersection points of $\gamma_0$ with its translate (resp. $\bar \gamma_0$ with its translate). In Figure~\ref{fig:4}, we show the four index 0 triangle classes which can contribute.
Two of the classes are in $\pi_2(\Theta^+,\Theta^+,\theta^+|\bar{\theta}{}^-)$ and two of the classes are in $\pi_2(\Theta^+,\Theta^+,\theta^-|\bar{\theta}{}^+)$.
If $\psi_1$ and $\psi_2$ denote the two classes in $\pi_2(\Theta^+,\Theta^+,\theta^+|\bar{\theta}{}^-)$, we claim
\begin{equation}
\# \mathcal{M}(\psi_1)+\# \mathcal{M}(\psi_2)\equiv 1\mod 2, \label{eq:sum-of-classes-is-1}
\end{equation}
and similarly for the two classes in $\pi_2(\Theta^+,\Theta^+,\theta^-|\bar{\theta}{}^+)$.
This claim is verified via a Gromov compactness argument which is illustrated and explained in Figure~\ref{fig:13}.
Hence $f_{\gamma_0\bar \gamma_0, \Delta_0, \gamma_0 \bar \gamma_0}(\Theta^+,\Theta^+)=\theta^+|\bar{\theta}^-+\theta^-|\bar{\theta}^+$, which is also easily seen to be the action $\tau$ on $\theta^+|\bar \theta^+$.
\end{proof}
\begin{figure}[ht]
\centering
\input{fig6.pdf_tex}
\caption{Verifying equation~\eqref{eq:1-handle-hypercube-involving-H1-unpacked} by counting triangles. Left: The diagram $(\mathbb{T}^2\# \mathbb{T}^2, \gamma_0\bar\gamma_0, \Delta_0)$. Right: The four classes of index 0 holomorphic triangles which contribute to the triangle map $f_{\gamma_0\bar \gamma_0,\Delta_0,\gamma_0\bar\gamma_0}(\Theta^+,\Theta^+)$.
}\label{fig:4}
\end{figure}
\begin{figure}[ht]
\centering
\input{fig13.pdf_tex}
\caption{A Gromov compactness argument which shows that the total count of holomorphic triangles of index 0 in $\pi_2(\Theta^+,\Theta^+,\theta^+|\bar{\theta}{}^-)$ is 1. In the left-most columns are index 1 classes of triangles. On the other columns of a given row, we show all decompositions of that class into an index 0 triangle class and an index 1 disk class (such that both classes have nonnegative multiplicities). Configurations $A$ and $E$ have the same total count as the $\theta^+|\bar{\theta}{}^-$ coefficient of our map. Configuration $B$ has total count 1. Configurations $C$ and $D$ have the same count. Gromov compactness gives $\# A+\# B+\# C\equiv 0$ and $\# D+\# E\equiv 0$, which quickly gives Equation~\eqref{eq:sum-of-classes-is-1}.}\label{fig:13}
\end{figure}
Finally, we verify that the bottom-most face of~\eqref{eq:1-handles} is a hypercube. Note that this face may be expanded into the diagram shown in~\eqref{eq:1-handles-expanded-bottom-portion}. In light of this, it suffices to show the following relation (note the similarity to Equation ~\eqref{eq:add-basepoint-hypercube}):
\begin{lem}\label{lem:formal-cube-1-handle}
As maps from $\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \bar \gamma_0)$ to $\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \bar \gamma_0)$, there is an equality
\[
F_3^{\gamma_0,\gamma_0}\circ B_\tau \circ F_1^{\gamma_0,\gamma_0}=\id.
\]
\end{lem}
\begin{proof}Compare \cite{ZemGraphTQFT}*{Lemma~7.10}. The relation follows from the form of the differential after attaching a 1-handle. Namely, it follows from \cite{ZemGraphTQFT}*{Proposition~8.5} that the differential satisfies
\[
\d(\ve{x}\times \theta^+_{\gamma_0,\gamma_0})= \d(\ve{x})\otimes \theta^+_{\gamma_0,\gamma_0}.
\]
Furthermore, the proof shows that the holomorphic curves which go from $\ve{x}\times \theta^+_{\gamma_0,\gamma_0}$ to $\ve{y}\times \theta^-_{\gamma_0,\gamma_0}$ for some $\ve{x}$ and $\ve{y}$ all have $\ve{x}=\ve{y}$ and have domain consisting of one of the two bigons contained in the 1-handle region. The 1-chain $\tau$ intersects the boundary of exactly one of these bigons, and hence $B_{\tau}(\ve{x}\times \theta_{\gamma_0,\gamma_0}^+)= \ve{x}\times \theta_{\gamma_0,\gamma_0}^-+\ve{z}\otimes \theta_{\gamma_0,\gamma_0}^+$, for some $\ve{z}\in \mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})$. The stated claim follows immediately.
\end{proof}
\subsection{Well-definedness of the 1-handle maps}
In this section, we prove that the stabilization maps are compatible with the maps for changing the alpha and beta curves via handleslides or isotopies, or changing the choice of doubling curves $\ve{\Delta}$ and $\ve{\Delta}_0$. We also prove that the involutive 1-handle maps are independent of the choice of the curve $\tau\subset \Sigma'\# \bar \Sigma'$, as well as the choice of chains used to construct the hypercubes in~\eqref{eq:1-handles}.
As a first step, we verify independence from the curve $\tau$:
\begin{lem}
\label{lem:1-handle-map-ind-tau}
Suppose that $\tau$ and $\tau'$ are two closed curves on $\Sigma'\# \bar \Sigma'$ obtained by doubling properly embedded arcs on $\Sigma'\setminus N(w)$ which intersect $\gamma_0$ transversely in a single point. Then there are choices of the additional auxiliary data in the construction so that the models of the 1-handle maps in Equation~\eqref{eq:1-handles} are chain homotopic.
\end{lem}
\begin{proof} In the compression of Equation~\eqref{eq:1-handles} the choice of $\tau$ only affects the diagonal map. There are two ways in which it appears. The first is via the choice of $\eta$, and the second is a summand which has a factor of
\begin{equation}
F_3^{\a \gamma_0,\a \gamma_0}\circ f_{\a \bar \b \gamma_0\bar \gamma_0}^{\Delta \gamma_0\bar \gamma_0\to \a \bar \a \gamma_0\bar \gamma_0\xrightarrow{\tau} \a \bar \a \gamma_0\bar \gamma_0}\circ F_{1}^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}.\label{eq:1-handle-map-tau-contribution}
\end{equation}
We consider this latter summand first. The map in Equation~\eqref{eq:1-handle-map-tau-contribution} counts holomorphic triangles weighted by $\# \tau \cap \d_{\a \bar \a \gamma_0\bar \gamma_0}(\psi)$. If $\psi\in \pi_2(\ve{x},\ve{y},\ve{z})$ is a class of triangles counted in the above composition, then both the $\ve{x}$ and $\ve{y}$ (the incoming intersection points) have $\theta_{\gamma_0,\gamma_0}^+$ as factors. By the stabilization result of Proposition~\ref{prop:multi-stabilization-counts}, if such a class has a representative, then $\ve{z}$ must have a factor of $\theta_{\gamma_0,\gamma_0}^+$ as well. However, such a term would evaluate trivially under $F_3^{\a\gamma_0,\a \gamma_0}$. Hence, Equation~\eqref{eq:1-handle-map-tau-contribution} vanishes.
We now consider the first summand, which involves the choice of $\eta$. Let $\tau$ and $\tau'$ be two choices of curves, as above, and let $\eta_\tau$ be some choice of diagonal chain for the hypercube from Equation~\eqref{eq:1-handle-hypercube-involving-H1}, when $\tau$ is used. It is straightforward to see that $[\tau]-[\tau']\in H_1(\Sigma'\# \bar \Sigma')$ lies in the span of the $\ve{\Delta}$ curves. Let $C$ be a 2-chain on $\Sigma$ such that $\d \Sigma$ is $\tau-\tau'$ plus a linear combination of small parallel pushoffs of the $\ve{\Delta}$ curves.
We define the 2-chain
\[
\eta_{\tau'}:=\eta_\tau+H_C(\Theta_{\Delta \gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0}).
\]
In the above, $H_C$ denotes the $\mathbb{F}[U]$-linear map whose value on an intersection point $\ve{x}\in \mathbb{T}_{\Delta\gamma_0\bar \gamma_0}\cap \mathbb{T}_{\a \bar \a \gamma_0\bar \gamma_0}$ is the sum of the multiplicities of $C$ at each factor of $\ve{x}$. It is straightforward to see that $\eta_{\tau'}$ is a valid choice of diagonal for the hypercube in Equation~\eqref{eq:1-handle-hypercube-involving-H1} which uses $\tau'$.
Since each summand of $H_C(\Theta_{\Delta \gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0})$ has $\theta_{\gamma_0,\gamma_0}^+$ as a factor, the same argument as above shows that
\[
F_3^{\gamma_0,\gamma_0}\circ f_{\a \bar \b\gamma_0\bar \gamma_0, \Delta \gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0}(F_1^{\gamma_0,\gamma_0}(\ve{x}), H_C(\Theta_{\Delta \gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0}))=0,
\]
so the diagonal maps computed using $\eta_\tau$ and $\eta_{\tau'}$ coincide.
\end{proof}
\begin{lem}\label{lem:1-handles-and-handleslides} Suppose that $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta})$ and $\mathcal{H}'=(\Sigma,\ve{\alpha}',\ve{\beta}')$ are two Heegaard diagrams for $Y$ with the same underlying Heegaard surface, such that $\ve{\alpha}'$ is obtained from $\ve{\alpha}$ by an elementary handleslide or isotopy. We assume $\ve{\beta}'$ is obtained similarly from $\ve{\beta}$. Let $\mathfrak{D}$ and $\mathfrak{D}'$ denote doubling enhancements of $\mathcal{H}$ and $\mathcal{H}'$ with choices of almost complex structures and doubling curves $\ve{\Delta}$ and $\ve{\Delta}'$. Let $\mathbb{S}_0$ denote a 0-sphere which is embedded in $\Sigma\setminus (\ve{\alpha}\cup \ve{\beta} \cup \ve{\alpha}'\cup \ve{\beta}')$ and let $\Sigma(\mathbb{S}_0)$ denote the result of attaching a 1-handle along $\mathbb{S}_0$ to $\Sigma$. Let $\ve{\Delta}_0$ and $\ve{\Delta}_0'$ denote extra choices of doubling arcs, as in the construction of the 1-handle map, and let $\mathfrak{D}(\mathbb{S}_0)$ and $\mathfrak{D}'(\mathbb{S}_0)$ denote the doubling enhancements of $(\Sigma(\mathbb{S}_0),\ve{\alpha}\gamma_0,\ve{\beta}\gamma_0)$ and $(\Sigma(\mathbb{S}_0),\ve{\alpha}'\gamma_0,\ve{\beta}'\gamma_0)$ by using the new doubling curves $\ve{\Delta}_0$ and $\ve{\Delta}_0'$, respectively. Let $\tau$ be a single choice of symmetric 1-cycle on $\Sigma(\mathbb{S}_0)\# \bar \Sigma(\mathbb{S}_0)$ which intersects $\gamma_0$ transversely at a single point. Finally, let $\mathit{CFI}(W(\mathbb{S}_0))$ and $\mathit{CFI}(W(\mathbb{S}_0))'$ denote the involutive 1-handle maps, computed with these two choices of data. Then
\begin{equation}
\Psi_{\mathfrak{D}(\mathbb{S}_0)\to \mathfrak{D}'(\mathbb{S}_0)}\circ \mathit{CFI}(W(\mathbb{S}_0))\simeq \mathit{CFI}(W(\mathbb{S}_0))'\circ \Psi_{\mathfrak{D}\to \mathfrak{D}'}.
\label{eq:transition-map-commutes-1-handles}
\end{equation}
\end{lem}
\begin{proof}
The proof is to enlarge the diagram in~\eqref{eq:1-handles} into a 3-dimensional hyperbox by adding an extra direction coming out of the page. One face will correspond to the map $\mathit{CFI}(W(\mathbb{S}_0))$ and an opposite face will correspond to the map $\mathit{CFI}(W(\mathbb{S}_0))'$. It will be straightforward to unpack the hypercube relations to obtain~\eqref{eq:transition-map-commutes-1-handles}. We presently describe the construction of the hyperbox.
We extend the top-most hypercube of~\eqref{eq:1-handles} to the hyperbox shown in equation~\eqref{eq:hyperbox-1-handles-top}. The unlabeled arrows are holomorphic triangle maps. There are no morphisms of length 2 or 3. The hypercube relations follow from the stabilization formula of Proposition~\ref{prop:multi-stabilization-counts}.
\begin{equation}
\begin{tikzcd}[
column sep={1.6cm,between origins},
row sep=1cm,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}]
\mathit{CF}(\ve{\alpha} ,\ve{\beta})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd, "F_1^{\bar \b,\bar \b}"]
\mathrm{a.r.}[rr,"F_1^{\gamma_0,\gamma_0}"]
&[.7 cm]
&\mathit{CF}(\ve{\alpha}\gamma_0,\ve{\beta} \gamma_0)
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd, "F_1^{\bar \b \bar\gamma_0, \bar \b \bar\gamma_0}"]
&[.7 cm]
\\
&[.7 cm]
|[alias=A]|\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr,"F_1^{\gamma_0,\gamma_0}"]
&\,
&[.7 cm]
|[alias=B]|
\mathit{CF} (\ve{\alpha}' \gamma_0, \ve{\beta}\gamma_0)
\mathrm{a.r.}[ddd,swap, pos=.35, "F_1^{\bar \b'\bar \gamma_0, \bar \b\bar \gamma_0}"]
\\
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&[.7 cm]\,&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr]
\\
& [.7 cm]
|[alias=D]|
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}', \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu, "F_1^{\bar \b', \bar \b}", pos=.35]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
&
&[.7 cm]
|[alias=C]|
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}'\gamma_0\bar \gamma_0, \ve{\beta} \bar \ve{\beta}\gamma_0\bar \gamma_0)
\end{tikzcd}
\hspace{-1.7cm}
\begin{tikzcd}[
column sep={1.6cm,between origins},
row sep=1cm,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}]
|[alias=A]|
\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd, "F_1^{\bar \b', \bar \b}"]
\mathrm{a.r.}[rr, "F_1^{\gamma_0,\gamma_0}"]
&[.7 cm]
&
|[alias=B]|
\mathit{CF}(\ve{\alpha}' \gamma_0, \ve{\beta} \gamma_0)
\mathrm{a.r.}[rd]
\mathrm{a.r.}[ddd, "F_1^{\bar \b'\bar \gamma_0,\bar \b \bar \gamma_0}"]
&[.7 cm]
\\
&[.7 cm]
\mathit{CF}(\ve{\alpha}',\ve{\beta}')
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr,"F_1^{\gamma_0,\gamma_0}"]
&\,
&[.7 cm]
\mathit{CF} (\ve{\alpha}'\gamma_0,\ve{\beta}'\gamma_0)
\mathrm{a.r.}[ddd, "F_1^{\bar \b' \bar \gamma_0, \bar \b'\bar \gamma_0} " ]
\\
\\
|[alias=D]|
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}', \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr,"F_1^{\gamma_0\bar \gamma_0,\gamma_0 \bar\gamma_0}"]
&[.7 cm]
&
|[alias=C]|
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}' \gamma_0 \bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr]
\\
& [.7 cm]
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}', \ve{\beta}' \bar \ve{\beta}')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu, "F_1^{\bar \b',\bar \b'}", pos=.4]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0 \bar\gamma_0}"]
&
&[.7 cm]
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}' \gamma_0 \bar \gamma_0, \ve{\beta}' \bar \ve{\beta}'\gamma_0 \bar \gamma_0)
\end{tikzcd}
\label{eq:hyperbox-1-handles-top}
\end{equation}
The hypercubes obtained by extending the two small left-most cubes of~\eqref{eq:1-handles} are similarly defined to be hyperboxes of stabilizations.
We now consider the right-most cubes of~\eqref{eq:1-handles} (i.e. the two cubes which have length 2-maps). The top-most cube is easily extended by building hypercubes of attaching curves where each length 1 chain is a cycle generating the top degree of $\mathit{HF}^-$. We leave the details to the reader.
We now consider the bottom-most cube of~\eqref{eq:1-handles} which has a length 2 map, and we describe its extension to a 3-dimensional hyperbox. (Recall that this hypercube involved more choices than usual, such as the choice of the chain $\eta$ and the curve $\tau$; see Remark~\ref{rem:1-handle-many-choices}.) The construction is obtained by pairing hypercubes of attaching curves. The main technical challenge in extending the 2-dimensional cube in~\eqref{eq:1-handle-hypercube-involving-H1} is building a 3-dimensional hypercube of attaching curves with the following form:
\begin{equation}
\begin{tikzcd}[
column sep={1.3cm,between origins},
row sep=.8cm,
labels=description,
fill opacity=.7,
text opacity=1,
]
\ve{\Delta} \gamma_0 \bar \gamma_0
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dddrr,dashed, "\eta"]
\mathrm{a.r.}[ddddrrr,dotted]
&[.7 cm]
& \ve{\Delta} \ve{\Delta}_0
\mathrm{a.r.}[rd]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[ddddr,dashed]
&[.7 cm]
\\
&[.7 cm]
\ve{\Delta}' \gamma_0 \bar \gamma_0
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr]
&\,
&[.7 cm]
\ve{\Delta}' \ve{\Delta}_0'
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
\\
\\
\ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr,"\tau"]
\mathrm{a.r.}[drrr,dashed, "h"]
&[.7 cm]\,&
\ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0
\mathrm{a.r.}[dr]
\\
& [.7 cm]
\ve{\alpha}' \bar \ve{\alpha}' \gamma_0 \bar \gamma_0
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
\mathrm{a.r.}[rr, "\tau"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul,dashed]
&
&[.7 cm]
\ve{\alpha}'\bar \ve{\alpha}' \gamma_0 \bar \gamma_0
\mathrm{a.r.}[from=uuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuull,dashed, "\eta'"]
\end{tikzcd}
\label{eq:hypercube-attaching-curves-1-handle-w-tau}
\end{equation}
In equation~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau}, the chains $\eta$ and $\eta'$ are the ones used to construct $\mathit{CFI}(W(\mathbb{S}_0))$ and $\mathit{CFI}(W(\mathbb{S}_0))'$, as constructed in Lemma~\ref{lem:1-handle-bottom-hypercube}. In particular, these morphisms are fixed, and similarly the top length 2 morphism has also already been chosen in the previous step, so we assume that chain is fixed. Note that the two morphisms from $\ve{\alpha} \bar \ve{\alpha} \gamma_0\bar\gamma_0$ to $\ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0$ are forced to coincide, since $\ve{\alpha}'$ is obtained from $\ve{\alpha}$ by an elementary handleslide, so there is a single top degree generator of $\mathit{CF}(\ve{\alpha}'\bar \ve{\alpha}'\gamma_0\bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}\gamma_0\bar \gamma_0)$, which is $\Theta_{\a'\bar \a', \a \bar \a}\otimes \theta_{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}^+$.
Assuming for the moment the existence of the hypercube in~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau}, one extends the bottom-most cube of~\eqref{eq:1-handles} by constructing two hypercubes. The first hypercube is obtained by pairing the back face of~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau} with the hypercube
\[
\begin{tikzcd}[column sep=2cm]
\ve{\alpha} \bar \ve{\beta} \mathrm{a.r.}[r, "\Theta_{\a \bar \b, \a'\bar \b'}"]& \ve{\alpha}'\bar \ve{\beta}'
\end{tikzcd}.
\]
The second hypercube is obtained by pairing the entirety of ~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau} with the 0-dimensional hypercube $\ve{\alpha}' \bar \ve{\beta}'$.
We now prove the existence of the hypercube in~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau}. The existence of length 2 chains in Equation~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau} along the top, left and right faces is straightforward.
We now consider the bottom face of~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau}. The desired relation is
\[
A_\tau(\Theta_{\a \bar \a \bar \gamma_0, \a'\bar \a' \gamma_0\bar \gamma_0})+B_\tau(\Theta_{\a \bar \a \bar \gamma_0, \a'\bar \a' \gamma_0\bar \gamma_0})=\d (h)
\]
where $h$ denotes the diagonal chain. Since $\tau$ is a closed curve, we have an equality $A_\tau=B_{\tau}$, the above equation is satisfied by $h=0$. (Compare Equation~\eqref{eq:telescope-A-lambda} for arcs instead of closed curves).
As well will see, setting $h=0$ might not allow the length 3 hypercube relation to be satisfied, so instead we use the chain
\begin{equation}
h=\varepsilon\cdot \Theta_{\a\bar \a \gamma_0\bar \gamma_0,\a'\bar \a'\gamma_0\bar \gamma_0},
\label{eq:length-2-map-1-handle-T}
\end{equation}
where $\varepsilon\in \mathbb{F}$ is to be determined.
Finally, we claim that we may make a choice of the constant $\varepsilon$ in equation~\eqref{eq:length-2-map-1-handle-T}, and also pick a length 3 chain for the cube in~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau} so that the hypercube relations are satisfied. To see this, let $C\in \mathit{CF}(\ve{\Delta} \gamma_0 \bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0 \bar \gamma_0)$ denote the sum of all compositions featuring in the length 3 hypercube relations, except for the length 3 map itself. Here, we momentarily set $\varepsilon=0$ in~\eqref{eq:length-2-map-1-handle-T}. The hypercube relations on the proper faces force $C$ to be a cycle. We note however that $C$ has the same grading as the top degree element generator of $\mathit{HF}^-(\ve{\Delta}\gamma_0\bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0)$. In particular, it either represents the top degree element or is null-homologous. If it is null-homologous, we set $\varepsilon=0$ in~\eqref{eq:hypercube-attaching-curves-1-handle-w-tau} and pick any primitive of $C$ in the appropriate grading to be the length 3 map in our hypercube. If $C$ represents the top degree generator, then we pick $\varepsilon=1$, and let the length 3 chain be any primitive of
\[
f_{\Delta\gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0, \a'\bar \a' \gamma_0\bar \gamma_0}(\Theta_{\Delta\gamma_0\bar \gamma_0, \a \bar \a \gamma_0\bar \gamma_0},\Theta_{\a \bar \a \gamma_0\bar \gamma_0, \a'\bar \a' \gamma_0\bar \gamma_0})+C.
\]
The hypercube relations are clearly satisfied.
Finally, it remains to extend the bottom-most cube of Figure~\ref{eq:1-handles}. To extend this hypercube, it is convenient to consider the expanded description in~\eqref{eq:1-handles-expanded-bottom-portion}. In fact, the only hypercube which requires explanation to extend is the cube
\[
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=1cm,
fill opacity=.7,
text opacity=1
]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \bar \gamma_0)
\mathrm{a.r.}[r, "F_1^{\gamma_0, \gamma_0}"]
\mathrm{a.r.}[d,equal]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0\bar \gamma_0)
\mathrm{a.r.}[r, "B_\tau"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0\bar \gamma_0)
\mathrm{a.r.}[d, "F_3^{\gamma_0,\gamma_0}"]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \bar \gamma_0)
\mathrm{a.r.}[rr, equal]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \bar \gamma_0)
\end{tikzcd}
\]
We define our 3-dimensional extension of this cube to have trivial length 3 map. The hypercube relations for the 3-dimensional extension of this cube are proven somewhat similarly to the proof of Proposition~\ref{prop:simple-expansion-hypercube}. The only non-trivial length 2 map in the above diagram is in the top subface. This subface has the following form
\[
\begin{tikzcd}[labels=description,column sep=1.3cm, row sep=1cm]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \bar \gamma_0,\ve{\alpha} \bar \ve{\alpha}\bar \gamma_0)
\mathrm{a.r.}[r, "F_1^{\gamma_0,\gamma_0}"]
\mathrm{a.r.}[d]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}\gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[r, "B_\tau"]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}\gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[d]\\
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}' \bar \gamma_0,\ve{\alpha} \bar \ve{\alpha}\bar \gamma_0)
\mathrm{a.r.}[r, "F_1^{\gamma_0,\gamma_0}"]
\mathrm{a.r.}[d]
&\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}'\gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[r, "B_{\tau}"]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}'\gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[d]\\
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}' \bar \gamma_0,\ve{\alpha}' \bar \ve{\alpha}'\bar \gamma_0)
\mathrm{a.r.}[r, "F_1^{\gamma_0,\gamma_0}"]&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}'\gamma_0 \bar \gamma_0, \ve{\alpha}' \bar \ve{\alpha}'\gamma_0\bar \gamma_0)
\mathrm{a.r.}[r, "B_{\tau}"]&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}'\gamma_0 \bar \gamma_0, \ve{\alpha}' \bar \ve{\alpha}'\gamma_0\bar \gamma_0)
\end{tikzcd}
\]
It is sufficient to show that the compression of this diagram composes trivially with the map $F_3^{\gamma_0, \gamma_0}$. The key tool we will use is that the triangle maps, when applied to elements of the form $(\ve{x}\times\theta_{\gamma_0,\gamma_0}^+)\otimes (\ve{y}\times \theta_{\gamma_0,\gamma_0}^+)$ have only summands of the form $\ve{z}\times \theta_{\gamma_0,\gamma_0}^+$. This fact follows from Proposition~\ref{prop:multi-stabilization-counts}. The top diagonal map counts holomorphic triangles which are weighted by the quantity $\# (\d_{\a\bar \a \gamma_0\bar \gamma_0}(\psi)\cap \tau)\in \mathbb{F}$. Similarly, the bottom triangle map is a sum of holomorphic triangle maps with certain weights. The special inputs for the length 1 triangle maps all have tensor factors of $\theta_{\gamma_0,\gamma_0}^+$, as does the special length 2 element used for the bottom diagonal map by equation~\eqref{eq:length-2-map-1-handle-T}. Hence, in compression of the diagram, the only holomorphic triangles which are counted have only tensor factors of $\theta_{\gamma_0,\gamma_0}^+$, and hence so does their output. In particular, the output vanishes when composed with $F_3^{\gamma_0,\gamma_0}$.
\end{proof}
\subsection{3-handles}
We define the 3-handle map by dualizing the construction of the 1-handle map. Concretely, this amounts to compressing the diagram shown in Equation~\eqref{eq:3-handles}.
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=2.6cm,
fill opacity=.7,
text opacity=1
]
\mathit{CF}(\ve{\alpha}\gamma_0,\ve{\beta}\gamma_0)
\mathrm{a.r.}[d, "F_{1}^{\bar{\b}\bar \gamma_0,\bar{\b}\bar \gamma_0}"]
\mathrm{a.r.}[rr, "F_3^{\gamma_0,\gamma_0}"]
&
&
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d,"F_{1}^{\bar{\b},\bar{\b}}"]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}} \gamma_0\bar \gamma_0, \ve{\beta} \bar{\ve{\beta}} \gamma_0\bar \gamma_0)
\mathrm{a.r.}[r, "B_\tau"]
\mathrm{a.r.}[d,"f^{\b \bar{\b} \gamma_0\bar \gamma_0\to \Delta\Dt_0}_{\a \bar{\b}\gamma_0 \bar \gamma_0}"]
\mathrm{a.r.}[dr,dashed]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[r,"F_3^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[d, "f^{\b \bar \b\gamma_0 \bar \gamma_0\to \Delta \gamma_0 \bar \gamma_0}_{\a \bar \b \gamma_0 \bar \gamma_0}"]
& \mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[d,"f^{\b \bar \b \to \Delta }_{\a \bar \b }"]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0,\ve{\Delta} \ve{\Delta}_0)
\mathrm{a.r.}[r, "f_{\a \bar \gamma_0\bar \gamma_0}^{\Delta\Dt_0\to \Delta \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[d, "f^{\Delta\to \a \bar{\a}}_{\a \bar{\b}}"]
\mathrm{a.r.}[dr,dashed]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[r, "F_3^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[d, "f_{\a \bar \b \gamma_0 \bar \gamma_0}^{\Delta \gamma_0 \bar \gamma_0\to \a \bar \a \gamma_0 \bar \gamma_0}"]
& \mathit{CF}(\ve{\alpha}\bar{\ve{\beta}},\ve{\Delta})
\mathrm{a.r.}[d, "f_{\a \bar \b}^{\Delta\to \a \bar \a}"]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}} \gamma_0\bar \gamma_0, \ve{\alpha} \bar{\ve{\alpha}} \gamma_0\bar \gamma_0)
\mathrm{a.r.}[r, equal]
\mathrm{a.r.}[d, "F_3^{\a \gamma_0,\a \gamma_0}"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[r, "F_3^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[d, "F_3^{\a,\a}"]
\\
\mathit{CF}(\bar{\ve{\beta}}\bar \gamma_0, \bar{\ve{\alpha}} \bar \gamma_0)
\mathrm{a.r.}[rr, "F_3^{\bar \gamma_0, \bar \gamma_0}"]
&
&
\mathit{CF}(\bar{\ve{\beta}},\bar{\ve{\alpha}})
\end{tikzcd}
\label{eq:3-handles}
\end{equation}
\subsection{Attaching 1-handles near the basepoint}
We may obtain a slightly simpler formula for the 1-handle map when we perform the attachment near the basepoint $w$.
\begin{lem}\label{lem:1-handles-simple} Suppose $\mathbb{S}_0\subset Y$ is a 0-sphere and $Y$ is connected. We have an isomorphism of groups
\[
\mathit{CFI}^-(Y(\mathbb{S}_0))\cong \mathit{CFI}^-(Y)\otimes_{\mathbb{F}} \langle \theta^+,\theta^-\rangle.
\]
With respect to this decomposition, the involutive 1-handle map takes the form
\[
\mathit{CFI}(W(\mathbb{S}_0))(\ve{x})=\ve{x}\times \theta^+,
\]
extended equivariantly over $\mathbb{F}[U,Q]/Q^2$. The 3-handle map takes the form $\ve{x}\times \theta^+\mapsto 0$ and $\ve{x}\times \theta^-\mapsto \ve{x}$, extended $\mathbb{F}[U,Q]/Q^2$-equivariantly.
\end{lem}
\begin{proof}We focus on the 1-handle map. We pick our diagram so that, before doubling, there is a special genus 1 region near the basepoint, and so the doubled arcs have the configuration shown in Figure~\ref{fig:4} in the double of this region. We form the compression of~\eqref{eq:1-handles} by first compression horizontally, then vertically.
We first show that for suitable almost complex structures, the diagonal maps make trivial contribution to compression of the diagram. We consider the top-most diagonal map. This map counts holomorphic quadrilaterals on the connected sum
\[
(\ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta}, \ve{\Delta}, \ve{\Delta})\# (\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0,\ve{\Delta}_0).
\]
Note that each Heegaard subdiagram of the right-handle quadruple has vanishing differential (since it has the minimal number of generators possible; see Figure~\ref{fig:4}). Hence the stabilization results for holomorphic quadrilaterals from Proposition~\ref{prop:multi-stabilization-counts} imply that
\[
h_{\a \bar \b \gamma_0\bar \gamma_0}^{\b \bar \b \gamma_0\bar \gamma_0\to \Delta \gamma_0\bar \gamma_0\to \Delta \Delta_0}\circ F_1^{\gamma_0 \bar \gamma_0, \gamma_0\bar \gamma_0}=F_1^{\gamma_0\bar \gamma_0, \Delta_0} \circ h_{\a\bar \b}^{\b \bar \b \to \Delta\to \Delta}.
\]
By the small translate theorem for holomorphic quadrilaterals \cite[Proposition 11.5]{HHSZExact}, the map $h_{\a\bar \b}^{\b \bar \b \to \Delta\to \Delta}$ vanishes.
We now consider the bottom-most diagonal map in~\eqref{eq:1-handles}. Because of our choice of diagram, the chain $\eta$ in~\eqref{eq:1-handle-hypercube-involving-H1} may be chosen to vanish, by the same model computation we used to verify Lemma~\ref{lem:1-handle-bottom-hypercube}. Hence, the dashed length 2 map is the sum
\begin{equation}
h^{\Delta \gamma_0\bar \gamma_0\to \Delta \Delta_0\to \a \bar \a \gamma_0\bar \gamma_0}_{\a \bar \b \gamma_0\bar \gamma_0}
+h_{\a \bar \b \gamma_0\bar \gamma_0}^{\Delta \gamma_0\bar \gamma_0\to \a \bar \a \gamma_0\bar \gamma_0\xrightarrow{\tau} \a \bar \a \gamma_0\bar \gamma_0}
\label{eq:holomorphic-quadrilaterals-1-handle-diagonal}
\end{equation}
Here, $h_{\a \bar \b \gamma_0\bar \gamma_0}^{\Delta \gamma_0\bar \gamma_0\to \a \bar \a \gamma_0\bar \gamma_0\xrightarrow{\tau} \a \bar \a \gamma_0\bar \gamma_0}$ denotes the map which counts holomorphic triangles on the diagram $(\Sigma'\# \bar \Sigma', \ve{\alpha} \bar \ve{\beta} \gamma_0, \bar \gamma_0, \ve{\Delta} \gamma_0\bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0\bar \gamma_0)$ which are weighted by an extra factor of $\d_{\a \bar \a \gamma_0\bar \gamma_0} D(\psi)\cdot \tau$.
We claim that both of these maps in~\eqref{eq:holomorphic-quadrilaterals-1-handle-diagonal} vanish when precomposed with $F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}$ and post-composed with $F_3^{\a \gamma_0,\a \gamma_0}$. Consider the first quadrilateral map. It counts curves on the diagram
\[
(\ve{\alpha} \bar \ve{\beta}, \ve{\Delta}, \ve{\Delta}, \ve{\alpha} \bar \ve{\alpha})\# (\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0, \ve{\Delta}_0, \gamma_0\bar \gamma_0 ).
\]
We can use Proposition~\ref{prop:multi-stabilization-counts} to destabilize the quadruple, noting that the evaluation of the right-hand triple on a degree 0 tree of $C_*^{\mathrm{cell}}(K_4)$ is $\theta^+|\bar \theta^-+\theta^-|\bar \theta^+$ by Lemma~\ref{lem:1-handle-bottom-hypercube}. Hence $\theta_{\gamma_0,\gamma_0}^-$ factor of the composition is
\[
h_{\a \bar \b}^{ \Delta \to \Delta\to \a \bar \a}(\ve{x})\otimes \theta_{\bar \gamma_0, \bar \gamma_0}^+.
\]
This expression vanishes by the small translate theorem for quadrilaterals \cite[Proposition 11.5]{HHSZExact}, since $\ve{\Delta}$ appears twice.
Finally, we argue that the triangle counting map $h_{\a \bar \b \gamma_0\bar \gamma_0}^{\Delta \gamma_0\bar \gamma_0\to \a \bar \a \gamma_0\bar \gamma_0\xrightarrow{\tau} \a \bar \a \gamma_0\bar \gamma_0}$ also makes trivial contribution. In this case, the triangle count occurs on
\[
(\ve{\alpha} \bar \ve{\beta}, \ve{\Delta}, \ve{\alpha} \bar \ve{\alpha})\#(\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0).
\]
Proposition~\ref{prop:multi-stabilization-counts} implies there are no triangles with $\theta_{\gamma_0,\gamma_0}^-$ as output when only terms involving $\theta_{\gamma_0,\gamma_0}^+$ are the input, so this triangle map makes trivial contribution to the compressed diagram.
The argument for 3-handles follows because the 3-handle map is defined dually to the 1-handle map, so the above argument immediately translates after taking duals.
\end{proof}
\subsection{Commutation of 1-handles}
\begin{lem}
\label{lem:commute-1/3-handles}
Suppose that $\mathbb{S}$ and $\mathbb{S}'$ are two embedded and disjoint spheres in $Y$, whose dimensions lie in $\{0,2\}$. Then
\[
\mathit{CFI}(W(Y(\mathbb{S}),\mathbb{S}'))\circ \mathit{CFI}(W(Y,\mathbb{S}))\simeq \mathit{CFI}(W(Y(\mathbb{S}'),\mathbb{S}))\circ
\mathit{CFI}(W(Y,\mathbb{S}')).
\]
\end{lem}
Lemma~\ref{lem:commute-1/3-handles} may be proven by using the model computations from Lemma~\ref{lem:1-handles-simple}. Alternatively, it may be proven by building a 3-dimensional hyperbox which has two faces corresponding to the definition of the 1-handle maps from~\eqref{eq:1-handles}.
\section{2-handles} \label{sec:2-handles}
In this section, we describe the map for a framed link $\mathbb{L}\subset Y$, with a self-conjugate $\Spin^c$ structure $\mathfrak{s}$ on $W(Y,\mathbb{L})$ (or more generally, a family $\mathfrak{S}\subset \Spin^c(W(Y,\mathbb{L}))$ which is closed under conjugation).
The existence of a map for 2-handles was described by Hendricks and Manolescu \cite{HMInvolutive}, though they were not able to prove invariance of the construction. The construction we present in this section should be thought of as a systematic version of the construction therein using the doubling operation.
\subsection{The construction}
Suppose $\mathbb{L}$ is a framed link in $Y$. We assume that the framing is Morse, so that there is a corresponding 2-handle cobordism $W(Y,\mathbb{L})$. We pick a bouquet for $\mathbb{L}$, in the sense of \cite{OSTriangles}*{Definition~4.1}, as well as a Heegaard triple $(\Sigma,\ve{\alpha}',\ve{\alpha},\ve{\beta})$ subordinate to this bouquet \cite{OSTriangles}*{Definition~4.2}.
\begin{rem}
Like the 2-handle maps from \cite{OSTriangles}, our 2-handle maps also depend on a bouquet for the link $\mathbb{L}$. Independence from the bouquet will be proven in Lemma~\ref{lem:well-defined-2handle}.
\end{rem}
We first construct several auxiliary hypercubes. We firstly have two hypercubes of alpha attaching circles
\begin{equation}
\begin{tikzcd}[column sep=1.7cm]
\ve{\alpha} \mathrm{a.r.}[r, "\Theta_{\a', \a}"] & \ve{\alpha}'
\end{tikzcd}
\quad
\text{and}
\quad
\begin{tikzcd}[column sep=1.7cm]
\ve{\alpha} \bar \ve{\beta} \mathrm{a.r.}[r, "\Theta_{\a' \bar \b, \a \bar \b}"] & \ve{\alpha}'\bar \ve{\beta}.
\end{tikzcd}
\label{eq:easy-aux-hypercubes-2-handles}
\end{equation}
and a hypercube of beta attaching circles
\begin{equation}
\begin{tikzcd}[column sep=1.7cm]
\ve{\beta}\bar \ve{\beta} \mathrm{a.r.}[r, "\Theta_{\b\bar \b, \Delta}"]& \ve{\Delta}
\end{tikzcd}
\label{eq:easy-aux-hypercubes-2-handles-beta}
\end{equation}
There is a unique choice of top degree cycle in the hypercubes in Equation~\eqref{eq:easy-aux-hypercubes-2-handles}, while the cycle $\Theta_{\b\bar \b,\Delta}$ in Equation~\eqref{eq:easy-aux-hypercubes-2-handles-beta} is only well-defined up to addition of a boundary.
Next we construct a more complicated hypercube of beta attaching curves:
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=3cm,
fill opacity=.7,
text opacity=1
]
\ve{\Delta}
\mathrm{a.r.}[r, "\Theta_{\Delta,\a'\bar\a'}"]
\mathrm{a.r.}[d, "\Theta_{\Delta, \a\bar \a}"]
\mathrm{a.r.}[dr,dashed, "\lambda_{\Delta,\a\bar\a'}"]
&
\ve{\alpha}'\bar \ve{\alpha}'
\mathrm{a.r.}[d, "\Theta_{\a'\bar\a', \a \bar \a'}"]
\\
\ve{\alpha} \bar \ve{\alpha}
\mathrm{a.r.}[r, "\Theta_{\a \bar \a,\a \bar \a'}"]
&
\ve{\alpha} \bar \ve{\alpha}',
\end{tikzcd}
\label{eq:second-auxiliary-hypercube-2-handle}
\end{equation}
We now prove that the hypercube in~\eqref{eq:second-auxiliary-hypercube-2-handle} may be constructed:
\begin{lem}\label{lem:lambda-cube} If $\Theta_{\Delta,\a'\bar \a'}$, $\Theta_{\a'\bar \a', \a \bar \a'}$, $\Theta_{\Delta,\a \bar \a}$ and $\Theta_{\a \bar \a, \a \bar \a'}$ are cycles which represent the top degree elements of their respective $\mathit{HF}^-$ groups, then
\[
\left[f_{\Delta, \a'\bar \a', \a \bar \a'}(\Theta_{\Delta, \a'\bar \a'}, \Theta_{\a'\bar \a', \a \bar \a'})\right]=\left[f_{\Delta, \a \bar \a, \a \bar \a'}(\Theta_{\Delta, \a \bar \a},\Theta_{\a \bar \a, \a \bar \a'})\right],
\]
where brackets denote the induced element in homology.
In particular, $\lambda_{\Delta,\a \bar \a'}\in \mathit{CF}(\ve{\Delta},\ve{\alpha}\bar \ve{\alpha}')$ may be chosen so that~\eqref{eq:second-auxiliary-hypercube-2-handle} is a hypercube of attaching curves.
\end{lem}
\begin{proof}
Both triples $(\Sigma\# \bar \Sigma,\ve{\Delta}, \ve{\alpha}'\bar \ve{\alpha}', \ve{\alpha} \bar \ve{\alpha}')$ and $(\Sigma\# \bar \Sigma,\ve{\Delta}, \ve{\alpha}\bar \ve{\alpha}, \ve{\alpha} \bar \ve{\alpha}')$ represent the same cobordism, topologically. Namely, they represent a cobordism from $(S^1\times S^2)^{\# g}$ to $(S^1\times S^2)^{g-|\mathbb{L}|}$ which is surgery on a $|\mathbb{L}|$ component link, each component of which is an $S^1$-fiber of a $S^1\times S^2$-summand.
\end{proof}
Since $\left[f_{\Delta, \a'\bar \a', \a \bar \a'}(\Theta_{\Delta, \a'\bar \a'}, \Theta_{\a'\bar \a', \a \bar \a'})\right]$ represents a top-degree class of $\mathit{CF}(S^1\times S^2)^{g-|\mathbb{L}|}$, we have that any two choices of $\lambda_{\Delta,\a\bar\a'}$ satisfying (\ref{eq:second-auxiliary-hypercube-2-handle}) are homologous.
Now suppose that $\mathfrak{S}\subset \Spin^c(W(Y,\mathbb{L}))$ is a finite set which is closed under conjugation. We now define the 2-handle map for $\mathbb{L}$, which we denote by $\mathit{CFI}(W(Y,\mathbb{L}),\mathfrak{S})$, using the hypercubes in~\eqref{eq:easy-aux-hypercubes-2-handles}, ~\eqref{eq:easy-aux-hypercubes-2-handles-beta} and~\eqref{eq:second-auxiliary-hypercube-2-handle}. We define the map $\mathit{CFI}(W(Y,\mathbb{L}),\mathfrak{S})$ to be the compression of the hyperbox shown in equation~\eqref{eq:hyperbox-2-handles}. A priori, $\mathit{CFI}(W(Y,\mathbb{L}), \mathfrak{S})$ depends on the choices in its construction; for well-definedness see Section \ref{subsec:independence-from-simple-handleslides}.
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=3cm,
fill opacity=.7,
text opacity=1
]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr, "f_{\a\to \a'}^{\b}"]
\mathrm{a.r.}[d, "F_1^{\bar \b, \bar \b}"]
&&
\mathit{CF}(\ve{\alpha}', \ve{\beta})
\mathrm{a.r.}[d, "F_1^{\bar \b, \bar \b}"]
\\
\mathit{CF}(\ve{\alpha}\bar\ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[rr, "f_{\a \bar \b\to \a' \bar \b}^{\b \bar \b}"]
\mathrm{a.r.}[d,"f_{\a \bar \b}^{\b \bar \b\to \Delta}"]
\mathrm{a.r.}[drr,dashed, "h_{\a\bar \b\to \a'\bar \b}^{\b\bar \b\to \Delta}"]
&&
\mathit{CF}(\ve{\alpha} '\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[d,"f_{\a' \bar \b}^{\b \bar \b\to \Delta}"]
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[rr,"f_{\a\bar \b\to \a'\bar \b}^{\Delta}"]
\mathrm{a.r.}[d,equal]
\mathrm{a.r.}[drr,dashed," h_{\a\bar\b\to \a'\bar \b}^{\Delta\to \a'\bar \a'}"]
&
&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[d,"f_{\a'\b}^{\Delta\to \a'\bar \a'}"]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[d,"f_{\a \bar{\b}}^{\Delta\to \a \bar \a}"]
\mathrm{a.r.}[dr,dashed, " H_{\a\bar \b}^{\Delta\to \a \bar\a'} "]
\mathrm{a.r.}[r, "f_{\a \bar \b}^{\Delta\to \a'\bar \a'} "]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[r, "f_{\a \bar \b\to \a'\bar \b}^{\a'\bar \a'}"]
\mathrm{a.r.}[d, "f_{\a \bar \b}^{\a' \bar \a'\to \a \bar \a'}"]
&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}, \ve{\alpha}' \bar \ve{\alpha}')
\mathrm{a.r.}[d, "F_3^{\a',\a'}"]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[r, "f_{\a \bar \b}^{\a\bar \a\to \a \bar \a'}"]
\mathrm{a.r.}[d,"F_3^{\a,\a}"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}')
\mathrm{a.r.}[r, "F_3^{\a, \a}"]
\mathrm{a.r.}[d, "F_3^{\a,\a}"]
&
\mathit{CF}( \bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[d,equal]
\\
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r,"f_{\bar \b}^{\bar \a\to \bar \a'}"]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[r,equal]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\end{tikzcd}
\label{eq:hyperbox-2-handles}
\end{equation}
\begin{lem}
If we compress each horizontal level in~\eqref{eq:hyperbox-2-handles}, then the resulting diagram is a hyperbox of chain complexes.
\end{lem}
\begin{proof} It suffices to show that each face is a hypercube of chain complexes. The subcubes with length 2 maps are all obtained by pairing hypercubes of attaching curves, so the claim is immediate for these faces. The top-most and bottom-most subcubes are hypercubes of stabilization, so the claim is immediate. The only remaining cube of interest is the second-bottom-most cube on the right. The hypercube relations amount to the relation
\begin{equation}
F_{3}^{\a',\a'}\circ f_{\a \bar \b\to \a'\bar \b}^{\a' \bar \a'}=F_3^{\a,\a}\circ f_{\a \bar \b}^{\a'\bar \a'\to \a \bar \a'}.
\label{eq:2-handle-map-relation-with-3-handles}
\end{equation}
This is proven as follows. The Heegaard triples for these triangle maps are $(\Sigma, \ve{\alpha}',\ve{\alpha},\ve{\alpha}')\#(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\beta}, \bar \ve{\alpha}')$ and $(\Sigma, \ve{\alpha}, \ve{\alpha}', \ve{\alpha})\#(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\alpha}', \bar \ve{\alpha}')$. In particular, both triples are stabilizations of a Heegaard triple for a small isotopy. Using the stabilization results of Proposition~\ref{prop:multi-stabilization-counts} and the small triangle theorem \cite{HHSZExact}*{Proposition~11.1}, we obtain ~\eqref{eq:2-handle-map-relation-with-3-handles}, completing the proof.
\end{proof}
\begin{rem}
If $\mathfrak{S}\subset \Spin^c(W(Y,\mathbb{L}))$ is a set of $\Spin^c$ structures which is closed under the conjugation action, then a cobordism map may be defined by summing the maps above for all $\mathfrak{s}\in \mathfrak{S}$. If $\mathfrak{S}$ is infinite, one may need to use coefficients in $\mathbb{F}[[U]]$ to obtain a well defined map.
\end{rem}
\subsection{Independence from simple handleslides}\label{subsec:independence-from-simple-handleslides}
We now prove that the construction above commutes with the maps for simple handleslides and changes of the doubling arcs. We phrase our result in enough generality that they also imply that the 2-handle maps are independent of the choice of bouquet for $\mathbb{L}$, and are independent from handleslides amongst the components of $\mathbb{L}$.
Suppose that $(\Sigma,\ve{\alpha}',\ve{\alpha},\ve{\beta})$ is a Heegaard triple subordinate to a bouquet of $\mathbb{L}\subset Y$. Suppose that $(\Sigma,\ve{\alpha}'_H,\ve{\alpha}_H,\ve{\beta}_H)$ is another Heegaard triple, such that $\ve{\alpha}'_H$, $\ve{\alpha}_H$ and $\ve{\beta}_H$ are obtained from $\ve{\alpha}',$ $\ve{\alpha}$ and $\ve{\beta}$ (respectively) by elementary handleslide equivalences, in the sense of Section~\ref{sec:elementary-equivalences}. We assume, furthermore that each of the following sets admits a unique top degree intersection point, which is a cycle in its respective Floer complex:
\[
\mathbb{T}_{\a_H}\cap \mathbb{T}_{\a}, \quad \mathbb{T}_{\a_H'}\cap \mathbb{T}_{\a'}, \quad \mathbb{T}_{\a_H'}\cap \mathbb{T}_{\a}, \quad \text{and} \quad \mathbb{T}_{\b}\cap \mathbb{T}_{\b_H}.
\]
We also suppose that we pick two sets of doubling arcs, $\ve{\Delta}$ and $\ve{\Delta}_H$. We require no particular relation between $\ve{\Delta}$ and $\ve{\Delta}_H$, except that the diagram $(\Sigma\# \bar \Sigma,\ve{\Delta},\ve{\Delta}_H,w)$ be admissible.
\begin{prop}\label{prop:2-handles-handle-slides} Suppose that $(\Sigma,\ve{\alpha}',\ve{\alpha},\ve{\beta})$ and $(\Sigma,\ve{\alpha}'_H,\ve{\alpha}_H,\ve{\beta}_H)$ are as above.
Let $\mathit{CFI}(W(\mathbb{L}))$ and $\mathit{CFI}(W(\mathbb{L}))^H$ denote the 2-handle maps, computed with the triples $(\Sigma,\ve{\alpha}',\ve{\alpha},\ve{\beta})$ and $(\Sigma,\ve{\alpha}'_H,\ve{\alpha}_H,\ve{\beta}_H)$, and the doubling arcs $\ve{\Delta}$ and $\ve{\Delta}_H$, respectively. Let $\mathfrak{D}$ denote the doubling data for $Y$ determined by $(\Sigma,\ve{\alpha},\ve{\beta})$, $\ve{\Delta}$, and some choices of almost-complex structures. Let $\mathfrak{D}_H$ denote the doubling data determined by $(\Sigma,\ve{\alpha}_H,\ve{\beta}_H)$, $\ve{\Delta}_H$, and some choices of almost-complex structures. Let $\mathfrak{D}'$ and $\mathfrak{D}_H'$ denote the analogous data for $Y(\mathbb{L})$. Then
\[
\Psi_{\mathfrak{D}'\to \mathfrak{D}_H'} \circ \mathit{CFI}(W(\mathbb{L}))\simeq \mathit{CFI}(W(\mathbb{L}))^H\circ \Psi_{\mathfrak{D}\to \mathfrak{D}_H}.
\]
Here, $\Psi_{\mathfrak{D}\to \mathfrak{D}_H}$ and $\Psi_{\mathfrak{D}'\to \mathfrak{D}'_H}$ denote the transition maps for elementary handleslides and changes of the doubling data, and $\simeq$ denotes $\mathbb{F}[U,Q]/Q^2$-equivariant chain homotopy.
\end{prop}
\begin{proof} Mirroring the construction of the maps $\Psi_{\mathfrak{D}\to \mathfrak{D}_H}$, we construct for each subcube of Equation~\eqref{eq:hyperbox-2-handles} a hyperbox of size $(2,1,1)$. The final coordinate corresponds to the involution. If we collapse this axis of the cube, we obtain a hyperbox of size $(2,1)$. If we compress this hyperbox, we obtain the following hypercube, which realizes the homotopy commutation in the statement.
\[
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=3cm,
fill opacity=.7,
text opacity=1
]
\mathit{CFI}(\mathfrak{D})
\mathrm{a.r.}[dr, "\Psi_{\mathfrak{D}\to \mathfrak{D}_H}"]
\mathrm{a.r.}[r,"F_{W(\mathbb{L})}"]
\mathrm{a.r.}[drr,dashed]
&
\mathit{CFI}(\mathfrak{D}')
\mathrm{a.r.}[dr, "\Psi_{\mathfrak{D}'\to \mathfrak{D}'_H}"]
&
\\
&\mathit{CFI}(\mathfrak{D}_H)
\mathrm{a.r.}[r,"F_{W(\mathbb{L})}^H"]
&
\mathit{CFI}(\mathfrak{D}'_H)
\end{tikzcd}
\]
The top-most hyperbox is obtained by extending the top-most hypercube of Equation~\eqref{eq:hyperbox-2-handles} out of the page, and is obtained by stacking two hypercubes of stabilization. We leave the details to the reader.
We now consider the hypercube in~\eqref{eq:hyperbox-2-handles} which is second from the top. The corresponding hyperbox is obtained by pairing hypercubes, as follow. We think of a hyperbox of size $(2,1,1)$ as being obtained by stacking two hypercubes. The first hypercube is obtained by constructing and pairing the following hypercubes of attaching curves, where the length-$1$ edges are obtained using the unique top-degree generators specified earlier; the $2$-faces are filled as in Lemma \ref{lem:lambda-cube}:
\[
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=2cm,
fill opacity=.7,
text opacity=1
]
\ve{\alpha} \bar \ve{\beta}
\mathrm{a.r.}[r]
\mathrm{a.r.}[drr,dashed]
\mathrm{a.r.}[dr]
&
\ve{\alpha}'\bar \ve{\beta}
\mathrm{a.r.}[dr]
&
\,
\\
&
\ve{\alpha}_H \bar \ve{\beta}_H
\mathrm{a.r.}[r]
&
\ve{\alpha}_H' \bar \ve{\beta}_H
\end{tikzcd}
\quad \text{and} \quad
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=3cm,
fill opacity=.7,
text opacity=1
]
\ve{\beta} \bar \ve{\beta}\mathrm{a.r.}[d]\\
\ve{\Delta}
\end{tikzcd}.
\]
The next hypercube is obtained by constructing and pairing the following hypercubes of attaching curves:
\[
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=2cm,
fill opacity=.7,
text opacity=1
]
\ve{\alpha}_H\bar \ve{\beta}_H \mathrm{a.r.}[r]
&
\ve{\alpha}_H'\bar \ve{\beta}_H
\end{tikzcd}
\quad
\text{and}
\quad
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=1.2cm,
fill opacity=.7,
text opacity=1
]
\ve{\beta} \bar \ve{\beta}
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddr,dashed]
&\,\\
\ve{\Delta}
\mathrm{a.r.}[dr]
& \ve{\beta}_H \bar \ve{\beta}_H
\mathrm{a.r.}[d]
\\
& \ve{\Delta}_H
\end{tikzcd}
\]
The remaining hypercubes which have a length 2 map are constructed by pairing hypercubes of attaching curves in a straightforward manner. We remark that the central-most hypercube of~\eqref{eq:hyperbox-2-handles} is obtained by a small rearrangement of the pairing of an alpha hypercube with a beta hypercube. The corresponding 3-dimensional hyperbox is obtained by performing the analogous rearrangement to two hypercubes obtained by pairing.
Finally, we consider the three bottom-most hypercubes of \eqref{eq:hyperbox-2-handles} which do not have a length-$2$ map. We consider the right hypercube in the second to bottom-most level.
\begin{equation}
\begin{tikzcd}[
column sep={1.6cm,between origins},
row sep=1cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr]
&[.7 cm]
&\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd, "F_3^{\a',\a'}"]
&[.7 cm]
\\
&[.7 cm]
|[alias=A]|\mathit{CF}(\ve{\alpha}_H \bar \ve{\beta}_H, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr]
&\,
&[.7 cm]
|[alias=B]|
\mathit{CF} (\ve{\alpha}'_H \bar \ve{\beta}_H, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[ddd, "F_3^{\a_H',\a'}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
\\
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr, "F_3^{\a,\a}"]
&[.7 cm]\,&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[dr]
\\
& [.7 cm]
|[alias=D]|
\mathit{CF}(\ve{\alpha}_H \bar \ve{\beta}_H, \ve{\alpha} \bar \ve{\alpha}')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
\mathrm{a.r.}[rr, "F_3^{\a_H,\a}"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul,dashed]
&
&[.7 cm]
|[alias=C]|
\mathit{CF}(\bar \ve{\beta}_H, \bar \ve{\alpha}')
\end{tikzcd}
\hspace{-1.2cm}
\begin{tikzcd}[
column sep={1.6cm,between origins},
row sep=1cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;
}]
|[alias=A]|
\mathit{CF}(\ve{\alpha}_H \bar \ve{\beta}_H, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr]
&[.7 cm]
&
|[alias=B]|
\mathit{CF}(\ve{\alpha}'_H \bar \ve{\beta}_H, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[rd]
\mathrm{a.r.}[ddd, "F_3^{\a_H',\a'}"]
&[.7 cm]
\\
&[.7 cm]
\mathit{CF}(\ve{\alpha}_H\bar \ve{\beta}_H, \ve{\alpha}'_H \bar \ve{\alpha}'_H)
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr]
&\,
&[.7 cm]
\mathit{CF} (\ve{\alpha}_H' \bar \ve{\beta}_H, \ve{\alpha}'_H\bar \ve{\alpha}'_H)
\mathrm{a.r.}[ddd, "F_3^{\a_H',\a_H'}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
\\
\\
|[alias=D]|
\mathit{CF}(\ve{\alpha}_H \bar \ve{\beta}_H, \ve{\alpha} \bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr, "F_3^{\a_H, \a}"]
&[.7 cm]&
|[alias=C]|
\mathit{CF}(\bar \ve{\beta}_H, \bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\\
& [.7 cm]
\mathit{CF}(\ve{\alpha}_H \bar \ve{\beta}_H, \ve{\alpha}_H \bar \ve{\alpha}_H')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
\mathrm{a.r.}[rr, "F_3^{\a_H,\a_H}"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul,dashed]
&
&[.7 cm]
\mathit{CF}(\bar \ve{\beta}_H, \bar \ve{\alpha}_H')
\end{tikzcd}
\label{eq:hyperbox-2-handle-bottom-right}
\end{equation}
In~\eqref{eq:hyperbox-2-handle-bottom-right}, all of the length 1 and 2 maps which are unlabeled are holomorphic triangle maps or quadrilateral maps. There are no length 3 maps.
The length 2 hypercube relations for the diagrams in~\eqref{eq:hyperbox-2-handle-bottom-right} are verified as follows. For the faces which have length 2 maps, the hypercube relations are obvious. For the other faces, the length 2 relations are proven similarly to the proof of~\eqref{eq:2-handle-map-relation-with-3-handles}.
It remains to prove the length 3 relations for the diagrams in ~\eqref{eq:hyperbox-2-handle-bottom-right}. To verify these, we note that each summand in the length 3 relation is a composition of a sum of quadrilateral maps followed by a 3-handle map. We claim that each summand vanishes identically. To see this, consider the length 2 map along the top face of the left cube of~\eqref{eq:hyperbox-2-handle-bottom-right}. This map is the sum of two quadrilateral maps, which take place on the diagrams $(\Sigma, \ve{\alpha}'_H,\ve{\alpha}',\ve{\alpha},\ve{\alpha}')\#(\bar \Sigma,\bar \ve{\beta}_H, \bar \ve{\beta}, \bar \ve{\beta}, \bar \ve{\alpha}')$ and $(\Sigma, \ve{\alpha}'_H,\ve{\alpha}_H,\ve{\alpha},\ve{\alpha}')\#(\bar \Sigma,\bar \ve{\beta}_H, \bar \ve{\beta}_H, \bar \ve{\beta}, \bar \ve{\alpha}') $. In particular, both diagrams are algebraically rigid and admissible multi-stabilizations. By our result on stabilizations in Proposition~\ref{prop:multi-stabilization-counts}, we see that
\[
F_3^{\a_H',\a_H}\circ h_{\a \bar \b\to \a'\bar \b\to \a_H'\bar \b_H}^{\a' \bar \a'}=h_{\bar \b \to \bar \b\to \bar \b_H}^{\bar \a'}\circ F_3^{\a, \a'},
\]
however $h_{\bar \b \to \bar \b\to \bar \b_H}^{\bar \a'}=0$ by the small translate theorem for quadrilaterals \cite{HHSZExact}*{Proposition~11.5}, since $\bar \ve{\beta}$ is repeated twice in the diagram. The same argument works for all of the other summands, verifying the hypercube relations for the diagrams in~\eqref{eq:hyperbox-2-handle-bottom-right}.
One now constructs hypercubes for the bottom two faces of~\eqref{eq:hyperbox-2-handles}. The construction of these hyperboxes is similar to the ones constructed above, and we leave the details to the reader.
There is one final hyperbox to construct, corresponding to the lowest level of Figure~\ref{def:transition-map-elementary-handleslide}. This hyperbox is constructed easily by pairing hypercubes of attaching curves and rearranging the result slightly. We leave the details to the reader.
\end{proof}
\subsection{The composition law}
In this section, we prove the composition law for cobordisms with only 2-handles.
\begin{prop}
\label{prop:composition-law}
Suppose that $\mathbb{L}_1$ and $\mathbb{L}_2$ are two framed links in $Y$, and $\mathfrak{s}_1\in \Spin^c(W(Y,\mathbb{L}_1))$ and $\mathfrak{s}_2\in \Spin^c(W(Y(\mathbb{L}_1),\mathbb{L}_2))$ are two self-conjugate $\Spin^c$ structures. Then
\[
\mathit{CFI}(W(Y(\mathbb{L}_1),\mathbb{L}_2),\mathfrak{s}_2)\circ \mathit{CFI}(W(Y,\mathbb{L}_1),\mathfrak{s}_1)\simeq \mathit{CFI}(W(Y,\mathbb{L}_1\cup \mathbb{L}_2),\mathfrak{S}(\mathfrak{s}_1,\mathfrak{s}_2))
\]
where $\mathfrak{S}(\mathfrak{s}_1,\mathfrak{s}_2)$ is the set of $\Spin^c$ structures on $W(Y,\mathbb{L}_1\cup \mathbb{L}_2)$ which restrict to $\mathfrak{s}_1$ and $\mathfrak{s}_2$.
\end{prop}
\begin{proof}
The proof of Proposition~\ref{prop:composition-law} is to construct a 3-dimensional hypercube with the following properties:
\begin{enumerate}
\item One direction corresponds to the involution.
\item One face corresponds to the cobordism map $\mathit{CFI}(W(Y,\mathbb{L}_1),\mathfrak{s}_1)$.
\item One face corresponds to the cobordism map $\mathit{CFI}(W(Y(\mathbb{L}_1),\mathbb{L}_2), \mathfrak{s}_2)$.
\item One face corresponds to the identity map from $\mathit{CFI}(Y)$ to itself.
\item One face corresponds to the map $\mathit{CFI}(W(Y,\mathbb{L}_1\cup \mathbb{L}_2), \mathfrak{S}(\mathfrak{s}_1,\mathfrak{s}_2))$.
\end{enumerate}
If we view the 3-dimensional hypercube as forming a 2-dimensional hypercube involving the $\mathit{CFI}$ complexes, then the corresponding shape is as follows:
\[
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=3cm,
fill opacity=.7,
text opacity=1
]
\mathit{CFI}(Y)
\mathrm{a.r.}[r, "F_{W(Y,\mathbb{L}_1),\mathfrak{s}_1}"]
\mathrm{a.r.}[dr,"F_{W(Y,\mathbb{L}_1\cup \mathbb{L}_2),\mathfrak{S}(\mathfrak{s}_1,\mathfrak{s}_2)}"]
\mathrm{a.r.}[drr,dashed]
&
\mathit{CFI}(Y(\mathbb{L}_1))
\mathrm{a.r.}[dr, "F_{W(Y(\mathbb{L}_1),\mathbb{L}_2),\mathfrak{s}_2}"]
&
\\
&\mathit{CFI}(Y(\mathbb{L}_1,\mathbb{L}_2))
\mathrm{a.r.}[r, "\id"]
&
\mathit{CFI}(Y(\mathbb{L}_1,\mathbb{L}_2))
\end{tikzcd}
\]
As a first step, pick a Heegaard quadruple $(\Sigma,\ve{\alpha}'',\ve{\alpha}',\ve{\alpha},\ve{\beta})$ satisfying the following:
\begin{enumerate}
\item $(\Sigma,\ve{\alpha}',\ve{\alpha},\ve{\beta})$ is subordinate to a bouquet of the link $\mathbb{L}_1\subset Y$.
\item $(\Sigma,\ve{\alpha}'',\ve{\alpha}',\ve{\beta})$ is subordinate to a bouquet of the link $\mathbb{L}_2\subset Y(\mathbb{L}_1)$.
\item $(\Sigma,\ve{\alpha}'',\ve{\alpha},\ve{\beta})$ is subordinate to a bouquet of the link $\mathbb{L}_1\cup \mathbb{L}_2\subset Y$.
\end{enumerate}
Let $\Theta_{\a',\a}$, $\Theta_{\a'',\a'}$ and $\Theta_{\a'',\a}$ be the canonical generators. As in Ozsv\'{a}th and Szab\'{o}'s original proof for $\mathit{HF}^-$, an easy model computation implies that the following diagram is a hypercube of alpha attaching curves:
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=3cm,
fill opacity=.7,
text opacity=1
]
\ve{\alpha}
\mathrm{a.r.}[r, "\Theta_{\a',\a}"]
\mathrm{a.r.}[d, "\Theta_{\a'', \a}"]
&
\ve{\alpha}'
\mathrm{a.r.}[d, "\Theta_{\a'',\a'}"]
\\
\ve{\alpha}''
\mathrm{a.r.}[r, "1"]
&
\ve{\alpha}''
\end{tikzcd}
\label{eq:hypercube-composition-law-non-inv}
\end{equation}
Pairing with the 0-dimensional hypercube $\ve{\beta}$ and restricting to the $\Spin^c$ structures in the statement gives the non-involutive version of Proposition~\ref{prop:composition-law}. Note that we can stabilize~\eqref{eq:hypercube-composition-law-non-inv} to obtain the following hypercube of attaching curves on $\Sigma\# \bar \Sigma$:
\begin{equation}
\begin{tikzcd}[
labels=description,
row sep=1cm,
column sep=3cm,
fill opacity=.7,
text opacity=1
]
\ve{\alpha}\bar \ve{\beta}
\mathrm{a.r.}[r, "\Theta_{\a'\bar \b,\a\bar \b}"]
\mathrm{a.r.}[d, "\Theta_{\a''\bar \b, \a\bar \b}"]
&
\ve{\alpha}'\bar\ve{\beta}
\mathrm{a.r.}[d, "\Theta_{\a''\bar \b,\a'\bar \b}"]
\\
\ve{\alpha}''\bar \ve{\beta}
\mathrm{a.r.}[r, "1"]
&
\ve{\alpha}''\bar \ve{\beta}
\end{tikzcd}
\label{eq:hypercube-composition-law-non-inv-stabilized}
\end{equation}
Our hyperbox for Proposition~\ref{prop:composition-law} will be of size $(1,1,6)$, which we organize into 6 hypercubes $\mathscr{C}_1,\dots, \mathscr{C}_6$, which we stack to obtain a hyperbox. The cubes $\mathscr{C}_1$ and $\mathscr{C}_2$ are shown in Figure~\ref{fig:composition-law-C1C2}.
\begin{figure}[ht!]
\[
\begin{tikzcd}[
column sep={2cm,between origins},
row sep=1.2cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;}
]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[dr, "f_{\a\to \a''}^{\b}"]
\mathrm{a.r.}[ddd, "F_1^{\bar \b, \bar \b}"]
\mathrm{a.r.}[rr, "f_{\a\to \a'}^{\b}"]
&[.7 cm]
&\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[rd, "f_{\a\to \a''}^{\b}"]
\mathrm{a.r.}[ddd,"F_1^{\bar \b, \bar \b}"]
&[.7 cm]
\\
&[.7 cm]
\mathit{CF}(\ve{\alpha}'',\ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "\id"]
&\,
&[.7 cm]
\mathit{CF} (\ve{\alpha}'', \ve{\beta})
\mathrm{a.r.}[ddd,"F_1^{\bar \b, \bar \b}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed, " h_{\a\to \a'\to \a''}^{\b}" ]
\\
\\
|[alias=A]|\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[dr, "f_{\a \bar \b\to \a'' \bar \b}^{\b \bar \b}"]
\mathrm{a.r.}[rr, "f_{\a \bar \b\to \a'\bar \b}^{\b \bar \b}"]
\mathrm{a.r.}[drrr,dashed, "h_{\a\bar \b \to \a'\bar \b\to \a''\bar \b}^{\b\bar\b}"]
&[.7 cm]\,&
|[alias=B]|\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[dr, "f_{\a'\bar \b\to\a'' \bar \b}^{\b \bar \b}"]
\\
& [.7 cm]
|[alias=D]|\mathit{CF}(\ve{\alpha}''\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"F_1^{\bar \b, \bar \b}"]
\mathrm{a.r.}[rr, "\id"]
&
&[.7 cm]
|[alias=C]|\mathit{CF}(\ve{\alpha}''\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\end{tikzcd}
\]
\[
\begin{tikzcd}[
column sep={2cm,between origins},
row sep=1.2cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;}
]
|[alias=A]|\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[dr, "f_{\a\bar\b\to \a'' \bar \b}^{\b \bar \b}"]
\mathrm{a.r.}[ddd, "f_{\a \bar \b}^{\b \bar \b\to \Delta}"]
\mathrm{a.r.}[rr, "f_{\a\bar \b\to \a'\bar \b}^{\b\bar \b}"]
\mathrm{a.r.}[dddrr,dashed]
&[.7 cm]
&|[alias=B]|\mathit{CF}(\ve{\alpha}'\bar \ve{\beta},\ve{\beta}\bar \ve{\beta})
\mathrm{a.r.}[rd, "f_{\a'\bar \b\to \a''\bar \b}^{\b\bar \b}"]
\mathrm{a.r.}[ddd,"f_{\a'\bar \b}^{\b \bar \b\to \Delta}"]
\mathrm{a.r.}[ddddr,dashed]
&[.7 cm]
\\
&[.7 cm]
|[alias=D]|\mathit{CF}(\ve{\alpha}''\bar \ve{\beta},\ve{\beta}\bar \ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "\id"]
&\,
&[.7 cm]
|[alias=C]|\mathit{CF} (\ve{\alpha}''\bar \ve{\beta}, \ve{\beta}\bar \ve{\beta})
\mathrm{a.r.}[ddd,"f_{\a''\bar \b}^{\b \bar \b\to \Delta}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed, " h_{\a\bar \b\to\a'\bar \b\to \a''\bar \b}^{\b\bar \b}" ]
\\
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[dr, "f_{\a \bar \b\to \a''\bar \b}^{\Delta}"]
\mathrm{a.r.}[rr, "f_{\a \bar \b\to \a'\bar \b}^{\Delta}"]
\mathrm{a.r.}[drrr,dashed, "h_{\a\bar \b \to \a'\bar \b\to \a''\bar \b}^{\b\bar\b}"]
&[.7 cm]\,&
\mathit{CF}(\ve{\alpha}' \bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[dr, "f_{\a'\bar \b\to\a'' \bar \b}^{\Delta}"]
\\
& [.7 cm]
\mathit{CF}(\ve{\alpha}''\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"f_{\a \bar \b}^{\b\bar \b\to \Delta}"]
\mathrm{a.r.}[rr, "\id"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul,dashed]
&
&[.7 cm]
\mathit{CF}(\ve{\alpha}''\bar \ve{\beta}, \ve{\Delta})
\end{tikzcd}
\]
\caption{The hypercubes $\mathscr{C}_1$ and $\mathscr{C}_2$. The gray faces are stacked.}
\label{fig:composition-law-C1C2}
\end{figure}
The second hypercube in Figure~\ref{fig:composition-law-C1C2} also has a length 3 map which is not drawn. In the figure, the top hypercube is a hypercube of stabilization, and the bottom hypercube is obtained by pairing the hypercube in~\eqref{eq:hypercube-composition-law-non-inv-stabilized} with the 1-dimensional hypercube
\[
\begin{tikzcd}[column sep=2cm, labels=description]
\ve{\beta} \bar \ve{\beta}\mathrm{a.r.}[r, "\Theta_{\b \bar \b, \Delta}"]& \ve{\Delta}.
\end{tikzcd}
\]
We now move to the construction of the hypercube $\mathscr{C}_3$ and $\mathscr{C}_4$, starting by constructing some auxiliary hypercubes. In Figure~\ref{fig:auxiliary-beta-hypercube-2-handle-map}, we construct the 3-dimensional analog of the hypercube of beta-attaching curves in \eqref{eq:second-auxiliary-hypercube-2-handle}. We next consider the hypercube of chain complexes obtained by pairing the cube in~\eqref{eq:hypercube-composition-law-non-inv-stabilized} with the 1-dimensional cube
\[
\begin{tikzcd}[column sep=2cm, labels=description]
\ve{\Delta}\mathrm{a.r.}[r, "\Theta_{\Delta, \a''\bar \a''}"]& \ve{\alpha}'' \bar \ve{\alpha}''.
\end{tikzcd}
\]
\begin{figure}
\begin{tikzcd}[
column sep={1.3cm,between origins},
row sep=.8cm,
labels=description,
fill opacity=.7,
text opacity=1,
]
\ve{\Delta}
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dddrr,dashed]
\mathrm{a.r.}[ddddrrr,dotted]
&[.7 cm]
& \ve{\alpha}' \bar \ve{\alpha}'
\mathrm{a.r.}[rd]
\mathrm{a.r.}[ddd]
&[.7 cm]
\\
&[.7 cm]
\ve{\alpha}'' \bar \ve{\alpha}''
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr]
&\,
&[.7 cm]
\ve{\alpha}' \bar \ve{\alpha}''
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
\\
\\
\ve{\alpha} \bar \ve{\alpha}
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr]
&[.7 cm]\,&
\ve{\alpha} \bar \ve{\alpha}'
\mathrm{a.r.}[dr]
\\
& [.7 cm]
\ve{\alpha} \bar \ve{\alpha}''
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
\mathrm{a.r.}[rr, "1"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul,dashed]
&
&[.7 cm]
\ve{\alpha} \bar \ve{\alpha}''
\end{tikzcd}
\caption{An auxiliary hypercube of $\beta$-attaching curves used in the construction of $\mathscr{C}_4$.}
\label{fig:auxiliary-beta-hypercube-2-handle-map}
\end{figure}
\noindent We reorganize the resulting cube slightly, to obtain the the diagram shown in Figure~\ref{fig:C3}. Although the diagram in the figure is not technically a hypercube, since the bottom face is subdivided, if we compress the bottom face, we obtain a hypercube of chain complexes, which we call $\mathscr{C}_3$.
\begin{figure}[ht]
\[
\begin{tikzcd}[
column sep={1.8cm,between origins},
row sep={1.5cm,between origins},
labels=description,
fill opacity=.7,
text opacity=1,
]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[ddd, equal]
\mathrm{a.r.}[rrrr, "f_{\a\bar \b\to \a'\bar \b}^{\Delta}"]
\mathrm{a.r.}[ddrr, "f_{\a\bar \b\to \a'' \bar \b}^{\Delta}"]
\mathrm{a.r.}[dddddrrrrrr,dotted]
&&&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[ddrr, "f_{\a'\bar \b \to \a''\bar \b}^{\Delta}"]
\mathrm{a.r.}[ddd,equal]
\mathrm{a.r.}[dddddrr,dashed, "h_{\a'\bar \b\to \a''\bar \b}^{\Delta\to \a''\bar \a''}" ]
&&\,
\\
\\
&&\mathit{CF}(\ve{\alpha}''\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[rrrr,equal]
&&&&
\mathit{CF}(\ve{\alpha}'' \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[from=uullllll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uullllll, sloped,"h_{\a\bar \b\to \a'\bar \b \to \a'' \bar \b}^{\Delta}",dashed]
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[dr, "f_{\a \bar \b}^{\Delta\to \a''\bar \a''}"]
\mathrm{a.r.}[rrrr, "f_{\a \bar \b\to \a'\bar \b}^{\Delta}"]
\mathrm{a.r.}[drrrrr,dashed, sloped, "h_{\a\bar \b \to \a'\bar \b}^{\Delta\to \a'' \bar \a''}", pos=.7]
&&&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[dr, "f_{\a'\bar \b}^{\Delta\to \a''\bar\a''}"]
\\
&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[rr, equal]
\mathrm{a.r.}[dr, "f_{\a\bar \b\to \a''\bar \b}^{\a''\bar \a''}"]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}''\bar \ve{\alpha}'')
\mathrm{a.r.}[rr, "f_{\a\bar \b\to \a'\bar \b}^{\a''\bar \a''}"]
\mathrm{a.r.}[dr, "f_{\a \bar \b\to \a''\bar \b}^{\a''\bar \a''}"]
\mathrm{a.r.}[drrr,dashed, "h_{\a \bar \b\to \a'\bar \b\to \a''\bar \b}^{\a''\bar \a''}"]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[dr, "f_{\a'\bar \b\to \a''\bar \b}^{\a''\bar \a''}"]
\\
\,&&
\mathit{CF}(\ve{\alpha}'' \bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[rr,equal]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"f_{\a''\bar \b}^{\Delta\to \a''\bar \a''}", pos=.6]
\mathrm{a.r.}[from=uuuuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuuull,"h_{\a\bar \b\to\a''\bar \b}^{\Delta\to \a''\bar \a''}",dashed]
&&
\mathit{CF}(\ve{\alpha}'' \bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[rr,equal]
&&
\mathit{CF}(\ve{\alpha}'' \bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"f_{\a''\bar \b}^{\Delta\to \a''\bar \a''}"]
\end{tikzcd}
\]
\caption{The diagram $\mathscr{C}_3$.}
\label{fig:C3}
\end{figure}
Next, we pair the top face of Figure~\ref{fig:auxiliary-beta-hypercube-2-handle-map} with the cube
\[
\begin{tikzcd}[labels=description, column sep=2cm]
\ve{\alpha}\bar \ve{\beta} \mathrm{a.r.}[r,"\Theta_{\a\bar \b, \a'\bar \b}"] &\ve{\alpha}' \bar \ve{\beta},
\end{tikzcd}
\]
and modify the resulting hypercube to obtain the back-most portion of the diagram shown on the top of Figure~\ref{fig:C456}. This diagram is not technically a hypercube, since some of the faces are subdivided. Nonetheless, we can form a hypercube $\mathscr{C}_4$ as follows. We view the diagram as being obtained by stacking two box-like diagrams (one on the back, and one on the front). Some of the faces of these boxes are subdivided, so we compress these faces. This gives two cube-diagrams. The hypercube relations for these cubes are straightforward to verify. For the back-most cube, they follow from the hypercube relations for the original cube described above. For the front cube, the hypercube relations are straightforward to verify directly, since all of the triangle maps involve stabilizations of Heegaard triples for small isotopies of the attaching curves. Hence, Proposition~\ref{prop:multi-stabilization-counts} may be used to destabilize the relevant counts, and then the small translate theorem for triangles \cite{HHSZExact}*{Proposition~11.1} may be used to identify the destabilized maps with nearest point maps.
We now describe the construction of the hypercube $\mathscr{C}_5$. We first pair the 0-dimensional hypercube $\ve{\alpha} \bar \ve{\beta}$ with the hypercube in Figure~\ref{fig:auxiliary-beta-hypercube-2-handle-map}. Then we rearrange, and obtain the diagram shown in the middle of Figure~\ref{fig:C456}.
There is a final hyperbox involving 3-handles, which we compress to obtain the hypercube $\mathscr{C}_6$, shown on the bottom of Figure~\ref{fig:C456}. Stacking and compressing the hypercubes $\mathscr{C}_1,\dots, \mathscr{C}_6$ proves the statement.
\end{proof}
\begin{figure}[p]
\[
\mathscr{C}_4:
\begin{tikzcd}[
column sep={1.8cm,between origins},
row sep=.5cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Number in {A,B}
{\coordinate (\Number) at (\Number.center);}
%
\filldraw[black] (A) node[rectangle,fill=white,fill opacity=.9, text opacity=1]{$\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}''\bar \ve{\alpha}'') $};
\filldraw[black] (B) node[rectangle,fill=white,fill opacity=.9, text opacity=1]{$ \mathit{CF}(\ve{\alpha}''\bar \ve{\beta},\ve{\alpha}''\bar \ve{\alpha}'')$};
}
]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[ddd, equal]
\mathrm{a.r.}[rrrr, ]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddddrrrrr,dotted]
\mathrm{a.r.}[dddrrrr,dashed]
&&&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[dr,]
\mathrm{a.r.}[ddd,]
\mathrm{a.r.}[ddddr,dashed, ]
&&\,
\\
&|[alias=A]| \phantom{\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}''\bar \ve{\alpha}'')}
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr,equal]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[rr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[dddrr,dashed]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[from=ulllll,dashed]
\\
&&|[alias=B]|
\phantom{\mathit{CF}(\ve{\alpha}''\bar \ve{\beta},\ve{\alpha}''\bar \ve{\alpha}'')}
\mathrm{a.r.}[rrrr,equal,crossing over]
&&&&
\mathit{CF}(\ve{\alpha}'' \bar \ve{\beta}, \ve{\alpha}''\bar \ve{\alpha}'')
\mathrm{a.r.}[from=ulllll,dashed,crossing over]
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr]
\mathrm{a.r.}[drrr,dashed]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[drrr,dashed]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta},\ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\\
&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[from=uuu, crossing over, equal]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}'')
\mathrm{a.r.}[rr]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=1,shorten <= 1.5cm]
\mathrm{a.r.}[from=uuu,shorten <= 1.5cm]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\alpha}' \bar \ve{\alpha}'')
\mathrm{a.r.}[dr, "F_3^{\a',\a'}"]
\\
\,&&
\mathit{CF}(\ve{\alpha}'' \bar \ve{\beta}, \ve{\alpha}'' \bar \ve{\alpha}'')
\mathrm{a.r.}[rrrr,"F_3^{\a'',\a''}"]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,equal]
&&
&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}'')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"F_3^{\a'',\a''}"]
\end{tikzcd}
\]
\[
\mathscr{C}_5:
\begin{tikzcd}[
column sep={1.8cm,between origins},
row sep=.5cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Number in {A,B}
{\coordinate (\Number) at (\Number.center);}
%
\filldraw[black] (A) node[rectangle,fill=white,fill opacity=.9, text opacity=1]{$\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}''\bar \ve{\alpha}'') $};
\filldraw[black] (B) node[rectangle,fill=white,fill opacity=.9, text opacity=1]{$\mathit{CF}(\ve{\alpha}''\bar \ve{\beta},\ve{\alpha}''\bar \ve{\alpha}'')$};
}
]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr ]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[dddrr,dashed]
\mathrm{a.r.}[ddddrrr,dotted]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr]
\mathrm{a.r.}[ddd,shift left]
\mathrm{a.r.}[ddddr,dashed]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[dr,]
\mathrm{a.r.}[ddd,"F_3^{\a',\a'}"]
&&\,
\\
&|[alias=A]|\phantom{\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}''\bar \ve{\alpha}'')}
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr,crossing over]
\mathrm{a.r.}[ddd]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}' \bar \ve{\alpha}'')
\mathrm{a.r.}[rr,crossing over]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[from=ulll,dashed,crossing over]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\alpha}' \bar \ve{\alpha}'')
\mathrm{a.r.}[dr, "F_3^{\a',\a'}"]
\mathrm{a.r.}[ddd, "F_3^{\a',\a'}"]
\mathrm{a.r.}[from=ulll,dashed,crossing over]
\\
&&
|[alias=B]|\phantom{\mathit{CF}(\ve{\alpha}''\bar \ve{\beta},\ve{\alpha}''\bar \ve{\alpha}'')}
\mathrm{a.r.}[rrrr,crossing over,"F_3^{\a'',\a''}", pos=.3]
&&&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}'')
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}\bar \ve{\alpha})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr]
\mathrm{a.r.}[drrr,dashed]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}\bar \ve{\alpha}')
\mathrm{a.r.}[rr, "F_3^{\a,\a}", pos=.7]
\mathrm{a.r.}[dr]
&&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\\
&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}'')
\mathrm{a.r.}[rr,equal]
\mathrm{a.r.}[dr, "F_3^{\a,\a}"]
\mathrm{a.r.}[from=uuu, crossing over]
\mathrm{a.r.}[from=uuuul,crossing over,dashed]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}\bar \ve{\alpha}'')
\mathrm{a.r.}[rr, "F_3^{\a,\a}"]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=1, shorten <= 2cm]
\mathrm{a.r.}[from=uuu, shorten <= 2cm]
&&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}'')
\mathrm{a.r.}[dr, equal]
\\
\,&&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}'')
\mathrm{a.r.}[rrrr,equal]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"F_3^{\a'',\a''}", pos=.85]
&&
&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}'')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,equal]
\end{tikzcd}
\]
\[
\mathscr{C}_6:
\begin{tikzcd}[
column sep={1.8cm,between origins},
row sep=.5cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Number in {A}
{\coordinate (\Number) at (\Number.center);}
%
\filldraw[black] (A) node[rectangle,fill=white,fill opacity=.9, text opacity=1]{$\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}'')$};
}]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[ddd, "F_3^{\a, \a}"]
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dr]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr, "F_3^{\a,\a}"]
\mathrm{a.r.}[ddd, "F_3^{\a,\a}",shift left]
&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[dr,]
\mathrm{a.r.}[ddd,equal]
&&\,
\\
& \mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}\bar \ve{\alpha}'')
\mathrm{a.r.}[dr, "F_3^{\a,\a}"]
\mathrm{a.r.}[rr,crossing over, equal]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}\bar \ve{\alpha}'')
\mathrm{a.r.}[rr, "F_3^{\a,\a}"]
\mathrm{a.r.}[ddd, "F_3^{\a,\a}"]
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}'')
\mathrm{a.r.}[dr,equal]
\mathrm{a.r.}[ddd,equal]
\\
&&
|[alias=A]|\phantom{\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}'')}
\mathrm{a.r.}[rrrr,crossing over,equal]
&&&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}'')
\\
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr]
\mathrm{a.r.}[drrr,dashed]
&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[rr, equal]
\mathrm{a.r.}[dr]
&&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}')
\mathrm{a.r.}[dr]
\\
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}'')
\mathrm{a.r.}[rr,equal]
\mathrm{a.r.}[dr,equal]
\mathrm{a.r.}[from=uuu,crossing over, "F_3^{\a,\a}"]
&&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}'')
\mathrm{a.r.}[rr, equal]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=1, shorten <= 1.7cm]
\mathrm{a.r.}[from=uuu,shorten <= 1.7cm]
&&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}'')
\mathrm{a.r.}[dr, equal]
\\
\,&&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha}'')
\mathrm{a.r.}[rrrr,equal]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7,shift right=.3mm]
\mathrm{a.r.}[from=uuu,equal, shift right=.3mm]
&&
&&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}'')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,equal]
\end{tikzcd}
\]
\caption{The diagrams used to construct $\mathscr{C}_4$, $\mathscr{C}_5$ and $\mathscr{C}_6$}
\label{fig:C456}
\end{figure}
\subsection{More on commutations of handles}
In this section, we prove that 1-handle and 3-handle maps may be commuted with 2-handle maps, when the corresponding topological handle manipulation is also possible:
\begin{prop}\label{prop:commute-1handles/2-handles}
Suppose that $\mathbb{L}$ is a framed 1-dimensional link in $Y$, and $\mathbb{S}$ is a 0- or 2-dimensional sphere in $Y\setminus \mathbb{L}$. Let $\mathfrak{S}\subset \Spin^c (W(Y,\mathbb{L}))$ be a set of $\Spin^c$ structures which is closed under conjugation, and let $\mathfrak{S}'\subset \Spin^c (W(Y(\mathbb{S}),\mathbb{L}))$ be the corresponding set. Then
\[
\mathit{CFI}(W(Y(\mathbb{L}),\mathbb{S}))\circ \mathit{CFI}(W(Y,\mathbb{L}),\mathfrak{S})\simeq \mathit{CFI}(W(Y(\mathbb{S}),\mathbb{L}),\mathfrak{S}')\circ \mathit{CFI}(W(Y,\mathbb{S})).
\]
\end{prop}
\begin{proof}
The proof is similar to the proof of Proposition~\ref{prop:composition-law}. We will build a 3-dimensional hypercube, four of whose faces correspond to the four maps $\mathit{CFI}(W(Y(\mathbb{L}),\mathbb{S}))$, $\mathit{CFI}(W(Y,\mathbb{L}),\mathfrak{S})$, $\mathit{CFI}(W(Y(\mathbb{S}),\mathbb{L}),\mathfrak{S}')$, and $\mathit{CFI}(W(Y,\mathbb{S}))$. The hypercube relations for the compression of this hyperbox translate exactly to the chain homotopy for the commutation described in the statement.
The hypercube realizing the commutation will be constructed as the compression of a hyperbox of size $(1,1,5)$, which we view as five hypercubes, stacked on top of each other. We write $\mathscr{B}_1,\cdots \mathscr{B}_5$ for these hypercubes.
The top-most hypercubes, $\mathscr{B}_1$ and $\mathscr{B}_2$, are shown in Figure~\ref{fig:B1B2}. They are constructed by stacking hypercubes constructed by pairing hypercubes of attaching curves, as well as hypercubes of stabilization.
\begin{figure}[H]
\[
\mathscr{B}_1:\hspace{-.5cm}
\begin{tikzcd}[
column sep={2.1cm,between origins},
row sep=1cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;}
]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[dr,"f_{\a\to \a'}^{\b}"]
\mathrm{a.r.}[ddd, "F_1^{\bar \b,\bar \b}"]
\mathrm{a.r.}[rr, "F_1^{\gamma_0,\gamma_0}"]
&[.7 cm]
&\mathit{CF}(\ve{\alpha}\gamma_0,\ve{\beta}\gamma_0)
\mathrm{a.r.}[rd, "f_{\a\gamma_0\to \a'\gamma_0}^{\b\gamma_0}"]
\mathrm{a.r.}[rr,equal]
\mathrm{a.r.}[ddd, "F_1^{\bar\b\bar \gamma_0,\bar\b\bar \gamma_0}"]
&[.7 cm]
&
\mathit{CF}(\ve{\alpha}\gamma_0,\ve{\beta} \gamma_0)
\mathrm{a.r.}[dr, "f_{\a\gamma_0\to \a'\gamma_0}^{\b\gamma_0}"]
\mathrm{a.r.}[ddd,"F_1^{\bar \b \bar \gamma_0, \bar \b \bar \gamma_0}"]
&[.7 cm]
\\
&[.7 cm]
\mathit{CF}(\ve{\alpha}',\ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr," F_1^{\gamma_0,\gamma_0}"]
&\,
&[.7 cm]
\mathit{CF}(\ve{\alpha}'\gamma_0,\ve{\beta}\gamma_0)
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, equal]
&&
\mathit{CF} (\ve{\alpha}'\gamma_0,\ve{\beta}\gamma_0)
\mathrm{a.r.}[ddd,"F_1^{\bar \b \bar \gamma_0, \bar \b \bar \gamma_0}"]
\\
\\
|[alias=A]| \mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar\ve{\beta})
\mathrm{a.r.}[dr, "f_{\a\bar \b\to \a'\bar \b}^{\b\bar \b}"]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&[.7 cm]\,&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr, "f_{\a\bar \b\gamma_0\bar \gamma_0\to \a'\bar \b \gamma_0\bar \gamma_0}^{\b \bar \b \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[rr,equal]
&[.7 cm]\,
&
|[alias=B]| \mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr,"f_{\a\bar \b\gamma_0\bar \gamma_0\to \a'\bar \b \gamma_0\bar \gamma_0}^{\b \bar \b \gamma_0\bar \gamma_0}"]
\\
& [.7 cm]
|[alias=D]| \mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"F_1^{\bar \b,\bar \b}"]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&
&[.7 cm]
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\beta} \bar \ve{\beta} \gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr,equal]
\mathrm{a.r.}[from=uuu, crossing over, "F_1^{\bar \b \bar \gamma_0, \bar \b \bar \gamma_0}"]
&&
|[alias=C]| \mathit{CF}(\ve{\alpha}'\bar \ve{\beta}\gamma_0\bar \gamma_0,\ve{\beta} \bar \ve{\beta} \gamma_0\bar \gamma_0)
\end{tikzcd}
\]
\[
\mathscr{B}_2:\hspace{-.5cm}
\begin{tikzcd}[
column sep={2.1cm,between origins},
row sep=1cm,
labels=description,
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;}
]
|[alias=A]| \mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\beta}\bar \ve{\beta})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&[.7 cm]
&\mathit{CF}(\ve{\alpha}\bar\ve{\beta} \gamma_0\bar \gamma_0,\ve{\beta}\bar \ve{\beta}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rd]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr,equal]
\mathrm{a.r.}[ddddr,dashed]
&[.7 cm]
&
|[alias=B]| \mathit{CF}(\ve{\alpha}\bar \ve{\beta}\gamma_0\bar \gamma_0,\ve{\beta}\bar \ve{\beta} \gamma_0\bar \gamma_0)
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[ddddr,dashed]
&[.7 cm]
\\
&[.7 cm]
|[alias=D]|\mathit{CF}(\ve{\alpha}'\bar \ve{\beta},\ve{\beta} \bar \ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr," F_1^{\gamma_0 \bar \gamma_0,\gamma_0 \bar \gamma_0}"]
&\,
&[.7 cm]
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}\gamma_0\bar \gamma_0,\ve{\beta}\bar \ve{\beta}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, equal]
&&
|[alias=C]| \mathit{CF}(\ve{\alpha}'\bar \ve{\beta}\gamma_0\bar \gamma_0,\ve{\beta}\bar \ve{\beta}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[ddd]
\\
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[dr,]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&[.7 cm]\,&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\Delta} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr,]
\mathrm{a.r.}[rr,]
\mathrm{a.r.}[drrr,dashed]
&[.7 cm]\,
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta} \ve{\Delta}_0)
\mathrm{a.r.}[dr]
\\
& [.7 cm]
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,]
\mathrm{a.r.}[rr,"F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul,dashed]
&
&[.7 cm]
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\Delta}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr]
\mathrm{a.r.}[from=uuu,crossing over]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}\gamma_0\bar \gamma_0,\ve{\Delta} \ve{\Delta}_0)
\end{tikzcd}
\]
\caption{The hypercubes $\mathscr{B}_1$ (top) and $\mathscr{B}_2$ (bottom) are obtained by compressing these two diagrams. The unlabeled arrows are all holomorphic triangle or quadrilateral maps. There is also a length 3 arrow in the right subcube for $\mathscr{B}_2$ which is not drawn.}
\label{fig:B1B2}
\end{figure}
The diagram used to construct $\mathscr{B}_3$ is shown in Figure~\ref{fig:B3}.
\begin{figure}[H]
\[
\begin{tikzcd}[
column sep={2.2cm,between origins},
row sep={1.5cm,between origins},
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Number in {A,B}
{\coordinate (\Number) at (\Number.center);}
%
\filldraw[black] (A) node[rectangle,fill=white, fill opacity=.9, text opacity=1]{$\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}, \ve{\Delta})$};
%
\filldraw[black] (B) node[rectangle,fill=white,fill opacity=.9, text opacity=1]{$\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\Delta} \gamma_0\bar \gamma_0)$};
%
}
]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[ddd,equal]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[ddrr]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr]
\mathrm{a.r.}[ddd,dash, shift left=2mm]
\mathrm{a.r.}[ddd,dash, shift left=1.3mm]
\mathrm{a.r.}[ddrr]
\mathrm{a.r.}[dddddrr,dashed]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta} \ve{\Delta}_0)
\mathrm{a.r.}[ddrr]
\mathrm{a.r.}[ddd,shift left=2mm,dash]
\mathrm{a.r.}[ddd,shift left=1.3mm, dash]
\mathrm{a.r.}[dddddrr,dashed]
&&\,
\\
\\
&&|[alias=A]|\phantom{\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}, \ve{\Delta})}
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&&
|[alias=B]|\phantom{\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\Delta} \gamma_0\bar \gamma_0)}
\mathrm{a.r.}[rr]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta}\Ds_0)
\mathrm{a.r.}[from=uullll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uullll,dashed]
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr,"F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta} \gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[drrr,dashed]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta} \ve{\Delta}_0)
\mathrm{a.r.}[dr]
\\
&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha}' \bar \ve{\alpha}')
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar\gamma_0,\gamma_0\bar\gamma_0}"]
\mathrm{a.r.}[dr]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha}' \bar \ve{\alpha}'\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr,"B_\tau"]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[drrr,dashed]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr]
\\
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[rr,"F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
\mathrm{a.r.}[from=uuuuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuuull,dashed]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr, "B_\tau"]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
\mathrm{a.r.}[from=uuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuull,dashed]
\end{tikzcd}
\]
\caption{The diagram used to construct the hypercube $\mathscr{B}_3$.}
\label{fig:B3}
\end{figure}
The hypercube $\mathscr{B}_4$ is displayed in Figure~\ref{fig:B4}. The construction is as follows. The back-left cube is a hypercube of stabilization for the map $F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}$. The back-right cube is constructed by pairing the 0-dimensional hypercube $\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0$ with a 3-dimensional generalized hypercube of beta attaching curves which has $\tau$ as one of the length 1 morphisms. The construction is similar to the definition of the 1-handle map, and the proof of well definedness is similar to Lemma~\ref{lem:1-handles-and-handleslides}. The front portion is verified to satisfy the hypercube relations (after compressing the top face) by observing that the length 2 map composes trivially with the 3-handle map $F_3^{\a'\gamma_0,\a'\gamma_0}$ by applying our stabilization result from Proposition~\ref{prop:multi-stabilization-counts} and applying the small translate theorems for triangles and quadrilaterals \cite[Propositions 11.1 and 11.5]{HHSZExact}.
Our final hypercube $\mathscr{B}_5$ is shown in Figure~\ref{fig:B5}. Stacking and compressing these hypercubes completes the proof.
\end{proof}
\begin{figure}[H]
\[
\begin{tikzcd}[
column sep={2.3cm,between origins},
row sep={1.8cm,between origins},
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;}
]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[dd]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[dr]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dd, shift left=1mm]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[dddr,dashed]
\mathrm{a.r.}[ddrr,dashed]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\Delta} \ve{\Delta}_0)
\mathrm{a.r.}[dr]
\mathrm{a.r.}[dd,shift left=1mm]
\mathrm{a.r.}[dddr,dashed]
&&\,
\\
&
|[alias=A]|\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr, "B_\tau"]
\mathrm{a.r.}[ddrr,dashed]
&&
|[alias=B]|\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
\mathrm{a.r.}[dd]
\\[.5cm]
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[dr]
\mathrm{a.r.}[rr,"F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[rr, "B_\tau"]
\mathrm{a.r.}[dr]
\mathrm{a.r.}[drrr,dashed]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0,\ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr]
\\
&
|[alias=D]|\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}')
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[from=uu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uu]
\mathrm{a.r.}[from=uuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuul,dashed]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}'\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr,"B_\tau"]
\mathrm{a.r.}[from=uu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uu]
&&
|[alias=C]|\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha}\bar \ve{\alpha}' \gamma_0 \bar \gamma_0)
\end{tikzcd}
\]
\[
\begin{tikzcd}[
column sep={2.3cm,between origins},
row sep={1.8cm,between origins},
fill opacity=.7,
text opacity=1,
execute at end picture={
\foreach \Nombre in {A,B,C,D}
{\coordinate (\Nombre) at (\Nombre.center);}
\fill[opacity=0.1]
(A) -- (B) -- (C) -- (D) -- cycle;},
]
|[alias=A]|\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr, "B_\tau"]
\mathrm{a.r.}[ddrr,dashed]
\mathrm{a.r.}[dd]
&&
|[alias=B]|\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[dd]
&\,
\\
&\mathit{CF}(\ve{\alpha}' \bar \ve{\beta}, \ve{\alpha}'\bar \ve{\alpha}')
\mathrm{a.r.}[rr,crossing over, "F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[from=ul]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr,crossing over,pos=.3, "B_\tau"]
\mathrm{a.r.}[from=ul]
&&
\mathit{CF}(\ve{\alpha}'\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha}'\bar \ve{\alpha}' \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
\mathrm{a.r.}[from=ul]
\\[.5cm]
|[alias=D]|\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}')
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0, \gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[dr, "F_3^{\a,\a}"]
\mathrm{a.r.}[from=uu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uu]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}'\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr,"B_\tau"]
&&
|[alias=C]|\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha}\bar \ve{\alpha}' \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr, "F_3^{\a \gamma_0,\a \gamma_0}"]
\\
\,&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[rrrr,"F_1^{\bar \gamma_0,\bar \gamma_0}"]
\mathrm{a.r.}[from=uu, "F_3^{\a',\a'}",crossing over, pos=.3]
&&
&&
\mathit{CF}(\bar \ve{\beta} \bar \gamma_0, \bar \ve{\alpha}' \bar \gamma_0)
\mathrm{a.r.}[from=uu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uu, "F_3^{\a'\gamma_0,\a'\gamma_0}"]
\end{tikzcd}
\]
\caption{The diagram used to build the hypercube $\mathscr{B}_4$. We stack the diagrams along the gray face. There is a length 3 map in the right portion of top-most diagram which is not drawn.}
\label{fig:B4}
\end{figure}
\begin{figure}[H]
\[
\begin{tikzcd}[
column sep={2.2cm,between origins},
row sep={1.1cm,between origins},
fill opacity=.7,
text opacity=1,
]
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[ddd, "F_3^{\a,\a}"]
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar \gamma_0,\gamma_0\bar \gamma_0}"]
\mathrm{a.r.}[dr]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha}\gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr, "B_\tau"]
\mathrm{a.r.}[dr]
&&
\mathit{CF}(\ve{\alpha}\bar \ve{\beta} \gamma_0 \bar \gamma_0, \ve{\alpha} \bar \ve{\alpha} \gamma_0 \bar \gamma_0)
\mathrm{a.r.}[dr]
\mathrm{a.r.}[ddd,shift left=1mm]
&&\,
\\
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha}')
\mathrm{a.r.}[rr, "F_1^{\gamma_0\bar\gamma_0,\gamma_0\bar \gamma_0}"]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha}\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[rr, "B_\tau",crossing over, pos=.3]
&&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta} \gamma_0\bar \gamma_0, \ve{\alpha}\bar \ve{\alpha}' \gamma_0\bar \gamma_0)
\mathrm{a.r.}[from=ulll,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=ulll,dashed]
\\
&&\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[rrrr, "F_1^{\bar \gamma_0, \bar \gamma_0}", pos=.45, crossing over]
\mathrm{a.r.}[from=ul, "F_3^{\a,\a}"]
&&
&&
\mathit{CF}(\bar \ve{\beta} \bar \gamma_0, \bar \ve{\alpha}' \bar \gamma_0)
\mathrm{a.r.}[from=ul, "F_3^{\a,\a}"]
\\
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[ddrr]
\mathrm{a.r.}[rrrr,"F_1^{\bar \gamma_0,\bar \gamma_0}", pos=.4]
&&
&&
\mathit{CF}(\bar \ve{\beta} \bar \gamma_0, \bar \ve{\alpha} \bar \gamma_0)
\mathrm{a.r.}[ddrr]
&&
\\
\,
\\
\,
&
\,&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[rrrr,"F_1^{\bar \gamma_0,\bar \gamma_0}"]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu, equal]
&&
&&
\mathit{CF}(\bar \ve{\beta} \bar \gamma_0, \bar \ve{\alpha}' \bar \gamma_0)
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,equal]
\end{tikzcd}
\]
\caption{The diagram used to construct the hypercube $\mathscr{B}_5$}
\label{fig:B5}
\end{figure}
\section{Naturality maps for stabilizations}\label{sec:stable}
In this section, we construct our naturality maps for stabilizations of the Heegaard surface, and prove some basic properties. Our definition of the naturality map for a stabilization is simple: we define the map for a stabilization as the composition of a 1-handle cobordism map with the 2-handle cobordism map for a 2-handle which topologically cancels the 1-handle.
The definition gives some diagrammatic flexibility for computing the stabilization map. In particular, for a stabilization we require two points $p_0$ and $p_1$ to be chosen on $\Sigma\setminus (\ve{\alpha} \cup \ve{\beta})$. We then add a tube to $\Sigma$ with feet at $p_0$ and $p_1$. We add one new alpha curve and one beta curve. The beta curve is a meridian of the tube, and the alpha curve is obtained by concatenating a longitude of the curve with an embedded curve on $\Sigma\setminus \ve{\alpha}$. If $\Sigma$ is embedded in $Y$, then we may view the construction as being determined by picking an embedded path on $\Sigma\setminus \ve{\alpha}$ which connects $p_0$ and $p_1$, and having the tube run parallel to this path. The new alpha circle runs parallel to this arc. We define the destabilization map as the composition of a 2-handle map for a 0-framed unknot, followed by the 3-handle cobordism map.
Note that this stabilization operation naturally determines a 3-ball in $Y$ which intersects $\Sigma$ in a disk. We say two stabilizations are \emph{disjoint} if there are two choices of corresponding 3-balls in $Y$ which are disjoint.
We write $\sigma$ for the stabilization map, and $\tau$ for the destabilization map. The main results of this section are summarized in the following proposition.
\begin{prop}\label{prop:stabilization-naturality}\,
\label{prop:properties-stabilization}
\begin{enumerate}
\item\label{stabilization-properties-1}The stabilization and destabilization maps commute with the transition maps for elementary equivalences of the doubling-enhanced Heegaard diagram.
\item\label{stabilization-properties-2} If $\sigma$ and $\sigma'$ are the maps for two disjoint stabilizations, then
\[
\sigma\circ \sigma'\simeq \sigma'\circ \sigma.
\]
\item\label{stabilization-properties-3} If $\sigma$ and $\tau$ are stabilizations and destabilizations for the same 3-ball, then
\[
\sigma\circ \tau\simeq \id\quad \text{and} \quad \tau\circ \sigma\simeq \id.
\]
\end{enumerate}
\end{prop}
Our proof of Parts \eqref{stabilization-properties-2} and \eqref{stabilization-properties-3} goes by way of computing basepoint-adjacent stabilizations and destabilizations (note that by handlesliding the original alpha and beta curves, such a configuration may always be achieved):
\begin{lem}
\label{lem:stabilization-near-basepoint}
Suppose that $p_1$ and $p_0$ are two points on $\Sigma\setminus (\ve{\alpha}\cup \ve{\beta})$ which both lie in the same component as the basepoint $w$. For a suitable choice of doubling curves on the stabilized diagram and suitably degenerated choices of almost complex structures, we have
\[
\sigma\simeq (\ve{x}\mapsto \ve{x}\times c)\quad \text{and} \quad \tau\simeq(\ve{x}\times c\mapsto \ve{x}),
\]
where $\{c\}=\a_0\cap \b_0$, and $(\ve{x}\mapsto \ve{x}\times c)$ denotes the $\mathbb{F}[U,Q]/Q^2$-equivariant map which sends $\ve{x}$ to $\ve{x}\times c$, and similarly for $(\ve{x}\times c\mapsto \ve{x})$.
\end{lem}
\begin{proof}Consider $\sigma$ (the proof for $\tau$ is essentially the same). By definition, $\sigma$ is the composition of the 1-handle cobordism map, followed by the map for a canceling 2-handle. By Lemma~\ref{lem:1-handles-simple}, for a suitable choice of almost complex structure, the 1-handle map takes the form $\ve{x}\mapsto \ve{x}\times \theta^+$, extended $\mathbb{F}[U,Q]/Q^2$-equivariantly. We will show that the 2-handle cobordism map is chain homotopic to a map which sends $\ve{x}\times \theta^+$ to $\ve{x}\times c$.
We recall that the 2-handle map is defined in Equation~\eqref{eq:hyperbox-2-handles}. We claim that the diagonal map of the compression is trivial. To see this, we make the following subclaims about the diagram in~\eqref{eq:hyperbox-2-handles} in the present context:
\begin{enumerate}
\item \label{model-comp-stabilization-1} The top-most length 2 map vanishes on elements of the form $\ve{x}\times \theta_{\b_0,\b_0}^+\times \Theta_{\bar \b, \bar \b}^+\times \theta_{\bar \b_0,\bar \b_0}^+$, for $\ve{x}\in \mathbb{T}_{\a}\cap \mathbb{T}_{\b}$.
\item\label{model-comp-stabilization-2} The middle length 2 map composes trivially with the 3-handle map $F_3^{\a \b_0, \a \b_0}$ (labeled $F_3^{\a',\a'}$ in Equation~\eqref{eq:hyperbox-2-handles}), and the bottom-most length 2 map composes trivially with the map $F_3^{\a \a_0, \a \a_0}$ (labeled $F_3^{\a,\a}$ in Equation~\eqref{eq:hyperbox-2-handles}).
\end{enumerate}
We consider claim~\eqref{model-comp-stabilization-1} first. There is a genus 2 portion of the Heegaard quadruple corresponding to the stabilization, which takes the following form:
\[
\mathcal{Q}_0=(\mathbb{T}^2\# \mathbb{T}^2, \a_0 \bar \b_0, \b_0 \bar \b_0, \b_0 \bar \b_0, \ve{\Delta}_0).
\]
We observe that
\[
\widehat{f}_{\a_0\bar \b_0, \b_0\bar \b_0, \Delta_0}(\widehat{f}_{\a_0\bar \b_0,\b_0\bar \b_0, \b_0\bar \b_0}(\Theta_{\a_0\bar \b_0, \b_0\bar \b_0}^+, \Theta_{\b_0\bar \b_0, \b_0\bar \b_0}^+), \Theta_{\b_0\bar \b_0, \Delta_0}^+)=\Theta_{\a_0\bar \b_0, \Delta_0}.
\]
This may be seen since the inner triangle map corresponds to the naturality map for a small perturbation of the curves $\b_0\bar \b_0$, while the outer triangle map corresponds to a factor of the doubling model of the involution for the diagram $(\mathbb{T}^2, \a_0,\b_0)$. Hence, by applying Proposition~\ref{prop:multi-stabilization-counts}, we obtain the equality
\[
h_{\a \b_0 \bar \b \bar \b_0\to \a \a_0 \bar \b \bar \b_0}^{\b \b_0 \bar \b \bar \b_0\to \Delta\Dt_0}(\ve{x}\times \theta_{\b_0,\b_0}^+\times \Theta_{\bar \b, \bar \b}^+\times \theta_{\bar \b_0,\bar \b_0}^+)=h_{\a \bar \b \to \a \bar \b}^{\b \bar \b \to \Delta}(\ve{x}\times \theta_{\b_0,\b_0}^+)\otimes \Theta_{\a_0\bar \b_0, \Delta_0}.
\]
which vanishes by the small translate theorem for holomorphic quadrilaterals \cite[Proposition 11.5]{HHSZExact}.
Next, we consider claim~\eqref{model-comp-stabilization-2}, from above. There are two quadrilateral maps which appear in this claim. The proof for both maps is essentially the same, so we focus on the map
\[
h_{\a \b_0\bar \b \bar \b_0\to \a \a_0 \bar \b\bar \b_0}^{\Delta\Dt_0\to \a \a_0\bar \a \bar \a_0}.
\]
This is the middle length 2 map in~\eqref{eq:hyperbox-2-handles}. We wish to show that this map composes trivially with $F_3^{\a',\a'}$. To see this, we consider the genus 2 Heegaard quadruple
\[
\mathcal{Q}_1= (\mathbb{T}^2\# \mathbb{T}^2, \a_0 \bar \a_0, \Delta_0, \b_0\bar \b_0, \a_0\bar \b_0).
\]
This is an admissible, algebraically rigid multi-stabilizing quadruple. Furthermore, we claim that
\begin{equation}
\widehat{f}_{\a_0\bar \a_0, \b_0\bar \b_0, \a_0\bar \b_0}(\widehat{f}_{\a_0\bar \a_0,\Delta_0, \b_0\bar \b_0}(\Theta^+_{\a_0\bar \a_0, \b_0\bar \b_0},\Theta^+_{\Delta_0, \b_0\bar \b_0}),\Theta^+_{\b_0\bar \b_0, \a_0\bar \b_0})=\Theta^-_{\a_0\bar \a_0, \a_0\bar \b_0}.
\label{eq:model-computation-stabilization-2-handle}
\end{equation}
The above equation is verified by noting that the inner triangle map may be interpreted as the cobordism map for a 2-handle attachment along a fiber of $S^1\times S^2$, while the outer triangle map may be interpreted as a 2-handle attached along a 0-framed unknot. The fact that $\Theta^-_{\a_0\bar \a_0, \a_0\bar \b_0}$ is the output may be viewed as a consequence of the above topological description.
Combining~\eqref{eq:model-computation-stabilization-2-handle} with Proposition~\ref{prop:multi-stabilization-counts}, we conclude that
\begin{equation}
h_{\a \b_0\bar \b \bar \b_0\to \a \a_0 \bar \b\bar \b_0}^{\Delta\Dt_0\to \a \a_0\bar \a \bar \a_0}(\ve{x} \times \Theta^+_{\b_0\bar \b_0,\Delta})= h_{\a \bar \b \to \a \bar \b}^{\Delta \to \a \bar \a}(\ve{x})\times \Theta^-_{\a_0\bar \b_0, \a_0 \bar \a_0}+\sum_{\ve{z}\in \mathbb{T}_{\a\bar \b}\cap \mathbb{T}_{\a \bar \a}} C_{\ve{z}}\cdot \ve{z}\times \Theta_{\a_0\bar \b_0,\a_0\bar \a_0}^+,
\label{eq:destabilize-quadrilatal-model-comp-stabilization}
\end{equation}
for some $C_{\ve{z}}\in \mathbb{F}[U]$. Equation ~\eqref{eq:destabilize-quadrilatal-model-comp-stabilization} vanishes once composed with $F_{3}^{\a \a_0,\a \a_0}$, by the small translate theorem for holomorphic quadrilaterals \cite[Proposition 11.5]{HHSZExact}.
The composition of the bottom-most length 2 map from Equation~\eqref{eq:hyperbox-2-handles} with $F_3^{\a \b_0,\a\b_0}$ vanishes in our present setting by a nearly identical argument.
\end{proof}
We can now prove Proposition~\ref{prop:properties-stabilization}:
\begin{proof}[Proof of Proposition~\ref{prop:properties-stabilization}] Claim~\eqref{stabilization-properties-1} follows immediately from the fact that 1-handle, 2-handle and 3-handle cobordism maps commute (up to chain homotopy) with the transition maps for elementary handleslides.
Claims~\eqref{stabilization-properties-2} and~\eqref{stabilization-properties-3} follow from Lemma~\ref{lem:stabilization-near-basepoint}, since by~\eqref{stabilization-properties-1}, we may perform stabilizations near the basepoint, and then observe that the formulas given therein clearly satisfy the stated relations.
\end{proof}
\section{Handleswaps} \label{sec:handleswap}
In this section, we prove handleswap invariance.
\subsection{Simple handleswaps}
Of central importance to the approach to naturality in \cite{JTNaturality} are the following loops in the graph of Heegaard diagrams for a 3-manifold $Y$:
\begin{define} \label{def:simple-handleswap}
Suppose $Y$ is a 3-manifold. A \emph{simple handleswap loop} is a triangle of Heegaard diagrams
\[
\begin{tikzcd}
&\mathcal{H}_1\arrow[rd,"e"] & \\
\mathcal{H}_3 \arrow[ur,"g"] && \mathcal{H}_2 \arrow[ll,"f"]
\end{tikzcd}
\]
which satisfies the following:
\begin{enumerate}
\item $\mathcal{H}_i $ is an embedded Heegaard diagram for $Y$, which has the same Heegaard surface for each $i\in \{1,2,3\}$.
\item $e$ is an $\a$-equivalence, $f$ is a $\b$-equivalence, and $g$
is a diffeomorphism.
\item The diagrams $\mathcal{H}_i$ decompose as connected sums of a fixed Heegaard diagram $\mathcal{H}=(\Sigma,\ve{\alpha}_u,\ve{\beta}_u)$ with genus two diagrams $\mathcal{G}_i=(\Sigma_0,\ve{\alpha}_i,\ve{\beta}_i)$. Furthermore, in the punctured genus two surface $(\Sigma\# \Sigma_0)\setminus \Sigma$, the above triangle is diffeomorphic to the triangle in Figure~\ref{fig:handlewap-elem}.
\end{enumerate}
\end{define}
In particular, in the above definition, $\ve{\alpha}_1$ consists of two closed curves $\a_1$ and $\a_2$, while $\ve{\beta}_1$ consists of two closed curves $\b_1$ and $\b_2$. The arrow $e$ corresponds to handlesliding $\a_1$ over $\a_2$, while the arrow $f$ corresponds to handlesliding $\b_1$ over $\b_2$.
\begin{figure}[H]
\centering
\input{handleswap-elem.pdf_tex}
\caption{A simple handleswap.}\label{fig:handlewap-elem}
\end{figure}
Recall that associated to handleslides and diffeomorphisms, we have defined morphisms, $\mathit{CFI}(e)$, $\mathit{CFI}(f)$ and $\mathit{CFI}(g)$, which are well-defined up to $\mathbb{F}[U,Q]/Q^2$-equivariant homotopy equivalence.
\begin{thm}
\label{thm:handleswap}
$
\mathit{CFI}(g)\circ \mathit{CFI}(f)\circ \mathit{CFI}(e)\simeq \Id_{\mathit{CFI}(\mathcal{H}_1)}.$
\end{thm}
The proof of Theorem~\ref{thm:handleswap} occupies the subsequent sections.
\subsection{Reduction to basepoint-adjacent handleswaps}
A simple handleswap need not occur in a region of the Heegaard diagram near a basepoint. However we now show that for the purposes of naturality, it is sufficient to consider handleswaps where the special genus 2 region is adjacent to the basepoint. We note that this fact was also observed by Sungkyung Kang.
\begin{lem}
Suppose that $F$ is a weak Heegaard invariant (in the sense of \cite{JTNaturality}*{Definition~2.24}) which satisfies all of the conditions of a strong Heegaard invariant (\cite{JTNaturality}*{Definition~3.23}) except possibly handleswap invariance. If $F$ has no monodromy around basepoint-adjacent handleswaps, then it has no monodromy around arbitrary handleswaps.
\end{lem}
\begin{proof}
Let $p\in \Sigma$ denote the connected sum point. Pick an arc $\lambda$ which connects $p$ to the basepoint $w$. We proceed by induction on the number of intersections of $(\ve{\alpha}_u\cup \ve{\beta}_u) \cap \lambda.$ The base case is the hypothesis. If $\mathcal{H}$ is arbitrary, there is an equivalence $j$ between $\mathcal{H}$ and a diagram $\mathcal{H}'$ (obtained by handlesliding an alpha or beta curve over $p$) such that on $\mathcal{H}'$, there is one fewer intersection of $(\ve{\alpha}_u \cup \ve{\beta}_u)\cap \lambda$. If $\mathcal{H}'_1,$ $\mathcal{H}_2'$ and $\mathcal{H}_3'$ denote the diagrams obtained by connect summing the standard genus 2 diagrams from Figure~\ref{fig:handlewap-elem}, there is an analogous simple handleswap loop involving $\mathcal{H}_1'$, $\mathcal{H}_2'$ and $\mathcal{H}_3'$. Write $e'$, $f'$ and $g'$ for the analogous equivalences. We have a diagram of equivalences
\[
\begin{tikzcd}[labels=description, column sep=.7cm, row sep=.9cm]
&&&\mathcal{H}_1
\mathrm{a.r.}[dddrrr, "e"]
\mathrm{a.r.}[d, "j_1"]
&&&\\
&&&\mathcal{H}_1'
\mathrm{a.r.}[dr, "e'"]
\\
&&\mathcal{H}_3'
\mathrm{a.r.}[ur, "g'"]
\mathrm{a.r.}[from=dll, "j_3"]
&& \mathcal{H}_2'
\mathrm{a.r.}[ll, "f'"]
\mathrm{a.r.}[drr, "j_2"]
\\[-.3cm]
\mathcal{H}_3
\mathrm{a.r.}[uuurrr, "g"]
&&&&&&
\mathcal{H}_2
\mathrm{a.r.}[llllll, "f"]
\end{tikzcd}
\]
The three outer rectanges are obtained from distinguished squares by inverting some of the arrows, which are isomorphsims. By hypothesis, $F$ commutes around these squares. Our inductive hypothesis is that $F$ has no monodromy around the central triangle. This clearly implies $F$ has no monodromy around the outer triangle, completing the proof.
\end{proof}
\subsection{Proof of Theorem~~\ref{thm:handleswap}}
Let us introduce some notation. For the curves in the handleswap region, we write $\ve{\alpha}_0$, $\ve{\alpha}_0^H$, $\ve{\beta}_0$ and $\ve{\beta}_0^H$. We write $\ve{\alpha}^H$ for $\ve{\alpha}_u\cup \ve{\alpha}_0^H$, and similarly for $\ve{\beta}_0^H$. Hence, for the three diagrams appearing in the handleswap loop, we write
\[
\mathcal{H}_1=(\Sigma\# \Sigma_0,\ve{\alpha},\ve{\beta}), \quad \mathcal{H}_2=(\Sigma\# \Sigma_0, \ve{\alpha}^H, \ve{\beta}),\quad \text{and} \quad \mathcal{H}_3=(\Sigma\# \Sigma_0, \ve{\alpha}^H, \ve{\beta}^H).
\]
Similarly, let $\ve{\Delta}_u$ be a collection of doubling curves on $\Sigma\# \bar \Sigma$. Let $\ve{\Delta}_0$ be choice of doubling curves on $\Sigma_0\# \bar \Sigma_0$, as shown in Figure~\ref{fig:involutive-handleswap}. Write $\ve{\Delta}$ for $\ve{\Delta}_u\cup \ve{\Delta}_0$.
We write $\mathfrak{D}$ for $\mathcal{H}$, decorated with $\ve{\Delta}_u$. We pick a collection of four doubling curves $\ve{\Delta}_0$ on $\Sigma_0\# \bar \Sigma_0$. We decorate $\mathcal{H}_1,$ $\mathcal{H}_2$, $\mathcal{H}_3$ with $\ve{\Delta}$ to form the doubling enhanced diagrams $\mathfrak{D}_1$, $\mathfrak{D}_2$, and $\mathfrak{D}_3$, respectively.
\begin{figure}[ht]
\centering
\input{handleswap-colortest.pdf_tex}
\caption{The double of the handleswap region. The left side is $\Sigma$, and the right is $\bar \Sigma$.}
\label{fig:involutive-handleswap}
\end{figure}
We note that for $i\in \{1,2,3\}$, there are canonical isomorphisms
\begin{equation}
\mathit{CFI}(\mathfrak{D})\cong \mathit{CFI}(\mathfrak{D}_i), \label{eq:canonical-iso}
\end{equation}
given by $\ve{x}\mapsto \ve{x}\times \ve{c}_i,$
extended $\mathbb{F}[U,Q]/Q^2$-equivariantly. Here, we write $\ve{c}_i$ for the unique intersection points
\[
\ve{c}_1\in \mathbb{T}_{\a_0}\cap \mathbb{T}_{\b_0}, \quad \ve{c}_2\in \mathbb{T}_{\a_0^H}\cap \mathbb{T}_{\b_0}, \quad \text{and} \quad \ve{c}_3\in \mathbb{T}_{\a_0^H}\cap \mathbb{T}_{\b_0^H}.
\]
In fact, we have the following:
\begin{lem}
\label{lem:admissibility-handleswap-1}
The genus 2 diagram shown in Figure~\ref{fig:involutive-handleswap} is weakly admissible. Furthermore, each of the following subdiagrams has vanishing differential.
\begin{enumerate}
\item $(\Sigma_0\# \bar \Sigma_0,\ve{\beta}_0\bar \ve{\beta}_0, \ve{\Delta}_0,w)$
\item $(\Sigma_0\# \bar \Sigma_0,\ve{\beta}^H_0\bar \ve{\beta}^H_0, \ve{\Delta}_0,w)$.
\item $(\Sigma_0\# \bar \Sigma_0,\ve{\Delta}_0, \ve{\alpha}_0 \bar \ve{\alpha}_0,w)$.
\item $(\Sigma_0\# \bar \Sigma_0,\ve{\Delta}_0, \ve{\alpha}^H_0\bar \ve{\alpha}^H_0,w)$.
\item $(\Sigma_0\# \bar \Sigma_0,\ve{\alpha}_0 \bar \ve{\beta}_0, \ve{\Delta}_0,w)$.
\item $(\Sigma_0\# \bar \Sigma_0,\ve{\alpha}^H_0 \bar \ve{\beta}^H_0, \ve{\Delta}_0,w)$.
\end{enumerate}
\end{lem}
\begin{proof} Admissibility is a straightforward computation which we leave to the reader. To check that each of the stated complexes has vanishing differential, it suffices to check that each has the same number of generators as its homology. The first four subdiagrams represent a complex for $\#^2 S^1\times S^2$ and also have exactly four generators. Similarly the final two subdiagrams represent $S^3$, and each have a unique intersection point.
\end{proof}
By definition,
\[
\mathit{CFI}(e)=\Psi_{\mathfrak{D}_1\to \mathfrak{D}_2}\quad \text{and} \quad \mathit{CFI}(f)=\Psi_{\mathfrak{D}_2\to \mathfrak{D}_3},
\]
where $\Psi_{\mathfrak{D}_i\to \mathfrak{D}_{i+1}}$ is obtained by compressing a hyperbox as in Figure~\ref{def:transition-map-elementary-handleslide}. Computing the composition $\Psi_{\mathfrak{D}_2\to \mathfrak{D}_3}\circ \Psi_{\mathfrak{D}_1\to \mathfrak{D}_2}$ is somewhat inefficient, however, since the involutive transition maps naturally allow for changing both the alpha and the beta curves simultaneously. Instead, we will consider the map $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$ obtained by by compressing the hyperbox in Figure~\ref{def:transition-map-elementary-handleslide}. (See Equation~\eqref{eq:Psi_1->3}, below). By Theorem~\ref{thm:naturality-fixed-Sigma},
\begin{equation}
\mathit{CFI}(f)\circ \mathit{CFI}(e)=\Psi_{\mathfrak{D}_2\to \mathfrak{D}_3}\circ \Psi_{\mathfrak{D}_1\to \mathfrak{D}_2}\simeq \Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}.
\label{eq:f-e-composed}
\end{equation}
To prove Theorem~\ref{thm:handleswap}, we also need to understand the diffeomorphism map $\mathit{CFI}(g)$. We note that the diffeomorphism $g$ does not fix $\ve{\Delta}_0$. Hence, by definition, this map is the composition
\begin{equation}
\mathit{CFI}(g):=\Psi_{g(\mathfrak{D}_3)\to \mathfrak{D}_1}\circ T_g,
\label{eq:CFI-g-def}
\end{equation}
which we explain presently. The map $T_g$ is the tautological map which sends an intersection point $\ve{x}\in \mathbb{T}_{\a \a_0^H}\cap \mathbb{T}_{\b \b_0^H}$ to its image in $\mathbb{T}_{\a \a_0}\cap \mathbb{T}_{\b \b_0}$ under $g$, extended $\mathbb{F}[U,Q]/Q^2$-equivariantly. Also, $g(\mathfrak{D}_3)$ is obtained by enhancing the Heegaard diagram $\mathcal{H}_1$ with the doubling curves $\ve{\Delta}_u\cup \ve{\Delta}_0'$, where $\ve{\Delta}_0'=g(\ve{\Delta}_0)$. The map $\Psi_{g(\mathfrak{D}_3)\to \mathfrak{D}_1}$ is built by compressing a hyperbox similar to the one in Figure~\ref{def:transition-map-elementary-handleslide} which realizes the change of doubling curves.
With respect to the canonical isomorphisms of~\eqref{eq:canonical-iso}, the map $T_g$ is clearly equal to the identity map. Theorem~\ref{thm:handleswap} follows from ~\eqref{eq:f-e-composed} and~\eqref{eq:CFI-g-def} and the following result.
\begin{prop}\label{prop:handleswap-factors=identity}
With respect to the canonical isomorphisms in~\eqref{eq:canonical-iso}, the maps $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$ and $\Psi_{g(\mathfrak{D}_3)\to \mathfrak{D}_1}$ are chain homotopic to the identity map.
\end{prop}
\begin{proof}
We focus first on the computation of $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$. By definition, this map is obtained by compressing the following hyperbox:
\begin{equation}
\begin{tikzcd}[labels=description, row sep=1cm]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{\bar{\b},\bar{\b}}"]
& \mathit{CF}(\ve{\alpha}^H,\ve{\beta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{\bar{\b}^H,\bar{\b}}"]
& \mathit{CF}(\ve{\alpha}^H,\ve{\beta}^H)
\mathrm{a.r.}[d,"F_{1}^{\bar{\b}^H,\bar{\b}^H}"]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}^H\bar{\ve{\beta}}^H, \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}^H \bar{\ve{\beta}}^H, \ve{\beta}^H \bar{\ve{\beta}}^H)
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}^H\bar \ve{\beta}^H,\ve{\Delta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}^H\bar \ve{\beta}^H,\ve{\Delta})
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a,\a }"]
&
\mathit{CF}(\ve{\alpha}^H \bar \ve{\beta}^H, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a^H,\a }"]
&
\mathit{CF}(\ve{\alpha}^H \bar{\ve{\beta}}^H, \ve{\alpha}^H \bar{\ve{\alpha}}^H)
\mathrm{a.r.}[d, "F_3^{\a^H,\a^H}"]
\\
\mathit{CF}(\bar{\ve{\beta}}, \bar{\ve{\alpha}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[drr,dashed]
\mathrm{a.r.}[d]
& \mathit{CF}(\bar{\ve{\beta}}^H,\bar{\ve{\alpha}})
\mathrm{a.r.}[r]
& \mathit{CF}(\bar{\ve{\beta}}^H,\bar{\ve{\alpha}}^H)
\mathrm{a.r.}[d]
\\
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha})
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha}^H)
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta}^H,\bar \ve{\alpha}^H)
\end{tikzcd}
\label{eq:Psi_1->3}
\end{equation}
We note that in the above diagram, each polygon counting map occurs on a connected sum. One summand is $\Sigma$ or $\Sigma\# \bar \Sigma$ while the other summand is the genus 2 subsurface $\Sigma_0$, or its double. We will argue by viewing the diagram as a multi-stabilization and applying Lemma~\ref{lem:admissibility-handleswap-1} and Proposition~\ref{prop:multi-stabilization-counts}.
There is a subtlety in that the almost complex structures used to construct the transition map Equation~\eqref{eq:Psi_1->3} are not naturally adapted to the genus four stabilization region where the handleswap occurs. Instead, we must first argue that we can change the above composition to one which is naturally adapted to the handleswap. To do so, we consider the closed curves on $\Sigma\# \bar \Sigma$ shown in Figure~\ref{fig:12}. We label these as $c_+,$ $c_-$, $c_{\delta}$ and $c_0$. These are as follows:
\begin{enumerate}
\item $c_0$ is a meridian of the connected sum tube used in the doubling operation.
\item $c_+$ is a closed curve on $\Sigma$ which bounds the handleswap region and the connected sum point.
\item $c_-$ is a symmetric closed curve on $\bar \Sigma$, and bounds the handleswap region as well as a the connected sum point.
\item $c_\delta$ is a closed curve on $\Sigma\# \bar \Sigma$ which bounds the union of both handleswap regions. Also, $c_{\delta}$ is disjoint from all $\ve{\Delta}$ curves.
\end{enumerate}
\begin{figure}[ht]
\centering
\input{fig12.pdf_tex}
\caption{Curves along which we degenerate the almost complex structure.}
\label{fig:12}
\end{figure}
Write $J_{c_0}$ for an almost complex structure or family of almost complex structures which is nodal along $c_0$, and use similar notation to define almost complex structures degenerated on another curve or collections of pairwise disjoint curves on $\Sigma\# \bar \Sigma$.
We introduce some shorthand. Write $\mathit{CF}(L_1),\dots, \mathit{CF}(L_6)$ for the six horizontal levels of Equation~\eqref{eq:Psi_1->3}. (These are themselves 1-dimensional hyperboxes of chain complexes). By using change of almost complex structure hypercubes, similar to \cite{HHSZExact}*{Section~13}, we may assume that the complexes and maps in Equation~\eqref{eq:Psi_1->3} are computed with respect to almost complex structures which are \emph{maximally pinched} along $c_0$, $c_+$ and $c_-$, in the sense of \cite{HHSZExact}*{Section~8.3}. This gives the map $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$, as defined in the earlier sections of the present paper. On the other hand, using similar techniques, we may use change of almost complex structure hypercubes to relate this model to one which is maximally pinched along $c_+$, $c_-$ and $c_\delta$. This gives a hyperbox which we schematically indicate by the following diagram:
\begin{equation}
\begin{tikzcd}[labels=description, row sep=.8cm, column sep=4cm]
\mathit{CF}_{J_{c_+}}(L_1)
\mathrm{a.r.}[r, "\id"]
\mathrm{a.r.}[d, "F_{1}"]
& \mathit{CF}_{J_{c_+}}(L_1)
\mathrm{a.r.}[d, "F_{1}"]
\\
\mathit{CF}_{J_{c_+,c_-,c_0}}(L_2)
\mathrm{a.r.}[r,"\Psi_{J_{c_+,c_-,c_0}\to J_{c_+,c_-,c_\delta}}"]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}_{J_{c_+,c_-,c_\delta}}(L_2)
\mathrm{a.r.}[d]
\\
\mathit{CF}_{J}(L_3)
\mathrm{a.r.}[r, "\Psi_{J\to J_{c_\delta}}"]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}_{J_{c_\delta}}(L_3)
\mathrm{a.r.}[d]
\\
\mathit{CF}_{J_{c_+,c_-,c_0}}(L_4)
\mathrm{a.r.}[r, "\Psi_{J_{c_+,c_-,c_0}\to J_{c_+,c_-,c_\delta}}"]
\mathrm{a.r.}[d, "F_3"]
&
\mathit{CF}_{J_{c_+,c_-,c_\delta}}(L_4)
\mathrm{a.r.}[d, "F_3"]
\\
\mathit{CF}_{J_{c_-}}(L_5)
\mathrm{a.r.}[r,"\id"]
\mathrm{a.r.}[dr,dashed]
\mathrm{a.r.}[d]
& \mathit{CF}_{J_{c_-}}(L_5)
\mathrm{a.r.}[d]
\\
\mathit{CF}_{J_{c_-}}(L_6)
\mathrm{a.r.}[r, "\id"]
&
\mathit{CF}_{J_{c_-}}(L_6)
\end{tikzcd}
\label{eq:Psi_1->3-cx-structures}
\end{equation}
The hyperbox in Equation~\eqref{eq:Psi_1->3-cx-structures} realizes a chain homotopy between the model of $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$, computed with respect to almost complex structures which are maximally pinched along $c_+,$ $c_-$ and $c_0$, and a model of $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$ which is computed using almost complex structures which are maximally pinched along $c_+,$ $c_-$ and $c_\delta$.
In the model of $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$ appearing along the right hand side of Equation~\eqref{eq:Psi_1->3-cx-structures}, the holomorphic polygon counts all occur on a multi-stabilization of the diagram on $\Sigma$ or $\Sigma\# \bar \Sigma$. The stabilization results for triangles imply that the two length 1 maps along the top of ~\eqref{eq:Psi_1->3} compose to give a map which is intertwined with the identity under the canonical isomorphisms $\mathit{CF}(\mathcal{H})\cong \mathit{CF}(\mathcal{H}_i)$ (similar to~\eqref{eq:canonical-iso}). (This is the content of \cite{JTNaturality}*{Propositions~9.31, 9.32}).
It remains to show that none of the diagonal maps make non-trivial contribution to the compression of~\eqref{eq:Psi_1->3}. To this end, we make the following claims, which together imply that the compression has trivial diagonal map:
\begin{enumerate}
\item\label{compression-claim-1} The compression of the top two levels of~\eqref{eq:Psi_1->3} has trivial diagonal map.
\item\label{compression-claim-2} The compression of the third and fourth levels of~\eqref{eq:Psi_1->3} has trivial diagonal map.
\item\label{compression-claim-3} The bottom-most level of ~\eqref{eq:Psi_1->3} has trivial diagonal map.
\end{enumerate}
Consider first the claim~\eqref{compression-claim-3}. The diagonal map is a quadrilateral counting map on the connected sum $(\Sigma,\ve{\alpha}_u,\ve{\alpha}_u,\ve{\beta}_u,\ve{\beta}_u)\#(\Sigma_0,\ve{\alpha}_0^H,\ve{\alpha}_0,\ve{\beta}_0,\ve{\beta}_0^H)$. Applying the stabilization result from Proposition~\ref{prop:multi-stabilization-counts} to the setting of holomorphic quadrilaterals, we may destabilize the map to the diagram $(\Sigma,\ve{\alpha}_u,\ve{\alpha}_u,\ve{\beta}_u,\ve{\beta}_u)$. By the small translate theorem for quadrilaterals \cite[Proposition 11.5]{HHSZExact}, this map vanishes.
Next, consider~\eqref{compression-claim-1}. For this claim, we observe that the stabilization results for holomorphic triangles imply that the length 1 holomorphic triangle maps preserve the subspaces spanned by the top-degree generator (and furthermore, act tensorially on these subspaces). Furthermore, the same argument as for claim~\eqref{compression-claim-3} implies that the holomorphic quadrilateral maps vanish on elements which have a factor which is the top degree generator. Since the top degree generator is the output of the 1-handle maps on the top row, this implies that the compression has trivial diagonal map. Claim~\eqref{compression-claim-2} follows from an analogous argument. This completes the proof of the claim about $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$.
We now consider the claim about $\Psi_{g(\mathfrak{D}_3)\to \mathfrak{D}_1}$. This map may be computed by compressing the following diagram
\begin{equation}
\begin{tikzcd}[labels=description, row sep=1cm]
\mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{\bar{\b},\bar{\b}}"]
& \mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_{1}^{\bar{\b},\bar{\b}}"]
& \mathit{CF}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d,"F_{1}^{\bar{\b},\bar{\b}}"]
\\
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}\bar{\ve{\beta}}, \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\beta} \bar{\ve{\beta}})
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha} \bar \ve{\beta},g(\ve{\Delta}))
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}\bar \ve{\beta},g(\ve{\Delta}))
\mathrm{a.r.}[r]
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed]
& \mathit{CF}(\ve{\alpha}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[d]
\\
\mathit{CF}(\ve{\alpha}\bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a,\a }"]
&
\mathit{CF}(\ve{\alpha} \bar \ve{\beta}, \ve{\alpha} \bar \ve{\alpha})
\mathrm{a.r.}[r]
\mathrm{a.r.}[d, "F_3^{\a,\a }"]
&
\mathit{CF}(\ve{\alpha} \bar{\ve{\beta}}, \ve{\alpha} \bar{\ve{\alpha}})
\mathrm{a.r.}[d, "F_3^{\a,\a}"]
\\
\mathit{CF}(\bar{\ve{\beta}}, \bar{\ve{\alpha}})
\mathrm{a.r.}[r]
\mathrm{a.r.}[drr,dashed]
\mathrm{a.r.}[d]
& \mathit{CF}(\bar{\ve{\beta}},\bar{\ve{\alpha}})
\mathrm{a.r.}[r]
& \mathit{CF}(\bar{\ve{\beta}},\bar{\ve{\alpha}})
\mathrm{a.r.}[d]
\\
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha})
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r]
&
\mathit{CF}(\bar \ve{\beta},\bar \ve{\alpha})
\end{tikzcd}
\label{eq:g(D)->D}
\end{equation}
The description in~\eqref{eq:g(D)->D} is not the most convenient for our purposes. Instead, we describe a sequence of doubling curves $g(\ve{\Delta}_0)=\ve{\Delta}_0^0,\dots, \ve{\Delta}_0^6=\ve{\Delta}_0$ on $\Sigma_0\# \bar \Sigma_0$, such that each $\ve{\Delta}_0^i$ is obtained from $\ve{\Delta}_0^{i-1}$ by a handleslide. Furthermore, we have the following properties for each $i\in \{0,\dots 6\}$:
\begin{enumerate}
\item The diagram
\[
(\Sigma_0\# \bar \Sigma_0,\ve{\alpha}_0 \bar \ve{\alpha}_0, \ve{\beta}_0 \bar \ve{\beta}_0, \ve{\Delta}_0^i,\ve{\Delta}_0^{i+1},w)
\]
is weakly admissible.
\item Each of the complexes $\mathit{CF}(\ve{\alpha}_0 \bar \ve{\beta}_0, \ve{\Delta}_0^i)$, $\mathit{CF}(\ve{\alpha} \bar \ve{\alpha}, \ve{\Delta}_0^i)$, $\mathit{CF}(\ve{\beta} \bar \ve{\beta}, \ve{\Delta}_0^i)$, and $\mathit{CF}(\ve{\Delta}_0^i,\ve{\Delta}_0^{i+1})$ has vanishing differential.
\end{enumerate}
The curves $\ve{\Delta}_0^0,\dots \ve{\Delta}_0^6$ are shown in Figure~\ref{fig:handleswaps-Delta}. We leave it to the reader to verify the above claims. (Compare Lemma~\ref{lem:admissibility-handleswap-1}).
We compute the map $\Psi_{g(\mathfrak{D}_3)\to \mathfrak{D}_1}$ as a composition of 6 maps, corresponding to the transition maps for changing $\ve{\Delta}_0^i$ to $\ve{\Delta}_0^{i+1}$, each of which is computed by compressing a hyperbox similar to~\eqref{eq:g(D)->D}. The same argument as for $\Psi_{\mathfrak{D}_1\to \mathfrak{D}_3}$ implies that each of these maps is intertwined with the identity map by the canonical isomorphisms in ~\eqref{eq:canonical-iso}. The proof is complete.
\end{proof}
\begin{figure}[p]
\centering
\input{handleswaps_delta-2.pdf_tex}
\caption{The curves $g(\ve{\Delta}_0)=\ve{\Delta}_0^0,\dots, \ve{\Delta}_0^6=\ve{\Delta}_0$, which relate $g(\ve{\Delta}_0)$ and $\ve{\Delta}_0$ by a sequence of handleslides. Each arrow indicates the subsequent handleslide.}
\label{fig:handleswaps-Delta}
\end{figure}
\section{The cobordism map for $S^2\times S^2$}
\label{sec:example}
In this section we compute the cobordism map for $S^2\times S^2$.
We prove the following:
\begin{prop}
Let $W$ denote $S^2\times S^2$, with two 4-balls removed. Let $\mathfrak{s}$ denote the unique self-conjugate $\Spin^c$ structure on $W$. Then the cobordism map
\[
\mathit{CFI}(W,\mathfrak{s})\colon \mathbb{F}[U,Q]/Q^2\to \mathbb{F}[U,Q]/Q^2
\]
is multiplication by $Q$.
\end{prop}
Using the composition law, we may decompose $W$ as two 2-handle cobordisms, $W=W_2\circ W_1$. The first 2-handle is attached along a 0-framed unknot in $S^3$ to form $W_1$. The second 2-handle is attached along an $S^1$ fiber of $S^1\times S^2$ to form $W_2$.
We start with the standard Heegaard triple $(\Sigma,\ve{\alpha}',\ve{\alpha},\ve{\beta})$ for $W_1$. In Figure~\ref{fig:s2s2-1}, we show the diagram $(\Sigma\# \bar{\Sigma},\ve{\alpha}'\bar{\ve{\beta}},\ve{\alpha}\bar{\ve{\beta}},\ve{\beta}\bar{\ve{\beta}},\ve{\Delta})$. Therein, the top half of the figure is $\Sigma$, and the bottom half is $\bar \Sigma$. This figure also encodes $(\Sigma,\ve{\alpha}',\ve{\alpha},\ve{\beta})$.
The first quadrilateral counting map in Equation~\eqref{eq:hyperbox-2-handles} is $h^{\b\bar{\b}\to \Delta}_{\a\bar{\b}\to \a'\bar{\b}}$. The Heegaard quadruple $(\Sigma\# \bar \Sigma, \ve{\alpha}' \bar \ve{\beta}, \ve{\alpha} \bar \ve{\beta}, \ve{\beta} \bar \ve{\beta} ,\ve{\Delta})$ is depicted in Figure \ref{fig:s2s2-1}. This diagram is weakly admissible. The input for $h^{\b\bar{\b}\to \Delta}_{\a\bar{\b}\to \a'\bar{\b}}$ from $\mathit{CF}(\ve{\beta}\bar{\ve{\beta}},\ve{\Delta})$ is $\Theta^+_{\b\bar{\b},\Delta}$, and the input from $\mathit{CF}(\ve{\alpha}'\bar{\ve{\beta}},\ve{\alpha}\bar{\ve{\beta}})$ is $\Theta^+_{\a'\bar{\b},\a\bar{\b}}$. We need only evaluate $h^{\b\bar{\b}\to \Delta}_{\a\bar{\b}\to \a'\bar{\b}}(\Theta_{\a\bar{\b},\b\bar{\b}}^+)$, since $F_1^{\bar{\b},\bar{\b}}(\Theta_{\a,\b})=\Theta_{\a\bar{\b},\b\bar{\b}}^+$, and for grading reasons $h^{\b\bar{\b}\to \Delta}_{\a\bar{\b}\to \a'\bar{\b}}(\Theta_{\a\bar{\b},\b\bar{\b}}^+)$ is a multiple of $\Theta_{\a'\bar{\b},\Delta}^+$. Hence, we need only consider holomorphic quadrilaterals in
\[
\pi_2(\Theta^+_{\a'\bar{\b},\a\bar{\b}},\Theta_{\a\bar{\b},\b\bar{\b}}^+,\Theta^+_{\b\bar{\b},\Delta},\Theta_{\a'\bar{\b},\Delta}^+)
\]
There is an index $-1$ domain (shaded). This class has a holomorphic representative for a unique conformal class of rectangle. Indeed, for each slit length (along the alpha and beta arcs in the interior of $P_2$) there is a map of a varying conformal class of rectangle to the domain of $P_2$. Meanwhile, there is a holomorphic rectangle mapping to $P_1$ for a unique conformal structure.
\begin{figure}[H]
\centering
\input{s2s2-1.pdf_tex}
\caption{The Heegaard quadruple $(\Sigma\# \bar{\Sigma},\ve{\alpha}'\bar{\ve{\beta}},\ve{\alpha}\bar{\ve{\beta}},\ve{\beta}\bar{\ve{\beta}},\ve{\Delta})$. The generators of interest are $\Theta^+_{\a'\bar{\b},\a\bar{\b}}=\{\phi,\phi^+\}$, $\Theta_{\a\bar{\b},\b\bar{\b}}^+=\{\theta,\theta^+\}$, $\Theta^+_{\b\bar{\b},\Delta}=\{\tau,\tau^+\}$ and $\Theta_{\a'\bar{\b},\Delta}^+=\{z,z^+\}$. }\label{fig:s2s2-1}
\end{figure}
By inspection, there are no other positive domains and so
\[
h^{\b\bar{\b}\to \Delta}_{\a\bar{\b}\to \a'\bar{\b}}(\Theta_{\a\bar{\b},\b\bar{\b}}^+)=\Theta_{\a'\bar{\b},\Delta}^+.
\]
For $h^{\Delta\to \a'\bar{\a}'}_{\a\bar{\b}\to \a'\bar{\b}}$ we need only consider summands in the image containing $\ve{x}=\{x_0^-,x_1^+\}$, since grading considerations imply that the only other generator that can appear in the image of $h^{\Delta\to \a'\bar{\a'}}_{\a\bar{\b}\to \a'\bar{\b}}\circ f^{\b\bar{\b}\to\Delta}_{\a\bar{\b}}\circ F_1^{\bar{\b},\bar{\b}}$ is $\{x_0^+,x_1^-\}$, which is annihilated by $F_3^{\a',\a'}$.
\begin{figure}[H]
\centering
\input{s2s2-2.pdf_tex}
\caption{The Heegaard quadruple $(\Sigma,\ve{\alpha}'\bar{\ve{\beta}},\ve{\alpha}\bar{\ve{\beta}},\ve{\Delta},\ve{\alpha}'\bar{\ve{\alpha}}')$ in the definition of $h^{\Delta\to \a'\bar{\a'}}_{\a\bar{\b}\to \a'\bar{\b}}$, with some intersection points labeled. The generators of interest are $\Theta^+_{\Delta,\a'\bar{\a}'}=\{\tau^+,\tau\}$, $\Theta^+_{\a\bar{\b},\Delta}=\{\theta,\theta'\}$, $\Theta^+_{\a'\bar{\b},\a\bar{\b}}=\{ \phi^+,\phi\}$ and $\ve{x}=\{x_0^-,x_1^+\}$. }\label{fig:s2s2-2}
\end{figure}
\begin{figure}[H]
\centering
\input{s2s2-3.pdf_tex}
\caption{The Heegaard quadruple $(\Sigma,\ve{\alpha}\bar{\ve{\beta}},\ve{\Delta},\ve{\alpha}'\bar{\ve{\alpha}}',\ve{\alpha}\bar{\ve{\alpha}}')$ in the definition of $H^{\Delta\to \a\bar{\a}'}_{\a\bar{\b}}$, with some intersection points labeled. The generators of interest are $\Theta^+_{\Delta,\a'\bar{\a}'}=\{\tau^+,\tau\}$, $\Theta^+_{\a'\bar{\a}',\a\bar{\a}'}=\ve{x}_2:=\{x,x_2\}$, $\ve{y}=\{ \phi_2,\phi_3\}$ and $\Theta_{\a\bar{\b},\Delta}=\{\theta,\theta'\}$. }\label{fig:s2s2-3}
\end{figure}
One checks directly that there are no positive domains among those generators for $h^{\Delta\to \a'\bar{\a'}}_{\a\bar{\b}\to \a'\bar{\b}}$ as well as $H^{\Delta\to \a\bar{\a}'}_{\a\bar{\b}}$.
It follows that the map $F_{W_1,\mathfrak{s}|_{W_1}}$ takes the form
\[\left(
\begin{tikzcd}[column sep=1cm]
\mathit{CFI}(S^3)
\mathrm{a.r.}[r, "F_{W_1,\mathfrak{s}|_{W_1}}"]
&
\mathit{CFI}(S^1\times S^2)
\end{tikzcd}\right)
\simeq
\begin{tikzcd}[column sep=2cm, row sep=1cm, labels=description]
\mathit{CF}(S^3)
\mathrm{a.r.}[r, "1\mapsto \Theta^-"]
\mathrm{a.r.}[d, "Q(1+\iota)"]
\mathrm{a.r.}[dr,dashed, "1\mapsto Q\Theta^+"]
&
\mathit{CF}(S^1\times S^2)
\mathrm{a.r.}[d, "Q(1+\iota)"]
\\
Q\cdot\mathit{CF}(S^3)
\mathrm{a.r.}[r, "1\mapsto \Theta^-"]
&
Q\cdot\mathit{CF}(S^1\times S^2)
\end{tikzcd}
\]
We now consider the $2$-handle cobordism $W$ from $S^1\times S^2\to S^3$. This is determined by the Heegaard triple $(\Sigma,\ve{\alpha}'',\ve{\alpha}',\ve{\beta})$, where $\ve{\alpha}',\ve{\beta}$ are as before, and $\ve{\alpha}''$ is a copy of $\ve{\alpha}$. Let $\theta^-_{\a',\b}$ be the lower-degree generator of $\mathit{CF}^-(\Sigma,\ve{\alpha}',\ve{\beta})$ and $\theta^+_{\bar\b,\bar\b}$ be the top-degree generator of $\mathit{CF}^-(\bar{\Sigma},\bar\ve{\beta},\bar\ve{\beta})$. We need to calculate $h^{\b\bar\b\to \Delta}_{\a'\bar\b\to \a''\bar\b}(\theta^-_{\a',\b}\times \theta^+_{\bar\b,\bar\b})$, $h^{\Delta,\a''\bar{\a}''}_{\a'\bar{\b}\to\a''\bar{\b}}(\Theta^-_{\a'\bar{\b},\Delta})$, and $H^{\Delta\to\a'\bar{\a}''}_{\a'\bar{\b}}(\Theta^-_{\a'\bar{\b},\Delta})$.
We consider diagrams for each of those quadruples. Indeed, the first quadruple is represented in Figure \ref{fig:s2s2-1} with the role of the $\ve{\alpha}'\bar{\ve{\beta}}$ and $\ve{\alpha}\bar{\ve{\beta}}$ curves exchanged. A direct calculation shows there are no positive domains for $\pi_2(\Theta^+_{\a''\bar{\b},\a'\bar{\b}},\theta^-_{\a',\b}\times \theta^+_{\bar\b,\bar\b},\Theta^+_{\b\bar{\b},\Delta},\Theta_{\a''\bar{\b},\Delta})$ so $h^{\b\bar\b\to \Delta}_{\a'\bar\b\to \a''\bar\b}(\theta^-_{\a',\b}\times\theta^+_{\bar\b,\bar\b})=0$. Entirely similar remarks apply for $h^{\Delta,\a''\bar{\a}''}_{\a'\bar{\b}\to\a''\bar{\b}}(\Theta^-_{\a'\bar{\b},\Delta})$, and $H^{\Delta\to\a'\bar{\a}''}_{\a'\bar{\b}}(\Theta^-_{\a'\bar{\b},\Delta})$.
Computing as above, we obtain that
\[\left(
\begin{tikzcd}[column sep=1cm]
\mathit{CFI}(S^1\times S^2)
\mathrm{a.r.}[r, "F_{W_2,\mathfrak{s}|_{W_2}}"]
&
\mathit{CFI}(S^3)
\end{tikzcd}\right)
\simeq
\begin{tikzcd}[column sep=2cm, row sep=1cm, labels=description]
\mathit{CF}(S^1\times S^2)
\mathrm{a.r.}[r, "\Theta^+\mapsto 1{,} \Theta^-\mapsto 0"]
\mathrm{a.r.}[d, "Q(1+\iota)"]
&
\mathit{CF}(S^3)
\mathrm{a.r.}[d, "Q(1+\iota)"]
\\
Q\cdot\mathit{CF}(S^1\times S^2)
\mathrm{a.r.}[r, "\Theta^+\mapsto 1{,} \Theta^-\mapsto 0"]
&
Q\cdot\mathit{CF}(S^3)
\end{tikzcd}
\]
Composing the $\mathit{CFI}(W_2,\mathfrak{s}_2)\circ \mathit{CFI}(W_1,\mathfrak{s}_1)$, we obtain multiplication by $Q$.
\begin{rem}
Although $W_1$ is obtained topologically by turning the cobordism $W_2$ around, the maps $\mathit{CFI}(W_1,\mathfrak{s}_1)$ and $\mathit{CFI}(W_2,\mathfrak{s}_2)$ are not dual over $\mathbb{F}[U,Q]/Q^2$. Compare \cite{OSTriangles}*{Theorem~3.5}. Note that our definition of the 2-handle map in Equation~\eqref{eq:hyperbox-2-handles} does not naturally dualize to another 2-handle map.
\end{rem}
\section{Knots and links} \label{sec:knots}
In this section, we describe how to adapt the naturality and functoriality results of the previous sections to the case of knots and links, and in particular prove Theorems \ref{thm:link-naturality} and \ref{thm:link-functoriality}.
If $L$ is a link, in this section we write $\mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(L)$ for the pair $(\mathcal{C\!F\!L}(L), \iota_L)$. Analogously to the case of knots, our construction requires a choice of orientation on $L$, so there are potentially $2^{\ell}$ different potential models for $\iota_L$. To simplify the notation, we assume that an orientation is fixed, and we consider only the two orientations which are either coherent, or opposite to our preferred orientation. We write $\iota_{L,+}$ and $\iota_{L,-}$ for these models of the link involution.
In analogy to our construction in the case of closed 3-manifolds, we will describe several expanded models of the link involution $\tilde{\iota}_{L,\pm}$ in Section~\ref{sec:expanded-involution-knots}. To streamline the presentation, we only define transition maps with respect to the expanded models (unlike in the case of closed 3-manifolds, where we defined both non-expanded and expanded models). Since we only consider the expanded models of the involution in this section, we will write $\mathfrak{D}$ for the expanded, doubling enhanced link diagram we consider in Section~\ref{sec:expanded-involution-knots}. This notation departs from previous sections, where we wrote $\tilde \mathfrak{D}$ for expanded models of the involution. Similarly, we write $\Psi_{ \mathfrak{D}_1\to\mathfrak{D}_2}$ for the expanded models of the transition maps in this section, instead of $\tilde \Psi_{\tilde \mathfrak{D}_1\to \tilde \mathfrak{D}_2}$.
In this section, we show how to adapt the techniques of the earlier sections to prove the following naturality theorem, which implies Theorem \ref{thm:link-naturality}.
\begin{thm}
\label{thm:naturality-knots-links}
Suppose that $L$ is a link in a 3-manifold.
\begin{enumerate}
\item The transition maps $ \Psi_{ \mathfrak{D}\to \mathfrak{D}}$ are well-defined up to homotopies of morphisms of $\iota_L$-complexes.
\item $\Psi_{ \mathfrak{D}\to \mathfrak{D}}\simeq \id_{\mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}( \mathfrak{D})}$.
\item If $\mathfrak{D}_1$, $ \mathfrak{D}_2$ and $\mathfrak{D}_3$ are three expanded, doubling enhanced Heegaard link diagrams for $(Y,L)$, then
\[
\Psi_{\mathfrak{D}_2\to\mathfrak{D}_3}\circ \Psi_{ \mathfrak{D}_1\to \mathfrak{D}_2}\simeq \Psi_{ \mathfrak{D}_1\to \mathfrak{D}_3}.
\]
\end{enumerate}
\end{thm}
Additionally, we prove the following functoriality theorem, which is an expanded version of Theorem \ref{thm:link-functoriality}.
\begin{thm}
\label{thm:functoriality}
Suppose that $(W,\Sigma)$ is a link cobordism from $(Y_1,L_1)$ to $(Y_2,L_2)$. Suppose further that $\Sigma$ consists of a collection of annuli, each with one boundary component in $Y_1$, and one boundary component in $Y_2$. Suppose further that each annulus is decorated with two parallel longitudinal arcs, $\mathfrak{s}\in W$ is a $\Spin^c$ structure such that $\bar \mathfrak{s}=\mathfrak{s}+\mathit{PD}[\Sigma]$. Then the cobordism map $\mathcal{C\hspace{-.5mm}F\hspace{-.4mm}\cL\hspace{-.1mm} \cI}(W,\Sigma,\mathfrak{s})$ defined via the doubling model is an enhanced $\iota_L$-homomorphism, and is well-defined up to enhanced $\iota_L$-homotopy.
\end{thm}
\subsection{Expanded models of the knot and link involutions}
\label{sec:expanded-involution-knots}
In this section, we describe our expanded model of the knot and link involution, similar to the expanded 3-manifold involution which appeared in Section~\ref{sec:expanded-def}. This is the model we use to prove Theorem~\ref{thm:naturality-knots-links}.
The construction is as follows. Suppose that $L$ is a link in $Y$, and $(\Sigma,\ve{\alpha},\ve{\beta},\ve{w},\ve{z})$ is a Heegaard link diagram for $(Y,L)$ such that each component of $L$ contains exactly two basepoints. We pick an auxiliary point $p\in \Sigma\setminus (\ve{\alpha}\cup \ve{\beta} \cup \{w,z\})$, at which we will form the connected sum of $\Sigma$ and $\bar \Sigma$.
We write $c$ for the curve which is a meridian of the connected sum tube, as in the 3-manifold case. There is a natural 1-handle map
\[
F_1^{c\bar \b,c \bar \b}\colon \mathcal{C\!F\!L}(\Sigma,\ve{\alpha},\ve{\beta},\ve{w},\ve{z})\to \mathcal{C\!F\!L}(\Sigma\# \bar \Sigma, \ve{\alpha} c \bar \ve{\beta}, \ve{\beta} c\bar \ve{\beta},\ve{w}\cup \bar \ve{z}, \ve{z}\cup \bar \ve{w}).
\]
We pick a collection of properly embedded and pairwise disjoint arcs $\delta_1,\dots, \delta_{2g+2|L|-1}$ on $\Sigma\setminus N(p)$, such that after cutting $\Sigma\setminus N(p)$ along $\delta_1,\dots, \delta_{2g+2|L|-1}$, we are left with $2|L|$ punctured disks. We assume that each disk contains exactly one basepoint from $\ve{w}\cup \ve{z}$. Equivalently, we may assume that $\delta_1,\dots, \delta_{2g+2|L|-1}$ forms a basis of $H_1(\Sigma\setminus (\ve{w}\cup \ve{z}), p)$. We double the arcs $\delta_1,\dots, \delta_{2g+2|L|-1}$ to obtain the attaching curves $\ve{\Delta}$. See Figure~\ref{fig:11} for an example.
The complex $\mathcal{C\!F\!L}(\Sigma\# \bar \Sigma, \ve{\beta} c \bar \ve{\beta}, \ve{\Delta}, \ve{w}\cup \bar \ve{z}, \ve{z} \cup \bar \ve{w})$ represents an $|L|$-component unlink in $\#^g S^1\times S^2$, where each component contains 4 basepoints. There are two canonical homology classes, represented by cycles $\Theta^{\ve{w}}_{\b c \bar \b, \Delta}$ and $\Theta^{\ve{z}}_{\b c \bar \b, \Delta}$. These are the generators of the top degree of homology with respect to the gradings $\gr_{\ve{w}}$ and $\gr_{\ve{z}}$, respectively. See \cite{ZemCFLTQFT}*{Lemma~3.7}. We define
\[
f_{\a c \bar \b}^{\b c \bar \b\to\Delta;\ve{z}}(-)=f_{\a c \bar \b, \b c \bar \b, \Delta}(-,\Theta^{\ve{w}}_{\b c \bar \b ,\Delta})
\]
and we similarly define $f_{\a c \bar \b}^{\b c \bar \b \to \Delta;\ve{w}}$ to be the holomorphic triangle map with special input $\Theta_{\b c \bar \b, \Delta}^{\ve{z}}$. (The reason that the $\ve{z}$-map has the $\ve{w}$-generator as its input is due to the interpretation of the map in terms of decorated link cobordisms from \cite{ZemCFLTQFT}).
Similarly, there are two distinguished classes $\Theta_{\Delta,\a c \bar \a}^{\ve{w}},\Theta_{\Delta,\a c \bar \a}^{\ve{z}}$ in $\mathcal{C\!F\!L}(\Sigma\# \bar \Sigma, \ve{\Delta}, \ve{\alpha} c \bar \ve{\alpha}, \ve{w}\cup \bar \ve{z}, \ve{z}\cup \bar \ve{w})$.
We define the following two expanded models of the knot involution:
\[
\begin{split}
\tilde\iota_{L,+}&=\eta_K\circ F_{3}^{\a c, \a c} \circ f_{\a c \bar \a}^{\Delta\to \a c \bar \a;\ve{w}} \circ f_{\a c \bar \a}^{\b c \bar \b \to \Delta;\ve{z}}\circ F_1^{c \bar \b, c \bar \b}\\
\tilde\iota_{L,-}&=\eta_K\circ F_{3}^{\a c, \a c} \circ f_{\a c \bar \a}^{\Delta\to \a c \bar \a;\ve{z}} \circ f_{\a c \bar \a}^{\b c \bar \b \to \Delta;\ve{w}}\circ F_1^{c \bar \b, c \bar \b}.
\end{split}
\]
\begin{figure}[h]
\centering
\input{fig11.pdf_tex}
\caption{A doubly pointed Heegaard knot diagram with a special point $p$ (left), and the double (right).}
\label{fig:11}
\end{figure}
\begin{lem}
\label{lem:expanded-transition-links}
The maps $\tilde \iota_{L,-}$ and $\tilde \iota_{L,+}$ are homotopic to the maps $\iota_{L,-}$ and $\iota_{L,+}$, respectively, which were defined in Section~\ref{sec:doubling-knots}.
\end{lem}
\begin{proof}
We focus on the case when $L$ is a doubly pointed knot $K$, since the proof for links is not substantially different.
The proof is to identify each of the above compositions with a cobordism map from \cite{ZemCFLTQFT}. The map $F_1^{c \bar \b, c \bar \b}$ is equal to a composition of $g$ 4-dimensional 1-handle maps, as well as the birth cobordism map for adding an unknot component. The map $F_3^{\a c, \a c}$ is similarly equal to the composition of $g$ 3-handle maps, as well as a death cobordism map which deletes an unknot.
We now consider the triangle maps. The map $f_{\a c \bar \a}^{\b c \bar \b\to \Delta;\ve{z}}$ is the composition of $g$ 4-dimensional 2-handle maps (as in the ordinary doubling model of the involution), as well as the saddle cobordism map which attaches a band which has both of its ends in a component of $K\setminus \{w,z\}$ which is oriented from $z$ to $w$ (i.e. which lies in the beta handlebody). Furthermore, the band is contained in the type-$\ve{z}$ subsurface of the link cobordism. The map $f_{\a c \bar \a}^{\Delta\to \a c \bar \a;\ve{w}}$ has a similar description, except that the band is instead in the type-$\ve{w}$ subsurface.
To make the identifications in the main statement, we argue as follows. By canceling the 1-handles and 2-handles, we may identify the composition $f_{\a c \bar \a}^{\b c \bar \b\to \Delta;\ve{z}}\circ F_1^{c \bar \b, c \bar \b}$ with the composition of a birth cobordism, followed by a $\ve{w}$-saddle map. By \cite{ZemCFLTQFT}*{Proposition~8.5}, we may identify this composition with the quasi-stabilization map $S^+_{\bar z, \bar w}$, from \cite{ZemCFLTQFT}*{Section~4} and \cite{ZemQuasi}. Analogously, we may identify the composition $F_3^{\a c, \a c}\circ f_{\a c \bar \a}^{\Delta\to \a c \bar \a;\ve{w}}$ with $T^-_{z,w}$. The composition $T^-_{z,w}\circ S^+_{\bar z, \bar w}$ is homotopic to the basepoint moving map which sends $w$ to $\bar z$ and $z$ to $\bar w$, both moving in the positive direction by \cite{ZemCFLTQFT}*{Section~4.4}. See Figure~\ref{fig:9}.
\end{proof}
\begin{figure}[h]
\centering
\input{fig9.pdf_tex}
\caption{Decorations on a cylindrical link cobordism $[0,1]\times K$ corresponding to a composition of a quasi-stabilization followed by a quasi-destabilization.}
\label{fig:9}
\end{figure}
\begin{rem} The compositions $\eta_K\circ F_{3}^{\a c, \a c} \circ f_{\a c \bar \a}^{\Delta\to \a c \bar \a;\ve{z}} \circ f_{\a c \bar \a}^{\b c \bar \b \to \Delta;\ve{z}}\circ F_1^{c \bar \b, c \bar \b}$ and $\eta_K\circ F_{3}^{\a c, \a c} \circ f_{\a c \bar \a}^{\Delta\to \a c \bar \a;\ve{w}} \circ f_{\a c \bar \a}^{\b c \bar \b \to \Delta;\ve{w}}\circ F_1^{c \bar \b, c \bar \b}$ are null-homotopic. Indeed the same argument as above identifies the two maps with the link cobordism maps for a cylinder which contains a null-homotopic loop in the dividing set. Such a cobordism map induces the trivial map, since it may be identified with one of the compositions $S_{w,z}^- S_{w,z}^+$ or $T_{w,z}^- T_{w,z}^+$, which both vanish by \cite{ZemCFLTQFT}*{Lemma~4.13}.
\end{rem}
\begin{rem}
As previously mentioned, if $L$ has $\ell$-components, there are $2^\ell$ natural models of the link involution, depending on how we twist each link component. The expanded model described above can also be used to realize any of these models. We consider the diagram $(\Sigma\# \bar \Sigma, \ve{\beta} c \bar \ve{\beta}, \ve{\Delta}, \ve{w}\cup \bar \ve{z}, \ve{z}\cup \bar \ve{w})$. Suppose that $\mathbb{O}$ is a map from the set of link components of $L$ to the set $\{+,-\}$ (corresponding to a choice of whether to twist in the positive or negative direction). We let $\mathbb{Z}(\mathbb{O})$ consist of the basepoints from $\ve{w}\cup \bar \ve{z}$, for components where $\mathbb{O}=-$, as well as the basepoints from $\ve{z}\cup \bar \ve{w}$ for components where $\mathbb{O}=+$. Let $\mathbb{W}(\mathbb{O})$ denote the complement. Then each collection $\mathbb{Z}(\mathbb{O})$ and $\mathbb{W}(\mathbb{O})$ determine Maslov gradings on $\mathcal{C\!F\!L}(\Sigma\# \bar \Sigma, \ve{\beta} c \bar \ve{\beta}, \ve{\Delta}, \ve{w}\cup \bar \ve{z}, \ve{z}\cup \bar \ve{w})$, and we let $\Theta^{\mathbb{Z}(\mathbb{O})}$ and $\Theta^{\mathbb{W}(\mathbb{O})}$ be the corresponding top-degree generators. A straightforward extension of Lemma~\ref{lem:expanded-transition-links} shows that the involution $\tilde \iota_{L,\mathbb{O}}$ is computed by using $\Theta^{\mathbb{Z}(\mathbb{O})}$ for the first triangle map, and $\Theta^{\mathbb{W}(\mathbb{O})}$ for the second.
\end{rem}
\subsection{Naturality for knots and links}
Most of the ideas from the previous sections carry over to define transition maps and cobordism maps for the model of involutive link Floer homology described in Section~\ref{sec:expanded-involution-knots}.
The main new technical subtlety is that the model of involutive link Floer homology in Section~\ref{sec:expanded-involution-knots} has an additional piece of data: the choice of connected sum point $p$. If $p$ and $p'$ are two choices of special points on $\Sigma$, a choice of path on $\Sigma$ from $p$ to $p'$ can be used to relate the two models where the connected sum is taken at $p$ or $p'$, by using a point-pushing diffeomorphism along this path. In principle, a transition map defined in this manner might depend on the choice of path. To handle this issue, we introduce further expansions of the doubling model, as follows.
Suppose that $\ve{p}=\{p_1,\dots, p_n\}$ is a collection of marked points on $\Sigma\setminus (\ve{\alpha}\cup \ve{\beta}\cup \ve{w}\cup \ve{z})$. We may form an expanded model of the involution, as follows. Firstly, we consider the surface $\Sigma\#_{\ve{p}} \bar \Sigma$, where we attach a connected sum tube at each point in $\ve{p}$. Let $\ve{c}=\{c_1,\dots, c_n\}$ denote a collection of meridians of the tubes. We pick properly embedded arcs $\delta_1,\dots, \delta_{2g+2|L|+|\ve{p}|-2}$, which form a basis of $H_1(\Sigma\setminus (\ve{w}\cup \ve{z}), \ve{p})$. It is straightforward to see that if $\ve{\Delta}$ is constructed by doubling $\delta_1,\dots, \delta_{2g+2|L|+|\ve{p}|-2}$, then $(\Sigma\#_{\ve{p}} \bar \Sigma, \ve{\alpha} \ve{c} \bar \ve{\beta}, \ve{\Delta}, \ve{w}\cup \bar \ve{z}, \ve{z}\cup \bar \ve{w})$ is a diagram for $(Y,L)$, where each component of $L$ is given four basepoints.
Given a collection of points $\ve{p}$ as well as doubling arcs, as above, the construction from Section~\ref{sec:expanded-involution-knots} adapts verbatim to produce a model of the link involution. It remains to relate the models for different choices of $\ve{p}$.
Suppose $\mathcal{H}=(\Sigma,\ve{\alpha},\ve{\beta},\ve{w},\ve{z})$ is a Heegaard link diagram, and $\ve{p}$ is a subset of $\Sigma\setminus (\ve{\alpha}\cup \ve{\beta}\cup \ve{w}\cup \ve{z})$, and $q$ is a point on $\Sigma\setminus (\ve{\alpha}\cup \ve{\beta} \cup \ve{w} \cup \ve{z} \cup \ve{p})$. Suppose that $\mathfrak{D}_{\ve{p}}$ and $\mathfrak{D}_{\ve{p}\cup \{q\}}$ are doubling enhancements of $\mathcal{H}$, which use the connected sum points $\ve{p}$ and $\ve{p}\cup \{q\}$, respectively. We now define \emph{expansion} and \emph{contraction} morphisms
\[
\mathscr{E}_{q}\colon \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}})\to \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}\cup \{q\}})\quad \text{and} \quad \mathscr{C}_{q}\colon \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}\cup \{q\}})\to \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}}).
\]
We define the map $\mathscr{E}_q$ in Equation~\eqref{eq:def-E_p}. Our definition of $\mathscr{E}_{q}$ is similar to our definition of the 1-handle map, and $\mathscr{C}_{q}$ is defined similarly to the 3-handle map.
\begin{equation}
\mathscr{E}_q:= \begin{tikzcd}[column sep=2cm, row sep=2cm, labels=description]
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr, "\id"]
\mathrm{a.r.}[d,"F_1^{c_p\bar \b,c_p\bar \b}"]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[d, "F_1^{c_p c_q \bar \b,c_p c_q\bar \b}"]
\\
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}\bar \ve{\beta},\ve{\beta}\ve{c}_{\ve{p}}\bar \ve{\beta})
\mathrm{a.r.}[r, "F_1^{c_q,c_q}"]
\mathrm{a.r.}[d, "f_{\a c_p \bar \b }^{\b c_p \bar \b\to \Delta;\ve{z}}"]
&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}c_{q}\bar \ve{\beta},\ve{\beta}\ve{c}_{\ve{p}}c_q\bar \ve{\beta})
\mathrm{a.r.}[r, "\id"]
\mathrm{a.r.}[d, "f_{\a c_p c_q \bar \b}^{\b c_p c_q \bar \b\to \Delta c_q;\ve{z}}"]
\mathrm{a.r.}[dr,dashed]
&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}c_{q}\bar \ve{\beta},\ve{\beta}\ve{c}_{\ve{p}}c_q\bar \ve{\beta})
\mathrm{a.r.}[d, "f_{\a c_p c_q \bar \b}^{\b c_p c_q \bar \b\to \Delta';\ve{z}} "]
\\
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}\bar \ve{\beta},\ve{\Delta})
\mathrm{a.r.}[r, "F_1^{c_q,c_q}"]
\mathrm{a.r.}[d,"f_{\a c_p \bar \b}^{\Delta\to \a c_p \bar \a;\ve{w}}"]
&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}c_q\bar \ve{\beta},\ve{\Delta} c_q)
\mathrm{a.r.}[r, "f_{\a c_p c_q \bar \b}^{\Delta c_q\to \Delta'}"]
\mathrm{a.r.}[d, "f_{\a c_p c_q \bar \b}^{\Delta c_q\to \a c_p c_q \bar \a;\ve{w}}"]
\mathrm{a.r.}[dr,dashed]
&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}c_q\bar \ve{\beta},\ve{\Delta}')
\mathrm{a.r.}[d,"f_{\a c_p c_q \bar \b}^{\Delta'\to \a c_p c_q \bar \a;\ve{w}}"]
\\
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}\bar \ve{\beta},\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\alpha})
\mathrm{a.r.}[r, "F_1^{c_q,c_q}"]
\mathrm{a.r.}[d,"F_3^{a c_p,a c_p}"]
&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}c_q\bar \ve{\beta},\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\alpha})
\mathrm{a.r.}[r, "B_\tau"]
&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}c_q\bar \ve{\beta},\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\alpha})
\mathrm{a.r.}[d, "F_3^{\a c_p c_q, \a c_p c_q}"]
\\
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[rr, "\id"]
&
&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\bar \ve{\beta}, \bar \ve{\alpha})
\end{tikzcd}
\label{eq:def-E_p}
\end{equation}
In the definition of $\mathscr{E}_p$, we write $\tau$ for a closed curve on $\Sigma\#_{\ve{p}\cup \{q\}} \bar \Sigma$ which is symmetric with respect to the $\mathbb{Z}_2$-symmetry, and has a single transverse intersection with $c_q$.
\begin{lem}\,
\begin{enumerate}
\item\label{eq:E-C-1} The maps $\mathscr{E}_{q}$ and $\mathscr{C}_{q}$ commute with the transition maps for handleslides and isotopies of the attaching curves. Furthermore, $\mathscr{E}_q$ and $\mathscr{C}_q$ are independent of the choice of $\tau$, up to chain homotopy.
\item\label{eq:E-C-2} If $p\neq q$, then
\[
[\mathscr{E}_p,\mathscr{C}_q]\simeq 0,\quad [\mathscr{E}_p,\mathscr{E}_q]\simeq 0\quad \text{and} \quad [\mathscr{C}_p,\mathscr{C}_q]\simeq 0.
\]
\item\label{eq:E-C-3} $\mathscr{E}_p \circ \mathscr{C}_p\simeq\id$ and $\mathscr{C}_p\circ \mathscr{E}_p\simeq \id$.
\item\label{eq:E-C-4} If $\phi$ is a surface isotopy of $\Sigma$ which moves $p$ to $q$ and fixes all of the other connected sum points and basepoints, then
\[
\phi_*\simeq \mathscr{C}_p\circ \mathscr{E}_q.
\]
\end{enumerate}
\end{lem}
\begin{proof}
The proof of part~\eqref{eq:E-C-1} is essentially the same as for the 1-handle and 3-handle maps. See Lemmas~\ref{lem:1-handle-map-ind-tau} and~\ref{lem:1-handles-and-handleslides}.
The proofs of claims~\eqref{eq:E-C-2} and \eqref{eq:E-C-3} amount to constructing five hyperboxes (one for each relation). All of these hyperboxes have a similar flavor, so we describe the hyperbox for the relation $\mathscr{C}_q\circ \mathscr{E}_q\simeq \id$, and leave the remaining hyperboxes to the reader.
The hyperbox we build will correspond to the hypercube
\[
\begin{tikzcd}[labels=description]\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}})\mathrm{a.r.}[r, "\mathscr{E}_q"] \mathrm{a.r.}[d, "\id"] \mathrm{a.r.}[dr,dashed]& \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}\cup \{q\}}) \mathrm{a.r.}[d, "\mathscr{C}_q"]\\
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}})\mathrm{a.r.}[r, "\id"]& \mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K\hspace{-.3mm} \cI}(\mathfrak{D}_{\ve{p}})
\end{tikzcd}
\]
We build four constituent hyperboxes, each corresponding to one of the four levels of Equation~\eqref{eq:def-E_p}. The highest level is the following diagram:
\[
\begin{tikzcd}[
column sep={2cm,between origins},
row sep=.8cm,
labels=description,
fill opacity=.7,
text opacity=1,
]
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[dr, "\id"]
\mathrm{a.r.}[ddd, "F_1^{c_p\bar \b,c_p \bar \b}"]
\mathrm{a.r.}[rr, "\id"]
&[.7 cm]
&\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rd, "\id"]
\mathrm{a.r.}[ddd,"F_1^{c_pc_q\bar \b,c_pc_q \bar \b}"]
&[.7 cm]
\\
&[.7 cm]
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha},\ve{\beta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "\id"]
&\,
&[.7 cm]
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}, \ve{\beta})
\mathrm{a.r.}[ddd,"F_1^{c_p\bar \b,c_p \bar \b}"]
\\
\\
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha}\ve{c}_{\ve{p}}\bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta})
\mathrm{a.r.}[dr, "\id"]
\mathrm{a.r.}[rr, "F_1^{c_q,c_q}"]
&[.7 cm]\,&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta})
\mathrm{a.r.}[dr, "F_3^{c_q,c_q}\circ B_\tau"]
\\
& [.7 cm]
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}}\bar \ve{\beta}, \ve{\beta}\ve{c}_{\ve{p}} \bar \ve{\beta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"F_1^{\bar \b, \bar \b}"]
\mathrm{a.r.}[rr, "\id"]
&
&[.7 cm]
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}}\bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}}\bar \ve{\beta})
\end{tikzcd}
\]
The bottom-most level is very similar to the above, so we omit it.
We now consider the hyperbox corresponding to the second level from the top in Equation~\eqref{eq:def-E_p}:
\begin{equation}
\begin{tikzcd}[
column sep={2.2cm,between origins},
row sep={1.5cm,between origins},
labels=description,
fill opacity=.8,
text opacity=1,
execute at end picture={
\foreach \Number in {A,B,C,D,E}
{\coordinate (\Number) at (\Number.center);}
%
\filldraw[black] (A) node[rectangle,fill=white, text opacity=1]{$\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta})$};
%
\filldraw[black] (B) node[rectangle,fill=white, text opacity=1]{$\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta})$};
%
\filldraw[black] (C) node[rectangle,fill=white, text opacity=1]{$\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta})$};
%
\filldraw[black] (D) node[rectangle,fill=white, text opacity=1]{$\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta})$};
\filldraw[black] (E) node[rectangle,fill=white, text opacity=1]{$\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta}) $};
}
]
|[alias=A]|\phantom{\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta})}
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[rr, "F_1^{c_q,c_q}"]
\mathrm{a.r.}[ddrr, "\id"]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}}c_{q}\bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}}c_{q} \bar \ve{\beta})
\mathrm{a.r.}[rr, "\id"]
\mathrm{a.r.}[ddd, shift right]
\mathrm{a.r.}[dr, "B_\tau"]
\mathrm{a.r.}[ddddr,dashed]
\mathrm{a.r.}[dddrr,dashed]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}}c_{q}\bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}}c_{q} \bar \ve{\beta})
\mathrm{a.r.}[dr, "B_\tau"]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[ddddr,dashed]
&&\,
\\
&&&
|[alias=B]|\phantom{\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta})}
\mathrm{a.r.}[rr, "\id", pos=.2]
\mathrm{a.r.}[ddd]
\mathrm{a.r.}[dr, "F_3^{c_q,c_q}"]
&&
|[alias=C]|\phantom{\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta})}
\mathrm{a.r.}[dr, "F_3^{c_q,c_q}"]
\mathrm{a.r.}[ddd]
\\
&&|[alias=D]|\phantom{\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta})}
\mathrm{a.r.}[rr, "\id",pos=.7]
&&
|[alias=E]|\phantom{\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta})}
\mathrm{a.r.}[rr, "\id",pos=.7]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\beta} \ve{c}_{\ve{p}} \bar \ve{\beta})
\\
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[ddrr, "\id"]
\mathrm{a.r.}[rr, "F_1^{c_q,c_q}"]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\Delta} c_q)
\mathrm{a.r.}[rr]
\mathrm{a.r.}[dr, "B_\tau"]
\mathrm{a.r.}[drrr,dashed]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\Delta}')
\mathrm{a.r.}[dr]
\\
&
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar\ve{\beta}, \ve{\Delta} c_q )
\mathrm{a.r.}[rr, "\id"]
\mathrm{a.r.}[dr, "F_3^{c_q,c_q}"]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\Delta} c_q)
\mathrm{a.r.}[dr, "F_3^{c_q,c_q}"]
\\
\,&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[rr,"\id"]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu, pos=.6]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[rr,"\id"]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7, shift left]
\mathrm{a.r.}[from=uuu, shift left]
&&
\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\alpha} \ve{c}_{\ve{p}} \bar \ve{\beta}, \ve{\Delta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu]
\end{tikzcd}
\label{eq:main-hypercube-C_q-E_q=id}
\end{equation}
In Equation~\eqref{eq:main-hypercube-C_q-E_q=id}, each unlabeled arrow is a triangle or quadrilateral map. The back right hypercube also has a length 3 map, which is not labeled. We now explain the back-right hypercube. This hypercube is obtained by pairing $\ve{\alpha} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}$ with the following hypercube of beta attaching curves:
\begin{equation}
\begin{tikzcd}[
column sep={2cm,between origins},
row sep=.8cm,
labels=description,
fill opacity=.7,
text opacity=1,
]
\ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}
\mathrm{a.r.}[dr, "\tau"]
\mathrm{a.r.}[ddd, "\Theta_{\b c_p c_q \bar \b, \Delta c_q}^{\ve{z}}"]
\mathrm{a.r.}[rr, "1"]
\mathrm{a.r.}[dddrr,dashed]
\mathrm{a.r.}[ddddrrr,dotted]
&[.7 cm]
&\ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}
\mathrm{a.r.}[rd, "\tau"]
\mathrm{a.r.}[ddd,"\Theta_{\b c_p c_q \bar \b, \Delta'}^{\ve{z}}"]
\mathrm{a.r.}[ddddr,dashed]
&[.7 cm]
\\
&[.7 cm]
\ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "1"]
&\,
&[.7 cm]
\ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}
\mathrm{a.r.}[ddd,"\Theta_{\b c_p c_q \bar \b, \Delta c_q}^{\ve{z}}"]
\\
\\
\ve{\Delta} c_q
\mathrm{a.r.}[dr, "\tau"]
\mathrm{a.r.}[rr, "\Theta_{\Delta c_q,\Delta'}", pos=.4]
\mathrm{a.r.}[drrr,dashed, "\eta"]
&[.7 cm]\,&
\ve{\Delta}'
\mathrm{a.r.}[dr, "\Theta_{\Delta',\Delta c_q}"]
\\
& [.7 cm]
\ve{\Delta} c_q
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"\Theta_{\b c_p c_q \bar \b, \Delta c_q}^{\ve{z}}"]
\mathrm{a.r.}[rr, "1"]
&
&[.7 cm]
\ve{\Delta} c_q
\end{tikzcd}
\label{eq:3-d-hyperucbe-[C,E]}
\end{equation}
The above hypercube of beta attaching curves is constructed as follows. Length 2 chains are constructed by adapting Lemma~\ref{lem:1-handle-bottom-hypercube}. Once these are chosen, a length 3 chain may be chosen by the standard cube-filling procedure. Namely, if $C$ is the sum of all other terms of the length 3 relation, we know $\d C=0$ because of the hypercube relations on the other proper faces, and $C$ is in $\mathcal{C\hspace{-.5mm}F\hspace{-.3mm}K}(\ve{\beta} \ve{c}_{\ve{p}} c_q \bar \ve{\beta}, \ve{\Delta} c_q)$, a complex for an unlink in a connected sum of several copies of $S^1\times S^2$, of $\gr_{\ve{z}}$-grading equal to that of the top degree generator of homology. By potentially adding a copy of the top degree generator to the 2-chain labeled $\eta$, in Equation~\eqref{eq:3-d-hyperucbe-[C,E]}, the hypercube relations may be satisfied.
The hyperbox for the third level of Equation~\eqref{eq:def-E_p} is constructed very similarly. Indeed, if we switch the roles of the alpha and beta curves, then the hypercube for the third level of Equation~\eqref{eq:def-E_p} is obtained by dualizing Equation~\eqref{eq:main-hypercube-C_q-E_q=id}. Stacking these hyperboxes and compressing gives the relation $\mathscr{C}_q\circ \mathscr{E}_q\simeq \id$.
One important point to check is that the diagonal maps on the second and third hypercubes can be chosen to match. For example, to build the hypercube in Equation~\eqref{eq:main-hypercube-C_q-E_q=id}, we may potentially have to add the top degree generator to $\eta$ so that the hypercube relations are satisfiable. A similar issue occurs in constructing the length 2 map on the top face of the third hypercube as well. Nonetheless, as in the proof of Lemma~\ref{lem:1-handles-and-handleslides}, addition of the top degree generator to $\eta$ in Equation~\eqref{eq:3-d-hyperucbe-[C,E]} has no effect on the diagonal map in the compression of Equation~\eqref{eq:main-hypercube-C_q-E_q=id}. Hence, the compression of Equation~\eqref{eq:main-hypercube-C_q-E_q=id} may be stacked on top of the corresponding hypercube below it.
We now consider part~\eqref{eq:E-C-4}. This follows from the first claims, as we now explain. Let $\lambda$ be a path from $p$ to $q$. Let $\phi_{q\to p}$ denote the surface isotopy which moves $q$ to $p$ along $\lambda$, and is supported in a small neighborhood of $\lambda$. It is sufficient to show that
\[
\phi_{q\to p}\circ \mathscr{C}_p\circ \mathscr{E}_q\simeq \id.
\]
Let $x_0$ be a point in the interior of $\lambda$. We compute as follows, using the previous relations:
\[
\begin{split}
\phi_{q\to p}\circ \mathscr{C}_p\circ \mathscr{E}_q&\simeq \phi_{q\to p}\circ \mathscr{C}_p\circ \mathscr{C}_{x_0} \circ \mathscr{E}_{x_0}\circ \mathscr{E}_q\\
&\simeq \phi_{q\to p} \circ \mathscr{C}_{x_0} \circ\mathscr{E}_{q} \circ \mathscr{C}_{p}\circ \mathscr{E}_{x_0}\\
&\simeq \mathscr{C}_{x_0} \circ \phi_{q\to p} \circ\mathscr{E}_{q} \circ \mathscr{C}_{p}\circ \mathscr{E}_{x_0}\\
&\simeq \mathscr{C}_{x_0} \circ \mathscr{E}_{p} \circ \mathscr{C}_{p}\circ \mathscr{E}_{x_0}\\
&\simeq \id.
\end{split}
\]
Note that commuting $\phi_{q\to p}$ past $\mathscr{C}_{x_0}$ in the second line is possible since the diffeomorphism $\phi_{q\to p}$ is isotopic to one which fixes $x_0$. Next, $\phi_{q\to p}\circ \mathscr{E}_q\simeq \mathscr{E}_p$ by naturality of the expansion and contraction maps with respect to changes of attaching curves, which is part~\eqref{eq:E-C-1} of the statement. This establishes part~\eqref{eq:E-C-4}.
\end{proof}
\subsection{Functoriality for knots and links}
In this section, we describe several aspects of the proof of Theorem~\ref{thm:functoriality}. Firstly, to construct cobordism maps for $(W,\Sigma)\colon (Y_1,L_1)\to (Y_2,L_2)$ when $\Sigma$ is a collection of annuli such that each component of $\Sigma$ intersects $Y_1$ and $Y_2$ non-trivially, it is sufficient to describe cobordism maps on link Floer homology for 4-dimensional handles attached to the complement of a link.
Cobordism maps for 4-dimensional handles attached in the complement of a link are defined similarly to the maps for cobordisms between closed 3-manifolds (cf. \cite{ZemCFLTQFT}*{Section~5}). The 1-handle and 3-handle maps are defined as in Section~\ref{sec:1-and-3-handles}, and the 2-handle maps are defined similarly to Section~\ref{sec:2-handles}. One caveat is that there is an asymmetry in the definition of the 2-handle maps with respect to the basepoints, in that we choose the 2-handle maps to count holomorphic triangles which represent classes satisfying $\mathfrak{s}_{\ve{w}}(\psi)=\mathfrak{s}$. Note that by \cite{ZemAbsoluteGradings}*{Lemma~3.8}, we have
\[
(\mathfrak{s}_{\ve{w}}-\mathfrak{s}_{\ve{z}})(\psi)=\mathit{PD}[\Sigma].
\]
In particular, the final hypercube we use in constructing the cobordism maps for our 2-handle map will have the following form:
\[
\begin{tikzcd}[column sep=3cm, labels=description]
\mathcal{C\!F\!L}(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\alpha})
\mathrm{a.r.}[r, "f_{\bar \b;\mathfrak{s}}^{\bar \a\to \bar \a'}"]
\mathrm{a.r.}[d, "\eta_L"]
&
\mathcal{C\!F\!L}(\bar \Sigma, \bar \ve{\beta}, \bar \ve{\alpha}')
\mathrm{a.r.}[d,"\eta_L"]
\\
\mathcal{C\!F\!L}(\Sigma,\ve{\alpha}, \ve{\beta})
\mathrm{a.r.}[r, "f_{\a\to \a', \bar \mathfrak{s}+\mathit{PD}[\Sigma]}^{\b}"]
&
\mathcal{C\!F\!L}(\Sigma,\ve{\alpha}',\ve{\beta})
\end{tikzcd}
\]
\section{2-handle maps and the knot surgery formula}
\label{sec:knot-surgery}
In this section, we compute another family of involutive cobordism maps. If $K\subset S^3$, and $n\in \mathbb{Z}$, we will compute the map for the 2-handle cobordism from $S_{2n}^3(K)$ to $S^3$. Our description will be in terms of the knot surgery formula of \cite{HHSZExact}, and is in a similar spirit to the non-involutive description of the 2-handle map of Ozsv\'{a}th and Szab\'{o} \cite{OSIntegerSurgeries}*{Theorem~1.1}.
We suppose that $K\subset S^3$ is a knot, and we consider the 2-handle cobordism $W'_m(K)$ from $S^3_m(K)$ to $S^3$. This cobordism is $\Spin$ if and only if $m$ is even. Furthermore, if $m=2n$, then the unique self-conjugate $\Spin^c$ structure on $W'_{2n}(K)$ is the one which satisfies $\langle c_1(\mathfrak{s}),\Sigma\rangle =0$, where $\Sigma$ is obtained by capping $K$ with the core of the 2-handle. This $\Spin^c$ structure restricts to $[n]\in \Spin^c(S^3_{2n}(K))$ under the isomorphism $\Spin^c(S^3_{2n}(K))\cong \mathbb{Z}/2n$.
Our formula is phrased in terms of the mapping cone formula of Ozsv\'{a}th and Szab\'{o} \cite{OSIntegerSurgeries} and its involutive analog from \cite{HHSZExact}. We review these constructions briefly. We write $A_s(K)$ for the subcomplex of $\mathit{CFK}^\infty(K)$ generated by elements $\mathscr{U}^i \mathscr{V}^j \ve{x}$ where $A(\ve{x})+j-i=0$ and $j\ge -s$ and $i\ge 0$. We write $B_s(K)$ for the subspace generated by similar monomials, with no restriction on $j$. We write $\tilde{B}_s(K)$ for the subspace generated by monomials which satisfy $j\ge -s$, but with no restriction on $i$. We write $v_s\colon A_s\to B_s$ and $\tilde{v}_s\colon A_s\to \tilde{B}_s$ for the canonical inclusions. We recall from \cite{OSIntegerSurgeries} that the mapping cone formula requires a choice of homotopy equivalence from $\tilde{B}_s(K)$ to $B_{s+m}(K)$. We write $\mathfrak{F}_m$ for this map. We write $\mathbb{A}(K)$ for the direct product over $s$ of $A_s(K)$, tensored with the power series ring $\mathbb{F}[[U]]$. Similarly $\mathbb{B}(K)$ denotes the direct product of the complexes $B_s(K)$, tensored with $\mathbb{F}[[U]]$.
We recall that Ozsv\'{a}th and Szab\'{o} proved that if $K\subset S^3$ and $m\in \mathbb{Z}$, then
\[
\ve{\mathit{CF}}^-(S_m^3(K))\cong \Cone(v+h_m\colon \mathbb{A}(K)\to \mathbb{B}(K)),
\]
where $h_m=\mathfrak{F}_m\tilde v$. In \cite{HHSZExact}, we extended this to the involutive setting, and proved that
\[
\ve{\mathit{CFI}}^-(S^3_m(K))\simeq \begin{tikzcd}[column sep=2cm]\mathbb{A}(K)\mathrm{a.r.}[dr,dashed, "H_m\tilde v"]\mathrm{a.r.}[r, "v+h_m"]\mathrm{a.r.}[d, "Q(\id+\iota_{\mathbb{A}})"]& \mathbb{B}(K)\mathrm{a.r.}[d, "Q(\id+\iota_{\mathbb{B}})"]\\
Q\mathbb{A}(K)\mathrm{a.r.}[r, "v+h_m"]& Q \mathbb{B}(K)
\end{tikzcd}
\]
We refer the reader to \cite{HHSZExact}*{Section~3.5} for the definitions of $\iota_{\mathbb{A}}$ and $\iota_{\mathbb{B}}$. The map $H_m$ is defined by picking a map $H_m\colon \tilde{B}_s(K)\to B_{-s}(K)$ satisfying
\[
U^s\iota_K+\mathfrak{F}_m U^s \iota_K \mathfrak{F}_m=[\d, H_m].
\]
See \cite{HHSZExact}*{Section~3.5}.
With the above notation in place, we now state our formula for the 2-handle map:
\begin{thm}\label{thm:knot-surgery-cobordism-map} Let $\mathfrak{s}$ be the unique self-conjugate $\Spin^c$ structure on the 2-handle cobordism $W'_{2n}(K)$ from $S^3_{2n}(K)$ to $S^3$. Write $\mathit{BI}_n(K):=\Cone(Q(\id+\iota_{\mathbb{B}})\colon B_n(K)\to Q B_n(K))$, and let
\[
J\colon \mathbb{XI}_n(K)\to \mathit{BI}_n(K)
\]
be the map $J:=v_n \Pi^A_n+Q\Pi_n^B\iota_{\mathbb{X}}$, where $\Pi_s^A$ is projection onto $A_s\oplus Q A_s$ and $\Pi_s^B$ is similar. Then there is a hypercube of $\mathbb{F}[U,Q]/Q^2$-equivariant maps
\[
\begin{tikzcd}[row sep=2cm, column sep=3cm, labels=description]
\mathit{CFI}^-(S_{2n}^3(K)) \mathrm{a.r.}[dr,dashed] \mathrm{a.r.}[d,"\simeq"]\mathrm{a.r.}[r, "{\mathit{CFI}(W'_{2n}(K),\mathfrak{s})}"]& \mathit{CFI}^-(S^3) \mathrm{a.r.}[d, "\simeq"]\\
\mathbb{XI}_n(K)\mathrm{a.r.}[r, "J"]& \mathit{BI}_n(K)
\end{tikzcd}
\]
\end{thm}
\begin{rem} In Theorem~\ref{thm:knot-surgery-cobordism-map}, $\iota_{\mathbb{X}}$ denotes the involution on the mapping cone, i.e. $\iota_{\mathbb{X}}=\iota_{\mathbb{A}}+\iota_{\mathbb{B}}+H_{2n} \tilde v$. Note $\iota_{\mathbb{A}}$ vanishes when composed with $\Pi_n^B$, and hence makes no contribution to the map $J$.
\end{rem}
We will focus on the case that $n\neq 0$. The case that $n=0$ is similar.
We begin by considering the non-involutive analog of the above theorem, which is similar to \cite{OSIntegerSurgeries}*{Theorem~1.1}. Ozsv\'{a}th and Szab\'{o} constructed a map from $\ve{\mathit{CF}}^-(S^3_{2n}(K))$ to $\Cone(\ve{\mathit{CF}}^-(S^3_{2n+m}(K))\to \underline{\ve{\mathit{CF}}}^-(S^3))$, which we think of as a hypercube
\begin{equation}
\begin{tikzcd}
\ve{\mathit{CF}}^-(S^3_{2n}(K))\mathrm{a.r.}[d] \mathrm{a.r.}[dr,dashed, "k"]\\
\ve{\mathit{CF}}^-(S^3_{2n+m}(K))\mathrm{a.r.}[r]& \underline{\ve{\mathit{CF}}}^-(S^3).
\label{eq:surgery-quasi-iso-main}
\end{tikzcd}
\end{equation}
In the above, $\underline{\ve{\mathit{CF}}}^-(S^3)$ denotes a version of twisted Floer homology which is isomorphic to $\ve{\mathit{CF}}^-(S^3)\otimes\mathbb{F}[T]/(T^m-1)$.
We may consider instead a version of the above hypercube with twisted coefficients lying in $\mathbb{F}[T]$. We will write $\tilde{\ve{\mathit{CF}}}{}^-(S^3)$ for this version. We can also construct an analog of Equation~\eqref{eq:surgery-quasi-iso-main} over $\mathbb{F}[T]$, though the resulting hypercube has a different shape:
\begin{equation}
\begin{tikzcd}
\ve{\mathit{CF}}^-(S^3_{2n}(K))\mathrm{a.r.}[d] \mathrm{a.r.}[dr,dashed, "k"] \mathrm{a.r.}[r, "F"]& \tilde{\ve{\mathit{CF}}}{}^-(S^3)\mathrm{a.r.}[d, "M"]\\
\ve{\mathit{CF}}^-(S^3_{2n+m}(K))\mathrm{a.r.}[r, "G"]& \tilde{\ve{\mathit{CF}}}{}^-(S^3).
\label{eq:surgery-quasi-iso-pre-quotient}
\end{tikzcd}
\end{equation}
In Equation~\eqref{eq:surgery-quasi-iso-pre-quotient}, $F$ denotes
\[
F=\sum_{\mathfrak{s}\in \Spin^c(W_{2n}'(K))} T^{A_{W_{2n}',\Sigma}(\mathfrak{s})}\mathit{CF}(W_{2n}'(K),\mathfrak{s})
\]
where
\[
A_{W_{2n}',\Sigma}(\mathfrak{s}):=\frac{\langle c_1(\mathfrak{s}),\widehat \Sigma\rangle +2n}{2}.
\]
Additionally,
\[
M=\sum_{s\in 2\mathbb{Z}+1} U^{m(s^2-1)/8}T^{m(s+1)/2}.
\]
Note that $M=0$ if we set $T^m=1$. Hence, the hypercube in Equation~\eqref{eq:surgery-quasi-iso-main} is obtained from Equation~\eqref{eq:surgery-quasi-iso-pre-quotient} if we set $T^m=1$ and remove the top right corner of the hypercube.
The formula for $F$ is derived as follows. Ozsv\'{a}th and Szab\'{o} proved Equation~\eqref{eq:surgery-quasi-iso-main} by considering a Heegaard quadruple $(\Sigma,\ve{\alpha}_3,\ve{\alpha}_2,\ve{\alpha}_1,\ve{\beta},w,z)$, with a special genus 1 surgery region, as in \cite{OSIntegerSurgeries}*{Section~3}. In the diagrams $(\Sigma,\ve{\alpha}_1,\ve{\beta})$ and $(\Sigma,\ve{\alpha}_2,\ve{\beta})$, the point $w$ and $z$ are immediately adjacent. The diagram $(\Sigma,\ve{\alpha}_1,\ve{\beta},w)$ represents $S^3_{2n}(K)$. The diagram $(\Sigma,\ve{\alpha}_2,\ve{\beta},w)$ represents $S^3_{2n+m}(K)$. The diagram $(\Sigma,\ve{\alpha}_3,\ve{\beta},w,z)$ represents $(S^3,K)$. One obtains the hypercube in Equation~\eqref{eq:surgery-quasi-iso-pre-quotient} by considering the degenerations of holomorphic quadrilaterals. The quantity $U^{m(s^2-1)/8}T^{m(s+1)/2}$ arises from a model count of holomorphic triangles on the diagram $(\Sigma,\ve{\alpha}_3,\ve{\alpha}_2,\ve{\alpha}_1,w,z)$ (see \cite{OSIntegerSurgeries}*{Section~3}).
We now fix $\delta\gg 0$, and we set $U^\delta=0$. Write $\mathit{CF}^\delta$ for $\mathit{CF}^-/U^\delta$, and make similar notation for maps. We assume that $m\gg 0$, so the only terms of $F$ with $U$-power less than $\delta$ have $s\in \{1,-1\}$. Hence
\[
F^\delta=(1+T^{m})\sum_{
\\ \mathfrak{s}\in \Spin^c(W_{2n}'(K))} T^{A_{W_{2n}',\Sigma}(\mathfrak{s})} \mathit{CF}^\delta(W_{2n}'(K),\mathfrak{s}).
\]
On $D(-m,1)$, $\Spin^c$ structures may be identified with $2\mathbb{Z}+1$, where $s\in 2\mathbb{Z}+1$ corresponds to the $\Spin^c$ structure $\mathfrak{t}$ satisfying
\[
\langle c_1(\mathfrak{t}),S^2\rangle=s\cdot m.
\]
We are most interested in the $\Spin^c$ structures corresponding to $s=\pm 1$, which are the $\Spin^c$ structures with maximal square. Note that the composition of the natural cobordisms from $S_{2n}^3(K)$ to $S^3_{2n+m}(K)$ (which has an extra boundary component $L(m,1)$) and the cobordism from $S^3_{2n+m}(K)$ to $S^3$ is $W'_{2n}(K)\# D(-m,1)$. We identify $\Spin^c$ structures on $W_{2n}'(K)\#D(-m,1)$ as pairs $\mathfrak{s}\#[s]$ where $\mathfrak{s}\in \Spin^c(W'_{2n}(K))$ and $s\in 2\mathbb{Z}+1$.
We recall the notation of Ozsv\'{a}th and Szab\'{o} for $\Spin^c$ structures on $W_{2n}'(K)$. They write $\mathfrak{x}_s$ and $\mathfrak{y}_s$ for the $\Spin^c$ structures which satisfy
\[
\langle c_1(\mathfrak{x}_s),\Sigma\rangle +2n=2s\quad \text{and} \quad \langle c_1(\mathfrak{y}_s),\Sigma\rangle -2n=2s,
\]
respectively.
There are also analogous $\Spin^c$ structures on $W_{2n+m}'(K)$, for which we write $\mathfrak{X}_s$ and $\mathfrak{Y}_s$. The $\Spin^c$ structure $\mathfrak{X}_s$ satisfies $\langle c_1(\mathfrak{X}_s),\Sigma\rangle +2n+m=2s$ and $\mathfrak{Y}_s$ is similar. We write
\[
G_{\mathfrak{X}_s},G_{\mathfrak{Y}_s}\colon \ve{\mathit{CF}}^-(S_{2n+m}^3(K))\to \ve{\mathit{CF}}^-(S^3)
\]
for the corresponding cobordism maps.
If $s\in \mathbb{Z}$, we may define a projection map $\Pi_{s}\colon \tilde{\mathit{CF}}{}^{\delta} (S^3)\to \mathit{CF}^\delta(S^3)$ by reading off only the component $T^s$. Using this map, we may form a hypercube
\[
\begin{tikzcd}
\mathit{CF}^\delta(S^3_{2n}(K),[s])
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed, "k"]
\mathrm{a.r.}[r, "F"]
& \tilde{\mathit{CF}}{}^\delta(S^3)
\mathrm{a.r.}[d, "M"]
\mathrm{a.r.}[r, "\Pi_s M"]
&\mathit{CF}^\delta(S^3)
\mathrm{a.r.}[d, "\id"]
\\
\mathit{CF}^\delta(S^3_{2n+m}(K))
\mathrm{a.r.}[r, "G"]
&\tilde{\mathit{CF}}{}^\delta(S^3)
\mathrm{a.r.}[r, "\Pi_s"]
&\mathit{CF}^\delta(S^3).
\end{tikzcd}
\]
If we compress the above hyperbox, we obtain the following diagram:
\[
\begin{tikzcd}[column sep=3cm]
\mathit{CF}^\delta(S^3_{2n}(K),[s])
\mathrm{a.r.}[d]
\mathrm{a.r.}[dr,dashed, "\Pi_s k"]
\mathrm{a.r.}[r, "{\mathit{CF}(W'_{2n},\mathfrak{x}_s)}"]
&
\mathit{CF}^\delta(S^3)
\mathrm{a.r.}[d, "\id"]
\\
\mathit{CF}^\delta(S^3_{2n+m}(K))
\mathrm{a.r.}[r, "G_{\mathfrak{X}_s}"]
&
\mathit{CF}^\delta(S^3)
\end{tikzcd}
\]
If we instead use $\Pi_{s+2n+m}$, we obtain a hypercube relating the maps $\mathit{CF}(W_{2n}',\mathfrak{y}_s)$ and $G_{\mathfrak{Y}_s}$.
The next step of the knot surgery formula is to construct a hypercube of the following form:
\begin{equation}
\begin{tikzcd}[column sep=1.5cm] \mathit{CF}^\delta(S_{2n+m}^3(K)) \mathrm{a.r.}[r, "G_{\mathfrak{X}_s}+G_{\mathfrak{Y}_s}"] \mathrm{a.r.}[d, "\Gamma"] \mathrm{a.r.}[dr,dashed,"j"]& \underline{\mathit{CF}}^\delta(S^3) \mathrm{a.r.}[d, "\theta_w"]\\
\mathbb{A}^\delta (K) \mathrm{a.r.}[r, "v+h_{2n}"]& \mathbb{B}^\delta (K)
\end{tikzcd}
\label{eq:cone-cobordism->AB}
\end{equation}
The map $\Gamma$ is a holomorphic triangle counting map. The map $\theta_w$ is the map for trivializing the twisted coefficients defining $\underline{\mathit{CF}}^\delta(S^3)$.
The above hypercube is built by gluing together hypercubes with the following shape:
\[
\begin{tikzcd} \mathit{CF}^\delta(S_{2n+m}^3(K)) \mathrm{a.r.}[r, "G_{\mathfrak{X}_s}"] \mathrm{a.r.}[d, "\Gamma"]& \underline{\mathit{CF}}^\delta(S^3) \mathrm{a.r.}[d, "\theta_w"]\\
\mathbb{A}^\delta (K) \mathrm{a.r.}[r, "v"]& \mathbb{B}^\delta (K)
\end{tikzcd}
\quad\text{and}\quad
\begin{tikzcd} \mathit{CF}^\delta(S_{2n+m}^3(K)) \mathrm{a.r.}[r, "G_{\mathfrak{Y}_s}"] \mathrm{a.r.}[d, "\Gamma"] \mathrm{a.r.}[dr,dashed,"j"]& \underline{\mathit{CF}}^\delta(S^3) \mathrm{a.r.}[d, "\theta_w"]\\
\mathbb{A}^\delta (K)\mathrm{a.r.}[r, "h_{2n}"]& \mathbb{B}^\delta (K)
\end{tikzcd}
\]
We may combine the above hypercubes into a hyperbox of the following shape:
\begin{equation}
\begin{tikzcd}[column sep=3cm,labels=description]
\mathit{CF}^\delta(S^3_{2n}(K),[s])
\mathrm{a.r.}[d, "{\mathit{CF}^\delta(W_{2n,m})}"]
\mathrm{a.r.}[dr, "k"]
\mathrm{a.r.}[drr,dashed, "k_{\mathfrak{x}_s\#[- 1]}"]
\mathrm{a.r.}[rr, "{\mathit{CF}(W_{2n}',\mathfrak{x}_s)}"]
&[-2.5cm]
& \mathit{CF}^\delta(S^3)\mathrm{a.r.}[d, "\id"]\\
\bigg(\mathit{CF}^\delta(S^3_{2n+m}(K))
\mathrm{a.r.}[d, "\Gamma"]
\mathrm{a.r.}[dr,"j"]
\mathrm{a.r.}[r, "G"]
\mathrm{a.r.}[rr, bend right=10, "G_{\mathfrak{X}_s}\Pi_{[s]}"]
&\underline{\mathit{CF}}^\delta(S^3) \mathrm{a.r.}[d, "\theta_w"]\bigg)
&
\mathit{CF}^\delta(S^3)\mathrm{a.r.}[d,"\theta_w"]
\\
\bigg(\mathbb{A}^\delta(K)\mathrm{a.r.}[r, "v+h_{2n}"]
\mathrm{a.r.}[rr, bend right=10, "v\Pi^A_s"]
&
\mathbb{B}^\delta(K)\bigg)& B_n^\delta(K)
\end{tikzcd}
\label{eq:non-involutive-cobordism-comp-1}
\end{equation}
In the above, $\Pi_{[s]}$ denotes projection of $\mathit{CF}^\delta(S^3_{2n+m}(K)$ onto the $\Spin^c$ structure identified with $[s]\in \mathbb{Z}/(2n+m)$.
Also, we view each parenthesized group as being a single point in the set $\mathbb{E}(1,2)\cong \{0,1\}\times \{0,1,2\}$.
Symmetrically, there is another hypercube of the following form:
\begin{equation}
\begin{tikzcd}[column sep=3cm,labels=description]
\mathit{CF}^\delta(S^3_{2n}(K),[s])
\mathrm{a.r.}[d, "{\mathit{CF}(W_{2n,m})}"]
\mathrm{a.r.}[dr, "k"]
\mathrm{a.r.}[drr,dashed, "k_{\mathfrak{y}_s\#[1]}"]
\mathrm{a.r.}[rr, "{\mathit{CF}(W_{2n}',\mathfrak{y}_s)}"]
&[-2.5cm]
& \mathit{CF}^\delta(S^3)\mathrm{a.r.}[d, "\id"]
\\
\bigg(\mathit{CF}^\delta(S^3_{2n+m}(K))
\mathrm{a.r.}[d, "\Gamma"]
\mathrm{a.r.}[dr,"j"]
\mathrm{a.r.}[drr,dashed, "\Pi_{s+2n+m}^B j"]
\mathrm{a.r.}[r, "G"]
\mathrm{a.r.}[rr, bend right=10, "G_{\mathfrak{Y}_s}\Pi_{[s]}"]
&\underline{\mathit{CF}}^\delta(S^3) \mathrm{a.r.}[d, "\theta_w"]\bigg)
& \mathit{CF}^\delta(S^3)\mathrm{a.r.}[d,"\theta_w"] \\
\bigg(\mathbb{A}^\delta(K)\mathrm{a.r.}[r, "v+h_{2n}"] \mathrm{a.r.}[rr, bend right=10, "h_{2n}\Pi^A_{s}"]& \mathbb{B}^\delta(K)\bigg)& B_n^\delta(K)
\end{tikzcd}
\label{eq:non-involutive-cobordism-comp-2}
\end{equation}
Equations~\eqref{eq:non-involutive-cobordism-comp-1} and~\eqref{eq:non-involutive-cobordism-comp-2} can be used to compute the cobordism maps for $W_{2n}'(K)$, very similarly to Ozsv\'{a}th and Szab\'{o}'s \cite{OSIntegerSurgeries}*{Theorem~1.1}.
In particular, the fundamental objects to construct are the hypercubes in Equations~\eqref{eq:surgery-quasi-iso-main} and~\eqref{eq:surgery-quasi-iso-pre-quotient}. The involutive extension of Equation~\eqref{eq:surgery-quasi-iso-main} is one of the main theorems of \cite{HHSZExact}. We now describe how to extend Equation~\eqref{eq:surgery-quasi-iso-pre-quotient} to an involutive hypercube:
\begin{prop}\label{prop:tilde-hypercubes} There is a hypercube of the form:
\[
\begin{tikzcd}
\ve{\mathit{CFI}}^-(S^3_{2n}(K))\mathrm{a.r.}[d] \mathrm{a.r.}[dr,dashed, "k"] \mathrm{a.r.}[r, "F"]& \tilde{\ve{\mathit{CFI}}}{}^-(S^3)\mathrm{a.r.}[d, "M"]\\
\ve{\mathit{CFI}}^-(S^3_{2n+m}(K))\mathrm{a.r.}[r, "G"]& \tilde{\ve{\mathit{CFI}}}{}^-(S^3).
\end{tikzcd}
\]
Furthermore, $G$ and $F$ are the involutive cobordism maps summed over all $\Spin^c$ structures (as defined in this paper) and $M$ is the map $\sum_{s\in 2\mathbb{Z}+1} U^{m(s^2-1)/8}T^{m(s+1)/2}\cdot \id+EQ$. The map $E$ decomposes into a sum
\[
E=\sum_{s\in 2\mathbb{Z}+1} E_s
\]
where each $E_s$ is of homogeneous grading $-m(s^2-1)/4+1$.
\end{prop}
\begin{proof}The proof is mostly by a straightforward reinterpretation of the techniques of \cite{HHSZExact}. There is one point where our description in \cite{HHSZExact} is not convenient for our present purpose, which is in the first subcube of the central hypercube. We recall the Heegaard quintuple $(\ve{\gamma}'',\ve{\gamma}',\ve{\gamma},\ve{\delta},\ve{\delta}')$ used to define the first central hypercube. There is a natural hypercube, obtained by pairing two hypercubes (in \cite{HHSZExact}, we referred to this hypercube as the first \emph{auxiliary} hypercube). Over $\mathbb{F}[T]$, this hypercube takes the following form:
\[
\begin{tikzcd}[
column sep={2cm,between origins},
row sep=1cm,
labels=description,
fill opacity=.7,
text opacity=1,
]
\ve{\mathit{CF}}^-(\ve{\gamma},\ve{\delta})
\mathrm{a.r.}[dr, "f_{\gamma\to \gamma'}^{\delta}"]
\mathrm{a.r.}[ddd, "f_{\gamma}^{\delta\to \delta'}"]
\mathrm{a.r.}[rr, "f_{\gamma\to \gamma''}^{\delta}"]
\mathrm{a.r.}[dddrr,dashed, "h_{\gamma\to \gamma''}^{\delta\to \delta'}"]
\mathrm{a.r.}[drrr,dashed, "h_{\gamma\to \gamma'\to \gamma''}^{\delta}"]
&[.7 cm]
&\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma}'',\ve{\delta})
\mathrm{a.r.}[dr, "M"]
\mathrm{a.r.}[ddd,"f_{\gamma''}^{\delta\to \delta'}"]
&[.7 cm]
\\
&[.7 cm]
\ve{\mathit{CF}}^-(\ve{\gamma}',\ve{\delta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "f_{\gamma'\to \gamma''}^{\delta}"]
&\,
&[.7 cm]
\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma}'',\ve{\delta})
\mathrm{a.r.}[ddd,"f_{\gamma''}^{\delta\to \delta'}"]
\\
\\
\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma},\ve{\delta}')
\mathrm{a.r.}[dr, "f_{\gamma\to \gamma'}^{\delta'}"]
\mathrm{a.r.}[rr, "f_{\gamma\to \gamma''}^{ \delta'}", pos=.4]
\mathrm{a.r.}[drrr,dashed, "h_{\gamma\to\gamma'\to \gamma''}^{\delta'}"]
&[.7 cm]\,&
\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma}'',\ve{\delta}')
\mathrm{a.r.}[dr, "M"]
\\
& [.7 cm]
\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma}',\ve{\delta}')
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"f_{\gamma'}^{\delta\to \delta'}"]
\mathrm{a.r.}[rr, "f_{\gamma'\to \gamma''}^{\delta'}"]
\mathrm{a.r.}[from=uuuul,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuuul,dashed, "h_{\gamma\to \gamma'}^{\delta\to \delta'}"]
&
&[.7 cm]
\tilde{ \ve{\mathit{CF}}}{}^-(\ve{\gamma}'',\ve{\delta}')
\mathrm{a.r.}[from=uuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuull,"h_{\gamma'\to \gamma''}^{\delta\to \delta'}", dashed, pos=.3]
\end{tikzcd}
\]
The length 3 map (not shown) is the pentagon counting map $p_{\gamma\to \gamma'\to \gamma''}^{\delta\to \delta'}$.
We recall that the diagram $(\ve{\gamma}',\ve{\delta}'')$ admits a 3-handle map to a diagram $\mathcal{H}_0$, and that both $(\ve{\gamma},\ve{\delta}')$ and $(\ve{\gamma}',\ve{\delta}')$ are genus 1 stabilizations of $\mathcal{H}_0$. Furthermore, the maps $f_{\gamma\to \gamma''}^{\delta'}$ and $f_{\gamma'\to \gamma''}^{\delta'}$ are the maps for surgering on a 0-framed unknot. Since the relevant Heegaard triples are both stabilizations, both of these maps admit left inverses, which are a 3-handle map followed by a stabilization. Write $\sigma$ for the stabilization map which has $(\ve{\gamma},\ve{\delta}')$ as its codomain, and write $\sigma'$ for the stabilization map which has $(\ve{\gamma}',\ve{\delta}')$ as its codomain.
As a replacement for the diagram $\mathcal{C}_{\textrm{cen}}^{(1)}$ in \cite{HHSZExact}, we build the following diagram:
\[
\begin{tikzcd}[
column sep={2.5cm,between origins},
row sep=1.4cm,
labels=description,
fill opacity=.7,
text opacity=1,
]
\ve{\mathit{CF}}^-(\ve{\gamma},\ve{\delta})
\mathrm{a.r.}[dr, "f_{\gamma\to \gamma'}^{\delta}"]
\mathrm{a.r.}[ddd, "\id"]
\mathrm{a.r.}[rr, "f_{\gamma\to \gamma''}^{\delta}"]
\mathrm{a.r.}[dddrr,dashed, "\sigma F_3 h_{\gamma\to \gamma''}^{\delta\to \delta'}"]
&[.7 cm]
&\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma}'',\ve{\delta})
\mathrm{a.r.}[dr, "M"]
\mathrm{a.r.}[ddd,"\sigma F_3 f_{\gamma''}^{\delta\to \delta'}"]
\mathrm{a.r.}[ddddr,dashed, "\sigma' F_3 h_1 \sigma F_3 f_{\gamma''}^{\delta\to \delta'}"]
&[.7 cm]
\\
&[.7 cm]
\ve{\mathit{CF}}^-(\ve{\gamma}',\ve{\delta})
\mathrm{a.r.}[rr,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rr, "f_{\gamma'\to \gamma''}^{\delta}"]
&\,
&[.7 cm]
\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma}'',\ve{\delta})
\mathrm{a.r.}[ddd,"\sigma' F_3 f_{\gamma''}^{\delta\to \delta'}"]
\mathrm{a.r.}[from=lllu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=lllu,dashed, "h_{\gamma\to \gamma'\to \gamma''}^{\delta}"]
\\
\\
\ve{\mathit{CF}}^-(\ve{\gamma},\ve{\delta})
\mathrm{a.r.}[dr, "f_{\gamma\to \gamma'}^{\delta}"]
\mathrm{a.r.}[rr, "f_{\gamma}^{\delta\to \delta'}", pos=.4]
\mathrm{a.r.}[drrr,dashed, "h_{\gamma\to \gamma'}^{\delta\to \delta'}"]
&[.7 cm]\,&
\tilde{\ve{\mathit{CF}}}{}^-(\ve{\gamma},\ve{\delta}')
\mathrm{a.r.}[dr, "f_{\gamma\to \gamma'}^{\delta'}"]
\\
& [.7 cm]
\ve{\mathit{CF}}^-(\ve{\gamma}',\ve{\delta})
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"\id"]
\mathrm{a.r.}[rr, "f_{\gamma'}^{\delta\to \delta'}"]
&
&[.7 cm]
\tilde{ \ve{\mathit{CF}}}{}^-(\ve{\gamma}',\ve{\delta}')
\mathrm{a.r.}[from=uuull,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuull,"\sigma' F_3 h_{\gamma'\to \gamma''}^{\delta \to \delta'}", dashed, pos=.3]
\end{tikzcd}
\]
The length 3 map (not shown) is $\sigma' F_3 p_{\gamma\to \gamma'\to \gamma''}^{\delta\to \delta'}+\sigma'F_3 h_1 \sigma F_3 h_{\gamma\to \gamma''}^{\delta\to \delta'}$. Following the notation of \cite{HHSZExact}, the map $h_1$ denotes $h_{\gamma\to \gamma'\to \gamma''}^{\delta'}$. The hypercube relations are straightforward to verify. Compare \cite{HHSZExact}*{Lemma~16.4}.
The hypercubes $\mathcal{C}_{\mathrm{cen}}^{(2)}$ and $\mathcal{C}_{\mathrm{cen}}^{(3)}$ are more straightforward to modify, and we leave the details to the reader. Upon compressing and stacking these hypercubes, we get the diagram in the statement. It remains to verify the listed claims. The identification of $J$ and $G$ as cobordism maps is straightforward, and we leave the details to the reader. Compare \cite{HHSZExact}*{Lemma~16.14}.
We now consider the map labeled $M$ in the statement, the length one component of this map (i.e. the non-$Q$ component) may be identified with $\sum_{s\in 2\mathbb{Z}+1} U^{m(s^2-1)/8}T^{m(s+1)/2}$ by the lattice point counting argument in \cite{OSIntegerSurgeries}*{Section~3}. It remains to identify the $Q$-term. The $Q$-term is essentially the same as the map $E$ considered in \cite{HHSZExact}*{Section~19}. Therein, we showed that $E$ is null-homotopic over $\mathbb{F}[U,T]/(T^m-1)$. Our argument used at several places that the coefficients were in $\mathbb{F}[U,T]/(T^m-1)$ instead of $\mathbb{F}[U,T]$, so we give an alternate argument that is sufficient for our present purposes.
The first observation is that the map $E$ decomposes over $\Spin^c(D(-m,1))$, which we have already identified with $2\mathbb{Z}+1$. The grading change formula is also a routine consequence of this identification, and the standard grading change formulas for the cobordism maps in Heegaard Floer homology \cite{OSIntersectionForms}.
\end{proof}
\begin{rem}
We write $M=M_0+Q M_1$. Since $\mathit{CF}^-(S^3)\simeq \mathbb{F}[U]$ is supported only in even gradings, and $M_1$ may be viewed as a chain map from $\ve{\mathit{CF}}^-(S^3)$ to itself which shifts the mod 2 grading by 1, we may conclude that $M_1\simeq 0$. In particular, by adding a term to the length 3 map of the hypercube, we may assume that $M_1=0$. Furthermore, fixing $\delta$ and letting $m$ be sufficiently large, we see that $E^\delta$ is homogeneously graded, ao the null-homotopy may also be taken to be homogeneously graded.
\end{rem}
In \cite{HHSZExact}*{Section~21}, we also described how to extend the hypercube in Equation~\eqref{eq:cone-cobordism->AB} into the involutive setting. Using this, together with Proposition~\ref{prop:tilde-hypercubes}, we may construct the involutive analog of the hyperboxes in Equation~\eqref{eq:non-involutive-cobordism-comp-1} and~\eqref{eq:non-involutive-cobordism-comp-2}. We construct this hyperbox so that the top face has the hyperbox from Equation~\eqref{eq:non-involutive-cobordism-comp-1}, which involves the $\Spin^c$ structures $\mathfrak{x}_s$ and $\mathfrak{X}_s$ with $s=n$. There is an important subtlety in that the bottom face of this hyperbox will be the hypercube from Equation~\eqref{eq:non-involutive-cobordism-comp-2}, involves the $\Spin^c$ structures $\mathfrak{y}_{-n}$ and $\mathfrak{Y}_{-n}$. (Note that $\mathfrak{y}_{-n}=\bar\mathfrak{x}_n$).
The front face of this hyperbox will compress to have the diagonal map $H_{2n}\tilde v$, where $H_{2n}$ is the lenth 2 map from the involutive mapping cone of \cite{HHSZExact}.
We add one more level to this hyperbox so that the top and bottom levels faces coincide. We add this level to the bottom of the hyperbox. We build this new level as a pair of hypercubes. The first hypercube has the following shape:
\[
\begin{tikzcd}[
column sep={2cm,between origins},
row sep=.8cm,
labels=description,
fill opacity=1,
text opacity=1,
]
\mathit{CF}^\delta(S_{2n}^3(K),[n])
\mathrm{a.r.}[dr, "{\mathit{CF}(W_{2n,2n+m})}"]
\mathrm{a.r.}[drr, "k"]
\mathrm{a.r.}[ddd, "\id"]
\mathrm{a.r.}[rrr, "{\mathit{CF}(W_{2n}',\mathfrak{y}_{-n})}"]
\mathrm{a.r.}[drrrr,dashed, "k_{\mathfrak{y}_{-n}\#[1]}"]
&
&[.7 cm]&\mathit{CF}^\delta(S^3)
\mathrm{a.r.}[rd, "\id"]
\mathrm{a.r.}[ddd,"\id", pos=.65]
&[.7 cm]
\\
&[.7 cm]
\bigg(\mathit{CF}^\delta(S_{2n+m}(K))
\mathrm{a.r.}[r]
\mathrm{a.r.}[rrr,bend right=12,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rrr,bend right=12, "G_{\mathfrak{Y}_{-n}}\Pi_{[-n]}"]
&
\underline{\mathit{CF}}^\delta(S^3)\bigg)
&\,
&[.7 cm]
\mathit{CF}^\delta(S^3)
\\
\\
\mathit{CF}^\delta(S_{2n}^3(K),[n])
\mathrm{a.r.}[dr, "{\mathit{CF}(W_{2n,2n+m})}"]
\mathrm{a.r.}[rrr, "{\mathit{CF}(W_{2n}',\mathfrak{x}_n)}"]
\mathrm{a.r.}[drr,"k"]
\mathrm{a.r.}[drrrr,dashed, pos=.4, "k_{\mathfrak{x}_n\#[-1]} "]
&[.7 cm]\,&&
\mathit{CF}^\delta(S^3)
\mathrm{a.r.}[dr, "\id"]
\\
& [.7 cm]
\bigg(\mathit{CF}^\delta(S_{2n+m}^3(K))
\mathrm{a.r.}[r]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"\id"]
\mathrm{a.r.}[rrr, bend right=12, "G_{\mathfrak{X}_n}\Pi_{[n]}"]
&
\underline{\mathit{CF}}^\delta(S^3)\bigg)
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"\id"]
&
&[.7 cm]
\mathit{CF}^\delta(S^3)
\mathrm{a.r.}[from=uuu, "\id"]
\mathrm{a.r.}[from=lluuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=lluuu, dashed,"\Pi_{n}"]
\end{tikzcd}
\]
In the above $\Pi_n$ denotes projection onto the $B_n$-summand of $\underline{\mathit{CF}}^-(S^3)$. The hypercube relations are straightforward.
We finally have one additional hypercube, as follows:
\[
\begin{tikzcd}[
column sep={3cm,between origins},
row sep=.8cm,
labels=description,
fill opacity=1,
text opacity=1,
]
\bigg(\mathit{CF}^\delta(S^3_{2n+m}(K))
\mathrm{a.r.}[dr, "\Gamma"]
\mathrm{a.r.}[ddd, "\id"]
\mathrm{a.r.}[r]
\mathrm{a.r.}[rrr, bend left=12, "G_{\mathfrak{Y}_{-n}}\Pi_{[-n]}"]
&[-.2cm]\underline{\mathit{CF}}^\delta(S^3)\bigg)
\mathrm{a.r.}[dr, "\theta_w"]
\mathrm{a.r.}[ddd,"\id", shift right=2mm, pos=.6]
\mathrm{a.r.}[dddrr,dashed, "\Pi_n", pos=.6]
&
&\mathit{CF}^\delta(S^3)
\mathrm{a.r.}[rd, "\theta_w"]
\mathrm{a.r.}[ddd,"\id"]
&\\
&\bigg(\mathbb{A}^\delta(K)
\mathrm{a.r.}[r]
\mathrm{a.r.}[rrr,bend right=12,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[rrr, bend right=12, "h \Pi^A_{-n}"]
&
\mathbb{B}^\delta(K)\bigg)
\mathrm{a.r.}[from=ull,crossing over,"j"]
&\,
&
B_n^\delta(K)
\mathrm{a.r.}[from=ullll, crossing over, "\Pi^B_{n} j",dashed]
\\
\\
\bigg(\mathit{CF}^\delta(S_{2n+m}^3(K))
\mathrm{a.r.}[dr, "\Gamma"]
\mathrm{a.r.}[r]
\mathrm{a.r.}[drr, "j"]
\mathrm{a.r.}[rrr,bend left=12, "G_{\mathfrak{X}_n} \Pi_{[n]}"]
&
\underline{\mathit{CF}}^\delta(S^3)\bigg)
\mathrm{a.r.}[dr,"\theta_w"]
\,&&
\mathit{CF}^\delta(S^3)
\mathrm{a.r.}[dr, "\theta_w"]
\\
& [.7 cm]
\bigg(\mathbb{A}^\delta(K)
\mathrm{a.r.}[r]
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"\id", pos=.4]
\mathrm{a.r.}[rrr, bend right=12, "v\Pi^A_n"]
&
\mathbb{B}^(K)\bigg)
\mathrm{a.r.}[from=uuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=uuu,"\id"]
&
&[.7 cm]
B_n^\delta(K)
\mathrm{a.r.}[from=uuu, "\id"]
\mathrm{a.r.}[from=lluuu,line width=2mm,dash,color=white,opacity=.7]
\mathrm{a.r.}[from=lluuu, dashed,"\Pi^B_n", pos=.4]
\end{tikzcd}
\]
The above is easily check to be a hypercube of chain complexes, so the proof is complete.
|
\section{Introduction}
With the advance of autonomous driving, self-driving cars will continue to encounter emergency situations \cite{Kegelman.2018}. In these situations the vehicle has to be controlled at its dynamic limit. A task at which most road car drivers are not trained at and lack the required experience. Expert race drivers in contrast exploit the vehicles capabilities to a much greater extent.
Gaining insight in race drivers' vehicle handling process might therefore lead to further development in the areas of automated driving, e.g. in case of critical scenarios at the vehicle's dynamic limit. Data from motorsport applications are ideally suited for that task since both engineers and race drivers spend a considerable amount of time on optimising the driver-vehicle system. However, most effort is put into developing the vehicle and not into analysing drivers.\\
The aim of this work is to provide evaluation methods for drivers which allow general, track independent conclusions on the driving style and point out areas for performance improvement. The best conditions for analysing driving styles provides a driver-in-the-loop simulator since all parameters and circumstances are exactly known at every time step and the simulated environment guarantees reproducible conditions. We develop our method to evaluate race drivers based on the BMW Motorsport simulator which has been in use since 2017. Great effort has been put into matching the simulator with reality, which is the foundation of this work to analyse the performance of race drivers.
\subsection{Background}
\label{chap_background}
The overall performance in motorsport is the result of driver-vehicle interaction. On a race track, overall performance is usually determined by the lap time. Given a desired trajectory around a race track, the lap time is minimised by maximising the average velocity. Two factors limit maximising the velocity. First, the vehicle can be restrained by the power of the engine to accelerate further. This is called 'power limited'. Second, there is an acceleration limit for the vehicle due to tire-road friction. Track sections where maximising the velocity would violate this limit are referred to as 'grip limited'.
Since the power limitation cannot be overcome by the driver we focus on the grip limited track sections in this work. Along a defined trajectory, a vehicle can either accelerate or decelerate. Deceleration to a lower speed is necessary before grip limited track sections (usually corners) that otherwise would require accelerations above the vehicle's capabilities to stay on the desired trajectory. In order to drive fast for as long as possible these deceleration phases need to be as short as possible and the amount of deceleration needs to be maximal. Thus, it is reasonable for the context of this work to define optimal performance as the maximum absolute horizontal acceleration achievable while retaining the spatial path of the vehicle's trajectory. In other words, for optimal performance the available grip should be completely exploited by the vehicle, it would then always be grip or power limited. \\
The system driver-vehicle has to be described as a closed-loop system. In order to attempt a separation within this system we divide it into four layers as presented in \autoref{fig_perf_overview}. Desirable would be a division into driver and vehicle related performance to determine specific areas of improvement. The \emph{overall performance} in layer 4, as defined above, results from layer 3 which contains the combined \emph{vehicle-trajectory performance} as well as the \emph{handling performance} to keep the vehicle on this trajectory. \\
On layer 2 we have the \emph{stabilisation performance} to account for the driver's skills to stabilise the vehicle on a given trajectory and the \emph{vehicle performance} to capture the vehicle's capabilities. Influenced by both driver and vehicle are \emph{trajectory} and \emph{driveability}. The trajectory results naturally from the driver's inputs and the vehicle's reaction. Depending on the vehicle setup different trajectories are more or less time-optimal. Driveability in the context of motorsport captures how feasible it is for a driver to extract the vehicle's potential.
In this work we focus on layer 3 and 4, that is the overall, vehicle-trajectory and handling performance.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig/perf_overview3.png}
\caption{Simplified model to describe how driver and vehicle interactions lead to the overall performance. The driver can choose a trajectory and has a stabilisation performance. The stabilisation performance with driveability lead to the total handling performance. In this work, we present a method to achieve the separation in handling and vehicle/trajectory performance in layer 3.}
\label{fig_perf_overview}
\end{center}
\end{figure}
\subsection{Related work}
\citet{Donges.1982} developed a three level approach of the driving task from an engineering point of view. There are the navigation level, guidance level and stabilisation level. The navigation level includes choosing the appropriate route from a given road infrastructure while also taking an approximate time needed for the route into account. For motorsport application, the course around a race track is predefined, therefore the navigation level is not relevant. The main driving process happens in the guidance and stabilisation level. The guidance level consists of deriving a target trajectory and target velocity, considering the planned driving route and the constantly changing conditions of the track environment. According to the derived targets, control actions are chosen in an anticipatory manner to yield the best possible initial condition with the least deviations from the targets. On the stabilisation level the driver has to manipulate the vehicle controls to keep the deviations to the target trajectory at a minimum level and interacts with the vehicle in a closed loop system. Thereby, the human driver's choice of actions does not solely depend on the current state, but also on prior knowledge and experience according to \citet{Macadam.2003}.
In \cite{Worle.2019, Woerle.,Kegelman.2017b} it is discussed that race drivers have different driving styles while achieving similar lap times.
\citet{Worle.2019} defined objective criteria for race drivers in professional motorsport, based on their driving control inputs and on their chosen driving trajectories. The algorithm automatically detects cornering manoeuvres and uses pattern recognition to classify different drivers.
Similarly, \citet{Schleinitz.2019} found that top level race drivers could be distinguished with high accuracy by analysing only the brake and throttle pedal signals from data of a single corner.
\citet{Kegelman.2017b} presented a case study where two professional race drivers employed different driving styles to achieve similar lap times. Accordingly, this supports the idea that driving at the dynamic limit allows a family of solutions in terms of paths and speed.
\citet{Segers.2014} evaluated driving behaviour of race car drivers, by introducing measures based on the performance, smoothness, response and consistency of the driver input signals. The measures were used to objectively detect differences between drivers.
\citet{Lockel.2020} proposed a probabilistic framework for imitating race drivers. They defined a set of metrics based on driver inputs, for example steering or braking aggressiveness to evaluate their framework. \\
Common motorsport metrics such as lap time, top speed or accelerations provide performance measures but do not give insights how they were achieved nor why certain traits are observed. Also properties which are linked to handling and therefore to driveability are difficult to extract from these metrics \cite{goy.16}. \\
Autonomous race driving is an area where understanding of vehicle handling at the limit is crucial.
\citet{Kegelman.2018} conducted an analysis of data from vintage race cars and quantified the dispersion of paths driven by drivers during races. An autonomous race car is used to research whether the variance observed in human driving is due to error or a consequence of purposefully exploiting the vehicle limits. It showed that expert human drivers operated the vehicle around and also beyond the stability limit on purpose, whereas the autonomous vehicle always remained within the stable region. This way the expert drivers were significantly faster than the autonomous vehicle.
\citet{Hermansdorfer.20.05.2020} compared an autonomous software stack to professional human race drivers using common motorsport analysis techniques and came to a similar conclusion. The autonomous software followed the trajectory smoothly but did not drive at the handling limit and was therefore inferior to human drivers. Both studies indicate that a very important talent of race drivers is constantly exploiting the available grip.
However, in reality this is hard to determine since the available grip changes constantly and with it the current dynamic limit of the vehicle. Here, driver-in-the-loop simulators have a big advantage for analysing drivers.
\citet{S.deGroot.2011} studied the effect of training inexperienced race drivers with different levels of grip on a simulator. They measured classical motorsport performance factors such as lap time, full throttle percentage and did questionnaires to assess the drivers confidence.
\citet{vanLeeuwen.2017} determined the differences between racing and non-racing drivers on a simulator using eye-tracking. They observed that race drivers showed a more variable gaze behaviour. \\
Studies about driver evaluation in terms of vehicle handling on a simulator are rare. A possible reason is that motorsport suited simulators are very expensive and motorsport teams usually do not publish their findings. \citet{Schwarzhuber.2020} introduced the TPER method (tire potential exploitation rating) whereby the difference between the vehicle's acceleration and an optimised acceleration based on a two-track vehicle and a Pacejka tire model is used as an evaluation criteria for different motion cueing algorithms. They investigated steady-state manoeuvres and concluded that TPER is a powerful tool for simulator fidelity investigations. \\
This work is based on the TPER method, but focuses on driver evaluation instead of simulator development. Therefore, not only steady-state manoeuvres, but complete laps are evaluated. In addition to optimising the vehicle's acceleration, we introduce four optimisers to determine the maximum force on each tire. Comparing the initial state to these optimised states allows to calculate scores for overall, vehicle-trajectory and handling performance.
\subsection{Overview}
\label{chap_problem}
In chapter 2, the simulator and the vehicle model are detailed and the data set is presented. \\
Chapter 3 describes the methodology.
We optimise the vehicle's acceleration in the center of gravity (Cog) for which we introduce the optCog\xspace optimiser and compare it to the initial state.
In order to evaluate the performance of the vehicle, a second optimiser which we named optTire\xspace evaluates the maximum possible force of each tire independently.
In the next step we define scores based on the outcome of the optimiser models that allow differentiation between overall, vehicle-trajectory and handling performance.
We also propose a LSTM-based machine learning architecture to directly determine the scores form the initial data.\\
Chapter 4 presents the results of this study. Having at hand the methods and the data set it is analysed how the drivers exploit the grip by studying the slip ratios and slip angles at the tires and comparing them for the optimization states.
We also compare the professional drivers to amateur drivers. Lastly, the prediction performance of the machine learning model is shown.
Summarized, the main contributions of this work are:
\begin{itemize}
\item A method to evaluate race drivers on a simulator based on tire potential exploitation
\item Separation of the overall performance in handling and vehicle-trajectory performance
\item Application of the introduced method to a data set from professional race drivers on a top-level motorsport simulator
\item Comparison of professional drivers to two amateur drivers
\item A machine learning model architecture which can replace the optimization based method while being much faster to compute
\end{itemize}
\section{Apparatus}
The experiment to evaluate the methodology as an objective driver analysis tool is carried out at a four degrees of freedom (DOF) driver-in-the-loop simulator (DiLS).
\subsection{Driver-in-the-loop simulator}
The DiLS used for the study is located in Munich and was designed by BMW Motorsport. Therefore, special focus has been put on its applicability for the motorsport environment. It has been operational since 2017. To this day, it is frequently used for vehicle development and driver training.
The mechanical assembly consists of a static $210^{\mathrm{\circ}}$ curved screen. The screen surrounds a four DOF motion platform on top of which a mock-up of the corresponding race car's chassis is attached. Three of the motion platform's DOFs are rendering angular velocities: yaw rate $\dot{\psi}$, pitch rate $\dot{\theta}$, and roll rate $\dot{\phi}$. Translational motion is displayed in vertical direction (heave) only by means of the heave acceleration $a_{\mathrm{z}}$. The range and dynamic capabilities of each DOF were identified using classical system identification techniques.
The motion platform is controlled by a custom motion cueing algorithm (MCA) \citep{Schwarzhuber.2020}. In addition to the motion and visualization system, stimuli are provided by means of a static sound system and the steering force feedback. Both provide relevant information about vehicle states, making them vital parts of the DiLS \cite{Liu.1995, Toffin.2003}. The steering wheel is directly driven by an electric motor which allows to display up to $24\,\mathrm{Nm}$ of steering torque. The torque demand is derived from the tire forces at the front axle and the vehicle's steering geometry.
All the above mentioned systems and their respective characteristics result in a certain simulator fidelity, which was proven to have an impact on driver-vehicle interaction \cite{Markkula.2019}. For the proposed methodology, the race drivers' behaviour is a fundamental part. In the context of driving simulation the term of validity refers to similar driving behaviour between simulator and in the real world driving tasks. An earlier study conducted at the BMW Motorsport DiLS quantitatively identified its validity \cite{Schwarzhuber.2020b}. Detailed results are not relevant for the present work but should be considered if an extension to real driving tasks is sought.
\subsection{Vehicle Model}
The vehicle and tire models are described separately as these are fundamental parts not only of the driving simulator but also of the methodology itself. For vehicle dynamics modelling, a custom two track model is implemented similar as described in \cite{Heiing.2013}. The proposed methodology relies on detailed knowledge of the models. It is a mandatory requirement to have all vehicle and tire model parameters available in order to make the methodology feasible. The models' inputs are twofold. Firstly, the driver controls the vehicle via steering wheel including gear-shifter functionality, the accelerator pedal, brake pedal and clutch pedal. Secondly, a terrain server provides road inputs by means of tires' contact patch coordinates. In order to further process contact patch and tire load information to tire forces, the simulation model is equipped with a Pacejka Magic Formula tire model \cite{Pacejka.1992}. The parameters of vehicle and tire models are identified individually by means of test bench data. Subsequently, the models are validated using mainly closed loop manoeuvres. Validity in this context is only defined for circuit racing close to the vehicle's maximum capabilities. The model in use was confirmed to provide relative validity, meaning that vehicle setup variations show the same sensitivities in the virtual and the real environment.
For the purpose of driving simulation, all models are required to be realtime executable. Deploying a compiled version of the models to a \emph{Speedgoat} real-time target machine, allows an execution frequency of $2$ kHz.
\subsection{Data}
The basis of this work is a data set from BMW motorsport. It consists of time series signals from a professional
motorsport racing series which are obtained from the aforementioned simulator.
\autoref{fig_data} provides an overview of the data set. It was logged with a sample rate of $100$ Hz and there are over $10$ million instances. The data were collected in 2020 as a part of race preparation and consist of seven drivers (D1-D7) on four tracks (T1-T4). Additionally, data from two amateur drivers (D8A and D9A) were collected on track T2.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig/data_overview.png}
\caption{Overview of the data set used in this work. The two amateur drivers D8A and D9A only drove on one race track. In total there are over $10$ million observations with a sample rate of $100$ Hz.}
\label{fig_data}
\end{center}
\end{figure}
\section{Methods}
\begin{figure}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig/optimization_states.png}
\caption{Path from the Init\xspace state \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace over the optCog\xspace state \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace to the optTire\xspace state \ensuremath{\dot{\boldsymbol{y}}_{\mathrm{as}}^{\mathrm{ot}}}\xspace. Using a vehicle or tire model forces $F$ and accelerations $a$ can be calculated from the state vectors $\dot{x}$ and $\dot{y}$. }
\label{fig_optimization_states}
\end{center}
\end{figure}
\subsection{Dynamic limit optimisation}
The purpose of the method described in this section is to evaluate how far the driver or the vehicle is from the optimum, which is given by the maximum horizontal acceleration while retaining the trajectory or having the highest possible tire forces, respectively. We split the driving task into two parts. According to the three-level model of driving tasks choosing a trajectory corresponds to the guidance level and stabilising the vehicle along this trajectory to the stabilisation level \cite{Worle.20,Donges.1982}.
\subsubsection{optCog optimisation}
\input{fig/OptCog_OptTire/TPER_combined.tex}
The optCog\xspace optimisation is equivalent to the \textit{TPER} method by \citet{Schwarzhuber.2020}, we provide a brief summary of the method in this section. \autoref{fig:TPERvehicle} shows a graphical overview.
\textit{TPER}, which stands for Tire Potential Exploitation Rating, is based on a two track vehicle model as described in \cite{Jazar.2017} combined with a Pacejka Magic Formula tire model \cite{Pacejka.1992}. The vehicle state vector \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace is composed according to \eqref{eq:xIn}, where the index as with 'a' $\in \{ \mathrm{f,r} \}$ and 's' $ \in \{ \mathrm{l,r}\}$ associates the corresponding wheel from front-left to rear-right on the vehicle. For simplicity the vehicle state vector is subdivided into a constant and a variable part, which changes during optimisation
\begin{equation}
\begin{aligned}
\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace =[ &\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace_{const} ~ \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace_{var}]^{\textrm{T}}, \text{ with}\\
\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace_{const} =[&\dot{\psi}~v~Fz_\mathrm{as}~\alpha_\mathrm{as,{n-1}}~\mu_\mathrm{as}~b_\mathrm{as}~D_{\mathrm{x}}~D_{\mathrm{y}}~\\
&\gamma_\mathrm{as}~r_\mathrm{as}~n_\mathrm{Engine}~i_\mathrm{Tot}~],\\
\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace_{var} =[ &\delta^\mathrm{in}~\beta^\mathrm{in}~\kappa_\mathrm{as}^\mathrm{in}].
\end{aligned}
\label{eq:xIn}
\end{equation}
\autoref{tab:symbols} provides an overview of the parameters. The aim of the optCog\xspace optimisation is to determine an optimised vehicle state vector from \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace
\begin{equation}
\begin{aligned}
\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace =[ &\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace_\mathrm{const} ~ \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace_\mathrm{var}]^{\textrm{T}},\\
\text{with } \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace_\mathrm{var} =[ &\delta^\mathrm{oc}~\beta^\mathrm{oc}~\kappa^\mathrm{oc}_\mathrm{as}], \\
\text{such that } L^{\mathrm{oc}}(\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace )= & \min L^{\mathrm{oc}}(\dot{\bm{x}} ),
\end{aligned}
\label{eq:xOpt}
\end{equation}
whereby $L^{\mathrm{oc}}$ is the loss function and $\ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace_\mathrm{var}$ are the optimisation variables. In consideration of the external forces $F_\mathrm{x,ext}$ and $F_\mathrm{y,ext}$, the longitudinal, lateral and angular momentum equalities of the vehicle model are defined.
\begin{equation}
\begin{aligned}
m \cdot a_\mathrm{x,cog}~=~&F_\mathrm{x,rl} \cdot \cos\delta_\mathrm{rl}+F_\mathrm{x,rr} \cdot \cos\delta_\mathrm{rr}-F_\mathrm{y,rl} \cdot \sin\delta_\mathrm{rl}-
F_\mathrm{y,rr} \cdot \sin\delta_\mathrm{rr}+ \\
&F_\mathrm{x,fl} \cdot \cos\delta_\mathrm{fl}+F_\mathrm{x,fr} \cdot \cos\delta_\mathrm{fr}-
F_\mathrm{y,fl} \cdot \sin\delta_\mathrm{fl}-F_\mathrm{y,fr} \cdot \sin\delta_\mathrm{fr}+ \\ &F_\mathrm{x,Ext}
\end{aligned}
\label{eq:ImpulseLong}
\end{equation}
\begin{equation}
\begin{aligned}
m \cdot a_\mathrm{y,cog}~=~&F_\mathrm{x,rl} \cdot \sin\delta_\mathrm{rl}+F_\mathrm{x,rr} \cdot \sin\delta_\mathrm{rr}+F_\mathrm{y,rl} \cdot \cos\delta_\mathrm{rl}+F_\mathrm{y,rr} \cdot \cos\delta_\mathrm{rr}+ \\ &F_\mathrm{x,fl} \cdot \sin\delta_\mathrm{fl} + F_\mathrm{x,fr} \cdot \sin\delta_\mathrm{fr}+F_\mathrm{y,fl} \cdot \cos\delta_\mathrm{fl}+F_\mathrm{y,fr} \cdot \cos\delta_\mathrm{fr}+ \\ &F_\mathrm{y,Ext}
\end{aligned}
\label{eq:ImpulseLat}
\end{equation}
\begin{equation}
\begin{aligned}
J_\mathrm{zz} \cdot \ddot{\psi}~=~&l_\mathrm{F} \cdot (F_\mathrm{x,fl} \cdot \sin\delta_\mathrm{fl}+F_\mathrm{x,fr} \cdot \sin\delta_\mathrm{fr}+F_\mathrm{y,fl} \cdot \cos\delta_\mathrm{fl}+F_\mathrm{y,fr} \cdot \cos\delta_\mathrm{fr})- \\ &l_\mathrm{R} \cdot (F_\mathrm{x,rl} \cdot \sin\delta_\mathrm{rl}+F_\mathrm{x,rr} \cdot \sin\delta_\mathrm{rr}+F_\mathrm{y,rl} \cdot \cos\delta_\mathrm{rl}+F_\mathrm{y,rr} \cdot \cos\delta_\mathrm{rr})+ \\ &b_\mathrm{fr} \cdot (F_\mathrm{x,fr} \cdot \cos\delta_\mathrm{fr}-F_\mathrm{y,fr} \cdot \sin\delta_\mathrm{fr})+ \\ &b_\mathrm{fl} \cdot (-F_\mathrm{x,fl} \cdot \cos\delta_\mathrm{fl}+F_\mathrm{y,fl} \cdot \sin\delta_\mathrm{fl})+ \\ &b_\mathrm{rr} \cdot (F_\mathrm{x,rr} \cdot \cos\delta_\mathrm{rr}-F_\mathrm{y,rr} \cdot \sin\delta_\mathrm{rr})+ \\ &b_\mathrm{rl} \cdot (-F_\mathrm{x,rl} \cdot \cos\delta_\mathrm{rl}+F_\mathrm{y,rl} \cdot \sin\delta_\mathrm{rl})+ \\ &M_\mathrm{z,fl}+M_\mathrm{z,fr}+M_\mathrm{z,rl}+M_\mathrm{z,rr}
\end{aligned}
\label{eq:ImpulseYaw}
\end{equation}
Equations \ref{eq:ImpulseLong} to \ref{eq:ImpulseYaw} enable to search for an optimal vehicle state vector \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace in the sense of
\begin{equation}
L^{\mathrm{oc}}\left(\ensuremath{\dot{\boldsymbol{x}}}\xspace\right)~=~-\sqrt{(a_\mathrm{x,cog})^{2}+(a_\mathrm{y,cog})^{2}}.
\label{eq:costfun}
\end{equation}
The objective of the optimisation is to maximise the norm of combined longitudinal ($a_{\mathrm{x,cog}}$) and lateral acceleration ($a_{\mathrm{y,cog}}$). The loss function $L^{\mathrm{oc}}\left(\ensuremath{\dot{\boldsymbol{x}}}\xspace\right)$ is formulated as a minimisation problem.
\begin{table}[t]
\centering
\caption{Description of the symbols used in this paper, whereby the index as with 'a' $\in \{ \mathrm{f,r} \}$ and 's' $ \in \{ \mathrm{l,r}\}$ associates the corresponding wheel from front-left to rear-right on the vehicle.}
\small
\begin{tabular}{l l}
\hline
\multicolumn{1}{c}{\bfseries Symbol} & \multicolumn{1}{c}{\bfseries Description} \\ \hline
$\dot{\psi}$ & angular velocity around the vertical axis \\ \hline
$v$ & resultant velocity of the vehicle \\ \hline
$Fz_\mathrm{as}$ & tire loads \\ \hline
$\alpha_\mathrm{as,{n-1}}$ & tire slip angles of the previous sample \\ \hline
$\mu_\mathrm{as}$ & dynamic toe angles \\ \hline
$b_\mathrm{as}$ & half track widths \\ \hline
$D_{\mathrm{x}}$, $D_{\mathrm{y}}$ & longitudinal, lateral aerodynamic drag force \\ \hline
$\gamma_\mathrm{as}$ & camber angles \\ \hline
$r_\mathrm{as}$ & dynamic tire radii \\ \hline
$\ensuremath{n_\mathrm{Engine}}\xspace$ & rotational speed of the engine \\ \hline
$i_\mathrm{Tot}$ & overall gearing from the engine to the tires\\ \hline
$\delta$ & steering angle \\ \hline
$\beta$ & body sideslip angle \\ \hline
$\kappa_\mathrm{as}$ & tires' longitudinal slip ratios \\ \hline
\ensuremath{\dot{\boldsymbol{x}}}\xspace & vehicle state vector \\ \hline
\ensuremath{\dot{\boldsymbol{y}}}\xspace & tire state vector \\ \hline
$a_\mathrm{x,cog},a_\mathrm{y,cog}$ & vehilce acceleration the center of gravity \\ \hline
$br_{\mathrm{dist}}$ & brake balance \\ \hline
$F_\mathrm{x,as},F_\mathrm{y,as}$ & forces at the tire contact patch \\ \hline
$\ensuremath{r_\mathrm{Throttle}}\xspace$ & throttle pedal actuation \\ \hline
$\ensuremath{r_\mathrm{Brake}}\xspace$ & brake pedal actuation \\ \hline
$\ensuremath{r_\mathrm{Rocker,as}}\xspace$ & rocker angle (suspension movement) \\ \hline
\end{tabular}
\label{tab:symbols}
\end{table}
For simplicity, the slip ratios are optimised directly, instead of modelling the entire powertrain and braking system. Though, powertrain limitations are formulated as constraints. The first constraint ensures driving trajectory compliance by keeping the approximated resultant angular acceleration $\ddot{\psi}$ constant
\begin{equation}
\ddot{\psi}^\mathrm{oc}-\ddot{\psi}^\mathrm{in}~=~0.
\label{eq:constraintsYawAcceleration}
\end{equation}
Furthermore, the direction of the resultant horizontal acceleration has to remain constant
\begin{equation}
\dfrac{a_\mathrm{x,cog}^\mathrm{oc}}{a_\mathrm{y,cog}^\mathrm{oc}}-\dfrac{a_\mathrm{x,cog}^\mathrm{in}}{a_\mathrm{y,cog}^\mathrm{in}}~=~0.
\label{eq:constraintsARatio}
\end{equation}
Next, constraints for the slip ratios $\kappa_{\mathrm{as}}$ are introduced to consider the dependencies in the vehicle model. The difference in driving and braking torque $T_{\mathrm{as}}$ between left and right tire has to remain constant on each axle
\begin{equation}
T^{\mathrm{oc}}_\mathrm{al}-T^{\mathrm{oc}}_\mathrm{ar}-(T^{\mathrm{in}}_\mathrm{al}-T^{\mathrm{in}}_\mathrm{ar})~=~0.
\label{eq:constraintsTorqueBraking}
\end{equation}
Since only rear wheel driven cars are considered, the front axle constraint refers to braking torque only. The initial distribution of braking torques between front and rear axle has to remain unchanged. Therefore, a continuously differentiable braking distribution ($br_\mathrm{dist}$) constraint is formulated
\begin{equation}
\begin{split}
&T_\mathrm{f}^{\mathrm{oc}}-(T_\mathrm{f}^{\mathrm{oc}}+T_\mathrm{r}^{\mathrm{oc}}) \cdot br_\mathrm{dist}~=~0,~\mathrm{with} \\
br_\mathrm{dist} ~\approx~ &br_\mathrm{drv} \cdot \left(\dfrac{\arctan(-\epsilon_\mathrm{1} \cdot (T_\mathrm{f}^{\mathrm{oc}}-\epsilon_\mathrm{2}))}{\pi}+\dfrac{1}{2}\right).
\end{split}
\label{eq:constraintsBrakingBias}
\end{equation}
The variable $br_\mathrm{drv}$ is the constant braking distribution selected by the driver and $\epsilon_1$ and $ \epsilon_2$ are small numerical parameters. The range of possible side slip angles and slip ratios is bound to
\begin{equation}
\left(\dfrac{\alpha^{\mathrm{oc}}_\mathrm{as}}{\alpha_\mathrm{max}}\right)^{2}-1~\leq~0,
\label{eq:ineqConstraintsSideSlipAngle}
\end{equation}
\begin{equation}
\left(\dfrac{\kappa^{\mathrm{oc}}_\mathrm{as}}{\kappa_\mathrm{max}}\right)^{2}-1~\leq~0.
\label{eq:ineqConstraintsSlipRatio}
\end{equation}
Lastly, the maximum driving torque $T_\mathrm{r}$ is limited to the engine capabilities $T_{\mathrm{Engine}_{\mathrm{max}}}$ by
\begin{equation}
T_\mathrm{r}^{\mathrm{oc}}-T_\mathrm{Engine_{max}} \cdot i_\mathrm{Tot}~\leq~0.
\label{eq:ineqConstraintsTorqueMax}
\end{equation}
The optimization is carried out with the \textit{fmincon} solver of the \textit{MathWorks$^{\textrm{\textregistered}}$ MATLAB optimization toolbox}.
The function \textit{fmincon} finds the minimum of a problem specified by
\begin{mini!}
{\substack{$\ensuremath{\dot{\boldsymbol{x}}}\xspace$}}
{L(\ensuremath{\dot{\boldsymbol{x}}}\xspace)}
{}{}
\addConstraint{c(\ensuremath{\dot{\boldsymbol{x}}}\xspace)}{\leq 0}{\quad \text{(nonlinear inequalities)}}
\addConstraint{ceq(\ensuremath{\dot{\boldsymbol{x}}}\xspace)}{= 0}{\quad \text{(nonlinear equations) }}
\label{eq_OptimisationF}
\end{mini!}
$c(\ensuremath{\dot{\boldsymbol{x}}}\xspace)$ and $ceq(\ensuremath{\dot{\boldsymbol{x}}}\xspace)$ are functions that return vectors and $L(\ensuremath{\dot{\boldsymbol{x}}}\xspace)$ is a function that returns a scalar. $L(\ensuremath{\dot{\boldsymbol{x}}}\xspace), ~c(\ensuremath{\dot{\boldsymbol{x}}}\xspace)$ and $ceq(\ensuremath{\dot{\boldsymbol{x}}}\xspace)$ can be nonlinear functions \cite{MathWorks.2020}. For the problem specified in this work, $ceq(\ensuremath{\dot{\boldsymbol{x}}}\xspace)$ contains the vehicle model's momentum equalities (\ref{eq:ImpulseLong},\ref{eq:ImpulseLat}, \ref{eq:ImpulseYaw}) as well as equality constraints (\ref{eq:constraintsYawAcceleration},\ref{eq:constraintsARatio},\ref{eq:constraintsTorqueBraking},\ref{eq:constraintsBrakingBias}).
The inequalities contained in $c(\ensuremath{\dot{\boldsymbol{x}}}\xspace)$ are described by equations (\ref{eq:ineqConstraintsSideSlipAngle},\ref{eq:ineqConstraintsSlipRatio},\ref{eq:ineqConstraintsTorqueMax}).
\subsubsection{optTire\xspace optimisation}
The optTire optimiser is similar to the optCog\xspace optimiser. Instead of the complete vehicle with a state vector \ensuremath{\dot{\boldsymbol{x}}}\xspace, every tire state vector $\ensuremath{\dot{\boldsymbol{y}}}\xspace_\mathrm{as}$ is optimised independently. \autoref{fig:TPERcombined}b visualizes the process. We base the optTire\xspace optimisation on the optCog\xspace state to make it less dependent on a possibly sub-optimal force distribution from the Init\xspace state (this is especially relevant for the amateur drivers). The optCog\xspace tire state vector is a subset of the optCog\xspace vehicle state vector
\begin{equation}
\ensuremath{\dot{\boldsymbol{y}}_{\mathrm{as}}^{\mathrm{oc}}}\xspace \subseteq \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{oc}}}\xspace.
\end{equation}
\ensuremath{\dot{\boldsymbol{y}}_{\mathrm{as}}^{\mathrm{oc}}}\xspace is subdivided into a constant part and a variable part, which is altered by the optimiser
\begin{equation}
\begin{aligned}
\ensuremath{\dot{\boldsymbol{y}}_{\mathrm{as}}^{\mathrm{oc}}}\xspace =[ &\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{oc}_\mathrm{as,const} ~ \ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{oc}_\mathrm{as,var}]^{\textrm{T}},\\
\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{oc}_\mathrm{as,const} =[&~Fz_\mathrm{as}~\mu_\mathrm{as}~
\gamma_\mathrm{as}~r_\mathrm{as}],\\
\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{oc}_\mathrm{as,var} =[ &\alpha_\mathrm{as}^\mathrm{oc}~\kappa_\mathrm{as}^\mathrm{oc}].
\end{aligned}
\label{eq:xInTire}
\end{equation}
The optTire\xspace optimisation results in a new tire state vector
\begin{equation}
\begin{aligned}
\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{ot}_\mathrm{as} =[ &\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{oc}_\mathrm{as,const} ~ \ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{ot}_\mathrm{as,var}]^{\textrm{T}},\\
\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{ot}_\mathrm{as,var} =[ &\alpha_\mathrm{as}^\mathrm{ot}~\kappa_\mathrm{as}^\mathrm{ot}],\\
\text{such that } L_{}^{\mathrm{ot}}(\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{ot}_\mathrm{as} )= & \min L_{}^{\mathrm{ot}}(\ensuremath{\dot{\boldsymbol{y}}}\xspace^\mathrm{}_\mathrm{as} ).
\end{aligned}
\label{eq:xOptTire}
\end{equation}
Similar to the vehicle model a tire model \ensuremath{\boldsymbol{T}}\xspace is used to calculate the relevant variables from the tire state vector \ensuremath{\dot{\boldsymbol{y}}}\xspace, in this case the forces at each tire in x and y direction
\begin{equation}
F_\mathrm{as} = \ensuremath{\boldsymbol{T}}\xspace_{F_\mathrm{as}} (\ensuremath{\dot{\boldsymbol{y}}}\xspace).
\label{eq:def_TM}
\end{equation}
The force generated by each tire is maximised individually
\begin{equation}
L^{\mathrm{ot}}_\mathrm{as}\left(\ensuremath{\dot{\boldsymbol{y}}}\xspace\right)~=~-\sqrt{(F_\mathrm{x,as})^{2}+(F_\mathrm{y,as})^{2}},
\label{eq:costfunTire}
\end{equation}
while keeping the direction of the force equal to the optTire\xspace state
\begin{equation}
\dfrac{F_\mathrm{x,as}^\mathrm{ot}}{F_\mathrm{y,as}^\mathrm{ot}}-\dfrac{F_\mathrm{x,as}^\mathrm{oc}}{F_\mathrm{y,as}^\mathrm{oc}}~=~0.
\label{eq:constraintsTire}
\end{equation}
Note that there does not have to exist a valid vehicle state for the optimised tire state nor a valid trajectory.
Similar to the optCog\xspace optimiser, the \textit{MathWorks$^{\textrm{\textregistered}}$ MATLAB optimisation toolbox} with the \textit{Active-Set} solver option is used.
\subsubsection{Scores}
\label{chap_metrics}
As a first step to obtain scores from the optCog\xspace results, the acceleration at the center of gravity and forces at the tires in the xy-plane in a polar coordinate system are calculated
\begin{equation}
\begin{gathered}
a_\mathrm{\rho,cog} = \sqrt{a_\mathrm{x,cog}^2+a_\mathrm{y,cog}^2},\\
a_\mathrm{\phi,cog} = \mathrm{atan}(\frac{a_\mathrm{y,cog}}{a_\mathrm{x,cog}}),\\
F_\mathrm{\rho,as} = \sqrt{F_\mathrm{x,as}^2+F_\mathrm{y,as}^2},\\
F_\mathrm{\phi,as} = \mathrm{atan}(\frac{F_\mathrm{y,as}}{F_\mathrm{x,as}}).\\
\end{gathered}
\end{equation}
Then, the sum of all tire forces is defined as
\begin{equation}
F_\mathrm{\rho,tires} =F_\mathrm{\rho,fl}+F_\mathrm{\rho,fr}+F_\mathrm{\rho,rl}+F_\mathrm{\rho,rr}.
\end{equation}
We define the following metrics which can be interpreted as scores for the system vehicle-driver.
The ratio between the amount of initial vehicle acceleration at the center of gravity to the optCog\xspace acceleration yields $S_\mathrm{handling}$
\begin{equation}
\begin{gathered}
S_\mathrm{handling} = \frac{a_\mathrm{\rho,cog}^{\mathrm{in}}}{a_\mathrm{\rho,cog}^{\mathrm{oc}}}.\\
\end{gathered}
\end{equation}
The vehicle score is defined as the ratio between the optCog\xspace and optTire\xspace forces
\begin{equation}
\begin{gathered}
S_\mathrm{veh-traj} = \frac{F_\mathrm{\rho,tires}^{\mathrm{oc}}}{F_\mathrm{\rho,tires}^{\mathrm{ot}}}.\\
\end{gathered}
\end{equation}
Finally, the overall score emerges from the Init\xspace forces to the optTire\xspace forces
\begin{equation}
\begin{gathered}
S_\mathrm{tot} = \frac{F_\mathrm{\rho,tires}^{\mathrm{in}}}{F_\mathrm{\rho,tires}^{\mathrm{ot}}}.\\
\end{gathered}
\end{equation}
\subsubsection{Limitations}
Since we apply the optimisation method on a relatively large data set and aim to use the methods for race preparation in future, computation time is a concern. This is also a main reason why we introduce the constraint of a constant driving line which is a simplification since the driver's performance on the stabilisation level also influences the driving line. For example when entering a corner too fast, the vehicle might get pushed towards the outside, although not being intended by the driver. However, for the professional drivers we can assume that they anticipate this effect. This is supported by \citet{Macadam.2003} who stated that the drivers choice of action is not only based on the current state but also prior knowledge and experience.\\
We execute the optimisation for approximately $3.5$ million data instances. The calculation time for one lap with an average length of $80$ seconds is approximately $200$ seconds on computer with two 10-core \emph{Intel\textregistered Xeon\textregistered Silver 4144} CPUs.
To reach this speed, both optimiser rely on parallel processing which significantly reduces calculation time. As a consequence, the models compute each data point isolated from its temporal surrounding. Dynamic effects due to changed accelerations from the optimised states are not respected, for example a change in load transfer. However, the closer the Init\xspace state is to the dynamic limit (the optCog\xspace state) the smaller the error will be, which is where race drivers usually operate the vehicle. Though, for some applications a calculation time of $200$ seconds per lap is still too high. For example in race simulations 50 and more laps can be driven and a very quick analysis is needed. The resulting calculation time of $10000$ seconds would make it infeasible in practice. Ideal would be a real-time calculation. Therefore, we propose an end-to-end machine learning approach to replace the computational expensive optimisation.
\subsection{Machine learning predictor}
We propose a machine learning model as an approximation to the aforementioned optimisation approach. Depending on the computer hardware and model size the runtime speed improvement is more than one order of magnitude and the model is real-time application suited. The three scores defined in section \ref{chap_metrics} are predicted simultaneously by an artificial neural network \ensuremath{\boldsymbol{P}}\xspace based on $\dot{\bm{x}}^\mathrm{p}$, which is a part of the initial vehicle state vector \ensuremath{\dot{\boldsymbol{x}}^{\mathrm{in}}}\xspace
\begin{equation*}
[\tilde{S}_\mathrm{handling}~\tilde{S}_\mathrm{veh-traj}~\tilde{S}_\mathrm{tot}] = \ensuremath{\boldsymbol{P}}\xspace (\dot{\bm{x}}^\mathrm{p}).
\end{equation*}
\subsubsection{Artificial neural network architecture}
The model is based on a LSTM neural network architecture, which has been used for time series prediction on a motorsport data set before \cite{Schleinitz_vasp.2021}. The hyperparameters are listed in \autoref{tab_lstm_params} and the architecture in \autoref{fig_lstm}. The model is built in R 3.6.2 \cite{RDevelopmentCoreTeam.2008} using the R interface to keras \cite{JJAllaire.2018} and tensorflow \cite{JJAllaire.2018b}. \\
The LSTM was introduced by \citet{Hochreiter.1997}. A LSTM consists of an \emph{input gate}, \emph{forget gate} and \emph{output gate}. With these, the LSTM can decide which information will be forgotten, learned or passed on to the next time step. The outputs of the previous time step serve as inputs for the current time step. Apart from the output, a LSTM passes a cell state to the next time step. The cell state makes it easier for information to flow unchanged through multiple time steps. This results in the ability of the LSTM cells to learn long-term dependencies while avoiding the vanishing /exploding gradient problem \cite{vanHoudt.2020, Schleinitz_vasp.2021}.
\begin{table}[b]
\centering
\caption{Hyperparameters for the LSTM-based predictor module}
\small
\begin{tabular}{l l}
\hline
\textbf{Parameter} & \textbf{Setting} \\
\hline
Learning rate & 0.001 \\ \hline
Batch size & 128 \\ \hline
Dropout & 0.3 \\ \hline
Recurrent dropout & 0.1 \\ \hline
Optimiser & "Adam" \cite{Kingma.22.12.2014} \\\hline
Input temporal dimension & 100\\ \hline
Input feature dimension & 24\\ \hline
Output temporal dimension & 100\\ \hline
Output feature dimension & 3\\
\hline
\end{tabular}
\label{tab_lstm_params}
\end{table}
\begin{figure}
\includegraphics[width=0.9\textwidth]{fig/lstm_pred.png}
\caption{Machine learning predictor model architecture. Two LSTM and two fully connected layers constitute the hidden layers. }
\label{fig_lstm}
\end{figure}
\subsubsection{Data preparation}
All time series data are normalized to increase the numerical stability. For this purpose let $\bm{w} \in \mathbb{R}^{m}$ be a raw time series signal. Then, $\bm{x} \in \mathbb{R}^{m}$ is calculated by
\begin{equation*}
\bm{x}= \frac{1}{\sigma(\bm{w})}(\bm{w}- \mu(\bm{w})\bm{e} ),
\end{equation*}
with the all-ones vector $\bm{e} \in \mathbb{R}^{m}$, whereby $\mu(\bm{w}) \in \mathbb{R}$ is the sample mean and $\sigma(\bm{w}) \in \mathbb{R}$ the sample standard deviation of $\bm{w}$.
The data is split in training ($80\%$), validation ($10\%$) and test set ($10\%$). Also, the Track T2 is excluded from the training set to examine the model accuracy on an unseen track.
\subsubsection{Model selection}
\label{chap_model_sel}
The root-mean square error (RMSE) between prediction and reference is used as an evaluation criteria for the neural network. The RMSE between a reference signal $\bm{yr}$ and a predicted signal $\bm{y}$ with $m$ time steps is
\begin{equation*}
\text{RMSE}(\bm{yr},\bm{y})=\sqrt{\frac{1}{m} \sum\limits_{t=0}^{m} (\bm{yr}^{(t)}-\bm{y}^{(t)})^2 }.
\end{equation*}
Three different models M1, M2 and M3 are trained on the data set. These models have different input time series. All inputs are obtained from the initial vehicle state.
\begin{itemize}
\item M1 (32 inputs): $\alpha_{\mathrm{as}}^{\mathrm{}}$, $\kappa_{\mathrm{as}}^{\mathrm{}}$, $Fx_{\mathrm{as}}^{\mathrm{}}$, $Fy_{\mathrm{as}}^{\mathrm{}}$, $Fz_{\mathrm{as}}^{\mathrm{}}$,
$\ensuremath{r_\mathrm{Brake}}\xspace$, $\ensuremath{r_\mathrm{Throttle}}\xspace$, $\dot{\psi}$, $a_\mathrm{x,cog}$, $a_\mathrm{y,cog}$, $\delta$, $v$, $\beta$, $\ensuremath{r_\mathrm{Rocker,as}}\xspace$
\item M2 (45 inputs): $\alpha_{\mathrm{as}}^{\mathrm{}}$, $\kappa_{\mathrm{as}}^{\mathrm{}}$, $Fx_{\mathrm{as}}^{\mathrm{}}$, $Fy_{\mathrm{as}}^{\mathrm{}}$, $Fz_{\mathrm{as}}^{\mathrm{}}$,
$\ensuremath{r_\mathrm{Brake}}\xspace$, $\ensuremath{r_\mathrm{Throttle}}\xspace$, $\dot{\psi}$, $a_\mathrm{x,cog}$, $a_\mathrm{y,cog}$, $\delta$, $v$, $\beta$, $\ensuremath{r_\mathrm{Rocker,as}}\xspace$, $\gamma_\mathrm{as}$, $r_\mathrm{as}$, $\mu_\mathrm{as}$
\item M3 (16 inputs): $Fz_{\mathrm{as}}^{\mathrm{}}$,
$\ensuremath{r_\mathrm{Brake}}\xspace$, $\ensuremath{r_\mathrm{Throttle}}\xspace$, $\dot{\psi}$, $a_\mathrm{x,cog}$, $a_\mathrm{y,cog}$, $\delta$, $v$, $\beta$, $\ensuremath{r_\mathrm{Rocker,as}}\xspace$
\end{itemize}
The models' accuracy should be consistent for different factors of influence. \autoref{fig_lstm_res} shows the RMSE in dependence of tracks as well drivers. To detect possible overfitting to the training data, we compare the RMSE on seen and unseen tracks (which were not a part of the training set) in \autoref{fig_lstm_res} a. None of the models showed overfitting to the seen tracks, the error on the unseen track is even a bit lower. To further test the models, they are also tested on the amateur drivers' data. The models were only trained on data from the professional drivers, which have distinctly higher scores. The amateur driver scores are therefore well outside the scope of the training data. \autoref{fig_lstm_res} depicts the resulting RMSE. The error is as expected higher but the models are very usable for the task at hand. This indicates that the models are robust.
In all cases the model M2 has the lowest error and is therefore chosen for this work. The additional inputs compared to M1 and M3 contained additional information which helped the model to predict more accurately. Since the advantage of M2 is also visible for unseen tracks and the amateur drivers it is not a result of overfitting.
\begin{figure}
\begin{subfigure}[c]{0.5\textwidth}
\includegraphics[width=1\textwidth]{fig/prediction_tracks.png}
\subcaption{Prediction model accuracy over tracks. The tracks are split into seen tracks, which were part of the training set and unseen tracks which were not. There is no significant RMSE difference between the two cases. }
\end{subfigure}
\begin{subfigure}[c]{0.5\textwidth}
\includegraphics[width=1\textwidth]{fig/prediction_drivers.png}
\subcaption{Model accuracy over drivers. The professional drivers 'Drivers Pro' were part of the training set, whereas the amateur drivers 'Drivers Am' were not. The error for the amateur drivers is clearly higher, however, still in the same range. }
\end{subfigure}
\caption{In all cases the model M2 has the lowest error. The error for the prediction of \ensuremath{S_\mathrm{veh-traj}}\xspace is generally higher than for \ensuremath{S_\mathrm{handling}}\xspace or \ensuremath{S_\mathrm{tot}}\xspace. }
\label{fig_lstm_res}
\end{figure}
\subsection{Control states}
In order to analyse driver performance independent from a specific track section, control states are defined based on the driver's inputs. Three boolean variables describe if the brake pedal, throttle pedal or steering wheel angle $\delta$ are active
\begin{equation}
B_{brake}=
\begin{cases}
1 ,& \text{if } \ensuremath{r_\mathrm{Brake}}\xspace>10 bar\\
0, & \text{otherwise,}
\end{cases}
\end{equation}
\begin{equation}
B_{throttle}=
\begin{cases}
1 ,& \text{if } \ensuremath{r_\mathrm{Throttle}}\xspace>10 \%\\
0, & \text{otherwise,}
\end{cases}
\end{equation}
\begin{equation}
B_{steer}=
\begin{cases}
1 ,& \text{if } \vert \delta \vert>10^\circ \\
1, & \text{if } \vert\dot{\delta}\vert > 500^\circ/s\\
0, & \text{otherwise.}
\end{cases}
\end{equation}
The second condition for $B_{steer}$ ensures that occasions where the steering angle is small but the change rate is high, are included. This can be for example during a quick counter steer. \\
The combination of the boolean variables results in the four control states. These relate to the commonly defined cornering sections \cite{goy.16}, which are in parentheses:
\begin{itemize}
\item Pure brake (Braking): $B_{brake}=1, ~ B_{throttle}=0, ~ B_{steer}= 0$
\item Trail brake (Turn Entry): $B_{brake}=1, ~ B_{throttle}=0, ~ B_{steer}= 1$
\item Pure steer (Mid Corner): $B_{brake}=0, ~ B_{throttle}=0, ~ B_{steer}= 1$
\item Throttle steer (Turn Exit) : $B_{brake}=0, ~ B_{throttle}=1, ~ B_{steer}= 1$
\end{itemize}
The other possible combinations are not relevant for the driver analysis in the context of this paper. For example under pure throttle the vehicle's acceleration is usually limited by the engine and not the driver. An overview of the control states in the time domain is e.g. depicted in \autoref{fig_kappa}.
\section{Results}
\subsection{optTire\xspace optimisation}
Comparing the Init\xspace state to the optTire\xspace state leads to insights how the vehicle is handled at the dynamic limit. We look at the initial and optimised slip ratios and slip angles of the tires.
\subsubsection{Slip ratio $\kappa$}
\label{chap_kappa}
\autoref{fig_kappa}a shows that the slip ratio in the optCog\xspace state $\kappa^\mathrm{oc}$ changes with a higher frequency in the Pure Brake zone than the Init\xspace state $\kappa^\mathrm{in}$. The optTire\xspace state $\kappa^\mathrm{ot}$ which depicts the optimum slip ratio for each tire provides a reference. This optimum cannot be reached for all tires simultaneously, therefore the optCog\xspace optimiser has to make trade-offs by alternating the tires that reach the optimum.
\autoref{fig_kappa}b shows the distribution of slip ratio differences from the Init\xspace to the optTire\xspace state per tire. It seems that all drivers are very careful not to exceed the optimal value for $\kappa$, since that would lead quickly to a so called 'tire lock-up' in reality. That means the tire is no longer rotating while the vehicle is still moving and just rubbing over the road surface. A tire lock-up not only decreases the braking acceleration but also causes significant tire wear due to the high relative velocities, which in turn can significantly decrease performance for the following laps. For comparison the amateur drivers are shown as well. The center of the distributions is further away from the optimum both on the front and on the rear axle. We conclude that the amateur drivers are more cautious and cannot reach the optimum as well as the the professional drivers.
\begin{figure}
\begin{subfigure}[c]{1\textwidth}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig/kappa_example_area_ano.png}
\end{center}
\subcaption{Exemplary overview about $\kappa$ in the time domain for the Init\xspace, optCog\xspace and optTire\xspace state.}
\end{subfigure}
\begin{subfigure}[c]{1\textwidth}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig/kappa_hist.png}
\end{center}
\subcaption{ $\kappa$ difference distribution between the Init\xspace and optTire\xspace state for the complete data set over the four tires and drivers. Negative values indicate that the absolute of $\kappa^\mathrm{in}$ is smaller than the absolute of $\kappa^\mathrm{ot}$.}
\end{subfigure}
\caption{The histograms (bottom plot) show the drivers very rarely exceed the optimal Slip Ratio $\kappa$. The top plot shows such rare occasion for the front right wheel. }
\label{fig_kappa}
\end{figure}
\subsubsection{Slip angle $\alpha$}
\autoref{fig_alpha} gives a similar overview for the slip angle $\alpha$. \autoref{fig_alpha}a shows an example in the time domain. For the front axle (the two topmost plots) the Init\xspace and optCog\xspace state are close to each other, on the rear axle seems to be a larger gap. Over the complete data set the histograms in \autoref{fig_alpha}b draw a similar picture. Instead of left and right the panel is split into inner and outer tire to make the analysis independent of a specific race track. What is more, during a cornering manoeuvre the outer tires have a higher normal force due to load transfer. \\
All professional drivers exceed the optimal slip angle $\alpha^\mathrm{ot}$ at the front axle most of the time. For the outer front the distribution is bimodal for most drivers with a maximum at approx. $-2^\circ$ and $1.5^\circ$. Interestingly, on the inner front there is a more pronounced maximum a $0^\circ$. On the outer rear all drivers clearly avoid slip angles higher than the optimum, resulting in a peak at $-2^\circ$. The distribution on the inner rear has a wider range and reaches slightly in the positive values.
The explanation for not exceeding the optimum on the outer rear axle is stability. Loosing grip on the rear axle would lead to an unstable oversteer and since the outer rear provides most of the grip it is more important to not exceed the optimum there than on the inner rear. On the front however, the drivers intentionally drive at higher slip angles than would be optimal because the generated lateral force stays relatively constant from the optimum towards higher slip angles but has a steeper slope from small values to the optimum. To better control the vehicle and anticipating it's reaction, the drivers stay in the more constant region above the optimum on purpose. \\
Between the professional drivers, some differences can be observed. On the outer and inner front tire, Driver D5 has a more unimodal distribution than the rest. On the inner rear tire, the same holds for Driver D4. The amateur drivers which are included as a reference have a distinctly different distribution. On the front, they have smaller slip angles than the optimum in general, however, instances where they have much higher slip angles than the optimum are also more frequent than for the professional drivers. On the rear axle the situation is similar. Analogical to the slip ratio, the amateur drivers are more cautious not to exceed the optimal slip angles. \\
\begin{figure}
\begin{subfigure}[c]{1\textwidth}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig/alpha_example_area_ano.png}
\end{center}
\subcaption{Exemplary overview about $\alpha$ in the time domain for the Init\xspace, optCog\xspace and optTire\xspace state.}
\end{subfigure}
\begin{subfigure}[c]{1\textwidth}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig/alpha_hist.png}
\end{center}
\subcaption{Distribution of slip angle differences between the Init\xspace and optTire\xspace state over the tires and drivers for all sections where $|gLat| \geq 1.5g$. Negative values indicate that the absolute of $\alpha^\mathrm{in}$ is smaller than the absolute of $\alpha^\mathrm{ot}$. The panel is split into outer and inner tire instead of left and right tire to make it track independent.}
\end{subfigure}
\caption{Both panels show that the drivers often exceed the optimal slip angle $\alpha$ on the front axle, while usually staying below the limit at the rear axle. Especially on the outer rear axle, the drivers avoid overshooting the optimum. Differences between the professional and amateur drivers are clearly visible. }
\label{fig_alpha}
\end{figure}
\subsection{optCog\xspace optimisation}
\autoref{fig_acc_example}a shows an exemplary overview of the acceleration from the optCog\xspace optimiser. It is close to the Init\xspace acceleration in the phases Pure Steer and Throttle Steer. For Trail Brake, the difference is slightly larger and for Pure Brake the optCog\xspace acceleration is significantly higher. The different control states are indicated by the coloured background. \\
A common visualization method in motorsport for accelerations is the 'GG-Diagram'; \autoref{fig_acc_example}b depicts the acceleration in x- and y-direction. It allows an understanding of the physical limits of the vehicle \cite{Kegelman.2018}, assuming it was driven on the dynamic limit. The shape results from the frictional properties of the tires and their kinematic relations to each other (aerodynamic effects like downforce play a role as well). The symmetry of the ellipse is broken for acceleration and deceleration since the acceleration of the vehicle is not always limited by tire grip but also the engine power. In contrast, the brakes are designed to exert enough torque to use the full friction potential. \\
The comparison between the Init\xspace and optCog\xspace state shows large differences under Pure Brake (Decelerate) and slight differences under Trail Brake (between Decelerate and Right or Left). The other areas match closely.\\
We expect two factors to be responsible for the unrealistic high braking accelerations of the optCog\xspace state. First, the optimiser could directly control the tire slip ratio $\kappa$ with the stated torque constraints. Since the optimiser has no information in the temporal domain, it leads to the high frequency changes in $\kappa$ which could not be achieved in reality. Brake system compliance is also not considered due to this approach. Second, the drivers leave a margin to the optimal tire slip ratio to avoid lock-ups as discussed in Section \ref{chap_kappa}.
\begin{figure}
\begin{center}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.8\textwidth]{fig/rho_cog_example_areas_ano.png}
\end{center}
\subcaption{Exemplary overview about the accelerations from the Init\xspace and optCog\xspace state in the time domain as well as the brake, throttle and steering inputs from the driver. The background is coloured with respect to the defined control states.}
\end{subfigure}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.9\textwidth]{fig/gg_cog_example_ano.png}
\end{center}
\subcaption{'GG-Diagram' of the accelerations in the xy-plane for the Init\xspace and optCog\xspace state. Under pure braking (decelerate) there is a significant difference between the two states.}
\end{subfigure}
\end{center}
\caption{optCog\xspace versus Init\xspace state accelerations.}
\label{fig_acc_example}
\end{figure}
\subsection{Scores evaluation}
\autoref{fig_scores_example} shows an overview of the scores and the corresponding predictions in the time domain as well the distributions for the complete data set. As analysed previously, the ratio between the Init\xspace and the optCog\xspace acceleration \ensuremath{S_\mathrm{handling}}\xspace is smaller for Pure Brake and Trail brake.
\autoref{fig_scores_example} b shows the differences given the control state. \ensuremath{S_\mathrm{handling}}\xspace is spread much wider under Pure Brake than for the other three control states. What is more, the center of the distribution is shifted towards the left. For $S_\mathrm{tot}$ the situation is similar only that for Pure Brake $100\%$ are almost never reached and that the Trail Brake score is skewed to the left.
This difference makes the control states, which were originally only defined to simplify the analysis for the drivers, necessary to compare different track sectors to each other. For example a track with more Pure Brake sections would lead to lower overall scores in comparison to a track with fewer of such sections. However, by analysing the scores for each control state individually, this will be compensated, resulting in a track independent benchmark. This is a valuable tool for engineers and race drivers which would not be possible without the presented methods.
\begin{figure}
\begin{center}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.96\textwidth]{fig/scores_example_area_ano.png}
\end{center}
\subcaption{Exemplary overview about the defined scores and the machine learning predictions in the time domain.}
\end{subfigure}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.96\textwidth]{fig/scores_hist_pred.png}
\end{center}
\subcaption{Scores overview for the complete data set over the control states. Scores smaller than $40\%$ are grouped in the leftmost bin. Pure Brake and Trail Brake have a higher density in the lower scores. }
\end{subfigure}
\end{center}
\caption{Under braking the distribution of all scores and especially \ensuremath{S_\mathrm{handling}}\xspace is skewed towards smaller values.}
\label{fig_scores_example}
\end{figure}
\subsection{Driving style evaluations}
We provide an exemplary evaluation of the motorsport dataset to show how the presented methods are used and to give an insight into the differences between professional race drivers when controlling a vehicle at its dynamic limit.
\subsubsection{Traditional lap-based metrics}
\begin{figure}
\begin{center}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.88\textwidth]{fig/all_driver/track_overview.png}
\end{center}
\subcaption{Professional drivers}
\end{subfigure}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.88\textwidth]{fig/bob/track_overview.png}
\end{center}
\subcaption{Amateur versus professional drivers}
\end{subfigure}
\end{center}
\caption{Important metrics on a per-lap basis. All metrics are normed, the units are deviations from the median in $\%$. $tLap$ is the lap time, $v_\mathrm{avg}$ the average speed, $c_\mathrm{sum}$ the sum of the trajectory's curvature and $sLap$ the lap distance. There are significant differences between professional and amateur drivers.}
\label{fig_track_metrics}
\end{figure}
It showed that the most useful representation are boxplots depicting certain characteristics from the distribution of scores. Of interest are the first quartile, the median and the third quartile, in other words the $25,~50$ and $75\%$ quantile. At first, traditional metrics are compared in \autoref{fig_track_metrics}a on a per-lap basis. The lap time $t_\mathrm{lap}$ is probably the most important criteria to evaluate performance, however it provides no insights on how it is achieved. Closely related are the average velocity $v_\mathrm{avg}$ and the lap distance $s_{lap}$. Additionally, the integral over the absolute curvature $c_\mathrm{sum}$ which is obtained from the trajectory, is shown.
Looking at the medians, Driver D7 achieved the best laptimes, followed by driver D6. However, the variance is much higher for D7 according to the high spread of the box and the whiskers. This example shows already that the question 'which driver is the best' cannot be answered in general but depends on the circumstances. Differences are also present in $s_\mathrm{lap}$. Driver D2 and D5 chose overall a longer driving line than the rest. The average speed results from lap time and lap distance naturally.
Although powerful, the disadvantage of these set of metrics is twofold. First, the most important metric $t_\mathrm{lap}$ can be used on a lap basis only. It is possible to use sector times by dividing the tracks in multiple parts, however, there is a limit on how short they can reasonably be. Second, these metrics do not always show where and why lap time differences occur.
\subsubsection{Professional race drivers}
We show a visualization that can be used for a single corner, but also for a data set spanning multiple tracks as presented here.
Figure \ref{fig_overall:scores1} depicts the scores for pure brake and pure steer. \\
In the following we exemplary compare the drivers D5 and D6. Under pure braking in \autoref{fig_overall:scores1}a, D6 has a higher total score \ensuremath{S_\mathrm{tot}}\xspace (Remark 1). We see that the difference arises because of a higher handling score \ensuremath{S_\mathrm{handling}}\xspace, i.e. D6 is able to take the vehicle closer to its dynamic limit (Remark 2). Looking at \ensuremath{S_\mathrm{veh-traj}}\xspace, D5 has a slightly higher score, indicating that the optimised vehicle state is closer to the ideal tire state. To sum up, the higher \ensuremath{S_\mathrm{veh-traj}}\xspace score cannot compensate the lower \ensuremath{S_\mathrm{handling}}\xspace score. \\
Under pure steer D5 has a lower \ensuremath{S_\mathrm{tot}}\xspace score, as shown in \autoref{fig_overall:scores1}b (Remark 3). Contrary to the difference in pure brake, the difference originates from the \ensuremath{S_\mathrm{veh-traj}}\xspace score (Remark 4), i.e. D5 has a less advantageous vehicle setup or driving line. The handling scores for both drivers are comparable. \\
In conclusion, D6 is overall better than D5 in pure brake and pure steer. The fact that D6 performed better than D5 in the analysed data set is supported by the better lap times and average speeds of D6 which are shown in \autoref{fig_track_metrics}a. However, in contrast to the lap time difference, the score analysis of this work allows much more detailed conclusions: D5 has to improve the handling skills under pure brake or needs a vehicle setup that has a better driveability.
\begin{figure}
\begin{center}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.88\textwidth]{fig/all_driver/PureBrake_an.png}
\end{center}
\end{subfigure}
\begin{subfigure}[c]{0.8\textwidth}
\begin{center}
\includegraphics[width=0.88\textwidth]{fig/all_driver/PureSteer_an.png}
\end{center}
\end{subfigure}
\end{center}
\caption{Exemplary score overview for the pure brake and pure steer states. The remarks show differences between the drivers D5 and D6.}
\label{fig_overall:scores1}
\end{figure}
\subsubsection{Amateur Drivers}
The differences in all of these scores were subtle for the professional drivers, which is natural given that they have been training for years to achieve the maximal performance.
To test our proposed method further and make it more interesting for a broader audience, we examine also amateur drivers with road car driving skills but no motorsport race experience. For that purpose, they drove in the same simulator environment as the professional drivers with a comparable vehicle setup on the track T2. In \autoref{fig_track_metrics}b significant differences between the two classes can be observed. For the amateurs, the lap time $t_\mathrm{lap}$ is 10 to $30\%$ higher and the average speed $v_\mathrm{avg}$ 10 to $20\%$ slower. The lap distance $s_\mathrm{lap}$ is shorter on average, which is a consequence of the much lower speed, allowing smaller cornering radii. D9A achieved quicker lap times than D8A, however, D8A was more consistent, especially in lap distance.
Unsurprisingly, significant differences in \ensuremath{S_\mathrm{tot}}\xspace are observed. Comparing the other two scores, the main difference comes from \ensuremath{S_\mathrm{handling}}\xspace. Since all drivers drove a comparable vehicle setup, the differences can be clearly attributed to the trajectory.\\
Instead of providing the diagrams for all control states we summarized the information in \autoref{fig_overall_bob}b. The mean difference between the professional and amateur drivers is depicted for every control state. Most of the amateurs' performance is lost under braking. Almost all of that is due to a worse handling score. The amateur drivers are worse at controlling the vehicle and we observed that they did not exploit the aerodynamic abilities of the vehicle. Because of high downforce, it is possible to brake much more at high speeds than at lower speeds.
The highest difference in \ensuremath{S_\mathrm{veh-traj}}\xspace is observed for throttle steer. Since the vehicle setups are equal, the reason has to be a difference in driving line. What is more, the difference in \ensuremath{S_\mathrm{handling}}\xspace is the second highest leading to the highest \ensuremath{S_\mathrm{tot}}\xspace difference after pure brake. \\
To sum up, the professional drivers' driving line is superior to the amateurs. However, the difference is much larger for the handling score. Most of the overall difference in \ensuremath{S_\mathrm{tot}}\xspace is a result of the much lower \ensuremath{S_\mathrm{handling}}\xspace score and not so much \ensuremath{S_\mathrm{veh-traj}}\xspace. For the amateur drivers to improve they would need to practice vehicle handling to be able to drive a higher speed along their chosen trajectory.
\begin{figure}
\begin{center}
\begin{subfigure}[c]{0.9\textwidth}
\begin{center}
\includegraphics[width=0.78\textwidth]{fig/bob/overall.png}
\end{center}
\caption{Score overview for track T2 with amateur and professional drivers. In the first row are the score obtained by the optimisation procedure and in the second row the scores estimated by the machine learning predictor. }
\end{subfigure}
\begin{subfigure}[c]{0.9\textwidth}
\begin{center}
\includegraphics[width=0.78\textwidth]{fig/bob/Diff_Pro_Am.png}
\end{center}
\caption{Difference between professional and amateur drivers for the control states. As expected the professional drivers have higher scores in all cases, therefore the difference is always positive. The difference in \ensuremath{S_\mathrm{tot}}\xspace comes mainly from \ensuremath{S_\mathrm{handling}}\xspace. In the second row are again the predictions of the machine learning model.}
\end{subfigure}
\end{center}
\caption{Professional versus amateur drivers. }
\label{fig_overall_bob}
\end{figure}
\subsection{Machine learning predictor evaluation}
We already showed in Section \ref{chap_model_sel} and \autoref{fig_lstm_res}b that the predictions of the machine learning model have a RMSE in the area of $2\%$ for \ensuremath{S_\mathrm{handling}}\xspace and \ensuremath{S_\mathrm{tot}}\xspace and around $3\%$ for \ensuremath{S_\mathrm{veh-traj}}\xspace on the considered data set with the professional drivers. Consequently, the predictions are in good accordance with the dynamic limit optimisation method while greatly decreasing computation time. This is supported by \autoref{fig_scores_example}b where the score distributions for reference and prediction are very similar for the whole data set.\\
The prediction also shows good results for the amateurs who were not part of the training set and whose driving styles were very different compared to the training data. In this case the RMSE is approximately $3\%$ for \ensuremath{S_\mathrm{handling}}\xspace and \ensuremath{S_\mathrm{tot}}\xspace and $5\%$ for \ensuremath{S_\mathrm{veh-traj}}\xspace.
Comparing the optimisation scores in the first row of \autoref{fig_overall_bob}b and the predictions in the second row they are very similar. Interpretation and analysis of both methods would lead to the same conclusion. What is more the difference between professional and amateur drivers is also very similar for both methods (\autoref{fig_overall_bob}c). \\
In terms of computational speed, the machine learning predictor needed $12$ seconds for inference (i.e. processing data through a trained neural network) on 1226 laps on a \emph{NVIDIA\textregistered Tesla\textregistered V100} GPU, which is a specialized deep learning graphics card. High-end consumer graphic cards will achieve a performance in the same order of magnitude. In contrast, the optimisation methods needed over $68$ hours for the same task on a computer with two 10-core \emph{Intel\textregistered Xeon\textregistered Silver 4144} CPUs. Due to the different hardware and architecture (machine learning on the GPU, optimisation the CPU), the comparison might not be completely fair, however, the results show the massive improvement in computation time very clearly. \\
In conclusion, the robustness of the machine learning model and the low prediction error allows to replace the computational expensive optimisation methods with the machine learning predictor for our specific use case. Given the very fast computation time, a real time execution is also possible.
\section{Conclusion}
The presented methods showed their practical use for analysing drivers on a race track. However, it could be applied to all scenarios where the goal is to handle a vehicle at the limit. For road car driving, this can be the case for evasion manoeuvres or under slippery conditions. What is more, control systems for autonomous vehicles could be judged by the presented scores in a simulated environment and compared to professional drivers.
However, what sets professional race drivers apart is their ability to not only handle a vehicles in a stable environment, but also in the real world under changing conditions. As a next step, we therefore propose a simulator study whereby the amount of grip is varied continuously. Having at hand the here presented driver evaluation methods, it could be investigated how the participants react to the grip changes and how long the adaptation time will be until the best possible performance is reached. We hope that further insight on how human experts achieve such high performance at controlling a vehicle at the limit helps the development of advanced autonomous control systems in the future.
\section*{Acknowledgement}
This work was generously supported by BMW AG.
\bibliographystyle{elsarticle-num-names}
|
\section{Introduction}
\label{introduction}
Mobile robots have been widely applied in various fields including but not limited to family service, logistics, and search-and-rescue applications. Motion planning is one of the key technologies for mobile robots to achieve full autonomy in these scenarios and plays an essential role in generating safe, smooth, and efficient motions \cite{palmieri2019dispertio,li2020efficient,le2019multi}.
\subsection{Path Searching}
Sampling-based planning and grid-based planning are two popular path searching approaches. Sampling-based planning algorithms such as the famous probabilistic roadmap (PRM) \cite{kavraki1996probabilistic} and rapidly-exploring random tree (RRT) \cite{lavalle2001randomized} have gained the popularity for the capability of efficiently searching in the configuration space (C-space). However, they are limited by completeness and optimality, and even some excellent variants such as RRT* \cite{karaman2011sampling} can only guarantee asymptotic optimality. In this letter, we focus on grid-based planning. Grid-based planning overlays a hyper-grid on the C-space and assumes each configuration is identified with the grid-cell center. Then, search algorithms such as A* \cite{hart1968formal} are used to find a path from the start to the goal. Grid-based planning can always find a resolution-optimal path if it exists in the discrete search space. However, these approaches depend on the space discretization and do not perform well as the environment dimension increases. In addition, most path searching algorithms only consider the path length in the cost function, which often results in the searched path being close to obstacles and poor safety.
\begin{figure}[t]
\vspace{0.2cm}
\centering
\includegraphics[scale=0.25]{fig/platform.pdf}
\caption{Our experimental mobile robot SUMMIT-XL HL is equipped with a Velodyne VLP-16 LiDAR, an Intel Realsense D435i depth camera, and an Intel NUC computer. Experiments are detailed in Section \ref{experiments} and the video is available at \url{https://youtu.be/V-S8cS2zv_U}.}
\label{fig:platform}
\vspace{-0.3cm}
\end{figure}
\subsection{Path Smoothing}
The path obtained by path searching approaches usually fails to meet the smoothness requirement for robot navigation and needs further smoothing \cite{ravankar2018path}. Two types of path smoothing approaches are investigated in this letter, namely, soft-constrained approaches and hard-constrained approaches.
\subsubsection{Soft-constrained Approaches}
These approaches formulate path smoothing as a non-linear unconstrained optimization problem, wherein the constraints are considered in the optimization objective in the form of penalty cost terms. Generally, the smoothness of the path and the clearance to obstacles are both taken into account. Then, gradient-based optimization algorithms are employed to solve the problem. Soft-constrained approaches utilize gradient information to push the path far from obstacles and can obtain a smooth path with a reasonable distance from obstacles. However, it is difficult for these approaches to guarantee the optimized path strictly satisfies the constraints. In addition, the optimization objective of soft-constrained approaches contains convex and non-convex terms, making these approaches suffer from local optima and time-consuming.
\subsubsection{Hard-constrained Approaches}
These approaches formulate path smoothing as a non-linear constrained optimization problem and can obtain a path that theoretically satisfies the constraints. The optimization objective usually considers the smoothness or the length of the path and thus is convex. This convexity allows the problem to be solved efficiently. However, hard-constrained approaches treat all free space equally, namely, distance from the free space to obstacles is ignored. As a result, the optimized path may be close to obstacles and is dangerous for robot navigation.
\subsection{Contributions}
Motivated by the aforementioned limitations of existing approaches, an efficient motion planning approach called G\ensuremath{^2}VD planner is newly proposed for mobile robots. Specifically, a grid-based generalized Voronoi diagram (G\ensuremath{^2}VD) is utilized to aid path searching and path smoothing to achieve better performance.
\subsubsection{Path Searching}
Given the start and the goal, we first employ the A* search to find the shortest grid path in a grid-based GVD. This shortest Voronoi path contains the topological information of the search direction and provides rough guidance for the subsequent fine search. Along the Voronoi grid path, a region called \emph{Voronoi corridor} is constructed by adding a bounding box to each path pixel. Furthermore, the cost function of path searching is redesigned based on a potential called \emph{Voronoi field} to make the searched path keep a reasonable distance from obstacles. Finally, the A* search combined with motion primitives is utilized to finely search a kinematically feasible path within the Voronoi corridor. Different from the original state lattice-based path planner \cite{likhachev2009planning} that takes the whole grid map as the search space, the proposed approach reduces the search space to the Voronoi corridor, thus considerable time is saved for path searching. In addition, the reasonable clearance to obstacles provides sufficient optimization margin for further path smoothing.
\subsubsection{Path Smoothing}
Taking the path searched above as the reference path, an efficient quadratic programming (QP)-based path smoothing approach is proposed, wherein the smoothness of the path and the deviation from the reference path are considered in the optimization objective. Our goal is to obtain a smooth path while minimizing the deviation between the optimized path and the reference path. Because the reference path searched above has a certain distance to obstacles, the clearance to obstacles is indirectly considered in the proposed approach, and the final smoothed path will not be close to obstacles, which is usually difficult for hard-constrained approaches. This is also the second reason that we introduce the Voronoi field to improve the path clearance of the searched path, in addition to providing wider optimization margin for path smoothing.
To summarize, the main contributions of this letter are as follows:
\begin{enumerate}[\hspace{1em}1)]
\item A new state lattice-based path planner is proposed, in which the search space is reduced to a Voronoi corridor to improve the search efficiency, along with a Voronoi potential constructed to make the searched path keep a reasonable distance from obstacles to provide sufficient optimization margin for further path smoothing.
\item An efficient QP-based path smoothing approach is presented, wherein the clearance to obstacles is considered in the form of the penalty of the deviation from the safe reference path to improve the path clearance of hard-constrained path smoothing approaches.
\item Autonomous navigation is realized in large-scale outdoor environments. Extensive simulation and experimental evaluations are presented to validate the effectiveness of the proposed approach. We will release the source code to the robotics community.
\end{enumerate}
The rest of this letter is organized as follows. We first review the related work in Section \ref{relatedwork}. The proposed G\ensuremath{^2}VD planner is detailed in Section \ref{approach}. Section \ref{implementationdetails} provides some implementation details. The results of simulations and experiments are presented in Sections \ref{simulations} and \ref{experiments}, respectively. Finally, this letter is concluded in Section \ref{conclusion}.
\section{Related work}
\label{relatedwork}
Grid-based planning obtains the resolution-optimal path by discretizing the C-space first and then using graph search algorithms to find the path. However, original grid-based planning approaches only visit the centers of grid cells and produce piecewise-linear paths that do not generally satisfy the kinodynamic constraints of the robot. To address this problem, Pivtoraiko \emph{et al.} \cite{pivtoraiko2009differentially} propose the state lattice approach for graph construction. In particular, the connectivity between two nodes in the graph is built from a pre-designed motion primitive that fulfills the kinodynamic constraints of the robot. Based on the work \cite{pivtoraiko2009differentially}, Likhachev \emph{et al.} present the state lattice-based path planner by combining AD* search with motion primitives \cite{likhachev2009planning}, which has been successfully applied in the DARPA Urban Challenge \cite{ferguson2008motion}. In this letter, a new state lattice-based path planner is proposed, wherein the search space is reduced to a Voronoi corridor derived based on a generalized Voronoi diagram (GVD) to further improve the search efficiency.
The use of GVDs has long been proposed in the context of robot motion planning. In \cite{choset2000sensor}, Choset and Burdick use the GVD to derive skeletons of the free space and then search on the graph. However, the shortest Voronoi path may be far from the actual optimal path. In \cite{ziegler2008navigating}, Ziegler \emph{et al.} utilize the GVD to inform the A* search by constructing a heuristic where the cost of a search state is the sum of the straight-line path to the closest Voronoi edge and the shortest path along the GVD edges. However, the resulting heuristic cannot be guaranteed to be admissible since the cost of the shortest Voronoi path may be greater than that of the actual shortest path. In \cite{dolgov2010path}, Dolgov \emph{et al.} present a potential field based on the GVD for path smoothing. This potential allows precise navigation in narrow passages while also effectively pushing the robot away from obstacles in wider areas. Inspired by the Voronoi field, we redesign the cost function of the state lattice-based path planner to obtain a path with a reasonable distance from obstacles.
The path obtained by path searching approaches usually needs further smoothing. In \cite{dolgov2010path}, Dolgov \emph{et al.} present a conjugate gradient (CG)-based path smoothing approach to smooth the path generated by the hybrid A* algorithm. In \cite{wen2020effmop}, Wen \emph{et al.} propose a gradient-based local path smoothing approach for mobile robots, wherein the sparse-banded system structure of the underlying optimization problem is fully exploited to efficiently solve the problem. The above soft-constrained approaches utilize gradient information to push the path far from obstacles and can obtain a path with better safety. However, it is difficult for these approaches to guarantee the optimized path strictly satisfies the constraints. Recently, Zhou \emph{et al.} propose a dual-loop iterative anchoring path smoothing (DL-IAPS) approach for autonomous driving \cite{zhou2020autonomous}, in which the nonlinear curvature constraint is linearized and sequential convex programming (SCP) is used to efficiently solve the path optimization problem. Such a hard-constrained approach can theoretically guarantee that the obtained path strictly satisfies the constraints. However, distance from the free space to obstacles is ignored, which often results in the optimized path being close to obstacles and poor safety. In this letter, a new QP-based path smoothing approach is proposed, in which the clearance to obstacles is indirectly considered in the form of the penalty of the deviation from the safe reference path to improve the path clearance of hard-constrained approaches.
\section{G\ensuremath{^2}VD Planner}
\label{approach}
\begin{figure}[t]
\vspace{0.2cm}
\centering
\includegraphics[scale=0.31]{fig/framework.pdf}
\caption{Flow chart of the proposed three-layer motion planning framework.}
\label{fig:frameork}
\vspace{-0.4cm}
\end{figure}
In this letter, a three-layer motion planning framework called G\ensuremath{^2}VD planner is carefully designed, which consists of path searching, path smoothing, and velocity planning, as shown in Fig. \ref{fig:frameork}. The path searching module is utilized to provide a safe reference path for the robot, and the path smoothing module combined with our previously proposed time-optimal velocity planning \cite{shen2017essential} is employed to generate safe, smooth, and efficient motion commands. In this section, we will detail the newly proposed path searching and path smoothing modules.
\subsection{Grid-based Generalized Voronoi Diagrams}
The GVD is defined as the set of points in the free space to which the two closest obstacles have the same distance \cite{choset2000sensor}. In this letter, an efficient, incrementally updatable GVD construction algorithm presented in \cite{lau2013efficient} is utilized to convert the occupancy grid map to a GVD in discrete form. This algorithm \cite{lau2013efficient} employs a wavefront algorithm to perform distance transforms. Therefore, every cell in the grid-based GVD stores not only the state of whether it belongs to the GVD but also the minimum Euclidean distance to the closest obstacle cell. Fig. \ref{fig:voronoi} illustrates the grid-based GVD of an office-like environment, in which the GVD edge pixels are colored in blue and red.
\begin{figure}[t]
\vspace{0.2cm}
\centering
\includegraphics[scale=0.3]{fig/gvd.pdf}
\caption{Grid-based GVD of an office-like environment. The GVD edge pixels are colored in blue and red, and the searched shortest Voronoi path is colored in red. The gray region surrounding the Voronoi path denotes the Voronoi corridor.}
\label{fig:voronoi}
\vspace{-0.2cm}
\end{figure}
\subsection{Voronoi Corridor}
After obtaining the grid-based GVD, breadth-first search is employed to find the cells $ C_1 $ and $ C_2 $ closest to the start cell $ C_{start} $ and the goal cell $ C_{goal} $ on the GVD, respectively. And the shortest grid paths $ L_1 $ from $ C_{start} $ to $ C_1 $ and $ L_2 $ from $ C_2 $ to $ C_{goal} $ are also computed in this process. Then, the A* search is utilized to search the shortest grid path $ L_3 $ from $ C_1 $ to $ C_2 $ along the GVD edge pixels. The final shortest Voronoi path from $ C_{start} $ to $ C_{goal} $ is composed of $ L_1 $, $ L_3 $, and $ L_2 $, as shown by the red path in Fig. \ref{fig:voronoi}.
Based on the shortest Voronoi grid path searched above, a region called Voronoi corridor is constructed as follows. For every cell $ C_k $ in the shortest Voronoi grid path, the minimum distance to the closest obstacle cell $ d_k $ is retrieved from the grid-based GVD. Then, a square bounding box with a side length of $ 2d_k $ is centered on $ C_k $. The obstacle-free cells covered by all bounding boxes make up the Voronoi corridor, as illustrated by the gray region in Fig. \ref{fig:voronoi}.
\subsection{Path Searching}
\label{pathsearching}
The shortest Voronoi path is far from the actual optimal path, and its piecewise-linear form also does not satisfy the kinodynamic constraints of the robot. To address these issues, a new state lattice-based path planner is proposed to perform fine path searching within the Voronoi corridor.
A typical state lattice-based path planner consists of two parts, namely, graph construction and graph search \cite{likhachev2009planning}. As for graph construction, the 3-D search space $ \left(x, y, \theta\right) $ is discretized first, where $ \left(x, y\right) $ denotes the position of the robot in the world and $ \theta $ represents the heading of the robot. In particular, the orientation space is discretized into 16 angles. Furthermore, the connectivity between two states in the graph is built from motion primitives which fulfill the kinematic constraints of the robot. In this letter, a quintic B\'{e}zier curve-based path generation approach described in \cite{zhang2018multilevel} is employed to generate motion primitives offline. A motion primitive $ \gamma\left(s, s'\right) $ consists of a sequence of internal robot poses when moving from state $ s $ to state $ s' $. As for graph search, the standard A* search is employed, where the 2-D heuristic $ h_{2D} $ proposed in \cite{likhachev2009planning} is utilized to guide the A* search away from those areas with dead-ends. This heuristic is constructed by computing the costs of shortest 2-D grid paths from the goal cell to other cells in the search space through Dijkstra's algorithm. Different from the original state lattice-based path planner \cite{likhachev2009planning} that takes the whole grid map as the search space, we reduce the search space to the Voronoi corridor to further improve the search efficiency. In addition, every time the grid map is updated, $ h_{2D} $ needs recalculation before performing path searching. Since the search space of the newly proposed state lattice-based path planner is reduced to the Voronoi corridor, Dijkstra's search is accordingly restricted to only compute the costs of shortest paths from the goal cell to those cells that are within the Voronoi corridor. Therefore, considerable time is also saved for the recalculation of $ h_{2D} $.
Although the path searched above is optimal in the search space, it may be very close to obstacles. The ideal path is to keep a reasonable distance from obstacles to provide sufficient optimization margin for subsequent path smoothing. To this end, inspired by the Voronoi field presented in \cite{dolgov2010path}, we design the following potential function:
\begin{equation}
\begin{array}{l}
{\rho _V}\left( {x,y} \right)\\
= \left\{ {\begin{aligned}
&\frac{{{d_{\mathcal{V}}}\left( {x,y} \right)}}{{{d_{\mathcal{O}}}\left( {x,y} \right) + {d_{\mathcal{V}}}\left( {x,y} \right)}}\frac{{{{\left( {{d_{\mathcal{O}}} - d_{\mathcal{O}}^{\min }} \right)}^2}}}{{{{\left( {d_{\mathcal{O}}^{\min }} \right)}^2}}} \ &{{d_{\mathcal{O}}} \le d_{\mathcal{O}}^{\min }}\\
&0 \ &{{d_{\mathcal{O}}} > d_{\mathcal{O}}^{\min }},
\end{aligned}} \right.
\end{array}
\end{equation}
where $ d_{\mathcal{V}}\left(x, y\right) $ and $ d_{\mathcal{O}}\left(x, y\right) $ denote the distances from the given path vertex $ \left(x, y\right) $ to the closest Voronoi edge and the closest obstacle, respectively. $ d_{\mathcal{O}}^{\min} $ is a threshold specifying the minimum safety distance to obstacles.
According to \cite{dolgov2010path}, this potential function has the following properties:
\begin{enumerate}[\hspace{1em}i)]
\item $ {\rho _V}\left( {x,y} \right) \in [0,1] $ and is continuous on $ \left(x, y\right) $ since we cannot simultaneously have $ d_{\mathcal{V}} = 0 $ and $ d_{\mathcal{O}} = 0 $;
\item $ {\rho _V}\left( {x,y} \right) $ reaches its maximum only when $ \left(x, y\right) $ is within obstacle areas;
\item $ {\rho _V}\left( {x,y} \right) $ reaches its minimum only when $ \left(x, y\right) $ is on the edges of the GVD or the distance from $ \left(x, y\right) $ to the closest obstacle is greater than $ d_{\mathcal{O}}^{\min} $.
\end{enumerate}
It is noteworthy that the potential cost is set to zero when the distance to the closest obstacle is greater than the safety threshold ($ {{d_{\mathcal{O}}} > d_{\mathcal{O}}^{\min }} $). The reason for this design is as follows. Our goal is to make the searched path close to Voronoi edges through the Voronoi field so that the obtained path can keep an appropriate distance from obstacles and provide sufficient optimization margin for subsequent path smoothing. However, if the environment is wider and the Voronoi edge is far from the obstacles on both sides, which may make the searched path far away from the optimal path. Therefore, we set a safety distance threshold $ d_{\mathcal{O}}^{\min } $ and regard the area where the distance to the closest obstacle exceeds $ d_{\mathcal{O}}^{\min } $ as a safe region. In this way, the searched path can keep a certain distance from obstacles and will not be far away from the optimal path through the Voronoi field.
Based on the above Voronoi field, the cost of motion primitives is defined as follows. For the sake of computational efficiency, we broadly follow the work \cite{dolgov2010path} and temporarily assume that the robot travels at constant linear and angular velocities. If a motion primitive $ \gamma\left(s, s'\right) $ collides with obstacles, the cost $ g\left( \gamma\left(s, s'\right) \right) $ is set to infinity. Otherwise, the cost of this motion primitive is defined as
\begin{equation}
g\left( {\gamma \left( {s,s'} \right)} \right) = t\left( {s,s'} \right) \cdot \left( {\mathop {\max }\limits_{(x,y) \in \mathcal{Q}} {\rho _V}\left( {x,y} \right) + 1} \right),
\end{equation}
where $ t\left( {s,s'} \right) $ is the travel time spent on $ \gamma\left(s, s'\right) $ assuming uniform motion, $ \mathcal{Q} $ is the set of 2-D cells covered by the robot when moving from state $ s $ to state $ s' $. Intuitively, the Voronoi field penalizes slightly more those actions for which the robot traverses high potential cost areas (e.g., obstacles) and makes the searched path as close as possible to the Voronoi edges or keep a certain distance from obstacles. In addition, the 2-D heuristic $ h_{2D} $ is also multiplied by the potential costs to be consistent with the cost definition of motion primitives.
\begin{figure}[t]
\vspace{0.2cm}
\centering
\subfigure[Path obtained by the state lattice-based path planner \cite{likhachev2009planning}.]{\includegraphics[scale=0.25]{fig/lattice_path_1.pdf}}
\centering
\subfigure[Path obtained by the proposed path planner.]{\includegraphics[scale=0.25]{fig/lattice_path_2.pdf}}
\caption{Illustration of path searching results.}
\label{fig:latticepath}
\vspace{-0.2cm}
\end{figure}
It should be noted that the Voronoi potential field in \cite{dolgov2010path} is used for path smoothing and the region of interest is the whole map. While the search space of the newly proposed state lattice-based path planner is reduced to the Voronoi corridor, and we only need to compute the Voronoi potential costs within the Voronoi corridor. In particular, an efficient distance transform algorithm presented in \cite{felzenszwalb2012distance} is employed to compute the Euclidean distance to the closest Voronoi edge for those cells within the Voronoi corridor. Figs. \ref{fig:latticepath}(a) and \ref{fig:latticepath}(b) illustrate the paths searched by the original state lattice-based path planner \cite{likhachev2009planning} and the newly proposed path planner, respectively. Compared with the original state lattice-based path planner, the path obtained by the proposed path planner has a certain distance from obstacles and can provide wider optimization margin (green bounding boxes) for further path smoothing, which will be detailed in Section \ref{pathsmoothing}.
\begin{figure}[t]
\vspace{0.2cm}
\centering
\includegraphics[scale=0.45]{fig/formulation.pdf}
\caption{Illustration of path smoothing formulation. $ \mathbf{x}_{i-1} $, $ \mathbf{x}_i $, and $ \mathbf{x}_{i+1} $ denote three successive path vertices. $ d_{i+1} $ is the distance from $ \mathbf{x}_{i+1} $ to the closest obstacle. The black blocks represent obstacles.}
\label{fig:pathsmoothing}
\vspace{-0.4cm}
\end{figure}
\subsection{Path Smoothing}
\label{pathsmoothing}
The path obtained by the state lattice-based path planner is kinematically feasible, but it is still piecewise-linear and not suitable for velocity planning. Therefore, an efficient QP-based path smoothing approach combined with cubic spline interpolation is employed to further smooth the path.
The input of path smoothing is several reference path vertices obtained by sampling in the path generated by the state lattice-based path planner with a fixed interval. In consideration of the smoothness of the path and the deviation from the reference path, a convex QP-based path smoothing formulation is defined as
\begin{equation}
\mathop {\min }\limits_{\mathbf{x}} \, {\omega _s}\sum\limits_{i = 2}^{n - 1} {{{\left\| {{\mathbf{x}_{i + 1}} - 2{\mathbf{x}_i} + {\mathbf{x}_{i - 1}}} \right\|}^2}} + {\omega _r}\sum\limits_{i = 1}^n {{{\left\| {{\mathbf{x}_i} - {\mathbf{x}_{{i_{ref}}}}} \right\|}^2}},
\label{eq3}
\end{equation}
subject to:
\begin{subequations}
\begin{align}
& \mathbf{x}_{1}=\mathbf{x}_{1_{ref}} \text { and } \mathbf{x}_{n}=\mathbf{x}_{n_{ref}}, \\
& \mathbf{x}_{i} \in \mathcal{B}_{i}, \text { for } i=2, \ldots, n-1. \label{eq4b}
\end{align}
\end{subequations}
Here, $ {\mathbf{x}} = {[{\mathbf{x}}_1^{\mathrm{T}}\;{\mathbf{x}}_2^{\mathrm{T}}\; \ldots \;{\mathbf{x}}_n^{\mathrm{T}}]^{\mathrm{T}}} $ is a $ 2n $-dimensional parameter vector, where $ \mathbf{x}_i = {\left( {x_i}, {y_i} \right)}^{\mathrm{T}}, 1 \le i \le n $ denotes the world coordinates of a path vertex. $ \mathbf{x}_{i_{ref}} = \left(x_{i_{ref}}, y_{i_{ref}}\right)^{\mathrm{T}}, 1 \le i \le n $ is the corresponding reference path vertex of $ \mathbf{x}_i $, and $ \omega_s $ and $ \omega_r $ are the weights of cost terms. $ \mathcal{B}_{i} $ is a state bubble constraining the feasible region of the path vertex $ \mathbf{x}_i $. In this letter, $ \mathcal{B}_{i} $ is approximated as an inscribed square of a circle centered on $ \mathbf{x}_{i_{ref}} $ with the radius of the distance $ d_i $ from $ \mathbf{x}_{i_{ref}} $ to the closest obstacle, as shown in Fig. \ref{fig:pathsmoothing}. Taking the circumscribed radius $ r_c $ of the robot into account, Eq. \eqref{eq4b} is approximated as:
\begin{subequations}
\begin{align}
& x_{i_{ref}} - b_i \le x_i \le x_{i_{ref}} + b_i, \\
& y_{i_{ref}} - b_i \le y_i \le y_{i_{ref}} + b_i,
\end{align}
\end{subequations}
where
\begin{equation}
b_i =
\left\{
\begin{aligned}
&\frac{{\sqrt 2 }}{2}{d_i} - {r_c} \ &\frac{{\sqrt 2 }}{2}{d_i} > {r_c}\\
&0 \ &\frac{{\sqrt 2 }}{2}{d_i} \le {r_c}.
\end{aligned}
\right.
\end{equation}
\begin{figure}[t]
\vspace{0.2cm}
\centering
\subfigure[]{\includegraphics[scale=0.17]{fig/local_smooth_path_1.pdf}}
\hspace{0.2cm}
\centering
\subfigure[]{\includegraphics[scale=0.17]{fig/local_smooth_path_2.pdf}}
\centering
\subfigure[]{\includegraphics[scale=0.17]{fig/local_smooth_path_3.pdf}}
\caption{Illustration of path smoothing results. Green circles denote the circumscribed radius of the robot. (a) Smoothed path with the path shown in Fig. \ref{fig:latticepath}(a) as input. (b) Smoothed path with the path shown in Fig. \ref{fig:latticepath}(b) as input but without penalizing the deviation from the reference path. (c) Smoothed path with the path shown in Fig. \ref{fig:latticepath}(b) as input and penalizing the deviation from the safe reference path.}
\label{fig:smoothpath}
\vspace{-0.4cm}
\end{figure}
The first term in Eq. \eqref{eq3} is a measure of the path smoothness. The cost function $ {{\mathbf{x}_{i + 1}} - 2{\mathbf{x}_i} + {\mathbf{x}_{i - 1}}} $ can be rewritten as $ \left( {{\mathbf{x}_{i + 1}} - {\mathbf{x}_i}} \right) + \left( {{\mathbf{x}_{i - 1}} - {\mathbf{x}_i}} \right) $. As shown in Fig. \ref{fig:pathsmoothing}, from a physical point of view, this cost term treats the path as a series spring system, where $ {{\mathbf{x}_{i + 1}} - {\mathbf{x}_i}} $ and $ {{\mathbf{x}_{i - 1}} - {\mathbf{x}_i}} $ are the forces on the two springs connecting the vertices $ {\mathbf{x}_{i + 1}} $, $ {\mathbf{x}_i} $ and $ {\mathbf{x}_{i - 1}} $, $ {\mathbf{x}_i} $, respectively. If the forces $ {{\mathbf{x}_{i + 1}} - {\mathbf{x}_i}} $ and $ {{\mathbf{x}_{i - 1}} - {\mathbf{x}_i}} $ are equal in size and opposite in direction, the resultant force $ \left( {{\mathbf{x}_{i + 1}} - {\mathbf{x}_i}} \right) + \left( {{\mathbf{x}_{i - 1}} - {\mathbf{x}_i}} \right) $ is a zero vector and the norm is minimum. If all the resultant forces are zero vectors, all the vertices would uniformly distribute in a straight line, and the path is ideally smooth.
The second term in Eq. \eqref{eq3} is used to penalize the deviation from the original safe reference path. As mentioned before, the hard-constrained formulation does not explicitly consider the path clearance, namely, distance from the free space to obstacles is ignored. As a result, the optimized path may be close to obstacles. To address this issue, the penalty of the deviation from the original reference path is introduced in the optimization objective. Our goal is to obtain a smooth path while minimizing the deviation between the optimized path and the reference path. Because the reference path searched by the proposed state lattice-based path planner has certain clearance to obstacles, the safety of the path is indirectly considered in the optimization formulation, and the final optimized path will not be close to obstacles. This is also the second reason that we introduce the Voronoi field to make the searched path keep a reasonable distance from obstacles in Section \ref{pathsearching}, in addition to providing wider optimization margin for path smoothing as mentioned before.
After path optimization, a path that is much smoother than the original reference path is obtained. However, the number of the path vertices is the same as that of the input reference path and the path is still piecewise-linear. Therefore, we further smooth the path via cubic spline interpolation to get a continuous curve.
Fig. \ref{fig:smoothpath}(a) depicts the smoothed path with the path generated by the original state lattice-based path planner (as shown in Fig. \ref{fig:latticepath}(a)) as input. Since the path obtained by the original state lattice-based path planner is close to obstacles, the path vertices near the corner are fixed during path optimization, and the final smoothed path is intuitively rough. On the contrary, the path generated by the proposed path planner (as shown in Fig. \ref{fig:latticepath}(b)) keeps a certain distance from obstacles and provides sufficient optimization margin for path smoothing. As a result, the smoothed paths illustrated in Figs. \ref{fig:smoothpath}(b) and \ref{fig:smoothpath}(c) are much smoother than the path shown in Fig. \ref{fig:smoothpath}(a). Compared with the path shown in Fig. \ref{fig:smoothpath}(b), the deviation from the safe reference path is considered in Fig. \ref{fig:smoothpath}(c), thus the path clearance of the final smoothed path is improved.
\section{Implementation details}
\label{implementationdetails}
\subsection{Setup}
G\ensuremath{^2}VD planner is implemented in C/C++. The convex QP problem described in Section \ref{pathsmoothing} is solved by an alternating direction method of multipliers (ADMM)-based QP solver, OSQP \cite{osqp}. The reference path of path smoothing is obtained by sampling in the path generated by the state lattice-based path planner with an interval of $ 0.1 $ $ \mathrm{m} $. In consideration of the computational efficiency and sensing range, the cumulative length of the reference path is set to $ 4.0 $ $ \mathrm{m} $. The weights $ \omega_s $ and $ \omega_r $ are empirically set to $ 10 $ and $ 1 $, respectively. In the future, we plan to use machine learning techniques to tune these weights adaptively.
\subsection{Metrics}
The proposed Voronoi corridor is integrated into the original state lattice-based path planner (A* + motion primitives) \cite{likhachev2009planning} to derive a new state lattice-based path planner (A* + motion primitives + Voronoi corridor). And the new path planner is compared with the original state lattice-based path planner to validate the effectiveness of the Voronoi corridor. The performance of path searching approaches is evaluated in terms of \emph{computational efficiency} and \emph{memory consumption}. In particular, the number of expanded states and the planning time are used to evaluate the computational efficiency, and the graph size, i.e., the number of created nodes in the search graph, is used to evaluate the memory consumption.
To validate the effectiveness of the proposed QP-based path smoothing approach, we compare it with a recent gradient-based path smoothing approach SBA \cite{wen2020effmop}. We comprehensively evaluate path smoothing approaches in terms of \emph{travel distance}, \emph{efficiency}, and \emph{smoothness}. In particular, the time $ T $ taken by the robot to complete the navigation task is employed to evaluate the motion efficiency, the computational efficiency $ C $ is measured as the time consumption of a single path optimization, and the smoothness is evaluated by the maximum curvature $ K_{max} $ and mean curvature $ K_{mean} $. The curvature is calculated by computing the turning radius of the robot based on the assumption that path vertices are uniformly and densely spread over the path \cite{zhou2020autonomous}. To satisfy this condition, cubic spline interpolation is utilized to smooth the traveling route of the robot to get a continuous curve. Then, we densely sample path vertices on the curve to compute the curvature.
\begin{table}[t]
\vspace{0.2cm}
\centering
\caption{Quantitative statistics of path planning results in simulation}
\label{tbl:pathsearching}
\begin{tabular}{ccccccc}
\toprule
\multicolumn{2}{c}{} & \# of & Time & Graph & Path \\
\multicolumn{2}{c}{} & expands & (secs) & size & cost \\
\midrule
\multirow{2}{*}{Test 1} & Lattice & 155,715 & 0.115 & 161,848 & 291,590 \\
& Ours & \textbf{140,909} & \textbf{0.097} & \textbf{144,943} & 291,590 \\
\midrule
\multirow{2}{*}{Test 2} & Lattice & 167,912 & 0.122 & 174,356 & 324,846 \\
& Ours & \textbf{146,300} & \textbf{0.103} & \textbf{149,912} & 324,846 \\
\midrule
\multirow{2}{*}{Test 3} & Lattice & 202,680 & 0.151 & 210,043 & 383,376 \\
& Ours & \textbf{169,949} & \textbf{0.121} & \textbf{173,716} & 383,376 \\
\bottomrule
\end{tabular}
\vspace{-0.2cm}
\end{table}
\section{Simulations}
\label{simulations}
In this section, we verify the applicability of the proposed G\ensuremath{^2}VD planner in simulation based on the popular Gazebo simulator \cite{koenig2004design}. The simulation scenarios designed in \cite{wen2021mrpb} are used for evaluation\footnote{These simulation models are available at \url{https://github.com/NKU-MobFly-Robotics/local-planning-benchmark}.}.
\subsection{Comparison on Pah Searching}
We choose the large-scale complex maze scenario in \cite{wen2021mrpb} to evaluate the performance of path searching approaches. Furthermore, we randomly select three sets of start and goal configurations in the maze for testing. Both the original state lattice-based path planner and the proposed path planner generate equal quality paths. Table \ref{tbl:pathsearching} enumerates some quantitative statistics of path searching results in these tests.
Since the search space is reduced to the Voronoi corridor, the search direction is focused towards the most promising search areas, and the computational efficiency of graph search is improved. Compared with the original state lattice-based path planner, the number of the expanded search states of the proposed path planner decreases by an average of $ 12.8\% $, and the computational efficiency is improved by $ 17.1\% $. Furthermore, since the search branches for searching in those unpromising search areas are reduced, the number of created nodes in the search graph decreases. Compared with the original state lattice-based path planner, the graph size of the proposed path planner decreases by an average of $ 13.9\% $. In conclusion, the newly proposed path planner generates equal quality paths with less time and memory consumption than the original state lattice-based path planner.
\begin{table*}[t]
\vspace{0.2cm}
\renewcommand\arraystretch{1.6}
\centering
\caption{Quantitative statistics of path smoothing results in simulation}
\label{tbl:pathsmoothing}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multicolumn{2}{|c|}{ \multirow{2}*{} } & \multicolumn{2}{c|}{\textbf{Travel Distance}} &\multicolumn{4}{c|}{\textbf{Efficiency}} &\multicolumn{4}{c|}{\textbf{Smoothness}} \\
\cline{3-12}
\multicolumn{2}{|c|}{} & \multicolumn{2}{c|}{$ S\left[ \mathrm{m} \right] $} & \multicolumn{2}{c|}{$ T\left[ \mathrm{s} \right] $} & \multicolumn{2}{c|}{$ C\left[ \mathrm{ms} \right] $} & \multicolumn{2}{c|}{$ K_{max} $} & \multicolumn{2}{c|}{$ K_{mean} $} \\
\cline{3-12}
\multicolumn{2}{|c|}{} & \textsl{\textbf{SBA}} & \textsl{\textbf{Ours}} & \textsl{\textbf{SBA}} & \textsl{\textbf{Ours}} & \textsl{\textbf{SBA}} & \textsl{\textbf{Ours}} & \textsl{\textbf{SBA}} & \textsl{\textbf{Ours}} & \textsl{\textbf{SBA}} & \textsl{\textbf{Ours}} \\
\cline{1-12}
\multicolumn{1}{|c|}{ \multirow{3}*{\textbf{Shopping Mall}} } & 1 & 29.28 & \textbf{28.92} & 73.23 & \textbf{71.98} & 3.19 & \textbf{0.50} & 1.90 & \textbf{1.54} & 0.33 & \textbf{0.32} \\
\cline{2-12}
\multicolumn{1}{|c|}{} & 2 & 25.43 & \textbf{25.05} & 64.18 & \textbf{62.78} & 2.86 & \textbf{0.50} & 2.68 & \textbf{2.13} & 0.37 & \textbf{0.34} \\
\cline{2-12}
\multicolumn{1}{|c|}{} & 3 & 58.52 & \textbf{57.84} & 141.39 & \textbf{136.64} & 2.65 & \textbf{0.48} & 1.76 & \textbf{1.33} & 0.24 & \textbf{0.22} \\
\cline{1-12}
\multicolumn{1}{|c|}{ \multirow{3}*{\textbf{Complex Maze}} } & 1 & 41.88 & \textbf{40.52} & 93.85 & \textbf{91.44} & 6.58 & \textbf{0.66} & \textbf{2.02} & 2.25 & 0.47 & \textbf{0.46} \\
\cline{2-12}
\multicolumn{1}{|c|}{} & 2 & 53.45 & \textbf{51.47} & 121.24 & \textbf{116.99} & 10.21 & \textbf{0.61} & 2.86 & \textbf{2.20} & \textbf{0.58} & 0.59 \\
\cline{2-12}
\multicolumn{1}{|c|}{} & 3 & 54.50 & \textbf{52.61} & 123.28 & \textbf{119.74} & 7.48 & \textbf{0.61} & 2.13 & \textbf{2.08} & 0.52 & \textbf{0.49} \\
\cline{1-12}
\multicolumn{1}{|c|}{ \multirow{2}*{\textbf{Narrow Passage}} } & 1 & 26.20 & \textbf{25.76} & 54.47 & \textbf{54.21} & 11.51 & \textbf{0.62} & 2.64 & \textbf{1.92} & 0.56 & \textbf{0.53} \\
\cline{2-12}
\multicolumn{1}{|c|}{} & 2 & 30.07 & \textbf{29.45} & 62.59 & \textbf{62.07} & 13.19 & \textbf{0.67} & 1.84 & 1.84 & 0.65 & \textbf{0.63} \\
\hline
\end{tabular}
\end{table*}
\subsection{Comparison on Path Smoothing}
We choose three challenging scenarios of large-scale shopping mall, complex maze, and narrow passage in \cite{wen2021mrpb} to evaluate path smoothing approaches. For each scenario, several sets of different start and goal configurations are selected for testing. Table \ref{tbl:pathsmoothing} presents some quantitative statistics of simulation results.
As shown in Table \ref{tbl:pathsmoothing}, the travel distance of the proposed approach is reduced by an average of $ 0.96 $ $ \mathrm{m} $. Accordingly, the motion efficiency of the proposed path smoothing approach is improved. Furthermore, because the proposed approach models the path smoothing problem as a convex QP problem, while the optimization objective of SBA contains both convex and non-convex terms, the computational efficiency of the proposed approach has been significantly improved. As shown in Table \ref{tbl:pathsmoothing}, the consumed times of the proposed approach are all less than $ 0.7 $ $ \mathrm{ms} $, and smoothing path with the proposed approach is averagely $ 11.86 $ times faster than smoothing path with SBA. Especially in the complex maze and narrow passage scenarios, the computational efficiency of the proposed approach is $ 12.99 $ and $ 19.13 $ times faster than that of SBA, respectively. In addition, the smoothness of motion guidance provided by the proposed approach is commensurate with SBA.
\section{Experiments}
\label{experiments}
In this section, outdoor experimental results in our campus are presented to validate the effectiveness of G\ensuremath{^2}VD planner.
\subsection{Experimental Setup}
As shown in Fig. \ref{fig:platform}, the mobile robot SUMMIT-XL HL is used as the experimental platform, which is equipped with a Velodyne VLP-16 LiDAR, an Intel Realsense D435i depth camera, and an Intel NUC computer. The maximum linear acceleration of the robot is $ 1.0 $ $ \mathrm{m/s^2} $. Considering the safety of robot navigation, the maximum linear velocity is set to $ 1.5 $ $ \mathrm{m/s} $ in the experiments.
To generate a prior global map for autonomous navigation, we first employ a LiDAR-inertial odometry algorithm described in \cite{liosam2020shan} to build a large-scale 3-D point cloud map. And then, a point clouds segmentation algorithm presented in \cite{liu2021point} is used to filter out point clouds that hit the ground and tree canopy. Finally, the filtered point clouds are projected to a 2-D plane to derive traversable regions in the form of occupancy grid map, as shown in Fig. \ref{fig:outdoor}. The dimension of the environment is approximately $ 210 \times 220 $ $ \mathrm{m^2} $. During robot navigation, the LiDAR-inertial odometry algorithm \cite{liosam2020shan} and the point clouds segmentation algorithm \cite{liu2021point} are also utilized to provide state estimation and local traversable regions for the robot, respectively.
\begin{figure}[t]
\vspace{0.2cm}
\centering
\includegraphics[scale=0.205]{fig/outdoor.pdf}
\caption{Occupancy grid map for outdoor navigation. The dimension of the environment is approximately $ 210 \times 220 $ $ \mathrm{m^2} $. The routes of two sets of robot navigation are colored in red and blue, respectively.}
\label{fig:outdoor}
\vspace{-0.4cm}
\end{figure}
\subsection{Autonomous Navigation}
As illustrated in Fig. \ref{fig:outdoor}, we select two sets of different start and goal configurations for outdoor navigation. The total travel distances of these two sets of outdoor navigation are approximately $ 118.3 $ $ \mathrm{m} $ and $ 165.9 $ $ \mathrm{m} $, respectively.
Here we summarize several representative experimental results of autonomous navigation to demonstrate the key characteristics of G\ensuremath{^2}VD planner. More details are included in the video.
\subsubsection{Dealing With Static Obstacles}
Fig. \ref{fig:navigation}(a) illustrates the scenario with static obstacles. The robot avoids a stationary person smoothly, according to the reliable path smoothing results.
\subsubsection{Dealing With Dynamic Obstacles}
Fig. \ref{fig:navigation}(b) shows the scenario with dynamic obstacles. The robot implements fast re-planning and avoids an oncoming person successfully, thanks to the efficient path smoothing approach.
\begin{figure}[t]
\vspace{0.2cm}
\centering
\subfigure[Dealing with static obstacles.]{
\includegraphics[scale=0.25]{fig/navigation_1.pdf}
}
\centering
\subfigure[Dealing with dynamic obstacles.]{
\includegraphics[scale=0.25]{fig/navigation_2.pdf}
}
\caption{Screenshots of outdoor navigation results. The green curve denotes the path generated by the state lattice-based path planner, and the red curve represents the path smoothed by the proposed path smoothing approach.}
\label{fig:navigation}
\vspace{-0.4cm}
\end{figure}
\section{Conclusion}
\label{conclusion}
In this letter, an efficient motion planning approach called G\ensuremath{^2}VD planner is newly proposed for mobile robots. Based on a grid-based GVD, a novel Voronoi corridor is proposed and fully integrated into the state lattice-based path planner to further improve the search efficiency, along with a Voronoi potential field constructed to make the searched path keep a reasonable distance from obstacles to provide sufficient optimization margin for further path smoothing. And an efficient QP-based path smoothing approach is presented, wherein the clearance to obstacles is considered in the form of the penalty of the deviation from the safe reference path to improve the path clearance of hard-constrained path smoothing approaches. The performance of G\ensuremath{^2}VD planner is validated in various complex simulation scenarios and large-scale outdoor environments.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
We consider the nonlinear wave equation with inverse-square potential
\begin{equation}\label{weq}
\left\{
\begin{aligned}
\partial_{t}^{2}u-\Delta u+\frac{a}{|x|^2}u&=|u|^{\frac{4}{N-2}}u\\
\vec{u}|_{t=0}=(u_0,u_1)&\in\dot{H}^{1}(\mathbb{R}^{N})\times L^{2}(\mathbb{R}^{N})
\end{aligned}
\right.
\end{equation}
We denote by $\mathcal{L}_a$ the operator
\begin{align}
\mathcal{L}_{a}:=-\Delta+\frac{a}{|x|^2},
\end{align}
which arises often in complicated mathematical and physical backgrounds, for example the Dirac equation with Coulomb potential, and the study of perturbations of spacetime metrics such as Schwartzshild and Reissner-Nordstr\"om (\cite{BPS, KSWW, VZ}). When $a>-\frac{(N-2)^2}{4}$ and $N\geq3$, we consider $\mathcal{L}_a$ as the Friedrichs extension of the quadratic form defined on $C_c^\infty(\Bbb R^N\backslash\{0\})$ via $$f\mapsto \int_{\Bbb R^N}\left(|\nabla f(x)|^2+a\frac{|f(x)|^2}{|x|^2}\right)dx.$$
The operator $\mathcal{L}_a$ is positive. Thus for $a>-\frac{(N-2)^2}{4}$, we define the function space $\dot{H}_a^1$ whose norm is $$\|f\|_{\dot{H}_a^1}^2=\langle\mathcal{L}_{a}f, f\rangle.$$ We denote the energy space $\mathcal{H}_{a}(\mathbb{R}^N)=\dot{H}^{1}_{a}(\mathbb{R}^N)\times L^2(\mathbb{R}^N)$.
The solution of \eqref{weq} admits the energy
\begin{align}\label{energy}
E(\vec{u}(t))=\int_{\Bbb R^N}\left(\frac12|\nabla_{t,x} u(t,x)|^2+a\frac{|u(t,x)|^2}{|x|^2}-\frac{N-2}{2N}|u(t,x)|^{\frac{2N}{N-2}}\right)dx,
\end{align}
which is conserved. For $\lambda>0$, $f\in\dot{H}^{1}_{a}(\mathbb{R}^N)$ and $g\in L^2(\mathbb{R}^{N})$, we let
\begin{align}\label{def: scalling}
f_{(\lambda)}(x)=\frac{1}{\lambda^{\frac{N-2}{2}}}f\left(\frac{x}{\lambda}\right),\qquad g_{[\lambda]}(x)=\frac{1}{\lambda^{\frac{N}{2}}}g\left(\frac{x}{\lambda}\right).
\end{align}
so that
\begin{align*}
\|f_{(\lambda)}\|_{\dot{H}_{a}^{1}}=\|f\|_{\dot{H}_{a}^{1}}\quad and\quad \|g_{[\lambda]}\|_{L^2}=\|g\|_{L^2}.
\end{align*}
If $u$ is a solution to \eqref{weq}, then $u_{(\lambda)}$ is also a solution. Moreover, $E((u_0)_{(\lambda)}, (u_1)_{[\lambda]})=E(u_0, u_1)$. Therefore, we call the equation \eqref{weq} energy-critical.
There has been increasing interest in the study of the energy critical wave equation with inverse-square potential \eqref{weq} in recent years. Dispersive and Strichartz estimates for radial solutions are established in \cite{BPS, PST} and global well-posedness were proved under the spherically symmetric assumption in critical spaces. The range of admissible pairs and corresponding well-posedness were improved in \cite{MZZ}. In focusing case, the equation \eqref{weq} admits a ground state as shown in \cite{MZZ}
\begin{align}\label{eq: the formula of Wa}
W_a(x)=[N(N-2)\beta^2]^{\frac{N-2}{4}}\left(\frac{|x|^{\beta-1}}{1+|x|^{2\beta}}\right)^{\frac{N-2}{2}},\quad \beta=\sqrt{1+\left(\frac{2}{N-2}\right)^2a}
\end{align}
which is the unique (up to symmetries of the equation) positive radial solution to $$\mathcal{L}_au=|u|^\frac{4}{N-2}u.$$ In \cite{MMZ}, the energy-critical case for large data in dimensions three and four were considered. They showed the global well-posedness and scattering for arbitrary data in the defocusing case and data below the ground states in the focusing case. The global well-posedness and scattering results were generalized in higher dimensions $N\geq5$ in \cite{DL}. However, there are very few results for solutions above the ground state. In this paper, we aim to establish the soliton resolution for radial solutions of \eqref{weq} in all dimensions $N\geq 3$.
Soliton resolution is one of the long-standing topics in nonlinear dispersive equations. In fact, for nonlinear dispersive equations, it is widely believed that coherent structures (e.g. solitons, kinks, monopoles, vortices, etc) and free radiation describe the long-term asymptotic behavior of generic solutions. This belief is known as the soliton resolution conjecture, which states that asymptotically in time, the evolution of generic solutions decouples as a sum of modulated traveling waves and a free radiation. The soliton resolution conjecture arose in the 70's and 80's from various numerical experiments \cite{K} and integrability theory (\cite{Eck, ES, Sch}). For non-integrable model, the soliton resolution conjecture for energy-critical wave equations has been verified in radial case. Duyckaerts, Kenig and Merle \cite{DKM1} proved the soliton resolution in the three dimensional case near the ground state. The radial assumption was removed in \cite{DKM6}. Without a smallness assumption in the radial case, the soliton resolution along time sequences for solutions bounded in the energy norm was proved in \cite{DKM8} in dimension three, in \cite{CR} for all other odd dimensions, in \cite{CKLS} in dimension four and in \cite{JK} in dimension six. In \cite{DJKM}, Duyckaerts, Jia, Kenig and Merle proved that the decomposition holds for any solution of the energy-critical wave equation in space dimensions 3,4,5 for a sequence of times going to its maximal time. Without small size assumption, the soliton resolution for all times was verified in \cite{DKM} in radial case in three dimensions. In a series of works \cite{DKM3, DKM2, DKM7}, Duyckaerts, Kenig and Merle extend the result in \cite{DKM} to all odd dimensions. In their proof, they overcome the fundamental obstruction for the 3D case by reducing the study of a multisoliton solution to a finite dimensional system of ordinary differential equations on the modulation parameters. The key ingredient is to show that this system of equations creates some radiation, contradicting the existence of pure multisolitons. Corresponding results in dimension four were established recently in \cite{DKMM} and extended to dimension six in \cite{CDKM}, although the channel of energy fails in even dimensions. For more details about the soliton resolution of energy critical wave equations, please refer to the survey article \cite{Kenig2}.
Recently there are also soliton resolution type results for other wave type equations. For wave maps, the soliton resolution along a time sequence was proved by Cote \cite{Cote} and Jia, Kenig \cite{JK}. Universality of blow up profiles for small blow-up solutions was established in \cite{DJKM2}. The resolution for all time was established in 1-equivariant case by Duyckaerts, Kenig, Martel and Merle \cite{DKMM} and $k$-equivariant case by Jendrej and Lawrie \cite{JL5}. For wave maps, the two bubble dynamics was studied in \cite{JL1}. More quantitative description and uniqueness results were given in \cite{JL2, JL3, JL4}. In addition, the defocusing wave equations with potentials was systematically studied. The soliton resolution were established by Jia, Liu and Xu in \cite{JBX1, JBX2} and with Schlag in \cite{JBSX1}. For focusing wave equation outside a ball in $\Bbb R^3$ with Dirichlet boundary condition and a superquintic power nonlinearity, Duychaerts and Yang \cite{DY} classified all radial stationary solutions, and proved that all radial global solutions are asymptotically the sum of a stationary solution and a free radiation.
\begin{theorem}\label{main thm}
Let $N\geq3$, $a>0$. Let $\beta=\sqrt{1+\frac{4a}{(N-2)^2}}$ such that $(N-2)\beta>2$ is odd. Let $u$ be a radial solution of \eqref{weq}, with maximal time of existence $T_+$, such that
\begin{equation}\label{bh}
\sup_{0\leq t<T_+}\|\vec{u}(t)\|_{\mathcal{H}_a}<\infty.
\end{equation}
Then there exists $J\geq0$, signs $(\iota_j)_j\in \{\pm1\}^J$, and scaling parameters $(\lambda_j)_j\in(0,\infty)^J$ such that
\begin{equation*}
\forall j\in \llbracket1, J-1\rrbracket, \ \ \lim_{t\rightarrow T_+}\frac{\lambda_j(t)}{\lambda_{j+1}(t)}=+\infty,
\end{equation*}
and
\begin{itemize}
\item (Type II blow-up case). If $T_+<\infty$, then $J\geq1$ and there exists $(v_0, v_1)\in\mathcal{H}_a$ such that
\begin{equation*}
\lim_{t\rightarrow T_+}\left\|\vec{u}(t)-(v_0, v_1)-\sum_{j=1}^J(\iota_j{W_a}_{(\lambda_j)},0)\right\|_{\mathcal{H}_a}=0,
\end{equation*}
where $W_{a(\lambda_j)}(x)=\frac{1}{\lambda_j^\frac{N-2}{2}}W_a\left(\frac{x}{\lambda_j}\right)$ is the $\dot{H}^1$ scaling of the ground state $W_a(x)$. Furthermore, \begin{equation*}
\lim_{t\rightarrow T_+}\frac{\lambda_1(t)}{T_+-t}=0.
\end{equation*}
\item (Global in time case). If $T_+=\infty$, then there exists a solution $v_L$ of the linear wave equation with inverse square potential such that
\begin{equation*}
\lim_{t\rightarrow +\infty}\left\|\vec{u}(t)-\vec{v}_L(t)-\sum_{j=1}^J(\iota_jW_{a(\lambda_j)},0)\right\|_{\mathcal{H}_a}=0.
\end{equation*}
Furthermore, \begin{equation*}
\lim_{t\rightarrow \infty}\frac{\lambda_1(t)}{t}=0.
\end{equation*}
\end{itemize}
\end{theorem}
\begin{remark}
{\it The restriction of $a$.} In order to establish the channel of energy, we will make use of Hankel transform, which is the corresponding distorted Fourier transform associated to the operator $\mathcal{L}_a$. During the derivation, we have to compute integrals involving the factor
\begin{align*}
\cos(r\rho_1-\sigma)\cos(r\rho_2-\sigma)&=\frac{1}{2}[\cos(r(\rho_1+\rho_2)-2\sigma)+\cos(r(\rho_1-\rho_2))],
\end{align*}
where \begin{equation*}
2\sigma=\big((N-2)\beta+1\big)\frac{\pi}{2}.
\end{equation*}
Only if $2\sigma\in\pi\mathbb{Z}$, we have \begin{align*}
\cos(r\rho_1-\sigma)\cos(r\rho_2-\sigma)&=\frac{1}{2}\big[(-1)^{k}\cos(r(\rho_1+\rho_2))+\cos(r(\rho_1-\rho_2))\big]
\end{align*}
and the computation of the integrals can be reduced to spherical Bessel functions. Compared with the classical wave equation case where $\beta=1$ and $2\sigma=((N-2)+1)\frac{\pi}{2}$, we see that the coefficients of inverse square potential plays the role of spatial dimensions partially. In addition, we impose the restriction $(N-2)\beta>2$ to ensure $W_a\in L^2(\Bbb R^N)$. It is an interesting open problem to remove the restriction on $a$.
\end{remark}
\begin{remark} {\it Improvement of the scattering threshold in the radial case. } In \cite{MMZ}, global well-posedness and scattering holds for solutions with energy and kinetic energy less than those of $W$. The essential ingredient in their proof is related to sharp Sobolev embedding
\begin{equation*}
\|f\|_{L^\frac{2N}{N-2}(\Bbb R^N)}\leq C_a\|f\|_{\dot{H}^1_a(\Bbb R^N)},
\end{equation*}
where $C_a$ denotes the sharp constant, see \cite{KMVZZ}
$$C_a=\frac{\|W_{a\wedge 0}\|_{\dot{H}^1_{a\wedge0}}}{\|W_{a\wedge 0}\|_{L^{\frac{2N}{N-2}}}}.$$
However, according to Theorem \ref{main thm}, if the initial data is radial and has energy and kinetic energy less than those of the ground state $W_a$, then the solution is global and scattering holds. Thus in the radial case, the threshold is improved since $C_a<\|W_{a}\|_{\dot{H}^1_{a}}/\|W_{a}\|_{L^{\frac{2N}{N-2}}}$ for $a>0$. In fact, under the radial assumption the equality in the Sobolev embedding holds exactly when $f=W_a$. We expect the improved scattering threshold holds for any $a>-\left(\frac{N-2}{2}\right)^2$ in the radial case.
\end{remark}
\medskip
In their series of works, Duyckaerts, Kenig and Merle have developed a systematic strategy to quantify the ejection of energy, that occurs as we approach the final time $T_+$ for nonlinear wave equations, through a method of energy channels. Exterior channel of energy estimates for the radial wave equation were first considered in three dimensions in \cite{DKM1}, for five dimensions in \cite{KLS}, and all odd dimensions in \cite{KLLS}. However, the exterior energy estimates were shown to fail in even dimensions in \cite{CKS}. Kenig \cite{Kenig} gave a survey on the method of energy channels. In this paper, we need to establish the exterior energy estimates for the radial solutions to the free wave equations with inverse-square potentials
\begin{equation}\label{lwp}
\left\{
\begin{aligned}
\partial_{t}^{2}u-\Delta u+\frac{a}{|x|^2}u&=0, \\
\vec{u}|_{t=0}=(u_0,u_1)&\in(\dot{H}^{1}_{a}\times L^{2})(\mathbb{R}^{N}).
\end{aligned}
\right.
\end{equation}
\medskip
In fact, we will prove
\begin{theorem}\label{thm:weq}\label{ce}
For $(N-2)\beta$ is odd and $(N-2)\beta\geq1$. Let $u_L$ be the solution of \eqref{lwp} with radial initial data $(f,g)$. Then
\begin{equation}\label{ineq: channel of energy}
\max\limits_{\pm}\lim\limits_{t\rightarrow\pm\infty}\int_{r\geq|t|+R}|\nabla_{t,x}u_{L}(t,r)|^2r^{N-1}dr\geq \frac{1}{2}\|\operatorname{\pi}_{P(R)}^{\perp}(f,g)\|_{\dot{H}^1_a\times L^2(r\geq R,r^{N-1}dr)}^2,
\end{equation}
where
\begin{align}\label{def: P(R)}
P(R)\triangleq\operatorname{span}\bigg\{(r^{\alpha+2i-2},0),(0,r^{\alpha+2j-2})\bigg|i=1,2,\cdots,\Big\lfloor\frac{(N-2)\beta+4}{4}\Big\rfloor;\nonumber\\
j=1,2,\cdots,\Big\lfloor\frac{(N-2)\beta+2}{4}\Big\rfloor;r\geq R\bigg\}.
\end{align}
Here $\operatorname{\pi}_{P(R)}^{\perp}$ denotes the orthogonal projection onto the complement of the plane $P(R)$ in $\dot{H}^1_a\times L^2(r\geq R,r^{N-1}dr)$.
\end{theorem}
\begin{remark}
The conditions that $1\leq i\leq \lfloor\frac{(N-2)\beta+4}{4}\rfloor$ and $1\leq j\leq \lfloor\frac{(N-2)\beta+2}{4}\rfloor$ guarantee that
\begin{align*}
(r^{\alpha+2i-2},0),(0,r^{\alpha+2j-2})\in(\dot{H}_{a}^{1}\times L^2)(r\geq R,r^{N-1}dr).
\end{align*}
\end{remark}
In order to establish Theorem \ref{ce}, an important observation is that for radial functions, the operator $\mathcal{L}_a$ can be represented as $\mathcal{L}_a=A^*A$, where $A=\partial_r+\frac{(N-2)(1-\beta)}{2r}$ and $A^*$ is the adjoint operator of $A$. Based on this observation, we define the norm \begin{equation*}\label{def norm}
\|f\|_{\dot{H}^{1}_{a}(r>R)}=\|A f\|_{L^2(r>R)}.
\end{equation*}
If we use other norms such as $ \|f\|_{\dot{H}^{1}_{a}(r>R)}=\|\sqrt{\mathcal{L}_a}f\|_{L^2(r>R)}$ or $\|f\|_{\dot{H}^1(r>R)}=\|\nabla f\|_{L^2(r>R)}$, then extra uncontrollable terms will appear, which makes the exterior energy estimates unavailable. We expect such observation also works for wave equations with other potentials.
In addition, we will prove analogous estimates for the linearized wave equation with inverse square potential at the ground state $W_a$ by using a completely different strategy from \cite{DKM2}. In fact, our method relies on the distorted Hankel transform. In our argument, we show that the linearized potential around $W_a$ produces a negligible error which does not affect the exterior energy estimates dramatically. Our strategy also works for the linearized wave equation around $W$ in \cite{DKM2}.
\medskip
We conclude the introduction with a brief outline of the paper. Section 2 introduces notations and reviews some preliminaries such as Strichartz estimates and profile decomposition. Section 3 concerns the exterior energy estimates for linear wave equations with inverse square potentials. Analogous estimates for linearized wave equation with inverse square potentials around the ground state and multi-solitons are given in Section 4. Then in Section 5, we show the soliton resolution along a time sequence. We consider the case of non-radiative solutions in Section 6. In Section 7, we reduce the proof of soliton resolution to the study of a finite dimensional dynamical system on the scaling parameters and some of the coefficients arising in the expansion of the solutions. Finally, in Section 8 we prove a blow-up/ejection result for this dynamical system and complete the proof.
{\bf Acknowledgement.} C. ~Miao was supported by the National Key Research and Development Program of China (No. 2020YFA0712900) and NSFC Grant of China (11831004). L.~Zhao was partially supported by the NSFC Grant of China (11771415) and the National Key Research and Development Program of China (2020YFA0713100).
\section{Preliminaries}
In this section, we will denote by $\dot{H}_{a}^1(\mathbb{R}^N)$ the homogeneous Sobolev space. Recording some standard facts about the Strichartz estimate, Profile decomposition and importing several relevant results about the ground state $W_a$ for general nonlinear wave equation with inverse-square potential \eqref{weq}.
\subsection{Structure of linear operator}
We define
\begin{equation*}
V(r)=\frac{(N-2)(1-\beta)}{2r}\triangleq\frac{c}{r},
\end{equation*}
where $c=\frac{(N-2)(1-\beta)}{2}$. For the radial case, there admits the decomposition $\mathcal{L}_{a}=A^\ast A$, where
\begin{equation}\label{def operator A}
A=\partial_{r}+V(r).
\end{equation}
Therefore, the homogeneous Sobolev space defined by $\mathcal{L}_a$ has the norm
\begin{equation}\label{def norm}
\|f\|_{\dot{H}^{1}_{a}}=\|(\partial_{r}+cr^{-1})f\|_{L^2}
\end{equation}
and for any $R>0$, \begin{equation}\label{def norm}
\|f\|_{\dot{H}^{1}_{a}(r>R)}=\|(\partial_{r}+cr^{-1})f\|_{L^2(r>R)}.
\end{equation}
\subsection{Strichartz estimates}
The Strichartz estimate and its generalized version \cite{GV} play a key role in the proof of well-posedness. Since the derivative of the nonlinearity is no longer Lipschitz continuous in the standard Strichartz space for dimensions $N>6$, Bulut, Czubak, Li, Pavlovic and Zhang \cite{BCLPZ} using certain “exotic Strichartz" spaces (which have same scaling with the standard Strichartz space but lower derivative) to prove the exotic Strichartz estimate. In \cite{MZZ}, the Strichartz-type estimates were proved for the solutions to the linear wave equation with inverse-square potential by C. Miao, J. Zhang and J. Zheng. For dimensions $N\geq3$ and any time interval $I\subset\mathbb{R}$, we introduce the following space
\begin{align*}
S(I)=L_{t,x}^{\frac{2(N+1)}{N-2}}(I\times\mathbb{R}^{N}),\qquad W(I)=L_{t}^{\frac{2(N+1)}{N-1}}\dot{B}_{\frac{2(N+1)}{N-1},2}^{\frac{1}{2}}(I\times\mathbb{R}^N).
\end{align*}
Based on these spaces, we state the Strichartz estimates for the wave equation with inverse-square potential, which we frequently use throughout the paper.
\begin{lemma}\label{Strichartz estimate}
If $f\in W(I)'$, $(u_0,u_1)\in\mathcal{H}_{a}$, then the solution $u$ of $\partial_{t}^2u+\mathcal{L}_{a}u=f$ with initial data $(u_0,u_1)$ is in $S(I)\cap W(I)$ and
\begin{align}\label{ineq: Strichartz estimate}
\sup\limits_{t\in\mathbb{R}}\|\vec{u}(t)\|_{\mathcal{H}_a}+\|\vec{u}\|_{S(I)}+\|\vec{u}\|_{W(I)}\lesssim\|(u_0,u_1)\|_{\mathcal{H}_a}+\|f\|_{W(I)'}.
\end{align}
\end{lemma}
\subsection{Profile decomposition}\label{profile}
Profile decomposition plays an important role in scattering theory and soliton resolutions. The linear profile decomposition for the wave equation was established in \cite{BG} for dimensions $N=3,4$ and extended to high dimensions $N\geq5$ in \cite{CR}. In \cite{MMZ} and \cite{DL} the linear profile decomposition was established for general solutions to the energy-critical wave equations with inverse square potential. Here we only recall the radial version. Let $\{u_{0,n}, u_{1,n}\}$ be a bounded sequence of radial functions in $\dot{H}^1_a\times L^2$. There exists a subsequence of $\{u_{0,n}, u_{1,n}\}$ (that we will still denote by $\{u_{0,n}, u_{1,n}\}$) with the following properties. There exists a sequence $\{U^j_L\}_{j\geq1}$ of radial solutions to the linear wave equation
\begin{equation}\label{lweq}
\partial_t^2u-\Delta u+\frac{a}{|x|^2}u=0
\end{equation}
with initial data $\{U_0^j, U_1^j\}_j$ in $\dot{H}_{a}^1\times L^2$ for $j\geq1$, sequences $\{\lambda_{j,n}\}_n$, $\{t_{j,n}\}_n$ with $\lambda_{j,n}>0$, $t_{j,n}\in\Bbb R$, and $(w_{0,n}^Jw_{1,n}^J)\in \dot{H}^1_a\times L^2$ satisfying
\begin{equation*}
u_{0,n}=\sum_{j=1}^J\frac{1}{\lambda_{j,n}^\frac{N-2}{2}}U^j_L\left(\frac{-t_{j,n}}{\lambda_{j,n}}, \frac{x}{\lambda_{j,n}}\right)+w_{0,n}^J,
\end{equation*}
and
\begin{equation*}
u_{1,n}=\sum_{j=1}^J\frac{1}{\lambda_{j,n}^\frac{N}{2}}(\partial_tU^j_L)\left(\frac{-t_{j,n}}{\lambda_{j,n}}, \frac{x}{\lambda_{j,n}}\right)+w_{1,n}^J,
\end{equation*}
where
\begin{itemize}
\item the pseudo-orthogonality property holds: for $j\neq k$,
\begin{equation*}
\lim_{n \rightarrow \infty}\Big|\log\frac{\lambda_{j,n}}{\lambda_{k,n}}\Big|+\frac{|t_{j,n}-t_{k,n}|^2}{\lambda_{j,n}\lambda_{k,n}}=\infty.
\end{equation*}
\item Let $w_n^J$ be the solution to \eqref{lweq} with data $(w_{0,n}^J w_{1,n}^J)$, then
$$\lim_{J\rightarrow\infty}\limsup\limits_{n\rightarrow\infty}\|w_n^J\|_{S(\Bbb R)}=0.$$
\item We have the decoupling property for each $J\geq1$:
\begin{align*}
\lim\limits_{n \rightarrow\infty}
\bigg(&\|u_{0,n}\|_{\dot{H}^1_a}^2+\|u_{1,n}\|_{L^2}^2-\sum\limits_{j=1}^J\left\|U_L^j\Big(\frac{-t_{j,n}}{\lambda_{j,n}}\Big)\right\|_{\dot{H}^1_a}^2\\
&-\sum\limits_{j=1}^J\left\|\partial_{t}U_L^j\Big(\frac{-t_{j,n}}{\lambda_{j,n}}\Big)\right\|_{L^2}^2-\|w_{0,n}^J\|_{\dot{H}^1_a}^2-\|w_{0,n}^J\|_{L^2}^2\bigg)=0\\
\lim_{n\rightarrow\infty}\bigg(E_a(u_{0,n},& u_{1,n})-\sum_{j=1}^JE_a
\left(U_L^j\left(\frac{-t_{j,n}}{\lambda_{j,n}}\right),\partial_tU_L^j\left(\frac{-t_{j,n}}{\lambda_{j,n}}\right)\right)-E_a(w_{0,n}^J, w_{1,n}^J)\bigg)=0
\end{align*}
\end{itemize}
\subsection{Ground state $W_a$}
In this subsection, we gather a few purely computational proofs. In $N$ dimensional space and for $a>0$, the ground state soliton is the unique (up to symmetries of the equation) positive solution of \eqref{lwp}. It was given in~\cite{KMVZZ} that the ground state for $a>-\frac{(N-2)^2}{4}$ can be given explicitly,
\begin{align}\label{eq: the formula of Wa}
W_a(x)=[N(N-2)\beta^2]^{\frac{N-2}{4}}\left(\frac{|x|^{\beta-1}}{1+|x|^{2\beta}}\right)^{\frac{N-2}{2}},\quad \beta=\sqrt{1+\left(\frac{2}{N-2}\right)^2a}.
\end{align}
Denote $\Lambda\triangleq\frac{N-2}{2}+x\cdot\nabla$, $\Lambda_0\triangleq\frac{N}{2}+x\cdot\nabla$. From \eqref{eq: the formula of Wa} we have
\begin{align}\label{asympototic1}
|\Lambda_0\Lambda W_a(x)|+|\Lambda W_a|+|W_a|\lesssim\min\left\{|x|^{\frac{(\beta-1)N}{2}-\beta+1},|x|^{-\frac{(\beta+1)N}{2}+\beta+1}\right\},
\end{align}
\begin{align}\label{asympototic2}
\left|(\partial_{r}+\frac{c}{r})\Lambda W_a\right|+\left|(\partial_{r}+\frac{c}{r}) W_a\right|\lesssim\min\left\{|x|^{\frac{(\beta-1)N}{2}-\beta},|x|^{-\frac{(\beta+1)N}{2}+\beta}\right\},
\end{align}
and
\begin{align}\label{asympototic3}
\left|(-\Delta+\frac{a}{r^2})\Lambda W_a\right|\lesssim\left\{|x|^{\frac{(\beta-1)N}{2}-\beta-1},|x|^{-\frac{(\beta+1)N}{2}+\beta-1}\right\}.
\end{align}
Moreover, for any $a,b,c,d>1$ with $b+d>N$, and $|x|=r$, under the assumption that $0<\lambda<\mu$, summing up over all estimates, we conclude
\begin{align}\label{main W}
&\quad\int_{\mathbb{R}^N}\min\bigg(\bigg(\frac{|x|}{\lambda}\bigg)^a,\bigg(\frac{\lambda}{|x|}\bigg)^b\bigg)\min\bigg(\bigg(\frac{|x|}{\mu}\bigg)^c,\bigg(\frac{\mu}{|x|}\bigg)^d\bigg)\ud x\nonumber\\
&=\bigg(\int\limits_{0}^{\lambda}+\int\limits_{\lambda}^{\mu}+\int\limits_{\mu}^{\infty}\bigg)\min\bigg(\bigg(\frac{r}{\lambda}\bigg)^a,\bigg(\frac{\lambda}{r}\bigg)^b\bigg)\min\bigg(\bigg(\frac{r}{\mu}\bigg)^c,\bigg(\frac{\mu}{r}\bigg)^d\bigg)r^{N-1}\ud r\nonumber\\
&=\int\limits_{0}^{\lambda}\bigg(\frac{r}{\lambda}\bigg)^a\bigg(\frac{r}{\mu}\bigg)^cr^{N-1}\ud r+\int\limits_{\lambda}^{\mu}\bigg(\frac{\lambda}{r}\bigg)^b\bigg(\frac{r}{\mu}\bigg)^cr^{N-1}\ud r+\int\limits_{\mu}^{\infty}\bigg(\frac{\lambda}{r}\bigg)^b\bigg(\frac{\mu}{r}\bigg)^dr^{N-1}\ud r\nonumber\\
&\lesssim\displaystyle\max\left\{\lambda^{b}\mu^{N-b},\lambda^{N+c}\mu^{-c}\right\}.
\end{align}
The following lemma introduces some useful inequalities by \eqref{main W}. The proof of this lemma may be found in \cite{DKM7}.
\begin{lemma}\emph{(\cite{DKM7})}
Let $0<\lambda<\mu$. Assume $(N-2)\beta\geq 2$, as a consequence of \eqref{asympototic1}, \eqref{asympototic2} and \eqref{asympototic3}, it holds that
\begin{align}\label{2.10}
\int_{\mathbb{R}^{N}}\Big|(\partial_r&+\frac{c}{r})\left(\Lambda {W_a}_{(\lambda)}\right) \cdot(\partial_r+\frac{c}{r})\left(\Lambda {W_a}_{(\mu)}\right)\Big|\nonumber\\
&+\int_{\mathbb{R}^{N}}\Big|(\partial_r+\frac{c}{r}) {W_a}_{(\lambda)} \cdot (\partial_r+\frac{c}{r}) {W_a}_{(\mu)} \Big|\lesssim\bigg(\frac{\lambda}{\mu}\bigg)^{\frac{(N-2)\beta}{2}},
\end{align}
\begin{align}\label{the bdd of Lambda Wa and Lambda Wa}
\int_{\mathbb{R}^{N}}\left|(\Lambda W_a)_{[\lambda]}(\Lambda W_a)_{[\mu]}\right|+\int_{\mathbb{R}^{N}}\left|(\Lambda W_a)_{[\lambda]}\left(\Lambda_{0} \Lambda W_a\right)_{[\mu]}\right| \lesssim\left(\frac{\lambda}{\mu}\right)^{\frac{(N-2)\beta}{2}-1},
\end{align}
\begin{align*}
&\left\|{W_a}_{(\lambda)} {W_a}_{(\mu)}^{\frac{4}{N-2}}\right\|_{L^{\frac{2 N}{N+2}}}
\lesssim\left\{
\begin{array}{l}
\left(\frac{\lambda}{\mu}\right)^{\frac{(N-2)\beta}{2}},\qquad \text{if}\quad N\geq6\\
\left(\frac{\lambda}{\mu}\right)^{\frac{(N-2)\beta+(N-6)}{2}},\quad \text{if}\quad N\leq5.
\end{array}\right.
\end{align*}
\begin{align}\label{2.12}
&\left\|{W_a}_{(\mu)} {W_a}_{(\lambda)}^{\frac{4}{N-2}}\right\|_{L^{\frac{2 N}{N+2}}}
\lesssim\left\{
\begin{array}{l}
\left(\frac{\lambda}{\mu}\right)^{2\beta},\qquad \text{if}\quad N\geq6\\
\left(\frac{\lambda}{\mu}\right)^{\frac{(N-2)\beta}{2}},\quad \text{if}\quad N\leq5.
\end{array}\right.
\end{align}
\begin{align*}
\int\Big|(\Lambda W_a)_{[\lambda]}\Big((-\Delta+\frac{a}{r^2})\Lambda W_a\Big)_{[\mu]}\Big| \lesssim\left(\frac{\lambda}{\mu}\right)^{\frac{(N-2)\beta}{2}-1},
\end{align*}
\begin{align}\label{2.13}
\int\Big|(\Lambda W_a)_{[\mu]}\Big((-\Delta+\frac{a}{r^2}) \Lambda W_a\Big)_{[\lambda]}\Big| \lesssim\left(\frac{\lambda}{\mu}\right)^{\frac{(N-2)\beta}{2}},
\end{align}
\begin{align}\label{ineq: Wa N N-2}
\int{W_a}_{(\lambda)}^{\frac{N}{N-2}}{W_a}_{(\mu)}^{\frac{N}{N-2}}\ud x\lesssim\left(\frac{\lambda}{\mu}\right)^{\frac{N\beta}{2}}.
\end{align}
\end{lemma}
\medskip
\begin{proposition}\label{prop: L1L2 bdd of Wa}\emph{(\cite{DKM7})}
Assume $N\geq 5$. Let $0<\lambda<\mu$. Then
\begin{align*}
\left\|\chi_{\{|x| \geq|t|\}} \min \left\{{W_a}_{(\lambda)}^{\frac{4}{N-2}} {W_a}_{(\mu)}, {W_a}_{(\mu)}^{\frac{4}{N-2}} {W_a}_{(\lambda)}\right\}\right\|_{L^{1}\left(\mathbb{R}, L^{2}\right)}\lesssim\left(\frac{\lambda}{\mu}\right)^{\frac{(N+2)\beta}{4}}.
\end{align*}
\end{proposition}
The proof of this proposition is similar to Claim A.3 in the appendix of \cite{DKM7}.
\medskip
\subsection{Pointwise bounds}
In this subsection, we record some inequalities associated with the nonlinear term of \eqref{weq}, which will be used in Section \ref{section 7}.
\begin{proposition}\emph{(\cite{DKM7})}
Assume $N\geq 5$, $J\geq 1$. For all $(y_1, y_2,\ldots,h)\in\mathbb{R}^{J+1}$, we have
\begin{align}\label{nonlinear estimate for y1 to yJ}
&\left|\frac{N-2}{2 N}\bigg|\sum_{j=1}^{J} y_{j}+h\bigg|^{\frac{2 N}{N-2}}-\frac{N-2}{2 N} \sum_{j=1}^{J}\left|y_{j}\right|^{\frac{2 N}{N-2}}-\sum_{j=1}^{J}\left|y_{j}\right|^{\frac{4}{N-2}} y_{j} h-\sum_{1 \leq j, k \leq J}\left|y_{j}\right|^{\frac{4}{N-2}} y_{j} y_{k}\right|\nonumber\\
\leq&\sum_{1 \leq j<k \leq J}\left(\min \left\{\left|y_{j}\right|^{\frac{4}{N-2}} y_{k}^{2},\left|y_{k}\right|^{\frac{4}{N-2}} y_{j}^{2}\right\}+\min \left\{\left|y_{j}\right|^{\frac{N+2}{N-2}}\left|y_{k}\right|,\left|y_{k}\right|^{\frac{N+2}{N-2}}\left|y_{j}\right|\right\}\right)\nonumber\\
\quad&+|h|^{\frac{2 N}{N-2}}+\sum_{j=1}^{J}\left|y_{j}\right|^{\frac{4}{N-2}}|h|^{2}.
\end{align}
\end{proposition}
\medskip
\begin{proposition}\emph{(\cite{DKM7})}
Let $(a,b,c)\in\mathbb{R}^3$ with $a\neq 0$, then
\begin{align}\label{eq: 2 nonlinear estimate}
\Big|F(a+b)-F(a)-F'(a)b\Big|\lesssim \chi_{\{|b|\leq|a|\}}|b|^{2}|a|^{\frac{6-N}{N-2}}+\chi_{\{|b|>|a|\}}|b|^{\frac{N+2}{N-2}},
\end{align}
and
\begin{align}\label{eq: estimate of F2}
&\Big|F(a+b+c)-F(a+b)-F(a+c)+F(a)\Big|
\lesssim\left\{\begin{array}{l}
|a|^{\frac{6-N}{2(N-2)}}|b|^{\frac{N+2}{2(N-2)}}|c|,\qquad if\quad N\geq6\\
|b||c|(|a|+|b|+|c|)^{\frac{6-N}{N-2}},\quad if\quad N\leq5.
\end{array}\right.
\end{align}
where $F(a)=|a|^{\frac{4}{N-2}}a$.
\end{proposition}
\medskip
\section{Exterior energy estimate for the linear equation}\label{section-linear}
In this section we consider the lower bound on the exterior energy for radial solution of the free wave equation with the inverse-square potential \eqref{lwp}. Our goal is to prove Theorem \ref{ce}.
\subsection{Spectrum of $-\Delta+\frac{a}{|x|^2}$ and Hankel transform}
\medskip
Before proving Theorem \ref{thm:weq}, we establish the Hankel transform of the operator $\mathcal{L}_a=-\Delta+\frac{a}{|x|^2}$ on $\mathbb{R}^N$, and recall a few standard facts about the solution to \eqref{lwp} with radial initial data.
Let us firstly consider the eigenvalue problem associated with the operator $\mathcal{L}_a$
\begin{align}\label{eigenvalue equation}
\left(-\Delta+\frac{a}{|x|^2}\right)u=\rho^2 u.
\end{align}
Since $u$ is radially symmetric, we let $u(x)=h(r)$, then
\begin{align*}
h''(r)+\frac{N-1}{r}h'(r)-(\frac{a}{r^2}-\rho^2)h(r)=0.
\end{align*}
Let $\lambda=\rho r$ and $h(r)=\lambda^{-\frac{N-2}{2}}\tilde{h}(\lambda)$, we obtain
\begin{align*}\label{bassel equation}
\tilde{h}''(\lambda)+\frac{1}{\lambda}\tilde{h}'(\lambda)+\left(1-\frac{(N-2)^2+4a}{4\lambda^2}\right)\tilde{h}(\lambda)=0.
\end{align*}
Define $\nu=\frac{\sqrt{(N-2)^2+4a}}{2}=\frac{N-2}{2}\beta$, then the Bessel function $J_{\nu}(\lambda)$ solves the Bessel equation \eqref{eigenvalue equation}, and the eigenfunction corresponding to the spectrum $\rho^2$ can be expressed by
\begin{align*}
h_{\rho}(r)=(r\rho)^{-\frac{N-2}{2}}J_{\nu}(r\rho).
\end{align*}
We define the following elliptic operator
\begin{align*}
A_{\nu}\triangleq-\partial_{r}^2-\frac{N-1}{r}\partial_{r}+\frac{a}{r^2}=-\partial_{r}^2-\frac{N-1}{r}\partial_{r}+\frac{\nu^2-(\frac{N-2}{2})^2}{r^2}.
\end{align*}
The Hankel transform of order $\nu$ is defined by:
\begin{equation}\label{Hankel transform}
(\mathcal{H}_{\nu}f)(\rho)\triangleq\tilde{f}(\rho)=\int_{0}^{\infty}(r\rho)^{-\frac{N-2}{2}}J_{\nu}(r\rho)f(r)r^{N-1}\ud r.
\end{equation}
We recall the following properties of the Hankel transform given in \cite{MZZ}.
\begin{lemma}
Let $\mathcal{H}_{\nu}$ and $\mathcal{A}_{\nu}$ be defined as above. Then
\begin{itemize}
\item $\mathcal{H}_{\nu}=\mathcal{H}_{\nu}^{-1}$,
\item $\mathcal{H}_{\nu}$ is self-adjoint, i.e. $\mathcal{H}_{\nu}=\mathcal{H}_{\nu}^{\ast}$,
\item $\mathcal{H}_{\nu}$ is an $L^2$ isometry, i.e. $\|\mathcal{H}_{\nu}\phi\|_{L_{\xi}^{2}}=\|\phi\|_{L_{x}^2}$,
\item $\mathcal{H}_{\nu}(A_{\nu}\phi)(\xi)=|\xi|^2(\mathcal{H}_{\nu}\phi)(\xi)$, for $\phi\in L^2$.
\end{itemize}
\end{lemma}
Taking the Hankel transformation of both sides of \eqref{lwp}, we have
\begin{equation}\label{new equation}
\left\{
\begin{aligned}
&\partial_{t}^{2}\tilde{u}+\rho^2\tilde{u}=0,\\
&\vec{u}|_{t=0}=(\tilde{f},\tilde{g}).\\
\end{aligned}
\right.
\end{equation}
The solution $\tilde{u}$ to \eqref{new equation} is given by
\begin{align}\label{Hankel transform for u}
\tilde{u}(t,\rho)=\frac{1}{2}\left(\tilde{f}(\rho)+\frac{\tilde{g}(\rho)}{i\rho}\right)e^{it\rho}+\frac{1}{2}\left(\tilde{f}(\rho)-\frac{\tilde{g}(\rho)}{i\rho}\right)e^{-it\rho}.
\end{align}
Substituting \eqref{Hankel transform for u} into \eqref{Hankel transform}, for the solution $u(t,r)$ this means that
\begin{align}\label{eq:u}
u(t,r)=\int_{0}^{\infty}(r\rho)^{-\frac{N-2}{2}}J_{\nu}(r\rho)\Big(\cos(t\rho)\tilde{f}(\rho)+\frac{1}{\rho}\sin(t\rho)\tilde{g}(\rho)\Big)\rho^{N-1}d\rho,\nonumber\\
\partial_{t}u(t,r)=\int_{0}^{\infty}(r\rho)^{-\frac{N-2}{2}}J_{\nu}(r\rho)\Big(-\sin(t\rho)\rho \tilde{f}(\rho)+\cos(t\rho)\tilde{g}(\rho)\Big)\rho^{N-1}d\rho.
\end{align}
The standard asymptotic behavior for $|z|\rightarrow\infty$ of the Bessel functions is given in \cite{AS} that
\begin{align}\label{eq:J}
&J_{\nu}(z)\sim\sqrt{\frac{2}{\pi z}}\Big[(1+\omega_2(z))\cos(z-\sigma)+\omega_{1}(z)\sin(z-\sigma)\Big],\nonumber\\
&J_{\nu}'(z)\sim\sqrt{\frac{2}{\pi z}}\Big[-(1+\tilde{\omega}_2(z))\sin(z-\sigma)+\tilde{\omega}_{1}(z)\cos(z-\sigma)\Big],
\end{align}
where the phase-shift $\sigma=\frac{(N-2)\beta+1}{2}\frac{\pi}{2}$, and for $n\in\mathbb{N}, z\geq1$,
\begin{align*}
&|\omega_1^{(n)}(z)|+|\tilde{\omega}_1^{(n)}(z)|\leq C_{n}z^{-1-n},\\
&|\omega_2^{(n)}(z)|+|\tilde{\omega}_2^{(n)}(z)|\leq C_{n}z^{-2-n}.
\end{align*}
We now sketch the proof of Theorem \ref{ce} and divide it into the following smaller steps.
\medskip
\subsection{Asymptotic formula}
Computing the asymptotic form of the exterior energy as in \cite{KLLS}, for $t\geq0$, from \eqref{eq:u} to \eqref{eq:J}, we have
\begin{align}\label{eq:u,ut}
&\|\partial_{t}u\|_{L^2(r\geq t+R,r^{N-1}\ud r)}^{2}+\|u\|_{\dot{H}_{a}^{1}(r\geq t+R,r^{N-1}\ud r)}^2\nonumber\\
=&\frac{2}{\pi}\lim\limits_{\varepsilon\rightarrow0}\int_{t+R}^{\infty}\iint\limits_{\rho_1,\rho_2>0}\Big(\sin(t\rho_1)\rho_1 \tilde{f}(\rho_1)-\cos(t\rho_1)\tilde{g}(\rho_1)\Big)\Big(\sin(t\rho_2)\rho_2\overline{\tilde{f}(\rho_2)}-\cos(t\rho_2)\overline{\tilde{g}(\rho_2)}\Big)\nonumber\\
&\times\Big[(1+\omega_2(r\rho_1))\cos(r\rho_1-\sigma)+\omega_1(r\rho_1)\sin(r\rho_1-\sigma)\Big]\nonumber\\
&\times\Big[(1+\omega_2(r\rho_2))\cos(r\rho_2-\sigma)+\omega_1(r\rho_2)\sin(r\rho_2-\sigma)\Big]\rho_1^{\frac{N-1}{2}}\rho_2^{\frac{N-1}{2}}e^{-\varepsilon r}\ud \rho_1\ud \rho_2\ud r\nonumber\\
&+\frac{2}{\pi}\lim\limits_{\varepsilon\rightarrow0}\int_{t+R}^{\infty}\iint\limits_{\rho_1,\rho_2>0}\Big(\cos(t\rho_1)\rho_1 \tilde{f}(\rho_1)+\sin(t\rho_1)\tilde{g}(\rho_1)\Big)\Big(\cos(t\rho_2)\rho_2\overline{\tilde{f}(\rho_2)}+\sin(t\rho_2)\overline{\tilde{g}(\rho_2)}\Big)\nonumber\\
&\times\Big[(1+\omega_2(r\rho_1))\cos(r\rho_1-\sigma)+\omega_1(r\rho_1)\sin(r\rho_1-\sigma)\Big]\nonumber\\
&\times\Big[(1+\omega_2(r\rho_2))\cos(r\rho_2-\sigma)+\omega_1(r\rho_2)\sin(r\rho_2-\sigma)\Big]\rho_1^{\frac{N-1}{2}}\rho_2^{\frac{N-1}{2}}e^{-\varepsilon r}\ud \rho_1\ud \rho_2\ud r.
\end{align}
For all $\varepsilon>0$, the error terms which contain $\omega_{i}$, $i=1,2$, example as
\begin{align*}\label{eq:error}
E_{1}(\varepsilon)\triangleq&\int_{t+R}^{\infty}\iint\limits_{\rho_1,\rho_2>0}\sin(t\rho_1)\sin(t\rho_2)\cos(r\rho_1-\sigma)\sin(r\rho_2-\sigma)\\
&\cdot\omega_1(r\rho_1)\cdot \tilde{f}(\rho_1)\overline{\tilde{g}(\rho_2)}(\rho_1\rho_2)^{\frac{N+1}{2}}e^{-\varepsilon r}\ud \rho_1\ud \rho_2\ud r.
\end{align*}
Using the asymptotic expansions of the Bessel functions, we can absorb all error terms in the $o(1)$ term. Therefore as in \cite{KLLS}, \eqref{eq:u,ut} can be reduced to
\begin{align*}
&\quad\|\partial_{t}u\|_{L^2(r\geq t+R,r^{N-1}\ud r)}^{2}+\|u\|_{\dot{H}_{a}^{1}(r\geq t+R,r^{N-1}\ud r)}^2\nonumber\\
=&\frac{2}{\pi}\lim\limits_{\varepsilon\rightarrow0}\int_{t+R}^{\infty}\iint\limits_{\rho_1,\rho_2>0}\Big(-\sin(t\rho_1)\rho_1\tilde{f}(\rho_1)+\cos(t\rho_1)\tilde{g}(\rho_1)\Big)\Big(-\sin(t\rho_2)\rho_2\overline{\tilde{f}(\rho_2)}+\cos(t\rho_2)\overline{\tilde{g}(\rho_2)}\Big)\nonumber\\
&\times\cos(r\rho_1-\sigma)\cos(r\rho_2-\sigma)\rho_1^{\frac{N-1}{2}}\rho_2^{\frac{N-1}{2}}e^{-\varepsilon r}\ud \rho_1\ud \rho_2\ud r\nonumber\\
&+\frac{2}{\pi}\lim\limits_{\varepsilon\rightarrow0}\int_{t+R}^{\infty}\iint\limits_{\rho_1,\rho_2>0}\Big(\cos(t\rho_1)\rho_1\tilde{f}(\rho_1)+\sin(t\rho_1)\nonumber\tilde{g}(\rho_1)\Big)\Big(\cos(t\rho_2)\rho_2\overline{\tilde{f}(\rho_2)}+\sin(t\rho_2)\overline{\tilde{g}(\rho_2)}\Big)\nonumber\\
&\times\sin(r\rho_1-\sigma)\sin(r\rho_2-\sigma)\rho_1^{\frac{N-1}{2}}\rho_2^{\frac{N-1}{2}}e^{-\varepsilon r}\ud \rho_1\ud \rho_2\ud r+o(1),
\end{align*}
where the $o(1)$ is for $t\rightarrow\infty$ and $\sigma=((N-2)\beta+1)\frac{\pi}{4}$. Since
\begin{align*}
\cos(r\rho_1-\sigma)\cos(r\rho_2-\sigma)&=\frac{1}{2}\Big[\cos(r(\rho_1+\rho_2)-2\sigma)+\cos(r(\rho_1-\rho_2))\Big],\\
\sin(r\rho_1-\sigma)\sin(r\rho_2-\sigma)&=\frac{1}{2}\Big[-\cos(r(\rho_1+\rho_2)-2\sigma)+\cos(r(\rho_1-\rho_2))\Big],
\end{align*}
for arbitrary $t\in\mathbb{R}$, it holds that
\begin{equation*}
\lim\limits_{\varepsilon\rightarrow 0+}\int_{t}^{\infty}\cos(a r)e^{\varepsilon r}\ud r=\lim\limits_{\varepsilon\rightarrow 0+}\frac{1}{2}\left(-\frac{e^{t(i a-\varepsilon)}}{i a-\varepsilon}+\frac{e^{-t(i a+\varepsilon)}}{i a+\varepsilon}\right)=\pi\delta_0(a)-\frac{\sin(ta)}{a}.
\end{equation*}
However, $\lim\limits_{\varepsilon\rightarrow 0+}\int_{t}^{\infty}\sin(a r)e^{\varepsilon r}\ud r$ is not well-defined. Therefore we introduce a restrictive condition
\begin{equation*}
2\sigma=\big((N-2)\beta+1\big)\frac{\pi}{2}\in\pi\mathbb{Z}.
\end{equation*}
This means $k\triangleq\frac{(N-2)\beta+1}{2}\in\mathbb{Z}$. Thus,
\begin{align*}
\cos(r\rho_1-\sigma)\cos(r\rho_2-\sigma)&=\frac{1}{2}\Big[(-1)^{k}\cos(r(\rho_1+\rho_2))+\cos(r(\rho_1-\rho_2))\Big],\\
\sin(r\rho_1-\sigma)\sin(r\rho_2-\sigma)&=\frac{1}{2}\Big[-(-1)^k\cos(r(\rho_1+\rho_2))+\cos(r(\rho_1-\rho_2))\Big].
\end{align*}
Since the order $\nu$ of Bessel function is a half-integer, we introduce the {\bf spherical Bessel functions} $j_{n}(z)$,
\begin{equation*}\label{def:jn}
j_{n-\frac{1}{2}}(z)=\left(\frac{\pi}{2z}\right)^{\frac{1}{2}}J_{n}(z).
\end{equation*}
Using $k=\frac{(N-2)\beta+1}{2}$, $\nu=\frac{(N-2)\beta}{2}$, it is easy to check that
$j_{k-1}(z)=(\frac{\pi}{2z})^{\frac{1}{2}}J_{\nu}(z)$.
For $n\geq0$ the spherical Bessel function $j_{n}(z)$ satisfies
\begin{equation*}\label{eq:jn}
j_{n}(z)=(-z)^{n}
\left(\frac{1}{z}\frac{d}{dz}\right)^{n}\frac{\sin z}{z}.
\end{equation*}
It is important to notice that the parity of the function $j_n(z)$ only depends on $n$ odd or even. Since $k\in\mathbb{Z}$, $J_{\nu}(z)$ can be replaced by $j_{k-1}(z)$. Treating $k$ with even and odd cases respectively, we conclude the following lemma as in \cite{KLLS},
\begin{lemma}\label{asf,asg}
Assume $(N-2)\beta\geq1$ and $(N-2)\beta$ is odd. Take $\kappa(x)$ to be a normalized bump function on $\mathbb{R}$, that is smooth, even, nonnegative, and such that $\kappa(|x|)$ is decreasing, supp$\kappa\subset[-1,1]$ and $\int_{\mathbb{R}}\kappa(x)\ud x=1$. Denote $\kappa_{\varepsilon}(x)=\frac{1}{\varepsilon}\kappa(\frac{x}{\varepsilon})$. For every $R>0$, every radial solution of \eqref{lwp} satisfies
\begin{align}\label{eq:lower bound}
\max\limits_{\pm}&\lim\limits_{t\rightarrow\pm\infty}\left(\|\partial_{t}u(t)\|_{L^2(r\geq|t|+R,r^{N-1}\ud r)}^2+\|u(t)\|_{\dot{H}^1_a(r\geq|t|+R,r^{N-1}\ud r)}^2\right)\nonumber\\
\geq& AS_{N}^{a}(f)+AS_{N}^{a}(g),
\end{align}
with
\begin{align}\label{eq:ASf}
AS_{N}^{a}(f)=&\frac{1}{2}\int_{0}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2r^{N-1}\ud r\nonumber\\
-&\frac{1}{2\pi^2}\lim\limits_{\varepsilon\rightarrow0}\iint\limits_{r_1,r_2>0}K_{R,\varepsilon}(r_1,r_2)u_0(r_1)\overline{u_0(r_2)}(r_1r_2)^{\frac{N-1}{2}}\ud r_1\ud r_2,
\end{align}
where
\begin{equation*}
K_{R,\varepsilon}(r_1,r_2)=\frac{1}{2r_1^2r_2^2}\int_{-R}^{R}(\kappa_{\varepsilon}\ast\mathcal{F}\psi_{k-1})(\frac{\xi}{r_1})\overline{(\kappa_{\varepsilon}\ast\mathcal{F}\psi_{k-1})}(\frac{\xi}{r_2})\ud \xi,
\end{equation*}
and
\begin{align}\label{eq:ASg}
AS_{N}^{a}(g)=&\frac{1}{2}\int_{0}^{\infty}|g(r)|^2r^{N-1}\ud r\nonumber\\
-&\frac{1}{2\pi^2}\lim\limits_{\varepsilon\rightarrow0}\iint\limits_{r_1,r_2>0}\tilde{K}_{R,\varepsilon}(r_1,r_2)u_1(r_1)\overline{u_1(r_2)}(r_1r_2)^{\frac{N-1}{2}}\ud r_1\ud r_2,
\end{align}
where
\begin{equation*}
\tilde{K}_{R,\varepsilon}(r_1,r_2)=\frac{1}{2r_1r_2}\int_{-R}^{R}(\kappa_{\varepsilon}\ast\mathcal{F}\varphi_{k-1})(\frac{\xi}{r_1})\overline{(\kappa_{\varepsilon}\ast\mathcal{F}\varphi_{k-1})}(\frac{\xi}{r_2})\ud \xi.
\end{equation*}
Here $k=\frac{(N-2)\beta+1}{2}$, $\varphi_{k-1}(z)=z\cdot j_{k-1}(z)$, $\psi_{k-1}(z)=z\varphi_{k-1}(z)=z^2j_{k-1}(z)$ with $j_{k-1}(z)$ being the spherical Bessel functions defined in \eqref{def:jn}. Moreover, the equality \eqref{eq:lower bound} holds when initial data is $(f,0)$ or $(0,g)$, and it holds for both time directions.
\end{lemma}
\subsection{Algebraic expression for the projection}
Before establishing the asymptotic behavior, we first introduce the Cauchy matrix as in \cite{S}.
Let $a_{1}$, $\cdots$, $a_{k}$ be linearly independent vectors in some inner product space $(V,\langle,\rangle)$ and let
$$
W=\operatorname{span}\left\{a_{1}, \cdots a_{k}\right\}.
$$
Taking any vector $u \in V$, the orthogonal projection onto $W^{\perp}$ is written as
\begin{equation*}
\operatorname{\pi}_{W^{\perp}} u=u-\left(\lambda_{1} a_{1}+\cdots \lambda_{k} a_{k}\right),
\end{equation*}
with the coefficients satisfying
\begin{equation*}
\left\langle\operatorname{\pi}_{W^{\perp}} u, a_{j}\right\rangle=\left\langle u, a_{j}\right\rangle-\sum\limits_{i} \lambda_{i}\left\langle a_{i}, a_{j}\right\rangle=0.
\end{equation*}
Denote
\begin{equation*}
U=\left[\left\langle u, a_{i}\right\rangle\right]_{1 \times k}, \quad \Lambda=\left[\lambda_{i}\right]_{1 \times k}, \quad A=\left[\left\langle a_{i}, a_{j}\right\rangle\right]_{k \times k}.
\end{equation*}
Using the fact that $A$ is symmetric, invertible and positive definite, setting $A=\left[a_{i j}\right], B=A^{-1}=\left[b_{i j}\right]$, then
\begin{align}\label{proj u norm}
\left\|\operatorname{\pi}_{W^{\perp}}u\right\|^{2}&=\langle u, u\rangle-\sum_{i,j=1}^{k}\lambda_{i}\lambda_{j}\left\langle a_{i}, a_{j}\right\rangle=\langle u, u\rangle-U A^{-1} U^{t}\nonumber\\
&=\langle u, u\rangle-\sum_{i,j=1}^{k}b_{i,j}\left\langle u,a_{i}\right\rangle\left\langle u,a_{j}\right\rangle.
\end{align}
In order to compute the matrix $B$, we need to introduce the {\bf Cauchy Matrix}. If $A$ is an $m\times m$ matrix of the form
\begin{equation*}
A=\left[\frac{1}{x_{i}-y_{j}}\right], \quad x_{i}-y_{j} \neq 0 ; \quad 1 \leq i, j \leq m,
\end{equation*}
then its determinant is
\begin{equation*}\label{det A}
\operatorname{det}(A)=\frac{\prod_{i<j}\left(x_{i}-x_{j}\right)\left(y_{j}-y_{i}\right)}{\prod_{i, j=1}^{m}\left(x_{i}-y_{j}\right)},
\end{equation*}
hence we conclude that the Cauchy matrix is invertible. Using Cramer's rule, we obtain an explicit formula for its inverse
\begin{equation}\label{def B}
B=\left[b_{i j}\right]=A^{-1}, \quad b_{i j}=\left(x_{j}-y_{i}\right) A_{j}\left(y_{i}\right) B_{i}\left(x_{j}\right),
\end{equation}
where $A_{i}(x)$ and $B_{j}(y)$ are the Lagrange polynomials for $x_{i}, y_{j}$ respectively, i.e.,
\begin{align}\label{def Ai}
A_{i}(x)&=\frac{\prod_{\ell \neq i}\left(x-x_{\ell}\right)}{\prod_{\ell \neq i}\left(x_{i}-x_{\ell}\right)}=\prod_{1 \leq \ell \leq m, \ell\neq i} \frac{x-x_{\ell}}{x_{i}-x_{\ell}},
\end{align}
\begin{align}\label{def Bj}
B_{j}(y) &=\prod_{1 \leq \ell\leq m,\ell\neq j} \frac{y-y_{\ell}}{y_{j}-y_{\ell}}.
\end{align}
Now we derive the explicit formula for the projection of initial data $(f,g)$ in space $\tilde{W}\times W$.
{\bf For initial data $(0,g)$}, we choose the space $V:=L^{2}(r\geq R,r^{N-1}dr)$, and the generalized kernel space $W$ of $\mathcal{L}_a$ in $V$ is
\begin{equation}\label{eq:span g}
W=\operatorname{span}\left\{r^{\alpha-2+2i}\bigg|i=1,2,\cdot\cdot\cdot,\tilde{k}_1=\Big\lfloor\frac{(N-2)\beta+2}{4}\Big\rfloor;r\geq R\right\}.
\end{equation}
Denote the Cauchy matrix $A=[\langle a_i,a_j\rangle]_{\tilde{k}_1\times\tilde{k}_1}:=(a_{ij})_{\tilde{k}_1\times\tilde{k}_1}$, where
\begin{align*}
a_{i j}(R)&\triangleq\left\langle r^{\alpha-2+2i},r^{\alpha-2+2j} \right\rangle_{V}=\int_{R}^{\infty}r^{2\alpha-4+2i+2j}r^{N-1}\ud r\nonumber\\
&=\frac{R^{-(N-2)\beta+2i+2j-2}}{(N-2)\beta-2i-2j+2}.
\end{align*}
In particular, when $R=1$, the Cauchy matrix is
\begin{equation*}
A(1)=\left[\frac{1}{(N-2)\beta-2i-2j+2}\right]_{\tilde{k}_1\times\tilde{k}_1}:=\left[\frac{1}{x_i-y_j}\right]_{\tilde{k}_1\times\tilde{k}_1},
\end{equation*}
here $x_i=(N-2)\beta-2i+2$, $y_j=2j$. By the inverse matrix formula \eqref{def B}, \eqref{def Ai} and \eqref{def Bj} we conclude
\begin{align}\label{bij for g}
b_{i j}(1)=&(x_j-y_i)A_{j}(y_i)B_{i}(x_j)\nonumber\\
=&\big((N-2)\beta-2i-2j+2\big)\prod_{1\leq \ell\leq\tilde{k}_1,\ell\neq j}\frac{2i-(N-2)\beta+2\ell-2}{2\ell-2j}\nonumber\\
&\times\prod_{1\leq \ell\leq\tilde{k}_1,\ell\neq i}\frac{(N-2)\beta-2i-2\ell+2}{2\ell-2i}\nonumber\\
=&\frac{\prod_{1\leq \ell\leq\tilde{k}_1}\big(2i+2\ell-2-(N-2)\beta\big)\prod_{1\leq \ell\leq\tilde{k}_1}\big(2j+2\ell-2-(N-2)\beta\big)}{\big((N-2)\beta-2i-2j+2\big)\prod_{1\leq \ell\leq\tilde{k}_1,\ell\neq j}(2\ell-2j)\prod_{1\leq \ell\leq\tilde{k}_1,\ell\neq i}(2\ell-2i)}.
\end{align}
Therefore we obtain the inverse Cauchy matrix
\begin{equation*}
B(R)=A(R)^{-1}=\langle b_{i j}(R)\rangle=b_{i j}(1)R^{(N-2)\beta-2i-2j+2}.
\end{equation*}
From \eqref{proj u norm} and \eqref{bij for g}, the projection formula is given by
\begin{align}\label{proj g}
&\quad\|\operatorname{\pi}_{W^{\perp}}g\|_{L^2(r\geq R,r^{N-1}\ud r)}\nonumber\\
&=\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{\tilde{k}_1}\frac{R^{(N-2)\beta-2i-2j+2}}{(N-2)\beta-2i-2j+2}\tilde{c}_i\tilde{c}_j\int_{R}^{\infty}g(r)r^{\alpha+2i-2}r^{N-1}\ud r\nonumber\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\times\int_{R}^{\infty}\overline{g(r)}r^{\alpha+2j-2}r^{N-1}\ud r\nonumber\\
&=\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{\tilde{k}_1}\frac{R^{-2c-2i-2j+N}}
{-2c-2i-2j+N}\tilde{c}_i\tilde{c}_j\int_{R}^{\infty}g(r)r^{c+2i-1}\ud r\int_{R}^{\infty}\overline{g(r)}r^{c+2j-1}\ud r,
\end{align}
with
\begin{equation}\label{eq:proj g}
\tilde{c}_{j}=\frac{\prod_{1\leq \ell\leq\tilde{k}_1}\big((N-2)\beta-2\ell-2j+2\big)}{\prod_{1\leq \ell\leq\tilde{k}_1,\ell\neq j}(2\ell-2j)}, \qquad \ell\leq j\leq \tilde{k}_1=\Big\lfloor\frac{(N-2)\beta+2}{4}\Big\rfloor.
\end{equation}
{\bf For initial data $(f,0)$.} Let $\tilde{V}=\dot{H}^{1}_{a}(r\geq R,r^{N-1}\ud r)$, and the corresponding generalized kernel space be
\begin{equation}\label{eq:span f}
\tilde{W}=\operatorname{span}\left\{r^{\alpha-2+2i}\bigg|i=1,2,\cdot\cdot\cdot,\tilde{k}_2=\Big\lfloor\frac{(N-2)\beta+4}{4}\Big\rfloor;r\geq R\right\}.
\end{equation}
The same computation as before yields that the elements in the matrix $A(R)$ are
\begin{align*}
\tilde{a}_{i j}(R)=\left(\alpha-2+2i\right)\left(\alpha-2+2j\right)\frac{R^{-(N-2)\beta+2i+2j-4}}{(N-2)\beta-2i-2j+4}.
\end{align*}
Direct computation shows the inverse of the Cauchy matrix $[\alpha_{i j}]:=[\frac{1}{(N-2)\beta-2i-2j+4}]_{\tilde{k}_2\times\tilde{k}_2}$. By \eqref{def B}, \eqref{def Ai} and \eqref{def Bj}, the elements $\beta_{i j}$ of $B_j$ have the form
\begin{align*}
\beta_{i j}=&\frac{1}{(N-2)\beta-2i-2j+4}\\
&\times\frac{\prod_{1\leq \ell\leq\tilde{k}_2}\big((N-2)\beta-2i-2\ell+4\big)\prod_{1\leq \ell\leq\tilde{k}_2}\big(N-2)\beta-2j-2\ell+4\big)}{\prod_{1\leq \ell\leq\tilde{k}_2,\ell\neq j}(2\ell-2j)\prod_{1\leq \ell\leq\tilde{k}_2,\ell\neq i}(2\ell-2i)}.
\end{align*}
Thus the inverse matrix $\tilde{B}(R)=[b_{i j}(R)]_{\tilde{k}_{2}\times\tilde{k}_2}$ can be written as
\begin{align}\label{def: tilde bij}
\tilde{b}_{i j}(R)=&\frac{1}{(N-2)\beta-2i-2j+4}\frac{R^{(N-2)\beta-2i-2j+4}}{(\alpha-2+2i)(\alpha-2+2j)}\nonumber\\
&\times\frac{\prod_{1\leq \ell\leq\tilde{k}_2}\big((N-2)\beta-2i-2\ell+4\big)\prod_{1\leq \ell\leq\tilde{k}_2}\big(N-2)\beta-2j-2\ell+4\big)}{\prod_{1\leq \ell\leq\tilde{k}_2,\ell\neq j}(2\ell-2j)\prod_{1\leq \ell\leq\tilde{k}_2,\ell\neq i}(2\ell-2i)}.
\end{align}
Applying \eqref{proj u norm} with \eqref{def: tilde bij}, we obtain the projection formula
\begin{align}\label{projection of f}
&\quad\|\operatorname{\pi}_{\tilde{W}^{\perp}}f\|^{2}_{\dot{H}^1_a(r\geq R,r^{N-1}\ud r)}\nonumber\\
=&\int_{R}^{\infty}\left|(\partial_{r}+\frac{c}{r})f(r)\right|^2r^{N-1}\ud r -\sum_{i,j=1}^{\tilde{k}_2}\tilde{d}_i\tilde{d}_j\frac{R^{(N-2)\beta-2i-2j+4}}{(N-2)\beta-2i-2j+4}\nonumber\\
&\times\int_{R}^{\infty}(\partial_{r}+\frac{c}{r})f(r)r^{\alpha+2i-3}r^{N-1}\ud r\int_{R}^{\infty}(\partial_{r}+\frac{c}{r})\overline{f(r)}r^{\alpha+2j-3}r^{N-1}\ud r\nonumber\\
=&\int_{R}^{\infty}\left|(\partial_{r}+\frac{c}{r})f(r)\right|^2r^{N-1}\ud r -\sum_{i,j=1}^{\tilde{k}_2}\tilde{d}_i\tilde{d}_j\frac{R^{-2c-2i-2j+N+2}}{-2c-2i-2j+N+2}\nonumber\\
&\times\int_{R}^{\infty}(\partial_{r}+\frac{c}{r})f(r)r^{c+2i-2}\ud r\int_{R}^{\infty}(\partial_{r}+\frac{c}{r})\overline{f(r)}r^{c+2j-2}\ud r\nonumber\\
=&\int_{R}^{\infty}\left|(\partial_{r}+\frac{c}{r})f(r)\right|^2r^{N-1}\ud r -\sum_{i,j=1}^{\tilde{k}_2}\tilde{d}_i\tilde{d}_j\frac{R^{-2c-2i-2j+N+2}}{-2c-2i-2j+N+2}\nonumber\\
&\times\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i-2}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j-2}\ud r,
\end{align}
where
\begin{equation}\label{eq:proj f}
\tilde{d}_{j}=\frac{\prod_{1\leq \ell\leq\tilde{k}_{2}}((N-2)\beta-2\ell-2j+4)}{\prod_{1\leq \ell\leq\tilde{k}_{2},\ell\neq j}(2\ell-2j)},\qquad 1\leq j\leq\tilde{k}_{2}=\Big\lfloor\frac{(N-2)\beta+4}{4}\Big\rfloor.
\end{equation}
\medskip
Let us recall some useful results about the coefficients $\tilde{c}_j$ and $\tilde{d}_j$.
\begin{lemma}\emph{(\cite{KLLS})}\label{coefficients results}
Given the coefficients $\tilde{c}_j$, $1\leq j\leq \tilde{k}_1=\lfloor\frac{(N-2)\beta+2}{4}\rfloor$ and $\tilde{d}_j$, $1\leq j\leq \tilde{k}_2=\lfloor\frac{(N-2)\beta+4}{4}\rfloor$ defined as in \eqref{eq:proj g} and \eqref{eq:proj f}, we have the following identities
\begin{align*
\sum\limits_{j=1}^{\tilde{k}_1}\frac{\tilde{c}_j}{(N-2)\beta-2m-2j+2}=1, \quad for\ any\ m\in\mathbb{Z}, 1\leq m\leq\tilde{k}_1,
\end{align*}
\begin{align}\label{prod cj}
\sum\limits_{j=1}^{\tilde{k}_1}\frac{\tilde{c}_j}{2j}+1=\prod\limits_{\ell=1}^{\tilde{k}_1}\frac{(N-2)\beta-2\ell+2}{2\ell}.
\end{align}
Similarly, for coefficients $\tilde{d}_j$ we have
\begin{align}\label{sum dj}
\sum\limits_{j=1}^{\tilde{k}_2}\frac{\tilde{d}_j}{(N-2)\beta-2m-2j+4}=1, \quad for\ any\ m\in\mathbb{Z}, 1\leq m\leq\tilde{k}_2,
\end{align}
\begin{align*
\sum\limits_{j=1}^{\tilde{k}_2}\frac{\tilde{d}_j}{2j}+1=\prod\limits_{\ell=1}^{\tilde{k}_2}\frac{(N-2)\beta-2\ell+4}{2\ell}.
\end{align*}
\end{lemma}
The following lemma in \cite{KLLS} collects various limits which we will use repeatedly in the calculation of $AS_{N}^{a}g$ and $AS_{N}^{a}f$. These limits all involve regularization by the mollifier $\kappa_{\varepsilon}(x)$ which we introduce in Lemma~\ref{asf,asg}. In the lemma we denote by $\delta_{ab}$ the Kronecker delta
\begin{equation*}
\delta_{ab}=\begin{cases}1, & \text { when } a=b, \\ 0, & \text { when } a \neq b.\end{cases}
\end{equation*}
In contrast, $\delta_{y}$ is the standard Dirac measure centered at $y \in \mathbb{R} .$ In other words, $\left(f * \delta_{y}\right)(x)=f(x-y)$.
\begin{lemma}\label{le:limits}
Let $i$, $j\geq 1$, $r_{1}$, $r_{2}>0$. For $a$, $b\in\{0,1\}$, we have the following limits for any test function $\phi\left(r_{1}, r_{2}\right)\in C_{0}^{\infty}\left(\mathbb{R}^{+}\times\mathbb{R}^{+}\right)$
\begin{align}\label{kappa xi kappa xi}
&\lim_{\varepsilon\rightarrow 0}\iint\limits_{r_{1},r_{2}>0}\left[\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{i} \chi_{(-1,1)}\right]\Big(\frac{\xi}{r_{1}}\Big)\left[\kappa_{\varepsilon}\ast\xi^{j}\chi_{(-1,1)}\right]\Big(\frac{\xi}{r_{2}}\Big)\ud \xi\right]\phi\left(r_{1}, r_{2}\right) \ud r_{1} \ud r_{2}\nonumber\\
&\quad=\left(1-(-1)^{i+j+1}\right)\iint\limits_{r_{1}, r_{2}>0} \frac{1}{i+j+1} \frac{\min \left(r_{1}, r_{2}, R\right)^{i+j+1}}{r_{1}^{i} r_{2}^{j}} \phi\left(r_{1}, r_{2}\right) \ud r_{1} \ud r_{2},
\end{align}
\begin{align}\label{delta,delta}
&\lim_{\varepsilon\rightarrow 0}\iint\limits_{r_{1},r_{2}>0}\left[\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast \delta_{(-1)^{a}}\right]\Big(\frac{\xi}{r_{1}}\Big)\left[\kappa_{\varepsilon}\ast\delta_{(-1)^{b}}\right]\Big(\frac{\xi}{r_{2}}\Big)\ud \xi\right]\phi\left(r_{1},r_{2}\right)\ud r_{1}\ud r_{2}\nonumber\\
&\quad=\delta_{ab}\int_{0}^{R}\xi^{2}\phi(\xi,\xi)\ud \xi,
\end{align}
\begin{align}\label{delta',delta'}
&\lim_{\varepsilon\rightarrow 0}\iint\limits_{r_{1},r_{2}>0}\left[\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast \delta_{(-1)^{a}}^{\prime}\right]\Big(\frac{\xi}{r_{1}}\Big)\left[\kappa_{\varepsilon}\ast \delta_{(-1)^{b}}^{\prime}\right]\Big(\frac{\xi}{r_{2}}\Big)\ud \xi\right]\phi\left(r_{1},r_{2}\right)\ud r_{1}\ud r_{2}\nonumber\\
&\quad=\bigg.\delta_{ab}\int_{0}^{R}\left[\partial_{x}\partial_{y}\left(x^{2}y^{2}\phi(x,y)\right)\right]\bigg|_{x=y=\xi} d \xi,
\end{align}
\begin{align}\label{xi,delta}
&\lim_{\varepsilon\rightarrow 0}\iint\limits_{r_{1},r_{2}>0}\left[\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{i} \chi_{(-1,1)}\right]\Big(\frac{\xi}{r_{1}}\Big)\left[\kappa_{\varepsilon}\ast\delta_{(-1)^{a}}\right]\Big(\frac{\xi}{r_{2}}\Big)\ud \xi\right] \phi\left(r_{1},r_{2}\right)\ud r_{1}\ud r_{2}\nonumber\\
&\quad=\int_{0}^{\infty}\int_{0}^{\min\left(r_{1},R\right)}r_{2}\left(\frac{(-1)^{a}r_{2}}{r_{1}}\right)^{i}\phi\left(r_{1},r_{2}\right)\ud r_{2}\ud r_{1},
\end{align}
\begin{align}\label{xi,delta'}
&\lim_{\varepsilon\rightarrow 0}\iint\limits_{r_{1},r_{2}>0}\left[\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{i} \chi(-1,1)\right]\Big(\frac{\xi}{r_{1}}\Big)\left[\kappa_{\varepsilon}\ast\delta_{(-1)^{a}}^{\prime}\right]\Big(\frac{\xi}{r_{2}}\Big)\ud \xi\right]\phi\left(r_{1},r_{2}\right)\ud r_{1}\ud r_{2}\nonumber\\
&\quad=(-1)^{a}\int_{0}^{\infty}\int_{0}^{\min\left(r_{1},R\right)}\left(\frac{(-1)^{a}r_{2}}{r_{1}}\right)^{i}\partial_{r_{2}}\left(r_{2}^{2} \phi\left(r_{1},r_{2}\right)\right)\ud r_{1}\ud r_{2},
\end{align}
\begin{align}\label{delta,delta'}
&\lim_{\varepsilon\rightarrow 0}\iint\limits_{r_{1},r_{2}>0}\left[\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast \delta_{(-1)^{a}}\right]\Big(\frac{\xi}{r_{1}}\Big)\left[\kappa_{\varepsilon}\ast\delta_{(-1)^{b}}^{\prime}\right]\Big(\frac{\xi}{r_{2}}\Big)\ud \xi\right]\phi\left(r_{1},r_{2}\right)\ud r_{1}\ud r_{2}\nonumber\\
&\quad=\bigg.(-1)^{a}\delta_{ab}\int_{0}^{R}\left[\partial_{y}\left(x y^{2}\phi(x,y)\right)\right]\Big|_{x=y=\xi}\ud \xi.
\end{align}
\end{lemma}
\subsection{Proof of Theorem \ref{thm:weq} for $(0,g)$ data}\label{SS:2.4}
By the formula of $AS_{N}^{a}g$ in \eqref{eq:ASg}, the Fourier expansion formula of $\varphi_{k-1}(x)$ needs to be used. Since the parity of $\varphi_j(x)$ is crucial, we have to discuss separately the two cases according to the order of the function.
\begin{lemma}\label{lem:fourier g}
Let $k_1\geq 1$. With $k-1=2k_1$, and $k-1$ is even, we have
\begin{align}\label{even varphi}
\mathcal{F}\varphi_{k-1}=&\mathcal{F}\varphi_{2k_1}\nonumber\\
=&\pi i(-1)^{k_1}\bigg[\sum\limits_{j=1}^{k_1}\frac{\prod_{\ell=1}^{k_1}(4k_1-2\ell-2j+3)}{\prod_{l\leq \ell\leq k,\ell\neq j}(2\ell-2j)}\xi^{2k_1-2j+1}\chi_{(-1,1)}(\xi)+(\delta_1+\delta_{-1})\bigg],
\end{align}
we denote
\begin{align*}
c_j=\frac{\prod_{\ell=1}^{k_1}(4k_1-2j-2\ell+3)}{\prod_{1\leq \ell\leq k_1,\ell\neq j}(2\ell-2j)}.
\end{align*}
With $k-1=2k_1-1$ odd
\begin{align}\label{odd varphi}
\mathcal{F}\varphi_{k-1}=&\mathcal{F}\varphi_{2k_1-1}\nonumber\\
=&\pi(-1)^{k_1-1}\bigg[\sum\limits_{j=1}^{k_1}\frac{\prod_{\ell=1}^{k_1}(4k_1-2j-2\ell+1)}{\prod_{1\leq \ell\leq k_1,\ell\neq j}(2\ell-2j)}\xi^{2k_1-2j}\chi_{(-1,1)}(\xi)+(-\delta_{1}-\delta_{-1})\bigg],
\end{align}
we also denote
\begin{align*}
c_{j}=\frac{\prod_{\ell=1}^{k_1}(4k_1-2j-2\ell+1)}{\prod_{1\leq \ell\leq k_1,\ell\neq j}(2\ell-2j)}.
\end{align*}
\end{lemma}
\medskip
Assume $r_1,r_2>R$ as before, since we consider initial data $(0,g)$, without loss of generality, assume that $g(r)=0$ for $r\leq R$. Hence by Lemma~\ref{le:limits}, the terms which integral region contain $(0,R)$ is equal zero and the terms created by the $\delta$-function will make no contribution to our integral.
If $k-1\in\mathbb{Z}$ is even, then there exists $k_1\in\mathbb{Z}$ such that $k-1=2k_1$. By definition of $c=\frac{(N-2)(1-\beta)}{2}$, then $2k_1=-c+\frac{N-3}{2}$. Using \eqref{eq:ASg}, \eqref{even varphi} and \eqref{kappa xi kappa xi}, we obtain
\begin{align*}
\quad&2AS_{N}^{a}g\nonumber\\
=&\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\frac{1}{2\pi^2}\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\big(\kappa_{\varepsilon}\ast\mathcal{F}\varphi_{k-1}\big)\big(\frac{\xi}{r_1}\big)
\overline{\big(\kappa_{\varepsilon}\ast\mathcal{F}\varphi_{k-1}\big)\big(\frac{\xi}{r_2}\big)}\ud \xi\nonumber\\
&\times g(r_1)\overline{g(r_2)}(r_1r_2)^{\frac{N-3}{2}}\ud r_1\ud r_2\nonumber\\
=&\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\frac{1}{2}\sum\limits_{i,j=1}^{k_1}\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{2k_1-2i+1}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_1}\big)\nonumber\\
&\times \left[\kappa_{\varepsilon}\ast\xi^{2k_1-2j+1}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_2}\big)\ud\xi~ g(r_1)\overline{g(r_2)}(r_1r_2)^{\frac{N-3}{2}}\ud r_1\ud r_2
\end{align*}
\begin{align*}
=&\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{k_1}\frac{c_i c_j R^{4k_1-2i-2j+3}}{4k_1-2i-2j+3}\int_{R}^{\infty}g(r)r^{-2k_1+2i+\frac{N-5}{2}}\ud r\nonumber\\
&\times\int_{R}^{\infty}\overline{g(r)}r^{-2k_1+2j+\frac{N-5}{2}}\ud r\nonumber\\
=&\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{k_1}\frac{c_i c_j R^{-2c-2i-2j+N}}{-2c-2i-2j+N}\int_{R}^{\infty}g(r)r^{c+2i-1}\ud r\int_{R}^{\infty}\overline{g(r)}r^{c+2j-1}\ud r.
\end{align*}
Due to the definition \eqref{eq:span g} of the space $W$, the dimension is $\tilde{k}_2=\lfloor\frac{(N-2)\beta+2}{4}\rfloor=\lfloor k_1+\frac{3}{4}\rfloor=k_1$, and $4k_1+3-2i-2j=(N-2)\beta+2-2i-2j$. Compared with \eqref{proj g}, we obtain
\begin{equation}\label{even ASg}
AS_{N}^{a}(g)=\frac{1}{2}\big\|\operatorname{\pi}_{W^{\perp}}g\big\|_{L^2(r\geq R,r^{N-1}\ud r)}^2, \qquad(N-2)\beta=4k_1+1.
\end{equation}
\medskip
Similarly, if $k-1$ is odd, then there exists $k_1\in\mathbb{Z}$ such that $k-1=2k_1-1$, then $k=2k_1=-c+\frac{N-1}{2}$. We thus see from \eqref{odd varphi} that
\begin{align*}
\quad&2AS_{N}^{a}g\nonumber\\
=&\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\frac{1}{2\pi^2}\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\big(\kappa_{\varepsilon}\ast\mathcal{F}\varphi_{2k_1-1}\big)\big(\frac{\xi}{r_1}\big)
\overline{\big(\kappa_{\varepsilon}\ast\mathcal{F}\varphi_{2k_1-1}\big)\big(\frac{\xi}{r_2}\big)}\ud \xi\nonumber\\
&\times g(r_1)\overline{g(r_2)}(r_1r_2)^{\frac{N-3}{2}}\ud r_1\ud r_2\nonumber\\
=&\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\frac{1}{2}\sum\limits_{i,j=1}^{k_1}c_i c_j\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{2k_1-2i}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_1}\big)\nonumber\\
&\times\left[\kappa_{\varepsilon}\ast\xi^{2k_1-2j}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_2}\big)\ud\xi g(r_1)\overline{g(r_2)}(r_1r_2)^{\frac{N-3}{2}}\ud r_1\ud r_2\nonumber\\
=&\int_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{k_1}\frac{c_i c_j R^{4k_1-2i-2j+1}}{4k_1-2i-2j+1}\int_{R}^{\infty}g(r)r^{-2k_1+2i+\frac{N-3}{2}}\ud r\nonumber\\
&\times\int_{R}^{\infty}\overline{g(r)}r^{-2k_1+2j+\frac{N-3}{2}}\ud r\nonumber\\
=&\int\limits_{R}^{\infty}|g(r)|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{k_1}\frac{c_i c_j R^{-2c-2i-2j+N}}{-2c-2i-2j+N}\int_{R}^{\infty}g(r)r^{c+2i-1}\ud r\int_{R}^{\infty}\overline{g(r)}r^{c+2j-1}\ud r.
\end{align*}
Compared with \eqref{proj g}, we have proved
\begin{align}\label{odd ASg}
AS_{N}^{a}(g)=\frac{1}{2}\big\|\operatorname{\pi}_{W^{\perp}}g\big\|_{L^2(r\geq R,r^{N-1}\ud r)}^2, \qquad(N-2)\beta=4k_1-1.
\end{align}
For each $k\in\mathbb{Z}$, together \eqref{even ASg} with \eqref{odd ASg}, we conclude the estimate \eqref{ineq: channel of energy} of Theorem~\ref{ce} for $(0,g)$.
\subsection{Proof of Theorem~\ref{thm:weq} for $(f,0)$ data}\label{SS:2.5}
Indeed, if we obtain the Fourier transform of $\psi_{k-1}$, we can calculate $AS_{N}^{a}f(r)$, and use the projection formula \eqref{projection of f} to show the estimate \eqref{ineq: channel of energy} for $(f,0)$. As before, we also divide into two situations.
\begin{lemma}\label{Fourier of psi}
Let $k_1\geq 2$. With $k-1=2k_1-2$ and $k-1$ is even, we have
\begin{align}\label{even Fourier psi}
\mathcal{F}\psi_{k-1}(\xi)=\mathcal{F}\psi_{2k_1-2}(\xi)=&\pi(-1)^{k_1}\bigg[\sum\limits_{j=1}^{k_1-1}d_j(2k_1-2j-1)\xi^{2k_1-2j-2}\chi_{(-1,1)}(\xi)\nonumber\\
&+\sum\limits_{j=1}^{k_1-1}d_j(-\delta_{1}-\delta_{-1})+(-\delta'_{1}+\delta'_{-1})\bigg],
\end{align}
where
\begin{align*}
d_j=\frac{\prod_{\ell=1}^{k_1-1}(4k_1-2j-2\ell-1)}{\prod_{1\leq \ell\leq k_1-1,\ell\neq j}(2\ell-2j)},\qquad 1\leq j\leq k_1-1.
\end{align*}
For $k-1=2k_1-1$ we have
\begin{align}\label{odd Fourier psi}
\mathcal{F}\psi_{k-1}(\xi)=\mathcal{F}\psi_{2k_1-1}(\xi)=&\pi(-1)^{k_1-1}i\bigg[\sum\limits_{j=1}^{k_1}d_j(2k_1-2j)\xi^{2k_1-2j-1}\chi_{(-1,1)}(\xi)\nonumber\\
&+\sum\limits_{j=1}^{k_1}d_j(-\delta_{1}+\delta_{-1})+(-\delta'_{1}-\delta'_{-1})\bigg],
\end{align}
where
\begin{align*}
d_j=\frac{\prod_{\ell=1}^{k_1}(4k_1-2j-2\ell+1)}{\prod_{1\leq \ell\leq k_1,\ell\neq j}(2\ell-2j)},\qquad 1\leq j\leq k_1.
\end{align*}
\end{lemma}
\subsubsection{\bf{Case $k-1=2k_1-2$}}
In this case, the dimension of the space $\tilde{W}$ is $\tilde{k}_1=\lfloor\frac{-2c+N+2}{4}\rfloor=\lfloor k_1+\frac{1}{4}\rfloor=k_1$. We also introduce the following notation to simplify the formula when we have the symmetric structure
$$\big\langle S(i,j)\big\rangle=S(i,j)+\overline{S(j,i)}.$$
Substituting the Fourier expansion of $\psi_{k-1}$ \eqref{even Fourier psi} into \eqref{eq:ASf}, combining Lemma~\ref{le:limits} \eqref{kappa xi kappa xi} to \eqref{delta,delta'} with the notation $F(r_1,r_2):=f(r_1)r_1^c\overline{f(r_2)r_2^c}(r_1r_2)^{-c+\frac{N-5}{2}}$, it holds
\begin{align*}
\quad&2AS_{N}^{a}f(r)\\
=&\int_0^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^{2}r^{N-1}\ud r-\lim\limits_{\varepsilon\rightarrow0}\iint\limits_{r_1,r_2>0}\bigg[\frac{1}{2\pi^2r_1^2r_2^2}\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast\mathcal{F}\psi_{2k_1-2}\big]\big(\frac{\xi}{r_1}\big)\\
&\times\overline{\big[\kappa_{\varepsilon}\ast\mathcal{F}\psi_{2k_1-2}\big]\big(\frac{\xi}{r_2}\big)}d\xi\bigg]f(r_1)\overline{f(r_2)}(r_1r_2)^{\frac{N-1}{2}}\ud r_1\ud r_2
\end{align*}
\begin{align*}
=&\int_{0}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^{2}r^{N-1}\ud r-\frac{1}{2}\sum\limits_{i,j=1}^{k_1-1}d_{i}d_{j}(2k_1-2i-1)(2k_1-2j-1)\\
&\times\lim\limits_{\varepsilon\rightarrow0}\iint\bigg[\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{2k_1-2i-2}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_1}\big)\left[\kappa_{\varepsilon}\ast\xi^{2k_1-2j-2}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_2}\big)\ud \xi\bigg]F(r_1,r_2)\ud r_1\ud r_2\\
&-\frac{1}{2}\sum\limits_{i,j=1}^{k_1-1}d_id_j\lim\limits_{\varepsilon\rightarrow0}\iint\bigg[\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast(\delta_{1}+\delta_{-1})\big]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast(\delta_{1}+\delta_{-1})\big]\big(\frac{\xi}{r_2}\big)\bigg]F(r_1,r_2)\ud r_1\ud r_2\\
&-\frac{1}{2}\lim\limits_{\varepsilon\rightarrow0}\iint\bigg[\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast(-\delta'_{1}+\delta'_{-1})\big]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast(-\delta'_{1}+\delta'_{-1})\big]\big(\frac{\xi}{r_2}\big)\ud\xi\bigg]F(r_1,r_2)\ud r_1\ud r_2\\
&+\frac{1}{2}\bigg\langle\sum_{i,j=1}^{k_1-1}(2k_1-2i-1)d_id_j\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{2k_1-2i-2}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_1}\big)\\
&\times\left[\kappa_{\varepsilon}\ast(\delta_{1}+\delta_{-1})\right]\big(\frac{\xi}{r_2}\big)\ud \xi F(r_1,r_2)\ud r_1\ud r_2\bigg\rangle\\
&-\frac{1}{2}\bigg\langle\sum\limits_{i=1}^{k_1-1}(2k_1-2i-1)d_i\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\left[\kappa_{\varepsilon}\ast\xi^{2k_1-2i-2}\chi_{(-1,1)}\right]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast(\delta'_{1}-\delta'_{-1})\big]\big(\frac{\xi}{r_2}\big)\ud \xi\\
&\times F(r_1,r_2)\ud r_1\ud r_2\bigg\rangle\\
&-\frac{1}{2}\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i\lim\limits_{\varepsilon\rightarrow0}\iint\bigg[\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast(\delta_{1}+\delta_{-1})\big]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast(\delta'_{1}-\delta'_{-1})\big]\big(\frac{\xi}{r_2}\big)\ud \xi\bigg]F(r_1,r_2)\ud r_1\ud r_2\bigg\rangle\\
=&\int_{0}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{k_1-1}d_id_j\iint\limits_{r_1,r_2>0}\frac{(2k_1-2i-1)(2k_1-2j-1)}{4k_1-2i-2j-3}\\
&\times\frac{\min(r_1,r_2,R)^{4k_1-2i-2j-3}}{r_1^{2k_1-2i-2}r_2^{2k_1-2j-2}}f(r_1)r_1^{c}\overline{f(r_2)r_2^c}(r_1r_2)^{-c+\frac{N-5}{2}}\ud r_1\ud r_2\\
&-\sum\limits_{i,j=1}^{k_1-1}d_id_j\int_{0}^{R}|f(r)r^c|^2r^{-2c+N-3}\ud r-\int_{0}^{R}\left|\partial_{r}\big(f(r)r^c\cdot r^{-c+\frac{N-1}{2}}\big)\right|^2\ud r\\
&+\bigg\langle\sum\limits_{i,j=1}^{k_1-1}d_id_j(2k_1-2i-1)\int_{0}^{\infty}\int_{0}^{\min(r_1,R)}r_2\big(\frac{r_2}{r_1}\big)^{2k_1-2i-2}f(r_1)r_1^{c}\overline{f(r_2)r_2^{c}}(r_1r_2)^{-c+\frac{N-5}{2}}\ud r_2\ud r_1\bigg\rangle\\
&-\bigg\langle-\sum\limits_{i=1}^{k_1-1}d_i(2k_1-2i-1)\int_{0}^{\infty}\int_{0}^{\min{(r_1,R)}}\big(\frac{r_2}{r_1}\big)^{2k_1-2i-2}f(r_1)r_1^{c}r_1^{-c+\frac{N-5}{2}}\\
&\times\partial_{r_2}\left(\overline{f(r_2)r_2^{c}}r_2^{-c+\frac{N-1}{2}}\right)\ud r_2\ud r_1\bigg\rangle\\
&-\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i\int_{0}^{R}f(r)r^{c}r^{-c+\frac{N-3}{2}}\partial_{r}\left(\overline{f(r)r^c}r^{-c+\frac{N-1}{2}}\right)\ud r\bigg\rangle\\
=&\int_{0}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2r^{N-1}\ud r-(S_1+S_2+S_3+S_4+S_5+S_6).
\end{align*}
Due to the symmetric position of $r_1,r_2$ and the structure of the operator $\mathcal{L}_{a}=A^{\ast} A$, we have
\begin{align*}
0=&\left\langle(-\Delta+\frac{a}{r^2})u,u\right\rangle=\left\langle\Big((-\partial_{r}+\frac{N-1}{r})+\frac{c}{r}\Big)\big(\partial_{r}+\frac{c}{r}\big)u,u\right\rangle\\
=&\left\langle(\partial_{r}+\frac{c}{r})u,(\partial_{r}+\frac{c}{r})u\right\rangle=\bigg\langle\frac{\partial_{r}\big(u(r)r^c\big)}{r^c},\frac{\partial_{r}\big(u(r)r^c\big)}{r^c}\bigg\rangle.
\end{align*}
We can assume $f(r)r^c=f(R)R^c$, for $r\leq R$. Notice that $2k_1=-c+\frac{N+1}{2}$, we compute each $S_i$, respectively. Integration by parts gives
\begin{align*}
S_1=&\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{(-2c-2i-2j+N+2)}\\
&\times\iint\limits_{r_1,r_2>0}\frac{\min(r_1,r_2,R)^{-2c-2i-2j+N-2}}{r_1^{-c-2i+\frac{N-3}{2}}r_2^{-c-2j+\frac{N-3}{2}}}(f(r_1)r_1^{c})(\overline{f(r_2)r_2^{c}})(r_1r_2)^{-c+\frac{N-5}{2}}\ud r_1\ud r_2\\
\quad=&\bigg\langle\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{(-2c-2i-2j+N+2)}\\
&\times\iint\limits_{0<r_1<r_2<R}r_1^{-2c-2j+N-3}r_2^{2j-1}(f(r_1)r_1^{c})(\overline{f(r_2)r_2^c})\ud r_1\ud r_2\bigg\rangle
\end{align*}
\begin{align*}
&+\bigg\langle\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{(-2c-2i-2j+N+2)}\\
&\times\iint\limits_{0<r_1<R<r_2}r_1^{-2c-2j+N-3}r_2^{2j-1}(f(r_1)r_1^{c})(\overline{f(r_2)r_2^c})\ud r_1\ud r_2\bigg\rangle\\
&+\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{(-2c-2i-2j+N+2)}R^{-2c-2i-2j+N-2}\\
&\times\int_{R}^{\infty}(f(r)r^c)r^{2i-1}\ud r\int_{R}^{\infty}(\overline{f(r)r^c})r^{2j-1}\ud r\\
=&\bigg\langle\sum\limits_{i,j=1}^{k_1-1}d_id_j\frac{(2k_1-2i-1)(2k_1-2j-1)}{(-2c-2i-2j+N-2)(-2c-2j+N-2)}\\
&\times\bigg[\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^{c}|^2+R^{-2c-2j+N-2}(f(R)R^c)\int_{R}^{\infty}(\overline{f(r)r^c})r^{2j-1}\ud r\bigg]\bigg\rangle\\
&+\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{(-2c-2i-2j+N+2)}R^{-2c-2i-2j+N-2}\\
&\times\int_{R}^{\infty}(f(r)r^c)r^{2i-1}\ud r\int_{R}^{\infty}(\overline{f(r)r^c})r^{2j-1}\ud r.
\end{align*}
From the formula,
\begin{equation*}
\int_{R}^{\infty}(f(r)r^c)r^{2j-1}\ud r=-\frac{1}{2j}\bigg[(f(R)R^c)R^{2j}+\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2j}\ud r\bigg],
\end{equation*}
we rewrite $S_1$-$S_6$ as
\begin{align}\label{S1}
S_1=&\bigg\langle-\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{2j(-2c-2i-2j+N-2)}\nonumber\\
&\times\bigg[\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2+\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\bigg]\bigg\rangle\nonumber\\
&+\bigg\langle\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{4i j(-2c-2i-2j+N-2)}R^{-2c-2j+N-2}\nonumber\\
&\times f(R)R^{c}\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\bigg\rangle\nonumber\\
&+\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{4i j}\frac{R^{-2c-2i-2j+N-2}}{-2c-2i-2j+N-2}\nonumber\\
&\times\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
&+\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{4i j}\frac{R^{-2c+N-2}}{-2c-2i-2j+N-2}\big|f(R)R^c\big|^2.
\end{align}
\begin{align}\label{S2}
S_2=&\sum\limits_{i,j=1}^{k_1-1}d_id_j\int_{0}^{R}\big|f(r)r^c\big|^2r^{-2c+N-3}\ud r=\sum\limits_{i,j=1}^{k_1-1}d_id_j\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2.
\end{align}
\begin{align}\label{S3}
S_3=\int_{0}^{R}\Big|\partial_{r}\big(f(r)r^c r^{-c+\frac{N-1}{2}}\big)\Big|^2\ud r=\Big(-c+\frac{N-1}{2}\Big)^2\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2.
\end{align}
\begin{align}\label{S4}
S_4=&-\bigg\langle\sum\limits_{i,j=1}^{k_1-1}d_id_j(2k_1-2i-1)\nonumber\\
&\times\int_{0}^{\infty}\int_{0}^{\min(r_1,R)}r_2^{-2c-2i+N-3}r_1^{2i-1}\big(f(r_1)r_1^{c}\big)\big(\overline{f(r_2)r_2^{c}}\big)\ud r_2\ud r_1\bigg\rangle\nonumber\\
=&-\bigg\langle\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)}{-2c-2i+N-2}\Big[\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\nonumber\\
&+R^{-2c-2i+N-2}\overline{f(R)R^c}\int_{R}^{\infty}\big(f(r)r^c\big)r^{2i-1}\ud r\Big]\bigg\rangle\nonumber\\
=&\bigg\langle\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)}{2i}\Big[\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\nonumber\\
&+\frac{R^{-2c-2i+N-2}}{-2c-2i+N-2}\overline{f(R)R^c}\int_{R}^{\infty}\partial_{r}\big(f(r)r^c\big)r^{2i}\ud r\Big]\bigg\rangle.
\end{align}
\begin{align*}
S_5=&\bigg\langle-\sum\limits_{i=1}^{k_1-1}d_i(2k_1-2i-1)\nonumber\\
&\times\int_{R}^{\infty}\int_{R}^{\min{(r_1,R)}}r_2^{-c-2i+\frac{N-3}{2}}r_1^{2i-1}\big(f(r_1)r_1^c\big)\partial_{r_2}\left(\overline{f(r_2)r_2^c}r_2^{-c+\frac{N-1}{2}}\right)\ud r_2\ud r_1\bigg\rangle\nonumber\\
=&\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i(2k_1-2i-1)(-c-2i+\frac{N-3}{2})\nonumber\\
&\times\int_{0}^{\infty}\int_{0}^{\min{(r_1,R)}}\big(f(r_1)r_1^c\big)r_1^{2i-1}\overline{f(r_2)r_2^c}r_2^{-2c-2i+N-3}\ud r_2\ud r_1\bigg\rangle\nonumber\\
&-\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i(2k_1-2i-1)\overline{f(R)R^c}\int_{0}^{\infty}\min(r_1,R)^{-2c-2i+N-2}\big(f(r_1)r_1^c\big)r_1^{2i-1}\ud r_1\bigg\rangle\nonumber\\
=&-\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i(2k_1-2i-1)\Big[\big|f(R)R^c\big|^2\frac{R^{-2c+N-2}}{-2c+N-2}\nonumber\\
&+R^{-2c-2i+N-2}\overline{f(R)R^c}\int_{R}^{\infty}\big(f(r)r^c\big)r^{2i-1}\ud r\Big]\bigg\rangle\\
&+\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i(2k_1-2i-1)(-c-2i+\frac{N-3}{2})\\
&\times\Big[\frac{|f(R)R^c|^2}{-2c-2i+N-2}\frac{R^{-2c+N-2}}{-2c+N-2}+\frac{R^{-2c-2i+N-2}}{-2c-2i+N-2}\overline{f(R)R^{c}}\int_{R}^{\infty}(f(r)r^c)r^{2i-1}\ud r\Big]\bigg\rangle\\
=&-\Big(-c+\frac{N-1}{2}\Big)\bigg\langle\sum\limits_{i=1}^{k_1-1}\frac{d_i(2k_1-2i-1)}{-2c-2i+N-2}\Big[\big|f(R)R^C\big|^2\frac{R^{-2c+N-2}}{-2c+N-2}\nonumber\\
&+\overline{f(R)R^c}R^{-2c-2i+N-2}\int_{R}^{\infty}\big(f(r)r^c)r^{2i-1}\ud r\Big]\bigg\rangle\nonumber
\end{align*}
\begin{align}\label{S5}
=&\Big(-c+\frac{N-1}{2}\Big)\bigg\langle\sum\limits_{i=1}^{k_1-1}\frac{d_1(2k_1-2i-1)}{2i}\Big[\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\nonumber\\
&+\frac{R^{-2c-2i+N-2}}{-2c-2i+N-2}\overline{f(R)R^c}\int_{R}^{\infty}\partial_{r}\big(f(r)r^c\big)r^{2i}\ud r\Big]\bigg\rangle.
\end{align}
\begin{align}\label{S6}
S_6=&\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i\int_{0}^{R}\big(f(r)r^c\big)r^{-c+\frac{N-3}{2}}\partial_{r}\big(\overline{f(r)r^c}r^{-c+\frac{N-1}{2}}\big)\ud r\bigg\rangle\nonumber\\
=&\Big(-c+\frac{N-1}{2}\Big)\bigg\langle\sum\limits_{i=1}^{k_1-1}d_i\big|f(R)R^c\big|^2\frac{R^{-2c+N-2}}{-2c+N-2}\bigg\rangle.
\end{align}
Direct computation shows that the terms from $S_1$ to $S_6$ can be divided into three categories based on the integral. Sorting out the formulas we got and collecting the same terms, we see that:
\medskip
{\bf The first type} The terms with $\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r$.
Recall from the assumption $k_1=\frac{(N-2)\beta+3}{4}$ and $\lfloor\frac{(N-2)\beta+4}{4}\rfloor=\lfloor k_1+\frac{1}{4}\rfloor=k_1$, we have
\begin{align*}
d_j=\frac{\prod_{1\leq \ell\leq k_1-1}(4k_1-2\ell-2j-1)}{\prod_{1\leq \ell\leq k_1-1,\ell\neq j}(2\ell-2j)}, \quad 1\leq j\leq k_1-1,
\end{align*}
and
\begin{align*}
\tilde{d_j}=\frac{\prod_{1\leq \ell\leq k_1}(4k_1-2\ell-2j+1)}{\prod_{1\leq \ell\leq k_1,\ell\neq j}(2\ell-2j)}, \quad 1\leq j\leq k_1.
\end{align*}
These facts gives
\begin{equation}\label{relationship of di}
d_j=\tilde{d}_{j+1}\frac{-2j}{2k_1-2j-1}, \quad 1\leq j\leq k_1-1.
\end{equation}
Thus we obtain terms of the first type from $S_1$ \eqref{S1},
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral1}}=&\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{4i j}\frac{R^{-2c-2i-2j+N-2}}{-2c-2i-2j+N-2}\\
&\times\int_{R}^{\infty}\partial_{r}\big(f(r)r^c\big)r^{2i}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\\
=&\sum\limits_{i',j'=2}^{k_1}\tilde{d}_i'\tilde{d}_j'\frac{R^{-2c-2i'-2j'+N-2}}{-2c-2i'-2j'+N-2}\int_{R}^{\infty}\partial_{r}\big(f(r)r^c\big)r^{2i'-2}\ud r\int_{R}^{\infty}\partial_{r}\big(\overline{f(r)r^c}\big)r^{2j'-2}\ud r.
\end{align*}
{\bf The second type} We collect the terms with $\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r$ from $S_1$, $S_4$, $S_5$, \eqref{S1}, \eqref{S4}, \eqref{S5}.
Since there are two types of terms: $\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r$ and $\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r$. Since
they are symmetric, we only treat the term $\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r$. We perform the same merge processing to the other case.
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral2}}=&-\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{2j(-2c-2i-2j+N-2)}\qquad\qquad\qquad\qquad\qquad\\
&\times\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\qquad\qquad\qquad\qquad\qquad\\
&+\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{4i j(-2c-2i-2j+N-2)}R^{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\\
&+\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2j-1)}{2j}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\\
&+\big(-c+\frac{N-1}{2}\big)\sum\limits_{j=1}^{k_1-1}\frac{d_j(2k_1-2j-1)}{2j}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\\
=&\sum\limits_{j=1}^{k_1-1}f(R)R^c R^{-2c-2j+N-2}\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\\
&\times\sum\limits_{i=1}^{k_1-1}\bigg[-\frac{d_id_j(2k_1-2i-1)(2k_2-2j-1)}{2j(-2c-2i-2j+N-2)}+\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{4i j(-2c-2i-2j+N-2)}\\
&+\frac{d_id_j(2k_1-2j-1)}{2j(-2c-2j+N-2)}\bigg]
\end{align*}
\begin{align*}
&+\big(-c+\frac{N-1}{2}\big)\sum\limits_{j=1}^{k_1-1}\frac{d_j(2k_1-2j-1)}{2j}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\\
=&\big(-c+\frac{N-1}{2}\big)\sum\limits_{j=1}^{k_1-1}f(R)R^c R^{-2c-2j+N-2}\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\\
&\times\bigg[\sum\limits_{i=1}^{k_1-1}\frac{d_id_j(2k_1-2j-1)}{4i j(-2c-2j+N-2)}+\frac{d_j(2k_1-2j-1)}{2j(-2c-2j+N-2)}\bigg].
\end{align*}
Define $i'=i+1$, $j'=j+1$, by the relationship \eqref{relationship of di}, we easily check that
\begin{align}\label{the type 2}
{\rm\uppercase\expandafter{\romannumeral2}}=&-\big(-c+\frac{N-1}{2}\big)\sum\limits_{j'=2}^{k_1}\tilde{d}_j'\bigg(1-\sum\limits_{i'=2}^{k_1}\frac{\tilde{d}_i'}{2k_1-2i'+1}\bigg)\nonumber\\
&\times \frac{R^{-2c+N-2}}{-2c+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j'-2}\ud r.
\end{align}
By \eqref{sum dj} and the fact that $2k_1=-c+\frac{N+1}{2}$, $(N-2)\beta+4=4k_1+1$, we choose $m=k_1$ and get
\begin{align}\label{relationship d1}
1-\sum\limits_{i'=2}^{k_1}\frac{\tilde{d}_i'}{2k_1-2i'+1}=\frac{\tilde{d}'_1}{2k_1-1}=\frac{\tilde{d}'_1}{-c+\frac{N-1}{2}}.
\end{align}
Finally, by substituting \eqref{relationship d1} into \eqref{the type 2}, the second type can be written as
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral2}}=&-\sum\limits_{j'=2}^{k_1}\tilde{d}_j'\tilde{d}'_1 \frac{R^{-2c-2j'+N}}{-2c-2j'+N}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j'-2}\ud r.
\end{align*}
{\bf The third type} We collect the terms with $|f(R)R^c|^2$, which come from $S_1$, $S_2$, $\cdots$, $S_6$, \eqref{S1} to \eqref{S6}.
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral3}}=&\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{4i j}\frac{R^{-2c+N-2}}{-2c-2i-2j+N-2}\big|f(R)R^c\big|^2\\
&-\bigg\langle\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)(2k_1-2j-1)}{2i(-2c-2i-2j+N-2)}\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\bigg\rangle
\end{align*}
\begin{align*}
&+\sum\limits_{i,j=1}^{k_1-1}d_id_j\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2+\big(-c+\frac{N-1}{2}\big)^2\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\\
&+\bigg\langle\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i-1)}{2i}\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\bigg\rangle\\
&+\bigg\langle\big(-c+\frac{N-1}{2}\big)\sum\limits_{i=1}^{k_1-1}\frac{d_i(2k_1-2i-1)}{2i}\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\bigg\rangle\\
&+\bigg\langle\big(-c+\frac{N-1}{2}\big)\sum\limits_{i=1}^{k_1-1}d_i\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\bigg\rangle.
\end{align*}
\noindent Taking $i'=i+1$, $j'=j+1$, we conclude by \eqref{relationship of di} and \eqref{sum dj}
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral3}}=&\sum\limits_{i',j'=2}^{k_1}\tilde{d}_{i'}\tilde{d}_{j'}\bigg[\frac{1}{-2c-2i'-2j'+N+2}\Big(1+\frac{-2j'+2}{(-2c+N-2)}+\frac{-2i'+2}{(-2c+N-2)}\Big)\\
&-\frac{-2i'+2}{(2k_1-2i'+1)(-2c+N-2)}-\frac{-2j'+2}{(2k_1-2j'+1)(-2c+N-2)}\\
&+\frac{(-2i'+2)(-2j'+2)}{(2k_1-2i'+1)(2k_1-2j'+1)(-2c+N-2)}\bigg]R^{-2c+N-2}\big|f(R)R^c\big|^2\\
&-(-c+\frac{N-1}{2})\sum\limits_{i'=2}^{k_1}\tilde{d}_{i'}\Big(1+\frac{2i'-2}{2k_1-2i'+1}\Big)\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\\
&-(-c+\frac{N-1}{2})\sum\limits_{j'=2}^{k_1}\tilde{d}_{j'}\Big(1+\frac{2j'-2}{2k_1-2j'+1}\Big)\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\\
&+(-c+\frac{N-1}{2})^2\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\\
=&(-c+\frac{N-1}{2})^2\bigg(\sum\limits_{i',j'=2}^{k_1}\frac{\tilde{d}_{i'}\tilde{d}_{j'}}{(2k_1-2i'+1)(2k_1-2j'+1)}-\sum\limits_{i'=2}^{k_1}\frac{\tilde{d}_{i'}}{2k_1-2i'+1}\\
&-\sum\limits_{j'=2}^{k_1}\frac{\tilde{d}_{j'}}{2k_1-2j'+1}+1\bigg)\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\\
=&(2k_1-1)^2\bigg[1-\sum\limits_{i'=2}^{k_1}\frac{\tilde{d}_{i'}}{2k_1-2i'+1}\bigg]\bigg[1-\sum\limits_{j'=2}^{k_1}\frac{\tilde{d}_{j'}}{2k_1-2j'+1}\bigg]\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2\\
=&\tilde{d}_{1'}\tilde{d}_{1'}\frac{R^{-2c+N-2}}{-2c+N-2}\big|f(R)R^c\big|^2.
\end{align*}
Summing up all the estimates and \eqref{projection of f}, we obtain that for $k-1$ is even
\begin{align*}
AS_{N}^{a}f
=&\int_{R}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2r^{N-1}\ud r -\sum\limits_{i,j=1}^{k_1}\tilde{d}_i\tilde{d}_j\frac{R^{-2c-2i-2j+N-2}}{-2c-2i-2j+N-2}\\
&\times\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i-2}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j-2}\ud r\\
=&\frac{1}{2}\big\|\operatorname{\pi}_{\tilde{W}^{\perp}}f\big\|^{2}_{\dot{H}^1_a(r\geq R,r^{N-1}\ud r)}.
\end{align*}
\subsubsection{\bf Case $k-1=2k_1-1$} Now we deal with the case that $k-1$ is odd, where the dimension of $\tilde{W}$ is $\tilde{k}_1=\lfloor\frac{(N-2)\beta}{4}\rfloor=\lfloor k_1+\frac{3}{4}\rfloor=k_1$ and $2k_1=k=-c+\frac{N-1}{2}$. Relying on the Fourier expansion of $\psi_{k-1}$ \eqref{odd Fourier psi} and the notation $F(r_1,r_2):=f(r_1)r_1^c\overline{f(r_2)r_2^c}(r_1r_2)^{-c+\frac{N-5}{2}}$, we infer that
\begin{align*}
\quad&2AS_{N}^{a}f(r)\\
=&\int_0^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^{2}r^{N-1}\ud r-\lim\limits_{\varepsilon\rightarrow0}\iint\limits_{r_1,r_2>0}\Big[\frac{1}{2\pi^2r_1^2r_2^2}\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast\mathcal{F}\psi_{2k_1-1}\big]\big(\frac{\xi}{r_1}\big)\\
&\times\overline{\big[\kappa_{\varepsilon}\ast\mathcal{F}\psi_{2k_1-1}\big]\big(\frac{\xi}{r_2}\big)}\ud \xi\bigg] f(r_1)\overline{f(r_2)}(r_1r_2)^{\frac{N-1}{2}}\ud r_1\ud r_2
\end{align*}
\begin{align*}
=&\int_{0}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^{2}r^{N-1}\ud r-\frac{1}{2}\sum\limits_{i,j=1}^{k_1}d_{i}d_{j}(2k_1-2i)(2k_1-2j)\\
&\times\lim\limits_{\varepsilon\rightarrow0}\iint\bigg[\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast\xi^{2k_1-2i-1}\chi_{(-1,1)}\big]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast\xi^{2k_1-2j-1}\chi_{(-1,1)}\big]\big(\frac{\xi}{r_2}\big)\ud \xi\bigg]\\
&\times F(r_1,r_2)\ud r_1\ud r_2\\
&-\frac{1}{2}\sum\limits_{i,j=1}^{k_1}d_id_j\lim\limits_{\varepsilon\rightarrow0}\iint\bigg[\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast(-\delta_{1}+\delta_{-1})\big]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast(-\delta_{1}+\delta_{-1})\big]\big(\frac{\xi}{r_2}\big)\bigg]\\
&\times F(r_1,r_2)\ud r_1\ud r_2\\
&-\frac{1}{2}\lim\limits_{\varepsilon\rightarrow0}\iint\Big[\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast(-\delta'_{1}-\delta'_{-1})\big]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast(-\delta'_{1}-\delta'_{-1})\big]\big(\frac{\xi}{r_2}\big)\ud\xi\Big]F(r_1,r_2)\ud r_1\ud r_2\\
&-\frac{1}{2}\bigg\langle\sum\limits_{i,j=1}^{k_1}(2k_1-2i)d_id_j\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast\xi^{2k_1-2i-1}\chi_{(-1,1)}\big]\big(\frac{\xi}{r_1}\big)\\
&\times\big[\kappa_{\varepsilon}\ast(-\delta_{1}+\delta_{-1})\big]\big(\frac{\xi}{r_2}\big)\ud\xi F(r_1,r_2)\ud r_1\ud r_2\bigg\rangle\\
&-\frac{1}{2}\bigg\langle\sum\limits_{i=1}^{k_1}(2k_1-2i)d_i\lim\limits_{\varepsilon\rightarrow0}\iint\int_{-R}^{R}\big[\kappa_{\varepsilon}\ast\xi^{2k_1-2i-1}\chi_{(-1,1)}\big]\big(\frac{\xi}{r_1}\big)\big[\kappa_{\varepsilon}\ast(-\delta'_{1}-\delta'_{-1})\big]\big(\frac{\xi}{r_2})\ud \xi\\
&\times F(r_1,r_2)\ud r_1\ud r_2\bigg\rangle\\
&-\frac{1}{2}\bigg\langle\sum\limits_{i=1}^{k_1}d_i\lim\limits_{\varepsilon\rightarrow0}\iint\Big[\int_{-R}^{R}[\kappa_{\varepsilon}\ast(\delta_{1}+\delta_{-1})](\frac{\xi}{r_1})[\kappa_{\varepsilon}\ast(-\delta'_{1}-\delta'_{-1})](\frac{\xi}{r_2})\ud \xi\Big]\\
&\times F(r_1,r_2)\ud r_1\ud r_2\bigg\rangle\\
=&\int_{0}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2r^{N-1}\ud r-\sum\limits_{i,j=1}^{k_1}d_id_j\iint\limits_{r_1,r_2>0}\frac{(2k_1-2i)(2k_1-2j)}{4k_1-2i-2j-1}\\
&\times\frac{\min(r_1,r_2,R)^{4k_1-2i-2j-1}}{r_1^{2k_1-2i-1}r_2^{2k_1-2j-1}}f(r_1)r_1^{c}\overline{f(r_2)r_2^c}(r_1r_2)^{-c+\frac{N-5}{2}}\ud r_1\ud r_2\\
&-\sum\limits_{i,j=1}^{k_1}d_id_j\int_{0}^{R}|f(r)r^c|^2r^{-2c+N-3}\ud r-\int_{0}^{R}\big|\partial_{r}(f(r)r^c\cdot r^{-c+\frac{N-1}{2}})\big|^2\ud r\\
&+\bigg\langle\sum\limits_{i,j=1}^{k_1}d_id_j(2k_1-2i)\int_{0}^{\infty}\int_{0}^{\min(r_1,R)}r_2\big(\frac{r_2}{r_1}\big)^{2k_1-2i-1}f(r_1)r_1^{c}\overline{f(r_2)r_2^{c}}(r_1r_2)^{-c+\frac{N-5}{2}}\ud r_2\ud r_1\bigg\rangle
&+\bigg\langle\sum\limits_{i=1}^{k_1}d_i(2k_1-2i)\int_{0}^{\infty}\int_{0}^{\min(r_1,R)}\big(\frac{r_2}{r_1}\big)^{2k_1-2i-1}f(r_1)r_1^{c}r_1^{-c+\frac{N-5}{2}}\partial_{r_2}\big(\overline{f(r_2)r_2^{c}}r_2^{-c+\frac{N-1}{2}}\big)\ud r_2\ud r_1\bigg\rangle\\
&-\bigg\langle\sum\limits_{i=1}^{k_1}d_i\int_{0}^{R}f(r)r^{c}r^{-c+\frac{N-3}{2}}\partial_{r}\big(\overline{f(r)r^c}r^{-c+\frac{N-1}{2}}\big)\ud r\bigg\rangle\\
=&\int_{0}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2r^{N-1}\ud r-(\tilde{S}_1+\tilde{S}_2+\tilde{S}_3+\tilde{S}_4+\tilde{S}_5+\tilde{S}_6).
\end{align*}
As before, we recall the relationship $2k_1=-c+\frac{N-1}{2}$ and assumption $f(r)r^c=f(R)R^c$ for $r\leq R$, we compute each $\tilde{S}_i$:
\begin{align*}
\tilde{S}_1=&\bigg\langle\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{(-2c-2i-2j+N-2)}\iint\limits_{0<r_1<r_2<R}r_1^{-2c-2j+N-3}r_2^{2j-1}(f(r_1)r_1^{c})(\overline{f(r_2)r_2^c})\ud r_1\ud r_2\bigg\rangle\\
&+\bigg\langle\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{(-2c-2i-2j+N-2)}\iint\limits_{0<r_1<R<r_2}r_1^{-2c-2j+N-3}r_2^{2j-1}(f(r_1)r_1^{c})(\overline{f(r_2)r_2^c})\ud r_1\ud r_2\bigg\rangle\\
&+\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{(-2c-2i-2j+N-2)}R^{-2c-2i-2j+N-2}\int_{R}^{\infty}(f(r)r^c)r^{2i-1}\ud r\int_{R}^{\infty}(\overline{f(r)r^c})r^{2j-1}\ud r\\
=&\bigg\langle\sum\limits_{i,j=1}^{k_1}d_id_j\frac{(2k_1-2i)(2k_1-2j)}{(-2c-2i-2j+N-2)(-2c-2j+N-2)}\\
&\times\Big[\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^{c}|^2+R^{-2c-2j+N-2}(f(R)R^c)\int_{R}^{\infty}(\overline{f(r)r^c})r^{2j-1}\ud r\Big]\bigg\rangle\\
&+\sum\limits_{i,j=1}^{k_1-1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{(-2c-2i-2j+N-2)}R^{-2c-2i-2j+N-2}\int_{R}^{\infty}(f(r)r^c)r^{2i-1}\ud r\int_{R}^{\infty}(\overline{f(r)r^c})r^{2j-1}\ud r.
\end{align*}
\noindent Integration by parts implies
\begin{equation*}
\int_{R}^{\infty}(\overline{f(r)r^c)}r^{2j-1}\ud r=-\frac{1}{2j}\Big[(\overline{f(R)R^c})R^{2j}+\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\Big],
\end{equation*}
so we have
\begin{align}\label{SS1}
\tilde{S}_1=&\bigg\langle-\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{2j(-2c-2i-2j+N-2)}\nonumber\\
&\times\Big[\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2+\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\Big]\bigg\rangle\nonumber\\
&+\bigg\langle\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{4i j(-2c-2i-2j+N-2)}R^{-2c-2j+N-2}f(R)R^{c}\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\bigg\rangle\nonumber\\
&+\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{4i j}\frac{R^{-2c-2i-2j+N-2}}{-2c-2i-2j+N-2}\nonumber\\
&\times\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
&+\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{4i j}\frac{R^{-2c+N-2}}{-2c-2i-2j+N-2}|f(R)R^c|^2.
\end{align}
\begin{align}\label{SS2}
\tilde{S}_2=&\sum\limits_{i,j=1}^{k_1}d_id_j\int_{0}^{R}|f(r)r^c|^2r^{-2c+N-3}\ud r=\sum\limits_{i,j=1}^{k_1}d_id_j\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2.
\end{align}
\begin{align}\label{SS3}
\tilde{S}_3=\int_{0}^{R}\left|\partial_{r}(f(r)r^c r^{-c+\frac{N-1}{2}})\right|^2\ud r=\big(-c+\frac{N-1}{2}\big)^2\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2.
\end{align}
\begin{align}\label{SS4}
\tilde{S}_4=&-\bigg\langle\sum\limits_{i,j=1}^{k_1}d_id_j(2k_1-2i)\int_{0}^{\infty}\int_{0}^{\min(r_1,R)}r_2^{-2c-2i+N-3}r_1^{2i-1}(f(r_1)r_1^{c})(\overline{f(r_2)r_2^{c}})\ud r_2\ud r_1\bigg\rangle\nonumber\\
=&-\bigg\langle\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)}{-2c-2i+N-2}\Big[\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\nonumber\\
&\qquad+R^{-2c-2i+N-2}\overline{f(R)R^c}\int_{R}^{\infty}(f(r)r^c)r^{2i-1}\ud r\Big]\bigg\rangle\nonumber\\
=&\bigg\langle\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)}{2i}\Big[\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\nonumber\\
&\quad+\frac{R^{-2c-2i+N-2}}{-2c-2i+N-2}\overline{f(R)R^c}\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r\Big]\bigg\rangle.
\end{align}
\begin{align}\label{SS5}
\tilde{S}_5=&\bigg\langle\sum\limits_{i=1}^{k_1}d_i(2k_1-2i)(-c-2i+\frac{N-3}{2})\nonumber\\
&\times\int_{0}^{\infty}\int_{0}^{\min{(r_1,R)}}(f(r_1)r_1^c)r_1^{2i-1}\overline{f(r_2)r_2^c}r_2^{-2c-2i+N-3}\ud r_2\ud r_1\bigg\rangle\nonumber\\
&-\bigg\langle\sum\limits_{i=1}^{k_1}d_i(2k_1-2i)\overline{f(R)R^c}\int_{0}^{\infty}(f(r_1)r_1^c)r_1^{2i-1}\min(r_1,R)^{-2c-2i+N-2}\ud r_1\bigg\rangle\nonumber\\
=&-\big(-c+\frac{N-1}{2}\big)\bigg\langle\sum\limits_{i=1}^{k_1}\frac{d_i(2k_1-2i)}{-2c-2i+N-2}\Big[|f(R)R^C|^2\frac{R^{-2c+N-2}}{-2c+N-2}\nonumber\\
&+\overline{f(R)R^c}R^{-2c-2i+N-2}\int_{R}^{\infty}(f(r)r^c)r^{2i-1}\ud r\Big]\bigg\rangle\nonumber\\
=&\big(-c+\frac{N-1}{2}\big)\bigg\langle\sum\limits_{i=1}^{k_1}\frac{d_1(2k_1-2i)}{2i}\Big[\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\nonumber\\
&+\frac{R^{-2c-2i+N-2}}{-2c-2i+N-2}\overline{f(R)R^c}\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r\Big]\bigg\rangle.
\end{align}
\begin{align}\label{SS6}
\tilde{S}_6=&\big(-c+\frac{N-1}{2}\big)\bigg\langle\sum\limits_{i=1}^{k_1}d_i|f(R)R^c|^2\frac{R^{-2c+N-2}}{-2c+N-2}\bigg\rangle.
\end{align}
Similarly, we sort out and collect the same terms in the above formulas.
{\bf Type ${\rm\uppercase\expandafter{\romannumeral1}}$}\quad The terms with $\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r$.
The assumption $k_1=\frac{(N-2)\beta+1}{4}$ and $\lfloor\frac{(N-2)\beta+4}{4}\rfloor=\lfloor k_1+\frac{3}{4}\rfloor=k_1$ yields
\begin{align*}
d_j=\frac{\prod_{1\leq \ell\leq k_1}(4k_1-2\ell-2j+1)}{\prod_{1\leq \ell\leq k_1-1,\ell\neq j}(2\ell-2j)}, \quad 1\leq j\leq k_1,\\
\tilde{d_j}=\frac{\prod_{1\leq \ell\leq k_1}(4k_1-2\ell-2j+3)}{\prod_{1\leq \ell\leq k_1,\ell\neq j}(2\ell-2j)}, \quad 1\leq j\leq k_1,
\end{align*}
hence
\begin{equation}\label{relationship of ddi}
d_j=\tilde{d}_{j+1}\frac{-2j}{2k_1-2j}, \quad 1\leq j\leq k_1-1.
\end{equation}
Since the term vanishes when the $i,j$ are equal to $k_1$, the summation is from $1$ to $k_1-1$. We preform the change of indices $i'=i+1$ and $j'=j+1$, according to \eqref{relationship of ddi} and the first type of term \eqref{SS1} from $\tilde{S}_1$, we conclude
\begin{align}\label{odd type first}
{\rm\uppercase\expandafter{\romannumeral1}}=&\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{4i j}\frac{R^{-2c-2i-2j+N-2}}{-2c-2i-2j+N-2}\nonumber\\
&\times\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
=&\sum\limits_{i',j'=2}^{k_1}\tilde{d}_i'\tilde{d}_j'\frac{R^{-2c-2i'-2j'+N-2}}{-2c-2i'-2j'+N-2}\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i'-2}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j'-2}\ud r.
\end{align}
\medskip
{\bf Type ${\rm\uppercase\expandafter{\romannumeral2}}$}\quad As before we only collect the terms with $\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i}\ud r$ from $\tilde{S}_1$, $\tilde{S}_4$, $\tilde{S}_5$, \eqref{SS1}, \eqref{SS4}, \eqref{SS5}, the other terms are symmetric.
\begin{align*}
{\rm \uppercase\expandafter{\romannumeral2}}=&-\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{2j(-2c-2i-2j+N-2)}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
&+\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{4i j(-2c-2i-2j+N-2)}R^{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
&+\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2j)}{2j}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
&+\big(-c+\frac{N-1}{2}\big)\sum\limits_{j=1}^{k_1}\frac{d_j(2k_1-2j)}{2j}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
=&2k_1\sum\limits_{j=1}^{k_1}d_j\Big(\sum\limits_{i=1}^{k_1}\frac{d_i}{2i}+1\Big)\frac{2k_1-2j}{2j}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r.
\end{align*}
Calculate the summation term by \eqref{prod cj}
\begin{align}\label{prod important}
1+\sum\limits_{i=1}^{k_1}\frac{d_i}{2i}&=\prod\limits_{\ell=1}^{k_1}\frac{4k_1-2\ell+1}{2\ell}=\frac{1}{2k_1}\frac{\prod_{\ell=1}^{k_1}(4k_1-2\ell+1)}{\prod_{\ell=2}^{k_1}(2\ell-2)}=\frac{\tilde{d}_1}{2k_1}.
\end{align}
Since $2k_1=-c+\frac{N-1}{2}$, the second type can be written equivalently as
\begin{align}\label{odd type second}
{\rm\uppercase\expandafter{\romannumeral2}}=&\sum\limits_{j=1}^{k_1}d_j\tilde{d}_1\frac{2k_1-2j}{2j}\frac{R^{-2c-2j+N-2}}{-2c-2j+N-2}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j}\ud r\nonumber\\
=&-\sum\limits_{j'=2}^{k_1}\tilde{d}_j'\tilde{d}'_1 \frac{R^{-2c-2j'+N}}{-2c-2j'+N}f(R)R^c\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j'-2}\ud r.
\end{align}
{\bf Type ${\rm\uppercase\expandafter{\romannumeral3}}$}\quad We collect the terms with $|f(R)R^c|^2$, which come from $\tilde{S}_1$, $\tilde{S}_2$, $\cdots$, $\tilde{S}_6$, \eqref{SS1} to \eqref{SS6}.
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral3}}=&\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{4i j}\frac{R^{-2c+N-2}}{-2c-2i-2j+N-2}|f(R)R^c|^2\\
&-\bigg\langle\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)(2k_1-2j)}{2i(-2c-2i-2j+N-2)}\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\bigg\rangle\\
&+\sum\limits_{i,j=1}^{k_1}d_id_j\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2+\big(-c+\frac{N-1}{2}\big)^2\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\\
&+\bigg\langle\sum\limits_{i,j=1}^{k_1}\frac{d_id_j(2k_1-2i)}{2i}\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\bigg\rangle\\
&+\bigg\langle\big(-c+\frac{N-1}{2}\big)\sum\limits_{i=1}^{k_1}\frac{d_i(2k_1-2i)}{2i}\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\bigg\rangle\\
&+\bigg\langle\big(-c+\frac{N-1}{2}\big)\sum\limits_{i=1}^{k_1}d_i\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2\bigg\rangle\\
=&\big(-c+\frac{N-1}{2}\big)^2\Big[\sum\limits_{i=1}^{k_1}\frac{d_i}{2i}+1\Big]\Big[\sum
\limits_{j=1}^{k_1}\frac{d_j}{2j}+1\Big]\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2.
\end{align*}
Using the identity \eqref{prod important}, and taking $i'=i+1$, $j'=j+1$, we rewrite
\begin{align}\label{odd typy three}
{\rm\uppercase\expandafter{\romannumeral3}}=&\tilde{d'_1}^2\frac{R^{-2c+N-2}}{-2c+N-2}|f(R)R^c|^2.
\end{align}
In conclusion, using the identities \eqref{odd type first}, \eqref{odd type second} and \eqref{odd typy three} we infer that when $k-1$ is odd
\begin{align*}
AS_{N}^{a}f
=&\int_{R}^{\infty}\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2r^{N-1}\ud r -\sum\limits_{i,j=1}^{k_1}\tilde{d}_i\tilde{d}_j\frac{R^{-2c-2i-2j+N-2}}{-2c-2i-2j+N-2}\\
&\times\int_{R}^{\infty}\partial_{r}(f(r)r^c)r^{2i-2}\ud r\int_{R}^{\infty}\partial_{r}(\overline{f(r)r^c})r^{2j-2}\ud r\\
=&\frac{1}{2}\big\|\operatorname{\pi}_{\tilde{W}^{\perp}}f\big\|^{2}_{\dot{H}^1_a(r\geq R,r^{N-1}\ud r)}.
\end{align*}
\medskip
Combining all the identities for data $(0,g)$ in Section~\ref{SS:2.4} and for $(f,0)$ in Section \ref{SS:2.5}, the proof of Theorem~\ref{ce} is completed.
\medskip
\section{Channel of energy for the linearized wave equation with inverse-square potential}\label{4}
\subsection{Exterior energy lower bound for the linearized wave equation around the ground state.}
We consider the linearized equation around the ground state
\begin{align}\label{eq:liearized wave}
\partial_{t}^{2}u+\mathcal{L}_{W}^{a}u=0,
\end{align}
where $\mathcal{L}_{W}^{a}$ is the linearized operator
\begin{align*}
\mathcal{L}_{W}^{a}=-\Delta-\frac{N+2}{N-2}W_{a}^{\frac{4}{N-2}}+\frac{a}{|x|^2}.
\end{align*}
Define:
\begin{align*}
\mathcal{Z}=\big\{Z\in\dot{H}_{a}^1:\mathcal{L}_{W}^a Z=0\big\}.
\end{align*}
Since \eqref{weq} is invariant by scaling translation, let
\begin{align*}
\Lambda W_{a}\triangleq x\cdot\nabla W_{a}+\frac{N-2}{2}W_{a},
\end{align*}
we have the following nondegeneracy property of $W_a$
\begin{align*}
\mathcal{Z}=\operatorname{span}\{\Lambda W_{a}\}.
\end{align*}
Since we assume $k=\frac{(N-2)\beta+1}{2}\in\mathbb{Z}$, $(N-2)\beta$ is an odd integer. Note that for $(N-2)\beta\geq 1$, $\Lambda W_{a}\in \dot{H}_{a}^1$ and that $\Lambda W_a\in L^2$ if and only if $(N-2)\beta\geq 3$.
Define
\begin{align}\label{def: mathbf mathcal Z}
\mathbf{\mathcal{Z}}=(\mathcal{Z}\times\mathcal{Z})\cap\mathcal{H}_a,
\end{align}
which is a finite dimensional subspace of $\mathcal{H}_a$. Note that if $u$ is a solution of \eqref{eq:liearized wave} with initial data $(u_0,u_1)\in\mathbf{\mathcal{Z}}$, then $u(t,x)=u_0(x)+t u_1(x)$.
Let us denote by $\operatorname{\pi}_{\mathbf{\mathcal{Z}}}(R)$ the orthogonal projection on $\mathbf{\mathcal{Z}}$ \eqref{def: mathbf mathcal Z} in the space $(\dot{H}_{a}^{1}\times L^2)(r\geq R,r^{N-1}\ud r)$. More precisely, for every $R>0$,
\begin{align}\label{def: projection of initial f g}
\operatorname{\pi}_{\mathbf{\mathcal{Z}}(R)}(f,g)=\frac{\int_{R}^{\infty}\left(\partial_{r}+\frac{c}{r}\right)f(r)\left(\partial_{r}+\frac{c}{r}\right)(\Lambda W_{a})r^{N-1}\ud r}{\int_{R}^{\infty}\left(\left(\partial_{r}+\frac{c}{r}\right)\Lambda W_{a}\right)^{2}r^{N-1}\ud r}\left(\Lambda W_{a},0\right).
\end{align}
The projection onto the orthogonal complement is
\begin{align*}
\operatorname{\pi}^{\bot}_{\mathbf{\mathcal{Z}}(R)}(f,g)=(f,g)-\operatorname{\pi}_{\mathbf{\mathcal{Z}}(R)}(f,g).
\end{align*}
We perform a spectral analysis associated with the operator $\mathcal{L}_{W}^{a}$ to find the solution of \eqref{eq:liearized wave} and establish the distorted Hankel transform corresponding to $\mathcal{L}_{W}^{a}$ on $\mathbb{R}^{N}$. Restricted to the subspace of radial functions, $\mathcal{L}_{W}^a$ can be represented as
\begin{align*}
\mathcal{L}_{W}^{a}=-\left(\partial_{r}^{2}+\frac{N-1}{r}\partial_{r}\right)+\frac{a}{r^{2}}+V(r),
\end{align*}
where $V(r)=-\frac{N+2}{N-2}W_{a}^{\frac{4}{N-2}}(r)$ is a smooth real valued even function of $r$ which decays with order $r^{-2(1+\beta)}$ as $r\rightarrow\infty$. Since the operator $\mathcal{L}_{W}^{a}$ admits absolutely continuous spectrum given by $[0,\infty)$, in order to define the distorted Hankel transformation, we consider $\mathcal{L}_{W}^{a}f=Ef$. Assume $E=\rho^{2}$, $\rho>0$ and let $\lambda=\rho r$, $\tilde{f}(r,\rho)=\lambda^{-\frac{N-2}{2}}\tilde{g}(\lambda)$. Then one can show that
\begin{align}\label{eq:spectrum equation}
\tilde{g}''(\lambda)+\frac{1}{\lambda}\tilde{g}'(\lambda)+\left(1-\frac{\nu^2}{\lambda^2}-\frac{N(N-2)\beta^2}{\lambda^2}\Big(\frac{\lambda^{\beta}\rho^{\beta}}{\lambda^{2\beta}+\rho^{2\beta}}\Big)^2\right)\tilde{g}(\lambda)=0,
\end{align}
where $\nu=\frac{(N-2)\beta}{2}$ is the order of the Bessel function $J_{\nu}$. The solution to $\eqref{eq:spectrum equation}$ with prescribed asymptotics at infinity can be constructed by the following lemma.
\begin{lemma}\label{lem:solution of spectrum}
For all $\rho\geq0$, there exists a real solution $\tilde{f}(r,\rho)$ to the equation
\begin{align*}
\mathcal{L}_{W}^{a}\tilde{f}(\cdot,\rho)=\rho^2\tilde{f}(\cdot,\rho),
\end{align*}
such that $\tilde{f}$ is a smooth function of $(r,\rho)\in\mathbb{R}\times\mathbb{R}_{+}$ of the form
\begin{align}\label{asymptotic of solution}
\tilde{f}(r,\rho)=(r\rho)^{-\frac{N-2}{2}}J_{\nu}(r\rho)+O\left((r\rho)^{-\frac{N+1}{2}}\right),\qquad as~r\rightarrow\infty.
\end{align}
\end{lemma}
\begin{proof}
Define the potential by
\begin{align*}
\tilde{V}(r)=\frac{N(N-2)\beta^2}{\lambda^2}\Big(\frac{\lambda^{\beta}\rho^{\beta}}{\lambda^{2\beta}+\rho^{2\beta}}\Big)^2.
\end{align*}
Then the equation \eqref{eq:spectrum equation} can be rewritten as
\begin{align*}
\tilde{g}''(\lambda)+\frac{1}{\lambda}\tilde{g}'(\lambda)+\Big(1-\frac{\nu^2}{\lambda^2}\Big)\tilde{g}(\lambda)=\tilde{V}(r)\tilde{g}(\lambda),
\end{align*}
which can be written as equivalent integral form
\begin{align*}
\tilde{g}(r,\rho)=J_{\nu}(\lambda)+&\int_{\lambda}^{\infty}{J_{\nu}}(\lambda-s)\tilde{V}(s)\tilde{g}(s)\ud s.
\end{align*}
Since for $\lambda\leq\rho$,
\begin{align*}
\left[\frac{\lambda^{\beta-1}\rho^{\beta}}{\rho^{2\beta}+\lambda^{2\beta}}\right]^2=\lambda^{-2\beta-2}\bigg[\frac{\rho^{\beta}}{1+\left(\frac{\rho}{\lambda}\right)^{2\beta}}\bigg]^2\approx\lambda^{-2}\left(\frac{\lambda}{\rho}\right)^{2\beta}\leq\lambda^{-2},
\end{align*}
and similarly, for $\lambda\geq\rho$,
\begin{align*}
\left[\frac{\lambda^{\beta-1}\rho^{\beta}}{\rho^{2\beta}+\lambda^{2\beta}}\right]^2\lesssim\lambda^{-4\beta}\lambda^{2\beta-2}\rho^{2\beta}=\lambda^{-2}\left(\frac{\rho}{\lambda}\right)^{2\beta}\leq\lambda^{-2},
\end{align*}
we obtain
\begin{align*}
|\tilde{V}(\lambda)|\leq C\lambda^{-2}.
\end{align*}
Meanwhile, the asymptotic series expansion of $J_{\nu}(z)$ in \eqref{eq:J} yields
\begin{align*}
|J_{\nu}(\lambda-s)|\leq C|\lambda-s|^{-\frac{1}{2}},
\end{align*}
and $\tilde{g}$ solves a Volterra integral equation hence $\sup\limits_{\lambda\in\mathbb{R}}|\tilde{g}(\lambda)|\leq C(\tilde{V})$. Therefore,
\begin{align*}
|\tilde{g}(\lambda)|\leq&J_{\nu}(\lambda)+C(V)\int_{\lambda}^{\infty}(\lambda-s)^{-\frac{1}{2}}s^{-2}\ud s\leq J_{\nu}(\lambda)+C(V)\lambda^{-\frac{3}{2}}.
\end{align*}
\noindent Thus, we get
\begin{align*}
\tilde{f}(r,\rho)=(r\rho)^{-\frac{N-2}{2}}J_{\nu}(r\rho)+O\left((r\rho)^{-\frac{N+1}{2}}\right).
\end{align*}
This concludes the proof of Lemma \ref{lem:solution of spectrum}.
\end{proof}
\medskip
Now, we define the distorted Hankel transform $H: L^{2}(\mathbb{R}^{N})\rightarrow L^{2}(\mathbb{R}_{+})$ corresponding to the operator $\mathcal{L}_{W}^{a}$.
\begin{align}\label{def:distorted hankel transform}
(H\varphi){(\rho)}=\frac{1}{(2\pi)^{\frac{N}{2}}}\int_{0}^{\infty}\tilde{f}(r,\rho)\varphi(r)\ud r,\qquad\qquad \varphi\in L_{rad}^{2}(\mathbb{R}^{N}).
\end{align}
Then we prove an exterior energy bound for the linearized equation (\ref{eq:liearized wave}) around the ground state.
\medskip
\begin{theorem}\label{thm: linearized wave}
Assume $(N-2)\beta\geq1$ and $(N-2)\beta$ is an odd integer. Then there exists a constant $C>0$ and $R>0$, such that for the radial solution to \eqref{eq:liearized wave} with initial data $(f,g)\in(\dot{H_{a}^{1}}\times L^{2})(\mathbb{R}^N)$, it holds that
\begin{align*}
\sum\limits_{\pm}\lim\limits_{t\rightarrow\pm\infty}\int_{|x|\geq t+R}\big|\nabla_{t,x}u(t,x)\big|^{2}\ud x\geq C\big\|\operatorname{\pi}_{\mathbf{\mathcal{Z}}}^{\perp}(f,g)\big\|_{(\dot{H_{a}^{1}}\times L^2)(r\geq R,r^{N-1}\ud r)}.
\end{align*}
where the space $\mathbf{\mathcal{Z}}$ is
\begin{equation*}
\mathbf{\mathcal{Z}}=\left\{
\begin{aligned}
&\operatorname{span}\{(\Lambda W_{a},0)\}\qquad &\text{if}~(N-2)\beta=1\\
&\operatorname{span}\{(\Lambda W_{a},0),(0,\Lambda W_{a})\}\qquad &\text{if}~(N-2)\beta\geq 3
\end{aligned}
\right.
\end{equation*}
\end{theorem}
\begin{proof}
Taking the distorted Hankel transform of the both sides of \eqref{eq:liearized wave}
\begin{equation}\label{eq:linearized spectrum}
\left\{
\begin{aligned}
&\partial_{t}^{2}H u-\rho^2H u=0,\\
&H u(0)=Hf,\quad\partial_{t}(H u)(0)=Hg.
\end{aligned}
\right.
\end{equation}
Then the solution $H u(t,\rho)$ for \eqref{eq:linearized spectrum} with initial data $(Hf,Hg)$ is given by
\begin{align*}
H u(t,\rho)=\cos(t\rho)Hf(\rho)+\frac{\sin(t\rho)}{\rho}Hg{(\rho)}.
\end{align*}
Note that if $f$, $g$ are radial functions, then $Hf$ and $Hg$ are also radial. Therefore, from \eqref{def:distorted hankel transform}, we get
\begin{align*}
u(t,\rho)=&\frac{1}{(2\pi)^{\frac{N}{2}}}\int_{0}^{\infty}\tilde{f}(r,\rho)H u(\rho)\ud\rho\\
=&\frac{1}{(2\pi)^{\frac{N}{2}}}\int_{0}^{\infty}\tilde{f}(r,\rho)\Big(\cos(t\rho)Hf(\rho)+\frac{\sin(t\rho)}{\rho}Hg{(\rho)}\Big)\ud\rho,
\end{align*}
and
\begin{align*}
\partial_{t}u(r,\rho)=&\frac{1}{(2\pi)^{\frac{N}{2}}}\int_{0}^{\infty}\tilde{f}(r,\rho)\Big(-\rho\sin(t\rho)Hf(\rho)+\cos(t\rho)Hg{(\rho)}\Big)\ud\rho,
\end{align*}
where the standard asymptotic for $\tilde{f}(r,\rho)$ is given by \eqref{asymptotic of solution}.\par
Now we start to compute the asymptotic form of the exterior energy. On the one hand, by \cite{KLLS}
\begin{align}\label{eq:lower bdd of u}
&(2\pi)^{N}\Big(\|\partial_{t}u(t,r)\|_{L^2(r>R+t, r^{N-1})}+\|(\partial_{r}+\frac{c}{r})u(t,r)\|_{L^2(r>R+t, r^{N-1})}\Big)\nonumber\\
\geq&\frac{1}{2}\int_{R}^{\infty}\Big(\Big|(\partial_{r}+\frac{c}{r})f(r)\Big|^2+\left|g(r)\right|^2\Big)r^{N-1}\ud r-|f(R)|^2R^{N-2}.
\end{align}
On the other hand, recalling the definition of $\operatorname{\pi}_{\mathcal{Z}}^{\bot}$ \eqref{def: projection of initial f g}, we obtain
\begin{align*}
&\left\|\operatorname{\pi}_{\mathcal{Z}}^{\bot}(f,g)\right\|_{\dot{H}_{a}^{1}\times L^2(r\geq R,r^{N-1}\ud r)}\\
=&\bigg\|\frac{\int_{R}^{\infty}\left(\partial_{r}+\frac{c}{r}\right)f(r)\left(\partial_{r}+\frac{c}{r}\right)(\Lambda W_{a})r^{N-1}\ud r}{\int_{R}^{\infty}\left(\left(\partial_{r}+\frac{c}{r}\right)\Lambda W_{a}\right)^{2}r^{N-1}\ud r}(\Lambda W_{a},0)\bigg\|_{\dot{H}_{a}^{1}\times L^2(r\geq R,r^{N-1}\ud r)}\\
=&\frac{\left|\int_{R}^{\infty}\left(\partial_{r}+\frac{c}{r}\right)f(r)\left(\partial_{r}+\frac{c}{r}\right)(\Lambda W_{a})r^{N-1}\ud r\right|^{2}}{\|\Lambda W_{a}\|_{\dot{H}_{a}^{1}(r\geq R,r^{N-1}\ud r)}},
\end{align*}
where
\begin{align*}
\Lambda W_{a}&=r\cdot\nabla W_{a}+\frac{N-2}{2}W_{a}\\
&=\frac{N-2}{2}\beta(N(N-2)\beta^2)^{\frac{N-2}{4}}\Big(\frac{r^{\beta-1}}{1+r^{2\beta}}\Big)^{\frac{N-2}{2}}\frac{1-r^{2\beta}}{1+r^{2\beta}},
\end{align*}
and
\begin{align*}
\partial_{r}(\Lambda W_{a}r^{c})=&\frac{N-2}{2}(N(N-2)\beta^2)^{\frac{N-2}{4}}\beta\Big(\frac{r^{\beta-1}}{1+r^{2\beta}}\Big)^{\frac{N}{2}+1}\\
&\Big[\big((N-2)(\beta-1)+2c\big)r^{c-2\beta+1}+\big(-(N-2)2\beta+2(c-4\beta)\big)r^{c+1}\\
&+\big((N-2)(1+\beta)-2(c+4\beta)\big)r^{c+2\beta+1}-2c r^{c+4\beta+1}\Big].
\end{align*}
Hence, for every $R>0$, direct computation shows
\begin{align}\label{eq:projection on lambda W}
&\|\Lambda W_{a}\|_{\dot{H}_{a}^{1}(r\geq R,r^{N-1}\ud r)}^2=\int_{R}^{\infty}\Big(\big(\partial_{r}+\frac{c}{r}\big)(\Lambda W_{a})\Big)^2r^{N-1}\ud r\nonumber\\
=&(\frac{N-2}{2})^2(N(N-2)\beta^2)^{\frac{N-2}{2}}\beta^2\int_{R}^{\infty}\Big|\Big(\frac{r^{\beta-1}}{1+r^{2\beta}}\Big)^{\frac{N+2}{2}}\nonumber\\
&\Big[\big((N-2)(\beta-1)+2c\big)r^{c-2\beta+1}+\big(-2(N-1)\beta+2(c-4\beta)\big)r^{c+1}\nonumber\\
&+\big((N-2)(1+\beta)-2(c+4\beta)\big)r^{c+2\beta+1}-2c r^{c+4\beta+1}\Big]\Big|^2r^{N-1-2c}\ud r\nonumber\\
\leq& CR^{-(N-2)\beta}.
\end{align}
And for sufficiently large $R>0$, we estimate by Taylor expansion,
\begin{align}\label{eq:projection on f}
&\int_{R}^{\infty}\Big(\partial_{r}+\frac{c}{r}\Big)f(r)\Big(\partial_{r}+\frac{c}{r}\Big)(\Lambda W_{a})r^{N-1}\ud r\nonumber\\
=&\frac{N-2}{2}(N(N-2)\beta^2)^{\frac{N-2}{4}}\beta\nonumber\\
&\times\Big[\big((N-2)(\beta-1)+2c\big)\int_{R}^{\infty}\partial_{r}(f(r)r^{c})(1+r^{2\beta})^{-\frac{N+2}{2}}r^{\frac{(\beta+1)N}{2}-\beta-1-c}\ud r\nonumber\\
&-\big(-2(N-2)\beta+2(c-4\beta)\big)\int_{R}^{\infty}\partial_{r}(f(r)r^{c})(1+r^{2\beta})^{-\frac{N+2}{2}}r^{\frac{(\beta+1)N}{2}-\beta-3-c}\ud r\nonumber\\
&+\big((N-2)(1+\beta)-2(c+4\beta)\big)\int_{R}^{\infty}\partial_{r}(f(r)r^{c})(1+r^{2\beta})^{-\frac{N+2}{2}}r^{\frac{(\beta+1)N}{2}+3\beta-3-c}\ud r\nonumber\\
&-2c\int_{R}^{\infty}\partial_{r}(f(r)r^{c})(1+r^{2\beta})^{-\frac{N+2}{2}}r^{\frac{(\beta+1)N}{2}+5\beta-1-c}\ud r\Big]\nonumber\\
\leq& C_1f(R)R^{c}+C_2R^{\frac{-(N-2)\beta}{2}}\Big(\int_{R}^{\infty}|\partial_{r}(f(r)r^c)|^2r^{N-1-2c}\ud r\Big)^{\frac{1}{2}}+o(1).
\end{align}
Thus, \eqref{eq:projection on lambda W} and \eqref{eq:projection on f} yield
\begin{align*}
&\|\operatorname{\pi}_{\mathcal{Z}}^{\bot}(f,g)\|_{\dot{H}_{a}^{1}\times L^{2}(r\geq R, r^{N-1})}^2\\
\leq&\|(f,g)\|_{\dot{H}_{a}^{1}\times L^{2}(r\geq R, r^{N-1})}^2-CR^{(N-2)\beta}\Big(C_1|f(R)R^{c}|^2\\
&+C_{2}R^{-(N-2)\beta}\int_{R}^{\infty}|(\partial_{r}+\frac{c}{r})f(r)|r^{N-1}\ud r\Big)\\
\leq&\|(f,g)\|_{\dot{H}_{a}^{1}\times L^{2}(r\geq R, r^{N-1})}^2-C_1|f(R)|^2R^{N-2}+C_{2}\int_{R}^{\infty}|(\partial_{r}+\frac{c}{r})f(r)|r^{N-1}\ud r\\
\leq& C\Big(\|(f,g)\|_{\dot{H}_{a}^{1}\times L^{2}(r\geq R, r^{N-1})}^2-|f(R)|^2R^{N-2}\Big).
\end{align*}
This estimate, combined with \eqref{eq:lower bdd of u}, implies
\begin{align*}
\|\operatorname{\pi}_{\mathcal{Z}}^{\bot}(f,g)\|_{\dot{H}_{a}^{1}\times L^{2}(r\geq R, r^{N-1})}^2\leq& C\max\limits_{\pm}\lim\limits_{t\rightarrow\pm\infty}\bigg(\|\partial_{t}u(t,r)\|_{L^2(r>R+t, r^{N-1})}\\
&+\big\|(\partial_{r}+\frac{c}{r})u(t,r)\big\|_{L^2(r>R+t, r^{N-1})}\bigg).
\end{align*}
Thus the proof is completed.
\end{proof}
\subsection{The exterior energy estimate close to a multi-soliton}
As a corollary of Theorem~\ref{thm:weq}, we will prove an exterior energy lower bound for the linearized operator close to a radial multi-soliton solution. First of all, we prove some estimates on space-time norms.
\begin{lemma}\label{ineq: theta N}
Assume $(N-2)\beta\geq 1$ and $0<\lambda<\mu$. Then it holds that
\begin{align}\label{estimate W, Lambda W}
&\Big\|\mathbf{1}_{\{|x|\geq|t|\}}{W_a}_{(\lambda)}^{\frac{4}{N-2}}{(\Lambda W_a)}_{(\mu)}\Big\|_{L^{1}\left(\mathbb{R}, L^{2}\right)}+\Big\|t\mathbf{1}_{\{|x|\geq|t|\}}{W_a}_{(\lambda)}^{\frac{4}{N-2}}{(\Lambda W_a)}_{[\mu]}\Big\|_{L^{1}\left(\mathbb{R}, L^{2}\right)}\nonumber\\
\lesssim&\left\{
\begin{aligned}
\left(\frac{\lambda}{\mu}\right)^{\frac{(N-2)\beta}{2}}, \qquad\qquad &\text{if}~(N-6)\beta\leq 0,\\
\left(\frac{\lambda}{\mu}\right)^{2\beta}, \qquad\qquad\qquad &\text{if}~(N-6)\beta>0.
\end{aligned}
\right.
\end{align}
\end{lemma}
\begin{proof}
Due to \eqref{asympototic1}, we have
\begin{align}\label{eq: Wa asymptotic}
|\Lambda W_{a}(x)|+|W_{a}(x)|\lesssim\min\left\{|x|^{\frac{(N-2)(\beta-1)}{2}},|x|^{-\frac{(N-2)(1+\beta)}{2}}\right\}.
\end{align}
By symmetry, it suffices to prove the $t\geq0$ case. Without loss of generality, we assume $\mu=1$. Now we estimate the norms by dividing the domain of integration with respect to $r$ into three parts: $(0,\lambda)$, $(\lambda,1)$ and $(1,\infty)$.
In order to deal with the case $\Big\|\mathbf{1}_{\{|x|\geq|t|\}}{W_a}_{(\lambda)}^{\frac{4}{N-2}}{(\Lambda W_a)}_{(\mu)}\Big\|_{L^{1}\left(\mathbb{R}, L^{2}\right)}$, by \eqref{eq: Wa asymptotic}, we define
\begin{align*}
f(r,t)=&\lambda^{-2}W_{a}^{\frac{4}{N-2}}\Big(\frac{x}{\lambda}\Big)W_{a}(x)\\
=&\lambda^{-2}\min\left\{\left(\frac{r}{\lambda}\right)^{2(\beta-1)},\left(\frac{\lambda}{r}\right)^{2(\beta+1)}\right\}\min\left\{r^{\frac{(N-2)(\beta-1)}{2}},r^{-\frac{(N-2)(\beta+1)}{2}}\right\}.
\end{align*}
Then it suffices to estimate
\begin{align*}
\int_{0}^{\infty}&\Big(\int_{t}^{\infty}f(r,t)r^{N-1}\ud r\Big)^{\frac{1}{2}}\ud t\lesssim\int_{0}^{\lambda}\Big(\int_{t}^{\lambda}f(r,t)r^{N-1}\ud r\Big)^{\frac{1}{2}}\ud t\\
&+\int_{0}^{1}\Big(\int_{\max\{t,\lambda\}}^{1}f(r,t)r^{N-1}\ud r\Big)^{\frac{1}{2}}\ud t+\int_{0}^{\infty}\Big(\int_{\max\{t,1\}}^{\infty}f(r,t)r^{N-1}\ud r\Big)^{\frac{1}{2}}\ud t\\
\lesssim&\lambda^{-2\beta}\int_{0}^{\lambda}\Big(\int_{t}^{\lambda}r^{\beta(N+2)-3}\ud r\Big)^{\frac{1}{2}}\ud t+\lambda^{2\beta}\int_{0}^{1}\Big(\int_{\max\{t,\lambda\}}^{1}r^{(N-6)\beta-3}\ud r\Big)^{\frac{1}{2}}\ud t\\
&+\lambda^{2\beta}\int_{0}^{\infty}\Big(\int_{\max\{t,1\}}^{\infty}r^{-\beta(N+2)-3}\ud r\Big)^{\frac{1}{2}}\ud t\\
=&{\rm\uppercase\expandafter{\romannumeral1}}+{\rm\uppercase\expandafter{\romannumeral2}}+{\rm\uppercase\expandafter{\romannumeral3}}.
\end{align*}
We estimate these three terms, respectively. First,
\begin{align}\label{eq: bdd of (1)}
{\rm\uppercase\expandafter{\romannumeral1}}\lesssim\lambda^{-2\beta}\lambda^{\frac{(N+2)\beta}{2}}\lesssim\lambda^{\frac{(N-2)\beta}{2}}.
\end{align}
Next we estimate ${\rm\uppercase\expandafter{\romannumeral2}}$. If $(N-6)\beta<3$, direct computation shows
\begin{align}\label{eq: bdd of (2.1)}
{\rm\uppercase\expandafter{\romannumeral2}}=&\lambda^{2\beta}\int_{0}^{1}\Big(\int_{\max\{t,\lambda\}}^{1}r^{(N-6)\beta-3}\ud r\Big)^{\frac{1}{2}}\ud t\nonumber\\
=&\lambda^{2\beta}\int_{0}^{\lambda}\Big(\int_{\lambda}^{1}r^{(N-6)\beta-3}\ud r\Big)^{\frac{1}{2}}\ud t+\lambda^{2\beta}\int_{\lambda}^{1}\Big(\int_{t}^{1}r^{(N-6)\beta-3}\ud r\Big)^{\frac{1}{2}}\ud t\nonumber\\
\lesssim&\lambda^{2\beta}\int_{0}^{\lambda}\lambda^{\frac{(N-6)\beta-2}{2}}\ud t+\lambda^{2\beta}\int_{\lambda}^{1}t^{\frac{(N-6)\beta-2}{2}}\ud t\nonumber\\
\lesssim&
\begin{cases}\lambda^{\frac{(N-2)\beta}{2}}, & \text {if}\quad(N-6)\beta\leq 0,\\ \lambda^{2\beta}, & \text {if}\quad(N-6)\beta> 0.\end{cases}
\end{align}
Furthermore, if $(N-6)\beta\geq3$, then
\begin{align}\label{eq: bdd of (2.2)}
{\rm\uppercase\expandafter{\romannumeral2}}=&\lambda^{2\beta}\int_{0}^{\lambda}\Big(\int_{\lambda}^{1}r^{(N-6)\beta-3}\ud r\Big)^{\frac{1}{2}}\ud t+\lambda^{2\beta}\int_{\lambda}^{1}\Big(\int_{t}^{1}r^{(N-6)\beta-3}\ud r\Big)^{\frac{1}{2}}\ud t\nonumber\\
\lesssim&\lambda^{2\beta}.
\end{align}
Finally, we give the bound on ${\rm\uppercase\expandafter{\romannumeral3}}$.
\begin{align}\label{eq: bdd of (3)}
{\rm\uppercase\expandafter{\romannumeral3}}=&\lambda^{2\beta}\int_{0}^{\infty}\Big(\int_{\max\{t,1\}}^{\infty}r^{-\beta(N+2)-3}\ud r\Big)^{\frac{1}{2}}\ud t\nonumber\\
=&\lambda^{2\beta}\int_{0}^{1}\Big(\int_{1}^{\infty}r^{-\beta(N+2)-3}\ud r\Big)^{\frac{1}{2}}\ud t+\lambda^{2\beta}\int_{1}^{\infty}\Big(\int_{1}^{\infty}r^{-\beta(N+2)-3}\ud r\Big)^{\frac{1}{2}}\ud t\nonumber\\
\lesssim&\lambda^{2\beta}.
\end{align}
For $(N-6)\beta\leq 0$, we have $\frac{\beta(N-2)}{2}\leq 2\beta$, then $\lambda^{2\beta}\leq\lambda^{\frac{\beta(N-2)}{2}}$. Combining \eqref{eq: bdd of (1)}, \eqref{eq: bdd of (2.1)}, \eqref{eq: bdd of (2.2)} and \eqref{eq: bdd of (3)}, we get
\begin{align}\label{eq: bdd of first}
\Big\|\mathbf{1}_{\{|x|\geq|t|\}}{W_a}_{(\lambda)}^{\frac{4}{N-2}}{(\Lambda W_a)}_{(\mu)}\Big\|_{L^{1}\left(\mathbb{R}, L^{2}\right)}\lesssim\begin{cases}\big(\frac{\lambda}{\mu}\big)^{\frac{(N-2)\beta}{2}}, & \text {if}\quad(N-6)\beta\leq 0,\\\big(\frac{\lambda}{\mu}\big)^{2\beta}, & \text {if}\quad(N-6)\beta> 0.\end{cases}
\end{align}
By analoguous arguments we obtain
\begin{align}\label{eq: bdd of second}
\Big\|t\mathbf{1}_{\{|x|\geq|t|\}}{W_a}_{(\lambda)}^{\frac{4}{N-2}}{(\Lambda W_a)}_{[\mu]}\Big\|_{L^{1}\left(\mathbb{R}, L^{2}\right)}\lesssim\begin{cases}\big(\frac{\lambda}{\mu}\big)^{\frac{(N-2)\beta+2}{2}}, & \text {if}\quad(N-6)\beta\leq -2,\\\big(\frac{\lambda}{\mu}\big)^{2\beta}, & \text {if}\quad(N-6)\beta>-2.\end{cases}
\end{align}
Due to \eqref{eq: bdd of first} and \eqref{eq: bdd of second}, for $-2\leq (N-6)\beta\leq 0$, we infer that $\lambda^{2\beta}\leq\lambda^{\frac{(N-2)\beta}{2}}$, and \eqref{estimate W, Lambda W} is obtained.
\end{proof}
\medskip
We will also use the following asymptotic property for the solution of the linear wave equation with inverse-square potential. Firstly, we introduce a function space on $\mathbb{R}\times S^{N-1}$:
\begin{align*}
\dot{H}_{\eta}^{1}(\mathbb{R}\times S^{N-1})=\left\{g\in C^{0}(\mathbb{R},L^2(S^{N-1})):\int_{\mathbb{R}\times S^{N-1}}|\partial_{\eta}g(\eta,\omega)|^2\ud \eta\ud \omega<\infty\right\}.
\end{align*}
\begin{definition}
Define $\dot{\mathcal{H}}_{\eta}^{1}$ as the quotient space of $\dot{H}_{\eta}^{1}$ by the equivalence relation:
\begin{align*}
g\sim\tilde{g}\Leftrightarrow\exists a(\omega)\in L^{2}(S^{N-1}):\forall\eta\in\mathbb{R}, g(\eta,\omega)-\tilde{g}(\eta,\omega)=a(\omega)\quad for\quad a.e.\quad\omega\in S^{N-1}.
\end{align*}
and denote by $\bar{g}\in\dot{\mathcal{H}_{\eta}^{1}}$ the equivalence class of $g\in\dot{H}_{\eta}^{1}$ with norm on $\dot{\mathcal{H}}_{\eta}^{1}$:
\begin{align*}
\|\bar{g}\|_{\dot{\mathcal{H}_{\eta}^{1}}(\mathbb{R}\times S^{N-1})}=\|\partial_{\eta}g\|_{L^2(\mathbb{R}\times S^{N-1})}.
\end{align*}
\end{definition}
\medskip
\begin{theorem}\label{thm: asymptotic liear wave}
Assume $(N-2)\beta\geq 1$ and let $u_L$ be a solution of the linear wave equation with inverse-square potential~\eqref{lwp}. Then we have
\begin{align*}
\lim\limits_{t\rightarrow+\infty}\Big\|\frac{1}{r}\nabla_{\omega}u_{L}(t)\Big\|_{L^2(\mathbb{R}\times S^{N-1})}+\Big\|\frac{1}{r}u_L(t)\Big\|_{L^2(\mathbb{R}\times S^{N-1})}=0,
\end{align*}
and there exists a unique $G_{+}\in L^{2}\left(\mathbb{R}\times S^{N-1}\right)$ such that
\begin{align*}
\lim\limits_{t\rightarrow+\infty}\int_{0}^{+\infty}\int_{S^{N-1}}\left|r^{\frac{N-1}{2}}\partial_{r}u_{L}(t,r\omega)-G_{+}(r-t,\omega)\right|^{2}\ud \omega\ud r&=0,\\
\lim\limits_{t\rightarrow+\infty}\int_{0}^{+\infty}\int_{S^{N-1}}\left|r^{\frac{N-1}{2}}\partial_{r}u_{L}(t,r\omega)+G_{+}(r-t,\omega)\right|^{2}\ud \omega\ud r&=0.
\end{align*}
Furthermore,
\begin{align*}
E_{L}(f,g)=\int_{\mathbb{R}\times S^{N-1}}\left|G_{+}(\eta,\omega)\right|^{2}\ud \eta\ud\omega=\left\|G_{+}\right\|_{L^2}^{2},
\end{align*}
and the map
\begin{align*}
(f,g)\mapsto&\sqrt{2}G_{+}\\
(\dot{H}_{a}^{1}\times L^{2})(\mathbb{R}^{N})\rightarrow&L^{2}(\mathbb{R}\times S^{N-1})
\end{align*}
is a bijective isometry.
\end{theorem}
\medskip
Before proving this theorem, we need some preparations. We first prove the result for smooth, compactly supported functions.
\begin{lemma}\label{lem 4.6}
Let $(N-2)\beta\geq1$, $(v_0,v_1)\in(C_{0}^{\infty}(\mathbb{R}^N))^2$ and let $v$ be the solution to the wave equation \eqref{lwp}. Then $\|v(t)\|_{L^2}$ is bounded and there exists $R>0$ such that
\begin{align}\label{lem3.6 ineq1}
\lim\limits_{t\rightarrow\infty}\int_{\Bbb R^N}\frac{1}{|x|^2}|v(t,x)|^2\ud x=0,
\end{align}
and
\begin{align}\label{lem3.6 ineq2}
\lim\limits_{R\rightarrow\infty}\limsup\limits_{t\rightarrow\infty}\int_{|x|\leq |t|-R}\Big|(\partial_{r}+\frac{c}{r})v(t,x)\Big|^2+|\partial_{t}v(t,x)|^2\ud x=0.
\end{align}
\end{lemma}
\begin{proof}
The proof of \eqref{lem3.6 ineq1} is just a repetition of Proposition $A.5$ in \cite{CKS}. So it remains to prove \eqref{lem3.6 ineq2}
\begin{align*}
&\frac{\ud }{\ud t}\left(\int_{\Bbb R^N} x\cdot\Big(\partial_{r}+\frac{c}{r}\Big)v\partial_{t}v+\frac{N-1}{2}\int_{\Bbb R^N} v\partial_{t}v\right)\\
=&\int_{\Bbb R^N} x\cdot\Big(\partial_{r}+\frac{c}{r}\Big)v_{t}v_t+x\cdot\Big(\partial_{r}+\frac{c}{r}\Big)v\cdot\left(\Delta+\frac{a}{r^2}\right)v+\frac{N-1}{2}\int_{\Bbb R^N}\left(\partial_{t}v\right)^2+v\cdot\Big(\Delta-\frac{a}{r^2}\Big)v\\
=&{\rm\uppercase\expandafter{\romannumeral1}}+{\rm\uppercase\expandafter{\romannumeral2}}+{\rm\uppercase\expandafter{\romannumeral3}}+{\rm\uppercase\expandafter{\romannumeral4}},
\end{align*}
where
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral1}}=&\int_0^{\infty} r\Big(\partial_{r}+\frac{c}{r}\Big)v_{t}\cdot v_t r^{N-1}\ud r=\int_0^\infty v_{t}\partial_{r}(v_{t}r^{c})r^{N-c}\ud r\\
=&-\frac{N-2c}{2}\int\limits_0^\infty (\partial_{t}v)^{2}r^{N-1}\ud r,
\end{align*}
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral2}}=&\int_0^{\infty} r\left(\partial_{r}+\frac{c}{r}\right)v\cdot\left(\Delta-\frac{a}{r^2}\right)v r^{N-1}\ud r=\int_0^{\infty}\partial_{r}(v r^{c})\cdot\Big(\partial_{r}+\frac{N-1}{r}-\frac{c}{r}\Big)\left(\partial_{r}+\frac{c}{r}\right)v r^{N-c}\ud r\\
=&\int_0^{\infty}\partial_{r}(v r^{c})\cdot\partial_{r}\left(\partial_{r}(v r^{c})r^{-2c}\right)r^{N}\ud r+\int_0^{\infty}\frac{N-1}{r}\cdot(\partial_{r}(v r^{c}))^2 r^{N-2c}\ud r\\
=&\int_0^{\infty}\partial_{r}(v r^{c})\Big(\partial_{r}+\frac{N-1}{r}\Big)\big(\partial_{r}(v
r^{c})\big)r^{N-2c}\ud r-2c\int_0^{\infty}\left(\partial_{r}(v r^{c})\right)^2 r^{N-1-2c}\ud r\\
=&\frac{1}{2}\int_0^{\infty}\partial_{r}\Big((\partial_{r}(v r^{c})r^{N-1})^2\Big)r^{-N-2c+2}\ud r-2c\int_0^{\infty}\int_0^{\infty}\left(\partial_{r}(v r^{c})\right)^2 r^{N-1-2c}\ud r\\
=&\frac{N-2-2c}{2}\int_0^{\infty}\big(\partial_{r}(v r^c)\big)^2 r^{N-1-2c}\ud r,
\end{align*}
and
\begin{align*}
{\rm\uppercase\expandafter{\romannumeral4}}=&\int_0^\infty v\cdot\left(\Delta-\frac{a}{r^2}\right)v r^{N-1}\ud r=\int_0^\infty v\partial_{r}\left(\partial_{r}(v r^c)r^{-2c}\right)r^{N-1+c}\ud r+\int_0^\infty v\frac{N-1}{r}\partial_{r}(v r^c)r^{N-1-c}\ud r\\
=&\int_0^\infty v\Big(\partial_{r}+\frac{N-1}{r}\Big)\partial_{r}(v r^c)r^{N-1-c}\ud r-2c\int_0^\infty v\partial_{r}(v r^c)r^{N-2-c}\ud r\\
=&\int_0^\infty v r^c\partial_{r}\Big(\partial_{r}(v r^c)r^{N-1}\Big)r^{-2c}\ud r-2c\int_0^\infty v\partial_{r}(v r^c)r^{N-2-c}\ud r\\
=&-\int_0^\infty \big(\partial_{r}(v r^c)\big)^2r^{N-1-2c}\ud r.
\end{align*}
Since note $c=\frac{(N-2)(1-\beta)}{2}<0$, we have
\begin{align*}
&\frac{\ud }{\ud t}\left(\int_{\Bbb R^N} x\cdot\left(\partial_{r}+\frac{c}{r}\right)v\partial_{t}v+\frac{N-1}{2}\int_{\Bbb R^N} v\partial_{t}v\right)\\
=&-\frac{N+2c}{2}\int_0^\infty(\partial_{t}v)^2r^{N-1}\ud r+\frac{N-2-2c}{2}\int_0^\infty\big(\partial_{r}(v r^c)\big)^2r^{N-1-2c}\ud r\\
&+\frac{N-1}{2}\int_0^\infty(\partial_{t}v)^2r^{N-1}\ud r-\frac{N-1}{2}\int_0^\infty\big(\partial_{r}(v r^c)\big)^2r^{N-1-2c}\ud r\\
=&-E_{a}(v_0,v_1)-2c E_{a}(v_0,v_1)\geq-E_{a}(v_0,v_1).
\end{align*}
Integrate from $0$ to $t$, yields
\begin{align*}
&\int_{\Bbb R^N} x\left(\partial_{r}+\frac{c}{r}\right)v\cdot \partial_{t}v\ud x+\frac{N-1}{2}\int_{\Bbb R^N} v\partial_{t}v\ud x\\
\geq&-t E_{a}(v_0,v_1)+\int_{\Bbb R^N} x\left(\partial_{r}+\frac{c}{r}\right)v_0\cdot v_1\ud x+\frac{N-1}{2}\int_{\Bbb R^N} v_0v_1\ud x.
\end{align*}
Since $(N-2)\beta\geq1$, $\dot{H}_{a}^{-1}$ is a Hilbert space, and $C_{0}^{\infty}(\mathbb{R}^{N})\subset\dot{H}_{a}^{-1}(\mathbb{R}^{N})$. By conservation of the $L^{2}\times\dot{H}_{a}^{-1}$ norm of $(v,\partial_{t}v)$,
\begin{align}\label{v bound}
\left|\int_{\Bbb R^N} x\cdot\left(\partial_{r}+\frac{c}{r}\right)v\cdot\partial_{t}v \ud x\right|\geq t E_{a}(v_0,v_1)-C(v_0,v_1).
\end{align}
Let $M>0$ be such that $|x|\leq M$ on the support of $(v_0,v_1)$. Then $|x|\leq M+|t|$ on the support of $(v(t),\partial_{t}v(t))$. By the finite speed of propagation and Cauchy-Schwarz inequality
\begin{align*}
&\left|\int_{\Bbb R^N} x\cdot\left(\partial_{r}+\frac{c}{r}\right)v\cdot\partial_{t}v\ud x\right|\\
\leq&(t-R)\int_{|x|\leq t-R}\Big|\left(\partial_{r}+\frac{c}{r}\right)v\cdot\partial_{t}v\Big|\ud x
+(t+M)\int_{|x|\geq t-R}\Big|\left(\partial_{r}+\frac{c}{r}\right)v\cdot\partial_{t}v\Big|\ud x\\
\leq&(t+M)E_{a}(v_0,v_1)-\frac{R}{2}\int_{|x|\leq t-R}|\nabla_{t,x}v|^2\ud x,
\end{align*}
then combining with \eqref{v bound}, yields
\begin{align*}
\frac{R}{2}\int_{|x|\leq t-R}|\nabla_{t,x}v|^2\ud x\leq C+ME_{a}(v_0,v_1),
\end{align*}
which implies
\begin{align*}
\lim\limits_{R\rightarrow\infty}\limsup\limits_{t\rightarrow\infty}\int_{|x|\leq t-R}|\nabla_{t,x}v|^2\ud x=0.
\end{align*}
\end{proof}
\medskip
As a consequence of Lemma \ref{lem 4.6}, as in \cite{CKS} we get
\begin{lemma}
Let $(N-2)\beta\geq 1$, $(v_0,v_1)\in(C_{0}^{\infty}(\mathbb{R}^N))^2$. Then
\begin{align}\label{eq:1/r v, L2}
\lim\limits_{t\rightarrow\infty}\Big\|\frac{1}{r}\nabla_{\omega}r(t)\Big\|_{L^{2}}+\Big\|\frac{1}{r}v(t)\Big\|_{L^2}=0,
\end{align}
and there exists $g\in\dot{H}_{\eta}^{1}$ such that
\begin{align}\label{eq: relatioinship between v and g}
\lim\limits_{t\rightarrow\infty}\left\|\nabla_{r,t}\Big(r^{\frac{N-1}{2}}v(t,r\omega)-g(r-t,\omega)\Big)\right\|_{L^2((0,\infty)\times S^{N-1})}=0.
\end{align}
\end{lemma}
\medskip
\begin{lemma}
Let $(v_0,v_1)$ and $g$ be as in above. Then
\begin{align*}
E(v_0,v_1)=\|\bar{g}\|_{\dot{H}_{\eta}^1(\mathbb{R}\times S^{N-1})}^2,
\end{align*}
where $$\|\bar{g}\|_{\dot{H}_{\eta}^1(\mathbb{R}\times S^{N-1})}^2=\int_{-\infty}^{\infty}\int_{S^{N-1}}\big|\partial_{\eta}g(\eta,\omega)\big|^2\ud\omega\ud\eta.$$
\end{lemma}
\begin{proof}
The energy of the equation \eqref{lwp} is given by
\begin{align*}
E_{a}(v,\partial_{t}v)=\frac{1}{2}\int_{0}^{\infty}\int_{S^{N-1}}\Big[\Big((\partial_{r}+\frac{c}{r})v\Big)^2+(\partial_{t}v)^2+\frac{1}{r^2}|\nabla_{\omega}v|^2+\frac{a}{r^{2}}|v|^2\Big]r^{N-1}\ud r.
\end{align*}
Relying on \eqref{eq:1/r v, L2}, we infer that
\begin{align*}
&\int_{0}^{\infty}\int_{S^{N-1}}\left((\partial_{r}+\frac{c}{r})v\right)^2r^{N-1}\ud \omega\ud r=\int_{0}^{\infty}\int_{S^{N-1}}\Big(\partial_{r}(r^{\frac{N-1}{2}}v)-(\frac{N-1}{2}-c)r^{\frac{N-3}{2}}v\Big)^2\ud r\\
=&\int_{0}^{\infty}\int_{S^{N-1}}\left(\partial_{r}(r^{\frac{N-1}{2}}v)\right)^2\ud r+o(1).
\end{align*}
Inserting \eqref{eq: relatioinship between v and g} into the expression of the energy, as $t\rightarrow\infty$,
\begin{align*}
E_{a}(v,\partial_{t}v)=&\frac{1}{2}\int_{0}^{\infty}\int_{S^{N-1}}\left(\partial_{r}(r^{\frac{N-1}{2}}v)\right)^2\ud \omega\ud r+\frac{1}{2}\int_{0}^{\infty}\int_{S^{N-1}}(\partial_{t}v)^2r^{N-1}\ud\omega\ud r+o(1)\\
=&\frac{1}{2}\int_{0}^{\infty}\int_{S^{N-1}}\big|\partial_{\eta}g(r-t,\omega)\big|^2\ud\omega\ud r+o(1)\\
=&\frac{1}{2}\int_{-t}^{\infty}\int_{S^{N-1}}\big|\partial_{\eta}g(\eta,\omega)\big|^2\ud\omega\ud\eta+o(1)\\
\rightarrow&\int_{-\infty}^{\infty}\int_{S^{N-1}}\big|\partial_{\eta}g(\eta,\omega)\big|^2\ud\omega\ud\eta.
\end{align*}
In view of the conservation law of the energy, implies
\begin{align*}
E_{a}(v_0,v_1)=\|\bar{g}\|_{\dot{H}_{\eta}^1(\mathbb{R}\times S^{N-1})}^2.
\end{align*}
\end{proof}
The proof of the general case and the surjectivity follows from \cite{CKS}.
\medskip
Finally, let us introduce the energy lower bound for the linearized operator close to an approximate radial multisoliton solution. For a sequence of monotonous parameters $\{\lambda_j\}$, we denote $G_J$ by a subset of $(0,\infty)^{J}$ such that
\begin{align*}
G_{J}=\Big\{\boldsymbol{\lambda}=(\lambda_{j})_{1\leq j\leq J},0<\lambda_{J}<\lambda_{J-1}<\ldots<\lambda_1\Big\}.
\end{align*}
If $\boldsymbol{\lambda}\in G_{J}$, we denote
$$
\begin{gathered}
\gamma(\boldsymbol{\lambda})=\max_{2\leq j\leq J}\frac{\lambda_{j}}{\lambda_{j-1}}\in(0,1),\\
\mathcal{L}_{\boldsymbol{\lambda}}^{a}=-\Delta-\sum_{j=1}^{J}\frac{N+2}{N-2}W_{a\left(\lambda_{j}\right)}^{\frac{4}{N-2}}+\frac{a}{|x|^2}, \\
Z_{\boldsymbol{\lambda}}=\operatorname{span}\Big\{\left((\Lambda W_a)_{\left(\lambda_{j}\right)},0\right),\left(0,(\Lambda W_a)_{\left[\lambda_{j}\right]}\right)\Big\},
\end{gathered}
$$
see \eqref{def: scalling} for the notations $(\Lambda W)_{\left(\lambda_{j}\right)}$ and $(\Lambda W)_{\left[\lambda_{j}\right]}$. Going back to the Lemma \ref{ineq: theta N} and using the asymptotic property for the solution of the linear wave equation with inverse-square potential, we prove the following Theorem \ref{thm:multisoliton energy} as in \cite{DKM7}
\medskip
\begin{theorem}\label{thm:multisoliton energy}\emph{(\cite{DKM7})}
Assume $(N-2)\beta\geq 1$ and $(N-2)\beta$ is odd. For any $J \geq 2$, there exists $\gamma_{*}>0$ and $C>0$ with the following property. For any $\boldsymbol{\lambda}$ with $\gamma(\boldsymbol{\lambda}) \leq \gamma_{*}$, for any solution $u$ of
$$
\partial_{t}^{2}u+\mathcal{L}_{\boldsymbol{\lambda}}^{a}u=0, \quad
\vec{u}\mid_{ t=0}=\left(f, g\right) \in \mathcal{H}_{a}.
$$
one has
\begin{align*}\label{ineq: multisoliton energy}
\quad&\left\|\operatorname{\pi}_{Z_{\lambda}^{\perp}}\left(f,g\right)\right\|_{\mathcal{H}_{a}}^{2}\nonumber\\
&\leq C\Big(\sum_{\pm}\lim\limits_{t\rightarrow\pm\infty}\int_{|x|\geq|t|}\left|\nabla_{t, x}u(t, x)\right|^{2} \ud x+\gamma(\boldsymbol{\lambda})^{2 \theta_{N}}\left\|\pi_{Z_{\lambda}}\left(f,g\right)\right\|_{\mathcal{H}_a}^{2}\Big),
\end{align*}
where
\begin{align*}
\theta_{N}=\left\{
\begin{aligned}
\frac{(N-2)\beta}{2},\qquad\qquad &\text{if}~(N-6)\beta\leq 0\\
2\beta \qquad\qquad\qquad &\text{if}~(N-6)\beta>0
\end{aligned}
\right.
\end{align*}
\end{theorem}
\section{Soliton resolution along a time sequence}\label{Soliton resolution along a time sequence}
In this section we will establish the soliton resolution along a time sequence. Because the channel of energy for wave equation with inverse square potential holds in any dimension, the proof is in the same spirit of \cite{CR}.
\begin{theorem}\label{timesequence}
Let $u$ be a radial solution to \eqref{lwp} with $T_+<\infty$ that satisfies \eqref{bh}. Then there exists $(v_0,v_1)\in \dot{H}^1_a\times L^2$, a sequence $T_+(u)$, an integer $J_0\geq1$, $J_0$ sequences of positive numbers $\{\lambda_{j,n}\}_n$, $j=1,\cdots, J_0$, $J_0$ signs $\iota_j\in\{\pm 1\}$ such that\begin{equation*}
\lambda_{1,n}\ll \lambda_{2,n}\ll \cdots \ll \lambda_{J_0, n}\ll T_+(u)-t_n
\end{equation*}
and \begin{equation*}
\vec{u}(t_n)=(v_0, v_1)+\sum_{j=1}^{J_0}\bigg(\frac{\iota_j}{\lambda_{j,n}^\frac{N-2}{2}}W_a\left(\frac{x}{\lambda_{j,n}}, 0\right)\bigg)+o(1)\ \ \mathrm{in}\ \ \dot{H}^1_a\times L^2\ \ \mathrm{as}\ n\rightarrow\infty.
\end{equation*}
\end{theorem}
\begin{lemma}
Let $\{u_{0,n}, u_{1,n}\}$, $\{U_L^j\}$, $\{t_{j,n}, \lambda_{j,n}\}$ and $\theta_n$ satisfy the assumptions of Subsection \ref{profile}. Let $\{\rho_n\}$, $\{\sigma_n\}$ be sequences with $0\leq \rho_n\leq \sigma_n\leq+\infty$. Then for $j\neq k$, we have
\begin{equation*}
\lim_{n\rightarrow\infty}\int_{\rho_n\leq |x|\leq \sigma_n}\nabla_{t,x}U_n^j(\theta_n,x)\cdot\nabla_{t,x}U_n^k(\theta_n,x)\ud x=0.
\end{equation*}
For $j\leq J$, it holds that
\begin{equation*}
\lim_{n\rightarrow\infty}\int_{\rho_n\leq |x|\leq \sigma_n}\nabla_{t,x}U_n^j(\theta_n,x)\cdot\nabla_{t,x}w_n^J(\theta_n,x)+\nabla_{t,x}U_n^j(\theta_n,x)\cdot\nabla_{t,x}r_n^J(\theta_n,x)\ud x=0
\end{equation*}
\end{lemma}
For the energy-critical wave equation, the same results were proved in \cite{DKM1} for $N=3$ which was extended to even dimensions in \cite{CKS} and all other odd dimensions in \cite{CR}.
As in \cite{DKM1}, there exists $(v_0,v_1)\in \mathcal{H}_a$ such that $$(u(t), \partial_t u(t))\mapsto (v_0, v_1)\ \ \mathrm{as}\ \ t\rightarrow1^-$$
weakly in $\mathcal{H}_a$. Let $v$ be the solution to \eqref{lwp} with initial data $(v_0, v_1)$ at time $t=1$. Let $d=u-v$ be the singular part of $u$. Then $d$ is well-defined on $[t_-, 1)$ for some $t_->\mathrm{max}\{T_-(u), T_-(v)\}$. By finite speed of propagation, it follows that $\vec{d}$ is supported in the cone
$$\Big\{(t,x): t_-\leq t<1 \ \ \mathrm{and}\ \ |x|\leq 1-t\Big\}.$$
\begin{proposition}\label{noenergy}
Let $u$ and $d$ be as above. Then
\begin{equation*}
\forall c_0\in(0,1), \ \ \lim_{t \rightarrow1^-}\int_{c_0(1-t)\leq|x|\leq 1-t}\Big(|\nabla d(t,x)|^2+|\partial_t d(t,x)|^2\Big)\ud x=0.
\end{equation*}
\end{proposition}
The key step in the proof of Proposition \ref{noenergy} is to establish the following
\begin{lemma}
Let $I_n^+=I_n\cap[\frac12, +\infty)$, and let $I_n^-=I_n\cap(-\infty, \frac12]$. Then there exists $\eta_1>0$ such that the following holds for large $n$; for all sequences $\{\theta_n\}_n$ with $\theta_n\in I_n^+$, or for all sequences $\{\theta_n\}_n$ with $\theta_n\in I_n^-$,
\begin{equation*}
\int_{|y|\geq |\theta_n|}\Big|\nabla_{\iota,y}(u_n-v_n)(\theta_n,y)\Big|^2\ud y\geq\eta_1.
\end{equation*}
\end{lemma}
Theorem \ref{timesequence} can be proved in exactly the same way as in \cite{CR} once the following proposition is established
\begin{proposition}\label{djt}
Let $u$ be a radial solution to \eqref{lwp} with $T_+<\infty$ that satisfies \eqref{bh}. Then there exists $(v_0,v_1)\in \dot{H}^1_a\times L^2$, a sequence $T_+(u)$, an integer $J_0\geq1$, $J_0$ sequences of positive numbers $\{\lambda_{j,n}\}_n$, $j=1,\cdots, J_0$, $J_0$ signs $\iota_j\in\{\pm 1\}$ such that\begin{equation*}
\lambda_{1,n}\ll \lambda_{2,n}\ll \cdots \ll \lambda_{J_0, n}\ll T_+(u)-t_n
\end{equation*}
and \begin{equation*}
\vec{u}(t_n)=(v_0, v_1)+\sum_{j=1}^{J_0}\Big(\frac{\iota_j}{\lambda_{j,n}^\frac{N-2}{2}}W_a(\frac{x}{\lambda_{j,n}}, 0)\Big)+(w_{0,n}, 0)+o(1)\ \ \mathrm{in}\ \ \dot{H}^1_a\times L^2\ \ \mathrm{as}\ n\rightarrow\infty,
\end{equation*}
where \begin{equation*}
\lim_{n\rightarrow\infty}\Big\|S(t)(w_{0,n}, 0)\Big\|_{S(\Bbb R)}=0.
\end{equation*}
\end{proposition}
As in \cite{DKM1}, we assume that there exist sequences $\{\tau_n\}$, $\{\lambda_n\}$ such that $\tau_n\in(0,1)$ $\tau_n\rightarrow1$ and for the singular part $d(x,t)=u(x,t)-v(x,t)$ it holds that $$\lambda_n\ll1-\tau_n,$$
and \begin{equation}\label{std}
\lim_{n\rightarrow\infty}\int_{|x|\geq \lambda_n}\Big[|\nabla d(\tau_n)|^2+(\partial_t d(\tau_n))^2+\frac{(d(\tau_n))^2}{|x|^2}\Big]\ud x=0.
\end{equation}
The proof is just a repetition of \cite{DKM1} verbatim (see Proposition 5.1) once we have established the following lemma:
\begin{lemma}\label{stdlemma}
Assume \eqref{std}. Then
\begin{equation}\label{stdl}
\lim_{n\rightarrow\infty}\frac{1}{1-\tau_n}\int_{\tau_n}^1\int_{\Bbb R^N}(\partial_t d)^2\ud x\ud t=0.
\end{equation}
\end{lemma}
\begin{proof} Since $d$ and $\partial_t d$ are compactly supported in the space variable, we can safely define
\begin{equation*}
z_1(t)=\int_{\Bbb R^N}(u\partial_t u-v\partial_t v)\ud x,\ \ z_2(t)=\int_{\Bbb R^N}(x\cdot\nabla u\partial_t u-x\cdot\nabla v\partial_t v)\ud x.
\end{equation*}
Direct computation gives that
\begin{align*}
z_1'(t)=&\int_{\Bbb R^N}\Big[(\partial_{t}u)^2-|\nabla u|^2-\frac{a}{|x|^2}|u|^2+|u|^\frac{2N}{N-2}\Big]\ud x\\
&-\int_{\Bbb R^N}\Big[(\partial_{t}v)^2-|\nabla v|^2-\frac{a}{|x|^2}|u|^2+|v|^\frac{2N}{N-2}\Big]\ud x,
\end{align*}
which can be further rewritten as
\begin{align*}
z_1'(t)=&\int_{\Bbb R^N}(\partial_t d)^2\ud x-\int_{\Bbb R^N}|\nabla d|^2\ud x-a\int_{\Bbb R^N}\frac{|d|^2}{|x|^2}\ud x+\int_{\Bbb R^N}|d|^\frac{2N}{N-2}\ud x\\
&+2\int_{\Bbb R^N}\partial_t d\partial_t v\ud x+2\int_{\Bbb R^N}\nabla d\nabla v\ud x+2a\int_{\Bbb R^N} \frac{d v}{|x|^2}\ud x\\
&+\int_{\Bbb R^N}\left(|d+v|^\frac{2N}{N-2}-|v|^\frac{2N}{N-2}-|d|^\frac{2N}{N-2}\right)\ud x.
\end{align*}
Noting that $|x|\leq 1-t$ on the support of $d$, that $v$ converges in $\dot{H}^1_a\times L^2$ as $t\rightarrow1$, and that $u$ is bounded in $\dot{H}^1_a\times L^2$, we get that as $t\rightarrow 1$,
\begin{align*}
z'_1(t)=&\int_{\Bbb R^N}(\partial_t d)^2 \ud x-\int_{\Bbb R^N}|\nabla d|^2\ud x-a\int_{\Bbb R^N}\frac{|d|^2}{|x|^2}\ud x+\int_{\Bbb R^N}|d|^\frac{2N}{N-2}\ud x+o(1).
\end{align*}
A similar computation shows that \begin{align*}
z_2'(t)=&-\frac{N}{2}\int_{\Bbb R^N} (\partial_t u)^2\ud x+\frac{N-2}{2}\int_{\Bbb R^N}\Big(|\nabla u|^2+a\frac{|u|^2}{|x|^2}-|u|^\frac{2N}{N-2}\Big)\ud x\\
&-\left[-\frac{N}{2}\int_{\Bbb R^N} (\partial_t v)^2\ud x+\frac{N-2}{2}\int_{\Bbb R^N}\Big(|\nabla v|^2+a\frac{|v|^2}{|x|^2}-|v|^\frac{2N}{N-2}\Big)\ud x\right]
\end{align*}
Therefore, we have
\begin{align*}
z'_2(t)=-&\frac{N}{2}\int_{\Bbb R^N} (\partial_t d)^2\ud x+\frac{N-2}{2}\int_{\Bbb R^N}\Big(|\nabla d|^2+a\frac{|d|^2}{|x|^2}-|d|^\frac{2N}{N-2}\Big)\ud x.
\end{align*}
We define $$Z(t)=\frac{N-2}{2}z_1(t)+z_2(t),$$ then
\begin{align*}
Z'(t)=-\int_{\Bbb R^N} (\partial_t d)^2\ud x+o(1),\ \ \mathrm{as}\ \ t\rightarrow1^-.
\end{align*}
Let $\varepsilon>0$, and $m$, $n$ be two large integers with $n<m$. Integrate the preceding inequality, we get
\begin{equation*}
\int_{\tau_n}^{\tau_m}\int_{\Bbb R^N}(\partial_t d)^2\ud x\ud t\leq |Z(\tau_m)-Z(\tau_n)|+\varepsilon|\tau_m-\tau_n|.
\end{equation*}
Let $m$ go to infinity, we obtain
\begin{equation*}
\int_{\tau_n}^{\tau_m}\int_{\Bbb R^N}(\partial_t d)^2\ud x\ud t\leq |Z(\tau_n)|+\varepsilon(1-\tau_n).
\end{equation*}
As in \cite{DKM1}, we have $$\lim_{n\rightarrow\infty}\frac{|Z(\tau_n)|}{1-\tau_n}=0,$$
which implies
\begin{equation*}
\lim_{n\rightarrow\infty}\frac{1}{1-\tau_n}\int_{\tau_n}^1\int_{\Bbb R^N}(\partial_t d)^2\ud x\ud t\leq\varepsilon.
\end{equation*}
Since $\varepsilon>0$ is arbitrary, we get \eqref{stdl}.
\end{proof}
As in \cite{DKM1}, based on the profile decomposition and rigidity of elliptic equation \cite{MMZ}, Proposition \ref{djt} follows from Lemma \ref{stdlemma} and Proposition \ref{noenergy}. Thus Theorem \ref{timesequence} is proved.
\section{ Non-radiative solutions and related estimates}
In this section, we shall choose proper parameters to impose the orthogonality conditions and consider a non-radiative solution close to a multi-soliton. In Subsection \ref{6.1}, we will use the channel of energy for the linearized equation in Section \ref{4} and the nonlinear term estimates to give a first order bound of the expansion of the solution. In Section \ref{6.2}, we will introduce the exterior scaling parameter $\iota$ of the multi-soliton and give a lower bound.
\subsection{Non-radiative solutions and orthogonality conditions}\label{6.1}
\begin{definition}\label{def: non-radiative solutions}
Let $t_{0}\in\mathbb{R}$, and let $u$ be a solution of the nonlinear wave equation with the inverse-square potential \eqref{weq}. We say that $u$ is non-radiative at $t=t_{0}$ if $u$ is defined on $\left\{|x|>\left|t-t_{0}\right|\right\}$ and
\begin{align*
\sum\limits_{\pm}\lim\limits_{t\rightarrow\pm\infty}\int_{|x| \geq\left|t-t_{0}\right|}\left|\nabla_{t, x}u(t, x)\right|^{2}\ud x=0.
\end{align*}
We say that $u$ is weakly non-radiative, if for large $R>0, u$ is defined on $\{|x|>|t|+R\}$, and
\begin{align*
\sum_{\pm} \lim _{t \rightarrow \pm \infty} \int_{|x| \geq|t|+R}\left|\nabla_{t, x} u(t, x)\right|^{2} \ud x=0.
\end{align*}
\end{definition}
\medskip
In this section, we assume that $(N-2)\beta\geq 3$ is odd and $u$ is the radial solution of \eqref{weq} with initial data $(u_0,u_1)$ at $t=0$ which is defined on $\{|x|>t\}$. Fix $J\geq 1$, $(\iota_{j})\in\{\pm1\}^{J}$, we denote the multi-soliton with scaling parameters $\boldsymbol{\lambda}=(\lambda_{j})_{j=1}^{J}\in G_{J}$ by $M_{a}=\sum\limits_{j=1}^{J}\iota_{j} {W_a}_{\left(\lambda_{j}\right)}$, where
\begin{align}\label{def GJ}
G_J=\{\boldsymbol{\lambda}=(\lambda_{j})_{j=1}^{J}:0<\lambda_J<\lambda_{J-1}<\cdots<\lambda_1\}.
\end{align}
We assume that there exists $\boldsymbol{\lambda}\in G_{J}$ such that
\begin{align}\label{def: delta}
\Big\|(u_0,u_1)-(M_{a}, 0)\Big\|_{\mathcal{H}_{a}}\triangleq \delta\ll 1
\end{align}
and
\begin{align*}
\gamma:=\gamma(\boldsymbol{\lambda})=\max\limits_{2\leq j\leq J}\frac{\lambda_j}{\lambda_{j-1}}\ll 1.
\end{align*}
In the following lemma, we choose the scaling parameters to get the orthogonality condition.
\begin{lemma}\label{lem: modulation}\emph{(\cite{DKM7})}
Let $J\geq 1$. There exists a small constant $\varepsilon_{J}>0$ and a large constant $C_{J}>0$ with the following property: for all monotonically decreasing $\boldsymbol{\mu}=(\mu_j)_{j}\in(0,\infty)^{J}$ such that $\gamma(\boldsymbol{\mu})<\varepsilon$, for all $(\iota_{j})_{j}\in\{\pm 1\}^{J}$, for all $f\in\dot{H}_{a}^{1}$ such that
\begin{align*}
\Big\|f-\sum\limits_{j=1}^{J}\iota_{j}{W_{a}}_{\left(\mu_{j}\right)}\Big\|_{\dot{H}_{a}^{1}}\leq\varepsilon,
\end{align*}
then there exists a unique $\boldsymbol{\lambda}\in(0,\infty)^{J}$ such that
\begin{align*}
\max\limits_{1\leq j\leq J}\Big|\frac{\lambda_{j}}{\mu_{j}}-1\Big|\leq C_{J}\varepsilon,
\end{align*}
and for all $1\leq j\leq J$,
\begin{align*}
\int\left(\partial_{r}+\frac{c}{r}\right)\Big(f-\sum\limits_{j=1}^{J}\iota_{j}{W_{a}}_{\left(\lambda_{j}\right)}\Big)\cdot\left(\partial_{r}+\frac{c}{r}\right)\left(\Lambda {W_{a}}\right)_{\left(\lambda_{j}\right)}\ud x=0.
\end{align*}
Furthermore, the map $f\mapsto\lambda$ is of class $C^1$.
\end{lemma}
\begin{proof}~Define the mapping
\begin{align*}
&\Phi:(0,\infty)^{J}\rightarrow \mathbb{R}^{J},\qquad \Phi=(\Phi_{\ell})_{1\leq \ell\leq J},\nonumber
\end{align*}
where
\begin{align*}
\Phi_{\ell}(\boldsymbol{\lambda})&=\lambda_{\ell}-\frac{\mu_{\ell}\iota_{\ell}}{\int\left|(\partial_{r}+\frac{c}{r})(\Lambda W_{a})\right|^{2}}\int\left(\partial_{r}+\frac{c}{r}\right)\Big(f-\sum\limits_{j=1}^{J}\iota_{j}{W_{a}}_{(\lambda_{j})}\Big)\cdot\left(\partial_{r}+\frac{c}{r}\right)\left(\Lambda {W_{a}}_{(\lambda_{j})}\right)\ud x.
\end{align*}
Thus it remains to prove the following claim
\medskip
\begin{claim}\label{claim: contraction}~$\Phi$ is a contraction mapping on the compact set
\begin{align*}
D_{\eta}\triangleq\left\{(\lambda_{j})_{1\leq j\leq J}:\max\limits_{1\leq j\leq J}\Big|1-\frac{\lambda_{\ell}}{\mu_{\ell}}\Big|\leq\eta\right\}.
\end{align*}
\end{claim}
\medskip
We now prove this claim. For any $\boldsymbol{\lambda}\in \textcolor{blue}{D_{\eta}}$,
\begin{align*}
\Big|\frac{\Phi_{\ell}(\boldsymbol{\lambda})}{\mu_{\ell}}-1\Big|&\leq\Big|1-\frac{\Phi_{\ell}(\mu)}{\mu_{\ell}}\Big|+\Big|\frac{\Phi_{\ell}(\lambda)-\Phi_{\ell}(\mu)}{\mu_{\ell}}\Big|\nonumber\\
&\leq\Big\|f-\sum\limits_{j=1}^{J}\iota_{j}{W_{a}}_({\mu_j})\Big\|_{\dot{H}_{a}^{1}}+\frac{1}{\mu_{\ell}}\sum\limits_{j=1}^{J}\frac{\ud \Phi_{\ell}}{\ud\lambda_{j}}\cdot(\mu_{j}-\lambda_{j}).
\end{align*}
If $\ell\neq j$, it is not hard to see that
\begin{align*}
\frac{\ud}{\ud \lambda_{j}}\Big(\partial_{r}+\frac{c}{r}\Big){W_{a}}_{(\lambda_{j})}&=\frac{\ud}{\ud \lambda_{j}}\Big(\partial_{r}+\frac{c}{r}\Big)\bigg({\lambda_{j}^{-\frac{N-2}{2}}W_{a}\Big(\frac{|x|}{\lambda_{j}}\Big)}\bigg)\\
&=\lambda_{j}^{-1}(\Lambda W_{a})_{(\lambda_{j})}.
\end{align*}
Using \eqref{main W}, the partial derivative of $\lambda_j$ is
\begin{align*}
\left|\frac{\ud \Phi_{\ell}}{\ud\lambda_{j}}\right|&=-\frac{\mu_{\ell}\iota_{\ell}}{\int\left|(\partial_{r}+\frac{c}{r})(\Lambda W_{a})\right|^{2}\ud x}\lambda_{j}^{-1}\left|\int\Big(\partial_{r}+\frac{c}{r}\Big)(\Lambda W_a)_{(\lambda_{j})}\cdot\Big(\partial_{r}+\frac{c}{r}\Big)(\Lambda W_{a})_{(\lambda_{\ell})}\ud x\right|\\
&\lesssim\frac{\mu_{\ell}}{\lambda_{j}}(\lambda_{j}\lambda_{\ell})^{-\frac{N-2}{2}}\max\left\{\lambda_{j}^{\frac{(\beta-1)N}{2}+N-\beta}\lambda_{\ell}^{-\frac{(\beta+1)N}{2}+\beta},\lambda_{j}^{\frac{(\beta+1)N}{2}-\beta}\lambda_{\ell}^{-\frac{(\beta-1)N}{2}+N+\beta}\right\}\\
&\lesssim\frac{\mu_{\ell}}{\mu_{j}}\varepsilon^{\frac{(N-2)\beta}{2}}.
\end{align*}
Therefore,
\begin{align}\label{l noteq j}
\left|\frac{1}{\mu_{\ell}}\frac{\ud \Phi_{\ell}}{\ud\lambda_{j}}(\mu_{j}-\lambda_{j})\right|\leq\frac{(\mu_{j}-\lambda_{j})}{\mu_{j}}\varepsilon^{\frac{(N-2)\beta}{2}}\leq\eta\varepsilon^{\frac{(N-2)\beta}{2}}.
\end{align}
If $\ell=j$, we observe that
\begin{align*}
\frac{\ud}{\ud\lambda_{\ell}}\left(\Lambda{W_{a}}_{(\lambda_{\ell})}\right)=\lambda_{\ell}^{-1}(\Lambda\Lambda{W_{a}})_{(\lambda_{\ell})},
\end{align*}
where we have used the expression of $W_a$ in the last line. Hence,
\begin{align}\label{l=j}
\left|\frac{\ud\Phi_{\ell}}{\ud\lambda_{\ell}}(\boldsymbol{\lambda})\right|\leq&\Big|1-\frac{\mu_{\ell}}{\lambda_{\ell}}\Big|+\int\left|\left(\partial_{r}+\frac{c}{r}\right)\left({W_{a}}_{(\lambda_{j})}-{W_{a}}_{(\mu_{j})}\right)\right|^2\ud x\nonumber\\
+&\bigg|\frac{\mu_{\ell}}{\lambda_{\ell}\int\left|(\partial_{r}+\frac{c}{r})(\Lambda W_{a})\right|^2\ud x}\int\big(\partial_{r}+\frac{c}{r}\big)\Big(f-\sum\limits_{j=1}^{J}\iota_{j}{W_{a}}_{(\lambda_{j})}\Big)\cdot\big(\partial_{r}+\frac{c}{r}\big)(\Lambda\Lambda W_{a})_{(\lambda_{\ell})}\ud x\bigg|\nonumber\\
\lesssim&\Big|1-\frac{\mu_{\ell}}{\lambda_{\ell}}\Big|+\Big\|f-\sum\limits_{j=1}^{J}\iota_{j}{W_{a}}_{(\mu_j)}\Big\|_{\dot{H}_{a}^{1}}+\Big\|\sum\limits_{j=1}^{J}\left(\iota_{j}{W_{a}}_{(\mu_{j})}-\iota_{j}{W_{a}}_{(\lambda_{j})}\right)\Big\|_{\dot{H}_{a}^{1}}\nonumber\\
\leq&\eta+\varepsilon,
\end{align}
and direct computation shows that
\begin{align*}
&\int\Big|\big(\partial_{r}+\frac{c}{r}\big)\left({W_{a}}_{(\lambda_{j})}-{W_{a}}_{(\mu_{j})}\right)\Big|^2\ud x\\
=&\int\Big|\Big(\big(\frac{\mu_j}{\lambda_j}\big)^{\frac{N}{2}}\partial_{r}W_a\big(\frac{\mu_j}{\lambda_j}r\big)-\partial_{r}W_a(r)\Big)+\frac{c}{r}\Big(\big(\frac{\mu_j}{\lambda_j}\big)^{\frac{N-2}{2}}W_a\big(\frac{\mu_j}{\lambda_j}r\big)-W_a(r)\Big)\Big|^2r^{N-1}\ud r\\
\lesssim&\eta^2+\int\Big|\big(\partial_{r}+\frac{c}{r}\big)W_a\big(\frac{\mu_j}{\lambda_j}r\big)-\big(\partial_{r}+\frac{c}{r}\big)W_{a}(r)\Big|^2r^{N-1}\ud r\\
\lesssim&\eta^2+\int_{0}^{\infty}\Big|r-\frac{\mu_j}{\lambda_j}r\Big|^2\frac{r^{N-1}}{\big(r^{\frac{(\beta-1)N}{2}-\beta-1}+r^{-\frac{(\beta+1)N}{2}+\beta-1}\big)^2}\ud r\lesssim\eta^2.
\end{align*}
Collecting \eqref{l noteq j} and \eqref{l=j}, if $\boldsymbol{\lambda}\in B_{\eta}$, and $\eta=M_{J}\varepsilon$ for some large constant $M_J$ such that $\varepsilon\leq\varepsilon_{J}\ll M_{J}^{-1}$, then
\begin{align*}
\Big|\frac{\Phi_{\ell}(\boldsymbol{\lambda})}{\mu_{\ell}}-1\Big|\lesssim\varepsilon+\eta\varepsilon^{\frac{(N-2)\beta}{2}}+\eta(\eta+\varepsilon)\leq\eta.
\end{align*}
Thus, $\Phi$ maps $D_{\eta}$ into $D_{\eta}$. Similarly, we can prove $\Phi$ is a contraction mapping on $D_{\eta}$. By the Banach fixed point theorem, concluding the proof of Claim \ref{claim: contraction}.
\end{proof}
Due to Lemma \ref{lem: modulation}, we can change the scaling parameters $(\lambda_j)_j$ such that
\begin{align*
u_{0}=M_a+h_{0},
\end{align*}
and
\begin{align}\label{def: g1}
u_1=\sum\limits_{j=1}^{J}\alpha_{j}\left(\Lambda W_a\right)_{[\lambda_j]}+g_1.
\end{align}
Here $(h_0,g_1)$ satisfies the following orthogonality conditions: for all $j$,
\begin{align}\label{orthogonality 1}
\int\big(\partial_{r}+\frac{c}{r}\big)h_0\cdot\big(\partial_{r}+\frac{c}{r}\big)(\Lambda W_a)_{(\lambda_j)}=0,
\end{align}
\begin{align}\label{orthogonality 2}
\int g_1\cdot(\Lambda W_{a})_{[\lambda_j]}=0.
\end{align}
Before proving the estimates of $(h_0,g_1)$, we first introduce the following lemma.
\begin{lemma}\label{Lemma 2.8}
Let $(N-2)\beta\geq1$, and $M$ be a positive constant. For all $V\in L_{\text{loc}}^{\frac{2(N+1)}{N+4}}\left(\mathbb{R}, L^{\frac{2(N+1)}{3}}\left(\mathbb{R}^{N}\right)\right)$ with
\begin{align*}
\Big\|\chi_{\{|x| \geq|t|\}} V\Big\|_{L^{\frac{2(N+1)}{N+4}}\big(\mathbb{R}, L^{\frac{2(N+1)}{3}}(\mathbb{R}^{N})\big)} \leq M,
\end{align*}
Let $u$ be the solution of
\begin{align*}
\partial_{t}^{2} u-\Delta u+V u=f_{1}+f_{2}, \quad \vec{u}\mid_{ t=0}=\left(u_{0}, u_{1}\right) \in \mathcal{H}_a,
\end{align*}
where $f_{1} \in L^{1}\left(\mathbb{R}, L^{2}\left(\mathbb{R}^{N}\right)\right)$ and $f_{2} \in W^{\prime}$, then there exists $C_{M}>0$ such that
\begin{align}\label{potential outside a wave cone}
&\left\|\chi_{\{|x| \geq|t|\}}u\right\|_{{L}^{\frac{2(N+1)}{N-2}}\left(\mathbb{R} \times \mathbb{R}^{N}\right)} +\sup _{t \in \mathbb{R}}\left\|\chi_{\{|x| \geq|t|\}} \nabla_{t, x} u(t)\right\|_{L^{2}} \nonumber\\
\leq & C_{M}\left(\left\|\left(u_{0}, u_{1}\right)\right\|_{\mathcal{H}_a}+\left\|\chi_{\{|x| \geq|t|\}} f_{1}\right\|_{L^{1}\left(\mathbb{R}, L^{2}\right)}+\left\|\chi_{\{|x| \geq|t|\}} f_{2}\right\|_{W^{\prime}}\right) .
\end{align}
\end{lemma}
The proof easily follows from the H\"{o}lder's inequality and Strichartz estimates. Therefore, a rigorous proof of Lemma \ref{Lemma 2.8} is referred to \cite{DKM7}. Then as a consequence, we obtain the following estimate.
\medskip
\begin{lemma}\label{lem:proj of initial data}
Let $u$ be the solution of \eqref{weq} defined on $\left\{(t,x)\in\mathbb{R}^{N}:|x|>|t|\right\}$, then
\begin{align}\label{projection of initial data}
\left\|\operatorname{\pi}_{Z_{\lambda}^{a}}^{\bot}\Big((u_0,u_1)-(M_{a},0)\Big)\right\|_{\mathcal{H}_{a}}\lesssim\gamma^{\frac{(N+2)\beta}{4}}+\delta^{\frac{N+1}{N-2}}\lesssim\gamma^{\frac{N\beta}{4}}+\delta^{\frac{N-2}{2}}.
\end{align}
\end{lemma}
\begin{proof}
Let $h(t)=u(t)-M_{a}$, then $h(t)$ satisfies
\begin{align*
\left\{\begin{array}{l}
\partial_{t}^2 h+\mathcal{L}_{\lambda}^{a}h=F(h)+\mathcal{N}(h),\\
h(0)=(u_0,u_1)-(M_{a},0),
\end{array}\right.
\end{align*}
where
\begin{align*}
\mathcal{N}(h)=F(M+h)-\sum_{j=1}^{J} F\left(\iota_{j} {W_{a}}_{\left(\lambda_{j}\right)}\right)-F(h)-\frac{N+2}{N-2}\sum\limits_{j=1}^{J} {W_{a}}_{\left(\lambda_{j}\right)}^{\frac{4}{N-2}} h.
\end{align*}
Define $\tilde{h}(t)=\chi_{\{|x|\geq|t|\}}h(t)$. Thus $\tilde{h}(t)-h(t)$ satisfies
\begin{align*}
\left\{\begin{array}{l}
\big(\partial_{t}^2+\mathcal{L}_{\lambda}^{a}\big)(\tilde{h}(t)-h(t))=\chi_{\{|x|\geq|t|\}}\big(F(h)+\mathcal{N}(h)\big)\\
\big(\tilde{h}-h\big)(0)=(0,0),
\end{array}\right.
\end{align*}
assume $T>0$ and denote by $\Gamma(T)=\{(t,x),|t|\leq\min\{|x|, T\}\}$. Employing the fractional chain rule
\begin{align}\label{the norm W' for Fh}
\Big\|F(h)\chi_{\{|x|\geq|t|\}}\Big\|_{W^{\prime}((0, T))}=\left\|F(\tilde{h})\chi_{\{|x|\geq|t|\}}\right\|_{W^{\prime}((0, T))} \lesssim\|\tilde{h}\|_{{W}((0, T))}\|\tilde{h}\|_{S(\Gamma(T))}^{\frac{4}{N-2}}.
\end{align}
Recall Claim $A.5$ in \cite{DKM7}
\begin{align*}
\bigg|F\Big(\sum_{j=1}^{J} y_{j}+h\Big) &-\sum_{j=1}^{J} F\left(y_{j}\right)-F(h)-\frac{N+2}{N-2} \sum_{j=1}^{J}\left|y_{j}\right|^{\frac{4}{N-2}} h\bigg| \\
&\lesssim\sum_{j\neq k}\min \left(\left|y_{j}\right|^{\frac{4}{N-2}}\left|y_{k}\right|,\left|y_{k}\right|^{\frac{4}{N-2}}\left|y_{j}\right|\right)
+\sum_{j=1}^{J}\left|y_{j}\right|^{\frac{1}{N-2}}|h|^{\frac{N+1}{N-2}}.
\end{align*}
Given $y_j={W_a}_{(\lambda_j)}^{\frac{4}{N-2}}$, it follows that
\begin{align*}
|\mathcal{N}(h)| \lesssim \sum_{j \neq k} \min \Big({W_a}_{(\lambda_{j})}^{\frac{4}{N-2}} {W_a}_{(\lambda_{k})},{W_a}_{(\lambda_{k})}^{\frac{4}{N-2}} {W_a}_{(\lambda_{j})}\Big)+\sum_{j=1}^{J} {W_a}_{\left(\lambda_{j}\right)}^{\frac{1}{N-2}}|h|^{\frac{N+1}{N-2}} .
\end{align*}
When $j \neq k$, by Proposition \ref{prop: L1L2 bdd of Wa},
\begin{align}\label{L1L2 norm of Wa}
\left\|\chi_{\{|x| \geq|t|\}} \min \Big\{{W_a}_{\left(\lambda_{j}\right)}^{\frac{4}{N-2}} {W_a}_{\left(\lambda_{k}\right)}, {W_a}_{\left(\lambda_{j}\right)}^{\frac{4}{N-2}} {W_a}_{\left(\lambda_{k}\right)}\Big\}\right\|_{L_{t}^{1} L_{x}^{2}} \lesssim \gamma^{\frac{(N+2)\beta}{4}}.
\end{align}
Since $W_a^{\frac{1}{N-2}}\lesssim|x|^{-\frac{1+\beta}{2}}$ and $\beta>1$, $W_a^{\frac{1}{N-2}}\chi_{\{|x|>|t|\}}\in L_{t}^2(\mathbb{R},L_{x}^{\infty}(\mathbb{R}^{N}))$, we deduce
\begin{align*}
\Big\|\chi_{\Gamma(T)} {W_a}_{\left(\lambda_{j}\right)}^{\frac{1}{N-2}}|h|^{\frac{N+1}{N-2}}\Big\|_{L_{t}^{1} L^{2}} & \leq\Big\|\chi_{\Gamma(T)} {W_a}_{\left(\lambda_{j}\right)}^{\frac{1}{N-2}}\Big\|_{L_{t}^{2} L_{x}^{\infty}}\Big\|\chi_{\Gamma(T)}|h|^{\frac{N+1}{N-2}}\Big\|_{L_{t, x}^{2}} \lesssim\|\tilde{h}\|_{S(\Gamma(T))}^{\frac{N+1}{N-2}},
\end{align*}
which, combined with \eqref{the norm W' for Fh}, yields,
\begin{align*
\|\mathcal{N}(h)\|_{L^1L^2}\lesssim\gamma^{\frac{(N+2)\beta}{4}}+\|\tilde{h}\|_{S(\Gamma(T))}^{\frac{N+1}{N-2}}.
\end{align*}
We define $h_{L}(t)$ to be the solution of linear equation
\begin{align*
\left\{\begin{array}{l}
\partial_{t}^2h_{L}+\mathcal{L}_{\lambda}^{a}h_{L}=0, \\
h_{L}(0)=(u_0,u_1)-(M_{a},0).
\end{array}\right.
\end{align*}
It is apparent from \eqref{potential outside a wave cone} that
\begin{align*}
\Big\|\tilde{h}-h_{L}\Big\|_{S(\Gamma(T))} \lesssim \gamma^{\frac{(N+2)\beta}{4}}+\|\tilde{h}\|_{S(\Gamma(T))}^{\frac{N+1}{N-2}}+\|\tilde{h}\|_{S(\Gamma(T))}^{\frac{4}{N-2}}\|\tilde{h}\|_{W((0, T))}.
\end{align*}
Using the Strichartz estimate \eqref{ineq: Strichartz estimate}, we obtain
\begin{align*}
&\sup_{-T \leq t \leq T}\Big\|\overrightarrow{\tilde{h}}(t)-\vec{h}_{L}(t)\Big\|_{\mathcal{H}_a}+\Big\|\overrightarrow{\tilde{h}}(t)-\vec{h}_{L}(t)\Big\|_{S((0,T))\cap W((0,T))}\\
\lesssim& \gamma^{\frac{(N+2)\beta}{4}}+\|\tilde{h}\|_{S(\Gamma(T))}^{\frac{N+1}{N-2}}+\|\tilde{h}\|_{S(\Gamma(T))}^{\frac{4}{N-2}}\|\tilde{h}\|_{W((0,T))}.
\end{align*}
On account of $\|h_{L}\|_{S((0,T))\cap W((0,T))}\lesssim\delta$ and
\begin{align*}
\big\|\tilde{h}\big\|_{S((0,T))\cap W((0,T))}\lesssim\big\|(u_0-M_a,u_1)\big\|_{\mathcal{H}_a}+\gamma^{\frac{(N+2)\beta}{4}}\lesssim\delta+\gamma^{\frac{(N+2)\beta}{4}},
\end{align*}
we have
\begin{align}\label{the bdd of h subtract hL}
\sup\limits_{-T\leq t\leq T}\big\|\vec{\tilde{h}}(t)-\vec{h}_{L}(t)\big\|_{\mathcal{H}_a}\lesssim\gamma^{\frac{(N+2)\beta}{4}}+\delta^{\frac{N+1}{N-2}}.
\end{align}
whose right hand side is uniform in $T$, that is
\begin{align*}
\sup _{t \in \mathbb{R}}\left\|\vec{\tilde{h}}(t)-\vec{h}_{L}(t)\right\|_{\mathcal{H}_a}\lesssim \gamma^{\frac{(N+2)\beta}{4}}+\delta^{\frac{N+1}{N-2}}.
\end{align*}
Going back to the assume that $u$ is non-radiative, it holds that
\begin{align*}
0=&\sum\limits_{\pm}\lim\limits_{t\rightarrow\pm\infty}\int_{|x|\geq|t|}\big|\nabla_{t,x}u(t,x)\big|^2\ud x=\sum\limits_{\pm}\lim\limits_{t\rightarrow\pm\infty}\int_{|x|\geq|t|}\big|\nabla_{t,x}(h(t,x)+M_a)\big|^2\ud x\\
=&\sum\limits_{\pm}\lim\limits_{t\rightarrow\pm\infty}\int\big|\nabla_{t,x}(\tilde{h}+M_a)\big|^2\ud x\\
\lesssim&\gamma^{\frac{(N+2)\beta}{4}}+\delta^{\frac{N+1}{N-2}}.
\end{align*}
Combining \eqref{the bdd of h subtract hL} and Theorem \ref{thm:multisoliton energy}, we divide
\begin{align*}
\big\|\operatorname{\pi}_{Z_{\lambda}^{\perp}}(h_0,h_1)\big\|_{\mathcal{H}_a}\lesssim\gamma^{\theta_{N}}\delta+\gamma^{\frac{(N+2)\beta}{4}}+\delta^{\frac{N+1}{N-2}}.
\end{align*}
This completes the proof of Lemma \ref{lem:proj of initial data}.
\end{proof}
Now we are in the position to show the estimate of $(h_0,g_1)$.
\medskip
\begin{proposition}\label{prop: coefficients estimate}
\begin{align}\label{bdd: initial data}
\big\|(h_0,g_1)\big\|_{\mathcal{H}_{a}}\lesssim\gamma^{\frac{N\beta}{4}}+\delta^{\frac{N}{N-2}},
\end{align}
\begin{align}\label{bdd: delta2}
\Big|\delta^{2}-\sum\limits_{j=1}^{\infty}\alpha_{j}^2\|\Lambda W_{a}\|_{L^2}^2\Big|\lesssim\gamma^{\frac{(N-2)\beta-2}{2}}+\delta^{\frac{2(N-1)}{N-2}}.
\end{align}
\end{proposition}
\begin{proof}
According to \eqref{projection of initial data}, in view of the orthogonality condition \eqref{orthogonality 1} and the expansion \eqref{def: g1} of $u_1$, we get \eqref{bdd: initial data}. By \eqref{def: delta}, a direct computation gives
\begin{align*}
\delta^2=&\big\|(u_0-M_{a},u_1)\big\|_{\mathcal{H}_{a}}^{2}=\big\|h_0\big\|_{2}^{2}+\big\|g_1\big\|_{2}^{2}+\Big\|\sum\limits_{j=1}^{J}\alpha_{j}(\Lambda{W_{a}})_{[\lambda_{j}]}\Big\|_{2}^{2}\\
=&\big\|h_0\big\|_{2}^{2}+\big\|g_1\big\|_{2}^{2}+\sum\limits_{j=1}^{J}\alpha_{j}\big\|\Lambda{W_{a}}\big\|_{2}^{2}+\sum\limits_{j\neq k}\int\big|\alpha_{j}\alpha_{k}(\Lambda{W_{a}})_{[\lambda_{j}]}(\Lambda{W_{a}})_{[\lambda_{k}]}\big|\ud x.
\end{align*}
Using the expression of $W_{a}$ and \eqref{main W}, we obtain
\begin{align*}
\Big|\delta^2-\sum\limits_{j=1}^{J}\alpha_{j}\big\|\Lambda W_{a}\big\|_{2}^{2}\Big|\lesssim&\gamma^{\frac{(N-2)\beta}{2}-1}\sum\limits_{j=1}^{J}\alpha_{j}^2+\delta^{\frac{2(N+1)}{N-2}}+\gamma^{\frac{(N+2)\beta}{2}}\\
\lesssim&\delta^{\frac{2(N-1)}{N-2}}+\gamma^{\frac{(N-2)\beta-2}{2}},
\end{align*}
which completes the proof of \eqref{bdd: delta2}.
\end{proof}
\medskip
\subsection{Estimates about the exterior scaling parameter}\label{6.2}
In this paper, the projection space $P(R)$ \eqref{def: P(R)} is a subspace of $\mathcal{H}_{a}(R)$ spanned by $\mathcal{P}$
\begin{align*}
\mathcal{P}=\operatorname{span}\bigg\{&\Big(r^{-\frac{(N-2)\beta+(N+2)}{2}+2k_1},0\Big),\Big(0,r^{-\frac{(N-2)\beta+(N+2)}{2}+2k_2}\Big):\\
&k_1=1,2,...,\Big\lfloor\frac{(N-2)\beta}{4}+1\Big\rfloor,k_2=1,2,...,\Big\lfloor\frac{(N-2)\beta}{4}+\frac{1}{2}\Big\rfloor\bigg\},
\end{align*}
then the number of the basis in $\mathcal{P}$ is
\begin{align*}
\Big\lfloor\frac{(N-2)(\beta-1)+(N+2)}{4}\Big\rfloor+\Big\lfloor\frac{(N-2)(\beta-1)+N}{4}\Big\rfloor=k,
\end{align*}
and the norm of the elements of $\mathcal{P}$ are
\begin{align}\label{the first norm of P}
\left\|\big(r^{-\frac{(N-2)\beta+(N+2)}{2}+2k_1},0\big)\right\|_{\mathcal{H}_{a}(R)}^2=\frac{\big((N-2)\beta+(N+2)-4k_1\big)^2}{4\big((N-2)\beta-4k_1+4\big)}R^{-(N-2)\beta+4k_1-4},
\end{align}
and
\begin{align}\label{the second norm of P}
\left\|\big(0,r^{-\frac{(N-2)\beta+(N+2)}{2}+2k_2}\big)\right\|_{\mathcal{H}_{a}(R)}^2=\frac{1}{\big((N-2)\beta-4k_2+2\big)}R^{-(N-2)\beta+4k_2-2}.
\end{align}
Rewrite $\mathcal{P}=\{\Xi_{m}\}_{m\in\llbracket1,k\rrbracket}$ and make the assumption $\|\Xi_{m}\|_{\mathcal{H}_{a}(R)}\approx\frac{c_m}{R^{m-a}}$. We distinguish between two cases.
\begin{itemize}
\item We first consider the case where $m=1$, the norm is
\begin{align*}
\|\Xi_{1}\|_{\mathcal{H}_{a}(R)}^2\approx\frac{c_1^2}{R^{2-2a}}.
\end{align*}
By employing \eqref{the first norm of P} and \eqref{the second norm of P}, choosing $a=-k+\frac{N}{2}$, we observe that
\begin{itemize}
\item If $k$ is odd,
\begin{align*}
(N-2)\beta-4k_1-4=2-2a,\quad r^{-\frac{(N-2)\beta+(N+2)}{2}+2k_1}=r^{-k}.
\end{align*}
\item If $k$ is even
\begin{align*}
(N-2)\beta-4k_2+2=2-2a,\quad r^{-\frac{(N-2)\beta+(N+2)}{2}+2k_2}=r^{-k-1}.
\end{align*}
\end{itemize}
\item We next assume that $m=k$. Then $\|\Xi_k\|_{\mathcal{H}_{a}(R)}^2=\frac{c_k^2}{R^{2k-2a}}$. By \eqref{the first norm of P} and the assumption $a=-k+\frac{N}{2}$, we have the element satisfies
\begin{align*}
r^{-\frac{(N-2)\beta+(N+2)}{2}+2k_1}=r^{-(N-2)\beta},
\end{align*}
this imply $\Xi_k(r)=(r^{-(N-2)\beta},0)$.
\end{itemize}
Summarize these two cases, we denote by $\mathcal{P}=\{\Xi_m\}_{m\in\llbracket1,k\rrbracket}$ and choosing $\Xi_m$ such that
\begin{align}\label{def: the norm of Xi}
\|\Xi_m\|_{\mathcal{H}_a(R)}=\frac{c_m}{R^{m+k-\frac{N}{2}}},
\end{align}
for some constant $c_m\neq 0$.
Incorporating with scaling transformation, we have the following lemma.
\begin{lemma}\label{lem: the norm of U}
Let $U\in P(R)$ and denote by $(\theta_{m}(R))_{1\leq m\leq k}$ its coordinates in $\mathcal{P}$. Then
\begin{align*}
\|U\|_{\mathcal{H}_{a}(R)}\approx\sum\limits_{m=1}^{k}\frac{|\theta_{m}(R)|}{R^{m+k-\frac{N}{2}}}.
\end{align*}
\end{lemma}
\begin{proof}
Assume $U=(u,v)\in\mathcal{H}_{a}(R)$, it follows that $\int_{R}^{\infty}\left(|(\partial_{r}+\frac{c}{r})u|^2+|v|^2\right)r^{N-1}\ud r<\infty$. Consider
\begin{align*}
U_{R^{-1}}(x)=\left(R^{\frac{N}{2}-1}u(Rx),R^{\frac{N}{2}}v(Rx)\right).
\end{align*}
A routine computation shows that
\begin{align*
\big\|U_{R^{-1}}\big\|_{\mathcal{H}_{a}(1)}=&\int_{1}^{\infty}\Big(\big|(\partial_{r}+\frac{c}{r})u(Rx)R^{-1}\big|^2+|v(Rx)|^2\Big)(R r)^{N-1}R\ud r\nonumber\\
=&\int_{R}^{\infty}\Big(\big|(\partial_{r}+\frac{c}{r})u(r)\big|^2+|v(r)|^2\Big)r^{N-1}\ud r\nonumber\\
=&\big\|U\big\|_{\mathcal{H}_a(R)}<\infty,
\end{align*}
thus $U_{R^{-1}}\in\mathcal{H}_{a}(1)$. According to \eqref{def: the norm of Xi} we deduce
\begin{align*}
\frac{c_m}{R^{m+k-\frac{N}{2}}}=\big\|\Xi_{m}\big\|_{\mathcal{H}_{a}(R)}=\big\|(\Xi_{m})_{R^{-1}}\big\|_{\mathcal{H}_{a}(1)},
\end{align*}
which implies $(\Xi_{m})_{R^{-1}}=\frac{1}{R^{m+k-\frac{N}{2}}}\Xi_{m}$.
From the assumption $U=\sum\limits_{m=1}^{k}\theta_{m}(R)\Xi_{m}$, it follows that
\begin{align*}
U_{R^{-1}}=\sum\limits_{m=1}^{k}\theta_{m}(R)(\Xi_{m})_{R^{-1}}=\sum\limits_{m=1}^{k}\frac{\theta_{m}(R)}{R^{m+k-\frac{N}{2}}}\Xi_m.
\end{align*}
Using $U_{R^{-1}}$ and \eqref{def: the norm of Xi}, we obtain
\begin{align*}
\big\|U\big\|_{\mathcal{H}_{a}(R)}=\big\|U_{R^{-1}}\big\|_{\mathcal{H}_{a}(1)}=\Big\|\sum\limits_{m=1}^{k}\frac{\theta_{m}(R)}{R^{m+k-\frac{N}{2}}}\Xi_{m}\Big\|_{\mathcal{H}_a(1)}\approx\sum\limits_{m=1}^{k}\frac{|\theta_{m}(R)|}{R^{m+k-\frac{N}{2}}},
\end{align*}
where the implicit constant is independent of $R>0$. This concludes the proof of Lemma \ref{lem: the norm of U}.
\end{proof}
\medskip
By the same arguments as in \cite{DKM3}, we get the following two theorems. The proofs are omitted.
\begin{theorem}\label{thm: main thm of Xi}\emph{(\cite{DKM3})}
Assume $(N-2)\beta\geq 3$ and $(N-2)\beta$ is odd. Let $u$ be a radial weakly non-radiative solution of \eqref{weq} and $\varepsilon_{0}>0$. Then there exist $m_{0}\in\llbracket1,m\rrbracket,\ell\in\mathbb{R}$ (with $\ell \neq 0$ if $m_{0}<k$) such that, for all $t_{0}\in\mathbb{R}$ and $R_{0}>0$, if $u$ is defined on
$\left\{(t,r): r>\left|t-t_{0}\right|+R_{0}\right\}$, $\left\|\vec{u}\left(t_{0}\right)\right\|_{\mathcal{H}_{a}\left(R_{0}\right)}<\varepsilon_{0}$ and
\begin{align*}
\sum_{\pm} \lim _{t \rightarrow \pm \infty} \int_{|x|>\left|t-t_{0}\right|+R_{0}}\big|\nabla_{t, x} u(t, r)\big|^{2} d x=0,
\end{align*}
then \textcolor{red}{for any} $R>R_{0}$,
\begin{align*}
\quad\big\|\vec{u}\left(t_{0}\right)-\ell\Xi_{m_{0}}\big\|_{\mathcal{H}_{a}(R)}\leq C\max \bigg\{\Big(\frac{R_{0}}{R}\Big)^{\left(m_{0}+k-\frac{N}{2}\right) \frac{N+2}{N-2}},\Big(\frac{R_{0}}{R}\Big)^{m_{0}+\frac{N}{2}-k}\bigg\}.
\end{align*}
\end{theorem}
\begin{theorem}\label{thm: main thm of Xi 2}
Assume $(N-2)\beta\geq 3$ and $(N-2)\beta$ is odd. Let $u$ be a radial non-radiative solution of \eqref{weq}. If $m_0=k$, then $u$ is a stationary solution.
\end{theorem}
\begin{proposition}\label{exterior profile estimate}
For $m_0\leq k-1$, there exist a constant $C>0$ such that if $u$ is a radial solution of \eqref{weq} defined on $\left\{(t,x)\in\mathbb{R}^{N}:|x|>t\right\}$, then
\begin{align*}
|\ell|\leq C\delta^{\frac{2N}{N-2}}\lambda_{1}^{m_0+k-\frac{N}{2}}.
\end{align*}
\end{proposition}
\begin{proof}
By scaling, we find that for any $j\in\llbracket1, J\rrbracket$
\begin{align*}
\big\|{W_a}_{(\lambda_j)}\big\|_{\dot{H}_a^1(R)}^2=&\int_{R}^{\infty}\Big|\big(\partial_{r}+\frac{c}{r}\big){W_a}_{(\lambda_j)}\Big|^2r^{N-1}\ud r=\big\|W_a\big\|_{\dot{H}_a^1(\frac{R}{\lambda_j})}^2.
\end{align*}
It follows from the definition of $\{\lambda_j\}_{1\leq j\leq J}$ that $\frac{R}{\lambda_j}\geq\frac{R}{\lambda_1}\gg1$, this means exactly that
\begin{align*}
W_a(r)\sim r^{-\frac{(N-2)(1+\beta)}{2}}.
\end{align*}
Then
\begin{align*}
\big\|{W_a}_{(\lambda_j)}\big\|_{\dot{H}_a^1(R)}=&\big\|W_{a}\big\|_{\dot{H}_a^1(\frac{R}{\lambda_j})}\\
=&\Big(\int_{R/{\lambda_j}}^{\infty}\Big(\big(\partial_{r}+\frac{c}{r}\big)r^{-\frac{(N-2)(1+\beta)}{2}}\Big)^2r^{N-1}\ud r\Big)^{\frac{1}{2}}=\Big(\frac{R}{\lambda_j}\Big)^{-\frac{(N-2)\beta}{2}}.
\end{align*}
For $(N-2)\beta\geq3$, it follows from $\|\vec{u}(0)-(M_a,0)\|_{\mathcal{H_a}(R)}\lesssim\delta$ that
\begin{align}\label{the bdd of the initial data}
\big\|\vec{u}(0)\big\|_{\mathcal{H}_{a}(R)}\lesssim&\big\|\vec{u}(0)-(M_a,0)\big\|_{\mathcal{H}_{a}(R)}+\left\|\sum\iota_j{W_a}_{(\lambda_j)}\right\|_{\dot{H}_a^1(R)}\nonumber\\
\lesssim&\delta+\Big(\frac{\lambda_1}{R}\Big)^{\frac{(N-2)\beta}{2}}=\delta+\Big(\frac{\lambda_1}{R}\Big)^{k-\frac{1}{2}}.
\end{align}
Let $\varepsilon_0>0$ such that $\|\vec{u}(0)\|_{\mathcal{H}_{a}(\lambda_1)}<\varepsilon_0$. Fix $B>0$ large enough, due to the fact that $\|\vec{u}(0)-(M_a,0)\|_{\mathcal{H}_{a}(R)}=\delta\leq\varepsilon_J\ll 1$, we have that $\|\vec{u}_0\|_{\mathcal{H}_{a}(R\lambda_1)}\leq\varepsilon_0$. For all $R>B\lambda_1$, from Theorem \ref{thm: main thm of Xi}, we get that for $m_0\leq k-1$,
\begin{align*}
\left|\big\|\vec{u}(0)\big\|_{\mathcal{H}_{a}(R)}-\frac{\ell c_{m_0}}{R^{m_0+k-\frac{N}{2}}}\right|\leq C\max\bigg\{\Big(\frac{B\lambda_1}{R}\Big)^{(m_0+k-\frac{N}{2})\frac{N+2}{N-2}},\Big(\frac{\lambda_1}{R}\Big)^{m_0-k+\frac{N}{2}}\bigg\},
\end{align*}
for all $R\geq B\lambda$, thus it follows from \eqref{the bdd of the initial data} that
\begin{align*}
\frac{\ell}{R^{m_0+k-\frac{N}{2}}}\lesssim&\delta+\Big(\frac{\lambda_1}{R}\Big)^{k-\frac{1}{2}}+\max\bigg\{\Big(\frac{\lambda_1}{R}\Big)^{(m_0+k-\frac{N}{2})\frac{N+2}{N-2}},\Big(\frac{\lambda_1}{R}\Big)^{m_0-k+\frac{N}{2}}\bigg\}\\
\lesssim&\delta+\Big(\frac{\lambda_1}{R}\Big)^{a_{m_0}},
\end{align*}
where $a_{m_0}=\min\left\{(m_0+k-\frac{N}{2})\frac{N+2}{N-2},m_0-k+\frac{N}{2}\right\}$.
Choose $R$ such that $\left(\frac{\lambda_1}{R}\right)^{a_{m_0}}=\delta$, then $R=\lambda_1\delta^{-\frac{1}{a_{m_0}}}$, and we have
\begin{align*}
|\ell|\lesssim R^{m_0+k-\frac{N}{2}}\delta=\left(\lambda_1\delta^{-\frac{1}{a_{m_0}}}\right)^{m_0+k-\frac{N}{2}}\delta\leq{\lambda_1}^{m_0+k-\frac{N}{2}}\delta^{1-\frac{m_0+k-\frac{N}{2}}{a_{m_0}}}.
\end{align*}
Owing to
\begin{align*}
\min\limits_{1\leq m_0\leq k-1}1-\frac{m_0+k-\frac{N}{2}}{a_{m_0}}=\frac{2N}{N-2},
\end{align*}
this means exactly that
\begin{align*}
|\ell|\lesssim{\lambda_1}^{m_0+k-\frac{N}{2}}\delta^{\frac{2N}{N-2}}.
\end{align*}
\end{proof}
\section{Differential inequalities of parameters}\label{section 7}
We now suppose that $u$ is a global solution of \eqref{weq} such that
\begin{align*}
\limsup _{t \rightarrow+\infty}\big\|\vec{u}(t)\big\|_{\mathcal{H}_{a}}<\infty.
\end{align*}
Let $v_{L}$ be the unique solution of the free wave equation \eqref{lwp} such that for any $A\in\mathbb{R}$
\begin{align*}
\lim _{t \rightarrow+\infty} \int_{|x| \geq A+|t|}\big|\nabla_{t, x}\left(u-v_{L}\right)(t, x)\big|^{2} d x=0.
\end{align*}
The existence of $v_L$ can be proved as in \cite{CR}.
For $J \geq 1, \iota_j \in\{\pm 1\}^{J},(f, g) \in \mathcal{H}_{a}$, we define its distance to the multi-soliton as
\begin{align*}
d_{J, \iota}(f, g)=\inf_{\lambda \in G_{J}}\bigg\{\Big\|(f, g)-\sum_{j=1}^{J} \iota_{j}\left({W_{a}}_{\left(\lambda_{j}\right)}, 0\right)\Big\|_{\mathcal{H}_{a}}+\gamma(\boldsymbol{\lambda})\bigg\},
\end{align*}
In Section \ref{Soliton resolution along a time sequence}, there exist $J\geq 1$, $\iota \in\{\pm 1\}^{J}$ and a time sequence $\{t_n\}_n$ such that
\begin{align*}
\lim_{n\rightarrow\infty}d_{J,\iota}\big(\vec{u}(t_{n})-\vec{v}_{L}(t_{n})\big)=0.
\end{align*}
In order to prove the soliton resolution for all time, we assume by contradiction that there exist a sequence $\{\tilde{t}_n\}$ satisfying $\tilde{t}_n<t_n$ for any $n$, and a small number $\varepsilon>0$ such that if $t\in\left(\tilde{t}_{n}, t_{n}\right]$, there holds
\begin{align}\label{distance d t}
d_{J, \iota}\big(\vec{u}(t)-\vec{v}_{L}(t)\big)<\varepsilon_{0},
\end{align}
and
\begin{align}\label{distance d tn}
d_{J,\iota}\big(\vec{u}(\tilde{t}_{n})-\vec{v}_{L}(\tilde{t}_{n})\big)=\varepsilon_{0}.
\end{align}
Since $h(t)=u(t)-v_{L}(t)-M_{a}(t)$, according to Lemma \ref{lem: modulation}, for $t \in\left[\tilde{t}_{n}, t_{n}\right]$ there exists $\boldsymbol{\lambda}(t)\in G_{J}$ such that the orthogonal conditions
\begin{align}\label{eq: orthogonal condition 1}
\int\big(\partial_{r}+\frac{c}{r}\big)h(t)\cdot\big(\partial_{r}+\frac{c}{r}\big)(\Lambda W_{a})_{(\lambda_{j}(t))}\ud x=0.
\end{align}
hold. Moreover, we also have
\begin{align}\label{def: dJ, gamma}
d_{J, \iota}\big(\vec{u}(t)-v_{L}(t)\big)\approx \big\|\vec{u}(t)-\vec{v}_{L}(t)-(M_{a}(t), 0)\big\|_{\mathcal{H}_{a}}+\gamma(\lambda).
\end{align}
By the definition of $\delta(t)$ \eqref{def: delta}, there holds
$$
\delta(t)=\sqrt{\big\|h(t)\big\|_{\dot{H}_{a}^{1}}^{2}+\big\|\partial_{t}\left(u-v_{L}\right)(t)\big\|_{L^{2}}^{2}} .
$$
Similar to \eqref{def: g1}, we expand $\partial_{t} U=\partial_{t} u-\partial_{t} v_{L}$ as
\begin{align}\label{the expansion of partial t U}
\partial_{t} U(t)=\sum_{j=1}^{J}\alpha_{j}(t)\iota_{j}\Lambda{W_{a}}_{\left[\lambda_{j}(t)\right]}+g_{1}(t).
\end{align}
where $g_1$ satisfies the orthogonality conditions
\begin{align*
\int g_{1}(t) \Lambda W_{\left[\lambda_{j}(t)\right]}=0,\quad \forall j \in\llbracket1, J\rrbracket.
\end{align*}
We also define
\begin{align}\label{def: beta j t}
\beta_{j}(t)=&-\iota_{j} \int(\Lambda W_a)_{[\lambda_{j}(t)]} \partial_{t} U(t)\ud x\nonumber\\
=&-\iota_{j}\sum\limits_{k\neq j}\iota_{k}\alpha_{k}(t)\int(\Lambda W_{a})_{[\lambda_{j}(t)]}(\Lambda W_{a})_{[\lambda_{k}(t)]}\ud x-\iota_{j}^{2}\alpha_{j}(t)\big\|\Lambda W_{a}\big\|_{2}^{2}.
\end{align}
\subsection{Non-radiative profile decomposition}
There exists $1=j_{1}<j_{2}<\ldots<j_{K+1}=J+1$ with $\llbracket1, J\rrbracket=\cup_{k=1}^{K}\llbracket j_{k}, j_{k+1}-1\rrbracket$ satisfying the following relationship
\begin{itemize}
\item For all $k\in \llbracket1, K\rrbracket$
\begin{align}\label{eq: in one term}
\nu_{j}=\lim _{n \rightarrow \infty} \frac{\lambda_{j}\left(s_{n}\right)}{\lambda_{j_{k}}\left(s_{n}\right)}>0,\quad \forall j \in \llbracket j_{k}, j_{k+1}-1\rrbracket.
\end{align}
\item For all $k\in \llbracket1, K-1\rrbracket$
\begin{align}\label{eq: not in a term}
\lim _{n \rightarrow \infty} \frac{\lambda_{j_{k+1}}\left(s_{n}\right)}{\lambda_{j_{k}}\left(s_{n}\right)}=0.
\end{align}
\end{itemize}
Thus we divide the interval $\llbracket1, J\rrbracket$ into $K$ sub-intervals.
We note that $\nu_{j_{k}}=1$. Employing the same arguments as in \cite{DKM7}, we can prove
\begin{lemma}\label{lem: nonradiative profile}
For all $k \in\llbracket1, K\rrbracket$, there exists $\left(V_{0}^{k}, V_{1}^{k}\right)\in\mathcal{H}^{a}_{rad}$ such that the solution $V^{k}$ of \eqref{weq} with initial data $\left(V_{0}^{k}, V_{1}^{k}\right)$ is defined on $\{|x|>|t|\}$ and is non-radiative. Furthermore,
Assume in the k-th sub-interval $J^{k}=$ $j_{k+1}-j_{k}, \iota^{k}=\left(\iota_{j_{k}}, \ldots, \iota_{j_{k+1}-1}\right)$ and
\begin{align*}
V_{n}^{k}(t, x)=\frac{1}{\lambda_{j_{k}}^{\frac{N-2}{2}}\left(s_{n}\right)} V^{k}\Big(\frac{t}{\lambda_{j_{k}}\left(s_{n}\right)}, \frac{x}{\lambda_{j_{k}}\left(s_{n}\right)}\Big),
\end{align*}
we have (extracting subsequences if necessary),
\begin{align*
\lim _{n \rightarrow \infty}\bigg\|\vec{u}\left(s_{n}\right)-\vec{v}_{L}\left(s_{n}\right)-\sum_{k=1}^{K} \vec{V}_{n}^{k}(0)\bigg\|_{\mathcal{H}_{a}}=0,
\end{align*}
and
\begin{align*}
\quad d_{J^{k}, \iota^{k}}\left(V_{0}^{k}, V_{1}^{k}\right) \leq C \varepsilon_{0}.
\end{align*}
More precisely, after extracting a subsequence, $(V_0^{k},V_1^k)$ can be decomposed into
\begin{align}\label{eq: initial date}
\left\{\begin{array}{l}
V_{0}^{k}=\sum\limits_{j=j_{k}}^{j_{k+1}-1} \iota_{j} {W_{a}}_{\left(\nu_{j}\right)}+\bar{h}_{0}^{k}, \\
V_{1}^{k}=\sum\limits_{j=j_{k}}^{j_{k+1}-1} \iota_{j} \bar{\alpha}_{j}(\Lambda W_{a})_{\left[\nu_{j}\right]}+\bar{g}_{1}^{k},
\end{array}\right.
\end{align}
where
\begin{align}\label{eq: weak limit}
\begin{array}{ll}
\bar{h}_{0}^{k}=&\underset{n\rightarrow\infty}{w-\lim}\lambda_{j_{k}}^{\frac{N-2}{2}}(s_{n})h(s_{n}, \lambda_{j_{k}}\left(s_{n}\right) \cdot),\\
\bar{\alpha}_{j}=&\lim\limits_{n\rightarrow\infty}\alpha_{j}\left(s_{n}\right),\\
\bar{g}_{1}^{k}=&\underset{n\rightarrow\infty}{w-\lim}\lambda_{j_{k}}^{\frac{N}{2}}\left(s_{n}\right)g_{1}\left(s_{n}, \lambda_{j_{k}}\left(s_{n}\right) \cdot\right).
\end{array}
\end{align}
Furthermore, we have that
\begin{align}\label{JE Wa}
JE(W_{a},0)=\sum\limits_{k=1}^{K} E\Big(\vec{V}^{k}(0)\Big).
\end{align}
where the energy is defined in \eqref{energy}. Note that the limits \eqref{eq: not in a term} and \eqref{eq: weak limit} imply for any $j$ in the same sub-interval, we have orthogonality conditions as before
\begin{align*}
\int\Big(\partial_{r}+\frac{c}{r}\Big)\bar{h}_{0}^{k}\cdot\Big(\partial_{r}+\frac{c}{r}\Big)(\Lambda W_{a})_{\left(\nu_{j}\right)}\ud x=\int \bar{g}_{1}^{k} \cdot\left(\partial_{r}+\frac{c}{r}\right)(\Lambda W_{a})_{\left[\nu_{j}\right]}\ud x=0.
\end{align*}
\end{lemma}
\subsection{Estimates of the parameters and their derivatives}
Using the profile decomposition in Lemma \ref{lem: nonradiative profile} and the estimates we have obtained in prop \ref{prop: coefficients estimate}, we can prove the following lemma by contradiction and we omit the proof because it is similar to \cite{DKM7}.
\begin{lemma}\label{lem: estimate of the parameters}
For $(N-2)\beta>2$. There exists a constant $C>0$, depending only on $J$ and $N$ and $\beta$, such that under the preceding assumptions,
\begin{align}
\big\|\left(h, g_{1}\right)\big\|_{\mathcal{H}_{a}} \leq& o_{n}(1)+C\Big(\gamma^{\frac{N}{4}\beta}+\delta^{\frac{N}{N-2}}\Big),\quad \forall t \in\left[\tilde{t}_{n}, t_{n}\right],\label{ineq: h0,g1}\\
\Big|\delta^2-\sum\limits_{j=1}^{J}\alpha_{j}^2\|\Lambda W_{a}\|_{2}^{2}\Big|\leq& o_{n}(1)+C\Big(\gamma^{\frac{(N-2)\beta-2}{2}}+\delta^{\frac{2(N-1)}{N-2}}\Big),\quad \forall t \in\left[\tilde{t}_{n}, t_{n}\right],\label{the bdd of alpha and delta}\\
\Big|\beta_{j}+\alpha_{j}\|\Lambda W\|_{L^{2}}^{2}\Big| \leq& o_{n}(1)+C\Big(\gamma^{\frac{(N-2)\beta-2}{4}}+\delta^{\frac{N}{N-2}}\Big). \quad \forall t \in\left[\tilde{t}_{n}, t_{n}\right].\label{the bdd of beta and alpha}
\end{align}
Here $o_{n}(1)$ represents a term that goes to 0 as $n$ goes to infinity, uniformly with respect to $t \in\left[\tilde{t}_{n}, t_{n}\right]$.
\end{lemma
\begin{lemma}\label{lem: expansion of energy}
For $(N-2)\beta>2$, then
\begin{align}
\big\|(h(t),g_{1}(t))\big\|_{\mathcal{H}_{a}}\lesssim& \gamma^{\frac{N\beta}{4}}+o_{n}(1),\label{the estimate of h,g}\\
\delta\lesssim& \gamma^{\frac{(N-2)\beta}{4}}+o_{n}(1), \label{the estimate of delta}\\
\bigg|\frac{1}{2}\delta^2-\kappa'_{1}\sum\limits_{1\leq j\leq J-1}\iota_{j}\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}\bigg|\leq&\gamma^{\frac{(N-2)\beta-2}{2}}+o_{n}(1),\label{the bdd of delta and kappa1}
\end{align}
where $o_{n}(1)\rightarrow 0$ as $n \rightarrow \infty$, uniformly with respect to $t\in\left[\tilde{t}_{n},t_{n}\right]$, and
\begin{align*}
\kappa'_{1}=\big(N(N-2)\beta^2\big)^{\frac{N-2}{2}}\int_{\Bbb R^N}\frac{1}{|x|^{\frac{(\beta+1)(N-2)}{2}}}{W_a}^{\frac{N+2}{N-2}}(x)\ud x=\frac{1}{\|\Lambda W_a\|_{L^2}^2}\kappa_{1}.
\end{align*}
\end{lemma
\begin{proof}
Recall that the energy of \eqref{weq} is given by
\begin{align}\label{def: erengy}
E_{a}[\vec{u}]=\int\frac{1}{2}|\nabla u|^2+\frac{1}{2}|\partial_{t}u|^2+\frac{1}{2}\frac{a}{|x|^2}|u|^2-\frac{N-2}{2N}|u|^{\frac{2N}{N-2}},
\end{align}
where $\vec{u}=(u,\partial_{t}u)$. Then we expand the energy \eqref{def: erengy} of $\vec{u}-\vec{v_{L}}$,
\begin{align*}
E_{a}(\vec{u}-\vec{v_L})=&E_{a}\Big(\sum\limits_{j=1}^{J}\iota_{j}{W_a}_{(\lambda_j)}+h_0,\sum\limits_{j=1}^{J}\alpha_{j}\iota_{j}\left(\Lambda W_a\right)_{[\lambda_j]}+g_1\Big)\\
=&\frac{1}{2}\int_{\Bbb R^N}\Big|\sum\limits_{j=1}^{J}\iota_j\nabla{W_a}_{(\lambda_j)}+\nabla h_0\Big|^2+\Big|\sum\limits_{j=1}^{J}\alpha_j\iota_j(\Lambda W_a)_{[\lambda_j]}+g_1\Big|^2\ud x\\
&+\frac{1}{2}\int_{\Bbb R^N}\frac{a}{|x|^2}\Big|\sum\limits_{j=1}^{J}\iota_j{W_a}_{(\lambda_j)}+h_0\Big|^2\ud x-\frac{N-2}{2N}\int_{\Bbb R^N}\Big|\sum\limits_{j=1}^{J}\iota_j{W_a}_{(\lambda_j)}+h_0\Big|^{\frac{2N}{N-2}}\ud x.
\end{align*}
Going back to the orthogonality \eqref{orthogonality 2}, integration by parts and using the equation
\begin{align*}
-\Delta {W_a}_{(\lambda_j)}+\frac{a}{|x|^2}{W_a}_{(\lambda_j)}=|{W_a}_{(\lambda_j)}|^{\frac{4}{N-2}}{W_a}_{(\lambda_j)},
\end{align*}
this energy can be further rewritten as
\begin{align*}
E_a(\vec{u}-\vec{v_L})=&\frac{J}{2}\big\|W_a\big\|_{\dot{H}^{1}_a}^2+\big\|h_0\big\|_{\dot{H}^{1}_a}^2+\frac{1}{2}\sum\limits_{j=1}^{J}\alpha_j^2\big\|\Lambda W_a\big\|_{L^2}^2+\big\|g_1\big\|_{L^2}^2\\
&+\sum\limits_{j=1}^{J}\iota_{j}\int_{\Bbb R^N}\Big(|{W_a}_{(\lambda_j)}|^{\frac{4}{N-2}}{W_a}_{(\lambda_j)}\Big)h_0\ud x+\sum\limits_{1\leq j,k\leq J}\int_{\Bbb R^N}\Big(|{W_a}_{(\lambda_j)}|^{\frac{4}{N-2}}{W_a}_{(\lambda_j)}\Big){W_a}_{(\lambda_k)}\ud x\\
&+\sum\limits_{1\leq j,k\leq J}\alpha_j\alpha_k\int_{\Bbb R^N}(\Lambda{W_a})_{[\lambda_j]}(\Lambda W_a)_{[\lambda_k]}\ud x-\frac{N-2}{2N}\int_{\Bbb R^N}\Big|\sum\limits_{j=1}^{J}\iota_j{W_a}_{(\lambda_j)}+h_0\Big|^{\frac{2N}{N-2}}\ud x.
\end{align*}
According to \eqref{JE Wa},
\begin{align*}
\lim\limits_{t\rightarrow\infty}E_{a}(\vec{u}-\vec{v_L})=\sum\limits_{k=1}^{K}\lim\limits_{t\rightarrow\infty}E_{a}(\vec{V}^{k}(0))=J E_a(W_a,0).
\end{align*}
Plugging \eqref{nonlinear estimate for y1 to yJ} into the energy term and taking $y_j=\iota_j {W_a}_{(\lambda_j)}$, it follows that
\begin{align*}
&\bigg|\frac{J}{2}\big\|W_a\big\|_{\dot{H}_{a}^1}^{2}+\frac{1}{2}\sum\limits_{j=1}^{J}\alpha_{j}^{2}\big\|\Lambda W_a\big\|_{L^2}^2-J E_{a}(W_a,0)-\frac{N-2}{2N}J\int{W_a}^{\frac{2N}{N-2}}\ud x\\
&+\sum\limits_{1\leq j<k\leq J}\iota_j\iota_k\int_{\Bbb R^N} {W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}{W_a}_{(\lambda_k)}\ud x\bigg|\\
\lesssim&
\sum_{1\leq j<k\leq J}\int_{\Bbb R^N}\bigg(\min\Big\{{W_a}_{(\lambda_{j})}^{\frac{4}{N-2}}{W_a}_{(\lambda_{k})}^{2}, {W_a}_{(\lambda_{k})}^{\frac{4}{N-2}} {W_a}_{(\lambda_{j})}^{2}\Big\}+\min\Big\{{W_a}_{(\lambda_{j})}^{\frac{N+2}{N-2}} {W_a}_{(\lambda_{k})}, {W_a}_{(\lambda_{k})}^{\frac{N+2}{N-2}} {W_a}_{(\lambda_{j})}\Big\}\bigg)\ud x \\
&+\big\|g_1\big\|_{L^2}^2+\big\|h_0\big\|_{\dot{H}_{a}^1}^2+\gamma^{\frac{(N-2)\beta-2}{2}}\delta^2+o_{n}(1),
\end{align*}
where we have used the fact $\int|(\Lambda{W_a})_{[\lambda_j]}(\Lambda{W_a})_{[\lambda_k]}|\lesssim\gamma^{\frac{(N-2)\beta-2}{2}}$.
On one hand, by the inequality \eqref{ineq: Wa N N-2} we get
\begin{align*}
&\int_{\Bbb R^N}\min\Big\{W_{(\lambda_{j})}^{\frac{4}{N-2}}{W_a}_{(\lambda_{k})}^{2}, {W_a}_{(\lambda_{k})}^{\frac{4}{N-2}} {W_a}_{(\lambda_{j})}^{2}\Big\}\ud x+\int_{\Bbb R^N}\min\Big\{{W_a}_{(\lambda_{j})}^{\frac{N+2}{N-2}} {W_a}_{(\lambda_{k})}, {W_a}_{(\lambda_{k})}^{\frac{N+2}{N-2}} {W_a}_{(\lambda_{j})}\Big\}\ud x\\
\lesssim& \int_{\Bbb R^N}{W_a}_{(\lambda_j)}^{\frac{N}{N-2}}{W_a}_{(\lambda_k)}^{\frac{N}{N-2}}\ud x\lesssim\gamma^{\frac{N\beta}{2}}.
\end{align*}
On the other hand, using the estimate \eqref{ineq: h0,g1} on $(h_0,g_1)$, we obtain
\begin{align}\label{the bdd of alpha and Wa}
\bigg|\frac{1}{2}\sum\limits_{j=1}^{J}\alpha_{j}^2\big\|\Lambda W_a\big\|_{L^2}^2+\sum\limits_{1\leq j<k\leq J}\iota_j\iota_k&\int_{\Bbb R^N} {W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}{W_a}_{(\lambda_k)}\ud x\bigg|\nonumber\\
\lesssim&\gamma^{\frac{N\beta}{2}}+\delta^{\frac{2N}{N-2}}+\gamma^{\frac{(N-2)\beta-2}{2}}\delta^2+o_{n}(1).
\end{align}
Now it remains to estimate $\int {W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}{W_a}_{(\lambda_k)}\ud x$.
For $j<k$, by the relationship of $\{\lambda_j\}_{1\leq j\leq J}$, we have $\lambda_k<\lambda_j$ and $\frac{\lambda_j}{\lambda_k}\leq\gamma^{-(k-j)}$. Thus
\begin{align*}
\int_{\Bbb R^N}{W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}{W_a}_{(\lambda_k)}\ud x=&\Big(\frac{\lambda_j}{\lambda_k}\Big)^{\frac{N-2}{2}}\int_{\Bbb R^N}{W_a}^{\frac{N+2}{N-2}}(x){W_a}\Big(\frac{\lambda_j x}{\lambda_k}\Big)\ud x.
\end{align*}
According to \eqref{eq: the formula of Wa},
\begin{align*}
\bigg|W_a(x)-\frac{\big(N(N-2)\beta^2\big)^{\frac{N-2}{4}}}{|x|^{\frac{(\beta+1)(N-2)}{2}}}\bigg|= O\bigg(\frac{1}{|x|^{\frac{3}{2}\beta+(N-2)}}\bigg),
\end{align*}
which leads to
\begin{align}\label{the bdd of Wa N}
\int_{\Bbb R^N}{W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}{W_a}_{(\lambda_k)}\ud x=&\Big(\frac{\lambda_k}{\lambda_j}\Big)^{\frac{\beta(N-2)}{2}}\int_{\Bbb R^N}\frac{(N(N-2)\beta^2)^{\frac{N-2}{4}}}{|x|^{\frac{(\beta+1)(N-2)}{2}}}{W_a}^{\frac{N+2}{N-2}}(x)\ud x\nonumber\\
&+O\bigg(\Big(\frac{\lambda_k}{\lambda_j}\Big)^{\frac{3\beta+(N-2)}{2}}\int_{\Bbb R^N}\frac{1}{|x|^{\frac{3}{2}\beta+(N-2)}}{W_a}^{\frac{N+2}{N-2}}(x)\ud x\bigg).
\end{align}
Therefore, if $j<k-1$, we get the bound
\begin{align*}
\int_{\Bbb R^N}{W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}{W_a}_{(\lambda_k)}\ud x\lesssim\gamma^{(N-2)\beta}.
\end{align*}
Combining \eqref{the bdd of alpha and delta} and \eqref{the bdd of alpha and Wa} we obtain
\begin{align*}
\delta^2\lesssim\gamma^{\frac{(N-2)\beta}{2}}+\gamma^{\frac{(N-2)\beta-2}{2}}\delta^2+\delta^{\frac{2(N-1)}{N-2}}+o_{n}(1),
\end{align*}
which yields $\delta\lesssim \gamma^{\frac{(N-2)\beta}{4}}+o_{n}(1)$. Hence
\begin{align*}
\bigg|\frac{1}{2}\delta^2-\big(N(N-2)\beta^2\big)^{\frac{N-2}{2}}\int_{\Bbb R^N}\frac{1}{|x|^{\frac{(\beta+1)(N-2)}{2}}}{W_a}^{\frac{N+2}{N-2}}(x)\ud x& \sum\limits_{1\leq j\leq J+1}\iota_j\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}\bigg|\\
&\lesssim\gamma^{\frac{(N-2)\beta-2}{2}}+o_{n}(1).
\end{align*}
Denote
\begin{align*}
\kappa'_{1}=\big(N(N-2)\beta^2\big)^{\frac{N-2}{2}}\int_{\Bbb R^N}\frac{1}{|x|^{\frac{(\beta+1)(N-2)}{2}}}{W_a}^{\frac{N+2}{N-2}}(x)\ud x,
\end{align*}
then we obtain \eqref{the bdd of delta and kappa1}.
\end{proof}
We set $U(t)=u(t)-v_{L}(t)$, and
\begin{align*}
h(t)=U(t)-M_a(t)=U(t)-\sum\limits_{j=1}^{J}\iota_{j}{W_{a}}_{(\lambda_{j}(t))}.
\end{align*}
If we denote the nonlinear term $F(u)=|u|^{\frac{4}{N-2}}u$, then $\left(h(t), \partial_{t} U(t)\right)$ satisfies the following system for $t\in\left[\tilde{t}_{n}, t_{n}\right]$,
\begin{align}\label{eq: function of h}
\left\{\begin{array}{l}
\frac{\partial h}{\partial t}=\frac{\partial U}{\partial t}+\sum\limits_{j=1}^{J}\iota_{j}\lambda_{j}^{\prime}(t)(\Lambda W_{a})_{[\lambda_{j}(t)]}\\
\frac{\partial}{\partial t}\left(\frac{\partial U}{\partial t}\right)-\Delta h+\frac{a}{|x|^2}h=F(U)-\sum\limits_{j=1}^{J} F\left(\iota_{j} {W_{a}}_{(\lambda_{j})}\right)+N\left(h, v_{L}\right),
\end{array}\right.
\end{align}
where
\begin{align*}
N\left(h, v_{L}\right)\triangleq F\left(M+h+v_{L}\right)-F(M+h),
\end{align*}
satisfies
\begin{align}\label{eq: estimate of sigma}
\big|N\left(h, v_{L}\right)\big| \lesssim\big|v_{L}(t)\big|^{\frac{N+2}{N-2}}+\sum_{j=1}^{J}\Big({W_{a}}_{\left(\lambda_{j}\right)}^{\frac{4}{N-2}}+|h(t)|^{\frac{4}{N-2}}\Big)\big|v_{L}(t)\big|.
\end{align}
Employing the orthogonality condition, we obtain the estimate on the first order derivative of $\lambda_j$.
\begin{lemma}\label{lem: lambeda first deravitive estimate}
\begin{align}\label{the bdd of alpha and gamma}
\left|\lambda_{j}^{\prime}+\alpha_{j}\right|\lesssim\gamma^{\frac{(N-1)\beta}{4}}+o_{n}(1).
\end{align}
where $o_{n}(1)$ represents a term goes to 0 as $n \rightarrow \infty$, uniformly with respect to $t \in\left[\tilde{t}_{n}, t_{n}\right]$.
\end{lemma
It is simple to show by the orthogonality condition \eqref{eq: orthogonal condition 1} and \eqref{2.13}.
\medskip
\begin{lemma}.\label{lem: lambda second deravitive estimate}
For $(N-2)\beta>2$, for all $j \in \llbracket1, J\rrbracket$,
\begin{align}\label{eq: lambda second deravitive estimate}
\bigg|\lambda_{j} \beta_{j}^{\prime}+\kappa_{0}\Big(\iota_{j} \iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}-\iota_{j} \iota_{j-1}\Big(\frac{\lambda_{j}}{\lambda_{j-1}}\Big)^{\frac{(N-2)\beta}{2}}\Big)\bigg| \lesssim \gamma^{\frac{N\beta}{2}}+o_{n}(1),
\end{align}
where
\begin{align*}
\kappa_{0}=\frac{(N(N-2))^{\frac{N-2}{4}}\beta^{\frac{N}{2}}}{2}\int\frac{1}{|x|^{\frac{(\beta+1)(N-2)}{2}}}W_{a}^{\frac{N+2}{N-2}}(x)\ud x.
\end{align*}
\end{lemma}
\begin{proof}
According to \eqref{def: beta j t}, $\beta_{j}(t)=-\iota_{j}\int(\Lambda W_{a})_{[\lambda_{j}(t)]}\partial_{t}U(t)\ud x$. Differentiating in $t$ gives
\begin{align}
\lambda_{j}\beta^{\prime}_{j}(t)=&-\iota_{j}\int\lambda_{j}^{-\frac{N}{2}}\lambda^{\prime}_{j}\Big(-\frac{N}{2}+\frac{x}{\lambda_{j}}\cdot\nabla\Big)(\Lambda W_{a})(\frac{x}{\lambda_{j}})\partial_{t}U(t)\ud t\nonumber\\
&-\iota_{j}\lambda_{j}\int(\Lambda W_{a})_{[\lambda_{j}]}\Big(\Delta h-\frac{a}{|x|^2}h+F(h)-\sum\limits_{j=1}^{J}F(\iota {W_{a}}_{\lambda_{j}})+\sigma(h,v_{L})\Big)\ud x\nonumber\\
=&-\iota_{j}\int\lambda^{\prime}_{j}(\Lambda_{0}\Lambda W_{a})_{[\lambda_{j}]}\partial_{t}U(t)\ud t\label{first term of sum}\\
&-\iota_{j}\int(\Lambda W_{a})_{(\lambda_{j})}\Big(\Delta h-\frac{a}{|x|^2}h+F(h)-\sum\limits_{j=1}^{J}F(\iota {W_{a}}_{\lambda_{j}})+\sigma(h,v_{L})\Big)\ud x\label{second term of sum}.
\end{align}
Our first goal is to show the bound of the first term \eqref{first term of sum} of this summation. Using the expansion \eqref{the expansion of partial t U} of $\partial_{t}U$, we obtain
\begin{align*}
\int(\Lambda_0\Lambda W_a)_{[\lambda_j]}\partial_{t}U\ud x=&\int(\Lambda_0\Lambda W_a)_{[\lambda_j]}g_1\ud x+\iota_j\alpha_j\int(\Lambda_0\Lambda W_a)_{[\lambda_j]}(\Lambda W_a)_{[\lambda_j]}\ud x\\
&+\sum\limits_{k\neq j}\iota_k\alpha_k\int(\Lambda_0\Lambda W_a)_{[\lambda_j]}(\Lambda W_a)_{[\lambda_k]}\ud x.
\end{align*}
Hence, \eqref{the estimate of h,g} gives
\begin{align*}
\int(\Lambda_0\Lambda W_a)_{[\lambda_j]}g_1\ud x\lesssim\big\|\Lambda_0\Lambda W_a\big\|_{L^2}\big\|g_1\big\|_{L^2}\lesssim\gamma^{\frac{N\beta}{4}}+o_{n}(1).
\end{align*}
Integration by parts shows that the second term equals $0$. Using \eqref{the bdd of Lambda Wa and Lambda Wa} we get
\begin{align*}
\int(\Lambda_0\Lambda W_a)_{[\lambda_j]}(\Lambda W_a)_{[\lambda_k]}\ud x\lesssim\gamma^{\frac{(N-2)\beta}{2}-1}.
\end{align*}
By the estimate of $\lambda'$ in Lemma \ref{lem: lambeda first deravitive estimate} and the assumption $(N-2)\beta\geq 1$, we obtain
\begin{align}\label{eq: lambda beta' first order}
\left|\int\lambda^{\prime}_{j}(\Lambda_{0}\Lambda W_{a})_{[\lambda_{j}]}\partial_{t}U(t)\ud t\right|\lesssim\gamma^{\frac{(3N-5)\beta}{4}-1}+o_{n}(1)\leq\gamma^{\frac{(N-2)\beta}{2}-1}+o_{n}(1).
\end{align}
It remains to estimate \eqref{second term of sum}. Plugging \eqref{eq: function of h} into \eqref{second term of sum} gives
\begin{align}
&\lambda_{j}\int(\Lambda W_{a})_{[\lambda_{j}]}\partial_{t}^2U\ud x\nonumber\\\label{eq:0}
=&-\int(\Lambda W_{a})_{[\lambda_{j}]}\Big(-\Delta+\frac{a}{|x|^2}-\frac{N+2}{N-2}{W_{a}}_{(\lambda_{j})}^{\frac{4}{N-2}}\Big)h\ud x\\\label{eq:1}
&+\int(\Lambda W_{a})_{(\lambda_{j})}\sigma(h,v_{L})\ud x\\\label{eq:2}
&+\int(\Lambda W_{a})_{(\lambda_{j})}\Big(F(\iota_{j}{W_{a}}_{(\lambda_{j})}+h)-F(\iota_{j}{W_{a}}_{(\lambda_{j})})-\frac{N+2}{N-2}{W_{a}}_{(\lambda_{j})}^{\frac{4}{N-2}}h\Big)\ud x\\\label{eq:3}
&+\int(\Lambda W_{a})_{(\lambda_{j})}\Big(F(M+h)-F(M)+F(\iota_j {W_{a}}_{(\lambda_{j})})-F(\iota_j {W_{a}}_{(\lambda_{j})}+h)\Big)\ud x\\\label{eq:4}
&+\int(\Lambda W_{a})_{(\lambda_{j})}\Big(F(M)-\sum\limits_{j=1}^{J}F(\iota_{k}{W_{a}}_{(\lambda_{k})})\Big)
\ud x.
\end{align}
Notice that the operator $\mathcal{L}_{W(\lambda_{j})}^{a}=-\Delta+\frac{a}{|x|^2}-\frac{N+2}{N-2}{W_{a}}_{(\lambda_{j})}^{\frac{4}{N-2}}$ is self-adjoint. It follows that
\begin{align*}
\eqref{eq:0}=\int(\Lambda W_{a})_{[\lambda_{j}]}\mathcal{L}_{W(\lambda_{j})}^{a}h\ud x=\int\mathcal{L}_{W(\lambda_{j})}^{a}(\Lambda W_{a})_{[\lambda_{j}]}h\ud x=0.
\end{align*}
From \eqref{eq: estimate of sigma} and H\"{o}lder's inequality, we have
\begin{align*}
|\eqref{eq:1}|\lesssim&\int(\Lambda W_{a})_{(\lambda_{j})}|v_{L}|^{\frac{N+2}{N-2}}\ud x+\sum\limits_{k=1}^{J}\int(\Lambda W_{a})_{(\lambda_{j})}{W_{a}}_{(\lambda_{k})}^{\frac{4}{N-2}}|v_{L}(t)|\ud x\\
&+J\int(\Lambda W_{a})_{(\lambda_{j})}|h(t)|^{\frac{4}{N-2}}|v_{L}(t)|\ud x\\
\lesssim&\|v_{L}\|_{L^{\frac{2N}{N-2}}}^{\frac{N+2}{N-2}}+\|v_{L}\|_{L^{\frac{2N}{N-2}}}.
\end{align*}
Since $v_{L}$ is a solution to \eqref{lwp}, we have $\lim\limits_{t\rightarrow\infty}\|v_{L}(t)\|_{L^{\frac{2N}{N-2}}}=0$, which implies that $|\eqref{eq:1}|=o_n(1)$.
According to the definition of nonlinear term $F(h)$, \eqref{eq: 2 nonlinear estimate} and
\begin{align}\label{eq: estimate of Lambda Wa}
\big|(\Lambda W_{a})_{(\lambda_{j})}\big|=\lambda_{j}^{-\frac{N-2}{2}}\bigg|\Big(\frac{N-2}{2}W_{a}+x\cdot\nabla W_{a}\Big)(\frac{x}{\lambda_{j}})\bigg|\lesssim\big|{W_{a}}_{(\lambda_{j})}\big|,
\end{align}
then by \eqref{eq: 2 nonlinear estimate} and Sobolev's inequality, we obtain
\begin{align}\label{the bdd of 2}
\eqref{eq:2}\lesssim&\int\big|{W_{a}}_{(\lambda_{j})}\big|h^2\chi_{\{|h|\leq{W_{a}}_{(\lambda_{j})}\}}{W_{a}}_{(\lambda_{j})}^{\frac{6-N}{N-2}}\ud x+\int\big|{W_{a}}_{(\lambda_{j})}\big|h^{\frac{N+2}{N-2}}\chi_{\{|h|>{W_{a}}_{(\lambda_{j})}\}}\ud x\nonumber\\
\lesssim&\int h^2{W_{a}}_{(\lambda_{j})}^{\frac{4}{N-2}}(x)\ud x+\int\big|h\big|^{\frac{2N}{N-2}}\ud x\nonumber\\
\lesssim&\gamma^{\frac{N\beta}{2}}+o_{n}(1).
\end{align}
Recalling \eqref{eq: estimate of F2}, the estimates of $\eqref{eq:3}$ are divided into two cases.
\begin{itemize}
\item If $N\geq6$.
We choose $a=\iota_{j}{W_{a}}_{(\lambda_j)}$, $b=\sum\limits_{k\neq j}\iota_{k}{W_{a}}_{(\lambda_{k})}$ and $c=h$ in \eqref{eq: estimate of F2}. From \eqref{eq: estimate of Lambda Wa} and H\"{o}lder's inequality, we obtain the estimate
\begin{align*}
&\bigg|\int(\Lambda W_{a})_{(\lambda_{j})}\Big[F(M+h)-F(M)+F(\iota_j {W_{a}}_{(\lambda_{j})})-F(\iota_j {W_{a}}_{(\lambda_{j})}+h)\Big]\ud x\bigg|\\
\lesssim&\sum\limits_{j\neq k}\int {W_{a}}_{(\lambda_{j})}^{\frac{N+2}{2(N-2)}}|h|{W_{a}}_{(\lambda_{k})}^{\frac{N+2}{2(N-2)}}\ud x\lesssim\|h\|_{L^{\frac{2N}{N-2}}}\sum\limits_{k\neq j}\Big(\int {W_{a}}_{(\lambda_{k})}^{\frac{N}{N-2}}{W_{a}}_{(\lambda_{j})}^{\frac{N}{N-2}}\Big)^{\frac{N+2}{2N}}\\
\lesssim&\gamma^{\frac{(N+1)\beta}{2}}+o_{n}(1).
\end{align*}
\item If $N\leq 5$, taking $a=\iota_{j}{W_{a}}_{(\lambda_j)}$, $b=\sum\limits_{k\neq j}\iota_{k}{W_{a}}_{(\lambda_{k})}$ and $c=h$, in \eqref{eq: estimate of F2} and by \eqref{2.12}, we have
\begin{align*}
&\bigg|\int(\Lambda W_{a})_{(\lambda_{j})}\Big[F(M+h)-F(M)+F(\iota_j {W_{a}}_{(\lambda_{j})})-F(\iota_j {W_{a}}_{(\lambda_{j})}+h)\Big]\ud x\bigg|\\
\lesssim&\bigg|\int(\Lambda W_{a})_{(\lambda_j)}\Big[\sum\limits_{k\neq j}{W_{a}}_{(\lambda_{k})}|h|\big({W_{a}}_{(\lambda_{j})}+\sum\limits_{k\neq j}{W_{a}}_{(\lambda_{k})}+|h|\big)^{\frac{6-N}{N-2}}\Big]\ud x\bigg|
\end{align*}
\begin{align*}
\lesssim&\int {W_{a}}_{(\lambda_{j})}\bigg[\sum\limits_{k\neq j}{W_{a}}_{(\lambda_{k})}|h|\big({W_{a}}_{(\lambda_{j})}^{\frac{6-N}{N-2}}+\sum\limits_{k\neq j}{W_{a}}_{(\lambda_{k})}^{\frac{6-N}{N-2}}+|h|^{\frac{6-N}{N-2}}\big)\bigg]\ud x\\
\lesssim&\sum\limits_{k\neq j}\Big[\int{W_{a}}_{(\lambda_{j})}^{\frac{4}{N-2}}|h|{W_{a}}_{(\lambda_{k})}+\int{W_{a}}_{(\lambda_j)}{W_a}_{(\lambda_k)}^{\frac{4}{N-2}}|h|+\int{W_a}_{(\lambda_j)}{W_a}_{(\lambda_k)}|h|^{\frac{4}{N-2}}\Big]\\
\lesssim&\big\|h\big\|_{L^{\frac{2N}{N-2}}}\sum\limits_{1\leq k,l\leq J,k\neq l}\big\|{W_a}_{(\lambda_{\ell})}{W_a}_{(\lambda_k)}^{\frac{4}{N-2}}\big\|_{L^{\frac{2N}{N+2}}}+\big\|h\big\|_{\frac{2N}{N-2}}^{\frac{4}{N-2}}\sum\limits_{1\leq k,l\leq J,k\neq l}\Big(\int{W_a}_{(\lambda_k)}^{\frac{N}{N-2}}{W_a}_{(\lambda_{\ell})}^{\frac{N}{N-2}}\ud x\Big)^{\frac{N-2}{N}}\\
\leq&\gamma^{\frac{(3N-4)\beta}{4}}+\gamma^{\frac{N^2\beta}{2(N-2)}-\beta}+o_{n}(1)\lesssim\gamma^{\frac{(3N-4)\beta}{4}}+o_n(1).
\end{align*}
\end{itemize}
Since $\gamma\ll1$, collecting the above estimates, we get
\begin{align}\label{the bound of 3}
|\eqref{eq:3}|\leq\max\Big\{\gamma^{\frac{(N+1)\beta}{2}},\gamma^{\frac{(3N-4)\beta}{4}}\Big\}+o_{n}(1).
\end{align}
As a consequence of \eqref{the bdd of 2} and \eqref{the bound of 3}, for $(N-2)\beta>2$, we have
\begin{align}\label{the bdd of the easy term}
|\eqref{eq:0}|+|\eqref{eq:1}|+|\eqref{eq:2}|+|\eqref{eq:3}|\lesssim\gamma^{\frac{N\beta}{2}}+o_{n}(1).
\end{align}
We finally return to the leading term,
\begin{align*}
\eqref{eq:4}
=&\int(\Lambda W_{a})_{(\lambda_{j})}\Big(F\big(\sum\limits_{k\neq j}\iota_{k}{W_{a}}_{(\lambda_k)}+\iota_{j}{W_{a}}_{(\lambda_j)}\big)-F(\iota_{j}{W_{a}}_{(\lambda_j)})\Big)\ud x\\
=&\int(\Lambda W_{a})_{(\lambda_{j})}\Big(F\big(\sum\limits_{k\neq j}\iota_{k}{W_{a}}_{(\lambda_k)}+\iota_{j}{W_{a}}_{(\lambda_j)}\big)-F(\iota_{j}{W_{a}}_{(\lambda_j)})-F'(\iota_{j}{W_{a}}_{(\lambda_j)})\sum\limits_{k\neq j}\iota_{k}{W_{a}}_{(\lambda_k)}\Big)\ud x\\
&+\int(\Lambda W_{a})_{(\lambda_j)}F'(\iota_j{W_a}_{(\lambda_j)})\sum\limits_{k\neq j}\iota_k{W_a}_{(\lambda_k)}\ud x.
\end{align*}
By \eqref{eq: estimate of Lambda Wa}, letting $y=x/\lambda_{j}$, we have
\begin{align*}
&\int(\Lambda W_{a})_{(\lambda_j)}F'(\iota_j{W_a}_{(\lambda_j)})\sum\limits_{k\neq j}\iota_k{W_a}_{(\lambda_k)}\ud x\lesssim\sum\limits_{k\neq j}\int(\Lambda W_a)_{(\lambda_j)}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}{W_{a}}_{(\lambda_k)}\ud x\\
=&\sum\limits_{k\neq j}\Big(\frac{\lambda_j}{\lambda_k}\Big)^{\frac{N-2}{2}}\int{W_a}^{\frac{N+2}{N-2}}(y)W_{a}\Big(\frac{\lambda_j}{\lambda_k}y\Big)\ud y\lesssim\sum\limits_{k\neq j}\Big(\frac{\lambda_k}{\lambda_j}\Big)^{\frac{(N-2)\beta}{2}}\int y^{-\frac{(\beta+1)(N-2)}{2}}W_a^{\frac{N+2}{N-2}}(y)\ud y.
\end{align*}
Therefore, it suffices to consider the case that $k=j\pm1$. To conclude the proof, it remains to prove
\begin{align}\label{eq: the second derivative of U}
&\bigg|\int_{\mathbb{R}^N}\Big(F(M)-\sum\limits_{k=1}^{J}F\big(\iota_k {W_a}_{(\lambda_k)}\big)\Big)(\Lambda{W_a})_{(\lambda_j)}\ud x\nonumber\\
&-\kappa_0\Big(\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_j}\Big)^{\frac{N-2}{2}}-\iota_{j-1}\Big(\frac{\lambda_j}{\lambda_{j-1}}\Big)^{\frac{N-2}{2}}\Big)\bigg|\lesssim\gamma^{\frac{N\beta}{2}}.
\end{align}
In order to prove this claim, it is equivalent to prove the following inequalities:
\begin{itemize}
\item For $1\leq j\leq J$,
\begin{align}\label{eq: the 1 bdd}
&\qquad\bigg|\int_{\mathbb{R}^{N}}(\Lambda{W_a})_{(\lambda_j)}\Big(F(M)-\sum\limits_{k=1}^{J}F(\iota_k{W_a}_{(\lambda_k)})\Big)\ud x\nonumber\\
-&\frac{N+2}{N-2}\int_{\mathbb{R}^N}(\Lambda{W_a})_{(\lambda_j)}\Big(\iota_{j+1}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}{W_a}_{(\lambda_{j+1})}+\iota_{j-1}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}{W_a}_{(\lambda_{j-1})}\Big)\ud x\bigg|\lesssim\gamma^{\frac{N\beta}{2}}.
\end{align}
\item For $1\leq j\leq J-1$,
\begin{align}\label{eq: the 2 bdd}
&\qquad\bigg|\int_{\mathbb{R}^{N}}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}{W_a}_{(\lambda_{j+1})}(\Lambda W_a)_{(\lambda_j)}\nonumber\\
-&\Big(\frac{\lambda_{j+1}}{\lambda_j}\Big)^{\frac{N-2}{2}}\frac{N^{\frac{N-2}{4}}(N-2)^{\frac{N+6}{4}}\beta^{\frac{N}{2}}}{2(N+2)}\int|x|^{-\frac{(\beta+1)(N-2)}{2}}W_{a}^{\frac{N+2}{N-2}}(x)\ud x\bigg|\lesssim\gamma^{\frac{N\beta}{2}}.
\end{align}
\item For $2\leq j\leq J$,
\begin{align}\label{eq: the 3 bdd}
&\qquad\bigg|\int_{\mathbb{R}^{N}}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}{W_a}_{(\lambda_{j-1})}(\Lambda W_a)_{(\lambda_j)}\nonumber\\
+&\Big(\frac{\lambda_{j}}{\lambda_{j-1}}\Big)^{\frac{N-2}{2}}\frac{N^{\frac{N-2}{4}}(N-2)^{\frac{N+6}{4}}\beta^{\frac{N}{2}}}{2(N+2)}\int|x|^{-\frac{(\beta+1)(N-2)}{2}}W_{a}^{\frac{N+2}{N-2}}(x)\ud x\bigg|\lesssim\gamma^{\frac{N\beta}{2}}.
\end{align}
\end{itemize}
Because $W_{a}(x)\sim|x|^{\frac{(\beta-1)(N-2)}{2}}(1+|x|^{2\beta})^{-\frac{N-2}{2}}$, ${W_{a}}_{(\lambda_{j})}<{W_{a}}_{(\lambda_{k})}$ is equivalent to
\begin{align*}
&\lambda_{j}^{-\frac{(N-2)\beta}{2}}|x|^{\frac{(\beta-1)(N-2)}{2}}\Big(\lambda_{j}^{\frac{\beta}{2}}+|x|^{\beta}\lambda_{j}^{-\frac{\beta}{2}}\Big)^{-(N-2)}\lambda_{j}^{\frac{(N-2)\beta}{2}}\\
\lesssim&\lambda_{k}^{-\frac{(N-2)\beta}{2}}|x|^{\frac{(\beta-1)(N-2)}{2}}\Big(\lambda_{k}^{\frac{\beta}{2}}+|x|^{\beta}\lambda_{k}^{-\frac{\beta}{2}}\Big)^{-(N-2)}\lambda_{k}^{\frac{(N-2)\beta}{2}}.
\end{align*}
The fact $(\beta-1)(N-2)>0$ means that
\begin{align*}
|x|^{\beta}\Big(\lambda_{k}^{-\frac{\beta}{2}}-\lambda_{j}^{-\frac{\beta}{2}}\Big)\lesssim\lambda_j^{\frac{\beta}{2}}-\lambda_{k}^{\frac{\beta}{2}},
\end{align*}
which can be rewritten as
\begin{align}\label{relationship of lambda k and lambda j}
\left\{\begin{array}{l}
|x|^{\beta}<(\lambda_j\lambda_k)^{\frac{\beta}{2}},\qquad if \quad\lambda_j>\lambda_k,\\
|x|^{\beta}>(\lambda_j\lambda_k)^{\frac{\beta}{2}},\qquad if \quad\lambda_j<\lambda_k.
\end{array}\right.
\end{align}
Then for any $k,j\in\llbracket1, J\rrbracket$, if $|x|\in\Big(\lambda_{k}^{\frac{1}{2}}\lambda_{k+1}^{\frac{1}{2}},\lambda_{k}^{\frac{1}{2}}\lambda_{k-1}^{\frac{1}{2}}\Big)$, there exists a constant $C$ such that ${W_a}_{(\lambda_j)}\lesssim{W_{a}}_{(\lambda_{k})}$.
{\bf Proof of \eqref{eq: the 1 bdd}}. Substituting \eqref{eq: the 1 bdd} into \eqref{relationship of lambda k and lambda j}, we obtain
\begin{align*}
&\int_{\mathbb{R}^{N}}(\Lambda{W_a})_{(\lambda_j)}\Big(F(M_a)-\sum\limits_{k=1}^{J}F(\iota_k{W_a}_{(\lambda_k)})\Big)\ud x\\
&-\frac{N+2}{N-2}\int_{\mathbb{R}^N}(\Lambda{W_a})_{(\lambda_j)}\Big(\iota_{j+1}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}{W_a}_{(\lambda_{j+1})}+\iota_{j-1}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}{W_a}_{(\lambda_{j-1})}\Big)\ud x\\
\triangleq&\int_{\mathbb{R}^{N}}P_{j}(x)\ud x=\sum\limits_{k=1}^{J}\int_{\sqrt{\lambda_{k}\lambda_{k+1}}\leq|x|\leq\sqrt{\lambda_k\lambda_{k-1}}}P_{j}(x)\ud x.
\end{align*}
If $j\neq k$, we have
\begin{align*}
\Big|F(M)-F\big(\iota_k{W_a}_{(\lambda_k)}\big)\Big|=\Big|F\Big(\sum\limits_{\ell=1}^{J}\iota_{\ell}{W_a}_{(\lambda_{\ell})}\Big)-F\big(\iota_k{W_a}_{(\lambda_k)}\big)\Big|
\lesssim\sum\limits_{\ell\neq k}{W_a}_{(\lambda_{\ell})}{W_a}_{(\lambda_k)}^{\frac{4}{N-2}}.
\end{align*}
We define the domain $B_k=\big\{x:\sqrt{\lambda_{k}\lambda_{k+1}}\leq|x|\leq\sqrt{\lambda_k\lambda_{k-1}}\big\}$. Since $j\neq k$, \eqref{relationship of lambda k and lambda j} implies ${W_a}_{(\lambda_{\ell})}\leq {W_a}_{(\lambda_k)}$ on $B_k$. Thus,
\begin{align*}
&\int_{B_k}|P_{j}(x)|\ud x\nonumber\\
\leq&\int_{B_k}{W_a}_{(\lambda_j)}\Big|F(M)-F(\iota_k{W_a}_{(\lambda_k)})-\sum\limits_{k\neq \ell}F(\iota_{\ell}{W_a}_{(\lambda_{\ell})})\\
&-\frac{N+2}{N-2}\Big(\iota_{j+1}{W_a}_{(\lambda_{j+1})}+\iota_{j-1}{W_a}_{(\lambda_{j-1})}\Big){W_a}_{(\lambda_j)}^{\frac{4}{N-2}}\Big|\ud x\\
\lesssim&\sum\limits_{\ell\neq k}\int_{B}{W_a}_{(\lambda_j)}{W_a}_{(\lambda_k)}^{\frac{4}{N-2}}{W_a}_{(\lambda_{\ell})}{W_a}_{(\lambda_j)}\ud x+\sum\limits_{l\neq k}\int_{B}{W_a}_{(\lambda_j)}{W_a}_{(\lambda_{\ell})}^{\frac{N+2}{N-2}}{W_a}_{(\lambda_j)}\ud x\\
&+\int_{B_k}{W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}\big({W_a}_{(\lambda_{j+1})}+{W_a}_{(\lambda_{j-1})}\big)\ud x\\
\lesssim&\sum\limits_{\ell\neq k}\int_{B_k}{W_a}_{(\lambda_j)}{W_a}_{(\lambda_{\ell})}^{\frac{N}{N-2}}{W_a}_{(\lambda_k)}^{\frac{N}{N-2}}\ud x+\int_{B_k}{W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}\big({W_a}_{(\lambda_{j+1})}+{W_a}_{(\lambda_{j-1})}\big)\ud x.
\end{align*}
By \eqref{eq: the formula of Wa}, we have
\begin{align*
&\int{W_a}_{(\lambda)}^{\frac{N}{N-2}}{W_a}_{(\mu)}^{\frac{N}{N-2}}\ud x\nonumber\\
=&(\lambda\mu)^{-\frac{N}{2}}\int_{0}^{\infty}\min\Big\{\Big(\frac{r}{\lambda}\Big)^{\frac{N(\beta-1)}{2}},\Big(\frac{\lambda}{r}\Big)^{\frac{N(\beta+1)}{2}}\Big\}\min\Big\{\Big(\frac{r}{\mu}\Big)^{\frac{N(\beta-1)}{2}},\Big(\frac{\mu}{r}\Big)^{\frac{N(\beta+1)}{2}}\Big\}r^{N-1}\ud r\nonumber\\
\lesssim&\Big(\frac{\lambda}{\mu}\Big)^{\frac{N\beta}{2}}.
\end{align*}
Thus
\begin{align}\label{the first estimate of 5.38}
\int_{B_k}\big|P_j(x)\big|\ud x\lesssim \gamma^{\frac{N\beta}{2}}.
\end{align}
If $k=j$, the domain of integration is $B_j=\big\{x:\sqrt{\lambda_{j}\lambda_{j+1}}\leq|x|\leq\sqrt{\lambda_j\lambda_{j-1}}\big\}$. By \eqref{eq: 2 nonlinear estimate}, writing
\begin{align*}
E_j\triangleq\Big\{x\in\mathbb{R}^N:\Big|\sum\limits_{\ell\neq j}\iota_{\ell}{W_a}_{(\lambda_{\ell})}(x)\Big|\leq{W_a}_{(\lambda_j)}(x)\Big\},
\end{align*}
then on the domain $B_j$, we choose $a=\iota_j{W_a}_{(\lambda_j)}$, $b=\sum\limits_{\ell\neq j}\iota_{\ell}{W_a}_{(\lambda_{\ell})}(x)$, and get
\begin{align*}
&\bigg|F(M)-F(\iota_j{W_a}_{(\lambda_j)})-\frac{N+2}{N-2}{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}\sum\limits_{\ell\neq j}\iota_{\ell}{W_a}_{(\lambda_{\ell})}\bigg|\\
\lesssim&\chi_{E_j}{W_a^2}_{(\lambda_{\ell})}{W_a}_{(\lambda_j)}^{\frac{6-N}{N-2}}+\chi_{\mathbb{R}^{N}\setminus E_j}\Big(\sum\limits_{\ell\neq j}\iota_j{W_a}_{(\lambda_{\ell})}\Big)^{\frac{N+2}{N-2}}\lesssim\sum\limits_{\ell\neq j}{W_a}_{(\lambda_{\ell})}^2{W_a}_{(\lambda_j)}^{\frac{6-N}{N-2}}.
\end{align*}
As a consequence of \eqref{eq: estimate of Lambda Wa}, it holds
\begin{align*}
\int_{B_j}|P_j(x)|\ud x\leq&\sum\limits_{\ell\neq j}\int_{B_j}{W_a}_{(\lambda_{\ell})}^2{W_a}_{(\lambda_j)}^{\frac{6-N}{N-2}}\ud x+\sum\limits_{\ell\neq\{j,j\pm1\}}\int_{B_j}{W_a}_{(\lambda_{\ell})}{W_a}_{(\lambda_j)}^{\frac{N+2}{N-2}}\ud x\\
\leq&\sum\limits_{\ell\neq j}\int_{B_j}{W_a}_{(\lambda_{\ell})}^{\frac{N}{N-2}}{W_a}_{(\lambda_j)}^{\frac{N}{N-2}}\ud x+\sum\limits_{\ell\neq\{j,j\pm 1\}}\int_{B_j}{W_a}_{(\lambda_{\ell})}\cdot\big(-\Delta+\frac{a}{|x|^2}\big){W_a}_{(\lambda_j)}\ud x.
\end{align*}
If $\ell\neq\{j,j\pm1\}$,\eqref{2.10} gives rise to
\begin{align*}
&\int_{B_j}{W_a}_{(\lambda_{\ell})}\cdot\big(-\Delta+\frac{a}{|x|^2}\big){W_a}_{(\lambda_j)}\ud x\\
=&\int_{B_j}\big(\partial_r+\frac{c}{r}\big){W_a}_{(\lambda_{\ell})}\cdot\big(\partial_r+\frac{c}{r}\big){W_a}_{(\lambda_j)}\ud x\\
\lesssim&\min\Big\{\Big(\frac{\lambda_j}{\lambda_{\ell}}\Big)^{\frac{(N-2)\beta}{2}},\Big(\frac{\lambda_{\ell}}{\lambda_j}\Big)^{\frac{(N-2)\beta}{2}}\Big\}\lesssim\gamma^{(N-2)\beta}.
\end{align*}
Collecting the above estimates, we get
\begin{align*}
\int_{\sqrt{\lambda_{j}\lambda_{j+1}}\leq|x|\leq\sqrt{\lambda_j\lambda_{j-1}}}|P_j(x)|\ud x\lesssim\gamma^{\frac{N\beta}{2}}.
\end{align*}
This estimate together with \eqref{the first estimate of 5.38} yields the desired estimate \eqref{eq: the 1 bdd}.
{\bf Proof of \eqref{eq: the 2 bdd}}.
Let $y=\frac{x}{\lambda_j}$, from the definition of $W_a$, we have
\begin{align*}
&\int(\Lambda W_a)_{(\lambda_j)}(x){W_a}_{(\lambda_j)}^{\frac{4}{N-2}}(x){W_a}_{(\lambda_{j+1})}(x)\ud x\\
=&\big(N(N-2)\beta^2\big)^{\frac{N-2}{4}}\Big(\frac{\lambda_j}{\lambda_{j+1}}\Big)^{\frac{(N-2)\beta}{2}}\int(\Lambda W_a)(y){W_a}^{\frac{4}{N-2}}(y)|y|^{-\frac{(\beta+1)(N-2)}{2}}\ud y+O(\gamma^{\frac{N\beta}{2}}).
\end{align*}
Noting that $\Lambda W_a(x)=(x\cdot\nabla+\frac{N-2}{2}){W_a}(x)$ and
\begin{align}\label{Delta 1}
&\Delta\left(|x|^{-\frac{(N-2)(\beta+1)}{2}+2}\right)\nonumber\\
=&\Big(2-\frac{(N-2)(\beta+1)}{2}\Big)\Big[-\frac{(N-2)(\beta+1)}{2}|x|^{-\frac{(N-2)(\beta+1)}{2}-1}-\frac{\sum x_i^2}{|x|}+N|x|^{-\frac{(N-2)(\beta+1)}{2}}\Big]\nonumber\\
=&\Big(2-\frac{(N-2)(\beta+1)}{2}\Big)\frac{(N+2)-\beta(N-2)}{2}|x|^{-\frac{(N-2)(\beta+1)}{2}}.
\end{align}
We easily show that by \eqref{Delta 1}
\begin{align*}
&\int{W_a}^{\frac{4}{N-2}}(x)x\cdot\nabla W_a(x)|x|^{-\frac{(N-2)(\beta+1)}{2}}\ud x\\
=&\frac{N-2}{(N+2)\big(2-\frac{(N-2)(\beta+1)}{2}\big)}\int\nabla\big({W_a}^{\frac{N+2}{N-2}}(x)\big)\nabla\Big(|x|^{2-\frac{(N-2)(\beta+1)}{2}}\Big)\ud x\\
=&-\frac{N-2}{(N+2)\big(2-\frac{(N-2)(\beta+1)}{2}\big)}\int\Delta\Big(|x|^{2-\frac{(N-2)(\beta+1)}{2}}\Big){W_a}^{\frac{N+2}{N-2}}\ud x\\
=&\Big(-\frac{N-2}{2}+\frac{(N-2)^2\beta}{2(N+2)}\Big)\int{W_a}^{\frac{N+2}{N-2}}(x)|x|^{-\frac{(N-2)(\beta+1)}{2}}\ud x.
\end{align*}
Then we obtain
\begin{align*}
\int(\Lambda W_a)(x){W_a}^{\frac{4}{N-2}}(x)|x|^{-\frac{(N-2)(\beta+1)}{2}}\ud x=\frac{(N-2)^2\beta}{2(N+2)}\int{W_a}^{\frac{N+2}{N-2}}(x)|x|^{-\frac{(N-2)(\beta+1)}{2}}\ud x
\end{align*}
and \eqref{eq: the 2 bdd} follows.
{\bf Proof of \eqref{eq: the 3 bdd}}. For $2\leq j\leq J$, it is easy to see that
\begin{align*}
W_a\Big(\frac{\lambda_j}{\lambda_{j-1}}x\Big)\sim\big(N(N-2)\beta^2\big)^{\frac{N-2}{4}}\Big(\frac{\lambda_j}{\lambda_{j-1}}x\Big)^{\frac{(\beta-1)(N-2)}{2}}.
\end{align*}
Therefore,
\begin{align*}
&\int{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}(x)(\Lambda W_a)_{(\lambda_j)}(x){W_a}_{(\lambda_{j-1})}(x)\ud x\\
=&\Big(\frac{\lambda_j}{\lambda_{j-1}}\Big)^{\frac{N-2}{2}}\int W_a^{\frac{4}{N-2}}\Big(\frac{x}{\lambda_j}\Big)(\Lambda W_a)\Big(\frac{x}{\lambda_j}\Big)W_a\Big(\frac{x}{\lambda_{j-1}}\Big)\ud x\\
=&\Big(\frac{\lambda_j}{\lambda_{j-1}}\Big)^{\frac{(N-2)\beta}{2}}\left((N(N-2)\beta^2\right)^{\frac{N-2}{4}}\int{W_a}^{\frac{4}{N-2}}(x)(\Lambda W_a)(x)|x|^{\frac{(\beta-1)(N-2)}{2}}\ud x+O(\gamma^{\frac{N\beta}{2}}).
\end{align*}
A similar calculation to equation \eqref{Delta 1} yields
\begin{align*}
\int{W_a}^{\frac{4}{N-2}}(x)(\Lambda W_a)(x)|x|^{\frac{(\beta-1)(N-2)}{2}}\ud x=-\frac{(N-2)^2\beta}{2(N+2)}\int{W_a}^{\frac{N+2}{N-2}}(x)|x|^{\frac{(\beta-1)(N-2)}{2}}\ud x,
\end{align*}
we have
\begin{align}\label{before the translate}
&\int{W_a}_{(\lambda_j)}^{\frac{4}{N-2}}(x){W_a}_{(\lambda_{j-1})}(x)(\Lambda W_a)_{(\lambda_j)}(x)\ud x\nonumber\\
=&-\Big(\frac{\lambda_j}{\lambda_{j-1}}\Big)^{\frac{(N-2)\beta}{2}}\frac{N^{\frac{N-2}{4}}(N-2)^{\frac{N+6}{4}}\beta^{\frac{N}{2}}}{2(N+2)}\int{W_a}^{\frac{N+2}{N-2}}(x)|x|^{\frac{(\beta-1)(N-2)}{2}}\ud x+O(\gamma^{\frac{N\beta}{2}}).
\end{align}
Assume $\sigma=\frac{1}{r}$, then direct computation shows
\begin{align*}
W_a^{\frac{N+2}{N-2}}(r)=&\big(N(N-2)\beta^2\big)^{\frac{N+2}{4}}|x|^{\frac{(\beta-1)(N+2)}{2}}\big(1+|x|^{2\beta}\big)^{-\frac{N+2}{2}}\\
=&\big(N(N-2)\beta^2\big)^{\frac{N+2}{4}}\sigma^{\frac{(\beta-1)(N+2)}{2}}\big(1+\sigma^{-2\beta}\big)^{-\frac{N+2}{2}}\\
=&\big(N(N-2)\beta^2\big)^{\frac{N+2}{4}}\sigma^{\frac{(\beta-1)(N+2)}{2}}\big(1+\sigma^{2\beta}\big)^{-\frac{N+2}{2}}\sigma^{N+2}\\
=&W_a^{\frac{N+2}{N-2}}(\sigma)\sigma^{N+2},
\end{align*}
and \eqref{before the translate} can be written as
\begin{align*}
\int{W_a}^{\frac{N+2}{N-2}}(r)r^{\frac{(\beta-1)(N-2)}{2}}r^{N-1}\ud r=&\int W_a^{\frac{N+2}{N-2}}(\sigma)\sigma^{\frac{(\beta-1)(N-2)}{2}}\sigma\ud \sigma\\
=&\int W_a^{\frac{N+2}{N-2}}(\sigma)\sigma^{-\frac{(\beta+1)(N-2)}{2}}\sigma^{N-1}\ud \sigma.
\end{align*}
This completes the proof of \eqref{eq: the 3 bdd} and \eqref{eq: the second derivative of U} follows.
Combining \eqref{eq: lambda beta' first order}, \eqref{the bdd of the easy term} and the fact that $\frac{(N-1)\beta-2}{2}<\frac{N\beta}{2}$, $\gamma\ll 1$, we reach \eqref{eq: lambda second deravitive estimate}.
\end{proof}
\begin{proposition}\label{prop: conclusion}
For $(N-2)\beta>2$. For all large $n$ and for all $t \in\left[\tilde{t}_{n}, t_{n}\right]$,
\begin{align}\label{the conclusion of delta}
&\delta\lesssim\gamma^{\frac{(N-2)\beta}{4}}+o_{n}(1)
\end{align}
\begin{align}\label{the conclusion of beta lambda}
\Big|\beta_{j}-\|\Lambda W_{a}\|_{2}^{2}\lambda_{j}^{\prime}\Big|\lesssim\gamma^{\frac{(N-2)\beta-2}{4}}+o_{n}(1)
\end{align}
\begin{align}\label{the conclusion of beta 2}
\bigg|\frac{1}{2}\sum\limits_{j=1}^{J}\beta_{j}^{2}-\kappa_{1}\sum\limits_{1\leq j\leq J-1}\iota_{j}\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}\bigg|\lesssim\gamma^{\frac{(N-2)\beta-2}{2}}+o_{n}(1)
\end{align}
\begin{align}\label{kappa1}
\left|\lambda_{j}\beta^{\prime}_{j}+\kappa_{0}\Big(\iota_{j}\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}-\iota_{j}\iota_{j-1}\Big(\frac{\lambda_{j}}{\lambda_{j-1}}\Big)^{\frac{(N-2)\beta}{2}}\Big)\right|\lesssim\gamma^{\frac{N\beta}{2}}+o_{n}(1)
\end{align}
where $o_{n}(1)$ goes to 0 as $n \rightarrow \infty$, uniformly with respect to $n$, $t \in\left[\tilde{t}_{n}, t_{n}\right]$, and
\begin{align*}
\kappa_{0}=&\frac{N(N-2)^{\frac{N-2}{4}}\beta^{\frac{N}{2}}}{2}\int|x|^{-\frac{(\beta+1)(N-2)}{2}}W_{a}^{\frac{N+2}{N-2}}(x)\ud x,\\
\kappa_{1}=&\|\Lambda W_{a}\|_{L^{2}}^{2}(N(N-2)\beta^2)^{\frac{N-2}{4}}\int|x|^{-\frac{(\beta+1)(N-2)}{2}} W^{\frac{N+2}{N-2}}(x)\ud x.
\end{align*}
\end{proposition}
\begin{proof}
\eqref{the conclusion of delta} follows from \eqref{the estimate of delta}. Due to \eqref{the bdd of beta and alpha}, \eqref{the bdd of alpha and gamma} and \eqref{the estimate of delta} we have
\begin{align*}
\left|\beta_j-\lambda_{j}'\|\Lambda W_a\|_{L^2}^2\right|\lesssim&\left|\beta_j+\alpha_{j}\|\Lambda W_{a}\|_{L^2}^2\right|+\left|\alpha_{j}+\lambda_{j}'\right|\|\Lambda W_a\|_{L^2}^2\\
\lesssim&\gamma^{\frac{(N-2)\beta-2}{4}}+\delta^{\frac{N}{N-2}}+\gamma^{\frac{N\beta}{4}}\lesssim\gamma^{\frac{(N-2)\beta-2}{4}}.
\end{align*}
Combining \eqref{the bdd of alpha and delta}, \eqref{the bdd of beta and alpha} \eqref{the bdd of delta and kappa1} and the estimate $|\alpha_j|+|\beta_j|\lesssim\gamma^{\frac{(N-2)\beta-2}{4}}$, for \eqref{the conclusion of beta lambda} it follows that
\begin{align}\label{the bdd alpha2 beta 2}
&\left|\beta_j^2-\|\Lambda W_a\|_{L^2}^4\alpha_j^2\right|=\left|(\beta_j-\|\Lambda W_{a}\|_{L^2}^2\alpha_j)(\beta_j+\|\Lambda W_{a}\|_{L^2}^2\alpha_j)\right|\nonumber\\
\lesssim&\gamma^{\frac{(N-2)\beta-2}{4}}\gamma^{\frac{(N-2)\beta-2}{2}}+o_{n}(1)\lesssim\gamma^{\frac{3(N-2)\beta}{4}-1}+o_{n}(1).
\end{align}
Replacing $\beta$ with $\alpha$, according to \eqref{the bdd alpha2 beta 2}, \eqref{the bdd of alpha and delta} and \eqref{the bdd of delta and kappa1} we get \eqref{the conclusion of beta 2}. Furthermore,
\begin{align*}
&\bigg|\frac{1}{2}\sum\limits_{j=1}^{J}\beta_{j}^2-\kappa_{1}\sum\limits_{j=1}^{J-1}\iota_{j}\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_j}\Big)^{\frac{(N-2)\beta}{2}}\bigg|\\
=&\frac{1}{2}\bigg|\sum\limits_{j=1}^{J}\beta_{j}^2-\|\Lambda W_a\|_{L^2}^2\kappa_{1}'\sum\limits_{j=1}^{J-1}\iota_{j}\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_j}\Big)^{\frac{(N-2)\beta}{2}}\bigg|\\
\lesssim&\frac{1}{2}\bigg|\|\Lambda W_a\|_{L^2}^4\sum\limits_{j=1}^{J}\alpha_j^2-\sum\limits_{j=1}^{J}\beta_j^2\bigg|+\frac{1}{2}\bigg|\delta^2-\|\Lambda W_a\|_{L^2}^2\sum\limits_{j=1}^{J}\alpha_j^2\bigg|\|\Lambda W_a\|_{L^2}^2\\
&+\bigg|\frac{1}{2}\delta^2-\kappa_1'\sum\limits_{j=1}^{J}\iota_j\iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_j}\Big)^{\frac{(N-2)\beta}{2}}\bigg|\big\|\Lambda W_a\big\|_{L^2}^2\\
\lesssim&\gamma^{\frac{(N-2)\beta-2}{2}}+o_{n}(1),
\end{align*}
which completes the proof of \eqref{kappa1}.
\end{proof}
\section{Proof of Theorem \ref{main thm}}
In this section, we will eliminate the $o_{n}(1)$ terms in the estimates of the scaling parameters and choose $\tilde{J}$ as the unique index in $\llbracket1, J\rrbracket$ such that the exterior profiles are not equal to $\pm W_a$. We first prove the lower bounds of the exterior scaling parameter $\ell$. Second, using the conditions we got in Section \ref{section 7} to prove the key estimates about the coefficients in Proposition \ref{the main estimate without o1} and get the contradiction. Finally we conclude the proof of Theorem \ref{main thm}.
Recalling from Section \ref{section 7} the definitions of $t_{n}$, $\tilde{t}_{n}$, $J$, and for $j\in\llbracket1, J\rrbracket$, $\iota_{j}$, $\alpha_{j}(t)$, $\beta_{j}(t)$, $\lambda_{j}(t)$, after extraction of subsequences, the following weak limits exists in $\mathcal{H}$,
\begin{align*}
\left(\widetilde{U}_{0}^{j}, \widetilde{U}_{1}^{j}\right)=\underset{n \rightarrow \infty}{\mathrm{w}-\lim }\left(\lambda_{j}\left(\tilde{t}_{n}\right)^{\frac{N}{2}-1} U\left(\tilde{t}_{n}, \lambda_{j}\left(\tilde{t}_{n}\right) \cdot\right), \lambda_{j}\left(\tilde{t}_{n}\right)^{\frac{N}{2}} \partial_{t} U\left(\tilde{t}_{n}, \lambda_{j}\left(\tilde{t}_{n}\right) \cdot\right)\right)
\end{align*}
where $U=u-v_{L}$ as before. We note that there exists $j \in\llbracket1, J\rrbracket$ such that $\left(\tilde{U}_{0}^{j}, \tilde{U}_{1}^{j}\right)\neq\left(\iota_{j} W_a, 0\right)$.
If not, for all $k\in\llbracket1, K-1\rrbracket$, denote $\mu_{k,n}=\lambda_{j_k}(s_n)$, after extracting subsequence, by the limits \eqref{eq: initial date}, \eqref{eq: weak limit}, we obtain for $n\rightarrow\infty$
\begin{align}
\left(\mu^{\frac{N-2}{2}}_{k,n}(u-v_{L})(s_n,\mu_{k,n}\cdot),\mu^{\frac{N}{2}}_{k,n}\partial_{t}(u-v_L)(s_n,\mu_{k,n}\cdot)\right)\rightharpoonup\left(V_0^k,V_1^k\right).
\end{align}
By the assumption \eqref{distance d t},
\begin{align}
\bigg\|\left(V_0^k,V_1^k\right)-\sum\limits_{j=j_k}^{j_{k+1}-1}\iota_j({W_a}_{(\lambda_j)},0)\bigg\|_{\mathcal{H}_a}\lesssim\epsilon_0.
\end{align}
And for $j\in\llbracket j_k,j_{k+1}-2\rrbracket$, \eqref{def GJ} and \eqref{def: dJ, gamma} imply
\begin{align}
\frac{\nu_{j+1}}{\nu_j}\leq\lim\limits_{n\rightarrow\infty}\gamma(\lambda(s_n))\leq \epsilon_0,
\end{align}
which means $j_{k+1}=j_k+1$. According to the partition of the interval $\llbracket 1,J\rrbracket$ in \eqref{eq: not in a term}, there is only one element in each sub-interval such that for all $j\in \llbracket 1,J\rrbracket$, $
\lim\limits_{n\rightarrow\infty}\frac{\lambda_{j+1}(\tilde{t}_n)}{\lambda_j(\tilde{t}_n)}=0$, which implies $\lim\limits_{n\rightarrow\infty}\gamma(\tilde{t}_n)=0$. Hence, \eqref{the conclusion of delta} yields $\lim\limits_{n\rightarrow\infty}\delta(\tilde{t}_n)=0$, which is a contradiction with the definition of $\delta$ and $\tilde{t}_n$. We assume that $\bar{J}$ is the smallest index of $\llbracket 1,J\rrbracket$ such that $(\tilde{U}^{\bar{J}}_0,\tilde{U}^{\bar{J}}_1)$ is not the ground state, namely, $\bar{J}$ is the unique index such that
\begin{align*}
\left(\tilde{U}_{0}^{j}, \tilde{U}_{1}^{j}\right)=&\left(\iota_{j} W_a, 0\right), \quad\forall j \in\llbracket 1,\bar{J}-1\rrbracket,\\
\left(\tilde{U}_{0}^{\bar{J}}, \tilde{U}_{1}^{\bar{J}}\right) \neq&\left(\iota_{\bar{J}}{W_a}, 0\right),\quad\forall j \in\llbracket \bar{J},J\rrbracket.
\end{align*}
For simplify, we denote
\begin{align*
\lambda_{\bar{J}, n}\triangleq\lambda_{\bar{J}}\left(\tilde{t}_{n}\right), \quad \tilde{\gamma}(t)\triangleq\gamma\left(\left(\lambda_{j}(t)\right)_{\bar{J} \leq j \leq J}\right)=\max _{\bar{J} \leq j \leq J-1} \frac{\lambda_{j+1}(t)}{\lambda_{j}(t)} .
\end{align*}
We introduce the following lemma. Readers can refer to \cite{DKM7} for the proof.
\begin{lemma}\emph{(\cite{DKM7})}\label{lem: the limits of t, lambda}
\begin{align*
\lim _{n \rightarrow \infty} \frac{t_{n}-\tilde{t}_{n}}{\lambda_{\widetilde{J}, n}}=+\infty.
\end{align*}
\end{lemma}
\begin{lemma}\emph{(\cite{DKM7})}\label{the tilde gamma bdd}Let $T>0$ and
\begin{align*}
t_{n}^{\prime}=\tilde{t}_{n}+T \lambda_{\tilde{J}, n}.
\end{align*}
Then for $(N-2)\beta>2$, for large $n$, for all $t \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right]$ and for all $j\in\llbracket\tilde{J},J\rrbracket$,
\begin{align}\label{the tilde bdd of beta lambda}
\Big|\beta_{j}-\|\Lambda W_a\|_{L^{2}}^{2} \lambda_{j}^{\prime}\Big| \leq C\tilde{\gamma}^{\frac{(N-2)\beta-2}{4}},
\end{align}
\begin{align}\label{the tilde bdd of the second derivative}
\bigg|\frac{1}{2} \sum_{j=\tilde{J}}^{\text {J }} \beta_{j}^{2}-\kappa_{1} \sum_{\tilde{J} \leq j \leq J-1} \iota_{j} \iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}\bigg| \leq C \tilde{\gamma}^{\frac{(N-2)\beta-2}{2}},
\end{align}
and for all $t \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right]$,
\begin{align}\label{the tilde bdd of the beta2}
\bigg|\lambda_{j} \beta_{j}^{\prime}+\kappa_{0}\Big(\iota_{j} \iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}-\iota_{j} \iota_{j-1}\Big(\frac{\lambda_{j}}{\lambda_{j-1}}\Big)^{\frac{(N-2)\beta}{2}}\Big)\bigg| \leq C \tilde{\gamma}^{\frac{N\beta}{2}}.
\end{align}
\end{lemma}
\begin{proof}
Since we have assumed $t'_n=\tilde{t}_{n}+T\lambda_{\tilde{J},n}$, by Lemma \ref{lem: the limits of t, lambda} and restricting $t$ to $[\tilde{t}_{n},t'_{n}]$, we can replace $\gamma(t)+o_{n}(1)$ by $\tilde{\gamma}(t)$, the following claim can be proved by exploiting the method in \cite{DKM7}.
\begin{itemize}
\item
\begin{align}\label{the equation 1 of cliam}
\lim\limits_{n\rightarrow\infty}\max\limits_{\tilde{t}_{n}\leq t\leq t'_{n}}|\beta_j(t)|+\frac{\lambda_{j+1}(t)}{\lambda_{j}(t)}=0,
\end{align}
\item
\begin{align}\label{the equation 2 of claim}
\liminf\limits_{n\rightarrow\infty}\min\limits_{\tilde{t}_n\leq t\leq t'_n}\tilde{\gamma}(t)>0.
\end{align}
\end{itemize}
Thus we can restrict the indices $j$ to $\llbracket\tilde{J},J\rrbracket$. According to Proposition \ref{prop: conclusion}, the proof is complete.
\end{proof}
As a consequence of Proposition \ref{exterior profile estimate}, we prove a lower bound on one of the scaling parameters $\lambda_j$.
\begin{lemma}\label{lem: the lower bdd of lambda j}
Let $\ell$ be defined by Theorem \ref{thm: main thm of Xi} and $m_{0}=\frac{(N-2)\beta+1}{2}$ . Then if $\varepsilon_{0}=\varepsilon_{0}(T)$ is chosen small enough, there exists a constant $C>0$ such that for large $n$,
\begin{align}\label{the lower bdd of lambda j}
\forall t \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right], \quad|\ell| \leq C\bigg(\frac{\lambda_{\widetilde{J}}(t)}{\lambda_{\widetilde{J}}\left(\tilde{t}_{n}\right)}\bigg)^{(m_{0}-\frac{1}{2})/\beta} \delta(t)^{\frac{2N}{N-2}}.
\end{align}
\end{lemma}
\begin{proof}
We prove this inequality by contradiction. Assuming (after extracting a subsequence) that there exists a large constant $M$ and a sequence $\left\{\tilde{s}_{n}\right\}_{n}$ with $\tilde{s}_{n} \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right]$ and for all $n$
\begin{align}\label{the contradiction assumption}
|\ell| \geq M\bigg(\frac{\lambda_{\tilde{J}}\left(\tilde{s}_{n}\right)}{\lambda_{\tilde{J}}\left(\tilde{t}_{n}\right)}\bigg)^{(m_{0}-\frac{1}{2})/{\beta}} \delta\left(\tilde{s}_{n}\right)^{\frac{2N}{N-2}}.
\end{align}
As in the proof in \cite{DKM7} and by the non-radiative profile decomposition Lemma \ref{lem: nonradiative profile}, we can find a constant $\tilde{\lambda}\in[0,\infty\}$ such that
\begin{align}\label{def: limits of tilde lambda}
\lim\limits_{n\rightarrow\infty}\frac{\lambda_{\tilde{J}}(\tilde{t}_n)}{\lambda_{\tilde{J}}(\tilde{s}_n)}=\tilde{\lambda}.
\end{align}
Let $m_0=\frac{(N-2)\beta+1}{2}$. By Theorem \ref{thm: main thm of Xi}, for large $R$,
\begin{align*}
\left\|\left(\tilde{\lambda}^{\frac{N-2}{2}} V_{0}^{\widetilde{J}}(\tilde{\lambda} \cdot), \tilde{\lambda}^{\frac{N}{2}} V_{1}^{\tilde{J}}(\tilde{\lambda} \cdot)\right)-\ell \Xi_{p_{0}}\right\|_{\mathcal{H}_a(R)}
\lesssim \max \bigg\{\frac{1}{R^{(2m_{0}-\frac{N}{2})\frac{N+2}{N-2}}}, \frac{1}{R^\frac{N}{2}}\bigg\},
\end{align*}
and after scaling, for large $R$,
\begin{align}\label{the contradiction}
\left\|\left(V_{0}^{\tilde{J}}, V_{1}^{\tilde{J}}\right)-\tilde{\lambda}^{(m_{0}-\frac{1}{2})/\beta} \ell \Xi_{p_{0}}\right\|_{\mathcal{H}_a(R)} \lesssim \max\bigg\{\frac{1}{R^{(2m_{0}-\frac{N}{2})\frac{N+2}{N-2}}}, \frac{1}{R^\frac{N}{2}}\bigg\}.
\end{align}
Replacing $|\iota|$ by $\tilde{\lambda}^{(m_0-\frac{1}{2})/\beta}|\iota|$ in Proposition \eqref{exterior profile estimate} gives
\begin{align*}
\tilde{\lambda}^{(m_0-\frac{1}{2})/\beta}|\ell|\leq 2C \delta(\tilde{s}_n)^{\frac{2N}{N-2}}.
\end{align*}
On the one hand, using \eqref{def: limits of tilde lambda}, for large $n$, we obtain
\begin{align*}
\bigg(\frac{\lambda_{\widetilde{J}}\left(\tilde{t}_{n}\right)}{\lambda_{\widetilde{J}}\left(\tilde{s}_{n}\right)}\bigg)^{(m_0-\frac{1}{2})/\beta}|\ell| \leq 2 C \delta\left(\tilde{s}_{n}\right)^{\frac{2N}{N-2}}.
\end{align*}
On the other hand, by \eqref{the contradiction assumption} we deduce for large $n$
\begin{align*}
\bigg(\frac{\lambda_{\widetilde{J}}\left(\tilde{t}_{n}\right)}{\lambda_{\widetilde{J}}\left(\tilde{s}_{n}\right)}\bigg)^{(m_0-\frac{1}{2})/\beta}|\ell| \geq M\delta\left(\tilde{s}_{n}\right)^{\frac{2N}{N-2}}.
\end{align*}
Since we can choose $M$ arbitrarily, let $M=3C$, we obtain that $\delta\left(\tilde{s}_{n}\right)=0$ for large $n$, which contradicts \eqref{the contradiction}.
\end{proof}
Using Proposition \ref{prop: conclusion}, the following proposition can be obtained as in \cite{DKM7}.
\begin{proposition}\emph{(\cite{DKM7})}\label{the main estimate without o1}
Let $(N-2)\beta>2$, $C>0$, $J_{0}\geq 2$, $a>0$. There exists $\varepsilon=\varepsilon\left(C, J_{0}, a\right)>0$, such that for all $L>0$, there exists $T^{*}=T^{*}\left(L, C, J_{0}, a\right)$ with the following property. For all $T>0$, for all $C^{1}$ functions
\begin{align*}
\boldsymbol{\lambda} &=\left(\lambda_{j}\right)_{j}:[0, T] \rightarrow G_{J_{0}} \\
\boldsymbol{\beta} &=\left(\beta_{j}\right)_{j}:[0, T] \rightarrow \mathbb{R}^{J_{0}},
\end{align*}
satisfying, for all $t \in[0, T]$, for any $j$,
\begin{align*}
\gamma(\boldsymbol{\lambda})\triangleq \gamma \leq \varepsilon,
\end{align*}
\begin{align*}
\Big|\beta_{j}-\|\Lambda W_a\|_{L^{2}}^{2} \lambda_{j}^{\prime}\Big| \leq C \gamma^{\frac{(N-2)\beta-2}{4}},
\end{align*}
\begin{align*}
\bigg|\frac{1}{2} \sum_{j=1}^{J_{0}} \beta_{j}^{2}-\kappa_{1} \sum_{1 \leq j \leq J_{0}-1} \iota_{j} \iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}\bigg| \leq C \gamma^{\frac{(N-2)\beta-2}{2}},
\end{align*}
\begin{align*}
\bigg|\lambda_{j} \beta_{j}^{\prime}+\kappa_{0}\Big(\iota_{j} \iota_{j+1}\Big(\frac{\lambda_{j+1}}{\lambda_{j}}\Big)^{\frac{(N-2)\beta}{2}}-\iota_{j} \iota_{j-1}\Big(\frac{\lambda_{j}}{\lambda_{j-1}}\Big)^{\frac{(N-2)\beta}{2}}\Big)\bigg| \leq C \gamma^{\frac{N\beta}{2}},
\end{align*}
\begin{align*}
L \leq C \gamma^{\frac{N-2}{2}}\left(\frac{\lambda_{1}}{\lambda_{1}(0)}\right)^{a},
\end{align*}
we have
\begin{align*}
T \leq T^{*} \lambda_{1}(0).
\end{align*}
\end{proposition}
\medskip
{\it Proof of Theorem \ref{main thm}} By \eqref{the estimate of delta}, for all $t \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right]$, it holds that
\begin{align*}
\delta(t)^{\frac{2N}{N-2}} \lesssim \gamma(t)^{\frac{N\beta}{2}}+o_{n}(1),
\end{align*}
which, combined with \eqref{the equation 1 of cliam} and \eqref{the equation 2 of claim}, implies that for all $t \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right]$,
\begin{align*}
\delta(t)^{\frac{2N}{N-2}} \lesssim \tilde{\gamma}(t)^{\frac{N\beta}{2}}.
\end{align*}
According to Lemma \ref{lem: the lower bdd of lambda j}, there exists a constant $C>0$ such that for any $t \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right]$
\begin{align}\label{the lower bdd of lambda j by gamma}
|\ell| \leq C\Big(\frac{\lambda_{\tilde{J}}(t)}{\lambda_{\tilde{J}}\left(\tilde{t}_{n}\right)}\Big)^{(m_{0}-\frac{1}{2})/{\beta}} \tilde{\gamma}(t)^{\frac{N\beta}{2}}.
\end{align}
By \eqref{the tilde bdd of beta lambda}, \eqref{the tilde bdd of the second derivative}, \eqref{the tilde bdd of the beta2} of Lemma \ref{the tilde gamma bdd}, the parameters $\left(\beta_{j}\right)_{\tilde{J} \leq j \leq J}$ and $\left(\lambda_{j}\right)_{\tilde{J} \leq j \leq J}$ satisfy the assumptions of Proposition \ref{the main estimate without o1} for times $t \in\left[\tilde{t}_{n}, t_{n}^{\prime}\right]$. Hence, Proposition \ref{the main estimate without o1} yields that for large $n$
\begin{align*}
t_{n}^{\prime}-\tilde{t}_{n} \leq T_{*} \lambda_{\tilde{J}}\left(\tilde{t}_{n}\right).
\end{align*}
Thus the assumption $t'_{n}-\tilde{t}_{n}=T\lambda_{\tilde{J}}(t_n)$ implies that $T\leq T_{*}$, for a constant $T_{*}$ depending only on the solution $u$ and the parameters $\ell, m_{0}$. Since $T$ can be taken arbitrarily large, we obtain a contradiction and complete the proof of Theorem \ref{main thm}.
|
\section{Introduction}
In the present paper, we consider positive characteristic analogues of $p$-adic multiple $L$-functions (\cite{Furusho2017b}).
\subsection{$p$-adic multipe $L$-functions in characteristic $0$}
Kubota and Leopoldt's $p$-adic $L$-functions are defined to be the $p$-adic continuous functions on $\mathbb{Z}_p$ interpolating certain complex analytic $L$-values at negative integers. It is known that $p$-adic $L$-functions can be written as $p$-adic integrations and are related tightly to ideal class groups of a certain kind of number fields
(cf. \cite{LangBook}).
In \cite{Furusho2017b}, Furusho, Komori, Matsumoto and Tsumura constructed $p$-adic versions of multiple zeta functions called $p$-adic multiple $L$-functions. Their method of construction is a generalization of Koblitz's integral expressions of $p$-adic $L$-functions. Their $p$-adic multiple $L$-functions are $p$-adic multivariable functions which interpolate the special values of complex multiple zeta ($L$-) functions at negative integers in a sense. They described the special values of $p$-adic multiple $L$-functions and obtain multiple Kummer congruence among them (\cite[Theorems 2.1 and 2.10]{Furusho2017b}).
\subsection{Positive characteristic analogues}
In positive characteristic case, Goss constructed the $v$-adic zeta function (\cite{Goss1979}), which interpolates the special values of Carlitz-Goss zeta function at negative integers $v$-adically ($v$ is a finite place of the rational function field over a finite field). This function is considered to be a positive characteristic analogue of the $p$-adic $L$-functions. Goss obtained the integral expression of $v$-adic zeta function similar to those of $p$-adic $L$-functions (\cite[Theorem 7.6, Addendum Theorem 2.1]{Goss1979}). It was shown that his $v$-adic zeta function plays an important role in a positive characteristic analogue of Iwasawa theory (\cite{Angles2020a}).
Angl\`{e}s, Ngo Dac and Tavares Ribeiro introduced the multiple generalizations of Carlitz-Goss zeta functions, which can be seen as positive characteristic analogues of multiple zeta functions, by following ideas of Goss and Thakur (\cite[Corollary 6.1]{Angles2016}). They also introduced multiple versions of Goss' $v$-adic zeta function. These $v$-adic functions can be considered as positive characteristic analogues of $p$-adic multiple $L$-functions. (Though Angl\`{e}s et al. considered their functions in quite generalized setting, we only deal with the simplest ones of their functions.)
\subsection{Results}
Let us take a power $q$ of a prime number $p$. The polynomial ring $\mathbb{F}_q[\theta]$ and the field of rational function field $\mathbb{F}_q(\theta)$ over the finite field are denoted by $A$ and $k$, respectively. We call $\infty$-adic muliple zeta functions (abbreviated as $\infty$MZFs, here $\infty$ is the place corresponding to $1/\theta \in k$) for the multivariable functions introduced in \cite[Corollary 6.1]{Angles2016} since they can be seen as positive characteristic analogues of multiple zeta functions, and denote them by $\zeta_\infty(s_1,\,\dots,\,s_r)$, see Definition \ref{DefMZF} for details.
We obtain integral expression of the special values of $\infty$MZFs:
\begin{ntheorem}[{Theorem \ref{IntegralExpression}}]
There exists $A_v$-valued measure $\mu$ on $A_v^r$ such that the special value of the $\infty$MZF at $(-m_1,\,\dots,-m_r)$ has the following integral expression for $m_1,\,\dots,m_r \in \mathbb{N}$:
\begin{equation}
\zeta_{\infty}(-m_1,\,\dots,-m_r)=\int_{A_v^r}x_1^{m_1}\cdots x_r^{m_r} \mu \in A.
\end{equation}
\end{ntheorem}
Here, we fix an irreducible monic polynomial $v \in A$ and write $A_v$ for the completion of $A$ at the place corresponding to $v$ with maximal ideal $\mathfrak{m}_v$. The field of fraction of $A_v$ is denoted by $k_v$.
Also in the $v$-adic situation, we call $v$-adic muliple zeta functions (abbreviated as $v$MZFs) for the multivariable functions introduced in \cite[Corollary 6.2]{Angles2016} since they can be seen as positive characteristic analogues of $p$-adic multiple $L$-functions, and denote them by $\zeta_v(s_1,\,\dots,\,s_r)$. These are $k_v$-valued multivariable functions on the set $X_v^r$ with $X_v\coloneqq k_v^\times \times {\mathbb{Z}_{d,\,p}}= k_v^\times \times \mathbb{Z}/(q^{\deg v}-1) \times \mathbb{Z}_p$, see Definition \ref{Def$v$MZF} for details.
While in the case of $v$MZFs not only the special values but also the functions themselves admit integral expression:
\begin{ntheorem}[{Theorem \ref{IntegralexpressionofStar}}]
For $t_1,\,\dots,\,t_r \in {\mathbb{Z}_{d,\,p}}$ and $\sigma_1,\,\dots,\,\sigma_r \in k_v \setminus \mathfrak{m}_v$, there exists $A_v$-valued measure $\mu^{\sigma_1,\,\dots,\,\sigma_r}$ on $A_v^r$ such that the value of the $v$MZF at $\big((\sigma_1,\, -t_1),$ $\dots,$ $(\sigma_r,\,-t_r) \big)\in X_v^r$ has the following integral expression:
\begin{equation}
\zeta_v((\sigma_1,\, -t_1),\,\dots, (\sigma_r,\,-t_r))=\int_{(A_v \setminus \mathfrak{m}_v)^r}x_1^{t_1}\cdots x_r^{t_r} \mu^{\sigma_1,\,\dots,\,\sigma_r} \in k_v.
\end{equation}
(The measure $\mu^{1,\,\dots,\,1}$ coincides with $\mu$ in the $\infty$-adic situation.)
\end{ntheorem}
These integral expressions imply Kummer-type congruence between special values of $v$MZFs:
\begin{ntheorem}[{Theorem \ref{KummerType}}]
If we take indices $(m_1,\,\dots,\,m_r)$ and $(l_1,\,\dots,\,l_r) \in \mathbb{Z}^r$ such that $m_i \equiv l_i \not \equiv 0 \mod (q^{\deg v}-1)q^{e\deg v}$ for all $1\leq i \leq r$ and some $e \in \mathbb{N}$, then both $\zeta_v(m_1,\,\dots,\,m_r)$ and $\zeta_v(l_1,\,\dots,\,l_r)$ are in $A_v$ and we have the following congruence between the special values of a $v$MZF at $(m_1,\,\dots,\,m_r)$ and $(l_1,\,\dots,\,l_r)$:
\begin{equation}
\zeta_v(m_1,\,\dots,\,m_r) \equiv \zeta_v(l_1,\,\dots,\,l_r) \mod \mathfrak{m}_v^r.
\end{equation}
\end{ntheorem}
Further, we obtain results on special values of $\infty$MZFs and $v$MZFs at negative integers:
\begin{ntheorem}[{Theorem \ref{recursiveformulae}}]
Special values of $\infty$MZFs and $v$MZFs at negative integers can be written in terms of special values of $\infty$MZFs with simpler arguments.
\end{ntheorem}
As a corollary, we obtain the following:
\begin{ntheorem}[{Thoerem \ref{recursiveformulaeCor}}]
Special values of $\infty$MZFs at negative integers can be written in terms of those of $v$MZFs, and vice versa.
\end{ntheorem}
We also consider star variants $\zeta_\infty^\star(s_1,\,\dots,\,s_r)$ and $\zeta_v^\star(s_1,\,\dots,\,s_r)$ of $\infty$MZFs and $v$MZFs in \eqref{EqDefMZSF} and \eqref{EqDefvMZSF}, respectively. In the present paper, we show all the results except for Theorem \ref{recursiveformulaeCor} for star versions in the parallel way to non-star versions.
Stimulated by the results of \cite[Theorem 3]{Zlobin2005} and \cite[Lemma 4.2.1]{Chang2019a}, we obtain the following relations between non-star variants and star-variants of $\infty$MZFs and $v$MZFs, which we call the orthogonal property:
\begin{ntheorem}[{Theorem \ref{orthogonality}}]
For $s_1,\,\dots,\,s_r \in X_\infty$ (see subsection \ref{2.1} for the definition), we have
\begin{equation}
\sum_{l=0}^{r}(-1)^{l}\zeta_\infty(s_r,\,\dots,\,s_{r-l+1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-l})=0.
\end{equation}
Similarly, for $s_1,\,\dots,\,s_r \in X_v$, we have
\begin{equation}
\sum_{l=0}^{r}(-1)^{l}\zeta_v(s_r,\,\dots,\,s_{r-l+1})\zeta_v^\star(s_1,\,\dots,\,s_{r-l})=0.
\end{equation}
\end{ntheorem}
The plan of the present paper goes as follows: Section \ref{SectionDefinitions} is on the definitions of $\infty$MZFs and $v$MZFs. The orthogonal properties (Theorem \ref{orthogonality}) are obtained in this section. In Section \ref{Construction of multiple zeta measures}, we obtain the integral expression of them (Theorem \ref{IntegralExpression}) and observe that they imply a Kummer-type congruence (Theorem \ref{KummerType}). We devote Section \ref{SectionINductiveFormulae} to recursive formulae which enable us to write down special values of $\infty$MZFs and $v$MZFs at negative integers in terms special values with simpler arguments (Theorems \ref{recursiveformulae}, and \ref{recursiveformulaev}). As a corollary, we obtain relationships between special values of $\infty$-adic and those of $v$-adic multipe zeta functions at negative integers (Theorem \ref{EqInductiveCor}).
\section{$\infty$-adic and $v$-adic multiple zeta functions}\label{SectionDefinitions}
This section is on the definitions of $\infty$MZFs and of $v$MZFs. In section \ref{2.1}, we review the definitions of Carlitz-Goss zeta function and of Goss' $v$-adic zeta function. In section \ref{2.2}, $\infty$MZFs and $v$MZFs are defined as certain infinite sums (Definitions \ref{DefMZF} and \ref{Def$v$MZF}), which appeared in \cite[\S 6]{Angles2016}.
Let us fix a power $q$ of a prime number $p$ and take indeterminate variable $\theta$. We consider the one-variable polynomial ring $A\coloneqq \mathbb{F}_q[\theta]$ over the finite field with $q$ elements to be a positive characteristic analogue of the ring $\mathbb{Z}$. The symbol $A_+$ denotes the set of monic polynomials in $A$, which is regarded as an analogue of the set $\mathbb{N}$.
The field of fractions $k\coloneqq \mathbb{F}_q(\theta)$ of $A$ can be regarded as an analogue of the field $\mathbb{Q}$.
\subsection{Single variable case} \label{2.1}
The field of Laurent series $k_\infty\coloneqq \mathbb{F}_q((1/\theta))$ is considered to be an analogue of the field $\mathbb{R}$ and the completion $\mathbb{C}_\infty$ of its algebraic closure $\overline{k}_\infty$ is considerd to be an analogue of the field $\mathbb{C}$.
Goss' character space at $\infty$ is defined by $X_\infty\coloneqq \mathbb{C}_\infty^\times \times \mathbb{Z}_p$ (\cite{Goss1983}) and, for $s=(\sigma,\,t)\in X_\infty$ and $n \in A_+$, we define
\begin{equation}n^s\coloneqq \sigma^{\deg n} (\theta^{-\deg n} n)^t. \label{powerXinfty}
\end{equation}
We regard the set $\mathbb{Z}$ as a subset of $X_\infty$ via the inclusion $i\mapsto (\theta^i,\,i)$, then the power $n^i$ in the sense of \eqref{powerXinfty} coincides with the power in the usual sense.
An analogue of Riemann zeta function is introduced by Carlitz (the case $s \in \mathbb{N}_{\geq1}$) and Goss (in general):
\begin{definition}[{\cite{Carlitz1935,Goss1979}}] \label{Defzeta}
The \textit{Carlitz-Goss zeta function} is the function $\zeta_\infty : X_\infty \rightarrow \mathbb{C}_\infty$ defined by
\begin{equation}
\zeta_{\infty}(s)\coloneqq\sum_{i\geq0}S_{i}(s) \label{eqrefCZV}
\end{equation}
where
\begin{equation}
S_i(s)\coloneqq \sum_{\substack{n \in A_+,\\\deg n =i}} \frac{1}{n^s}.
\end{equation}
Especially, we call the value $\zeta_{\infty}(m)$ \textit{the Carlitz zeta value} if $m \in \mathbb{N}$.
\end{definition}
Let us confirm the convergence of \eqref{eqrefCZV} in the case $s \in \mathbb{Z}$. Since $\mathbb{C}_\infty$ is ultrametric, the sum \eqref{eqrefCZV} converges if $s\geq1$. In the case $s\leq0$, the following lemma implies that the sum is finite.
\begin{lemma}[{\cite[Proposition 4.1]{Goss1979}}] \label{firstlemma}
For $m\in \mathbb{N}_{\geq0}$, $S_i(-m)=0$ if $i\geq m+1$.
\end{lemma}
\begin{proof}This lemma in proven by induction. If $m=0$, we have
\begin{equation*}
S_i(0)=1\cdot(q^{i})=0.
\end{equation*}
For $m\geq1$, the assertion is shown as follows:
\begin{align*}
S_{i}(-m)&=\sum_{\substack{\deg a =i-1\\a:\text{monic}\\b\in\mathbb{F}_q}}(\theta a+b)^m
=\sum_{\substack{\deg a =i-1\\a:\text{monic}\\b\in\mathbb{F}_q}} \sum_{l=0}^m (\theta a)^l b^{m-l} \dbinom{m}{l}\\
&=\sum_{\substack{\deg a =i-1\\a:\text{monic}}} \theta^m a^m \sum_{b\in\mathbb{F}_q}b^0+ \sum_{\substack{0\leq l\leq m-1\\b\in \mathbb{F}_q}}\theta^l b^{m-l}\dbinom{m}{l}\left(\sum_{\substack{\deg a =i-1\\a:\text{ monic}}}a^l\right)=0.
\end{align*}
\end{proof}
\begin{remarks}
Goss proved that the sum \eqref{eqrefCZV} converges for any $s \in X_\infty$ (\cite[Corollary 2.3.4]{Goss1983}).
It is known that this function satisfies some properties similar to those of the Riemann zeta function. For example, Carlitz zeta values $\zeta_{\infty}(s)$ with $s\in \mathbb{N},\, (q-1)|s$ can be written down in terms of analogues of Bernoulli numbers and fundamental period $2\pi i$ of exponential function (see \cite[Theorem 9.4]{Carlitz1935} for details) and they have Euler product expression.
\end{remarks}
Irreducible monic polynomials are consider to be an analogue of prime numbers. The symbol $v$ stands for a fixed irreducible monic polynomial in $A$ of degree $d>0$.
We denote the completions of $A$ and $k$ corresponding to $v$ by $A_v$ and $k_v$, respectively. They are consider to be analogues of $\mathbb{Z}_p$ and $\mathbb{Q}_p$, respectively. We write $\mathfrak{m}_v$ for the maximal ideals of $A$ and $A_v$ generated by $v$.
We put
\begin{equation}
\mathbb{Z}_{d,\,p}\coloneqq \lim_{\substack{\leftarrow\\e}}\mathbb{Z}/(q^d-1)q^{de}=\mathbb{Z}/(q^d-1) \times \mathbb{Z}_p.
\end{equation}
We can define $n^t$ for $n \in A_v \setminus \mathfrak{m}_v$ and $t \in {\mathbb{Z}_{d,\,p}}$ (\cite[Proposition 6.2]{Goss1979}).
We further put $X_v\coloneqq k_v^\times \times {\mathbb{Z}_{d,\,p}}$ and, for $s=(\sigma,\,t)\in X_v$ and $n\in A_v \setminus \mathfrak{m}_v$, define $n^s\coloneqq \sigma^{\deg n}n^{t}$. The set $\mathbb{Z}$ can be seen as a subset of $X_v$ via the inclusion $i\mapsto (1,\,i)$.
Goss introduced $v$-adic zeta function, which interpolates the special values of Carlitz-Goss zeta function at negative integers $v$-adically and hence can be seen as a positive characteristivc analogue of $p$-adic $L$ function:
\begin{definition}[{\cite[Addendum, Definition 1.1]{Goss1979}}]\label{Defvzeta}
The \textit{$v$-adic zeta function} $\zeta_v(s)\,(s=(\sigma,t) \in X_v)$ is the function on $X_v$ with values in $k_v$ defined by
\begin{equation}
\zeta_v(s)\coloneqq \sum_{i \geq0}\sigma^{-i} \tilde{S}_i(t)=\sum_{i\geq0}\tilde{S}_i(s) \in k_v,\label{EqDefzvetafunction}
\end{equation}
where
\begin{equation}
\tilde{S}_i(s)\coloneqq \sum_{\substack{n \in A_+\\\deg n =i\\(n,v)=1}}\frac{1}{n^s} \in k_v.
\end{equation}
\end{definition}
Since the equality
\begin{equation}
\tilde{S}_i(s)=\begin{cases}
S_i(s)-v^{-s}S_{i-d}(s) &(i\geq d),\\
S_i(s) & (i<d),\\
\end{cases}
\end{equation}
holds if $s \in \mathbb{Z}$, we have
\begin{equation}\zeta_v(-m)=(1-v^{m})\zeta_{\infty}(-m) \in k
\end{equation}
for any $m\in \mathbb{Z}_{\geq0}$, i.e., $v$-adic zeta function interpolates the special values of Carlitz-Goss zeta function at negative integers $v$-adically and hence can be seen as a positive characteristivc analogue of $p$-adic $L$-function. Goss proved that the sum \eqref{EqDefzvetafunction} converges (\cite[Theorem 5.4.7]{Goss1983}) and has an integral expression (Theorem \ref{SingleIntegralExpression}).
\begin{remarks}
Goss considered $\infty$-adic and $v$-adic zeta functions in more general setting where the ring $A$ is allowed to be the ring of functions of arbitrary smooth curve over $\mathbb{F}_q$ regular outside a fixed rational point $\infty$ (\cite{Goss1979, Goss1983}).
\end{remarks}
\subsection{Multivariable case} \label{2.2}
We consider positive characteristic analogues of multiple zeta functions:
\begin{definition}[{\cite[\S 6]{Angles2016}}]\label{DefMZF}
For $s_1=(\sigma_1,\,t_1),\,\dots,\,s_r=(\sigma_r,\,t_r) \in X_\infty$, we define
\begin{align}
\zeta_{\infty}(s_1,\,\dots,\,s_r)&\coloneqq \sum_{i_1>\cdots>i_r\geq 0} S_{i_1}(s_1)\cdots S_{i_r}(s_r) \in \mathbb{C}_\infty, \label{EqDefMZF} \\
\zeta_{\infty}^\star(s_1,\,\dots,\,s_r)&\coloneqq \sum_{i_1\geq\cdots\geq i_r\geq 0} S_{i_1}(s_1)\cdots S_{i_r}(s_r) \in \mathbb{C}_\infty.\label{EqDefMZSF} \\
\end{align}
We call these functions on $X_\infty^r$ \textit{$\infty$-adic multiple zeta function} (\textit{$\infty$MZF} in short) and \textit{$\infty$-adic multiple zeta star function} (\textit{$\infty$MZSF} in short), respectively.
\end{definition}
Though it follows from the arguments in \cite[Corollary 6.1]{Angles2016} in their more general setting that the series \eqref{EqDefMZF} and \eqref{EqDefMZSF} defining $\infty$MZFs and $\infty$MZSFs converge and are rigid analytic, we present proofs in order to make our paper self-contained in Appendix \ref{AppA}.
\begin{remarks}
(1) In \cite{Angles2016}, Angl\`{e}s et al. worked in more general setting where the ring $A$ is allowed to be the ring of functions of arbitrary smooth curve over $\mathbb{F}_q$ regular outside a fixed rational point $\infty$. They also considered twisting of the numerators of summands.
(2) If $r=1$, both of $\infty$MZF and $\infty$MZSF coincide with Carlitz-Goss zeta function in Definition \ref{Defzeta}.
(3) For $m_1,\,\dots ,\, m_r \in \mathbb{Z}_{\geq 0}$, $\zeta_{\infty}(-m_1,\,\dots,\,-m_r),\, \zeta_{\infty}^\star(-m_1,\,\dots,\,-m_r) \in A$ since Lemma \ref{firstlemma} implies that the sum in \eqref{EqDefMZF} and \eqref{EqDefMZSF} are finite sums of elements in $A$.
(4) The special values of these functions at positive integers coincide with Thakur's multiple zeta values and multiple zeta star values, respectively (\cite[\S 5.10]{ThakurBook}).
\end{remarks}
Next we consider positive characteristic analogues of $p$-adic multiple $L$-functions:
\begin{definition}[{\cite[\S 6]{Angles2016}}] \label{Def$v$MZF}
For $s_1=(\sigma_1,\,t_1),\,\dots,\,s_r=(\sigma_r,\,t_r) \in X_v$, we define
\begin{align}
\zeta_v(s_1,\,\dots,\,s_r)&\coloneqq \sum_{i_1>\cdots>i_r\geq 0} \sigma_1^{-i_1} \cdots\sigma_r^{-i_r}\tilde{S}_{i_1}(t_1)\cdots \tilde{S}_{i_r}(t_r) \label{EqDef$v$MZF}\\
&=\sum_{i_1>\cdots>i_r\geq 0} \tilde{S}_{i_1}(s_1)\cdots \tilde{S}_{i_r}(s_r) \in k_v ,\\
\zeta_v^\star (s_1,\,\dots,\,s_r)&\coloneqq \sum_{i_1\geq \cdots \geq
i_r\geq 0} \sigma_1^{-i_1} \cdots\sigma_r^{-i_r}\tilde{S}_{i_1}(t_1)\cdots \tilde{S}_{i_r}(t_r) \label{EqDefvMZSF}\\
&=\sum_{i_1\geq\cdots\geq i_r\geq 0} \tilde{S}_{i_1}(s_1)\cdots \tilde{S}_{i_r}(s_r) \in k_v.\\
\end{align}
We call these functions on $X_v^r$ by \textit{$v$-adic multiple zeta functions} (\textit{$v$MZFs} in short) and \textit{$v$-adic multiple zeta star functions} (\textit{$v$MZSFs} in short), respectively.
\end{definition}
Though it follows from the arguments in \cite[Corollary 6.2]{Angles2016} that the series \eqref{EqDef$v$MZF} and \eqref{EqDefvMZSF} converge and are rigid analytic, we present proofs in order to make our paper self-contained in Appendix \ref{AppA}.
\begin{remarks}\label{remarkink}
(1) In \cite{Angles2016}, Angl\`{e}s et al worked in more general setting where the ring $A$ is allowed to be the ring of functions of arbitrary smooth curve over $\mathbb{F}_q$ regular outside a fixed rational point $\infty$. They also considered twisting of the numerators of summands.
(2) In $r=1$, both of the $v$MZF and $v$MZSF coincide with $v$-adic zeta function in Definition \ref{Defvzeta}.
(3) For $m_1,\,\dots,\, m_r \in \mathbb{Z}_{\geq 0}$, we can see that the special values $\zeta_v(-m_1,\,\dots,\,-m_r)$ and $ \zeta_v^\star(-m_1,\,\dots,\,-m_r)$ are in $A$ since Lemma \ref{firstlemma} implies that the sum in \eqref{EqDef$v$MZF} and \eqref{EqDefvMZSF} are finite sums of elements in $A$.
(4) The values of $\zeta_v(s_1,\,\dots,\,s_r)$ at integers coincide with (interpolated) $v$-adic multiple zeta values introduced by Thakur (\cite[\S 5.10]{ThakurBook}). It should be noticed that he suggested that they can be interpolated to $v$-adic continuous functions on $\mathbb{Z}_{d,\,p}^r$. Q.~Shen also studied these function on $\mathbb{Z}_{d,\,p}^r$ (\cite{Shenthesis}). Another formulation of $v$-adic multiple zeta values was also introduced by Chang and Mishiba (\cite{Chang2021}). Their $v$-adic multiple zeta values are different from those introduced by Thakur, but it is expected that $v$-adic multiple zeta values introduced by Thakur and those introduced by Chang and Mishiba are related each other via certain linear relations.
\end{remarks}
\subsection{Orthognal propeties of MZFs}
The following theorem, which we call the orthogonal property, is the positive characteristic analogues of \cite[Theorem 3]{Zlobin2005}.
\begin{theorem}\label{orthogonality}
For $s_1,\,\dots,\,s_r \in X_\infty$, we have
\begin{equation}
\sum_{l=0}^{r}(-1)^{l}\zeta_\infty(s_r,\,\dots,\,s_{r-l+1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-l})=0 \label{star and nonstar}
\end{equation}
(here, the summand with $l=0$ is $\zeta_\infty^\star(s_1,\,\dots,\,s_{r})$ and the term with $l=r$ is $(-1)^r\zeta_\infty(s_r,\,\dots,\,s_{1})$). Similarly, for $s_1,\,\dots,\,s_r \in X_v$, we have
\begin{equation}
\sum_{l=0}^{r}(-1)^{l}\zeta_v(s_r,\,\dots,\,s_{r-l+1})\zeta_v^\star(s_1,\,\dots,\,s_{r-l})=0
\end{equation}
\end{theorem}
\begin{proof}
The vanishing of the summation of the $\infty$-adic functions can be confirmed as follows:
\begin{align}
&\quad \ \zeta_\infty^\star (s_1,\,\dots,\,s_r)=\sum_{i_1\geq\cdots \geq i_r\geq0} S_{i_1}(s_1)\cdots S_{i_r}(s_r)\\
\intertext{abbreviating $S_{i_j}(s_j)$ as $S_j$ for $1\leq j \leq r$ (each $S_j$ depends on $i_j \in \mathbb{Z}_{\geq 0}$),}
&=\sum_{i_r=0}^\infty S_r \sum_{i_1\geq\cdots \geq i_{r-1}}S_1 \cdots S_{r-1} -\sum_{\substack{i_1\geq\cdots \geq i_{r-1}\\i_{r}>i_{r-1}}} S_1\cdots S_r\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1}) -\sum_{\substack{i_1\geq\cdots \geq i_{r-1}\\i_{r}>i_{r-1}}} S_1\cdots S_r\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-\sum_{i_{r}>i_{r-1}}S_{r-1}S_r \sum_{i_1\geq\cdots\geq
i_{r-2}}S_1\cdots S_{r-2}\\
&\quad +\sum_{\substack{i_1\geq\cdots \geq i_{r-2}\\i_{r}>i_{r-1}>i_{r-2}}} S_{i_1}\cdots S_{i_r}\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})\\
&\quad +\sum_{\substack{i_1\geq\cdots \geq i_{r-2}\\i_{r}>i_{r-1}>i_{r-2}}} S_{i_1}\cdots S_{i_r}\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})\\
&\quad +\sum_{i_{r}>i_{r-1}>i_{r-2}}S_{r-2}S_{r-1}S_r \sum_{i_1\geq\cdots \geq i_{r-3}}S_1\cdots S_{r-3}\\
&\quad -\sum_{\substack{i_1\geq\cdots \geq i_{r-3}\\i_{r}>i_{r-1}>i_{r-2}>i_{r-3}}} S_{i_1}\cdots S_{i_r}\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})\\
&\quad +\zeta_\infty(s_r,\,s_{r-1},\,s_{r-2})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-3}) -\sum_{\substack{i_1\geq\cdots \geq i_{r-3} \\i_{r}>i_{r-1}>i_{r-2}>i_{r-3}}} S_{i_1}\cdots S_{i_r}\\
&\ \ \vdots\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})+\cdots\\
&\quad \cdots-(-1)^{r-2} \sum_{\substack{i_1\geq i_2\\i_r>\cdots>i_2}}S_1 \cdots S_r\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})+\cdots\\
&\quad \cdots-(-1)^{r-2} \sum_{i_r>\cdots>i_3}S_3 \cdots S_r \sum_{i_1\geq i_2}S_1 S_2 -(-1)^{r-1} \sum_{\substack{i_1\\i_r>\cdots>i_2}}S_1\cdots S_r\\
&=\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})+\cdots\\
&\quad \cdots-(-1)^{r-2} \zeta_\infty(s_r,\,\dots,\,s_3)\zeta_\infty^\star(s_1,\,s_2) -(-1)^{r-1} \sum_{\substack{i_1\\i_r>\cdots>i_2}}S_1\cdots S_r\\
&=-(-1)^r\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-(-1)^2\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})+\cdots\\
&\quad \cdots-(-1)^{r-2} \zeta_\infty(s_r,\,\dots,\,s_3)\zeta_\infty^\star(s_1,\,s_2) -(-1)^{r-1} \sum_{i_r>\cdots>i_2}S_2\cdots S_r \sum_{i_1}S_1 \\
&\quad - (-1)^r\sum_{i_r>\cdots>i_1}S_1\cdots S_r\\
&=-(-1)^r\zeta_\infty(s_r)\zeta_\infty^\star(s_1,\,\dots,\,s_{r-1})-(-1)^2\zeta_\infty(s_r,\,s_{r-1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-2})+\cdots\\
&\quad \cdots-(-1)^{r-2} \zeta_\infty(s_r,\,\dots,\,s_3)\zeta_\infty^\star(s_1,\,s_2) -(-1)^{r-1} \zeta_\infty(s_r,\,\dots,\,s_2)\zeta_\infty^\star(s_1)\\
&\quad - (-1)^r \zeta_\infty(s_r,\,\dots,\,s_1)\\
&=-\sum_{l=1}^{r}(-1)^{l}\zeta_\infty(s_r,\,\dots,\,s_{r-l+1})\zeta_\infty^\star(s_1,\,\dots,\,s_{r-l}).
\end{align}
Replacing $S_j=S_{i_j}(s_j)$ by $\tilde{S}_{i_j}(s_j)$, we can obtain the formula for $v$MZFs and $v$MZSFs.
\end{proof}
\begin{remarks}
In characteristic $0$ case, it is known that the completely same relations betsween multiple zeta values and multiple zeta star values holds, see \cite{Zlobin2005} for details.
\end{remarks}
\begin{remarks}
In positive characteristic case, Chang and Mishiba obtained similar relations
\begin{equation}
\sum_{l=1}^r (-1)^l \operatorname{Li}_{m_r,\,\dots,\,m_{r-l-1}}(z_r,\,\dots,\,z_{r-l-1})\operatorname{Li}_{m_1,\,\dots,\,m_{r-l}}^\star(z_1,\,\dots,\,z_{r-l})=0 \label{CMstarnonstar}
\end{equation}
between Carlitz multiple polylogarithms and Carlitz multiple star polylogarithms (see \cite[Lemma 4.2.1]{Chang2019a} for the definitions).
Based on the work of Anderson and Thakur in \cite{Anderson1990}, Chang showed in \cite[Theorem 5.5.2]{Chang2014} that there are $u_{m,\,0},\,\dots,\,u_{m\,e_m} \in A$ such that we have
\begin{align*}
\Gamma_{m_1}\cdots \Gamma_{m_r} \zeta(m_1,\,\dots,\,m_r)
&=\sum_{(j_1,\,\dots,\,j_r)\in \mathfrak{J}}\theta^{j_1+\cdots+j_r}\operatorname{Li}_{m_1,\,\dots,\,m_r}(u_{m_1,\,j_1},\,\dots,\,u_{m_r,\,j_r}), \\
\Gamma_{m_1}\cdots \Gamma_{m_r} \zeta^\star(m_1,\,\dots,\,m_r)
&=\sum_{(j_1,\,\dots,\,j_r)\in \mathfrak{J}}\theta^{j_1+\cdots+j_r}\operatorname{Li}_{m_1,\,\dots,\,m_r}^\star(u_{m_1,\,j_1},\,\dots,\,u_{m_r,\,j_r})
\end{align*}
for $m_1,\,\dots,\,m_r \in \mathbb{Z}_{>0}$, where $\Gamma_{m_1},\dots, \Gamma_{m_r}$ are non-zero elements of $A$ called the Carlitz gammas and $\mathfrak{J}=\mathfrak{J}_{m_1,\,\dots,m_r}$ is the product set $\{0,\,1,\,\dots,\,e_{m_1}\}\times\cdots\times\{0,\,1,\,\dots,\,e_{m_r}\}$ for some $e_{m_j} \in \mathbb{Z}_{\geq0}$. Using these equations, we can show that \eqref{star and nonstar} can be deduced from \eqref{CMstarnonstar} in the case where $s_j=m_j \in \mathbb{Z}_{>0}$ $(j=1,\,\dots,\,r)$ as follows:
\begin{align*}
&\sum_{(j_1,\,\dots,\,j_r)\in \mathfrak{J}_{m_1,\dots,m_r}} \theta^{j_1+\cdots+j_r}\Bigg\{\sum_{l=0}^r (-1)^l \operatorname{Li}_{m_r,\,\dots,\,m_{r-l+1}}(u_{m_r,\,j_r},\,\dots,\,u_{m_{r-l+1},\,j_{r-l+1}})\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\cdot\operatorname{Li}_{m_1,\,\dots,\,m_{r-l}}^\star(u_{m_1,\,j_1},\,\dots,\,u_{m_{r-l},\,j_{r-1}}) \Bigg\}\\
&=\sum_{l=0}^r (-1)^l \sum_{\substack{(j_1,\,\dots,\,j_r) \\ \in \mathfrak{J}_{m_1,\dots,m_r}}} \Bigg\{\theta^{j_r+\cdots+j_{r-l+1}}\operatorname{Li}_{m_r,\,\dots,\,m_{r-l+1}}(u_{m_r,\,j_r},\,\dots,\,u_{m_{r-l+1},\,j_{r-l+1}})\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\cdot\theta^{j_1+\cdots+j_l}\operatorname{Li}_{m_1,\,\dots,\,m_{r-l}}^\star(u_{m_1,\,j_1},\,\dots,\,u_{m_{r-l},\,j_{r-l}}) \Bigg\}\\&=\sum_{l=0}^r (-1)^l \sum_{\substack{(j_r,\,\dots,\,j_{r-l+1})\\ \in \mathfrak{J}_{m_r,\dots,m_{r-l+1}}}}\Bigg\{\theta^{j_r+\cdots+j_{r-l+1}}\operatorname{Li}_{m_r,\,\dots,\,m_{r-l+1}}(u_{m_r,\,j_r},\dots,u_{m_{r-l+1},\,j_{r-l+1}})\Bigg\}\\
&\qquad\qquad\qquad\qquad\quad\cdot\Bigg\{\sum_{\substack{(j_1,\,\dots,\,j_l) \\ \in\mathfrak{J}_{m_1,\dots,m_l}}}\theta^{j_1+\cdots+j_l}\operatorname{Li}_{m_1,\,\dots,\,m_{r-l}}^\star(u_{m_1,\,j_1},\,\dots,\,u_{m_{r-l},\,j_{r-l}}) \Bigg\}\\
&=\sum_{l=0}^{r}(-1)^{l}\Gamma_{m_r}\cdots \Gamma_{m_{r-l+1}} \zeta_\infty(m_r,\,\dots,\,m_{r-l+1})\Gamma_{m_1}\cdots \Gamma_{m_{r-l}} \zeta_\infty^\star(m_1,\,\dots,\,m_{r-l})\\
&=\Gamma_{m_1}\cdots \Gamma_{m_r} \sum_{l=0}^{r}(-1)^{l}\zeta_\infty(m_r,\,\dots,\,m_{r-l+1})\zeta_\infty^\star(m_1,\,\dots,\,m_{r-l}).
\end{align*}
\end{remarks}
\section{Integral expressions}\label{Construction of multiple zeta measures}
In subsection \ref{Single valued case}, we review the integral expression of special values of Carlitz-Goss zeta function at negative integers and of $v$-adic zeta functions obtained by Goss (Theorem \ref{SingleIntegralExpression}). We obtain similar integral expressions of special values of $\infty$MZFs at negative integers and $v$MZFs in subsection \ref{Multivariable case} (Theorem \ref{IntegralExpression}). From these expression of $v$MZFs, we obtain some congruence between special values of $v$MZFs in subection \ref{Multiple Kummer-type congruence}.
\subsection{Single variable case}\label{Single valued case}
In this subsection, we recall Goss' construction of $v$-adic measure which gives integral expressions of special values of Carlitz-Goss zeta function at negative integers and integral expressions of $v$-adic zeta functions (\cite{Goss1979}).
\begin{definition}[{\cite[Addendum Definition 2.1]{Goss1979}}] \label{Bernoulli-Goss measure}
For $m \in \mathbb{Z}_{\geq 0}$ and $\sigma \in k_v \setminus \mathfrak{m}_v$, we define $A_v$-valued measure $\mu_m^\sigma$ as follows:
\begin{equation}
\mu_m^\sigma(a+\mathfrak{m}_v^e)\coloneqq \sum_{j\geq0} \sigma^{-j} \sum_{\substack{n\in A_+\\ \deg n=j\\ n \equiv a \mod \mathfrak{m}_v^e}}n^m.
\end{equation}
We simply write $\mu^{\sigma}$ for $\mu_0^{\sigma}$.
\end{definition}
\begin{remarks}
The classical \textit{$m$-th Bernoulli distribution} $B_m$ on $\mathbb{Z}_p$ is defined by
\begin{equation}
B_m(a+p^e\mathbb{Z}_p)\coloneqq (p^e)^{m-1}\mathbb{B}_m\left(\frac{a}{p^e}\right)=m\zeta_{\equiv a (p^e)}(1-m)\quad(0\leq a <p^e),
\end{equation}
where $\mathbb{B}_m(x)$ is the $m$-th Bernoulli polynomial and $\zeta_{\equiv a (p^e)}(s)$ is the partial zeta function defined by
\begin{equation*}
\zeta_{\equiv a (p^e)}(s)\coloneqq \sum_{n \equiv a \mod p^e}\frac{1}{n^s} \quad \operatorname{Re}(s)>1,
\end{equation*}
(cf. \cite[\S 12.1, Example (3)]{WashingtonBook}). Therefore Goss' measures can be seen as positive characteristic analogues of the Bernoulli distributions. Since the Bernoulli distributions are not measures but are just distributions, we have to modify them in order to consider the integration. In contrast to the classical case, we do not need such modifications in the positive characteristic case.
\end{remarks}
Goss obtained the following integral expressions of special values of Carlitz-Goss zeta function at negative integers and of $v$-adic zeta function:
\begin{theorem}[{\cite[Theorem 7.6, Addendum Theorem 2.1]{Goss1979}}]\label{SingleIntegralExpression}
(1) For $m \in \mathbb{N}_{\geq0}$, the following equation holds:
\begin{equation}
\zeta_{\infty}(-m)\coloneqq \int_{A_v}x^m \mu^1.
\end{equation}
(2) For $t \in {\mathbb{Z}_{d,\,p}}$ and $\sigma \in k_v \setminus \mathfrak{m}_v$, we have the following equation:
\begin{equation}
\zeta_v((\sigma,\,-t))\coloneqq \int_{A_v \setminus \mathfrak{m}_v}x^t \mu^\sigma.
\end{equation}
\end{theorem}
\subsection{Multivariable case}\label{Multivariable case}
In this subsection, we generalize the result by Goss dealt with in the previous subsection to the multivariable setting.
\begin{definition}\label{Def multiple zeta measure}
For $\sigma_1,\,\dots,\,\sigma_r \in k_v \setminus \mathfrak{m}_v$, we define $A_v$-valued measures $\mu=\mu^{\sigma_1,\,\dots,\,\sigma_r}$ and $\mu^\star=\mu^{\star\,\sigma_1,\,\dots,\,\sigma_r}$ on $A_v^r$ as follows: for $e_1,\,\dots,\,e_r \in \mathbb{N}$ and $ \alpha_1,\,\dots,\,\alpha_r \in A$,
\begin{align}
\mu((\alpha_1,\,\dots,\,\alpha_r)+\underbrace{\mathfrak{m}_v^e \times \cdots\times \mathfrak{m}_v^e}_{r \text{ times}})
\coloneqq \sum_{i_1>\cdots>i_r\geq0} \sigma_1^{-i_1}\cdots\sigma_r^{-i_r}
\sum_{\substack{n_j \in A_+\, \deg n_j=i_j\\ \alpha_j \equiv n_j \mod \mathfrak{m}_v^{e_j} \\ (\text{for } r \geq j \geq 1)}}1, \label{defMZM}\\
\mu^\star((\alpha_1,\,\dots,\,\alpha_r)+\underbrace{\mathfrak{m}_v^e \times \cdots\times \mathfrak{m}_v^e}_{r \text{ times}})
\coloneqq \sum_{i_1\geq\cdots\geq i_r \geq0} \sigma_1^{-i_1}\cdots\sigma_r^{-i_r}
\sum_{\substack{n_j \in A_+\, \deg n_j=i_j\\ \alpha_j \equiv n_j \mod \mathfrak{m}_v^{e_j} \\ (\text{for } r \geq j \geq 1)}}1.
\end{align}
\end{definition}
Let us take an open set $(\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e_1}\times\cdots\times \mathfrak{m}_v^{e_r} \subset A_v^r$ where $\alpha_1,\,\dots,\,\alpha_r \in A$ and $e_1,\,\dots,\,e_r \in \mathbb{Z}_{\geq0}$, and let $J \in \{1,\,\dots,\,r\}$. For fixed $i_1>\cdots>i_r$, we have
\begin{equation*}
\sum_{\substack{\deg n_j =i_j\\n_j \equiv \alpha_j (\mathfrak{m}_v^{e_j})\\(\text{for } r \geq j \geq 1)}}1=\sum_{\substack{\beta \in A_v/(\mathfrak{m}_v^{e_J+1})\\ \beta \equiv \alpha_J (\mathfrak{m}_v^{e_J})}}\sum_{\substack{\deg n_j =i_j,\,(\text{for } r \geq j \geq 1)\\n_j \equiv \alpha_j (\mathfrak{m}_v^{e_j}),\,(j \neq J) \\n_J\equiv \beta (\mathfrak{m}_v^{e_J+1})}}1.
\end{equation*}
Multiplying $\sigma_1^{-i_1}\cdots\sigma_r^{-i_r}$ and summing up all $i_1>\cdots>i_r$, we have
\begin{align*}
&\mu((\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e_1}\times\cdots\times \mathfrak{m}_v^{e_r})\\
=&\sum_{\substack{\beta \in A_v/(\mathfrak{m}_v^{e_J+1})\\ \beta \equiv \alpha_J (\mathfrak{m}_v^{e_J})}}\mu((\alpha_1,\,\dots,\,\beta,\,\dots,\alpha_r)+\mathfrak{m}_v^{e_1}\times\cdots \times \mathfrak{m}_v^{e_J+1}\times \cdots \times \mathfrak{m}_v^{e_r}).
\end{align*}
Therefore, we have
\begin{equation}
\mu\left(\bigcup_{i\in I}U_i\right)=\sum_{i\in I}\mu(U_i),\
\mu^\star \left(\bigcup_{i\in I}U_i \right)=\sum_{i\in I}\mu^\star(U_i),
\end{equation}
for any pairwise disjoint finite family $\{ U_i\}_{i\in I}$ of compact open sets in $A_v^r$.
If $r=1$, the measures defined above coincides with that defined in Definition \ref{Bernoulli-Goss measure}.
\begin{prop}
If we fix $e \in \mathbb{N}$, take $\sigma_1,\,\dots,\,\sigma_r \in k_v \setminus \mathfrak{m}_v$ and take $\alpha_1,\,\dots,\,\alpha_r \in A$ such that $\deg \alpha_1,\,\dots,\, \deg \alpha_r < \deg v^e=de$, then the volume of the compact open set $(\alpha_1,\,\dots,\,\alpha_r)+\underbrace{\mathfrak{m}_v^e \times \cdots\times \mathfrak{m}_v^e}_{r \text{ times}}$ defined by the measure $\mu$ can be calculated as follows:
\begin{enumerate}
\item \label{main terms}
If $\deg \alpha_1>\cdots> \deg \alpha_r$ and $\alpha_1,\,\dots,\,\alpha_r \in A_+$, then
\begin{equation}
\mu((\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e}\times\cdots\times \mathfrak{m}_v^{e})
=\sigma_1^{-\deg \alpha_1}\cdots \sigma_r^{-\deg \alpha_r} + \sigma_1^{-de}\sigma_2^{-\deg \alpha_2}\cdots \sigma_r^{-\deg \alpha_r}.
\end{equation}
\item\label{vanishing term1}
If a tuple $(\alpha_1,\,\dots,\,\alpha_r)$ satisfies the following conditions:
\begin{itemize}
\item $\deg \alpha_2>\cdots> \deg \alpha_r$,
\item $\alpha_2,\,\dots,\,\alpha_r \in A_+$ and
\item $\deg \alpha_1 \leq \deg \alpha_2$ or $\alpha_1 \in A \setminus A_+$,
\end{itemize}
then we have
\begin{equation}
\mu((\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e}\times\cdots\times \mathfrak{m}_v^{e})
=\sigma_1^{-de}\sigma_2^{-\deg \alpha_2}\cdots \sigma_r^{-\deg \alpha_r}.
\end{equation}
\item\label{vanishing term2}
Otherwise, we have
\begin{equation}
\mu((\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e}\times\cdots\times \mathfrak{m}_v^{e})
=0.
\end{equation}
\end{enumerate}
\end{prop}
\begin{proof}
For $i_1,\,\dots,\,i_r \in \mathbb{N}$, let $X_{i_1,\,\dots,\,i_r}$ be the set
\begin{equation}
\{(n_1,\,\dots,\,n_r) \in A_+^r \mid \deg n_j =i_j,\, n_j \equiv \alpha_j \mod \mathfrak{m}_v^e \text{ for }j=1,\,\dots,r\}.
\end{equation}
Now we can write
\begin{equation}
\mu((\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e}\times\cdots\times \mathfrak{m}_v^{e})
=\sum_{i_1>\cdots>i_r} \left(\#X_{i_1,\,\dots,\,i_r} \right) \sigma_1^{-i_1}\,\dots,\,\sigma_r^{-i_r}.
\end{equation}
Let us suppose $i_1>de$. If $(n_1,\,\dots,\,n_r) \in X_{i_1,\,\dots,\,i_r}$, then we can see that $(n_1+fv^e,\,\dots,\,n_r) \in X_{i_1,\,\dots,\,i_r}$ for all $f \in \mathbb{F}_q$, so we have $q|(\#X_{i_1,\,\dots,\,i_r})$. Therefore we assume $i_1\leq de$ in what follows.
In the case (1), we have
\begin{align}
X_{\deg \alpha_1,\,\dots,\,\deg\alpha_r}&=\{ (\alpha_1,\,\dots,\,\alpha_r) \},\\
X_{de,\,\deg\alpha_2,\dots,\,\deg\alpha_r}&=\{ (\alpha_1+v^e,\,\alpha_2,\,\dots,\,\alpha_r) \}
\end{align}
and $X_{i_1,\,\dots,\,i_r}=\emptyset$ for other $i_1>\cdots >i_r$.
In the case (2), we have
\begin{equation}
X_{de,\,\deg\alpha_2,\dots,\,\deg\alpha_r}=\{ (\alpha_1+v^e,\,\alpha_2,\,\dots,\,\alpha_r) \}
\end{equation}
and $X_{i_1,\,\dots,\,i_r}=\emptyset$ for other $i_1>\cdots >i_r$. In the case (3), $X_{i_1,\,\dots,\,i_r}=\emptyset$ for any $i_1>\cdots >i_r$. Hence the result follows.
\end{proof}
In the same way, we can calculate the measure $\mu^\star$ as follows:
\begin{prop}
Let us fix $e \in \mathbb{N}$, take $\sigma_1,\,\dots,\,\sigma_r \in A_v^*$ and take $\alpha_1,\,\dots,\,\alpha_r \in A$ such that $\deg \alpha_1,\,\dots,\, \deg \alpha_r < \deg v^e =de$. We say that a tuple $(\alpha_1,\,\dots,\,\alpha_r)$ satisfies the condition $C_i$ if $\deg \alpha_i \geq \cdots\geq \deg \alpha_r$ and $\alpha_i,\,\dots,\,\alpha_r \in A_+$. If $i \in \mathbb{N}$ is the smallest such that the condition $C_i$ is satisfied, then
\begin{equation*}
\mu^{\star}((\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e}\times\cdots\times \mathfrak{m}_v^{e})=\sum_{j=i}^{r}\sigma_1^{-de}\cdots \sigma_{j-1}^{-de }\sigma_j^{-\deg \alpha_j}\cdots\sigma_r^{-\deg \alpha_r}.
\end{equation*}
If $(\alpha_1,\,\dots,\,\alpha_r)$ does not satisfy $C_r$, then $\mu^{\star}((\alpha_1,\,\dots,\,\alpha_r)+\mathfrak{m}_v^{e}\times\cdots\times \mathfrak{m}_v^{e})=0$.
\end{prop}
We can generalize Theorem \ref{SingleIntegralExpression} to multivariable setting as follows:
\begin{theorem}\label{IntegralExpression}
For $m_1,\,\dots,\,m_r \in \mathbb{Z}_{\geq0}$, the special value $\zeta_{\infty}(-m_1,\,\dots,-m_r)$ has the following integral expression:
\begin{equation*}
\zeta_{\infty}(-m_1,\,\dots,-m_r)=\int_{A_v^r}x_1^{m_1}\cdots x_r^{m_r} \mu^{1,\,\dots,\,1},
\end{equation*}
and the special value $\zeta_{\infty}^\star (-m_1,\,\dots,-m_r)$ has the following integral expression:
\begin{equation*}
\zeta_{\infty}^\star(-m_1,\,\dots,-m_r)=\int_{A_v^r}x_1^{m_1}\cdots x_r^{m_r} \mu^{\star\ 1,\,\dots,\,1}.
\end{equation*}
\end{theorem}
\begin{proof}
By definitions, we can calculate Riemann sum as follows:
\begin{align}
&\qquad \quad\int_{A_v^r}x_1^{m_1}\cdots x_r^{m_r} \mu^{1,\,\dots,\,1}\\
&=\lim_{e\to \infty} \left[2 \sum_{\substack{de>\deg n_1> \cdots > \deg n_r \\n_1,\,\dots,\,n_r \in A_+}}n_1^{m_1}\, \cdots\, n_r^{m_r}+\sum_{\substack{de>\deg n_2> \cdots > \deg n_{r}, \\ n_2,\,\dots,\,n_{r} \in A_+,\, de>\deg n_1,\\ \deg n_2 \geq \deg n_{1} \text{ or } n_1 \in A \setminus A_+}} n_1^{m_1}\, \cdots\, n_r^{m_r} \right]\\
&=\lim_{e\to \infty} \left[ \sum_{\substack{de>\deg n_1> \cdots > \deg n_r \\n_1,\,\dots,\,n_r \in A_+}}n_1^{m_1}\, \cdots\, n_r^{m_r}+
\sum_{\substack{de>\deg n_2> \cdots > \deg n_r, \\n_2,\,\dots,\,n_r \in A_+,\\ de>\deg n_1,\, n_1 \in A}} n_1^{m_1}\, \cdots\, n_r^{m_r} \right]\\
&=\zeta_{\infty}(-m_1,\,\dots,-m_r)+\zeta_{\infty}(-m_2,\,\dots,-m_r)\zeta_{\infty}(-m_1) \sum_{f\in\mathbb{F}_q^\times} f^{m_1}\\
&=\zeta_{\infty}(-m_1,\,\dots,-m_r).
\end{align}
To prove the last equality, we used the facts that $\zeta_{\infty}(-m_1)=0$ if $(q-1)|m_1$ and that $\sum_{f\in\mathbb{F}_q^\times} f^{m_1}=0$ if $(q-1)\nmid m_1$.
The second assertion can be confirmed as follows:
\begin{align}
\begin{autobreak}
\int_{A_v^r}x_1^{m_1}\cdots x_r^{m_r} \mu^{\star\ 1,\,\dots,\,1}
=\zeta_{\infty}^\star(-m_1,\,\dots,-m_r)
\ +\zeta_{\infty}^\star(-m_2,\,\dots,-s_{r})\zeta_{\infty}^\star(-m_1)\sum_{f_1\in\mathbb{F}_q^\times} f_1^{m_1}
\ +\zeta_{\infty}^\star(-m_2,\,\dots,-m_{r})\zeta_{\infty}^\star(-m_{1})\zeta_{\infty}^\star(-m_2)\sum_{f_{1}\,f_2 \in\mathbb{F}_q^\times} f_{1}^{m_1}f_2^{m_2}
\ +\cdots
\ +\zeta_{\infty}(-m_1)\cdots\zeta_{\infty}(-m_r)\sum_{f_1,\,\dots,\,f_r \in \mathbb{F}_q^\times}f_1^{m_1}\cdots f_r^{m_r}
=\zeta_\infty^\star(-m_1,\,\dots,-m_r).
\end{autobreak}
\end{align}
\end{proof}
Contrary to the case of $\infty$MZFs and $\infty$MZSFs, we can obtain integral expressions of $v$MZFs and of $v$MZSFs not just the expressions of special values:
\begin{theorem}\label{IntegralexpressionofStar}
Let $t_1,\,\dots,\,t_r \in {\mathbb{Z}_{d,\,p}}$ and $\sigma_1,\,\dots,\,\sigma_r \in A_v \setminus \mathfrak{m}_v$. Then we have the following integral expressions of the value $\zeta_v((\sigma_1,\, -t_1),$ $\dots,$ $(\sigma_r,\,-t_r))$ of $v$MZF:
\begin{equation}
\zeta_v((\sigma_1,\, -t_1),\,\dots, (\sigma_r,\,-t_r))=\int_{(A_v \setminus \mathfrak{m}_v)^r}x_1^{t_1}\cdots x_r^{t_r} \mu^{\sigma_1,\,\dots,\,\sigma_r}, \label{eqvintexpression}
\end{equation}
and of the value $\zeta_v^\star((\sigma_1,\, -t_1),$ $\dots,$ $(\sigma_r,\,-t_r))$of $v$MZSFs:
\begin{equation*}
\zeta_v^\star((\sigma_1,\, -t_1),\,\dots, (\sigma_r,\,-t_r))=\int_{(A_v \setminus \mathfrak{m}_v)^r}x_1^{t_1}\cdots x_r^{t_r} \mu^{\star\ \sigma_1,\,\dots,\,\sigma_r}.
\end{equation*}
\end{theorem}
\begin{proof}
For the proof of the assertion, we can assume that $t_1,\,\dots,\,t_r \in \mathbb{N}$ since $-\mathbb{N}$ is dense in ${\mathbb{Z}_{d,\,p}}$. Then our desired equality is obtained as follows:
\begin{align}
&\int_{(A_v \setminus \mathfrak{m}_v)^r}x_1^{t_1}\cdots x_r^{t_r} \mu^{\sigma_1,\,\dots,\,\sigma_r}\\
=&
\begin{autobreak}\lim_{e \to \infty} \left\{ \sum_{\substack{de>\deg n_1> \cdots > \deg n_r \\n_1,\,\dots,\,n_r \in A_+\\(n_1,\,v)=\cdots=(n_r,\,v)=1}} \sigma_1^{-\deg n_1}\cdots \sigma_r^{-\deg n_r} n_1^{t_1}\, \cdots\, n_r^{t_r} \right.
+ \left. \sum_{\substack{de>\deg n_1> \cdots > \deg n_r \\n_1,\,\dots,\,n_r \in A_+\\(n_1,\,v)=\cdots=(n_r,\,v)=1}}\sigma_1^{-de}\sigma_2^{-\deg n_2}\cdots \sigma_{r}^{-\deg n_{r}} n_1^{t_1}\, \cdots\, n_r^{t_r} \right.
+ \left. \sum_{\substack{de>\deg n_2 > \cdots > \deg n_r, \\n_2,\,\dots,\,n_{r-1} \in A_+,\, \deg n_1<de \\ \deg n_2\geq \deg n_1 \text{ or } n_1 \in A \setminus A_+\\(n_1,\,v)=\cdots=(n_r,\,v)=1}}\sigma_1^{-de}\sigma_2^{-\deg n_2}\cdots \sigma_{r}^{-\deg n_{r}} n_1^{t_1}\, \cdots\, n_r^{t_r} \right\}
\end{autobreak}\\
=&\begin{autobreak}\zeta_v((\sigma_1,\, -t_1),\,\dots, (\sigma_r,\,-t_r))+\zeta_v((\sigma_2,\, -t_2),\,\dots, (\sigma_r,\,-t_r))
\cdot\lim_{e\to\infty} \left\{\sigma_1^{-de} \zeta_v(-t_1) \right\}\sum_{f\in\mathbb{F}_q^\times} f^{t_1}=\zeta_v((\sigma_1,\, -t_1),\,\dots, (\sigma_r,\,-t_r)).
\end{autobreak}
\end{align}
The integral expression of $v$MZSVs are confirmed in the same way.
\end{proof}
\subsection{Multiple Kummer-type congruence}\label{Multiple Kummer-type congruence}
We obtain some congruence in Theorem \ref{KummerType} between values of $v$MZFs by the integral expression of Theorem \ref{IntegralexpressionofStar} which are analogous to \cite[Theorem 2.10]{Furusho2017b}
\begin{lemma}\label{Fermat type congruences}
If we take $m,\,l \in \mathbb{Z}$ such that $m \equiv l \mod (q^d-1)q^{(e-1)d}$ for $ e\in \mathbb{N}$, then we have the congruence $x^m \equiv x^l \mod \mathfrak{m}_v^e$ for all $x \in A_v \setminus \mathfrak{m}_v$.
\end{lemma}
This follows from the fact that $\# (A_v/\mathfrak{m}_v^e)^\times=\# (A/\mathfrak{m}_v^e)^\times=(q^d-1)q^{(e-1)d}$.
\begin{theorem}\label{KummerType}
For indices $(m_1,\,\dots,\,m_r) ,\,(l_1,\,\dots,\,l_r) \in \mathbb{Z}^r$ with $m_i \equiv l_i \mod (q^d-1)q^{(e-1)d}$ for all $1\leq i \leq r$, the following congruence hold:
\begin{align}
\zeta_v(m_1,\,\dots,\,m_r) \equiv \zeta_v(l_1,\,\dots,\,l_r) \mod \mathfrak{m}_v^e,\\
\zeta_v^\star(m_1,\,\dots,\,m_r) \equiv \zeta_v^\star(l_1,\,\dots,\,l_r) \mod \mathfrak{m}_v^e.
\end{align}
\end{theorem}
\begin{proof}
By Lemma \ref{Fermat type congruences}, the function $x_1^{m_1}\cdots x_r^{m_r}-x_1^{l_1}\cdots x_r^{l_r}$ on $(A_v \setminus \mathfrak{m}_v)^r$ has values in $\mathfrak{m}_v^e$. Since $\mu$ is $A_v$-valued, we have the formula
\begin{equation*}
\int_{(A_v \setminus \mathfrak{m}_v)^r} (x_1^{m_1}\cdots x_r^{m_r}-x_1^{l_1}\cdots x_r^{l_r}) \mu \in \mathfrak{m}_v^e,
\end{equation*}
which yields the desired congruence $\zeta_v(m_1,\,\dots,\,m_r) \equiv \zeta_v(l_1,\,\dots,\,l_r)$. The congruence for special values of the $v$MZSV is proven by the same manner.
\end{proof}
\begin{example}
(1). For $m,\,l \in \mathbb{Z}$ such that $m \equiv l \mod (q^d-1)q^{(e-1)d}$, we have
\begin{equation}
\zeta_v(m)\equiv \zeta_v(l) \mod m_v^e.
\end{equation}
If we take $m_1,\,m_2,\,l_1,\,l_2 \in \mathbb{Z}$ such that $m_1 \equiv l_1 \mod (q^d-1)q^{(e-1)d}$ and $m_2 \equiv l_2 \mod (q^d-1)q^{(e-1)d}$, then we have
\begin{equation}\zeta_v(m_1,\,m_2)\equiv \zeta_v(l_1,\,l_2) \text{ and } \zeta_v^\star(m_1,\,m_2)\equiv \zeta_v^\star(l_1.\,l_2) \mod m_v^e. \end{equation}
\end{example}
\section{Inductive formulae for special values of multiple zeta functions at negative integers}\label{SectionINductiveFormulae}
In subsection \ref{4.1}, we obtain recursive formula for special values of $\infty$MZFs at negative integers, which enable us to write down the special values in terms of those with simpler arguments. We also obtain similar formulae for the special values of $v$MZFs in subsection \ref{4.2}. Comparing these two results, we obtain a relationship between special values of $\infty$MZFs and those of $v$MZFs in subsection \ref{4.3}.
\subsection{Inductive formulae for special values of $\infty$MZFs at negative integers}\label{4.1}
We obtain recursive formula for special values of $\infty$MZFs at negative integers, which enable us to write down the special values in terms of those with simpler arguments. Similar results for star versions are also shown.
For $d\geq1$ and for $m_1,\,\dots,\,m_r \in \mathbb{Z}$, we write
\begin{align}
\zeta_{<d}(m_1,\,\dots,\,m_r)\coloneqq \sum_{d>i_1>\cdots>i_r\geq0}S_{i_1}(m_1)\cdots S_{i_r}(m_r),\\
\zeta_{< d}^\star(m_1,\,\dots,\,m_r)\coloneqq \sum_{d > i_1 \geq \cdots \geq i_r\geq0}S_{i_1}(m_1)\cdots S_{i_r}(m_r).
\end{align}
If $r\geq d$ we put $\zeta_{<d}(m_1,\,\dots,\,m_r)\coloneqq 0$.
Using these notations, we can write down special values of $\infty$MZFs and $\infty$MZSFs at negative integers in terms of simpler ones:
\begin{theorem}\label{recursiveformulae}
Let $m_1\,\dots,\,m_r \in \mathbb{N}_{\geq0}$ and let $v$ be a monic irreducible polynomial with degree $d$. Then we have the equalities
\begin{align}
& \quad \zeta_{\infty}(-m_1,\,\dots,\,-m_r)\label{inftyInductive}
\\
&=\sum_{l=0}^{r}\Biggr\{(-1)^l \zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}} v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots\\
&\quad \cdots \zeta_{<d}(a_l-m_l)\cdot \dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l)\Biggr \},
\end{align}
and
\begin{align}
& \quad \zeta_{\infty}^\star(-m_1,\,\dots,\,-m_r) \label{inftystarrec}\\
&=\sum_{l=0}^{r}\Biggr\{ (-1)^l \zeta_{< d}^\star (-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots\\
&\quad \cdots \zeta_{<d}(a_l-m_l)\cdot \dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}^\star(-a_1,\,\dots,\,-a_l)\Biggr \}, \label{inftystarInductive}
\end{align}
If $m_j=0$, then the sums for $0\leq a_j <m_j$ are interpreted as empty sums. The summands with $l=0$ are $\zeta_{< d}^\star (-m_{1},\,\dots,\,-m_r)$ and $\zeta_{< d}^\star (-m_{1},\,\dots,\,-m_r)$, respectively.
\end{theorem}
The right hand sides of \eqref{inftyInductive} and \eqref{inftystarrec} do not depend on the choice of monic prime element $v$ of degree $d$.
\begin{proof}
We have
\begin{align}
&\zeta_{\infty}(-m_1,\,\dots,\,-m_r)\\
=&\sum_{l=0}^{r} \left[\sum_{ i_1> \cdots>i_l \geq d > i_{l+1}>\cdots>i_r\geq0 } \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\(\text{for } r \geq j \geq 1)}}n_1^{m_1}\cdots n_r^{m_r}\right\}\right]\\
=&\sum_{l=0}^{r} \left[\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r) \sum_{i_1>\cdots>i_l\geq d} \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\(\text{for }1\leq j \leq l)}}n_1^{m_1}\cdots n_l^{m_l}\right\}\right]\\
\intertext{putting $n_j=vh_j+\alpha_j$,}
=&\sum_{l=0}^{r} \left[\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r) \sum_{i_1>\cdots>i_l\geq d} \left\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ \deg \alpha_j<d \\(\text{for }1\leq j \leq l)}}(vh_1+\alpha_1)^{m_1}\cdots (vh_l+\alpha_l)^{m_l}\right\}\right]\\
\intertext{by expanding each $(vh_j+\alpha_j)^{m_j}$,}
=&\sum_{l=0}^{r} \Biggr[\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r) \sum_{i_1>\cdots>i_l\geq d} \Biggr\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ \deg \alpha_j<d \\(\text{for }1\leq j \leq l)}}
\Biggr(\sum_{\substack{0\leq a_j\leq m_j\\(\text{for }1\leq j \leq l)}} (vh_1)^{a_1}\cdots \\ &\cdots(vh_l)^{a_l}\cdot \alpha_1^{m_1-a_1}\cdots \alpha_l^{m_l-a_l}\cdot\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l}
\Biggr)\Biggr\}\Biggr]\\
\intertext{since $\sum_{\alpha_j}\alpha_j^0=\sum_{\alpha_j}1=q^{d}=0$,}
=&\sum_{l=0}^{r} \Biggr[\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r) \sum_{i_1>\cdots>i_l\geq d} \Biggr\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ \deg \alpha_j<d \\(\text{for }1\leq j \leq l)}}
\Biggr(\sum_{\substack{0\leq a_j< m_j\\(\text{for }1\leq j \leq l)}} (vh_1)^{a_1}\cdots \\ &\cdots(vh_l)^{a_l}\cdot \alpha_1^{m_1-a_1}\cdots \alpha_l^{m_l-a_l}\cdot\dbinom{m_1}{a_1}\cdots\dbinom{m_l}{a_l}
\Biggr)\Biggr\}\Biggr]\\
=&\sum_{l=0}^{r}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\\deg \alpha_j<d\\(\text{for }1\leq j \leq l)}}v^{a_1+\cdots+a_l} \alpha_1^{m_1-a_1}\cdots \alpha_l^{m_l-a_l}\cdot
\dbinom{m_1}{a_1}\cdots \\
&\,\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l)\Biggr \} \\
=&\sum_{l=0}^{r}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\f_j \in \mathbb{F}_q^\times \\(\text{for }1\leq j \leq l)}}v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots \zeta_{<d}(a_l-m_l)\cdot \\
&\qquad \quad \cdot f_1^{m_1-a_1}\cdots f_l^{m_l-a_l} \cdot \dbinom{m_1}{a_1}\,\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l)\Biggr \} \\
=&\sum_{l=0}^{r}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1)\cdots \\
&\qquad \quad \cdots \zeta_{<d}(a_l-m_l)\cdot \dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l)\Biggr \},
\end{align}
here, the last equality follows from the fact that $\sum_{f\in\mathbb{F}_q^\times} f^{m}=0$ for $m$ not divisible by $(q-1)$.
The second formula can be confirmed as follows:
\begin{align}
& \quad \ \zeta_{\infty}^\star(-m_1,\,\dots,\,-m_r)\\
&=\sum_{l=0}^{r} \left[\sum_{ i_1\geq \cdots \geq i_l \geq d > i_{l+1}\geq \cdots \geq i_r\geq0 } \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\(\text{for } r \geq j \geq 1)}}n_1^{m_1}\cdots n_r^{m_r}\right\}\right]\\
&=\sum_{l=0}^{r} \left[\zeta_{< d}^\star (-m_{l+1},\,\dots,\,-m_r) \sum_{i_1 \geq \cdots \geq i_l\geq d} \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\(\text{for }1\leq j \leq l)}}n_1^{m_1}\cdots n_l^{m_l}\right\}\right]\\
\intertext{putting $n_j=vh_j+\alpha_j$,}
&=\sum_{l=0}^{r} \left[\zeta_{< d}^\star (-m_{l+1},\dots,-m_r) \sum_{i_1 \geq \cdots \geq i_l\geq d} \left\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ \deg \alpha_j<d \\(\text{for }1\leq j \leq l)}}(vh_1+\alpha_1)^{m_1}\cdots (vh_l+\alpha_l)^{m_l}\right\}\right]\\
&=\sum_{l=0}^{r} \Biggr[\zeta_{< d}^\star (-m_{l+1},\,\dots,\,-m_r) \sum_{i_1 \geq \cdots \geq i_l\geq d} \Biggr\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ \deg \alpha_j<d \\(\text{for }1\leq j \leq l)}}
\Biggr(\sum_{\substack{0\leq a_j\leq m_j\\(\text{for }1\leq j \leq l)}} (vh_1)^{a_1}\cdots \\ & \quad \cdots(vh_l)^{a_l}\cdot \alpha_1^{m_1-a_1}\cdots \alpha_l^{m_l-a_l}\cdot \dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l}
\Biggr)\Biggr\}\Biggr]\\
&=\sum_{l=0}^{r}\Biggr\{\zeta_{< d}^\star (-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\\deg \alpha_j<d\\(\text{for }1\leq j \leq l)}}v^{a_1+\cdots+a_l} \alpha_1^{m_1-a_1}\cdots \alpha_l^{m_l-a_l}\cdot
\dbinom{m_1}{a_1}\cdots \\
&\,\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}^\star (-a_1,\,\dots,\,-a_l)\Biggr \} \\
&=\sum_{l=0}^{r}\Biggr\{\zeta_{< d}^\star (-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\f_j \in \mathbb{F}_q^\times \\(\text{for }1\leq j \leq l)}}v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots \\ &\cdots \zeta_{<d}(a_l-m_l)\cdot f_1^{m_1-a_1}\cdots f_l^{m_l-a_l} \cdot
\dbinom{m_1}{a_1}\,\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}^\star (-a_1,\,\dots,\,-a_l)\Biggr \} \\
&=\sum_{l=0}^{r}\Biggr\{\zeta_{< d}^\star (-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j <m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots \\
&\quad \ \cdots \zeta_{<d}(a_l-m_l)\cdot
\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}^\star(-a_1,\,\dots,\,-a_l)\Biggr \}.
\end{align}
\end{proof}
\begin{example}
(1). For $m \in \mathbb{Z}_{\geq0}$, we have
\begin{align}
\zeta_{\infty}(-m)&=\zeta_{<d}(-m)-\sum_{\substack{0\leq a<m \\(q-1)|(m-a)}}v^{a}\zeta_{<d}(a-m)\binom{m}{a}\zeta_{\infty}(-a),\\
\end{align}
(This appeared in \cite[Theorem 5.6]{Goss1979} and \cite[\S 5.3]{ThakurBook}.)
(2). For $m_1,\,m_2\in \mathbb{Z}_{\geq0}$, we have
\begin{align}
&\quad \ \zeta_\infty(-m_1,\,-m_2)\\
&=\zeta_{<d}(-m_1,\,-m_2)-\zeta_{<d}(-m_2)\sum_{\substack{0\leq a_1<m_1 \\(q-1)|(m_1-a_1)}}v^{a_1}\zeta_{<d}(a_1-m_1)\binom{m_1}{a_1}\zeta_{\infty}(-a_1)\\
&\quad \ +\sum_{\substack{0\leq a_1<m_1 \\0\leq a_2<m_2\\(q-1)|(m_1-a_1)\\(q-1)| (m_2-a_2)}}v^{a_1+a_2}\zeta_{<d}(a_1-m_1)\zeta_{<d}(a_2-m_2)\binom{m_1}{a_1}\binom{m_2}{a_2}\zeta_{\infty}(-a_1,\,-a_2),
\end{align}
and
\begin{align}
&\quad \ \zeta_\infty^\star(-m_1,\,-m_2)\\
&=\zeta_{<d}^\star(-m_1,\,-m_2)-\zeta_{<d}^\star(-m_2)\sum_{\substack{0\leq a_1<m_1 \\(q-1)|(m_1-a_1)}}v^{a_1}\zeta_{<d}(a_1-m_1)\binom{m_1}{a_1}\zeta_{\infty}(-a_1)\\
&\quad \ +\sum_{\substack{0\leq a_1<m_1 \\0\leq a_2<m_2\\(q-1)|(m_1-a_1)\\(q-1)| (m_2-a_2)}}v^{a_1+a_2}\zeta_{<d}(a_1-m_1)\zeta_{<d}(a_2-m_2)\binom{m_1}{a_1}\binom{m_2}{a_2}\zeta_{\infty}^\star(-a_1,\,-a_2).
\end{align}
(3). For $m_1,\,\dots,\,m_{r-1} \in \mathbb{Z}_{\geq 0}$, we have
\begin{align}
&\quad \ \zeta_\infty(-m_1,\,\dots,\,-m_{r-1},\,0) \label{depthdown}
\\&=\sum_{l=0}^{r-1}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots\\
&\quad \cdots \zeta_{<d}(a_l-m_l)\cdot
\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l) \Biggr \} ,
\end{align}
(here $l$ runs from $0$ to $r-1$, not to $r$).
\end{example}
\begin{remarks}\label{reminduction}
Since $S_i(0)=0$ for $i \geq 1$, we have $\zeta_\infty(-m_1,\,\dots,\,-m_r)=0$ if one of $m_1,\,\dots,\,m_{r-1}$ is zero for $m_1,\,\dots,\,m_r \in \mathbb{Z}_{\geq0}$. Combining this with $\eqref{depthdown}$, we can see that the special value $\zeta_\infty(-m_1,\,\dots,\,-m_r)$ can be written down $\zeta_\infty(-l_1,\,\dots,\,-l_{r^\prime})$'s with $r^\prime<r$ or $l_j<m_j$ for some $1\leq j\leq r$.
\end{remarks}
\subsection{The formulae for special values of $v$MZFs at negative integers}\label{4.2}
We show that special values of $v$MZFs at negative integers can be written in terms of those of $\infty$MZFs. Similar results for star versions is also shown, i.e. we write down special values of $v$MZSFs at negative integers in terms of those of $\infty$MZSFs.
\begin{theorem}\label{recursiveformulaev}
Let $m_1\,\dots,\,m_r \in \mathbb{N}_{\geq0}$ and let $v$ be a monic irreducible polynomial with degree $d$. Then we have the equalities
\begin{align}
& \quad \zeta_v(-m_1,\,\dots,\,-m_r)\label{vInductive} \\
&=\sum_{l=0}^{r}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots\\
&\quad \cdots \zeta_{<d}(a_l-m_l)\cdot
\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l) \Biggr \} .
\end{align}
and
\begin{align}
& \qquad \zeta_v^\star(-m_1,\,\dots,\,-m_r)\\
&=\sum_{l=0}^{r}\Biggr\{\zeta_{< d}^\star (-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}(-1)^l v^{a_1+\cdots+a_l} \cdot\zeta_{<d}(a_1-m_1) \cdots \\
&\quad \cdots \zeta_{<d}(a_l-m_l)\cdot
\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}^\star(-a_1,\,\dots,\,-a_l) \Biggr \}. \label{vstarInductive}
\end{align}
\end{theorem}
Let us recall that these special values are in $k$ (Remark \ref{remarkink}). The difference between Theorems \ref{recursiveformulae} and \ref{recursiveformulaev} appears in the ranges which $a_j$'s run through.
\begin{proof}
We have
\begin{align}
&\zeta_v(-m_1,\,\dots,\,m_l)
=\sum_{l=0}^{r} \left[\sum_{ i_1> \cdots>i_l \geq d > i_{l+1}>\cdots>i_r\geq0 } \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\ (n_j,\,v)=1\\(\text{for } r \geq j \geq 1)}}n_1^{m_1}\cdots n_r^{m_l}\right\}\right]\\
=&\sum_{l=0}^{r} \left[\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r) \sum_{i_1>\cdots>i_l\geq d} \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\ (n_j,\,v)=1\\(\text{for }1\leq j \leq l)}}n_1^{m_1}\cdots n_l^{m_l}\right\}\right]\\
\intertext{here, let us recall that $d$ is the degree of $v$. Putting $n_j=vh_j+\alpha_j$, we have}
=&\sum_{l=0}^{r} \left[\zeta_{<d}(-m_{l+1},\dots,-m_r) \sum_{i_1>\cdots>i_l\geq d} \left\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ \deg \alpha_j<d, \,\alpha_j \in A \setminus \{0\} \\(\text{for }1\leq j \leq l)}}(vh_1+\alpha_1)^{m_1}\cdots (vh_l+\alpha_l)^{m_l}\right\}\right]\\
\intertext{(The ranges which $\alpha_j$'s run through are different from those in the proof of Theorem \ref{recursiveformulae}.)}
=&\sum_{l=0}^{r} \Biggr[\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r) \sum_{i_1>\cdots>i_l\geq d} \Biggr\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ 0 \leq \deg \alpha_j<d \\(\text{for }1\leq j \leq l)}}
\Biggr(\sum_{\substack{0\leq a_j\leq m_j\\(\text{for }1\leq j \leq l)}} (vh_1)^{a_1}\cdots \\ &\cdots(vh_l)^{a_l}\cdot \alpha_1^{m_1-a_1}\cdots \alpha_l^{m_l-a_l}\cdot \dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l}
\Biggr)\Biggr\}\Biggr]\\
=&\sum_{l=0}^{r}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\f_j \in \mathbb{F}_q^\times \\(\text{for }1\leq j \leq l)}}v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots\zeta_{<d}(a_l-m_l) \\
& \cdot f_1^{m_1-a_1} \cdots f_l^{m_l-a_l} \cdot
\dbinom{m_1}{a_1}\,\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l)\Biggr \} \\
=&\sum_{l=0}^{r}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots \\
&\qquad \qquad \cdots \zeta_{<d}(a_l-m_l)\cdot \dbinom{m_1}{a_1}\cdots\dbinom{m_l}{a_l} \cdot \zeta_{\infty}(-a_1,\,\dots,\,-a_l)\Biggr \}.
\end{align}
The second formula can be confirmed as follows:
\begin{align}
&\zeta_v^\star(-m_1,\,\dots,\,m_l)
=\sum_{l=0}^{r} \left[\sum_{ i_1\geq \cdots \geq i_l \geq d > i_{l+1} \geq \cdots \geq i_r\geq0 } \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\ (n_j,\,v)=1\\(\text{for } r \geq j \geq 1)}}n_1^{m_1}\cdots n_r^{m_l}\right\}\right]\\
=&\sum_{l=0}^{r} \left[\zeta_{< d}^\star(-m_{l+1},\,\dots,\,-m_r) \sum_{i_1 \geq \cdots \geq i_l\geq d} \left\{\sum_{\substack{\deg n_j= i_j\\n_j \in A_+\\ (n_j,\,v)=1\\(\text{for }1\leq j \leq l)}}n_1^{m_1}\cdots n_l^{m_l}\right\}\right]\\
=&\sum_{l=0}^{r} \left[\zeta_{< d}^\star(-m_{l+1},\dots,-m_r) \sum_{i_1 \geq \cdots \geq i_l\geq d} \left\{\sum_{\substack{\deg h_j= i_j-d\\h_j \in A_+\\ \deg \alpha_j<d, \,\alpha_j \neq 0 \\(\text{for }1\leq j \leq l)}}(vh_1+\alpha_1)^{m_1}\cdots (vh_l+\alpha_l)^{m_l}\right\}\right]\\
=&\sum_{l=0}^{r}\Biggr\{\zeta_{< d}^\star(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\f_j \in \mathbb{F}_q^\times \\(\text{for }1\leq j \leq l)}}v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots \\
&\quad\cdots \zeta_{<d}(a_l-m_l)\cdot f_1^{m_1-a_1} \cdots f_l^{m_l-a_l}\cdot
\dbinom{m_1}{a_1}\,\cdots \dbinom{m_l}{a_l}\cdot \zeta_{\infty}^\star(-a_1,\,\dots,\,-a_l)\Biggr \} \\
=&\sum_{l=0}^{r}\Biggr\{\zeta_{< d}^\star(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\(q-1)|(m_j-a_j)\\(\text{for }1\leq j \leq l)}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots \\
&\qquad \cdots \zeta_{<d}(a_l-m_l)\cdot
\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l} \cdot \zeta_{\infty}^\star(-a_1,\,\dots,\,-a_l)\Biggr \}.
\end{align}
\end{proof}
\begin{example}
(1). For $m \in \mathbb{Z}_{\geq0}$, we have
\begin{align}
&\quad\zeta_{v}(-m)=\zeta_{<d}(-m)-\sum_{\substack{0\leq a\leq m \\(q-1)|(m-a)}}v^{a}\zeta_{<d}(a-m)\binom{m}{a}\zeta_{\infty}(-a).
\end{align}
(2). For $m_1,\,m_2\in \mathbb{Z}_{\geq0}$, we have
\begin{align}
&\quad \zeta_v(-m_1,\,-m_2)\\
&=\zeta_{<d}(-m_1,\,-m_2)-\zeta_{<d}(-m_2)\sum_{\substack{0\leq a_1 \leq m_1 \\(q-1)|(m_1-a_1)}}v^{a_1}\zeta_{<d}(a_1-m_1)\binom{m_1}{a_1}\zeta_{\infty}(-a_1)\\
&\quad +\sum_{\substack{0\leq a_1 \leq m_1 \\0\leq a_2 \leq m_2\\(q-1)|(m_1-a_1)\\(q-1)| (m_2-a_2)}}v^{a_1+a_2}\zeta_{<d}(a_1-m_1)\zeta_{<d}(a_2-m_2)\binom{m_1}{a_1}\binom{m_2}{a_2}\zeta_{\infty}(-a_1,\,-a_2),
\end{align}
and
\begin{align}
&\quad \zeta_v^\star(-m_1,\,-m_2)\\
&=\zeta_{< d}^\star(-m_1,\,-m_2)-\zeta_{< d}^\star(-m_2)\sum_{\substack{0\leq a_1 \leq m_1 \\(q-1)|(m_1-a_1)}}v^{a_1}\zeta_{<d}(a_1-m_1)\binom{m_1}{a_1}\zeta_{\infty}(-a_1)\\
&\quad +\sum_{\substack{0\leq a_1 \leq m_1 \\0\leq a_2 \leq m_2\\(q-1)|(m_1-a_1)\\(q-1)| (m_2-a_2)}}v^{a_1+a_2}\zeta_{<d}(a_1-m_1)\zeta_{<d}(a_2-m_2)\binom{m_1}{a_1}\binom{m_2}{a_2}\zeta_{\infty}^\star(-a_1,\,-a_2).
\end{align}
\end{example}
\subsection{Relationship between $\infty$MZFs and $v$MZFs} \label{4.3}
In the case of characteristic $0$, there is a formula connecting ``special values'' of multiple zeta functions at negative integers with those of $p$-adic $L$-functions shown in \cite[Remark 2.2]{Furusho2017b}.
Using the results of previous subsections, we obtain similar relationship between special values of $\infty$MZFs and $v$MZFs at negative integers.
\begin{theorem}\label{recursiveformulaeCor}
Special values of $v$MZFs at negative integers can be written in terms of those of $\infty$MZFs by the following identity which holds for $m_1\,\dots,\,m_r \in \mathbb{N}_{\geq0}$:
\begin{align}
\quad &\quad \zeta_v(-m_1,\,\dots,\,-m_r)-\zeta_{\infty}(-m_1,\,\dots,\,-m_r) \label{EqInductiveCor}\\
&=\sum_{l=0}^{r}\Biggr\{\zeta_{<d}(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\(q-1)|(m_j-a_j)\\(1\leq j \leq l)\\a_j=m_j \text{ for some $j$}}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots\\
&\ \cdots \zeta_{<d}(a_l-m_l)\cdot
\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l}\cdot\zeta_{\infty}(-a_1,\,\dots,\,-a_l) \Biggr \} .
\end{align}
Conversely, special values of $\infty$MZFs at negative integers can be written in terms of those of $v$MZFs.
\end{theorem}
\begin{proof}
Comparing the equations \eqref{inftyInductive} and \eqref{vInductive}, we obtain \eqref{EqInductiveCor}. As \eqref{EqInductiveCor} implies
\begin{equation}
\zeta_v(-m)=(1-v^{m})\zeta_{\infty}(-m), \ (m \in \mathbb{N}), \label{Dep1interpolate}
\end{equation}
we can write down special values of $\infty$MZFs at negative integers in terms of those of $v$MZFs if $r=1$, and we can reduced to the case $r=1$ by Remark \ref{reminduction}.
\end{proof}
By virtue of Theorem \ref{recursiveformulaeCor}, we can write down special values of $\infty$MZFs at negative integers in terms of those of $v$MZFs, and vice versa.
\begin{example}
In the case of $r=1$, the corollary above yields the well-known formula \eqref{Dep1interpolate}.
If $r=2$, we can write down special values of the $v$MZF in terms of special values of $\infty$MZFs as follows:
\begin{align}
\quad \ \zeta_v(-m_1,\,-m_2)&=(1+v^{m_1+m_2})\zeta_\infty(-m_1,\,-m_2)-\zeta_{<d}(-m_2)v^{m_1}\zeta_\infty(-m_1) \label{dep2ex1}\\
&\, +\sum_{\substack{0\leq a_1 <m_1\\(q-1)|(m_1-a_1)}}v^{a_1+m_2}\zeta_{<d}(a_1-m_1)\dbinom{m_1}{a_1}\zeta_\infty(-a_1,\, -m_2)\\
&\, +\sum_{\substack{0\leq a_2 < m_2\\(q-1)|(m_2-a_2)}}v^{m_1+a_2}\zeta_{<d}(a_2-m_2)\dbinom{m_2}{a_2}\zeta_\infty(-m_1,\, -a_2).
\end{align}
Conversely, recursive applications of \eqref{dep2ex1} induces the following formula, which expresses special values of $\infty$MZFs at negative integers in terms of those of $v$MZFs, for $m_1,\,m_1^\prime \in \mathbb{Z}_{\geq0}$,
\begin{align*}
& \quad \ v^{m_1+m_1^\prime}\zeta_\infty(-m_1,\,-m_1^\prime)\\
&=\sum_{\substack{l,\,l^\prime \geq 1\\m_1>m_2>\cdots\,>m_l\geq 0\\m_1^\prime>m_2^\prime>\cdots\,>m_{l^\prime}^\prime \geq0 \\ (q-1)|(m_{i-1}-m_i)\\(q-1)|(m_{i-1}^\prime-m_i^\prime)}}
\left\{ \zeta_v(-m_l,\,-m_{l^\prime}^\prime)+ \zeta_{<d}(-m_{l^\prime}^\prime) \frac{v^{m_l}}{1-v^{m_l}} \zeta_v(-m_l)\right\}\\
&\quad \cdot \Biggl[ -\sum_{ [\dots,(j_i,\,j_i^\prime),\dots] \in \operatorname{S}_{l,\,l^\prime}} \prod_{i=1}^{l+l^\prime-1}\Biggl\{\zeta_{<d}(m_{j_{i-1}}-m_{j_i})\zeta_{<d}(m_{j_{i-1}^\prime}^\prime-m_{j_i^\prime}^\prime) \\ &\quad \quad \cdot\frac{-v^{m_{j_i}+m_{{j_i^\prime}}^\prime}}{1+v^{m_{j_i}+m_{{j_i^\prime}}^\prime}}\dbinom{m_{j_{i-1}}}{m_{j_i}}\dbinom{m_{{j_{i-1}^\prime}}^\prime}{m_{{j_i^\prime}}^\prime}\ \Biggr\} \Biggr]\\
&=-\sum_{\substack{l,\,l^\prime \geq 1\\ [\dots,(j_i,\,j_i^\prime),\dots] \in \operatorname{S}_{l,\,l^\prime}}}\sum_{\substack{m_1>m_2>\cdots\,>m_l\geq 0\\m_1^\prime>m_2^\prime>\cdots\,>m_{l^\prime}^\prime \geq0 \\ (q-1)|(m_{i-1}-m_i),(m_{i-1}^\prime-m_i^\prime)}}
\Biggl[ \prod_{i=1}^{l+l^\prime-1}\Biggl\{\zeta_{<d}(m_{j_{i-1}}-m_{j_i}) \\
&\quad \quad \cdot\zeta_{<d}(m_{j_{i-1}^\prime}^\prime-m_{j_i^\prime}^\prime)\frac{-v^{m_{j_i}+m_{{j_i^\prime}}^\prime}}{1+v^{m_{j_i}+m_{{j_i^\prime}}^\prime}}\dbinom{m_{j_{i-1}}}{m_{j_i}}\dbinom{m_{{j_{i-1}^\prime}}^\prime}{m_{{j_i^\prime}}^\prime}\ \Biggr\} \Biggr]\\
& \quad \ \cdot \left\{ \zeta_v(-m_l,\,-m_{l^\prime}^\prime)+ \zeta_{<d}(-m_{l^\prime}^\prime) \frac{v^{m_l}}{1-v^{m_l}} \zeta_v(-m_l)\right\}.\\
\end{align*}
Here, we define $\zeta_{<d}(m_{j_0}-m_{j_1})=\zeta_{<d}(m_{j_0}^\prime- m_{j_1}^\prime)=\binom{m_{j_0}}{m_{j_1}}=\binom{m_{j_0}^\prime}{m_{j_1}^\prime} \coloneqq 1$ and define $\operatorname{S}_{l,\,l^\prime}$ to be the set of maximal strictly increasing sequences in $\{1,\,\dots,\,l\} \cdot \{1,\,\dots,\,l ^\prime\}$. In other words, $\operatorname{S}_{l,\,l^\prime}$ is the set of sequences $[(j_1,\,j_1^\prime)$, $\dots$,$(j_{l+l^\prime-1}$, $j_{l+l^\prime-1}^\prime)]$ of length $l+l^\prime-1$ such that $j_1=j_1^\prime=1$, $(j_i,\,j_i^\prime)\neq (j_{i+1},\,j_{i+1}^\prime)$, $j_{i+1}=j_i,\,j_i+1$ and $j_{i+1}^\prime=j_i^\prime,\,j_i^\prime+1$ for all $1 \leq i \leq l+l^\prime-2$ (these conditions imply that $j_{l+l^\prime-1}=l$ and $j_{l+l^\prime-1}^\prime=l^\prime$). For example, $[(1,1),\,(1,2),\,(1,3),\,(2,3),\,\dots,\,(l-1,l^\prime),\,(l,l^\prime)] \in \operatorname{S}_{l,\,l^\prime}$.
\end{example}
By the similar arguments, we obain the following formulae for star-versions:
\begin{theorem}\label{recursiveformulaeCorstar}
For $m_1\,\dots,\,m_r \in \mathbb{N}_{\geq0}$, we have the following identity which enable us to write down special values of $v$MZSFs at negative integers in terms of those of $\infty$MZSFs:
\begin{align}
\quad &\quad \zeta_v^\star(-m_1,\,\dots,\,-m_r)-\zeta_{\infty}^\star(-m_1,\,\dots,\,-m_r) \label{EqInductiveCorstar}\\
&=\sum_{l=0}^{r}\Biggr\{\zeta_{<d}^\star(-m_{l+1},\,\dots,\,-m_r)\sum_{\substack{0\leq a_j \leq m_j\\(q-1)|(m_j-a_j)\\(1\leq j \leq l)\\a_j=m_j \text{ for some $j$}}}(-1)^l v^{a_1+\cdots+a_l} \zeta_{<d}(a_1-m_1) \cdots\\
&\ \cdots \zeta_{<d}(a_l-m_l)\cdot
\dbinom{m_1}{a_1}\cdots \dbinom{m_l}{a_l}\cdot\zeta_{\infty}^\star(-a_1,\,\dots,\,-a_l) \Biggr \} .
\end{align}
Conversely, we can write down special values of $\infty$MZSFs at negative integers in terms of those of $v$MZSFs.
\end{theorem}
Theorem \ref{recursiveformulaeCorstar} enable us to write down special values of $\infty$MZSFs at negative integers in terms of those of $v$MZSFs, and vice versa.
\vskip\baselineskip
\textbf{Acknowledgement}
The author is deeply grateful to his doctoral advisor, Professor H. Furusho for his profound instruction and continuous encouragements. He also grateful to Proffessors R.~Harada, Y.~Mishiba and T.~Ngo~Dac who gave very variable comments on manuscript of this paper.
|
\section{Introduction}
One class of groups that is studied from algebraic, geometric and combinatoric perspective is the class consisting of Artin groups (also known as Artin-Tits groups).
Given a finite simplicial graph $\Gamma$ with the vertex set $V(\Gamma)$, the edge set $E(\Gamma)$ and with an edge-labeling $m\colon E(\Gamma)\to\left\{2,3,4,\ldots\right\}$, the associated \emph{Artin group} $A_\Gamma$ is defined as
$$A_\Gamma:= \langle V(\Gamma)\mid \underbrace{vwv\ldots}_{\substack{m(\left\{v,w\right\})-letters}}=\underbrace{wvw\ldots}_{\substack{m(\left\{v,w\right\})-letters}}\text{ whenever } \left\{v,w\right\}\in E(\Gamma)\rangle\,.$$
The most common examples of Artin groups are braid groups and Artin groups $A_\Gamma$ where $E(\Gamma)=\emptyset$ or $m(E(\Gamma))= \left\{2\right\}$, those are called \emph{right-angled} Artin groups.
\subsection{Intersections of parabolic subgroups}
Given a subset $X\subset V(\Gamma)$ of the vertex set, we write $A_X$ for the subgroup generated by $X$. It was proven by van der Lek in \cite{Vanderlek} that $A_X$ is canonically isomorphic to the Artin group $A_{\langle X\rangle}$ where we denote by $\langle X\rangle$ the induced subgraph of $X$ in $\Gamma$. A group of the form $A_X$ is called a \emph{standard parabolic subgroup} of $A_\Gamma$, and a subgroup conjugate to a standard parabolic subgroup is simply called a \emph{parabolic subgroup}. We see the parabolic subgroups of an Artin group as puzzle pieces of the whole group and we are in particular interested in their intersection behavior. It was proven by van der Lek in \cite{Vanderlek} that the class of standard parabolic subgroups is closed under intersection and it is conjectured that the same result holds for the class consisting of all parabolic subgroups. Here our focus is mainly on parabolic subgroups where the diameter of the defining graph is small.
We say that $X\subset V(\Gamma)$ is \emph{free of infinity} if $\left\{v,w\right\}\in E(\Gamma)$ for all $v,w\in X, v\neq w$.
In this case the subgroup $A_X$ is called \emph{complete standard parabolic subgroup}.
A subgroup conjugate to a complete standard parabolic subgroup is called a \emph{complete parabolic subgroup}.
Usually, it is quite hard to prove that an Artin group $A_\Gamma$ has a specific property, but it is sometimes possible to reduce a conjecture about $A_\Gamma$ to complete standard parabolic subgroups of $A_\Gamma$. The following reduction principle was formulated by Godelle and the second author in \cite{GodelleParisA}.
\begin{principle}
Let $\mathcal{P}$ be a property of a group and let $A_\Gamma$ be an Artin group. If all complete standard parabolic subgroups of $A_\Gamma$ have property $\mathcal{P}$, then $A_\Gamma$ has property $\mathcal{P}$.
\end{principle}
Here we are interested in intersections of parabolic subgroups of $A_\Gamma$ and our aim is to reduce the intersection conjecture of parabolic subgroups to complete standard parabolic subgroups of $A_\Gamma$. Hence we are interested in the following properties of $A_\Gamma$:
\begin{itemize}
\item {\bf Property (Int):}
For each free of infinity subset $Y\subset V(\Gamma)$ and for all parabolic subgroups $P_1,P_2$ of $A_Y$, the intersection $P_1\cap P_2$ is a parabolic subgroup.
\item {\bf Property (Int+):}
For all complete parabolic subgroups $P_1,P_2$ of $A_\Gamma$, the intersection $P_1\cap P_2$ is a parabolic subgroup.
\item {\bf Property (Int++):}
For all parabolic subgroups $P_1,P_2$ of $A_\Gamma$, the intersection $P_1\cap P_2$ is a parabolic subgroup.
\end{itemize}
Unfortunately, we cannot prove that property (Int) implies property (Int++).
However, we can prove that property (Int) implies property (Int+--), which is a property between properties (Int+) and (Int++), and which is defined as follows.
\begin{itemize}
\item {\bf Property (Int+--):}
For each complete parabolic subgroup $P_1$ and for each parabolic subgroup $P_2$ of $A_\Gamma$, the intersection $P_1\cap P_2$ is a parabolic subgroup.
\end{itemize}
It is easily checked that property (Int++) implies property (Int+) and that property (Int+) implies property (Int).
We think that all these properties are actually equivalent and, even more, that they always hold. We show:
\begin{theorem}\label{Int}
Let $A_\Gamma$ be an Artin group. If $A_\Gamma$ has property (Int), then $A_\Gamma$ has property (Int+--).
\end{theorem}
The crucial ingredient of the proof of Theorem \ref{Int} is Bass-Serre theory. If $\Gamma$ is not complete, then $A_\Gamma$ is an amalgam of smaller standard parabolic subgroups and we use the action on the corresponding Bass-Serre tree to show the result of the theorem.
For each Artin group $A_\Gamma$ there is an associated \emph{Coxeter group} $W_\Gamma$. It is obtained by adding the relations $v^2=1$ for all $v\in V(\Gamma)$. Hence, the Coxeter group $W_\Gamma$ associated to $A_\Gamma$ is given by the following presentation
$$W_\Gamma:= \langle V(\Gamma)\mid v^2=1, (vw)^{m(\left\{v,w\right\})}=1\text{ for all } v\in V(\Gamma), \left\{v,w\right\}\in E(\Gamma)\rangle\,.$$
An Artin group is of \emph{spherical type} if the associated Coxeter group is finite and an Artin group is of \emph{FC-type} if all complete standard parabolic subgroups are of spherical type. It was proven by Cumplido et al. in \cite{CGG} that intersections of parabolic subgroups in a finite type Artin group are parabolic, therefore as an immediately corollary we obtain the following result.
\begin{corollary}
Let $A_\Gamma$ be an Artin group of FC-type and $P_1, P_2$ be two parabolic subgroups. If $P_1$ is complete, then $P_1\cap P_2$ is parabolic.
\end{corollary}
This is a generalisation of Theorem 3.1 in \cite{MorrisWright}, which states that the intersection of two complete parabolic subgroups of an Artin group of FC-type is parabolic.
\subsection{Automatic continuity}
In a remarkable article \cite{Dudley} Dudley was interested in the relation between locally compact Hausdorff groups and free (abelian) groups. Using a special length function on the target groups he showed that any group homomorphism from a locally compact Hausdorff group into a free (abelian) group is continuous. Inspired by this result Conner and Corson defined the notion of $lcH$-slenderness \cite{ConnerCorson}. A discrete group $G$ is called \emph{lcH-slender} if any group homomorphism from a locally compact Hausdorff group into $G$ is continuous. Our focus here is on continuity of group homomorphisms from locally compact Hausdorff groups into Artin groups. Many types of Artin groups are known to be $lcH$-slender such as right-angled Artin groups \cite{KramerVarghese}, \cite{CorsonKazachkov}, \cite{MoellerVarghese}, Artin groups of spherical type \cite{KramerVarghese} and more generally Artin groups of FC-type \cite{KeppelerMoellerVarghese}. We conjecture:
\begin{conjecture}
All Artin groups are $lcH$-slender.
\end{conjecture}
Since automatic continuity of group homomorphisms from locally compact Hausdorff groups into ``geometric'' groups and fixed point properties of these geometric groups are strongly connected, see \cite{MoellerVarghese}, we are interested in fixed point properties of subgroups of Artin groups, in particular in properties ${\rm F}\mathcal{A}'$ and ${\rm F}\mathcal{C}'$.
Recall, a group $G$ is called an ${\rm F}\mathcal{A}'$-group if any simplicial action of $G$ on a tree without inversion is locally elliptic, i.e. any element in $G$ acts as an elliptic isometry, that means every isometry has a fixed point. Finite groups are special cases of groups having property ${\rm F}\mathcal{A}'$ but there are also many examples of infinite groups having this property, for instance divisible and compact groups \cite{CapraceMarquis}. We conjecture:
\begin{conjecture}
All Artin groups do not have non-trivial ${\rm F}\mathcal{A}'$-subgroups.
\end{conjecture}
We define a class $\mathcal{A}$ of Artin groups as follows: an Artin group $A_\Gamma$ is contained in the class $\mathcal{A}$ iff $A_\Gamma$ has property (Int++).
Examples of Artin groups in the class $\mathcal{A}$ are right-angled Artin groups \cite{DuncanKazachkovRemeslennikov}, Artin groups of spherical type \cite{CGG}, and large-type Artin groups \cite{LargeType}.
We reduce the above conjecture for the class $\mathcal{A}$ to complete Artin groups.
\begin{proposition}\label{FA}
Let $A_\Gamma$ be an Artin group in the class $\mathcal{A}$ and let $H\subset A_\Gamma$ be a subgroup. If $H$ is an ${\rm F}\mathcal{A}'$-group, then $H$ is contained in a complete parabolic subgroup of $A_\Gamma$.
\end{proposition}
Since a complete right-angled Artin group is isomorphic to a free abelian group and free abelian groups do not have non-trivial ${\rm F}\mathcal{A}'$-subgroups we obtain:
\begin{corollary}
Right-angled Artin groups do not have non-trivial ${\rm F}\mathcal{A}'$-subgroups.
\end{corollary}
The proof of Proposition \ref{FA} is of geometric nature. It is known that if $A_\Gamma$ is not complete, then $A_\Gamma$ in an amalgam of non-trivial parabolic subgroups $A_{\Gamma_1}\ast_{A_{\Gamma_3}} A_{\Gamma_2}$ and this group acts on the Bass-Serre tree corresponding to this splitting. We show that the subgroup $H$ has a global fixed vertex and therefore it is contained in a conjugate of one of the factors in the amalgam. We proceed to decompose the factors in the amalgam until the defining graphs of these subgroups are complete. The main tool for showing that $H$ has a global fixed vertex is the following algebraic result concerning parabolic subgroups of Artin groups. Note that, as in many other cases, the corresponding result for Coxeter groups is known to be true (see e.g. \cite{Qi}).
\begin{proposition}\label{IntVert}
Let $A_\Gamma$ be an Artin group and $gA_\Omega g^{-1}$, $hA_\Delta h^{-1}$ be two parabolic subgroups such that
$gA_\Omega g^{-1}\subset hA_\Delta h^{-1}$. Then the cardinalities of $V(\Omega)$ and $V(\Delta)$ satisfy $|V(\Omega)|\le |V(\Delta)|$ and, if $|V(\Omega)|= |V(\Delta)|$, then $gA_\Omega g^{-1}= hA_\Delta h^{-1}$.
\end{proposition}
As immediate corollary we have:
\begin{corollary}
Let $A_\Gamma$ be an Artin group. If $A_\Gamma$ is in the class $\mathcal{A}$, then an arbitrary intersection of parabolic subgroups is a parabolic subgroup. In particular, for a subset $B\subset A_\Gamma$ there exists a unique minimal (with respect to inclusion) parabolic subgroup containing $B$.
\end{corollary}
By definition, a locally compact Hausdorff group $L$ is called \emph{almost connected} if the quotient $L/L^\circ$, where $L^\circ$ is the connected component of $L$, is compact. Using the fact that any almost connected locally compact Hausdorff group has property ${\rm F}\mathcal{A}'$ \cite{Alperin} we show:
\begin{theorem}\label{MainTheorem1}
Let $A_\Gamma$ be an Artin group in the class $\mathcal{A}$.
\begin{enumerate}
\item Let $\psi\colon L\to A_\Gamma$ be a group homomorphism from a locally compact Hausdorff group $L$ into $A_\Gamma$. If $L$ is almost connected, then $\psi(L)$ is contained in a complete parabolic subgroup of $A_\Gamma$.
\item If all complete standard parabolic subgroups of $A_\Gamma$ are $lcH$-slender, then $A_\Gamma$ is $lcH$-slender.
\end{enumerate}
\end{theorem}
Associated to an Artin group $A_\Gamma$ is a generalization of the Deligne complex, the so called clique-cube complex $C_\Gamma$, whose vertices are cosets of complete standard parabolic subgroups of $A_\Gamma$.
We describe the construction of this cube complex and some important properties of it in Section 4. The group $A_\Gamma$ acts on $C_\Gamma$ via left-multiplication and preserves the cubical structure of $C_\Gamma$. We use this action to show under ``weaker'' assumptions on parabolic subgroups of an Artin group than on Artin groups in the class $\mathcal{A}$ that principle RP holds for the property of being $lcH$-slender. Since the puzzle pieces of the complex $C_\Gamma$ are cosets of complete standard parabolic subgroups of $A_\Gamma$ and hence the stabilizers of these vertices are complete parabolic subgroups, this proof relies heavily on their ``good'' behavior.
A generalization of the fixed point property ${\rm F}\mathcal{A}'$ is the property ${\rm F}\mathcal{C}'$. A group $G$ has property ${\rm F}\mathcal{C}'$ if every cellular action of $G$ on a finite dimensional CAT(0) cube complex is locally elliptic. Examples of ${\rm F}\mathcal{C}'$-groups are finite, divisible and compact groups, see \cite{CapraceMarquis}. It is known that torsion free ${\rm CAT}(0)$ cubical groups do not have non-trivial ${\rm F}\mathcal{C}'$-subgroups. Hence, any ${\rm F}\mathcal{C}'$-subgroup of a right-angled Artin group is trivial. We conjecture:
\begin{conjecture}
All Artin groups do not have non-trivial ${\rm F}\mathcal{C}'$-subgroups.
\end{conjecture}
We define a class $\mathcal{B}$ of Artin groups as follows: $A_\Gamma$ is in the class $\mathcal{B}$ iff $A_\Gamma$ has property (Int).
We show:
\begin{proposition}\label{FC}
Let $A_\Gamma$ be an Artin group in the class $\mathcal{B}$. Let $H\subset A_\Gamma$ be a subgroup. If $H$ is a ${\rm F}\mathcal{C}'$-group, then $H$ is contained in a complete parabolic subgroup of $A_\Gamma$.
\end{proposition}
Using the result of the Main Theorem in \cite{MoellerVarghese} we reduce the assumptions regarding parabolic subgroups in Theorem \ref{MainTheorem1} to complete parabolic subgroups.
\begin{theorem}\label{MainTheorem2}
Let $A_\Gamma$ be an Artin group in the class $\mathcal{B}$.
\begin{enumerate}
\item Let $\psi\colon L\to A_\Gamma$ be a group homomorphism from a locally compact Hausdorff group $L$ into $A_\Gamma$. If $L$ is almost connected, then $\psi(L)$ is contained in a parabolic complete subgroup of $A_\Gamma$.
\item If all complete standard parabolic subgroups of $A_\Gamma$ are $lcH$-slender, then $A_\Gamma$ is $lcH$-slender.
\end{enumerate}
\end{theorem}
In particular principle RP holds for Artin groups in the class $\mathcal{B}$ for the property of being $lcH$-slender.
\section{Groups associated to edge-labeled graphs}
In this section we review the basics of simplicial graphs, Coxeter groups and Artin groups, which are relevant for our applications.
\subsection{Graphs} We will be working with non-empty simplicial graphs. The basic, standard definitions can be found in \cite{Diestel}.
Since the definition of the star and the link of a vertex sometimes vary we recall their definitions. Given a graph $\Gamma=(V,E)$ and a vertex $v\in V$ we define two subgraphs of $\Gamma$, the \textit{star} of $v$, denoted by st$(v)$ and the \textit{link} of $v$ denoted by \textit{lk}$(v)$ in the following way: The star of $v$ is the subgraph generated by all vertices connected to $v$ and $v$, i.e. the subgraph induced by $\{w\in V| w=v\textit{ or } \{v,w\}\in E\}$. We obtain the link of $v$ from the star of $v$ by removing the vertex $v$ and all edges that have $v$ as an element. Recall, if $X\subseteq V$ is a subset of the vertex set, then the subgraph \textit{generated or induced} by $X$, denoted by $\langle X\rangle$, is defined as the graph $(X,F)$, where $\left\{v,w\right\}\in F$ if and only if $\left\{v,w\right\}\in E$.
\begin{center}
\begin{tikzpicture}[scale=0.8]
\coordinate (AB) at (-10,0);
\coordinate[label=above: $v$] (AC) at (-8,0);
\coordinate (AD) at (-5.5,0);
\coordinate (AE) at (-4,0);
\coordinate[label=above: $w$] (AF) at (-2,0);
\coordinate (AG) at (2,0);
\coordinate (AH) at (6,0);
\coordinate (AI) at (8,0);
\coordinate (AJ) at (10,0);
\coordinate (AK) at (-9,2);
\coordinate (AL) at (-7,2);
\coordinate (AM) at (-3,2);
\coordinate (AN) at (-1,2);
\coordinate (AO) at (-9,-2);
\coordinate (AP) at (-7,-2);
\coordinate (AQ) at (-3,-2);
\coordinate (AR) at (-1,-2);
\coordinate (AS) at (2,4);
\coordinate (AT) at (2,-4);
\coordinate (AU) at (1,-5);
\coordinate (AV) at (2,-6);
\coordinate (AW) at (3,-5);
\coordinate (AX) at (3,-6);
\coordinate (AY) at (1,5);
\coordinate (AZ) at (3,5);
\coordinate (BA) at (1.5,6);
\coordinate (BB) at (2.5,6);
\coordinate (BC) at (6,2);
\coordinate (BD) at (6,4);
\coordinate (BE) at (5,2);
\coordinate (BF) at (7,2);
\coordinate (BG) at (6,-2);
\coordinate (BH) at (8,1.5);
\coordinate (BI) at (8,-1.5);
\coordinate (BJ) at (9,1);
\coordinate (BK) at (9,-1);
\fill[color=blue] (AB) circle (2pt);
\fill[color=blue] (AC) circle (2pt);
\fill[color=blue] (AD) circle (2pt);
\fill[color=red] (AE) circle (2pt);
\fill (AF) circle (2pt);
\fill[color=red] (AM) circle (2pt);
\fill[color=red] (AN) circle (2pt);
\fill (AO) circle (2pt);
\fill[color=blue] (AP) circle (2pt);
\fill (AQ) circle (2pt);
\fill[color=red] (AR) circle (2pt);
\draw[color=blue] (AB) -- (AC);
\draw[color=blue] (AB) -- (AK);
\draw (AB) -- (AO);
\draw[color=blue] (AC) -- (AD);
\draw[color=blue] (AC) -- (AL);
\draw[color=blue] (AC) -- (AP);
\draw (AO) -- (AP);
\draw[color=blue] (AK) -- (AC);
\draw (AD) -- (AE);
\draw[color=red] (AE) -- (AM);
\draw (AE) -- (AF);
\draw (AE) -- (AQ);
\draw (AQ) -- (AR);
\draw (AF) -- (AN);
\draw (AF) -- (AR);
\draw (AM) -- (AF);
\draw[color=blue] (-8,2) -- (-8,2) node[above]{$st(v)$};
\draw[color=red] (-2,2) -- (-2,2) node[above]{$lk(w)$};
\end{tikzpicture}
\end{center}
We say a graph $\Gamma=(V,E)$ is \textit{complete} if every pair of vertices $v, w\in V$, $v\neq w$ is connected by an edge, that is $\left\{v,w\right\}\in E$. We often denote the vertex set of $\Gamma$ by $V(\Gamma)$ and the edge set by $E(\Gamma)$.
\subsection{Coxeter and Artin groups}
Let $\Gamma$ be a finite simplicial graph with an edge-labeling $m\colon E(\Gamma)\to\mathbb{N}_{\geq 2}$.
\begin{enumerate}
\item The \emph{Coxeter group} $W_\Gamma$ is defined as
$$W_\Gamma:=\langle V(\Gamma)\mid v^2, (vw)^{m(\left\{v,w\right\})} \text{ for all }v\in V(\Gamma) \text{ and whenever} \left\{v,w\right\}\in E(\Gamma)\rangle\,.$$
\item The \emph{Artin group} $A_\Gamma$ is defined as
$$A_\Gamma:= \langle V(\Gamma)\mid \underbrace{vwv\ldots}_{\substack{m(\left\{v,w\right\})-letters}}=\underbrace{wvw\ldots}_{\substack{m(\left\{v,w\right\})-letters}}\text{ whenever } \left\{v,w\right\}\in E(\Gamma)\rangle\,.$$
\end{enumerate}
In particular, $W_\Gamma$ is the quotient of $A_\Gamma$ by the subgroup normally generated by the set $\{v^2\mid v\in
\break
V(\Gamma)\}$. The epimorphism $\theta\colon A_\Gamma\to W_\Gamma$ induced by $v\mapsto v$ is called \emph{natural projection} and the kernel of $\theta$, denoted by $CA_\Gamma$, is called \emph{colored Artin group}.
Given a subset $X\subset V(\Gamma)$ of the vertex set, we write $A_X$ resp. $W_X$ for the subgroup generated by $X$, and we set $CA_X=CA_\Gamma \cap A_X$. The group $A_X$ is called \emph{standard parabolic subgroup} of $A_\Gamma$ resp. $W_X$ is called \emph{standard parabolic subgroup} of $W_\Gamma$.
\begin{proposition}
Let $\Gamma$ be a finite simplicial graph with edge-labeling $m\colon E(\Gamma)\to\left\{2,3,\ldots\right\}$ and $A_\Gamma$ resp. $W_\Gamma$ be the associated Artin resp. Coxeter group.
Let $X$ be a subset of $V(\Gamma)$.
\begin{enumerate}
\item The subgroup $W_X$ is canonically isomorphic to $W_{\langle X\rangle}$ \cite{Bourbaki2}.
\item The subgroup $A_X$ is canonically isomorphic to $A_{\langle X\rangle}$ \cite{Vanderlek}.
\end{enumerate}
\end{proposition}
As a consequence we get:
\begin{corollary}
Let $\Gamma$ be a finite simplicial graph with edge-labeling $m\colon E(\Gamma)\to\left\{2,3,\ldots\right\}$. Let $X$ be a subset of $V(\Gamma)$. Then the subgroup $CA_X$ of $CA_\Gamma$ is canonically isomorphic to $CA_{\langle X\rangle}$.
\end{corollary}
For colored Artin groups there exists a very useful tool to project to standard parabolic subgroups:
\begin{proposition}[\cite{GodelleParisB}]
Let $A_\Gamma$ denote an Artin group and $X\subset V(\Gamma)$ a subset. Then $CA_X$ is a retract of $CA_\Gamma$ in the sense that there is a homomorphism $\pi_X\colon CA_\Gamma\to CA_X$ satisfying $\pi_X(a)=a$ for all $a\in CA_X$.
\end{proposition}
We say that $A_X$ resp. $W_X$ is a \emph{complete} standard parabolic subgroup if the subgraph $\langle X\rangle$ is complete. From our point of view the puzzle pieces of a general Artin group are complete standard parabolic subgroups, in the sense that using amalgamation one can decompose any Artin group in complete standard parabolic subgroups.
\begin{lemma}\label{decomposition}
Let $A_\Gamma$ be an Artin group and $\Gamma_1$, $\Gamma_2$ two induced subgraphs of $\Gamma$. If $\Gamma_1\cup \Gamma_2=\Gamma$, then
$$A_\Gamma\cong A_{\Gamma_1}*_{A_{\Gamma_1\cap\Gamma_2}} A_{\Gamma_2}\,.$$
\end{lemma}
The proof of this lemma follows easily by analyzing the presentation of $A_\Gamma$ and the canonical presentation of the amalgam.
In particular, if there exist two vertices $v,w\in V(\Gamma)$ such that $\left\{v,w\right\}\notin E(\Gamma)$, then
\begin{enumerate}
\item $A_\Gamma\cong A_{st(v)}*_{A_{lk(v)}} A_{ V-\left\{v\right\}}$,
\item $A_\Gamma\cong A_{V-\left\{v\right\}}*_{A_{V-\left\{v,w\right\}}} A_{V-\left\{w\right\}}$.
\end{enumerate}
Before we proceed to introduce further properties of Coxeter and Artin groups we discuss one example of a decomposition into an amalgam. Let $\Gamma$ be as follows:
\begin{center}
\begin{tikzpicture}
\fill[color=black] (0,0) circle (2pt) node[below]{$v$};
\fill[color=black] (2,0) circle (2pt) node[below]{$w$};
\fill[color=black] (2,2) circle (2pt) node[above]{$x$};
\fill[color=black] (0,2) circle (2pt) node[above]{$y$};
\draw[color=black] (0,0) -- (2,0);
\node at (1, -0.2) {$2$};
\draw[color=black] (2,0) -- (2,2);
\node at (2.2, 1) {$3$};
\draw[color=black] (2,2) -- (0,2);
\node at (1,2.2) {$4$};
\draw[color=black] (0,0) -- (0,2);
\node at (-0.2, 1) {$5$};
\end{tikzpicture}
\end{center}
\noindent
The associated Artin group is given by the presentation
$$A_\Gamma=\langle v,w,x,y\mid vw=wv, wxw=xwx, xyxy=yxyx, yvyvy=vyvyv\rangle\,.$$
The vertices $v$ and $x$ in $\Gamma$ are not connected by an edge, hence
$A_\Gamma\cong A_{st(v)}\ast_{A_{lk(v)}} A_{ V-\left\{v\right\}}\,,$
where
\begin{gather*}
A_{st(v)}=\langle v,w,y\mid vw=wv, vyvyv=yvyvy\rangle\,,\ A_{lk(v)}=\langle w,y\rangle\,,\\
A_{V-\left\{v\right\}}=\langle w,x,y\mid wxw=xwx, xyxy=yxyx\rangle\,.
\end{gather*}
We continue to decompose the amalgamated parts until the special subgroups are complete. We obtain
\begin{gather*}
A_\Gamma\cong(\langle v,w\mid vw=wv\rangle\ast_{\langle v\rangle} \langle v,y\mid vyvyv=yvyvy\rangle)\ast_{\langle w,y\rangle}\\
(\langle w,x\mid wxw=xwx\rangle\ast_{\langle x\rangle}\langle x,y\mid xyxy=yxyx\rangle)\,.
\end{gather*}
Hence Lemma \ref{decomposition} gives us an algebraic tool to reduce some questions about Artin groups to complete standard parabolic subgroups.
\subsection{Parabolic subgroups of Coxeter and Artin groups}
Coxeter groups are fundamental, well understood objects in group theory, but there are many open questions concerning Artin groups. It is conjectured that most properties of Coxeter groups carry over to Artin groups. Remember that the conjugates of standard parabolic subgroups in $A_\Gamma$ resp. $W_\Gamma$ are called \emph{parabolic} subgroups.
\begin{proposition}\cite[proof of Lemma 3.3]{Qi}\label{CoxeterVertices}
Let $W_\Gamma$ be a Coxeter group and $gW_\Omega g^{-1}$, $hW_\Delta h^{-1}$ be two parabolic subgroups such that $gW_\Omega g^{-1}\subset hW_\Delta h^{-1}$.
Then the cardinalities of $V(\Omega)$ and $V(\Delta)$ satisfy $|V(\Omega)|\le |V(\Delta)|$ and, if $|V(\Omega)|= |V(\Delta)|$, then $gW_\Omega g^{-1}= hW_\Delta h^{-1}$.
\end{proposition}
We now move to Proposition \ref{IntVert} from the introduction, namely the corresponding version of the above proposition for Artin groups.
\begin{proposition}\label{ArtinVertices}
Let $A_\Gamma$ be an Artin group and $gA_\Omega g^{-1}$, $hA_\Delta h^{-1}$ be two parabolic subgroups such that $gA_\Omega g^{-1}\subset hA_\Delta h^{-1}$.
Then $|V(\Omega)|\le |V(\Delta)|$ and, if $|V(\Omega)|= |V(\Delta)|$, then $gA_\Omega g^{-1}= hA_\Delta h^{-1}$.
\end{proposition}
The remaining of the subsection is dedicated to the proof of the above proposition.
First, we need to understand how to lift an inclusion of the form $gW_\Omega g^{-1}\subset W_\Delta$ to the corresponding Artin group (see Lemma \ref{lift}).
For that we will use the following classical result on Coxeter groups. If $W_\Gamma$ is a Coxeter group, then we denote by $\lg : W_\Gamma\to \N$ the word length with respect to $V(\Gamma)$. For more information on the length function of Coxeter groups see \cite{Humphreys}.
\begin{proposition}\cite{Bourbaki2}\label{minimality}
Let $W_\Gamma$ be a Coxeter group and $X,Y$ subsets of $V(\Gamma)$.
Let $g\in W_\Gamma$.
There exists a unique element $g_0$ in the double coset $W_X\cdot g\cdot W_Y$ of minimal length, and each element $g'\in W_X\cdot g\cdot W_Y$ can be written in the form $g'=h_1g_0h_2$ with $h_1\in W_X$, $h_2\in W_Y$ and $\lg(g') = \lg(h_1) + \lg(g_0) + \lg (h_2)$.
Moreover, $\lg(g_0h)=\lg(g_0)+\lg(h)$ for all $h\in W_Y$ and $\lg(hg_0)=\lg(h)+\lg(g_0)$ for all $h\in W_X$.
\end{proposition}
Let $\Gamma$ be a finite simplicial graph with edge-labeling $m\colon E(\Gamma)\to\left\{2,3,\ldots\right\}$.
We consider a set-section $\iota: W_\Gamma \to A_\Gamma$ of the natural projection $\theta: A_\Gamma\to W_\Gamma$ which is defined as follows.
Let $g\in W_\Gamma$.
Recall that an expression $g=v_1\cdots v_l$ over $V(\Gamma)$ is called \emph{reduced} if $\lg(g)=l$.
Choose a reduced expression $g=v_1 \cdots v_l$ of $g$ and define $\iota(g)$ to be the element of $A_\Gamma$ represented by the same word $v_1 \cdots v_l$.
By Tits \cite{Tits1} this definition does not depend on the choice of the reduced expression.
Note that $\iota$ is not a group homomorphism, but, if $g,h\in W_\Gamma$ are such that $\lg(gh)=\lg(g)+\lg(h)$, then $\iota (gh) = \iota(g)\,\iota(h)$.
\begin{lemma}\label{lift}
Let $\Gamma$ be a finite simplicial graph with edge-labeling $m\colon E(\Gamma)\to\left\{2,3,\ldots\right\}$.
Let $X,Y$ be two subsets of $V(\Gamma)$ and $g\in W_\Gamma$ such that $gW_X g^{-1} \subset W_Y$.
Then $\iota(g)A_X\iota(g)^{-1} \subset A_Y$, and, if $|X|=|Y|$, then $\iota(g)A_X\iota(g)^{-1}=A_Y$.
\end{lemma}
\begin{proof} By Proposition \ref{minimality},
we can write $g$ in the form $g=h_1g_0h_2$, where $h_1\in W_Y$, $h_2 \in W_X$, $g_0$ is the element of minimal length in the double coset $W_Y\cdot g\cdot W_X$, and $\lg(g) = \lg(h_1)+\lg(g_0)+lg(h_2)$.
Then $\iota(g) = \iota(h_1)\iota(g_0)\iota(h_2)$, and, since $h_1\in W_Y$ and $h_2\in W_X$, we also obtain $g_0W_X g_0^{-1} \subset W_Y$.
Let $v\in X$.
There exists $f_v\in W_Y$ such that $g_0v=f_vg_0$.
By Proposition \ref{minimality}, $\lg(g_0)+1 = \lg(g_0v) = \lg(f_vg_0)=\lg(f_v)+\lg(g_0)$, hence $\lg(f_v)=1$, that is, $f_v\in Y$.
Moreover, by the definition itself of $\iota$, $\iota(g_0) v = \iota(g_0v) = \iota(f_vg_0) = f_v \iota(g_0)$, hence $\iota(g_0)v\iota(g_0)^{-1}=f_v$.
So, $\iota(g_0) X \iota(g_0)^{-1} \subset Y$.
This implies that, on the one hand, $\iota(g_0) A_X \iota(g_0)^{-1} \subset A_Y$, and, on the other hand, if $|X|=|Y|$, then $\iota(g_0)X\iota(g_0)^{-1}=Y$ and therefore $\iota(g_0)A_X\iota(g_0)^{-1}=A_Y$.
So, since $\iota(h_1)\in A_Y$ and $\iota(h_2)\in A_X$,
\begin{align*}
\iota(g) A_X \iota(g)^{-1} &=
\iota(h_1)\iota(g_0)\iota(h_2)A_X\iota(h_2)^{-1}\iota(g_0)^{-1}\iota(h_1)^{-1}\\&=
\iota(h_1)\iota(g_0)A_X\iota(g_0)^{-1}\iota(h_1)^{-1}\\&\subset
\iota(h_1)A_Y\iota(h_1)^{-1}\\&=
A_Y\,,
\end{align*}
and we have equality if $|X|=|Y|$.
\end{proof}
With this tool we can now prove the main proposition of this section.
\begin{proof}[Proof of Proposition \ref{ArtinVertices}]
Recall that $\theta\colon A_\Gamma \to W_\Gamma$ denotes the natural projection to the corresponding Coxeter group.
By applying $\theta$ we have $\theta(g)W_\Omega\theta(g)^{-1} \subset \theta(h)W_\Delta \theta(h)^{-1}$, hence, by Proposition \ref{CoxeterVertices}, $|V(\Omega)| \le |V(\Delta)|$.
Now, we assume that $|V(\Omega)|=|V(\Delta)|$ and we prove that $gA_\Omega g^{-1}=hA_\Delta h^{-1}$.
Without loss of generality we can assume that $h=1$, else we conjugate both sides with $h^{-1}$ and replace $g$ by $h^{-1}g$.
Set $\bar g=\theta(g)$.
We have $\bar g W_\Omega \bar g^{-1} \subset W_\Delta$ and $|V(\Omega)|=|V(\Delta)|$, hence, by Lemma \ref{lift}, $\iota(\bar g)A_\Omega \iota(\bar g)^{-1}=A_\Delta$.
Thus, we obtain $gA_{\Omega}g^{-1}\subset \iota(\bar g)A_{\Omega}\iota(\bar g)^{-1}$ and after conjugating with $\iota(\bar g)^{-1}$ we obtain $kA_{\Omega}k^{-1}\subset A_{\Omega}$, where $k=\iota(\bar g)^{-1}g$.
Note that $\theta(k)=1$, that is, $k\in CA_\Gamma$, hence we can apply the retraction map $\pi_\Omega\colon CA_\Gamma\to CA_\Omega$ to $k$.
We have the following commutative diagram:
\hspace*{5.5cm}
\begin{xy}
\xymatrix{
CA_\Omega \ar[r] \ar[d] & A_\Omega \ar[r] \ar[d] & W_\Omega \ar[d] \\
CA_\Omega \ar[r] & A_\Omega \ar[r]& W_\Omega
}
\end{xy}
\noindent
where the vertical arrow $W_\Omega\to W_\Omega$ is the identity and the vertical arrows $CA_\Omega\to CA_\Omega$ and $A_\Omega\to A_\Omega$ are conjugations by $k$.
Using the retraction map $\pi_\Omega\colon CA_\Gamma\to CA_\Omega$ we show that the left vertical arrow is an isomorphism.
For $a\in CA_\Omega$ we define $a':=k\pi_\Omega(k^{-1})a\pi_\Omega(k)k^{-1}$.
Then $a'\in k CA_\Omega k^{-1} \subset CA_\Omega$, and $a'=\pi_\Omega(a')=a$.
Hence, the arrow is surjective.
Note that the arrow is also injective, since the map is a conjugation.
Since we know that the arrows $CA_\Omega\to CA_\Omega$ and $W_\Omega\to W_\Omega$ are isomorphisms, by applying the five lemma we get that the arrow $A_\Omega\to A_\Omega$ is an isomorphism as well, which means that $k A_\Omega k^{-1} = A_\Omega$.
By conjugating with $\iota(\bar g)$ we conclude that $g A_\Omega g^{-1}=A_\Delta$.
\end{proof}
\subsection{Parabolic closure}
Let $A_\Gamma$ be an Artin group. Suppose that a subset $B\subset A_\Gamma$ is contained in a unique minimal parabolic subgroup of $A_\Gamma$ (minimal with respect to $\subset$). Then this parabolic subgroup is called the \emph{parabolic closure of B} and is denoted by $PC_\Gamma(B)$.
In the case that the intersection of two parabolic subgroups is parabolic, we show that the parabolic closure exists.
\begin{proposition}\label{PC}
Let $A_\Gamma$ be an Artin group. If $A_\Gamma$ has property (Int++), then an arbitrary intersection of parabolic subgroups is a parabolic subgroup. In particular, for a subset $B\subset A_\Gamma$ there exists $PC_\Gamma(B)$.
\end{proposition}
\begin{proof}
Let $N$ be a set consisting of parabolic subgroups of $A_\Gamma$. Our goal is to show that $\bigcap N$ is parabolic. We write $N$ using indices in an index set $I$ as $N=\bigcup\limits_{i\in I} \left\{P_i\right\}$ where each $P_i$ is a parabolic subgroup of $A_\Gamma$. We construct a chain:
$$P_{i_1}\supset P_{i_1}\cap P_{i_2}\supset P_{i_1}\cap P_{i_2}\cap P_{i_3}\supset...$$
We show that this chain stabilizes: Each proper inclusion in the chain is of the form $gA_{\Delta_1}g^{-1}\supsetneq hA_{\Delta_2}h^{-1}$ since finite intersection of parabolic subgroups are assumed to be parabolic. Due to Proposition \ref{ArtinVertices} the cardinality of the vertices in the defining graphs has to strictly reduce at each step. Since $\Gamma$ is finite, this can only happen finitely many times. Thus there exists $n\in\mathbb{N}$ such that $\bigcap N=P_{i_1}\cap\ldots\cap P_{i_n}$ and by assumption this finite intersection is parabolic.
For the in particular statement, we consider the set
$$M=\{P\subset A_\Gamma| P\text{ is a parabolic subgroup and } B\subset P\}$$
and we write $M$ using indices in an index set $J$ as $M=\bigcup\limits_{j\in J} \left\{Q_j\right\}$, where each $Q_j$ is a parabolic subgroup of $A_\Gamma$. We have
$$Q_{j_1}\supset Q_{j_1}\cap Q_{j_2}\supset Q_{j_1}\cap Q_{j_2}\cap Q_{j_3}\supset...$$
This chain stabilizes by the above proof.
We define $R=\bigcap\limits_{j\in J} Q_j$ which is a parabolic subgroup since the chain stabilizes. That $R$ is minimal is obvious from its definition. Note that $R$ is also unique, more precisely: assume that there exists another minimal parabolic subgroup $R'$ with $B\subset R'$. By assumption $R\cap R'$ is a parabolic subgroup containing $B$. Since $R$ and $R'$ are minimal, we have $R\cap R'=R$ and $R\cap R'=R'$, hence $R=R'$.
\end{proof}
In particular, the parabolic closure of a subset $B\subset A_\Gamma$ is the parabolic subgroup $a A_{\Gamma'} a^{-1}$ containing $B$ for $|V(\Gamma')|$ minimal.
The proofs of the following lemmata are of the same flavor as the proof of the above proposition so we omit the details.
\begin{lemma} \label{PCcomplete}
Let $A_\Gamma$ be an Artin group and $B$ be a subset of a complete parabolic subgroup. If $A_\Gamma$ has (Int+--), then there exists $PC_\Gamma(B)$.
\end{lemma}
\begin{lemma}\label{subsetPC}
Let $A_\Gamma$ be an Artin group and $B_1$ and $B_2$ be subsets of $A_\Gamma$. Assume that $PC_\Gamma(B_1)$ and $PC_\Gamma(B_2)$ exist. If $B_1\subset B_2$, then $PC_\Gamma(B_1)\subset PC_\Gamma(B_2)$.
\end{lemma}
Remember that we defined two classes of Artin groups in the introduction as follows:
\begin{enumerate}
\item
An Artin group $A_\Gamma$ is contained in the class $\mathcal{A}$ iff $A_\Gamma$ has property (Int++).
\item
An Artin group $A_\Gamma$ is contained in the class $\mathcal{B}$ iff $A_\Gamma$ has property (Int).
\end{enumerate}
Due to previous work of multiple authors, we know that many classes of Artin groups fall into $\mathcal{A}$ or $\mathcal{B}$, for example:
\begin{enumerate}
\item
Artin groups of spherical type, right-angled Artin groups and large type Artin groups are in the class $\mathcal{A}$ \cite{CGG,DuncanKazachkovRemeslennikov,LargeType}.
\item
Artin groups of FC type are in the class $\mathcal{B}$ \cite{MorrisWright}.
\end{enumerate}
In the next section we show that an Artin group in the class $\mathcal{B}$ already has property (Int+--). This implies:
\begin{corollary}\
\begin{enumerate}
\item
Let $A_\Gamma$ be in the class $\mathcal{A}$. Then for any subset $B\subset A_\Gamma$ there exists $PC_\Gamma(B)$.
\item
Let $A_\Gamma$ be in the class $\mathcal{B}$. For a subset $B\subset A_\Gamma$ of a complete parabolic subgroup there exists $PC_\Gamma(B)$.
\end{enumerate}
\end{corollary}
\section{Bass-Serre theory meets Artin groups}
Let $\psi\colon G\to{\rm Isom}(T)$ be a group action via simplicial isometries on a tree $T$ without inversion. One can consider a tree as a metric space by assigning each edge the length one. If we write `$x\in T$' we implicitly assume that we metrized the tree in that way. For a subset $A\subset G$ we denote by ${\rm Fix}(\psi(A)):=\left\{x\in T\mid \psi(a)(x)=x \text{ for all }a\in A\right\}$ the fixed point set of $\psi(A)$. Note that if ${\rm Fix}(\psi(A))$ is non-empty, then it is a subtree of $T$. Further, for an edge $e\in E(T)$ the stabilizer of $e$, denoted by ${\rm stab}(e)$ is defined as ${\rm stab}(e):=\left\{g\in G\mid \psi(g)(e)=e\right\}$ and resp. for a vertex $v\in V(T)$ the stabilizer ${\rm stab}(v):=\left\{g\in G\mid \psi(g)(v)=v\right\}$.
Given an amalgam $A*_C B$, by the Bass-Serre theory, there is a simplicial tree
$T_{A*_C B}$ on which $A*_C B$ acts simplicially without a global fixed point. The Bass-Serre
tree $T_{A*_C B}$ is constructed as follows: the vertices of $T_{A*_C B}$ are cosets of $A$ and $B$ and
two vertices $gA$ and $hB$, $g,h\in A*_C B$ are connected by an edge iff $gA\cap hB=gC$. For more information about amalgamated products and Bass-Serre theory see \cite{Serre}. For the very small example of $G=\mathbb{Z}/4\mathbb{Z}\ast\mathbb{Z}/6\mathbb{Z}$ (a portion of) the Bass-Serre tree looks like this (suppose $G=\langle a,b\rangle, o(a)=4, o(b)=6$):
\begin{center}
\begin{tikzpicture} \coordinate[label={[xshift=0.3cm]:${1\langle a\rangle}$}] (A) at (-2,0);
\coordinate[label={[xshift=-0.3cm]:${1\langle b\rangle}$}] (B) at (2,0);
\coordinate[label=left :${b\langle a\rangle}$] (C) at (3,2);
\coordinate[label={[xshift=-0.3cm,yshift=-0.1cm]:${b^2\langle a\rangle}$}] (D) at (3.5,1);
\coordinate[label=above :${b^3\langle a\rangle}$] (E) at (4,0);
\coordinate[label={[xshift=0.1cm]:${b^4\langle a\rangle}$}] (F) at (3.5,-1);
\coordinate[label=left :${b^5\langle a\rangle}$] (G) at (3,-2);
\coordinate[label=right:${a\langle b\rangle}$] (H) at (-3,1);
\coordinate[label={[xshift=0.2cm]:${a^2\langle b\rangle}$}] (I) at (-4,0);
\coordinate[label=right:${a^3\langle b\rangle}$] (J) at (-3,-1);
\fill (A) circle (2pt);
\fill (B) circle (2pt);
\fill (C) circle (2pt);
\fill (D) circle (2pt);
\fill (E) circle (2pt);
\fill (F) circle (2pt);
\fill (G) circle (2pt);
\fill (H) circle (2pt);
\fill (I) circle (2pt);
\fill (J) circle (2pt);
\draw (A) -- (B);
\draw (B) -- (C);
\draw (B) -- (D);
\draw (B) -- (E);
\draw (B) -- (F);
\draw (B) -- (G);
\draw (A) -- (H);
\draw (A) -- (I);
\draw (A) -- (J);
\draw[dotted] (J) -- (-2.5,-2);
\draw[dotted] (J) -- (-2.25,-1.7);
\draw[dotted] (J) -- (-3,-2.25);
\draw[dotted] (J) -- (-3.5,-2);
\draw[dotted] (J) -- (-3.75,-1.7);
\draw[dotted] (H) -- (-2.5,2);
\draw[dotted] (H) -- (-2.25,1.7);
\draw[dotted] (H) -- (-3,2.25);
\draw[dotted] (H) -- (-3.5,2);
\draw[dotted] (H) -- (-3.75,1.7);
\draw[dotted] (I) -- (-4.5,0.5);
\draw[dotted] (I) -- (-4.75,0.3);
\draw[dotted] (I) -- (-5,0);
\draw[dotted] (I) -- (-4.75,-0.3);
\draw[dotted] (I) -- (-4.5,-0.5);
\draw[dotted] (C) -- (2.5,2.8);
\draw[dotted] (C) -- (3,3);
\draw[dotted] (C) -- (3.5,2.8);
\draw[dotted] (D) -- (4.5,1.5);
\draw[dotted] (D) -- (4.6,1);
\draw[dotted] (D) -- (4.2,2);
\draw[dotted] (E) -- (5,0.5);
\draw[dotted] (E) -- (5.1,0);
\draw[dotted] (E) -- (5,-0.5);
\draw[dotted] (F) -- (4.1,-1.9);
\draw[dotted] (F) -- (4.6,-1);
\draw[dotted] (F) -- (4.5,-1.5);
\draw[dotted] (G) -- (2.5,-2.8);
\draw[dotted] (G) -- (3,-3);
\draw[dotted] (G) -- (3.5,-2.8);
\end{tikzpicture}
\end{center}
How the Bass-Serre tree changes with amalgamation can be seen in \cite[p. 35]{Serre}, where the Bass-Serre tree is drawn for $\mathbb{Z}/4\mathbb{Z}\ast_{\mathbb{Z}/2\mathbb{Z}}\mathbb{Z}/6\mathbb{Z}$.
The group $G=A\ast_C B$ acts on its Bass-Serre tree via left-multiplication. For this action the stabilizer of a vertex $gA$ is given by $gAg^{-1}$ and the stabilizer of an edge between $gA$ and $hB$ is given by $gCg^{-1}$ (since $gA\cap hB=gC$).
\subsection{Intersection of parabolic subgroups}
In this section we prove Theorem \ref{Int} from the introduction. Our proof is of geometric nature with main tool being Bass-Serre theory. We will be using a slightly different notation for the vertices and edges of the Bass-Serre tree associated to $A_\Gamma=A_X\ast_{A_Z}A_Y$.
\begin{notation}
Given an Artin group $A_\Gamma=A_X\ast_{A_Z}A_Y$, we write vertices in the Bass-Serre tree $T$ as $v(a,X):=aA_X$ and $v(b,Y):=bA_Y$. For the edges we simply associate an edge $e(a)$ to each $a\in A_\Gamma$ with endpoints $v(a,X)$ and $v(a,Y)$. We have $e(a)=e(b)$ if and only if $aA_Z=bA_Z$ and every edge of $T$ has this form.
\end{notation}
\begin{proof}[Proof of Theorem \ref{Int}]
We assume that $A_\Gamma$ has Property (Int).
We take a complete parabolic subgroup $P_1$ and a parabolic subgroup $P_2$ of $A_\Gamma$, and we prove that $P_1\cap P_2$ is a parabolic subgroup by induction on the number of pairs $\{s,t\} \subset V(\Gamma)$ satisfying $\left\{s,t\right\}\notin E(\Gamma)$.
If there is no such a pair, then $\Gamma$ itself is complete and then property (Int) suffices for saying that $P_1 \cap P_2$ is a parabolic subgroup.
So, we can assume that there exists a pair $\{s,t\} \subset V(\Gamma)$ such that $\{s,t\}\notin E(\Gamma)$ and that the inductive hypothesis holds.
We set $I=V(\Gamma)\setminus \{s\}$, $J=V(\Gamma)\setminus\{t\}$, and $K=V(\Gamma)\setminus\{s,t\}$. By Lemma \ref{decomposition} we have the amalgamated product $A_\Gamma\cong A_I*_{A_K} A_J$, which leads to the construction of the Bass-Serre tree $T$ associated to this splitting.
There exist subsets $X,Y\subset V(\Gamma)$ and elements $g,h\in A_\Gamma$ such that $X$ is free of infinity, $P_1 = g A_X g^{-1}$, and $P_2 = h A_Y h^{-1}$.
Since $X$ is free of infinity, we have either $X\subset I$ or $X\subset J$, hence there exists a vertex $u_0=v(a_0,U_0)$ of $T$ such that
\[
P_1=g A_X g^{-1} \subset {\rm stab}(u_0)=a_0 A_{U_0} a_0^{-1}\,.
\]
From here the proof is divided into two cases.
\bigskip\noindent
{\it Case 1 :}
$\{s,t\}\not\subset Y$.
Then either $Y\subset I$ or $Y\subset J$, hence there exists a vertex $v_0=v(b_0,V_0)$ of $T$ such that
\[
P_2=h A_Y h^{-1} \subset {\rm stab}(v_0)=b_0 A_{V_0} b_0^{-1}\,.
\]
We denote by $d$ the distance in $T$ between $u_0$ and $v_0$ and we argue by induction on $d$.
Suppose $d=0$, that is, $u_0=v_0$.
Then $a_0^{-1} P_1 a_0$ is a complete parabolic subgroup of $A_{U_0}$ and $a_0^{-1} P_2 a_0$ is a parabolic subgroup of $A_{U_0}$.
The number of pairs $\{i,j\}$ satisfying $\{i,j\} \not\in E(\Gamma)$ is strictly smaller in $U_0$ than in $V(\Gamma)$, hence, by the induction hypothesis, $(a_0^{-1} P_1 a_0) \cap (a_0^{-1} P_2 a_0) = a_0^{-1} (P_1 \cap P_2) a_0$ is a parabolic subgroup of $A_{U_0}$, and therefore a parabolic subgroup of $A_\Gamma$.
So, $P_1 \cap P_2$ is a parabolic subgroup of $A_\Gamma$.
Now we assume that $d\ge 1$ and that the inductive hypothesis on $d$ holds.
Let $(u_0,u_1,\dots,u_d)$ be the unique geodesic in $T$ connecting $u_0$ with $v_0=u_d$.
For $i\in\{1,\dots,d\}$ we denote by $e_i$ the edge connecting $u_{i-1}$ with $u_i$.
Note that, since $P_1 \subset {\rm stab}(u_0)$ and $P_2\subset {\rm stab}(v_0)$, we have $P_1\cap P_2 \subset {\rm stab}(u_i)$ for all $i\in\{0,1,\dots, d\}$ and $P_1\cap P_2\subset {\rm stab}(e_i)$ for all $i\in\{1,\dots, d\}$.
We set $u_1=v(a_1,U_1)$ and $e_1=e(c_1)$.
\begin{center}
\begin{tikzpicture}
\coordinate[label=below:${u_0=v(a_0,U_0)}$] (A) at (-5,0);
\coordinate[label=above:${u_1=v(a_1,U_1)}$] (B) at (-3.8,1);
\coordinate (C) at (-2.6,-1);
\coordinate (D) at (-1.4,1);
\coordinate (E) at (-0.2,-1);
\coordinate (F) at (0,1);
\coordinate[label=right:${v_0=u_d}$] (G) at (1,0);
\draw (A) -- (B) node[midway,above,sloped]{$e(c_1)$};
\draw (B) -- (C);
\draw (C) -- (D);
\draw (D) -- (E);
\draw (E) -- (G);
\draw[dotted] (-6,1.5) -- (A);
\draw[dotted] (-6,-1.5) -- (A);
\draw[dotted] (2,1.5) -- (G);
\draw[dotted] (2,-1.5) -- (G);
\fill (A) circle (2pt);
\fill (B) circle (2pt);
\fill (C) circle (2pt);
\fill (D) circle (2pt);
\fill (E) circle (2pt);
\fill (G) circle (2pt);
\end{tikzpicture}
\end{center}
The group $a_0^{-1}g A_X g^{-1} a_0$ is a complete parabolic subgroup of $A_{U_0}$ and $a_0^{-1} c_1 A_K c_1^{-1} a_0$ is a parabolic subgroup of $A_{U_0}$, hence, by the induction hypothesis (on the number of pairs $\{i,j\}$ such that $\{i,j\}$ is not an edge), $(a_0^{-1}g A_X g^{-1} a_0) \cap (a_0^{-1} c_1 A_K c_1^{-1} a_0) = a_0^{-1} ( (g A_X g^{-1}) \cap (c_1 A_K c_1^{-1}) )a_0$ is a parabolic subgroup, and therefore $(g A_X g^{-1}) \cap (c_1 A_K c_1^{-1})$ is a parabolic subgroup.
It is complete because it is contained in $P_1= g A_X g^{-1}$.
So, there exist $g_1\in A_\Gamma$ and $X_1\subset V(\Gamma)$ such that $X_1$ is free of infinity and $(g A_X g^{-1}) \cap (c_1 A_K c_1^{-1}) = g_1 A_{X_1}g_1^{-1}$.
Moreover,
\[
g_1 A_{X_1} g_1^{-1} \subset c_1 A_K c_1^{-1} \subset a_1 A_{U_1} a_1^{-1} = {\rm stab}(u_1)\,.
\]
By the induction hypothesis (on $d$), it follows that
\begin{gather*}
P_1 \cap P_2 =
(g A_X g^{-1}) \cap (h A_Y h^{-1}) =
(g A_X g^{-1}) \cap (c_1 A_K c_1^{-1}) \cap (h A_Y h^{-1}) =\\
(g_1 A_{X_1} g_1^{-1}) \cap (h A_Y h^{-1})
\end{gather*}
is a parabolic subgroup of $A_\Gamma$.
This finishes the proof of Case 1.
\bigskip\noindent
{\it Case 2:}
$\{s,t\} \subset Y$.
We set $Y_I=Y\setminus \{s\}=Y\cap I$, $Y_J=Y\setminus \{t\} = Y\cap J$ and $Y_K=Y\setminus \{s,t\} = Y\cap K$.
As for $A_\Gamma$ we have the amalgamated product $A_Y=A_{Y_I} *_{A_{Y_K}} A_{Y_J}$.
We denote by $T_Y$ the Bass-Serre tree associated to this amalgamated product.
\bigskip\noindent
{\it Claim.}
{\it We have an embedding of $T_Y$ into $T$ which, for each $a\in A_Y$, sends $v(a,Y_I)$ to $v(a,I)$, $v(a,Y_J)$ to $v(a,J)$, and $e(a)$ to $e(a)$.}
\bigskip\noindent
{\it Proof of the Claim.}
Let $a,b \in A_Y$.
We need to show the following equivalencies:
\begin{gather*}
aA_{Y_I}=bA_{Y_I}\ \Leftrightarrow\ aA_I = bA_I\,,\quad
aA_{Y_J}=bA_{Y_J}\ \Leftrightarrow\ aA_J = bA_J\,,\\
aA_{Y_K}=bA_{Y_K}\ \Leftrightarrow\ aA_K = bA_K\,.
\end{gather*}
We prove the first one.
The others can be proved in the same way.
Recall that, by Van der Lek \cite{Vanderlek}, $A_{Y_I}=A_{Y\cap I} = A_Y \cap A_I$.
So, since $a^{-1}b\in A_Y$,
\[
aA_{Y_I}=bA_{Y_I}\ \Leftrightarrow\
a^{-1}b \in A_{Y_I}=A_Y \cap A_I\ \Leftrightarrow\
a^{-1}b \in A_I\ \Leftrightarrow\
aA_I = bA_I\,.
\]
This completes the proof of the Claim.
\bigskip\noindent
Let $v_0$ be the unique vertex of $h T_Y$ at minimal distance from $u_0$.
Note that $v_0$ may be equal to $u_0$.
\begin{center}
\begin{tikzpicture}
\coordinate (A) at (2,0);
\coordinate (B) at (0,0);
\coordinate (C) at (3,1);
\coordinate (D) at (3,-1);
\coordinate (E) at (2.5,1.5);
\coordinate (F) at (3.5,1.5);
\coordinate (G) at (2.5,-1.5);
\coordinate (H) at (3.5,-1.5);
\coordinate (I) at (-1,1);
\coordinate (J) at (-1,-1);
\coordinate (K) at (-0.5,1.5);
\coordinate (L) at (-1.5,1.5);
\coordinate (M) at (-0.5,-1.5);
\coordinate (N) at (-1.5,-1.5);
\fill (B) circle (2pt);
\fill (D) circle (2pt);
\fill (G) circle (2pt);
\fill (H) circle (2pt);
\fill (I) circle (2pt);
\fill (K) circle (2pt);
\fill (L) circle (2pt);
\fill (M) circle (2pt);
\fill (N) circle (2pt);
\draw (A) -- (B);
\draw[color=red] (A) -- (C);
\draw (A) -- (D);
\draw[color=red] (C) -- (E);
\draw[color=red] (C) -- (F);
\draw (D) -- (G);
\draw (D) -- (H);
\draw (B) -- (I);
\draw (B) -- (J);
\draw (I) -- (K);
\draw (I) -- (L);
\draw (J) -- (M);
\draw (J) -- (N);
\fill[color=blue] (J) circle (2pt);
\draw[color=blue] (J) -- (J) node[midway, right]{$u_0$};
\fill[color=red] (A) circle (2pt);
\fill[color=red] (C) circle (2pt);
\fill[color=red] (E) circle (2pt);
\fill[color=red] (F) circle (2pt);
\draw[color=red] (3.2,0.1) -- (3.2,0.1) node[midway,above]{$h T_Y$};
\draw[color=red] (A) -- (A) node[midway, below]{$v_0$};
\end{tikzpicture}
\end{center}
Since $P_1$ stabilizes $u_0$ and $P_2$ stabilizes $h T_Y$ (as a set, not pointwise), $P_1 \cap P_2$ stabilizes $v_0$, that is, $P_1\cap P_2 \subset {\rm stab}(v_0)$ due to the minimality of the distance between $u_0$ and $v_0$.
Let $a\in A_Y$ and $V\in \{I,J\}$ such that $v_0=v(h a,V)$.
We have
\begin{gather*}
{\rm stab}(v_0) \cap P_2 =
(h a A_V a^{-1} h^{-1}) \cap (h A_Y h^{-1}) =
(h a A_V a^{-1} h^{-1}) \cap (h a A_Y a^{-1} h^{-1}) =\\
h a (A_V \cap A_Y) a^{-1} h^{-1} =
h a A_{V\cap Y} a^{-1} h^{-1}\,,
\end{gather*}
hence
\[
P_1 \cap P_2 =
P_1 \cap {\rm stab}(v_0) \cap P_2 =
P_1 \cap (h a A_{V\cap Y} a^{-1} h^{-1})\,.
\]
Since $\{s,t\}\not\subset V\cap Y$, we conclude by Case 1 that $P_1 \cap P_2$ is a parabolic subgroup.
\end{proof}
\subsection{Serre's property FA and a generalization of this fixed point property}
We use Bass-Serre theory in this section to prove Proposition \ref{FA} and Theorem \ref{MainTheorem1}. Let us first recall the definitions of properties ${\rm F}\mathcal{A}$ and ${\rm F}\mathcal{A}'$ and the basic implications we need later on.
A group $G$ is said to have property ${\rm F}\mathcal{A}$ if every simplicial action of $G$ on any tree without inversions has a global fixed vertex. A weaker property is property ${\rm F}\mathcal{A'}$, here we require that for every action of $G$ on any tree without inversions every element has a fixed point.
\begin{proposition}
Artin groups do not have property ${\rm F}\mathcal{A}$.
\end{proposition}
\begin{proof}
Due to \cite[Thm. 15]{Serre} a countable group cannot have property F$\mathcal{A}$ if it has a quotient isomorphic to $\mathbb{Z}$. Given an Artin group $A_\Gamma$ with standard generating set $V=\{v_1,...,v_n\}$ we define a homomorphism $\varphi\colon A_\Gamma\to \mathbb{Z}$ by setting $\varphi(v_i)=1$ for every $i\in\{1,...,n\}$. Due to the universal property of group presentations this defines a homomorphism, since all relations are in the kernel. Furthermore this map is surjective, which means that $A_\Gamma$ indeed has a quotient isomorphic to $\mathbb{Z}$, namely $A_\Gamma/\ker(\varphi)$.
\end{proof}
Moving on to property ${\rm F}\mathcal{A'}$ we need the following lemma to show that ${\rm F}\mathcal{A'}$ subgroups of Artin groups have to be contained in complete parabolic subgroups.
\begin{lemma}[\cite{Serre}, Prop. 26]
\label{twogroups}
Let $\psi\colon G\to{\rm Isom}(T)$ be a simplicial action on a tree $T$ without inversion. Let $A$ and $B$ be subgroups of $G$. If ${\rm Fix}(\psi(A))\neq\emptyset, {\rm Fix}(\psi(B))\neq\emptyset$ and ${\rm Fix}(\psi(ab))\neq\emptyset$ for all $a\in A$ and $b\in B$, then ${\rm Fix}(\psi(\langle A, B\rangle))\neq\emptyset$.
\end{lemma}
This allows us to prove the following proposition:
\begin{proposition}\label{Prop FA}
Let $\psi\colon G\to{\rm Isom}(T)$ be a simplicial action on a tree $T$ without inversion. If ${\rm Fix}(\psi(g))\neq\emptyset$ for all $g\in G$, then either
\begin{enumerate}
\item ${\rm Fix}(\psi(G))\neq\emptyset$, or
\item there exists a sequence of edges $(e_i)_{i\in\mathbb{N}}$ such that
$${\rm stab}(e_1)\subsetneq {\rm stab}(e_2)\subsetneq ... .$$
Additionally, if $G$ is countable, then $G=\bigcup\limits_{i\in\mathbb{N}} {\rm stab}(e_i)$.
\end{enumerate}
\end{proposition}
\begin{proof}
We differentiate two cases:
\bigskip\noindent
\textit{Case 1: There exists a subgroup $H\subset G$ such that ${\rm Fix}(\psi(H))$ consists of exactly one vertex $v$.}
Then for $g\in G$ we know by Lemma \ref{twogroups} that ${\rm Fix}(\psi(\langle H, g\rangle))\neq\emptyset$, hence $\psi(g)(v)=v$. Since $g$ is arbitrary we have ${\rm Fix}(\psi(G))=\left\{v\right\}$.
\bigskip\noindent
\textit{Case 2: For any subgroup $H\subset G$ such that ${\rm Fix}(\psi(H))\neq \emptyset$ the fixed point set ${\rm Fix}(\psi(H))$ contains always an edge.}
Let $g_1$ be in $G$. By assumption the fixed point set of $\psi(g_1)$ contains at least one edge $e_1\in {\rm Fix}(\psi(g_1))$. If $e_1\in{\rm Fix}(\psi(G))$, then we are done. Otherwise there exists $g_2\in G-{\rm stab}(e_1)$. By Lemma \ref{twogroups} the fixed point set ${\rm Fix}(\psi(\langle {\rm stab}(e_1), g_2\rangle))$ is non-empty and by assumption this set contains at least one edge $e_2\in{\rm Fix}(\psi(\langle{\rm stab}(e_1), g_2\rangle))$. We obtain
$${\rm stab}(e_1)\subsetneq {\rm stab}(e_2).$$
If $e_2\in {\rm Fix}(\psi(G))\neq\emptyset$, then we are done. Otherwise there exists $g_3\in G-{\rm stab}(e_2)$ and we proceed as before. By this construction we obtain an ascending chain of ${\rm stab}(e_i)$. Now there are two possibilities: this chain either stabilizes or it does not.
If it stabilizes, then there exists an edge $e_n$ such that $G={\rm stab}(e_n)$ and thus $e_n\in{\rm Fix}(\psi(G))$. If the chain does not stabilize, then we end up in case (2) of Proposition \ref{Prop FA} and if $G$ is countable, then it is straightforward to verify that it is possible to write $G=\bigcup\limits_{i\in\mathbb{N}} {\rm stab}(e_i)$. This can be done with a slight modification of the chain construction by always picking a specific element $g_m\in G-{\rm stab}(e_{m-1})$.
\end{proof}
\begin{corollary}\label{AmalgamFA}
Let $G=A\ast_{C}B$ denote an amalgamated free product and $H\subset G$ be a subgroup. If $H$ is an ${\rm F}\mathcal{A}'$-group, then
\begin{enumerate}
\item $H$ is contained in a conjugate of $A$ or $B$ or
\item there exists a sequence of elements $(g_i)_{i\in\mathbb{N}}$, $g_i\in G$ such that
$$g_1Cg_1^{-1}\cap H\subsetneq g_2Cg_2^{-1}\cap H\subsetneq\ldots$$
Additionally, if $H$ is countable, then $H=\bigcup\limits_{i\in\mathbb{N}} g_iCg_i^{-1}\cap H$.
\end{enumerate}
\end{corollary}
\begin{proof}
The group $G$ acts on its Bass-Serre tree without inversion. Restricting this action to $H$ and applying Proposition \ref{Prop FA} shows the corollary since the stabilizers of vertices are conjugates of $A$ or $B$ and the stabilizers of edges are conjugates of $C$ intersected with $H$.
\end{proof}
\begin{proposition}
Let $A_\Gamma$ be an Artin group in the class $\mathcal{A}$ and let $H\subset A_\Gamma$ be a subgroup. If $H$ is an ${\rm F}\mathcal{A}'$-group, then $H$ is contained in a conjugate of a special complete subgroup of $A_\Gamma$.
\end{proposition}
\begin{proof}
Let $A_\Gamma$ denote an Artin group in the class $\mathcal{A}$ and $H\subset A_\Gamma$ an $\mathcal{FA'}$ subgroup. If $\Gamma$ is not complete, then decompose $A_\Gamma$ as an amalgamated product according to Lemma \ref{decomposition}, thus $A_\Gamma=A_{st(v)}\ast_{A_{lk(v)}} A_{ V-v}$. We now apply Corollary \ref{AmalgamFA} to the amalgamated product $A_\Gamma=A_{st(v)}\ast_{A{_{lk(v)}}} A_{ V-v}$.
If we are in case (1), we are in a conjugate of one of the factors and we repeat the process of decomposing the factor into an amalgam.
If we are in case (2), we obtain a proper infinite chain
$$g_1A_{lk(v)}g_1^{-1}\cap H\subsetneq g_2A_{lk(v)}g_2^{-1}\cap H\subsetneq...$$
and since $A_\Gamma$ is countable we know that $H$ is countable, hence $H=\bigcup\limits_{i\in\mathbb{N}} g_i A_{lk(v)}g_i^{-1}\cap H$.
By Lemma \ref{subsetPC} we can transform the above chain into the following chain of parabolic subgroups
$$PC_\Gamma(g_1A_{lk(v)}g_1^{-1}\cap H)\subset PC_\Gamma(g_2A_{lk(v)}g_2^{-1}\cap H)\subset...$$
and we also have
$$H\subset \bigcup\limits_{i\in\mathbb{N}} PC_\Gamma(g_i A_{lk(v)}g_i^{-1}\cap H)\,.$$
We can now apply Proposition \ref{ArtinVertices} to see that the above chain of parabolic subgroups stabilizes, say at $PC_\Gamma(g_n A_{lk(v)} g_n^{-1}\cap H)$. Then we have
$$H\subset PC_\Gamma(g_n A_{lk(v)} g_n^{-1}\cap H)\subset g_n A_{lk(v)}g_n^{-1}\subset g_n A_{st(v)}g_n^{-1}\,.$$
If the subgraph $st(v)$ is complete we are done, if not then we decompose $A_{st(v)}$ again and proceed as before. This process stops at some point since the graph $\Gamma$ is finite and we remove at least one vertex at every step of the decomposition.
\end{proof}
Before moving to the proof of Theorem \ref{MainTheorem1} we recall a basic result about $lcH$-slender groups. By definition, a discrete group $G$ is called $lcH$-slender if every group homomorphism from a locally compact Hausdorff group into $G$ is continuous. From here on we will be dealing with locally compact Hausdorff groups and their basic properties. Everything that will be used can be found in \cite{Stroppel} and theorems will be cited, however not every basic property will be cited in that way.
\begin{lemma}\label{lchtorsion}
Let $G$ be an lcH-slender group. Then $G$ is torsion free.
\end{lemma}
\begin{proof}
Suppose there exists a non-trivial torsion element $g\in G$. Without loss of generality we can assume that $g$ has order $p\in \mathbb{N}$ and $p$ is a prime number (if it is composite, there is a power of $g$ having a prime order). Then $\langle g\rangle\cong \mathbb{Z}/p\mathbb{Z}$. Now consider the group $\prod_{\mathbb{N}}\mathbb{Z}/p\mathbb{Z}$. This is a compact topological group and also a vector space. The space $\bigoplus_{\mathbb{N}} \mathbb{Z}/p\mathbb{Z}$ is a linear subspace.
We define $\psi\colon\bigoplus_{\mathbb{N}} \mathbb{Z}/p\mathbb{Z}\to \mathbb{Z}/p\mathbb{Z}$ by setting $\psi (x):=\sum_{m\in \mathbb{N}} x_m$. Now we take a basis $B$ of $\bigoplus_{\mathbb{N}} \mathbb{Z}/p\mathbb{Z}$ and extend it to a basis $C$ of $\prod_{\mathbb{N}}\mathbb{Z}/p\mathbb{Z}$. We then extend the map $\psi$ to a map $\varphi\colon \prod_{\mathbb{N}}\mathbb{Z}/p\mathbb{Z}\to \mathbb{Z}/p\mathbb{Z}$ via linear extension of:
$$\varphi(c_j):=\begin{cases} \psi(c_j)&\text{ if }c_j\in B\\
0&\text{ if }c_j\in C-B\\
\end{cases}$$
The linearity of this map guarantees this is a group homomorphism. However this map cannot be continuous because of the following reason: If it was continuous, then $\varphi^{-1}(0)$ would be an open neighborhood $V$ of the identity, since $\{0\}\subset \mathbb{Z}/p\mathbb{Z}$ is an open subset. So (after reordering the factors) $V$ contains a set $A_1\times A_2\times...\times A_m \times \prod_{n>m}\mathbb{Z}/p\mathbb{Z}$. But then the elements $x=(0,0,...)$ and $y=(0,...,0,1,0,0,...)$, where the $1$ is in coordinate $m+1$, both lie in $V$ and we have $ 0=\varphi(x)=\varphi(y)=\varphi(x)+1=1$ which is a contradiction.
\end{proof}
\begin{theorem}
Let $A_\Gamma$ be an Artin group in the class $\mathcal{A}$.
\begin{enumerate}
\item Let $\psi\colon L\to A_\Gamma$ be a group homomorphism from a locally compact Hausdorff group $L$ into $A_\Gamma$. If $L$ is almost connected, then $\psi(L)$ is contained in a conjugate of a special complete subgroup of $A_\Gamma$.
\item If all special complete subgroups of $A_\Gamma$ are $lcH$-slender, then $A_\Gamma$ is $lcH$-slender.
\end{enumerate}
\end{theorem}
\begin{proof}
It is known that an almost connected locally compact Hausdorff group has property $\mathcal{FA'}$ \cite[Cor. 1]{Alperin}. Thus, by Proposition \ref{FA}, it follows that $\psi(L)$ is contained in a complete parabolic subgroup since property $ \mathcal{FA'}$ is preserved under images of homomorphisms.
Let $\varphi\colon L\to A_\Gamma$ be a group homomorphism from a locally compact Hausdorff group $L$ into an Artin group $A_\Gamma$ that lies in the class $\mathcal{A}$. We give $A_\Gamma$ the discrete topology. By (1) we know that the image of the connected component $L^\circ$ under $\varphi$ is contained in a parabolic complete subgroup $gA_\Delta g ^{-1}$ of $A_\Gamma$. By assumption, any special complete subgroup of $A_\Gamma$ is $lcH$-slender, therefore the restricted group homomorphism $\varphi_{\mid L^\circ}\colon L^\circ\to gA_\Delta g^{-1}$ is continuous. Since the image of a connected group under a continuous group homomorphism is connected, the group $\varphi(L^\circ)$ is trivial and therefore the map $\varphi$ factors through $\psi\colon L/L^\circ\to A_\Gamma$. Since the group $L/L^\circ$ is totally disconnected there exists a compact open subgroup $K\subset L/L^\circ$ by van Danzig's Theorem \cite[III §4, No.6]{Bourbaki}. By (1) we know that $\psi(K)$ is contained in a parabolic complete subgroup $hA_\Omega h^{-1}$. By assumption we know that $\psi_{\mid K}$ is continuous. The image of a compact set under a continuous map is always compact, hence $\psi(K)$ is finite. We also know that a $lcH$-slender group is always torsion free due to Lemma \ref{lchtorsion}, thus $\psi(K)$ is trivial. Hence the map $\psi$ has open kernel, since it contains the compact open group $K$ and is therefore continuous. Since the quotient map $\pi\colon L\to L/L^\circ$ is continuous, so is $\varphi=\psi\circ \pi$.
\end{proof}
\section{The clique-cube complex}
In this section we will be using CAT$(0)$ spaces and their basic properties. For the definition and further properties of these metric spaces we refer to \cite{BridsonHaefliger}.
Associated to an Artin group $A_\Gamma$ is a CAT$(0)$ cube complex $C_\Gamma$ where the dimension is bounded above by the cardinality of $V(\Gamma)$. We describe the construction of this cube complex that is closely related to the Deligne complex introduced in \cite{CharneyDavis}.
For an Artin group $A_\Gamma$ we consider the poset
$$\left\{aA_\Delta\mid a\in A_\Gamma\text{ and }\Delta\text{ is a complete subgraph of }\Gamma \text{ or }\Delta=\emptyset\right\}\,.$$
This poset is ordered by inclusion. We now construct the cube complex $C_\Gamma$ in the usual way: The vertices are the elements in the poset and two vertices $aA_{\Delta_1}\subsetneq bA_{\Delta_2}$ span an $n$-cube if $|\Delta_2|-|\Delta_1|=n$.
The group $A_\Gamma$ acts on $C_\Gamma$ by left multiplication and preserves the cubical structure. Moreover, the action is \emph{strongly cellular} i.e. the stabilizer group of any cube fixes that cube pointwise. Therefore, if a subgroup $H\subset A_\Gamma$ has a global fixed point in $C_\Gamma$, then there exists a vertex in $C_\Gamma$ which is fixed by $H$.
The action is cocompact with the fundamental domain $K$, which is the subcomplex spanned by all cubes with vertices $1A_\Delta$ for $\Delta\subset \Gamma$ a complete subgraph. However the action will in general not be proper, since the stabilizer of a vertex $gA_X$ for $X\neq \emptyset$ is the parabolic complete subgroup $gA_Xg^{-1}$.
\begin{theorem}[\cite{GodelleParisB}, Thm. 4.2]
The clique-cube complex $C_\Gamma$ is a finite-dimensional CAT$(0)$ cube complex.
\end{theorem}
It was proven in \cite[Thm. A]{LederVarghese} that any cellular action of a finitely generated group on a finite dimensional CAT(0) cube complex via elliptic isometries always has a global fixed point. In general, this result is not true for not finitely generated groups, not even for actions on trees.
\begin{example}[\cite{Serre} Thm. 15] Consider the group $\mathbb{Q}$ and pick an infinite sequence of elements $\left(g_i\right)_{i\in \mathbb{N}}$ such that $\langle g_1\rangle \subsetneq \langle g_1,g_2\rangle\subsetneq...$. This is possible since $\mathbb{Q}$ is not finitely generated. Set $G_i:=\langle g_1,g_2,...,g_i\rangle$ for $i\in \mathbb{N}$ and define a graph $\Gamma$ in the following way. The set of vertices of $\Gamma$ is the disjoint union of $\mathbb{Q}/G_n$, i.e. the vertices are the cosets $qG_n$ and there is an edge between two vertices if and only if they correspond to consecutive $\mathbb{Q}/G_n$ and $\mathbb{Q}/G_{n+1}$ and correspond under the canonical projection $\mathbb{Q}/G_n\to \mathbb{Q}/G_{n+1}$.
One can now check that this is a tree with a natural action of $\mathbb{Q}$. The key feature is that if there was a fixed point $P$, then there exists an $n\in \mathbb{N}$ such that $P\in \mathbb{Q}/G_n$ and therefore $\mathbb{Q}=G_n$, a contradiction.
\end{example}
We conjecture that the structure of Artin groups does not allow such an example, that means:
\begin{conjecture}
Let $A_\Gamma$ be an Artin group and $\Phi\colon A_\Gamma\to{\rm Isom}(C_\Gamma)$ be the action on the associated clique-complex via left multiplication. Let $H\subset A_\Gamma$ be a subgroup. If $\Phi(h)$ is elliptic for all $h\in H$, then ${\rm Fix}(\Phi(H))$ is non-empty, thus $\Phi(H)$ is contained in a complete parabolic subgroup.
\end{conjecture}
Before we prove that this conjecture holds for Artin groups in the class $\mathcal{B}$ we discuss a tool for proving that some fixed point sets are non-empty.
We have
$${\rm Fix}(\Phi(H))=\bigcap\limits_{h\in H}{\rm Fix}(\Phi(h))\,.$$
By definition, a family of subsets $(A_i)_{i\in I}$ of a metric space is said to have the finite intersection property if the intersection of each finite subfamily is non-empty. Monod proved in \cite[Thm. 14]{Monod} that a family consisting of bounded closed convex subsets of a complete CAT$(0)$ space with the finite intersection property has a non-empty intersection.
We consider the family consisting of fixed point sets $({\rm Fix}(\phi(h)))_{h\in H}$. Any fixed point set is closed and convex. Since the CAT(0) space $C_\Gamma$ is a finite dimensional cubical complex we know by \cite[Thm. A]{LederVarghese} that this family has the finite intersection property, but in general a fixed point set does not need to be bounded as the following example shows. Thus we need a different strategy in order to prove that ${\rm Fix}(\phi(H))$ is non-empty.
\begin{example}
Let $\Gamma$ be the following graph:
\begin{center}
\begin{tikzpicture}
\coordinate[label=above: $a$] (A) at (-2,0);
\coordinate[label=above: $b$] (B) at (0,0);
\coordinate[label=above: $c$] (C) at (2,0);
\draw (A) -- (B) node[midway,above]{$2$} ;
\draw (B) -- (C) node[midway, above]{$2$} ;
\fill (A) circle (2pt);
\fill (B) circle (2pt);
\fill (C) circle (2pt);
\end{tikzpicture}
\end{center}
Now we consider the corresponding Artin group $A_\Gamma$ and the clique-complex $C_\Gamma$. For the clique-complex the fundamental domain $K$ is given by
\begin{center}
\begin{tikzpicture}[scale=0.5]
\coordinate[label=below:{$A_{\{a\}}$}] (A) at (-4,0);
\coordinate[label=below:{$A_{\emptyset}$}] (B) at (0,0);
\coordinate[label=below:{$A_{\{c\}}$}] (C) at (4,0);
\coordinate[label=above:{$A_{\{a,b\}}$}] (D) at (-4,4);
\coordinate[label=above:{$A_{\{b\}}$}] (E) at (0,4);
\coordinate[label=above:{$A_{\{b,c\}}$}] (F) at (4,4);
\fill (A) circle (2pt);
\fill (B) circle (2pt);
\fill (C) circle (2pt);
\fill (D) circle (2pt);
\fill (E) circle (2pt);
\fill (F) circle (2pt);
\draw (A) -- (B);
\draw (B) -- (C);
\draw (A) --(D);
\draw (D) -- (E);
\draw (E) -- (F);
\draw (C) -- (F);
\draw (B) -- (E);
\fill[color=black, opacity=0.3] (A) -- (B) -- (E) -- (D) ;
\fill[color=black, opacity=0.3] (B) -- (C) -- (F) -- (E) ;
\end{tikzpicture}
\end{center}
Let $\Phi\colon A_\Gamma \to {\rm Isom}(C_\Gamma)$ denote the natural action by left multiplication. We want to show that ${\rm Fix}(\Phi(b))$ is unbounded. First notice that since $\Gamma$ is not complete, the clique-complex is unbounded itself. First we calculate $K\cap {\rm Fix}(\Phi(b))$:
Since $b\neq a^n$ and $b\neq c^k$ for any $n,k\in\mathbb{Z}$, $b$ does not fix $A_{\{a\}}$ or $A_{\{c\}}$ and it also can't fix $A_\emptyset$. However $b$ fixes all the vertices in the top row, that is $\Phi(b)(A_{\{a,b\}})=A_{\{a,b\}}$, $\Phi(b)(A_{\{b\}})=A_{\{b\}}$ and $\Phi(b)(A_{\{b,c\}})=A_{\{b,c\}}$ and also the edges between those.
In general we know that $b$ lies in the center of $A_\Gamma$, so following the pattern from above it is easy to see that $\Phi(b)$ fixes vertices of the following form: $wA_\Delta$, where $w\in A_\Gamma$ is an arbitrary element and $\Delta$ is a subgraph of $\Gamma$ containing the vertex $b$. That means any copy of the `top edge' in the fundamental domain is fixed by $\Phi(b)$. Since $K$ is a fundamental domain for the action and $C_\Gamma$ is unbounded, it immediately follows that ${\rm Fix}(\Phi(b))$ is unbounded, too.
\end{example}
Let $A_\Gamma$ be in the class $\mathcal{B}$ and $H$ be a subgroup that is contained in a parabolic complete subgroup. Since the action of $A_\Gamma$ on $C_\Gamma$ is strongly cellular, we can write the parabolic closure of a subgroup $H$ as $$PC_\Gamma(H):=\bigcap\limits_{v\in V(C_\Gamma)}\{{\rm stab}(v)\mid H\subset {\rm stab}(v)\},$$ which coincides with the notion defined in Chapter 2.
\begin{proposition}
\label{globalfixedpoint}
Let $A_\Gamma$ be in $\mathcal{B}$ and $\Phi\colon A_\Gamma\to{\rm Isom}(C_\Gamma)$ be the action on the associated clique-complex via left multiplication.
Let $H\subset A_\Gamma$ be a subgroup. If $\Phi(h)$ is elliptic for all $h\in H$, then ${\rm Fix}(\Phi(H))$ is non-empty and therefore $\Phi(H)$ is contained in a complete parabolic subgroup.
\end{proposition}
\begin{proof}
If $\Phi(H)$ had no global fixed point, then we could construct an infinite chain
$${\rm Fix}(\Phi(H_1))\supsetneq {\rm Fix}(\Phi(H_2))\supsetneq...,$$
where $H_i:=\langle h_1,h_2,...,h_i\rangle$, since any finitely generated group acting locally elliptically already has a global fixed point due to \cite[Thm. A]{LederVarghese}.
Now we claim that for a subset $B$ of a parabolic complete subgroup we have
$${\rm Fix}(\Phi(B))={\rm Fix}(\Phi(PC_\Gamma(B)))\,.$$
The inclusion $\supset$ is clear since $B\subset PC_\Gamma(B)$. For the other directions, let $v\in {\rm Fix}(\Phi(B))$. It suffices to check that $v\in {\rm Fix}(\Phi(PC_\Gamma(B)))$ since the action is strongly cellular. Since $B$ fixes the vertex $v$, we have $B\subset {\rm stab}(v)$. Due to the definition of the parabolic closure, we therefore obtain $PC_\Gamma(B)\subset {\rm stab}(v)$, and thus $PC_\Gamma(B)$ fixes $v$ as well, or in other words, $v\in {\rm Fix}(\Phi(PC_\Gamma(B)))$.
So, our chain transforms into
$${\rm Fix}(\Phi(PC_\Gamma(H_1)))\supsetneq {\rm Fix}(\Phi(PC_\Gamma(H_2)))\supsetneq...$$
On the other hand, since $H_i \subset H_{i+1}$ for all $i$, we have the chain
$$PC_\Gamma(H_1)\subset PC_\Gamma(H_2)\subset...$$
Since $\Gamma$ is finite, this chain stabilizes by Proposition \ref{ArtinVertices}.
So, there exists an index $j$ such that $PC_\Gamma(H_i)=PC_\Gamma(H_j)$ for all $i\ge j$, hence
$${\rm Fix}(\Phi(PC_\Gamma(H_i)))={\rm Fix}(\Phi(PC_\Gamma(H_j))) $$
for all $i\ge j$, which is a contradiction.
Therefore $\Phi(H)$ has a global fixed vertex. Since the stabilizer of a vertex is a complete parabolic subgroup, $\Phi(H)$ is contained in such a group.
\end{proof}
This implies Proposition \ref{FC} from the introduction, since an F$\mathcal{C}'$ group acts locally elliptically on every CAT$(0)$ cube complex.
\begin{corollary}\label{compact}
Let $K$ denote a compact group and $A_\Gamma$ an Artin group in the class $\mathcal{B}$, further let $\phi\colon K\to A_\Gamma$ be a group homomorphism. Then $\phi(K)$ is contained in a complete parabolic subgroup.
\end{corollary}
\begin{proof}
Due to \cite[Cor. 4.4]{MoellerVarghese} the compact group $K$ acts locally elliptically on $C_\Gamma$. Thus by Proposition \ref{globalfixedpoint} the image of $K$ under $\phi$ is contained in a complete parabolic subgroup.
\end{proof}
\subsection{Proof of Theorem \ref{MainTheorem2}}
First we recall the result of the Main Theorem in \cite{MoellerVarghese}.
\begin{theorem}
Let $\Phi\colon L\to{\rm Isom}(X)$ be a group action of an almost connected locally compact Hausdorff group $L$ on a complete CAT$(0)$ space $X$ of finite flat rank. If
\begin{enumerate}
\item the action is semi-simple,
\item the infimum of the translation lengths of hyperbolic isometries is positive,
\item any finitely generated subgroup of $L$ which acts on $X$ via elliptic isometries has a global fixed point,
\item any subfamily of $\left\{{\rm Fix}(\Phi(l)) \mid l\in L\right\}$ with the finite intersection property
has a non-empty intersection,
\end{enumerate}
then $\Phi$ has a global fixed point.
\end{theorem}
\begin{proof}[Proof of Theorem \ref{MainTheorem2}]
The proof of the second part is very similar to the proof of Theorem \ref{MainTheorem1}.
Let $\psi\colon L\to A_\Gamma$ be a group homomorphism from an almost connected locally compact Hausdorff group $L$ into an Artin group $A_\Gamma$ that is contained in the class $\mathcal{B}$. Further, let $\Phi\colon A_\Gamma\to{\rm Isom}(C_\Gamma)$ be the action on the associated clique-cube complex via left-multiplication.
Since the dimension of $C_\Gamma$ is finite we know that the flat rank of $C_\Gamma$ is also finite. Further, the first three conditions are satisfied by any cellular action on a finite dimensional CAT$(0)$ cube complex \cite[Thm. A and Prop.]{Bridson}, \cite[Thm. A]{LederVarghese}, hence also the action $\psi\circ\Phi\colon L\to A_\Gamma\to{\rm Isom}(C_\Gamma)$ satisfies these conditions. By Proposition \ref{globalfixedpoint} it follows that any subfamily of $\left\{{\rm Fix}(\psi\circ\Phi(l)) \mid l\in L\right\}$ with the finite intersection property has a non-empty intersection. Hence, the action $\psi\circ\Phi$ has a global fixed point. Since the action is strongly cellular there exists a vertex $gA_\Delta \in C_\Gamma$ that is fixed by this action and therefore $\psi(L)$ is contained in the stabilizer of this vertex that is equal to $gA_\Delta g^{-1}$.
Let $\varphi\colon L\to A_\Gamma$ be a group homomorphism from a locally compact Hausdorff group $L$ into an Artin group $A_\Gamma$ that lies in the class $\mathcal{B}$. Once again we give $A_\Gamma$ the discrete topology. By the above paragraph we know that the image of the connected component $L^\circ$ under $\varphi$ is contained in a parabolic complete subgroup $gA_\Delta g ^{-1}$ of $A_\Gamma$. Now follow the same argument as given in the proof of Theorem \ref{MainTheorem1} (2).
\end{proof}
|
\section{Introduction} \label{section_introduction}
The voracious data demands have promoted the rapid development of wireless communications. Although terrestrial cellular systems have made great success in the past decades, nearly half of the places on the earth have deficient Internet connectivity, e.g., in the sea, dessert, mountain and rural areas \cite{Di2019UltraDenseLEO}.
To complement and extend existing terrestrial cellular systems, satellite communication (SATCOM) emerges as a crucial technology, and has been considered in 5G new radio (NR) for the non-terrestrial networks (NTN) \cite{3GPP_NonTerrestrial}.
In particular, the low earth orbit (LEO) satellites, deployed between $500$ km to $2000$ km altitudes, are highly promising to provide low-latency broadband Internet access for the user terminals (UTs) around the world, on account of their superiority on the short round-trip delay, reduced pathloss, low manufacture and launch cost, etc.
Multibeam satellites have been a popular research topic in SATCOM.
In multibeam satellite systems, multiple color reuse schemes are usually employed to reduce the inter-beam interference, in which the orthogonal polarizations or different frequency bands are assigned to the adjacent spot beams. However, these schemes only allow the spot beams located far from each other to share the frequency bands, thus restricting the system capacity.
To fully exploit the scarce spectrum, full frequency reuse (FFR) schemes, where the frequency bands are reused among all spot beams, have been adopted in multibeam SATCOM, and result in higher system capacity if the inter-beam interference is properly treated \cite{Vazquez2016PrecodingChallenges}.
In the uplink (UL) SATCOM, multiple UTs on ground send messages to the satellite simultaneously, and the sum rate capacity can be achieved by performing the successive interference cancellation (SIC) decoding for the received signals \cite{Goldsmith2003CapacityMIMO}.
The UL multibeam SATCOM has been considered in some existing works, most focusing on the capacity analysis, e.g., \cite{Letzepis2008CapacitySateRician,Christopoulos2011CapacityMultibeamUplink,Yang2016UpperBoundGEOMSCUplink,Arnau2014PerformanceReturnLink}.
In \cite{Letzepis2008CapacitySateRician}, the sum rate capacity for both non-fading and Rician fading UL multibeam satellite channels is studied.
The lower and upper bounds on the ergodic sum rate capacity of the UL multibeam SATCOM are derived in \cite{Christopoulos2011CapacityMultibeamUplink} and \cite{Yang2016UpperBoundGEOMSCUplink}, respectively. In \cite{Arnau2014PerformanceReturnLink}, the ergodic capacity and outage capacity are analyzed by considering the spatially correlated rain attenuation in the UL multibeam satellite channels.
Conventional multibeam satellites usually adopt the fixed or infrequently modified beamforming \cite{PerezNeira2019SPforHTS}, which cannot adapt to the real-time link conditions.
Nowadays, massive multiple-input multiple-output (MIMO) has been a pivotal technique in terrestrial 5G systems \cite{Marzetta2010Noncooperative}. With a large number of antennas at the base station (BS), massive MIMO can exploit multiple reconfigurable beams with high angular resolution to further improve the spectrum and energy efficiency \cite{Hien2013ESEfficiency}.
Enlightened by the success of massive MIMO in terrestrial 5G communications, in this paper, we consider massive MIMO LEO SATCOM where an LEO satellite is equipped with a large number of antennas and generates reconfigurable beams by using digital beamforming, and explore an further improvement of the UL SATCOM system capacity.
The massive MIMO transmission in LEO SATCOM depends on the channel state information (CSI). Although the receivers may accurately estimate the instantaneous CSI (iCSI) via channel estimation techniques, it is quite challenging for the transmitters to track the iCSI due to the inherent LEO satellite channel impediments, such as the large propagation delays and Doppler shifts compared to the terrestrial wireless channels \cite{Guidotti2017SatelliteLTELEO}. In contrast to the iCSI, the statistical CSI (sCSI) is usually stable for longer time intervals \cite{You2019MassiveMIMOLEO,KX2021DownlinkMassiveMIMOLEO}, and thus can be more accurately acquired. Hence, it is more practical to use the sCSI at the transmitters (sCSIT) for massive MIMO transmission in LEO SATCOM, which is considered in the present paper.
The problem of UL transmit designs with only sCSIT has been extensively studied for multiuser MIMO (MU-MIMO) \cite{Soysal2007PowerMIMOMACPartCSI,Soysal2009OptBeamformingMIMOMAC,Li2010CapacityBoundMIMOMAC,Li2010CapacityMIMOMACLowSNR,Wen2011SumRateMIMORician,Wen2013DEMIMOMAC} and massive MIMO \cite{Zhang2013CapacityMIMOMAC,Lu2016FreeDeterministic} communications.
However, these works do not consider the LEO satellite channel characteristics and more importantly have high implementation complexity.
Thus, it is of great importance to design UL transmit strategies, tailored for massive MIMO LEO SATCOM systems.
Recently, a massive MIMO LEO satellite downlink (DL) transmission approach was proposed in \cite{You2019MassiveMIMOLEO}, where the LEO satellite channel model, the DL linear precoders and the user grouping strategy were carefully designed, by assuming that each UT has only a single antenna. By deploying multiple antennas at UTs, the DL transmit design and its low-complexity implementations for massive MIMO LEO satellite communications were studied in \cite{KX2021DownlinkMassiveMIMOLEO}.
The UL massive MIMIO transmit strategies for LEO SATCOM with multiple antennas at both the satellite and UTs have not been investigated.
In this paper, we propose a practical UL transmit design that uses sCSIT in massive MIMO LEO SATCOM systems, where both the satellite and the UTs are equipped with uniform planar arrays (UPAs).
First, we derive the UL massive MIMO LEO satellite channel model under the UPA configurations at both the satellite and the UTs. Based on the presented massive MIMO LEO satellite channel model, we study the optimal UL transmit design that achieves the ergodic sum rate capacity and its low-complexity implementations, by exploiting the long-term sCSIT. Our major contributions are summarized as follows.
\begin{itemize}
\item We show that the rank of each UT's transmit covariance matrix that achieves the UL ergodic sum rate capacity is no larger than that of its own channel correlation matrix at the UT side, thus revealing the maximum number of independent data streams that can be delivered from each multi-antenna UT to the satellite.
\item Then, we find that each UT's transmit covariance matrix can be fully determined by a lower-dimensional matrix. Furthermore, by deriving the structure of the optimal lower-dimensional matrices, we develop an algorithm based on the stochastic programming to compute these lower-dimensional matrices.
\item To avoid the computationally expensive sample average in the stochastic programming, we approximate the ergodic sum rate with its asymptotic expression and develop a computationally efficient algorithm based on asymptotic programming to compute the UTs' transmit covariance matrices.
\item Simulation results show that the proposed UL transmit designs provide better performance than the existing UL LEO SATCOM transmit strategies, while the low-complexity asymptotic programming based UL transmit design can achieve near-optimal performance in massive MIMO LEO SATCOM systems.
\end{itemize}
The remainder of this paper is organized as follows. \Cref{section_system_model} introduces the system model, where the UL channel model is established for the satellite and the UTs both equipped with UPAs. \Cref{section_capacity_achieve_transmit_design} shows the properties of the transmit covariance matrices, and proposes two algorithms to compute them by maximizing the UL ergodic sum rate. \Cref{section_simulation} provides the simulation results, and \Cref{section_conclusion} concludes this paper.
\textit{Notations:} Throughout this paper, lower case letters denote scalars, and boldface lower (upper) letters denote vectors (matrices). The set of all $n$-by-$m$ complex (real) matrices is denoted as $\bbC^{n\times m}$ ($\bbR^{n\times m}$). The trace, determinant, rank, conjugate, transpose, and conjugate transpose for matrix are represented by $\trace(\cdot)$, $\det(\cdot)$, $\rank(\cdot)$, $(\cdot)^*$, $(\cdot)^T$, and $(\cdot)^H$, respectively. The Euclidean norm of vector $\bdx$ is denoted as $\norm{\bdx} = \sqrt{\bdx^H \bdx}$. The identity matrix is represented by $\bdI$ or $\bdI_n$. $\bdone$ and $\bdzro$ denote all-one and all-zero vectors, respectively. Denote $\otimes$ as the Kronecker product. Let $[\bdA]_{n,m}$ represent the $(n,m)$th element of matrix $\bdA$. The diagonal matrix with $\bdx$ along its main diagonal is denoted as $\diag(\bdx)$. $\bdx^{1/2}$ and $\bdx^{-1}$ represent the element-wise square root and reciprocal of $\bdx$, respectively. $\bbE \{ \cdot \}$ means mathematical expectation. $\clCN(\bdm,\bdC)$ denotes the proper complex Gaussian random vector with mean vector $\bdm$ and covariance matrix $\bdC$.
The uniform distribution between $a$ and $b$ is denoted as $\mathrm{U}\ (a,b)$.
\section{System Model} \label{section_system_model}
We consider the UL transmission in FFR massive MIMO LEO SATCOM systems operating at the lower frequency bands, e.g., L/S/C bands. As depicted in \Cref{fig_uplink_UPA}, the mobile UTs on the ground send messages to an LEO satellite at an altitude of $H$.
The satellite and the mobile UTs use UPAs with digital active antennas, which allow the amplitude and phase of each antenna element of the UPAs to be digitally adjusted. The UPA at the satellite has $\Mx$ and $\My$ directional elements in the $\rx$-axis and $\ry$-axis, respectively. Thus, the satellite has $\Mx \My \triangleq M$ antennas.
Meanwhile, each UT uses the UPA consisting of $\Nx$ and $\Ny$ omnidirectional antenna elements in the $\rx'$-axis and $\ry'$-axis, respectively. Hence, there are $\Nx \Ny \triangleq N$ antennas at each UT.
Note that the antenna configurations can be extended to the cases that the UPA of each UT has a distinct number of antenna elements.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\textwidth]{satellite_uplink.eps}
\caption{The UL of massive MIMO LEO SATCOM.}
\label{fig_uplink_UPA}
\end{figure}
\subsection{Signal and Channel Models in Analog Baseband} \label{subsection_UL_channel_model}
In this subsection, we elaborate the UL signal and channel models in the analog baseband for LEO SATCOM.
The received signal at the satellite at time instant $t$ is given by
\begin{equation}
\bdy(t) = \sum_{k} \int_{-\infty}^{\infty} \ckbdH_k(t,\tau) \bdx_k(t-\tau) \dint \tau + \bdz(t)\comma \label{RxSignal_time-domain}
\end{equation}
where $\ckbdH_k(t,\tau) \in \Complex{M}{N}$ and $\bdx_k(t) \in \Complex{N}{1}$ are the channel impulse response and the transmit signal of UT $k$, and $\bdz(t) \in \Complex{M}{1}$ is the additive noise signal at the satellite.
The time-varying channel impulse response $\ckbdH_k(t,\tau)$ can be written as
\begin{equation}
\ckbdH_k(t,\tau) = \sum_{\ell=0}^{L_k-1} a_{k,\ell} e^{j2\pi \nu_{k,\ell} t } \delta(\tau-\tau_{k,\ell}) \bdg_{k,\ell} \bdd_{k,\ell}^H\comma \label{channel_model_UT_k}
\end{equation}
where $\delta(x)$ is the Dirac delta function, $j\triangleq \sqrt{-1}$, $L_k$ is the multipath number, $a_{k,\ell}$, $\nu_{k,\ell}$ and $\tau_{k,\ell}$ are the complex channel gain, the Doppler shift and the propagation delay for the $\ell$th path of UT $k$'s channel. Besides, $\bdg_{k,\ell} \in \Complex{M}{1}$ and $\bdd_{k,\ell} \in \Complex{N}{1}$ are the array response vectors for the $\ell$th path of UT $k$'s channel at the satellite and the UT sides, respectively.
It is worth noting that in LEO SATCOM, the Doppler shifts $\nu_{k,\ell}$'s and the propagation delays $\tau_{k,\ell}$'s are much larger than those in terrestrial wireless communications. The Doppler shift $\nu_{k,\ell}$ in \eqref{channel_model_UT_k} can be separated as $\nu_{k,\ell} = \nu_{k,\ell}^{\sat} + \nu_{k,\ell}^{\ut}$, where $\nu_{k,\ell}^{\sat}$ and $\nu_{k,\ell}^{\ut}$ are the Doppler shifts due to the motion of the satellite and UT $k$, respectively. Furthermore, the Doppler shifts $\nu_{k,\ell}^{\sat}$, $0\le \ell \le L_k-1$, tend to be identical for different paths of UT $k$'s channel \cite{Papath2001Acomparison}. Hence, we can rewrite $\nu_{k,\ell}^{\sat} = \nu_{k}^{\sat}$, $0\le \ell \le L_k-1$. The minimal and maximal propagation delays of UT $k$'s channel are denoted by $\tau_{k}^{\min} \triangleq \min_{\ell} \tau_{k,\ell}$ and $\tau_k^{\max} \triangleq \max_{\ell} \tau_{k,\ell} $, respectively.
Denote $\bdtheta_{k,\ell} = (\theta_{k,\ell}^{\rx},\theta_{k,\ell}^{\ry})$ and $\bdvphi_{k,\ell} = (\vphi_{k,\ell}^{\rx'},\vphi_{k,\ell}^{\ry'})$ as the paired angles-of-arrival (AoAs) and angles-of-departure (AoDs) related to the $\ell$th path of UT $k$'s channel. Then, $\bdg_{k,\ell}$ and $\bdd_{k,\ell}$ in \eqref{channel_model_UT_k} can be expressed as $\bdg_{k,\ell} = \bdg(\bdtheta_{k,\ell})$ and $\bdd_{k,\ell} = \bdd(\bdvphi_{k,\ell})$, respectively. Here, $\bdg(\bdtheta)$ and $\bdd(\bdvphi)$ for arbitrary $\bdtheta=(\theta_{\rx},\theta_{\ry})$ and $\bdvphi=(\vphi_{\rx'},\vphi_{\ry'})$ are defined as
\begin{subequations}
\begin{align}
\bdg(\bdtheta) &= \bda_{\Mx} \left( \sin \theta_{\ry} \cos \theta_{\rx} \right) \otimes \bda_{\My} \left( \cos \theta_{\ry} \right)\comma \\
\bdd(\bdvphi) &= \bda_{\Nx} \left( \sin \vphi_{\ry'} \cos \vphi_{\rx'} \right) \otimes \bda_{\Ny} \left( \cos \vphi_{\ry'} \right)\comma
\end{align}
\end{subequations}
respectively.
Also, $\bda_{\nv}(x) \in \Complex{\nv}{1}$ is defined as
\begin{align}
\bda_{\nv} \left( x \right) = \frac{1}{\sqrt{\nv}} \left( 1, e^{-j \frac{2\pi d_{\rv} }{ \lambda } x }, \dots, e^{-j\frac{2\pi d_{\rv} }{ \lambda } (\nv-1) x } \right)^T\comma
\end{align}
where $\lambda=c/f_c$ is the carrier wavelength, $c$ is the speed of the light, $f_c$ is the carrier frequency, $d_{\rv}$ is the spacing between adjacent antennas along the $\rv$-axis with $\rv \in \left\{ \rx,\ry,\rx',\ry' \right\}$.
Moreover, owing to the high altitude of the satellite, the paired AoAs for different paths of UT $k$'s channel are nearly identical, i.e., $\bdtheta_{k,\ell} = \bdtheta_k$, $0 \le \ell \le L_k-1$ \cite{3GPP_NonTerrestrial}. Thus, we can discard the subscript of the path $\ell$ in $\bdg_{k,\ell}$ and rewrite it as $\bdg_{k,\ell} = \bdg_{k} = \bdg(\bdtheta_k)$,
where $\bdtheta_k = (\theta_k^{\rx},\theta_k^{\ry})$ is referred to as the physical AoA pair of UT $k$.
It is worth noting that in LEO SATCOM, $\bdg_k$ changes quite slowly due to the long distance between the satellite and UT $k$.
Furthermore, we define $\bdalpha_k = (\alpha_k^{\rx},\alpha_k^{\ry})$ as the space angle pair of UT $k$, where $\alpha_k^{\rx} = \sin \theta_k^{\ry} \cos \theta_k^{\rx} $ and $\alpha_k^{\ry} = \cos \theta_k^{\ry} $.
The nadir angle $\vtheta_k$ is given by $ \vtheta_k = \cos^{-1} \left(\sin \theta_k^{\ry} \sin \theta_k^{\rx}\right)$.
\subsection{Signal and Channel Models for OFDM Based Transmission} \label{subsection_UL_Doppler_delay_compensation}
We consider that the orthogonal frequency division multiplex (OFDM) technique is adopted in the wideband satellite systems.
The number of subcarriers and cyclic prefix (CP) length are represented by $\Nsc$ and $\Ncp$, respectively. Let us denote $\Ts$ as the system sampling period. Then, the time duration of CP is given by $\Tcp = \Ncp \Ts$. Besides, the time durations for each OFDM symbol without and with CP are given by $\Tsc = \Nsc \Ts$ and $T = \Tcp + \Tsc$, respectively.
Let $\{ \bdx_{k,s,r} \}_{r=0}^{\Nsc-1}$ be the frequency-domain transmit signal of UT $k$ within the $s$th OFDM symbol. Then, the time-domain transmit signal is given by $\bdx_{k,s}(t) = \sum_{r=0}^{\Nsc-1} \bdx_{k,s,r} e^{j2\pi r \Delta f \cdot t}$, $-\Tcp \le t-sT < \Tsc$ \cite{Hwang2009OFDMSurvey}, where $\Delta f = 1/\Tsc$ is the frequency interval between adjacent subcarriers.
Let $\nu_k^{\syn} = \nu_{k}^{\sat}$ and $\tau_k^{\syn} = \tau_{k}^{\min}$. After the joint Doppler and delay compensation \cite{You2019MassiveMIMOLEO}, the transmit signal of UT $k$ in the $s$th OFDM symbol is given by $\bdx_{k,s}^{\syn}(t) = \bdx_{k,s} \left( t+\tau_k^{\syn} \right) e^{-j2\pi \nu_k^{\syn} \left( t + \tau_k^{\syn} \right) }.$
Then, the time-domain received signal at the satellite in OFDM symbol $s$ can be written as
$\bdy_{s}^{\syn}(t) = \sum_{k} \int_{-\infty}^{\infty} \ckbdH_{k}(t,\tau) \bdx_{k,s}^{\syn}(t-\tau) \dint \tau + \bdz_{s}(t)$, where $\bdz_{s}(t)\in\Complex{M}{1}$ is the additive noise signal.
Thus, the frequency-domain received signal at the satellite over subcarrier $r$ in OFDM symbol $s$ can be written as $\bdy_{s,r} = \xinv{\Tsc} \int_{sT}^{sT+\Tsc} \bdy_{s}^{\syn}(t) e^{-j2\pi r \Delta f \cdot t } \dint t$ \cite{Hwang2009OFDMSurvey}.
Let us now denote the effective channel frequency response $\bdH_k(t,f)$ of UT $k$ as
\begin{equation}
\bdH_k(t,f) = \bdg_k \left(\bdd_k(t,f)\right)^H\comma \label{Channel_matrix_Hktf_auxiliary}
\end{equation}
where $\bdd_k(t,f) = \sum_{\ell=0}^{L_k-1} a_{k,\ell}^* e^{ -j2\pi\left( \nu_{k,\ell}^{\ut} t - f \tau_{k,\ell}^{\ut} \right) } \bdd_{k,\ell} \in \Complex{N}{1}$ and $\tau_{k,\ell}^{\ut} = \tau_{k,\ell} - \tau_{k}^{\min}$.
Consequently, the frequency-domain received signal $\bdy_{s,r}$ can be expressed as
\begin{equation}
\bdy_{s,r} = \sum_{k} \bdH_{k,s,r} \bdx_{k,s,r} + \bdz_{s,r} \comma \label{RxSignal_ysr_Hksr_xksr}
\end{equation}
where $\bdH_{k,s,r}$ is the channel matrix of UT $k$, $\bdz_{s,r}$ is the additive Gaussian noise, both on subcarrier $r$ in OFDM symbol $s$. Furthermore, $\bdH_{k,s,r}$ in \eqref{RxSignal_ysr_Hksr_xksr} can be written as
\begin{equation}
\bdH_{k,s,r} = \bdH_k(sT,r\Delta f) = \bdg_k \bdd_{k,s,r}^H\comma
\end{equation}
where $\bdd_{k,s,r} = \bdd_k(sT,r\Delta f)$.
Now that the Doppler and delay effects have been compensated at each UT, the time and frequency at the satellite and the UTs can be viewed as perfectly synchronized.
\subsection{Satellite Channel's Statistical Properties}
For convenience, we omit the subscripts of OFDM symbol $s$ and subcarrier $r$ in $\bdH_{k,s,r} = \bdg_k \bdd_{k,s,r}^H$ and denote $\bdH_k = \bdg_k \bdd_k^H$ as the flat fading channel matrix of UT $k$ over a specific subcarrier.
Therein, the random vector $\bdd_k$ can be written as
\begin{equation}
\bdd_k = \sqrt{\frac{\kappa_k \beta_k}{\kappa_k+1}} \bdd_{k,0} + \sqrt{\frac{\beta_k}{\kappa_k + 1}} \tdbdd_k\comma \label{Rician_fading_dk}
\end{equation}
where $\kappa_k$ is the Rician factor, $\beta_k = \bbE \left\{ \trace(\bdH_k \bdH_k^H) \right\} = \bbE \left\{ \snorm{\bdd_k}^2 \right\}$ is the average channel power, $\bdd_{k,0}$ represents the direction of the LoS path seen at UT $k$'s side, and $\tdbdd_k$ is a circularly symmetric complex Gaussian (CSCG) random vector distributed as $\tdbdd_k \sim \clCN(\bdzro, \bdSigma_k)$ with $\trace(\bdSigma_k) = 1$. Thus, the channel matrix $\bdH_k$ with Rician fading can be expressed as follows
\begin{equation}
\bdH_k = \bdg_k \bdd_k^H = \sqrt{\frac{\kappa_k \beta_k}{\kappa_k+1}} \bdH_k^{\LoS} + \sqrt{\frac{\beta_k}{\kappa_k + 1}} \bdH_k^{\NLoS}\comma \label{Rician_fading_Hk}
\end{equation}
where $\bdH_k^{\LoS} = \bdg_k \bdd_{k,0}^H$ is the deterministic LoS component, $\bdH_k^{\NLoS} = \bdg_k \tdbdd_k^H$ is the random scattering component. Furthermore, we consider that the LEO satellite channel undergoes the correlated Rician fading, which means that the covariance matrices $\bdSigma_k$'s can be arbitrary positive semidefinite matrices.
The sCSI refers to the long-term channel parameters $\clH \triangleq \{\beta_k,\kappa_k,\bdg_k,\bdd_{k,0},\bdSigma_k\}_{\forall k}$, which are influenced by the operating frequency bands, the propagation environment around each UT, and so on. The satellite and UTs are assumed to move within such a range that the sCSI is nearly unchanged. Once the satellite or some UT moves beyond this range, these channel parameters in $\clH$ should be accordingly updated.
The channel correlation matrices of UT $k$ at the satellite and the UT sides are given by
\begin{subequations}
\begin{align}
\bdR_k^{\sat} & = \bbE \left\{ \bdH_k \bdH_k^H \right\} = \beta_k \bdg_k \bdg_k^H\comma \label{Rk_sat} \\
\bdR_k^{\ut} & = \bbE \left\{ \bdH_k^H \bdH_k \right\} = \frac{\kappa_k \beta_k}{\kappa_k+1} \bdd_{k,0} \bdd_{k,0}^H + \frac{\beta_k}{\kappa_k + 1} \bdSigma_k\comma \label{Rk_ut}
\end{align}
\end{subequations}
respectively.
It is worth noting that $\bdR_k^{\sat}$ is rank-one indicating that the arriving signals on different antennas at the satellite are highly correlated. Meanwhile, the rank of $\bdR_k^{\ut}$ is determined by the propagation environment around UT $k$.
\section{UL Transmit Design} \label{section_capacity_achieve_transmit_design}
In this section, building on the massive MIMO LEO satellite channel model in \Cref{section_system_model}, we investigate the transmit covariance matrix design that achieves the UL ergodic sum rate capacity in massive MIMO LEO SATCOM systems. First, we show that the rank of each UT's transmit covariance matrix should not exceed that of its channel correlation matrix at the UT side. Then, we show that each UT's transmit covariance matrix can be fully determined by a lower-dimensional matrix. Finally, by invoking the stochastic programming and asymptotic programming approaches, we develop two algorithms to compute the lower-dimensional matrices.
\subsection{Rank Property of Transmit Covariance Matrices}
Here, we omit the subscripts of OFDM symbol $s$ and subcarrier $r$ in $\bdx_{k,s,r}$, and denote $\bdx_k \in \Complex{N}{1}$ as the transmit signal of UT $k$ on a specific subcarrier.
In this paper, we consider a general design of the transmit signals $\bdx_k$'s, where $\bdx_k$ is a CSCG random vector with zero mean and covariance matrix $\bdQ_k = \bbE \{ \bdx_k \bdx_k^H \} \in \Complex{N}{N}$. Let us denote the eigenvalue decomposition (EVD) of $\bdQ_k$ as $\bdQ_k = \bdV_{\rQ,k} \diag(\bdxi_{\rQ,k}) \bdV_{\rQ,k}^H$, where the columns in $\bdV_{\rQ,k} \in \Complex{N}{N}$ and the elements in $\bdxi_{\rQ,k} \in \Real{N}{1}$ are the eigenvectors and eigenvalues of $\bdQ_k$, respectively. Then, the transmit signal $\bdx_k$ can be written as
\begin{equation}
\bdx_k = \bdV_{\rQ,k} \diag(\bdxi_{\rQ,k}^{1/2}) \bds_{\rQ,k}\comma
\end{equation}
where $\bds_{\rQ,k} \sim \clCN(\bdzro,\bdI_N)$.
Moreover, we also consider the sum power constraint $\trace (\bdQ_k) \le P_k$ for each UT $ k$ in the UL transmission.
We assume that $K$ mobile UTs send messages to the satellite simultaneously. The UT index set is denoted by $\clK = \left\{1,\dots,K\right\}$.
Thus, the received signal $\bdy \in \Complex{M}{1}$ at the satellite is expressed as
\begin{equation}
\bdy = \sum_{k=1}^K \bdH_k \bdx_k + \bdz\comma \label{Data_Transmission_Model_UL}
\end{equation}
where $\bdz \in \Complex{M}{1}$ is the additive Gaussian noise at the satellite. The noise signal $\bdz$ is distributed as $\bdz \sim \clCN (\bdzro, \sigma^2 \bdI_M)$.
We assume that perfect iCSI is known by the receiver at the satellite side, while only sCSI is known by the transmitters at the UTs' side.
The UL ergodic sum rate is given by
\begin{align}
I_{\SUM} & = \bbE \left\{ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^{K} \bdH_k \bdQ_k \bdH_k^H \right) \right\} \notag \\
& \stackeq{a} \bbE \left\{ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^{K} \bdd_k^H \bdQ_k \bdd_k \cdot \bdg_k \bdg_k^H \right) \right\}\comma \label{UL_Capacity}
\end{align}
where (a) follows from $\bdH_k = \bdg_k \bdd_k^H $ in \eqref{Rician_fading_Hk}.
The UL ergodic sum rate capacity in massive MIMO LEO SATCOM systems is derived via the maximization of the following problem
\begin{equation}
\clS_{\rQ}:\ \max_{ \bdQ_k \succeq \bdzro\comma\ \trace(\bdQ_k) \le P_k\comma\ \forall k \in \clK }\ \bbE \left\{ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^{K} \bdd_k^H \bdQ_k \bdd_k \cdot \bdg_k \bdg_k^H \right) \right\}. \label{Problem_UL_sum_rate_capacity}
\end{equation}
Even if the problem $\clS_{\rQ}$ is a convex program \cite{BoydConvexOptimization}, the mathematical expectation in the ergodic sum rate renders it difficult to be solved.
As a matter of fact, the rank of the matrix $\bdQ_k$ reveals the maximum independent data streams that can be delivered through UT $k$'s channel. In the following theorem, we show the rank property of the transmit covariance matrices $\{\bdQ_k\}_{k=1}^K$, which facilitates the consequent UL transmit design in massive MIMO LEO SATCOM systems.
\begin{mytheorem} \label{Proposition_1}
The transmit covariance matrices $\{\bdQ_k\}_{k=1}^K$ that achieve the UL ergodic sum rate capacity should satisfy
\begin{equation}
\rank(\bdQ_k) \le \rank \left( \bdR_k^{\ut} \right)\comma \ \forall k \in \clK. \label{Rank_Qk_UL_Edd}
\end{equation}
\end{mytheorem}
\begin{IEEEproof}
Please refer to \Cref{appendix_Proposition_1_Proof}.
\end{IEEEproof}
The rank condition of $\bdQ_k$ for the $k$th UT in \Cref{Rank_Qk_UL_Edd} holds independently of other UTs' channel correlation matrices. From \Cref{Proposition_1}, the maximum number of independent data streams transmitted from UT $k$ to the satellite should be no larger than the rank of its own channel correlation matrix $\bdR_k^{\ut}$. It can be anticipated that if there are only sparse scatterers distributed around UT $k$, $\rank(\bdQ_k)$ can be much lower than the number of antennas $N$ at UT $k$.
Next, we show that in some extreme cases including the low signal-to-noise ratio (SNR) case and the high Rician factor case, the optimal $\{\bdQ_k\}_{k=1}^K$ to the problem $\clS_{\rQ}$ are of rank-one.
\subsubsection{Low SNR Case} \label{subsubsection_low_SNR_Qmat}
If $P_k \rightarrow 0$, $\forall k \in \clK$, holds, $I_{\SUM}$ can be approximated by
\begin{align}
I_{\SUM}
\rightarrow \xinv{\sigma^2} \sum_{k=1}^K \trace \left( \bdR_k^{\ut} \bdQ_k \right). \label{sum_rate_low_SNR}
\end{align}
Then, the problem $\clS_{\rQ}$ can be simplified into
\begin{equation}
\clS_{\rQ}^{\LowSNR}:\ \max_{ \bdQ_k \succeq \bdzro\comma\ \trace(\bdQ_k) \le P_k\comma\ \forall k \in \clK }\ \sum_{k=1}^{K} \trace\left( \bdR_k^{\ut} \bdQ_k \right). \label{Problem_UL_sum_rate_capacity_lowSNR}
\end{equation}
The optimal matrices $\{\bdQ_k\}_{k=1}^K$ to the problem $\clS_{\rQ}^{\LowSNR}$ are given by
\begin{equation}
\bdQ_k = P_k \cdot \bdv_{\rQ,k} \bdv_{\rQ,k}^H\comma\ \forall k \in \clK\comma \label{Qk_optimal_low_SNR}
\end{equation}
where $\bdv_{\rQ,k} \in \Complex{N}{1}$ is the unit-norm eigenvector of $\bdR_k^{\ut}$ associated with its maximum eigenvalue.
\subsubsection{High Rician Factor Case} \label{subsubsection_high_Rician_factor_Qmat}
If $\kappa_k \rightarrow \infty$ holds for each UT $ k \in \clK$, the problem $\clS_{\rQ}$ is reduced into
\begin{equation}
\clS_{\rQ}^{\HighRiceK}:\ \max_{ \bdQ_k \succeq \bdzro\comma\ \trace(\bdQ_k) \le P_k\comma\ \forall k \in \clK }\ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^{K} \beta_k \bdd_{k,0}^H \bdQ_k \bdd_{k,0} \cdot \bdg_k \bdg_k^H \right). \label{Problem_UL_sum_rate_capacity_infRician}
\end{equation}
The optimal matrices $\{\bdQ_k\}_{k=1}^K$ to the problem $\clS_{\rQ}^{\HighRiceK}$ can be derived as follows
\begin{equation}
\bdQ_k = P_k \cdot \bdd_{k,0} \bdd_{k,0}^H,\ \forall k \in \clK. \label{Qk_optimal_high_Rician}
\end{equation}
The optimal transmit strategy for the high Rician factor case is to perform the transmit beamforming along each UT's LoS direction seen at the UT side.
In this case, owing to $\bdd_{k,0} = \bdd(\bdvphi_{k,0})$, only the paired AoDs $\bdvphi_{k,0}=(\vphi_{k,0}^{\rx'},\vphi_{k,0}^{\ry'})$ for the LoS path is required to be known at UT $k$.
Moreover, the relatively simple phased array antennas (PAAs) can be used at the UT sides to implement the beamformers $\{\sqrt{P_k} \bdd_{k,0}\}_{k=1}^K$, which can significantly reduce the hardware complexity.
\subsection{Reduce Transmit Covariance Matrix Design to Lower-Dimensional Matrix Design} \label{subsection_structure_transmit_matrix}
In this subsection, we show that the design of each UT $k$'s transmit covariance matrix can be reduced into that of a lower-dimensional matrix, which can facilitate the UL transmit design.
Let us denote the EVD of $\bdSigma_k$ as $\bdSigma_k = \bdU_k \diag(\bdlambda_k) \bdU_k^H$. The columns in $\bdU_k = [ \bdu_{k,1}\ \cdots \ \bdu_{k,S_k} ] \in \Complex{N}{S_k}$ are the eigenvectors and the elements in $\bdlambda_k =[ \lambda_{k,1},\dots,\lambda_{k,S_k}]^T$ are the corresponding positive eigenvalues in non-increasing order, where $S_k = \rank(\bdSigma_k)$.
Let us further denote the linear subspace spanned by the columns in $\bdU_k$ as $\spann(\bdU_k)$.
We can separate $\bdd_{k,0}$ into two orthogonal terms as
\begin{equation}
\bdd_{k,0} = \udbdu_{k,0} + \bdU_k \bdp_{k,0} \comma \label{dk0_uk0_pk0}
\end{equation}
where $\udbdu_{k,0} \triangleq (\bdI - \bdU_k \bdU_k^H) \bdd_{k,0}$ and $\bdp_{k,0} \triangleq \bdU_k^H \bdd_{k,0}$. The first term in \eqref{dk0_uk0_pk0} is orthogonal to the linear subspace $\spann(\bdU_k)$ and the second term lies in $\spann(\bdU_k)$.
We rewrite $\tdbdd_k$ as $\tdbdd_k = \bdU_k \udtdbdc_k$, where the elements in $\udtdbdc_k = \left[ \tdc_{k,1},\dots,\tdc_{k,S_k} \right]^T \in \Complex{S_k}{1}$ are independent CSCG random variables with distinct variance values. Indeed, $\udtdbdc_k$ is distributed as $\udtdbdc_k \sim \clCN\left(\bdzro,\diag(\bdlambda_k)\right)$.
Henceforth, $\bdd_k$ in \eqref{Rician_fading_dk} can be rewritten as
\begin{align}
\bdd_k & = \sqrt{\frac{\kappa_k \beta_k}{\kappa_k+1}} \bdd_{k,0} + \sqrt{\frac{\beta_k}{\kappa_k + 1}} \bdU_k \udtdbdc_k \notag \\
& = \sqrt{\frac{\kappa_k \beta_k}{\kappa_k+1}} \udbdu_{k,0} + \sqrt{\frac{\kappa_k \beta_k}{\kappa_k+1}} \bdU_k \bdp_{k,0} + \sqrt{\frac{\beta_k}{\kappa_k + 1}} \bdU_k \udtdbdc_k \notag \\
& = \sqrt{\frac{\beta_k}{\kappa_k + 1}} \left( \sqrt{\kappa_k} \udbdu_{k,0} + \bdU_k \left( \sqrt{\kappa_k} \bdp_{k,0} + \udtdbdc_k \right) \right). \label{dk_Bkck}
\end{align}
Let $\xi_{k,0} = \snorm{\udbdu_{k,0}}^2$ and $\bdu_{k,0} = \frac{\udbdu_{k,0}}{\snorm{\udbdu_{k,0}}}$. Define $\bdB_k \in \Complex{N}{\Sone_k}$ and $\bdc_k \in \Complex{\Sone_k}{1}$ as follows
\begin{subequations} \label{Bk_ck}
\begin{align}
\bdB_k & =
\begin{cases}
\left[ \bdu_{k,0} \ \bdU_k \right]\comma & \text{ if } \udbdu_{k,0} \ne \bdzro \\
\bdU_k\comma & \text{ if } \udbdu_{k,0} = \bdzro\comma
\end{cases} \\
\bdc_k & =
\begin{cases}
\left[
\begin{matrix}
\sqrt{\kappa_k \xi_{k,0}} \\
\sqrt{\kappa_k} \bdp_{k,0} + \udtdbdc_{k}
\end{matrix} \right] \sqrt{\frac{\beta_k}{\kappa_k + 1}}\comma & \text{ if } \udbdu_{k,0} \ne \bdzro \\
\left(\sqrt{\kappa_k} \bdp_{k,0} + \udtdbdc_{k}\right) \sqrt{\frac{\beta_k}{\kappa_k + 1}}\comma & \text{ if } \udbdu_{k,0} = \bdzro\comma
\end{cases}
\end{align}
\end{subequations}
respectively, where $\Sone_k = S_k + 1$ if $\udbdu_{k,0} \ne \bdzro$, and otherwise, $ \Sone_k = S_k $.
From \Cref{dk_Bkck,Bk_ck}, we can rewrite $\bdd_k$ as
\begin{equation}
\bdd_k = \bdB_k \bdc_k\comma
\end{equation}
where the columns in $\bdB_k$ are orthogonal to each other, i.e., $\bdB_k^H \bdB_k = \bdI$. Moreover, $\bdR_k^{\ut}$ in \eqref{Rk_ut} can be rewritten as
\begin{equation}
\bdR_k^{\ut} = \bbE \left\{ \bdH_k^H \bdH_k \right\} = \bbE \{ \bdd_k \bdd_k^H \} = \bdB_k \bdOmega_k \bdB_k^H\comma \label{Rkut_Bk_Omegak_Bk}
\end{equation}
where $\bdOmega_k \triangleq \bbE\{\bdc_k \bdc_k^H\} \in \Complex{\Sone_k}{\Sone_k}$ is given by
\begin{equation}
\bdOmega_k =
\begin{cases}
\left[ \begin{matrix}
\kappa_k \xi_{k,0} & \kappa_k \sqrt{\xi_{k,0}} \bdp_{k,0}^H \\
\kappa_k \sqrt{\xi_{k,0}} \bdp_{k,0} & \kappa_k \bdp_{k,0} \bdp_{k,0}^H + \diag(\bdlambda_k)
\end{matrix}\right] \frac{\beta_k}{\kappa_k + 1} \comma & \text{ if } \udbdu_{k,0} \ne \bdzro \\
\left( \kappa_k \bdp_{k,0} \bdp_{k,0}^H + \diag(\bdlambda_k) \right) \frac{\beta_k}{\kappa_k + 1} \comma & \text{ if } \udbdu_{k,0} = \bdzro.
\end{cases}
\end{equation}
Note that $\bdOmega_k$ is positive definite. Thus, we have
\begin{equation}
\rank\left( \bdR_k^{\ut} \right) \stackeq{a} \rank(\bdB_k \bdOmega_k \bdB_k^H) \stackeq{b} \rank(\bdB_k) \stackeq{c} \Sone_k \comma \label{rank_relation_Sk_Edd}
\end{equation}
where (a) comes from \eqref{Rkut_Bk_Omegak_Bk}, (b) is from \cite[Observation 7.1.8(b)]{Horn2013MatrixAnalysis}, and (c) follows from the fact that $\bdB_k$ has orthogonal columns.
We can rewrite the UL ergodic sum rate as
\begin{align}
I_{\SUM}
=\bbE \left \{ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^K \bdc_k^H \bdT_k \bdc_k \cdot \bdg_k \bdg_k^H \right) \right\} \triangleq R_{\SUM} \comma \label{Rsum_Qk_Sk}
\end{align}
where $\bdT_k = \bdB_k^H \bdQ_k \bdB_k \in \Complex{\Sone_k}{\Sone_k}$. Here, we use $R_{\SUM}$ to denote the UL ergodic sum rate, since $R_{\SUM}$ has become a function of $\{\bdT_k\}_{k=1}^K$.
\begin{mytheorem} \label{Proposition_2}
The optimal solution $\{\bdQ_k^{\star}\}_{k=1}^K$ to the problem $\clS_{\rQ}$ can be obtained via
\begin{equation} \label{Relation_Qk_Tk_optimal}
\bdQ_k^{\star} =
\bdB_k \bdT_k^{\star} \bdB_k^H\comma\ \forall k \in \clK.
\end{equation}
Here, $\{\bdT_k^{\star}\}_{k=1}^K$ is the optimal solution to the following problem
\begin{equation} \label{Problem_UL_sum_rate_Tk}
\clS_{\rT}:\ \max_{\bdT_k\succeq \bdzro\comma\ \trace \left( \bdT_k \right) \le P_k\comma\ \forall k \in \clK}\ \bbE \left\{ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^K \bdc_k^H \bdT_k \bdc_k \cdot \bdg_k \bdg_k^H \right) \right\}.
\end{equation}
\end{mytheorem}
\begin{IEEEproof}
Please refer to \Cref{appendix_Proposition_2_proof}.
\end{IEEEproof}
\Cref{Proposition_2} reveals that the design of each UT $k$'s $N \times N$ transmit covariance matrix $\bdQ_k$ can be converted into that of an $\Sone_k \times \Sone_k$ lower dimensional matrix $\bdT_k$, whose dimension is exactly equal to $\rank(\bdR_k^\ut)$ as shown by \Cref{rank_relation_Sk_Edd}. Interestingly, this is in accord with the results in \Cref{Proposition_1}.
After the optimal lower-dimensional matrices $\{\bdT_k^{\star}\}_{k=1}^K$ are obtained, the optimal transmit covariance matrices $\{\bdQ_k^{\star}\}_{k=1}^K$ can be derived immediately by using \eqref{Relation_Qk_Tk_optimal}. Henceforth, we only need to concentrate on the optimization of the lower-dimensional matrices $\{\bdT_k\}_{k=1}^K$.
Notice that the problem $\clS_{\rT}$ keeps the convex property, and the optimization variables therein have reduced dimensions.
The following theorem shows that the solution to the problem $\clS_{\rT}$ can be further simplified under some special conditions.
\begin{mytheorem} \label{Proposition_3}
If $\bdp_{k,0} = \bdzro$ holds for UT $k$, then $\bdT_k^{\star}$ is a diagonal matrix.
\end{mytheorem}
\begin{IEEEproof}
Please refer to \Cref{appendix_Prop_Tk_diagonal_proof}.
\end{IEEEproof}
In fact, $\bdp_{k,0} = \bdzro$ indicates that $\bdd_{k,0}$ is orthogonal to $\spann(\bdU_k)$. In this case, $\bdB_k$ reduces to $\bdB_k = \left[\bdd_{k,0}\ \bdU_k\right]$, whose column vectors actually become the eigenvectors of $\bdQ_k$.
As shown in \Cref{Proposition_3}, if $\bdp_{k,0} = \bdzro$ holds, the optimal transmit strategy of UT $k$ would be sending independent data streams along the directions determined by the columns in $\bdB_k = \left[\bdd_{k,0}\ \bdU_k\right]$.
Let $\bdT_k = \bdV_{\rT,k} \diag(\bdxi_{\rT,k}) \bdV_{\rT,k}^H$ denote the EVD of $\bdT_k$, where the unitary matrix $\bdV_{\rT,k} \in \Complex{\Sone_k}{\Sone_k}$ consists of eigenvectors and $\bdxi_{\rT,k} \in \Real{\Sone_k}{1}$ includes the corresponding non-negative eigenvalues. The transmit signal $\bdx_k$ should be given by
\begin{equation}
\bdx_k = \bdB_k \bdV_{\rT,k} \diag(\bdxi_{\rT,k}^{1/2}) \bds_{\rT,k}\comma
\end{equation}
where $\bds_{\rT,k} \sim \clCN(\bdzro,\bdI_{\Sone_k})$. Similarly, we can show that under some limit cases, the optimal matrices $\{\bdT_k\}_{k=1}^K$ to the problem $\clS_{\rT}$ have unit rank.
\subsubsection{Low SNR Case} \label{subsubsection_low_SNR_Tmat}
When $P_k \rightarrow 0$, $\forall k \in \clK$, $R_{\SUM}$ can be approximated by
\begin{align}
R_{\SUM}
\rightarrow \xinv{\sigma^2} \sum_{k=1}^K \trace \left( \bdOmega_k \bdT_k \right). \label{sum_rate_low_SNR_Tmat}
\end{align}
Thus, the problem $\clS_{\rT}$ is reduced into
\begin{equation}
\clS_{\rT}^{\LowSNR}:\ \max_{ \bdT_k \succeq \bdzro\comma\ \trace(\bdT_k) \le P_k\comma\ \forall k \in \clK }\ \sum_{k=1}^{K} \trace\left( \bdOmega_k \bdT_k \right). \label{Problem_UL_sum_rate_capacity_lowSNR_Tmat}
\end{equation}
The optimal matrices $\{\bdT_k\}_{k=1}^K$ to the problem $\clS_{\rT}^{\LowSNR}$ are given by
\begin{equation}
\bdT_k = P_k \cdot \bdv_{\rT,k} \bdv_{\rT,k}^H\comma\ \forall k \in \clK\comma \label{Tk_optimal_low_SNR}
\end{equation}
where $\bdv_{\rT,k} \in \Complex{\Sone_k}{1}$ is the unit-norm eigenvector of $\bdOmega_k$ associated with its maximum eigenvalue.
In fact, $\bdB_k \bdv_{\rT,k}$ is a unit-norm eigenvector of $\bdR_k^{\ut}$ corresponding to its maximum eigenvalue. Thus, \Cref{Tk_optimal_low_SNR} is in accord with the results in \Cref{Qk_optimal_low_SNR}.
\subsubsection{High Rician Factor Case} \label{subsubsection_high_Rician_factor_Tmat}
If $\kappa_k \rightarrow \infty$ holds for each UT $ k \in \clK$, the UL sum rate capacity is given by maximizing the following problem
\begin{equation}
\clS_{\rT}^{\HighRiceK}:\ \max_{ \bdT_k \succeq \bdzro\comma\ \trace(\bdT_k) \le P_k\comma\ \forall k \in \clK }\ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^{K} \beta_k \bdc_{k,0}^H \bdT_k \bdc_{k,0} \cdot \bdg_k \bdg_k^H \right)\comma \label{Problem_UL_sum_rate_capacity_infRician_Tmat}
\end{equation}
where $\bdc_{k,0} = [ \sqrt{\xi_{k,0}} \ \bdp_{k,0}^T ]^T$ if $\udbdu_{k,0} \ne \bdzro$, and otherwise, $\bdc_{k,0} = \bdp_{k,0}$.
The optimal transmit covariance matrices $\{\bdT_k\}_{k=1}^K$ to the problem $\clS_{\rT}^{\HighRiceK}$ is given by
\begin{equation}
\bdT_k = P_k \cdot \bdc_{k,0} \bdc_{k,0}^H,\ \forall k \in \clK. \label{Tk_optimal_high_Rician}
\end{equation}
Note that $\bdd_{k,0} = \bdB_k \bdc_{k,0}$. Thus, \eqref{Tk_optimal_high_Rician} coincides with the results in \Cref{Qk_optimal_high_Rician}.
Even though rank-one transmit covariance matrices achieve the ergodic sum rate capacity in the aforementioned special cases, the optimal transmit covariance matrices for the most general cases are still unknown so far. Next, we elaborate the general design approaches for the transmit covariance matrices, which is applicable to arbitrary link conditions in UL massive MIMO LEO SATCOM systems.
\subsection{General Transmit Design with Stochastic Programming} \label{subsection_General_Design}
As mentioned in \Cref{subsection_structure_transmit_matrix}, the design of transmit covariance matrices $\{\bdQ_k\}_{k=1}^K$ has been converted into that of lower-dimensional of matrices $\{\bdT_k\}_{k=1}^K$. In this subsection, we will show the lower-dimensional matrix design approach. By revealing the structure of the optimal lower-dimensional matrices, we develop a stochastic programming based iterative algorithm to compute the lower-dimensional matrices. Then, we use \eqref{Relation_Qk_Tk_optimal} to obtain the transmit covariance matrices $\{\bdQ_k\}_{k=1}^K$ from the lower-dimensional matrices $\{\bdT_k\}_{k=1}^K$.
The Lagrangian function of the problem $\clS_{\rT}$ can be written as
\begin{align}
\clL_{\rT} = R_{\SUM} - \sum_{k=1}^K \trace(\bdT_k \bdZ_k) - \sum_{k=1}^K \zeta_k \left( \trace(\bdT_k) - P_k \right)\comma
\end{align}
where $\bdZ_k \succeq \bdzro$ and $\zeta_k \ge 0$ are the Lagrange multipliers associated with the constraints $\bdT_k \succeq \bdzro$ and $\trace(\bdT_k) \le P_k$, respectively.
For the optimal solution $\{\bdT_k^{\star}\}_{k=1}^K$ to the problem $\clS_{\rT}$, there must exist some $\{\bdZ_k^{\star}, \zeta_k^{\star}\}_{k=1}^K$ satisfying the Karush-Kuhn-Tucker (KKT) conditions as follows
\begin{subequations} \label{KKT_Tk_Rsum}
\begin{align}
& \bdM_k^{\star} - \bdZ_k^{\star} - \zeta_k^{\star} \bdI = \bdzro \comma \label{KKT_grad_Tk_Rsum} \\
& \bdT_k^{\star} \bdZ_k^{\star} = \bdzro, \bdT_k^{\star} \succeq \bdzro, \bdZ_k^{\star} \succeq \bdzro\comma \label{KKT_Tk_Zk} \\
& \zeta_k^{\star} \left( \trace(\bdT_k^{\star}) - P_k \right) = 0, \trace(\bdT_k^{\star}) \le P_k, \zeta_k^{\star} \ge 0\comma \label{KKT_Tk_Pk}
\end{align}
\end{subequations}
$k\in\clK$, where $\bdM_k^{\star} = \bdM_k|_{\bdT_k = \bdT_k^{\star}, \forall k\in\clK}$, and $\bdM_k \triangleq \nabla_{\bdT_k} R_{\SUM}$ is given by
\begin{align}
\bdM_k
& = \bbE \left\{ \xinv{\sigma^2} \bdg_k^H \Xinv{ \bdI_M + \xinv{\sigma^2} \sum_{k=1}^K \bdc_k^H \bdT_k \bdc_k \cdot \bdg_k \bdg_k^H } \bdg_k \cdot \bdc_k \bdc_k^H \right\} \notag \\
& \stackeq{a} \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_k \bdc_k } \cdot \bdc_k \bdc_k^H \right\} \comma \label{Rsum_gradient_Tk}
\end{align}
where (a) follows from the matrix inversion lemma \cite{Horn2013MatrixAnalysis}, and $\bdJ_k = \bdI_M + \xinv{\sigma^2} \sum_{i\ne k} \bdc_i^H \bdT_i \bdc_i \cdot \bdg_i \bdg_i^H$. To further proceed, we express $\bdM_k$ as follows
\begin{align} \label{Mk_Sk_Tk}
\bdM_k &= (\bdI - \bdM_k \bdT_k) \Xinv{\bdI - \bdM_k \bdT_k} \bdM_k \notag \\
&= \left[ \bdI + \Xinv{\bdI - \bdM_k \bdT_k} - \bdI \right]^{-1} \Xinv{\bdI - \bdM_k \bdT_k} \bdM_k \notag \\
&= \left[ \bdI + \Xinv{\bdI - \bdM_k \bdT_k} \left( \bdI - \left( \bdI - \bdM_k \bdT_k \right) \right) \right]^{-1} \Xinv{\bdI - \bdM_k \bdT_k} \bdM_k \notag \\
&= \Xinv{ \bdI + \bdS_k \bdT_k } \bdS_k\comma
\end{align}
where $\bdS_k \in \Complex{\Sone_k}{\Sone_k}$ is defined as
\begin{equation}
\bdS_k = \Xinv{\bdI - \bdM_k \bdT_k} \bdM_k. \label{definition_Sk}
\end{equation}
Thus, the condition in \eqref{KKT_grad_Tk_Rsum} can be further expressed as
\begin{equation}
\Xinv{ \bdI + \bdS_k^{\star} \bdT_k^{\star} } \bdS_k^{\star} - \bdZ_k^{\star} - \zeta_k^{\star} \bdI = \bdzro\comma \label{KKT_grad_Tk_Rsum_Sk}
\end{equation}
where $\bdS_k^{\star} = \bdS_k|_{\bdT_k=\bdT_k^{\star},\forall k\in\clK}$.
Now, we can see that \Cref{KKT_grad_Tk_Rsum_Sk,KKT_Tk_Zk,KKT_Tk_Pk} for each UT $k$ can be regarded as the KKT conditions of the following problem
\begin{equation}
\max_{\bdT_k\succeq \bdzro, \trace(\bdT_k)\le P_k}\ \log \det \left( \bdI + \bdS_k^{\star} \bdT_k \right). \label{problem_log_det_Sk_Tk}
\end{equation}
The problem in \eqref{problem_log_det_Sk_Tk} paves the way for better characterizing the structure of the optimal solution $\{\bdT_k^{\star}\}_{k=1}^K$ to the problem $\clS_{\rT}$, which will be formally stated in the following.
\begin{mytheorem} \label{Proposition_4}
Let us denote the EVD of $\bdS_k^{\star}$ as $\bdS_k^{\star} = \bdV_{\rS^{\star},k} \diag(\bdxi_{\rS^{\star},k}) \bdV_{\rS^{\star},k}^H$. The optimal solution $\{\bdT_k^{\star}\}_{k=1}^K$ to the problem $\clS_{\rT}$ should have the following form
\begin{equation} \label{solution_problem_log_det_Sk_Tk}
\bdT_k^{\star} = \bdV_{\rS^{\star},k} \diag\left( \left[ \Xinv{\zeta_k^{\star}} \cdot \bdone - \bdxi_{\rS^{\star},k}^{-1} \right]^+ \right) \bdV_{\rS^{\star},k}^H\comma \ \forall k \in \clK\comma
\end{equation}
where $[\bdx]^+ = \max\{\bdx,0\}$, and $\zeta_k^{\star}$ is chosen to make the power constraint $\trace(\bdT_k^{\star}) = P_k$ satisfied.
\end{mytheorem}
\begin{IEEEproof}
It has been revealed that the log-determinant maximization problem in \eqref{problem_log_det_Sk_Tk} has the water-filling solution form as in \eqref{solution_problem_log_det_Sk_Tk} \cite{Telatar1999MIMO}. In addition, since the KKT conditions of the problem $\clS_{\rT}$ for each UT $k$ are identical to that of the problem in \eqref{problem_log_det_Sk_Tk}, the optimal solution $\{\bdT_k^{\star}\}_{k=1}^K$ to the problem $\clS_{\rT}$ should also be characterized by the water-filling form in \eqref{solution_problem_log_det_Sk_Tk}. This concludes the proof.
\end{IEEEproof}
\Cref{Proposition_4} describes the structure of the optimal lower-dimensional matrices $\{\bdT_k^{\star}\}_{k=1}^K$. In more detail, for each UT $k$, the eigenvectors of $\bdT_k^{\star}$ should be identical to those of $\bdS_k^{\star}$. Meanwhile, the eigenvalues of $\bdT_k^{\star}$ are obtained by performing water-filling on the virtual subchannels, where the subchannel gains are given by the eigenvalues of $\bdS_k^{\star}$. Inspired by this structure, we develop an iterative algorithm based on the stochastic programming to compute the lower-dimensional matrices in the following.
It is worth noting that $\bdT_k$ can be determined by $\bdS_k$, which in its turn also depends on $\bdT_k$. Thus, we conceive an iterative procedure to compute $\{\bdS_k\}_{k=1}^K$ and $\{\bdT_k\}_{k=1}^K$ in an alternating manner. First, we compute $\{\bdS_k\}_{k=1}^K$ for given $\{\bdT_k\}_{k=1}^K$. Then, we update $\{\bdT_k\}_{k=1}^K$ by using \Cref{Proposition_4}, and so on.
Once the optimal lower-dimensional $\{\bdT_k\}_{k=1}^K$ are obtained, the optimal $\{\bdQ_k\}_{k=1}^K$ can be derived immediately by using \Cref{Relation_Qk_Tk_optimal}. The stochastic programming based algorithm for solving the problem $\clS_{\rQ}$ is summarized in \Cref{algorithm_UL_sum_rate_Tk}, which can serve as the baseline for the ergodic sum rate capacity achieving UL transmit design in massive MIMO LEO SATCOM systems.
\begin{algorithm}[!t]
\SetAlgoNoLine
\caption{Algorithm for UL transmit design with stochastic programming.}
\label{algorithm_UL_sum_rate_Tk}
\textbf{Initialization:} $\bdT_k^{(0)} = (P_k / \Sone_k) \cdot \bdI$, $\forall k \in \clK$, and iteration index $n = 0$. \\
\KwOut{Transmit covariance matrices $\{\bdQ_k\}_{k=1}^K$.}
\While{$1$}{
Compute $\bdS_k^{(n)}$ with \eqref{definition_Sk}, $\forall k \in \clK$. \\
Update $\{\bdT_k^{(n+1)}\}_{k=1}^K$ according to \Cref{Proposition_4}. \\
\eIf{$n\ge \Niter - 1$ \rm or $\sabs{ R_{\SUM}^{(n+1)} - R_{\SUM}^{(n)} } < \epsilon$ }{
Set $\bdT_k: = \bdT_k^{(n+1)}$, $\forall k \in \clK$, \textbf{break}.}{
Set $n:=n+1$.}}
Compute $\bdQ_k = \bdB_k \bdT_k \bdB_k^H$, $\forall k \in \clK$.
\end{algorithm}
Due to the expectation operation in the ergodic sum rate, the Monte-Carlo method is necessary to compute $\{\bdT_k\}_{k=1}^K$ in \Cref{algorithm_UL_sum_rate_Tk}, which brings considerable computational burden when averaging the channel samples. Next, to avoid the sample average, we resort to the asymptotic programming approach to compute $\{\bdT_k\}_{k=1}^K$, which has low complexity and can achieve near-optimal performance.
\subsection{General Transmit Design with Asymptotic Programming} \label{subsubsection_asyprog_Tk}
In this subsection, we present another UL transmit design approach by approximating the ergodic sum rate with its asymptotic expression. In this case, the structure of the optimal lower-dimensional matrices analogous to that in \Cref{subsection_General_Design} can also be derived. Then, a low-complexity iterative algorithm based on the asymptotic programming is developed to compute the lower-dimensional matrices. Once the lower-dimensional matrices $\{\bdT_k\}_{k=1}^K$ are obtained, the transmit covariance matrices $\{\bdQ_k\}_{k=1}^K$ can be computed by \Cref{Relation_Qk_Tk_optimal}.
Let $\bdomega_k \triangleq \frac{\beta_k}{\kappa_k + 1} \bbE \left\{ \tdbdc_k \odot \tdbdc_k \right\} \in \Complex{\Sone_k}{1}$, where $\tdbdc_k = [0 \ \udtdbdc_k^T]^T$ if $\udbdu_{k,0} \ne \bdzro$, and otherwise, $\tdbdc_k = \udtdbdc_k$.
Define $\brbdH_{\re} \in \Complex{M}{S_{\re}}$ as follows
\begin{align}
\brbdH_{\re} = \left[ \sqrt{\frac{\kappa_1 \beta_1}{\kappa_1+1}} \bdg_1 \bdc_{1,0}^H \ \cdots \ \sqrt{\frac{\kappa_K \beta_K}{\kappa_K+1}} \bdg_K \bdc_{K,0}^H \right]\comma
\end{align}
where $S_{\re} \triangleq \sum_{k=1}^{K} \Sone_k$.
An asymptotic expression, a.k.a. the deterministic equivalent, of $R_{\SUM}$ can be written as \cite{Wen2011SumRateMIMORician,Lu2016FreeDeterministic}
\begin{equation}
\udR_{\SUM} = \log \det \left( \bdI + \bdXi \bdT \right) + \log \det \left( \bdI + \bdC_{\rR} \right) - \sum_{k=1}^{K} \gamma_k \bdomega_k^T \bdpsi_k\comma \label{DE_sum_rate}
\end{equation}
where $\bdT = \diag\left( \bdT_1,\dots,\bdT_K \right)$ and $\bdpsi_k = [\psi_{k,1} \ \cdots \ \psi_{k,\Sone_k}]^T \in \Real{\Sone_k}{1}$. In \eqref{DE_sum_rate}, $\{(\gamma_k, \bdpsi_k)\}_{k=1}^K$ is the unique solution of the following equations
\begin{subequations} \label{gamma_psi_equations_DE}
\begin{align}
\gamma_k & = \bdg_k^H \Xinv{\bdI+\bdPsi} \bdg_k\comma \\
\bdpsi_k & = \diag\left( \bdT_k \left\langle \Xinv{\bdI + \bdXi \bdT} \right\rangle_k \right)\comma
\end{align}
\end{subequations}
$k\in\clK$, where $\langle \cdot \rangle_k$ means the operation of taking the $k$th sub-block along the diagonal of the matrix argument. On the other hand, $\bdXi \in \Complex{S_{\re}}{S_{\re}}$ and $\bdPsi \in \Complex{M}{M}$ in \eqref{gamma_psi_equations_DE} are dependent on $\{(\gamma_k, \bdpsi_k)\}_{k=1}^K$ by
\begin{subequations} \label{DE_Matrix_Xi_Psi}
\begin{align}
\bdXi & = \bdC_{\rT} + \brbdH_{\re}^H \Xinv{\bdI+\bdC_{\rR}} \brbdH_{\re}\comma \label{DE_Matrix_Xi} \\
\bdPsi & = \bdC_{\rR} + \brbdH_{\re} \bdT \Xinv{\bdI+\bdC_{\rT} \bdT} \brbdH_{\re}^H\comma \label{DE_Matrix_Psi}
\end{align}
\end{subequations}
respectively, where $\bdC_{\rT} = \diag\left( \bdC_{\rT,1},\dots,\bdC_{\rT,K} \right)$ and $\bdC_{\rR} = \sum_{k=1}^{K} \bdC_{\rR,k}$, $\bdC_{\rT,k} \in \Complex{\Sone_k}{\Sone_k}$ and $\bdC_{\rR,k} \in \Complex{M}{M}$ are given by
\begin{subequations}
\begin{align}
\bdC_{\rT,k} & = \gamma_k \cdot \diag(\bdomega_k)\comma \\
\bdC_{\rR,k} & = \bdomega_k^T \bdpsi_k \cdot \bdg_k \bdg_k^H\comma
\end{align}
\end{subequations}
respectively. It has been claimed that the unique solution $\{(\gamma_k,\bdpsi_k)\}_{k=1}^K$ to the equations in \eqref{gamma_psi_equations_DE} can be obtained by performing a fixed-point iterative procedure until convergence \cite{Wen2011SumRateMIMORician,Lu2016FreeDeterministic}.
Then, we consider the asymptotic approximation of the problem $\clS_{\rT}$ as follows
\begin{equation} \label{Problem_UL_sum_rate_Tk_DE}
\udclS_{\rT}:\ \max_{\bdT_k\succeq \bdzro\comma\ \trace \left( \bdT_k \right) \le P_k\comma\ \forall k \in \clK}\ \udR_{\SUM}.
\end{equation}
The Lagrangian function of the problem $\udclS_{\rT}$ can be written as
\begin{align}
\clL_{\rT} = \udR_{\SUM} - \sum_{k=1}^K \trace(\bdT_k \bdTheta_k) - \sum_{k=1}^K \eta_k \left( \trace(\bdT_k) - P_k \right)\comma
\end{align}
where $\bdTheta_k \succeq \bdzro$ and $\eta_k \ge 0$ are the Lagrange multipliers corresponding to the constraints $\bdT_k \succeq \bdzro$ and $\trace(\bdT_k) \le P_k$, respectively.
For the optimal solution $\{\bdT_k^{\circ}\}_{k=1}^K$ to the problem $\udclS_{\rT}$, there must exist some Lagrange multipliers $\{\bdTheta_k^{\circ}, \eta_k^{\circ}\}_{k=1}^K$, such that the following KKT conditions are satisfied \cite{Wen2011SumRateMIMORician,Lu2016FreeDeterministic}
\begin{subequations} \label{KKT_Tk_Rsum_DE}
\begin{align}
& \left\langle \bdW^{\circ} \right\rangle_k - \bdTheta_k^{\circ} - \eta_k^{\circ} \bdI = \bdzro \comma \label{KKT_grad_Tk_RsumDE} \\
& \bdT_k^{\circ} \bdTheta_k^{\circ} = \bdzro, \bdT_k^{\circ} \succeq \bdzro, \bdTheta_k^{\circ} \succeq \bdzro\comma \label{KKT_Tk_Thetak_DE} \\
& \eta_k^{\circ} \left( \trace(\bdT_k^{\circ}) - P_k \right) = 0, \trace(\bdT_k^{\circ}) \le P_k, \eta_k^{\circ} \ge 0\comma \label{KKT_Tk_Pk_DE}
\end{align}
\end{subequations}
$k\in\clK$, where $\bdW^{\circ} = \bdW|_{\bdT_k = \bdT_k^{\circ},\forall k\in\clK}$, $\bdW = \Xinv{ \bdI + \bdXi \bdT } \bdXi$.
By following similar techniques as in \eqref{Mk_Sk_Tk}, we can rewrite $\langle \bdW \rangle_k$ as follows
\begin{equation}
\langle \bdW \rangle_k = \Xinv{ \bdI + \bdXi_k \bdT_k } \bdXi_k\comma
\end{equation}
where $\bdXi_k \in \Complex{\Sone_k}{\Sone_k}$ is given by
\begin{align} \label{definition_Xik}
\bdXi_k = \Xinv{ \bdI - \langle \bdW \rangle_k \bdT_k } \langle \bdW \rangle_k.
\end{align}
As a result, the condition in \eqref{KKT_grad_Tk_RsumDE} can be expressed as
\begin{align} \label{KKT_grad_Tk_Rsum_Xik_DE}
\Xinv{\bdI + \bdXi_k^{\circ} \bdT_k^{\circ}} \bdXi_k^{\circ} - \bdTheta_k^{\circ} - \eta_k^{\circ} \bdI = \bdzro\comma
\end{align}
where $\bdXi_k^{\circ} = \bdXi_k|_{\bdT_k = \bdT_k^{\circ},\forall k\in\clK}$.
Hence, \Cref{KKT_Tk_Thetak_DE,KKT_Tk_Pk_DE,KKT_grad_Tk_Rsum_Xik_DE} for each UT $k$ can be viewed as the KKT conditions of the following problem
\begin{equation}
\max_{\bdT_k\succeq \bdzro, \trace(\bdT_k)\le P_k}\ \log \det \left( \bdI + \bdXi_k^{\circ} \bdT_k \right). \label{problem_log_det_Xik_Tk_DE}
\end{equation}
Similarly, the problem in \Cref{problem_log_det_Xik_Tk_DE} provides a more convenient way to describe the structure of the optimal solution $\{\bdT_k^{\circ}\}_{k=1}^K$ to the problem $\udclS_{\rT}$, which is formally stated in the following.
\begin{mytheorem} \label{Proposition_5}
Let us denote the EVD of $\bdXi_k^{\circ}$ as $\bdXi_k^{\circ} = \bdV_{\Xi^{\circ},k} \diag(\bdxi_{\Xi^{\circ},k}) \bdV_{\Xi^{\circ},k}^H$. The optimal solution $\{\bdT_k^{\circ}\}_{k=1}^K$ to the problem $\udclS_{\rT}$ can be written as
\begin{equation}
\bdT_k^{\circ} = \bdV_{\Xi^{\circ},k} \diag \left( \left[ \Xinv{\eta_k^{\circ}} \cdot \bdone - \bdxi_{\Xi^{\circ},k}^{-1} \right]^+ \right) \bdV_{\Xi^{\circ},k}^H\comma \ \forall k \in \clK\comma
\end{equation}
where $\eta_k^{\circ}$ is chosen such that the power constraint $\trace( \bdT_k^{\circ} ) = P_k$ is satisfied.
\end{mytheorem}
\begin{IEEEproof}
The proof is similar to that of \Cref{Proposition_4}. Thus, it is omitted here.
\end{IEEEproof}
Similar to the results in \Cref{Proposition_4}, \Cref{Proposition_5} characterizes the eigenvectors and eigenvalues of the optimal lower-dimensional matrices $\{\bdT_k^\circ\}_{k=1}^K$. Hence, the iterative procedure to compute $\{\bdT_k\}_{k=1}^K$ based on the asymptotic programming is given as follows. In each iteration, for given $\{\bdT_k\}_{k=1}^K$, we can derive $\{(\gamma_k,\bdpsi_k)\}_{k=1}^K$ by solving the fixed-point equations in \eqref{gamma_psi_equations_DE}, and then compute $\{\bdXi_k\}_{k=1}^K$ by \eqref{definition_Xik}. After that, we update $\{\bdT_k\}_{k=1}^K$ based on \Cref{Proposition_5}, and then the next iteration follows.
After $\{\bdT_k\}_{k=1}^K$ is derived, $\{\bdQ_k\}_{k=1}^K$ can be obtained with the help of \eqref{Relation_Qk_Tk_optimal}. The iterative procedure to compute $\{\bdQ_k\}_{k=1}^K$ via the asymptotic programming is presented in \Cref{algorithm_UL_sum_rate_Tk_DE}.
\begin{algorithm}[!t]
\SetAlgoNoLine
\caption{Algorithm for UL transmit design with asymptotic programming.}
\label{algorithm_UL_sum_rate_Tk_DE}
\textbf{Initialization:} $\bdT_k^{(0)} = (P_k / \Sone_k) \cdot \bdI$, $\forall k \in \clK$, and iteration index $n = 0$. \\
\KwOut{ Transmit covariance matrices $\{\bdQ_k\}_{k=1}^K$.}
\While{$1$}{
Compute $\bdXi_k^{(n)}$ with \eqref{definition_Xik}, $\forall k \in \clK$. \\
Update $\{\bdT_k^{(n+1)}\}_{k=1}^K$ according to \Cref{Proposition_5}. \\
\eIf{$n\ge \Niter - 1$ \rm or $\sabs{ \udR_{\SUM}^{(n+1)} - \udR_{\SUM}^{(n)} } < \epsilon$}{
Set $\bdT_k: = \bdT_k^{(n+1)}$, $\forall k \in \clK$, \textbf{break}.}{
Set $n:=n+1$.}}
Compute $\bdQ_k = \bdB_k \bdT_k \bdB_k^H$, $\forall k \in \clK$.
\end{algorithm}
The computational complexity of \Cref{algorithm_UL_sum_rate_Tk_DE} in terms of the number of multiplication operations is given by $\clO(N_{\rf}(S_{\re}^3 + K^3 + K^2 M))$, where $N_{\rf}$ is the number of fixed-point iterations to solve the equations in \eqref{gamma_psi_equations_DE}.
\section{Simulation Results} \label{section_simulation}
\begin{table}[!t]
\centering
\footnotesize
\renewcommand\arraystretch{1.2}
\captionof{table}{Simulation Parameters}
\label{table_simulation}
\begin{tabular}{Lc}
\toprule
Parameters & Values \\
\midrule
Earth radius $R_e$ & $6378$ km \\
Orbit altitude $H$ & $1000$ km \\
Central frequency $f_c$ & $2$ GHz \\
Bandwidth $B$ & $20$ MHz \\
Noise temperature $T_\rn$ & $273$ K \\
Number of antennas $\Mx(\My)$, $\Nx(\Ny)$ & $18/16/14/12$, $6$ \\
Antenna spacing $d_{\rx}(d_{\ry})$, $d_{\rx'}(d_{\ry'})$ & $\lambda$, $\frac{\lambda}{2}$ \\
Antenna gain $G_{\sat}$, $G_{\ut}$ & $7$ dBi, $0$ dBi \\
Maximum nadir angle $\vtheta_{\max}$ & $\xdeg{30}$ \\
Number of UTs $K$ & $100/80/60/40$ \\
Transmit power per UT & $10$ dBm -- $40$ dBm \\
\bottomrule
\end{tabular}
\end{table}
In this section, we provide the simulation results to verify the performance of the proposed UL transmit designs in massive MIMO LEO SATCOM.
The simulation parameters are summarized in \Cref{table_simulation}.
We denote the maximum nadir angle of the UTs as $\vtheta_{\max}$. The space angle pair $\bdalpha_k = (\alpha_k^{\rx},\alpha_k^{\ry})$ is generated according to the uniform distribution in the circle $\{(x,y)|x^2 + y^2 \le \sin^2 \vtheta_{\max}\}$, on account of the relation
$\cos \vtheta_k = \sin \theta_k^{\ry} \sin \theta_k^{\rx} = \sqrt{1 - (\alpha_k^{\ry})^2 - (\alpha_k^{\rx})^2} \ge \cos \vtheta_{\max}$.
The elevation angle of UT $k$ in \Cref{fig_uplink_UPA} is given by $v_k = \cos^{-1} \left( \frac{R_s}{R_e} \sin \vtheta_k \right)$, where $R_e$ is the earth radius, $R_s = R_e + H$ is the orbit radius \cite{Lutz2000SatSysPerson}.
The distance between the satellite and UT $k$ in \Cref{fig_uplink_UPA} is given by $D_k = \sqrt{R_e^2 \sin^2 v_k + H^2 + 2 H R_e} - R_e \sin v_k$ \cite{3GPP_NonTerrestrial}.
The per-antenna gains of the UPAs at the satellite and the UT sides are denoted as $G_{\sat}$ and $G_{\ut}$, respectively.
The pathloss and shadow fading are computed according to the model parameters in \cite[Section 6]{3GPP_NonTerrestrial}, and the ionospheric loss is set as $2$ dB approximately \cite[Section 6]{3GPP_NonTerrestrial}.
For simplicity, we assume that each UT's UPA is placed horizontally, which implies that the paired AoDs $\bdvphi_{k,0}=(\vphi_{k,0}^{\rx'},\vphi_{k,0}^{\ry'})$ of the LoS path satisfies $\sin\vphi_{k,0}^{\ry'} \sin \vphi_{k,0}^{\rx'} = \sin v_k$.
In order to generate the covariance matrices $\bdSigma_k$'s, the eigenvectors $\{\bdu_{k,i}\}_{i=1}^{S_k}$ are constructed by $\bdu_{k,i} = \bda_{\Nx}(\phi_k^{\rx'} + 2i/\Nx) \otimes \bda_{\Ny}(\phi_k^{\ry'} + 2i/\Ny)$ with $\phi_k^{\rx'} = \sin\vphi_{k,0}^{\ry'} \cos \vphi_{k,0}^{\rx'}$ and $\phi_k^{\ry'} = \cos \vphi_{k,0}^{\ry'}$,
while the eigenvalues $\{\lambda_{k,i}\}_{i=1}^{S_k}$ are first randomly chosen according to the uniform distribution $\mathrm{U}(0,1)$ and then rescaled so that $\sum_{i=1}^{S_k} \lambda_{k,i} = 1$.
The noise variance is given by $\sigma_k^2 = k_\rB T_\rn B$, where $k_\rB = 1.38 \times 10^{-23} \text{ J} \cdot \text{K}^{-1}$ is the Boltzmann constant, $T_{\rn}$ is the noise temperature and $B$ is the system bandwidth.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\textwidth]{Fig_Convergence_Rx_12x12_Tx_6x6_Rank_1_NumUser_100_StaProg_AsyProg_HighRiceK_10dB.eps}
\caption{Convergence of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} at different transmit power.}
\label{fig_convergence}
\end{figure}
In \Cref{fig_convergence}, the convergence performance of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} is depicted, where $M = \Mx \times \My = 12 \times 12$, $K= 100$ and $S_k = 1$, $\forall k \in \clK$. The transmit power $P_k$ of each UT $k$ takes different values including $20$ dBm, $30$ dBm, and $40$ dBm.
The Rician factor $\kappa_k$ of each UT $k$ is set as $\kappa_k = 10$ dB, $\forall k \in \clK$. To solve the fixed-point equations in \eqref{gamma_psi_equations_DE}, the number of iterations $N_{\rf}$ is set as $N_{\rf} = 10$ hereafter.
As we can see, both \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} can converge to the optimal points within a very small number of iterations.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\textwidth]{Fig_Rx_12x12_Tx_6x6_Rank_1_NumUser_100_StaProg_AsyProg_diffRice.eps}
\caption{Performance of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} for different Rician factors.}
\label{fig_performance_diffRice}
\end{figure}
\Cref{fig_performance_diffRice} shows the performance of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} for different values of Rician factors, where $M = \Mx \times \My = 12 \times 12$, $K= 100$ and $S_k = 1$, $\forall k \in \clK$. It is worth noting that the proposed FFR UL transmit strategies can dramatically improve the UL sum rate performance compared to the conventional four-color frequency reuse (FR4) schemes. The performance loss between \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} is negligible for both high and low Rician factor cases.
In addition, we notice that lower Rician factors can decrease the UL sum rate in massive MIMO LEO SATCOM, which means that the LoS components in LEO satellite channels are more beneficial for UL transmission.
\begin{figure}[!t]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{Fig_Rx_12x12_Tx_6x6_Rank_1_NumUser_100_StaProg_AsyProg_diffMt_HighRiceK_10dB.eps}
\caption{High Rician factor case with $\kappa_k = 10$ dB.}
\label{fig_performance_diffMt_HighRiceK}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{Fig_Rx_12x12_Tx_6x6_Rank_1_NumUser_100_StaProg_AsyProg_diffMt_LowRiceK_0dB.eps}
\caption{Low Rician factor case with $\kappa_k = 0$ dB.}
\label{fig_performance_diffMt_LowRiceK}
\end{subfigure}
\caption{Performance of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} for different values of $M$.}
\label{fig_performance_diffMt}
\end{figure}
\Cref{fig_performance_diffMt} illustrates the performance of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} for different values of $M$, where $K = 100$ and $S_k = 1$, $\forall k \in \clK$.
We can observe that increasing the number of antennas at the satellite can significantly improve the UL sum rate performance in FFR LEO SATCOM.
In \Cref{fig_performance_diffK}, the performance of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} for different values of $K$ is depicted, with $M = \Mx\times\My = 12\times 12$ and $S_k = 1$, $\forall k \in \clK$.
It can be seen that by serving a larger number of UTs, the proposed FFR UL transmit designs can remarkably enhance the UL sum rate performance for massive MIMO LEO SATCOM.
\begin{figure}[!t]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{Fig_Rx_12x12_Tx_6x6_Rank_1_NumUser_100_StaProg_AsyProg_diffK_HighRiceK_10dB.eps}
\caption{High Rician factor case with $\kappa_k = 10$ dB.}
\label{fig_performance_diffK_HighRiceK}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{Fig_Rx_12x12_Tx_6x6_Rank_1_NumUser_100_StaProg_AsyProg_diffK_LowRiceK_0dB.eps}
\caption{Low Rician factor case with $\kappa_k = 0$ dB.}
\label{fig_performance_diffK_LowRiceK}
\end{subfigure}
\caption{Performance of \Cref{algorithm_UL_sum_rate_Tk,algorithm_UL_sum_rate_Tk_DE} for different values of $K$.}
\label{fig_performance_diffK}
\end{figure}
\section{Conclusion} \label{section_conclusion}
In this paper, we have investigated the UL transmit design that achieves the ergodic sum rate capacity in massive MIMO LEO SATCOM systems. The UL massive MIMO LEO satellite channel model is presented, where the satellite and the UTs both employ the UPAs. It is shown that the rank of each UT's transmit covariance matrix that achieves the ergodic sum rate capacity does not exceed that of its channel correlation matrix at the UT side. This reveals the maximum number of independent data streams that can be supported through each UT's UL satellite channel. The design of transmit covariance matrices is reduced into that of lower-dimensional matrices. By deriving the structure of the optimal lower-dimensional matrices, a stochastic programming based algorithm is developed for the lower-dimensional matrix design. To reduce the computational complexity, we resort to the asymptotic programming and develop a low-complexity algorithm to compute the transmit covariance matrices. The effectiveness of the proposed UL transmit designs is verified in the simulation results.
\appendices
\section{Proof of \Cref{Proposition_1}} \label[secinapp]{appendix_Proposition_1_Proof}
The Lagrangian to the problem $\clS_{\rQ}$ is given by
\begin{align}
\clL_{\rQ} ={} & \bbE \left\{ \log \det \left( \bdI_M + \xinv{\sigma^2} \sum_{k=1}^{K} \bdd_k^H \bdQ_k \bdd_k \cdot \bdg_k \bdg_k^H \right) \right\} + \sum_{k=1}^K \trace (\bdPhi_k \bdQ_k) \notag \\
& \quad - \sum_{k=1}^K \mu_k \left( \trace(\bdQ_k) - P_k \right)\comma
\end{align}
where $\mu_k \ge 0$ and $\bdPhi_k \succeq \bdzro$ are Lagrange multipliers associated with the constraints $\trace(\bdQ_k) \le P_k$ and $\bdQ_k \succeq \bdzro$, respectively. At the optimum to the problem $\clS_{\rQ}$, the gradient of $\clL_{\rQ}$ will vanish, i.e.,
\begin{equation}
\nabla_{\bdQ_k} \clL_{\rQ} = \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \right\} - \mu_k \bdI_N + \bdPhi_k = \bdzro\comma \label{sum_rate_capacity_Lagrangian_grad}
\end{equation}
where $\bdA_k = \bdI_M + \xinv{\sigma^2} \sum_{i \ne k} \bdd_i^H \bdQ_i \bdd_i \cdot \bdg_i \bdg_i^H$.
By multiplying $\bdQ_k$ with the equation in \eqref{sum_rate_capacity_Lagrangian_grad} from the right side, it yields
\begin{equation}
\bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \right\} \bdQ_k - \mu_k \bdQ_k = \bdzro\comma \label{sum_rate_capacity_Lagrangian_grad_zero}
\end{equation}
where the complementary slackness condition $\bdPhi_k \bdQ_k = \bdzro$ is applied. With some rearrangement to \eqref{sum_rate_capacity_Lagrangian_grad_zero}, we can obtain the following relation
\begin{equation}
\bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \right\} \bdQ_k = \mu_k \bdQ_k. \label{sum_rate_capacity_relation}
\end{equation}
We may now notice that the Lagrange multiplier $\mu_k$ must be strictly positive. Otherwise, the zero matrices $\{\bdQ_k|_{\bdQ_k = \bdzro}\}_{k=1}^K$ will be an optimal solution to the problem $\clS_{\rQ}$, which is apparently impossible. From the equality in \eqref{sum_rate_capacity_relation}, we can derive
\begin{align}
\rank(\bdQ_k) & = \rank \left( \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \right\} \bdQ_k \right) \notag \\
& \stackleq{a} \rank \left( \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \right\} \right)\comma \label{sum_rate_capacity_rank_Edd}
\end{align}
where (a) follows from the inequality \cite[0.4.5(c)]{Horn2013MatrixAnalysis} $\rank(\bdA \bdB) \le \min\left\{ \rank(\bdA),\rank(\bdB) \right\}$.
In addition, for every random vector $\bdd_k$, we have
\begin{equation}
\frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \stackpreceq{a} \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k \bdd_k^H\comma \label{sum_rate_capacity_dd}
\end{equation}
where (a) follows from $\bdd_k^H \bdQ_k \bdd_k \ge 0$. By taking the mathematical expectation for all random vector $\bdd_k$ in \eqref{sum_rate_capacity_dd}, we have
\begin{equation}
\bbE \left\{\frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \right\} \preceq \xinv{\sigma^2} \bbE \{ \bdg_k^H \bdA_k^{-1} \bdg_k \} \cdot \bdR_k^{\ut}\comma \label{sum_rate_capacity_expectation_dd}
\end{equation}
where $\bbE \{ \bdg_k^H \bdA_k^{-1} \bdg_k \} > 0$ and $\bdR_k^{\ut} = \bbE \{\bdd_k \bdd_k^H\}$.
By applying the inequality $\rank(\bdA) \ge \rank(\bdB)$ for $\bdA \succeq \bdB \succeq \bdzro$ \cite[Theorem 7.8]{Zhang2011MatrixTheory}, the relation in \eqref{sum_rate_capacity_expectation_dd} implies that
\begin{equation}
\rank \left( \bbE \left\{\frac{ \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdA_k^{-1} \bdg_k \cdot \bdd_k^H \bdQ_k \bdd_k } \cdot \bdd_k \bdd_k^H \right\} \right) \le \rank \left( \bdR_k^{\ut} \right). \label{sum_rate_capacity_rank_final}
\end{equation}
After combining the results in \eqref{sum_rate_capacity_rank_Edd} and \eqref{sum_rate_capacity_rank_final}, we can conclude the proof.
\section{Proof of \Cref{Proposition_2}} \label[secinapp]{appendix_Proposition_2_proof}
For any given positive semidefinite matrix $\bdT_k$, the linear matrix equation $\bdB_k^H \bdQ_k \bdB_k = \bdT_k$ always has a positive semidefinite solution $\bdQ_k$ if and only if $\bdT_k \bdB_k^{-} \bdB_k = \bdT_k$ \cite{GRO2000AXA=B},
where $\bdB_k^{-} \in \Complex{\Sone_k}{M}$ is an arbitrary generalized inverse (g-inverse) of $\bdB_k$. The g-inverse $\bdA^{-}$ of $\bdA$ is defined as the matrix that satisfies the equation $\bdA \bdA^- \bdA = \bdA$. In terms of \cite[pp.~47]{Adi2003GeneralInverse}, one of the g-inverses of $\bdB_k$ is given by $\bdB_k^- = \Xginv{\bdB_k^H \bdB_k} \bdB_k^H$.
Thus, we have
\begin{align} \label{Bk-Bk_Uk-Uk}
\bdT_k \bdB_k^- \bdB_k = \bdT_k \bdB_k^H \bdB_k \stackeq{a} \bdT_k\comma
\end{align}
where (a) follows from $\bdB_k^H \bdB_k = \bdI$.
From the relation in \eqref{Bk-Bk_Uk-Uk}, we can conclude that for any given positive semidefinite $\bdT_k$, there always exists a positive semidefinite solution $\bdQ_k$ to the equation $\bdB_k^H \bdQ_k \bdB_k = \bdT_k$.
According to \cite[Lemma 2.1]{Khatri1976HermitianMatrixEquations}, for given positive semidefinite matrix $\bdT_k$, the positive semidefinite solution $\bdQ_k$ to the equation $\bdB_k^H \bdQ_k \bdB_k = \bdT_k$ can be represented by
\begin{equation}
\begin{aligned}
\bdQ_{N,k} = \bdB_k \bdT_k \bdB_k^H + \left( \bdI - \bdB_k \bdB_k^H \right) \bdN_k \left( \bdI - \bdB_k \bdB_k^H \right)\comma \label{Qk_g_inverse_Tk}
\end{aligned}
\end{equation}
where $\bdN_k \in \Complex{N}{N}$ is an arbitrary positive semidefinite matrix. Then, the trace of matrix $\bdQ_{N,k}$ in \eqref{Qk_g_inverse_Tk} is given by
\begin{equation}
\trace \left( \bdQ_{N,k} \right) = \trace \left( \bdT_k \right) + \Delta_k\comma
\end{equation}
where $\Delta_k = \trace \left( \left( \bdI - \bdB_k \bdB_k^H \right) \bdN_k \left( \bdI - \bdB_k \bdB_k^H \right) \right)$.
If $\Delta_k > 0$, we can always construct another matrix $\bdQ_{T,k} = \bdB_k \bdT_k \bdB_k^H$. Then, the matrix $\frac{\trace(\bdQ_{N,k})}{\trace(\bdT_{k})} \bdQ_{T,k}$ can attain a larger UL sum rate $I_{\SUM}$ in \eqref{UL_Capacity} under the same power constraint with matrix $\bdQ_{N,k}$.
On the other hand, if $\Delta_k = 0$, which indicates that $\left( \bdI - \bdB_k \bdB_k^H \right) \bdN_k \left( \bdI - \bdB_k \bdB_k^H \right) = \bdzro$, then $\bdQ_{N,k}$ is exactly equal to $\bdB_k \bdT_k \bdB_k^H$.
Consequently, for a given matrix $\bdT_k$, we only need to consider the solutions $\bdQ_k$ to the linear matrix equation $\bdB_k^H \bdQ_k \bdB_k = \bdT_k$ with the following form
\begin{equation}
\bdQ_k = \bdB_k \bdT_k \bdB_k^H. \label{Relation_Qk_Tk}
\end{equation}
Hence, the power constraint $\trace(\bdQ_k) \le P_k $ can be rewritten as
\begin{align}
\trace(\bdQ_k) \stackeq{a} \trace(\bdB_k \bdT_k \bdB_k^H) \stackeq{b} \trace(\bdB_k^H \bdB_k \bdT_k) \stackeq{c} \trace(\bdT_k) \le P_k\comma
\end{align}
where (a) follows from \eqref{Relation_Qk_Tk}, (b) follows from $\trace(\bdA \bdB) = \trace(\bdB \bdA)$, and (c) follows from $\bdB_k^H \bdB_k = \bdI$.
With the above statement, the problem $\clS_{\rQ}$ can be equivalently transformed into the problem $\clS_{\rT}$. Furthermore, once the optimal solution $\{\bdT_k^{\star}\}_{k=1}^K$ to the problem $\clS_{\rT}$ is obtained, the optimal matrices $\{\bdQ_k^{\star}\}_{k=1}^K$ to the problem $\clS_{\rQ}$ can be directly derived with the aid of the relation in \eqref{Relation_Qk_Tk_optimal}.
This concludes the proof.
\section{Proof of \Cref{Proposition_3}} \label[secinapp]{appendix_Prop_Tk_diagonal_proof}
In the proof, we follow the similar techniques as in \cite{Veeravalli2005CorrelatedMIMO,Visotsky2001SpaceTime,Tulino2006CapacityAchieving}.
Let $\bdT_{D,k}^{\circ}$ be the maximizer of $R_{\SUM}$ for all $\bdT_{D,k} \in \clD_k $, where the set $\clD_k$ is defined as
\begin{align}
\clD_k = \left\{ \bdT_{k}: \bdT_k \succeq \bdzro\comma\ \trace(\bdT_k) \le P_k \text{ and } \bdT_{k} \text{ is diagonal} \right\}.
\end{align}
For any $\bdT_{D,k} \in \clD_k$, the first order optimality condition must hold at $\bdT_{D,k}^{\circ}$ as follows
\begin{equation}
\trace \left\{ \left. \nabla_{\bdT_{D,k}}R_{\SUM} \right\rvert_{\bdT_{D,k} = \bdT_{D,k}^{\circ}} \left( \bdT_{D,k} - \bdT_{D,k}^{\circ} \right) \right\} \le 0. \label{Proof_OptimalityCondition_Pk_trace}
\end{equation}
Consequently, the inequality in \eqref{Proof_OptimalityCondition_Pk_trace} can be further written as
\begin{equation} \label{Proof_OptimalityCondition_Pk_scalar}
\bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H ( \bdT_{D,k} - \bdT_{D,k}^{\circ} ) \bdc_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{D,k}^{\circ} \bdc_k } \right\} \le 0.
\end{equation}
Then, we show that $\bdT_{D,k}^{\circ}$ is also the maximizer of $R_{\SUM}$ for all $\bdT_k \in \clN_k$, where the set $\clN_k$ is defined as
\begin{align}
\clN_k = \left\{ \bdT_k: \bdT_k \succeq \bdzro\comma\ \trace(\bdT_k) \le P_k \right\}.
\end{align}
To explain this, we separate $\bdT_k$ as $\bdT_k = \bdT_{G,k} + \bdT_{F,k}$, where $\bdT_{G,k}$ and $\bdT_{F,k}$ contain the diagonal and off-diagonal entries in $\bdT_k$, respectively.
Then, we can obtain that
\begin{align}
& \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H ( \bdT_k - \bdT_{D,k}^{\circ} ) \bdc_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{D,k}^{\circ} \bdc_k } \right\} \notag \\
={}& \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H ( \bdT_{G,k} - \bdT_{D,k}^{\circ} ) \bdc_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{D,k}^{\circ} \bdc_k } \right\}
+ \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{F,k} \bdc_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{D,k}^{\circ} \bdc_k } \right\}. \label{Proof_OptimalityCondition_Pk_sum}
\end{align}
Notice that $\bdT_{G,k} \in \clD_k$ holds, because of $\trace(\bdT_k) = \trace(\bdT_{G,k}) \le P_k$. Combining \eqref{Proof_OptimalityCondition_Pk_scalar} and $\bdT_{G,k} \in \clD_k$, we can conclude that the first term in \eqref{Proof_OptimalityCondition_Pk_sum} is non-positive.
On the other hand, the second term in \eqref{Proof_OptimalityCondition_Pk_sum} can be rewritten as
\begin{align}
\bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{F,k} \bdc_k }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{D,k}^{\circ} \bdc_k } \right\}
=\sum_{p=1}^{\Sone_k} \sum_{\substack{q=1, q\ne p}}^{\Sone_k} \bbE \left\{ \frac{ \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot [\bdT_{F,k}]_{p,q} c_{k,p}^* c_{k,q} }{ 1 + \xinv{\sigma^2} \bdg_k^H \bdJ_k^{-1} \bdg_k \cdot \bdc_k^H \bdT_{D,k}^{\circ} \bdc_k } \right\} \stackeq{a} 0 \comma
\end{align}
where $c_{k,p}$ is the $p$th element in $\bdc_k$, (a) follows from the fact that the odd functions of CSCG random variables are equal to zeros.
Thus, the following inequality holds for all $\bdT_k \in \clN_k$
\begin{equation}
\trace \left\{ \left. \nabla_{\bdT_k} R_{\SUM} \right\rvert_{\bdT_k = \bdT_{D,k}^{\circ}} \left( \bdT_k - \bdT_{D,k}^{\circ} \right) \right\} \le 0. \label{Proof_OptimalityCondition_Pk_diagonal}
\end{equation}
The condition in \eqref{Proof_OptimalityCondition_Pk_diagonal} means that $\bdT_k^{\star}$ will be a diagonal matrix under the condition $\bdp_{k,0} = \bdzro$.
This concludes the proof.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{sec1}
Nematic mesophases are probably the simplest states of matter observed in liquid-crystalline systems that exhibit long-range orientational order in the absence of translational symmetry breaking \cite{deGennes_Book, FigueiredoNeto2005, Singh2000, Palffy2007}. Indeed, uniaxial nematic structures are characterized macroscopically by the existence of orientation-dependent physical properties (for example, optical or magnetic anisotropies), which lead to the definition of the director of a nematic phase.
Notwithstanding, the breaking of isotropy in the plane perpendicular to the uniaxial director may lead to the elusive biaxial state, whose possibility was theoretically pointed out by Freiser \cite{freiser1970ordered} about 50 years ago.
Experimentally, the existence of the biaxial phase was initially
confirmed for lyotropic systems \cite{yu1980observation}. More recently, there have been claims
of the identification of the phase in thermotropic systems composed of bent-core molecules, although this remains debatable (see Ref.\,\cite{Jakli2018} and references therein). In any case, these claims catalyzed various experimental, computational, and theoretical investigations \cite{Luckhurst2015biaxial,Akpinar2019} of candidate biaxial systems.
Most theoretical and computational studies looking for biaxial phases focus on the orientational order, leaving aside effects associated with a varying density of nematogens. Approaches based on the phenomenological Landau--de Gennes expansion \cite{GRAMSBERGEN1986} are able to partially remedy this situation, by exploiting variations in the expansion coefficients, although these are difficult to connect with microscopic parameters. Our aim in this paper is to investigate the equilibrium phase diagrams of a statistical model in which nematogens with non-cylindrical symmetry can move from site to site in a lattice whose occupation can be controlled. In our model, pairs of nematogens interact via an isotropic potential which can be repulsive or attractive, as well as via an anisotropic potential which favors a biaxial arrangement, leading, at sufficiently high occupation and sufficiently low temperature, to a biaxial phase.
Lattice models of nematic order have a long history in the literature. For uniaxial systems, the pioneering work of Lebwohl and Lasher \cite{Lebwohl1972} inspired a number of other investigations, including a lattice-gas extension by Bates \cite{bates2001computer,bates2002phase}. For biaxial systems, the Luckhurst--Romano model \cite{Luckhurst1980}, based on the truncation of an anisotropic potential to second-rank terms, has been likewise influential. As a rule, Monte Carlo calculations for nearest-neighbor versions of these models on fully-occupied cubic lattices lead to the same qualitative predictions as those obtained from mean-field versions \cite{maier1958einfache,Longa2005}, despite sometimes substantial quantitative discrepancies \cite{Lebwohl1972,Biscarini1995}.
A quite general bilinear anisotropic interaction potential $V_{12}$ between two nematogens labeled as $1$ and $2$ was proposed by Straley \cite{Straley1974}. In the two-tensor formulation of Sonnet, Virga and Durand \cite{sonnet2003dielectric}, it takes the form
\begin{equation}
V_{12} = -\frac{9}{4}A\left\{ \mathbf{q}_1\mathbf{:}\mathbf{q}_2 + \zeta\left( \mathbf{q}_1\mathbf{:}\mathbf{b}_2 + \mathbf{b}_1\mathbf{:}\mathbf{q}_2 \right) + \lambda \mathbf{b}_1\mathbf{:}\mathbf{b}_2\right\}.
\label{eq:V12}
\end{equation}
In Eq.\,\eqref{eq:V12}, $A>0$ sets the energy scale, while the second-rank tensors $\mathbf{q}$ and $\mathbf{b}$ are defined in terms of mutually orthogonal unit vectors
$\hat{n}_1$, $\hat{n}_2$ and $\hat{n}_3$ pointing along the first, second and third principal axes of each nematogen as
\begin{equation}
\mathbf{q}=\hat{n}_{1}\otimes\hat{n}_{1}-\frac{1}{3}\mathbf{I}\quad\text{and}\quad\mathbf{b}=\hat{n}_{2}\otimes\hat{n}_{2}-\hat{n}_{3}\otimes\hat{n}_{3},
\end{equation}
$\mathbf{I}$ being the $3\times3$ identity matrix.
The operation $\mathbf{q}_1\mathbf{:}\mathbf{q}_2$
is the Frobenius inner product \cite{Horn_Book}, given by $\Tr\left(\mathbf{q}_1\mathbf{q}_2\right)$, where $\Tr\textbf{M}$ is the trace of matrix $\textbf{M}$.
The adimensional parameters $\zeta$ and $\lambda$ gauge the importance of biaxial couplings. If $\zeta=\lambda=0$, Eq.\,\eqref{eq:V12} is reduced to the Maier--Saupe interaction energy \cite{maier1958einfache}, defined solely by the relative orientation of the first principal axes of both nematogens. This is appropriate when dealing with nematogens whose form may be properly approximated as uniaxial. Otherwise, if the nematogens are intrinsically biaxial, a proper description of the interaction energy requires setting either $\zeta$ or $\lambda$ to nonzero values, so that the relative orientations of other principal axes are also relevant.
Here
we work with the condition $\lambda=\zeta^2$, corresponding to the London approximation for dispersion forces \cite{sonnet2003dielectric}, which allows us to write $V_{12}$ in the form
\begin{equation}
V_{12} = -\frac{9}{4}A \left(\mathbf{q}_1 + \frac{\Delta}{3}\mathbf{b}_1 \right)\mathbf{:} \left(\mathbf{q}_2 + \frac{\Delta}{3}\mathbf{b}_2 \right).
\label{eq:V12Delta}
\end{equation}
By resorting to a simplified view of a biaxial nematogen as a rectangular platelet, the biaxiality parameter $\Delta = 3\zeta$ can be interpreted in terms of the sides of the platelet, so that $\Delta=0$ would correspond to a ``rod-like'' object, $\Delta=3$ to a ``disk-like'' object and $\Delta=1$ to a maximally biaxial object \cite{nascimento2015maier}.
In the same spirit as the lattice-gas version of the Lebwohl--Lasher model investigated by Bates \cite{bates2001computer}, we allow each site of regular lattice to be empty or occupied by a single nematogen, adding an isotropic interaction to the potential in Eq.\,\eqref{eq:V12Delta} to obtain the contribution of two neighboring sites $i$ and $j$ to the total interaction energy of the system,
\begin{equation}
V_{ij} = \gamma_i\gamma_j\left\{ U -\frac{9}{4}A \left(\mathbf{q}_i + \frac{\Delta}{3}\mathbf{b}_i \right)\mathbf{:} \left(\mathbf{q}_j + \frac{\Delta}{3}\mathbf{b}_j \right) \right\}.
\label{eq:V12Full}
\end{equation}
The occupation variable $\gamma_i$ is equal to $0$ if site $i$ is empty and to $1$ if the site is occupied. In this work we allow the isotropic interaction parameter $U$ to be either negative, representing attractive interactions, or positive, representing repulsion. This last case could lead to long-range sublattice ordering in cubic lattices, an unphysical feature for a fluid phase. At the mean-field level, however, describing such kind of arrangement would require the explicit introduction of sublattices. Instead, we proceed with the simplest mean-field strategy, which would be appropriate for describing a frustrated lattice or, for that matter, a fluid phase.
In order to perform detailed calculations, besides using Eq.\,\eqref{eq:V12Full} to describe the pair interactions, we also employ the Zwanzig approximation \cite{Zwanzig1963}, which restricts the possible orientations of a nematogen to the coordinate axes. This approximation has been applied in different contexts \cite{de1986reentrant, do2010statistical, do2011phase, liarte2012enhancement, nascimento2015maier, Sauerwein2016, nascimento2016lattice, petri2018field, rodrigues2020magnetic, dosSantos2021}, always leading to qualitative results which fully agree with continuous versions of the corresponding models when a comparison is possible. In particular, when dealing with intrinsically biaxial nematogens, these models are capable of reproducing the qualitative characteristics of nematic phase diagrams, such as sequences of biaxial-uniaxial-isotropic phase transitions with increasing temperature, and a well-defined Landau multicritical point, which signals a direct transition between the isotropic and the biaxial phases \cite{nascimento2016lattice,dosSantos2021}.
Therefore, in this work we investigate the phase diagrams of what may be characterized as a lattice-gas (LG) extension of the Maier--Saupe--Zwanzig model (MSZ), which from now on we will call the LGMSZ model. The LG extension introduces dilution as an extra ingredient in our model, allowing the study of phenomena such as vapor-liquid, vapor-nematic and nematic-nematic (low-high concentration) coexistence. The study of such coexistence is not possible if we treat a model based on a fully-occupied lattice.
This paper is organized as follows. Sec.\,\ref{sec2} presents the model description and sketches its mean-field solution. In Sec.\,\ref{sec3} we present a detailed analysis of the dilution effects, in the absence of isotropic interactions. Sec.\,\ref{sec4} is dedicated to the study of the effects of isotropic interactions for molecular systems with fixed degrees of biaxiality. In Sec.\,\ref{sec5} we present an analysis of the effects of the biaxiality degree in the multicritical points present in the phase diagrams. Conclusions are drawn in Sec.\,\ref{sec6}. A few technical details are relegated to Appendices \ref{appendix1} and \ref{appendix2}.
\section{The LGMSZ model}\label{sec2}
We consider a lattice system with $N$ sites and $N_m$ non-spherical objects such that $N \ge N_m$. Each lattice site can be either empty or occupied by an asymmetric object, the state of site $i$ being described by an occupation variable $\gamma_i$ taking the values $0$ (empty site) or $1$ (occupied site). Then, based on Eq.\,\eqref{eq:V12Full}, we define the LGMSZ model by means of effective the Hamiltonian
\begin{equation}\label{eq:MSZUmodel}
\mathcal{H}=\sum_{(i,j)}V_{ij}=-A\sum_{(i,j)}\gamma_{i}\gamma_{j}\mathbf{\Omega}_{i}\mathbf{:}\mathbf{\Omega}_{j}+U\sum_{(i,j)}\gamma_{i}\gamma_{j},
\end{equation}
where $A$ and $U$ are coupling parameters, with $A>0$, the sum is performed over pairs $(i,j)$ of neighboring sites $i$ and $j$ in the lattice, and the quantity $\mathbf{\Omega}_{i}$ is a second-rank tensor associated with the nematogen at site $i$. Specifically, $\mathbf{\Omega}_{i}$ is represented by a $3\times 3$ square matrix with real entries. For nematogens, $\mathbf{\Omega}_{i}$ is a symmetric traceless matrix, its eigenvalues $\omega_i$ are real and their sum is zero \cite{deGennes_Book,Luckhurst2015biaxial}. Then, we can assume that $\omega_{1}=(-1+\Delta)/2$, $\omega_{2}=(-1-\Delta)/2$, and $\omega_{3}=1$, where the parameter $\Delta$ gauges the asymmetry or biaxiality degree of the object \cite{nascimento2015maier}: $\Delta=0$ for rod-like shapes, $\Delta=3$ for plate-like shapes, and $\Delta \neq 0,3$ for brick-like shapes. Biaxial objects with $\Delta=1$ present a maximal degree of asymmetry. Instead of working with continuum orientational states, we follow the Zwanzig prescription \cite{Zwanzig1963} in assuming that the principal axes of a nematogen are restricted to align in the directions of the Cartesian axes, which leads to an effective spin-like model with six states described by diagonal matrices $\bm{\Omega}_i$ \cite{nascimento2015maier}. Notice that, in the limit $\Delta=0$, Eq.\,\eqref{eq:MSZUmodel} reduces to a discretized version of the Lebwohl--Lasher lattice-gas
model introduced by Bates in Ref.\,\cite{bates2001computer} (with a rescaling of energy, as our parameter $A$ would be equivalent to $2\epsilon/3$, $\epsilon$ being the energy scale of the anisotropic interaction in Ref.\,\cite{bates2001computer}).
The first term in Eq.\,\eqref{eq:MSZUmodel} represents a dilute version of the MSZ model, and the orientation-dependent interaction may give rise to distinct nematic phases. The second term is the isotropic contribution to the pair potential. For the particular case of $U<0$, representing attractive isotropic interactions, one can find phase transitions between isotropic fluid states, in analogy with previous studies \cite{bates2001computer,bates2002phase}. In the present work we assume that the parameter $U$ can also be positive, representing repulsive interactions. In this latter case, as we are interested in modeling fluid phases only, we refrain from trying to account for any kind of sublattice ordering whatsoever.
Determining the thermodynamic properties of the lattice system defined by Eq. \eqref{eq:MSZUmodel} is rather intricate, due to the complex interplay between the various interactions. Therefore, we think it is appropriate to study the model in a mean-field treatment, which is equivalent to considering the fully-connected Hamiltonian
\begin{equation}\label{MFmodel}
\mathcal{H}_\text{mf}=-\frac{A}{2N}\sum_{i,j=1}^{N}\gamma_{i}\gamma_{j}\mathbf{\Omega}_{i}\mathbf{:}\mathbf{\Omega}_{j}+ \frac{U}{2N}\sum_{i,j=1}^{N}\gamma_{j}\gamma_{j},
\end{equation}
where the sums over pairs of neighboring sites are replaced by sums over all pairs of sites, and the coupling parameters are replaced by new ones that are inversely proportional to the number of sites to ensure that energy is extensive.
This form of effective, long-range model has been proposed to investigate the phase behavior of statistical models with nematic-like phases \cite{do2010statistical,nascimento2015maier,nascimento2016lattice,SalinasNascimento2017,
petri2018field,rodrigues2020magnetic}.
Therefore, our main interest is to study the thermodynamics of phases transitions of the mean-field model in Eq. \eqref{MFmodel}.
The canonical ensemble is the usual route to investigate the macroscopic behavior of Hamiltonian systems in statistical mechanics. Nevertheless, because of its lattice-gas character, the configurations of microscopic variables of our model are subject to the restriction that the sum of $\gamma_i$ over all lattice sites should be equal to $N_m$,
\begin{comment}
\begin{equation}\label{eq:relaxsum}
\sum^{N}_{i=1}\gamma_{i}=N_{m},
\end{equation}
\end{comment}
which leads to complications in evaluating the canonical partition function. As a result, it is more convenient to consider the formalism of the grand canonical ensemble, where the number of nematogens may fluctuate due to the coupling to a particle reservoir \cite{do2010statistical,nascimento2015maier,rodrigues2020magnetic}. Then, we must determine the grand partition function
\begin{equation}\label{eq:gpflgmsz}
\begin{split}
\Xi=&\sum_{\{\gamma_{i}\}}\sum_{\{\mathbf{\Omega}_{i}\}}\exp\left(\beta \mathcal{H}_\text{mf} + \beta\mu\sum_{i}\gamma_{i}\right),
\end{split}
\end{equation}
where $\beta = 1/k_B T$, $k_B$ is the Boltzmann constant (which we take to be equal to $1$ in suitable units), $T$ is the temperature and $\mu$ is the chemical potential.
In this ensemble, the sum over configurations in Eq. \eqref{eq:gpflgmsz} is no longer restricted, and mean-field calculations are feasible, as indicated in Appendix \ref{appendix1}.
As a result, we obtain the Landau--de Gennes free-energy functional
\begin{equation}\label{eq:landaufef}
\begin{split}
\psi(S,\eta,\phi) &= \dfrac{A}{4}\left(3S^{2}+\eta^{2}\right)+\dfrac{U}{2}\phi^{2}-\mu\phi\\
&\quad +\dfrac{1}{\beta}\left[(1-\phi)\ln{\left(\dfrac{1-\phi}{6}\right)}+\phi\ln{(\phi)}\right] \\
& \quad -\dfrac{\phi}{\beta}\ln{\left[\Lambda(S,\eta)\right]},
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
\Lambda(S,\eta) &= 2\exp\left[ -\frac{3\beta A}{4}(S+\eta) \right] \cosh{\left[\dfrac{3\beta A}{4}\left(S-\dfrac{\eta}{3}\right)\Delta\right]} \\
&\quad + 2\exp\left[ -\frac{3\beta A}{4}(S-\eta)\right] \cosh{\left[\dfrac{3\beta A}{4}\left(S+\dfrac{\eta}{3}\right)\Delta\right]} \\
&\quad + 2\exp\left( \frac{3\beta A}{2}S\right) \cosh{\left(\dfrac{\beta A}{2}\eta \Delta\right)}.
\end{split}
\end{equation}
and the scalar parameters $S$ and $\eta$ are associated with the symmetric and traceless tensor order parameter \cite{GRAMSBERGEN1986}
\begin{equation} \label{Qtensor}
\mathbf{Q} = \left< \bm{\Omega} \right> = \frac{1}{2}
\begin{pmatrix}
-S - \eta & 0 & 0 \\
0 & -S + \eta & 0 \\
0 & 0 & 2S
\end{pmatrix},
\end{equation}
in which $\left< \cdot \right>$ denotes the ensemble average.
The equilibrium values of $S$, $\eta$ and $\phi$ are determined by locating the absolute minima of $\psi(S,\eta,\phi)$, leading to the mean-field (MF) equations
\begin{equation}
\frac{\partial\psi}{\partial S} = \frac{\partial\psi}{\partial\eta} = \frac{\partial\psi}{\partial\phi}=0,
\label{eq:mf}
\end{equation}
which take the self-consistent forms $S=F_{1}(S,\eta,\phi;\beta,\mu,\Delta)$, $\eta=F_{2}(S,\eta,\phi;\beta,\mu,\Delta)$, and $\phi=F_{3}(S,\eta,\phi;\beta,\mu,\Delta)$.
Depending on the solutions to these mean-field equations, the structure of the eigenvalues ${Q_x,Q_y,Q_z}$ of $\mathbf{Q}$ may be such that (i) $Q_x=Q_y=Q_z=0$, corresponding to the isotropic phase; (ii) $Q_x=Q_y\neq Q_z$ (or similar relations with permutations of the indices $x$, $y$ and $z$), corresponding to an uniaxial nematic phase; (iii) $Q_x$, $Q_y$ and $Q_z$ all distinct, corresponding to a biaxial nematic phase. If the eigenvalue with the largest absolute value is positive (negative), the nematic solution is calamitic (discotic). We use this terminology for both uniaxial and biaxial cases throughout the paper. In terms of the quantities $S$ and $\eta$, the isotropic solution is given by $S=\eta=0$, uniaxial solutions are such that $S\neq0$ with $\eta=0$ or $\eta = \pm 3S$, while the remaining cases represent biaxial solutions.
We emphasize that the values of $S$, $\eta$ and $\phi$ at the absolute minima of $\psi$ represent thermodynamic equilibrium values for fixed reciprocal temperature $\beta$, chemical potential $\mu$ and biaxiality degree $\Delta$. The free-energy $\mathcal{F}=\mathcal{F}(\beta,\mu,\Delta)$ of the system corresponds to the convex envelope of $\psi$ determined after inserting values of $S$, $\eta$ and $\phi$ associated with the minima of the free-energy functional.
\section{Behavior in the absence of the isotropic interaction}\label{sec3}
We start the investigation by assuming zero isotropic interaction, $U=0$, which simplifies the analysis of the problem by reducing the number of parameters. Some aspects of this case were discussed by Rodrigues \textit{et al.} \cite{rodrigues2020magnetic}, but taking into account only a specific range of model parameters. Here, we will present phase diagrams with many distinct topologies by exploring a wider range of values of thermodynamic fields. The results with zero isotropic interaction are helpful in understanding the situation involving both isotropic and anisotropic couplings, to be analyzed in the next section.
By considering intrinsically rod-like nematogens, for which $\Delta=0$, we find the phase diagram shown in Fig.\,\ref{fig:1}, which is qualitatively equivalent and quantitatively similar to the one obtained by Monte Carlo simulations for the
Lebwohl--Lasher lattice-gas model of Ref.\,\cite{bates2001computer}, in the absence of isotropic interactions.
At high concentration ($\phi\gtrsim 0.75$), as $T$ decreases, the observed phase sequence is isotropic (ISO), followed by a biphasic region of coexisting rod-rich uniaxial nematic (N$_\mathrm{U}^{+}$) and rod-poor isotropic phases, followed by a pure uniaxial nematic and finally a reentrant coexistence region. At lower rod concentration the coexistence region is stable at low temperatures.
The coexistence lines signaling the discontinuous transition from the isotropic phase to the uniaxial nematic phase is determined by Eq.\,(\ref{eq:mf}) evaluated at $\left(S,\eta,\phi\right)=\left( S_\mathrm{U}, 0, \phi_\mathrm{U} \right)$ and at $\left(S,\eta,\phi\right)=\left( 0, 0, \phi_\mathrm{I} \right)$, supplemented by $\psi(S_{\mathrm{U}},0,\phi_{\mathrm{U}})=\psi(0,0,\phi_{\mathrm{I}})$,
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{U}}, 0,\phi_{\mathrm{U}})}=\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(S_{\mathrm{U}}, 0,\phi_{\mathrm{U}})}=\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(0,0,\phi_{\mathrm{I}})}=0, \quad \psi(S_{\mathrm{U}},0,\phi_{\mathrm{U}})=\psi(0,0,\phi_{\mathrm{I}}),
\end{equation}
\end{comment}
where $\phi_{\mathrm{I}}$ and $\phi_{\mathrm{U}}$ are respectively the concentrations of the isotropic and uniaxial phases at the transition point, and $S_{\mathrm{U}}$ is the value of $S$ at that point. Notice that, since the nematogens are intrinsically uniaxial, we can assume $\eta = 0$ without loss of generality.
\begin{figure}[ht]
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_1a.pdf
\caption{$\Delta=0$}
\label{fig:1}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_1b.pdf
\caption{$\Delta=19/20$}
\label{fig:2}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_1c.pdf
\caption{$\Delta=1$}
\label{fig:3}
\end{subfigure}
\caption{Phase diagram in terms of temperature $T$ (in units of $A$) and concentration $\phi$ of nematogens, for different values of biaxiality degree and in the absence of isotropic interactions ($U=0$). ISO: isotropic phase. N$_{\mathrm{U}}^{+}$: calamitic uniaxial nematic phase. N$_\mathrm{B}$: biaxial nematic phase. Dashed lines are the boundaries of biphasic region (grey). Red solid line: critical end point ($\mathrm{CE}$). LTC is a Landau tricritical point.}
\end{figure}
For the case of objects that are non-cylindrical, $\Delta \neq 0$ and $\Delta \neq 3$, it is possible to observe stable biaxial phases (N$_\mathrm{B}$), as shown in Fig.\,\ref{fig:2} for biaxiality degree $\Delta=19/20$.
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figurePhi_2.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a biaxiality degree $\Delta=19/20$. Red solid-line: critical end points ($\mathrm{CE}$). Dashed line: first-order phase transitions. Solid line: second-order phase transitions. Coexistence region is shown in grey. }
\label{fig:2}
\end{figure}
\end{comment}
In this diagram, at high concentrations and high temperatures, there is a small biphasic region of coexisting uniaxial and isotropic phases. As temperature decreases, we have a second-order transition from the N$_{\mathrm{U}}^{+}$ phase to a pure N$_\mathrm{B}$ phase, and finally the biphasic region ISO-N$_\mathrm{B}$ appears. The conditions for determining the first-order transition between ISO and N$_\mathrm{B}$ are given by Eq.\,(\ref{eq:mf}) evaluated at $\left(S,\eta,\phi\right)=\left( S_\mathrm{B}, \eta_\mathrm{B}, \phi_\mathrm{B} \right)$ and at $\left(S,\eta,\phi\right)=\left( 0, 0, \phi_\mathrm{I} \right)$, as well as $\psi\left( S_\mathrm{B}, \eta_\mathrm{B}, \phi_\mathrm{B} \right) = \psi\left( 0, 0, \phi_\mathrm{I} \right) $,
where $S_{\mathrm{B}}$ and $\eta_{\mathrm{B}}$ are the values taken by $S$ and $\eta$ in the biaxial state at the transition point. On the other hand, the second-order transition between uniaxial and biaxial phases is located by Eq.\,(\ref{eq:mf}) and $\partial^2 \psi / \partial \eta^2 = 0$, all evaluated at $\left(S,\eta,\phi\right)=\left( S_o, 0 , \phi_o\right)$, where $S_o$ and $\phi_o$ are the values of $S$ and $\phi$ at the transition point.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{o},0,\phi_{o})}=\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(S_{o},0,\phi_{o})}=0, \quad \dfrac{\partial\psi}{\partial \eta}\Bigg|_{(S_{o},0,\phi_{o})}=\dfrac{\partial^{2}\psi}{\partial \eta^{2}}\Bigg|_{(S_{o},0,\phi_{o})}=0.
\end{equation}
\end{comment}
We also find that the N$^{+}_{\mathrm{U}}$-N$_{\mathrm{B}}$ transition line meets the ISO-N$_{\mathrm{B}}$ biphase region at a critical end point ($\mathrm{CE}$), in which a critical nematic state separating uniaxial and biaxial phases coexists with a noncritical isotropic state. Critical end points are possible multicritical points that can be found in thermodynamics systems with many components \cite{Uzunov_Book,Mario_Book}. In our case, we have critical end points related to nematic transitions in a lattice-gas model with orientation-dependent interactions. These kinds of multicritical points were also reported in a Maier--Saupe model that mimics binary mixtures of uniaxial and biaxial nematogens \cite{nascimento2015maier}.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=\dfrac{\partial\psi}{\partial \eta}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=0, \quad
\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(0,0,\phi_{\mathrm{I}})}=0,
\end{equation}
\begin{equation}
\psi(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})=\phi(0,0,\phi_{\mathrm{I}}),
\end{equation}
\end{comment}
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figurePhi_3.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for maximal biaxiality degree, $\Delta=1$. Dashed lines are the boundary of the ISO-N$_B$ coexistence region (grey). The solid line is a line of Landau points, which meets the biphasic region at Landau tricritical point (LTC).}
\label{fig:3}
\end{figure}
\end{comment}
For anisotropic objects with maximal biaxiality degree, $\Delta = 1$, stable uniaxial phases are absent and the phase diagrams present the general aspect shown in Fig.\,\ref{fig:3}.
In this case, for high temperatures and high concentration, the $\mathrm{ISO}$-$\mathrm{N_{B}}$ transition is continuous and determined by the conditions $\partial \psi / \partial \phi = \partial \psi / \partial \eta = \partial^2 \psi / \partial \eta^2 = 0$, evaluated at the transition point $\left(S,\eta,\phi\right)=\left(0,0,\phi_o \right)$.
\begin{comment}
\begin{equation}\label{eq:condd1}
\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(0,0,\phi_{o})}=\dfrac{\partial\psi}{\partial \eta}\Bigg|_{(0,0,\phi_{o})}=\dfrac{\partial^{2}\psi}{\partial \eta^{2}}\Bigg|_{(0,0,\phi_{o})}=0.
\end{equation}
\end{comment}
This line of continuous transitions is actually a line of multicritical Landau points. On the other hand, for low $T$ and intermediate concentrations, we observe an ISO-N$_{\mathrm{B}}$ coexistence region associated with a first-order transition at which $\partial \psi / \partial \eta = \partial \psi / \partial \phi = 0$ at $\left(S,\eta,\phi\right)=\left( 0, \eta_\mathrm{B},\phi_\mathrm{B}\right)$, $\partial \psi / \partial \phi =0$ at $\left(S,\eta,\phi\right)=\left( 0, 0,\phi_\mathrm{I}\right)$, and $\psi(0,\eta_{\mathrm{B}},\phi_{\mathrm{B}})=\psi(0,0,\phi_{\mathrm{I}})$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial \eta}\Bigg|_{(0,\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(0,\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=\dfrac{\partial\psi}{\partial \phi}\Bigg|_{(0,0,\phi_{\mathrm{I}})}=0, \quad \psi(0,\eta_{\mathrm{B}},\phi_{\mathrm{B}})=\psi(0,0,\phi_{\mathrm{I}}).
\end{equation}
\end{comment}
The discontinuous and continuous transitions meet at a multicritical point which we call Landau tricritical (LTC) point. Roughly speaking, according to the solutions of mean-field equations, the multicritical point LTC has properties common to both Landau points \cite{GRAMSBERGEN1986} and tricritical points \cite{Uzunov_Book,Mario_Book}. Notice that in the limit of a pure system (i.e. $\phi \to 1$) consisting of biaxial objects with $\Delta =1$, our findings are in agreement with earlier mean-field results, which shown a direct ISO-N$_\mathrm{B}$ transition through a single, isolated Landau point in the $\Delta$-$T$ phase diagram \cite{nascimento2015maier,nascimento2016lattice}.
It is possible to determine the conditions that characterize a Landau tricritical point by following the discussion presented by Rodrigues \textit{et al.} \cite{rodrigues2020magnetic}. Indeed, in our context, an LTC point
is the endpoint of a line of Landau points, and a Landau point happens when the stable solutions of MF equations for ISO and N$_{\mathrm{U}}^{\pm}$ become degenerate. Each point on a Landau line satisfies $\partial \psi / \partial \phi = d^2 \psi / d S^2 = d^3 \psi / d S^3 = 0$, evaluated at $\left(S,\eta,\phi\right)=\left( 0,0,\phi_L\right)$.
\begin{comment}
\begin{equation}\label{eq:firstL}
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(0,0,\phi_{\mathrm{L}})}=\dfrac{\,d^{2}\psi}{\,dS^{2}}\Bigg|_{(0,0,\phi_{\mathrm{L}})}=\dfrac{\,d^{3}\psi}{\,dS^{3}}\Bigg|_{(0,0,\phi_{\mathrm{L}})}=0,
\end{equation}
\end{comment}
Observe that these conditions involve partial derivatives as well as total derivatives (with respect to $S$) of the free-energy functional $\psi$. We must treat $\phi$ as an implicit function of $S$ while calculating the total derivatives. Thus, one can find $\Delta=1$, $(\beta A-1)e^{\beta\mu}-1=0$ and $ \beta A \phi_\mathrm{L} = 1 $, which are the same results obtained in Ref. \cite{rodrigues2020magnetic}. The solutions to these equations define a line of Landau points, which is represented by a solid line in Fig.\,\ref{fig:3}. In the limiting case of maximum concentration of biaxial objects, i.e. $\beta\mu\gg1$ or equivalently $\phi_{\mathrm{L}}\rightarrow1$, we recover the results obtained in previous treatments \cite{nascimento2015maier,nascimento2016lattice}, apart from differences in the definitions of parameters.
Nevertheless, we also have to check whether the solution leading to a Landau point corresponds to a minimum of the free-energy functional. This can be done by analyzing the behavior of the total fourth-order derivative of $\psi$ with respect to $\eta$ at $\left(S,\eta,\phi\right)=\left( 0,0,\phi_ \mathrm{L}\right)$, which gives $d^4 \psi / d \eta^4 = -3 A^4 \beta^3\phi_\mathrm{L} \left( 1 - 2\phi_\mathrm{L} \right)/8 $.
\begin{comment}
\begin{equation} \label{D4psieta}
\dfrac{\,d^{4}\phi}{\,d\eta^{4}}\Bigg|_{(0,0,\phi_{\mathrm{L}})}=-\dfrac{3}{8}A^{4}\beta^{3}\phi_{\mathrm{L}}(1-2\phi_{\mathrm{L}}),
\end{equation}
\end{comment}
This total derivative should be positive for stable states, but we notice that it may change its sign from positive, for $\phi_{\mathrm{L}}>1/2$, to negative, for $\phi_{\mathrm{L}}<1/2$, indicating that the Landau point is stable only if $\phi_{\mathrm{L}}>1/2$ (implying $\beta A<2$). Thus, precisely at $\phi_{\mathrm{L}}=1/2$, both $d^2 \psi / d \eta^2$ and $d^4 \psi / d \eta^4$ are zero, setting the conditions for locating a tricritical point that is also a Landau point. The coordinates of the LTC point are $\left( \beta A \right)_\mathrm{LTC}=2$, $\phi_{\mathrm{LTC}}=1/2$, and $\mu_{\mathrm{LTC}}=0$. The stability of the LTC point can be checked by looking at the sixth-order derivative of $\psi$ with respect to $\eta$, which gives $d^{6}\psi/\,d\eta^{6}=2A >0 $, therefore corresponding to a free-energy minimum.
\begin{comment}
Clearly, to assure stable solutions, the free energy landscape have to exhibit the features expected for a minimum in the vicinity of the LTC point.
\end{comment}
\begin{figure}[ht]
\includegraphics[width=0.6\textwidth]{figures/figure_2.pdf
\caption{Lines of multicritical points in the plane $\Delta$-$T$ for zero isotropic interaction. The line of Landau points (black) meets the lines of critical end points (red) at a Landau tricritical (LTC) point, which is only present for maximal biaxiality degree $\Delta=1$.}
\label{fig:4}
\end{figure}
We plotted all the lines of multicritical points obtained until now in the $\Delta$-$T$ plane shown in Fig.\,\ref{fig:4}. It is worth mentioning that, as we are assuming zero isotropic interaction, the space of thermodynamic fields is spanned by temperature $T$, chemical potential $\mu$ and biaxiality $\Delta$. Due to that, the lines presented in Fig.\,\ref{fig:4} are critical solutions of MF equations with varying chemical potential. Besides, although we have focused the discussion on calamitic nematic phases, for which $0<\Delta<1$, the results for discotic nematics (see e.g. Ref. \cite{Luders2021}) with $1<\Delta<3$ lead to phase diagrams with analogous topologies. Observe that for systems with maximal biaxiality degree, the LTC point occurs when the line of Landau points meets the two lines of critical end points. The Landau tricritical point is present only for maximal biaxiality $\Delta = 1$.
\begin{comment}
We find that For biaxiality degree $0<\Delta<1$ we existence of a line of critical-end points, which indicates, for each $\Delta$, the stability limit of a $\mathrm{N^{+}_{U}}$-$\mathrm{N_{B}}$ continuous phase transition, as shown in the phase diagrams of this section. A qualitatively similar situation occurs when $1<\Delta<3$, with the difference that we will have a discotic uniaxial phase instead of a calamitic one ($\mathrm{N^{-}_{U}}\leftrightarrow\mathrm{N^{+}_{U}}$) . This work is limited to presenting phase diagrams of calamitic systems because the phase diagrams topologies, in the $\phi$-$T$ planes, for discotic systems are somehow present in calamitic systems.
\end{comment}
\section{Behavior in the presence of the isotropic interaction}\label{sec4}
We now discuss phase diagrams in the presence of an isotropic interaction $U\neq 0$. In addition to uniaxial and biaxial structures, we may observe coexistence between isotropic fluid-like phases, which we call isotropic liquid ($\mathrm{IL}$) and isotropic vapor ($\mathrm{IV}$), as well as between nematic phases with different nematogen concentrations.
\subsection{Phase diagrams for uniaxial prolate nematogens (\texorpdfstring{$\Delta=0$}{}) }\label{subsecA}
For intrinsically uniaxial, rod-like objects, a sufficiently attractive ($U<0$) isotropic interaction leads to the appearance of a vapor-liquid (or a high-density--low-density transition) coexistence analogous to the van der Walls condensation, see Fig.\,\ref{fig:5}.
The vapor-liquid transition is determined by $\partial \psi / \partial \phi = 0$ at $\left(S,\eta,\phi\right)=(0,0,\phi_{\mathrm{IV}})$ and at $\left(S,\eta,\phi\right)=(0,0,\phi_{\mathrm{IL}})$, in addition to $\psi(0,0,\phi_{\mathrm{IV}})=\psi(0,0,\phi_{\mathrm{IL}})$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(0,0,\phi_{\mathrm{IV}})}=\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(0,0,\phi_{\mathrm{IL}})}=0, \quad \psi(0,0,\phi_{\mathrm{IV}})=\psi(0,0,\phi_{\mathrm{IL}}).
\end{equation}
\end{comment}
These first-order lines meet at a simple critical point ($\mathrm{C}$), located at
$\phi_{\mathrm{C}}=1/2$, $\beta_{\mathrm{C}}=-4/U$, $\mu_{\mathrm{C}}=U/2$ with $\psi_{\mathrm{C}}=U[2\ln{(12)}-1]/8$.
\begin{figure}[ht]
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_3a.pdf
\caption{$(A, U)=(1,-3)$}
\label{fig:5}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_3b.pdf
\caption{$(A,U)=(1,11/10)$}
\label{fig:6}
\end{subfigure}
\caption{Phase diagrams in terms of temperature $T$ and concentration $\phi$ of nematogens, for an intrinsically uniaxial system ($\Delta=0$). Red dashed line: triple point. Black dashed line: first-order transitions. $\mathrm{C}$ is a simple critical point.}
\end{figure}
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figurePhi_4.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a intrinsically uniaxial system with parameters $(A, U)=(1,-3)$. Red dashed: vapor-liquid-uniaxial triple point. Black dashed line: first-order transition. $\mathrm{C}$ is a simple critical point. }
\label{fig:5}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figurePhi_5.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a intrinsically uniaxial objects with parameters $(A, U)=(1,11/10)$. The red dashed line represents an isotropic-uniaxial-uniaxial triple point. The black dashed-line represents a first-order transition. $\mathrm{C}$ is a simple critical point. The inset is a zoom of the uniaxial-uniaxial coexistence region.}
\label{fig:6}
\end{figure}
\end{comment}
We also find a vapor-liquid-uniaxial triple point, which is determined by evaluating Eq.\,\eqref{eq:mf} at $\left(S,\eta,\phi\right)=\left(0,0,\phi_\mathrm{IV}\right)$, at $\left(S,\eta,\phi\right)=\left(0,0,\phi_\mathrm{IL}\right)$ and at $\left(S,\eta,\phi\right)=\left(S_\mathrm{U},0,\phi_\mathrm{U}\right)$, in addition to imposing $\psi\left(0,0,\phi_\mathrm{IV}\right)=\psi\left(0,0,\phi_\mathrm{IL}\right)=\psi\left(S_\mathrm{U},0,\phi_\mathrm{U}\right)$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(0,0,\phi_{\mathrm{IV}})}=\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(0,0,\phi_{\mathrm{IL}})}=0, \quad
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(0,0,\phi_{\mathrm{U}})}=\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{U}},0,\phi_{\mathrm{U}})}=0,
\end{equation}
\begin{equation}
\psi(0,0,\phi_{\mathrm{IV}})=\psi(0,0,\phi_{\mathrm{IL}})=\psi(S_{\mathrm{U}},0,\phi_{\mathrm{U}}).
\end{equation}
\end{comment}
For $T$ values lower than the triple-point temperature, the IV-IL discontinuous transition becomes metastable with respect to the IV-N$_{\mathrm{U}}^{+}$ first-order transition. As the strength $|U|$ of the attractive interaction increases, the region of stability of N$_{\mathrm{U}}^{+}$ decreases and tends to become limited to a very small region near $\phi=1$; see Fig.\,\ref{fig:5}.
This reduction in the area of the uniaxial phase was observed by Bates, using Monte Carlo simulations, in a lattice-gas extension of the Lebwohl--Lasher model \cite{bates2001computer} and later in the model proposed by Humphries, Luckhurst and Romano \cite{bates2002phase}.
For repulsive isotropic interactions ($U>0$), it is possible to notice the appearance of a very narrow coexistence region between uniaxial nematic phases, as shown in Fig.\,\ref{fig:6}. This biphasic coexistence region between uniaxial structures presents an ordinary critical point C, which can be found by imposing the conditions $\partial \psi / \partial S = \partial \psi / \partial \phi = d^2 \psi / d \phi^2 = d^3 \psi / d \phi^3 = 0$, evaluated at $\left(S,\eta,\phi \right) = \left( S_\mathrm{C}, 0, \phi_\mathrm{C} \right)$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{C}},0,\phi_{\mathrm{C}})}=\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(S_{\mathrm{C}},0,\phi_{\mathrm{C}})}=0, \quad
\dfrac{\,d^{2}\psi}{\,d\phi^{2}}\Bigg|_{(S_{\mathrm{C}},0,\phi_{\mathrm{C}})}=\dfrac{\,d^{3}\psi}{\,d\phi^{3}}\Bigg|_{(S_{\mathrm{C}},0,\phi_{\mathrm{C}})}=0.
\end{equation}
\end{comment}
We plot the lines of critical points and of triple points in the $U$-$T$ plane in Fig.\,\ref{fig:7}. These lines meet at higher-order critical points, which we call multicritical end points (MCE), in analogy with critical end points appearing when lines of first-order and second-order transitions meet. For $U<U^{(1)}_{\mathrm{MCE}}\approx-2.596$, we find phase diagrams with a simple critical point related to an IV-IL biphase region, in addition to a vapor-liquid-uniaxial triple point. This kind of phase phenomenon is associated with an attractive character of the isotropic interaction.
\begin{comment}
, and the increase in modulus of $U$ to more attractive systems implies an increase of the temperature of both points.
Nevertheless, when $U=U^{(1)}_{\mathrm{MCE}}$, the critical point and the vapor-liquid-uniaxial triple point meet at the temperature $T^{(1)}_{\mathrm{ MCE}}\approx0.649$.
\end{comment}
Nevertheless, for $U>U^{(1)}_{\mathrm{MCE}}$, it is no longer possible to distinguish between the IV and IL phases, and from a thermodynamic perspective there is a single isotropic phase. Then, we have phase diagrams which only show ISO-N$_\mathrm{U}$ coexistence regions.
In the case of repulsive isotropic interactions with $U<U^{(2)}_{\mathrm{MCE}}\approx1.035$, the phase diagrams also exhibit first-order transitions between isotropic and uniaxial phases. However, for $U^{(2)}_{\mathrm{MCE}}<U<U^{(3)}_{\mathrm{MCE}}=3/2$, as illustrated in Fig.\,\ref{fig:6}, it is possible to find phase diagrams exhibiting a coexistence region between uniaxial structures, with an associated critical point, as well as a triple point connecting one isotropic and two uniaxial states. As $U$ increases, we notice a decrease in the area of the low-temperature isotropic-uniaxial coexistence region, together with the decrease in the temperature of the critical and the triple points, until the ISO-N$_\mathrm{U}$ coexistence disappears completely as $U\to U^{(3) }_{\mathrm{MCE}}$. For this limiting value of $U$, both the temperatures of the critical point and of the triple point become zero.
\begin{figure}[ht]
\includegraphics[width=0.6\textwidth]{figures/figure_4.pdf
\caption{Lines of critical and of triple points in the $U$-$T$ plane, for the case of rod-like nematogens ($\Delta=0$). We notice that the lines of critical points meet the lines of triple points at higher-order multicritical end points ($\mathrm{MCE}$). The inset shows the case for repulsive isotropic interaction ($U>0$).}
\label{fig:7}
\end{figure}
\subsection{Phase diagrams for \texorpdfstring{$0<\Delta<1$}{}}
As previously mentioned, the discrete-state Maier--Saupe model presents phase diagrams with stable biaxial structures when the nematogens are intrinsically biaxial \cite{nascimento2015maier,nascimento2016lattice}. Then, we expect that the presence of dilution and isotropic interactions may lead to phase diagrams with more elaborate topologies. Indeed, for systems with attractive isotropic interactions, we obtain phase diagrams of the type shown in Fig.\,\ref{fig:66}. In this case, we have a critical point $\mathrm{C}$ associated with an IV-IL biphasic region, and an IV-IL-N$_\mathrm{U}^{+}$ triple point, analogous to those discussed in Sec.\,\ref{subsecA} for intrinsically uniaxial nematogens. We also find an IV-N$_\mathrm{B}$ discontinuous transition, determined by the conditions in Eq.\,\eqref{eq:mf}, evaluated at $\left(S,\eta,\phi\right)=\left( 0,0, \phi_\mathrm{IV}\right)$ and at $\left(S,\eta,\phi\right)=\left( S_\mathrm{B}, \eta_\mathrm{B}, \phi_\mathrm{B}\right)$, supplemented by $\psi(0,0,\phi_{\mathrm{IV}})=\psi(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})$. The coexistence between the biaxial phase and the isotropic vapor is verified at low temperatures, below the temperature of a critical end point CE ($T_{\mathrm{CE}}\approx0.56$ in the figure), whose location is set by Eq.\,\eqref{eq:mf}, evaluated at $\left(S,\eta,\phi\right)=\left( 0,0, \phi_\mathrm{IV}\right)$ and at $\left(S,\eta,\phi\right)=\left( S_\mathrm{CE}, 0, \phi_\mathrm{CE}\right)$, supplemented by $\psi(0,0,\phi_{\mathrm{IV}})=\psi(S_{\mathrm{CE}},0,\phi_{\mathrm{CE}})$ and $\partial^2\psi/\partial\eta^2=0$ at $\left(S,\eta,\phi\right)=\left( S_\mathrm{CE}, 0, \phi_\mathrm{CE}\right)$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(0,0,\phi_{\mathrm{IV}})}=\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=0, \quad
\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=\dfrac{\partial\psi}{\partial\eta}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=0,
\end{equation}
\begin{equation}
\psi(0,0,\phi_{\mathrm{IV}})=\psi(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}}).
\end{equation}
\end{comment}
The biaxial nematic phase is stable for high concentrations and small temperatures.
\begin{comment}
In fact, for $-3 <U\leq0 $, the region of stability of N$_\mathrm{B}$ increases as the isotropic interaction is attractive interaction decreases.
\end{comment}
\begin{figure}[ht]
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_5a.pdf
\subcaption{$\Delta=19/20$ and $(A, U)=(1,-3)$}
\label{fig:66}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_5b.pdf
\subcaption{$\Delta=4/5$ and $(A, U)=(1,1)$}
\label{fig:8}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_5c.pdf
\subcaption{$\Delta=9/20$ and $(A, U)=(1,13/10)$}
\label{fig:9}
\end{subfigure}
\caption{Phase diagrams in terms of temperature $T$ (in units of $A$) and concentration $\phi$ of nematogens, for different values of the biaxiality degree. The red dashed line represents a triple point. The red solid line is associated with a critical end point ($\mathrm{CE}$), while $\mathrm{C}$ is a critical point and $\mathrm{TC}$ is a tricritical point. The inset shows the isotropic-uniaxial coexistence region.}
\end{figure}
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figurePhi_6.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a intrinsically biaxial system with $\Delta=19/20$ and parameters $(A, U)=(1,-3)$. The red dashed line represents an vapor-liquid-uniaxial triple point. The red solid line represents an critical-end point ($\mathrm{CE}$). $\mathrm{C}$ is a critical point.}
\label{fig:66}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figurePhi_8.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a intrinsically biaxial system with $\Delta=4/5$ and parameters $(A, U)=(1,1)$. The red dashed line represents an isotropic-uniaxial-biaxial triple point. $\mathrm{TC}$ is a tricritical point.}
\label{fig:8}
\end{figure}
\end{comment}
Let us now consider repulsive isotropic interactions, $(A, U)=(1,1)$, with biaxiality degree $\Delta=4/5$. The phase diagram is shown in Fig.\,\ref{fig:8}, where it is possible to identify a triple point in which isotropic, uniaxial, and biaxial phases coexist, as well as a tricritical point (TC), which satisfies the conditions $\partial \psi / \partial \phi = \partial \psi / \partial S = d^2 \psi / d \eta^2 = d^4 \psi / d \eta^4 = 0$, evaluated at $\left(S,\eta,\phi\right)=\left( S_\mathrm{TC}, 0, \phi_\mathrm{TC}\right)$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(S_{\mathrm{TC}},0,\phi_{\mathrm{TC}})}= \dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{TC}},0,\phi_{\mathrm{TC}})}=0, \quad
\dfrac{\,d^{2}\psi}{\,d\eta^{2}}\Bigg|_{(S_{\mathrm{TC}},0,\phi_{\mathrm{TC}})}= \dfrac{\,d^{4}\psi}{\,d \eta^{4}}\Bigg|_{(S_{\mathrm{TC}},0,\phi_{\mathrm{TC}})}=0.
\end{equation}
\end{comment}
The total derivatives are determined by treating $S$ and $\phi$ as implicit functions of $\eta$. The boundaries of the coexistence region associated with uniaxial and biaxial phases are determined by Eq.\,\eqref{eq:mf} evaluated at $\left(S,\eta,\phi\right)=\left( S_\mathrm{U}, 0, \phi_\mathrm{U} \right)$ and at $\left(S,\eta,\phi\right)=\left( S_\mathrm{B}, \eta_\mathrm{B}, \phi_\mathrm{B} \right)$, as well as $\psi\left( S_\mathrm{U}, 0, \phi_\mathrm{U} \right) = \psi\left( S_\mathrm{B}, \eta_\mathrm{B} , \phi_\mathrm{B} \right)$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(S_{\mathrm{U}},0,\phi_{\mathrm{U}})}=\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=0,
\end{equation}
\begin{equation}
\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{U}},0,\phi_{\mathrm{U}})}=\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=\dfrac{\partial\psi}{\partial\eta}\Bigg|_{(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}})}=0,
\end{equation}
and
\begin{equation}
\psi(S_{\mathrm{U}},0,\phi_{\mathrm{U}})=\psi(S_{\mathrm{B}},\eta_{\mathrm{B}},\phi_{\mathrm{B}}).
\end{equation}
\end{comment}
We also show in Fig.\,\ref{fig:9} the phase diagram corresponding to the repulsive case with $\Delta=19/20$ and $(A,U)=(1,13/10)$. There are biphasic regions associated with ISO and N$_\mathrm{U}^{+}$, N$_\mathrm{U}^{+}$ and N$_\mathrm{B}$, and ISO and N$_\mathrm{B}$. Besides, there is a triple point marking the coexistence of ISO, N$_\mathrm{U}^{+}$ and N$_\mathrm{B}$. Finally, we observe the presence of a biaxial-biaxial coexistence region, whose boundaries are determined by Eq.\,\eqref{eq:mf} evaluated at $\left(S,\eta,\phi\right)=\left( S_1, \eta_1, \phi_1 \right)$ and at $\left(S,\eta,\phi\right)=\left( S_2, \eta_2, \phi_2 \right)$, supplemented by $\psi\left( S_1, \eta_1, \phi_1 \right)=\psi\left( S_2, \eta_2, \phi_2 \right)$.
\begin{comment}
\begin{equation}
\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(S_{\mathrm{LB}},\eta_{\mathrm{LB}},\phi_{\mathrm{LB}})}=\dfrac{\partial\psi}{\partial\phi}\Bigg|_{(S_{\mathrm{HB}},\eta_{\mathrm{HB}},\phi_{\mathrm{HB}})}=0,
\end{equation}
\begin{equation}
\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{LB}},\eta_{\mathrm{LB}},\phi_{\mathrm{LB}})}=\dfrac{\partial\psi}{\partial S}\Bigg|_{(S_{\mathrm{HB}},\eta_{\mathrm{HB}},\phi_{\mathrm{HB}})}=0,
\end{equation}
\begin{equation}
\dfrac{\partial\psi}{\partial\eta}\Bigg|_{(S_{\mathrm{LB}},\eta_{\mathrm{LB}},\phi_{\mathrm{LB}})}=\dfrac{\partial\psi}{\partial\eta}\Bigg|_{(S_{\mathrm{HB}},\eta_{\mathrm{HB}},\phi_{\mathrm{HB}})}=0,
\end{equation}
\begin{equation}
\psi(S_{\mathrm{LB}},\eta_{\mathrm{LB}},\phi_{\mathrm{LB}})=\psi(S_{\mathrm{HB}},\eta_{\mathrm{HB}},\phi_{\mathrm{HB}}),
\end{equation}
where $\mathrm{LB}$ and $\mathrm{HB}$ in subscripts represent the biaxial phase of low an high concentration of nematogens, respectively.
\end{comment}
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figurePhi_9.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a intrinsically biaxial system with $\Delta=19/20$ and parameters $(A, U)=(1,13/10)$. The red dashed line represents an isotropic-uniaxial-biaxial triple point. The red solid line represents a critical-end ($\mathrm{CE}$) point. $\mathrm{C}$ is a critical point related to a biaxial-biaxial discontinuous transition. The inset shows the isotropic-uniaxial coexistence region.}
\label{fig:9}
\end{figure}
\end{comment}
\subsection{Phase diagrams for \texorpdfstring{$\Delta=1$}{}}\label{subsecC}
Following our discussion in Sec.\,\ref{sec3}, we can obtain the conditions leading to Landau points for the maximal biaxiality degree and investigate the possible presence of Landau tricritical points. Indeed, we find analogous features when nonzero isotropic interactions are considered. Nevertheless, the parameter $U$ plays an important role in the criteria for determining the LTC point. After performing the calculation, we find that the coordinates of the Landau point satisfy $(\beta A-1)e^{\beta \mu}=e^{U/A}$ and $\beta A \phi = 1$.
For $\mu\rightarrow\infty$, i.e. in the limit of a fully occupied lattice, we recover the expected phase diagram with $\beta A=1$ at the Landau point, whereas for $U/A\rightarrow0$, we obtain the results discussed in Sec.\,\ref{sec3}. As we already know, the stability of a Landau point is related to the existence of an absolute minimum of the free-energy functional, and high-order derivatives should be considered because we are dealing with a multicritical point. The fourth-order derivative is
\begin{equation}
\left. \dfrac{d^4\psi}{d\eta^4} \right|_{(0,0,\phi)} = -\dfrac{3}{8}A^{3}\beta^2\left[\dfrac{U+A^{2}\beta-A(2+\beta U)}{A^{2}\beta+U(\beta A-1)}\right],
\end{equation}
This fourth-order derivative changes sign when $A\left( A-U\right)\beta=2A-U$,
which sets the condition for a possible LTC point. Notice that, as long as the isotropic interaction is attractive ($U<0$), there is always a candidate Landau tricritical point (since $\beta$ must be positive). However, as in the case $U=0$, the stability of that point for $U\neq0$ must be checked by looking at the sixth-order derivative of $\psi$ with respect to $\eta$,
\begin{equation}
\left.\dfrac{\,d^{6}\psi}{\,d\eta^{6}}\right|_{(0,0,\phi)}=\dfrac{(U-2A)^{4}(8A^{2}-30AU+15U^{2})}{64A(A-U)^{4}}.
\end{equation}
We then note that, since $A>0$, any LTC points are locally unstable if the isotropic interaction is repulsive ($U>0$) and such that $0.32\lesssim U/A \lesssim 1.68$.
\begin{figure}[ht]
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_6a.pdf
\subcaption{$(A,U)=(1,-5)$}
\label{fig:11}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_6b.pdf
\subcaption{$(A,U)=(1,-3)$}
\label{fig:12}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{figures/figure_6c.pdf
\subcaption{$(A,U)=(1,13/10)$}
\label{fig:13}
\end{subfigure}
\caption{Phase diagram in terms of temperature $T$ (in units of $A$) and concentration $\phi$ of nematogens, for maximal biaxiality degree $\Delta=1$. The red dashed line represents a triple point. The red solid line represents a Landau critical end point (LE). $\mathrm{C}$ is a critical point.}
\end{figure}
For $U<0$, the LTC point is always locally stable, although it may not correspond to an absolute minimum of the free-energy functional. This is the case for $U=-5$, as shown by the phase diagram in Fig.\,\ref{fig:11}. There is a wide coexistence region associated with isotropic phases of vapor and liquid, and an ordinary critical point ($\mathrm{C}$). For high concentrations, as $T$ decreases, there exists a continuous transition from the IL phase to the N$_\mathrm{B}$ phase. Additionally, for a fixed sufficiently low temperature, by varying the concentration we enter a coexistence region between the IV and the N$_\mathrm{B}$ phases. The line of continuous transition consists of Landau points, and that line meets the coexistence regions at a Landau critical end point (LE).
On the other hand, for isotropic interaction $U=-3$, we obtain the phase diagram exhibited in Fig.\,\ref{fig:12}. In this diagram, we now observe an LTC point, i.e. the LE point is not stable, and there also exists a triple point related to the IL, IV and N$_\mathrm{B}$ phases. When the isotropic interaction is sufficiently repulsive, we have a biaxial-biaxial coexistence region, as shown in Fig.\,\ref{fig:13}. This biphasic region presents a critical point C and a Landau critical end point LE. For phase diagrams with $U>2$, there are no coexistence regions and we only observe second-order transitions between the ISO and N$_\mathrm{B}$ phases; see Appendix \ref{appendix2}.
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figures/figurePhiDELTA1_0.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a system with maximal biaxiality degree ($\Delta=1$) and parameters $(A,U)=(1,-5)$. The red solid line represent a Landau-end point ($\mathrm{LE}$). The black dashed line and solid line represent first-order and second-order transitions respectively. $\mathrm{C}$ is a simple critical point.}
\label{fig:11}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figures/figurePhiDELTA1_1.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for maximal biaxiality degree $\Delta=1$ and $(A,U)=(1,-3)$. The red dashed line represent a vapor-liquid-biaxial triple point. The black dashed lines and solid line are respectively first-order and second-order transitions. $\mathrm{C}$ is a critical point.}
\label{fig:12}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figures/figurePhiDELTA1_2.png}
\caption{Phase diagram in terms of temperature $T$ and concentration $\phi$ of nematogens, for a system with maximal biaxiality degree ($\Delta=1$) and parameters $(A,U)=(1,13/10)$. The red solid line represent a Landau-end ($\mathrm{LE}$). The black dashed and solid lines represent first-order and second-order transitions respectively. $\mathrm{C}$ is a simple critical point.}
\label{fig:13}
\end{figure}
\end{comment}
\begin{figure}[ht]
\includegraphics[width=0.6\textwidth]{figures/figure_7.pdf
\caption{Lines of multicritical points in the $U$-$T$ plane for the case of maximal biaxiality parameter $\Delta=1$. The gray region marked as `Landau zone' consists of Landau points associated with different concentrations. ML$_2$ and ML$_3$ are higher-order Landau multicritical points. MCE is a higher-order multicritical end point.}
\label{fig:14}
\end{figure}
For the particular case $\Delta=1$ we can plot a graph in the $U$-$T$ plane showing the multicritical points found for maximal biaxiality; see Fig.\,\ref{fig:14}. The corresponding phase diagrams in the $\phi$-$T$ plane present a line of Landau points regardless of the character of the isotropic interaction. The stability limits of points belonging to these Landau lines are (i) at high temperatures, the point $(\phi,T)=(1,1)$ (fully-occupied lattice) and (ii) at low temperatures, a multicritical point whose nature depends on the value of $U$. In the $U$-$T$ plane, the stable Landau points occupy an extensive region which we call the Landau zone. The boundaries of this region are the line $(\phi,T)=(1,1)$ and the lines of Landau critical end points and Landau tricritical points, which meet at multicritical Landau points ML$_2$ and ML$_3$. We also find a higher-order multicritical end point MCE related to a line of triple points. These triple points are associated with coexisting vapor, liquid and biaxial phases. Observe that the MCE point occurs when the line of triple points meets a line of critical points.
\begin{comment}
When the system presents attractive isotropic interaction such as $U<U_{\mathrm{M_{2}}}\approx-4.034$ the phase diagrams are characterized by the presence of a high temperature vapor-liquid $\mathrm{C}$ point and a $\mathrm{LE}$ point. Then, by increasing the modulus of $U$, with $U<0$, we have a decreasing of the region of stability of N$_\mathrm{B}$. The $\mathrm{M_{2}}$ point implies the destabilization of a line of Landau end points in a line of Landau tricritical point and a line of vapor-liquid-biaxial triple points. For $U_{\mathrm{M_{2}}}<U<U_{\mathrm{MCE}}$, the increase of $U$ implies the approximation of the triple point and the simple critical point temperatures, that is, the decrease of vapor-liquid coexistence region. Just at $U=U_{\mathrm{MCE}}\approx-2.861$ we have the destabilization of the vapor-liquid coexistence when the temperature of both points is equal, $T_{\mathrm{MCE}}\approx 0.7153$. In phase diagrams with $U_{\mathrm{MCE}} <U<U_{\mathrm{M_{3}}}$ it is not possible to distinguish the vapor and liquid phases, so they are characterized by the existence of an isotropic-biaxial coexistence superiorly limited by an LTC point which is the beginning of a Landau-line. The increase of $U$ implies decreasing the temperatures of the $\mathrm{LE}$ point and the $\mathrm{C}$ point, and for $U>2$ there are no multicritical points.
\end{comment}
\section{Multicritical points in the biaxiality-temperature plane}
\label{sec5}We may summarize the different topologies of the $\phi$-$T$ phase diagrams of the model by constructing diagrams of multicritical points in the plane $\Delta$-$T$ for a fixed value of $U$, as shown in Fig.\,\ref{figmulti}. Thus, given a nematic-like system with parameters $(A, U)$, we can determine the multicritical points in the $\phi$-$T$ phase diagrams for different values of $\Delta$. Due to the large parameter space, we only focus on some representative values of the isotropic interaction $U$.
\begin{figure}
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{figures/figure_8a.pdf
\subcaption{$U=-3$}
\label{fig:16}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.481\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/figure_8b.pdf
\subcaption{$U=1$}
\label{fig:15}
\end{subfigure}
\caption{Lines of multicritical points in the $\Delta$-$T$ plane. MCE: higher-order critical end point. M$^\pm_1$, M$^\pm_2$, and M$^\pm_3$ are higher-order multicritical points.}
\label{figmulti}
\end{figure}
\subsection{Case with $U<0$}
By assuming attractive isotropic interactions with $(A,U)=(1,-3)$, we obtain the $\Delta$-$T$ diagram shown in Fig.\,\ref{fig:16}. We notice that the line of triple points meets the lines of critical end points at higher-order multicritical points M$_1^{\pm}$. Besides, the line of ordinary critical points meets the line of triple points at the higher-order multicritical end point MCE. For $\Delta<\Delta^{+}_1\approx0.994$, where $\Delta^{\pm}_1$ are the values of $\Delta$ at M$_1^{\pm}$, phase diagrams in the $\phi$-$T$ plane exhibit ordinary critical points related to vapor-liquid biphasic regions, critical end points (CE), and vapor-liquid-uniaxial triple points, a topology exemplified in Fig.\,\ref{fig:66}. Precisely at $\Delta=\Delta^{+}_1$, the lines of CE and triple points meet at the temperature $T^{+}_1\approx0.7298$. For values of model parameters corresponding to $\mathrm{M_{1}^\pm}$, $\phi$-$T$ phase diagrams do not exhibit a coexistence region between the isotropic vapor and the uniaxial phases. In the range $\Delta^{+}_1<\Delta<1$, the temperature of the CE point is higher than that of the triple point, which now represents a coexistence of isotropic (vapor and liquid) and biaxial phases. For maximal biaxiality $\Delta=1$, only isotropic and biaxial phases are stable, and $\phi$-$T$ phase diagrams are characterized by an ordinary vapor-liquid critical point, a Landau line, and, dependending on the value of $U<0$, a Landau tricritical point, as in Fig.\,\ref{fig:12}, or a Landau end point, as in Fig.\,\ref{fig:11}.
On the other hand, for $1<\Delta<\Delta^{-}_1\approx1.006$, the $\phi$-$T$ phase diagrams may exhibit uniaxial discotic phases, whose region of stability increases with $\Delta$. In addition, we have CE points and vapor-liquid-biaxial triple points, producing the same topology as in Fig.\,\ref{fig:66}. When $\Delta=\Delta^{-}_1$, the lines of $\mathrm{CE}$ and triple points meet at the temperature $T^{-}_1\approx0.7322$. For $\Delta^{-}_1<\Delta<\Delta_{\mathrm{MCE}}\approx1.063$, the $\phi$-$T$ phase diagrams also present $\mathrm{CE}$ points and vapor-liquid-biaxial triple points whose temperature approaches that of the vapor-liquid critical point as $\Delta\rightarrow\Delta_{\mathrm{MCE}}$. For biaxiality degree $\Delta=\Delta_{\mathrm{MCE}}$, the vapor-liquid-biaxial triple point and the ordinary vapor-liquid critical point meet at the temperature $T_{\mathrm{MCE}}=3/4$, and we cannot distinguish isotropic vapor and liquid phases.
For nematic systems with $\Delta_{\mathrm{MCE}}<\Delta<3$, the topology of the $\phi$-$T$ phase diagrams is the same as the one shown in Fig.\,\ref{fig:2}, the only multicritical point being a CE point separating regions of coexistence between the low-concentration isotropic phase and the high-concentration biaxial (at low temperatures) or uniaxial (at higher temperatures) phases.
For the intrinsically uniaxial cases $\Delta=0$ or $\Delta = 3$, the phase diagrams exhibit only isotropic and uniaxial phases, as exemplified in Sec.\,\ref{subsecA}.
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figures/phaseDiagramUn300.png}
\caption{Lines of multicritical points in the $\Delta$-$T$ plane, for $U=-3$. $\mathrm{LTC}$: Landau tricritical point. $\mathrm{MCE}$ represents a multicritical-end point. $\mathrm{M_{1}}$ represents a multicritical point in which a line of triple points and a line of critical-end points meet.}
\label{fig:16}
\end{figure}
\end{comment}
\subsection{Case with $U>0$}
Now, by considering repulsive isotropic interactions with $(A,U)=(1,1)$, we obtain the multicritical lines shown in Fig.\,\ref{fig:15}. Here, lines of CE, tricritical and triple points meet at multicritical points M$_2^{\pm}$. We also have the multicritical points M$_3^{\pm}$, where lines of CE, triple and ordinary critical points meet.
The topology of the $\phi$-$T$ phase diagrams is essentially symmetric with respect to the axis $\Delta=1$, except for the change in character of the uniaxial phases, from calamitic (for $0\leq\Delta<1$) to discotic (for $1<\Delta\leq 3$).
In the ranges $0<\Delta<\Delta^{+}_2\approx0.525$ or $\Delta_2^{-}\approx 1.3743 < \Delta < 3$, where $\Delta^{\pm}_{i}$ is the biaxiality parameter at $\mathrm{M}_i^{\pm}$, the topology of the $\phi$-$T$ phase diagrams is the same as the one shown in Fig.\,\ref{fig:2}, and the temperature of the CE point increases as the value of $\Delta$ becomes closer to $1$. For biaxiality in the ranges $\Delta^{+}_2<\Delta<\Delta^{+}_3\approx0.872$ or $\Delta^{-}_3\approx1.115 <\Delta<\Delta^{-}_2$, there exist isotropic-uniaxial and uniaxial-biaxial coexistence regions, as well as a tricritical (TC) point, as illustrated in Fig.\,\ref{fig:8}. Finally, for $\Delta^{+}_3 < \Delta < \Delta^{-}_3$, the TC point is replaced by a low-concentration CE point (or a Landau end point if $\Delta=1$) and an ordinary critical point associated with a biaxial-biaxial coexistence region, a topology exemplified in Fig.\,\ref{fig:13}. For biaxiality exactly equal to $\Delta^{+}_3$ or $\Delta^{-}_3$, the lines of critical and $\mathrm{CE}$ points meet the line of $\mathrm{TC}$ points and the biaxial-biaxial coexistence region is absent.
\begin{comment}
When the value of $\Delta$ increases in the range $\Delta^{+}_{\mathrm{M_{3}}}<\Delta<1$ the temperature of the $\mathrm{CE}$ and simple $\mathrm{C}$ points decreases and increases respectively. The region where the uniaxial (calamitic) phase is stable decreases until it disappears when $\Delta=1$.
\end{comment}
\begin{comment}
A $\mathrm{TC}$ point marks the upper limit of uniaxial-biaxial coexistence, the lower limit is determined by an isotropic-uniaxial-biaxial triple point. This behavior is similar for any phase diagram in the range.
\end{comment}
\begin{comment}
\begin{figure}[ht]
\centering
\includegraphics[width=8.89cm, height=8.0cm]{figures/phaseDiagramU100.png}
\caption{Lines of multicritical points in the $\Delta$-$T$ plane, for $U=1$. $\mathrm{LE}$ represents a Landau-end point. $\mathrm{M_{2}}$ represents a multicritical point in which a line of critical-end points become metastable with respect to a line of tricritical points and a line of triple point. $\mathrm{M_{3}}$ represents a multicritical point in which a line of tricritical points become metastable with respect to a line of critical-end points and a line of simple critical points.}
\label{fig:15}
\end{figure}
\end{comment}
\begin{comment}
For $1<\Delta<\Delta^{-}_3$ we have uniaxial and biaxial discotic structures as stable phases, and the stability region of uniaxial states increases with $\Delta$. In fact, the increase in the biaxiality degree also produces an increase in the temperatures of critical end points CE and biaxial-biaxial critical points C. For biaxiality $\Delta^{-}_3\approx1.115$ and temperature $T^{-}_3\approx0.2814$, the lines of $\mathrm{C}$ and $\mathrm{CE}$ points meet the line of $\mathrm{TC}$ points and the biaxial-biaxial coexistence region is absent.
\end{comment}
\section{Conclusions}\label{sec6}
We considered a lattice-gas version of the Maier--Saupe model for biaxial nematics with discrete orientations, in addition to an energetic term that described an isotropic interaction. The model is investigated in mean-field theory through a fully-connected spin-like system with inclusion of dilution effects. The free energy functional and the mean-field equations were obtained exactly.
For systems without isotropic interactions, $U=0$, we have drawn phase diagrams in terms of temperature and concentration of nematogens, with fixed value of $\Delta$. The case $\Delta = 1$ is particularly interesting due to the absence of a nematic uniaxial phase, and we find a line of Landau points which is limited by a Landau tricritical point ($\mathrm{LTC}$). In the cases $\Delta=0$ or $3$ the nematogens are intrinsically uniaxial, so that the phase diagrams show no biaxial nematic phase. Any other value of $\Delta$ leads to a diagram which presents a critical end point ($\mathrm{CE}$) at high concentration.
Systems with $U\neq0$ present a great variety of multicritical points depending on the character of the isotropic interaction and the biaxiality degree of the nematogens. To clarify this idea, diagrams of multicritical points were constructed in the $U$-$T$ plane for some values of $\Delta$, these diagrams show the different multicritical points that can be found in the phase diagrams.
It would be interesting to extend the present work to deal with the limit in which the orientational interactions are described by the potential in Eq.\,\eqref{eq:V12} with $\zeta=0$ and $\lambda\neq0$. This would allow comparison with the results obtained by Skutnik \emph{et al.} \cite{Skutnik2020} for a three-dimensional model with short-range interactions via constant-pressure Monte Carlo simulations. Such a comparison would point to possible multicritical phenomena which could be further investigated via simulations.
Finally, we point out that our model could in principle be used to fit experimental data from lyotropic systems, providing estimates of coupling energies and biaxiality parameters, if we allow for variation of the parameter $\Delta$ with both temperature and concentration of components in a lyotropic mixture. Models for this variation should be informed by calculations similar to those provided by Amaral \emph{et al.} for the change in micelle form induced by cosurfactant addition \cite{Amaral1997}.
\section{Acknowledgment}
This work was funded by CNPq, FAPESP, INCT/FCx, NAP/FCx, and Coordenação de Aperfeiçoamento de Pessoal de Nível Superior – Brasil (CAPES) – Finance Code 001.
|
\section*{Table of contents}
\noindent \textbf{Supplementary material on proofs and fairness models} \\
\noindent $\hookrightarrow$ Additional \textbf{Equality of Opportunity} Details \hrulefill Pg
\pageref{sec:sup_eoo}\\
\noindent $\hookrightarrow$ Handling \textbf{Statistical parity} \hrulefill Pg
\pageref{sec-stat-par}\\
\noindent $\hookrightarrow$ \textbf{Weak Hypothesis Assumption} Discussion \hrulefill Pg
\pageref{sec-wha-disc}\\
\noindent $\hookrightarrow$ General KL Distortion \ref{sec-kl-dist} \hrulefill Pg
\pageref{sec-kl-dist}\\
\noindent $\hookrightarrow$ Proof of Theorem \ref{thm-distortion-to-blackbox} \hrulefill Pg
\pageref{proof-thm-distortion-to-blackbox}\\
\noindent $\hookrightarrow$ Proof of Theorem \ref{cor-distortion-to-blackbox} and \ref{cor-distortion-to-blackbox-b} \hrulefill Pg
\pageref{proof-cor-distortion-to-blackbox}\\
\noindent $\hookrightarrow$ Proof of Theorem \ref{thm:entropyupper} and \ref{thm-boosting} \hrulefill Pg
\pageref{proof-thm-boosting}\\
\noindent $\hookrightarrow$ Proof of Theorem \ref{thm-alternative-alpha-in-alpha-tree} \hrulefill Pg
\pageref{proof-thm-alternative-alpha-in-alpha-tree}\\
\noindent $\hookrightarrow$ Proof of Lemma \ref{lem-entropies} \hrulefill Pg
\pageref{proof-lem-entropies}\\
\noindent $\hookrightarrow$ Proof of Theorem \ref{thm-eoo} \hrulefill Pg
\pageref{proof-thm-eoo}\\
\noindent \textbf{Supplementary material on experiments} \\
\noindent $\hookrightarrow$ \supplement Experiment Settings \hrulefill Pg
\pageref{app-exp-settings}\\
\noindent $\hookrightarrow$ Experiments on Statistical Parity \hrulefill Pg
\pageref{app-exp-SP-main}\\
\noindent $\hookrightarrow$ Additional Main Text Experiments \hrulefill Pg
\pageref{app-exp-rest}\\
\noindent $\hookrightarrow$ Neural Network Experiments \hrulefill Pg
\pageref{app-exp-nn}\\
\noindent $\hookrightarrow$ Proxy Sensitive Attributes \hrulefill Pg
\pageref{app-exp-proxy}\\
\noindent $\hookrightarrow$ Distribution Shift \hrulefill Pg
\pageref{app-exp-shift}\\
\noindent $\hookrightarrow$ High Clip Value \hrulefill Pg
\pageref{app-exp-clip}\\
\noindent $\hookrightarrow$ Example Alpha-Tree \hrulefill Pg
\pageref{app-exp-example}\\
\newpage
\input{supplementary/sec-extended-EOO}
\input{supplementary/statistical-parity}
\input{supplementary/weak-hypothesis-assumption}
\input{supplementary/general-kl-distortion.tex}
\input{supplementary/proof-distortion-black-box}
\input{supplementary/proof-distortion-corollary}
\input{supplementary/proof-boosting}
\input{supplementary/proof-audacious-scoring}
\input{supplementary/proof-bound-entropy}
\input{supplementary/proof-equality-of-opportunity}
\newpage
\input{supplementary/experiments-settings}
\newpage
\input{supplementary/experiments-SP-main-to-SI}
\newpage
\input{supplementary/experiments-restofmain}
\newpage
\input{supplementary/experiments-mlp}
\newpage
\input{supplementary/experiments-proxy}
\newpage
\input{supplementary/experiments-distribution-shift}
\newpage
\input{supplementary/experiments-clip}
\newpage
\input{supplementary/experiments-example-tree}
\end{document}
\section{Introduction}
\label{introduction}
The social impact of Machine Learning (ML) has seen a dramatic increase over the past decade -- enough so that the bias of model outputs must be accounted for alongside accuracy \citep{alAN,hpsEO,zvggFC}. Considering the various number of fairness targets \citep{mmslgAS} and the energy and CO2 footprint of ML \citep{mSD,sgmEA}, the combinatorics of training accurate \emph{and} fair models is non-trivial. This is especially so given the inherit incompatibilities of fairness constraints~\citep{kleinberg2017inherent} and the underlying tension of satisfying fairness whilst maintaining accuracy. One trend in the field "decouples" the two constraints by post-processing \emph{pretrained} (accurate) models to achieve fairer outputs~\citep{zvggFC}. Post-processing may be the only option if we have no access to the model's training data / algorithm / hyperparameters (etc.).
\newcommand{\des}[1]{\textbf{(#1)}}
Within the post-processing approach, three trends have emerged: learning a new fair model close to the black-box, tweaking the output subject to fairness constraints, and exploiting sets of classifiers.
If the task is class probability estimation \citep{rwID}, the estimated black-box is an accurate but potentially unfair posterior $\posunfair : \mathcal{X} \rightarrow [0,1]$ which neither can be opened nor trained further. The goal is then to learn a fair posterior $\posfair$ from it. In addition to the black-box constraint, a number of \emph{desiderata} can be considered for post-processing.
Ideally in correcting a black-box, we would want the approach to have \des{flexibility} in satisfying substantially different fairness criteria, \des{proximity} of the learnt \( \posfair \) to the original \( \posunfair \), and meaningful \des{composability} properties if, \emph{e.g.}, $\posfair$ was later treated as a new black-box to be post-processed. To facilitate a specific style of correction, we may also want the representation of the correction to facilitate \des{explainability} for auditing the post-processing procedure and bounds on the increased model \des{complexity} of the final classifier \( \posfair \). In training the correction, algorithmically we would also want guarantees for \des{convergence}.
\textbf{Our contribution} satisfies the aforementioned desiderata in its correction, representation, and algorithmic guarantees.
By leveraging recent theory in \emph{im}proper loss functions, we utilize a \emph{universal} correction of black-box posteriors defined by the \( \alpha \)-loss.
This allows for a flexible correction which yields convenient divergence bounds between $\posfair$ and $\posunfair$, a convenient form for the Rademacher complexity of the class of $\posfair$, and a simple composability property. Representation-wise, the corrections we learn are easy-to-interpret tree-shaped functions that we define as \textit{$\alpha$-trees}. Algorithmically speaking, we provide two formal boosting algorithms to learn $\alpha$-trees building upon seminal results \citep{kmOT}.
We demonstrate our algorithm for conditional value-at-risk (\(\cvar{}\)) \citep{wmFR}, equality of opportunity (EOO), and statistical parity; as depicted in \cref{fig:teaser}. Experiments are provided against five baselines on readily available datasets. All proofs and more experiments are in an Appendix denoted as \supplement.
\begin{figure}
\centering
\includegraphics[trim=35bp 602bp 505bp 61bp,clip,width=\textwidth]{Figs/FigSummary}%
\caption{Summary of using different \( \alpha \)-correction wrappers to obtain different fairness criteria guarantees. See \cref{sec-formal,sec-fairness} for full details on guarantees.}
\label{fig:teaser}
\fnegativespace
\end{figure}
\section{Related Work}
\label{sec-rel}
Post-processing models to achieve fairness is one of three different categories in tackling the ML + fairness challenge \citep[Section 6.2]{zvggFC}. Although other notions exist, \emph{e.g.} individual fairness \citep{dwork2012fairness}, we limit our analysis to group fairness, which concerns itself with ensuring that statistics of sub-populations are similar.
We further segment this cluster into three subsets: (I) approaches learning a new model with two constraints: being close to the pretrained model and being fair \citep{kgzMB,pmsyPP,wrcOS,yang2020fairness}; (II) approaches biasing the output of the pretrained model at classification time, modifying observations for fairer outcomes \citep{alAN,hpsEO,lrbsvpBM,mwTC,wgosLN,yang2020fairness}; and (III) techniques consisting of exploiting sets of models to achieve fairness \citep{diklDC}. None of these approaches formulates substantial guarantees on all of the desiderata in the introduction. Some bring contributions with the \des{flexibility} of being applicable to more than two fairness notions \citep{cpfghAD,wrcOS,diklDC,yang2020fairness}. Two of which provide the convenience of analytic conditions on new fairness notions to fit in the approach \citep{wrcOS,diklDC}. However, for all of them, the algorithmic price-tag is unclear \citep{cpfghAD,diklDC} or heavily depends on convex optimization routines \citep{wrcOS}. \citet{alAN,yang2020fairness} provide strong guarantees regarding \textbf{(proximity)}, w.r.t. \textit{consistency and generalization}. To our knowledge, our approach of correcting prediction unfairness through improper losses \citep{nssBP} is new.
\section{Setting and Motivating Example}
\label{sec-example}
Let $\mathcal{X}$ be a domain of observations, $\mathcal{Y} \defeq \{-1,1\}$ be labels and $S$ is a sensitive attribute in $\mathcal{X}$. We assume that the modalities of $S$ induce a partition of $\mathcal{X}$. We further let \( \meas{D} \) denote the joint measure over \( \mathcal{X} \times \mathcal{Y} \), \( \meas{M} \) denote the marginal measure over \( \mathcal{X} \), and \( \prior \defeq \pr[\Y = 1] \) being the prior. We denote conditioning of \( \meas{M} \) through a subscript, \emph{e.g.}, \( \meas{M}_{s} \) for \( s \in S \) denotes \( \meas{M} \) conditioned on a sensitive attribute subgroup \( \SSS = s \). We leave the \( \sigma \)-algebra to be implicit (which is assumed to be the same everywhere).
As is often assumed in ML,
sampling is i.i.d.; we make no notational distinction between empirical and true measures to simplify exposition -- most of our results apply for both
Consider the task of learning a function \( \posterior \in [0,1]^{\mathcal{X}} \) to estimate the true posterior \( \bayespo({\bm{e}}{x}) = \pr[\Y=1 | \X = {\bm{e}}{x}] \) in binary classification.
For instance, we may want to predict the probability of hiring an applicant for a company.
Given the pointwise loss \( \poirisk(\posterior(\X),\bayespo(\X)) \), which determines the loss of a single example, the \emph{(total) risk} is defined as
\begin{eqnarray}\label{eq:risk_unfair}
\totrisk(\posterior; \meas{M}, \bayespo) & \defeq & \expect_{\X \sim \meas{M}} \left[\poirisk(\posterior(\X),\bayespo(\X))\right]
\end{eqnarray}
(with slight abuse of notation). A low risk corresponds to good classification performance. In this paper, we consider the risk determined by the \emph{log-loss}:
\begin{equation}
\totrisk(\posterior; \meas{M}, \bayespo)
\defeq
\mathop{\expect}_{\X \sim \meas{M}}
\left[
\bayespo(\X)\cdot - \log \posterior(\X) + (1-\bayespo(\X))\cdot - \log (1-\posterior(\X)) \right].\label{lossgen}
\end{equation}
We now consider a simple fairness problem, centered around the example of \cref{fig:alpha-trees}.
Suppose that we are given a black-box \( \posunfair \) which predicts hiring probabilities without considering fairness. Although the minimized total risk of \eqref{eq:risk_unfair} might be small, there can be discrepancies in the performance between different subgroups.
Instead of considering the total risk, the predictive performance of specific subgroups can be examined through the \emph{subgroup risk} \( \totrisk(\posunfair; \meas{M}_{s}, \bayespo) \), for a subgroup \( s \in S \). For instance, we might want to examine the discrepancy of subgroup risks among the age of applications. A natural fairness task would be to improve the worst performing subgroup, say \( s_{\mathrm{w}} \in S \).
To post-process unfairness, we want to learn a function \( \alpha: \mathcal{X} \to \mathbb{R}\) which ``wraps'' \( \posunfair \) and \emph{lowers} the worst subgroup risk.
We propose the following ``wrapping'', inspired by improper loss functions \citep{nssBP}
\begin{figure}[!t]
\centering
\includegraphics[trim=0pt 528pt 217pt 0pt,clip,width=\textwidth]{Figs/synth_example}%
\caption{%
Improving \( \cvar{} \) for a toy hiring task with \(\alpha\)-trees.
An \( \alpha\)-tree (left) transforms the posterior via \eqref{ourposfair} (middle);
resulting in an input-dependent fairness correction of the posterior \( \posunfair\) (right).
}
\label{fig:alpha-trees}
\fnegativespace
\end{figure}%
\begin{eqnarray}
\posfair({\bm{e}}{x}) & \defeq & \frac{\posunfair({\bm{e}}{x})^{\alpha({\bm{e}}{x})}}{\posunfair({\bm{e}}{x})^{\alpha({\bm{e}}{x})}+(1-\posunfair({\bm{e}}{x}))^{\alpha({\bm{e}}{x})}} \; \in [0,1].\label{ourposfair}
\end{eqnarray}
Notice when \( \alpha({\bm{e}}{x}) = 1 \) the resulting posterior is the original \( \posunfair({\bm{e}}{x}) \).
Importantly, \eqref{ourposfair} is flexible enough to transform any input black-box \( \posunfair \) to any needed \( \posfair \). Looking at \cref{fig:alpha-trees} (left and middle), intuitively by setting different \( \alpha({\bm{e}}{x}) \) values, \eqref{ourposfair} {\color{orange} ``sharpens''} (yellow, \( \alpha > 1\)), {\color{blue}``dampens''} (blue, \( 0 < \alpha < 1 \)), or {\color{red}``polarity reverses''} (red, \( \alpha < 0 \)) the original posterior \( \posunfair \).
To improve fairness, we need a combination of these corrections to accommodate different subsets of the input domain (thus learning \( \alpha(.) \) as a function).
We specifically learn \( \alpha(.) \) to be a tree structure, which allows an interpretable correction alongside other formal properties (\cref{sec-formal}).
\begin{definition}\label{def-tree}
An \textbf{$\alpha$-tree} is a rooted, directed binary tree, with internal nodes labeled with observation variables. Outgoing arcs are labeled with tests over the nodes' variable. Leaves are real valued. $\leafset(\alphatree)$ is the leafset of $\alpha$-tree $\alphatree$. An \( \alpha\)-tree induces a correction over posteriors as per \eqref{ourposfair} with \( \alpha = \alphatree \).
\end{definition}
Our fairness problem is now learning an \( \alpha \)-tree \( \posterior \) which provides a corresponding correction that improves the worst subgroup risk, \emph{i.e.}, \( \totrisk(\posunfair; \meas{M}_{s_{\mathrm{w}}}, \bayespo) > \totrisk(\posfair; \meas{M}_{s_{\mathrm{w}}}, \bayespo) \). The entirety of \cref{fig:alpha-trees} presents such a process. In this hiring task example, the ground truth hiring rate is constant \emph{w.r.t.} the inputs, \( \bayespo({\bm{e}}{x}) = 0.7 \). Despite this, the black-box \( \posunfair(.) \) unfairly depends on the age of applicants and incorrectly depends on a noise feature \( \mathcal{Z} \). By choosing \( \alpha({\bm{e}}{x}) \) as per \cref{fig:alpha-trees} (left), the correction
changes
\( \posunfair \) to be closer to \( \bayespo \), improving the risk of the worst subgroup (alongside the other subgroup in this example): the worse-case loss improves from \( 1.09 \) to \( 0.62 \).
Although the fairness criteria discussed might be considered simple, the procedure of iteratively minimizing the worse performing subgroup can be used to improve the \emph{conditional value-at-risk} (lower is better) fairness criteria \citep{wmFR}:
\begin{eqnarray} \label{eq:cvar}
\cvar{\beta}(\posfair) &\defeq& {\expect}_{\SSS} [\totrisk(\posfair; \meas{M}_{\SSS}, \bayespo) \mid \totrisk(\posfair; \meas{M}_{\SSS}, \bayespo) \geq L_\beta],
\end{eqnarray}
where $L_\beta$ is the risk value for the $\beta$ quantile among subgroups, which is user-defined.
The difference is that \( \cvar{\beta} \) not only considers the worse case subgroup, but all subgroups above the \( L_{\beta} \) risk value (let these subgroups be \(\mathcal{S}_{\beta}\)). One can simply iteratively improve all \( s \in \mathcal{S}_{\beta} \), as done in the example of \cref{fig:alpha-trees}. Indeed, in the example, \( \cvar{\beta} \) with \( \beta = 0.9 \) improves from \( 1.09 \) to \( 0.62 \) (which is equivalent to worse-case loss in this case).
\begin{figure}[!t]
\begin{minipage}{0.55\linewidth}
\vspace{-\fboxsep}
\begin{algorithm}[H]
\caption{\topdown($\meas{M}_{\mbox{\tiny{\textup{t}}}},\postarget, \alphatree_0, B$)}\label{alg:topdown}
\begin{algorithmic}
\STATE \textbf{Input} mixture $\meas{M}_{\mbox{\tiny{\textup{t}}}}$, posterior $\postarget$, $\alpha$-tree $\alphatree_0$, $B\in \mathbb{R}_{+*}$;
\STATE Step 1: $\alphatree \leftarrow \alphatree_0$;
\STATE Step 2 : \textbf{while} stopping condition not met \textbf{do}
\STATE \hspace{0.2cm} Step 2.1 : pick leaf $\leaf^\star \in \leafset(\alphatree)$; // \emph{i.e.} heaviest leaf
\STATE \hspace{0.2cm} Step 2.2 : $h^\star \leftarrow \arg\min_{h\in \mathcal{H}} \entropy(\alphatree(\leaf^\star, h); \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$;
\STATE \hspace{0.2cm} Step 2.3 : $\alphatree \leftarrow \alphatree(\leaf^\star, h^\star)$; // split using $h^\star$ at $\leaf^\star$
\STATE Step 3 : label leaves \( \forall \leaf \in \leafset(\alphatree) \):
\negativespace
\begin{eqnarray}
\alphatree(\leaf) &\defeq& \normalizedlogit\left(\frac{1+\edge(\meas{M}_{\leaf}, \postarget)}{2}\right), \;\; \textrm{// \(\alpha\)-value}\label{alphatreeoutputCONS}
\end{eqnarray}%
\STATE \textbf{Output} $\alphatree$;
\end{algorithmic}
\end{algorithm}
\end{minipage}%
\hfill%
\begin{minipage}{0.4\linewidth}
\begin{figure}[H]
\centering
\includegraphics[trim=25bp 520bp 450bp 25bp,clip,width=\textwidth]{Figs/FigBigAlphaTree2}%
\caption{Picking $\alphatree_0$ a stump on the fairness attribute allows to finely tune growths of sub-$\alpha$-trees to the fairness criterion at hand.}
\label{fig:bigalpha}
\vspace{-0.5cm}
\end{figure}
\end{minipage
\fnegativespace
\end{figure}
\section{Fairness and Societal Considerations}
\label{sec-fairness}
In this section, we present the fairness guarantees \topdown~can achieve. In particular, we provide a discussion about how \cref{thm-boosting} can guarantee minimization of the \( \cvar{} \) criteria. Furthermore, we provide alternative inputs to \topdown~which allows for EOO to be targeted as a fairness criteria. In the \supplement Section \ref{sec-stat-par}, we further present a treatment of statistical parity. Lastly, we discuss how using \( \alpha \)-trees provides explainable corrections and how utilization of the sensitive attribute (as per \cref{fig:bigalpha}) can be circumvented.
\csubsec{Guarantees on CVaR}
As discussed in previous sections, one way to improve the \( \cvar{} \) fairness criteria (as per \eqref{eq:cvar}) is to focus optimization on the worst treated subgroups. Given a specified quantile group \( \beta \) and the set of worse subgroups \( \mathcal{S}_{\beta} \), we can repeat \eqref{topdown_cvar} until \( \cvar{\beta}(\posfair) \) gets below a threshold or (more specifically) its worst tread group gets a risk below a threshold (\emph{i.e.}, a stopping criterion). Importantly, \cref{thm-boosting} provides a guarantee: to ensure \( \cvar{\beta} \) is below \( \varepsilon \), we simply need to boost for \( \vert S \vert \) times the tree size bound \( \vert \leafset(\alphatree) \vert \) given in \cref{thm-boosting}.
\csubsec{Guarantees on EOO}
EOO requires to smooth discrimination within an ``advantaged'' group, modeled by the label $\Y=1$ \citep{hpsEO}. We say that $\posfair$ achieves $\varepsilon$-equality of opportunity iff a mapping $\predfair$ of $\posfair$ to $\mathcal{Y}$ (\textit{e.g.} using the sign of its logit \( \logit \)) satisfies
\begin{eqnarray}
\max_{s\in S} \mathop{\pr}_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right] - \min_{s\in S} \mathop{\pr}_{\X \sim \meas{P}_{s}} \left[\predfair(\X) = 1 \right] &\leq& \varepsilon\label{eqEOO},
\end{eqnarray}
where $\meas{P}_s$ is the positive observations' measure conditioned to value $\SSS = s$ for the sensitive attribute.
It is clear that EOO can be antagonistic to accuracy: the rate of advantage in the data \( \meas{D} \) may not be equal among the subgroups. As such, unlike \( \cvar{} \), we do not want to target the Bayes posterior \( \postarget \not\leftarrow \bayespo \) for EOO. Instead, we target a skewed posterior which aims to improve the least advantaged subgroup, \emph{i.e.,} increasing
$s^\circ \in \arg\min_{s\in S} \pr_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right]$.
Our strategy consists of picking a target posterior which skews part of the original \( \bayespo \) to be more advantaged, thus reducing the LHS of \eqref{eqEOO} until \eqref{eqEOO} is satisfied%
\footnote{If we instead \textit{reduce} $\arg\max_{s\in S} \pr_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right]$ we get a symmetric strategy.
The application informs which to use: if positive class implies money spending (\textit{e.g.} loan prediction), then our strategy implies spending more money; while the latter aims to reduce money lent to achieve fairness.}.
For this, we create a $(p,\delta)$-\textit{pushup} of $\bayespo$, defined in \supplement \cref{sec:sup_eoo}.
Fig.~\ref{fig:eta-diff-eoo} (right) presents an example of a pushup map.
Notice that the pushup only changes the predicted probability of example which do not have a ``confident prediction'' (the interval $[\eta(p), 1/2 + \delta]$).
Intuitively, \( p \) controls how many examples are corrected and \( \delta \) controls how much the correction ``pushes up'' advantage, further discussion in \supplement \cref{sec:sup_eoo}.
We then run \topdown~using as mixture the \textit{positive} measure conditioned to $S = s^\circ$ and $p \defeq \pr_{\X \sim \meas{P}_{s^*}} \left[\predfair(\X) = 1 \right] + \epsilon/(K-1), \delta \defeq K\epsilon/(K-1)$, with $K>1$ user-fixed. Thus, we do
\begin{tcolorbox}[colframe=blue,boxrule=0.5pt,arc=4pt,left=6pt,right=6pt,top=2pt,bottom=4pt,boxsep=0pt]
\begin{equation}
\textrm{
Use \topdown~with $\meas{M}_{\mbox{\tiny{\textup{t}}}} \leftarrow \meas{P}_{s^\circ}$ and $\postarget \leftarrow \bayespo_{p,\delta}$
}.
\tag{EOO}
\end{equation}
\end{tcolorbox}
\begin{theorem}\label{thm-eoo}
If \topdown~is run until $\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \leq (\epsilon^4/2) + \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[ \entropy(\postarget(\X))\right]$, then after the run we observe $\pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right] - \pr_{\X \sim \meas{P}_{s^\circ}} \left[\predfair(\X) = 1 \right] \leq \epsilon$.
\end{theorem}%
In the full context of EOO, in the optimization we should not wait to get the bound on $\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$.
Rather, we should make sure (a) we update $\arg\min_{s\in S} \pr_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right]$ (and thus $s^\circ$) after each split in the $\alpha$-tree and (b) we keep $\arg\max_{s\in S} \pr_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right]$ as is, to prevent switching targets and eventually composing pushup transformations for the same $S=s^\circ$, which would not necessarily comply with our theory.
Notably, the guarantee presented in \cref{thm-eoo} depends on the mapping \( \predfair \) and not the direct posterior \( \posfair \), as typically considered~\citep{hpsEO}.
When taking a threshold (sign of the logit), \( \predfair \) can be interpreted as forcing the original posterior to be extreme values of 0 or 1.
Unlike the \( \cvar{} \) case, EOO (as per \eqref{eoo_mapping}, \supplement) requires an explicit approximation of \( \bayespo \). In practice, we find that taking a simple approximation of \( \bayespo \) still can yield fairness gains. However, if one does not want to make such an approximation, one can adapt the statistical parity approach (detailed in \supplement, Section \ref{sec-stat-par}).
Similarly, if wants to consider the typical EOO definitions depending on posterior values, the target measure can be replaced (\emph{i.e.}, swapping measure \( \meas{M}_{\mbox{\tiny{\textup{t}}}} \) with the positive examples \( \meas{P} \)).
\csubsec{Explainability}
\( \alpha \)-trees using the initialization proposed in \cref{sec:growing_alpha_trees} (and \cref{fig:bigalpha}) allows for \des{explainability} properties similar to that of decision tree classifiers.
Fixing a sensitive attribute \( \SSS = s \), the corresponding sub-\(\alpha\)-tree \( \alphatree_{s} \) can be examined to scrutinize the correction done for the corresponding subgroup. If the splits of the \( \alpha \)-tree are simple, similarly to standard decision tree classifiers, corresponding partitions of the input domain can be examined. Furthermore, the type of corrections can also be examined, as discussed in \cref{sec-example}; where corrections can be classified as ``sharpening'', ``dampening'', or ``polarity flipping'' depending on the leaves' \( \alpha \)-values.
\csubsec{Usage of sensitive attribute}
Post-processing methods have been flagged in the context of fair classification for the fact that they require explicit access to the sensitive feature at classification time \citep[$\S$ 6.2.3]{zvggFC}. Our basic approach to the induction of $\alpha$-trees falls in this category (Fig. \ref{fig:bigalpha}), but there is a simple way to \textit{mask} the use of the sensitive attribute and the polarity of disparate treatment it induces: it consists in first inducing a decision tree to \textit{predict} the sensitive feature based on the other features and use this decision tree as an alternative initialization to naively splitting on subgroups. We thus also \textit{redefine} sensitive groups based on this decision tree -- thus alleviating the need to use the sensitive attribute in the \( \alpha \)-tree.
The use of \emph{proxy sensitive attributes} in a similar manner has seen ample use in a various domain such as health care~\citep{bUP,bkbkUB} and finance~\citep{fbwbsehrUO}. We however note that its application in post-process and \( \alpha \)-trees may not be appropriate across all domains~\citep{dfkmsUP}.
\section{Experiments}
\section{Experiments}
\label{sec-experiments}
To evaluate \topdown\footnote{Implementation public at: \url{https://github.com/alexandersoen/alpha-tree-fair-wrappers}}, we consider three datasets presenting a range of different size / feature types, Bank and German Credit (preprocessed by \aif~\citep{aif360}) and the American Community Survey (ACS) dataset preprocessed by \folktables\footnote{Public at: \url{https://github.com/zykls/folktables}} \citep{dhmsRA}. The \supplement~(pg \pageref{app-exp-settings}) presents all results at length (including considerations on proxy sensitive attributes, distribution shift, and interpretability), along with the different black-boxes considered (random forests and neural nets). We concentrate in the Section on the ACS dataset for income prediction in the state of CA and evaluate \topdown's application to various fairness criteria (as per \cref{sec-fairness} and \supplement pg \pageref{sec-stat-par}) with Random Forests (RF).
For these experiments, we consider \( \emph{age} \) as a binary sensitive attribute with a bin split at $25$ (a trinary modality is deferred to the \supplement).
For the black-box, we consider a clipped (\cref{assumptionBUNF} with \(B=1\)) random forest (RF) from \sklearn calibrated using Platt's method \cite{platt1999probabilistic}. The RF consists of an ensemble of 50 decision trees with a maximum depth of 4 and a random selection of 10\% of the training samples per decision tree.
Data is split into 3 subsets for black-box training, post-processing training, and testing; consisting of 40:40:20 splits in 5 fold cross validation.
For \EOO, we utilize an out-of-the-box Gaussian Naive Bayes classifier from \sklearn to approximate \( \bayespo \).
\input{Figs/fairness_models_figure}
\noindent\textbf{Multiple fairness criteria}
We evaluate \topdown for \( \cvar{} \), equality of opportunity \( \EOO \), and statistical parity \( \SP \).
The complete treatment of SP is pushed to \supplement~(Sections \ref{sec-stat-par}, \ref{app-exp-SP-main}).
\( \SP \) aims to make subgroup's expected posteriors similar and is popular in a various post-processing methods~\citep{wrcOS,alAN}. The definition can be found in \supplement (pg \pageref{sec-stat-par}) along with the strategy used in \topdown.
Conservative and audacious updates rules are also tested. For each of these \topdown~configurations, we boost for \( 32 \) iterations. The initial \(\alpha\)-tree is initialized as in \cref{fig:bigalpha}.
We compare against 5 baseline approaches. For \( \cvar{} \) we consider the in-processing approach (\basecvara) presented in \citet{wmFR}. For \EOO, we consider a derived predictor (\baseeooa) \citep{hpsEO}. Our SP baselines include an optimized score transformation approach (\baseeoob)
~\citep{wrcOS}; a derived predictor modified for \SP (\basespa) \citep{hpsEO}; and a randomized threshold optimizer approach (\basespb) \citep{alAN}.
We denote the clipped black-box as \basebb.
The experiments for \( \cvar{} \) and \( \EOO \) are summarized in \cref{fig:main_text_fairness}; the full plot with \( \SP \) is presented at \supplement \cref{fig:rf_acs_fairness}. For clarity we only plot the baselines and wrappers which are directly associated to each fairness criteria. We also plot the posterior mean difference between the data and debiased posteriors \meandiff (0/1 loss) to examine the effects on accuracy.
\noindent\textbf{For \( \cvar{} \)}, both conservative and audacious approaches decreases \( \cvar{} \), which results in better \( \cvar{} \) values than both the original \basebb and in-processing baseline \basecvara~-- which is good news since \basecvara~directly optimizes $\cvar{}$. We note that there are cases in which the in-processing approach is better than ours (trinary sensitive attributes in \supplement), but this is expected given \basecvara's optimization goal. Interestingly, the audacious update is superior in both \( \cvar{} \) and \meandiff~than the conservative update. This is also consistent for trinary sensitive attributes. Thus, the audacious update is desirable when optimizing \( \cvar{} \). Another observation is that only one sensitive attribute subgroup's \( \alpha \)-tree is initialized (only one `\( \times \)'). This indicates that after \( 32 \) iterations the worse case subgroup does not change in the binary case.
\noindent\textbf{For \EOO}, there is a huge difference between conservative and audacious updates as the former gets to the most fair outcomes of all baselines. Even if we used early stopping or pruning of the $\alpha$-tree (taking an earlier iterations) the audacious update would fail at producing outcomes as fair as its conservative counterpart. Furthermore, the audacious update comes with a significant degradation of accuracy \meandiff. Furthermore, by looking at the iterations in which subgroup \( \alpha \)-trees are initialized, the audacious update causes large (primarily bad) jumps in performance. This rejoins our remark on the interest of having a conservative update in Section \ref{sec-formal}. When compared to \baseeooa, we find that the conservative \topdown approach produces lower \EOO. However, \baseeooa tend to have better accuracy scores in \meandiff. These observations are consistent with the trinary sensitive attribute (\supplement).
\noindent\textbf{For \SP}, we can observe fairness results that can be on par with contenders for the conservative update, but observe a substantial degradation of MD. This, we believe, follows from a simple plug-in instantiation of $\meas{M}, \postarget$ for the fairness notion in \supplement Section \ref{sec-stat-par}, resulting in potentially harsh updates. In \supplement~(pg \pageref{sec-stat-par}), we discuss an alternative approach using ties with optimal transport.
\section{Growing Alpha-Trees}%
\label{sec:growing_alpha_trees}%
We now introduce the procedure to grow an \( \alpha \)-tree via a boosting algorithm \topdown, \cref{alg:topdown}. \topdown~can be thought of as a generalization of the standard decision tree induction used for classification \citep{kmOT}.
We first introduce relevant concepts from decision tree induction to explain \topdown. We contextualize \topdown through its application in improving the \( \cvar{} \) criteria
We first introduce a technical assumption for the black-box \( \posunfair \) to be post-processed:
\begin{assumption}\label{assumptionBUNF}
The black-box prediction is bounded away from the extremes: $\exists B>0$ such that
\begin{eqnarray}
\mathrm{Im} (\posunfair) \subseteq \mathbb{I} & \defeq & \left[({1+\exp (B)})^{-1}, ({1+\exp (-B)})^{-1}\right] \quad \mbox{(a.s.)} \label{eqclipSUP}.
\end{eqnarray}
\end{assumption}
Compliance with Assumption \ref{assumptionBUNF} can be done by clipping the black-box's output with a user-fixed $B$ or making sure it is calibrated and then finding $B$.
\csubsec{Entropy-based updates for fairness}
An important component in standard decision tree induction is the \emph{edge} function, which measures the \emph{label purity} (proportion of positive examples) of a decision tree node. We introduce a generalization which considers the \emph{alignment purity} of a black-box.
\begin{definition}\label{def:edge}
Let $\logit(u) \defeq \log(u/(1-u))$ the logit of $u \in (0,1)$ and $\normalizedlogit(u) \defeq \logit(u)/B$ a normalization which satisfies $\normalizedlogit(\mathbb{I}) = [-1,1]$.
The \textbf{alignment edge} of $\meas{M}_{\mbox{\tiny{\textup{t}}}}$ and $\postarget$ is defined as,
\begin{eqnarray}
\edge(\meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \defeq & \expect_{(\X, \Y) \sim \measdtarget} \left[\Y \normalizedlogit(\posunfair(\X))\right],\label{defEDGEGEN}
\end{eqnarray}%
where \( \measdtarget \) is the joint measure induced by \( \meas{M}_{\mbox{\tiny{\textup{t}}}} \) and \( \postarget \).
With Assumption~\ref{assumptionBUNF}, $\edge(\meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \in [-1,1]$.
\end{definition}
By replacing the normalized logit \( \normalizedlogit \) with a constant \( 1 \), \eqref{defEDGEGEN} reduces to a measure of label purity used in regular classification. In our case, \eqref{defEDGEGEN} measures how well the black-box \( \posunfair \) ``aligns'' with the true labels \( \Y \) through the logit. This also takes into account the ``confidence'' of the black-box's predictions: for a high alignment purity, predictions not only need be correct but also to be highly confident (\( \posunfair \) close to the endpoints of \(\mathbb{I}\)).
Similar to how the splits of a decision tree classifier are determined by the entropy of a tree's label purity, an \( \alpha \)-tree splits based on its alignment entropy.
\begin{definition}\label{defENTROPYA}
Given an $\alpha$-tree $\alphatree$ with leafset $\leafset$, when Assumption \ref{assumptionBUNF} is satisfied, the \textbf{entropy of} $\alphatree$ is:
\begin{eqnarray}
\entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \defeq & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[\entropy_1(\leaf; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)\right],\label{defBINENT}
\end{eqnarray}%
where $\entropy(q) \defeq -q \log(q) - (1-q)\log(1-q)$,
$\entropy_1(\leaf; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \defeq \entropy\left((1+\edge(\meas{M}_\leaf, \postarget))/2\right)$, $\meas{M}_\leaf$ is $\meas{M}_{\mbox{\tiny{\textup{t}}}}$ conditioned to leaf $\leaf \in \leafset(\alphatree)$, and $\meas{M}_{\leafset(\alphatree)}$ is a measure induced on $\leafset(\alphatree)$ by leaf weights on $\meas{M}_{\mbox{\tiny{\textup{t}}}}$.
\end{definition}
\begin{theorem}\label{thm:entropyupper}
For any \( \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget \),
let \( \alphatree \)'s leaves follow \eqref{alphatreeoutputCONS}. Then
%
\( \totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \leq \entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \).
\end{theorem}
\cref{alg:topdown} can now be explained by repeatedly leveraging \cref{thm:entropyupper}.
Suppose that we have a hypothesis set of possible splits \( \mathcal{H} \) to grow our \( \alpha \)-tree. Denote \( \alphatree(\leaf, h) \) as the \( \alpha \)-tree \( \alphatree \) split at leaf \( \leaf \in \leafset(\alphatree) \) using test \( h \). The inner loop within Step 2 is the process of finding the best possible leaf splits
which helps to minimize the \( \alpha \)-tree's entropy and to reduce
the risk as per \cref{thm:entropyupper}. The \( \alpha \)-values of \eqref{alphatreeoutputCONS} calculated in step 3 are those used to ensure \cref{thm:entropyupper} holds.
By setting \( \meas{M}_{\mbox{\tiny{\textup{t}}}} \leftarrow \meas{M}_{s_{\mathrm{w}}} \) and \( \postarget \leftarrow \bayespo \),
\cref{alg:topdown} improves \( \cvar{} \) by iteratively improving the \( \alpha\)-tree's worst subgroup entropy \( \entropy(\alphatree; \meas{M}_{s_{\mathrm{w}}}, \bayespo) \), which as a surrogate improves the worst subgroup risk \( \totrisk(\posunfair; \meas{M}_{s_{\mathrm{w}}}, \bayespo) \). To accommodate for different \( \beta \) quantiles values for \( \cvar{} \), \topdown can be run repeatedly (replacing the initial input tree \( \alphatree_{0} \)) to progressively improve all \( s\in \mathcal{S}_{\beta}\). Hence, to reduce $\cvar{}$, we basically
\begin{tcolorbox}[colframe=blue,boxrule=0.5pt,arc=4pt,left=6pt,right=6pt,top=2pt,bottom=4pt,boxsep=0pt]
\begin{equation} \label{topdown_cvar}
\textrm{
use
\topdown~with $\meas{M}_{\mbox{\tiny{\textup{t}}}} \leftarrow \meas{M}_s$ $(s \in \mathcal{S}_\beta)$ and $\postarget \leftarrow \bayespo$.}
\tag{\(\cvar{}\)}
\end{equation}
\end{tcolorbox}%
As alluded to by the notation used to instantiate \topdown, the inputs of the algorithm can be instantiated to optimize for fairness criteria beyond \( \cvar{} \). This is discussed in \cref{sec-fairness}. In the usual ML setting, \( \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget \) can be \textit{estimated} from a training sample (see \cref{sec-experiments}).
\csubsec{Initialization}
In the procedure of improving \( \cvar{} \), the worst subgroups can be iteratively improved. However, we also need to make sure that improvement of a subgroup does not adversely affect another subgroup (which could potentially harm \( \cvar{} \) instead). As such, we introduce an additional structure to the \( \alpha \)-tree \( \alphatree \) by tweaking the initial tree structure \( \alphatree_{0} \) used in \topdown.
Since the fairness attribute \( \SSS \) partitions the dataset, a convenient choice of initializing the \( \alpha \)-tree is to split by the subgroup modalities, as depicted in \cref{fig:bigalpha}. As such, we grow separate sub-\(\alpha\)-trees for each of the sensitive modalities. For \( \cvar{} \), this allows the subgroup risk of individual subgroups to be tweaked without adversely affecting other subgroups.
\section{Formal Properties}
\label{sec-formal}
We move to the formal properties of our approach. We first detail the background of improper loss functions which motivates our correction given in \cref{ourposfair}. We then present the formal properties of this correction. The useful properties of having \( \alpha(.) \) represented by a tree structure is then discussed. Finally, we present a convergence analysis of \cref{alg:topdown} and an alternative boosting scheme.
\paragraph{Can \( \posfair \) as per \eqref{ourposfair} correct (any) potential unfairness? Yes.}
In short, this comes from recent theory in \emph{improper loss functions} for class probability estimation (CPE) \citep{nssBP}. We are interested in the pointwise minimizer (eventually set-valued) of:
\begin{eqnarray}
\pseudob(\posterior) & \defeq & \arg\inf_{u\in[0,1]} L(u,\posterior).
\end{eqnarray}
Dubbed as the \emph{Bayes tilted estimate} of a loss \(\loss\) \citep{nssBP}, \( \pseudob(\posterior) \) is the set of optimal ``responses'' given a ground truth (pointwise) posterior \( \posterior \). Common loss functions are \emph{proper}: the ground truth value \( \posterior \in \pseudob(\posterior) \) is an optimal response. However, in the case where \( \posterior \) cannot be trusted (for instance when it is \emph{unfair}), we may not want to default to imitating \( \posterior \). In addition we also want to make sure that the Bayes tilted estimate can fit to any desired (in our context, \emph{fair}) target. The so-called $\alpha$-loss \( \loss^{\alpha} \), which generalizes the (proper) log-loss, is a good candidate parameterized by a variable \( \alpha \). Its Bayes tilted estimate is the pointwise version of \eqref{ourposfair}, for \( \alpha \not \in \{0, \infty\}\) and \( \posterior \neq 1/2 \):
\begin{eqnarray}
\apseudob(\posterior) & = & \left\{\posterior^\alpha/(\posterior^\alpha+(1-\posterior)^\alpha)\right\} .\label{eqCORRECT}
\end{eqnarray}
Importantly for \( \alpha \)-losses, for any $\posterior \not\in \{0,1/2,1\}$ and any $\posterior' \in (0,1)$, there exists $\alpha \in \mathbb{R}$ in \eqref{eqCORRECT} such that $\apseudob(\posterior) = \{\posterior'\}$. This property, called \emph{twist-properness} \citep{nssBP}, allows for any pointwise correction. By extending \( \alpha \) to a function (of \( {\bm{e}}{x} \in \mathcal{X} \), as per \eqref{ourposfair}), twist-properness ensures that given an initial unfair posterior an appropriately learned \( \alpha(.) \) can correct any unfairness. This allows for \des{flexible} fairness post-processing -- different \( \alpha \) functions can be learned for different criteria (\emph{i.e.}, \cref{fig:teaser}).
\paragraph{Why use the Log-Loss?}
As per \cref{sec-example}, we minimize the log-loss. We choose the log-loss for two reasons: \ding{172} it is strictly proper and so minimizing \( \totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \) (\emph{i.e.}, via \topdown) ``pushes'' $\posfair$ towards target $\postarget$; and \ding{173} it is the $\alpha$-loss for $\alpha = 1$, so we are guaranteed that for the minimizer
$\posfair \rightarrow \posunfair \iff \alpha \rightarrow 1$.
With alternative (\emph{i.e.} non-strictly proper) loss, we might have only ``$\Rightarrow$''.
\paragraph{Do we have guarantees on some proximity of \( \posfair \) with respect to \( \posunfair \)? Yes, with light assumptions.}
We examine the \des{proximity} of black-box and post-processed posteriors with the KL divergence \citep{anMO}:
\begin{eqnarray}
\kl(\posunfair, \posfair; \meas{M}) & \defeq & \expect_{(\X, \Y)\sim \measdunfair} \left[\log \left({\dmeasdunfair((\X,\Y))} \big/ {\dmeasdfair((\X,\Y))}\right) \right],\label{defKL}
\end{eqnarray}
where $\measdunfair, \measdfair$ are the product measures defined from $\meas{M}$ and their respective posteriors. To bound the proximity \eqref{defKL}, we present setting \textbf{(S1)}.
\begin{enumerate}
\item [\textbf{(S1)}] Assumption \ref{assumptionBUNF} holds for some $0<B\leq 3$ and function $\alpha$ satisfies $|\alpha({\bm{e}}{x})-1| \leq 1/B$ (a.s.).
\end{enumerate}
This setting lead to the following \emph{data independent} proximity bound.
\begin{theorem}\label{cor-distortion-to-blackbox}
For any \( \meas{M} \),
\textup{\textbf{(S1)}} implies
\(
\kl(\posunfair, \posfair; \meas{M}) \leq \pi^2/(6\cdot (2 + \exp (B) + \exp (-B))) \).
\end{theorem}
As an example, fix $B=3$ for \textbf{(S1)}. In this case, we want $\alpha(.) \in [2/3, 4/3]$ (a.s.) which is a reasonable sized interval centered at 1. The clamped black-box posterior's interval is approximately $[0.04, 0.96]$, which is quite flexible and the distortion is upperbounded as $\kl(\posunfair, \posfair; \meas{M}) \leq 7.5E-2$.
\paragraph{Is the composition of transformations meaningful? Yes.}
The analytical form in \eqref{ourposfair} brings the following easy-to-check \textbf{(composability)} property.
\begin{lemma}\label{lemcomposability}
The composition of any two wrapping transformations $\posunfair \raisebox{-3pt}{ $\stackrel{\alpha}{\mapsto}$ } \posfair \raisebox{-3pt}{ $\stackrel{\alpha'}{\mapsto}$ } \posfair'$ following \eqref{ourposfair} is equivalent to the single transformation $\posunfair \raisebox{-3pt}{ $\stackrel{\alpha \cdot \alpha'}{\mapsto}$ } \posfair'$.
\end{lemma}
This gives an \emph{invertibility} condition -- wrapping $\posfair$ with $\alpha' = 1/\alpha$ recovers the original black-box \( \posunfair \).
\paragraph{Given some capacity parameter for \( \posunfair \), can we easily compute that of \( \posfair \)? Yes, \emph{e.g.}, for decision trees.}
Such a question is particularly relevant for generalization. As we are using the log-loss \eqref{lossgen}, a relevant capacity notion to assess the uniform convergence of risk minimization for the whole wrapped model is the Rademacher \textbf{(complexity)} \citep{bmRA}.
We examine the following set of functions:
\begin{eqnarray}
\mathcal{H}_{\mbox{\tiny{f}}} & \defeq & \left\{ \posfair : \posfair({\bm{e}}{x}) \textrm{ given by \eqref{ourposfair} with } \alpha, \posunfair; \forall (\alpha, \posunfair)\right\},\label{defHR}
\end{eqnarray}
where we assume known the set of functions from which $\posunfair$ was trained.
We now assume we have a $m$-training sample $\mathcal{S} \defeq \{({\bm{e}}{x}_i, y_i)\sim \meas{D}\}_{i=1}^{m} $. The empirical Rademacher complexity of a set of functions $\mathcal{H}$ from $\mathcal{X}$ to $\mathbb{R}$, $\radcomp_{\mathcal{S}}(\mathcal{H}) \defeq \expect_{{\bm{e}}{\sigma}} \sup_{h \in \mathcal{H}} \expect_{i} [\sigma_i h({\bm{e}}{x}_i)]$
(sampling uniform with $\sigma_i \in \{-1,1\}$), is a capacity parameter that yields efficient control of uniform convergence when the loss used is Lipschitz \citep[Theorem 7]{bmRA}, which is the case of the log-loss.
To see how the $\alpha$-tree affects the Rademacher complexity of classification using $\posfair$ instead of $\posunfair$, suppose real-valued prediction based on $\posunfair$ is achieved via logit mapping, $\logit \circ \posunfair$ \eqref{defHR}.
Such mappings are common for decision trees \citep{ssIB}.
\begin{lemma} \label{lem:complexity}
Suppose $\{\posunfair\}$ is the set of decision trees of depth $\leq d$ and denote $\radcomp_{\mathcal{S}}(\textsc{dt}(d))$ the empirical Rademacher complexity of decision trees of depth $\leq d$ \citep{bmRA} and $d'$ the maximum depth allowed for $\alpha$-trees. Then we have for $\mathcal{H}_{\mbox{\tiny{\textup{f}}}}$ in \eqref{defHR}: $\radcomp_{\mathcal{S}}(\mathcal{H}_{\mbox{\tiny{\textup{f}}}}) \leq \radcomp_{\mathcal{S}}(\textsc{dt}(d+d'))$.
\end{lemma}%
The proof is straightforward once we remark that elements in $\mathcal{H}_{\mbox{\tiny{f}}}$ can be represented as decision trees, where we plug at each leaf of $\posunfair$ a copy of the $\alpha$-tree $\alphatree$.
\paragraph{Does \cref{alg:topdown} have any convergence properties? Yes, it is a boosting algorithm.}
Following a similar blueprint to classical decision tree induction, it comes with no surprise that \topdown~can achieve boosting compliant convergence. To show that \topdown~is a boosting algorithm, we need a \emph{Weak Hypothesis Assumption (WHA)}, which postulates informally that each chosen split brings a small edge over random splits for a tailored distribution.
\begin{definition}\label{balancedDist}
Let $\leaf \in \leafset(\alphatree)$ and ${\measdtarget}_\leaf$ be the product measure on $\mathcal{X} \times\mathcal{Y}$ conditioned on $\leaf$. The \textbf{balanced} product measure ${\rebmeasdtarget}_\leaf$ at leaf $\leaf$ is defined as ($z \defeq ({\bm{e}}{x},y)$ for short):
\begin{eqnarray}
{\rebmeasdtarget}_\leaf (z) & \defeq & \frac{1 - \edge(\meas{M}_\leaf, \postarget) \cdot y \normalizedlogit(\posunfair({\bm{e}}{x}))}{1 - \edge(\meas{M}_\leaf, \postarget)^2} \cdot {\measdtarget}_\leaf (z).\label{defDMEASTARGET2}
\end{eqnarray}
\end{definition}
We check that $\int_{\leaf} \mathrm{d} {\rebmeasdtarget}_\leaf = 1$ because of Def.~\ref{def:edge}. Our balanced distribution is named after \citet{kmOT}'s: ours indeed generalizes theirs.
The key difference comes from the change in setting, where we consider the alignment purity of a leaf and not its label purity. The ``\textit{fairness-free case}'' where $\normalizedlogit(.)$ is replaced by constant $1$ yields the original balanced distribution \citep{kmOT}. We now state our WHA.
\begin{assumption}\label{defWHA}
\begin{flushleft}
Let $h : \mathcal{X} \rightarrow \mathcal{Y}$ be the function splitting leaf $\leaf$. For $\upgamma > 0$, then \textbf{$h$ $\upgamma$-witnesses} the Weak Hypothesis Assumption (\textbf{WHA}) at $\leaf$ iff
%
\begin{minipage}{.4\linewidth}%
\begin{equation*}
\mbox{\bf (i)} \; \bigg\vert\mathop{\expect}_{(\X, \Y) \sim {\rebmeasdtarget}_\leaf} \left[\Y \normalizedlogit(\posunfair(\X)) \cdot h(\X)\right]\bigg\vert \geq \upgamma;
\end{equation*}
\end{minipage}%
\begin{minipage}{.6\linewidth}
\begin{equation*}
\quad \mbox{\bf (ii)} \; \edge(\meas{M}_\leaf, \postarget) \cdot \mathop{\expect}_{(\X, \Y) \sim {\measdtarget}_{\leaf}} \left[(1-\normalizedlogit^2(\posunfair(\X)))\cdot h(\X)\right] \leq 0.
\end{equation*}
\end{minipage}
\end{flushleft}%
\end{assumption}
Intuitively, (i) gives a condition on the split \( h \)'s correlation with unfair posterior \( \posunfair \) agreement with labels and (ii) does the same for the confidence of \( \posunfair \) predictions. Similarly to the balanced distribution, in the fairness-free case where we only care about the label purity of splits, the WHA simplified to that of \citet{kmOT}'s -- \emph{i.e.}, only (i) matters.
A further discussion on our balanced distribution and WHA is in the \supplement, Section \ref{sec-wha-disc}. We now state \topdown's boosting compliant \des{convergence}.
\begin{theorem}\label{thm-boosting}
Suppose (a) Assumption \ref{assumptionBUNF} holds, (b) we pick the heaviest leaf to split at each iteration in Step 2.1 of \topdown and (c) $\exists \upgamma > 0$ such that each split $h^\star$ (Step 2.2) in $\alphatree$ $\upgamma$-witnesses the WHA. Then there exists a constant $c>0$ such that $\forall \varepsilon > 0$, if the number of leaves of $\alphatree$ satisfies $|\leafset(\alphatree)| \geq \left(1/\varepsilon\right)^{c \log \left(\frac{1}{\varepsilon}\right)/\upgamma^2}$,
then $\posfair$ crafted from \eqref{ourposfair} using \topdown's $\alphatree$ achieves $\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \leq \varepsilon$.
\end{theorem}
\begin{figure}[t]
\centering
\includegraphics[trim=0bp 400bp 560bp 20bp,clip,height=0.28\textwidth]{Figs/FigAggressiveConservative-2}%
\hspace{2cm}
\includegraphics[trim=5bp 455bp 690bp 15bp,clip,height=0.27\textwidth]{Figs/FigEOO-1}%
\caption{\textit{Left}: Difference between the per-leaf bounds on risk $\totrisk(\posunfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$ using \eqref{defBINENT} and \cref{thm:entropyupper} (conservative scoring) and \eqref{totriskB2} (audacious scoring). Details in the proof of Lemma \ref{lem-entropies}. \textit{Right}: A representation of the $(p,\delta)$-pushup of $\bayespo$, where $\eta(p) \defeq \inf \bayespo(\mathcal{X}_p) < 1/2$ (Def. \ref{def-pushup}). All posteriors in $[\eta(p), 1/2 + \delta]$ are mapped to $1/2 + \delta$; others do not change. New posterior $\bayespo_{p,\delta}$ eventually reduces the accuracy of classification for observations whose posterior lands in the thick red interval ($x$-axis).}
\label{fig:eta-diff-eoo}
\fnegativespace
\end{figure}%
\paragraph{Are there alternative ways of growing \( \alpha\)-trees? Yes.}
Let us call \textit{conservative} the scoring scheme in \eqref{alphatreeoutputCONS}. There is an alternative scoring scheme, which can lead to substantially larger corrections in absolute values, hence the naming, and yields better entropic bounds for the $\alpha$-tree.
\begin{definition}\label{audaciousScore}
For any mixture $\meas{M}_{\mbox{\tiny{\textup{t}}}}$ and posteriors $\posunfair, \postarget$, let \( \edge^+(\meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \) and \( \edge^-(\meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \) be defined by
%
\begin{eqnarray}
\edge^{\pm}(\meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \defeq & {\expect}_{(\X, \Y) \sim \measdtarget} \left[\max\{0, \pm \Y \normalizedlogit(\posunfair(\X))\}\right]. \label{defEDGEGEPNMF}
\end{eqnarray}
The \textbf{audacious} scoring schemes at the leaves of the $\alpha$-tree replaces \eqref{alphatreeoutputCONS} in Step 3 by:
\begin{eqnarray*}
\alphatree(\leaf) & \defeq & \normalizedlogit\left(\frac{\edge^+(\meas{M}_{\leaf}, \postarget)}{\edge^+(\meas{M}_{\leaf}, \postarget) + \edge^-(\meas{M}_{\leaf}, \postarget)}\right), \quad \forall \leaf \in \leafset(\alphatree).
\end{eqnarray*}
\end{definition}
\begin{theorem}\label{thm-alternative-alpha-in-alpha-tree}
Suppose Assumption \ref{assumptionBUNF} holds and let $\entropy_2(q) \defeq \entropy(q) / \log 2$ ($\in [0,1]$), $H$ being defined in Definition \ref{defENTROPYA}. For any leaf $\leaf \in \leafset(\alphatree)$, denote for short:
%
\begin{eqnarray*}
\entropy_2(\leaf; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \defeq & \log(2) \cdot \left( 1 + (\edge^+_{\leaf} + \edge^-_{\leaf}) \cdot \left( \entropy_2 \left( \frac{\edge^+_{\leaf}}{\edge^+_{\leaf}+\edge^-_{\leaf}} \right) - 1 \right)\right),
\end{eqnarray*}%
%
where let $\edge^b_{\leaf} \defeq \edge^b(\meas{M}_{\leaf}, \postarget), \forall b \in \{+,-\}$. Using audacious scoring, we get instead of
\cref{thm:entropyupper}:
\begin{eqnarray}
\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \leq & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[\entropy_2(\leaf; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)\right]. \label{totriskB2}
\end{eqnarray}
\end{theorem}
While upperbounds in \cref{thm:entropyupper}
and \cref{thm-alternative-alpha-in-alpha-tree} may look incomparable, it takes a simple argument to show that \eqref{totriskB2} is never worse and can be much tighter.
\begin{lemma}\label{lem-entropies}
$\forall \alpha$-tree $\alphatree$, $\expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[\entropy_2(\leaf; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)\right] \leq \entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$.
\end{lemma}
It thus comes at no surprise that using the audacious scoring also results in a boosting result for \topdown~guaranteeing the same rates as in Theorem \ref{thm-boosting}. It also takes a simple picture to show that the per-leaf slack in Lemma \ref{lem-entropies} can be substantial, a slack which can be represented using a simple picture, see Figure \ref{fig:eta-diff-eoo} (left), following from the use of Jensen's inequality in the Lemma's proof.
\paragraph{As audacious scoring is better boosting-wise, is conservative scoring useful? Yes.}
If we only cared about accuracy, we would barely have any reason to use the conservative correction. Even thinking about generalization, the Rademacher complexity of decision trees is a function of their depth so faster the convergence, the better \citep[Section 4.1]{bmRA}. Adding fairness substantially changes the picture: some constraints, like equality of opportunity (Section \ref{sec-fairness}) can antagonize accuracy to some extent. In such a case, using the conservative correction can keep posteriors $\posunfair$ and $\postarget$ close enough (\cref{cor-distortion-to-blackbox}%
) so that fairness can be achieved without substantial sacrifice on accuracy.
\section*{Acknowledgments}
YM received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (grant agreement No. 882396), the Israel Science Foundation(grant number 993/17), Tel Aviv University Center for AI and Data Science (TAD), and the Yandex Initiative for Machine Learning at Tel Aviv University. AS and LX thank members of the ANU Humanising Machine Intelligence program for discussions on fairness and ethical concerns in AI, and the
NeCTAR Research Cloud for providing computational resources, an Australian research platform
supported by the National Collaborative Research Infrastructure Strategy.
\section{Limitations and Conclusion}
Given the context of fairness, it is important to highlight possible limitations of our approach and the potential social harm from such misuse. We highlight two of these for our \topdown approach. Firstly, our approach has shown to have failure cases in \emph{small data cases}. This can be seen in the experiments on German Credit dataset (\supplement, Section \ref{app-exp-rest}-\ref{app-exp-proxy}). This, we believe, has a formal basis in our approach. For instances, \EOO requires accurate data posterior estimation which may be difficult in small data regimes. Secondly, \topdown is of course not unilaterally better than all other post-processing fairness approaches -- there is \emph{No Free Lunch}. As such, we do not claim that our instantiation of \topdown is optimal in \( \cvar{} \), \EOO, or \SP. However, considering that such different fairness models instantiated in the \textit{same} algorithm can lead to competitive results with the respective state of the art, the avenue for improved instantiations or accurate extensions to new fairness constraints appears promising. We leave these for future work.
\section{Neural Network Experiments} \label{app-exp-nn}
In this \supplement section, we repeat all experiments evaluating different fairness models and proxy sensitive attributes using the neural network (NN) black-box.
\cref{fig:mlp_german_fairness,fig:mlp_bank_fairness,fig:mlp_acs_fairness} presents neural network equivalent plots to that of \cref{fig:rf_german_fairness,fig:rf_bank_fairness,fig:rf_acs_fairness}.
In particular:
\begin{itemize}
\item \cref{fig:mlp_german_fairness} presents the evaluation using a NN black-box with \( B = 1 \) clipping on the German Credit dataset.
\item \cref{fig:mlp_bank_fairness} presents the evaluation using a NN black-box with \( B = 1 \) clipping on the Bank dataset.
\item \cref{fig:mlp_acs_fairness} presents the evaluation using a NN black-box with \( B = 1 \) clipping on the ACS dataset.
\end{itemize}
When comparing the NN experiments to the experiments corresponding to that of the random forest (RF) black-box experiments, only minor deviation can be seen with most trends staying the same. One consistent deviation is that the \( \cvar{} \) criterion and accuracy measures (\meandiff and \auc) are frequently smaller at the initial and final point of boosting. This comes from the strong representation power of the NN black-box being translated from the initial black-box to the final wrapper classifier. In this regard, switching to a NN did not help the optimization of \( \cvar{} \) for the German Credit dataset, see \cref{fig:mlp_german_fairness}.
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/german_credit_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/german_credit_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{\topdown~optimized for different fairness models evaluated on German Credit with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:mlp_german_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/bank_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/bank_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{\topdown~optimized for different fairness models evaluated on Bank with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:mlp_bank_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/acs_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/acs_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{\topdown~optimized for different fairness models evaluated on Bank with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:mlp_acs_fairness}
\end{sidewaysfigure}
\clearpage
\section{Neutral Loss} \label{sec-neutral-loss}
The condition of ``neutrality'' as presented in the main-text is better known as the condition of a loss being \emph{fair} \cite{rwID}. Notably, this fair loss condition is distinct from that of algorithmic fairness and the corresponding fairness criteria over, \emph{i.e.}, predictors. Reiterating the condition for a fair loss: \( \partialloss{1}(1) = \partialloss{-1}(0) = 0 \) and \( 0 = \min \partialloss{1} = \min \partialloss{-1} \). As such, we deliberately break convention and rename this condition as the ``neutrality'' of a loss: the condition can be summarized as the condition that correct prediction does give a loss (positive value) nor a gain (negative value), and thus is neutral in its response (zero value).
\section{Proof of Theorem \ref{thm-alternative-alpha-in-alpha-tree}}\label{proof-thm-alternative-alpha-in-alpha-tree}
The proof is obtained via a generalisation of Lemma \ref{lemGB1-NSS}.
\begin{lemma}\label{lemGB2}
Fix any $B>0$. For any $\alpha \in \mathbb{R}$, any $\theta, z \in [-B,B]$, if we let
\begin{eqnarray}
\vartheta(z) & \defeq & (z-\theta) \cdot \left\{ \begin{array}{ccl}
\frac{1}{B+\theta} & \mbox{ if } & z < \theta,\\
0 & \mbox{ if } & z = \theta,\\
\frac{1}{B-\theta}& \mbox{ if } & z > \theta.
\end{array}\right.,
\end{eqnarray}
then we have
\begin{eqnarray*}
\log(1+\exp(\alpha z)) & \leq & \log\left(\frac{1+\exp(B \alpha)}{1+\exp(\theta \alpha)}\right) \cdot |\vartheta(z)| - B\alpha \max\{0, -\vartheta(z)\} + \log(1+\exp(\theta \alpha)).
\end{eqnarray*}
\end{lemma}
\textbf{Remark}: Lemma \ref{lemGB1-NSS} is obtained for the choices $\theta = \pm B$.\\
\begin{myproof}
We fix any $\theta' \in [-1,1]$ and let
\begin{eqnarray}
\mathrm{l} & \defeq & (-1, \log(1+\exp(-\alpha))),\\
\mathrm{c} & \defeq & (\theta', \log(1+\exp(\alpha\theta'))),\\
\mathrm{r} & \defeq & (1, \log(1+\exp(\alpha))).
\end{eqnarray}
The equation of the line passing through $\mathrm{l}, \mathrm{c}$ is
\begin{align}
&f_{\mathrm{l}}(z) \nonumber \\
& = \frac{\log\left(\frac{1+\exp(\theta' \alpha)}{1+\exp(-\alpha)}\right)}{1+\theta'} \cdot z + \frac{\log\left(\frac{1+\exp(\theta' \alpha)}{1+\exp(-\alpha)}\right)}{1+\theta'} + \log(1+\exp(-\alpha))\\
& = -\frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1+\theta'} \cdot z + \frac{\alpha z}{1+\theta'} -\frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1+\theta'} + \frac{\alpha}{1+\theta'}+ \log(1+\exp(-\alpha))\\
& = \frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1+\theta'} \cdot (\theta' - z) + \frac{\alpha (z-\theta')}{1+\theta'} - \log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right) + \log(1+\exp(\alpha))\\
& = \frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1+\theta'} \cdot (\theta' - z) + \frac{\alpha (z-\theta')}{1+\theta'} + \log(1+\exp(\theta' \alpha))
\end{align}
and the equation of the line passing through $\mathrm{c},\mathrm{r}$ is
\begin{eqnarray}
f_{\mathrm{r}}(z) & = & \frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1-\theta'} \cdot z - \frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1-\theta'} + \log(1+\exp(\alpha))\\
& = & \frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1-\theta'} \cdot (z-\theta') - \log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right) + \log(1+\exp(\alpha))\\
& = & \frac{\log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right)}{1-\theta'} \cdot (z-\theta') + \log(1+\exp(\theta' \alpha)).
\end{eqnarray}
For any $z \in [-1,1]$, define $\vartheta'(z) \in [-1,1]$ to be:
\begin{eqnarray}
\vartheta'(z) & \defeq & (z-\theta') \cdot \left\{ \begin{array}{ccl}
\frac{1}{1+\theta'} & \mbox{ if } & z < \theta',\\
0 & \mbox{ if } & z = \theta',\\
\frac{1}{1-\theta'}& \mbox{ if } & z > \theta'.
\end{array}\right.
\end{eqnarray}
Function $z\mapsto \log(1+\exp(\alpha z))$ being convex, we thus get the secant upperbound:
\begin{align}
&\log(1+\exp(\alpha z)) \nonumber \\
& \leq \log\left(\frac{1+\exp(\alpha)}{1+\exp(\theta' \alpha)}\right) \cdot |\vartheta'(z)| + \alpha \min\{0, \vartheta'(z)\} + \log(1+\exp(\theta' \alpha)),
\end{align}
and this holds for $z \in [-1,1]$. If instead $z\in [-B,B]$, then letting $\theta \defeq B \theta' \in [-B, B]$, we note:
\begin{align}
&\log(1+\exp(\alpha z)) \nonumber \\
& = \log(1+\exp(\alpha B \cdot (z/B))) \nonumber\\
& \leq \log\left(\frac{1+\exp(\alpha B)}{1+\exp(\theta' \alpha B)}\right) \cdot |\vartheta'(z/B)| + \alpha B \min\{0, \vartheta'(z/B)\} + \log(1+\exp(\theta' \alpha B)),
\end{align}
where this time,
\begin{eqnarray}
\vartheta'\left(\frac{z}{B}\right) & \defeq & \left(\frac{z}{B}-\theta'\right) \cdot \left\{ \begin{array}{ccl}
\frac{1}{1+\theta'} & \mbox{ if } & z < B\theta',\\
0 & \mbox{ if } & z = B\theta',\\
\frac{1}{1-\theta'}& \mbox{ if } & z > B\theta'.
\end{array}\right.\nonumber\\
& = & (z-\theta) \cdot \left\{ \begin{array}{ccl}
\frac{1}{B+\theta} & \mbox{ if } & z < \theta,\\
0 & \mbox{ if } & z = \theta,\\
\frac{1}{B-\theta}& \mbox{ if } & z > \theta.
\end{array}\right. \defeq \vartheta(z).
\end{eqnarray}
We thus get
\begin{align}
&\log(1+\exp(\alpha z)) \nonumber \\
& \leq \log\left(\frac{1+\exp(B \alpha)}{1+\exp(\theta \alpha)}\right) \cdot |\vartheta(z)| + B\alpha \min\{0, \vartheta(z)\} + \log(1+\exp(\theta \alpha)),
\end{align}
and since $\min\{0,z\} = -\max\{0,-z\}$, we get the statement of the Lemma.
\end{myproof}\\
We use Lemma \ref{lemGB2} with $\theta = 0$, which yields $\vartheta(z) = z/B$; using notations from the proof of Theorem \ref{thm-alpha-in-alpha-tree}, we thus get (using the same notations as in the proof of Theorem \ref{thm-boosting}),
\begin{align}
&- \log \posfair \nonumber \\
& = \log\left(1 + \exp\left(\alphatree \cdot -\logit(\posunfair)\right)\right)\nonumber\\
& \leq \frac{1}{B}\cdot \log\left(\frac{1+\exp(B\alphatree)}{2}\right) \cdot |\logit(\posunfair(\X))| + \alphatree \min\{0, -\logit(\posunfair(\X))\}+ \log(2)\nonumber\\
& = \frac{1}{B}\cdot \log\left(\frac{1+\exp(B\alphatree)}{2}\right) \cdot |\logit(\posunfair(\X))| - \alphatree \max\{0, \logit(\posunfair(\X))\}+ \log(2)\label{blogb1}\\
\nonumber \\
&- \log (1-\posfair) \nonumber \\
& = \log \left(1 + \exp\left(\alphatree \cdot \logit(\posunfair(\X))\right)\right)\nonumber\\
& \leq \frac{1}{B}\cdot \log\left(\frac{1+\exp(B\alphatree)}{2}\right) \cdot |\logit(\posunfair(\X))| + \alphatree \min\{0, \logit(\posunfair(\X))\} + \log(2)\nonumber\\
& = \frac{1}{B}\cdot \log\left(\frac{1+\exp(B\alphatree)}{2}\right) \cdot |\logit(\posunfair(\X))| - \alphatree \max\{0, -\logit(\posunfair(\X))\} + \log(2).\label{blogb2}
\end{align}
We get that the inequality in \eqref{bENT1} now reads (for \textit{any} values $\{\alphatree(\leaf), \leaf \in \leafset(\alphatree)\}$) $\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) = \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ J(\leaf)\right]$ with $J(\leaf)$ satisfying:
\begin{align}
&J(\leaf) \nonumber \\
& \leq \expect_{\X \sim \meas{M}_{\leaf}} \left[\begin{array}{l}
\postarget(\X)\cdot\left(\frac{1}{B}\cdot \log\left(\frac{1+\exp(B\alphatree(\leaf))}{2}\right) \cdot |\logit(\posunfair(\X))| - \alphatree(\leaf) \max\{0, \logit(\posunfair(\X))\}+ \log(2)\right)\\
+(1-\postarget(\X))\cdot \left(\frac{1}{B}\cdot \log\left(\frac{1+\exp(B\alphatree(\leaf))}{2}\right) \cdot |\logit(\posunfair(\X))| - \alphatree(\leaf) \max\{0, -\logit(\posunfair(\X))\} + \log(2)\right)
\end{array}\right]\nonumber\\
& = \log(2) - B\alphatree(\leaf)\cdot \edge^+(\meas{M}_{\leaf}, \postarget) \nonumber \\ & \quad \quad +\log\left(\frac{1+\exp(B\alphatree(\leaf))}{2}\right) \cdot (\edge^+(\meas{M}_{\leaf}, \postarget) + \edge^-(\meas{M}_{\leaf}, \postarget)),
\end{align}
and the bound takes its minimum on $\alphatree(\leaf)$ for
\begin{eqnarray}
\alphatree(\leaf) & = & \frac{1}{B} \cdot \log\left(\frac{\edge^+(\meas{M}_{\leaf}, \postarget)}{\edge^-(\meas{M}_{\leaf}, \postarget)}\right) = \normalizedlogit\left(\frac{\edge^+(\meas{M}_{\leaf}, \postarget)}{\edge^+(\meas{M}_{\leaf}, \postarget) + \edge^-(\meas{M}_{\leaf}, \postarget)}\right),
\end{eqnarray}
yielding (using notations from Theorem \ref{thm-alternative-alpha-in-alpha-tree}),
\begin{eqnarray}
J(\leaf) & \leq & \log(2)\cdot\left(1- \edge^-_{\leaf} -\edge^+_{\leaf}\right) - \edge^+_{\leaf}\cdot \log\left(\frac{\edge^+_{\leaf}}{\edge^-_{\leaf}}\right) + \log\left(\frac{\edge^-_{\leaf} + \edge^+_{\leaf}}{\edge^-_{\leaf}}\right) \cdot (\edge^-_{\leaf} +\edge^+_{\leaf})\nonumber\\
& & = \log(2) \cdot \left(1+(\edge^-_{\leaf} + \edge^+_{\leaf})\cdot \left(H_2 \left(\frac{\edge^+_{\leaf}}{\edge^+_{\leaf} +\edge^-_{\leaf}}\right) - 1\right)\right),
\end{eqnarray}
and brings the statement of Theorem \ref{thm-alternative-alpha-in-alpha-tree} after plugging the bound in the expectation.
\section{Handling \textbf{Statistical parity} }\label{sec-stat-par}
\noindent\textbf{Statistical parity} (SP) is a group fairness notion \cite{dhprzFT}, implemented recently in a context similar to ours \cite{alAN} as the constraint that per-group expected treatments must not be too far from each other. We say that $\posfair$ achieves $\varepsilon$-statistical parity (across all groups induced by sensitive attribute $S$) iff
\begin{eqnarray}
\max_{s\in S} \expect_{\X \sim \meas{M}_s} \left[\posfair(\X)\right] - \min_{s\in S} \expect_{\X \sim \meas{M}_s} \left[\posfair(\X)\right] & \leq & \varepsilon. \label{eqSTATPAR}
\end{eqnarray}
Denote $s^\circ \defeq \arg\min_{s\in S} \expect_{\X \sim \meas{M}_s} \left[\posfair(\X)\right], s^* \defeq \arg\max_{s\in S} \expect_{\X \sim \meas{M}_s} \left[\posfair(\X)\right]$. Since the risk we minimize in \eqref{lossgen} involves a proper loss, the most straightforward use of \topdown~is to train the sub-$\alpha$-tree for one of these two groups, giving as target posterior the \textit{expected} posterior of the other group, \textit{i.e.} we use $\postarget({\bm{e}}{x}) = \expect_{\X \sim \meas{M}_{s^*}} \left[\posunfair(\X)\right] \defeq \overline{\posunfair}_{s^*}$ if we grow the $\alpha$-tree of $s^\circ$ and thus iterate
\begin{tcolorbox}[colframe=blue,boxrule=0.5pt,arc=4pt,left=6pt,right=6pt,top=6pt,bottom=6pt,boxsep=0pt]
\begin{center}
\topdown~with $\meas{M}_{\mbox{\tiny{\textup{t}}}} \leftarrow \meas{M}_{s^\circ}$ and $\postarget \leftarrow \overline{\posunfair}_{s^*}$,
\end{center}
\end{tcolorbox}
and we repeat until $s^\circ$ does not achieve anymore the smallest expected posterior. We then update the group and repeat the procedure until a given slack $\varepsilon$ is achieved between the extremes in \eqref{eqSTATPAR}. More sophisticated / gentle approaches are possible, including using the links between statistical parity and optimal transport (OT, \citet[Section 3.2]{dhprzFT}), suggesting to use as target posterior the expected posterior obtained from an OT plan between groups $s^\circ$ and $s^*$.
\section{Proof of Theorem \ref{cor-distortion-to-blackbox} and \ref{cor-distortion-to-blackbox-b}}\label{proof-cor-distortion-to-blackbox}
We start by \textbf{(S1)}. We study function
\begin{eqnarray}
f_k(u) & \defeq & u(1-u) \left|\log\left(\frac{1-u}{u}\right)\right|^k, \forall u \in \left[\frac{1}{1+\exp (B)}, \frac{1}{1+\exp (-B)}\right] .\label{defFK}
\end{eqnarray}
$f_k$ being symmetric around $u=1/2$ and zeroing in $1/2$, we consider wlog $u< 1/2$ to find its maximum, so we can drop the absolute value.
We have
\begin{eqnarray}
f_k'(u) & = & \log^{k-1}\left(\frac{1-u}{u}\right)\cdot\left((1-2u)\cdot \log\left(\frac{1-u}{u}\right)-k\right).
\end{eqnarray}
Function $u \mapsto (1-2u)\cdot \log\left(\frac{1-u}{u}\right)$ is strictly decreasing on $(0,1/2)$ and has limit $+\infty$ on $0^+$, so the unique maximum of $f$ on $[0,1/2)$ (we close by continuity the interval in $0$ since $\lim_{0^+} f = 0$) is attained at the only solution $u_k$ of
\begin{eqnarray}
(1-2u_k)\cdot \log\left(\frac{1-u_k}{u_k}\right) & = & k,\label{eqUK}
\end{eqnarray}
and such a solution always exist for any $k \ll \infty$. It also follows $u_{k+1} < u_k$, so if we denote as $k^*$ the smallest $k$ such that
\begin{eqnarray}
u_{k^*} & \leq & \frac{1}{1+\exp (B)},\label{eqUKStar}
\end{eqnarray}
then we will have the upperbound:
\begin{eqnarray}
f_k(u) & \leq & \frac{1}{1+\exp (B)} \cdot \frac{1}{1+\exp (-B)} \cdot B^k\nonumber\\
& & = \frac{B^k}{2 + \exp (B) + \exp (-B)}, \forall k \geq k^*.
\end{eqnarray}
We can also compute $k^*$ exactly as it boils down to taking the integer part of the solution of \eqref{eqUK} where $u_k$ is picked as in \eqref{eqUKStar}:
\begin{eqnarray}
k^* & = & \left\lfloor \frac{\exp(B) - 1}{\exp(B) + 1} \cdot B \right\rfloor,
\end{eqnarray}
to get $k^* = 2$, it is sufficient that $B\leq 3$, which thus gives:
\begin{eqnarray}
\kl(\posunfair, \posfair; \meas{M}) & \leq & \sum_{k= 2}^K \frac{\expect_{\X \sim \meas{M}} \left[(B\cdot |{\alpha(\X)}-1|)^k\right]}{(2 + \exp (B) + \exp (-B))k(k-1)} + G,
\end{eqnarray}
and if $|{\alpha({\bm{e}}{x})}-1| \leq 1/B = 1/3, \forall {\bm{e}}{x} \in \mathcal{X}$, then we can include all terms for all $k\geq 2$ in the upperbound, which makes the little-oh remainder vanish and we get:
\begin{eqnarray}
\kl(\posunfair, \posfair; \meas{M}) & \leq & \lim_{K\rightarrow +\infty} \frac{1}{2 + \exp (B) + \exp (-B)} \cdot \sum_{k= 2}^K \frac{1}{k(k-1)} \\
& \leq & \frac{1}{2 + \exp (B) + \exp (-B)} \cdot \sum_{k\geq 1} \frac{1}{k^2} \\
& = & \frac{\pi^2}{6(2 + \exp (B) + \exp (-B))},
\end{eqnarray}
which is \eqref{boundDelta1} and proves the Corollary for setting \textbf{(S1)}. The proof for setting \textbf{(S2)} is direct as in this case we get:
\begin{eqnarray}
\kl(\posunfair, \posfair; \meas{M}) & \leq & \lim_{K\rightarrow +\infty} \expect_{\X \sim \meas{M}} \left[\sum_{k= 2}^K \frac{\posunfair(\X)(1-\posunfair(\X)) f^k (\alpha(\X),\posunfair(\X)) }{k(k-1)}\right]\nonumber\\
& & = \expect_{\X \sim \meas{M}} \left[\sum_{k= 2}^K \frac{\posunfair(\X)(1-\posunfair(\X)) f^k (\alpha(\X),\posunfair(\X)) }{k(k-1)}\right]\nonumber\\
& \leq & \expect_{\X \sim \meas{M}} \left[\sum_{k= 2}^K \frac{\posunfair(\X)(1-\posunfair(\X))}{k(k-1)}\right]\\
& \leq & \frac{1}{4} \cdot \sum_{k= 2}^K \frac{1}{k(k-1)} \\
& \leq & \frac{1}{4} \cdot \sum_{k\geq 1} \frac{1}{k^2} \\
& = & \frac{\pi^2}{24},
\end{eqnarray}
as claimed.\\
\begin{figure}[h]
\centering
\centerline{\includegraphics[trim=20bp 550bp 540bp 10bp,clip,width=0.7\textwidth]{Figs/FigAuthorized-range}}
\caption{Admissible couples of values $(g, \posunfair)$ (in blue) complying with setting \textbf{(S2)}. For example, any couple $(g,0.1)$ with $g \in [0.54, 1.46]$ is admissible.}
\label{fig:authorized-range}
\end{figure}
Figure \ref{fig:authorized-range} provides an idea of the set of \textit{admissible} couples (correction, black-box posterior) that comply with \textbf{(S2)}, from which we see that the range of admissible corrections is quite flexible, even when $\posunfair$ comes quite close to $\{0,1\}$.
\section{Proofs}
\subsection*{Proof of \cref{thm:cvar_to_l1}}
To prove the theorem, we use the ``robust representation'' given by \citet{ru13}.
\begin{theorem}[Robust Representation~\citep{ru13}]
\begin{equation}\label{eq:robust_representation}
\cvar{\beta}(Z) = \sup_{Q \in \mathcal{Q}} \expect[Z Q],
\end{equation}
where \( \mathcal{Q} = \{ Q \mid 0 \leq Q \leq \nicefrac{1}{1 - \beta}, \expect{Q} = 1 \} \).
\end{theorem}
\begin{proof}
\end{proof}
\section{\textbf{Weak Hypothesis Assumption} Discussion}\label{sec-wha-disc}
In this Section, we discuss the balanced distribution introduced in Definition \ref{balancedDist} and the Weak Hypothesis Assumption (WHA) introduced in Assumption \ref{defWHA}. In particular, we note that these definitions are a generalization of those introduced in ``classical'' decision tree boosting \cite{kmOT}. The primary difference in our case and the ``classical'' case, is that we must consider the posterior values of a base black-box \( \posunfair \), whereas the usual top-down tree induction only cares about minimizing its loss function to produce an accurate \( \postarget \). In-fact, we can even interpret our variants as a case where we replace the labels in a classification task \( \Y \) with ``labels'' (in \( [0, 1]\)) of how well another classifier does in classification \( \Y \normalizedlogit(\posunfair(\X)) \). We will explore various cases to show that our definitions reduce to the classical case; and further provide settings to intuitive describe our more complicated assumption.
\noindent\textbf{The balanced distribution} in Definition \ref{balancedDist} acts as a re-weighting mechanism for particular samples of the original distribution \( {\measdtarget}_\leaf (z) \). Let us first consider the reduction to the balanced distribution introduced by \citet{kmOT}. This comes from the fairness-free case, where we ignore the confidence of the black-box \( \posunfair \) and set \( \normalizedlogit(\posunfair) \leftarrow 1 \). This provides two simplifications. Firstly, the edge value becomes \( \edge(\meas{M}_\leaf, \postarget) = \expect_{(\X, \Y) \sim {\measdtarget}_\leaf} \left[\Y\right] = 2q_{\leaf} - 1 \), where $q_{\leaf}$ the local proportion of positive examples in $\leaf$. That is, we have \( q_{\leaf} = \pr(\Y = +1 | \X \textrm{ at leaf } \leaf) \). The second simplification is in the numerator of Equation \ref{defDMEASTARGET2}. In summary, we get:
\begin{equation} \label{eq:balanced_kmot}
{\rebmeasdtarget}_\leaf (z) \leftarrow \frac{1 - \edge(\meas{M}_\leaf, \postarget) \cdot y}{1 - \edge(\meas{M}_\leaf, \postarget)^2} \cdot {\measdtarget}_\leaf (z) = \frac{1 - y\cdot(2q_{\leaf} - 1)}{4q_{\leaf}(1 - q_{\leaf})} \cdot {\measdtarget}_\leaf (z) =
{\measdtarget}_\leaf (z) \cdot
\begin{cases}
\frac{1}{2q_{\leaf}} & \textrm{if } y = +1 \\
\frac{1}{2(1 - q_{\leaf})} & \textrm{if } y = -1
\end{cases},
\end{equation}
which is indeed the balanced distribution of \citet{kmOT}. Intuitively, this balanced distribution simply ensures that regardless of \( {\bm{e}}{x} \), it is equally likely for \( y \) be either \( -1 \) or \( +1 \) (hence balanced in its prediction value \( \Y \)). This can be seen by calculating the edge with the new distribution \( \expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf}}[\Y] = 0 \).
To consider our balanced distribution, let us start with \eqref{eq:balanced_kmot}. In particular, we will replace the labels \( \Y \in \{ -1, +1 \} \) by the predictive ability of black-box \( \posunfair \). This predictive ability is summarized by \( \Y \normalizedlogit(\posunfair(\X)) \in [-1, 1] \), where larger is better. In particular, taking from the boosting jargon, the \textit{confidences} ($|\normalizedlogit|$, \citet{ssIBj}) of the black-box \( \posunfair \) are considered. For instance, a highly confident \( \vert \normalizedlogit(\posunfair(\X)) \vert \rightarrow 1 \) and correct \( \mathrm{sign}(\Y) = \mathrm{sign}(\logit(\posunfair(\X))) \) prediction will lead to a value close to \( +1 \) (positive label). If the prediction is confident but incorrect \( \mathrm{sign}(\Y) \neq \mathrm{sign}(\logit(\posunfair(\X))) \), then the value will be close to \( -1 \) (negative label). If the prediction is not confident \( \vert \normalizedlogit(\posunfair(\X)) \vert \rightarrow 0 \), then we get a neutral response \( 0 \).
With this in mind, the balanced distribution in Definition \ref{balancedDist} is ``balanced'' in the prediction score \( \Y \normalizedlogit(\posunfair(\X)) \). Of course, the distribution \( {\rebmeasdtarget}_\leaf (z) \) is with respect to \( \mathcal{X} \times\mathcal{Y} \) and not \( \mathcal{X} \times [-1, 1] \), and thus we are not re-balancing for each possible \( \Y \normalizedlogit(\posunfair(\X)) \in [-1, 1] \). The balance of the distribution takes into account the confidence of the predictions, this can be seen in examination of the balanced edge in this setting:
\begin{equation} \label{eq:bal_new_edge}
\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf}}[\Y \cdot \normalizedlogit(\posunfair(\X)) ] = \frac{1}{1 - \edge(\meas{M}_\leaf, \postarget)^{2}} \cdot \edge(\meas{M}_\leaf, \postarget) \cdot \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf}}[1 - \normalizedlogit^{2}(\posunfair(\X))].
\end{equation}
\eqref{eq:bal_new_edge} is in general non-zero. However, in the case that the \( \posunfair \) is confident such that \( \normalizedlogit(\posunfair(\X)) \in \{ -1 , +1 \} \), then the new edge goes to 0. In other words, when we have a maximally confident black-box, the balanced edge will be 0; otherwise, the edge is scaled with respect to a function of the confidence \( \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf}}[1 - \normalizedlogit^{2}(\posunfair(\X))] \).
\noindent\textbf{Our Weak Learning Assumption (WHA)} in Assumption \ref{defWHA} has a similar analysis as that of the balanced distribution. Condition (i) of the assumption is exactly the same \( \gamma \)-witness condition in \citet{kmOT}. Condition (ii), similar to the analysis of \eqref{eq:bal_new_edge}, vanishes when we have a maximally confident black-box \( \posunfair \). In general, condition (ii) ensures that one side of the split induced by \( h \) are skewed to one side.
Let us exemplify how our WHA works if we have a leaf $\leaf$ where local ``treatments due to the black-box" are bad ($y \normalizedlogit(\posunfair({\bm{e}}{x})) < 0$ often). In such a case, $\edge(\meas{M}_\leaf, \postarget) < 0$ so the balanced distribution (Definition \ref{balancedDist}) reweights higher examples whose treatment is better than average, \textit{i.e.} the local minority. Suppose (i) holds as is without the $|.|$. In such a case, the split ``aligns" the treatment quality with $h$, so $h = +1$ for a substantial part of this minority. (ii) imposes $\expect_{(\X, \Y) \sim {\measdtarget}_{\leaf}} \left[h(\X)\right] \geq \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf}} \left[ \normalizedlogit^2(\posunfair(\X)) \cdot h(\X)\right] $: $h = -1$ for a substantial part of large confidence treatment. The split thus tends to separate mostly large confidence but bad treatments (left) and mostly good treatments (right). Before the split, the value $\alphatree(\leaf)$ would be negative \eqref{alphatreeoutputCONS} and thus reverse the polarity of the black-box, which would be good for badly treated examples but catastrophic for the local minority of adequately treated examples. After the split however, we still have the left ($h = -1$) leaf where this would eventually happen, but the minority at $\leaf$ would have disproportionately ended in the right ($h = +1$) leaf, where it would be likely that $\alphatree(.)$ would this time be \textit{positive} and thus preserve the polarity of the treatment of the black-box.
\noindent\textbf{Why are the assumptions states using the balanced distribution?} A point worth clarifying is the ``why's'' of consider a WHA in the way it is formulated and the focus on the balanced distribution. In typical boosting algorithms, a weak learning hypothesis is needed: the assumption that a classifier / split / weak learner exists for \emph{any} distribution which is better than random. In top-down tree induction, it is in-fact sufficient to only have this assumption hold for one type of distribution --- the balanced distribution \cite{kmOT}. As such the WHA presented, is actually a weaker assumption than the ''distributionally global'' weak learning assumption used in other boosting algorithms.
\section{Proxy Sensitive Attributes}\label{app-exp-proxy}
We examine the use of sensitive attribute proxies to remove sensitive attribute requirements at test time. In particular, we use a decision tree with a maximum depth of \( 8 \) to predict sensitive attributes (from other features) as a proxy to the true sensitive attribute. We present results for both RF and NN black-boxes.
In particular:
\begin{itemize}
\item \cref{fig:german_stree} presents the proxy sensitive attribute evaluation using a RF and NN black-box with \( B = 1 \) clipping on the German Credit dataset.
\item \cref{fig:bank_stree} presents the proxy sensitive attribute evaluation using a RF and NN black-box with \( B = 1 \) clipping on the Bank dataset.
\item \cref{fig:acs_stree} presents the proxy sensitive attribute evaluation using a RF and NN black-box with \( B = 1 \) clipping on the ACS dataset.
\end{itemize}
\cref{fig:acs_stree} (top) presents the RF \topdown proxy sensitive attribute experiments results of the ACS 2015 dataset not present in the main text. We focus on the binary case (left). Unsurprisingly, the proxy increases the variance of \( \cvar{} \) and \auc whilst also being worse than their non-proxy counterparts; but still manages to improve $\cvar{}$ and \auc at the end (with an initial dip quickly erased for the latter criterion). Remark the non-trivial nature of the proxy approach, as growing the $\alpha$-tree is based on groups learned at the decision tree leaves \textit{but} the $\cvar{}$ computation still relies on the \textit{original} sensitive grouping.
\cref{fig:german_stree,fig:bank_stree} (top) presents the RF \topdown proxy sensitive attribute results of the German Credit and Bank datasets. The ACS and Bank experiments presented here are similar to that presented in the main text. For German Credit, similar degradation in \(\cvar{}\) in the non-proxy case can be seen for \topdown results using proxy attributes.
When comparing to the MLP variants (\cref{fig:german_stree,fig:bank_stree,fig:acs_stree} bottom), results are quite similar with slight increases in \( \cvar{} \) from the change in black-box. One notable difference can be seen in \cref{fig:acs_stree}. In particular, the proxy and regular curves do not ``cross". This indicates that (given that the sensitive attribute proxy used is the same as RF) the black-box being post-processed is an important consideration in the use of proxies. In particular, as RF has a higher / worse initial \( \cvar{} \), which is highly tied to the loss / cross entropy of the black-box, the robustness of the black-box needs to be considered.
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/rf/german_credit_stree.pdf}
&
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/rf/german_credit_multi_stree.pdf}
\\
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/mlp/german_credit_stree.pdf}
&
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/mlp/german_credit_multi_stree.pdf}
\end{tabular}
\caption{RF (top) and MLP (bottom) evaluation of replacing sensitive attributes with a proxy decision tree on the German Credit datasets.}
\label{fig:german_stree}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/rf/bank_stree.pdf}
&
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/rf/bank_multi_stree.pdf}
\\
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/mlp/bank_stree.pdf}
&
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/mlp/bank_multi_stree.pdf}
\end{tabular}
\caption{RF (top) and MLP (bottom) evaluation of replacing sensitive attributes with a proxy decision tree on the Bank datasets.}
\label{fig:bank_stree}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/rf/acs_stree.pdf}
&
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/rf/acs_multi_stree.pdf}
\\
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/mlp/acs_stree.pdf}
&
\includegraphics[width=0.45\columnwidth,trim={0, 5pt, 0, 0},clip]{Figs/experiments_plots/mlp/acs_multi_stree.pdf}
\end{tabular}
\caption{RF (top) and MLP (bottom) evaluation of replacing sensitive attributes with a proxy decision tree on the ACS 2015 datasets.}
\label{fig:acs_stree}
\end{sidewaysfigure}
\clearpage
\section{Proof of Theorem \ref{thm:entropyupper} and \ref{thm-boosting}}\label{proof-thm-boosting}
We proceed in two steps, first showing that the loss we care about for fairness \eqref{lossgen} (main file) is upperbounded by the entropy of the $\alpha$-tree $\alphatree$, then developing the boosting result from the minimization of the entropy itself. We thus start with the following Theorem.
\begin{theorem}[Theorem \ref{thm:entropyupper} in Main Paper]\label{thm-alpha-in-alpha-tree}
Suppose Assumption \ref{assumptionBUNF} holds and the outputs of $\alphatree$ are:
\begin{eqnarray}
\alphatree({\bm{e}}{x}) & \defeq & \normalizedlogit\left(\frac{1+\edge(\meas{M}_{\leaf({\bm{e}}{x})}, \postarget)}{2}\right), \forall {\bm{e}}{x} \in \mathcal{X},\label{alphatreeoutput}
\end{eqnarray}
where $\leaf({\bm{e}}{x})$ is the leaf reached by ${\bm{e}}{x}$ in $\alphatree$. Then the following bound holds for the risk \eqref{lossgen}:
\begin{eqnarray}
\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \leq & \entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \label{cvarcomp-deg3-bound}.
\end{eqnarray}
\end{theorem}
\begin{myproof}
We need a simple Lemma, see \textit{e.g.} \citet{nssBP}.
\begin{lemma}\label{lemGB1-NSS}
$\forall \kappa \in \mathbb{R}, \forall B\geq 0, \forall |z| \leq B$,
\begin{eqnarray}
\log(1+\exp(\kappa z)) & \leq & \log(1+\exp(\kappa B)) - \kappa \cdot \frac{B-z}{2}.\label{boundBZ}
\end{eqnarray}
\end{lemma}
We then note, using $z \defeq
\log\left(\frac{1-\posunfair}{\posunfair}\right)$ (stripping variables
for readability) and Assumption \ref{assumptionBUNF},
\begin{eqnarray}
- \log \posfair & = & -\log \left(\frac{\posunfair^\alphatree}{\posunfair^\alphatree+(1-\posunfair)^\alphatree}\right)\nonumber\\
& = & -\log \left(\frac{1}{1+\left(\frac{1-\posunfair}{\posunfair}\right)^\alphatree}\right)\nonumber\\
& = & \log \left(1+\left(\frac{1-\posunfair}{\posunfair}\right)^\alphatree\right)\nonumber\\
& = & \log\left(1 + \exp\left(\alphatree \log\left(\frac{1-\posunfair}{\posunfair}\right)\right)\right)\nonumber\\
& \leq & \log(1+\exp(\alphatree B)) -\alphatree \cdot \frac{B - \log\left(\frac{1-\posunfair}{\posunfair}\right)}{2}\label{bound1}\\
& & = \log(1+\exp(\alphatree B)) -\alphatree \cdot \frac{B + \logit(\posunfair)}{2},\label{blog1}
\end{eqnarray}
where in \eqref{bound1} we have used \eqref{boundBZ} with $\kappa \defeq \alphatree$, using Assumption \ref{assumptionBUNF} guaranteeing $| \logit(\posunfair)| \leq B$. We also get, using this time $\kappa \defeq -\alphatree$,
\begin{eqnarray}
- \log (1-\posfair) & = & \log \left(1 + \exp\left(-\alphatree \log\left(\frac{1-\posunfair}{\posunfair}\right)\right)\right)\nonumber\\
& \leq & \log(1+\exp(-\alphatree B)) +\alphatree \cdot \frac{B + \logit(\posunfair)}{2}\nonumber\\
& & = \log(1+\exp(\alphatree B)) - \alphatree B +\alphatree \cdot \frac{B + \logit(\posunfair)}{2}\nonumber\\
& & = \log(1+\exp(\alphatree B)) - \alphatree \cdot \frac{B - \logit(\posunfair)}{2}.\label{blog2}
\end{eqnarray}
Assembling \eqref{blog1} and \eqref{blog2} for an upperbound to $\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$, we get, using the fact that an $\alpha$-tree partitions $\mathcal{X}$ into regions with constant predictions,
\begin{eqnarray}
\lefteqn{\totrisk(\posfair; \meas{M}, \postarget)}\nonumber\\
& \defeq & \expect_{\X \sim \meas{M}} \left[\postarget(\X)\cdot - \log \posfair(\X)+(1-\postarget(\X))\cdot - \log (1-\posfair(\X))\right] \nonumber\\
& \leq & \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[
\begin{array}{c}
\postarget(\X)\cdot \left(\log(1+\exp(\alphatree(\X) B)) -\alphatree(\X) \cdot \frac{B + \logit(\posunfair(\X))}{2}\right)\nonumber\\
+(1-\postarget(\X))\cdot \left(\log(1+\exp(\alphatree(\X) B)) -\alphatree(\X) \cdot \frac{B - \logit(\posunfair(\X))}{2}\right)
\end{array}\right] \nonumber\\
& & = \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[ \log(1+\exp(\alphatree(\X) B)) - \alphatree(\X) \cdot \left(\begin{array}{c}
\postarget(\X)\cdot \frac{B + \logit(\posunfair(\X))}{2}\\
+(1-\postarget(\X))\cdot \frac{B - \logit(\posunfair(\X))}{2}
\end{array}\right)\right] \nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) \cdot \expect_{\X \sim \meas{M}_\leaf} \left[\left(\begin{array}{c}
\postarget(\X)\cdot \frac{B + \logit(\posunfair(\X))}{2}\\
+(1-\postarget(\X))\cdot \frac{B - \logit(\posunfair(\X))}{2}
\end{array}\right)\right]\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) \cdot \expect_{(\X,\Y) \sim {\measdtarget}_\leaf} \left[\frac{B + \Y \cdot \logit(\posunfair(\X))}{2}\right]\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) \cdot \frac{B + \expect_{(\X,\Y) \sim {\measdtarget}_\leaf} \left[\Y \cdot \logit(\posunfair(\X))\right]}{2}\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) B \cdot \frac{1 + \edge(\meas{M}_\leaf, \postarget)}{2}\right],\label{bENT1}
\end{eqnarray}
where we have used index notation for leaves introduced in the Theorem's statement, used the definition of $\edge(\meas{M}_\leaf, \postarget)$ and let $\alphatree(\leaf)$ denote $\leaf$'s leaf value in $\alphatree$. Looking at \eqref{bENT1}, we see that we can design the leaf values to minimize each contribution to the expectation (noting the convexity of the relevant functions in $\alphatree(\leaf)$), which for any $\leaf \in \leafset(\alphatree)$ we define with a slight abuse of notations as:
\begin{eqnarray}
L(\alphatree(\leaf)) & \defeq & \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) B \cdot \frac{1 + \edge(\meas{M}_\leaf, \postarget)}{2}.
\end{eqnarray}
We note
\begin{eqnarray*}
L'(\alphatree(\leaf)) & = & B \cdot \left(\frac{\exp(\alphatree(\leaf) B)}{1+\exp(\alphatree(\leaf) B)} - \frac{1 + \edge(\meas{M}_\leaf, \postarget)}{2}\right),
\end{eqnarray*}
which zeroes for
\begin{eqnarray*}
\alphatree(\leaf) & = & \frac{1}{B} \cdot \log\left(\frac{1+\edge(\meas{M}_\leaf, \postarget)}{1-\edge(\meas{M}_\leaf, \postarget)}\right) = \normalizedlogit\left(\frac{1+\edge(\meas{M}_{\leaf}, \postarget)}{2}\right),
\end{eqnarray*}
yielding the bound (we use $\edge(\leaf)$ as a shorthand for $\edge(\meas{M}_\leaf, \postarget)$):
\begin{eqnarray}
\lefteqn{\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)}\nonumber\\
& \leq & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ \log\left(1+\frac{1+\edge(\leaf)}{1-\edge(\leaf)}\right) - \log\left(\frac{1+\edge(\leaf)}{1-\edge(\leaf)}\right) \cdot \frac{1 + \edge(\leaf)}{2}\right]\nonumber\\
& & = \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ - \log\left(\frac{1-\edge(\leaf)}{2}\right)- \log\left(\frac{1+\edge(\leaf)}{1-\edge(\leaf)}\right) \cdot \frac{1 + \edge(\leaf)}{2}\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ - \log\left(\frac{1 - \edge(\leaf)}{2}\right) + \frac{1 + \edge(\leaf)}{2}\cdot \log\left(\frac{1 - \edge(\leaf)}{2}\right) - \frac{1 + \edge(\leaf)}{2} \cdot \log\left(\frac{1+\edge(\leaf)}{2}\right) \right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset(\alphatree)}} \left[ -\frac{1 - \edge(\leaf)}{2}\cdot \log\left(\frac{1 - \edge(\leaf)}{2}\right) - \frac{1 + \edge(\leaf)}{2} \cdot \log\left(\frac{1+\edge(\leaf)}{2}\right) \right]\nonumber\\
& = & \entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget),
\end{eqnarray}
which is the statement of Theorem \ref{thm-alpha-in-alpha-tree}.
\end{myproof}\\
Armed with Theorem \ref{thm-alpha-in-alpha-tree}, what we now show is the boosting compliant convergence on the entropy of the $\alpha$-tree. For the informed reader, the proof of our result relies on a generalisation of \citet[Lemma 2]{kmOT}, then branching on the proofs of \citet[Lemma 6, Theorem 9]{kmOT} to complete our result. For this objective, we first introduce notations, summarized in Figure \ref{fig:wha-1}, for the split of a leaf $\leaf_q$ in a subtree with two new leaves $\leaf_p, \leaf_r$. Here, we make use of simplified notation
\begin{eqnarray}
\edge_p & \defeq & \edge(\meas{M}_{\leaf_p}, \postarget),
\end{eqnarray}
and similarly for $\edge_q$ and $\edge_r$. Quantities $p, q, r\in [0,1]$\footnote{Under Assumption \ref{assumptionBUNF}.} are computed from the corresponding $\edge_.$. $\tau$ is the probability, measured from ${\measdtarget}_{\leaf_q}$, that an example has $h(.) = +1$, where $h$ is the split function at $\leaf_q$.
\begin{figure}[h]
\centering
\centerline{\includegraphics[trim=20bp 580bp 700bp 40bp,clip,width=0.8\textwidth]{Figs/FigWHA-1}}
\caption{Main notations used in the proof of Theorem \ref{thm-boosting}, closely following some notations of \citet[Fig. 4]{kmOT}.}
\label{fig:wha-1}
\end{figure}
We state and prove our generalisation to \citet[Lemma 2]{kmOT}.
\begin{lemma}\label{lemWHA-C}
Assuming notations in Figure \ref{fig:wha-1} for the split $h$ investigated at a leaf $\leaf_q$, and letting $\delta \defeq r - p$, if for some $\upgamma > 0$ the split $h$ $\upgamma$-witnesses the WHA at $\leaf$, then $\tau (1-\tau) \delta \geq \upgamma \cdot q(1-q)$.
\end{lemma}
\begin{myproof}
Using the definition of the rebalanced distribution, we have:
\begin{eqnarray}
\lefteqn{\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h(\X)\right]}\nonumber\\
& = & \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\frac{1 - \edge_q \cdot \Y \cdot \normalizedlogit(\posunfair(\X))}{1 - \edge_q^2} \cdot \Y h(\X) \normalizedlogit(\posunfair(\X))\right] \nonumber\\
& = & \frac{\expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\Y h(\X) \normalizedlogit(\posunfair(\X))\right] - \edge_q \cdot \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\normalizedlogit^2(\posunfair(\X)) h(\X)\right]}{1 - \edge_q^2}, \label{eqwha1}
\end{eqnarray}
since $y^2 = 1, \forall y \in \mathcal{Y}$. We also have, by definition of the partition induced by $h$ and the definition of $\tau$,
\begin{eqnarray}
\tau \edge_r - (1-\tau) \edge_p & = & \tau \cdot \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_r}} \left[\Y \cdot \normalizedlogit(\posunfair(\X))\right] - (1-\tau) \cdot \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_p}} \left[\Y \cdot \normalizedlogit(\posunfair(\X))\right]\nonumber\\
& = & \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\Y h(\X) \normalizedlogit(\posunfair(\X))\right].\label{eqwha2}
\end{eqnarray}
We can thus write:
\begin{eqnarray}
\lefteqn{\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h(\X)\right]}\nonumber\\
& = & \frac{\tau \edge_r - (1-\tau) \edge_p - \edge_q \cdot \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\normalizedlogit^2(\posunfair(\X)) h(\X)\right]}{1 - \edge_q^2}\label{eqwha3}\\
& = & \frac{2 \tau \edge_r - \edge_q\cdot \left(1+\expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\normalizedlogit^2(\posunfair(\X)) h(\X)\right]\right)}{1 - \edge_q^2}\label{eqwha4}\\
& = & \frac{2 \tau \edge_r - 2 \tau \edge_q}{1 - \edge_q^2} - \edge_q\cdot \frac{\left(1 - 2\tau +\expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\normalizedlogit^2(\posunfair(\X)) h(\X)\right]\right)}{1 - \edge_q^2}\label{eqwha5}\\
& = & \frac{2 \tau \edge_r - 2 \tau \edge_q}{1 - \edge_q^2} + \edge_q\cdot \frac{\left(2\tau -1 - \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[\normalizedlogit^2(\posunfair(\X)) h(\X)\right]\right)}{1 - \edge_q^2}\label{eqwha6}\\
& = & \frac{2 \tau \edge_r - 2 \tau \edge_q}{1 - \edge_q^2} + \frac{\edge_q\cdot \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[(1-\normalizedlogit^2(\posunfair(\X)))\cdot h(\X)\right]}{1 - \edge_q^2}\label{eqwha7}.
\end{eqnarray}
Here, \eqref{eqwha3} follows from \eqref{eqwha1} and \eqref{eqwha2}, \eqref{eqwha4} uses the fact that $\edge_q = (1-\tau)\edge_p + \tau \edge_r$, \eqref{eqwha5} and \eqref{eqwha6} are convenient reformulations after adding $2 \tau \edge_q-2 \tau \edge_q$ and \eqref{eqwha7} follows from $\expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[h(\X)\right] = 2\tau - 1$ by definition of $\tau$ and $h\in \{-1,1\}$. Let
\begin{eqnarray}
\Delta(h) & \defeq & \edge_q\cdot \expect_{(\X, \Y) \sim {\measdtarget}_{\leaf_q}} \left[(1-\normalizedlogit^2(\posunfair(\X)))\cdot h(\X)\right].
\end{eqnarray}
We have $p = (1 + \edge_p)/2$ (and similarly for $q = (1 + \edge_q)/2$ and $r = (1 + \edge_r)/2$), so we reformulate \eqref{eqwha6} as:
\begin{eqnarray}
\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h(\X)\right]& = & \frac{2\tau(2r-2q)}{4q (1-q)} + \frac{\Delta(h)}{4q (1-q)} \nonumber\\
& = & \frac{\tau(r-q)}{q (1-q)} + \frac{\Delta(h)}{4q (1-q)} \nonumber\\
& = & \frac{\tau(1-\tau) \delta}{q (1-q)} + \frac{\Delta(h)}{4q (1-q)}, \label{lastEQ}
\end{eqnarray}
where the last identity comes from the fact that $r = q + (1-\tau) \delta$. We now have two cases depending on what removing the absolute value in the WHA leads to:\\
\noindent \textbf{Case 1} (i) is $\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h(\X)\right] \geq \upgamma$. We get from \eqref{lastEQ}:
\begin{eqnarray}
\tau(1-\tau) \delta & \geq & \upgamma \cdot q(1-q) - \frac{\Delta(h)}{4}, \label{propH1}
\end{eqnarray}
and since (ii) brings $\Delta(h) \leq 0$, we obtain $\tau(1-\tau) \delta \geq \upgamma \cdot q(1-q) $, as claimed.\\
\noindent \textbf{Case 2} (i) is $\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h(\X)\right] \leq -\upgamma$. Since $\mathcal{H}$ is closed by negation we replace $h$ by $h' \defeq -h$, which satisfies $\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h'(\X)\right] = - \expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h(\X)\right]$. The change switches the sign of $\delta$ by its definition and also $\Delta(h') = -\Delta(h)$ so \eqref{propH1} becomes $-\tau(1-\tau) \delta \leq -\upgamma \cdot q(1-q) + \Delta(h')/4$, \textit{i.e.}
\begin{eqnarray}
\tau(1-\tau) \delta & \geq & \upgamma \cdot q(1-q) - \frac{\Delta(h')}{4}, \label{propH2}
\end{eqnarray}
which brings us back to Case 1 with the switch $h \leftrightarrow h'$ as $h'$ satisfies $\expect_{(\X, \Y) \sim {\rebmeasdtarget}_{\leaf_q}} \left[\Y \normalizedlogit(\posunfair(\X)) h'(\X)\right] \geq \upgamma$. This ends the proof of Lemma \ref{lemWHA-C}.
\end{myproof}\\
Branching Lemma \ref{lemWHA-C} to the proof of Theorem \ref{thm-boosting} via the results of \cite{kmOT} is simple as all major parameters $p, q, r, \delta, \tau$ are either the same or satisfy the same key relationships (linked to the linearity of the expectation). This is why, if we compute the decrease $\entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) - \entropy(\alphatree(\leaf, h); \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$, $\alphatree(\leaf, h)$ being the $\alpha$-tree $\alphatree$ with the split in Figure \ref{fig:wha-1} performed with $h$ at $\leaf$, then we immediately get
\begin{eqnarray}
\entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) - \entropy(\alphatree(\leaf, h); \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \geq & \upgamma^2 q(1-q), \label{eqDEC1}
\end{eqnarray}
which comes from \citet[Lemma 6]{kmOT}, and \eqref{eqDEC1} can be directly used in the proof of \citet[Theorem 9]{kmOT} -- which unravels the local decrease of $\entropy(.; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$ to get to the global decrease of the criterion for the whole of $\alphatree$'s induction --, and to get $\entropy(\alphatree; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) \leq \varepsilon$, it is sufficient that
\begin{eqnarray}
|\leafset(g)| & \geq & \left(\frac{1}{\varepsilon}\right)^{\frac{c \log \left(\frac{1}{\varepsilon}\right)}{\upgamma^2}},\label{eqLAST-th}
\end{eqnarray}
as claimed, for $c>0$ a constant. This ends the proof of Theorem \ref{thm-boosting}.
\begin{remark}
Lemma \ref{lemGB1-NSS} reveals an interesting property: instead of requesting $\faircorr_{S', \leaf} (h) \leq 0$ in split-fair-compliance, suppose we strengthen the assumption, requesting for some $\beta > 0$ that
\begin{eqnarray}
\faircorr_{S', \leaf} (h) & \leq & - \upbeta \cdot (1 - \edge_q^2), \label{betaPROP}
\end{eqnarray}
then the "advantage" $\upgamma$ becomes an advantage $\upgamma + \upbeta$ in \eqref{eqLAST-th}. Since we have $\faircorr_{S', \leaf} (h) \defeq \edge_q\cdot \expect_{(\X, \Y) \sim \meas{D}_{S', \leaf_q}} \left[(1-\normalizedlogit^2(\posunfair(\X)))\cdot h(\X)\right]$, constraint \eqref{betaPROP} quickly vanishes as $|\edge_q| \rightarrow 1$, \textit{i.e.} as the black-box gest very good --or-- very bad (in this last case, we remark that $1-\posunfair$ becomes very good, so this is not a surprise). For example, if $\edge_q \geq 1 - \varepsilon'$ for small $\varepsilon'$, then we just need
\begin{eqnarray}
\expect_{(\X, \Y) \sim \meas{D}_{S', \leaf_q}} \left[(1-\normalizedlogit^2(\posunfair(\X)))\cdot h(\X)\right] & \leq & -\varepsilon' \upbeta \cdot \frac{2-\varepsilon'}{1-\varepsilon'}.
\end{eqnarray}
\end{remark}
\section{General KL Distortion}\label{sec-kl-dist}
We introduce a general Theorem for upperbounding the distortion an \( \alpha \)-correction can make, as promised in Section~\ref{sec-formal}. Notably, the following result does not rely on any setting.
\begin{theorem}\label{thm-distortion-to-blackbox}
For any function $\alpha:\mathcal{X} \rightarrow \mathbb{R}$, any black-box posterior $\posunfair$, and any integer $K\geq 2$, using \eqref{ourposfair} yields the following bound on the KL divergence:
\begin{equation}
\kl(\posunfair, \posfair; \meas{M})
\leq \expect_{\X \sim \meas{M}} \left[\sum_{k= 2}^K \frac{\posunfair(\X)(1-\posunfair(\X)) f^k (\alpha(\X),\posunfair(\X)) }{k(k-1)}\right]
+ o\left(\expect_{\X \sim \meas{M}} \left[(\alpha(\X)-1)^K\right]\right), \label{boundDelta1}
\end{equation}
where we have used function $f: \mathbb{R} \times [0,1]\rightarrow \mathbb{R}$ defined as:
\negativespace
\begin{eqnarray}
f(z,u) & \defeq & \left| \log \left(u/(1-u)\right) \cdot (1-z)\right| = |\logit(u) \cdot (1 - z)|.\label{defFKL}
\negativespace
\end{eqnarray}
\end{theorem}
The proof is presented in Section \ref{proof-thm-distortion-to-blackbox}.
In general, one can see that if \( \alpha(\cdot) \) does not differ from \( 1 \) too much, the KL divergence will be small. This intuitively makes sense as \( \alpha = 1 \) causes \( \posfair = \posunfair \) --- there is no untwisting.
Theorem \ref{thm-distortion-to-blackbox} can be further weakened to provide easier to understand bounds, as per Corollary~\ref{cor-distortion-to-blackbox}.
We further present an alternative setting to that in the main-text, which provides another bound on distortion. In this setting \textbf{(S2)}, we ensure that when the predictions of \( \posunfair \) are confident, then the corresponding \( \alpha \) correction are small (close to \( \alpha = 1 \)).
\begin{enumerate}
\item [\textbf{(S2)}]
$ f(\alpha({\bm{e}}{x}), \posunfair({\bm{e}}{x})) = \left| \logit(\posunfair({\bm{e}}{x})) \cdot (1 - \alpha({\bm{e}}{x}))\right| \leq 1$
(a.s.)
, $f$ being in \eqref{defFKL}.
\negativespace
\end{enumerate}
This provides the alternative upperbound on the distortion.
\begin{corollary}\label{cor-distortion-to-blackbox-b}
Under setting \textbf{(S2)}, we have the upperbound
\begin{eqnarray}
\kl(\posunfair, \posfair; \meas{M}) & \leq & \pi^2/24 \approx 0.41.\label{boundDelta2}
\end{eqnarray}
\end{corollary}
The proof of the Corollary is in \supplement, Section \ref{proof-cor-distortion-to-blackbox} and includes a graphical view of the values of \( \posunfair(.) \) and \( \alpha(.) \) complying with \textbf{(S2)}.
\section{\supplement Experiment Settings} \label{app-exp-settings}
In this \supplement section, we briefly discuss the additional datasets\footnote{Public at: \url{https://github.com/Trusted-AI/AIF360}} and experimental settings included in the subsequent sections. In particular, we highlight the datasets used, the black-boxes post-processed, and specifics of the \topdown algorithm. German Credit and Bank are standard public benchmark datasets in the literature. ACS is a newer dataset with curation details listed in \citep[Section 3]{dhmsRA}.
\subsection*{Datasets}
\begin{itemize}
\item \textbf{German Credit.} In the \supplement, we additionally consider the German Credit dataset, preprocessed by \aif~\citep{aif360}. The dataset consists of only 1000 examples, which is the smallest of the 3 datasets considered. On the other hand, the dataset provided by \aif contains 57 features, primarily from one-hot encoding.
\item \textbf{Bank.} Another dataset we consider in the \supplement is the Bank dataset, preprocessed by \aif~\citep{aif360}. The dataset consists 30488 examples, above the German Credit dataset but below the ACS datasets. The dataset also has 57 features which is largely from one-hot encoding.
\item \textbf{ACS.} The American Community Survey dataset is the dataset we present in the main text. More specifically, we consider the income prediction task (as depicted in the \folktables \python package~\citep{dhmsRA}) over 1-year survey periods in the state of CA. Our of the 3 datasets, ACS provides the largest dataset, with 187475 examples for the 2015 sample of the dataset. Despite this, \folktables only provides 10 features for its prediction task. Through one-hot encoding, this is extended to 29 features.
\end{itemize}
\aif uses a Apache License 2.0 and \folktables uses a MIT License.
Additional \( Z \)-score normalization was used where appropriate. Sensitive attributes are binned into binary and trinary modalities, as specified in the main text (and one-hot encoded for the trinary case).
Each experiment / dataset is used with 5-fold cross-validation and further split such that there are subset partitions for: (1) training the black-box; (2) training a post-processing method; and (3) testing and evaluation.
In particular, we utilize standard cross-validation to split the data into a 80:20 training testing split. The training split is then split randomly equally for separate training of the black-box and post-processing method. The final data splits result in 40:40:20 partitions.
\subsection*{Black-boxes}
\begin{itemize}
\item \textbf{Random Forest.} As per the main text, we primarily consider a calibrated random forest classifier provided by the \sklearn \python package. The un-calibrated random forest classifier consists of 50 decision trees in an ensemble. Each decision tree has a maximum depth of 4 and is trained on a 10\% subset of the black-box training data. In calibration, 5 cross validation folds are used for Platt scaling.
\item \textbf{Neural Network.} Additionally to random forests, we consider a calibrated neural network in the \supplement, also provided by \sklearn. The un-calibrated neural network is trained using mostly default parameters provided by \sklearn. The exception to this is the specification of 300 training iterations and the specification of 10\% of the training set to be used for early stopping.
\end{itemize}
The black-boxes are additionally clipped to adhere to \cref{assumptionBUNF} with \( B = 1 \) for all sections except for \cref{app-exp-clip}.
For the criteria we evaluate \topdown and baselines to in the \supplement, we consider those introduced in the main-text alongside \auc as an additional metric for accuracy.
\subsection*{Compute}
Compute was done with no GPUs. Virtual machines were used with RAM 16GB VCPUs 8 VCPU Disk 30GB from [a local HPC facility] (to be named after publication).
\subsection*{Code}
The code used in this submission is attached in the supplementary material, which will be released upon publication.
\subsection*{\topdown Specifics}
The \( \alpha \)-trees learnt by \topdown are initialized as per \cref{fig:bigalpha}. That is, we initialize sub-\( \alpha \)-trees with \( \alpha = 1 \) for each of the modalities of the sensitive attribute. In addition, each split of the \( \alpha \)-tree consists of projects to a specific feature / attribute. The split is either a modality of the discrete feature or a single linear threshold of a continuous feature. In addition, to avoid over-fitting we restrict splits to only those which result in children node that have at least 10\% of the parent node's examples; and at a minimum have at least \( 30 \) examples for each child node.
For \EOO we utilize a Gaussian Naive Bayes classifier from \sklearn with default parameters to fit \( \bayespo \) from data. We note that no fine-tuning was done for this classifier.
In the \supplement, we consider all variants \topdown examined in the main-text. Additionally we consider the symmetric variant of the \SP \topdown approach. We reiterate the two ways of enforcing \SP: as per Section \ref{sec-fairness}, there are two symmetric strategies for \SP. In particular, we aim to match the either the maximum or minimum subgroup to the opposite extreme (conditionally) expected posterior. As such, we can either match to the largest posterior, which we denote as (\( \uparrow \)), or we can match to the smallest posterior (\( \downarrow \)). We already present \SP \( \uparrow \) in the main-text and additionally present evaluation for \SP \( \downarrow \) here.
\section{Proof of Theorem \ref{thm-alpha-in-alpha-tree}}\label{proof-thm-alpha-in-alpha-tree}
We need a simple Lemma, see \textit{e.g.} \citet{nssBP}.
\begin{lemma}\label{lemGB1-NSS}
$\forall \kappa \in \mathbb{R}, \forall B\geq 0, \forall |z| \leq B$,
\begin{eqnarray}
\log(1+\exp(\kappa z)) & \leq & \log(1+\exp(\kappa B)) - \kappa \cdot \frac{B-z}{2}.\label{boundBZ}
\end{eqnarray}
\end{lemma}
We then note, using $z \defeq
\log\left(\frac{1-\posunfair}{\posunfair}\right)$ (stripping variables
for readability) and Assumption \ref{assumptionBUNF},
\begin{eqnarray}
- \log \posfair & = & -\log \left(\frac{\posunfair^\alphatree}{\posunfair^\alphatree+(1-\posunfair)^\alphatree}\right)\nonumber\\
& = & -\log \left(\frac{1}{1+\left(\frac{1-\posunfair}{\posunfair}\right)^\alphatree}\right)\nonumber\\
& = & \log \left(1+\left(\frac{1-\posunfair}{\posunfair}\right)^\alphatree\right)\nonumber\\
& = & \log\left(1 + \exp\left(\alphatree \log\left(\frac{1-\posunfair}{\posunfair}\right)\right)\right)\nonumber\\
& \leq & \log(1+\exp(\alphatree B)) -\alphatree \cdot \frac{B - \log\left(\frac{1-\posunfair}{\posunfair}\right)}{2}\label{bound1}\\
& & = \log(1+\exp(\alphatree B)) -\alphatree \cdot \frac{B + \logit(\posunfair)}{2},\label{blog1}
\end{eqnarray}
where in \eqref{bound1} we have used \eqref{boundBZ} with $\kappa \defeq \alphatree$, using Assumption \ref{assumptionBUNF} guaranteeing $| \logit(\posunfair)| \leq B$. We also get, using this time $\kappa \defeq -\alphatree$,
\begin{eqnarray}
- \log (1-\posunfairalpha) & = & \log \left(1 + \exp\left(-\alphatree \log\left(\frac{1-\posunfair}{\posunfair}\right)\right)\right)\nonumber\\
& \leq & \log(1+\exp(-\alphatree B)) +\alphatree \cdot \frac{B + \logit(\posunfair)}{2}\nonumber\\
& & = \log(1+\exp(\alphatree B)) - \alphatree B +\alphatree \cdot \frac{B + \logit(\posunfair)}{2}\nonumber\\
& & = \log(1+\exp(\alphatree B)) - \alphatree \cdot \frac{B - \logit(\posunfair)}{2}.\label{blog2}
\end{eqnarray}
Assembling \eqref{blog1} and \eqref{blog2} for an upperbound to $\totrisk(\posfair; \meas{M}, \postarget)$, we get, using the fact that an $\alpha$-tree partitions $\mathcal{X}$ into regions with constant predictions,
\begin{eqnarray}
\lefteqn{\totrisk(\posfair; \meas{M}, \postarget)}\nonumber\\
& \defeq & \expect_{\X \sim \meas{M}} \left[\postarget(\X)\cdot - \log \posfair(\X)+(1-\postarget(\X))\cdot - \log (1-\posfair(\X))\right] \nonumber\\
& \leq & \expect_{\X \sim \meas{M}} \left[
\begin{array}{c}
\postarget(\X)\cdot \left(\log(1+\exp(\alphatree(\X) B)) -\alphatree(\X) \cdot \frac{B + \logit(\posunfair(\X))}{2}\right)\nonumber\\
+(1-\postarget(\X))\cdot \left(\log(1+\exp(\alphatree(\X) B)) -\alphatree(\X) \cdot \frac{B - \logit(\posunfair(\X))}{2}\right)
\end{array}\right] \nonumber\\
& & = \expect_{\X \sim \meas{M}} \left[ \log(1+\exp(\alphatree(\X) B)) - \alphatree(\X) \cdot \left(\begin{array}{c}
\postarget(\X)\cdot \frac{B + \logit(\posunfair(\X))}{2}\\
+(1-\postarget(\X))\cdot \frac{B - \logit(\posunfair(\X))}{2}
\end{array}\right)\right] \nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) \cdot \expect_{\X \sim \meas{M}_\leaf} \left[\left(\begin{array}{c}
\postarget(\X)\cdot \frac{B + \logit(\posunfair(\X))}{2}\\
+(1-\postarget(\X))\cdot \frac{B - \logit(\posunfair(\X))}{2}
\end{array}\right)\right]\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) \cdot \expect_{(\X,\Y) \sim {\measdtarget}_\leaf} \left[\frac{B + \Y \cdot \logit(\posunfair(\X))}{2}\right]\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) \cdot \frac{B + \expect_{(\X,\Y) \sim {\measdtarget}_\leaf} \left[\Y \cdot \logit(\posunfair(\X))\right]}{2}\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) B \cdot \frac{1 + \edge(\meas{M}_\leaf, \postarget)}{2}\right],\label{bENT1}
\end{eqnarray}
where we have used index notation for leaves introduced in the Theorem's statement, used the definition of $\edge(\meas{M}_\leaf, \postarget)$ and let $\alphatree(\leaf)$ denote $\leaf$'s leaf value in $\alphatree$. Looking at \eqref{bENT1}, we see that we can design the leaf values to minimize each contribution to the expectation (noting the convexity of the relevant functions in $\alphatree(\leaf)$), which for any $\leaf \in \leafset(\alphatree)$ we define with a slight abuse of notations as:
\begin{eqnarray}
L(\alphatree(\leaf)) & \defeq & \log(1+\exp(\alphatree(\leaf) B)) - \alphatree(\leaf) B \cdot \frac{1 + \edge(\meas{M}_\leaf, \postarget)}{2}.
\end{eqnarray}
We note
\begin{eqnarray*}
L'(\alphatree(\leaf)) & = & B \cdot \left(\frac{\exp(\alphatree(\leaf) B)}{1+\exp(\alphatree(\leaf) B)} - \frac{1 + \edge(\meas{M}_\leaf, \postarget)}{2}\right),
\end{eqnarray*}
which zeroes for
\begin{eqnarray*}
\alphatree(\leaf) & = & \frac{1}{B} \cdot \log\left(\frac{1+\edge(\meas{M}_\leaf, \postarget)}{1-\edge(\meas{M}_\leaf, \postarget)}\right) = \normalizedlogit\left(\frac{1+\edge(\meas{M}_{\leaf}, \postarget)}{2}\right),
\end{eqnarray*}
yielding the bound (we use $\edge(\leaf)$ as a shorthand for $\edge(\meas{M}_\leaf, \postarget)$):
\begin{eqnarray}
\lefteqn{\totrisk(\posfair; \meas{M}, \postarget)}\nonumber\\
& \leq & \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ \log\left(1+\frac{1+\edge(\leaf)}{1-\edge(\leaf)}\right) - \log\left(\frac{1+\edge(\leaf)}{1-\edge(\leaf)}\right) \cdot \frac{1 + \edge(\leaf)}{2}\right]\nonumber\\
& & = \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ - \log\left(\frac{1-\edge(\leaf)}{2}\right)- \log\left(\frac{1+\edge(\leaf)}{1-\edge(\leaf)}\right) \cdot \frac{1 + \edge(\leaf)}{2}\right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ - \log\left(\frac{1 - \edge(\leaf)}{2}\right) + \frac{1 + \edge(\leaf)}{2}\cdot \log\left(\frac{1 - \edge(\leaf)}{2}\right) - \frac{1 + \edge(\leaf)}{2} \cdot \log\left(\frac{1+\edge(\leaf)}{2}\right) \right]\nonumber\\
& = & \expect_{\leaf \sim \meas{M}_{\leafset}} \left[ -\frac{1 - \edge(\leaf)}{2}\cdot \log\left(\frac{1 - \edge(\leaf)}{2}\right) - \frac{1 + \edge(\leaf)}{2} \cdot \log\left(\frac{1+\edge(\leaf)}{2}\right) \right]\nonumber\\
& = & \entropy(\alphatree; \meas{M}, \postarget),
\end{eqnarray}
which is the statement of Theorem \ref{thm-alpha-in-alpha-tree}.
\section{Distribution shift}\label{app-exp-shift}
To examine how \topdown is effected by distribution shift, we train various wrappers over multiple years of the ACS dataset. In particular, we train and evaluate \( \cvar{} \) wrappers over the ACS dataset from years 2015 to 2018. \cref{fig:rf_yeareval_conservative,fig:rf_yeareval_aggressive} report the \( \cvar{} \) values over the multiple years for the random forest (RF) black-box. \cref{fig:mlp_yeareval_conservative,fig:mlp_yeareval_aggressive} likewise reports corresponding results for neural network (NN) black-boxes.
In particular:
\begin{itemize}
\item \cref{fig:rf_yeareval_conservative} presents the conservative update distribution shift evaluation using a RF black-box with \( B = 1 \) clipping on the ACS dataset over years 2015 to 2018.
\item \cref{fig:rf_yeareval_aggressive} presents the aggressive update distribution shift evaluation using a RF black-box with \( B = 1 \) clipping on the ACS dataset over years 2015 to 2018.
\item \cref{fig:mlp_yeareval_conservative} presents the conservative update distribution shift evaluation using a MLP black-box with \( B = 1 \) clipping on the ACS dataset over years 2015 to 2018.
\item \cref{fig:mlp_yeareval_aggressive} presents the aggressive update distribution shift evaluation using a MLP black-box with \( B = 1 \) clipping on the ACS dataset over years 2015 to 2018.
\end{itemize}
As the ACS dataset consists of census data, one could expect that prior years of the data will be (somewhat) represented in subsequent years of the data. This is further emphasised in the plots, where curves become more closely group together as the training year used to train \topdown~increases, \emph{i.e.}, 2018 containing enough example which are indicative of prior years' distributions. Unsurprisingly, we can see that most circumstances the largest decrease in \( \cvar{} \) (mostly) comes from instances where the data matches the evaluation. \emph{i.e.,} the 2015 curve in (top) \cref{fig:rf_yeareval_conservative}.
Nevertheless, we can see that despite the training data, all evaluation curves decrease from their initial values in all plots; where a slight 'break' in `monotonicity' occurs in some instances of miss-matching data --- most prominently in (top) \cref{fig:rf_yeareval_conservative} for the 2015 plot around 21 boosting iterations. We also remark, perhaps surprisingly, that there is no crossing between curves (\textit{e.g.} as could be expected for the test-2015 and test-2016 curves on training from 2016's data in Figure \ref{fig:rf_yeareval_conservative}), but if test-2015 remains best, we also remark that it does become slightly worse for train-2016 while test-2016 expectedly improves with train-2016 compared to train-2015. Ultimately, all test-* curves converge to a 'midway baseline' on train-2018.
In general, there is little change when comparing the two different black-boxes. The only consist pattern in comparison is that the NN approaches start and end with a smaller \( \cvar{} \) value than their RF counter parts. When comparing binary versus trinary results, there is a distinct larger spread between evaluation curves (between each year within a plot) for the trinary counterparts. This is expected as in the trinary sensitive attribute modality, \( \cvar{} \) is sensitive to additional partitions of the dataset. The spread is further strengthened as the final \( \alpha \)-tree in \topdown often does not provide an \( \alpha \)-correction for all subgroups, \emph{i.e.}, at least one subgroup is not changed by the \( \alpha \)-tree with \( \alpha = 1 \). When comparing conservative versus aggressive approaches, it can also be seen that there is a larger spread between evaluation curves for the aggressive variant.
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/conservative_yeareval_no_legend}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/conservative_multi_yeareval}
\vspace{-0.5cm}
\caption{Random forest black-box conservative \( \cvar{} \) wrapper trained for ACS 2015 to 2018 datasets Each plot is trained on a different dataset year. Each curve colour, indicates the data being used to evaluate the wrapper.}
\label{fig:rf_yeareval_conservative}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/aggressive_yeareval_no_legend}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/aggressive_multi_yeareval}
\vspace{-0.5cm}
\caption{Random forest black-box aggressive \( \cvar{} \) wrapper trained for ACS 2015 to 2018 datasets Each plot is trained on a different dataset year. Each curve colour, indicates the data being used to evaluate the wrapper.}
\label{fig:rf_yeareval_aggressive}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/conservative_yeareval_no_legend}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/conservative_multi_yeareval}
\vspace{-0.5cm}
\caption{Neural Network black-box conservative \( \cvar{} \) wrapper trained for ACS 2015 to 2018 datasets Each plot is trained on a different dataset year. Each curve colour, indicates the data being used to evaluate the wrapper.}
\label{fig:mlp_yeareval_conservative}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/aggressive_yeareval_no_legend}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp/aggressive_multi_yeareval}
\vspace{-0.5cm}
\caption{Neural Network black-box aggressive \( \cvar{} \) wrapper trained for ACS 2015 to 2018 datasets Each plot is trained on a different dataset year. Each curve colour, indicates the data being used to evaluate the wrapper.}
\label{fig:mlp_yeareval_aggressive}
\end{sidewaysfigure}
\clearpage
\section{Experiments on Statistical Parity} \label{app-exp-SP-main}
We refer to \supplement, Section \ref{sec-stat-par} for the formal aspects of handling statistical parity.
\noindent\textbf{For \SP}, a similar pattern to that of \EOO follows, except the differences are more extreme. In particular, the audacious approach fails to optimize for \SP and instead harms it significantly, but does slightly improve \meandiff. The audacious update is problematic here as the target \( \postarget \) in the \SP strategy is a constant (and does not take into consideration of the subgroup being updated). As such the audacious approach should not be used for \SP as it will optimize to match the constant \( \postarget \) more harshly. On the other hand, the conservative update variant provides an improvement to \SP whilst antagonizing \meandiff accuracy. Notably, the ``best'' iteration for \SP and \meandiff occurs at its first iteration (which shows interest in early stopping / pruning the \( \alpha \)-tree). This is expected, as there is large shift when changing from an \( \alpha = 1 \) to the initial rooted value (\emph{e.g.}, \eqref{alphatreeoutputCONS}). The pattern of conservative updates being superior to the audacious counterparts is consistent in other datasets and sensitive attribute modalities. Comparing the conservative \SP \topdown to the baselines, discounting \baseeoob for \meandiff, we find that \basespa and \basespb result in lower \SP and \meandiff. This is unsurprising: our \topdown treatment \SP can result in harsh updates; in \supplement~(pg \pageref{sec-stat-par}), we discuss an alternative approach using ties with optimal transport.
\section{Example Alpha-Tree} \label{app-exp-example}
In this section, we provide an example of an \( \alpha \)-tree generated using \topdown. In particular, we look at one example from training \( \cvar{} \) \topdown on the Bank dataset with binary sensitive attributes. \cref{fig:example_alpha_tree} presents the example \( \alpha \)-tree. The tree contains information about the attributes in which splits are made and the \( \alpha \)-correction made at leaf nodes (and their induced partition). In the example, could note that the \( \alpha \) trees for modalities of the age sensitive attribute are imbalanced. The right tree is significantly smaller than the left. One could also note the high reliance on ``education" based attributes for determining partitions. These factors could be used to scrutinise the original blackbox; and eventually, even provide constraints on the growth of an \( \alpha \)-tree which would aim to avoid certain combinations of attribute. We leave these factors for future work.
\begin{sidewaysfigure}
\centering
\includegraphics[width=\textwidth]{Figs/example_tree.pdf}
\caption{Example tree generated in the optimization of \topdown for \( \cvar{} \) in the Bank dataset with binary sensitive attributes.}
\label{fig:example_alpha_tree}
\end{sidewaysfigure}
\section{Proof of Theorem \ref{thm-distortion-to-blackbox}}\label{proof-thm-distortion-to-blackbox}
We first show two technical Lemmata.
\begin{lemma}\label{lemTE}
For any $a \geq 0$, let
\begin{eqnarray}
h(z) & \defeq & \log\left(\frac{1}{1+a^{1+z}}\right) .
\end{eqnarray}
We have
\begin{eqnarray}
h^{(k)}(z) & = & -\frac{\log^k (a) \cdot a^{1+z}}{(1+a^{1+z})^{k}}\cdot P_{k-1}(a^{1+z}),
\end{eqnarray}
where $P_k(x)$ is a degree-$k-1$ polynomial. Letting $c_{k,j}$ the constant factor of monomial $x^j$ in $P_k(x)$, for $j\leq k-1$, we have the following recursive definitions: $c_{1,0} = 1$ ($k=1$) and
\begin{eqnarray}
c_{k+1,k} & = & (-1)^{k},\\
c_{k+1,j} & = & (j+1) \cdot c_{k,j} - (k+1-j)\cdot c_{k,j-1}, \forall 0<j<k,\\
c_{k+1,0} & = & 1.
\end{eqnarray}
Hence, we have for example $P_1(x) = 1, P_2(x) = -x + 1, P_3(x) = x^2 - 4x + 1, P_4(x) = -x^3 + 11x^2 - 11x + 1, ...$.
\end{lemma}
\begin{myproof}
We let
\begin{eqnarray}
f(z) & \defeq & \frac{a^{1+z}}{1+a^{1+z}},
\end{eqnarray}
so that $h'(z) = - \log(a) \cdot g(z)$ and we show
\begin{eqnarray}
f^{(k)}(z) & = & \frac{\log^k (a) \cdot a^{1+z}}{(1+a^{1+z})^{k+1}}\cdot P_k(a^{1+z}).\label{propGder}
\end{eqnarray}
We first check
\begin{eqnarray}
f'(z) & = & \frac{\log(a) \cdot a^{1+z}}{(1+a^{1+z})^2},
\end{eqnarray}
which shows $P_1(x) = 1$. We then note that for any $k\in \mathbb{N}_*$,
\begin{eqnarray}
\frac{\mathrm{d}}{\mathrm{d} z} \frac{a^{1+z}}{(1+a^{1+z})^k} & = & \frac{\log(a) \cdot a^{1+z}}{(1+a^{1+z})^{k+1}}\cdot(-(k-1) a^{1+z} + 1),
\end{eqnarray}
so the induction case yields $f^{(k+1)}(z) \defeq {f^{(k)}}'(z)$, that is:
\begin{eqnarray}
\lefteqn{f^{(k+1)}(z)}\nonumber\\
& = & \log^k (a) \cdot \frac{\mathrm{d}}{\mathrm{d} z} \left(\frac{a^{1+z}}{(1+a^{1+z})^{k+1}}\cdot P_k(a^{1+z})\right)\nonumber\\
& = & \log^k (a) \cdot \left(\frac{\log(a) \cdot a^{1+z}}{(1+a^{1+z})^{k+2}}\cdot(-k a^{1+z} + 1)\cdot P_k(a^{1+z}) + \frac{a^{1+z}\cdot \log(a) }{(1+a^{1+z})^{k+1}}\cdot a^{1+z}\cdot \left.\frac{\mathrm{d} P_k(x) }{\mathrm{d} x}\right|_{x=a^{1+z}}\right)\nonumber\\
& = & \frac{\log^{k+1} (a) \cdot a^{1+z}}{(1+a^{1+z})^{k+2}}\cdot \underbrace{\left((-k a^{1+z} + 1)\cdot P_k(a^{1+z}) + a^{1+z}(1+a^{1+z})\cdot \left.\frac{\mathrm{d} P_k(x) }{\mathrm{d} x}\right|_{x=a^{1+z}}\right)}_{\defeq P_{k+1}(a^{1+z})},
\end{eqnarray}
from which we check that $P_{k+1}$ is indeed a polynomial and its coefficients are obtained via identification from $P_k$, which establishes \eqref{propGder} and yields to the statement of the Lemma.
\end{myproof}
\begin{lemma}\label{lemCK}
Coefficient $c_{k,j}$ admits the following bound, for any $0\leq j\leq k$:
\begin{eqnarray}
|c_{k,j}| & \leq & (k-1)! \binom{k-1}{j} .
\end{eqnarray}
\end{lemma}
\begin{myproof}
First, we have the following recursive definition for the absolute value of the leveraging coefficients in $c_{.,.}$ (we call them $a_{.,.}$ for short): $|c_{.,.}| = a_{.,.}$ with
\begin{eqnarray}
a_{k+1,k} & = & 1,\\
a_{k+1,j} & = & (j+1) \cdot a_{k,j} + (k+1-j)\cdot a_{k,j-1}, \forall 0<j<k,\\
a_{k+1,0} & = & 1.
\end{eqnarray}
We now show by induction that $a_{k+1,j} \leq k! \binom{k}{j} \defeq b_{k+1, j}$. For $j=0$, $b_{k+1,0} = k ! \geq a_{k+1,0}$ ($k\geq 2$) and for $j=k$, $b_{k+1,k} = k! \geq a_{k+1,0}$ as well. We now check, assuming the property holds at all ranks $k$, that for ranks $k+1$, we have
\begin{eqnarray}
a_{k+1,j} & = & (j+1) \cdot a_{k,j} + (k+1-j)\cdot a_{k,j-1}\nonumber\\
& \leq & (j+1) (k-1)! \binom{k-1}{j} + (k+1-j)(k-1)! \binom{k-1}{j-1},
\end{eqnarray}
and we want to check that the RHS is $\leq k! \binom{k}{j}$ for any $0<j<k$. Simplifying yields the equivalent inequality
\begin{eqnarray}
(j+1)(k-j) + (k+1-j)j & \leq & k^2.
\end{eqnarray}
finding the worst case bound for $j$ yields $j = k/2$ (we disregard the fact that $j$ is an integer) and plugging in the bound yields the constraint on $k$: $k \geq 2$, which indeed holds.
\end{myproof}\\
We also check that $h$ in Lemma \ref{lemTE} is infinitely differentiable. As a consequence, we get from Lemma \ref{lemTE} the Taylor expansion around $g=1$ (for any $a\geq 0$) at any order $K\geq 2$,
\begin{eqnarray}
\lefteqn{\log\left(\frac{1}{1+a^g}\right)}\nonumber\\
& = & \log\left(\frac{1}{1+a}\right) - \frac{a \log a}{1+a}\cdot (g-1) - \underbrace{\sum_{k=2}^K \frac{a \log^k (a) P_{k-1}(a) }{k! (1+a)^{k}}\cdot (g-1)^k}_{\defeq R_{K,a}(g)} + o((g-1)^K).\label{taylExp1}
\end{eqnarray}
The choice to start the summation at $k=2$ is done for technical simplifications to come. We thus have
\begin{eqnarray*}
\log \posfair({\bm{e}}{x}) & = & \log\left(\frac{1}{1+\left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right)^{\alpha({\bm{e}}{x})}}\right)\\
& = & \log \posunfair({\bm{e}}{x}) - (1-\posunfair({\bm{e}}{x})) \log\left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right) \cdot ({\alpha({\bm{e}}{x})}-1) - R_{\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})},K}({\alpha({\bm{e}}{x})}) \\
& & + o((\alpha({\bm{e}}{x})-1)^K),\\
\log (1-\posfair({\bm{e}}{x})) & = & \log (1-\posunfair({\bm{e}}{x})) - \posunfair({\bm{e}}{x}) \log\left(\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right) \cdot ({\alpha({\bm{e}}{x})}-1) - R_{\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})},K}({\alpha({\bm{e}}{x})})\\
& & + o((\alpha({\bm{e}}{x})-1)^K).
\end{eqnarray*}
Define for short $\Delta_{\mbox{\tiny{u}}}({\bm{e}}{x}) \defeq \posunfair({\bm{e}}{x})\cdot - \log \posfair({\bm{e}}{x}) + (1-\posunfair({\bm{e}}{x}))\cdot - \log (1-\posfair({\bm{e}}{x})) - (\posunfair({\bm{e}}{x})\cdot - \log \posunfair({\bm{e}}{x}) + (1-\posunfair({\bm{e}}{x}))\cdot - \log (1-\posunfair({\bm{e}}{x})))$, so that $\kl(\posunfair, \posfair; \meas{M}) = \expect_{\X \sim \meas{M}} \left[\Delta_{\mbox{\tiny{u}}}(\X)\right]$. The Taylor expansion \eqref{taylExp1} unveils an interesting simplification:
\begin{eqnarray*}
\Delta_{\mbox{\tiny{u}}}({\bm{e}}{x}) & = & -\posunfair({\bm{e}}{x})\log \posunfair({\bm{e}}{x}) +\posunfair({\bm{e}}{x}) (1-\posunfair({\bm{e}}{x})) \log\left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right) \cdot ({\alpha({\bm{e}}{x})}-1) \\
& & +\posunfair({\bm{e}}{x})\cdot R_{\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})},K}({\alpha({\bm{e}}{x})})\\
& & - (1-\posunfair({\bm{e}}{x}))\log (1-\posunfair({\bm{e}}{x})) + (1-\posunfair({\bm{e}}{x}))\posunfair({\bm{e}}{x}) \log\left(\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right) \cdot ({\alpha({\bm{e}}{x})}-1) \\
& & +(1-\posunfair({\bm{e}}{x}))\cdot R_{\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})},K}({\alpha({\bm{e}}{x})})\\
& & - (\posunfair({\bm{e}}{x})\cdot - \log \posunfair({\bm{e}}{x}) + (1-\posunfair({\bm{e}}{x}))\cdot - \log (1-\posunfair({\bm{e}}{x})))+ o((\alpha({\bm{e}}{x})-1)^K)\\
& = & \posunfair({\bm{e}}{x})\cdot R_{\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}}({\alpha({\bm{e}}{x})}) +(1-\posunfair({\bm{e}}{x}))\cdot R_{\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})},K}({\alpha({\bm{e}}{x})}) + o((\alpha({\bm{e}}{x})-1)^K), \forall {\bm{e}}{x} \in \mathcal{X},
\end{eqnarray*}
so the divergence to the black-box prediction simplifies as well, this time using Lemma \ref{lemCK}:
\begin{eqnarray}
\kl(\posunfair, \posfair; \meas{M}) & = & \expect_{\X \sim \meas{M}} \left[\posunfair(\X)\cdot R_{\frac{1-\posunfair(\X)}{\posunfair(\X)},K}({\alpha(\X)}) +(1-\posunfair(\X))\cdot R_{\frac{\posunfair(\X)}{1-\posunfair(\X)},K}({\alpha(\X)})\right] \nonumber\\
& & + o\left(\expect_{\X \sim \meas{M}} \left[(\alpha(\X)-1)^K\right]\right).\label{deltaug}
\end{eqnarray}
Not touching the little-oh term, we simplify further and bound the term in the expectation: for any ${\bm{e}}{x} \in \mathcal{X}$,
\begin{eqnarray}
\lefteqn{\posunfair({\bm{e}}{x})\cdot R_{\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})},K}({\alpha({\bm{e}}{x})}) +(1-\posunfair({\bm{e}}{x}))\cdot R_{\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})},K}({\alpha({\bm{e}}{x})})} \nonumber\\
& = & \posunfair({\bm{e}}{x})\cdot \sum_{k=2}^K \frac{\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\cdot \log^k \left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right) P_{k-1}\left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right) }{k! \left(1+\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right)^{k}}\cdot ({\alpha({\bm{e}}{x})}-1)^k\nonumber\\
& & + (1-\posunfair({\bm{e}}{x}))\cdot \sum_{k=2}^K \frac{\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\cdot \log^k \left(\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right) P_{k-1}\left(\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right) }{k! \left(1+\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right)^{k}}\cdot ({\alpha({\bm{e}}{x})}-1)^k\nonumber\\
& = & \posunfair({\bm{e}}{x})\cdot \sum_{k=2}^K \frac{\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\cdot \log^k \left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right) \cdot \sum_{j=0}^{k-2} c_{k-1,j}\left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right)^j }{k! \left(1+\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right)^{k}}\cdot ({\alpha({\bm{e}}{x})}-1)^k\nonumber\\
& & + (1-\posunfair({\bm{e}}{x}))\cdot \sum_{k=2}^K \frac{\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\cdot \log^k \left(\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right) \cdot \sum_{j=0}^{k-2} c_{k-1,j}\left(\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right)^j }{k! \left(1+\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right)^{k}}\cdot ({\alpha({\bm{e}}{x})}-1)^k\nonumber\\
& = & \sum_{k=2}^K \frac{\log^k \left(\frac{1-\posunfair({\bm{e}}{x})}{\posunfair({\bm{e}}{x})}\right) \cdot \sum_{j=0}^{k-2} c_{k-1,j} \cdot \posunfair^{k-j}({\bm{e}}{x}) (1-\posunfair({\bm{e}}{x}))^{j+1}}{k!}\cdot ({\alpha({\bm{e}}{x})}-1)^k\nonumber\\
& & + \sum_{k=2}^K \frac{\log^k \left(\frac{\posunfair({\bm{e}}{x})}{1-\posunfair({\bm{e}}{x})}\right) \cdot \sum_{j=0}^{k-2} c_{k-1,j} \cdot (1-\posunfair({\bm{e}}{x}))^{k-j} \posunfair^{j+1}({\bm{e}}{x}) }{k!}\cdot ({\alpha({\bm{e}}{x})}-1)^k\label{eqBDeltaX}
\end{eqnarray}
We now note, using Lemma \ref{lemCK} that for any ${\bm{e}}{x} \in \mathcal{X}$,
\begin{eqnarray*}
\lefteqn{\sum_{j=0}^{k-2} |c_{k-1,j}| \cdot \posunfair^{k-j}({\bm{e}}{x}) (1-\posunfair({\bm{e}}{x}))^{j+1}}\nonumber\\
& = & \posunfair^{2}({\bm{e}}{x})(1-\posunfair({\bm{e}}{x}))\cdot \sum_{j=0}^{k-2} |c_{k-1,j}| \cdot \posunfair^{k-2-j}({\bm{e}}{x}) (1-\posunfair({\bm{e}}{x}))^{j}\\
& \leq & \posunfair^{2}({\bm{e}}{x})(1-\posunfair({\bm{e}}{x}))\cdot \sum_{j=0}^{k-2}(k-2)! \binom{k-2}{j} \posunfair^{k-2-j}({\bm{e}}{x}) (1-\posunfair({\bm{e}}{x}))^{j}\nonumber \\
& & = (k-2)!\cdot \posunfair^{2}({\bm{e}}{x})(1-\posunfair({\bm{e}}{x}))\cdot \underbrace{\sum_{j=0}^{k-2} \binom{k-2}{j} \posunfair^{k-2-j}({\bm{e}}{x}) (1-\posunfair({\bm{e}}{x}))^{j}}_{=(1-\posunfair({\bm{e}}{x}) + \posunfair({\bm{e}}{x}))^{k-2} = 1}\nonumber\\
& = & (k-2)! \cdot \posunfair^{2}({\bm{e}}{x})(1-\posunfair({\bm{e}}{x})),
\end{eqnarray*}
and similarly
\begin{eqnarray*}
\sum_{j=0}^{k-2} |c_{k-1,j}| \cdot (1-\posunfair({\bm{e}}{x}))^{k-j} \posunfair^{j+1}({\bm{e}}{x}) & \leq & (k-2)! \cdot \posunfair({\bm{e}}{x})(1-\posunfair({\bm{e}}{x}))^{2},
\end{eqnarray*}
so plugging the two last bounds on \eqref{eqBDeltaX} yields the bound on $\kl(\posunfair, \posfair; \meas{M})$ from \eqref{deltaug}:
\begin{eqnarray}
\kl(\posunfair, \posfair; \meas{M}) & \leq & \expect_{\X \sim \meas{M}} \left[\sum_{k= 2}^K \frac{(\posunfair^{2}(\X)(1-\posunfair(\X))+\posunfair(\X)(1-\posunfair(\X))^2) \left| \log \left(\frac{1-\posunfair(\X)}{\posunfair(\X)}\right)\right|^k }{k(k-1)}\cdot |{\alpha(\X)}-1|^k\right] \nonumber\\
& & + o\left(\expect_{\X \sim \meas{M}} \left[(\alpha(\X)-1)^K\right]\right)\nonumber\\
& = & \expect_{\X \sim \meas{M}} \left[\sum_{k= 2}^K \frac{\posunfair(\X)(1-\posunfair(\X))\left| \log \left(\frac{1-\posunfair(\X)}{\posunfair(\X)}\right)\right|^k }{k(k-1)}\cdot |{\alpha(\X)}-1|^k\right] \nonumber\\
& & + o\left(\expect_{\X \sim \meas{M}} \left[(\alpha(\X)-1)^K\right]\right),
\end{eqnarray}
which yields the statement of Theorem \ref{thm-distortion-to-blackbox}.
\section{Additional Main Text Experiments} \label{app-exp-rest}
In this section, we report the experiments identical to that presented in the main-text, including missing criteria, settings, and the additional German Credit and Bank datasets. Each plot we present provides the binary and trinary sensitive attribute settings over all criteria discussed in the previous setting.
In particular:
\begin{itemize}
\item \cref{fig:rf_german_fairness} presents the evaluation using a RF black-box with \( B = 1 \) clipping on the German Credit dataset.
\item \cref{fig:rf_bank_fairness} presents the evaluation using a RF black-box with \( B = 1 \) clipping on the Bank dataset.
\item \cref{fig:rf_acs_fairness} presents the evaluation using a RF black-box with \( B = 1 \) clipping on the ACS dataset.
\end{itemize}
\subsection*{Fairness Models}
In comparison to ACS, \cref{fig:rf_bank_fairness} for the Bank dataset performs similarly to the main text figure. There are only slight deviations in the ordering of which \topdown settings perform best. For example, the \( \cvar{} \) optimization of audacious and conservative updates are a lot closer in the Bank dataset than that of the ACS 2015 dataset.
In comparison, the result's of \topdown on the German Credit largely deviate from that of the other experiments. This can be clearly seen in the number of boosting iteration \topdown completes being significantly lower before the entropy stops being decreased (and thus terminating the algorithm). Another major deviation is that \( \cvar{} \) fails to get lowered for both binary and trinary sensitive attribute modalities in the German Credit dataset. Despite this, \EOO and \SP both have slight improvements for the best corresponding \topdown setting (conservative \EOO and conservative \( \SP \uparrow \)), which is consistent with other datasets. This is despite the original classifier's \EOO and \SP being significantly lower than the ACS dataset. However, there is a major cost in the case of \EOO, where the accuracy (both for \meandiff and \auc) is harmed significantly.
A reason for the significantly worse performance, predominantly in \( \cvar{} \) optimization, of \topdown for the German Credit is likely the significantly smaller number of example available in the dataset. Given that there are only 1000 examples and 57 features variables, the 40:40:20 split of the dataset results in the subsets to not be representative of the entire dataset's support. Additionally, \( \cvar{} \) is strongly tied to the cross-entropy loss function and empirical risk minimization~\citep{wmFR,ruOO}. As such, given the nonrepresentative subsets of the dataset used for training \topdown, minimizing the \( \cvar{} \) for low sample inputs is difficult.
Thus for such a failure cases, one should confirm that \topdown is not decreasing fairness to prevent social harms.
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/german_credit_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/german_credit_multi_fairness.pdf}
\caption{\topdown~optimized for different fairness models evaluated on German Credit with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:rf_german_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/bank_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/bank_multi_fairness.pdf}
\caption{\topdown~optimized for different fairness models evaluated on Bank with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:rf_bank_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/acs_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf/acs_multi_fairness.pdf}
\caption{\topdown~optimized over boosting iterations for different fairness models evaluated on ACS 2015 with binary (up) and trinary (down) sensitive attributes. ``c" on the x-axis denotes the clipped black-box. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:rf_acs_fairness}
\end{sidewaysfigure}
\clearpage
\section{Additional Equality of Opportunity Details}
\label{sec:sup_eoo}
In this section, we present additional details for the equality of opportunity (EOO) strategy presented in \cref{sec-fairness}. In particular, we present the full definition of the $(p,\delta)$-\textit{pushup} of $\bayespo$.
\begin{definition}\label{def-pushup}
Fix $p\in [0,1]$ and let $\mathcal{X}_p$ be a subset of $\mathcal{X}$ such that (i) $\inf \bayespo(\mathcal{X}_p) \geq \sup \bayespo(\mathcal{X}\backslash \mathcal{X}_p)$ and (ii) $\int_{\mathcal{X}_p} \dmeas{M} = p$. For any $\delta \geq 0$, the $(p,\delta)$-pushup of $\bayespo$, $\bayespo_{p,\delta}$, is the posterior defined as $\bayespo_{p,\delta} = \bayespo$ if $\inf \bayespo(\mathcal{X}_p) \geq 1/2$ and otherwise:
\begin{eqnarray}
\bayespo_{p,\delta}({\bm{e}}{x}) & \defeq & \left\{
\begin{array}{ccl}
\frac{1}{2}+\delta & \mbox{ if } \bayespo({\bm{e}}{x}) \in [\inf \bayespo(\mathcal{X}_p), 1/2 + \delta] \\
\bayespo({\bm{e}}{x}) & \multicolumn{2}{l}{\mbox{ otherwise.}}
\end{array}
\right. \label{eoo_mapping}
\end{eqnarray}
\end{definition}
Following Fig.~\ref{fig:eta-diff-eoo} (right) and discussion in the main-text, we make a couple more observation. From Def.~\ref{def-pushup}, the selection of the set \( \mathcal{X}_{p} \) (\emph{i.e.} choice of \( p \)) presents a tradeoff between accuracy and the fairness objective. As \( p \) increases, the size of \( \mathcal{X}_{p} \) necessarily increases. For instance, taking \( \mathcal{X}_{p} = \{ {\bm{e}}{x} : \bayespo({\bm{e}}{x}) \in > l \} \), the larger \( \mathcal{X}_{p} \) is the more negative prediction are flipped. Thus intuitively, \( p \) measures the size of correct posterior values and \( \delta \) defines the size of the correction.
\section{Proof of Theorem \ref{thm-eoo}}\label{proof-thm-eoo}
We remind that we craft product measures using a mixture and a posterior that shall be implicit from context: we thus note that the KL divergence
\begin{eqnarray}
\kl(\postarget, \posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}) & \defeq & \expect_{(\X, \Y)\sim \measdtarget} \left[\log \left(\frac{\dmeasdtarget((\X,\Y))}{\dmeasdfair((\X,\Y))}\right) \right]\\
& = & -\expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[\postarget(\X)\cdot \log \left(\frac{\posfair(\X)}{\postarget(\X)}\right) + (1-\postarget(\X))\cdot \log \left(\frac{1-\posfair(\X)}{1-\postarget(\X)}\right) \right]\\
& = & \totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) - \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[ H(\postarget(\X))\right],\label{defKL2}
\end{eqnarray}
where $\measdtarget$ (resp. $\measdfair$) is obtained from couple $(\meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget)$ (resp. $(\meas{M}_{\mbox{\tiny{\textup{t}}}}, \posfair)$). Denote
\begin{eqnarray}
s^\circ & \defeq & \arg\min_s \pr_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right],
\end{eqnarray}
where $\predfair$ is the $+1/-1$ prediction obtained from the posterior $\posfair$ using \textit{e.g.} the sign of its logit. We define the total variation divergence:
\begin{eqnarray}
\tv(\postarget, \posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}) & \defeq & \int_{\mathcal{X}\times \mathcal{Y}} |\dmeasdtarget((\X,\Y))-\dmeasdfair((\X,\Y))|,
\end{eqnarray}
which, because of the definition of the product measures, is also equal to:
\begin{eqnarray}
\tv(\postarget, \posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}) & = & \int_{\mathcal{X}} | \postarget(\X) \dmeas{M}_{\mbox{\tiny{\textup{t}}}}(\X) - \posfair(\X) \dmeas{M}_{\mbox{\tiny{\textup{t}}}}(\X)| \\
& & + \int_{\mathcal{X}} | (1-\postarget(\X)) \dmeas{M}_{\mbox{\tiny{\textup{t}}}}(\X) - (1-\posfair(\X)) \dmeas{M}_{\mbox{\tiny{\textup{t}}}}(\X)|\\
& = & 2 \int_{\mathcal{X}} | \postarget(\X) - \posfair(\X) | \dmeas{M}_{\mbox{\tiny{\textup{t}}}}(\X).\label{defTV2}
\end{eqnarray}
We have Pinsker's inequality, $\tv(\postarget, \posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}) \leq \sqrt{2 \kl(\postarget, \posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}})}$ (see \textit{e.g.} \cite{vehRD}), so if we run \topdown~until
\begin{eqnarray}
\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \leq & \frac{\tau^2}{2} + \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[ H(\postarget(\X))\right],\label{condTOPDOWNa}
\end{eqnarray}
then because of \eqref{defKL2} and \eqref{defTV2},
\begin{eqnarray}
\int_{\mathcal{X}} | \postarget(\X) - \posfair(\X) | \dmeas{M}_{\mbox{\tiny{\textup{t}}}}(\X) & \leq & \tau.\label{firstapprox}
\end{eqnarray}
Denote subgroups $s^\star \defeq \arg\max_s \pr_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right]$ and $s^\circ \defeq \arg\min_s \pr_{\X \sim \meas{P}_s} \left[\predfair(\X) = 1 \right]$. We pick
\begin{eqnarray}
\meas{M}_{\mbox{\tiny{\textup{t}}}} & \leftarrow & \meas{P}_{s^\circ}
\end{eqnarray}
for $\topdown$ and the $(p, \delta)$-push up posterior $\postarget$, with
\begin{eqnarray}
p & \defeq & \pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right] + \frac{\delta}{2},\label{defP}
\end{eqnarray}
assuming the RHS is $\leq 1$.
Denote $\mathcal{X}_{p, s^\circ}$ the subset of the support of $\meas{P}_{s^\circ}$ such that $\postarget(\X) \geq (1/2) + \delta$. Notice that by definition,
\begin{eqnarray}
\int_{\mathcal{X}_{p, s^\circ}} \dmeas{P}_{s^\circ}(\X) & = & p.\label{defXCIRC}
\end{eqnarray}
We have two possible outcomes for $\posfair$ of relevance on $\mathcal{X}_{p, s^\circ}$: (i) $\posfair(\X) \leq 1/2$ and (ii) $\posfair(\X) > 1/2$. Notice that in this latter case, we are guaranteed that $\predfair(\X) = 1$, which counts towards bringing closer $\pr_{\X \sim \meas{P}_{s^\circ}} \left[\predfair(\X) = 1 \right]$ to $\pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right]$, so we have to make sure that (i) occurs with sufficiently small probability, and this is achieved via guarantee \eqref{firstapprox}.
If the total weight on $\mathcal{X}_{p, s^\circ}$ of the event (i) $\posfair(\X) \leq 1/2$ is more than $\delta$, then
\begin{eqnarray}
\int_{\mathcal{X}} | \postarget(\X) - \posfair(\X) | \dmeas{P}_{s^\circ}(\X) & \geq & \int_{\mathcal{X}_{p, s^\circ}} | \postarget(\X) - \posfair(\X) | \dmeas{P}_{s^\circ}(\X)\nonumber\\
& \geq & \left|\frac{1}{2}+\delta - \frac{1}{2}\right| \cdot \int_{\mathcal{X}_{p, s^\circ}} \iver{\posfair(\X) \leq 1/2} \dmeas{P}_{s^\circ}(\X) \nonumber\\
& > & \left|\frac{1}{2}+\delta - \frac{1}{2}\right| \cdot \delta \nonumber\\
& & = \delta^2. \label{secondapprox}
\end{eqnarray}
If we have the relationship $\delta = \sqrt{\tau}$, then we get a contradiction with \eqref{firstapprox}. In conclusion, if \eqref{condTOPDOWNa} holds, then
\begin{eqnarray}
\int_{\mathcal{X}_{p, s^\circ}} \iver{\posfair(\X) \leq 1/2} \dmeas{P}_{s^\circ}(\X) & \leq & \delta.\label{boundBAD}
\end{eqnarray}
In summary, for any $\tau > 0$, if we run \topdown~with the choices $\meas{M}_{\mbox{\tiny{\textup{t}}}} \leftarrow \meas{P}_{s^\circ}$ (which corresponds to the "worst treated" subgroup with respect to EOO) and craft the $(p, \delta)$-push up posterior $\postarget$ with $p$ as in \eqref{defP}, then
\begin{eqnarray}
\pr_{\X \sim \meas{P}_{s^\circ}} \left[\predfair(\X) = 1 \right] & \geq & \int_{\mathcal{X}_{p, s^\circ}} \iver{\posfair(\X) > 1/2} \dmeas{P}_{s^\circ}(\X)\\
& & = \int_{\mathcal{X}_{p, s^\circ}} (1-\iver{\posfair(\X) \leq 1/2}) \dmeas{P}_{s^\circ}(\X)\\
& = & \int_{\mathcal{X}_{p, s^\circ}} \dmeas{P}_{s^\circ}(\X) - \int_{\mathcal{X}_{p, s^\circ}} \iver{\posfair(\X) \leq 1/2} \dmeas{P}_{s^\circ}(\X)\\
& \geq & p - \delta \label{useb}\\
& & = \pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right] - \frac{\delta}{2},
\end{eqnarray}
where \eqref{useb} makes use of \eqref{defXCIRC} and \eqref{boundBAD}. Fixing $\delta \defeq 2 \epsilon$, $\epsilon$ being used in \eqref{eqEOO} (main file), we obtain
\begin{eqnarray}
\pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right] - \pr_{\X \sim \meas{P}_{s^\circ}} \left[\predfair(\X) = 1 \right] & \leq & \epsilon,\label{finEQ}
\end{eqnarray}
and via relationship $\delta = \sqrt{\tau}$, we check that \eqref{condTOPDOWNa} becomes the following function of $\epsilon$:
\begin{eqnarray}
\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \leq & 8\epsilon^4 + \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[ H(\postarget(\X))\right],\label{condTOPDOWN}
\end{eqnarray}
and we get the statement of the Theorem for the choice \eqref{defP}, which corresponds to $K=2$ and reads
\begin{eqnarray}
p & \defeq & \pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right] + \epsilon.\label{defPEPS}
\end{eqnarray}
If the RHS in \eqref{defPEPS} is not $\leq 1$, we can opt for an alternative with one more free variable, $K\geq 1$,
\begin{eqnarray}
p & \defeq & \pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right] + \frac{\delta}{K},\label{defP2}
\end{eqnarray}
where $K$ is large enough for the constraint to hold. In this case, to keep \eqref{finEQ} we must have $\delta(K-1)/K = \epsilon$, which elicitates
\begin{eqnarray}
\delta & = & \frac{K\epsilon}{K-1}
\end{eqnarray}
instead of $\delta \defeq 2 \epsilon$, bringing
\begin{eqnarray}
p & \defeq & \pr_{\X \sim \meas{P}_{s^\star}} \left[\predfair(\X) = 1 \right] + \frac{\epsilon}{K-1},\label{defP3}
\end{eqnarray}
and a desired approximation guarantee for \topdown~of:
\begin{eqnarray}
\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \leq & \frac{K^4}{2(K-1)^4} \cdot \epsilon^4 + \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[ H(\postarget(\X))\right].\label{condTOPDOWN2}
\end{eqnarray}
Since $K > 1$, $K^4/(K-1)^4\geq 1$, so we are guaranteed that \eqref{condTOPDOWN2} holds if we ask for
\begin{eqnarray}
\totrisk(\posfair; \meas{M}_{\mbox{\tiny{\textup{t}}}}, \postarget) & \leq & \frac{\epsilon^4}{2} + \expect_{\X \sim \meas{M}_{\mbox{\tiny{\textup{t}}}}} \left[ H(\postarget(\X))\right],\label{condTOPDOWN3}
\end{eqnarray}
\section{High Clip Value} \label{app-exp-clip}
In this section, we consider a higher clipping value than that used in other experiments. In other sections, we consider a \( B = 1 \) clipping value which results in posterior restricted between roughly \( [0.27, 0.73] \). Although this clipping seems harsh, from the prior experiments one can see that \topdown provides a lot of improvement across all fairness criterion (and we will see \( B = 1\) allows \topdown to improve beyond optimization for a large clip value).
We will now consider \topdown experiments which correspond to evaluation over \( \cvar{} \), \EOO, and \SP criterion with clipping \( B = 3 \) (as discussed in theory sections of the main text). This restricts the posterior to be between roughly \( [0.05, 0.95] \). \cref{fig:rf3_german_fairness,fig:rf3_bank_fairness,fig:rf3_acs_fairness} presents RF plots over German, Bank, and ACS datasets; and \cref{fig:mlp3_german_fairness,fig:mlp3_bank_fairness,fig:mlp3_acs_fairness} presents equivalent MLP plots.
In particular:
\begin{itemize}
\item \cref{fig:rf3_german_fairness} presents the evaluation using a RF black-box with \( B = 3 \) clipping on the German Credit dataset.
\item \cref{fig:rf3_bank_fairness} presents the evaluation using a RF black-box with \( B = 3 \) clipping on the Bank dataset.
\item \cref{fig:rf3_acs_fairness} presents the evaluation using a RF black-box with \( B = 3 \) clipping on the ACS dataset.
\item \cref{fig:mlp3_german_fairness} presents the evaluation using a NN black-box with \( B = 3 \) clipping on the German Credit dataset.
\item \cref{fig:mlp3_bank_fairness} presents the evaluation using a NN black-box with \( B = 3 \) clipping on the Bank dataset.
\item \cref{fig:mlp3_acs_fairness} presents the evaluation using a NN black-box with \( B = 3 \) clipping on the ACS dataset.
\end{itemize}
In general, there is only a slight difference between the RF and MLP plots in this clipping setting.
We focus on the RF ACS plot of the higher clipping value, \cref{fig:rf3_acs_fairness}. The most striking issue is that the minimization of \( \cvar{} \) is a lot worse than when using clipping \( B = 1 \). In particular, \basebb (which in \cref{fig:rf3_acs_fairness} has \( B = 3 \)) is not beaten by the final wrapped classifier produced by either update of \topdown. However, for \EOO and \SP there is still a reduction in criterion, although a lower reduction for some cases, \emph{i.e.}, conservative \EOO. It is unsurprising that \( \cvar{} \) is more difficult to optimize in this case as the black-box would be closer to an optimal accuracy / cross-entropy value without larger clipping. As a result, \( \cvar{} \) would be more difficult to improve on as it depends on subgroup / partition cross-entropy. In particular, the large spike in the first iteration of boosting is striking. This comes from the fact that we are no directly minimizing a partition's cross-entropy directly, but an upper-bound, where the theory specifies that the upper-bound requires that the original black-box is already an \( \alpha \)-tree with correct corrections. However, as the the original black-box is not an \( \alpha \)-tree with correction specified by the update, the initial update can cause an increase in the \( \cvar{} \) (which appears to be more common with higher clipping values).
Despite the initial ``jump" and in-ability to recover, let us compare the \( B = 3 \) plot to the original \( B = 1 \) RF \topdown plot given in \cref{fig:rf_acs_fairness}. From comparing the results, one can see that the final boosting iteration for the \( B = 1 \) aggressive updates beats the \( B = 3 \) black-box classifiers. Thus, even when comparing against \( \cvar{} \) which is highly influenced by accuracy (thus a higher clipping value is desired), a smaller clipping value resulting in a more clipped black-box posterior is potentially more useful in \( \cvar{} \) \topdown. If one looks at the conservative curves in \cref{fig:rf_acs_fairness}, these do not beat the \( B = 3 \) black-box. This further strengthens the argument that the aggressive update is preferred in \( \cvar{} \) \topdown; and is further emphasized by the increase cap between curves with \( B = 3 \) black-boxes, as shown in \cref{fig:rf3_acs_fairness}.
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf3/german_credit_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf3/german_credit_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{RF with \( B = 3 \) \topdown~optimized for different fairness models evaluated on German Credit with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:rf3_german_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf3/bank_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf3/bank_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{RF with \( B = 3 \) \topdown~optimized for different fairness models evaluated on Bank with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:rf3_bank_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf3/acs_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/rf3/acs_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{RF with \( B = 3 \) \topdown~optimized for different fairness models evaluated on Bank with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:rf3_acs_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp3/german_credit_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp3/german_credit_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{MLP with \( B = 3 \) \topdown~optimized for different fairness models evaluated on German Credit with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:mlp3_german_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp3/bank_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp3/bank_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{MLP with \( B = 3 \) \topdown~optimized for different fairness models evaluated on Bank with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:mlp3_bank_fairness}
\end{sidewaysfigure}
\clearpage
\begin{sidewaysfigure}[!ht]
\centering
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp3/acs_fairness_no_legend.pdf}
\includegraphics[width=\textwidth,trim={0, 6pt, 0, 0},clip]{Figs/experiments_plots/mlp3/acs_multi_fairness.pdf}
\vspace{-0.5cm}
\caption{MLP with \( B = 3 \) \topdown~optimized for different fairness models evaluated on Bank with binary (up) and trinary (down) sensitive attributes. Crosses denote when a subgroup's \( \alpha \)-tree is initiated (over any fold). The shade depicts \( \pm \) a standard deviation from the mean. However, this disappears in the case where other folds stop early.}
\label{fig:mlp3_acs_fairness}
\end{sidewaysfigure}
\clearpage
\section{Proof of Lemma \ref{lem-entropies}}\label{proof-lem-entropies}
We note that $H_2(1/2) = 1$, so we can reformulate:
\begin{eqnarray}
\frac{\entropy_2(\leaf; \meas{M}, \postarget)}{\log 2} & = & (1- (\edge^+_{\leaf} + \edge^-_{\leaf}))\cdot H_2\left(\frac{1}{2}\right) +(\edge^+_{\leaf} + \edge^-_{\leaf})\cdot H_2 \left(\frac{\edge^+_{\leaf}}{\edge^+_{\leaf}+\edge^-_{\leaf}}\right),\label{propH21}
\end{eqnarray}
and we also have $\edge^+_{\leaf} \leq 0, \edge^-_{\leaf} \geq 0, \edge^+_{\leaf}+\edge^-_{\leaf} \leq 1$, plus
\begin{eqnarray}
(1- (\edge^+_{\leaf} + \edge^-_{\leaf}))\cdot\left(\frac{1}{2}\right) +(\edge^+_{\leaf} + \edge^-_{\leaf})\cdot \left(\frac{\edge^+_{\leaf}}{\edge^+_{\leaf}+\edge^-_{\leaf}}\right) = \frac{1 + \edge^+_{\leaf} - \edge^-_{\leaf} }{2} = \frac{1+\edge(\meas{M}_\leaf, \postarget)}{2},
\end{eqnarray}
as indeed $\edge(\meas{M}_\leaf, \postarget) = \edge^+_{\leaf} - \edge^-_{\leaf}$ from its definition. Thus, by Jensen's inequality, since $H$ is concave,
\begin{eqnarray*}
\lefteqn{\log(2) \cdot \left(1+(\edge^+_{\leaf}+\edge^-_{\leaf})\cdot \left(H_2 \left(\frac{\edge^+_{\leaf}}{\edge^+_{\leaf}+\edge^-_{\leaf}}\right) - 1\right)\right)}\nonumber\\
& = & \log(2) \cdot \left((1- (\edge^+_{\leaf}+\edge^-_{\leaf}))\cdot H_2\left(\frac{1}{2}\right) +(\edge^-_{\rho, \leaf} + \edge^+_{\rho, \leaf})\cdot H_2 \left(\frac{\edge^+_{\leaf}}{\edge^+_{\leaf}+\edge^-_{\leaf}}\right)\right)\\
& \leq & \log(2) \cdot H_2\left( (1- (\edge^+_{\leaf}+\edge^-_{\leaf})) \cdot \frac{1}{2} + (\edge^+_{\leaf}+\edge^-_{\leaf})\cdot \frac{\edge^+_{\leaf}}{\edge^+_{\leaf}+\edge^-_{\leaf}}\right)\\
& & = \log(2) \cdot H_2\left(\frac{1+\edge(\meas{M}_\leaf, \postarget)}{2}\right)\\
& = & H\left(\frac{1+\edge(\meas{M}_\leaf, \postarget)}{2}\right),
\end{eqnarray*}
which, after plugging in expectations and simplifying, yields the statement of Lemma \ref{lem-entropies}.
|
\section{Introduction}
Multihop wireless relaying in a cellular network has been a long due necessity. The use of relaying has shown to be a promising deployment solution for extending coverage area and sometimes boosting the network capacity of a cellular network \cite{hoymann2012}. It becomes more crucial in 5G, where ultradense deployment of millimeter-wave (mmWave) cells is the key enabler to meet 5G requirements such as high spectral efficiency, low energy consumption, and low latency. Ultradense networks pose high operation and capital costs for network operators because each base station (BS) needs to connect to the 5G core (5GC) through fiber backhaul. Furthermore, wired backhaul may face deployment restrictions in many urban as well as hard-to-reach areas. Wireless backhauling helps deploy ultradense networks easily and quickly without incurring additional wired backhaul costs.
It is appealing that the mmWave spectrum is also a prime candidate for wireless backhaul. It exhibits large bandwidth, high directional links, and almost noise-limited characteristics \cite{akdeniz2014millimeter}. Hence, operators can leverage a part of the spectrum to provide fiber-like reliability. By not employing a separate spectrum license for backhaul, we can further reduce radio hardware and deployment costs. The above merits allow operators to envision diverse deployment scenarios in cellular networks, such as outdoor-to-indoor, outdoor small-cell, and group mobility (e.g., cells on buses or trains).
Recognizing its importance to facilitate faster 5G deployments, the 3rd Generation Partnership Project (3GPP) standardized a solution for multihop relaying support over 5G new radio (NR) in Release 16 \cite{ts38300} called integrated access and backhaul (IAB). This feature supports both out-band and in-band backhauling. For the latter case, access and backhaul links can share the same spectrum in either time, frequency, or spatial domain. Many research activities have demonstrated the feasibility and potential of mmWave-based IAB networks using either end-to-end simulations \cite{polese2020integrated}, field trials \cite{tian2019field, cudak2021integrated}, or investigating routing and resource scheduling schemes \cite{saha2019millimeter, li2017radio}.
An IAB network has two types of BSs: IAB-donors are the BSs that have fiber (or wired) backhaul to the 5GC, while IAB-nodes are the relays that use one or more wireless backhaul links to reach IAB-donors. The IAB-donors thus act as access gateways for IAB-nodes and provide core connectivity to them. An IAB-node, therefore, plays a dual role -- as a user equipment (UE) from the perspective of upstream BSs and as a gNB (base station in NR) from the perspective of downstream BSs and UEs. Figure \ref{fig:illustration} represents an example of a multihop IAB network where macro BS is the IAB-donor, and small cell BSs are IAB-nodes.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{images/iab-network}
\caption{An illustration of a cellular network structure that supports multihop relaying. Each relay (fixed or moving) creates its own small cell and the macro BS act as the anchor point for these cells.}
\label{fig:illustration}
\end{figure}
The 3GPP 5G IAB architecture, however, has certain limitations that we will discuss later. This article aims to take a step in this direction by presenting a flexible IAB architecture for beyond 5G (B5G) networks. As we would observe later, the proposed architecture is flexible enough to handle almost every IAB design requirement and relaying deployments scenarios. Although the 3GPP is currently discussing some of them, the timeline for others is unknown at this point. The proposed architecture has minimal impact on the existing solutions for orchestration and management (OAM) and network service provisioning as 5GC is untouched. In essence, we design an IAB architecture that would help operators to incremental deploy 5G cells or extend an already deployed cellular network anytime, anywhere without worrying much about geographical restrictions and spectrum crunch.
\section{IAB Use Cases and Architecture Design Requirements} \label{sec:iab_requirements}
To enable the IAB feature, one needs modifications to the standard cellular system. These modifications can be, for example, a fresh design of the protocol stack or addition of new layers into the existing protocol stack to realize backhaul-related signaling and management. The 3GPP opted for the latter approach, which is more convenient and plausible. We may need to (re)design interfaces among IAB-nodes and between IAB-nodes and IAB-donors to carry signaling across multihop wireless links. We may also have to rethink C- and U-plane procedures such as bearer and mobility management, handover, and radio link failure (RLF) recovery. Incorporating these aspects transparently into the existing 5G radio access network (RAN) would achieve a robust IAB architecture.
The architecture should have provisions to enable a smooth transition and flexible integration from/to the legacy deployments. By designing an architecture for legacy terminals, relaying operations become transparent to procedures for UEs, gNBs, and 5GC elements. Enforcing quality of service (QoS) for each bearer across IAB RAN is another facet that is paramount to architecture success. Since mmWave backhaul links are subject to blocking from moving objects or infrastructure changes, we also need topology adaptation and multi-connectivity capabilities. Under these features, it is possible to add or remove IAB-nodes to the network topology autonomously, recover from backhaul link overload/failure and reconfigure backhaul under local congestion in the event of variable traffic conditions.
Support for mobile relays or vehicle mounted relays (VMRs) is necessary to improve the user experience for the onboard passengers in moving vehicles like buses or trains. Another merit of mobile relaying is group mobility, which prevents enormous signaling that may arise due to concurrent handovers of in-vehicle UEs. Mobile relay can also facilitate new deployment opportunities. For example, VMRs can be used to reach UEs with no or poor cellular coverage. Or vehicle-to-vehicle relaying can form a dynamic network and provide access to vehicles that become out of macro coverage (e.g., due to blocking) \cite{noh2020}.
Lastly, NR coexistence with other radio access technologies (RATs) would allow operators to achieve better spectrum efficiency and wider coverage at lower costs. More importantly, in early deployments where NR might not have full coverage, the inter-RAT mobility between LTE and NR will often occur. With multi-RAT coexistence, IAB RAN can be deployed in areas with partial or no core connectivity by utilizing other RATs (e.g., LTE or wireless LAN (WLAN)) as overlay networks. Such a capability for IAB may give rise to newer deployment scenarios in the future. Both network and users would also benefit from faster mobility, enhanced service continuity, and augmented capacity when relays can maintain simultaneous connections to multiple cells belonging to different RATs. By aggregating these design aspects, we obtain a blueprint of an IAB network that can be deployed on sites where mobile traffic demand may arise.
\section{A Primer on 3GPP IAB Architecture} \label{sec:3gpp_arch}
The Release $16$ IAB specification defines radio protocols, upper layer management, and physical layer aspects for the 3GPP IAB architecture. The architecture has been designed to have low processing and design complexity at an IAB-node and minimal impact on the 5GC (and its related signaling overhead). Each IAB-node hosts a distributed unit (DU) and a UE part called mobile termination (MT). It is responsible for the lower layers of the radio interface to UEs or downstream MTs (of other IAB-nodes). The IAB-donor hosts a DU and the centralized unit (CU) that handles the control and upper layers of the radio interfaces. The IAB-donor-CU is responsible for IAB network topology and route management. It also manages radio resources if they are supervised centrally. Before an IAB-node can provide DU functionality, its MT registers to the 5GC and establish an NR connection with its parent node. Afterward, each IAB-node-DU maintains a logical F1 interface with the IAB-donor-CU regardless of its hop level. An additional layer called the backhaul adaptation protocol (BAP) layer on top of the radio link control (RLC) layer carries information on network topology and routing between the IAB-donor-CU and IAB-node-DUs. Specifically, the IAB-donor-CU configures BAP routing ID (carried in BAP header) on each IAB-node so that the BAP layer can route IP packets to the appropriate next node. Release 17, which may be available by March $2022$, is working on a few IAB enhancements such as improving load balancing, spectral efficiency, and multihop latency. The specifications \cite{ts38300, tr38874, ts38340} offer more details to interested readers on the architecture and its working.
\begin{figure*}
\begin{minipage}{0.39884719156\textwidth}
\begin{subfigure}{\linewidth}
\includegraphics[width=\linewidth]{images/part1}
\caption{} \label{subfig:part1}
\end{subfigure}
\vfill
\begin{subfigure}{\linewidth}
\includegraphics[width=\linewidth]{images/part2}
\caption{} \label{subfig:part2}
\end{subfigure}
\end{minipage}
\hfill
\begin{subfigure}{0.60115280843\textwidth}
\includegraphics[width=\linewidth]{images/part3}
\caption{} \label{subfig:part3}
\end{subfigure}
\caption{Evolution of the proposed IAB architecture from the standard 5G architecture along with network elements and their interfaces. The additional (IAB related) nodes and interfaces are in orange color. (a) Abstract view of the standard 3GPP 5G architecture. (b) Splitting of gNB as gNB-DU and gNB-CU units. Even if these units are not colocated but are connected via an IP network, together, they still provide gNB functionality. (c) Abstract view of the proposed architecture. The two units communicate via the 5G IP network to which they belong.}
\label{fig:arch}
\end{figure*}
That said, the 3GPP IAB architecture has a few limitations that may hinder its widespread commercial deployments. First, the architecture only supports fixed relays; IAB-nodes cannot move seamlessly across different parent nodes while retaining contexts of downstream MTs and UEs. Although migration of an IAB-node to another parent node
underneath the same IAB-donor-CU is supported but only in the events of backhaul RLF recovery. Second, operators must deploy IAB-donors in a planned way before installing any IAB-node for proper relay operations. IAB-nodes can only work with and anchor through IAB-donors (that have special gNB-DUs to support relaying) and not with the traditional gNBs. This defeats the purpose of extending network coverage and deploying 5G cells flexibly wherever needed. The possibility of dual- or multi-connectivity for IAB-nodes also reduces because IAB-donors typically would have sparser deployment than gNBs. Third, the BAP layer requires a meticulous design for related C- and U-plane signaling. For example, we need enhancements in the radio resource control (RRC) layer to provide BAP addresses to IAB-nodes and treat an MT and UE differently at layer-2 and layer-3. Finally, the architecture does not support multi-RAT coexistence yet. As an example, we cannot use different RATs on access and backhaul at an IAB-node (e.g., LTE on access and NR on backhaul). This feature would act as a perfect platform for multi-RAT convergence in B5G networks where a UE can use any RAT for access, independent of RAT used for backhaul and associated core network.
As part of future releases, the 3GPP may work on new use cases and possibly mitigate some of these limitations. Some design issues, however, would remain the same. For example, even if Release $18$ supports mobile IAB, the movement would still be restricted to deployment regions having IAB-donors.
\section{Proposed IAB Architecture}
We provide an alternate architecture to support multihop relaying by connecting gNB-CU and gNB-DU over the IP connection provided by the 5G network itself. Figure \ref{fig:arch} contains a graphic summary of how our IAB architecture has evolved from the standard 5G RAN architecture. To maintain consistency with the 3GPP, we have kept the same nomenclature for MT, IAB-node, and IAB-donor. The noticeable difference is that IAB-donor is now moved to the edge cloud and hosts the gNB-CU and a dedicated user plane function (dUPF). The latter forwards RLC and lower layers packets from IAB-node-DUs to the IAB-node-CU through its IP plane. Accordingly, the IAB-donor-CU acts as a server and is connected to the dUPF over the N$6$ interface. The IAB-node-DU and IAB-donor-CU communicate via the logical F1 interface over the 5G IP connection. The edge cloud is in the proximity of RAN, ensuring that C- and U-plane delays between MT and IAB-donor are affordable. We only need a single IAB-donor-CU at the edge cloud for all IAB-node-DUs within the network. We can place multiple IAB-donor-CUs as well if required for maintaining load balancing (or backup) and network robustness. In contrast, the 3GPP solution requires relatively more IAB-donors in the IAB RAN. Note that the logical separation of a gNB into two units is only for visualization purposes and understanding the architecture workings. In reality, the gNB (as shown in Fig. \ref{subfig:part3}) is indeed a BS and can exist as a single logical unit (i.e., a full gNB) or composition of CU and DU(s) apart from the IAB RAN elements.
\begin{figure*}
\centering
\begin{subfigure}{0.35779\textwidth}
\includegraphics[width=\linewidth]{images/two_stacks}
\caption{} \label{subfig:two_stacks}
\end{subfigure}
\hfill
\begin{subfigure}{0.54221\textwidth}
\includegraphics[width=\linewidth]{images/multi_stacks}
\caption{} \label{subfig:multi_stacks}
\end{subfigure}
\caption{Protocol stack at IAB-nodes in the proposed architecture for (a) two-hop and (b) multihop IAB system.}
\label{fig:stacks}
\end{figure*}
The MT makes an NR connection to a gNB and reaches IAB-donor over the 5G IP network. The IAB-node-DU and the IAB-donor-CU then, in conjunction, provide gNB functionalities to prospective UEs. Therefore, the IAB-node forwards/receives C- and U-plane packets of UE to/from the 5GC. These packets are exchanged between the IAB-node-DU and the IAB-donor-CU through the F1 interface over the 5G IP connection established earlier. The IAB-donor-CU then communicates the packets with the 5GC over its N2/N3 interface. In the end, the UE packets follow the path IAB-node$\rightarrow$gNB$\rightarrow$dUPF$\rightarrow$IAB-donor-CU$\rightarrow$5GC.
The U-plane path of MT terminates at the IAB-donor-CU irrespective of its hop level. The MT obtains the IP address of its endpoint (i.e., IAB-donor-CU) from the OAM server during the initial setup procedure. If multiple IAB-donor-CUs are available, the OAM server can select the best one based on criteria like load situation and quality of backhaul links. An IAB-node, acting as an IP router for the previous (child or parent) node, forwards packets based on their destination IP addresses and selects appropriate RLC channels. Because of this flat IP design, the IAB network can support IP-related functionality efficiently, such as multicast, multipath, differentiated services, and well-known routing protocols. Many of these functions are taken care of by the BAP layer in the 3GPP IAB architecture.
\subsection{Protocol Stack}
Figure \ref{fig:stacks} shows the protocol stack of the proposed IAB architecture for two-hop relaying (corresponding to the network in Fig. \ref{subfig:part3}). We see another difference between our architecture and the 3GPP IAB architecture -- MT has the complete UE stack instead of the partial UE stack as in the latter case. Having complete UE functionalities simplify many UE-related procedures for an MT. The figure also illustrates that the architecture uses existing/standard interfaces and protocol layers, making them economical and easier to develop and maintain. They also facilitate installing IAB-nodes anywhere without requiring additional interface modules at available cell sites, that is, they can connect to any gNB.
It may seem that an IAB-node needs more processing in our architecture than in the 3GPP counterpart due to an additional layer (i.e., PDCP + SDAP vs. BAP). However, the 3GPP solution would need packet data convergence protocol (PDCP) layer at IAB-nodes anyway for security protection of the F1 interface \cite{ts38300}. Also, when VMR and multi-RAT capabilities are to be supported in future releases, implementing the BAP layer may become more challenging than it already is.
The two-hop IAB architecture can easily accommodate multiple intermediate IAB-nodes to support multihop relaying operations. For example, Fig. \ref{subfig:multi_stacks} depicts the protocol stack for a $3$-hop relaying case. Note that the access IAB-node (IAB-node2), i.e., the serving BS for UE, has a different protocol stack than that of intermediate IAB-node (IAB-node1). By employing a full gNB Uu stack, IAB-node1 configures and maintains NR connection facing its downstream MT (of IAB-node2) as if it were a UE. The MT then uses its backhaul data radio bearer (DRB) to send C- and U-plane messages to the 5GC or communicate to the IAB-donor. Similarly, the UE communicates to the 5GC over path IAB-node2$\rightarrow$IAB-node1$\rightarrow$gNB$\rightarrow$dUPF$\rightarrow$IAB-donor-CU for both its C- and U-plane messages.
Of course, intermediate IAB-nodes and anchor gNBs can have their own UEs. How does then IAB-node1 differentiate whether a child node is MT or UE? After registration, like a UE usually does, an MT asks the 5GC to reach its endpoint (the IAB-donor-CU) over IP. The 5GC, in response, tells the concerned gNB-CU (coincidently the IAB-donor-CU) to establish a data path to the dUPF. Because the IAB-donor-CU has been already using the dUPF to exchange messages with the MT (via gNB and the IAB-node1), it becomes aware that the UE context is actually from an MT. Next, it communicates this information to IAB-node1 over the F1 interface. The IAB-node1 marks the UE and MT contexts so that the UE data is forwarded via the F1 interface but the MT data via PDCP+SDAP layers. How does a gNB differentiate between MTs and UEs? It is straightforward -- IAB-donor is contacted by MTs only. Consequently, the IAB-node1 and gNB can treat MTs and UEs differently and manage their traffic based on certain precedence rules. Similarly, IAB-donor receives F1 packets only from the access IAB-node (and not from intermediate IAB-nodes), making it aware of UE's location within the IAB topology.
\textbf{Bearer mapping and QoS support: } Like a UE, each MT can establish multiple backhaul DRBs supporting diverse QoS needs. We notice that the C-plane signaling of a UE is sent via the U-plane of the IAB-node. To guarantee its QoS, we can reserve a DRB only to carry C-plane signaling at IAB-nodes and assign it the highest priority among DRBs. For U-plane data, an IAB-node selects an appropriate DRB to meet the QoS requirements of the data flow. Further, the IAB-node can multiplex several UE bearers (may belong to different UEs) onto a single IAB-node DRB on a similar QoS profile, typically termed many-to-one bearer mapping. Whenever the existing IAB-node bearer can no longer satisfy the QoS requirements, the IAB-donor-CU reprograms the related QoS parameters for the corresponding IAB-node bearer and sends the updated configuration to the intermediate IAB-nodes as well.
\textbf{Mobile IAB and VMR: } The proposed architecture can handle MT mobility by leveraging its full UE stack and performing the standard handover procedures. This ensures that MT mobility is transparent to the traditional network elements (and other UEs/MTs) and does not depend on whether the target BS is IAB-donor, IAB-node, or gNB. Therefore, a handover between IAB-nodes is the same as the standard inter-gNB-DU handover. In contrast, a handover between a gNB and IAB-node is actually an inter-gNB-CU handover as each IAB-node is a DU of the IAB-donor-CU. As an outcome, our architecture would experience significantly lower MT handovers and signaling overhead than the 3GPP architecture (if future releases support mobile IAB), specifically in urban deployments.
Probably the most promising aspect of our architecture is that when a VMR moves from one gNB to another, onboard UEs do not experience any handover. As mentioned above, several UE bearers are carried through the end-to-end IP tunnel between VMR and the IAB-donor-CU. From the 5G system perspective, the UEs are still connected to the same gNB (i.e., VMR + IAB-donor-CU). In other words, there is no impact on UEs' connectivity to 5GC as neither its gNB-DU nor gNB-CU changes. Similarly, when an IAB-node performs handover between parent nodes belonging to different anchor gNBs, no handover signaling would occur for its child nodes (UEs and MTs). In the existing 3GPP solution, however, the onboard UEs will have RRC connections with the anchor BS (i.e., IAB-donor) itself. Therefore, simultaneous handovers will occur if the VMR moves across different anchor BSs (or IAB-donors), which may lead to prohibitive signaling overhead.
\textbf{Topology Adaptation and Multi-connectivity: } We can reuse existing F1AP functionality to support route redundancy in the C-plane. For example, an MT can send C-plane signaling on each backhaul link and perform RRC connection re-establishment only when all serving links become unavailable. Alternatively, MT can simultaneously connect to multiple parent nodes and manage any addition/removal of redundant routes after considering propagation conditions and traffic load on each backhaul link. Currently, 5G only supports NR-NR dual connectivity (NR-DC) at the PDCP level. Researchers are now working on extensions of NR-DC to multi-connectivity. We can additionally provide multi-connectivity at the IP level by having multiple MT instances at an IAB-node.
\section{Multi-RAT Extension of the Architecture}
The proposed architecture is flexible enough to allow multi-RAT coexistence where access and backhaul can belong to different RATs (5G, LTE, or WLAN) without the need for inter-operability interfaces. For example, to provide NR access over LTE backhaul, a UE connects to IAB-node over NR interface, but the IAB-node connects to its parent node (an eNB in such case) over LTE backhaul. This necessitates that the IAB-node should possess a gNB-DU radio stack towards the UE and an LTE UE radio stack towards the eNB as depicted in Fig. \ref{subfig:5g_over_lte}. The IAB-node then establishes an IP tunnel to the IAB-donor and provides 5GC connectivity to the UE as described earlier. Consequently, a connection is created between NR UE and 5GC over an LTE IP network. Similarly, LTE access over 5G backhaul is possible if the IAB-node has an LTE eNB radio stack towards UE and an NR UE radio stack towards gNB, and the LTE network supports control and user plane separation (CUPS) mechanism \cite{ts23214}. Figure \ref{subfig:wifi_over_5g} shows the protocol stack of an IAB-node if we want to realize WLAN access (or any non-3GPP access for that matter) over 5G backhaul. The WLAN access point (WAP) part of the IAB-node establishes a Y2 interface to a non-3GPP interworking function (N3IWF) \cite{ts23501} over 5G IP network. These new envisaged deployments will support UE mobility across different RATs if the UE has radio interfaces of those RATs.
\begin{figure}[h]
\begin{minipage}{0.5\textwidth}
\begin{subfigure}{\linewidth}
\includegraphics[width=\linewidth]{images/5g_over_lte}
\caption{} \label{subfig:5g_over_lte}
\end{subfigure}
\vfill
\begin{subfigure}{\linewidth}
\includegraphics[width=\linewidth]{images/wifi_over_5g}
\caption{} \label{subfig:wifi_over_5g}
\end{subfigure}
\vfill
\begin{subfigure}{\linewidth}
\includegraphics[width=\linewidth]{images/hybrid_arch}
\caption{} \label{subfig:hybrid_arch}
\end{subfigure}
\end{minipage}
\caption{Examples of some of the multi-RAT scenarios that are possible to realize by extending the proposed IAB architecture. These scenarios are (a) 5G access over LTE backhaul, (b) WLAN access over 5G backhaul, and (c) Hybrid architecture with NR, LTE, and WLAN in the RAN, and 5GC and EPC in the core network. Any combination of RATs can be used for access and backhaul.}
\label{fig:multi_rat}
\end{figure}
Note that 5G access over LTE/WLAN backhaul may not provide requisite QoS to some of the DRBs established for time-sensitive or data-intensive services. This deployment is meant to be used only when anticipated services can run on LTE/WLAN infrastructures. A logical extension of multi-RAT capability is to enable dual- and multi-connectivity where UEs/MTs have multiple access links that may belong to different RATs, for example, E-UTRA-NR dual connectivity. Interestingly, it is widely accepted that the initial rollout of 5G is non-standalone (NSA) deployment, which means UEs would use NR cells for U-plane and LTE for C-plane operations. Many operators believe that NSA deployment can provide enhanced mobile broadband with reliable connectivity. Through the multi-connectivity feature, our architecture can expedite the rollout of NSA deployments before deploying 5GC and allow users to take advantage of 5G technology sooner.
Lastly, we can design a multi-RAT multihop IAB system where a UE would have the flexibility to utilize a particular RAT, for example, based on service requirements and RAT load level. Such a hybrid cellular system is beneficial when a specific network (5G, LTE or WLAN) is partially available or unavailable (e.g., due to natural disaster) and an on-demand network is needed. A crucial aspect for realizing this, as shown in Fig. \ref{subfig:hybrid_arch}, is to have both core networks (i.e., evolved packet core (EPC) and 5GC) simultaneously in the system. We can then readily set up a relay where a core network can become independent of RAN.
\section{Performance Analysis}
To illustrate how the proposed architecture performs against the 3GPP architecture, we assess their performance on two aspects: 1) average and cell-edge throughput, and 2) mobility robustness. We evaluate these metrics using the 3GPP channel model \cite{tr36873} for the mmWave spectrum and the NYU model \cite{akdeniz2014millimeter} for beamforming gains. We assume heterogeneous deployment with $7$ hexagonal macro cell sites with inter-site distance (ISD) of $500$ m. Each cell site has $4$ outdoor small cells (IAB-nodes) installed randomly within its coverage. Each cell site also has one open-space office with dimensions $50\text{~m}\times120\text{~m}\times 3$ m that is covered by $4$ ceiling mounted indoor small cells (IAB-nodes) with ISD of $20$ m. There are $N$ UEs randomly dropped within each macrocell, of which almost half are indoor. Each BS has uniform rectangular antenna arrays with $256$ elements at both transmitter and receiver sides, whereas each UE has only one element. The rest of the system-level evaluation parameters (as recommended by \cite{tr38874, ts38214}) are summarized in Table \ref{tab:parameters}.
\begin{table}[h]
\scriptsize
\begin{tabular}{|p{0.34\linewidth}|p{0.56\linewidth}|}
\hline
Parameters & Values \\\hline
Carrier frequency & $30$ GHz \\\hline
System bandwidth & $1$ GHz \\\hline
Subcarrier spacing & $15$ KHz \\\hline
Thermal noise density & $-174$ dBm/Hz \\ \hline
Pathloss model & Macro to MT/UE: 5GCM UMa \newline
Outdoor IAB-node to MT/UE: UMi Street Canyon \newline
Indoor IAB-node to MT/UE: InH\\\hline
Antenna height & Macro: $25$ m, ~Outdoor IAB-node: $10$ m, \newline
Indoor IAB-node: $3$ m, ~UE: $1.5$ m \\\hline
Transmit power & Macro: $40$ dBm, ~Outdoor IAB-node: $33$ dBm, \newline
Indoor IAB-node: $23$ dBm, ~UE: $23$ dBm \\\hline
Noise margin & BS: $7$ dB, ~UE: $10$ dB \\\hline
Outdoor-to-indoor \newline penetration loss & $50\%$ low loss model, $50\%$ high loss model\\\hline
MCS index table & based on \cite{ts38214} \\\hline
\end{tabular}
\caption{Simulation parameters for IAB evaluations.} \label{tab:parameters}
\end{table}
We consider tree topology where each MT/UE has only one parent node. A UE/MT connects to the parent node having the strongest signal-to-noise ratio (SNR) in access links. We focus on in-band backhauling and use a static time division multiplexing scheme for coordination between parent and child BSs -- in each timeslot, either odd-hop BSs (including IAB-donors) transmit and even-hop BSs receive or vice versa. Subsequently, each BS performs round-robin algorithm to serve downlink traffic to its child nodes. We assume that control signalings are instantaneous and do not occupy radio resources. Finally, Monte Carlo simulations are performed with $20$ independent runs where each one lasts for $10^6$ timeslots.
\subsection{Average and Cell-edge Throughput}
For the proposed architecture, MTs can connect to any gNB as they have a complete UE protocol stack. For the 3GPP architecture, we assume that only $N_d$ gNBs are deployed as IAB-donors. We use a full buffer traffic model where the network continuously sends downlink data for each UE.
Figure \ref{fig:throughput} compares the average and cell-edge throughputs for the two architectures. The cell-edge throughput represents the $5$th percentile of user throughputs. We observe that the proposed architecture exhibits around $2-20$ percent better average UE throughput than that of the 3GPP architecture. Further, cell-edge throughput is significantly higher for the proposed architecture. In fact, we found atleast $70$ percent improvement in cell-edge throughput compared to the $N_d=5$ case. The figure also confirms the notion that with increasing $N_d$, cell-edge capacity increases multifold.
\begin{figure}[h]
\centering
\begin{subfigure}{\linewidth}
\includegraphics[width=\linewidth]{images/throughput}
\caption{}\label{fig:throughput}
\end{subfigure}
\hfill
\begin{subfigure}{\linewidth}
\includegraphics[width=\linewidth]{images/hof}
\caption{}
\label{fig:mobility}
\end{subfigure}
\caption{Throughput and mobility robustness comparison between the proposed architecture and the 3GPP architecture having $N_d$ IAB-donors. The plots shows (a) Average and Cell-edge throughput with different UE density per macrocell and (b) Average outage rate and HFR when an MT is moving at $120$ kmph.} \label{fig:results}
\end{figure}
\subsection{Mobility Robustness}
In order to assess mobility robustness, a small cell (IAB-node) is mounted on a moving bus. We assume that the bus is $3$ m high, and its transmitting and receiving antennas have $64$ elements. The bus moves according to the random waypoint mobility model within the deployment coverage at a constant speed of $120$ kmph. We further assume that beamforming vectors between the bus relay or VMR and BSs are perfectly aligned by using position estimation techniques. We do not consider the Doppler effect in this work.
The NR handover procedure used in our simulation is as follows. The VMR periodically measures SNR from serving and neighbor cells and initiates a handover process when a target cell becomes $3$ dB better than the serving cell and maintains it for $80$ ms. The serving cell then handover the VMR to the target cell if the handover preparation and execution phases complete successfully. A handover fails, however, when an RLF occurs during the handover process. The VMR then begins RLF recovery, and it takes $100$ ms to re-establish a link. The VMR declares RLF when the SNR falls below $2$ dB and does not exceed $5$ dB within $1000$ ms. We assume handover preparation delay for inter-gNB-DU handover as $20$ ms and inter-gNB-CU handover as $40$ ms, and handover execution delay for inter-gNB-DU and inter-gNB-CU handover as $25$ and $50$ ms, respectively.
We demonstrate VMR's mobility robustness through outage rate (i.e., the percentage of time it was in an outage) and handover failure ratio or HFR (i.e., the percentage of handovers that failed). The VMR is in an outage when performing RLF recovery, or SNR of its active connection is below $2$ dB. Figure \ref{fig:mobility} demonstrates that the VMR experiences no outage conditions and negligible HFR in the proposed architecture. In contrast, the 3GPP architecture indicates a much higher outage rate and HFR values. Like before, when there are fewer IAB-donors, the VMR's mobility robustness may degrade to unacceptable levels.
We emphasize that we obtained these results with simplified modeling. In reality, IAB-nodes would be deployed with careful network planning. Optimal cell selection and resource scheduling (like \cite{saha2019millimeter}) can provide better gains than presented above. Similarly, optimizing mobility-related parameters would improve mobility robustness further.
\section{Conclusion and Future Trends}
The 3GPP has standardized IAB in Release 16 as a promising cost-effective solution for ultra-densification in mmWave 5G networks. IAB enables multihop relaying, which makes deploying 5G networks in areas with partial wired backhaul or 5GC connectivity in a scalable and quick manner. The 3GPP IAB solution, however, suffers from some inherent limitations that may hinder making it flexible and self-organizing architecture by design. This article has proposed a novel alternative for B5G networks that mitigates these shortcomings. Our architecture supports VMR and multi-RAT coexistence (e.g., access and backhaul may belong to different RATs) that are not currently enabled by the existing 3GPP architecture. Our architecture does not modify any standard network element and is transparent to legacy operations of UEs, gNB, and 5GC. In fact, we have shown that our architecture fulfills almost every architectural design aspect related to IAB networks. Due to its robust design, the architecture can be rapidly standardized and deployed within the existing 5G system. Simulations have shown that the proposed architecture possesses significant gains in cell-edge throughput over that of the 3GPP architecture. Mobility robustness in terms of outage rate and HFR is also superior in the proposed architecture, especially when fewer IAB-donors are available. We believe these capabilities would give rise to a new paradigm of use-cases in the future and makes a case for standardization for B5G network.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Transformer-based architecture has achieved tremendous success in the field of natural language processing (NLP) \cite{vaswani2017attention} \cite{devlin2018bert}. Inspired by the great success of transformer in the language domain, vision transformer\cite{dosovitskiy2020image} has been proposed and achieved superior performance on the ImageNet dataset. The vision transformer splits the image into patches and feeds into the transformer, the same way as words token in NLP, and passes through several multi-head self-attention layers of the transformer to establish the long-range dependencies.
\begin{figure}[!t]
\centerline{\includegraphics[width=0.9\linewidth]{chart}}
\caption{Graph of accuracy vs. number of parameters for various local transformer-based models. It shows that our all versions of the model: MOA-T, MOA-S, and MOA-B have higher accuracy and comparatively fewer number of parameters.}
\label{fig:fig1}
\end{figure}
Unlike the word token, a high-resolution image contains more pixels compared to words in the passage. This leads to an increase in the computation cost as self-attention in the transformer has quadratic complexity. To alleviate this problem, various local attention-based transformers \cite{liu2021swin} \cite{vaswani2021scaling} \cite{zhang2021nested} have been proposed with a linear computation complexity. However, all the proposed approaches could not establish long-range dependencies and some of them are very complicated.
To overcome these issues in the local transformers, we develop a very simple module, named multi-resolution overlapped attention(MOA), to generate global features. The proposed module only consists of multiplication and addition operations and is embedded after each stage in the transformer before the downsampling operation. As the module is added only after each stage instead of each transformer layer, it does not add much computation cost and the number of parameters. Our experiments show that aggregating the resultant features of this module to the local transformer establish the long-range dependencies and hence significantly increases the accuracy in contrast to the total number of parameters as shown in Figure~\ref{fig:fig1}
\begin{figure*}
\centerline{\includegraphics[width=0.85\linewidth]{Blockpaper}}
\caption{The architecture of the MOA-T is composed of three stages. Each stage consists of a patch embedding/merging layer and local transformer block along with a global multi-resolution overlapped attention module after each stage except the last stage. In the beginning patch, a partition layer is included to divide the image into a fixed number of patches. }
\label{fig:fig2}
\end{figure*}
Our proposed MOA module takes the output generated by the group of local window-based attention as an input. It first converts it to a 2D feature map, and projects it to a new low-dimension feature map. Similar to ViT \cite{dosovitskiy2020image}, the projected feature map is divided into a fixed number of patches except for a few modifications. In contrast to ViT \cite{dosovitskiy2020image}, the patch sizes of query and key-value are different. The resolution of the patches in the query is the same as the window size used in the local transformer layer. In contrast, the resolution of patches in key-value is slightly larger than the query patch and overlapped. The hidden dimension of the MOA global attention module is kept the same as the previous transformer layer. Therefore, the resultant features are directly aggregated to the output of the previous transformer layer.
Extensive experiments show that keeping the key-value patches slightly larger with overlap to each other leads to significant performance gain due to small information exchange between two neighborhood windows. In short, our method exploits the neighborhood information along with global information exchange between all non-local windows by embedding the proposed MOA mechanism in the local transformer.
The contributions of the proposed approach are summarized as below:
\begin{enumerate}
\item We propose a multi-resolution overlapped attention (MOA) module that can be plugged in after each stage in the local Transformer to promote information communication along with nearby windows and all non-local windows.
\item We thoroughly study the impact of global information in local transformer using the proposed MOA module.
\item We investigate the effect of the dimension of essential architecture components through extensive experiments and discover the optimum architecture for image classification.
\item We train the proposed model from scratch on CIFAR-10/CIFAR-100 \cite{cifar} and ImageNet-1K \cite{deng2009imagenet} datasets and achieve state-of-the-art accuracy using a local transformer.
\end{enumerate}
\section{Related Work}
\subsection{Convolutional Neural Networks}
After the revolutionary invention of AlexNet \cite{krizhevsky2012imagenet}, convolutional neural network (CNN) has become a standard network for all computer vision tasks, such as image classification \cite{ma2022semantic}\cite{patel2022discriminative}, object detection \cite{li2021colonoscopy}, tracking \cite{zhang2020efficient}, segmentation \cite{he2021sosd}\cite{patel2021enhanced}, counting \cite{sajid2021audio}, and image generation \cite{xu2021domain}. Various versions of CNNs have been proposed to improve the performance by making it deeper and/or broader, such as VGG network \cite{simonyan2014very}, ResNet \cite{he2016deep}, Wide-ResNet \cite{zagoruyko2016wide}, DenseNet \cite{huang2017densely}, etc. There are also several works proposed to make it more efficient by modifying the individual convolutional layer, such as dilated convolution\cite{yu2015multi}, depth-wise separable convolution \cite{chollet2017xception}, group convolution \cite{krizhevsky2012imagenet}, etc. In our work, we employ the convolutional layer along with the transformer layer to reduce the overall dimension of the feature map. Our experiments show that the combination of convolutions and multi-head attention increases the performance.
\subsection{Self Attention in CNN}
Self-attention mechanisms have become ubiquitous in the field of computer vision tasks. Various works \cite{gajurel2021fine}\cite{wang2018non}\cite{cao2019gcnet}\cite{woo2018cbam}\cite{sajid2021towards}\cite{fu2019dual}\cite{zhao2018psanet}\cite{ma2021miti} have been proposed that utilize either channel-based or position based self-attention layers to augment the convolution network. Non-local network\cite{wang2018non} and PSANet\cite{zhao2018psanet} model the spatial relationship between all the pixels in the feature map and are embedded the attention module after each block in CNN, whereas SENet \cite{hu2018squeeze} establishes a channel relationship in the convolution network by squeezing the features using global average pooling. CBAM \cite{woo2018cbam}, BAM \cite{park2018bam} and dual attention network \cite{fu2019dual} employ both channel and position based attention mechanisms separately, then combine the resultant features from both attention modules using either element-wise addition or concatenation and uses the resultant features into convolution output after each stage, whereas GCNet \cite{cao2019gcnet} combines SENet \cite{hu2018squeeze} and non-local network \cite{wang2018non} together and propose the hybrid attention mechanism that aggregates the information of both channel and spatial relationships in the same attention module.
\subsection{Vision Transformers}
Similar to AlexNet, vision Transformer (ViT) \cite{dosovitskiy2020image} has changed the perspective of researchers towards solving computer vision problems. Since then, many vision transformer-based networks have been proposed to improve accuracy or efficiency. The ViT needs to be pre-trained on large datasets such as JFT300M \cite{sun2017revisiting} to achieve high performance. DeiT \cite{touvron2021training} solves this problem by student-teacher setup, substantial augmentation, and regularization techniques. To train the transformer on the mid-sized dataset like ImageNet-1K from scratch, the token-to-token vision transformer \cite{yuan2021tokens} recursively aggregate neighboring tokens (patches) into one token (patch) to reduce the number of tokens. A Cross-ViT \cite{chen2021crossvit} comes up with a dual branch approach with multi-scale patch size to produce robust image features and pyramid vision Transformer (PVT) \cite{wang2021pyramid} introduces a multi-scale-based spatial dimension design similar to FPN \cite{lin2017feature} in CNN and demonstrated good performance. Furthermore, PVT introduced a spatial reduction in key to reduce the computation cost in multi-head attention.
Various local attention-based transformers have been introduced to alleviate the quadratic complexity issues \cite{vaswani2021scaling}\cite{liu2021swin}\cite{zhang2021nested}. The HaloNet \cite{vaswani2021scaling} introduces the idea of a slightly larger window of key than the query in a local attention mechanism and proves its effectiveness through various experiments. In our model, the key is also calculated using a slightly larger patch, but in the context of global attention, the idea of a larger key is different from the HaloNet. A
swin Transformer \cite{liu2021swin} proposes a non-overlapping window-based local self-attention mechanism to avoid quadratic complexity and achieve comparable performance and aggregated nested Transformer \cite{zhang2021nested} come with the multi-scale approach with block-aggregation mechanism after each stage.
Some Transformer-based works have been proposed to utilize both local and global features \cite{han2021transformer} \cite{chu2021twins}.
A Transformer in Transformer (TNT) \cite{han2021transformer} further divides the local patches (visual sentences) into smaller patches (visual words). The MHA on visual words embedding is calculated and aggregated to the sentence embedding to establish the global relationship. The twin Transformer \cite{chu2021twins} is quite the same as ours. However, global attention is applied after each local Transformer layer, increasing the computation cost significantly. In contrast, we apply it after each stage, and we have slightly larger and overlapped patches in key in multi-head attention. The proposed network efficiently utilizes global information in the local transformer and achieves higher accuracy than the above-mentioned transformer-based models.
\begin{figure}
\centerline{\includegraphics[width=1.0\linewidth]{updatedpatch.png}}
\caption{Patch creation for query embedding is shown in the blue, and key/value is shown in the red for feature map size $9 \times 9$ and window size $3 \times 3$. Blue patches have the same size as the window and are non-overlapped to each other. In contrast, red patches are larger and slightly overlapped with each other. Appropriate padding is applied while creating the key-value patches.}
\label{fig:fig3}
\end{figure}
\section{Proposed Method}
We aim to provide global information exchange across all windows in the local transformer by increasing the minimal computation cost and a number of parameters. An overview of our proposed model is shown in Figure~\ref{fig:fig2}, which shows MOA module after each stage. All stages have a similar architecture design, including patch merging layer and local transformer block except the first stage. The first stage consists of patch partition, linear embedding layer, and local transformer block. Our global MOA module is applied between each stage before the patch merging layer.
Specifically, the model takes an RGB image as an input and splits it into fix number of patches. Here each patch is treated as a token. In our experiment on the ImageNet dataset, we set the patch size to $4\times4$, which leads to $4\times4\times3 = 48$ feature dimensions for each patch. These row features are projected to a specific dimension C using the patch embedding layer in the first stage. The resultant features are then passed through consecutive stages consisting of patch merging layer, local transformer block, and MOA module in-between each stage. Unlike swin Transformer \cite{liu2021swin}, our Transformer block employs the same self-attention mechanism as ViT \cite{dosovitskiy2020image} without any shifted window approach. Similar to swin Transformer, the number of tokens is reduced, and the output dimension is doubled in the patch merging layer after each stage. For example, the resolution after the first, second and third stage is $\frac{H}{2} \times \frac{W}{2}$, $\frac{H}{4} \times \frac{W}{4}$, and $\frac{H}{8} \times \frac{W}{8}$, respectively. The average pooling layer is inserted at the end of the last stage, followed by a linear layer to generate a classification score. The detailed explanation of each element of architecture are as follows:
\subsection{Patch embedding layer}
It is a basic linear embedding layer applied to the row features of patches to project it to a specific dimension C.
\subsection{Patch merging layer}
Patch merging layer reduces the number of tokens by concatenating the features of $2 \times 2$ neighboring patches and doubles the number of hidden dimensions by applying a linear layer on the concatenated 4C - dimensional features.
\subsection{Local Transformer Block}
The local transformer block consists of a local window-based standard multi-head attention module, followed by a two-layer MLP with GELU non-linearity. A layer norm is used before each multi-head attention module and each MLP with residual connection after each module.
\subsection{Multi-resolution Overlapped Attention Block}
To utilize the advantage of global information in local transformer, we apply a global attention module named multi-resolution overlapped attention (MOA) in-between each stage. The architecture of the
MOA mechanism is the same as the standard multi-head attention except for a few modifications. Similar to standard MHA, it first divides the feature map into the fixed size of patches. However, unlike the standard MHA, patches for generating key and value embeddings are slightly larger and overlapped, while the patches for query embedding are non-overlapped as shown in Figure~\ref{fig:fig3}.
As shown in the Figure~\ref{fig:fig3}, the input to MOA block is of size $W \times H \times$ hidden dim, Where $W = \frac{W}{2}$, $\frac{W}{4}$ or $\frac{W}{8}$, $H = \frac{H}{2}$, $\frac{H}{4}$ or $\frac{H}{8}$, and hidden dim = 96, 192, or 384. Calculating query, key, and value embeddings directly from the input is quite expensive in computation. For example, in context to the ImageNet dataset, the feature map size of the input to MOA block after the first stage is $56 \times 56 \times 96$. Deriving query embedding directly from the input feature with a patch size 14 will lead to the resultant feature of dimension $14 \times 14 \times 96 = 18816$. Therefore, we first reduce the hidden dimension with factor R by applying $1 \times 1$ convolution, which reduces the computation cost. The resultant feature dimension after applying the convolution is $ H \times W \times \frac{hidden dim}{R}$. This leads to feature size in one query patch is $14 \times 14 \times \frac {hidden dim}{R}$, which is projected to the one-dimensional vector of size: $1 \times 1 \times hiddendim$. The total number of the query is $\frac{H}{14} \times \frac{W}{14}$. Similarly, the key and value vector are projected, but the patch size is slightly larger than the query as shown in Figure~\ref{fig:fig3}. In our model, we set the key-value patch size to 16. Therefore, the number of key-value will be according to the equation: $ (\frac{ H - 16 + (2 \times padding)}{stride} + 1 , \frac{ W - 16 + (2 \times padding)}{stride} + 1 )$. Multi-head attention is applied to this query, key, and value embedding, followed by two-layer MLP with GELU non-linearity in between. Similar to the Transformer block, layer norm is applied along with residual connection after each MOA module. At last, on the resultant features, $1 \times 1$ convolution is applied, followed by broadcast addition of resultant features with the output of the previous transformer block, which contains the local information.
\subsection{Relative Position Index}
We use relative position bias $B \in R ^{M^{2} \times N^{2}}$, as used by \cite{bao2020unilmv2} \cite{hu2019local} \cite{hu2018relation} \cite{raffel2019exploring} , in the heads of both local and global attentions during similarity computation:
\begin{equation}
Attention(Q, K,V) = Softmax(\frac{QK^{T}}{\sqrt{d}} + B)V
\end{equation}
where $Q \in R^{M^{2} \times d}$ is a query matrix, $K, V \in R^{N^{2} \times d}$ are the key and value matrices; d is the hidden dimension, $M^{2}$ is the total number of patches in the queue and $N^{2}$ is total the number of patches in the key.
\subsection{Architecture Detail}
By following the previous works\cite{liu2021swin}\cite{chu2021twins}, we build three versions of the model: MOA-T, MOA-S, and MOA-B for the ImageNet dataset and only two versions of the model: MOA-T and MOA-B for the CIFAR -10/100 dataset as it is quite smaller. Table ~\ref{tab: tab1} shows the architecture configurations for the CIFAR and the ImageNet datasets. In the CIFAR based models, both MOA-T and MOA-B contain the same number of Transformer layers: 12, but have a different number of hidden dimensions. In context to the ImageNet based models, the total number of layers for MOA-T and MOA-S is 12 and 24 respectively, but the hidden dimension is kept the same, whereas MOA-S and MOA-B have the same number of Transformer layers: 24, with contrast hidden dimensions 96 and 124 respectively.
\begin{table*}
\caption{ Model configuration for CIFAR/ImageNet dataset}
\begin{center}
\begin{tabular}{ c c c c c c c c }
\hline
Model & Dataset & Input-Size & Window-Size& No. of Layers& No. of Heads&Hidden Dim& Patch -Size \\
\hline
MOA-T & CIFAR &$32 \times 32$ & $4 \times 4$ & [2, 2, 6, 2] & [3, 6, 12, 24]& [96, 192, 384, 768]& 1 \\
MOA-B & CIFAR & $32 \times 32$ & $4 \times 4$ & [2, 2, 6, 2] & [4, 8, 16, 32]& [128, 256, 512, 1024]& 1 \\
MOA-T & ImageNet & $224 \times 224$ & $14 \times 14$ & [2, 2, 8] & [3, 6, 12]& [96, 192, 384]& 4 \\
MOA-S & ImageNet & $224 \times 224$ & $14 \times 14$ & [2, 2, 20] & [3, 6, 12]& [96, 192, 384]& 4 \\
MOA-B & ImageNet & $224 \times 224$ & $14 \times 14$ & [2, 2, 20] & [4, 8, 16]& [128, 256, 512]& 4 \\
\hline
\label{tab: tab1}
\end{tabular}
\end{center}
\end{table*}
\section{Experimental Evaluations}
We verify our model through extensive experiments on CIFAR-10/CIFAR-100 and ImageNet-1K for image classification. We design three architecture versions: MOA-T, MOA-S, and MOA-B, for the classification tasks.
\subsection{CIFAR-10/100 Results}
CIFAR-10 and CIFAR-100 datasets consist of 50,000 training and 10,000 test images of resolution $32 \times 32$ with the total number of classes 10 and 100, respectively. We train the network for 300 epochs using AdamW \cite{kingma2014adam} optimizer with an initial learning rate of 0.009 and weight decay of 0.05. We utilize a cosine decay learning rate schedular along with 20 warm-up epochs. We implemented two models: MOA-T and MOA-B for the CIFAR dataset with total batch-size 128 and stochastic drop-rate 0.2 \cite{larsson2016fractalnet}.
Table ~\ref{tab: tab3} shows the performance of our model on the CIFAR-10 and CIFAR-100 datasets. We presented only two models with the same number of layers but with different hidden dimensions for this dataset. As shown in the table, it can be seen that both models outperform all the previous Transformer-based models by a significant amount. It improves the performance by 0.59\% and 0.98\% on CIFAR-10 and 0.56\% and 0.23\% on CIFAR-100 for the Tiny and Base models, respectively, compared to Swin Transformer. For the Base model, our model achieves state-of-the-art accuracy on local vision Transformer with a comparatively fewer number of parameters and GFLOPs. The accuracy of other models is reported by training the models from scratch with the same training setting reported in the papers \cite{liu2021swin} \cite{touvron2021training} \cite{wang2021pyramid}.
\begin{table}[!htb]
\caption{Results on CIFAR - 10/100 }
\begin{center}
\begin{tabular}{ c c c c}
\hline
Model & CIFAR-100(\%) & CIFAR-10(\%)& Parameters \\
\hline
Deit-T&70.33&89.2&5M \\
PVT-T&72.80&91&13M \\
Swin-T& 78.07 & 94.41 & 27.5M \\
MOA-T& \textbf{78.63} &\textbf{95}& 30M \\
\hline
DeiT-B&71.54&93&85M \\
PVT-B&70.1 &89.87&61M \\
Swin-B&78.45&94.47&86.7M \\
MOA-B&\textbf{78.68}&\textbf{95.05}& 53M \\
\hline
\label{tab: tab3}
\end{tabular}
\end{center}
\end{table}
\subsection{ImageNet Results}
ImageNet-1K dataset consists of around 1.28M training images and 50K validation images with 1000 classes. We resize all the images to the resolution $224 \times 224$ during training. We follow the same training technique, like Swin and Twin, and train the network for 300 epochs using AdamW \cite{kingma2014adam} optimizer with a cosine learning rate schedular and 20 warmup epochs. We keep the batch-size 128 for MOA-T and 64 for MOA-S and MOA-B models per GPU. We employ a total of four GPUs together during training leads to a total batch-size 512 for MOA-T and 256 for MOA-S and MOA-B models. We utilize the same augmentation technique used by \cite{liu2021swin} such as a mixture of cutmix \cite{yun2019cutmix} and mixup \cite{zhang2017mixup} and regularization technique stochastic drop rate. We set the drop rate \cite{larsson2016fractalnet} of 0.2, 0.3, and 0.5 respectively for MOA-T, MOA-S, and MOA-B.
Table ~\ref{tab: tab4} shows our model's result and a similar Transformer-based model on the ImageNet-1K classification task. Our proposed models: MOA-T, MOA-S, and MOA-B, achieve higher accuracy than most of the Transformer-based models with significant parameter reduction. MOA-T outperforms Twin-S and Swin-T by 0.34\% with around 22\% fewer parameters. Our MOA-S improves the performance by 0.5\% and 0.3\% compared to Swin-S and Twin-M respectively, even with the lower batch size during training. Our MOA-B achieves the state-of-the-art accuracy of 83.7\% on ImageNet-1K with comparatively fewer parameters with a smaller batch size than the remaining vision transformers. Our model increases the computation cost by a negligible amount, but the performance improvement and parameter reduction are highly rewardable.
\begin{table}[t]
\caption{Results on ImageNet-1K}
\begin{center}
\begin{tabular}{ c c c c}
\hline
Model & Accuracy(\%) & Parameters & GFLOPs \\
\hline
Deit-Small/16 & 79.9 & 22.1M & 4.6\\
CrossViT-S&81.0&26.7M&5.6\\
T2T-ViT-14&81.5&22M&5.2\\
PVT-Small&79.8&24.5M&3.8\\
TNT-T&73.9&6.1M&1.4\\
Twins-PCPVT-S&81.2&24.1M&3.8\\
Swin-T&81.3&29M&4.5\\
Twins-SVT-S&81.7&24M&2.9\\
MOA-T&\textbf{82.05}&17M&4.8\\
\hline
T2T-ViT-19&81.9&39.2M&8.9\\
PVT-Medium&81.2&44.2M&6.7\\
TNT-S&81.5&23.8M&5.2\\
Twins-PCPVT-B&82.7&43.8&6.7\\
Swin-S&83.0&50M&8.7\\
Twins-SVT-B&83.2&56M&8.6\\
MOA-S&\textbf{83.5}&39M&9.4\\
\hline
ViT-Base/16 &77.9&86.6M&17.6\\
Deit-Base/16&81.8&86.6M&17.6\\
T2T-ViT-24&82.3&64.1M&14.1\\
CrossViT-B&82.2&104.7M&21.2\\
PVT-Large&81.7&61.4M&9.8\\
TNT-B&82.9&65.6M&14.1\\
Swin-B&83.3&15.4M&83.7\\
Twins-SVT-L&\textbf{83.7}&99.2M&15.1\\
MOA-B&\textbf{83.7}&68M&16.2\\
\hline
\label{tab: tab4}
\end{tabular}
\end{center}
\end{table}
\section{Ablation Study}
In this section, we conduct ablation experiments to understand the effect of the dimension of each component, such as window size, the overlapped area between the key-value patches, and the reduction factor in global attention, in our model. We employ the Tiny model to perform all ablation experiments, and all the experiments are performed either on CIFAR-100 or ImageNet dataset. The training configurations remain the same as reported in the experiment section.
\subsection{Window-size}
The sequence length of the local-Transformer is one of the essential factors on which computation cost relies. As the sequence length increases, computation cost in the self-attention mechanism increases as well. In a local vision Transformer, sequence length depends on the window size. There is always a trade-off between the accuracy and computation cost based on the sequence length. We perform experiments with various window sizes in our model and find that $4 \times 4$ and $14 \times 14 $ window size works well on CIFAR-100 and ImageNet datasets, respectively, as shown in Table ~\ref{tab: tab5}. Furthermore, we remove the stages where the window size is greater than the feature map size to significantly reduce the number of parameters.
\begin{table}[t]
\caption{Results with different window-size on ImageNet }
\begin{center}
\begin{tabular}{ c c c c c}
\hline
Window-Size & Dataset & No. of Stage & Accuracy & Parameters \\
\hline
$2 \times 2$ & CIFAR -100& 4 & 76.04 & 29.7M\\
$4 \times 4$ & CIFAR-100& 4 & 78.61 & 30M \\
$8 \times 8$ & CIFAR-100& 3 & 76.02 & 16M\\
$7 \times 7$ &ImageNet& 4 & 81.4 & 31M\\
$14 \times 14$ & ImageNet&3 & 82.07 & 17M \\
$28 \times 28$ & ImageNet&2 & 78.2 & 6M\\
\hline
\label{tab: tab5}
\end{tabular}
\end{center}
\end{table}
\subsection{Overlapped Portion}
To initiate the neighborhood information transmission, we propose to use slightly larger and overlapped keys. To investigate the effect of the portion of the overlapped area, we perform experiments with different percentages of overlapped portions in keys as shown in Table ~\ref{tab: tab6}. It can be seen from the results that the performance is increased in terms of accuracy as the percentage decreases, which means only slight information exchange between the neighborhood windows are required to improve the performance. Furthermore, fewer overlapped portions decrease the sequence length, which reduces the number of parameters and GFLOPs.
\begin{table}[t]
\caption{Results on CIFAR-100 with different percentage of the overlapped portion}
\begin{center}
\begin{tabular}{ c c c}
\hline
\% Overlap &Accuracy & Parameters \\
\hline
~17\% & 78.63 & 30.05M\\
~33\% & 78.52 & 30.06M\\
~50\% & 78.38 & 30.08M\\
~66\% & 78.38 & 30.59M\\
\hline
\label{tab: tab6}
\end{tabular}
\end{center}
\end{table}
\subsection{Reduction}
Before the MOA global attention, the hidden dimension is reduced to decrease the number of parameters and computation cost. Table ~\ref{tab: tab7} shows the performance of our model with various values of R. From the result, it is evident that R = 32 achieves the best result with comparatively less number of parameters and computation cost than a smaller value of R.
\begin{table}[t]
\caption{Results with different window-size on CIFAR-100}
\begin{center}
\begin{tabular}{ c c c}
\hline
Reduction &Accuracy & Parameters \\
\hline
8 & 78.38 & 31.67M\\
16 & 78.34 & 30.59M\\
32 & 78.63 & 30.06M\\
64 & 78.51& 29.78M\\
num-heads&78.41&31.43M\\
\hline
\label{tab: tab7}
\end{tabular}
\end{center}
\end{table}
\begin{table}[!t]
\caption{Significance of global attention and overlapped patches}
\begin{center}
\begin{tabular}{ c c c}
\hline
Model&Accuracy & Parameters \\
\hline
Without Global & 75.56 & 27M \\
With Global (ViT) & 78.34 & 30.59M\\
With Global (Ours) & 78.63 & 30.06M\\
\hline
\label{tab: tab8}
\end{tabular}
\end{center}
\end{table}
\subsection{Effect of Overlapped Key-Value}
To verify the effect of overlapped and larger key-value patches, we train the model without overlapping patches and compare the results. Furthermore, we also conduct an experiment without applying global attention in-between each stage to verify the significance of global information exchange. From the result in Table ~\ref{tab: tab8}, it can be seen that including global attention and overlapped key-value patches achieve the best performance.
\section{Conclusion}
The paper has investigated the effect of aggregating global information in local Transformer after each stage and neighborhood pixel information transmission. We have also proposed a multi-resolution overlapped attention (MOA) module that can be plugged in after each stage in the local transformer to promote information communication along with nearby windows. Our results show that both types of features: global and local, are crucial for image classification. As a result, exploiting both features leads to significant performance gain on the standard classification datasets such as CIFAR10/100 and the ImageNet with comparatively fewer parameters.
\balance
{\small
\bibliographystyle{IEEEtranS}
|
\section{Introduction}
\label{sec:introduction}
A central challenge in natural language processing is the design of models and learning algorithms that are simultaneously \emph{flexible} enough to capture the variability of human language and \emph{structured} enough to generalize in predictable and human-like ways.
One important source of structure is the \textbf{principle of compositionality}, which (in one formulation)
states that sentence meanings can be computed from a \emph{lexicon} of word meanings and a set of \emph{composition rules} governing how meanings combine \cite{montague1970universal}.
A long line of language processing research has operationalized the principle of compositionality as a \textbf{constraint on model architectures}, via independence assumptions (e.g.\ in synchronous grammars; \citealp{lewis1968syntax}) or parameter tying schemes (e.g.\ in modular neural networks; \citealp{socher2013recursive,andreas2016neural}) that ensure a compositional process of sentence interpretation.
Compositional models enjoy sample-efficient learning and strong forms of out-of-distribution generalization in tasks ranging from machine translation to question answering \cite{mccoy2020does}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\linewidth]{figures/teaser_vqa.pdf}
\caption{Example of our approach to compositional modeling in the visual question answering domain. Given a dataset of (image, question, answer) triples, we extract a lexicon that relates words to their visual groundings. We then find \emph{homomorphic transformations} (\cref{sec:theory}) of this lexicon that, when applied to training examples, produce new, well-formed examples.} %
\label{fig:teaser}
\end{figure}
But much of human language is not (or at least not straightforwardly) compositional. Idioms, disfluencies, and context-sensitive meanings present major challenges to models in which all predictions must be derivable from a sequence of local composition operations.
In recent years, more generic model architectures such as recurrent neural networks (RNNs) and transformers, with no explicit compositional scaffolding,
have consistently outperformed compositional models in language processing tasks with natural data distributions \cite{googlemt}.
At the same time, these models capture linguistic regularities only when trained on enormous amounts of data, and make surprising or problematic predictions when presented with novel word collocations or syntactic structures \cite{lake2018generalization}.
How can we train unstructured neural sequence models that generalize compositionally?
In this paper, we describe how to inject compositional inductive bias into general function approximators by operationalizing compositionality as a \textbf{constraint on data distributions} rather than model architectures.
\cref{fig:teaser} gives the intuition behind our approach: in a visual domain in which the compatibility of images and captions can be computed using a lexicon that relates words like \emph{yellow} and \emph{green} to their perceptual groundings, we show how to transform a structural regularity of the lexicon (the fact that the words \emph{yellow} and \emph{green} are of the same type and play the same role in downstream computation) into a structural regularity of the dataset itself (in any well-formed image--caption pair, swapping the words \emph{yellow} and \emph{green} and their associated colors yields a new pair that is still well-formed).
More formally, we prove that for any dataset in which the well-formedness of examples can be computed compositionally (i.e.\ factored into a lexicon and a composition function), every \emph{homomorphism} of the lexicon
corresponds to a \emph{symmetry} of the dataset.
Using this characterization, we describe a procedure that uses statistical alignment algorithms to build lexicons for problems of interest, then uses these lexicons to construct data augmentation procedures \emph{guaranteed} to produce semantically acceptable examples.
An important benefit of this data-centric approach to compositionality is that it does not require a complete compositional theory for a given problem domain, or even a lexicon entry for every word:
partial information about a problem's lexicon suffices to derive nontrivial data transformations.
Applied to transformers and RNNs,
this approach to data augmentation improves performance on tests of compositional generalization in two problem domains: the CLEVR CoGenT visual question answering dataset, and the COGS semantic parsing dataset. %
This paper thus offers two contributions: first, a new lens on the principle of compositionality via symmetries of data distributions; second, a concrete instantiation of this framework as a data augmentation scheme that improves model performance on multiple tasks. Our results highlight the fact that machine learning with compositional inductive bias need not require new models: there is a natural ``data-centric'' view of the principle of compositionality
that may be more robust and broadly applicable than the traditional model-centric one.\footnote{Code for the experiments is released under \url{https://github.com/ekinakyurek/lexsym}}
\section{Background \& Approach}
We begin with a discussion on the more general role of \emph{symmetry} in machine learning applications.
\begin{definition}\label{def:symmetry}
A \textbf{symmetry} of a set $X$ is a function $f$ satisfying:
\begin{equation}
\{ f(x) : x \in X \} = X
\end{equation}
That is, applying $f$ to each element of $X$ leaves $X$ unchanged.
\end{definition}
A familiar example from computer vision is \emph{reflection symmetry}: in object recognition problems, image classes are generally invariant under reflection (a zebra seen in a mirror is still a zebra). The set of (image, class) pairs thus has as a symmetry the function $(x, y) \mapsto (\texttt{reflect}(x), y)$.
In many problems, especially those (like computer vision and computational chemistry) that are constrained by physical laws, knowledge of the symmetries exhibited by a problem domain can dramatically reduce the difficulty of learning in that domain.
Past work has incorporated symmetry into machine learning problems in two ways. \textbf{Invariant and equivariant modeling} approaches structurally enforce symmetries via specialized architectures (improving generalization by decreasing the size of the hypothesis class; \citealp{cohen2016group}). \textbf{Data augmentation} approaches generate new training examples by applying known symmetries like reflections directly to training data (improving generalization by increasing dataset size; \citealp{shorten2019survey}). Data augmentation, the focus of this paper, is model-agnostic, and can be used in conjunction with pretraining (even on non-symmetric domains) while producing the same asymptotic effects as specialized model architectures \cite{chen2020group}.
The question this paper aims to answer is whether compositionality, like other domain-specific constraints, can be formalized in the language of symmetry.
We are not the first to consider this question: \citet{kiddon2015symmetry} define a theory of semantic equivalence in terms of symmetries of the set of natural language sentences, and \citet{Gordon2020Permutation} propose a model architecture for compositional semantic parsing via a symmetry that enforces \emph{permutation invariance} of lexical items. As described in \cref{sec:theory}, our approach also derives symmetries from lexicons. It builds on the work of \citeauthor{Gordon2020Permutation} by (1) characterizing the relationship between compositionality and symmetry for general problem domains, (2) discovering lexicons and symmetries automatically, and (3) showing how to leverage them in a model-agnostic way via data augmentation.
Additional related work is discussed in \cref{sec:otherrelated}.
\section{Compositionality as Lexical Symmetry}
\label{sec:theory}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/tasks.pdf}
\caption{Idealized compositional semantic parser following \cref{def:abstraction}. A (sentence, logical form) pair is translated into a \emph{lexical abstraction} containing information about each token's type and semantic equivalences. We then determine whether the sentence evaluates to the logical form using \emph{only} the type and equivalence matrices, using types to assign the sentence an abstract logical form, and equivalences to determine whether it matches the target.
}
\label{fig:semantic-parser}
\end{figure}
In this section, we formalize the relationship between \emph{compositional sentence interpretation} and \emph{symmetries of datasets}.
We consider problems defined by a space of possible examples $\mathcal{X}$, of which a subset of examples $X$ are \textbf{well-formed}.
We assume each example $\mathbf{x} \in \mathcal{X}$ is a discrete sequence $[x_1, \ldots, x_n]$, with $x_i$ drawn from a vocabulary $\Sigma$.
Finally, we assume that well-formedness can be computed by a \textbf{interpretation function} $\mathcal{I} : \mathcal{X} \to \{0, 1\}$ with $\mathcal{I}(\mathbf{x}) = 1$ iff $\mathbf{x} \in X$.
\newcommand{\inp_\textrm{NL}}{\mathbf{x}_\textrm{NL}}
\newcommand{\inp_\textrm{LF}}{\mathbf{x}_\textrm{LF}}
\begin{example}{1(a)}{Semantic Parsing}
Examples $\mathbf{x}$ are pairs $(\inp_\textrm{NL}, \inp_\textrm{LF})$, where $\inp_\textrm{NL}$ is an sentence, $\inp_\textrm{LF}$ is a logical form, and $\mathcal{I}(\inp_\textrm{NL}, \inp_\textrm{LF}) = 1$ iff $\inp_\textrm{LF}$ represents a possible meaning of $\inp_\textrm{NL}$.
\end{example}
\begin{example}{1(b)}{Paraphrase Detection}
Examples $\mathbf{x}$ are sentence pairs $(\mathbf{x}', \mathbf{x}'')$, with $\mathcal{I}(\mathbf{x}', \mathbf{x}'') = 1$ iff $\mathbf{x}'$ and $\mathbf{x}''$ have the same meaning.
\end{example}
\begin{example}{1(c)}{Visual Question Answering}
\newcommand{\inp_\textrm{Q}}{\mathbf{x}_\textrm{Q}}
\newcommand{\inp_\textrm{I}}{\mathbf{x}_\textrm{I}}
\newcommand{\inp_\textrm{A}}{\mathbf{x}_\textrm{A}}
Examples $\mathbf{x}$ are triples $(\inp_\textrm{Q}, \inp_\textrm{I}, \inp_\textrm{A})$, where $\inp_\textrm{Q}$ is a question, $\inp_\textrm{I}$ is a (discretized) image, $\inp_\textrm{A}$ is an answer, and $\mathcal{I}(\inp_\textrm{Q}, \inp_\textrm{I}, \inp_\textrm{A}) = 1$ iff $\inp_\textrm{A}$ is the answer to $\inp_\textrm{Q}$ in $\inp_\textrm{I}$.
\end{example}
We aim to prove that \emph{if $\mathcal{I}$ can be computed compositionally, $X$ exhibits identifiable symmetries in the sense of \cref{def:symmetry}}.
We begin by formalizing what it means for $\mathcal{I}$ to be compositional.
\begin{definition}
A \textbf{lexicon} $\mathcal{L}$ is a pair $(\tau, \epsilon)$ where $\tau : \Sigma \to \mathcal{T}$ assigns each symbol a \textbf{type} from some set $\mathcal{T}$, and $\epsilon : \Sigma \times \Sigma \to \{ 0, 1 \}$ is a relation indicating which symbols are \textbf{semantically equivalent}.
\end{definition}
\begin{definition}\label{def:abstraction}
Denote the \textbf{lexical abstraction} $\mathcal{L}(\mathbf{x}) = (\typefn(\inp), \equivfn(\inp))$ where $\typefn(\inp)$ is a vector with $\typefn(\inp)_i = \tau(x_i)$ and $\equivfn(\inp)$ is a matrix with $\equivfn(\inp)_{ij} = \epsilon(x_i, x_j)$ (see \cref{fig:semantic-parser}). Then we say that the interpretation function $\mathcal{I}$ (or equivalently the set $X$) is $\mathcal{L}$-\textbf{compositional} iff we can write $\mathcal{I}(\mathbf{x}) = \mathcal{C}(\mathcal{L}(\mathbf{x}))$ for some \textbf{composition procedure} $\mathcal{C}$. In other words, $\mathcal{I}$ (and $X$) are compositional if they can be defined purely in terms of the types of and relations between the symbols in $\mathbf{x}$, without any information about the identity of the symbols themselves.\footnote{Note that every $\mathcal{I}$ is trivially $\mathcal{L}$-compositional under an $\mathcal{L}$ that assigns each symbol a unique type and labels no symbols as equivalent.}
\end{definition}
We emphasize that this definition makes \emph{no} assumptions about how $\mathcal{C}$ is implemented, beyond the fact that it can be defined purely in terms of $\mathcal{L}(\mathbf{x})$. It thus characterizes many different tasks:
\begin{example}{2(a)}{Semantic Parsing}
A sketch of a semantic parser factorizable into a lexicon and an abstract composition function is depicted in \cref{fig:semantic-parser}.
As a real-world example, in the factored CCG semantic parser of \citet{kwiatkowski2011lexical}, words are assigned types and fragments of logical forms via a lexicon that pairs words with single logical functions and a set of word-general templates. These logical fragments are then composed by a parsing algorithm whose behavior that depends only the types of the words selected.
\end{example}
\begin{example}{2(b)}{Paraphrase Detection}
\citet{maccartney2014natural}'s Natural Logic framework provides a procedure for determining whether two sentences are logically equivalent via a set of sentence rewriting operations that use only word-level information about semantic equivalence.\footnote{The full NL framework of \citeauthor{maccartney2014natural} can model a rich set of sentence relations beyond equivalence, including contradiction and entailment, via a similarly enriched set of word-level relations. Our framework has a natural generalization that replaces $\epsilon$ with a set of n-ary relations and modifies \cref{def:lhom} to require that $f$ be a homomorphism with respect to each (see \cref{app:generalization} for more details).}
\end{example}
\cref{def:abstraction} captures the intuition that a sentence interpretation procedure is compositional if the meaning of a sentence can be derived in a generic way ($\mathcal{C}$) from the meanings of its words ($\mathcal{L}$). As shown in Example 2(a), it can be used to implement a language-to-logical form mapping, and thus generalizes the Motagovian definition of compositionality as a homomorphism from sentences to meanings.
We remark, finally, that the parsing procedure depicted in \cref{fig:semantic-parser} is an idealization used to \emph{motivate} a particular approach to data generation; the actual models used in our experiments are more conventional neural sequence models.
We are now ready to describe how, for compositional $\mathcal{I}$, structure in $\mathcal{L}$ translates into structure in the set of well-formed examples $X$.
\begin{definition}\label{def:lhom}
A function $f$ is a \textbf{homomorphism of $\Sigma$ with respect to $\mathcal{L}$} (an ``$\mathcal{L}$-homomorphism'') if:
\begin{align}
\forall x \in \Sigma&: \tau(x) = \tau\big(f(x)\big) \\
\forall x, x' \in \Sigma&: \epsilon(x, x') = \epsilon\big(f(x), f(x')\big) \label{eq:lhom}
\end{align}
$f$ ``preserves the structure'' of $\mathcal{L}$: it replaces symbols with other symbols of the same type, and ensures that pairs of symbols maintain equivalence relationships. An example is depicted in \cref{fig:teaser}; note that both the words \emph{yellow} and \emph{green} and the corresponding meanings must be swapped in order to satisfy \cref{eq:lhom}.
\end{definition}
Our main claim is then as follows:
\begin{proposition}\label{prop:lhomsymm}
If $X$ is $\mathcal{L}$-compositional, $f$ is an $\mathcal{L}$-homomorphism, and $\mathbf{x} \in X$, then $f(\mathbf{x}) = [f(x_1), \ldots, f(x_n)] \in X$. Thus, every homomorphism of $\mathcal{L}$ corresponds to a symmetry of $X$.
\end{proposition}
\begin{proof}
\phantom{\qedhere}
From \cref{def:abstraction} and \ref{def:lhom}, $\typefn(\inp) = \tau(f(\mathbf{x}))$ and $\equivfn(\inp) = \epsilon(f(\mathbf{x}))$. Then,
\begin{align*}
\mathds{1}_{[f(\mathbf{x}) \in X]} &= \mathcal{I}(f(\mathbf{x})) \\
&= \mathcal{C}(\mathcal{L}(f(\mathbf{x}))) \\
&= \mathcal{C}(\mathcal{L}(\tau(f(\mathbf{x})), \epsilon(f(\mathbf{x})))) \\
&= \mathcal{C}(\mathcal{L}(\tau(f(\mathbf{x}), \epsilon(f(\mathbf{x})))) \\
&= \mathcal{I}(\mathbf{x}) = \mathds{1}_{[\mathbf{x} \in X]}
\tag*{\qed}
\end{align*}
\end{proof}
Despite its simplicity, \cref{prop:lhomsymm} has an important consequence: if we can identify candidate entries in $\mathcal{L}$, \emph{even if $\mathcal{C}$ is unknown}, we can construct new examples $\mathbf{x} \in X$ that respect (and provide evidence for) the compositional structure of $X$.
\section{Discovering Symmetries Automatically}
Given a lexicon of symbols identified with types and equivalence relations, we have shown how to turn isomorphisms of a lexicon into symmetries of a dataset. These symmetries are functions that take examples as input and return other examples; given a finite training set, they may be applied to training data to produce additional examples. Next, we show how to \emph{automate} this process: we give procedures for identifying types, equivalences, and homomorphisms, turning \cref{prop:lhomsymm} into a practical data augmentation scheme. %
\subsection{Inferring equivalence relations}
As noted in Examples 2a--b and \cref{fig:teaser}, the semantic equivalence function $\epsilon$ might model a variety of relations: between words and their logical forms, synonyms or even perceptual groundings.
Our approach to inferring $\epsilon$ focuses on problems in which the interpretation function $\mathcal{I}$ must judge the compatibility of two ``views'' of the same underlying meaning---e.g.\ compatibility of sentences and logical forms, or between images and (question, answer) pairs.
Across applications \cite{chiang2005hiero, akyurek2021lexicon}, existing models with lexicons are typically initialized by finding \emph{alignments} between the two views that identify pairs of symbols that frequently co-occur. Building on this past work, we identify candidate equivalences using a statistical alignment algorithm.
\paragraph{Learning Alignments}
Given an input $\mathbf{x}$ consisting of two views $(\mathbf{x}_\text{text}, \mathbf{x}_\text{other})$
we use the IBM model (see \cref{app:ibm} for details) to obtain forward and reverse alignments
between between views of individual training examples. For every text symbol $x_i$ and other symbol $x_j \in \Sigma$, we count the total number of times those symbols are aligned in both the forward and reverse directions across the dataset. For every $x_j$ that is aligned to at least 1\% of occurrences of $x_i$ in the dataset, we add the pair $(x_i, x_j)$ to the lexicon.
Finally, we produce a \emph{one-to-many} lexicon by deleting lexicon entries $(x_i, x_j)$ and $(x_i', x_j)$ when both exist.
\paragraph{Aside: Continous Domains}\label{par:continous}
Up to this point, we have assumed that inputs $\mathbf{x}$ consist of a discrete sequence of symbols drawn from a finite vocabulary. In many grounded language learning problems, however, examples $\mathbf{x}$ may contain images or torques defined over continuous domains. Recent work has shown that it is possible to \emph{learn} high-quality discrete representations of continuous input data. For example, in the VQ-VAE model of \citet{oord2017neural}, a continuous image is transformed into a grid of categorical codes, with individual codes reprsenting color, and in some cases materials and illumination (real examples shown in \cref{tab:samples}). We use this discretization procedure for VQA experiments in this paper (see \cref{app:vqvae} for details).
\subsection{Inferring types}
Following \citet{clark2007polynomial} and \citet{geca}, we assign type labels to individual symbols based on the environments in which they occur. For each symbol $x \in \Sigma$, we define the context $\kappa(x) = \{ (\alpha, \beta) : \alpha x \beta \in X \}$: that is, the set of contexts $(\alpha, \beta)$ that appear surrounding $x$ in the training set. (If the two examples in \cref{fig:teaser} formed the entire training set, we would have $\kappa(\textit{yellow}) = \kappa(\textit{green}) = \{ (\textit{Q: How many}, ~\textit{objects? A: 1}) \}$.) We then represent $\Sigma$ as a graph with an edge between each $x_i$ and $x_j$ where $\kappa(x_i) \cap \kappa(x_j) \neq \emptyset$ (Clark and Eyraud's \emph{syntactic congruence} relation). Finally, we assign each connected component of this graph a distinct type.
In practice, we impute types in this fashion only for symbols corresponding to natural language words. In semantic parsing and visual question answering applications we then \emph{project} these types into the logical or visual domain: if $\tau(x_i) = t$ and $\epsilon(x_i, x_j)$, we introduce a new symbol $t^\dagger$ and define $\tau(x_j) = \tau^\dagger(x_i) = t^\dagger$. We note that this is only one possible approach to typing; alternatives might use part-of-speech taggers or unsupervised clustering of distributed representations.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/model_vqa.pdf}
\caption{Example of our approach in VQA. We discretize images using a VQVAE \cite{oord2017neural} learned from the training data. This discretization represents every image as a sequence of categorical codes. (a) We run a statistical aligner on ($\mathbf{x}_{\textrm{text}}$, $\mathbf{x}_{\textrm{img}}$) pairs to find word--visual token alignments within individual examples, then use these alignments to construct a global lexicon. (b) Each entry in the lexicon is assigned a type based on the context in which it occurs. (c) Next, we find \emph{homomorphisms} of this lexicon, and use these as data augmentation functions to generate new training examples. (d) Finally, we train a neural sequence model on the augmented dataset.}
\label{fig:method}
\end{figure*}
\subsection{Constructing Homomorphisms}
Having inferred semantic equivalence relations and types above, isomorphisms can be generated by \emph{swapping} pairs of equivalent source and target tokens. In particular:
\label{sec:constrhom}
\begin{proposition}
Let $x_i$ and $x_j$ be tokens in $\Sigma$ with the same type $\tau(x_i) = \tau(x_j)$. For convenience, let $E_i=\{x : \epsilon(x_i, x) = 1\}$. Then the function:
\begin{equation}
\label{eq:fswap}
f(x) = \begin{cases}
x_j & \textrm{if } x = x_i \\
x_i & \textrm{if } x = x_j \\
x' \in E_j & \textrm{if } x \in E_i\\
x' \in E_i & \textrm{if } x \in E_j\\
x & \textrm{otherwise}
\end{cases}
\end{equation}
is an $\mathcal{L}$-homomorphism.
\end{proposition}\label{prop:homex}
\begin{proof}
Since $\tau(x_i) = \tau(x_j)$ and $\tau(x' \in E_i) = \tau^{\dagger}(x_i) = \tau^{\dagger}(x_j) = \tau(x' \in E_j)$, we have $\tau(f(x)) = \tau(x)$ for any $x$.
Now consider the $i^{\text{th}}$ row of the matrix $\equivfn(\inp)$:
\begin{itemize}
\item If $x \in E_i$, then $\epsilon(x_i, x) = 1$ and $\epsilon(f(x_i)$ $f(x)) = \epsilon(x_j, x' \in E_j) =1$.
\item If $x \in E_j$, then $\epsilon(x_i, x) = \mathds{1}_{E_i=E_j}$ and $\epsilon(f(x_i),$ $f(x)) = \epsilon(x_j, x' \in E_i)=\mathds{1}_{E_i=E_j}$. (Because the lexicon is one-to-many, either $E_i=E_j$ or $E_i \cap E_j = \emptyset$.)
\item If $x=x_j$, then $ \epsilon(x_i, x_j) = \epsilon(x_j, x_i) = \epsilon(f(x_i), f(x_j)) $.
\item Else $\epsilon(x, x_i) = 0=\epsilon(f(x), f(x_i)) = 0$.
\end{itemize}
By symmetry of $x_i$ and $x_j$, $\epsilon(f(x)) = \epsilon(x)$.%
\end{proof}
The homomorphism $f$
is closely related to
\emph{swap-based} data augmentation procedures used in the bias-fairness literature in NLP \cite{sun2019mitigating}, but generalizes it to problems that require modeling an (input, output) relation. Another related data augmentation procedure in NLP is random \emph{substitution} \cite{wang2018switchout} of aligned word pairs with different pairs, which corresponds to removing the second and fourth cases in \cref{eq:fswap}.
Substitution, unlike swapping, is not an $\mathcal{L}$-homomorphism, and can create examples that are not well-formed. (Consider substituting \emph{yellow} for \emph{green} without swapping in \cref{fig:teaser}, which changes the answer from \emph{1} to \emph{2}.) Our work thus offers a formal justification for these data augmentation procedures, and shows how to modify them to ensure that they preserve the well-formedness of data.
\subsection{Limitations}
The procedure described above is not capable of producing every data transformation commonly associated with compositional generalization: it will not exchange substructures larger than a single token, and thus will not synthesize examples \emph{longer} than those already present in the training set \citep{lake2019human}. In images represented as discrete symbol sequences, if a single symbol simultaneously encodes multiple visual features (e.g.\ color and texture), these features will remain entangled in synthesized examples. We will see that even with these limitations, data augmentation with $\mathcal{L}$-homomorphisms can greatly improve generalization across models and problem domains; future work might generalize this approach to extract even more general families of data transformation.
\section{Experiments}
Our experiments aim to evaluate whether the procedure given above is capable of extracting non-trivial lexicons across problem domains, and whether these lexicons in turn give rise to data augmentation schemes that improve compositional generalization in downstream models.
We combine our approach with standard RNN and transformer architectures, and compare it to baselines and state-of-the-art models in two problem domains: semantic parsing and visual question answering.
\subsection{Datasets}
\paragraph{CLEVR-CoGenT}
The CLEVR dataset \citep{johnson2017clevr} contains English-language questions about generated 3D scenes containing multiple objects. Questions involve a number of complex computational operations, including quantification, comparison, and spatial reasoning. CLEVR has been a popular testbed for evaluation visual question answering models' robustness to novel word and concept combinations. Our main experiment uses the CoGenT split of the dataset, which focuses on compositional generalization.
In the CLEVR-CoGenT training set (Split A), which contains roughly $700K$ (question, image, answer) triples, all cubes are gray, blue, brown or yellow, while all cylinders are red, green, purple or cyan. In the test set (validation set of Split B), these colors are reversed. %
\paragraph{COGS}
We additionally evaluate using the COGS dataset \cite{kim2020cogs}, which tests compositional generalization in semantic parsing.
It includes 24,155 automatically generated English (sentence, logical form) pairs, with systematic differences between train and test sets in the syntactic and semantic roles filled by individual lexical items.
\input{cogent_table.tex}
\subsection{Models}
\paragraph{VQA Transformer}
Most prior work on visual question answering has used pre-trained convolutional networks to encode images, and recurrent networks to encode questions and generate answers. For our experiments on CLEVR, we use a simplified model architecture in which both questions and images are mapped to answers by a transformer model, similarly to \citet{ramesh2021zero}. The model has the following form (dimensions in \cref{app:vqatransformer}):
\begin{align}\label{eq:vqatransformer}
\begin{split}
c_{\mathbf{x}_{I}} &= \operatorname{VQVAE_{\textrm{enc}}}({\mathbf{x}_{I}}) \\
e_Q &= W_Q \mathbf{x}_Q + \operatorname{1D_{\textrm{positional}}}(\mathbf{x}_Q) \\
e_{\mathbf{x}_{I}} &= W_c c_{\mathbf{x}_{I}} + \operatorname{2D_{\textrm{positional}}}(c_{\mathbf{x}}) \\
h &= \operatorname{Transformer}([e_Q \,e_{\mathbf{x}_I}]) \\
p(\mathbf{x}_A \mid \mathbf{x}_Q, \mathbf{x}_{I}) &= \textrm{softmax}(W_{\textrm{proj}} h_{\textrm{start}})
\end{split}
\end{align}
Recall that the lexicon and data augmentation procedure itself operates over sequences of discrete visual codes produced by a vector-quantized variational autoencoder (\cref{par:continous}). Once discrete representations $c_{\mathbf{x}_{I}}$ have been produced, we infer lexicons and perform data augmentation directly to these representations, without re-synthesizing images $\mathbf{x}_I$ (though such synthesis is possible, as in \cref{fig:method}, to interpret model behavior).
\paragraph{LSTM}
For semantic parsing experiments, we use LSTM networks \cite{hochreiter1997long} equipped with a copy mechanism \cite{see2017get} using a baseline model provided by \citet{akyurek2021lexicon}.
\subsection{Data Augmentation}
We train these base models by performing data augmentation with $\mathcal{L}$-homomorphisms.
Data augmentation is applied online during model training. We construct transformation functions by (1) sampling $x_i$ uniformly from the tokens in the lexicon, (2) sampling $x_j$ from $\{x_j \mid \tau(x_j)=\tau(x_i)\}$ in inverse probability to its frequency in the training set, (3) sampling $x_i' \sim \textrm{Unif}(E_i)$ and $x_j' \sim \textrm{Unif}(E_j)$, and (4) constructing an $\mathcal{L}$-homomorphism as in \cref{eq:fswap}. In CLEVR, we apply one transformation uniformly to each batch; in COGS, we apply one transformation to each example with 20\% probability.
\subsection{Results}
\begin{table}[t]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{ll}
\toprule
Model & Generalization \\
\midrule
LSTM+copy & 0.66 \stderr{0.03} \\
\boxSpace\boxDownRight Substitute & 0.81 \stderr{0.02} \\
\boxSpace\boxDownRight Swap & 0.81 \stderr{0.01} \\
\midrule
LexLSTM \cite{akyurek2021lexicon} & 0.82 \stderr{0.00} \\
LAER \cite{liu2021learning} & 0.98 \stderr{0.00} \\
T5+CSL-Aug \cite{qiu2021improving} & \textbf{0.99} \\
\bottomrule
\end{tabular}
}
\caption{Semantic parsing results on COGS dataset for generalization set. We provide mean and standard deviations over 10 random seeds. Augmentation improves significantly over the LSTM baseline and obtains performance on-par with the LexLSTM model but behind specialized semantic parsing approaches.}
\label{tab:generalization_comparison}
\end{table}
\begin{table*}[t]
\label{tab:samples}
\input{samples_cogs}
\input{samples_vqa}
\caption{Generated samples for CLEVR-CoGent and COGS datasets.
}
\end{table*}
\paragraph{VQA}
In \cref{tab:clevr_cogent}, a transformer model trained with symmetry-based data augmentation achieves state-of-the-art results on the CLEVR-CoGenT dataset, improving accuracy by more than 7\% relative to the best existing system.
$\mathcal{L}$-homomorphism-based data augmentation also outperforms data augmentation based on synchronous token substitution (\cref{sec:constrhom}), which is not guaranteed to preserve meaning, and performs particularly well on questions involving quantification (\emph{Count}, \emph{Exist} and \emph{Compare} columns).
On the standard CLEVR dataset, our approach achieves performance comparable to human accuracy, and slightly behind that of other machine learning approaches.
\paragraph{Semantic Parsing}
On the COGS semantic parsing dataset, $\mathcal{L}$-homomorphism-based data augmentation substantially outperforms a baseline LSTM model. Its results are worse than can be achieved with training schemes specifically tailored toward semantic parsing tasks,
but comparable to performance of a generic neural sequence architectures specially targeted at compositional generalization (the LexLSTM model of \citealp{akyurek2021lexicon}). We note that unlike data augmentation with $\mathcal{L}$-homomorphisms, none of COGS models can be applied to VQA. These experiments highlight the flexibility of our approach, but do not rule out the possibility that stronger forms of compositional generlization may be obtained with approaches more narrowly tailored to specific problem domains.
\section{Other Related Work}
\label{sec:otherrelated}
\paragraph{Lexicalized neural models}
Word-level alignments between input and output sequences were an essential feature of statistical phrase- and tree-based sequence models \cite{chiang2005hiero, Koehn2003StatisticalPT}.
Neural scoring functions were sometimes integrated into these models \cite{misra2016neural}.
Neural models with attention \cite{bahdanau2014neural} do not require explicit alignment, though several pieces of past work have shown that incorporating explicit token-level correspondences improves generalization \cite{akyurek2021lexicon, Prabhu2020MakingAP, pham2018towards}.
To the best of our knowledge, this paper is the first to make use of token-level alignments in joint neural models of text and images.
\paragraph{Data Augmentation}
Data augmentation approaches are widely used across machine learning application domains featuring known invariances of the data distribution \cite{japkowicz2000learning, data-recombination-copy, shaw2020compositional}.
Substitution-based data augmentation schemes that replace words with synonyms, or synchronously replace words and their translations, are widely
used in tasks like machine translation \cite{wang2018switchout, wei2019eda}. Previous work use substitution of pronouns to de-bias NLP models for gender. Back-translation \citep{sennrich2015improving}, which encourages cycle-consistency in sequence transduction, may also be viewed as a form of data augmentation.
\section{Conclusion}
We have presented a new lexicon-based data augmentation method that improves compositional generalization of neural models in multiple domains. This method is derived from a new characterization of the principle of compositionality as a constraint on the symmetries of data distributions, and a procedure for automatically identifying these symmetries using token-level alignments.
Our results highlight the fact that many of the inductive biases targeted by specialized models in NLP can be alternatively, and often more flexibly, expressed as a hypothesis about the structure of the dataset to be modeled.
\section*{Acknowledgements}
This work was supported by the MachineLearningApplications initiative at MIT CSAIL and the MIT--IBM Watson AI lab. Computing resources were provided by a gift from NVIDIA through the NVAIL program and by the Lincoln Laboratory Supercloud.
\bibliographystyle{acl_natbib}
\section{Introduction}
\label{sec:introduction}
A central challenge in natural language processing is the design of models and learning algorithms that are simultaneously \emph{flexible} enough to capture the variability of human language and \emph{structured} enough to generalize in predictable and human-like ways.
One important source of structure is the \textbf{principle of compositionality}, which (in one formulation)
states that sentence meanings can be computed from a \emph{lexicon} of word meanings and a set of \emph{composition rules} governing how meanings combine \cite{montague1970universal}.
A long line of language processing research has operationalized the principle of compositionality as a \textbf{constraint on model architectures}, via independence assumptions (e.g.\ in synchronous grammars; \citealp{lewis1968syntax}) or parameter tying schemes (e.g.\ in modular neural networks; \citealp{socher2013recursive,andreas2016neural}) that ensure a compositional process of sentence interpretation.
Compositional models enjoy sample-efficient learning and strong forms of out-of-distribution generalization in tasks ranging from machine translation to question answering \cite{mccoy2020does}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\linewidth]{figures/teaser_vqa.pdf}
\caption{Example of our approach to compositional modeling in the visual question answering domain. Given a dataset of (image, question, answer) triples, we extract a lexicon that relates words to their visual groundings. We then find \emph{homomorphic transformations} (\cref{sec:theory}) of this lexicon that, when applied to training examples, produce new, well-formed examples.} %
\label{fig:teaser}
\end{figure}
But much of human language is not (or at least not straightforwardly) compositional. Idioms, disfluencies, and context-sensitive meanings present major challenges to models in which all predictions must be derivable from a sequence of local composition operations.
In recent years, more generic model architectures such as recurrent neural networks (RNNs) and transformers, with no explicit compositional scaffolding,
have consistently outperformed compositional models in language processing tasks with natural data distributions \cite{googlemt}.
At the same time, these models capture linguistic regularities only when trained on enormous amounts of data, and make surprising or problematic predictions when presented with novel word collocations or syntactic structures \cite{lake2018generalization}.
How can we train unstructured neural sequence models that generalize compositionally?
In this paper, we describe how to inject compositional inductive bias into general function approximators by operationalizing compositionality as a \textbf{constraint on data distributions} rather than model architectures.
\cref{fig:teaser} gives the intuition behind our approach: in a visual domain in which the compatibility of images and captions can be computed using a lexicon that relates words like \emph{yellow} and \emph{green} to their perceptual groundings, we show how to transform a structural regularity of the lexicon (the fact that the words \emph{yellow} and \emph{green} are of the same type and play the same role in downstream computation) into a structural regularity of the dataset itself (in any well-formed image--caption pair, swapping the words \emph{yellow} and \emph{green} and their associated colors yields a new pair that is still well-formed).
More formally, we prove that for any dataset in which the well-formedness of examples can be computed compositionally (i.e.\ factored into a lexicon and a composition function), every \emph{homomorphism} of the lexicon
corresponds to a \emph{symmetry} of the dataset.
Using this characterization, we describe a procedure that uses statistical alignment algorithms to build lexicons for problems of interest, then uses these lexicons to construct data augmentation procedures \emph{guaranteed} to produce semantically acceptable examples.
An important benefit of this data-centric approach to compositionality is that it does not require a complete compositional theory for a given problem domain, or even a lexicon entry for every word:
partial information about a problem's lexicon suffices to derive nontrivial data transformations.
Applied to transformers and RNNs,
this approach to data augmentation improves performance on tests of compositional generalization in two problem domains: the CLEVR CoGenT visual question answering dataset, and the COGS semantic parsing dataset. %
This paper thus offers two contributions: first, a new lens on the principle of compositionality via symmetries of data distributions; second, a concrete instantiation of this framework as a data augmentation scheme that improves model performance on multiple tasks. Our results highlight the fact that machine learning with compositional inductive bias need not require new models: there is a natural ``data-centric'' view of the principle of compositionality
that may be more robust and broadly applicable than the traditional model-centric one.\footnote{Code for the experiments is released under \url{https://github.com/ekinakyurek/lexsym}}
\section{Background \& Approach}
We begin with a discussion on the more general role of \emph{symmetry} in machine learning applications.
\begin{definition}\label{def:symmetry}
A \textbf{symmetry} of a set $X$ is a function $f$ satisfying:
\begin{equation}
\{ f(x) : x \in X \} = X
\end{equation}
That is, applying $f$ to each element of $X$ leaves $X$ unchanged.
\end{definition}
A familiar example from computer vision is \emph{reflection symmetry}: in object recognition problems, image classes are generally invariant under reflection (a zebra seen in a mirror is still a zebra). The set of (image, class) pairs thus has as a symmetry the function $(x, y) \mapsto (\texttt{reflect}(x), y)$.
In many problems, especially those (like computer vision and computational chemistry) that are constrained by physical laws, knowledge of the symmetries exhibited by a problem domain can dramatically reduce the difficulty of learning in that domain.
Past work has incorporated symmetry into machine learning problems in two ways. \textbf{Invariant and equivariant modeling} approaches structurally enforce symmetries via specialized architectures (improving generalization by decreasing the size of the hypothesis class; \citealp{cohen2016group}). \textbf{Data augmentation} approaches generate new training examples by applying known symmetries like reflections directly to training data (improving generalization by increasing dataset size; \citealp{shorten2019survey}). Data augmentation, the focus of this paper, is model-agnostic, and can be used in conjunction with pretraining (even on non-symmetric domains) while producing the same asymptotic effects as specialized model architectures \cite{chen2020group}.
The question this paper aims to answer is whether compositionality, like other domain-specific constraints, can be formalized in the language of symmetry.
We are not the first to consider this question: \citet{kiddon2015symmetry} define a theory of semantic equivalence in terms of symmetries of the set of natural language sentences, and \citet{Gordon2020Permutation} propose a model architecture for compositional semantic parsing via a symmetry that enforces \emph{permutation invariance} of lexical items. As described in \cref{sec:theory}, our approach also derives symmetries from lexicons. It builds on the work of \citeauthor{Gordon2020Permutation} by (1) characterizing the relationship between compositionality and symmetry for general problem domains, (2) discovering lexicons and symmetries automatically, and (3) showing how to leverage them in a model-agnostic way via data augmentation.
Additional related work is discussed in \cref{sec:otherrelated}.
\section{Compositionality as Lexical Symmetry}
\label{sec:theory}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/tasks.pdf}
\caption{Idealized compositional semantic parser following \cref{def:abstraction}. A (sentence, logical form) pair is translated into a \emph{lexical abstraction} containing information about each token's type and semantic equivalences. We then determine whether the sentence evaluates to the logical form using \emph{only} the type and equivalence matrices, using types to assign the sentence an abstract logical form, and equivalences to determine whether it matches the target.
}
\label{fig:semantic-parser}
\end{figure}
In this section, we formalize the relationship between \emph{compositional sentence interpretation} and \emph{symmetries of datasets}.
We consider problems defined by a space of possible examples $\mathcal{X}$, of which a subset of examples $X$ are \textbf{well-formed}.
We assume each example $\mathbf{x} \in \mathcal{X}$ is a discrete sequence $[x_1, \ldots, x_n]$, with $x_i$ drawn from a vocabulary $\Sigma$.
Finally, we assume that well-formedness can be computed by a \textbf{interpretation function} $\mathcal{I} : \mathcal{X} \to \{0, 1\}$ with $\mathcal{I}(\mathbf{x}) = 1$ iff $\mathbf{x} \in X$.
\newcommand{\inp_\textrm{NL}}{\mathbf{x}_\textrm{NL}}
\newcommand{\inp_\textrm{LF}}{\mathbf{x}_\textrm{LF}}
\begin{example}{1(a)}{Semantic Parsing}
Examples $\mathbf{x}$ are pairs $(\inp_\textrm{NL}, \inp_\textrm{LF})$, where $\inp_\textrm{NL}$ is an sentence, $\inp_\textrm{LF}$ is a logical form, and $\mathcal{I}(\inp_\textrm{NL}, \inp_\textrm{LF}) = 1$ iff $\inp_\textrm{LF}$ represents a possible meaning of $\inp_\textrm{NL}$.
\end{example}
\begin{example}{1(b)}{Paraphrase Detection}
Examples $\mathbf{x}$ are sentence pairs $(\mathbf{x}', \mathbf{x}'')$, with $\mathcal{I}(\mathbf{x}', \mathbf{x}'') = 1$ iff $\mathbf{x}'$ and $\mathbf{x}''$ have the same meaning.
\end{example}
\begin{example}{1(c)}{Visual Question Answering}
\newcommand{\inp_\textrm{Q}}{\mathbf{x}_\textrm{Q}}
\newcommand{\inp_\textrm{I}}{\mathbf{x}_\textrm{I}}
\newcommand{\inp_\textrm{A}}{\mathbf{x}_\textrm{A}}
Examples $\mathbf{x}$ are triples $(\inp_\textrm{Q}, \inp_\textrm{I}, \inp_\textrm{A})$, where $\inp_\textrm{Q}$ is a question, $\inp_\textrm{I}$ is a (discretized) image, $\inp_\textrm{A}$ is an answer, and $\mathcal{I}(\inp_\textrm{Q}, \inp_\textrm{I}, \inp_\textrm{A}) = 1$ iff $\inp_\textrm{A}$ is the answer to $\inp_\textrm{Q}$ in $\inp_\textrm{I}$.
\end{example}
We aim to prove that \emph{if $\mathcal{I}$ can be computed compositionally, $X$ exhibits identifiable symmetries in the sense of \cref{def:symmetry}}.
We begin by formalizing what it means for $\mathcal{I}$ to be compositional.
\begin{definition}
A \textbf{lexicon} $\mathcal{L}$ is a pair $(\tau, \epsilon)$ where $\tau : \Sigma \to \mathcal{T}$ assigns each symbol a \textbf{type} from some set $\mathcal{T}$, and $\epsilon : \Sigma \times \Sigma \to \{ 0, 1 \}$ is a relation indicating which symbols are \textbf{semantically equivalent}.
\end{definition}
\begin{definition}\label{def:abstraction}
Denote the \textbf{lexical abstraction} $\mathcal{L}(\mathbf{x}) = (\typefn(\inp), \equivfn(\inp))$ where $\typefn(\inp)$ is a vector with $\typefn(\inp)_i = \tau(x_i)$ and $\equivfn(\inp)$ is a matrix with $\equivfn(\inp)_{ij} = \epsilon(x_i, x_j)$ (see \cref{fig:semantic-parser}). Then we say that the interpretation function $\mathcal{I}$ (or equivalently the set $X$) is $\mathcal{L}$-\textbf{compositional} iff we can write $\mathcal{I}(\mathbf{x}) = \mathcal{C}(\mathcal{L}(\mathbf{x}))$ for some \textbf{composition procedure} $\mathcal{C}$. In other words, $\mathcal{I}$ (and $X$) are compositional if they can be defined purely in terms of the types of and relations between the symbols in $\mathbf{x}$, without any information about the identity of the symbols themselves.\footnote{Note that every $\mathcal{I}$ is trivially $\mathcal{L}$-compositional under an $\mathcal{L}$ that assigns each symbol a unique type and labels no symbols as equivalent.}
\end{definition}
We emphasize that this definition makes \emph{no} assumptions about how $\mathcal{C}$ is implemented, beyond the fact that it can be defined purely in terms of $\mathcal{L}(\mathbf{x})$. It thus characterizes many different tasks:
\begin{example}{2(a)}{Semantic Parsing}
A sketch of a semantic parser factorizable into a lexicon and an abstract composition function is depicted in \cref{fig:semantic-parser}.
As a real-world example, in the factored CCG semantic parser of \citet{kwiatkowski2011lexical}, words are assigned types and fragments of logical forms via a lexicon that pairs words with single logical functions and a set of word-general templates. These logical fragments are then composed by a parsing algorithm whose behavior that depends only the types of the words selected.
\end{example}
\begin{example}{2(b)}{Paraphrase Detection}
\citet{maccartney2014natural}'s Natural Logic framework provides a procedure for determining whether two sentences are logically equivalent via a set of sentence rewriting operations that use only word-level information about semantic equivalence.\footnote{The full NL framework of \citeauthor{maccartney2014natural} can model a rich set of sentence relations beyond equivalence, including contradiction and entailment, via a similarly enriched set of word-level relations. Our framework has a natural generalization that replaces $\epsilon$ with a set of n-ary relations and modifies \cref{def:lhom} to require that $f$ be a homomorphism with respect to each (see \cref{app:generalization} for more details).}
\end{example}
\cref{def:abstraction} captures the intuition that a sentence interpretation procedure is compositional if the meaning of a sentence can be derived in a generic way ($\mathcal{C}$) from the meanings of its words ($\mathcal{L}$). As shown in Example 2(a), it can be used to implement a language-to-logical form mapping, and thus generalizes the Motagovian definition of compositionality as a homomorphism from sentences to meanings.
We remark, finally, that the parsing procedure depicted in \cref{fig:semantic-parser} is an idealization used to \emph{motivate} a particular approach to data generation; the actual models used in our experiments are more conventional neural sequence models.
We are now ready to describe how, for compositional $\mathcal{I}$, structure in $\mathcal{L}$ translates into structure in the set of well-formed examples $X$.
\begin{definition}\label{def:lhom}
A function $f$ is a \textbf{homomorphism of $\Sigma$ with respect to $\mathcal{L}$} (an ``$\mathcal{L}$-homomorphism'') if:
\begin{align}
\forall x \in \Sigma&: \tau(x) = \tau\big(f(x)\big) \\
\forall x, x' \in \Sigma&: \epsilon(x, x') = \epsilon\big(f(x), f(x')\big) \label{eq:lhom}
\end{align}
$f$ ``preserves the structure'' of $\mathcal{L}$: it replaces symbols with other symbols of the same type, and ensures that pairs of symbols maintain equivalence relationships. An example is depicted in \cref{fig:teaser}; note that both the words \emph{yellow} and \emph{green} and the corresponding meanings must be swapped in order to satisfy \cref{eq:lhom}.
\end{definition}
Our main claim is then as follows:
\begin{proposition}\label{prop:lhomsymm}
If $X$ is $\mathcal{L}$-compositional, $f$ is an $\mathcal{L}$-homomorphism, and $\mathbf{x} \in X$, then $f(\mathbf{x}) = [f(x_1), \ldots, f(x_n)] \in X$. Thus, every homomorphism of $\mathcal{L}$ corresponds to a symmetry of $X$.
\end{proposition}
\begin{proof}
\phantom{\qedhere}
From \cref{def:abstraction} and \ref{def:lhom}, $\typefn(\inp) = \tau(f(\mathbf{x}))$ and $\equivfn(\inp) = \epsilon(f(\mathbf{x}))$. Then,
\begin{align*}
\mathds{1}_{[f(\mathbf{x}) \in X]} &= \mathcal{I}(f(\mathbf{x})) \\
&= \mathcal{C}(\mathcal{L}(f(\mathbf{x}))) \\
&= \mathcal{C}(\mathcal{L}(\tau(f(\mathbf{x})), \epsilon(f(\mathbf{x})))) \\
&= \mathcal{C}(\mathcal{L}(\tau(f(\mathbf{x}), \epsilon(f(\mathbf{x})))) \\
&= \mathcal{I}(\mathbf{x}) = \mathds{1}_{[\mathbf{x} \in X]}
\tag*{\qed}
\end{align*}
\end{proof}
Despite its simplicity, \cref{prop:lhomsymm} has an important consequence: if we can identify candidate entries in $\mathcal{L}$, \emph{even if $\mathcal{C}$ is unknown}, we can construct new examples $\mathbf{x} \in X$ that respect (and provide evidence for) the compositional structure of $X$.
\section{Discovering Symmetries Automatically}
Given a lexicon of symbols identified with types and equivalence relations, we have shown how to turn isomorphisms of a lexicon into symmetries of a dataset. These symmetries are functions that take examples as input and return other examples; given a finite training set, they may be applied to training data to produce additional examples. Next, we show how to \emph{automate} this process: we give procedures for identifying types, equivalences, and homomorphisms, turning \cref{prop:lhomsymm} into a practical data augmentation scheme. %
\subsection{Inferring equivalence relations}
As noted in Examples 2a--b and \cref{fig:teaser}, the semantic equivalence function $\epsilon$ might model a variety of relations: between words and their logical forms, synonyms or even perceptual groundings.
Our approach to inferring $\epsilon$ focuses on problems in which the interpretation function $\mathcal{I}$ must judge the compatibility of two ``views'' of the same underlying meaning---e.g.\ compatibility of sentences and logical forms, or between images and (question, answer) pairs.
Across applications \cite{chiang2005hiero, akyurek2021lexicon}, existing models with lexicons are typically initialized by finding \emph{alignments} between the two views that identify pairs of symbols that frequently co-occur. Building on this past work, we identify candidate equivalences using a statistical alignment algorithm.
\paragraph{Learning Alignments}
Given an input $\mathbf{x}$ consisting of two views $(\mathbf{x}_\text{text}, \mathbf{x}_\text{other})$
we use the IBM model (see \cref{app:ibm} for details) to obtain forward and reverse alignments
between between views of individual training examples. For every text symbol $x_i$ and other symbol $x_j \in \Sigma$, we count the total number of times those symbols are aligned in both the forward and reverse directions across the dataset. For every $x_j$ that is aligned to at least 1\% of occurrences of $x_i$ in the dataset, we add the pair $(x_i, x_j)$ to the lexicon.
Finally, we produce a \emph{one-to-many} lexicon by deleting lexicon entries $(x_i, x_j)$ and $(x_i', x_j)$ when both exist.
\paragraph{Aside: Continous Domains}\label{par:continous}
Up to this point, we have assumed that inputs $\mathbf{x}$ consist of a discrete sequence of symbols drawn from a finite vocabulary. In many grounded language learning problems, however, examples $\mathbf{x}$ may contain images or torques defined over continuous domains. Recent work has shown that it is possible to \emph{learn} high-quality discrete representations of continuous input data. For example, in the VQ-VAE model of \citet{oord2017neural}, a continuous image is transformed into a grid of categorical codes, with individual codes reprsenting color, and in some cases materials and illumination (real examples shown in \cref{tab:samples}). We use this discretization procedure for VQA experiments in this paper (see \cref{app:vqvae} for details).
\subsection{Inferring types}
Following \citet{clark2007polynomial} and \citet{geca}, we assign type labels to individual symbols based on the environments in which they occur. For each symbol $x \in \Sigma$, we define the context $\kappa(x) = \{ (\alpha, \beta) : \alpha x \beta \in X \}$: that is, the set of contexts $(\alpha, \beta)$ that appear surrounding $x$ in the training set. (If the two examples in \cref{fig:teaser} formed the entire training set, we would have $\kappa(\textit{yellow}) = \kappa(\textit{green}) = \{ (\textit{Q: How many}, ~\textit{objects? A: 1}) \}$.) We then represent $\Sigma$ as a graph with an edge between each $x_i$ and $x_j$ where $\kappa(x_i) \cap \kappa(x_j) \neq \emptyset$ (Clark and Eyraud's \emph{syntactic congruence} relation). Finally, we assign each connected component of this graph a distinct type.
In practice, we impute types in this fashion only for symbols corresponding to natural language words. In semantic parsing and visual question answering applications we then \emph{project} these types into the logical or visual domain: if $\tau(x_i) = t$ and $\epsilon(x_i, x_j)$, we introduce a new symbol $t^\dagger$ and define $\tau(x_j) = \tau^\dagger(x_i) = t^\dagger$. We note that this is only one possible approach to typing; alternatives might use part-of-speech taggers or unsupervised clustering of distributed representations.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/model_vqa.pdf}
\caption{Example of our approach in VQA. We discretize images using a VQVAE \cite{oord2017neural} learned from the training data. This discretization represents every image as a sequence of categorical codes. (a) We run a statistical aligner on ($\mathbf{x}_{\textrm{text}}$, $\mathbf{x}_{\textrm{img}}$) pairs to find word--visual token alignments within individual examples, then use these alignments to construct a global lexicon. (b) Each entry in the lexicon is assigned a type based on the context in which it occurs. (c) Next, we find \emph{homomorphisms} of this lexicon, and use these as data augmentation functions to generate new training examples. (d) Finally, we train a neural sequence model on the augmented dataset.}
\label{fig:method}
\end{figure*}
\subsection{Constructing Homomorphisms}
Having inferred semantic equivalence relations and types above, isomorphisms can be generated by \emph{swapping} pairs of equivalent source and target tokens. In particular:
\label{sec:constrhom}
\begin{proposition}
Let $x_i$ and $x_j$ be tokens in $\Sigma$ with the same type $\tau(x_i) = \tau(x_j)$. For convenience, let $E_i=\{x : \epsilon(x_i, x) = 1\}$. Then the function:
\begin{equation}
\label{eq:fswap}
f(x) = \begin{cases}
x_j & \textrm{if } x = x_i \\
x_i & \textrm{if } x = x_j \\
x' \in E_j & \textrm{if } x \in E_i\\
x' \in E_i & \textrm{if } x \in E_j\\
x & \textrm{otherwise}
\end{cases}
\end{equation}
is an $\mathcal{L}$-homomorphism.
\end{proposition}\label{prop:homex}
\begin{proof}
Since $\tau(x_i) = \tau(x_j)$ and $\tau(x' \in E_i) = \tau^{\dagger}(x_i) = \tau^{\dagger}(x_j) = \tau(x' \in E_j)$, we have $\tau(f(x)) = \tau(x)$ for any $x$.
Now consider the $i^{\text{th}}$ row of the matrix $\equivfn(\inp)$:
\begin{itemize}
\item If $x \in E_i$, then $\epsilon(x_i, x) = 1$ and $\epsilon(f(x_i)$ $f(x)) = \epsilon(x_j, x' \in E_j) =1$.
\item If $x \in E_j$, then $\epsilon(x_i, x) = \mathds{1}_{E_i=E_j}$ and $\epsilon(f(x_i),$ $f(x)) = \epsilon(x_j, x' \in E_i)=\mathds{1}_{E_i=E_j}$. (Because the lexicon is one-to-many, either $E_i=E_j$ or $E_i \cap E_j = \emptyset$.)
\item If $x=x_j$, then $ \epsilon(x_i, x_j) = \epsilon(x_j, x_i) = \epsilon(f(x_i), f(x_j)) $.
\item Else $\epsilon(x, x_i) = 0=\epsilon(f(x), f(x_i)) = 0$.
\end{itemize}
By symmetry of $x_i$ and $x_j$, $\epsilon(f(x)) = \epsilon(x)$.%
\end{proof}
The homomorphism $f$
is closely related to
\emph{swap-based} data augmentation procedures used in the bias-fairness literature in NLP \cite{sun2019mitigating}, but generalizes it to problems that require modeling an (input, output) relation. Another related data augmentation procedure in NLP is random \emph{substitution} \cite{wang2018switchout} of aligned word pairs with different pairs, which corresponds to removing the second and fourth cases in \cref{eq:fswap}.
Substitution, unlike swapping, is not an $\mathcal{L}$-homomorphism, and can create examples that are not well-formed. (Consider substituting \emph{yellow} for \emph{green} without swapping in \cref{fig:teaser}, which changes the answer from \emph{1} to \emph{2}.) Our work thus offers a formal justification for these data augmentation procedures, and shows how to modify them to ensure that they preserve the well-formedness of data.
\subsection{Limitations}
The procedure described above is not capable of producing every data transformation commonly associated with compositional generalization: it will not exchange substructures larger than a single token, and thus will not synthesize examples \emph{longer} than those already present in the training set \citep{lake2019human}. In images represented as discrete symbol sequences, if a single symbol simultaneously encodes multiple visual features (e.g.\ color and texture), these features will remain entangled in synthesized examples. We will see that even with these limitations, data augmentation with $\mathcal{L}$-homomorphisms can greatly improve generalization across models and problem domains; future work might generalize this approach to extract even more general families of data transformation.
\section{Experiments}
Our experiments aim to evaluate whether the procedure given above is capable of extracting non-trivial lexicons across problem domains, and whether these lexicons in turn give rise to data augmentation schemes that improve compositional generalization in downstream models.
We combine our approach with standard RNN and transformer architectures, and compare it to baselines and state-of-the-art models in two problem domains: semantic parsing and visual question answering.
\subsection{Datasets}
\paragraph{CLEVR-CoGenT}
The CLEVR dataset \citep{johnson2017clevr} contains English-language questions about generated 3D scenes containing multiple objects. Questions involve a number of complex computational operations, including quantification, comparison, and spatial reasoning. CLEVR has been a popular testbed for evaluation visual question answering models' robustness to novel word and concept combinations. Our main experiment uses the CoGenT split of the dataset, which focuses on compositional generalization.
In the CLEVR-CoGenT training set (Split A), which contains roughly $700K$ (question, image, answer) triples, all cubes are gray, blue, brown or yellow, while all cylinders are red, green, purple or cyan. In the test set (validation set of Split B), these colors are reversed. %
\paragraph{COGS}
We additionally evaluate using the COGS dataset \cite{kim2020cogs}, which tests compositional generalization in semantic parsing.
It includes 24,155 automatically generated English (sentence, logical form) pairs, with systematic differences between train and test sets in the syntactic and semantic roles filled by individual lexical items.
\input{cogent_table.tex}
\subsection{Models}
\paragraph{VQA Transformer}
Most prior work on visual question answering has used pre-trained convolutional networks to encode images, and recurrent networks to encode questions and generate answers. For our experiments on CLEVR, we use a simplified model architecture in which both questions and images are mapped to answers by a transformer model, similarly to \citet{ramesh2021zero}. The model has the following form (dimensions in \cref{app:vqatransformer}):
\begin{align}\label{eq:vqatransformer}
\begin{split}
c_{\mathbf{x}_{I}} &= \operatorname{VQVAE_{\textrm{enc}}}({\mathbf{x}_{I}}) \\
e_Q &= W_Q \mathbf{x}_Q + \operatorname{1D_{\textrm{positional}}}(\mathbf{x}_Q) \\
e_{\mathbf{x}_{I}} &= W_c c_{\mathbf{x}_{I}} + \operatorname{2D_{\textrm{positional}}}(c_{\mathbf{x}}) \\
h &= \operatorname{Transformer}([e_Q \,e_{\mathbf{x}_I}]) \\
p(\mathbf{x}_A \mid \mathbf{x}_Q, \mathbf{x}_{I}) &= \textrm{softmax}(W_{\textrm{proj}} h_{\textrm{start}})
\end{split}
\end{align}
Recall that the lexicon and data augmentation procedure itself operates over sequences of discrete visual codes produced by a vector-quantized variational autoencoder (\cref{par:continous}). Once discrete representations $c_{\mathbf{x}_{I}}$ have been produced, we infer lexicons and perform data augmentation directly to these representations, without re-synthesizing images $\mathbf{x}_I$ (though such synthesis is possible, as in \cref{fig:method}, to interpret model behavior).
\paragraph{LSTM}
For semantic parsing experiments, we use LSTM networks \cite{hochreiter1997long} equipped with a copy mechanism \cite{see2017get} using a baseline model provided by \citet{akyurek2021lexicon}.
\subsection{Data Augmentation}
We train these base models by performing data augmentation with $\mathcal{L}$-homomorphisms.
Data augmentation is applied online during model training. We construct transformation functions by (1) sampling $x_i$ uniformly from the tokens in the lexicon, (2) sampling $x_j$ from $\{x_j \mid \tau(x_j)=\tau(x_i)\}$ in inverse probability to its frequency in the training set, (3) sampling $x_i' \sim \textrm{Unif}(E_i)$ and $x_j' \sim \textrm{Unif}(E_j)$, and (4) constructing an $\mathcal{L}$-homomorphism as in \cref{eq:fswap}. In CLEVR, we apply one transformation uniformly to each batch; in COGS, we apply one transformation to each example with 20\% probability.
\subsection{Results}
\begin{table}[t]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{ll}
\toprule
Model & Generalization \\
\midrule
LSTM+copy & 0.66 \stderr{0.03} \\
\boxSpace\boxDownRight Substitute & 0.81 \stderr{0.02} \\
\boxSpace\boxDownRight Swap & 0.81 \stderr{0.01} \\
\midrule
LexLSTM \cite{akyurek2021lexicon} & 0.82 \stderr{0.00} \\
LAER \cite{liu2021learning} & 0.98 \stderr{0.00} \\
T5+CSL-Aug \cite{qiu2021improving} & \textbf{0.99} \\
\bottomrule
\end{tabular}
}
\caption{Semantic parsing results on COGS dataset for generalization set. We provide mean and standard deviations over 10 random seeds. Augmentation improves significantly over the LSTM baseline and obtains performance on-par with the LexLSTM model but behind specialized semantic parsing approaches.}
\label{tab:generalization_comparison}
\end{table}
\begin{table*}[t]
\label{tab:samples}
\input{samples_cogs}
\input{samples_vqa}
\caption{Generated samples for CLEVR-CoGent and COGS datasets.
}
\end{table*}
\paragraph{VQA}
In \cref{tab:clevr_cogent}, a transformer model trained with symmetry-based data augmentation achieves state-of-the-art results on the CLEVR-CoGenT dataset, improving accuracy by more than 7\% relative to the best existing system.
$\mathcal{L}$-homomorphism-based data augmentation also outperforms data augmentation based on synchronous token substitution (\cref{sec:constrhom}), which is not guaranteed to preserve meaning, and performs particularly well on questions involving quantification (\emph{Count}, \emph{Exist} and \emph{Compare} columns).
On the standard CLEVR dataset, our approach achieves performance comparable to human accuracy, and slightly behind that of other machine learning approaches.
\paragraph{Semantic Parsing}
On the COGS semantic parsing dataset, $\mathcal{L}$-homomorphism-based data augmentation substantially outperforms a baseline LSTM model. Its results are worse than can be achieved with training schemes specifically tailored toward semantic parsing tasks,
but comparable to performance of a generic neural sequence architectures specially targeted at compositional generalization (the LexLSTM model of \citealp{akyurek2021lexicon}). We note that unlike data augmentation with $\mathcal{L}$-homomorphisms, none of COGS models can be applied to VQA. These experiments highlight the flexibility of our approach, but do not rule out the possibility that stronger forms of compositional generlization may be obtained with approaches more narrowly tailored to specific problem domains.
\section{Other Related Work}
\label{sec:otherrelated}
\paragraph{Lexicalized neural models}
Word-level alignments between input and output sequences were an essential feature of statistical phrase- and tree-based sequence models \cite{chiang2005hiero, Koehn2003StatisticalPT}.
Neural scoring functions were sometimes integrated into these models \cite{misra2016neural}.
Neural models with attention \cite{bahdanau2014neural} do not require explicit alignment, though several pieces of past work have shown that incorporating explicit token-level correspondences improves generalization \cite{akyurek2021lexicon, Prabhu2020MakingAP, pham2018towards}.
To the best of our knowledge, this paper is the first to make use of token-level alignments in joint neural models of text and images.
\paragraph{Data Augmentation}
Data augmentation approaches are widely used across machine learning application domains featuring known invariances of the data distribution \cite{japkowicz2000learning, data-recombination-copy, shaw2020compositional}.
Substitution-based data augmentation schemes that replace words with synonyms, or synchronously replace words and their translations, are widely
used in tasks like machine translation \cite{wang2018switchout, wei2019eda}. Previous work use substitution of pronouns to de-bias NLP models for gender. Back-translation \citep{sennrich2015improving}, which encourages cycle-consistency in sequence transduction, may also be viewed as a form of data augmentation.
\section{Conclusion}
We have presented a new lexicon-based data augmentation method that improves compositional generalization of neural models in multiple domains. This method is derived from a new characterization of the principle of compositionality as a constraint on the symmetries of data distributions, and a procedure for automatically identifying these symmetries using token-level alignments.
Our results highlight the fact that many of the inductive biases targeted by specialized models in NLP can be alternatively, and often more flexibly, expressed as a hypothesis about the structure of the dataset to be modeled.
\section*{Acknowledgements}
This work was supported by the MachineLearningApplications initiative at MIT CSAIL and the MIT--IBM Watson AI lab. Computing resources were provided by a gift from NVIDIA through the NVAIL program and by the Lincoln Laboratory Supercloud.
\bibliographystyle{acl_natbib}
|
\section{Introduction}
\label{sec:intro}
The recently commercialized persistent memory (PM) devices, represented by Intel Optane Persistent Memory (Optane PMem)~\cite{DCPMM}, deliver persistence, high capacity, lower cost and fast speed on the memory bus.
There have been many PM-inspired (re)designs in data-intensive systems~\cite{Viper,Spitfire,Zen,Hymem,SOFORT,Arulraj2015,FEDB}.
In particular, much exciting progress has been made on devising single-level persistent OLTP indexes that directly operate and store data on PM without involving the storage stack, even before real PM devices became available~\cite{Hwang2018,Chen2015,HiKV,NV-Tree,WORT,Venkataraman2011,PMwCAS,DBPCM,FPTree,BzTree}.
As we have shown in the past~\cite{PiBench}, although these pre-Optane proposals do not perform as expected (e.g., much slower) on real PMem devices due to inaccurate assumptions and emulation, several building blocks (unsorted nodes, fingerprinting, hybrid DRAM-PM structures) have proved useful for devising new indexes on PM.
The availability of real PM devices has further enabled a new breed of PM-based indexes~\cite{Dash,roart,utree,dptree,lbtree,pactree}, which are tailor-made for Intel Optane PMem.
Although they appear/claim to perform better than pre-Optane proposals on real PM, it remains unclear (1) \textit{how they compare against each other}, as these new indexes appeared roughly concurrently and/or were proposed by different research communities (e.g., VLDB/SIGMOD vs. SOSP/OSDI), (2) \textit{how they are different from (or similar to) the pre-Optane proposals} (i.e., what ``legacy'' has pre-Optane proposals left for the new breed?), and (3) \textit{what further challenges and opportunities remain in this area}.
The goal of this paper is to answer these questions, which is the key to pushing actual adoption of these new indexes and PM-based systems in general.
We do so by conducting a comprehensive evaluation of five state-of-the-art PM indexes, including \dptree~\cite{dptree}, \utree~\cite{utree}, \lbtree~\cite{lbtree}, \roart~\cite{roart} and \pactree~\cite{pactree}.
We also include as a reference FPTree~\cite{FPTree}, a representative and arguably the best-performing design from the pre-Optane era~\cite{PiBench}.
This work can be seen as a sequel to the ``first episode'' of our past work which benchmarked and distilled the aforementioned useful building blocks from pre-Optane era PM indexes~\cite{PiBench}, but is not a mere repeat of what was done before with newer indexes.
On the one hand, we follow the similar methodology and focus on representative range indexes.
On the other hand, in this paper we give a snapshot of the latest state of this area and highlight new findings, observations and perspectives that were often omitted in the past, especially variable-length key support, NUMA-awareness and new potential impact of existing PM indexes on future work.
We summarize our findings below.
\circled{1} The key optimization target remains to be reducing read and (more often) write operations on PM to preserve PM bandwidth (and thus achieving higher performance).
\circled{2} As we predicted in past work, several pre-Optane techniques are effective on real PM hardware and widely adopted by new proposals, including fingerprinting~\cite{FPTree}, unsorted nodes and leveraging DRAM.
\circled{3} Although modern PM-enabled servers are often multi-socket, most state-of-the-art PM indexes are still only optimized for single-socket machines and consciously avoided NUMA effect in their experimental evaluations;
handling NUMA effect remains an unsolved problem.
\circled{4} PM programming infrastructure (e.g., PM allocators and runtime) is still far from ideal in many cases and requires further improvements.
\circled{5} Finally and perhaps most profoundly, for the first time, we observe that techniques employed by PM indexes can also be useful in devising high-performance volatile in-memory indexes.
Surprisingly, when running on DRAM without the extra fences and cacheline flushes, under certain workloads a PM index can match or even outperform well-tuned indexes specifically designed for DRAM, such as HOT~\cite{HOT} and Masstree~\cite{Masstree}.
This highlights the potential of unifying persistent and volatile indexing to simplify the design and implementation of future systems.
The remaining sections expand on more details and insights.
We give the necessary background on PM in Section~\ref{sec:bg}, followed by a review of pre-Optane PM range indexes in Section~\ref{sec:pre-optane}.
Sections~\ref{sec:new}--\ref{sec:analyze-new} then survey and analyze state-of-the-art PM range indexes.
Sections~\ref{sec:eval}--\ref{sec:observe} present our experimental results and observations.
We present an outlook of future indexes in Section~\ref{sec:next}.
Section~\ref{sec:related} covers related work and Section~\ref{sec:conclusion} concludes this paper.
We have open-sourced our code and results at \url{\vldbavailabilityurl}.
\section{Persistent Memory}
\label{sec:bg}
We briefly review the properties of PM (in particular Intel Optane PMem\footnote{Also known as Optane DC Persistent Memory Module (DCPMM)~\cite{DCPMMBrief}.}) and its implications on software; readers already familiar with the literature may skim and fast forward to Section~\ref{sec:pre-optane}.
Collectively, ``PM'' refers to a class of devices that offer byte-addressability (like DRAM), high capacity and persistence (like SSDs) on the memory bus.
They can be built using various materials and techniques, such as PCM~\cite{PCM}, STT-RAM~\cite{STT-RAM} and memristor~\cite{Memristor}.
However, only the 3D XPoint~\cite{Intel3DXP} based Optane PMem so far has delivered high capacity promised by the PM vision.
Thus, we target Optane PMem in this paper.
Optane PMem's performance generally fall between DRAM and SSDs.\footnote{Exceptions apply under certain workloads and hardware configurations~\cite{SSDStrikingBack}.}
The first generation (100 series) exhibits $\sim$300ns latency and $\sim$5--40GB/s bandwidth depending on the access pattern, and with sequential/read accesses being faster than random/write accesses.
The latest 200 series further gives $\sim$30\% higher bandwidth~\cite{Optane200}.
Although PM offers persistence,
it is still behind multiple levels of CPU caches.
Software normally goes through CPU caches to access data stored on PM using \load and \store instructions, and explicitly issues cacheline flush instructions (e.g., \CLWB and \CLFLUSHOPT) and fences to ensure data is correctly persisted~\cite{IntelManual}.
However, PM-capable platforms use asynchronous DRAM refresh (ADR) to guarantee data flushed from the CPU caches will first land on the CPU's write buffer, which is power failure protected.
As a result, a PM write is considered complete once the data is forced to the ADR domain, without necessarily having arrived at physical PM media.
More recent platforms further feature enhanced ADR (eADR) which also protects the CPU cache, effectively providing durable CPU caches and potentially sparing the need of cacheline flush instructions (although fences are still needed for correct ordering).
eADR is only available for the more recent 200-series Optane PMem and Skylake platforms~\cite{IntelManual}.
Our experiments are based on the 100-series PMem and Cascade Lake platform without eADR (detailed setups in Section~\ref{sec:eval}).
However, we do not expect our conclusions to change based on recent results obtained from evaluating the impact of eADR~\cite{PMIdio}.
We leave it as future work to document eADR's detailed behavior and impact on PM range indexes.
Optane PMem can operate under the Memory, App Direct, or Dual modes~\cite{IntelManual}.
The Memory mode uses the system's DRAM as a hardware-controlled cache and presents bigger but volatile memory.
The App Direct mode enables persistence, thus allowing building persistent indexes.
The Dual mode combines both by allowing part of PM to be configured for the Memory or App Direct mode.
Same as other work, we focus on the App Direct mode as it gives software the flexibility to use DRAM and PM as needed.
\section{Previously on PM Range Indexes}
\label{sec:pre-optane}
Indexing has received much attention even before real PM was available~\cite{FPTree,BzTree,NV-Tree,Chen2015,Venkataraman2011,DBPCM,WORT,PMwCAS,Hwang2018}.
This section reviews pre-Optane designs to set the stage for our new evaluations.
\subsection{Early Assumptions in the Pre-Optane Era}
Due to the lack of real devices, early proposals had to ``guesstimate'' the properties of PM based on prototypes and simulations~\cite{PCM,ZhouPCM,LeePCM}.\footnote{Various materials can be used to build PM, yet they may perform differently. This partially forced past work to make general assumptions for potentially wider applicability.}
Common assumptions included (1) limited write endurance, (2) 3--5$\times$ higher latency but similar bandwidth to DRAM's~\cite{NVM-DLog,Pelley2014}, (3) writes slower than reads, (4) 8-byte atomic PM writes~\cite{BPFS}, and (5) volatile CPU caches and reorderings by the CPU.
Out of these, the assumption on bandwidth turned out to be inaccurate: Optane PMem's lower-than-DRAM bandwidth is a major factor that limits performance~\cite{PiBench}.
The speed gap between sequential and random accesses was also largely left out.
Endurance so far has not been a major issue for Optane PMem which warrants virtually unlimited endurance during the usual replacement cycle of 3--5 years~\cite{DCPMMEnduranceReport,DCPMMBrief}.
Nevertheless, these assumptions rightfully suggested that classic in-memory indexes would not guarantee correctness (customized recovery protocols are necessary), nor perform well on PM.
Concurrency control must be carefully considered given PM's higher latency.
These led to the development of numerous PM indexes even before actual PM hardware was available.
The key is reducing PM read/write operations and avoiding unnecessary cacheline flushes and fences to both improve performance and reduce wear.
\subsection{Presumed Designs and Building Blocks}
\label{subsec:pre-optane-designs}
To reduce PM accesses, several building blocks have been proposed around re-designing the tree architecture, node structure and concurrency control.
Now we discuss them in the context of previously evaluated PM indexes (\wbtree~\cite{Chen2015}, \bztree~\cite{BzTree}, \fptree~\cite{FPTree} and \nvtree~\cite{NV-Tree}).
Table~\ref{tbl:indexes} lists their main design choices.
The first three dimensions (architecture, node structure and concurrency) received the most attention in the past;
in this work, we also consider variable-length keys, PM management and NUMA-awareness.
\textbf{Tree Architecture: PM-Only $\rightarrow$ DRAM-PM Hybrid.}
Most pre-Optane proposals adapt in-memory B+-trees.
Some (e.g., \wbtree and \bztree) place the entire tree on PM to allow instant recovery.
But doing so leads to much slower traversal speed compared to DRAM indexes.
A key contribution by \fptree and \nvtree was to leverage the fact that B+-tree's inner nodes only guide search traffic and are reconstructible using leaf nodes.
This allows improved traversal speed by loosening the consistency requirements for inner nodes, by omitting flushes/fences~\cite{NV-Tree} and/or placing all the inner nodes in DRAM~\cite{FPTree}.
Upon restart, the inner nodes can be rebuilt using B+-tree bulk loading algorithms.
The downside is recovery time can scale with data size, sacrificing instant recovery.
\textbf{Node Structure: Sorted $\rightarrow$ Unsorted.}
Traditional B+-trees keep keys sorted for fast binary search.
The drawback of inheriting this design on PM is insertions may shift keys, causing excessive PM reads and writes, while having the risk of incomplete updates upon hardware failure.
Moreover, binary search becomes less (or not at all) beneficial with small nodes commonly used by in-memory B+-trees.
A popular solution is to keep nodes unsorted and use a linear scan to retrieve the target key.
Most of the surveyed pre-Optane indexes adopted this technique.
To mitigate the impact of linear search, BzTree periodically consolidates nodes to become sorted.
FPTree proposed fingerprinting which maintains one-byte hashes of the keys in the node and a lookup starts by checking the fingerprints.
Only the keys with matched fingerprints will be further examined.
This greatly reduces PM accesses, especially for negative search where the key does not exist.
Another approach is to keep an indirection array~\cite{Chen2015} that stores sorted index positions of keys, allowing binary search using the indirection array.
\textbf{Concurrency Control: Pessimistic $\rightarrow$ Optimistic.}
PM indexes often prefer lightweight concurrency control over pessimistic lock coupling~\cite{CowBook}.
\fptree uses separate strategies for inner and leaf nodes: For the former it leverages hardware transactional memory (HTM) to reduce traversal costs, and for the latter it uses traditional locking because inserting into PM-resident leaf nodes may involve cacheline flushes, which in turn will abort HTM transactions.
\bztree uses lock-free multi-word compare-and-swap (PMwCAS)~\cite{PMwCAS} that can atomically modify multiple 8-byte words.
Without pessimistic locking, \fptree scales to high core count but exhibits high tail latency and low throughput under contention, due to HTM's inherent limitations~\cite{PiBench}.
\fptree delegates the detailed HTM-locking interactions
to Intel TBB which uses a fast-path that uses HTM and a slow-path that serves as a fallback when HTM abort rate becomes higher than a threshold (10 by default).\footnote{Details at \url{https://github.com/oneapi-src/oneTBB/blob/v2021.5.0/src/tbb/rtm\_mutex.cpp\#L33}. Our evaluation (Section~\ref{sec:eval}) sets the threshold to 256 for better performance.}
\subsection{Functionality and PM Management}
\label{subsec:pre-optane-func}
Now we turn to the remaining three dimensions under consideration in Table~\ref{tbl:indexes}: support for variable-length keys, NUMA awareness and PM programming infrastructure support.
Most pre-Optane proposals~\cite{FPTree,Chen2015,NV-Tree} focused on handling fixed-length, 8-byte keys, so were past evaluation efforts.
Variable-length keys are usually supported using pointers to keys stored in the (persistent) heap.
Some proposals differentiate pointers from inlined values by designating a special ``type'' bit in the 8-byte key area, effectively limiting the maximum length of keys to be 63 bits.
Others, e.g., \fptree, require compile-time customization by specifying whether keys are 8-byte integers or pointers to support full 64-bit keys.
Pointers are used in case both types are required.
As a result, accessing shorter keys $\le$ 8 bytes is also prone to cache misses caused by pointer chasing.
Out of the four indexes, only \bztree provides inlined support for variable-length keys with slotted pages.
None of the surveyed pre-Optane indexes handle NUMA effect.
Most of them also do not have a well thought-out design for managing PM.
\fptree uses PMDK~\cite{PMDK}, the current de facto standard PM library, to avoid issues such as PM leaks.
For better performance, \fptree has to use customized slabs (large chunks allocated from PMDK allocator) to amortize PM allocation cost.
Still, our previous work~\cite{PiBench} has identified PM allocation in all these indexes as a main bottleneck that should be removed by future designs.
\section{State-of-the-Art PM Range Indexes}
\label{sec:new}
We survey five representatives of recent PM indexes optimized for Optane PMem: \lbtree~\cite{lbtree}, \dptree~\cite{dptree}, \utree~\cite{utree}, \roart~\cite{roart} and \pactree~\cite{pactree}.
They can be categorized as B+-tree based, trie based and hybrid which makes use of both B+-trees and tries.\footnote{In addition to traditional indexes, learned indexes~\cite{LearnedIndex} are also being adapted for PM~\cite{APEX}. However, they are still in very early stage. We thus leave it as future work to evaluate them to avoid pre-mature conclusions.}
\subsection{B+-Tree based: \lbtree and \utree}
\label{subsec:btree}
We survey two representative B+-tree based PM range indexes, \lbtree~\cite{lbtree} and \utree~\cite{utree}, which represent designs that mainly optimize for high throughput and low tail latency, respectively.
\textbf{\lbtree.} As evaluated by previous work~\cite{PiBench,UCSDMeasurement,UCSDGuide}, Optane PMem uses 256-byte granularity internally: accesses smaller than 256 bytes will still incur 256-byte of traffic to/from the physical media.
It then becomes important to (1) coordinate PM accesses in 256-byte granularity and (2) reduce unnecessary PM accesses (to save PM bandwidth)~\cite{UCSDGuide,UCSDMeasurement}.
To reach these goals, as shown in Tables~\ref{tbl:indexes} and \ref{tbl:new-indexes}, \lbtree starts with several useful techniques proposed by \fptree: fingerprinting, DRAM+PM architecture and optimistic concurrency with HTM and locking.
Compared to \fptree, it uses hand-rolled RTM transactions instead of TBB.
We highlight the impact of this design decision later in Section~\ref{sec:eval}.
On top of these existing techniques, \lbtree proposes new ones to further optimize PM accesses.
First, node sizes are set to align with and be multiples of 256 bytes for better CPU cache and PM utilization.
Second, to minimize PM writes during inserts, \lbtree proposes entry moving to bound the number of cacheline writes per insert.
As Figure~\ref{fig:trees}(a) shows, a 256-byte leaf node is divided into four 64-byte cachelines.
Upon insert, \lbtree first attempts to insert the record into \texttt{Line 0} if an empty slot is available in it.
Since the header is also in \texttt{Line 0}, only one cacheline write to PM is needed to persist both the record and header.
Otherwise, \lbtree moves data from \texttt{Line 0} to another line where the new entry is inserted.
This proactively spares empty slots in the first cacheline, which will reduce flush operations incurred by future inserts.
Write-ahead logging (WAL) is widely used for crash recovery ~\cite{PMDK}, but incurs additional PM writes.
\lbtree disposes of WAL with logless node splits by storing two \texttt{sibling} pointers in each leaf node and uses an alt bit in node header to indicate the valid pointer.
Upon split, a new leaf node is first allocated and tracked by the unused pointer, followed by a redistribution of entries to the last two cachelines of the new node.
Then the alt bit and bitmap in the original node are updated using an 8-byte atomic PM write
to ensure the tree is always in a consistent state even across failures.
\textbf{\utree.}
Different from most PM indexes, \utree mainly optimizes for tail latency caused by PM's high latency.
It places B+-tree inner nodes in DRAM, but redesigns leaf nodes to use both DRAM and PM.
As shown in Figure~\ref{fig:trees}(b), leaf nodes in \utree consist of two layers: an array layer and a list layer.
The former consists of traditional B+-tree nodes which store pointers to list layer nodes; the latter is a PM-resident singly-linked list where each node stores a key-value pair.
\utree uses optimistic locking~\cite{ARTOLC} for the DRAM-resident B+-tree and manages the PM-resident linked list in a lock-free manner.
To insert a record, the thread first traverses the in-DRAM B+-tree optimistically without holding any locks, and then inserts a node representing the key-value pair in the linked list using the compare-and-swap (\CAS) instruction~\cite{PracticalLockFree,HarrisLockFreeLinkedList}.
It then acquires the lock in the corresponding leaf node in the B+-tree to insert the key, which may trigger splits that will in turn acquire locks from the bottom up in DRAM.
This way, B+-tree structural modification operations (SMOs) and actual key-value inserts/removals (which incur PM flushes in the list layer) are decoupled, and multiple threads inserting into the same leaf node could proceed in parallel in the list layer.
This could allow more concurrency and removes flushes from the critical path, thus reducing tail latency.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{paper-plots/indexes/trees.pdf}
\caption{Architecture of five state-of-the-art PM indexes.}
\label{fig:trees}
\end{figure}
\subsection{Trie based: \roart and \pactree}
\label{subsec:trie}
Newer designs have also explored tries on PM.
They are mostly PM-only (instead of DRAM-PM hybrids) and based on ART~\cite{ART}.
We survey two representative proposals, \roart~\cite{roart} and \pactree~\cite{pactree}.
\textbf{\roart.}
Based on ART, \roart optimizes for range scans, which are known to be a weak point of tries~\cite{roart}.
Figure~\ref{fig:trees}(c) shows the overall architecture of \roart.
It is purely on PM but selectively persists metadata entries and reconstructs the inconsistent ones upon recovery (similar to \nvtree's selective consistency~\cite{NV-Tree}).
To optimize scan, \roart compacts small sub-trees ($<$ 64 entries) into leaf arrays that store pointers to records.
This reduces pointer chasing overhead during scan and makes the index shallower.
The downside is splits become expensive, as all the keys in a leaf array are divided into subsets based on the first differentiating byte, followed by a node allocation for each subset.
This requires more PM writes and fences, and puts higher pressure on the PM allocator.
To this end, \roart reduces the number of fences by relaxing the order of split steps and using a depth field to detect and resolve inconsistency.
It also proposes delayed check memory management (DCMM) to cope with the high PM allocation demand.
DCMM performs fast allocations using thread-local pools but delays garbage collection by traversing the entire index at a later time, potentially increasing PM usage.
Since \roart is fully in PM, it supports true instant recovery, and natively supports variable-length keys without extra pointer chasing using its trie-based design.
\begin{table*}[t]
\caption{Main design choices of state-of-the-art PM range indexes (\lbtree, \dptree, \utree, \roart and \pactree).
They base on Intel Optane PMem and inherit designs from pre-Optane proposals, by using DRAM (sometimes more aggressively), lightweight concurrency control and unsorted nodes.
Some also advocate customized PM allocators to reduce PM accesses.
Variable-length keys and NUMA-awareness are still less considered.
}
\label{tbl:new-indexes}
\begin{tabular}{@{}m{1.9cm}m{3.2cm}m{3.4cm}m{2.8cm}m{1cm}m{2.3cm}m{1cm}@{}}
\toprule
\textbf{} & \textbf{Architecture} & \textbf{Node Structure} & \textbf{Concurrency} & \textbf{Var. Keys} & \textbf{PM Allocator} & \textbf{NUMA-Aware} \\ \midrule
\textbf{\lbtree}~\cite{lbtree} & B+-tree; DRAM (inner) + PM (leaf) & Unsorted leaf; fingerprints; extra metadata& HTM (inner) + locking (leaf) & Pointer & Customized/PMDK & No \\
\rowcolor[HTML]{EFEFEF}
\textbf{\utree}~\cite{utree} & B+-tree; DRAM (B+-tree) + PM (linked list) & Sorted & Locking (array layer) + lock-free (list layer) & Pointer & PMDK & No \\
\textbf{\dptree}~\cite{dptree} & Hybrid; DRAM (B+-tree, trie inner) + PM (trie leaf) & Unsorted leaf; fingerprints; indirection; extra metadata & Optimistic lock~\cite{ARTOLC} + async. updates & Pointer & PMDK & No\\
\rowcolor[HTML]{EFEFEF}
\textbf{\roart}~\cite{roart} & Trie; PM-only & B+-tree like unsorted leaf; fingerprints & ROWEX~\cite{ARTOLC} & Inlined & Customized/PMDK & No\\
\textbf{\pactree}~\cite{pactree} & Trie; PM-only or optionally DRAM+PM & Unsorted leaf; fingerprints; indirection & Optimistic lock~\cite{ARTOLC} + async. update & Inlined & Customized/PMDK & Yes \\ \bottomrule
\end{tabular}
\end{table*}
\textbf{\pactree.}
\pactree is a PM-only two-layer persistent trie with B+-tree styled leaf nodes.
As shown in Figure~\ref{fig:trees}(d), \pactree consists of a search layer and a data layer.
The search layer is a durable trie based on concurrent ART that uses read-optimized write exclusion (ROWEX)~\cite{ARTOLC}.
The data layer is a doubly-linked list of B+-tree leaf nodes, each of which contains 64 key-value pairs and an anchor key to indicate the smallest key in the node.
\pactree stores fingerprints and indirection arrays in leaf nodes to facilitate search and scan, but they are not persisted to reduce PM writes.
Upon split, the target leaf node in the data layer is first locked, and then a log entry is written to a per-thread SMO log in PM.
The thread then splits the leaf node and commits without modifying inner nodes in the search layer.
A background thread will then finish the remaining SMO in the search layer.
This allows worker threads to commit early right after modifying leaf nodes.
However, it also creates inconsistencies between the search and data layers.
Thus, query threads may need to perform a ``last-mile'' search after arriving at the data layer to find the correct leaf node using anchor keys.
\pactree is the only index that mitigates NUMA effect.
It uses separate pools for the search layer, data layer and logs in each NUMA node.
It also advocates the use of snooping-based CPU coherence protocols (instead of the default directory-based protocol on most platforms) to avoid poor performance when PM accesses cross NUMA boundaries.
\subsection{Hybrid: \dptree}
\label{subsec:hybrid}
As shown in Figure~\ref{fig:trees}(e), \dptree is a PM-DRAM hybrid index that combines up to two B+-trees (a \textit{front} and a \textit{middle} buffer tree) in DRAM, a trie (\textit{base} tree) that places inner nodes in DRAM and leaf nodes in PM.
To search for a key, \dptree first visits the front buffer tree.
If the target key is not found, the middle buffer tree (if exists) will be further searched.
If the key does not exist in the buffer trees, the base tree will have to be searched.
To reduce unnecessary traversals, \dptree maintains a bloom filter per buffer tree.
For range queries, \dptree has to search and merge results from all the trees.
When the size ratio between the front buffer tree and base tree reaches a pre-defined threshold, \dptree creates a new front buffer tree and turns the previous front buffer tree into a middle buffer tree.
Tree merge operations are triggered when the size ratio between the front buffer tree and the base tree reaches a pre-defined threshold, and are performed by background threads.
Using the version number and extra set of metadata, \dptree ensures changes are invisible to concurrent queries when a merge is in progress.
After merging, the middle buffer tree is destroyed and the inner nodes of the base tree (ART) are rebuilt.
Then the global version bit is flipped to expose changes to incoming requests.
\dptree also uses selective metadata persistence with reconstructible metadata (e.g., record count and fingerprints) in DRAM.
\section{Analyzing the State-of-the-Art}
\label{sec:analyze-new}
With the high-level designs laid out, now we analyze the new PM indexes in detail and distill common building blocks which can be useful for future PM indexes.
We analyze them from the six dimensions in Table~\ref{tbl:new-indexes}, followed by empirical evaluation in Section~\ref{sec:eval}.
\subsection{Index Architecture}
New PM indexes often inherit the PM+DRAM architecture with new optimizations, and consider tries and hybrid structures.
\textbf{(More Extensive) Use of DRAM.}
New PM indexes based on B+-tree and hybrid structures (\lbtree, \dptree and \utree) continue to use DRAM to store part of the index.
We also observe more aggressive use of DRAM, e.g., \dptree and \utree place entire tree structures in DRAM to get more performance gains.
The tradeoffs are (1) longer recovery time, (2) more complex programming and (3) higher DRAM consumption which we quantify in Section~\ref{sec:eval}.
\textbf{Beyond B+-Trees and Monolithic Indexes.}
New PM indexes also adapt tries (\pactree and \roart), but they default to pure PM designs, potentially leading to suboptimal performance but preserving instant recovery.
Notably, \pactree stores keys in both inner and leaf nodes, but it is possible to place inner nodes in DRAM to improve performance.
We expect to see more B+-tree based PM indexes utilize DRAM for faster access/write speed, as PM servers will still feature DRAM in the foreseeable future.\footnote{DRAM must be present for PMem to work, even if the software does not need it~\cite{SSDStrikingBack}.}
As a major departure from just adapting one type of data structure, \dptree combines B+-trees and tries.
When it comes to node structure, new indexes are also introducing new designs that no longer use pure trie or B+-tree nodes, which we highlight next.
\subsection{Node Structure}
New PM indexes base their designs on Optane PMem with node alignment of 256 bytes to reduce unnecessary PM accesses.
Several pre-Optane designs---fingerprinting, unsorted (leaf) nodes and selective consistency for metadata---continue to be used by new PM indexes.
But they are further optimized with new techniques.
\textbf{Fingerprinting on Steroids.}
As PM accesses are slow, fingerprinting becomes the most popular approach used by four out of the five new PM indexes.
Meanwhile, new techniques are introduced to better store and use fingerprints.
\lbtree uses SIMD instructions to compare up to 64 one-byte fingerprints in one instruction.
\roart embeds a two-byte fingerprint inside pointers to key-value pairs, minimizing pointer chasing at the leaf level.
\textbf{Extra and Selectively Persisted Metadata.}
\lbtree and \dptree both use an extra set of metadata per leaf node to avoid logging (thus reducing PM writes).
For \lbtree, this allows it to achieve logless split.
\dptree uses the extra metadata to track PM allocations and hide incomplete changes during tree merge operations.
Not all metadata entries have to be persisted when modified.
For instance, version locks in \pactree leaf nodes are only meaningful at runtime;
fingerprints and indirection arrays can be rebuilt during recovery (\dptree and \pactree) or on demand at runtime by query threads (\roart).
Keeping them volatile can significantly reduce PM writes and improve performance, at the cost of slower recovery.
\textbf{Hybrid Leaf Nodes.}
All the surveyed PM indexes that adopt trie (\dptree, \roart and \pactree) use B+-tree like leaf nodes where a node stores multiple records to reduce insert overhead, as the leaf node is only split when it is full.
This design also reduces pressure on the PM allocator, as trie-based indexes typically incur more frequent allocations with varying node sizes compared to B+-trees.
Scan performance is also improved with less pointer chasing.
Different from other proposals, \utree introduces a linked list layer in PM, making its ``leaf node'' logical.
Although this design decouples SMOs and data movement to potentially enable more parallelism, it adds more overhead to scans due to more pointer chasing.
\subsection{Concurrency Control}
All the surveyed new PM indexes use optimistic concurrency control.
They optimize traversals using lock-free read or HTM for inner nodes;
locks are only acquired at the leaf level and/or as needed in inner nodes to reduce PM writes.
Further, they tend to use background threads for SMOs (e.g., \pactree and \dptree).
The benefit of offloading SMOs to the background is potentially lower latency for index operations.
However, it can be tricky to determine the appropriate number of background threads.
Also, with a given CPU budget (e.g., in the cloud), the machine may not have enough resources to spare for the background threads, which may then fall behind the foreground threads and affect the overall progress.
\subsection{Functionality and PM Management}
As Table~\ref{tbl:new-indexes} lists, among the new PM indexes, only trie-based \roart and \pactree natively support variable-length keys; the others follow pre-Optane proposals to use pointers to keys.
With real hardware and libraries like PMDK, all the indexes have taken into account PM management issues (e.g., avoiding persistent leaks and optimizing allocation performance).
However, many need a customized allocator for performance reasons.
Finally, only \pactree is designed to mitigate NUMA effect.
Other indexes would have to use general-purpose approaches~\cite{NAP} that can be applied on any PM index, but they come with limitations (e.g., focus on certain workloads).
Such facts indicate that in terms of functionality, new PM indexes have been mainly sticking with the status quo.
\section{Evaluating the State-of-the-Art}
\label{sec:eval}
Now we empirically evaluate new PM indexes and compare them with \fptree, the best-performing pre-Optane PM range index.
\subsection{Experimental Setup}
We run experiments on a 40-core (80-hyperthread) server equipped with two Intel Xeon Gold 6242R CPUs clocked at 3.10 GHz with 36MB of cache.
The server is fully populated with 12$\times$32GB DRAM DIMMs (384GB in total) and 12$\times$128GB Optane PMem DIMMs (1.5TB in total)
for maximum bandwidth.
Both DRAM and PMem run at 2666MT/s.\footnote{DRAM has to be clocked down from 3200MT/s to 2666MT/s for PMem to work~\cite{DCPMMBrief}.}
The server runs Arch Linux with kernel 5.14.9.
All the code is compiled using GCC 11.1 with all the optimizations.
Unless otherwise specified, we use PMDK~\cite{PMDK}/\jemalloc~\cite{jemalloc} for PM/DRAM allocations.
\textbf{Benchmarking Framework.}
We use \pibench~\cite{PiBench}, a unified framework for benchmarking PM indexes, to stress test the indexes.
\pibench generates and issues synthetic workloads of given distributions that consist of user-specified operations (lookup/insert/update/delete/scan).
It requires each index implement a set of common interfaces, by extending an abstract C++ class.
We create a wrapper for each index that uses \pibench's interfaces to invoke the index's internal operations.
The wrapper is compiled as a shared library and loaded into \pibench's address space at runtime.
\textbf{Metrics and Workloads.}
We measure throughput (operations per second) and latency at various thread counts.
Using \pibench, we collect statistics such as cache misses and bandwidth to aid analysis.
We test both fixed-length (8-byte) integer and variable-length keys, following the setups used by previous work~\cite{PiBench}.
For point queries we use keys chosen randomly under uniform or skewed distributions;
for range scans, we uniform randomly choose a start key $K$ and scan 100 records following $K$.
We prefill each index with 100 million key-value pairs, after which we start to run individual and mixes of index operations.
The results across runs vary little and so we report the average of three 10-second runs.
\subsection{Index Implementations and Parameters}
For all the indexes (except \fptree which we had to implement), we use the original authors' code obtained from their public repositories.
We use parameters that lead to the best performance (described below) and make necessary changes to each index for correctness, functionality and fairness.
\textbf{\lbtree}.
(1) The original insert and delete functions do not guarantee persistence;
we followed the authors' suggestions to fix them.\footnote{Details at \url{https://github.com/schencoding/lbtree/issues/2}.}
(2) We implemented range scan (similar to \fptree) and update with locking in leaf nodes.
(3) We found the PMDK allocator can provide sufficient performance after tuning PMDK parameters and allocating 256-byte aligned leaf nodes, so we also use PMDK for \lbtree.
The inner/leaf nodes contain 15/14 entries (256-byte).
\textbf{\utree}.
Instead of using chunk-based allocation, the original code in fact uses PMDK's \texttt{POBJ\_ZALLOC}.
We changed it to use \texttt{pmemobj\_alloc} for much better performance.
We were not able to verify the correctness of multi-threaded inserts (with keys missing after successful inserts),
so we only include \utree in single-threaded experiments.
Inner/leaf node sizes are both set to 29 entries.
\textbf{\roart}.
The open-source code of \roart supports both PMDK and its customized DCMM allocator~\cite{roart}; we present numbers under both allocators.
Leaf node size is set to 64 entries.
\textbf{\pactree}.
We use \pactree's own NUMA-aware PM allocation.
For fair comparison, we pin the background and worker threads to the same CPU cores so that all the indexes use the same amount of CPU cores.
Node size is set to 64 entries.
\textbf{\dptree}.
The original code misses PM allocator support, so we ported it to use PMDK.
We follow the original paper to use an equal number of worker and merge threads.
For fair comparison, we also pin the merge and worker threads to the same cores (similar to \pactree's setup).
Inner and leaf nodes contain 31 entries for buffer trees;
the base tree uses 256-entry leaf nodes.
\textbf{\fptree}.
Since the original implementation is proprietary, we implemented \fptree by strictly following the paper.\footnote{Code available at \url{https://github.com/sfu-dis/fptree}.}
We have verified that our implementation performs similarly to the original author's binary does.
We follow past work's recommendations to set inner/leaf node sizes to 128/64 entries~\cite{PiBench,FPTree}.
\subsection{Single-threaded Performance}
\label{subsec:single-thread-exp}
We begin with single-threaded experiments to avoid concurrency complicating our analysis.
We run lookup/insert/update/scan operations under the uniform random distribution and report throughput.
\textbf{Point Queries. }
\lbtree, \dptree and \utree perform similarly for lookups in Figure~\ref{fig:pm-uniform-1t}(a).
They are up to $\sim$2$\times$ faster than \fptree, \roart and \pactree, which also perform similarly.
Overall, \dptree performs the best under single-threaded lookups, largely because of its extensive use of DRAM:
if the search key is in the DRAM-resident buffer tree, the entire query can finish without ever accessing PM.
If the base tree needs to be visited, only the search in leaf node will incur PM access, which is mitigated by binary search.
The other two faster indexes (\lbtree and \utree) also benefit from placing inner nodes and the array leaf layer in DRAM.
\lbtree also extensively uses SIMD instructions and prefetching, which as shown in Figure~\ref{fig:cache_misses_pmem}(a) drastically reduces cache misses and our factor analysis showed that prefetching alone improves performance by $\sim$10\%.
For inserts, B+-tree based \fptree/\lbtree and hybrid \dptree are up to $\sim$2$\times$ faster than trie-based \roart and \pactree in Figure~\ref{fig:pm-uniform-1t}(b).
\utree's use of linked lists in the leaf level cancels out some of B+-tree's advantages due to high cache miss rates in Figures~\ref{fig:cache_misses_pmem}(b--c).
The performance of PM allocators is critical for \roart as for each update it needs to allocate a new PM block, and using its own DCMM can double the throughput for updates in Figure~\ref{fig:pm-uniform-1t}(c).
Compared to lookups, updates incur additional PM writes to update the payload, but will not trigger SMOs compared to inserts.
As expected, the update performance of all indexes falls between their lookup and insert performance, but follows the trend of insert performance more closely because
(1) PM exhibits higher write latency, and
(2) similar to inserts, leaf-level locks can only be released after the new value is flushed, adding delays (although being a constant amount of overhead under a single thread).
\textbf{Range Scans.}
As Figure~\ref{fig:pm-uniform-1t}(d) shows, range scan performance depends largely on the cost of scanning within and across leaf nodes, i.e., whether the nodes are sorted and they are big or small.
Although \dptree needs to search multiple trees and combine results, it is still the fastest.
\dptree's leaf nodes use indirection arrays, so the result can be returned directly without sorting, as oppose to trees that use unsorted nodes, e.g., \fptree.
\lbtree inherited a lot from \fptree, but is $\sim$30\% slower than \fptree for scans, because its leaf nodes are smaller (14 entries).
To scan for the same number of records, compared to \fptree which uses 64-entry nodes, more leaf nodes have to be visited by \lbtree, causing more cache misses in Figure~\ref{fig:cache_misses_pmem}(d).
This result highlights the tradeoff between hardware consciousness and optimization goals:
using small nodes allows \lbtree to perform well in point queries, but can penalize scans.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{paper-plots/plots/single_pmem_indexes_uniform_8b_100m_10s.pdf}
\caption{Single-threaded throughput (uniform distribution).
Overall, \dptree and \lbtree perform the best.
\fptree can be very competitive to (or even better than) newer indexes. }
\label{fig:pm-uniform-1t}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/cache_misses_pmem}
\caption{Last-level cache misses per operation under a single (a--d) and 20 threads (e--h).}
\label{fig:cache_misses_pmem}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/pmem_indexes_uniform_skewed_8b_100m_10s.pdf}
\caption{Throughput under uniform (a--d) and skewed (e--g, self-similar with 80\% accesses on 20\% of keys) distributions.}
\label{fig:pm-mt}
\end{figure*}
All the tested indexes except \roart first copy the scan results to an array and then optionally sort them before they are returned to the user.
\roart, however, first returns an array of leaf pointers without copying.
Moreover, to support variable-length keys, it stores in leaf nodes pointers to keys.
This mandates the sorting pass to dereference pointers to keys, causing extra cache misses.
We note that as an optimization, if key size is known, one may change \roart to copy keys first, which in our tests can double the performance at the cost of supporting variable-length keys.
Finally, \utree exhibits low scan performance because every record is stored in a linked list node, traversing them results in many cache misses.
\textbf{Summary.}
The most effective technique to achieve high performance remains leveraging DRAM,
which newer PM indexes adopt more aggressively, by putting more components or even complete trees in DRAM.
This is at the cost of higher memory consumption, more complex recovery protocol and higher cost of ownership.
Importantly, \fptree remains very competitive.
\pactree and \roart are only marginally faster than \fptree for lookups.
In Figure~\ref{fig:pm-uniform-1t}(b), \pactree, \roart and \utree are even slower than \fptree.
Only \dptree and \pactree perform better than \fptree for scans.
\subsection{Multi-threaded Experiments}
\label{subsec:multi-thread-exp}
Now we measure index throughput with different thread counts under uniform and skewed distributions.
We start with one socket and expand to NUMA with two sockets in Section~\ref{subsec:numa-exp}.
\textbf{Individual Operations.}
Figures~\ref{fig:pm-mt}(a--d) show the throughput of lookup/insert/update/scan under the uniform distribution;
the shaded areas (over 20 threads) indicate numbers obtained when hyperthreads are also used.
All indexes scale well for pure lookups in Figure~\ref{fig:pm-mt}(a), with \dptree and \lbtree achieving higher raw throughput than others.
This result aligns with that obtained in Section~\ref{subsec:single-thread-exp} under a single thread.
\dptree uses optimistic lock coupling for its buffer/base trees, and readers can traverse without incurring PM accesses.
Its bloom filter also helps avoid unnecessary lookups in the buffer trees.
\lbtree uses HTM which without write operations exhibits little/no aborts.
For inserts, although \lbtree does not perform the best under a single thread, it scales the best under multiple threads, by being 1.55$\times$/1.96$\times$/3.07$\times$/2.23$\times$ faster than \roart-DCMM/\roart-PMDK/\dptree/\pactree.
Although \lbtree inherits many designs from \fptree, its logless split, new node layout and inner node locks to avoid re-traversals during split further make it 2.09$\times$ faster than \fptree.
\dptree's performance stops scaling beyond 10 threads, mainly due to its 7-phase merge:
each time a merge occurs, records in the middle buffer tree will be moved into the base tree, causing the base tree's inner nodes to be rebuilt.
This in turn incurs high garbage collection costs.
\roart-DCMM achieves 1.27$\times$ higher performance using its customized PM allocator compared to using PMDK.
Not leveraging DRAM also contributes to its lower performance compared to others that do leverage DRAM.
For updates, \lbtree outperforms others in Figure~\ref{fig:pm-mt}(c), thanks to its fast traversal and node layout design.
\dptree is also very competitive, as updates are served in-place without triggering merge operations.
Similar to the single-threaded results, for all indexes, updates behave more similarly to inserts (than lookups) as leaf-level locks must be retained until the new value is flushed.
As shown in Figure~\ref{fig:pm-mt}(d), under multiple threads the relative merits of different indexes on range scan are similar to the single-threaded results.
The only exception and best performing index is \pactree.
It scales to 40 threads, thanks to the combination of
(1) its leaf node design that inlines key-value pairs and leverages PM's fast sequential read,
(2) indirection that avoids sorting,
and (3) optimistic concurrency that incurs no PM writes for reads.
The other trie-based \roart performs the worst although it specifically optimizes for scan since it does not use DRAM and incurs more cache misses (Section~\ref{subsec:single-thread-exp});
under high core counts, cache misses are further exacerbated in Figure~\ref{fig:cache_misses_pmem}(h).
For B+-tree variants, \lbtree performs much worse for scans due to its use of small nodes (more cache misses).
In contrast to the single-threaded results, \fptree performs poorly:
using larger leaf reduces cache misses, but increases lock contention on leaf nodes.
\dptree takes no locks for scans (OLC), but needs to visit multiple indexes and merge results, which contributes to its lower performance than \pactree.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/mixed_workload_pmem.pdf}
\caption{Throughput of mixed workloads (lookups + inserts) under uniform distribution.}
\label{fig:mixed_workload_pmem}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/1t_pmem_tail_latency.pdf}
\caption{Tail latency of PM indexes under uniform distribution and a single thread (a--c) and 20 threads (d--f).}
\label{fig:pm-latency}
\end{figure*}
\textbf{Skewed Accesses.}
Under the self-similar distribution where 80\% of accesses are focused on 20\% of all the keys~\cite{BillionSynthetic},
lookups as shown in Figure~\ref{fig:pm-mt}(e) exhibit a similar trend to uniform distribution but with higher raw throughput because of better CPU cache utilization (the working set is smaller).
For updates in Figure~\ref{fig:pm-mt}(f),
\dptree remains scalable as updates are all performed in the DRAM buffer tree.
\fptree shows unstable and low throughput for updates due to frequent HTM aborts and more PM writes caused by SMOs during update: updates to the same record are appended without deduplication, so the node can become full and get split during an update.
No index scales under update workloads with hyperthreading.
Under contention, locking takes over to become the main bottleneck in \fptree, despite the working set is smaller.
Like \fptree, \lbtree also uses locking for leaf nodes, but scales under scan because of their different ways of using HTM:
\fptree delegates HTM and locking to TBB (Section~\ref{subsec:pre-optane-designs}) which has a global fallback path after a pre-defined number (256) of aborts of HTM transactions, whereas \lbtree directly uses HTM instructions (e.g., \texttt{xbegin}/\texttt{xend}).
Although scan is read-only, the first leaf node lock (in PM) is acquired inside the HTM transaction at the end of traversal.
This incurs much contention under skewed workloads and triggers HTM aborts, leading \fptree to use the slow fallback path protected by a global mutex.
\dptree's optimistic locks allow high scan throughput under skewed accesses as no locks are acquired.
\pactree scales with the best performance under skewed accesses for scans, although the gain diminishes with more hyperthreads.
\roart scales slightly worse in skewed update: for each update it creates a new leaf and replaces the original leaf pointer inside the leaf array, which becomes more expensive under contention.
\textbf{Mixed Workloads.}
We test mixed workloads with different read/insert ratios: read heavy (90\% lookups + 10\% inserts), balanced (50\% lookups + 50\% inserts) and write heavy (10\% lookups + 90\% inserts).
As Figure~\ref{fig:mixed_workload_pmem} shows, \lbtree exhibits the best performance and scalability.
\dptree scales worse with more inserts as tree merge becomes a major overhead.
Finally, \fptree again remains very competitive with \roart, \pactree and \dptree.
\subsection{Tail Latency}
\label{subsec:latency}
We measure tail latency using the same approach from previous work to strike a balance between overhead and accuracy~\cite{PiBench}.
In detail, we sample 10\% of all the operations during each run under uniform distribution to rule out the impact of CPU caches.
Figure~\ref{fig:pm-latency} shows the tail latency at varying percentiles under one thread (a--c) and 20 threads (d--f).
As expected, we observe no obvious differences between one and 20 threads for lookups.
For inserts, lookups and scans in Figures~\ref{fig:pm-latency}(a--c), \utree shows consistently higher latency (and skyrockets at 99.99\% for inserts), although its key design goal is to reduce tail latency (Section~\ref{subsec:btree}).
We observe the reason is in its use of PM-resident linked lists with per-record nodes.
To handle an insert, \utree uses $\sim$2900 cycles to allocate a list node and $\sim$2200 cycles to complete a \CAS and cacheline flush to insert the allocated node into the linked list.
In contrast, \lbtree only needs around 80 cycles to insert a key into a leaf node.
Moreover, the use of linked lists in the leaf layer causes many cache misses during scans:
as shown in Figure~\ref{fig:cache_misses_pmem}(d), \utree exhibits the highest cache miss ratio.
\roart-DCMM exhibits lower latency than \roart-PMDK for inserts, thanks to the better DCMM allocator: in the worst case, a split in \roart could allocate 63 leaf arrays and one inner node.
\roart has relatively higher latency for scans in Figures~\ref{fig:pm-latency}(c) and \ref{fig:pm-latency}(f), due to pointer chasing at the leaf level to dereference pointers to keys.
The other trie-based \pactree directly stores keys in leaf nodes, hence exhibiting lower latency.
\pactree also shows relatively low latency for all operations in most cases (except inserts under 20 threads) because SMOs are offloaded to background threads.
Under 20 threads, the background threads start to fall behind, requiring worker threads to traverse extra nodes to reach the correct leaf node, thus increasing latency.
\subsection{Support for Variable-Length Keys}
\label{subsec:varkey}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/var_key_pmem.pdf}
\caption{Throughput under variable-length keys using synthetic and real-world datasets~\cite{LastMile}. }
\label{fig:var-key-pmem}
\end{figure}
Now evaluate variable-length key support.
We first run the same experiment as Section~\ref{subsec:multi-thread-exp} with 8-byte keys, but force the indexes to use their variable-length key support.
This may limit the depth of trie-based indexes (thus giving them advantages), but allows us to reason about the efficiency of the variable-length key support by comparing with the fixed-length key experiments; we use three real-world datasets later.
For most indexes, including \fptree, \lbtree and \dptree this means allocating the key in the heap and storing a pointer to the key in the index.
\roart uses its own native support to store keys in index nodes themselves.
All indexes are plotted except \pactree.\footnote{\pactree assumes null-terminated strings. This is incompatible with \pibench which may generate keys with \texttt{$\backslash$0}, which will be wrongly treated as short keys.}
As shown in Figures~\ref{fig:var-key-pmem}(a) and~\ref{fig:var-key-pmem}(e), \roart performs the best in all cases.
This is exactly opposite to the case using fixed-length keys (cf. Figure~\ref{fig:pm-mt}).
In particular, cache misses caused by pointer chasing (to access keys) dominate the performance of \fptree, \lbtree and \dptree.
We further test the indexes with three representative real-world datasets~\cite{LastMile}: Reddit usernames (\texttt{names})~\cite{dataset-reddit-usernames}, Wikipedia (\texttt{wiki})~\cite{dataset-wiki} and URLs (\texttt{uk-2005})~\cite{dataset-uk-urls}.
They respectively consist of string keys of up to 25, 256 and 2029 bytes; for space limitation we omit more details about the datasets which can be found in the Appendix A.
As shown in Figures~\ref{fig:var-key-pmem}(b--d) and~\ref{fig:var-key-pmem}(f--h), the gaps between \roart and \fptree/\lbtree shrink, although \roart remains advantageous.
With longer keys, trie-based indexes will build more inner nodes to form deeper traversal paths, while the depth of B+-tree variants is not affected as they store pointers to keys.
Among all the indexes, \dptree performs much worse with longer keys.
We found a main reason is that
most lookups (94\% according to our profiling results) needed to traverse the base tree and search leaf nodes.
Since \dptree maintains multiple indexes, it also requires more complex traversal logic that incurs extensive key comparisons (on average 34 \texttt{memcmp} calls vs. 22 in \fptree), further lowering its performance.
Overall, these results highlight the need to enhance variable-length key support in future PM indexes, for example by combining the best of tries and B+-trees without tradeoffs.
\subsection{Impact of NUMA Effect}
\label{subsec:numa-exp}
Now we extend our experiments to use both NUMA nodes on the server.
Except for \pactree, we allocate PM and DRAM from the first socket.
All the threads are pinned, and we first use all the 40 physical cores across two sockets, before using hyperthreading beyond 40 threads.
This allows us to stress the indexes with inter-socket traffic and contrast their behavior with and without NUMA effect.
For \pactree, we include two variants which respectively enable and disable its NUMA-aware per-node PM pools.
As shown in Figure~\ref{fig:numa-pm}, NUMA effect has major impact on all indexes' throughput, and no index scales well beyond one socket for all operations.
Although not specifically designed for NUMA, \lbtree achieves the best scalability for lookups, and the highest throughput for inserts and updates (with a dropping trend beyond one socket).
We attribute the reason to its frugal use of cacheline flushes and careful node layout designs.
Both reduce PM accesses and cross-socket traffic.
For lookups, \fptree also does not collapse, whereas the performance of other indexes fluctuate and/or drop beyond 20 threads.
This implies HTM is robust to NUMA effect for read-only workloads, thanks to its lightweight conflict detection mechanism that piggybacks on the coherence protocol.
Moreover, HTM can use the extra last-level cache in the second NUMA node to track reads~\cite{ParkHTM2018,UnderstandingHTMCapacity}.
Other approaches (OLC, locking, ROWEX) are unable to take good advantage of the coherence protocol like HTM, contributing to more severe NUMA effect.
\pactree is the only PM index that takes NUMA effect into account by (1) using separate PM pools for each NUMA node and (2) leveraging snooping coherence protocol.
With separate PM pools (\texttt{\pactree-NUMA}), \pactree maintains performance beyond one socket without collapsing, but still does not scale as expected.
The main culprit is the directory-based coherence protocol that incurs additional PM accesses to update the directory.
Thus, \pactree advocates using snooping protocols for PM, which broadcasts coherence traffic across all cores, instead of using a directory to record cacheline status, hence does not incur additional PM accesses.
However, most platforms default to a directory-based protocol because snooping may not scale to high core counts.
Since our server
does not allow changing coherence protocols, we were unable to verify the performance of PM indexes using snooping protocols; we leave it as future work.
As we have noted in Section~\ref{subsec:trie}, requiring a certain coherence protocol may inflict issues with other applications and limit the applicability of the index.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/numa_effect_pmem_80t.pdf}
\caption{Impact of NUMA effect for PM indexes.
No index scales well under all operations due to additional PM accesses by the directory-based CPU coherence protocol.
}
\label{fig:numa-pm}
\end{figure}
\subsection{PM and DRAM Space Consumption}
New PM indexes are using DRAM more extensively to achieve high performance.
The downsides are (1) more complex recovery protocols and (2) higher DRAM space consumption (hence higher cost of ownership).
Table~\ref{tbl:mem-usage} lists the amount of DRAM and PM used by each index after loading 100 million records of 8-byte key and values (1.6GB).
\dptree and \utree practically store complete trees in DRAM, resulting in up to $\sim$18$\times$ higher DRAM consumption when compared with \fptree and \lbtree.
\pactree uses a similar amount of PM to \utree's and is among the most frugal in using DRAM (second to \fptree) because of its packed design.
Surprisingly, \roart uses 18.92GB to index 1.6GB of data, while others need 2.5--4.8GB.
The reason is it requires cacheline-aligned nodes and overprovisions leaf arrays:
for 100 million records, it allocates space for around one billion records (17 million leaf arrays) occupying 10.89GB of PM as each leaf array is 640-byte.
However, most space (reserved for leaf pointers) are unused.
Such overprovisioning is due to \roart's split mechanism.
In our experiment, keys are uniform randomly generated, so the 64 records in a full leaf array are usually very different despite they share a common prefix.
Then a split operation could allocate 63 leaf arrays in the worst case, leaving each new leaf array with only few records, leading to an average occupancy of $\sim$9\% and a waste of over 9GB of PM.
\begin{table}[t]
\smallskip\noindent
\setlength\tabcolsep{0.6mm}
\caption{PM and DRAM consumption (GB) after loading 100 million records with 8-byte keys and 8-byte values.}
\label{tbl:mem-usage}
\begin{tabular}{lccccccc}
\toprule
& \bf\fptree & \bf\lbtree & \bf\roart & \bf\dptree & \bf\pactree & \bf\utree \\\midrule
{\bf DRAM} & 0.14 & 0.34 & 0.14 & 1.14 & 0.18 & 2.63 &\\
{\bf PM} & 2.69 & 2.54 & 18.92 & 4.79 & 3.2 & 3.2 &\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Bandwidth Utilization and Requirements}
Previous evaluation~\cite{PiBench} has shown that PM bandwidth is scarce.
Since a database system also uses various other components, it is desirable to keep the bandwidth consumption low for indexes.
This has been the main focus of newer indexes.
We observe that the peak usage across all indexes and operations does not reach the limit ($\sim$10GB/s/$\sim$40GB/s for random write/sequential read).
This shows the effectiveness of the bandwidth saving techniques proposed by the surveyed indexes.
Due to space limitation, we omit the details on total bandwidth utilization and focus on the bandwidth used per operation, which is more indicative on how frugal (or not) an index uses PM bandwidth.
Figure~\ref{fig:pmbw} shows the results
with 20 threads when running lookups, inserts and scans under 8-byte keys and 8-byte values under the uniform distribution.
Overall, \lbtree exhibits the lowest number of bytes per operation, thanks to its node layout and logless design.
\roart exhibits the highest PM reads as it overprovisions node space.
In contrast, \pactree, which is also trie-based, has a similar bandwidth requirement to \lbtree's, because of its layout designed to reduce unnecessary PM accesses.
\section{Observations and Insights}
\label{sec:observe}
In this section, we summarize the major findings based on our experimental results and analysis of the surveyed PM indexes.
\textbf{1. The rule of thumb remains reducing PM accesses, which was first set in the pre-Optane era.}
Almost all the design choices (both pre-Optane and new ones) center around this goal, due to PM's lower bandwidth and higher latency.
If the properties of future PM hardware changes, the principles may be revisited.
\textbf{2. Some building blocks from the pre-Optane era continue to work well and are further optimized by new PM indexes.}
Most indexes use DRAM to accelerate traversal; some (e.g., \dptree) even place entire trees in DRAM.
However, tries typically cannot take the full advantage of DRAM to store reconstructible data.
Fingerprints are also widely used and enhanced by placing them in the spare bits of pointers and accessing them using SIMD instructions.
\textbf{3. Using extra metadata and selective persistence of metadata can further accelerate performance.}
The main reason is these approaches can avoid using WAL, which may incur additional PM writes and complicate code logic.
\textbf{4. Newly Proposed $\ne$ Better.}
Pre-Optane \fptree is still very competitive and sometimes can even outperform newer indexes.
Although \utree optimizes for tail latency, it exhibits the highest latency in many cases.
Such results call for careful benchmarking and comprehensive evaluations.
\textbf{5. All the new PM indexes are tailor-made for one product (Intel Optane PMem), which can be a double-edged sword.}
While this can deliver high performance, as exemplified by \lbtree which performs the best in most cases, it could pose challenges when the PM hardware landscape becomes more diverse.
\textbf{6. Support for NUMA-awareness, efficient PM management and variable-length keys remains inadequate.}
There have been initial attempts (e.g., using pointers for variable-length keys), but they are usually ad hoc or partial solutions with practical limitations (e.g., requiring a specific coherence protocol).
\textbf{7. There is no clear ``winner'' index architecture, but the choice may affect how (efficiently) functionality can be supported.}
For example, B+-tree (trie) variants perform well for fixed-length (variable-length) keys.
But
it remains to be explored whether it is easier to add efficient variable-length key support in \lbtree or to optimize \pactree to match \lbtree's performance.
\textbf{8. Linked lists with small nodes are a bad fit for PM indexes, and cache misses in general should be minimized or hidden.}
Accessing and scanning through a linked list of individual records incur many cache misses which can dominate the performance and lead to high latency (e.g., in \utree), canceling out the positive effects brought by other optimizations.
\textbf{9. HTM can perform well under NUMA for read-only workloads, but is challenging to handle contention and debug.}
In particular, the programmability issue
is further complicated with other system-level infrastructure: we observed extremely high abort rates under \texttt{glibc} version 2.33 which does not use the right instructions that can work with HTM in \texttt{memcpy}.\footnote{Details at \url{https://sourceware.org/bugzilla/show_bug.cgi?id=28033}.}
The bug was fixed in \texttt{glibc} 2.34~\cite{glibc234} which is used in our experiments.
It is noticeable that the best performing \lbtree is based on HTM;
it therefore remains to be seen in future work whether other approaches can overcome these issues while maintaining high performance.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/dram_indexes_uniform_skewed_8b_100m_10s.pdf}
\caption{Throughput of PM and DRAM indexes running on DRAM.
Without the extra fences, cacheline flushes and PM management code, PM-tailored indexes at least match the performance of \hot and \masstree.
\lbtree performs even better than \hot and \masstree on certain insert workloads (b), and \fptree tops scan performance under uniform distribution (d).
}
\label{fig:dram-tp}
\end{figure*}
\section{On the Next PM and DRAM Indexes}
\label{sec:next}
We give the outlook of the PM indexing space and describe promising future directions for PM and DRAM indexing.
\subsection{Future PM Indexes}
We identify three promising areas of future work for PM indexes.
\textbf{1. Efficient Support for Full Functionality.}
As we have discussed previously, variable-length keys and NUMA-awareness remain open problems for future PM indexes.
Importantly, it is desirable to maintain the high performance obtained by existing designs while better supporting full functionality.
\textbf{2. Wider Applicability/Less Tailor-Made.}
There are various ways to realize PM, by using new materials (e.g., memristor~\cite{Memristor}, STT-RAM~\cite{STT-RAM} and Intel 3D XPoint which PMem is based upon) or NVDIMMs which combine flash and DRAM~\cite{AgigaNVDIMM,VikingNVDIMM}.
However, most (if not all) indexes aiming for real PM are tailor-made for Intel Optane PMem;
yet certain properties like 256-byte alignment may even change across generations of the same product, and designs based on them may not work well on NVDIMMs, diminishing their applicability.
Although some of the hardware efforts are in their early stage, we argue it is important to consider applicability of future designs on different PM devices.
\textbf{3. Real-World Adoption and Cost-Effectiveness.}
Although there have been numerous PM index proposals, we are yet to see major adoption in real systems and applications.
Part of the reason is the low cost effectiveness of PM-based servers as identified by other work~\cite{SSDStrikingBack}, especially when compared to modern SSDs which can deliver high bandwidth and microsecond-level latency.
Therefore, on the hardware side, we hope future work to lower the per GB cost of PM servers.
On the software side, PM indexes and data structures in general should focus more on cost/performance.
\subsection{Unifying PM and DRAM Indexing}
In a similar vein to the point on wider applicability, we observe techniques proposed for PM indexes can also be effective for DRAM.
We conduct preliminary experiments to compare the surveyed state-of-the-art PM indexes (with the extra cacheline flushes and fences removed) and two representative DRAM-optimized volatile indexes (\hot~\cite{HOT} and \masstree~\cite{Masstree}).
Figure~\ref{fig:dram-tp} shows the throughput obtained by running the same workload as Section~\ref{subsec:multi-thread-exp}, but purely on DRAM.
Under both uniform and skewed distributions, PM indexes perform competitively with DRAM-optimized indexes.
In certain cases, PM indexes perform even better than \hot and \masstree which are specifically optimized for DRAM, e.g.,
\lbtree for inserts in Figure~\ref{fig:dram-tp}(b) and \fptree (despite being a pre-Optane proposal) in Figure~\ref{fig:dram-tp}(d) for scans.
Although it remains to be seen how the optimizations for PM and DRAM indexes compare, and how PM techniques may be used by DRAM indexes (and vice versa), our results indicate it is promising to devise future indexes that would work on both volatile and non-volatile memory.
This could greatly simplify implementation and widen the applicability of techniques proposed by both types of indexes.
\section{Related Work}
\label{sec:related}
Our work is most related to performance studies for PM devices and data structures, PM indexes and PM management issues.
\textbf{Performance Studies.}
Early work~\cite{UCSDMeasurement,UCSDGuide} characterized the performance of Optane PMem, exposing a set of properties different from what were previously assumed by emulations.
Gugnani et. al~\cite{PMIdio} exposed more properties of Optane PMem under various scenarios, e.g., eADR and NUMA, along with case studies.
Beyond range indexes, Hu et. al~\cite{HuPMHI} evaluated PM-optimized hash indexes on Optane PMem.
Koutsoukos et. al~\cite{HtuPMiD} analyzed the performance of PM-enhanced database engines under TPC-C and TPC-H, and came up with guidelines of tuning the system for best performance.
\textbf{PM Indexes.}
In addition to adapting specific indexes, general-purpose approaches such as
RECIPE~\cite{RECIPE}, NAP~\cite{NAP} and TIPS~\cite{TIPS} present principled methods for converting DRAM indexes into PM indexes.
It is interesting future work to evaluate these approaches.
Some early efforts have adapted learned indexes~\cite{LearnedIndex} for PM.
Chen et. al~\cite{PMLearnedEval} observe that the bigger nodes used by learned indexes can cause excessive PM accesses.
APEX~\cite{APEX} transforms the DRAM-based updatable ALEX~\cite{ALEX} with concurrency and instant recovery on PM.
Hash tables are also being re-designed for PM.
CCEH~\cite{CCEH} is a failure-atomic variant of extendible hashing that reduces directory management overhead.
Dash~\cite{Dash} integrates a set of useful techniques to adapt extendible and linear hashing for PM;
the key insight is that both PM reads and writes should be minimized.
Clevel~\cite{CLEVEL} is a lock-free version of level hashing~\cite{LevelHashing} that performs asynchronous resizing in the background.
\textbf{PM Libraries.}
PMDK~\cite{PMDK} is the de facto standard, but may not be the optimal solution:
\roart, \dptree and \pactree all devise their own approaches.
Designing better PM libraries remains an open area, as seen by many recent alternatives~\cite{Pangolin,Ralloc,Metall,Makalu,nvmMalloc}.
\section{Conclusion}
\label{sec:conclusion}
We conducted a comprehensive evaluation of representative PM range indexes proposed based on real Intel Optane PMem.
These new indexes inherited many useful designs from pre-Optane PM indexes and proposed new building blocks
that can be useful for building future PM indexes.
Based on our evaluation, we gave a list of observations, insights and future directions.
We found the new indexes do not necessarily outperform the pre-Optane proposals, and efficient designs for variable-length keys, PM management and NUMA awareness are still lacking.
We also discovered for the first time that a PM range index can match or even outperform DRAM-optimized indexes, highlighting the potential of unifying PM and DRAM indexing to save design and implementation efforts.
\section*{Appendix}
\section{Datasets Used and Clarification}
Here we present the details of all three datasets used in the variable-length key experiments.
The \texttt{names} dataset contains 25 million unique Reddit usernames of length up to 25
characters.
\texttt{wiki} contains 55 million unique article titles of length up to 256 characters.
\texttt{uk-2005} contains 40 million unique URLs of up to 2029 characters.
The key distributions and statistics of the three datasets are shown in Figure~\ref{fig:dist}, covering a wide range of different key lengths in reality.
As a trie-based index, ROART already supports variable length keys.
So no modification is required.
For \lbtree, \fptree and \dptree, we modified their lookup and insert operations to support variable-length keys.
The parameters of the related operations are unchanged but we allocate the variable-length keys from volatile/persistent heap and store the key pointers in inner/leaf nodes.
Key comparison is performed using \texttt{memcmp} over the two key pointers.
We also allocate each key by the maximum length regardless of its actual length, to ease implementation.
This approach does results in wasted memory especially in the case of skewed datasets.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/long_run.pdf}
\caption{\lbtree performance over time.}
\label{fig:time-perf}
\end{figure}
\section{Long-Running Workloads}
Although PM's bandwidth exhausts quickly with increasing number of writers~\cite{PiBench}, its performance tends to be stable across the short-lived benchmarks.
To simulate the impact of long-running systems and high space utilization on PM performance, we conducted two stress tests using \lbtree.
A pool of 485GB is first allocated from PM, which is the maximum available PM space that can be allocated on the first node of our machine.
We then populated the entire reserved space with 8-byte keys/values generated from a uniform distribution, and ran a read-only workload for 20 minutes with 40 threads.
The lookup throughput averaged by minutes is shown in Figure~\ref{fig:time-perf}(a).
As depicted, the number of per-second read operations completed is very stable across the entire experiment phase.
We further conducted another insert experiment, where 20 billion uniformly distributed 8-byte key/values are inserted with 40 threads.
The experiment kept running until the PM space is exhausted (it took around 20 minutes).
As shown by Figure~\ref{fig:time-perf}(b), the insert throughput was also stable with minor gradual degradation ($\sim$12\%) as we insert more keys.
This is expected because the index gets larger and deeper with more keys inserted, thus resulting in slower traversal and SMOs.
Overall, PMem demonstrates robustness and stable performance throughout the long-running read/write intensive benchmarks while under high space utilization.
Different from traditional SSDs that are known to perform worse as approaching full space utilization, PMem has shown no obvious hardware/software performance degradation (except due to the characteristics of the workload itself, e.g., inserts) in our experiments.
However, we acknowledge that these tests are preliminary and further investigation with even longer experiments would be needed to thoroughly explore this topic in future work.
\section{NUMA Experiments on DRAM}
This set of experiments repeats our earlier NUMA experiments in the main text but were done using DRAM to explore the impact of NUMA effect.
As shown by Figure~\ref{fig:numa-pm}, most PM indexes are not scalable and their throughput can drop drastically beyond one socket.
However, it is not the case on DRAM.
As Figure~\ref{fig:dram-numa} shows, when running on DRAM, \fptree, \lbtree and \roart are still scalable across two NUMA nodes.
Thus, the main PM-specific concern lies in the impact of PM's higher latency/lower bandwidth on the interconnect traffic and cache coherence protocol~\cite{pactree}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{paper-plots/plots/numa_effect_dram_80t.pdf}
\caption{Impact of NUMA effect for indexes on DRAM.}
\label{fig:dram-numa}
\end{figure}
\end{appendix}
|
\section{Introduction}
Let $M=(M,g)$ be an $m$-dimensional Riemannian manifold and $\nabla$ be the Levi-Civita connection of $g$. The curvature tensor $R$ on $M$ is defined by $R(X, Y)Z=[\nabla_X, \nabla_Y]Z-\nabla_{[X,Y]}Z$ for $X$, $Y$, $Z\in \mathfrak{X}(M)$, where $\mathfrak{X}(M)$ is the Lie algebra of all smooth vector fields on $M$. The Ricci tensor of $M$ is defined by $\rho(X,Y)=\Tr(Z\rightarrow R(Z,X)Y)$ and the scalar curvature of $M$ is obtained by $\tau = \Tr \rho$. Throughout the paper, we assume that the components of the tensor fields are with respect to a local orthonormal frame $\{e_i\}$ and we also adopt the Einstein convention on sum over repeated indices unless otherwise specified.
An $m$-dimensional Riemannian manifold $(M,g)$ is said to be {\it Einstein} if $\rho = \frac{\tau}{m}g$. The $m$-dimensional Einstein manifold $(M,g)$ is said to be {\it super-Einstein} if the following condition is satisfied on $M$
\begin{equation}\label{eq:supEin}
{\tR}(X,Y)= \sum_{a, b, c =1}^{m} R(X,e_{a},e_{b},e_{c})R(Y,e_{a},e_{b},e_{c})=\dfrac{||R||^2}{m} g(X,Y)
\end{equation} with constant $||R||^2$ (see \cite{BV,EPS13,GW}). The condition \eqref{eq:supEin} has some geometric meanings: For a compact manifold, an Einstein metric is critical for the functional $\int_{M}||R||^2 dv_{g}$ restricted to $vol(M)=1$ if and only if $\tR(X,Y) = \frac{||R||^2}{m}g(X,Y)$ (see \cite[Corollary~4.72]{Bes1}).
{{Boeckx and Vanhecke \cite{BV} showed that an Einstein manifold $M$ is super-Einstein if and only if the unit tangent sphere bundle $T_1 M$ equipped with the standard contact metric structure has constant scalar curvature (\cite[Proposition~3.6]{BV}).}}
Let $M_{q}(r,f)$ denote the mean-value of a real-valued function $f$ over a geodesic sphere $S(q;r)$ with center $q$ and radius $r$ in an $m$-dimensional Riemannian manifold $M$:
\begin{equation*}\label{avr}
M_{q}(r,f) := \frac{1}{\int_{S(q;r)} dv_{S(q;r)}}\, \int_{p\in S(q;r)} f(p) \ d v_{S(q;r)}.
\end{equation*}
Gray and Willmore \cite{GW} showed the mean-value properties for an Einstein and super-Einstein manifold: They proved, from the expansions of $M_{q}(r,f)$, that the harmonic function $f$ near $q$ has the mean-value properties
\begin{equation*}
M_{q}(r,f) = f(q) + O(r^6),\quad {\text{as}}\quad r\rightarrow 0
\end{equation*}
for an Einstein manifold (\cite[Theorem~1.1]{GW}) and
\begin{equation*}
M_{q}(r,f) = f(q) + O(r^8),\quad {\text{as}}\quad r\rightarrow 0
\end{equation*}
for a super-Einstein manifold (\cite[Theorem~6.1]{GW}).
A Riemannian manifold {{$M$}} is said to be {\it $2$-stein} if there exist two functions $\mu_1, \mu_2$ on $M$ such that $\Tr R_X = \mu_1 ||X||^2$ and $ \Tr (R_X^2) = \mu_2 ||X||^4$, for all $p \in M$ and all $X \in T_pM$. Here, the Jacobi operator $R_X$ is defined by $R_X Y=R(Y,X)X$ for a tangent vectors $X$, $Y$ at a point $p\in M$. A unit vector field $V$ on $M$ is said to be a {\it harmonic vector field} if it is a critical point for the energy functional in
the set of all unit vector fields of $M$ \cite{Wood}. A contact
metric manifold whose characteristic vector field $\xi$ is a
harmonic vector field is called an {\it H-contact manifold}. Nikolayevsky and Park \cite{NP} showed that for a Riemannian manifold $M$, $T_1 M$ equipped with the standard contact metric structure is $H$-contact if and only if $M$ is 2-stein. Gilkey, Swann, and Vanhecke \cite{GSV} showed that a 4-dimensional manifold $M$ is 2-stein if and only if locally there is a choice of orientation of $M$ for which the metric is self-dual and Einstein (\cite[Theorem~2.6]{GSV}). From the definition of a 2-stein manifold, we can derive the super-Einstein conditions (see \cite[Chap.~6, \S E]{Bes2}). Thus, a 2-stein manifold is necessarily super-Einstein.
Euh, Park, and Sekigawa \cite{EPS13} dervied a curvature identity on any 4-dimensional manifold from the Chern-Gauss-Bonnet theorem. There are many applications of this identity (see \cite{EPS11,EPS013,GHMV}). On the other hand, Deszcz, Hotlo\'s, and Sent\"urk \cite{DHS} gave some curvature properties of 4-dimensional semi-Riemannian manifolds as an application of Patterson's curvature identity.
In this paper, we introduce the curvature identities on some $5$- and $6$-dimensional Riemannian manifolds such as Einstein, super-Einstein manifolds. Our main results are the following.
\smallskip
{{\begin{theorema}
Let $M=(M,g)$ be a 5-dimensional Riemannian manifold.
\begin{enumerate}[label={{\rm{(\alph*)}}}]
\item\label{eq:5-einsteinid} If $M$ is Einstein, then the following curvature identity holds on $M${\rm :}
\begin{equation*}
2\tau\tR_{ij}+4\check{R}_{ij}+4\hat{R}_{ij}-8\mathring{R}_{ij} = \Big(\frac{\tau}{5}||R||^2 + \frac{\tau^3}{25}\Big)g_{ij}.
\end{equation*}
\item\label{eq:5-supEinid} If $M$ is super-Einstein, then the following curvature identity holds on $M${\rm :}
{{\begin{equation*}
4\mathring{R}_{ij} - 2\hat{R}_{ij} =\Big(\frac{9}{50}\tau||R||^2 -\frac{\tau^3}{50}\Big)g_{ij}.
\end{equation*}}}
\end{enumerate}
\end{theorema}
\smallskip
\begin{theoremb}
Let $M=(M,g)$ be a 6-dimensional Riemannian manifold.
\begin{enumerate}[label={{\rm{(\alph*)}}}]
\item\label{eq:6-einstein} If $M$ is Einstein, then the following curvature identity holds on $M${\rm :}
{{\begin{equation*}
4\tau {{\tR_{ij}}}+12\check{R}_{ij}+12\hat{R}_{ij}-24\mathring{R}_{ij}=(\tau ||R||^2 -4\mathring{R}+2\hat{R})g_{ij}.
\end{equation*}}}
\item\label{eq:6-dim2stein} If $M$ is super-Einstein, then the following curvature identity holds on $M${\rm :}
\begin{equation*}
2\mathring{R}_{ij}-\hat{R}_{ij}=\frac{1}{6}(2\mathring{R}-\hat{R})g_{ij}.
\end{equation*}
\end{enumerate}
Here, we set
\begin{gather*}
\check{R}_{ij}=R_{iuvj}R_{abcu}R_{abcv},\quad \hat{R}_{ij}=R_{ibcd}R_{jbuv}R_{cduv},\quad \mathring{R}_{ij}=R_{ibcd}R_{jucv}R_{budv},\\
\hat{R}= R_{abcd}R_{abuv}R_{cduv},\quad \mathring{R}=R_{abcd}R_{aucv}R_{budv}.
\end{gather*}
\end{theoremb}}}
We derive the curvature identities from Patterson's curvature identity. In Section 2, we recall the previous results for a 6-dimensional Riemannian manifold \cite{EPS17} and introduce Patterson's curvature identity. In Section 3 and Section 4, we prove Theorem A and Theorem B, respectively. We also give some examples supporting the theorems.
In the appendix, we attach the detailed computation for the proof of Lemma \ref{thm:6-einstien(0,6)-tensor}.
\section{{{Curvature identities}}}
On a 6-dimensional compact oriented Riemannian manifold, Chern-Gauss-Bonnet theorem states that Euler characteristic $\chi(M)$ of $M$ is given by the following integral formula.
\begin{prop}[\cite{Sa}]
Let $M=(M,g)$ be a $6$-dimensional compact oriented Riemannian manifold. Then
the Euler characteristic $\chi(M)$ of $M$ is given by
\begin{equation*}\label{eq:id6}
\begin{aligned}
\chi(M)
=\frac{1}{384\pi^3}\int_M\{&\tau^3-12\tau||\rho||^2+3\tau||R||^2+16\rho_{ab}\rho_{ac}\rho_{bc}\\
&-24\rho_{ab}\rho_{cd}R_{acbd}
-24\rho_{uv}R_{abcu}R_{abcv}\\
&+8R_{abcd}R_{aucv}R_{bvdu}-2R_{abcd}R_{abuv}R_{cduv}\}dV_g.
\end{aligned}
\end{equation*}
\end{prop}
In \cite{EPS17}, the authors consider the one-parameter deformation $g(t)$ of $g$, using the fact that
Euler characteristic is a topological invariant for the deformation, they got the universal
curvature identity which holds on any 6-dimensional Riemannian manifold. In particular, they have the following.
\begin{thm}[\cite{EPS17}]\label{th:6-eins}
Let $M=(M,g)$ be a 6-dimensional Einstein manifold. Then the following identity holds on $M${\rm :}
\begin{equation*}
(-\tau ||R||^2+4\mathring{R}-2\hat{R})g_{ij}+12\check{R}_{ij}+12\hat{R}_{ij}-24\mathring{R}_{ij}+4\tau {{\tR_{ij}}}=0.
\end{equation*}
\end{thm}
\smallskip
{{For an integer $N \ge 1$, the {\textit{generalized Kronecker delta}} is given by
\begin{equation*}
\delta^{j_1j_2\cdots j_N}_{i_1i_2\cdots i_N}=
\begin{vmatrix}
\delta^{j_1}_{i_1}&\delta^{j_2}_{i_1}&\cdots&\delta^{j_N}_{i_1}\\
\delta^{j_1}_{i_2}&\delta^{j_2}_{i_2}&\cdots &\delta^{j_N}_{i_2}\\
\vdots&\vdots&\vdots&\vdots\\
\delta^{j_1}_{i_N}&\delta^{j_2}_{i_N}&\cdots&\delta^{j_N}_{i_N}
\end{vmatrix}.
\end{equation*}
By using the skew-symmetric properties of the generalized Kronecker delta, Patterson [14] obtained the curvature identity for an $m$-dimensional Riemannian manifold as follows:}}
\begin{equation}\label{eq:Pa_n}
\delta^{jj_1j_2\cdots j_m}_{ii_1i_2\cdots i_m}{R^{i_1i_2}}_{j_1j_2} \cdots {R^{i_{2r-1}i_{2r}}}_{j_{2r-1}j_{2r}}=0,
\end{equation}
where $r$ is any integer such that $1\leq r\leq \frac{m}{2}$ ($m$ even) or $1\leq r\leq \frac{m-1}{2}$ ($m$ odd).
We note that the identity~\eqref{eq:Pa_n} holds with respect to the Weyl curvature tensor $W$ even if we replace $R$ by $W$ in \eqref{eq:Pa_n} (\cite[Section~8]{Pa}).
\section{Proof of Theorem A}
Let $(M,g)$ be a 5-dimensional Riemannian manifold. Making use of \eqref{eq:Pa_n}, we have the curvature identity with respect to the Weyl curvature tensor $W$ for the case when $r=2$ as follows:
\begin{equation}\label{eq:5-Weyl}
\begin{aligned}
&||W||^2(g_{ik}g_{jl}-g_{il}g_{jk})-4\Big\{W_{abcj}W_{abcl}g_{ik}-W_{abcj}W_{abck}g_{il}+W_{abci}W_{abck}g_{jl}\\
&-W_{abci}W_{abcl}g_{jk}-2W_{iabl}W_{kabj}+2W_{iabk}W_{labj}-W_{abij}W_{abkl}\Big\}=0.
\end{aligned}
\end{equation}
Since the Weyl tensor $W$ on a 5-dimensional Riemannian manifold is given by
\begin{equation*}
W_{pqrs}=R_{pqrs}-\frac{1}{3}(\rho_{ps}g_{qr}+\rho_{qr}g_{ps}-\rho_{pr}g_{qs}-\rho_{qs}g_{pr})+\frac{\tau}{12}(g_{ps}g_{qr}-g_{pr}g_{qs}),
\end{equation*}
we obtain the explicit formula of the curvature identity on $5$-dimensional Einstein manifolds.
\begin{lem}
Let $M=(M,g)$ be a 5-dimensional Einstein manifold. The following curvature identity holds on $M${\rm :}
\begin{equation}\label{eq:5-einstein}
\begin{aligned}
&\Big(||R||^2 + \frac{\tau^2}{5} \Big)(g_{ik}g_{jl} - g_{il}g_{jk})-4(\tR_{ik}g_{jl} + \tR_{jl}g_{ik} -\tR_{il}g_{jk} -\tR_{jk}g_{il})\\
&~+8(R_{iabl}R_{kabj}-R_{iabk}R_{labj}) +4R_{abij}R_{abkl} + \frac{12}{5}\tau R_{ijkl} = 0.
\end{aligned}
\end{equation}
\begin{proof}
Since $\rho_{ps}=\frac{\tau}{5}g_{ps}$, we obtain $W_{pqrs}=R_{pqrs}-\frac{\tau}{20}(g_{ps}g_{qr}-g_{pr}g_{ps})$. Now we apply this $W$ to the identity~\eqref{eq:5-Weyl}. Then we have
\begin{gather*}
||W||^2=||R||^2-\frac{\tau}{10},\quad W_{abcj}W_{abcl}=R_{abcj}R_{abcl}-\frac{\tau^2}{50}g_{jl},\\
W_{iabj}W_{kabj}=R_{iabl}R_{kabj}-\frac{\tau}{10}R_{ijlk}-\frac{\tau^2}{80}g_{il}g_{jk}+\frac{\tau^2}{400}g_{ik}g_{jl},\\
W_{abij}W_{abkl}=R_{abij}R_{abkl}+\frac{\tau}{5}R_{ijkl}+\frac{\tau^2}{200}\big(g_{ik}g_{jl}-g_{il}g_{jk}\big).
\end{gather*}
Similarly, we can obtain remaining terms. Then, by rearranging all terms, we complete the proof.
\end{proof}
\end{lem}
{{We note that there are useful formulae as follows:
\begin{equation}\label{eq:re}
R_{ibcd}R_{jbuv}R_{cudv}=\frac{1}{2}\hat{R}_{ij},\quad R_{ibcd}R_{jubv}R_{cudv}=\frac{1}{4}\hat{R}_{ij},\quad R_{ibcd}R_{jucv}R_{bvdu}=\mathring{R}_{ij}-\frac{1}{4}\hat{R}_{ij}.
\end{equation}}}
Now, we transvect each term of ~\eqref{eq:5-einstein} with $R_{pjkl}$ and use $\rho_{ip}=\frac{\tau}{5}g_{ip}$. Then, we obtain
\begin{equation*}
\begin{aligned}
&\Big(||R||^2 + \frac{\tau^2}{5} \Big)(g_{ik}g_{jl} - g_{il}g_{jk})R_{pjkl} = -\frac{2}{5}\tau\Big(||R||^2 + \frac{\tau^2}{5}\Big)g_{ip},\\
-&4(\tR_{ik}g_{jl} + \tR_{jl}g_{ik} -\tR_{il}g_{jk} -\tR_{jk}g_{il})R_{pjkl} = -4\Big(-\frac{2}{5}\tau\tR_{ip} - 2\check{R}_{ip}\Big),\\
&~8(R_{iabl}R_{kabj}-R_{iabk}R_{labj})R_{pjkl} = -16\mathring{R}_{ip} + 4\hat{R}_{ip},\\
&~4R_{abij}R_{abkl}R_{pjkl} = 4\hat{R}_{ip},\\
&~\frac{12}{5}\tau R_{ijkl}R_{pjkl} = \frac{12}{5}\tau\tR_{ip}.
\end{aligned}
\end{equation*}
For the third equation above, we use~\eqref{eq:re}. Hence, we complete the proof of Theorem~A-\ref{eq:5-einsteinid}.
\smallskip
{{Now we prove Theorem~A-\ref{eq:5-supEinid}. Applying the condition~\eqref{eq:supEin} to~\eqref{eq:5-einstein}, we obtain}}
\begin{equation}\label{eq:Pa5}
R_{ijab}R_{abkl}+2R_{iabl}R_{kabj}-2R_{iabk}R_{labj}+\frac{3}{5}\tau R_{ijkl}=\Big(\frac{3}{20}||R||^2-\frac{1}{20}\tau^2\Big) (g_{ik}g_{jl}-g_{il}g_{jk}).
\end{equation}
Transvecting each term of \eqref{eq:Pa5} with $R_{pjkl}$ and using~\eqref{eq:supEin}, we obtain
\begin{equation*}\label{eq:5-supEin}
\begin{aligned}
&R_{ijab}R_{abkl}R_{pjkl}=\hat{R}_{ip},\\
&2R_{iabl}R_{kabj}R_{pjkl}= -2\Big(\mathring{R}_{ip} - \frac{1}{4}\hat{R}_{ip}\Big),\\
-&2R_{iabk}R_{labj}R_{pjkl}=-2\Big(\mathring{R}_{ip} - \frac{1}{4}\hat{R}_{ip}\Big),\\
&\frac{3}{5}\tau R_{ijkl}R_{pjkl}=\frac{3}{25}\tau||R||^2g_{ip},\\
&(g_{ik}g_{jl}-g_{il}g_{jk}) R_{pjkl}=-\frac{2}{5}\tau g_{ip}.
\end{aligned}
\end{equation*}
For the second and third equations, we use \eqref{eq:re}.
Therefore, we complete the proof of Theorem~A-\ref{eq:5-supEinid}.
\smallskip
For a 5-dimensional 2-stein manifold, there is the following orthonormal basis introduced by Nikolayevsky \cite{Nik}.
\begin{prop}\big(\cite[Proposition~4]{Nik}\big)
Let $M=(M,g)$ be a 2-stein manifold of dimension 5. Then there exists an orthonomal basis $\{e_i\}$ at each point $p\in M$ such that
\begin{equation*}\label{eq:basis5}
\begin{aligned}
R_{1212}&=R_{1313}=R_{2323}=R_{2424}=R_{3434}=\alpha-\beta,~
R_{1414}=\alpha-4\beta,\\ R_{1515}&=R_{4545}=\alpha,~ R_{2525}=R_{3535}=\alpha-3\beta,\\
R_{1234}&=\beta,~R_{1235}=\sqrt{3}\beta,~R_{1324}=-\beta,~R_{1325}=\sqrt{3}\beta,\\
R_{1423}&=-2\beta, ~R_{2425}=\sqrt{3}\beta,~R_{3435}=-\sqrt{3}\beta,
\end{aligned}
\end{equation*}
and all the other components of $R$ vanish.
\end{prop}
By using {{Nikolayevsky's}} basis, we show the following theorem.
\begin{thm}\label{thm:3.4}
Let $M=(M,g)$ be a 5-dimensional 2-stein manifold. The identity \eqref{eq:Pa5} holds on $M$.
\begin{proof}
Each term of the left hand side of \eqref{eq:Pa5} in the case of $i=1$, $j=2$, $k=3$, $l=4$ is as follows:
\begin{equation*}
\begin{aligned}
R_{12ab}R_{ab34}&=2(2 \alpha - 5 \beta)\beta,\\
2R_{1ab4}R_{3ab2}& = 2(2\alpha-5\beta)\beta,\\
-2R_{1ab3}R_{4ab2}& = 2(2\alpha+\beta)\beta,\\
\frac{3}{5}\tau R_{1234}& = -6(2\alpha-3\beta)\beta.\\
\end{aligned}
\end{equation*}
Since $R_{12ab}R_{ab34}+2R_{1ab4}R_{3ab2}-2R_{1ab3}R_{4ab2}+\frac{3}{5}\tau R_{1234}=0$ and
$(g_{13}g_{24}-g_{14}g_{23})=0$ in \eqref{eq:Pa5}, the equation \eqref{eq:Pa5} holds for $i=1$, $j=2$, $k=3$, $l=4$. For other choice of $i, j, k, l$,
similar processes can be done for the proof of Theorem \ref{thm:3.4}.
\end{proof}
\end{thm}
Now we give the examples for Theorem A.
\begin{ex}
Let $M$ be a Riemannian product manifold of a 3-dimensional Riemannian manifold $M_{1}$ of constant sectional curvature $k$ and a surface $M_{2}$ of constant sectional curvature $2k$ $(k\neq 0)$. Then $M$ is not super-Einstein but Einstein manifold. Let $\{e_{i}\}$, $i=1,\ldots,5$ be an orthonormal basis of $M$, where $\{e_{1},e_{2},e_{3}\}$ and $\{e_{4},e_{5}\}$ are bases for $M_1 $ and $M_2 $, respectively. Then we have
\begin{equation}\label{eq:ex_5EinnotsupEin}
R_{1221}=R_{1331}=R_{2332}=k,\quad R_{4554}=2k,
\end{equation}
and all the other components of $R$ vanish. From \eqref{eq:ex_5EinnotsupEin} we have $\tau = 10k$, $||R||^2 = 28k^2$, and
\begin{gather*}
\tR_{ij}=
\begin{cases}
4k^2 & {\text{if}}\;\, i=j=1,2,3\\
8k^2 & {\text{if}}\;\, i=j=4,5\\
0 & {\text{otherwise}}
\end{cases},\qquad
\check{R}_{ij}=
\begin{cases}
8k^3 & {\text{if}}\;\, i=j=1,2,3\\
16k^3 & {\text{if}}\;\, i=j=4,5\\
0 & {\text{otherwise}}
\end{cases},\\
\hat{R}_{ij}=
\begin{cases}
-8k^3 & {\text{if}}\;\, i=j=1,2,3\\
-32k^3 & {\text{if}}\;\, i=j=4,5\\
0 & {\text{otherwise}}
\end{cases},\qquad
\mathring{R}_{ij}=
\begin{cases}
-2k^3 & {\text{if}}\;\, i=j=1,2,3\\
0 & {\text{otherwise}}
\end{cases}.
\end{gather*}
Therefore, we find that the curvature identity of Theorem ~A-\ref{eq:5-supEinid} does not hold, but that of Theorem~A-\ref{eq:5-einsteinid} holds on $M$.
\end{ex}
\begin{ex
{{Let}} $M=SL(3)/SO(3)$. Then $M$ is a 5-dimensional 2-stein manifold.
The inner product and the curvature tensor are given by
\begin{equation*}
<X,Y>=\Tr{XY},\quad R(X,Y)Z= - [[X,Y],Z]
\end{equation*}
for $X$, $Y$, $Z\in \mathfrak{X}(M)$. Then, with orthonormal basis of $SL(3)/SO(3)$
\begin{equation*}
\begin{gathered}
X_1=\frac{1}{\sqrt{2}}\begin{pmatrix}
0 & 1 & 0 \\
1& 0 & 0 \\
0 & 0 & 0
\end{pmatrix}, \quad
X_2=\frac{1}{\sqrt{2}}\begin{pmatrix}
0&0&1\\
0&0&0\\
1&0&0
\end{pmatrix},\quad
X_3=\frac{1}{\sqrt{2}}\begin{pmatrix}
0&0&0\\
0&0&1\\
0&1&0
\end{pmatrix}\\
X_4=\frac{1}{\sqrt{2}}\begin{pmatrix}
-1&0&0\\
0&1&0\\
0&0&0
\end{pmatrix},\quad
X_5=\frac{1}{\sqrt{6}}\begin{pmatrix}
1&0&0\\
0&1&0\\
0&0&-2
\end{pmatrix},
\end{gathered}
\end{equation*}
we have the components of the curvature tensor:
\begin{equation}\label{eq:ex_5-2stein}
\begin{aligned}
&~R_{1221}=R_{1331}=R_{2332}=R_{2442}=R_{3443}= -\frac{1}{2},~
R_{1441}= - 2,\\
&~R_{1551}=R_{4554}=0,~R_{2552}=R_{3553}=-\frac{3}{2},\\
&~R_{1234}=-\frac{1}{2},~
R_{1235}=-\frac{\sqrt{3}}{2},~R_{1324}=\frac{1}{2},~R_{1325}=-\frac{\sqrt{3}}{2},\\
&~R_{1423}=1,~R_{2425}=-\frac{\sqrt{3}}{2},~R_{3435}=\frac{\sqrt{3}}{2},
\end{aligned}
\end{equation}
and all the other components vanish.
{{From \eqref{eq:ex_5-2stein}, we compute the Ricci tensor $\rho_{ij}$. For $i\neq j$, $\rho_{ij}=0$ and
$\rho_{11}=\sum\limits_{a=1}^{5}R_{1aa1}=-3$. Similarly, $\rho_{22}= \rho_{33}= \rho_{44}= \rho_{55}= - 3$}} and so we get the scalar curvature $\tau$ is $-15$. Making use of \eqref{eq:ex_5-2stein}, ${{||R||^2=75}}$, $\hat{R}_{ij}=75\delta_{ij}$, and $\mathring{R}_{ij}=\frac{15}{4}\delta_{ij}$, where $\delta_{ij}$ denotes the Kronecker delta. Therefore, we find that the curvature identity of Theorem~A-\ref{eq:5-supEinid} holds on $M$.
\end{ex}
\section{Proof of Theorem B}
We obtain the explicit formula of the curvature identity on 6-dimensional Einstein manifolds for the case when $m=6$ and $r=2$ of the identity~\eqref{eq:Pa_n}. To prove the Theorem B, we need Lemma \ref{thm:6-einstien(0,6)-tensor}, which is proved in the Appendix.
\begin{lem}\label{thm:6-einstien(0,6)-tensor}
Let $M=(M,g)$ be a $6$-dimensional Einstein manifold. The following curvature identity holds on $M${\rm :}
\begin{equation}\label{eq:6-einstein(0,6)-tensor}
\begin{aligned}
&\Big\{\frac{1}{8}\Big(||R||^2 +\frac{\tau^2}{3}\Big)\Big(g_{ij}g_{hk}g_{lm}-g_{ij}g_{hm}g_{lk}-g_{ik}g_{hj}g_{lm}+g_{ik}g_{hm}g_{lj}+g_{im}g_{hj}g_{lk}-g_{im}g_{hk}g_{lj}\Big)\Big\}\\
&+ \Big\{-\frac{1}{2}\Big(\tR_{ij}(g_{hk}g_{lm}-g_{hm}g_{lk})-\tR_{ik}(g_{hj}g_{lm}-g_{hm}g_{lj})+\tR_{im}(g_{hj}g_{lk}-g_{hk}g_{lj})\\
&\qquad-\tR_{hj}(g_{ik}g_{lm}-g_{im}g_{lk})+\tR_{hk}(g_{ij}g_{lm}-g_{im}g_{lj})-\tR_{hm}(g_{ij}g_{lk}-g_{ik}g_{lj})\\
&\qquad+\tR_{lj}(g_{ik}g_{hm}-g_{im}g_{hk})-\tR_{lk}(g_{ij}g_{hm}-g_{im}g_{hj})+\tR_{lm}(g_{ij}g_{hk}-g_{ik}g_{hj})\Big)\Big\}\\
&+\Big\{\Big(-T_{ijkh} + T_{ikjh} +\frac{1}{2}S_{ihjk}+\frac{\tau}{3} R_{ihjk}\Big)g_{lm}-\Big(-T_{ijkl} +T_{ikjl}+\frac{1}{2}S_{iljk}+\frac{\tau}{3} R_{iljk}\Big)g_{hm}\\
&-\Big(-T_{ijmh}+T_{imjh}+\frac{1}{2}S_{ihjm}+\frac{\tau}{3} R_{ihjm}\Big)g_{lk}+\Big(-T_{ijml} +T_{imjl}+\frac{1}{2}S_{iljm}+\frac{\tau}{3} R_{iljm}\Big)g_{hk}\\
&+\Big(-T_{ikmh} +T_{imkh}+\frac{1}{2}S_{ihkm}+\frac{\tau}{3} R_{ihkm}\Big)g_{lj}-\Big(-T_{ikml} +T_{imkl}+\frac{1}{2}S_{ilkm}+\frac{\tau}{3}R_{ilkm}\Big)g_{hj}\\
&-\Big(-T_{hjml} +T_{hmjl}+\frac{1}{2}S_{hljm}+\frac{\tau}{3} R_{hljm}\Big)g_{ik}+\Big(-T_{hjkl} +T_{hkjl}+\frac{1}{2}S_{hljk}+\frac{\tau}{3} R_{hljk}\Big)g_{im}\\
&+\Big(-T_{hkml} +T_{hmkl}+\frac{1}{2}S_{hlkm}+\frac{\tau}{3} R_{hlkm}\Big)g_{ij}\Big\}\\
&+A_{hjkmil}-A_{ljkmih}-A_{ijkmhl}+A_{ijmkhl}-A_{hjmkil}+A_{ljmkih}-A_{ikmjhl}+A_{hkmjil}-A_{lkmjih}\\
&=0,
\end{aligned}
\end{equation}
where $T_{pqrs} = R_{pabq}R_{rabs}$, $S_{pqrs} = R_{abpq}R_{abrs}$, and $A_{pqrstu} = R_{apqr}R_{astu}$.
\begin{proof}
The proof is similar to that of Lemma 3.1 for the case when $m=6$ and $r=2$. Here, the Weyl tensor W is given by
\begin{equation*}
W_{abcd} = R_{abcd} - \frac{1}{4}(\rho_{ad}g_{bc}+\rho_{bc}g_{ad} - \rho_{ac}g_{bd} -\rho_{bd}g_{ac}) + \frac{\tau}{20}(g_{ad}g_{bc} - g_{ac}g_{bd}).
\end{equation*}
Next we expand \eqref{eq:Pa_n} for $W$ and rearrange them. Using the Einstein condition $\rho_{ij} = \frac{\tau}{6}g_{ij}$, we obtain 34 terms. Here we deal with a few terms among them. For more details, we refer to appendix. The numbers shown below correspond to those in appendix.
\begin{enumerate}[label={(\arabic*)},leftmargin=0cm,align=left]
\item[(1)] $\big(||R||^2 - 4||\rho||^2 + \tau^2 \big)g_{ij}g_{hk}g_{lm}=\left(||R||^2 +\dfrac{\tau^2}{3} \right)g_{ij}g_{hk}g_{lm}$
\item[(7)]
$\big(-4R_{iabc}R_{jabc} + 8\rho_{ia}\rho_{ja} +8R_{iabj}\rho_{ab}-4\tau\rho_{ij}\big)g_{hk}g_{lm}=-4\tR_{ij}g_{hk}g_{lm} -\dfrac{2}{9}\tau^2 g_{ij}g_{hk}g_{lm}$
\item[(25)]
\begin{align*}
&\big(-8R_{iabj}R_{kabh} +8R_{iabk}R_{jabh} +4R_{abih}R_{abjk} -8R_{aijk}\rho_{ah} +8R_{akih}\rho_{aj} -8R_{ajih}\rho_{ak}\\
&\quad+8R_{ahjk}\rho_{ai} +8\rho_{ij}\rho_{hk} -8\rho_{ik}\rho_{hj} -4\tau R_{ihjk}\big)g_{lm}\\
&=\Big(-8T_{ijkh} +8T_{ikjh} +4S_{ihjk} +\dfrac{4}{3}\tau R_{ihjk} +\dfrac{2}{9}\tau^2 g_{ij}g_{hk} - \dfrac{2}{9}\tau^2 g_{ik}g_{hj} \Big)g_{lm}
\end{align*}
\item[(34)]
\begin{align*}
&8R_{ahjk}R_{amil}-8R_{aljk}R_{amih}-8R_{aijk}R_{amhl}+8R_{aijm}R_{akhl}-8R_{ahjm}R_{akil}+8R_{aljm}R_{akih}\\
&\;-8R_{aikm}R_{ajhl}+8R_{ahkm}R_{ajil}-8R_{alkm}R_{ajih} -8R_{ilkm}\rho_{hj}+8R_{iljm}\rho_{hk}+8R_{hljk}\rho_{im}\\
&\;+8R_{ihjk}\rho_{lm}+8R_{ihkm}\rho_{lj}-8R_{ihjm}\rho_{lk}-8R_{iljk}\rho_{hm}+8R_{hlkm}\rho_{ij}-8R_{hljm}\rho_{ik}\\
&= 8(A_{hjkmil}-A_{ljkmih}-A_{ijkmhl}+A_{ijmkhl}-A_{hjmkil}+A_{ljmkih}-A_{ikmjhl}+A_{hkmjil}\\
&\quad-A_{lkmjih}){{+\frac{4}{3}\tau}} \Big({{-}}R_{ilkm}g_{hj}+ R_{iljm}g_{hk}+ R_{hljk}g_{im}+ R_{ihjk}g_{lm}+ R_{ihkm}g_{lj}- R_{ihjm}g_{lk}\\
&\quad- R_{iljk}g_{hm}+R_{hlkm}g_{ij}- R_{hljm}g_{ik}\Big).
\end{align*}
\end{enumerate}
By rearranging all terms, we complete the proof of Lemma~\ref{thm:6-einstien(0,6)-tensor}.
\end{proof}
\end{lem}
Now, we transvect {{each term}} of \eqref{eq:6-einstein(0,6)-tensor} with $R_{ihjk}$.
For the first term of \eqref{eq:6-einstein(0,6)-tensor}, we have
\begin{equation*}
\begin{aligned}
&\frac{1}{8}\Big(||R||^2 +\frac{\tau^2}{3}\Big)\Big(g_{ij}g_{hk}g_{lm}-g_{ij}g_{hm}g_{lk}-g_{ik}g_{hj}g_{lm}+g_{ik}g_{hm}g_{lj}+g_{im}g_{hj}g_{lk}-g_{im}g_{hk}g_{lj}\Big) R_{ihjk}\\
&=-\frac{1}{6}\Big(\tau||R||^2 +\frac{\tau^3}{3}\Big) g_{lm}.
\end{aligned}
\end{equation*}
\noindent For the second term, we have
\begin{equation*}
-\frac{1}{2}\tR_{ij}(g_{hk}g_{lm}-g_{hm}g_{lk})R_{ihjk}=\frac{\tau}{12}||R||^2g_{lm}-\frac{1}{2}\check{R}_{lm},
\end{equation*}
\begin{equation*}
\frac{1}{2}\tR_{ik}(g_{hj}g_{lm}-g_{hm}g_{lj}) R_{ihjk}=\frac{\tau}{12}||R||^2g_{lm}-\frac{1}{2}\check{R}_{lm},\; -\frac{1}{2}\tR_{im}(g_{hj}g_{lk}-g_{hk}g_{lj}) R_{ihjk}=-\frac{\tau}{6}\tR_{lm},
\end{equation*}
\begin{equation*}
\frac{1}{2}\tR_{hj}(g_{ik}g_{lm}-g_{im}g_{lk})R_{ihjk}=\frac{\tau}{12}||R||^2g_{lm}-\frac{1}{2}\check{R}_{lm},
\end{equation*}
\begin{equation*}
-\frac{1}{2}\tR_{hk}(g_{ij}g_{lm}-g_{im}g_{lj})R_{ihjk}=\frac{\tau}{12}||R||^2g_{lm}-\frac{1}{2}\check{R}_{lm},
\end{equation*}
\begin{equation*}
\frac{1}{2}\tR_{hm}(g_{ij}g_{lk}-g_{ik}g_{lj})R_{ihjk}=-\frac{\tau}{6}\tR_{lm}, \;-\frac{1}{2}\tR_{lj}(g_{ik}g_{hm}-g_{im}g_{hk}) R_{ihjk}=-\frac{\tau}{6}\tR_{lm},
\end{equation*}
\begin{equation*}
\frac{1}{2}\tR_{lk}(g_{ij}g_{hm}-g_{im}g_{hj}) R_{ihjk}=-\frac{\tau}{6}\tR_{lm},\; -\frac{1}{2}\tR_{lm}(g_{ij}g_{hk}-g_{ik}g_{hj})R_{ihjk}=\tau\tR_{lm}.
\end{equation*}
Thus, the second term of \eqref{eq:6-einstein(0,6)-tensor} transvecting with $R_{ihjk}$ becomes $\dfrac{\tau}{3}||R||^2 g_{lm}-2\check{R}_{lm}+\dfrac{\tau}{3}\tR_{lm}$.\\
For the third term, we have
\begin{equation*}
\begin{aligned}
&\Big(-T_{ijkh} + T_{ikjh} +\frac{1}{2}S_{ihjk}+\frac{\tau}{3} R_{ihjk}\Big)g_{lm}{{R_{ihjk}}} \\
&=\left(-R_{iabj}R_{kabh} R_{ihjk} +R_{iabk}R_{jabh} R_{ihjk} +\frac{1}{2}R_{abih}R_{abjk} R_{ihjk} +\dfrac{1}{3}\tau R_{ihjk} R_{ihjk} \right)g_{lm}\\
&=\left(-2\mathring{R}+\hat{R}+\frac{1}{3}\tau||R||^2\right)g_{lm},
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}
-&\Big(-T_{ijkl} +T_{ikjl}+\frac{1}{2}S_{iljk}+\frac{\tau}{3} R_{iljk}\Big)g_{hm}{{R_{ihjk}}}\\
&=\Big(R_{iabj}R_{kabl} R_{imjk} -R_{iabk}R_{jabl} R_{imjk}-\frac{1}{2}R_{abil}R_{abjk} R_{imjk}-\dfrac{\tau}{3} R_{iljk} R_{imjk}\Big)\\
&=2\mathring{R}_{lm}-\hat{R}_{lm}-\frac{\tau}{3}\tR_{lm},
\end{aligned}
\end{equation*}
\begin{equation*}
-\Big(-T_{ijmh}+T_{imjh}+\frac{1}{2}S_{ihjm}+\frac{\tau}{3} R_{ihjm}\Big)g_{lk}{{R_{ihjk}}}=2\mathring{R}_{lm}-\hat{R}_{lm}-\frac{\tau}{3}\tR_{lm},
\end{equation*}
\begin{align*}
&\Big(-T_{ijml} +T_{imjl}+\frac{1}{2}S_{iljm}+\frac{\tau}{3} R_{iljm}\Big)g_{hk}{{R_{ihjk}}}\\
&=\left(-R_{iabj}R_{mabl}+R_{iabm}R_{jabl}+\frac{1}{2}R_{abil}R_{abjm}+\frac{\tau}{3}R_{iljm}\right)(-\rho_{ij})\\
&=-\frac{\tau}{6}\left(-\frac{\tau^2}{36}g_{lm}+\tR_{lm}+\frac{1}{2}\tR_{lm}-\frac{\tau^2}{18}g_{lm}\right)\\
&=\frac{\tau^3}{72}g_{lm}-\frac{\tau}{4}\tR_{lm},
\end{align*}
\begin{equation*}
\Big(-T_{ikmh} +T_{imkh}+\frac{1}{2}S_{ihkm}+\frac{\tau}{3} R_{ihkm}\Big)g_{lj}{{ R_{ihjk}}}
=2\mathring{R}_{lm}-\hat{R}_{lm}-\frac{\tau}{3}\tR_{lm},
\end{equation*}
\begin{equation*}
-\Big(-T_{ikml}+T_{imkl}+\frac{1}{2}S_{ilkm}+\frac{\tau}{3} R_{ilkm}\Big)g_{hj}{{ R_{ihjk}}}
=\frac{\tau^3}{72}g_{lm}-\frac{\tau}{4}\tR_{lm},
\end{equation*}
\begin{equation*}
-\Big(-T_{hjml} +T_{hmjl}+\frac{1}{2}S_{hljm}+\frac{\tau}{3} R_{hljm}\Big)g_{ik}R_{ihjk} =\frac{\tau^3}{72}g_{lm}-\frac{\tau}{4}\tR_{lm},
\end{equation*}
\begin{equation*}
\Big(-T_{hjkl} +T_{hkjl}+\frac{1}{2}S_{hljk}+\frac{\tau}{3} R_{hljk}\Big)g_{im}{{ R_{ihjk}}}
=2\mathring{R}_{lm}-\hat{R}_{lm}-\frac{\tau}{3}\tR_{lm},
\end{equation*}
\begin{equation*}
\Big(-T_{hkml} +T_{hmkl}+\frac{1}{2}S_{hlkm}+\frac{\tau}{3} R_{hlkm}\Big)g_{ij}{{ R_{ihjk}}} =\frac{\tau^3}{72}g_{lm}-\frac{\tau}{4}\tR_{lm}.
\end{equation*}
Thus, the third {{term}} of \eqref{eq:6-einstein(0,6)-tensor} transvecting with $R_{ihjk}$ becomes
\begin{equation*}
\left(-2\mathring{R}+\hat{R}+\frac{\tau}{3}||R||^2 + \frac{\tau^3}{18}\right)g_{lm}+8\mathring{R}_{lm}-4\hat{R}_{lm}-\frac{7}{3}\tau\tR_{lm}.
\end{equation*}
Making use of \eqref{eq:re}, the remaining terms in \eqref{eq:6-einstein(0,6)-tensor} transvecting with $R_{ihjk}$, we have
\begin{equation*}
\begin{aligned}
&(A_{hjkmil}-A_{ljkmih}-A_{ijkmhl}+A_{ijmkhl}-A_{hjmkil}\\
&+A_{ljmkih}-A_{ikmjhl}+A_{hkmjil}-A_{lkmjih}){{ R_{ihjk}}}\\
=\;& R_{ahjk}R_{amil} R_{ihjk}-R_{aljk}R_{amih} R_{ihjk}-R_{aijk}R_{amhl} R_{ihjk}\\
&+R_{aijm}R_{akhl} R_{ihjk}-R_{ahjm}R_{akil} R_{ihjk}+R_{aljm}R_{akih} R_{ihjk}\\
&-R_{aikm}R_{ajhl} R_{ihjk}+R_{ahkm}R_{ajil} R_{ihjk}-R_{alkm}R_{ajih} R_{ihjk}\\
=&-4\check{R}_{lm}-2\hat{R}_{lm}+4\mathring{R}_{lm}.
\end{aligned}
\end{equation*}
{{Summing all terms, then we complete the proof of Theorem~B-\ref{eq:6-einstein}.
\smallskip
Now, we prove Theorem~B-\ref{eq:6-dim2stein}. Applying~\eqref{eq:supEin} to \eqref{eq:6-einstein(0,6)-tensor}, we have}}
\begin{equation}\label{eq:6-2stein(0,6)-tensor}
\begin{aligned}
&-\frac{1}{8}\Big(||R||^2 -\frac{\tau^2}{3}\Big)\Big(g_{ij}g_{hk}g_{lm}-g_{ij}g_{hm}g_{lk}-g_{ik}g_{hj}g_{lm}+g_{ik}g_{hm}g_{lj}+g_{im}g_{hj}g_{lk}-g_{im}g_{hk}g_{lj}\Big)\\
&+\Big(-T_{ijkh} + T_{ikjh} +\frac{1}{2}S_{ihjk}+\frac{\tau}{3} R_{ihjk}\Big)g_{lm}-\Big(-T_{ijkl} +T_{ikjl}+\frac{1}{2}S_{iljk}+\frac{\tau}{3} R_{iljk}\Big)g_{hm}\\
&-\Big(-T_{ijmh}+T_{imjh}+\frac{1}{2}S_{ihjm}+\frac{\tau}{3} R_{ihjm}\Big)g_{lk}+\Big(-T_{ijml} +T_{imjl}+\frac{1}{2}S_{iljm}+\frac{\tau}{3} R_{iljm}\Big)g_{hk}\\
&+\Big(-T_{ikmh} +T_{imkh}+\frac{1}{2}S_{ihkm}+\frac{\tau}{3} R_{ihkm}\Big)g_{lj}-\Big(-T_{ikml} +T_{imkl}+\frac{1}{2}S_{ilkm}+\frac{\tau}{3}R_{ilkm}\Big)g_{hj}\\
&-\Big(-T_{hjml} +T_{hmjl}+\frac{1}{2}S_{hljm}+\frac{\tau}{3} R_{hljm}\Big)g_{ik}+\Big(-T_{hjkl} +T_{hkjl}+\frac{1}{2}S_{hljk}+\frac{\tau}{3} R_{hljk}\Big)g_{im}\\
&+\Big(-T_{hkml} +T_{hmkl}+\frac{1}{2}S_{hlkm}+\frac{\tau}{3} R_{hlkm}\Big)g_{ij}+A_{hjkmil}-A_{ljkmih}-A_{ijkmhl}+A_{ijmkhl}\\
&-A_{hjmkil}+A_{ljmkih}-A_{ikmjhl}+A_{hkmjil}-A_{lkmjih}=0.
\end{aligned}
\end{equation}
{{We transvect {{each term}} of \eqref{eq:6-2stein(0,6)-tensor} with $R_{ihjk}$.}} Here, we give a few representative terms as follows:
\begin{equation*}
\begin{aligned}
&-\frac{1}{8}\Big(||R||^2 -\frac{\tau^2}{3}\Big)\Big(g_{ij}g_{hk}g_{lm}-g_{ij}g_{hm}g_{lk}-g_{ik}g_{hj}g_{lm}+g_{ik}g_{hm}g_{lj}+g_{im}g_{hj}g_{lk}-g_{im}g_{hk}g_{lj}\Big){{R_{ihjk}}}\\&=\frac{1}{6}\tau\Big(||R||^2 -\frac{\tau^2}{3}\Big) g_{lm},
\end{aligned}
\end{equation*}
\begin{equation*}
\Big(-T_{ijkh} + T_{ikjh} +\frac{1}{2}S_{ihjk}+\frac{\tau}{3} R_{ihjk}\Big)g_{lm}{{R_{ihjk}}} =\left(-2\mathring{R}+\hat{R}+\frac{8}{3}\tau||R||^2\right)g_{lm},
\end{equation*}
\begin{equation*}
-\Big(-T_{ijkl} +T_{ikjl}+\frac{1}{2}S_{iljk}+\frac{\tau}{3} R_{iljk}\Big)g_{hm}{{R_{ihjk}}}=2\mathring{R}_{lm}-\hat{R}_{lm}-\frac{1}{18}\tau||R||^2g_{lm},
\end{equation*}
{{\begin{equation*}
\begin{aligned}
&(A_{hjkmil}-A_{ljkmih}-A_{ijkmhl}+A_{ijmkhl}-A_{hjmkil}+A_{ljmkih}-A_{ikmjhl}+A_{hkmjil}-A_{lkmjih})R_{ihjk}\\
&=-4\check{R}_{lm}-2\hat{R}_{lm}+4\mathring{R}_{lm}\\
&=-\frac{\tau}{9}||R||^2 g_{lm}-2\hat{R}_{lm}+4\mathring{R}_{lm}.
\end{aligned}
\end{equation*}}}
Similarly, we can obtain remaining terms by transvecting with $R_{ihjk}$. Then, by rearranging all terms, we have Theorem~B-\ref{eq:6-dim2stein}.
\smallskip
Now we give an example of Theorem B.
\begin{ex}
Let $M$ be a Riemannian product manifold of 3-dimensional Riemannian manifolds $M_1(k)$ and $M_2(k)$ of constant sectional curvature $k$ ($k \neq 0$). {{Then we can easily check that $M$ is an Einstein manifold. But $M$ can never be 2-stein: Let $X=(X_1 , X_2)$ is a tangent vector with $X_1$, $X_2$ its components tangent to $M_1 (k)$ and $M_2 (k)$, respectively. Then, $\Tr(R_X^2) = 2k^2 (||X_1 ||^4 + ||X_2 ||^4)$ which cannot be equal to $\mu_2(||X_1 ||^2 + ||X_2 ||^2)^2$.}} Let $\{e_{i}\}, \; i=1,2,\ldots,6$ be an orthonormal basis of $M$, where $\{e_{1},e_{2},e_{3}\}$ and $\{e_{4},e_{5},e_{6}\}$ are bases for $M_1(k)$ and $M_2(k)$, respectively. Then, we have
\begin{equation}\label{eq:ex_Ein}
R_{1221}=R_{1331}=R_{2332}=R_{4554}=R_{4664}=R_{5665}=k,
\end{equation}
and all the other components of $R$ vanish. From \eqref{eq:ex_Ein} we have
\begin{gather*}
\tau = 12k,\quad ||R||^2 = 24 k^2,\quad \mathring{R} = -12k^3,\quad \hat{R} = -48k^3,\\
\check{R}_{ij}=8k^3 \delta_{ij},\quad \hat{R}_{ij}=-8k^3 \delta_{ij},\quad \mathring{R}_{ij} = -2k^3 \delta_{ij},\quad \tR_{ij}=4k^2 \delta_{ij}.
\end{gather*}
Therefore, we find that the curvature identities of Theorem B hold on $M$. Here we note that $M$ is a super-Einstein manifold.
\end{ex}
\section{Appendix}
In this appendix we give the proof of Lemma \ref{thm:6-einstien(0,6)-tensor}.
{{By replacing}} the curvature tensor $R$ by the Weyl curvature tensor $W$ in (\ref{eq:Pa_n}), we can also obtain the curvature identity of $W$. In the case of $m=6$ and $r=2$, making use of the fact that the Weyl curvature tensor is traceless, we have the following curvature identity.
\begin{prop}The Weyl curvature tensor $W$ of any 6-dimensional Riemannian manifold satisfies the following identity{\rm :}
\begin{equation}\label{eq:weyl_iden}
\begin{aligned}
0=||W||^2\big\{&g_{ij}g_{hk}g_{lm}+g_{ik}g_{hm}g_{lj}+g_{im}g_{hj}g_{lk}-g_{ik}g_{hj}g_{lm}-g_{ij}g_{hm}g_{lk}-g_{im}g_{hk}g_{lj}\big\}\\
-4\big\{&W_{iabc}W_{jabc}g_{hk}g_{lm}+W_{iabc}W_{kabc}g_{hm}g_{lj}+W_{iabc}W_{mabc}g_{hj}g_{lk}\\
&-W_{iabc}W_{kabc}g_{hj}g_{lm}-W_{iabc}W_{jabc}g_{hm}g_{lk}-W_{iabc}W_{mabc}g_{hk}g_{lj}\\
&-W_{habc}W_{jabc}g_{ik}g_{lm}-W_{habc}W_{kabc}g_{im}g_{lj}-W_{habc}W_{mabc}g_{ij}g_{lk}\\
&+W_{habc}W_{kabc}g_{ij}g_{lm}+W_{habc}W_{jabc}g_{im}g_{lk}+W_{habc}W_{mabc}g_{ik}g_{lj}\\
&+W_{labc}W_{jabc}g_{hk}g_{im}+W_{labc}W_{kabc}g_{hm}g_{ij}+W_{labc}W_{mabc}g_{hj}g_{ik}\\
&-W_{labc}W_{kabc}g_{hj}g_{im}-W_{labc}W_{jabc}g_{hm}g_{ik}-W_{labc}W_{mabc}g_{hk}g_{ij}\big\}\\
-8\big\{&(W_{iabj}W_{kabh}-W_{iabk}W_{jabh})g_{lm}-(W_{iabj}W_{mabh}-W_{iabm}W_{jabh})g_{lk}\\
&+(W_{iabk}W_{mabh}-W_{iabm}W_{kabh})g_{lj}-(W_{iabj}W_{kabl}-W_{iabk}W_{jabl})g_{hm}\\
&+(W_{iabj}W_{mabl}-W_{iabm}W_{jabl})g_{hk}-(W_{iabk}W_{mabl}-W_{iabm}W_{kabl})g_{hj}\\
&+(W_{habj}W_{kabl}-W_{habk}W_{jabl})g_{im}-(W_{habj}W_{mabl}-W_{habm}W_{jabl})g_{ik}\\
&+(W_{habk}W_{mabl}-W_{habm}W_{kabl})g_{ij}\big\}
\end{aligned}
\end{equation}
\begin{align*}
\quad\qquad\qquad\;+4\big\{&W_{abih}W_{abjk}g_{lm}-W_{abih}W_{abjm}g_{lk}+W_{abih}W_{abkm}g_{lj}-W_{abil}W_{abjk}g_{hm}\\
&+W_{abil}W_{abjm}g_{hk}-W_{abil}W_{abkm}g_{hj}+W_{abhl}W_{abjk}g_{im}-W_{abhl}W_{abjm}g_{ik}\\
&+W_{abhl}W_{abkm}g_{ij}\big\}\\
\quad\qquad\qquad\;+8\big\{&W_{ahjk}W_{amil}-W_{aljk}W_{amih}-W_{ajhl}W_{aikm}-W_{aijk}W_{amhl}+W_{ajih}W_{almk}\\
&+W_{ajil}W_{ahkm}-W_{ahjm}W_{akil}+W_{aljm}W_{akih}+W_{aijm}W_{akhl}\big\}.
\end{align*}
\end{prop}
Since the Weyl tensor W on a 6-dimensional Riemannian manifold is given by
\begin{equation*}
W_{abcd} = R_{abcd} - \frac{1}{4}(\rho_{ad}g_{bc}+\rho_{bc}g_{ad} - \rho_{ac}g_{bd} -\rho_{bd}g_{ac}) + \frac{\tau}{20}(g_{ad}g_{bc} - g_{ac}g_{bd}),
\end{equation*}
we substitute the above Weyl tensor into \eqref{eq:weyl_iden} and use the Einstein condition $\rho_{ij} = \frac{\tau}{6}g_{ij}$. Then, we obtain the following 34 terms.
{{\begin{enumerate}[label={(\arabic*)},leftmargin=0cm,align=left]
\item $\big(||R||^2 - 4||\rho||^2 + \tau^2 \big)g_{ij}g_{hk}g_{lm}=\left(||R||^2 +\dfrac{\tau^2}{3} \right)g_{ij}g_{hk}g_{lm}$
\item $-\big(||R||^2 - 4||\rho||^2 + \tau^2 \big)g_{ij}g_{hm}g_{lk}=-\left(||R||^2 +\dfrac{\tau^2}{3} \right)g_{ij}g_{hm}g_{lk}$
\item $-\big(||R||^2 - 4||\rho||^2 + \tau^2 \big)g_{ik}g_{hj}g_{lm}=-\left(||R||^2 +\dfrac{\tau^2}{3} \right)g_{ik}g_{hj}g_{lm}$
\item $\big(||R||^2 - 4||\rho||^2 + \tau^2 \big)g_{ik}g_{hm}g_{lj}=\left(||R||^2 +\dfrac{\tau^2}{3} \right)g_{ik}g_{hm}g_{lj}$
\item $\big(||R||^2 - 4||\rho||^2 + \tau^2 \big)g_{im}g_{hj}g_{lk}=\left(||R||^2 +\dfrac{\tau^2}{3} \right)g_{im}g_{hj}g_{lk}$
\item $-\big(||R||^2 - 4||\rho||^2 + \tau^2 \big)g_{im}g_{hk}g_{lj}=-\left(||R||^2 +\dfrac{\tau^2}{3} \right)g_{im}g_{hk}g_{lj}$
\smallskip
\item $\big(-4R_{iabc}R_{jabc} + 8\rho_{ia}\rho_{ja} +8R_{iabj}\rho_{ab}-4\tau\rho_{ij}\big)g_{hk}g_{lm}=\left(-4R_{iabc}R_{jabc} -\dfrac{2}{9}\tau^2 g_{ij}\right)g_{hk}g_{lm}$
\item $-\big(-4R_{iabc}R_{jabc} + 8\rho_{ia}\rho_{ja} +8R_{iabj}\rho_{ab}-4\tau\rho_{ij}\big)g_{hm}g_{lk}=-\left(-4R_{iabc}R_{jabc} -\dfrac{2}{9}\tau^2 g_{ij}\right)g_{hm}g_{lk}$
\item $-\big(-4R_{iabc}R_{kabc} + 8\rho_{ia}\rho_{ka} +8R_{iabk}\rho_{ab}-4\tau\rho_{ik}\big)g_{hj}g_{lm}=-\left(-4R_{iabc}R_{kabc} -\dfrac{2}{9}\tau^2 g_{ik}\right)g_{hj}g_{lm}$
\item $\big(-4R_{iabc}R_{kabc} + 8\rho_{ia}\rho_{ka} +8R_{iabk}\rho_{ab}-4\tau\rho_{ik}\big)g_{hm}g_{lj}=\left(-4R_{iabc}R_{kabc} -\dfrac{2}{9}\tau^2 g_{ik}\right)g_{hm}g_{lj}$
\item $\big(-4R_{iabc}R_{mabc} + 8\rho_{ia}\rho_{ma} +8R_{iabm}\rho_{ab}-4\tau\rho_{im}\big)g_{hj}g_{lk}=\left(-4R_{iabc}R_{mabc} -\dfrac{2}{9}\tau^2 g_{im}\right)g_{hj}g_{lk}$
\item $-\big(-4R_{iabc}R_{mabc} + 8\rho_{ia}\rho_{ma} +8R_{iabm}\rho_{ab}-4\tau\rho_{im}\big)g_{hk}g_{lj}=-\left(-4R_{iabc}R_{mabc} -\dfrac{2}{9}\tau^2 g_{im}\right)g_{hk}g_{lj}$
\item $-\big(-4R_{habc}R_{jabc} + 8\rho_{ha}\rho_{ja} +8R_{habj}\rho_{ab}-4\tau\rho_{hj}\big)g_{ik}g_{lm}=-\left(-4R_{habc}R_{jabc} -\dfrac{2}{9}\tau^2 g_{hj}\right)g_{ik}g_{lm}$
\item $\big(-4R_{habc}R_{jabc} + 8\rho_{ha}\rho_{ja} +8R_{habj}\rho_{ab}-4\tau\rho_{hj}\big)g_{im}g_{lk}=\left(-4R_{habc}R_{jabc} -\dfrac{2}{9}\tau^2 g_{hj}\right)g_{im}g_{lk}$
\item $\big(-4R_{habc}R_{kabc} + 8\rho_{ha}\rho_{ka} +8R_{habk}\rho_{ab}-4\tau\rho_{hk}\big)g_{ij}g_{lm}=\left(-4R_{habc}R_{kabc} -\dfrac{2}{9}\tau^2 g_{hk}\right)g_{ij}g_{lm}$
\item $-\big(-4R_{habc}R_{kabc} + 8\rho_{ha}\rho_{ka} +8R_{habk}\rho_{ab}-4\tau\rho_{hk}\big)g_{im}g_{lj}=-\left(-4R_{habc}R_{kabc} -\dfrac{2}{9}\tau^2 g_{hk}\right)g_{im}g_{lj}$
\item $-\big(-4R_{habc}R_{mabc} + 8\rho_{ha}\rho_{ma} +8R_{habm}\rho_{ab}-4\tau\rho_{hm}\big)g_{ij}g_{lk}=-\left(-4R_{habc}R_{mabc} -\dfrac{2}{9}\tau^2 g_{hm}\right)g_{ij}g_{lk}$
\item $\big(-4R_{habc}R_{mabc} + 8\rho_{ha}\rho_{ma} +8R_{habm}\rho_{ab}-4\tau\rho_{hm}\big)g_{ik}g_{lj}
=\left(-4R_{habc}R_{mabc} -\dfrac{2}{9}\tau^2 g_{hm}\right)g_{ik}g_{lj}$
\item $\big(-4R_{labc}R_{jabc} + 8\rho_{la}\rho_{ja} +8R_{labj}\rho_{ab}-4\tau\rho_{lj}\big)g_{ik}g_{hm}
=\left(-4R_{labc}R_{jabc} -\dfrac{2}{9}\tau^2 g_{lj}\right)g_{ik}g_{hm}$
\item $-\big(-4R_{labc}R_{jabc} + 8\rho_{la}\rho_{ja} +8R_{labj}\rho_{ab}-4\tau\rho_{lj}\big)g_{im}g_{hk}
=-\left(-4R_{labc}R_{jabc} -\dfrac{2}{9}\tau^2 g_{lj}\right)g_{im}g_{hk}$
\item $-\big(-4R_{labc}R_{kabc} + 8\rho_{la}\rho_{ka} +8R_{labk}\rho_{ab}-4\tau\rho_{lk}\big)g_{ij}g_{hm}
=-\left(-4R_{labc}R_{kabc} -\dfrac{2}{9}\tau^2 g_{lk}\right)g_{ij}g_{hm}$
\item $\big(-4R_{labc}R_{kabc} + 8\rho_{la}\rho_{ka} +8R_{labk}\rho_{ab}-4\tau\rho_{lk}\big)g_{im}g_{hj}
=\left(-4R_{labc}R_{kabc} -\dfrac{2}{9}\tau^2 g_{lk}\right)g_{im}g_{hj}$
\item $\big(-4R_{labc}R_{mabc} + 8\rho_{la}\rho_{ma} +8R_{labm}\rho_{ab}-4\tau\rho_{lm}\big)g_{ij}g_{hk}
=\left(-4R_{labc}R_{mabc} -\dfrac{2}{9}\tau^2 g_{lm}\right)g_{ij}g_{hk}$
\item $-\big(-4R_{labc}R_{mabc} + 8\rho_{la}\rho_{ma} +8R_{labm}\rho_{ab}-4\tau\rho_{lm}\big)g_{ik}g_{hj}
=-\left(-4R_{labc}R_{mabc} -\dfrac{2}{9}\tau^2 g_{lm}\right)g_{ik}g_{hj}$
\smallskip
\item \begin{align*}
&\big(-8R_{iabj}R_{kabh} +8R_{iabk}R_{jabh} +4R_{abih}R_{abjk} -8R_{aijk}\rho_{ah} +8R_{akih}\rho_{aj} -8R_{ajih}\rho_{ak}\\
&\quad+8R_{ahjk}\rho_{ai} +8\rho_{ij}\rho_{hk} -8\rho_{ik}\rho_{hj} -4\tau R_{ihjk}\big)g_{lm}\\
&=\Big(-8R_{iabj}R_{kabh} +8R_{iabk}R_{jabh} +4R_{abih}R_{abjk} +\dfrac{4}{3}\tau R_{ihjk} +\dfrac{2}{9}\tau^2 g_{ij}g_{hk} - \dfrac{2}{9}\tau^2 g_{ik}g_{hj} \Big)g_{lm}
\end{align*}
\item \begin{align*}
&-\big(-8R_{iabj}R_{kabl} +8R_{iabk}R_{jabl} +4R_{abil}R_{abjk} -8R_{aijk}\rho_{al} +8R_{akil}\rho_{aj} -8R_{ajil}\rho_{ak}\\
&\quad+8R_{aljk}\rho_{ai} +8\rho_{ij}\rho_{lk} -8\rho_{ik}\rho_{lj} -4\tau R_{iljk}\big)g_{hm}\\
&=-\Big(-8R_{iabj}R_{kabl} +8R_{iabk}R_{jabl} +4R_{abil}R_{abjk} +\dfrac{4}{3}\tau R_{iljk} +\dfrac{2}{9}\tau^2 g_{ij}g_{lk} - \dfrac{2}{9}\tau^2 g_{ik}g_{lj}\Big)g_{hm}
\end{align*}
\item \begin{align*}
&-\big(-8R_{iabj}R_{mabh} +8R_{iabm}R_{jabh} +4R_{abih}R_{abjm} -8R_{aijm}\rho_{ah} +8R_{amih}\rho_{aj} -8R_{ajih}\rho_{am}\\
&\quad+8R_{ahjm}\rho_{ai} +8\rho_{ij}\rho_{hm} -8\rho_{im}\rho_{hj} -4\tau R_{ihjm}\big)g_{lk}\\
&=-\Big(-8R_{iabj}R_{mabh} +8R_{iabm}R_{jabh} +4R_{abih}R_{abjm} +\dfrac{4}{3}\tau R_{ihjm} +\dfrac{2}{9}\tau^2 g_{ij}g_{hm} - \dfrac{2}{9}\tau^2 g_{im}g_{hj}\Big)g_{lk}
\end{align*}
\item \begin{align*}
&\big(-8R_{iabj}R_{mabl} +8R_{iabm}R_{jabl} +4R_{abil}R_{abjm} -8R_{aijm}\rho_{al} +8R_{amil}\rho_{aj} -8R_{ajil}\rho_{am}\\
&\quad+8R_{aljm}\rho_{ai} +8\rho_{ij}\rho_{lm} -8\rho_{im}\rho_{lj} -4\tau R_{iljm}\big)g_{hk}\\
&=\Big(-8R_{iabj}R_{mabl} +8R_{iabm}R_{jabl} +4R_{abil}R_{abjm} +\dfrac{4}{3}\tau R_{iljm} +\dfrac{2}{9}\tau^2 g_{ij}g_{lm} - \dfrac{2}{9}\tau^2 g_{im}g_{lj}\Big)g_{hk}
\end{align*}
\item \begin{align*}
&\big(-8R_{iabk}R_{mabh} +8R_{iabm}R_{kabh} +4R_{abih}R_{abkm} -8R_{aikm}\rho_{ah} +8R_{amih}\rho_{ak} -8R_{akih}\rho_{am} \\
&\quad+8R_{ahkm}\rho_{ai} +8\rho_{ik}\rho_{hm} -8\rho_{im}\rho_{hk} -4\tau R_{ihkm}\big)g_{lj}\\
&=\Big(-8R_{iabk}R_{mabh} +8R_{iabm}R_{kabh} +4R_{abih}R_{abkm} +\dfrac{4}{3}\tau R_{ihkm} +\dfrac{2}{9}\tau^2 g_{ik}g_{hm} - \dfrac{2}{9}\tau^2 g_{im}g_{hk}\Big)g_{lj}
\end{align*}
\item \begin{align*}
&-\big(-8R_{iabk}R_{mabl} +8R_{iabm}R_{kabl} +4R_{abil}R_{abkm} -8R_{aikm}\rho_{al} +8R_{amil}\rho_{ak} -8R_{akil}\rho_{am} \\
&\quad+8R_{alkm}\rho_{ai} +8\rho_{ik}\rho_{lm} -8\rho_{im}\rho_{lk} -4\tau R_{ilkm}\big)g_{hj}\\
&=-\Big(-8R_{iabk}R_{mabl} +8R_{iabm}R_{kabl} +4R_{abil}R_{abkm} +\dfrac{4}{3}\tau R_{ilkm} +\dfrac{2}{9}\tau^2 g_{ik}g_{lm} - \dfrac{2}{9}\tau^2 g_{im}g_{lk}\Big)g_{hj}
\end{align*}
\item \begin{align*}
&-\big(-8R_{habj}R_{mabl} +8R_{habm}R_{jabl} +4R_{abhl}R_{abjm} -8R_{ahjm}\rho_{al} +8R_{amhl}\rho_{aj} -8R_{ajhl}\rho_{am}\\
&\quad+8R_{aljm}\rho_{ah} +8\rho_{hj}\rho_{lm} -8\rho_{hm}\rho_{lj} -4\tau R_{hljm}\big)g_{ik}\\
&=-\Big(-8R_{habj}R_{mabl} +8R_{habm}R_{jabl} +4R_{abhl}R_{abjm} +\dfrac{4}{3}\tau R_{hljm} +\dfrac{2}{9}\tau^2 g_{hj}g_{lm} - \dfrac{2}{9}\tau^2 g_{hm}g_{lj}\Big)g_{ik}
\end{align*}
\item \begin{align*}
&\big(-8R_{habj}R_{kabl} +8R_{habk}R_{jabl} +4R_{abhl}R_{abjk} -8R_{ahjk}\rho_{al} +8R_{akhl}\rho_{aj} -8R_{ajhl}\rho_{ak} \\
&\quad+8R_{aljk}\rho_{ah} +8\rho_{hj}\rho_{lk} -8\rho_{hk}\rho_{lj} -4\tau R_{hljk}\big)g_{im}\\
&=\Big(-8R_{habj}R_{kabl} +8R_{habk}R_{jabl} +4R_{abhl}R_{abjk} +\dfrac{4}{3}\tau R_{hljk} +\dfrac{2}{9}\tau^2 g_{hj}g_{lk} - \dfrac{2}{9}\tau^2 g_{hk}g_{lj}\Big)g_{im}
\end{align*}
\item \begin{align*}
&\big(-8R_{habk}R_{mabl} +8R_{habm}R_{kabl} +4R_{abhl}R_{abkm} -8R_{ahkm}\rho_{al} +8R_{amhl}\rho_{ak} -8R_{akhl}\rho_{am} \\
&\quad+8R_{alkm}\rho_{ah} +8\rho_{hk}\rho_{lm} -8\rho_{hm}\rho_{lk} -4\tau R_{hlkm}\big)g_{ij}\\
&=\Big(-8R_{habk}R_{mabl} +8R_{habm}R_{kabl} +4R_{abhl}R_{abkm} +\dfrac{4}{3}\tau R_{hlkm} +\dfrac{2}{9}\tau^2 g_{hk}g_{lm} - \dfrac{2}{9}\tau^2 g_{hm}g_{lk}\Big)g_{ij}
\end{align*}
\smallskip
\item \begin{align*}
&8\Big(R_{ahjk}R_{amil}-R_{aljk}R_{amih}-R_{aijk}R_{amhl}+R_{aijm}R_{akhl}-R_{ahjm}R_{akil}+R_{aljm}R_{akih}\\
&\;-R_{aikm}R_{ajhl}+8R_{ahkm}R_{ajil}-R_{alkm}R_{ajih} -R_{ilkm}\rho_{hj}+R_{iljm}\rho_{hk}+R_{hljk}\rho_{im}\\
&\;+R_{ihjk}\rho_{lm}+R_{ihkm}\rho_{lj}-R_{ihjm}\rho_{lk}-R_{iljk}\rho_{hm}+R_{hlkm}\rho_{ij}-R_{hljm}\rho_{ik}\Big)\\
&=8\big(R_{ahjk}R_{amil}-R_{aljk}R_{amih}-R_{aijk}R_{amhl}+R_{aijm}R_{akhl}-R_{ahjm}R_{akil}+R_{aljm}R_{akih}\\
&\;-R_{aikm}R_{ajhl}+R_{ahkm}R_{ajil}-R_{alkm}R_{ajih}\Big)-\dfrac{4}{3}\tau \Big(R_{ilkm}g_{hj}- R_{iljm}g_{hk}- R_{hljk}g_{im}\\
&\;- R_{ihjk}g_{lm}- R_{ihkm}g_{lj}+ R_{ihjm}g_{lk}+ R_{iljk}g_{hm}- R_{hlkm}g_{ij}+ R_{hljm}g_{ik}\Big).
\end{align*}
\end{enumerate}}}
By rearranging all terms, we complete the proof of Lemma~\ref{thm:6-einstien(0,6)-tensor}.
\vskip 0.3cm
\section*{{Acknowledgements}}
The authors thank Prof. Nikolayevsky for several useful discussions.
|
\section{Introduction}
Throughout this note, all complex spaces are always assumed to be reduced and paracompact unless otherwise mentioned; we mainly refer to \cite{GR84, Richberg68} for basic references on the theory of complex spaces. For various terminologies and basic facts in algebraic geometry, we refer to the standard expositions \cite{Kollar97, Kollar_MMP, K-M98, La04}.
Let $X$ be an $n$-dimensional normal complex space with a canonical (Weil) divisor/class $K_X$ and $\Delta=\sum d_k\Delta_k$ a Weil $\mathbb{R}$-divisor on $X$, where $\Delta_k$ are distinct prime divisors. The pair $(X,\Delta)$ is called to be a \emph{log pair} if the $\mathbb{R}$-divisor $K_X+\Delta$ is an $\mathbb{R}$-Cartier divisor on $X$. Let $\pi:\widetilde X\to X$ be a log resolution of $(X,\Delta)$, i.e., a proper modification $\pi:\widetilde X\to X$ such that $\widetilde X$ is smooth, the exceptional locus $\text{Ex}(\pi)$ of $\pi$ is a divisor and $\pi_*^{-1}\Delta+\text{Ex}(\pi)$ has simple normal crossings (SNC). Write $K_{\widetilde X}=\pi^*(K_X+\Delta)+\sum_{i}a(E_i,X,\Delta)E_i$, where $a(E_i,X,\Delta)\in\mathbb{R}$ and $E_i\subseteq\widetilde X$ are distinct prime divisors. We say that a log pair $(X,\Delta)$ is \emph{KLT} (respectively, \emph{LC}) if $a(E_i,X,\Delta)>-1$ (respectively, $a(E_i,X,\Delta)\geq-1$) for all $i$. Similarly, a log pair $(X,\Delta)$ is called to be \emph{PLT} if $d_k\leq1$ and for all log resolutions $\pi:\widetilde X\to X$ of $(X,\Delta)$ we have $a(E_i,X,\Delta)>-1$ for every exceptional divisor $E_i$.
Suppose that $(X,S+B)$ is a log pair, where $S\subseteq X$ is a reduced Weil divisor (complex subspace of pure codimension one) with $\xi:\widehat S\to S$ the normalization of $S$, and $B\subseteq X$ is an $\mathbb{R}$-divisor on $X$ which has no irreducible components in common with $S$. Then, there exists a unique $\mathbb{R}$-divisor $\text{Diff}_{\widehat S}(B)$ (called the \emph{different} of $B$) on $\widehat S$ such that $(K_X+S+B)|_{\widehat S}=K_{\widehat S}+\text{Diff}_{\widehat S}(B)$. In addition, if $\widetilde S=\pi_*^{-1}S$ is the strict transform (may be disconnected) of $S$ on $\widetilde X$ and $\mu:\widetilde S\to\widehat S$ is the induced morphism, then $\text{Diff}_{\widehat S}(B)=\mu_*\big((\Delta_{\widetilde X}-\widetilde S)|_{\widetilde S}\big)$, where $\Delta_{\widetilde X}=\pi^*(K_X+S+B)-K_{\widetilde X}$; see \cite{Kollar92, Kollar_MMP} for more details.
\subsection{Multiplier and adjoint ideal sheaves for log pairs}
Let $(X,S+B)$ be a log pair as above and $\varphi\in L_\text{loc}^1(X_\text{reg})$ be a weight function on $X$ such that $\varphi|_S\not\equiv-\infty$ on every irreducible component of $S$. In the present note, by introducing an Ohsawa-type measure on $S$ via adapted volume forms for the pair $(X,S+B)$, we define the Nadel-Ohsawa multiplier ideal sheaves on $S$ associated to the weight $\varphi$ for the pair $(X,S+B)$; see Definition \ref{MIS_NO} for more details. Relying on the notion of Nadel-Ohsawa multiplier ideal sheaf and an $L^2$ extension theorem of measure version, we are able to obtain a reasonable generalization of adjoint ideal sheaves to the analytic setting for log pairs and establish the adjunction exact sequence as follows:
\begin{theorem} \label{Aadjoint}
With the same notations as above, $B$ an effective $\mathbb{R}$-divisor and $\varphi\in\emph{QPsh}(X)$ a quasi-psh function on $X$. Then, there exists an ideal sheaf $$\emph{Adj}_S(X,S+B;\varphi)\subseteq\mathcal{O}_X,$$
called the \emph{analytic adjoint ideal sheaf} associated to the triple $(X,S+B;\varphi)$ along $S$, sitting in an exact sequence:
$$0\longrightarrow\mathscr{I}(X,S+B;\varphi)\stackrel{\iota}{\longrightarrow}\emph{Adj}_S(X,S+B;\varphi)\stackrel{\rho}{\longrightarrow} i_*\mathscr{I}_\emph{NO}(\varphi|_S)\longrightarrow0, \eqno(\star)$$
where $\mathscr{I}(X,S+B;\varphi)$ is the usual multiplier ideal sheaf associated to the weight $\varphi$ for the pair $(X,S+B)$ on $X$, and $i:S\hookrightarrow X,\ \iota$ together with $\rho$ are the natural inclusion and restriction map respectively.
\end{theorem}
\begin{remark}
$(1)$ By the construction of analytic adjoint ideal sheaf $\text{Adj}_S(X,S+B;\varphi)$, the surjectivity of $\rho$ relies heavily on the $L^2$ extension theorem and strong openness of multiplier ideals (cf. the proof of Theorem \ref{Aadjoint}). Moreover, we can deduce the coherence of $\text{Adj}_S(X,S+B;\varphi)$ from the adjunction exact sequence $(\star)$.
$(2)$ If $X$ is smooth, the above result has been established in \cite{Li_multiplier, Li_adjoint}; refer to \cite{Taka10, Eisen10} and \cite{Gue12, Kim15a, G-L_adjoint} for related topics.
$(3)$ Whenever $\varphi$ has analytic singularities, it follows that $\text{Adj}_S(X,S+B;\varphi)$ coincides with the algebraic adjoint ideal sheaf defined by Takagi and Eisenstein in \cite{Taka10, Eisen10}, which is independent of the choice of log resolution.
\end{remark}
\subsection{Inversion of adjunction}
In higher dimensional birational geometry, the adjunction/inversion of adjunction theorems have been important tools for studying singularities of pairs. Thanks to a crucial connectedness result (cf. \cite{Kollar92}, Theorem 17.4), the following inversion of adjunction was established by \cite{Shokurov92} in dimension 3 and by \cite{Kollar92} in general.
\begin{theorem} \emph{(\cite{Kollar92}, Theorem 17.6).} \label{Kollar_inversion}
Let $X$ be a normal complex space and $S\subseteq X$ an irreducible reduced Weil divisor. Let $B$ be an effective $\mathbb{Q}$-divisor on $X$ such that $K_X+S+B$ is $\mathbb{Q}$-Cartier. Then, $(X,S+B)$ is \emph{PLT} in a neighborhood of $S$ if and only if $(\widehat S,\emph{Diff}_{\widehat S}(B))$ is \emph{KLT}.
\end{theorem}
In a recent paper \cite{KimSeo21}, the authors present the following natural question posed by Koll\'ar whether we can find an analytic proof of Theorem \ref{Kollar_inversion} relying on an $L^2$ extension theorem. Concretely, we have
\begin{question} \emph{(\cite{KimSeo21}, Question B).} \label{Question_Kollar}
Does the inversion of adjunction Theorem \ref{Kollar_inversion} have an analytic proof using $L^2$ extension theorems?
\end{question}
As a corollary of Theorem \ref{Aadjoint}, we can deduce a generalized analytic inversion of adjunction as follows, by which we answer Question \ref{Question_Kollar} in the affirmative, i.e.,
\begin{theorem} \label{Answer_Kollar}
With the same hypotheses as before, then we have along $S$,
$$\emph{Adj}_S(X,S+B;\varphi)=\mathcal{O}_X\iff\mathscr{I}_\emph{NO}(\varphi|_S)=\mathcal{O}_S.\eqno{(\spadesuit)}$$
In particular, the solution to Question \ref{Question_Kollar} is positive.
\end{theorem}
As we will present, $(X,S+B)$ is PLT in a neighborhood of $S$ if and only if the analytic adjoint ideal sheaf $\text{Adj}_S(X,S+B;0)=\mathcal{O}_X$ along $S$, and $(\widehat S,\text{Diff}_{\widehat S}(B))$ is KLT if and only if the Nadel-Ohsawa multiplier ideal sheaf $\mathscr{I}_\text{NO}(0)=\mathcal{O}_S$. Therefore, in order to answer Question \ref{Question_Kollar}, it is equivalent to show that $\text{Adj}_S(X,S+B;0)=\mathcal{O}_X$ along $S$ if and only if $\mathscr{I}_\text{NO}(0)=\mathcal{O}_S$, which is an immediate consequence of $(\spadesuit)$ by taking a trivial weight $\varphi=0$.
\begin{remark}
When $X$ is smooth, Question \ref{Question_Kollar} has been answered implicitly by the adjunction exact sequence $(\star)$ established in \cite{Li_multiplier} or \cite{Li_adjoint}; see also Theorem 1.4 in \cite{KimSeo21} for an explicit presentation based on an analogous argument.
\end{remark}
\section{Preliminaries}
In this section, we will state some notions and useful results that will be used throughout this paper.
\begin{definition}
Let $X$ be a normal complex space and $\Delta=\sum d_k\Delta_k$ a Weil $\mathbb{R}$-divisor on $X$, where $\Delta_k$ are distinct prime divisors. The \emph{round-up} $\lceil\Delta\rceil$ and \emph{round-down} (or \emph{integral part}) $\lfloor\Delta\rfloor=[\Delta]$ of $\Delta$ are the integral divisors
$$\lceil\Delta\rceil=\sum\lceil d_k\rceil\Delta_k,\quad \lfloor\Delta\rfloor=\sum\lfloor d_k\rfloor\Delta_k,$$
where for any $c\in\mathbb{R}$ we denote the least integer $\geq c$ by $\lceil c\rceil$, and the largest integer $\leq c$ by $\lfloor c\rfloor=[c]$.
The \emph{fractional part} $\{\Delta\}$ and \emph{co-fractional part} $\langle\Delta\rangle$ of $\Delta$ are defined by $$\{\Delta\}=\Delta-\lfloor\Delta\rfloor,\quad \langle\Delta\rangle=\lceil\Delta\rceil-\Delta.$$
\end{definition}
\begin{definition} \label{MIS_Pair}
Let $(X,\Delta)$ be a log pair with $\varphi\in L_\text{loc}^1(X_\text{reg})$ . A positive definite \emph{adapted measure} $\upsilon$ on $X_\text{reg}$ for the pair $(X,\Delta)$ is defined as a positive measure locally of the form $\upsilon=e^{\phi}\cdot\Big((\sqrt{-1})^{\gamma n^2}\sigma\wedge\overline\sigma\Big)^{\frac{1}{\gamma}}$ on $X$, where $\phi$ is a bounded measurable function and $\sigma$ is a local generator of $\mathcal{O}_X\big(\gamma(K_X+\Delta)\big)$ on an open subset of $X$ for some real number $\gamma>0$ (cf. \cite{EGZ09}, Definition 6.5).
The \emph{multiplier ideal sheaf} associated to $\varphi$ on $X$ for the pair $(X,\Delta)$ is defined to be the $\mathcal{O}_X$-submudle $\mathscr{I}(X,\Delta;\varphi)\subseteq\mathscr{M}_X$ of germs of meromorphic functions $f\in\mathscr{M}_{X,x}$ such that $|f|^2e^{-2\varphi}$
is locally integrable at $x\in X$ with respect to an adapted measure $\upsilon$ defined as above. Here, $\varphi$ is regarded as the weight function and $\mathscr{I}(X,\Delta;\varphi)$ is independent of the choice of $\upsilon$; one can refer to \cite{La04} for an algebraic counterpart.
If $\varphi$ has analytic singularities on $X$, i.e., locally $\varphi=c\cdot\varphi_\mathfrak{a}:=c\cdot\log|\mathfrak{a}|+O(1)$ for some ideal sheaf $\mathfrak{a}\subseteq\mathcal{O}_X$ and $c\in\mathbb{R}$ on $X$, the multiplier ideal sheaf $\mathscr{I}(X,\Delta;\varphi)$ is also denoted by $\mathscr{I}(X,\Delta;c\cdot\varphi_\mathfrak{a})$ or $\mathscr{I}(X,\Delta;\mathfrak{a}^c)$. In the context, we associate every Weil $\mathbb{R}$-divisor $\Delta=\sum d_k\Delta_k$ with a natural weight $\varphi_\Delta$, which can be locally written as $\varphi_\Delta=\sum d_k\log|\mathscr{I}_{\Delta_k}|$.
\end{definition}
\begin{remark}
$(1)$ If $X$ is smooth and $\varphi\in\text{QPsh}(X)$, it follows that $\mathscr{I}(X,\Delta=0;\varphi)\subseteq\mathcal{O}_X$ is nothing but the multiplier ideal sheaf $\mathscr{I}(\varphi)$ on $X$ introduced by Nadel (cf. \cite{De10}).
$(2)$ If $\Delta$ is effective and $\varphi\in\text{QPsh}(X)$, then $\mathscr{I}(X,\Delta;\varphi)\subseteq\mathcal{O}_X$ is a coherent ideal sheaf.
$(3)$ If $\Delta$ is an $\mathbb{R}$-Cartier divisor on $X$, then we have $\mathscr{I}(X,\Delta;\varphi)=\mathscr{I}(X,0;\varphi+\varphi_\Delta)$.
\end{remark}
Similar to the smooth case, we have the following bimeromorphic transformation rule for pairs (see \cite{La04}, Proposition 9.3.62).
\begin{proposition}
Let $\pi:(\widetilde X,\Delta_{\widetilde X})\to(X,\Delta)$ be a bimeromorphic morphisms between log pairs such that $K_{\widetilde X}+\Delta_{\widetilde X}=\pi^*(K_X+\Delta)$ and $\pi_*\Delta_{\widetilde X}=\Delta$. Then, it follows that
$$\mathscr{I}(X,\Delta;\varphi)=\pi_*\left(\mathscr{I}(\widetilde X,\Delta_{\widetilde X};\varphi\circ\pi)\right).$$
Moreover, if $\widetilde X$ is smooth, we have
$$\mathscr{I}(X,\Delta;\varphi)=\pi_*\left(\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}\rceil)\otimes\mathscr{I}(\widetilde X,0;\varphi\circ\pi+\varphi_{\langle-\Delta_{\widetilde X}\rangle})\right).$$
\end{proposition}
\begin{remark}
Let $(X,\Delta)$ be a log pair and the weight $\varphi_A=c_1\log|\mathfrak{a}_1|-c_2\log|\mathfrak{a}_2|$ be a difference of two quasi-psh functions with analytic singularities on $X$. Let $\pi:\widetilde X\to X$ be a common log resolution of the pair $(X,\Delta)$ and $\mathfrak{a}_k\ (k=1,2)$ with $\mathfrak{a}_k\cdot\mathcal{O}_{\widetilde X}=\mathcal{O}_{\widetilde X}(-D_k)$ and $K_{\widetilde X}-\pi^*(K_X+\Delta)+D_1+D_2$ having SNC supports. Then, it follows that
$$\mathscr{I}(X,\Delta;\varphi_A)=\pi_*\mathcal{O}_{\widetilde X}\left(\lceil K_{\widetilde X}-\pi^*(K_X+\Delta)-c_1D_1+c_2D_2\rceil\right).$$
\end{remark}
In the context, we will adopt a relative version of Grauert-Riemenschneider vanishing theorem for the higher direct images stated as below.
\begin{theorem} \emph{(\cite{Matsu_morphism}, Corollary 1.5).} \label{G-R_vanishing}
Let $\pi:X\to Y$ be a surjective proper (locally) K\"ahler morphism from a complex manifold $X$ to a complex space $Y$, and $(L,e^{-2\varphi_L})$ be a (possibly singular) Hermitian line bundle on $X$ with semi-positive curvature. Then, the higher direct image sheaf $$R^q\pi_*\Big(\omega_X\otimes\mathcal{O}_X(L)\otimes\mathscr{I}(\varphi_L)\Big)=0,$$
for every $q>\dim X-\dim Y$.
\end{theorem}
For the sake of convenience, we state a special case of the Ohsawa--Takegoshi-type $L^2$ extension theorem obtained by Guan and Zhou (Theorem 2.2 in \cite{G-Z_optimal}) as follows.
\begin{theorem} \label{OT}
Let $\Omega\subseteq\mathbb{C}^{n+r}$ be a pseudoconvex domain and $Z\subseteq\Omega$ an (closed) complex subspace of pure codimension $r$ such that $\mathscr{I}_Z$ is globally generated by holomorphic functions $g_1,\dots,g_m\in\mathcal{O}(\Omega)$.
Then, there is a constant $C>0$ (depending only on $r$) such that for any $\varphi\in\text{\emph{Psh}}(\Omega)$ and any $f\in\mathcal{O}_Z(Z)$ with $\int_Z|f|^2e^{-2\varphi}dV_{X}[\Psi]<+\infty$, there exists a holomorphic extension $\tilde f$ of $f$ to $\Omega$ such that
$$\int_\Omega\frac{|\tilde f|^2}{|g|^{2r}(-\log|g|)^2}e^{-2\varphi}dV_\Omega\leq C\cdot\int_Z|f|^2e^{-2\varphi}dV_{X}[\Psi],$$
where $\Psi=r\log|g|^2=r\log(|g_1|^2+\cdots+|g_m|^2)$.
\end{theorem}
In \cite{B-V03, Wlo08} and \cite{Bravo13, Eisen10}, the following desingularization theorem is established, which is necessary for our construction of the Nadel-Ohsawa multiplier/adjoint ideal sheaves for the higher codimensional case.
\begin{theorem} \emph{(Strong factorizing desingularization).} \label{SFD}
Let $X$ be a complex space of pure dimension and $Z\subseteq X$ a (closed) complex subspace with defining sheaf of ideals $\mathscr{I}_Z$ and no irreducible components contained in $X_\emph{sing}$. Then there exists an embedded resolution of singularities $\pi:\widetilde X\to X$ of $Z$ such that
$$\pi^*(\mathscr{I}_Z)=\mathscr{I}_{\widetilde Z}\cdot\mathscr{I}_{R_Z},$$
where $\mathscr{I}_{\widetilde Z}$ is the defining sheaf of ideals of the strict transform $\widetilde Z$ of $Z$ in $\widetilde X$ which has simple normal crossings with the exceptional divisor $\emph{Ex}(\pi)$ of $\pi$, and $\mathscr{I}_{R_Z}$ is the sheaf of ideals of a simple normal crossing divisor $R_Z$ supported on $\emph{Ex}(\pi)$.
Furthermore, we can choose $\pi$ to be a log resolution of any $\mathbb{R}$-linear combination of subschemes on $X$ not containing any component of $Z$ in its support (cf. \cite{Eisen10}, Corollary 3.2).
\end{theorem}
\section{Nadel-Ohsawa multiplier ideal sheaves on divisors for log pairs}
In this section, we present an Ohsawa-type measure on divisors arising from the research of so-called Ohsawa-Takigoshi $L^2$ extension theorem, by which will construct one type of multiplier ideal sheaves on divisors for log pairs.
\subsection{Ohsawa-type measure}
In order to establish a general $L^2$ extension theorem, Ohsawa \cite{Ohsawa5} introduced a positive measure on regular part of the associated complex subspace. Afterwards, associated with the same measure, the authors established various general $L^2$ extension theorems in \cite{G-Z_optimal, De16, CDM17} and so on. In the following, we will introduce an Ohsawa-type measure on divisors for log pairs to study singularities of pairs associated to psh functions in the sense of multiplier and adjoint ideal sheaves.
Let $(X,S+B)$ be a log pair with an adapted measure $\upsilon$ on $X_\text{reg}$, where $S\subseteq X$ is a reduced Weil divisor and $B\subseteq X$ is an $\mathbb{R}$-divisor on $X$ which has no irreducible components in common with $S$. Then, for any polar function $\Psi:X\to[-\infty,+\infty)$ along $S$ (i.e., locally written as $\log|\mathscr{I}_S|^2+O(1)$), we can associate a positive measure $dV_{S}[\Psi]$ on $S_\text{reg}$ defined by
$$\int_{S_\text{reg}}fdV_{S}[\Psi]=\limsup_{t\to+\infty}\int_{X_\text{reg}}f\cdot\mathbbm{1}_{\{-t-1<\Psi<-t\}}d\upsilon$$
for any nonnegative continuous function $f$ with $\text{Supp}\,f\,{\subset\subset}\,X_\text{reg}\backslash S_\text{sing}$, where $\mathbbm{1}_{\{-t-1<\Psi<-t\}}$ denotes the characteristic function of the set $\{z\in X\,|-t-1<\Psi(z)<-t\}$.
\begin{remark} \label{computation}
Let $\pi:\widetilde X\to X$ be a log resolution of $(X,S+B)$ or a strong factorizing desingularization of $S$ by Theorem \ref{SFD} such that $\pi^*(\mathscr{I}_S)=\mathscr{I}_ {\widetilde S}\cdot\mathscr{I}_{R_S}$, where $\widetilde S$ is the strict transform (may be disconnected) of $S$ in $\widetilde X$ and $R_S$ is an effective divisor supported on the exceptional divisor $\text{Ex}(\pi)$ of $\pi$. Write $$K_{\widetilde X}+\Delta_{\widetilde X}=\pi^*(K_X+S+B)=\pi^*(K_X+B)+\widetilde S+R_S.$$
Following the same argument as Remark 2.1 in \cite{Li_multiplier}, one can check that the measure $dV_{S}[\Psi]$ is locally the direct image of measures defined upstairs by
$$f\longmapsto\int_{\widetilde S}\left|f\circ\pi|_{\widetilde S}\right|^2\cdot\left|\frac{\text{Jac}(\pi)}{h_{R_S}}\Big|_{\widetilde S}\right|^2dV_{\widetilde S},$$
up to multiplicative bounded factors; where $\big(\text{Jac}(\pi)\big)=K_{\widetilde X}-\pi^*(K_X+B)=-\Delta_{\widetilde X}+\widetilde S+R_S$ and $h_{R_S}$ is the (local) defining function of $R_S$.
\end{remark}
\subsection{Nadel-Ohsawa multiplier ideals on divisors}
The usual analytic multiplier ideals are constructed by the integrability with respect to the Lebesgue measure or an adapted measure through the associated pluricanonical forms (e.g., Definition \ref{MIS_Pair}). In the analytic setting, the main difficulty of extending the notion of multiplier ideals to any singular complex (sub-)space is to choose a suitable measure in the sense of integrability. Thanks to the Ohsawa-type measure on divisors defined above, we are able to construct one type of multiplier ideals on divisors for log pairs in the following.
\begin{definition} \label{MIS_NO}
Let $(X,S+B)$ a log pair with an adapted measure $\upsilon$ on $X_\text{reg}$ as before. Let $\Psi:X\to[-\infty,+\infty)$ be a polar function along $S$ and $\varphi\in L_\text{loc}^1(X_\text{reg})$ such that $\varphi|_S\not\equiv-\infty$ on every irreducible component of $S$.
Then, the \emph{Nadel-Ohsawa multiplier ideal sheaf} $\mathscr{I}_\text{NO}(\varphi|_S)$ on the divisor $S$ associated to weight $\varphi$ for the pair $(X,S+B)$ is defined to be the fractional ideal sheaf of germs of meromorphic functions $f\in\mathscr{M}_{S,x}$ such that $|f|^2e^{-2\varphi}$ is locally integrable at $x$ on $S$ with respect to the measure $dV_{S}[\Psi]$. One can check that $\mathscr{I}_\text{NO}(\varphi|_S)$ is independent of the choices of $\upsilon$ and $\Psi$.
\end{definition}
\begin{theorem} \label{property}
With the same notations as above and $\varphi\in\emph{QPsh}(X)$, then the Nadel-Ohsawa multiplier ideal sheaf $\mathscr{I}_\emph{NO}(\varphi|_S)\subseteq\mathscr{M}_{S}$ is a coherent fractional ideal sheaf and satisfies the strong openness property, i.e.,
$$\mathscr{I}_\emph{NO}(\varphi|_S)=\bigcup\limits_{\varepsilon>0}\mathscr{I}_\emph{NO}\big((1+\varepsilon)\varphi|_S\big).$$
\end{theorem}
\begin{proof}
Let $\pi:\widetilde X\to X$ be a log resolution of $(X,S+B)$ or a strong factorizing desingularization of $S$ by Theorem \ref{SFD} such that $\pi^*(\mathscr{I}_S)=\mathscr{I}_ {\widetilde S}\cdot\mathscr{I}_{R_S}$, where $\widetilde S$ is the strict transform (may be disconnected) of $X$ in $\widetilde X$ and $R_S$ is an effective divisor supported on the exceptional divisor $\text{Ex}(\pi)$ of $\pi$. Write $$K_{\widetilde X}+\Delta_{\widetilde X}=\pi^*(K_X+S+B)=\pi^*(K_X+B)+\widetilde S+R_S.$$
Then, by Remark \ref{computation} for any $f\in\mathscr{I}_\text{NO}(\varphi|_S)_x$ defined on a small enough neighborhood $U$ of $x$ in $X$, we have
\begin{equation*}
\begin{split}
+\infty>\int_U|f|^2e^{-2\varphi}dV_{X}[\Psi]=&\int_{\widetilde U=\pi|_{\widetilde S}^{-1}(U)}\left|f\circ\pi|_{\widetilde S}\right|^2\cdot e^{-2\varphi\circ\pi|_{\widetilde S}}\cdot\left|\frac{\text{Jac}(\pi)}{h_{R_S}}\Big|_{\widetilde S}\right|^2dV_{\widetilde S},
\end{split}
\end{equation*}
which implies that $$f\in\pi_*\left(\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}+\widetilde S\rceil)|_{\widetilde S}\otimes\mathscr{I}(\varphi\circ\pi|_{\widetilde S}+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle}|_{\widetilde S})\right)_x,$$
where $\big(\text{Jac}(\pi)\big)=K_{\widetilde X}-\pi^*(K_X+B)=-\Delta_{\widetilde X}+\widetilde S+R_S$ and $h_{R_S}$ is the (local) defining function of $R_S$. On the other hand,
$$\pi_*\left(\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}+\widetilde S\rceil)|_{\widetilde S}\otimes\mathscr{I}(\varphi\circ\pi|_{\widetilde S}+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle}|_{\widetilde S})\right)\subseteq\mathscr{M}_X$$
is an $\mathcal{O}_X$-submodule of germs of meromorphic functions $f\in\mathscr{M}_{S,x}$ such that $|f|^2e^{-2\varphi}$ is locally integrable at $x$ on $S$ with respect to the measure $dV_{S}[\Psi]$. Then, it follows that
$$\mathscr{I}_\text{NO}(\varphi|_S)=\pi_*\left(\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}+\widetilde S\rceil)|_{\widetilde S}\otimes\mathscr{I}(\varphi\circ\pi|_{\widetilde S}+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle}|_{\widetilde S})\right),\eqno{(\heartsuit)}$$
which is coherent by Grauert's direct image theorem and Nadel's coherence theorem.
In addition, the strong openness of Nadel-Ohsawa multiplier ideals immediately follows from $(\heartsuit)$ and the result established by Guan and Zhou in \cite{G-Z_open} (see also \cite{Pham14}).
\end{proof}
\begin{remark}
$(1)$ When both of $X$ and $S$ are smooth, the Nadel-Ohsawa multiplier ideal sheaf $\mathscr{I}_\text{NO}(\varphi|_S)$ above is nothing but the usual multiplier ideal sheaf $\mathscr{I}(\varphi|_S+\varphi_B|_S)$ on $S$ introduced by Nadel.
$(2)$ In addition, if $\varphi_A$ is a quasi-psh function on $X$ possessing analytic singularities and $\varphi_A|_S\not\equiv-\infty$ on every irreducible component of $S$, by combining with an argument of log resolution, we can deduce that the fractional ideal sheaf $\mathscr{I}_\text{NO}\big((\varphi-\varphi_A)|_S\big)\subseteq\mathscr{M}_S$ is coherent.
\end{remark}
For the convenience of readers, we state the following result on Nadel-Ohsawa multiplier ideal sheaves established in \cite{Li_multiplier} for smooth ambient spaces (see also \cite{Li_adjoint} for the case of divisors), relying on the $L^2$ extension theorem and strong openness of multiplier ideal sheaves.
\begin{theorem} \emph{(\cite{Li_multiplier}, Theorem 1.4).} \label{Aadjoint_smooth}
Let $X$ be a complex manifold, $Z\subseteq X$ a complex subspace of pure codimension $r$ and $\varphi\in\emph{QPsh}(X)$ such that $\varphi|_Z\not\equiv-\infty$ on every irreducible component of $Z$.
Then, there exists an ideal sheaf $\emph{Adj}_Z(\varphi)\subseteq\mathcal{O}_X$ such that the following sequence of ideal sheaves is exact:
$$0\longrightarrow\mathscr{I}(\varphi+r\log|\mathscr{I}_Z|)\stackrel{\iota}{\longrightarrow}\emph{Adj}_Z(\varphi)\stackrel{\rho}{\longrightarrow} i_*\mathscr{I}_{\emph{NO}}(\varphi|_Z)\longrightarrow0.$$
\end{theorem}
\begin{remark} \label{Aadjoint_smoothRe}
In particular, when $r=1$ or $Z\subseteq X$ is smooth, the ideal sheaf $\text{Adj}_Z(\varphi)$ could be constructed as follows.
As the statement is local, we may assume that $X$ is a bounded Stein domain in some $\mathbb{C}^{N}$. Let $\mathscr{J}\subseteq\mathcal{O}_X$ be an ideal sheaf such that $\mathscr{J}|_Z=\mathscr{I}_\text{NO}(\varphi|_Z)\ $ (=$\mathscr{I}(\varphi|_Z)$ if $Z$ is smooth), which implies that $\mathscr{J}+\mathscr{I}_Z$ is independent of the choice of $\mathscr{J}$. Then, it follows that $$\text{Adj}_Z(\varphi):=\bigcup\limits_{\varepsilon>0}\text{Adj}_Z^0\Big((1+\varepsilon)\varphi\Big)\cap\Big(\mathscr{J}+\mathscr{I}_Z\Big),$$
where $\text{Adj}_Z^0(\varphi)\subseteq \mathcal{O}_X$ is an ideal sheaf of germs of holomorphic functions $f\in\mathcal{O}_{X,x}$ such that $$\frac{|f|^{2}e^{-2\varphi}}{|\mathscr{I}_Z|^{2r}\log^2|\mathscr{I}_Z|}$$ is locally integrable with respect to the Lebesgue measure near $x$ on $X$.
\end{remark}
\section{Proof of main results}
\subsection{Proof of Theroem \ref{Aadjoint}}
Let $\pi:\widetilde X\to X$ be a log resolution of $(X,S+B)$ or a strong factorizing desingularization of $S$ by Theorem \ref{SFD} such that $\pi^*(\mathscr{I}_S)=\mathscr{I}_ {\widetilde S}\cdot\mathscr{I}_{R_S}$, where $\widetilde S$ is the strict transform (may be disconnected) of $X$ in $\widetilde X$ and $R_S$ is an effective divisor supported on the exceptional divisor $\text{Ex}(\pi)$ of $\pi$. Write $$K_{\widetilde X}+\Delta_{\widetilde X}=\pi^*(K_X+S+B)=\pi^*(K_X+B)+\widetilde S+R_S.$$
Then, it follows from Theorem \ref{Aadjoint_smooth} that the following sequence of ideal sheaves is exact
\begin{equation*}
\begin{split}
0\longrightarrow\mathscr{I}(\varphi\circ\pi+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle})\otimes\mathcal{O}_{\widetilde X}(-\widetilde S)&\stackrel{\iota}{\longrightarrow}\text{Adj}_{\widetilde S}(\varphi\circ\pi+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle})\\
&\stackrel{\rho}{\longrightarrow} i_*\mathscr{I}(\varphi\circ\pi|_{\widetilde S}+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle}|_{\widetilde S})\longrightarrow0.
\end{split}
\end{equation*}
Here, the surjectivity of $\rho$ is due to the strong openness of multiplier ideals and an $L^2$ extension theorem, e.g., Theorem \ref{OT}.
Twist the exact sequence by $\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}+\widetilde S\rceil)$, and then we deduce that
\begin{equation*}
\begin{split}
0\to\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}\rceil)\otimes\mathscr{I}(\varphi\circ\pi+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle})&\to\text{Adj}_{\widetilde S}(\varphi\circ\pi+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle})\otimes\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}+\widetilde S\rceil)\\
&\to i_*\mathscr{I}(\varphi\circ\pi|_{\widetilde S}+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle}|_{\widetilde S})\otimes\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}+\widetilde S\rceil)\to0.
\end{split}
\end{equation*}
Taking
$$\text{Adj}_S(X,S+B;\varphi):=\pi_*\left(\text{Adj}_{\widetilde S}(\varphi\circ\pi+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle})\otimes\mathcal{O}_{\widetilde X}(\lceil-\Delta_{\widetilde X}+\widetilde S\rceil)\right)$$
and pushing forward the above exact sequence, we will infer the desired adjunction exact sequence
$$0\longrightarrow\mathscr{I}(X,S+B;\varphi)\stackrel{\iota}{\longrightarrow}\text{Adj}_S(X,S+B;\varphi)\stackrel{\rho}{\longrightarrow} i_*\mathscr{I}_\text{NO}(\varphi|_S)\longrightarrow0 \eqno(\star)$$
from the local vanishing (e.g., Theorem \ref{G-R_vanishing}) on the higher direct images of the term on the left; the proof is thereby concluded.
\hfill $\Box$
\subsection{Proof of Theorem \ref{Answer_Kollar}}
Note that $$\mathscr{I}(X,S+B;\varphi)\subseteq\mathscr{I}(X,S+B;0)\subseteq\mathscr{I}_S$$ by the definition. Then, by twisting the exact sequence $(\star)$ through by $\mathcal{O}_S$, we can deduce a restriction formula $\mathscr{I}_\text{NO}(\varphi|_S)=\text{Adj}_S(X,S+B;\varphi)\cdot\mathcal{O}_S$ from the right exactness of tensor functor. In particular, we have along $S$,
$$\text{Adj}_S(X,S+B;\varphi)=\mathcal{O}_X\iff\mathscr{I}_\text{NO}(\varphi|_S)=\mathcal{O}_S.\eqno{(\spadesuit)}$$
Let $\mathfrak{a}\subseteq\mathcal{O}_X$ be an ideal sheaf on $X$ with zeros not containing any irreducible component of $S$ and put $\varphi=c\cdot\varphi_{\mathfrak{a}}=c\cdot\log|\mathfrak{a}|$ for some $c\in\mathbb{R}_{\geq0}$. Let $\pi:\widetilde X\to X$ be a common log resolution of $(X,S+B)$ and $\mathfrak{a}$ such that $\pi^*(\mathscr{I}_S)=\mathscr{I}_ {\widetilde S}\cdot\mathscr{I}_{R_S}$ as before, and $\mathfrak{a}\cdot\mathcal{O}_{\widetilde X}=\mathcal{O}_{\widetilde X}(-F)$ for some effective divisor $F$ on $\widetilde X$. By a direct calculation of $\text{Adj}_{\widetilde S}(c\cdot\varphi_{\mathfrak{a}}\circ\pi+\varphi_{\langle-\Delta_{\widetilde X}+\widetilde S\rangle})$, we could obtain that $\text{Adj}_S(X,S+B;\mathfrak{a}^c)=\pi_*\mathcal{O}_{\widetilde X}\left(\lceil-\Delta_{\widetilde X}+\widetilde S-c\cdot F\rceil\right)$, which implies that $(X,S+B;\mathfrak{a}^c)$ is PLT in a neighborhood of $S$ if and only if the analytic adjoint ideal sheaf $\text{Adj}_S(X,S+B;\mathfrak{a}^c)=\mathcal{O}_X$ along $S$.
On the other hand, by the definition of different $\text{Diff}_{\widehat S}(B)$ it follows that $(\widehat S,\text{Diff}_{\widehat S}(B);\mathfrak{a}^c|_S)$ is KLT if and only if the divisor $\lceil-\Delta_{\widetilde X}+\widetilde S-c\cdot F\rceil\big|_{\widetilde S}$ is effective on $\widetilde S$, which is equivalent to triviality of the Nadel-Ohsawa multiplier ideal sheaf $\mathscr{I}_\text{NO}(\mathfrak{a}^c|_S)=\mathcal{O}_S$. Therefore, by taking $c=0$ it follows from $(\spadesuit)$ that $(X,S+B)$ is PLT in a neighborhood of $S$ if and only if $(\widehat S,\text{Diff}_{\widehat S}(B))$ is KLT.
\hfill $\Box$
|
\section{Introduction}
\label{sec:intro}
The system of nonlinear shallow water equations (SWEs) is a mathematical model for the fluid movement in
various shallow water environments, where the horizontal scales of motion are
much greater than the vertical scale. The SWEs have been widely
used to model flow in the river, near-shore ocean, and earth's atmosphere, etc.
In two dimensions, the inviscid SWEs take the following form:
\begin{subequations}
\label{swe}
\begin{alignat}{2}
\label{swe1}
h_t + \nabla\cdot (h\bld u)=&\;0,\\
\label{swe2}
(h\bld u)_t + \nabla\cdot (h\bld u\otimes \bld u)
+\frac12g \nabla (h^2) =&\;-g h\nabla b,
\end{alignat}
\end{subequations}
where $h$ is the water height, $\bld u = (u, v)$ is the velocity field,
$b(x,y)$ represents the bottom topography and $g$ is the gravitational
constant.
Below we review the four important properties that the SWEs satisfy, namely the
entropy condition, the {\it lake-at-rest} well-balanced property,
the
positivity of the water height $h$, and the conservation property.
The system \eqref{swe} is a system of {\it balance laws},
\begin{align}
\label{bal}
U_t + F(U)_x+G(U)_y = -s(x,y,U),
\end{align}
where $U=[h, hu, hv]^T$ is the vector of unknowns,
$F=[hu, hu^2+\frac12 gh^2, huv]^T$ and $G=[hv, huv, hv^2+\frac12 gh^2]^T$ are
flux vectors, and $s=[0, ghb_x, ghb_y]^T$ is the source vector.
It is well-known that solutions of the balance law \eqref{bal} can develop shock
discontinuities in a finite time, independent of whether the initial data is
smooth or not. Hence, the solution of \eqref{swe} are considered in the
weak sense \cite{D16}, which are in general not unique.
\textbf{(i) The entropy condition}.
To address the issue of non-uniqueness for weak solutions,
an additional admissibility criterion based on the so-called {\it entropy
condition} is imposed.
For the SWEs, the total energy
\[
E(U) :=\frac12h(u^2+v^2)+\frac12gh^2+ghb
\]
serves as an entropy function, which contains the kinetic energy
$\frac12h(u^2+v^2)$ and the gravitational potential energy
$\frac12gh^2+ghb$. If the bottom topography $b$ and
the solution of \eqref{swe} is smooth,
a straightforward calculation \cite{FMT11} reveals that
\begin{align}
\label{entro}
E(U)_t
+ \left(
\frac12(hu^3+huv^2)
+ghu(h+b)
\right)_x+\left(
\frac12(hu^2v+hv^3)
+ghv(h+b)
\right)_y = 0,
\end{align}
which is obtained by taking the inner product of
the SWEs \eqref{swe} with the
{\it entropy variable}
\begin{align}
\label{entro-v}
V:=\partial_U E = [g(h+b)-\frac12(u^2+v^2), u, v]^T
\end{align}
and applying the chain rule.
Since entropy should be dissipated across shock discontinuities,
the entropy conservation equation \eqref{entro} needs to be replaced by the following
entropy dissipation postulate
\begin{align}
\label{entroD}
E(U)_t
+ \left(
\frac12(hu^3+huv^2)
+ghu(h+b)
\right)_x+\left(
\frac12(hu^2v+hv^3)
+ghv(h+b)
\right)_y \le 0,
\end{align}
\textbf{(ii) Steady states and well-balanced property}.
Another important issue which arises in SWEs \eqref{swe} is the simulation of
their steady states, which are solutions that are independent of the time
variables. The most import example of a steady state solution to \eqref{swe}
is the so-called {\it lake at rest}, given by
\begin{align}
\label{wb}
u=v\equiv 0, \quad h+b \equiv Const.
\end{align}
Many interesting applications, such as
waves on a lake or tsunami waves in deep ocean,
involve computing perturbations of the lake at rest.
A numerical scheme which preserves a discrete version of the steady state
\eqref{wb}
is termed {\it well-balanced} with respect to the steady state.
\textbf{ (iii) Positivity of the water height}.
The water height $h$ in the SWEs needs to remain positive (non-negative)
for the system \eqref{swe} to remain well-posed.
If the water height becomes negative, the system \eqref{swe} will be
non-hyperbolic and non-physical, and the problem will be ill-posed.
\textbf{(iv) The conservation property}.
For constant topography $b\equiv Const$, the balance law \eqref{bal}
reduces to a hyperbolic system of conservation laws
\[
U_t + F(U)_x + G(U)_y=0.
\]
Integrating the above equation over any control volume $D\subset \mathbb{R}^2$
and applying the Gauss law, there holds the following conservation property:
\[
\frac{d}{dt}\int_{D}U\,\mathrm{dx} = -\int_{\partial D}
[F(U), G(U)]^T\cdot \bld n\, \mathrm{ds},
\]
where $\bld n$ is the outward unit normal direction on the boundary
$\partial D$.
The main focus of this work is to construct high-order numerical schemes for the
SWEs \eqref{swe} on unstructured meshes that respect the above mentioned four
properties.
All these properties are important in practice:
\begin{itemize}
\item
The Lax-Wendroff theorem \cite{LW60} ensures that if
a conservative scheme produces a discrete solution
$U_h(x,t)$ that uniformly converges to $\bar{U}(x,t)$, then ${\bar{U}}(x,t)$
is a weak solution to the continuous equation. Non-conservative schemes may converge to wrong solutions.
\item
Many shallow water applications involve rapidly moving interfaces between wet
and dry areas, where no water is present.
If no special attention is paid to maintain the positivity (non-negativity)
of the water height,
standard numerical methods may produce unacceptable negative water height
near the dry/wet front, which crashes of the numerical
simulation.
\item Well-balanced schemes are essential for computing perturbations of steady
states.
\item Entropy stability \eqref{entroD} provides additional stabilization
mechanism to the
scheme which further enhance its robustness.
\end{itemize}
Various numerical schemes satisfying (part of) these properties
for hyperbolic conservation laws or balance laws have been proposed in the
literature. We refer to the review articles \cite{XS14, Xing17} for a survey of
numerical schemes for the SWEs, in particular high-order well-balanced and
positivity-preserving schemes; to the review articles \cite{ZS11, XZ17}
for a survey of maximum-principle-satisfying and positivity-preserving
high-order schemes for conservation laws; and to the review article
\cite{Tadmor16} for entropy stable schemes.
Of particular relevance to the current work is the class of
entropy stable schemes for the SWEs, which respect the entropy dissipation
postulate
\eqref{entroD}. First-order entropy stable finite volume (FV) schemes for the SWEs were
proposed in \cite{TadmorZhong08,FMT09, FMT11} where the key concepts of entropy
variable, entropy conservative/stable numerical fluxes were discussed.
Similar entropy conservative/stable numerical fluxes were adopted in the
high-order {\it nodal} DG literature, which, in combination with
discrete derivative operators
using Gauss-Lobatto quadrature points that
satisfy the summation-by-parts (SBP) property,
yield entropy conservative/stable DG discretizations
\cite{GWK16, WWAGK17, WWAGW18}.
High-order entropy stable DG schemes were more recently extended to {\it modal}
formulations \cite{WDGX20, WKC21}, following the work of Chan \cite{Chan18}.
All these works are also well-balanced and conservative.
These entropy stable spatial discretizations were then combined with
{\it explicit} strong
stability preserving Runge-Kutta (SSP-RK) time integrators \cite{GST01}
to yield fully discrete conservative and well-balanced schemes.
We note that, in all these works, the entropy stability was proven in the
semi-discrete level where only spatial discretization was involved, which
does not hold theoretically for the explicit fully discretizations.
We further note that
the schemes \cite{WWAGW18, WDGX20} can preserve the positivity of the
water height with the aid of a positivity preserving scaling limiter \cite{ZS11}.
Moreover, most of the above cited works use structured/rectangular meshes,
with the exceptions of \cite{WWAGW18} which uses unstructured quadrilateral
meshes and \cite{WDGX20} which works on unstructured triangular meshes.
In this work, we construct high-order locally conservative, positivity
preserving, well-balanced, and entropy stable DG schemes for the SWEs on general
triangular meshes. Our approach is very different from the above cited entropy
stable DG schemes which relies on the SBP property of the underling difference
operators.
In our semi-discrete scheme,
instead of directly approximating the conservative variables,
we use the water height and velocity as the solution unknowns.
As a result, entropy stability is achieved naturally within the
weak formulation without the need to convert to the
strong form or work with difference
operators/matrices.
Our approach is more closely related to the entropy variable approach
to entropy stable schemes proposed by Tadmor \cite{Tadmor86} back in 1986, as
the velocity is part of the entropy variable \eqref{entro-v}.
For the purpose of efficient explicit time integration, we still keep
the water height as the solution unknown. As a result, we need to use the skew-symmetric formulation of the momentum equation
\cite{GWK16} to guarantee entropy stability of the semi-discrete scheme.
Similar to the works \cite{WDGX20, WKC21}, our proposed scheme achieves
entropy stability regardless of the choice of underlying numerical integration rules,
although the proofs are very different.
Actually, all integrals in our scheme involve polynomials only, which can be easily computed exactly if one wishes.
The proposed DG semi-discretization is then carefully combined with a classical SSP-RK
time integrator \cite{GST01}, in combination with a positivity-preserving
scaling limiter to ensure positivity of the water height.
Here special attention is paid to the Runge-Kutta inner stage reconstructions to
maintain local conservation of the fully discrete scheme.
To do so, we re-introduce the discharge (momentum)
as an auxiliary solution unknown and
reconstruct inner stage values based on the
the conservative variables, i.e.,
water height and discharge. We prove that this water height-velocity-discharge three-field formulation
is mathematically equivalent to the water height-velocity two-field formulation
in the semi-discrete level.
The advantage of this three-field formulation over the two-field formulation is that
standard slope limiting procedures can now be applied on
the conservative variables (water height and discharge)
to suppress numerical
oscillations
near discontinuities
without violating the local conservation property.
Here we apply the characteristic-wise TVB limiter \cite{CS98} with the
Fu-Shu troubled cell indicator \cite{FS17} using the total height as the indicating function.
The last ingredient of our fully discrete scheme is a proper wetting/drying treatment for problems with (moving) dry areas. The above mentioned positivity preserving limiter and TVB limiter do not directly work on the velocity approximation. As a result, the scheme may produce arbitrarily large velocity approximations near dry cells where the water height is very small.
Without any special treatment near these regions, the large velocity near dry areas will dictate the time step size, and may even crash the code due to too large velocity values.
We looked into a couple of wetting/drying treatments in the literature, but didn't find a good one yet that works for our scheme. Hence we introduce a new wetting/drying treatment that at least works for our numerical examples; see details in Remark \ref{rk:dry} below.
The rest of the paper is organized as follows. In Section 2, we introduce
the reformation of SWEs \eqref{swe}, and used it
to design a conservative, well-balanced and entropy stable DG spatial discretization.
In Section 3,
we present the explicit temporal discretization, and then prove the
posivitity preservation property. The implementation of a characteristic-wise
TVD slope limiter with an efficient troubled cell indicator
is then discussed. We further remark on the proper wetting/drying treatment in the velocity calculation.
Numerical results in one- and two-dimensions are then reported
in Section 4. We draw concluding remarks in Section 5.
\section{Reformulation of SWEs and the DG semi-discretization}
In this section, we first reformulate the SWEs \eqref{swe} into an equivalent
skew-symmetric form, c.f. \cite{GWK16}, and then introduce the associated conservative, entropy stable and
well-balanced DG semi-discretization. Although using the same skew-symmetric
form, we emphasis that our entropy stable DG discretization is completely
different from the work \cite{GWK16}, where we use velocity as independent
solution unknowns.
\subsection{The skew-symmetric form of the SWEs}
Multiplying the mass conservation equation \eqref{swe1} by $\frac12\bld u$, and
subtract it from the momentum balance equation \eqref{swe2}, we get
the following equivalent form of the SWEs:
\begin{subequations}
\label{sweS}
\begin{align}
\label{sweS1}
h_t +\nabla\cdot(h\bld u)=&\;0,\\
\label{sweS2}
(h\bld u)_t+ \nabla\cdot (h\bld u\otimes \bld u)
+g h\nabla (h+b)
-\frac12 h_t\bld u-\frac12\nabla\cdot(h\bld u)\bld u
=&\;0.
\end{align}
\end{subequations}
Here \eqref{sweS2} is referred to as the skew-symmetric form of the momentum
balance equation \eqref{swe2}, c.f. \cite{GWK16}.
Multiplying \eqref{sweS1} with $g(h+b)$ and \eqref{sweS2} with $\bld u$ and
adding, we immediate get the entropy conservation equality \eqref{entro}.
This suggests to use finite elements to directly approximate the quantities
$g(h+b)$ and $\bld u$ in order to design a Galerkin method that respect the
entropy conservation property \eqref{entro}, which is the approach we take in this article.
In practice, we use a discontinuous finite element space to directly approximate
the water height
$h(x,t)$ and use the same finite element space to approximate the
bottom topography $b(x)$, so that $g(h+b)$ can be taken as a test function in the
Galerkin formulation.
\subsection{The conservative, entropy-stable, and well-balanced DG spatial
discretization}
Without loss of generality, we formulate the DG spatial discretization for
the SWEs \eqref{sweS} on a periodic domain $\Omega\subset \mathbb{R}^2$.
Other standard boundary conditions will be used in the numerical experiments.
Here we formulate the scheme on a general unstructured triangular mesh, while
noting that the proposed method works on any standard meshes.
To this end, let $\Omega_h:=\{K\}$ be a conforming triangular discretization of the
domain $\Omega$. Denote $\partial \Omega_h:=\{\partial K\}$ as the
collection of element boundaries with $\bld n_K$ the associated outward unit
normal direction. Let $\mathcal{E}_h:=\{F\}$ be the collection
of edges of the triangulation $\Omega_h$. For any polynomial degree $k\ge 0$, let
\begin{align}
\label{space}
V_h^k :=\{v\in L^2(\Omega):\quad v|_K\in\mathcal{P}_k(K), \quad \forall
K\in\Omega_h\},
\end{align}
where $\mathcal{P}_k(K)$ is the space of polynomials of degree at most $k$ on the
element $K$.
Furthermore, let $\bld V_h^k$ be the vectorial version of the space $V_h^k$.
Given an edge $F=K^+\cap K^-\in\mathcal{E}_h$ which is shared by two elements $K^+$
and $K^-$, we denote $\bld n$ as the unit normal direction on $F$ pointing
towards $K^-$, and denote
$\jmp{\phi}|_F :=\phi^+-\phi^-$ and
$\avg{\phi}|_F:=\frac12(\phi^++\phi^-)$ as the standard jump and average on $F$
for any function $\phi\in V_h^k$, where $\phi^\pm:=\phi|_{K^\pm}$.
The proposed DG spatial discretization of \eqref{sweS} on the periodic domain
$\Omega$ reads as follows:
find $(h_h,\bld u_h)\in V_h^k\times \bld V_h^k$ such that
\begin{subequations}
\label{dg}
\begin{align}
\label{dg1}
M_h\left((h_h)_t, e_h\right) + A_h((h_h, \bld u_h), e_h) &\;= 0,\\
\label{dg2}
\bld M_h\left((h_h\bld u_h)_t, \bld v_h\right)
+ B_h((h_h, \bld u_h), \bld v_h)
+C_h((h_h, \bld u_h), \bld v_h) \;\;&\;\nonumber\\
-
M_h\left((h_h)_t, \frac12\bld u_h\cdot\bld v_h\right) - A_h\left((h_h, \bld u_h), \frac12\bld u_h\cdot\bld v_h\right) &\;= 0,
\end{align}
for all $(e_h, \bld v_h)\in V_h^k\times \bld V_h^k$,
where
$b_h\in V_h^k$ is a proper approximation of the bottom topography, and
the associated operators are given below:
\begin{align}
\label{op-1}
M_h((h_h)_t, e_h):= & \sum_{K\in\Omega_h}\int_K(h_h)_te_h\,\mathrm{dx},\\
A_h((h_h, \bld u_h), e_h):= &
\sum_{K\in\Omega_h}\left(-\int_Kh_h\bld u_h\cdot\nabla e_h\,\mathrm{dx}+
\int_{\partial K}\widehat{h_h\bld u_h}\cdot \bld
n_Ke_h\,\mathrm{ds}\right),\nonumber\\
\label{op-2}
= &-\sum_{K\in\Omega_h}\int_Kh_h\bld u_h\cdot\nabla e_h\,\mathrm{dx}+
\sum_{F\in\mathcal{E}_h}\int_F\widehat{h_h\bld u_h}\cdot \bld
n\jmp{e_h}\,\mathrm{ds} ,\\
\label{op-5}
\bld M_h\Big((h_h\bld u_h)_t,\bld v_h\Big)
:= &\;
\sum_{K\in\Omega_h}\int_K(h_h\bld u_h)_t
\cdot\bld v_h\,\mathrm{dx},\\
B_h((h_h, \bld u_h), \bld v_h):= &
-\sum_{K\in\Omega_h}\int_K
h_h(\bld u_h\otimes\bld u_h):\nabla \bld v_h
\,\mathrm{dx}+
\sum_{F\in\mathcal{E}_h}
\int_F
\widehat{(h_h\bld u_h\bld u_h)}\bld n
\cdot \jmp{\bld v_h}\,\mathrm{ds}, \label{op-3}\\
C_h((h_h, \bld u_h), \bld v_h) := &\;
\sum_{K\in\Omega_h}\int_Kgh_h\nabla(h_h+b_h)\cdot\bld v_h
\,\mathrm{dx}
-
\sum_{F\in\mathcal{E}_h}\int_Fg\jmp{h_h+b_h}\avg{h_h\bld v_h}\cdot\bld n
\,\mathrm{dx},
\label{op-4}
\end{align}
where the numerical fluxes $\widehat{h_h\bld u_h}\cdot\bld n$
and $\widehat{(h_h\bld u_h\bld u_h)}\bld n$
in the operators \eqref{op-2} and
\eqref{op-3} are defined as follows:
\begin{align}
\label{flux}
\widehat{h_h\bld u_h}\cdot\bld n:= &
\avg{h_h\bld u_h}\cdot\bld n + \frac12\alpha_h \jmp{h_h+b_h},\\
\label{flux2}
\widehat{(h_h\bld u_h\bld u_h)}\bld n:= &
\avg{h_h\bld u_h}\cdot\bld n\,\avg{\bld u_h} +
\frac12\alpha_h \jmp{(h_h+b_h)\bld u_h},
\end{align}
with estimated maximum speed
\begin{align}
\label{speed}
\alpha_h|_F :=\max\left\{\sqrt{gh_h^+}+|\bld u_h^+\cdot \bld n|,
\sqrt{gh_h^-}+|\bld u_h^-\cdot \bld n|\right\}.
\end{align}
We will show below that these local Lax-Friedrichs type
numerical fluxes are entropy-stable.
We note that the above operators are very natural DG scretizations of
the corresponding PDE operators in \eqref{sweS}, in particular,
\begin{itemize}
\item The operators $M_h$ in \eqref{op-1} and $\bld M_h$ in \eqref{op-5} are the weak forms associated with
the time derivative term $h_t$ in \eqref{sweS1}, and
$
(h_h\bld u_h)_t $ in \eqref{sweS2}, respectively;
\item The operator $A_h$ in \eqref{op-2}
is the DG discretization of the convection term
$\nabla\cdot(h\bld u)$ in
\eqref{sweS1}, with the numerical flux \eqref{flux}, and
The operator $B_h$ in \eqref{op-3} is the DG discretization of the convection term $\nabla\cdot(h\bld u\otimes\bld u)$
in \eqref{sweS2}, with the numerical flux \eqref{flux2}.
Here the particular choice of the numerical fluxes \eqref{flux} and \eqref{flux2} is crucial for the {\it entropy stability} of the semi-discrete scheme \eqref{dg}. Similar numerical fluxes have been used in the literature, c.f. \cite{FMT11,GWK16};
\item The operators $M_h$ and $A_h$ in \eqref{dg2} are the operators associated with the
skew-symmetric terms $\frac12 h_t\bld u$ and $\frac12\nabla\cdot(h\bld u)\bld u$ in \eqref{sweS2}, respectively;
\item The operator $C_h$ in \eqref{op-4} is a DG discretization of
the
gravitational
term $gh\nabla(h+b)$ in \eqref{sweS2} using a central numerical flux.
To see this, we note that the DG discretization with central numerical flux
for this operator reads as follows:
\begin{align*}
- \sum_{K\in\Omega_h}\int_Kg(h_h+b_h)\nabla\cdot(h_h\bld v_h)
\,\mathrm{dx}
+
\sum_{F\in\mathcal{E}_h}\int_Fg\avg{h_h+b_h}\jmp{h_h\bld v_h}\cdot\bld n
\,\mathrm{dx},
\end{align*}
which is equivalent to $C_h$ by integration by parts.
We mention that the two gravitational terms in momentum balance \eqref{swe2}
are combined into a single non-conservative product, which is another key to
the entropy stability and well-balanced property of our scheme.
We note that such non-conservative product has been explored in the literature, e.g.,
\cite{RBV08}, to design
well-balanced DG schemes.
\end{itemize}
\end{subequations}
We conclude this section with the main properties of our proposed DG
discretization \eqref{dg}, namely, local conservation, entropy stability and
the well-balanced property.
\begin{theorem}
\label{thm1}
The semi-discrete scheme \eqref{dg} is
\begin{subequations}
\label{semi-p}
\begin{itemize}
\item locally conservative in the sense of the following equalities:
\begin{align}
\label{e1}
\frac{d}{dt} \int_K h_h\,\mathrm{dx} = &\; -\int_{\partial K}\widehat{h_h\bld u_h}\cdot\bld n_K\,\mathrm{ds}, \\
\label{e2}
\frac{d}{dt} \int_K h_h\bld u_h\,\mathrm{dx} = &\;
-\int_{\partial K} \widehat{(h_h\bld u_h\bld u_h)}\bld n_K\,\mathrm{ds}\\
&\;-
\int_{K} gh_h\nabla(b_h)\,\mathrm{dx}
+\int_{\partial K} \frac12gh_hb_h\bld n_K\,\mathrm{ds}\nonumber
\end{align}
\item entropy stable in the sense of the following equality:
\begin{align}
\label{e3}
\frac{d}{dt}E_h = &\;
-\sum_{F\in\mathcal{E}_h}\int_{F}\frac12\alpha_h
\left(g\jmp{h_h+b_h}^2+\avg{h_h+b_h}\jmp{\bld u_h}\cdot\jmp{\bld u_h}\right)\,\mathrm{ds}
\le 0 ,
\end{align}
where the discrete entropy (total energy) is
\[
E_h:= \sum_{K\in\Omega_h} \int_K(\frac12\bld h_h|\bld
u_h|^2+\frac12gh^2+ghb)\mathrm{dx},
\]
\item well-balanced in the sense that it preserve the lake-at-rest steady state:
if the initial condition satisfies
\begin{align}
\label{eX}
\bld u_h(0) = 0, \text{and}\quad h_h(0)+b_h = C,
\end{align}
the solution to the semi-discrete scheme \eqref{dg} satisfies
\begin{align}
\label{e4}
\bld u_h(t) = 0, \text{and}\quad h_h(t)+b_h = C.
\end{align}
\end{itemize}
\end{subequations}
\end{theorem}
\begin{proof}
Taking test function $e_h:=1$ on element $K$ and {\it zero} elsewhere in \eqref{dg1}, we get mass conservation in \eqref{e1}.
Denote $u_h$ and $v_h$ as the two components of the velocity approximation $\bld u_h$.
Taking test function $e_h:=\frac12u_h$ on element $K$ and {\it zero} elsewhere in \eqref{dg1}, and $\bld v_h:=(1,0)$ on element $K$ and {\it zero} elsewhere in \eqref{dg2}
and adding, we get the following:
\begin{align*}
\frac{d}{dt} \int_K h_hu_h\,\mathrm{dx} =
-\int_{\partial K} \widehat{(h_h\bld u_h\bld u_h)}n_x\,\mathrm{ds}-
\int_{K} gh_h\partial_x(h_h+b_h)\,\mathrm{dx}
+\int_{\partial K} \frac12gh_h(h_h+b_h) n_x\,\mathrm{ds},
\end{align*}
where $n_x$ is the first component of the normal direction $\bld n_K$.
Combining the above identity with the fact that
\begin{align*}
\int_{K} gh_h\partial_x(h_h)\,\mathrm{dx}
-\int_{\partial K} \frac12 gh_h^2 n_x\,\mathrm{ds}
=0,
\end{align*}
we get the first component of the momentum balance identity \eqref{e2}.
We can apply the same argument to obtain the second component of the momentum balance identity \eqref{e2}.
Taking test function $e_h=g(h_h+b_h)$ in \eqref{dg1} and
$\bld v_h=\bld u_h$ in \eqref{dg2} and adding, we get
\begin{align*}
&\; M_h((h_h)_t, g(h_h+b_h))
+\bld M_h((h_h\bld u_h)_t, \bld u_h)
-\frac12M_h((h_h)_t, \bld u_h\cdot\bld u_h)\\
=&\; \underbrace{-B_h((h_h, \bld u_h), \bld u_h)
+ A_h\Big((h_h, \bld u_h), \frac12\bld u_h\cdot\bld u_h\Big)}_{:=I_1} \underbrace{-C_h((h_h, \bld u_h), \bld u_h) -
A_h\Big((h_h, \bld u_h), g(h_h+b_h)\Big)}_{:=I_2}.
\end{align*}
Simplifying the above equality, we yield the entropy dissipation equality \eqref{e3}.
More specifically, it is easy to show that the left hand side of the above equality is
the entropy dissipation rate
$\frac{d}{dt}E_h$, and the first term in the above right hand side
\begin{align}
\label{i1}
I_1 = -\sum_{F\in\mathcal{E}_h}\int_{F}\frac12\alpha_h\avg{h_h+b_h}\jmp{\bld u_h}\cdot\jmp{\bld u_h}\,\mathrm{ds},
\end{align}
and the second term
\begin{align*}
I_2 = -\sum_{F\in\mathcal{E}_h}\int_{F}\frac12\alpha_hg\jmp{h_h+b_h}^2\,\mathrm{ds},
\end{align*}
Below we only give detailed proof of the identity for $I_1$.
We have
\begin{align*}
I_1 = &\; -B_h((h_h, \bld u_h), \bld u_h)
+ A_h\Big((h_h, \bld u_h), \frac12\bld u_h\cdot\bld u_h\Big)\\
= &\;
\underbrace{ \sum_{K\in\Omega_h}\int_K
\left( h_h(\bld u_h\otimes\bld u_h):\nabla \bld u_h
-\frac12 h_h\bld u_h\cdot \nabla (\bld u_h\cdot\bld u_h)
\right)\,\mathrm{dx}}_{\equiv 0} \\
&\;-\sum_{F\in\mathcal{E}_h}
\int_F
\left(\widehat{(h_h\bld u_h\bld u_h)}\bld n
\cdot \jmp{\bld u_h}
-\widehat{h_h\bld u_h}\cdot\bld n
\jmp{\frac12\bld u_h\cdot\bld u_h}
\right)
\,\mathrm{ds}.
\end{align*}
Now by the definition of the numerical fluxes in \eqref{flux}--\eqref{flux2} and the
simple fact that $\jmp{ab} = \avg{a}\jmp{b}+\jmp{a}\avg{b}$, we have
\[
\widehat{(h_h\bld u_h\bld u_h)}\bld n
\cdot \jmp{\bld u_h}
-\widehat{h_h\bld u_h}\cdot\bld n
\jmp{\frac12\bld u_h\cdot\bld u_h} = \frac12\alpha_h\avg{h_h+b_h}\jmp{\bld u_h}\cdot\jmp{\bld u_h},
\]
which proves the identity for $I_1$ in \eqref{i1}.
Finally, under the assumption \eqref{eX}, it is trivial to show that the spatial operators $A_h, B_h, C_h$ in \eqref{dg} all stays {\it zero}.
Hence, $(h_h)_t \equiv 0$ from equation \eqref{dg1}, and
$(h_h\bld u_h)_t = \frac12(h_h)_t\bld u_h \equiv 0$ from \eqref{dg2}.
This implies the well-balanced property \eqref{e4}.
\end{proof}
\begin{remark}[Comparison with other entropy stable DG schemes]
\label{rk:es}
Our first order scheme with polynomial degree $k=0$ is closely related to the first order finite volume entropy stable schemes \cite{FMT09, FMT11}, as both approaches use the concept of entropy conservative/stable fluxes, cf. \eqref{flux}--\eqref{flux2}. The work \cite{FMT09,FMT11} promote to use a Roe-type dissipation operator in the numerical flux, while our numerical dissipation is of the more dissipative Lax-Friedrichs type.
For our high-order DG scheme with $k\ge 1$, we were not able to find similar work in the literature. There are two main approaches to construct high-order entropy stable schemes for nonlinear conservation laws, both stem from Tadmor's pioneer work on entropy variables and entropy conservative/stable fluxes \cite{Tadmor87, Tadmor86}.
The first approach directly discretizes the conservation equations using the entropy variables \cite{Tadmor86}, see also \cite{Hughes86}. The major drawback of this approach is that explicit time stepping is usually not applicable to these schemes due to the highly nonlinear mapping between the entropy variables and the conservative variables.
Hence they are generally more expensive than other explicit schemes.
The second approach is based on the (quadrature-based) SBP operator concept, and has undergo a major development in the past few years, see, e.g., the entropy stable DG schemes \cite{GWK16, WWAGK17, WWAGW18,WDGX20, WKC21} for SWEs. These quadrature-based approaches may lead to accuracy loss, and they may be more cumbersome to implement on unstructured triangular meshes than classical DG schemes; see the more discussion in the recent review work \cite{CS20}.
Our scheme \eqref{dg} combines both advantages of the above mentioned approaches:
\begin{itemize}
\item the proof of entropy stability can be performed directly on the variational formation \eqref{dg} without converting to any matrix-vector form.
This is made possible due to the use of velocity approximation and
the skew-symmetrization of the momentum balance equation.
In particular, we have the entropy stability result \eqref{e3} for the DG scheme \eqref{dg} with any choice of numerical quadrature rule.
This is possible because the proof of \eqref{e3} does not rely on integration by parts as the operators have already been properly skew-symmetrized.
Hence, the quadrature rule can be chosen only for accuracy considerations. In our numerical experiments, we simply use Gauss quadrature rules that are exact for integrating polynomials of degree $2k$.
\item
the semi-discrete scheme \eqref{dg} can be discretized in time using classical explicit time stepping schemes. This is possible because we discretize the water height $h_h$ as the solution unknown.
\end{itemize}
\end{remark}
\section{Fully discrete scheme: local conservation, well-balanced property,
positivity preservation, and slope limiting}
In this section, we discrete the semi-discrete scheme \eqref{dg} in time using
explicit SSP-RK time integrators.
Special attention is paid to maintain the local conservation, well-balanceness,
and positivity preservation properties.
We also discuss the use of a characteristic-wise TVB slope limiter
\cite{CS98} in combination with the recent troubled-cell indicator proposed in
\cite{FS17} to improve its efficiency. The slope limiter, which suppress
numerical oscillations near shock discontinuities, is a crucial component
for the accuracy and robustness of the overall scheme for polynomial degree $k\ge 1$,
c.f. \cite{CS01}.
Moreover, we propose a simple wetting/drying treatment for the velocity calculation near dry cells where water height is small.
\subsection{A three-field reformulation of the semi-discrete scheme \eqref{dg}}
Here we introduce a three-field reformation of the semi-discrete DG scheme \eqref{dg} by using the discharge $\bld m:= h\bld u$ as an additional independent unknown, which is then discretized in time using the explicit SSP-RK method.
The three-field DG scheme reads as follows:
find $(h_h, \bld u_h, \bld m_h)\in V_h^k\times\bld V_h^k\times\bld V_h^k$ such that
\begin{subequations}
\label{dgr}
\begin{align}
\label{dgr1}
M_h\left((h_h)_t, e_h\right) + A_h((h_h, \bld u_h), e_h) = &\;0,\\
\label{dgr2}
\bld M_h\left((\bld m_h)_t, \bld v_h\right)
+ B_h((h_h, \bld u_h), \bld v_h)
+C_h((h_h, \bld u_h), \bld v_h) \;\;\quad&\\
-M_h\left((h_h)_t, \frac12\bld u_h\cdot\bld v_h\right) - A_h\left((h_h, \bld u_h), \frac12\bld u_h\cdot\bld v_h\right) = &\;0,\\
\label{dgr3}
\bld M_h\left(h_h\bld u_h-\bld m_h, \bld w_h\right) = &\;0,
\end{align}
\end{subequations}
for all $(e_h, \bld v_h,\bld w_h)\in V_h^k\times\bld V_h^k\times \bld V_h^k$.
We have the following equivalence of the two formulations \eqref{dg} and \eqref{dgr}.
\begin{theorem}
Let $(h_h, \bld u_h, \bld m_h)$ be the solution to the three-field DG formulation \eqref{dgr}.
Then $(h_h, \bld u_h)$ is the solution to the two-field DG formulation \eqref{dg}.
\end{theorem}
\begin{proof}
Taking the time derivative of equation \eqref{dgr3}, we can replace the auxiliary variable $\bld m_h$ in \eqref{dgr2} by $h_h\bld u_h$. This means the solution $(h_h,\bld u_h)$ solves the system \eqref{dg}.
\end{proof}
The advantage of this reformulation will be clear next when we discuss SSP-RK time discretizations and slope limiting.
\subsection{High order SSP-RK discretization and
inner stage reconstruction}
The semi-discrete scheme \eqref{dgr} is not a standard ODE system
$U_t + \mathcal{F}(U) = 0$, with $U$ being the solution vector and $\mathcal{F}(U)$ the spatial operator, as the time derivative terms in \eqref{dgr2} involve the nonlinear product $h_h\bld u_h$, and \eqref{dgr3} is an algebraic equation. As a result, special care is need in design locally conservative high-order time discretizations. Here we apply the third order SSP-RK3 scheme to \eqref{dgr} which preserves the local conservation property. It is built on top of a plain forward Euler discretization in Algorithm \ref{alg1},
a velocity update in Algorithm \ref{algV}, and
a convex combination step in Algorithm \ref{alg2}. The full plain SSP-RK3 algorithm without slope limiting is given in Algorithm \ref{alg3}.
\newcommand{\ALOOP}[1]{\ALC@it\algorithmicloop\ #1%
\begin{ALC@loop}}
\newcommand{\end{ALC@loop}\ALC@it\algorithmicendloop}{\end{ALC@loop}\ALC@it\algorithmicendloop}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\newcommand{\textbf{break}}{\textbf{break}}
\newcommand{\STATE \algorithmicbreak}{\STATE \textbf{break}}
\begin{algorithm}
\caption{Plain Forward Euler + DG}
\begin{algorithmic}[1]
\REQUIRE $h_h^{old}\in V_h^k$, $\bld u_h^{old}, \bld m_h^{old}\in \bld V_h^k$, and time step size $\Delta t>0$.
\ENSURE $h_h^{new}\in V_h^k$, and $\bld m_h^{new}\in \bld V_h^k$.
\STATE Compute $h_h^{new}$ by the following equation:
\[
M_h(h_h^{new}, e_h) = M_h(h_h^{old}, e_h) - \Delta tA_h((h_h^{old}, \bld u_h^{old}), e_h),\quad \forall
e_h\in V_h^k,
\]
\STATE Compute $\bld m_h^{new}$ by the following equation:
\begin{align*}
\bld M_h(\bld m_h^{new}, \bld v_h) =&\; \bld M_h(\bld m_h^{old}, \bld v_h) - \Delta tB_h((h_h^{old}, \bld u_h^{old}), \bld v_h)
- \Delta tC_h((h_h^{old}, \bld u_h^{old}), \bld v_h)\\
&\hspace{-1cm} +\frac12M_h\left(h_h^{new}-h_h^{old}, \bld u_h^{old}\cdot\bld v_h\right) +\frac12\Delta tA_h\left((h_h^{old}, \bld u_h^{old}), \bld u_h^{old}\cdot\bld v_h\right),\quad \forall
\bld v_h\in \bld V_h^k,
\end{align*}
\end{algorithmic}
\label{alg1}
\end{algorithm}
\begin{algorithm}
\caption{Velocity update}
\begin{algorithmic}[1]
\REQUIRE $h_h\in V_h^k$, $\bld m_h\in \bld V_h^k$.
\ENSURE $\bld u_h\in \bld V_h^k$.
\STATE Compute $\bld u_h$ by the following equation:
\[
\bld M_h(h_h\bld u_h, \bld w_h) =\bld M_h(\bld m_h, \bld w_h),\quad \forall
\bld w_h\in \bld V_h^k.
\]
\end{algorithmic}
\label{algV}
\end{algorithm}
\begin{algorithm}
\caption{Convex combination}
\begin{algorithmic}[1]
\REQUIRE Data $h_h^{1}, h_h^{2}\in V_h^k$, $\bld m_h^{1}, \bld m_h^{2}\in \bld V_h^k$. Positive weights $w_1, w_2$ with $w_1+w_2=1$.
\ENSURE $h_h\in V_h^k$, and $\bld m_h\in \bld V_h^k$.
\STATE Compute $h_h$ and $\bld m_h$ using convex combination:
\[
h_h\leftarrow w_1h_h^1+w_2h_h^2, \quad
\bld m_h\leftarrow w_1\bld m_h^1+w_2\bld m_h^2.
\]
\end{algorithmic}
\label{alg2}
\end{algorithm}
\begin{algorithm}
\caption{Plain SSP-RK3 + DG}
\begin{algorithmic}[1]
\REQUIRE $h_h^{n}\in V_h^k$, $\bld u_h^{n}, \bld m_h^{n}\in \bld V_h^k$ at time level $t^n$, and time step size $\Delta t>0$.
\ENSURE $h_h^{n+1}\in V_h^k$, and $\bld u_h^{n+1}, \bld m_h^{n+1}\in \bld V_h^k$ at next time level $t^{n+1}:=t^n+\Delta t$.
\STATE Apply Algorithm \ref{alg1} with inputs $h_h^n, \bld u_h^n, \bld m_h^n$ and $\Delta t$.
Denote the outputs as $h_h^{(1)}, \bld m_h^{(1)}$.
\STATE Apply Algorithm \ref{algV} with inputs $h_h^{(1)}, \bld m_h^{(1)}$.
Denote the velocity output as $\bld u_h^{(1)}$.
\STATE Apply Algorithm \ref{alg1} with inputs $h_h^{(1)}, \bld u_h^{(1)}, \bld m_h^{(1)}$ and $\Delta t$. Denote outputs as $h_h^{(2*)}, \bld m_h^{(2*)}$.
\STATE
Apply Algorithm \ref{alg2} with inputs $h_h^{n}, h_h^{(2*)}, \bld m_h^{n}, \bld m_h^{(2*)}$ and weights $w_1=0.75, w_2 = 0.25$.
Denote outputs as $h_h^{(2)}, \bld m_h^{(2)}$.
\STATE Apply Algorithm \ref{algV} with inputs $h_h^{(2)}, \bld m_h^{(2)}$.
Denote the velocity output as $\bld u_h^{(2)}$.
\STATE Apply Algorithm \ref{alg1} with inputs $h_h^{(2)}, \bld u_h^{(2)}, \bld m_h^{(2)}$ and $\Delta t$. Denote outputs as $h_h^{(3*)}, \bld m_h^{(3*)}$.
\STATE
Apply Algorithm \ref{alg2} with inputs $h_h^{n}, h_h^{(3*)}, \bld m_h^{n}, \bld m_h^{(3*)}$ and weights $w_1=1/3, w_2 = 2/3$.
Denote outputs as $h_h^{n+1}, \bld m_h^{n+1}$.
\STATE Apply Algorithm \ref{algV} with inputs $h_h^{n+1}, \bld m_h^{n+1}$.
Denote the velocity output as $\bld u_h^{n+1}$.
\end{algorithmic}
\label{alg3}
\end{algorithm}
\begin{remark}[Local conservation and well-balanced property]
Similar to the proof of local conservation in Theorem \ref{thm1}, the
forward Euler algorithm Algorithm \ref{alg1} is also locally conservative.
Meanwhile, the local conservation property is not polluted by the convex combination step in Algorithm \ref{alg2} due to the use of conservative variables in convex combination.
If the velocity $\bld u_h$ were to be used in the convex combination step, local conservation for the discharge would be lost.
This is the major reason that the discharge $\bld m_h$ is re-introduced as an independent unknown in the DG formulation. Hence the overall algorithm
Algorithm \ref{alg3} is locally conservative.
Moreover, it is easy to see that Algorithm \ref{alg3}
preserve the steady state solution \eqref{wb}, hence it is also well-balanced.
\end{remark}
\begin{remark}[Computational cost and positivity of water height]
The computational cost of Algorithm \ref{alg1} involves the evaluation of the right hand side operators, and solving the associated linear system for the mass matrix for $V_h^k$ in Step 1,
and for $\bld V_h^k$ in Step 2.
The mass matrices can be made diagonal if orthogonal $L^2$-basis is used in the computation.
The computational cost of Algorithm \ref{algV} involves the linear system solve of a water height-weighted mass matrix, which is block diagonal and can be computed very efficiently.
The Algorithm \ref{alg2} is simply a vector update.
Hence, the computational cost in Algorithm \ref{alg3} is of linear complexity with respect to the total number of elements, which is similar to, but slightly more expensive than (due to the velocity computation in Algorithm \ref{algV}), the cost of a classical DG scheme with SSP-RK3 time stepping.
Note that in Algorithm \ref{algV}, we need to invert the water height-weighted mass matrix to compute the velocity approximation $\bld u_h$.
This weighted mass matrix may fail to be invertible if the water height approximation $h_h$ become negative in parts of the domain.
In practice, requiring positivity of water height on the volume integration points used to compute these mass matrices suffice to ensure its invertibility, which, however, is not guaranteed in the plain Algorithm \ref{alg3}.
In the next subsection, we apply the positivity-preserving limiting approach used in \cite{XZ13} to guarantee such positivity requirement.
\end{remark}
\subsection{Hydrostatic reconstruction and posivitivity-preserving limiter}
The key idea of the posivity-preserving limiter in \cite{XZ13} is to ensure the cell average of water height is positive after one step of forward Euler time stepping under a usual CFL condition.
Given solution $h_h^n, \bld u_h^n$ at time $t^n$, and time step size $\Delta t$, the water height $h_h^{n+1}$ at next time level for the forward Euler Algorithm \ref{alg1} reads as follows:
\begin{align*}
M_h(h_h^{n+1}, e_h) = M_h(h_h^{n}, e_h) - \Delta tA_h((h_h^{n}, \bld u_h^{n}), e_h),\quad \forall
e_h\in V_h^k.
\end{align*}
Taking test function $e_h=1$ on a single element $K\in\Omega_h$, we get the evolution equation for the water height cell average:
\begin{align}
\label{cell}
\bar h_K^{n+1} = \bar h_K^{n} - \frac{\Delta t}{|K|}\int_{\partial K}\widehat{h_h^n\bld u_h^n}\cdot\bld n_K\,\mathrm{ds}
\end{align}
where $\bar{h}_K^n$ stands for the cell average of $h_h$ on the triangle $K$ at time level $t^n$, and $|K|$ is the area of the element $K$.
Due to the fact that the numerical flux \eqref{flux} contains the jump of bottom topography $b_h$, which can arbitrarily large, we can not prove positivity of $\bar h_K^{n+1}$ in the above equation \eqref{cell} under the condition of positivity of $\bar h_K^n$ and a reasonable time step size restriction. To fix this, we slightly modify the numerical fluxes using the idea of hydrostatic reconstruction \cite{ABF04,XZ13}.
In particular, introducing the following hydrostatic reconstructed version of the water height:
\begin{subequations}
\label{hydro}
\begin{align}
h_h^{*,+} := \max\left\{0, h^+ +\min\{0, \jmp{b_h}\} \right\},\\
h_h^{*,-} := \max\left\{0, h^- -\max\{0, \jmp{b_h}\} \right\},
\end{align}
\end{subequations}
we replace $h_h^{\pm}$ in the flux terms in the scheme \eqref{dg}
by $h_h^{*,\pm}$, and replace the associated the jump term $\jmp{h_h+b_h}$ by $\jmp{h_h^{*}} =
h_h^{*,+}-h_h^{*,-}$.
For example, the flux \eqref{flux} is now replaced by the following one:
\begin{align}
\label{fluxH}
\widehat{h_h^*\bld u_h}\cdot\bld n:=\avg{h_h^*\bld u_h}\cdot\bld n
+\frac12\alpha_h^*\jmp{h_h^*},
\end{align}
with
\begin{align}
\label{speed2}
\alpha_h^*|_F :=\max\left\{\sqrt{gh_h^{*,+}}+|\bld u_h^+\cdot n|,
\sqrt{gh_h^{*,-}}+|\bld u_h^-\cdot n|\right\}.
\end{align}
It is clear that if $h_h$ satisfies $h_h+b_h=Const$ with $h_h>0$, then
\begin{align*}
h_h^{*,+} = h^+ +\min\{0, \jmp{b_h}\},\\
h_h^{*,-} = h^- -\max\{0, \jmp{b_h}\},
\end{align*}
and $\jmp{h_h^*} \equiv \jmp{h_h+b_h}$, which implies that the modified fluxes will not pollute the well-balanced property of the original fluxes.
With this modification, the forward Euler discretization lead to the following cell average evolution for water height:
\begin{align}
\label{cell2}
\bar h_K^{n+1} =&\; \bar h_K^{n} - \frac{\Delta t}{|K|}\int_{\partial K}\widehat{h_h^{*,n}\bld u_h^n}\cdot\bld n_K\,\mathrm{ds}\nonumber\\
=&\; \bar h_K^{n} - \frac{\Delta t}{|K|}\int_{\partial K}
\left(\avg{h_h^{*,n}\bld u_h^{n}}\cdot\bld n_K + \frac12\alpha_h^{*,n}\jmp{h_h^{*,n}}_K\right)
\,\mathrm{ds},
\end{align}
where $\jmp{h_h}_K:= h_h^{int(K)} - h_h^{ext(K)}$ is the jump, with
$h_h^{int(K)}$ and $h_h^{ext(K)}$ being the approximations
obtained from the interior and the exterior of $K$.
Note that by definition, on any edge $F=K^+\cap K^-$ shared by two elements, there holds
\begin{align*}
\jmp{h_h} = h_h|_{K+}-h_h|_{K^-} = \jmp{h_h}_{K^+} = -\jmp{h_h}_{K^-}.
\end{align*}
The cell average evolution equation \eqref{cell} now has a similar form as
\cite[Equation 3.1]{XZ13}. Hence, we can follow the same analysis in \cite[Section 3]{XZ13} to ensure positivity of the water height cell average at next time level.
The following result is Theorem 3.2 in \cite{XZ13}. The proof is almost identical, hence we omit it for simplicity.
\begin{theorem}[Theorem 3.2 in \cite{XZ13}]
\label{thm:pp}
For the scheme \eqref{cell2} to be positivity preserving, i.e., $\bar h_K^{n+1}\ge 0$, a sufficient condition is that $h_K(\bld x)\ge 0$,
$\forall \bld x\in S_K$ for all $K$, under the CFL condition
\begin{align}
\label{cfl}
\alpha \frac{\Delta t}{|K|}|\partial K| \le \frac23\widehat w_1.
\end{align}
Here
$h_K(\bld x)$ denotes the polynomial for water height at time level $n$,
$S_K$ is a set of (symmetric) quadrature points on $K$ that includes
$k+1$ Gauss quadrature points on each boundary edge,
$\alpha$ is the maximum estimated speed \eqref{speed2},
$|\partial K|$ is the perimeter of element $K$, and
$\widehat w_1$ is the quadrature weight of the $\lceil \frac{k+3}{2}\rceil$-point Gauss-Lobatto rule on
$[-1/2, 1/2]$ for the first quadrature point.
\end{theorem}
At time level $n$, given the water height DG polynomial $h_K(\bld x)$ with its cell average $\bar h_K^n\ge0$, we use the simple scaling limiter \cite[Section 3.4]{XZ13} to ensure the above sufficient condition $h_K(\bld x)\ge0$ for all
$\bld x\in S_K$, i.e., replacing
$h_K(\bld x)$ by a linear scaling around the cell average:
\begin{align}
\label{pplimit}
\widetilde{h}_K(\bld x) = \theta_K(h_K(\bld x)-\bar h_K^n)+\bar h_K^n,
\end{align}
where $\theta_K\in[0,1]$ is determined by
\begin{align}
\label{scale}
\theta_K := \min_{\bld x\in S_K}\theta_{\bld x},
\quad\theta_{\bld x} =\min\left\{1,
\frac{\bar h_K^n }{\bar h_K^n-h_K(\bld x)}
\right\}.
\end{align}
A slightly more efficient and less restrictive scaling parameter $\theta_K$ can be obtained using a reduced set of quadrature points, see \cite[Section 3.4]{XZ13} for more details.
In practice, the positivity preserving limiter \eqref{pplimit} is applied in each inner stage of the RKDG algorithm \ref{alg3}.
We notice that the well-balanced property is also not affected by this positivity preserving limiter.
\begin{remark}[On hydrostatic reconstruction and bottom topography approximation]
The proof of Theorem \ref{thm:pp} requires the use of hydrostatic reconstruction \eqref{hydro}, which is needs due to the lack of control of the bottom topography jump $\jmp{b_h}$ across edges.
When the polynomial degree $k\ge 1$ in the DG scheme \eqref{dg},
one can approximate the bottom topography using a continuous approximation
$b_h\in V_h^k\cap H^1(\Omega)$, which implies $\jmp{b_h}=0$.
In this case, under the positivity assumption of Theorem \ref{thm:pp}, we have
$h^{*,\pm}_h = h^{\pm}_h$, hence equivalence of the original scheme \eqref{cell} and the reconstructed version \eqref{cell2}.
For this reason, we prefer to use a continuous bottom topography approximation for $k\ge 1$, where the hydrostatic reconstruction \eqref{hydro} is not necessary anymore.
\end{remark}
\subsection{The troubled-cell indicator and slope limiter}
Another important ingredient of the DG methods is the slope
limiter procedure which is needed to suppress spurious oscillations near solution discontinuities.
We follow the standard slope limiting procedure in
RKDG methods \cite{CS01, QS05}:
\begin{itemize}
\item [(1)] First we identify the {\it troubled cells}, namely, those cells which might need the limiting procedure.
\item [(2)] Second we replace the solution polynomials in those troubled cells by reconstructed polynomials
with limited slopes that maintain the original
cell averages (conservation).
\end{itemize}
We use the Fu-Shu troubled-cell indicator proposed in \cite{FS17} to identify the troubled cells, with a scaling modification to improve its performance and computational efficiency.
Given a discontinuous function $p\in V_h^k$, the
troubled-cell indicator \cite{FS17} is given as follows:
\begin{align}
\label{indicator}
I_{K}(p) = \frac{\sum_{T\in \omega(K)}|\bar{\bar p}_T-\bar p_K|}{\bar p_{\max}-\bar p_{\min}},
\end{align}
where $\omega(K)$ is the union of cells that share a common edge with
$K$, including $K$ itself, and $\bar{\bar p}_T$ is the cell average of the polynomial $p|_T$ extended to the target cell $K$, and
$\bar p_{\max}$ and $\bar p_{\min}$ are the global maximal and minimal cell average on the domain.
Relying on the {\it assumption} \cite{CS98, CS01} that spurious oscillations are present in the solution $p_h$ only if they are present in its linear part $p_h^1$, which is its $L^2$-projection into the space of piecewise linear functions
$V_h^1$, we use use the linear $L^2$-projection of the total height $h_h+b_h$ as the indicating function in \eqref{indicator}, which simplifies the implementation of the extended cell average $\bar{\bar h}_h$ for high-order case where the polynomial degree $k>1$. The cell $K$ is marked as a troubled cell if
\begin{align}
\label{tol}
I_K(h_h^1+b_h^1) > tol,
\end{align}
where $tol$ is a user defined parameter.
Note that this indicator is of $\mathcal{O}(h^2)$ in smooth regions, and of $\mathcal{O}(1)$ near discontinuities, hence is expected to be effective to detect troubled cells near discontinuities with a proper choice of $tol$.
Our numerical experiments suggest that the indicator is not too sensitive to the tolerance $tol$. The indicator with $tol\in (0.01, 0.1)$ performs similarly for most of the examples, where $tol=0.01$ leads to a slightly larger number of detected troubled cells than $tol=0.1$ as expected. In our implementation, we take $tol=0.02$ for all the reported results.
\begin{remark}[On scaling of the indicator \eqref{indicator}]
The original indicator proposed in \cite{FS17} use the local maximal cell average $\max_{T\in \omega(K)}\{\bar p_T\}$ as the scaling denominator. This scaling has the drawback of not able to detect any troubled cells for small perturbation tests where the total height is a very small perturbation of a constant state. In particular, the original indicator with a local maximum scaling will produce a completely different result when the indicating function is perturbed by a global constant $p(x)\leftarrow p(x)+Const$.
The new global difference scaling denominator $\bar p_{\max}-\bar p_{\min}$ now produce the same indicating value when the indicating function is perturbed by a global constant. It performs quite well for all the numerical examples reported here.
We further mention that this global scaling is suggested to us by Prof. Chi-Wang Shu from Brown University in a private communication.
\end{remark}
After the troubled cells have been detected, we apply the characteristic-wise TVB limiter \cite{CS98, CS01} on the
conservative variables $(h_h+b_h, \bld m_h)$
with TVB parameter $M=0$.
To save space,
we leave out the derivation of this limiter and refer to \cite{CS98, CS01} for details.
We mention that while this TVB limiter is compatible with the well-balanced property of the DG scheme as $h_h+b_h$ is used in the limiting process.
In practice, we first apply this TVB limiter then apply the positivity preserving limiter \eqref{pplimit} for each inner Runge-Kutta stage values.
\subsection{Velocity computation and dry cell treatment}
\label{rk:dry}
We note that while Theorem \ref{thm:pp} and the limiter
\eqref{pplimit} ensures non-negativity of the water height cell average $\bar h_K^{n+1}$, and water height on the quadrature points $S_K$ at the next time level, this in general is not enough for the invertibility of the water height-weighted mass matrix, which is needed to compute the velocity approximation.
The invertibility of this weighted mass matrix is not a big issue as one can compute the scaling factor in \eqref{scale} such that it ensures posivitity of water height on all volume integration points in each cell.
A more serious issue is the velocity computation on dry cells with nearly zero water height, this weighted mass matrix is nonsingular but now close to zero, and the computed velocity approximation may be unphysically large and not reliable anymore.
Without a special velocity treatment on dry cells, the scheme (with TVB and posivity preserving limiters) may still fail to solve challenging problems with moving interface with wet and dry areas.
There are various wetting/drying treatment available in the literature \cite{KP07, BKW09}. However, our preliminary numerical experiments suggest that the most common approaches may not work well for our velocity based DG scheme.
For example, the simple trick of setting zero velocity when the water height $h_h$ is less than a given threshold, e.g. $10^{-6}$, which worked in \cite{XZS10}, or using
a regularized water height
\[
h^*:= \frac12 h+\frac12 \max\{h, (tol)^2/h\},
\]
with $tol$ a given small tolerance,
to compute the weighted mass matrix in Algorithm \ref{algV} as suggested
in \cite{KP07}
were not enough for our scheme with polynomial degree $k=2$ to solve a dam break problem on a dry bed.
After some initial testing, we come up with a relative simple velocity limiting approach that works for the numerical results reported in this manuscript. We apply the following two steps after an inner stage
water height $h_h$ and discharge $\bld m_h$ has been computed by Algorithm \ref{alg1}:
(1) Given a threshold percentage $0<\epsilon_{d}\ll 1$, we first mark cells with cell average $\bar h_K \le \epsilon_d\times h_{\max}^0$ as {\it dry} cells, where $h_{\max}^0$ is the maximum water height at initial time. Then, we remove the {\it high order} information on these dry cells by reverting to piecewise constant approximation of water height and discharge:
\begin{align}
\label{dry}
h_K\leftarrow \bar h_K, \quad
\bld m_K\leftarrow \bar{\bld m}_K, \quad
\text{ for all } K\in \Omega_h \text{ such that }
\bar h_K \le \epsilon_d\times h_{\max}^0,
\end{align}
where $h_K$ and $\bld m_K$ are the polynomial data in cell $K$, and
$\bar h_K$ and $\bar {\bld m_K}$ are the cell averages.
Note that this approach does not affect the local conservation property, but may lead to accuracy loss. However, since there are only a small amount of water in dry cells, such loss of accuracy may not be too significant if $\epsilon_d$ is taken small enough. We note that similar treatment was used in \cite{BKW09}.
(2)
The above approach may not be enough to control the velocity magnitude for high order schemes when $\epsilon_d$ is taken to be too small. We further propose a velocity limiter to smooth out extreme velocity values.
Given a user tunable value $V_{\max}$, which is an estimation of maximal allowed velocity approximation, we do the following two steps
for each component of the velocity approximation:
\begin{itemize}
\item [(i)]
Identify the collection of {\it troubled velocity cells}, denoted as $\omega(u_h)$, for the velocity
component $u_h$ as the cells where the maximum of the absolute velocity is larger than $V_{\max}$. For polynomial degree $k=2$ on triangles, the maximal value in the triangle is estimated as the maximal value on three vertices and three mid points of each edge:
\begin{align}
\label{tx}
\omega(u_h):=\{K\in \Omega_h:\quad
\max_{x\in v(K)}{|u_h(x)| > V_{\max}},
\}
\end{align}
where $v(K)$ is the collection of three vertices and three edge midpoints of cell $K$.
\item [(ii)]
On each of these troubled cells, we remove the velocity data, and replace it by the average of cell averages of its immediate neighboring cells which are not marked as troubled cells.
We repeat this procedure until all troubled cells have an updated (constant) velocity value:
\begin{align}
\label{tm}
& \text{While $\omega(u_h)$ is not empty, do the following: }\nonumber\\
&\;\;\quad\quad u_K\leftarrow \text{average of } \{\bar u_T\}
\text{ for } T\in \omega(K) \text{ and } T\not \in \omega(u_h).\\
&\;\;\quad\quad \text{remove cell $K$ from $\omega(u_h)$ if its value has been updated.}\nonumber
\end{align}
\end{itemize}
We note that the above velocity limiting procedure does not affect the local conservation property as the water height and discharge cell averages were never changed.
The above two approaches introduce two tunable parameters, namely $\epsilon_d$ in \eqref{dry}, and $V_{\max}$ in \eqref{tx}.
They will be chosen accordingly for specific wetting/drying examples.
For example, we can take $\epsilon_d = 5\times 10^{-3}$, and take $V_{\max}$ based on the maximum velocity magnitude for the lowest order scheme with $k=0$ for problems with moving wet/dry interfaces. With the above wetting/drying treatment, we are able to run simulation for the circular dam break problem with a dry bed, and
the dam bream problem with three mounds on unstructured triangular grids.
We mention that the above treatments are far away from perfect yet, as they need parameter tuning, and may lead to accuracy loss near dry cells.
They only serves as initial approaches for a successful simulation of SWEs with moving wet/dry interfaces. We are planning to further investigate more
robust and accurate wetting and drying treatments for our velocity based DG scheme in the near future.
For completeness, we list the final form of the fully discrete scheme below.
This method is locally conservative, well-balanced, and positivity preserving provided the time step size $\Delta t$ satisfy the CFL
condition \eqref{cfl}.
In practice, we take the time step size to be
\begin{align}
\Delta t = cfl \min_{K\in\Omega_h}\{\tau_K/\alpha_K^{\max}\},
\end{align}
where $cfl$ is the CFL number which depends on the polynomial degree $k$,
$\tau_K$ is the mesh size, and $\alpha_K^{\max}$ is the estimated maximum speed on the cell $K$.
If we detect a water height cell average $\bar h_K< \epsilon=10^{-12}$ in the inner stages in Step 1/2/4 of Algorithm \ref{alg3pt}, which means the time step size does not satisfy the condition \eqref{cfl}, we simply decrease $\Delta t$ by a half and redo the whole computation.
\begin{algorithm}
\caption{Posivity-preserving SSP-RK3 + DG + TVB limiter + wetting/drying treatment}
\begin{algorithmic}[1]
\REQUIRE $h_h^{n}\in V_h^k$, $\bld u_h^{n}, \bld m_h^{n}\in \bld V_h^k$ at time level $t^n$, and time step size $\Delta t>0$.
$tol >0$ for TVB limiter indicator \eqref{indicator},
$\epsilon_d>0$ for dry cell indicator \eqref{dry},
and
$V_{\max}>0$ for troubled velocity cell indicator \eqref{tx},
\ENSURE $h_h^{n+1}\in V_h^k$, and $\bld u_h^{n+1}, \bld m_h^{n+1}\in \bld V_h^k$ at next time level $t^{n+1}:=t^n+\Delta t$.
\STATE Apply Algorithm \ref{alg1} with inputs $h_h^n, \bld u_h^n, \bld m_h^n$ and $\Delta t$.
(If bottom topography $b_h$ is discontinuous, apply the hydrostatic reconstruction \eqref{hydro} in flux evaluations.)
Denote the outputs as $h_h^{(1)}, \bld m_h^{(1)}$.
\STATE Apply the dry cell limiter \eqref{dry}
for $h_h^{(1)}$ and $\bld m_h^{(1)}$;
Apply the characteristic-wise TVB limiter for the variables $(h_h^{(1)}+b_h, \bld m_h^{(1)})$ using indicator \eqref{indicator}
with indicating function
$h_h^{(1)}+b_h$;
Apply the positivity preserving limiter for $h_h^{(1)}$.
\STATE Apply Algorithm \ref{algV} with inputs $h_h^{(1)}, \bld m_h^{(1)}$.
Denote the velocity output as $\bld u_h^{(1)}$.
Then apply the velocity limiter in \eqref{tm}.
\STATE Apply Algorithm \ref{alg1} with inputs $h_h^{(1)}, \bld u_h^{(1)}, \bld m_h^{(1)}$ and $\Delta t$. Denote outputs as $h_h^{(2*)}, \bld m_h^{(2*)}$.
\STATE
Apply Algorithm \ref{alg2} with inputs $h_h^{n}, h_h^{(2*)}, \bld m_h^{n}, \bld m_h^{(2*)}$ and weights $w_1=0.75, w_2 = 0.25$.
Denote outputs as $h_h^{(2)}, \bld m_h^{(2)}$.
\STATE Apply the dry cell limiter \eqref{dry}
for $h_h^{(2)}$ and $\bld m_h^{(2)}$;
Apply the characteristic-wise TVB limiter for the variables $(h_h^{(2)}+b_h, \bld m_h^{(2)})$ using indicator \eqref{indicator}
with indicating function
$h_h^{(2)}+b_h$;
Apply the positivity preserving limiter for $h_h^{(2)}$.
\STATE Apply Algorithm \ref{algV} with inputs $h_h^{(2)}, \bld m_h^{(2)}$.
Denote the velocity output as $\bld u_h^{(2)}$.
Then apply the velocity limiter in \eqref{tm}.
\STATE Apply Algorithm \ref{alg1} with inputs $h_h^{(2)}, \bld u_h^{(2)}, \bld m_h^{(2)}$ and $\Delta t$. Denote outputs as $h_h^{(3*)}, \bld m_h^{(3*)}$.
\STATE
Apply Algorithm \ref{alg2} with inputs $h_h^{n}, h_h^{(3*)}, \bld m_h^{n}, \bld m_h^{(3*)}$ and weights $w_1=1/3, w_2 = 2/3$.
Denote outputs as $h_h^{n+1}, \bld m_h^{n+1}$.
\STATE Apply the dry cell limiter \eqref{dry}
for $h_h^{n+1}$ and $\bld m_h^{n+1}$;
Apply the characteristic-wise TVB limiter for the variables $(h_h^{n+1}+b_h, \bld m_h^{n+1})$ using indicator \eqref{indicator}
with indicating function
$h_h^{n+1}+b_h$;
Apply the positivity preserving limiter for $h_h^{n+1}$.
\STATE Apply Algorithm \ref{algV} with inputs $h_h^{n+1}, \bld m_h^{n+1}$.
Denote the velocity output as $\bld u_h^{n+1}$.
Then apply the velocity limiter in \eqref{tm}.
\end{algorithmic}
\label{alg3pt}
\end{algorithm}
\section{Numerical results}
In this section we present numerical results of our velocity based DG scheme Algorithm \eqref{alg3pt}. We report results using the third order DG method with $k = 2$. The CFL number is taken to be $cfl = 0.1$ for 1D examples, and $cfl=0.05$ for 2D examples.
The gravitation constant $g$ is fixed as 9.812 except the test in Example 4.6, where it is taken to be $g=10$.
We take the tolerance $tol = 0.02$ in the Fu-Shu indicator \eqref{indicator} for all examples.
Moreover, unless explicitly mentioned, we turn off the dry cell limiter \eqref{dry} and the velocity limiter \eqref{tx}--\eqref{tm}, which are only needed when the problem has a moving dry/wet interface.
The implementation is based on the python interface of the NGSolve
software \cite{Schoberl16}, \url{https://ngsolve.org/}.
Source code for all the examples can be found in the git repository, \url{https://github.com/gridfunction/SWE}.
\subsection*{Example 4.1: Accuracy Test in 1D}
We start with an accuracy test to demonstrate the high order accuracy of our schemes for a smooth solution of the SWEs. Following the setup in \cite{WDGX20}, we take the following bottom topography and initial
conditions:
\begin{align*}
&b(x) =\; \sin^2(\pi x),\quad h(x,0) = \; 5 + e^{\cos(2\pi x)},\quad hu(x,0) = \;\sin(\cos(2\pi x)).
\end{align*}
The computation domain is a periodic unit interval $[0, 1]$, and final time is $t=0.1$ where the solution is still smooth.
We apply the plain Algorithm \ref{alg3} without limiter, and compute the $L^2$-errors of water height $h_h$, velocity $u_h$, and discharge $m_h$ on a sequence of uniform meshes with $N=50\times 2^l$ cells for $l=0,1,2,3,4$. We take the solution on $N=25\times 2^5=1600$ cells
as the {\it reference} solution when computing these $L^2$-errors.
The results are recorded in Table \ref{tab:1}.
We clearly observe the expected third order convergence, and the error magnitude are comparable to the results reported in \cite{WDGX20}
for another third order entropy stable DG scheme.
\begin{table}[ht!]
\centering
\begin{tabular}{c|cc|cc|cc}
N & $L^2$-err in $h_h$ & rate
& $L^2$-err in $u_h$ & rate
& $L^2$-err in $m_h$ & rate\\
\hline
50& 2.997e-04 & -- & 3.583e-04 & -- & 2.577e-03 &-- \\
100& 2.730e-05 & 3.46 & 3.273e-05 & 3.45 & 2.352e-04 & 3.45 \\
200& 2.949e-06 & 3.21 & 3.538e-06 & 3.21 & 2.542e-05 & 3.21 \\
400& 3.600e-07 & 3.03 & 4.323e-07 & 3.03 & 3.103e-06 & 3.03 \\
800& 4.408e-08 & 3.03 & 5.296e-08 & 3.03 & 3.798e-07 & 3.03 \\
\hline
\end{tabular}
\vspace{1ex}
\caption{$L^2$ errors and
convergence rate at time $t=0.1$ for Example 4.1.}
\label{tab:1}
\end{table}
\subsection*{Example 4.2: The Well-Balanced Test in 1D}
In this example, we test the well-balanced property of our proposed methods to ensure that
the still-water steady state is exactly preserved.
We consider two different choices of the bottom topography as used in \cite{WDGX20}:
a smooth bottom with
\[
b(x) = 5\exp\left(-0.4(x-5)^2
\right),
\]
and a discontinuous bottom with
\[
b(x) = \left\{
\begin{tabular}{ll}
4, & if $4\le x\le 8$,\\[.3ex]
0, & otherwise.
\end{tabular}
\right.
\]
The computational domain is $[0, 10]$ with wall boundary conditions.
The initial condition is taken as the stationary state
\[
h+b = 10, u = 0.
\]
We solve the problem until time $t=0.5$ on three meshes with 100, 2000, and 400 uniform cells, and record the $L^2$-errors in Table \ref{tab:2}. We observe all errors are at the level of
round-off errors, which verifies the well-balanced property.
\begin{table}[ht!]
\centering
\begin{tabular}{c|ccc|ccc}
& & smooth bot. & && disc. bot. &
\\
\hline
N & $L^2$-err in $h_h$
& $L^2$-err in $u_h$
& $L^2$-err in $m_h$
& $L^2$-err in $h_h$
& $L^2$-err in $u_h$
& $L^2$-err in $m_h$ \\
\hline
100& 9.819e-14 & 7.380e-14 & 5.025e-13 & 7.700e-14 & 7.627e-14 & 5.024e-13 \\
200& 1.747e-13 & 8.344e-14 & 6.331e-13 & 1.805e-13 & 1.042e-13 & 8.494e-13 \\
400& 3.740e-13 & 1.328e-13 & 1.029e-12 & 2.787e-13 & 2.153e-13 & 1.355e-12 \\
\hline
\end{tabular}
\vspace{1ex}
\caption{$L^2$ errors at time $t=0.5$ for Example 4.2.}
\label{tab:2}
\end{table}
\subsection*{Example 4.3: A Small Perturbation Test in 1D}
We test the following quasi-stationary test case proposed by LeVeque \cite{LeVeque98}, which is a small
perturbation of the steady state solution.
The computational domain is $[0, 2]$, and the bottom topography b(x) is given by
\[
b(x) = \left\{
\begin{tabular}{ll}
$\frac14(\cos(10\pi(x-1.5))+1)$, & if $1.4\le x\le 1.6$,\\[.7ex]
0, & otherwise.
\end{tabular}
\right.
\]
The initial conditions are
\[
h(x,0) = \left\{
\begin{tabular}{ll}
$1-b(x)+\epsilon$, & if $1.1\le x\le 1.2$,\\[.7ex]
$1-b(x)$, & otherwise.
\end{tabular}
\right., \quad u(x,0) = 0,
\]
where $\epsilon$ is a given constant representing the size of the perturbation.
Following \cite{LeVeque98},
we consider a
case with a big pulse ($\epsilon = 0.2$) and case with a small pulse ($\epsilon=0.001$). The final time of simulation is $t=0.2$.
We compare our scheme on a uniform coarse mesh with $N=200$ cells and
a uniform fine mesh with $N=2000$ cells.
The results at final time
for the total water surface $h+b$ and discharge $m=hu$ for the big pulse case
are shown in Figure \ref{fig:sp1}, and
those for the small pulse case
are shown in Figure \ref{fig:sp2}.
In these figures, blue squares indicate the troubled cells identified by our indicator \eqref{indicator}.
We observe good agreement of the results on coarse and fine meshes without spurious
numerical oscillations, which also agrees well with results in the literature.
Moreover, we observe that the indicator \eqref{indicator} with $tol=0.02$ only activates cells close to the moving shocks for both case, with slightly more cells being identified as troubled cells for the small perturbation test.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{rho2b}
\includegraphics[width=0.45\textwidth]{m2b}
\caption{Example 4.3 with
with a big pulse
$\epsilon=0.2$ at time $t = 0.2$. Left: the water
surface $h + b$; right: the discharge $m$. Blue squares indicate cells where TVB limiter are used at the final time.
}
\label{fig:sp1}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{rho2a}
\includegraphics[width=0.45\textwidth]{m2a}
\caption{Example 4.3 with
with a small pulse
$\epsilon=0.001$ at time $t = 0.2$. Left: the water
surface $h + b$; right: the discharge $m$. Blue squares indicate cells where TVB limiter are used at the final time.
}
\label{fig:sp2}
\end{figure}
\subsection*{Example 4.4: A Dam Breaking Problem over a Bump in 1D}
We consider a one-dimensional dam breaking problem over a rectangular bump.
It involves a rapidly varying flow over a discontinuous bottom topography.
Following \cite{WDGX20}, we take the computational domain as [0, 1500], and use the following discontinuous bottom topography:
\[
b(x) = \left\{
\begin{tabular}{ll}
8, & if $|x-750|<1800/8$,\\[.7ex]
0, & otherwise.
\end{tabular}
\right.
\]
We use outflow boundary conditions, and record the results at time $t=60$ in Figure \ref{fig:db}, again using a uniform coarse mesh with $N=200$ cells, and a uniform fine mesh with $N=2000$ cells.
We observe good agreement of the results on two meshes, which also agrees well with results reported in the literature. We note that the discharge has a small kink near
$x=937.5$, where the discontinuous of bottom topography happens. Also, the indicator is successful in identifying solution discontinuities.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{rho3a}
\includegraphics[width=0.45\textwidth]{m3a}
\caption{Example 4.4 at time $t = 60$. Left: the water
surface $h + b$; right: the discharge $m$. Blue squares indicate cells where TVB limiter are used at the final time.
}
\label{fig:db}
\end{figure}
\subsection*{Example 4.5: Entropy Glitch Test in 1D}
We consider the Riemann problem with a flat bottom considered in \cite{WWAGW18}.
The computational domain is $[-1,1]$, the bottom topography $b(x)=0$, and initial condition is
\[
h(x,0) = \left\{
\begin{tabular}{ll}
1, & if $x<0$,\\[.7ex]
0.1, & otherwise.
\end{tabular}
\right.\quad u(x,0) = 0.
\]
The gravitational constant is taken to be $g=10$, and final time of simulation is
$t=0.2$.
It was shown in \cite[Fig. 8]{WWAGW18} that standard DG method with a local Lax–Friedrichs numerical flux develops an unphysical discontinuity, called an “entropy glitch”, at $x = 0$, while the entropy stable DG method is able
to capture the solution well on the coarse mesh.
The results in a uniform mesh with $200$ cells are shown in Figure \ref{fig:gt}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{rho4c}
\includegraphics[width=0.45\textwidth]{m4c}
\caption{Example 4.5 at time $t=0.2$.
Left: the water
surface $h + b$; right: the discharge $m$.
Blue squares indicate cells where TVB limiter are used at the final time.
}
\label{fig:gt}
\end{figure}
\subsection*{Example 4.6: A Dam Break Problem with a Dry Bed in 1D}
We consider a Riemann Problem with a constant bottom used in \cite{XZS10}.
Here a dried river bed is used to examine the performance of our scheme in case of moving wet/dry interface. The computation domain is taken to be $[-300, 300]$, and initial condition is
\[
h(x,0) = \left\{
\begin{tabular}{ll}
10, & if $x<0$,\\[.7ex]
$10^{-12}$, & otherwise.
\end{tabular}
\right.\quad u(x,0) = 0.
\]
Here we use a tiny positive value $10^{-12}$ to indicate the dry bed.
This avoids division by zero in the velocity computation.
This is a very challenging problem as our default algorithm without dry cell limiter or
velocity limiter fails after a couple of time steps due to an excessive large velocity approximation.
Here we activate the dry cell limiter \eqref{dry} with $\epsilon_d=5\times 10^{-3}$ to avoid excessive large velocity approximations. The velocity limiter \eqref{tx}--\eqref{tm} is not needed for this example.
The results at times $t=4$, $t=8$, and $t=12$ on the uniform mesh with $N=200$ cells are plotted in
Figure \ref{alg3}.
Again, we observe good agreement with results in the literature.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{rho4a}
\includegraphics[width=0.45\textwidth]{m4a}
\caption{Example 4.6 at times $t=4$, $t=8$, and $t=12$.
Left: the water
surface $h + b$; right: the discharge $m$.
}
\label{fig:pp}
\end{figure}
\subsection*{Example 4.7: Accuracy Test in 2D}
Now we turn to the performance of our scheme on 2D triangular meshes.
This is a 2D version of the 1D accuracy test considered in Example 4.1.
The domain is a periodic unit square $[0,1]\times [0, 1]$.
The bottom topography and the initial conditions are given as follows:
\begin{align*}
&b(x) =\; \sin(2\pi x)+\sin(2\pi y),\\
& h(x,0) = \; 10 + e^{\sin(2\pi x)}\cos(2\pi y),\\
& hu(x,0) = \;\sin(\cos(2\pi x))\sin(2\pi y),\\
& hv(x,0) = \;\cos(2\pi x)\cos(\sin(2\pi y)).
\end{align*}
The final time is $t=0.05$ where the solution is still smooth.
We apply the plain Algorithm \ref{alg3} without limiter, and compute the $L^2$-errors of water height $h_h$, velocity $u_h$, and discharge $m_h$ on a sequence of uniform structured triangular meshes with $N\times N\times 2$ cells where $N=25\times 2^l$ cells for $l=0,1,2,3$. We take the solution on
uniform structured triangular meshes with $400\times 400\times 2$
as the {\it reference} solution when computing these $L^2$-errors.
The results are recorded in Table \ref{tab:3}.
We again observe the expected third order convergence for the water height, and nearly third order convergence for the velocity and discharge.
\begin{table}[ht!]
\centering
\begin{tabular}{c|cc|cc|cc}
N & $L^2$-err in $h_h$ & rate
& $L^2$-err in $u_h$ & rate
& $L^2$-err in $m_h$ & rate\\
\hline
25& 1.420e-03 & -- & 1.670e-03 & -- & 1.379e-02 &-- \\
50& 1.567e-04 & 3.18 & 2.083e-04 & 3.00 & 1.683e-03 & 3.03 \\
100& 1.917e-05 & 3.03 & 2.820e-05 & 2.89 & 2.371e-04 & 2.83 \\
200& 2.363e-06 & 3.02 & 3.880e-06 & 2.86 & 3.403e-05 & 2.80 \\
\hline
\end{tabular}
\vspace{1ex}
\caption{$L^2$ errors and
convergence rate at time $t=0.05$ for Example 4.7.}
\label{tab:3}
\end{table}
\subsection*{Example 4.8: A Small Perturbation Test in 2D}
We test the following 2D quasi-stationary test case proposed by LeVeque \cite{LeVeque98}, which is a small
perturbation of the steady state solution.
The computation domain is $[0, 2] \times [0, 1]$. The bottom topography consists of an elliptical shaped hump
\[
b(x,y) = 0.8\exp(-5(x-0.9)^2-50(x-0.5)^2),
\]
and the initial conditions are
\[
h(x,y,0) = \left\{
\begin{tabular}{ll}
$1-b(x,y)+0.01$, & if $0.05\le x\le 0.15$,\\[.7ex]
$1-b(x,y)$, & otherwise.
\end{tabular}
\right., \quad u(x,y,0) = v(x,y,0) =0.
\]
Outflow boundary conditions are imposed on the left and right boundary while wall boundary condition (symmetry) are imposed on the top and bottom boundaries.
Due to symmetry, we perform the calculation on half of the domain
$\Omega = [0, 2]\times [0, 0.5]$ with symmetry boundary condition on the top boundary
$y=0.5$. We consider our scheme on an unstructured triangular mesh with mesh size $\tau_K = 0.01$. The water surface contour at times $t=0.12, 0.24,0.36,0.48,0.60$ are recorded in Figure \ref{fig:sp3}. Our scheme produce non oscillatory solutions and the results agrees well with those in the literature.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{wb0}
\includegraphics[width=0.45\textwidth]{wb1}
\includegraphics[width=0.45\textwidth]{wb2}
\includegraphics[width=0.45\textwidth]{wb3}
\includegraphics[width=0.45\textwidth]{wb4}
\caption{Example 4.8.
Contour plot for $h+b-1$. 30 uniform contour lines from
$h+b-1 = -0.008$ to
$h+b-1 = 0.008$.
From left to right, top to bottom: $t=0.12, 0.24,0.36,0.48,0.6$.
}
\label{fig:sp3}
\end{figure}
\subsection*{Example 4.9: Circular Dam Break Test in 2D}
We consider the circular dam break problem used in \cite[Sect. 3.2.1]{M09}.
The space domain is a $50\times 50$ square with a cylindrical dam with radius $r=11$ and centred in the square. The initial water height is 10 inside the dam, and is either
$1$ outside the dam (a wet bed), or $10^{-12}$ outside the dam (a dry bed). The final time of the simulation is $t=0.69$.
The bottom topography is set to be zero.
Here the dry bed case need special care, where we applied the dry cell limiter \eqref{dry} with $\epsilon_d=5\times 10^{-3}$, and activated the velocity limiter
\eqref{tx}--\eqref{tm} with $V_{\max}=15$.
These treatment were not used for the wet bed case.
Due to symmetry of the problem, we perform the computation only on a quarter of the domain with symmetric boundary conditions. We take an unstructured triangular mesh with
mesh size $\tau_K = 0.5$. The results of two cases are presented in Figure \ref{fig:cd}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{cm1}
\includegraphics[width=0.45\textwidth]{cm2}
\caption{Example 4.9.
Contour and surface plots of water height for the circular dam-break problem at $t=0.69$.
Left: web bed. 11 uniform contour lines from $2$ to $9.4$;
Right dry bed. 12 uniform contour lines from $0.01$ to $8.9$.
}
\label{fig:cd}
\end{figure}
\subsection*{Example 4.10: Dam Break on a Closed Channel in 2D}
We consider the problem \cite[Sect. 8.9]{GPC07}
which model dam break on a closed channel.
The domain represents a channel of 75 length and 30 width with
three mounds will wall boundary conditions.
The shape of the mounds is defined by the function
$b(x, y) = \max(0, m1, m2, m3)$, where
\begin{align*}
m_1 =&\; 1- 0.1\sqrt{
(x - 30)^2 + (y-22.5)^2},\\
m_2 =&\; 1 - 0.1\sqrt{
(x - 30)^2 + (y-7.5)^2},\\
m_3 =&\; 2.8 - 0.28\sqrt{
(x-47.5)^2 + (y-15)^2}.
\end{align*}
The initial conditions are
\[
h(x,y,0) = \left\{
\begin{tabular}{ll}
$1.875$, & if $x<16$\\[.7ex]
$10^{-12}$, & otherwise.
\end{tabular}\quad u(x,y,0)=v(x,y,0) = 0
\right.
\]
Due to the moving wet/dry interface, we activate the dry cell limiter with
$\epsilon_d = 10^{-3}$, and the velocity limiter \eqref{tx}--\eqref{tm} with
$V_{\max}=9$.
Due to symmetry, we only perform the calculation on half of the domain $[0,75]\times [0,15]$, and apply symmetry boundary conditions on all the boundaries.
Contour plots of the water surface for the simulation results on an unstructured triangular mesh with mesh size $\tau_K = 0.5$ are shown in Figure \ref{fig:xx}
for various times. We observe complex flow structures for this problem, and our scheme
produces satisfactory results compared with those from \cite[Fig. 15]{GPC07}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{tm0}
\includegraphics[width=0.45\textwidth]{tm1}
\includegraphics[width=0.45\textwidth]{tm2}
\includegraphics[width=0.45\textwidth]{tm3}
\includegraphics[width=0.45\textwidth]{tm4}
\includegraphics[width=0.45\textwidth]{tm5}
\includegraphics[width=0.45\textwidth]{tm6}
\includegraphics[width=0.45\textwidth]{tm7}
\caption{Example 4.10.
Contour and surface plots of water surface for the dam-break problem on a closed channel.
20 uniform contour lines from $0$ to $1.2$.
Left to right, top to bottom: $t=5, 10, 15, 20, 25, 30, 35, 40$.
}
\label{fig:xx}
\end{figure}
\section{Conclusion}
We proposed a novel velocity-based DG scheme for the SWEs. Our semidiscrete DG scheme is
locally conservative, entropy stable, and well-balanced.
We then apply the SSP-RK3 time stepping for the time discretzation, and obtained an explicit locally conservative, well-balanced, and positivity-preserving fully discrete scheme in Algorithm \ref{alg3pt}, where the treatment of strong shocks via a
characteristic-wise TVB limiter and proper wetting/drying treatment near dry cells was also discussed. Ample numerical examples in 1D and 2D illustrated the good performance of our scheme. Our entropy stable scheme is particularly simple and competitive compared with existing entropy stable DG schemes for SWEs in the literature.
The velocity-based DG scheme can be used to construct robust entropy/energy stable DG schemes for other compressible flow problems, which will be carried out in our future studies.
\
\textbf{Acknowledgement:} The author would like to thank Yulong Xing from Ohio State University for fruitful discussions on the topic.
\bibliographystyle{siam}
|
\section{Introduction}
An infinite matrix with real entries $A=(a_{n,k})$ is said to be \emph{regular} if it transforms convergent sequences into convergent sequences and preserves the corresponding limits (details will be given in Section \ref{sec:mainresults}).
A classical result due to Silverman--Toeplitz provides necessary and sufficient conditions, depending only on the entries of $A$, which characterize the class of regular matrices, see e.g. \cite[Theorem 2.3.7]{MR1817226}:
\begin{thm}\label{thm:SilvermanToeplizsimple}
An infinite real matrix $A=(a_{n,k})$ is regular if and only if\textup{:}
\begin{enumerate}[label={\rm (\roman*)}]
\item $\sup_n \sum_k|a_{n,k}|<\infty$\textup{;}
\item $\lim_n\sum_ka_{n,k}=1$\textup{;}
\item $\lim_na_{n,k}=0$ for all $k$\textup{.}
\end{enumerate}
\end{thm}
Several extensions and analogues
of the characterization above
can be found in the literature.
First, a \textquotedblleft multidimensional\textquotedblright\,version of Theorem \ref{thm:SilvermanToeplizsimple} for double sequences has been proved by George M.
Robinson \cite{Robinson26} and Hugh J. Hamilton \cite{MR1545904}.
Second, in 1950 Abraham Robinson \cite{MR37371} proved the operator analogue of Theorem \ref{thm:SilvermanToeplizsimple} replacing each $a_{n,k}$ with a (possibly unbounded) linear operator $A_{n,k}$ acting on a given Banach space, cf. Theorem \ref{thm:originalIJREGULAR} below.
Third, on a different direction, the author and Jeff Connor \cite{ConnorLeo} recently studied the ideal/filter version of the notion of regularity in the scalar case and proved that the analogue of Theorem \ref{thm:SilvermanToeplizsimple} holds in several, but not all, cases.
The aim of this work is to provide a unifying framework which allows to shed light on the relationship between all the above results, to extend the latter ones, and to obtain, as a byproduct of the employed methods, several related characterizations.
This will require us to deal with the theory of infinite matrices of linear operators and to prove a certain number of intermediate lemmas.
Most results are formulated in the context of ideal convergence. We remark that this choice is not done for the sake of generality: indeed, for the above purposes, we will exploit the simple facts that $c(\mathcal{I}) \cap \ell_\infty$ is $c$ if $\mathcal{I}=\mathrm{Fin}$ and equals to $\ell_\infty$ if $\mathcal{I}$ is maximal, that the Pringsheim convergence of double sequences coincides with $\mathcal{I}$-convergence for a suitable ideal $\mathcal{I}$, etc.
An additional motivation comes from the fact that
that the study of ideals on countable sets and their representability
may have some relevant potential for the study of the geometry of Banach spaces, see e.g. \cite{MR4124855, MR3436368, KwelaLeonetti, LeonettiCaprio}.
Informally, we provide an operator version of the characterization of regular matrices in the context of ideal convergence, together with some sufficient conditions which allow for several substantial simplifications.
The results depend on the boundedness assumption on the sequence spaces in the domain and/or codomain of such matrices.
In addition, we provide a characterization of the matrix classes
$(\ell_\infty, c_0(\mathcal{J})\cap \ell_\infty)$, $(c(\mathcal{I}), c_0(\mathcal{J})\cap \ell_\infty)$ $(\ell_\infty, \ell_\infty(\mathcal{J}))$, and $(c,\ell_\infty(\mathcal{J}))$ for certain ideals $\mathcal{I},\mathcal{J}$ on $\omega$,
see Corollary \ref{cor:llinftyc0}, Theorem \ref{thm:cIc_0bJ}, Theorem \ref{thm:maddoxmain}, and Theorem \ref{thm:maddoxmaincXellinfty}, respectively.
Lastly, we obtain an ideal version of the Hahn--Schur theorem (which is used to prove that weak and norm convergence coincide on $\ell_1$), see Theorem \ref{thm:hahnschur}.
The proofs of the main results are given in Section \ref{sec:mainproofs}.
\section{Notations and Main Results}\label{sec:mainresults}
Let $\mathcal{I}$ be an ideal on the nonnegative integers $\omega$, that is, a collection of subsets of $\omega$ which is closed under subsets and finite unions.
Unless otherwise stated, it is assumed that it contains the collection $\mathrm{Fin}$ of finite sets and it is different from the power set.
Denote its dual filter by $\mathcal{I}^\star:=\{S\subseteq \omega: S^c \in \mathcal{I}\}$ and define $\mathcal{I}^+:=\{S\subseteq \omega: S\notin \mathcal{I}\}$.
Among the most important ideals, we find the family of asymptotic density zero sets
\begin{equation}\label{eq:definitionZ}
\mathcal{Z}:=\left\{S\subseteq \omega: \lim_{n\to \infty} \frac{|S \cap [0,n]|}{n+1}=0\right\}.
\end{equation}
We refer to \cite{MR2777744}
for a recent survey on ideals and associated filters.
Let $V$ be a real Banach space, and denote its closed unit ball by $B_V$ and its unit sphere by $S_V$.
Given a sequence $\bm{x}=(x_n)$ taking values in $V$ and an ideal $\mathcal{I}$ on $\omega$, we say that $\bm{x}$ is $\mathcal{I}$\emph{-convergent to} $\eta \in V$, shortened as $\mathcal{I}\text{-}\lim \bm{x}=\eta$ or $\mathcal{I}\text{-}\lim_n x_n=\eta$, if $\{n \in \omega: x_n \notin U\} \in \mathcal{I}$ for all neighborhood $U$ of $\eta$; for the clarify of exposition, all sequences taking values in $V$ will be written in bold.
Note that $\mathcal{Z}$-convergence is usually called \emph{statistical convergence}, see e.g. \cite{MR1181163}.
As remarked in \cite[Example 3.4]{MR3671266}, the notion of $\mathcal{I}$-convergence include the well-known uniform, Pringsheim, and Hardy convergences for double sequences.
In addition, if $\bm{y}$ is a real nonnegative sequence, we write $\mathcal{I}\text{-}\limsup \bm{y}:=\inf\{r \in \mathbf{R}\cup \{\infty\}: \{n \in \omega: y_n\ge r\}\in \mathcal{I}\}$.
Now, define the following sequence spaces
\begin{displaymath}
\begin{split}
\ell_\infty(V)&:=\left\{\bm{x} \in V^\omega: \|\bm{x}\|<\infty\right\},\\
\ell_\infty(V,\mathcal{I})&:=\left\{\bm{x} \in V^\omega: \mathcal{I}\text{-}\limsup\nolimits_n\|x_n\|<\infty\right\},\\
c(V,\mathcal{I})&:=\left\{\bm{x} \in V^\omega: \mathcal{I}\text{-}\lim \bm{x}=\eta \text{ for some }\eta \in V\right\},\\
c_0(V,\mathcal{I})&:=\left\{\bm{x} \in V^\omega: \mathcal{I}\text{-}\lim \bm{x}=0\right\},\\
c_{00}(V,\mathcal{I})&:=\left\{\bm{x} \in V^\omega: \mathrm{supp}\,\bm{x} \in \mathcal{I}\right\},
\end{split}
\end{displaymath}
where $\|\bm{x}\|:=\sup_n\|x_n\|$ stands for the supremum norm and $\mathrm{supp}\,\bm{x}$ for the support $\{n \in \omega: x_n\neq 0\}$.
Clearly $\ell_\infty(V)=\ell_\infty(V, \mathrm{Fin})$;
sequences in $\ell_\infty(V,\mathcal{I})$ are usually called $\mathcal{I}$-bounded.
If $V=\mathbf{R}$ and $\mathcal{I}=\mathrm{Fin}$, the above sequence spaces correspond to the usual $\ell_\infty$, $c$, $c_0$, and $c_{00}$, respectively.
Every subspace of $\ell_\infty(V)$ will be endowed with the supremum norm.
It is clear that $c_{00}(V,\mathcal{I})\subseteq c_{0}(V,\mathcal{I})\subseteq c(V,\mathcal{I})\subseteq \ell_\infty(V,\mathcal{I})$. However, unless $\mathcal{I}=\mathrm{Fin}$ or $V=\{0\}$, $c_{00}(V,\mathcal{I})$ is not contained in $\ell_\infty(V)$.
Hence it makes sense to define the subspace
$$
c^b(V,\mathcal{I}):=c(V,\mathcal{I})\cap \ell_\infty(V),
$$
and, similarly, $c_{0}^b(V,\mathcal{I})$ and $c_{00}^b(V,\mathcal{I})$.
The symbol $V$ will be removed from the notation if it is understood from the context so that, e.g., $c_{00}(\mathcal{I})=c_{00}(V, \mathcal{I})$. Similarly, we may remove $\mathcal{I}$ in the case $\mathcal{I}=\mathrm{Fin}$.
At this point, let $X,Y$ be two Banach space and denote by $\mathcal{L}(X,Y)$ and $\mathcal{B}(X,Y)$ the vector spaces of linear operators from $X$ to $Y$ and its subspace of bounded linear operators, respectively.
We assume that $\mathcal{L}(X,Y)$ and all its subspaces are endowed with the strong operator topology, so that a sequence $(T_n)$ of linear operators in $\mathcal{L}(X,Y)$ converges to $T \in \mathcal{L}(X,Y)$ if and only if $(T_nx)$ is convergent in the norm of $Y$ to $Tx$ for all $x \in X$.
Let $A=(A_{n,k}: n,k \in \omega)$ be an infinite matrix of linear operators $A_{n,k}\in \mathcal{L}(X,Y)$.
Moreover,
for each $n \in \omega$ and $E\subseteq \omega$, let us write
$$
A_{n,E}:=(A_{n,k}: k \in E)
$$
and $A_{n,\ge k}:=A_{n,\{k,k+1,\ldots\}}$ for the $k$th tail of the $n$th row of $A$.
In particular, $A_{n,\omega}$ is the $n$th row of $A$ (and use an analogue notation for a sequence $(T_k)$ of operators so that, for instance, $T_{\ge 2}=(T_2,T_3,\ldots)$).
For each $n \in \omega$ and $E\subseteq \omega$, define the the group norm
$$
\|A_{n,E}\|:=
\sup\left\{\left\|\sum\nolimits_{k\in F}A_{n,k}x_k\right\|: F\subseteq E \text{ is finite and each }x_k \in B_X\right\},
$$
cf. \cite{MR0447877, MR568707}
(in fact, \emph{every} $x_k$ can be chosen on the unit sphere $S_X$: this depends on the fact that, given distinct $a,b \in X$, the function $f: [0,1] \to \mathbf{R}$ defined by $f(t):=\|a+t(b-a)\|$ has a point of maximum in $t=0$ or in $t=1$: indeed, the segment $\{a+t(b-a): t \in [0,1]\}$ is contained in the closed ball with center $0$ and radius $\max\{\|a\|,\|b\|\}$, which is convex).
Note that the value $\|A_{n,E}\|$ is possibly not finite.
In addition, if $X=\mathbf{R}$ and $A_{n,E}$ is represented by the real sequence $(a_{n,k}: k \in E)$ then $\|A_{n,E}\|=\sum_{k \in E}|a_{n,k}|$.
Given an $X$-valued sequence $\bm{x}=(x_n)$, let $A\bm{x}$ be its $A$-transform, that is, the sequence $A\bm{x}:=(A_n\bm{x}: n \in \omega)$ where
$$
\forall n \in \omega, \quad
A_n\bm{x}:=\sum\nolimits_{k}A_{n,k}x_k,
$$
provided that each series is convergent in the norm of $Y$. Accordingly, let $\mathrm{dom}(A)$ be the domain of $A$, that is, the family of those sequences $\bm{x}$ such that $A\bm{x}$ is well defined.
For each sequence subspace $\mathscr{A}\subseteq X^\omega$ and $\mathscr{B}\subseteq Y^\omega$, let $(\mathscr{A}, \mathscr{B})$ be the set of matrices $A=(A_{n,k})$ of
(not necessarily bounded)
linear operators in $\mathcal{L}(X,Y)$ such that
$$
\mathscr{A} \subseteq \mathrm{dom}(A)
\,\,\,\text{ and }\,\,\,
A\bm{x} \in \mathscr{B} \,\text{ for all }\bm{x} \in \mathscr{A}.
$$
We refer to \cite{MR568707} for the theory of infinite matrices of operators.
In the scalar case, the relationship between summability and ideal convergence has been recently studied in \cite{Filipow18}.
\subsection{Bounded to Bounded case}\label{subsec:boundedbounded}
The main definition of this work follows:
\begin{defi}\label{def:mainIJregular}
Let $\mathcal{I}, \mathcal{J}$ be ideals on $\omega$ and fix $T \in \mathcal{L}(X,Y)$.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ is said to be $(\mathcal{I}, \mathcal{J})$\emph{-regular with respect to $T$} if
$$
A \in (c^b(X,\mathcal{I}),c^b(Y,\mathcal{J}))
\,\,\, \text{ and }\,\,\,
\mathcal{J}\text{-}\lim A\bm{x}=T(\,\mathcal{I}\text{-}\lim \bm{x})
\,\text{ for all }\bm{x} \in c^b(X,\mathcal{I}).
$$
\end{defi}
If $T$ is the identity operator $I$ on $X$ (namely, $Ix=x$ for all $x \in X$), we will simply say that $A$ is $(\mathcal{I},\mathcal{J})$\emph{-regular}.
Note that if $\mathcal{I}=\mathcal{J}=\mathrm{Fin}$, $X=Y=\mathbf{R}$,
and $T=I$,
then Definition \ref{def:mainIJregular} corresponds to the ordinary regular matrices.
The following result, essentially due to Robinson \cite[Theorem VII]{MR37371}, is the (unbounded) operator version of Theorem \ref{thm:SilvermanToeplizsimple},
cf. also \cite[Theorem 1]{MR361524}.
\begin{thm}\label{thm:originalIJREGULAR}
Fix a linear operator $T \in \mathcal{B}(X,Y)$, where $X,Y$ are Banach spaces.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ is $(\mathrm{Fin}, \mathrm{Fin})$-regular with respect to $T$ if and only if there exists $k_0 \in \omega$ such that\textup{:}
\begin{enumerate}[label={\rm (\textsc{S}\arabic{*})}]
\item \label{item:S1} $\sup_n\|A_{n,\ge k_0}\|<\infty$\textup{;}
\item \label{item:S2} $\lim_n \sum_{k}A_{n,k}=T$\textup{;}
\item \label{item:S3} $\lim_n A_{n,k}=0$ for all $k \in \omega$\textup{.}
\end{enumerate}
\end{thm}
A variant for continuous linear operators between Fr\'{e}chet spaces has been proved by Ramanujan in \cite{MR216203}, cf. also \cite[Corollary 6]{MR1936721}.
Notice that \ref{item:S3} can be rewritten as $\lim_n \sum_{k \in E}A_{n,k}x_k=0$ for all sequences $\bm{x} \in \ell_\infty(X)$ and $E \in \mathrm{Fin}$, which is also equivalent to $A \in (c_{00}(X), c_0(Y))$.
\begin{rmk}\label{rmk:evilgroupnorm}
Condition \ref{item:S3} may look equivalent also to:
\begin{enumerate}[label={\rm (\textsc{S}\arabic{*})}
\setcounter{enumi}{3}
\item \label{item:S3sharp} $\lim_n \|A_{n,k}\|=0$ for all $k \in \omega$.
\end{enumerate}
This is correct if $X$ is finite dimensional, cf. Lemma \ref{lem:finitedimensionalnormkjfdhgd} below.
However, \ref{item:S3sharp} is strictly stronger in general. For, set $X=\ell_2=\{x \in \ell_\infty: \sum_tx_t^2<\infty\}$, and define $A_{n,0}x=(0,\ldots,0,x_{n+1},x_{n+2},\ldots)$ for all $n \in \omega$ and $x \in \ell_2$, and $A_{n,k}=0$ whenever $k>0$.
Then $\|A_{n,0}\|=1$ for all $n\in \omega$, and $\lim_n A_{n,0}x=0$ for all $x \in \ell_2$.
\end{rmk}
\begin{rmk}\label{rmk:regulardoesnotimplylinftylinfty}
Another difference from the finite dimensional case is that a $(\mathrm{Fin}, \mathrm{Fin})$-regular matrix
does not necessarily belong to $(\ell_\infty(X),\ell_\infty(Y))$.
For, suppose that $X=Y=\ell_2$,
and let $e_k$ be the $k$th unit vector of $X$ for each $k \in \omega$.
Building on the above example, consider the matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(\ell_2,\ell_2)$ such that $A=\mathrm{Id}+B$, where $\mathrm{Id}$ is the identity matrix and
\begin{displaymath}
\forall n,k \in \omega, \forall x \in \ell_2, \quad
B_{n,k}(x):=
\begin{cases}
\,(0,\ldots,0,x_{n+1},x_{n+2},\ldots)\,\,& \text{if }k=0;\\
\,-x_{n+k}e_{n+k} & \text{if }k>0.
\end{cases}
\end{displaymath}
Then $A$ satisfies conditions \ref{item:S1}-\ref{item:S3} with $T=I$ and $k_0=0$, hence by Theorem \ref{thm:SilvermanToeplizsimple} $A$ is a $(\mathrm{Fin}, \mathrm{Fin})$-regular matrix.
However, the sequence $\bm{x}:=(e_0,e_1,\ldots) \in \ell_\infty(\ell_2)$ does not belong to $\mathrm{dom}(A)$, indeed $A_0\bm{x}=e_0-e_1-e_2-\cdots$ is not norm convergent in $\ell_2$.
Therefore $A\notin (\ell_\infty(X),\ell_\infty(Y))$, cf. Theorem \ref{thm:maddoxmain} below.
\end{rmk}
Our first main result, which corresponds to the operator version of \cite[Theorem 1.2]{ConnorLeo}, follows.
\begin{thm}\label{main:IJREGULAR}
Fix a linear operator $T \in \mathcal{L}(X,Y)$, where $X,Y$ are Banach spaces.
Let also $\mathcal{I}, \mathcal{J}$ be ideals on $\omega$.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$ if and only if
there exists $k_0 \in \omega$ such that\textup{:}
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*})}]
\item \label{item:T1} $\sup_n\|A_{n,\ge k_0}\|<\infty$\textup{;}
\item \label{item:T2}
$\sup_n\|A_{n,k}x\|<\infty$ for all $x \in X$ and $k<k_0$\textup{;}
\item \label{item:T3} $\sum_kA_{n,k}x_k$ converges in the norm of $Y$ for all $\bm{x} \in c^b(X,\mathcal{I})$ and $n \in \omega$\textup{;}
\item \label{item:T4} $\mathcal{J}\text{-}\lim_n \sum_{k}A_{n,k}=T$\textup{;}
\item \label{item:T5} $A \in (c_{00}^b(X,\mathcal{I}), c_0(Y,\mathcal{J}))$\textup{.}
\end{enumerate}
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=0$.
\end{thm}
For the sake of clarity,
condition \ref{item:T4} means that $A_n(x,x,\ldots)$ is norm convergent for all $n \in \omega$ and $x \in X$ (which is weaker than \ref{item:T3}) and, in addition, $\mathcal{J}\text{-}\lim_n A_n(x,x,\ldots)=Tx$ for all $x \in X$.
Lastly, condition \ref{item:T5} can be rephrased as: $A_n\bm{x}$ is norm convergent for all $n \in \omega$ and all bounded sequences $\bm{x}$ supported on $\mathcal{I}$ and, for such sequences, $\mathcal{J}\text{-}\lim A\bm{x}=0$.
\begin{rmk}\label{rmk:conditionT5}
Note that \ref{item:T5} could be replaced also with the stronger condition:
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^\prime$)}]
\setcounter{enumi}{4}
\item \label{item:T5prime} $A \in (c_{00}^b(X,\mathcal{I}), c_0^b(Y,\mathcal{J}))$.
\end{enumerate}
Indeed by Definition \ref{def:mainIJregular} the transformed sequences $A\bm{x}$ is necessarily bounded for all $\bm{x} \in c^b(X,\mathcal{I})$.
Also, the latter condition \ref{item:T5prime} would imply automatically \ref{item:T2}.
Therefore, $A$ is $(\mathcal{I},\mathcal{J})$-regular with respect to $T$ if and only if \ref{item:T1}, \ref{item:T3}, \ref{item:T4}, and \ref{item:T5prime} hold.
However, we chose to state it in the former version for two reasons.
First, if each $A_{n,k}$ is bounded, then \ref{item:T2} is void so that our characterization holds with the weaker condition \ref{item:T5}.
Second, most importantly, condition \ref{item:T5} will be used also in the unbounded analogue given in Theorem \ref{main:IJREGULARboundedUnbounded} below: this allows to highlight the differences between the two cases.
\end{rmk}
Even if Theorem \ref{main:IJREGULAR} may look quite complicated, the reader should keep in mind that it deals with (possibly unbounded) linear operators and general ideal/filter convergence.
We are going to see that, in some special circumstances, it may be considerably simplified because either some of the conditions \ref{item:T1}-\ref{item:T5} are automatically satisfied or the latter ones collapse to simpler properties (in particular, recovering the classical ones).
Several related results
may be found in the literature in the case $X=Y$ equal to $\mathbf{R}$ or $\mathbf{C}$, $T$ equals to the identity operator $I$ or the zero operator, and $\mathcal{I},\mathcal{J}$ being certain $F_{\sigma\delta}$-ideals (where ideals are regarded as subsets of the Cantor space $\{0,1\}^\omega$), see e.g. \cite{MR3511151,
MR1963462, MR3671266, MR3911031, MR1433948}.
We remark also that, if $T$ is not bounded, then an $(\mathcal{I},\mathcal{J})$-regular matrix $A$ with respect to $T$ may not exist: indeed, if each $A_{n,k}$ is bounded and $\mathcal{J}=\mathrm{Fin}$, condition \ref{item:T4} and the Banach--Steinhaus theorem imply that $T$ is necessarily bounded.
In the case that $T=0$, we obtain the following immediate consequence, cf. also Corollary \ref{cor:llinftyc0} below for the finite dimensional case with $\mathcal{I}$ maximal.
\begin{cor}\label{cor:ctoc0}
Let $X,Y$ be Banach spaces, and let also $\mathcal{I}, \mathcal{J}$ be ideals on $\omega$.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ belongs to $(c^b(X,\mathcal{I}), c^b_0(Y,\mathcal{J}))$ if and only if there exists $k_0 \in \omega$ such that \ref{item:T1}-\ref{item:T5} hold, with $T=0$.
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=0$.
\end{cor}
It will be useful to define also the following properties:
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^{\natural}$)}]
\setcounter{enumi}{2}
\item \label{item:T3natural} $\lim_k \|A_{n,\ge k}\|=0$ for all $n\in \omega$;
\end{enumerate}
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^\flat$)}]
\setcounter{enumi}{5}
\item \label{item:T6flat} $\mathcal{J}\text{-}\lim_n \|A_{n,k}\|=0$ for all $k\in \omega$.
\end{enumerate}
It is clear that \ref{item:S3sharp} corresponds to \ref{item:T6flat} in the case $\mathcal{J}=\mathrm{Fin}$.
Some implications between the above-mentioned conditions
are collected below.
\begin{prop}\label{prop:implications}
With the same hypothesis of Theorem \ref{main:IJREGULAR}, the following hold\textup{:}
\begin{enumerate}[label={\rm (\roman*)}]
\item \label{item:1simplification} If $\mathcal{I}=\mathrm{Fin}$ then \ref{item:T1} and \ref{item:T4} imply \ref{item:T3}\textup{;}
\item \label{item:2simplification} If $\mathcal{J}=\mathrm{Fin}$ then \ref{item:T5} implies \ref{item:T2}\textup{;}
\item \label{item:3simplification} \ref{item:T1} and \ref{item:T3natural} imply \ref{item:T3}\textup{;}
\item \label{item:4simplification} If each $A_{n,k}$ is bounded, it is possible to choose $k_0=0$, hence \ref{item:T2} holds\textup{;}
\item \label{item:5simplification} If $\mathrm{dim}(X)<\infty$ each $A_{n,k}$ is bounded. Moreover, \ref{item:T1} implies \ref{item:T3natural}\textup{;}
\item \label{item:6simplification} If $\mathrm{dim}(X)<\infty$ and $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$, then \ref{item:T6flat} holds\textup{;}
\item \label{item:7simplification} If each $A_{n,k}$ is a multiple of $A_0 \in \mathcal{L}(X,Y)$, then \ref{item:T1} implies \ref{item:T3natural} and $A_0\in \mathcal{B}(X,Y)$\textup{;}
\item \label{item:8simplification} If each $A_{n,k}$ is a multiple of $A_0 \in \mathcal{L}(X,Y)$, then \ref{item:T1} and \ref{item:T5} imply \ref{item:T6flat}\textup{.}
\end{enumerate}
\end{prop}
It is immediate to check that Theorem \ref{thm:originalIJREGULAR} comes as a corollary, putting together Theorem \ref{main:IJREGULAR} and Proposition \ref{prop:implications}.\ref{item:1simplification} and \ref{item:2simplification}.
However, the
usefulness of a characterization of $(\mathcal{I}, \mathcal{J})$-regular matrices with respect to $T$ comes from the practical easiness to check whether conditions \ref{item:T1}-\ref{item:T5} hold together.
Taking into account the implications given in Proposition \ref{prop:implications}, it is evident that \ref{item:T5} is the most demanding in this direction.
Hence it makes sense to search for sufficient conditions which allows us to simplify it.
In the same spirit of \cite[Theorem 1.3]{ConnorLeo}, which studies the case $X=Y=\mathbf{R}$ and $T=I$, we obtain
characterizations of
such matrices
which avoid condition \ref{item:T5}.
We will need the new and much \textquotedblleft easier\textquotedblright\,condition:
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*})}]
\setcounter{enumi}{5}
\item \label{item:T6} $\mathcal{J}\text{-}\lim_n \|A_{n,E}\|=0$ for all $E \in \mathcal{I}$.
\end{enumerate}
Directly by the definition of group norm, it is clear that \ref{item:T6} implies \ref{item:T5} (and also \ref{item:T6flat}).
This means that we are allowed to replaced \ref{item:T5} with the stronger condition \ref{item:T6} provided that the latter is satisfied for matrices $A$ which are $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$, possibly under some additional constraints.
\begin{thm}\label{thm:JfinIJregularmain}
With the same hypotheses of Theorem \ref{main:IJREGULAR}, suppose, in addition, that $\mathcal{J}$ is countably generated and that conditions \ref{item:T3natural} and \ref{item:T6flat} hold.
Then $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$ if and only if there exists $k_0 \in \omega$ such that \ref{item:T1}, \ref{item:T4}, and \ref{item:T6} hold.
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=0$.
\end{thm}
On a similar direction, recall that, if $X,Y$ are vector lattices, then a linear operator $T \in \mathcal{L}(X,Y)$ is said to be \emph{positive} if
$Tx\ge 0$
whenever $x\ge 0$.
In addition, a Banach space $V$ is called an \emph{AM-space} if $V$ is also a vector lattice such that $0\le x\le y$ implies $\|x\| \le \|y\|$, and $\|x \vee y\|=\max\{\|x\|,\|y\|\}$ for all $x,y \ge 0$;
we say that $e \in V$ is an \emph{order unit} if, for all $x \in V$ there exists $n \in \omega$ such that $-ne \le x\le ne$.
Accordingly, if $V\neq \{0\}$, then necessarily $e>0$.
Examples of AM-spaces with order units include $\ell_\infty$ and $C(K)$ spaces, for some compact Hausdorff space $K$.
We refer to \cite{MR2011364, MR2262133} for the underlying theory on vector lattices.
\begin{thm}\label{thm:AMspaceregularmain}
With the same hypotheses of Theorem \ref{main:IJREGULAR}, suppose, in addition, that $X$ is an AM-space with order unit $e$, $Y$ is a Banach lattice, each $A_{n,k}$ is a positive linear operator, and that condition \ref{item:T3natural} holds.
Then $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$ if and only if \ref{item:T1}, \ref{item:T4}, and \ref{item:T6} hold with $k_0=0$.
\end{thm}
In the finite dimensional case, everything is simpler. Indeed, suppose that $X=\mathbf{R}^d$ and $Y=\mathbf{R}^m$, for some integers $d,m\ge 1$. Then each linear operator $A_{n,k}$ is represented by the real matrix $[\,a_{n,k}(i,j): 1\le i\le m, 1\le j\le d\,]$ and $T$ is represented by the real matrix $[\,t(i,j): 1\le i\le m, 1\le j\le d\,]$ .
\begin{cor}\label{cor:finitedimensionmain}
With the same hypotheses of Theorem \ref{main:IJREGULAR}, suppose that $X=\mathbf{R}^d$, $Y=\mathbf{R}^m$, and that $\mathcal{I}=\mathrm{Fin}$ or $\mathcal{J}$ is countably generated or $a_{n,k}(i,j)\ge 0$ for all $1\le i \le m$, $1\le j\le d$, and $n,k \in \omega$.
Then $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$ if and only if\textup{:}
\begin{enumerate}[label={\rm (\textsc{F}\arabic{*})}]
\item \label{item:F1} $\sup_n \sum_k\sum_{i,j}\left|a_{n,k}(i,j)\right|<\infty$\textup{;}
\item \label{item:F4} $\mathcal{J}\text{-}\lim_n \sum_k a_{n,k}(i,j)=t(i,j)$ for all $1\le i \le m$ and $1\le j\le d$\textup{;}
\item \label{item:F6} $\mathcal{J}\text{-}\lim_n \sum_{k\in E}\sum_{i,j}\left|a_{n,k}(i,j)\right|=0$ for all $E \in \mathcal{I}$\textup{.}
\end{enumerate}
\end{cor}
It is remarkable that the \textquotedblleft easier\textquotedblright\,characterization with condition \ref{item:F6} does \emph{not} hold uniformly for all ideals $\mathcal{I}, \mathcal{J}$: indeed, it has been proved in \cite[Theorem 1.4]{ConnorLeo} that, even in the simplest case $X=Y=\mathbf{R}$ and $T=I$, there exists a $(\mathcal{Z}, \mathcal{Z})$-regular matrix which does not satisfy \ref{item:F6}, where $\mathcal{Z}$ is the asymptotic density zero ideal defined in \eqref{eq:definitionZ}.
In addition, condition \ref{item:F6} can be simplified if $T$ is the zero operator and $\mathcal{I}$ a maximal ideal:
\begin{cor}\label{cor:llinftyc0}
With the same hypotheses of Corollary \ref{cor:finitedimensionmain}, $A \in (\ell_\infty(\mathbf{R}^d), c_0^b(\mathbf{R}^m, \mathcal{J}))$ if and only if condition \ref{item:F1} holds, together with:
\begin{enumerate}[label={\rm (\textsc{F}\arabic{*}$^\prime$)}]
\setcounter{enumi}{2}
\item \label{item:F6prime} $\mathcal{J}\text{-}\lim_n \sum_{k}\sum_{i,j}\left|a_{n,k}(i,j)\right|=0$.
\end{enumerate}
\end{cor}
This provides a generalization of \cite[Lemma 3.2]{MR2209588} in the case $d=m=1$, $T=0$, and $\mathcal{J}$ equal to the countably generated ideal $\mathcal{I}_{\mathrm{P}}$ defined below in \eqref{eq:definitionIP}.
A similar result can be obtained if each $A_{n,k}$ is a multiple of a given linear operator:
\begin{cor}\label{cor:eachAnkmultiple}
With the same hypotheses of Theorem \ref{main:IJREGULAR}, suppose that each $A_{n,k}$ is a multiple of a nonzero $A_0 \in \mathcal{L}(X,Y)$.
In addition, assume that $\mathcal{I}=\mathrm{Fin}$, or $\mathcal{J}$ is countably generated, or that $X$ is an AM-space with order unit $e$, $Y$ is a Banach lattice, and $a_{n,k}\ge 0$ for all $n,k \in \omega$.
Then $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$ if and only if\textup{:}
\begin{enumerate}[label={\rm (\textsc{M}\arabic{*})}]
\item \label{item:M0} $A_0 \in \mathcal{B}(X,Y)$\textup{;}
\item \label{item:M1} $\sup_n \sum_k\left|a_{n,k}\right|<\infty$\textup{;}
\item \label{item:M4} $T=\kappa A_0$, with $\kappa=\mathcal{J}\text{-}\lim_n \sum_k a_{n,k}$\textup{;}
\item \label{item:M6} $\mathcal{J}\text{-}\lim_n \sum_{k\in E}\left|a_{n,k}\right|=0$ for all $E \in \mathcal{I}$\textup{.}
\end{enumerate}
\end{cor}
In particular, under the hypotheses of Corollary \ref{cor:eachAnkmultiple},
if $A$ is $(\mathcal{I},\mathcal{J})$-regular with respect to a linear operator $T$, then $T$ is necessarily bounded.
In the following sections, we obtain the analogues of Theorem \ref{main:IJREGULAR} where we replace the bounded sequence spaces $c^b(X,\mathcal{I})$ and $c^b(Y,\mathcal{J})$ of Definition \ref{def:mainIJregular} with their unbounded versions $c(X,\mathcal{I})$ and $c(Y,\mathcal{J})$.
\subsection{Bounded to Unbounded case}
An ideal $\mathcal{J}$ on $\omega$ is said to be a \emph{rapid}$^+$\emph{-ideal} if,
for every $S \in \mathcal{I}^+$ and $F \in \mathrm{Fin}^+$,
there exists $S^\prime \subseteq S$ such that $S^\prime \in \mathcal{J}^+$ and $|S^\prime \cap [0,n]|\le |F \cap [0,n]|$ for all $n\in \omega$.
Moreover, $\mathcal{J}$ is called a $P^+$\emph{-ideal} if, for every decreasing sequence $(S_n)$ in $\mathcal{J}^+$, there exists $S \in \mathcal{J}^+$ such that $S\setminus S_n$ is finite for all $n \in \omega$. The class of rapid$^+$ and $P^+$-ideals have been studied also, e.g., in \cite{MR4172859, MR2876731, MR2777744, MR1649074}.
The ideal $\mathcal{J}$ is said to be \emph{countably generated} if there exists a sequence $(Q_j)$ of subsets of $\omega$ such that $S \in \mathcal{J}$ if and only if $S\subseteq \bigcup_{j \in F}Q_j$ for some $F \in \mathrm{Fin}$.
As it has been observed in \cite[Section 2]{MR4172859}, every countably generated ideal is a rapid$^+$ $P^+$-ideal.
In particular, $\mathrm{Fin}$ is a rapid $^+$ $P^+$-ideal.
It is a known fact that every $F_\sigma$-ideal is a $P^+$-ideal, see e.g. \cite{MR748847}.
However, the summable ideal $\mathcal{I}_{1/n}:=\{S\subseteq \omega: \sum_{n \in S}1/(n+1)<\infty\}$ is an $F_\sigma$-ideal which is not rapid$^+$, and $\mathcal{Z}$ is neither a rapid$^+$ nor $P^+$-ideal.
This does not mean that the topological complexity of rapid$^+$ $P^+$-ideals is low: indeed, under Martin's axiom for countable posets, there exist uncountably many nonisomorphic maximal ideals with such property (on the other hand, their existence is not provable in ZFC), see \cite[Section 5.1]{MR4172859}.
Lastly, we need to following weakening of condition \ref{item:T1} (and they coincide if $\mathcal{J}=\mathrm{Fin}$):
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^\flat$)}]
\setcounter{enumi}{0}
\item \label{item:T1flat} There exists $J_0 \in \mathcal{J}^\star$
for which $\sup_{n \in J_0}\|A_{n,\ge k_0}\|<\infty$ and, for all $n \in \omega\setminus J_0$, there exists $f(n) \in \omega$ such that $\|A_{n,\ge f(n)}\|<\infty$\textup{.}
\end{enumerate}
This condition has been suggested by the example given in \cite[Section 4]{ConnorLeo}.
With these premises, we state the analogue of Theorem \ref{main:IJREGULAR} for the unbounded codomain sequence spaces.
\begin{thm}\label{main:IJREGULARboundedUnbounded}
Fix a linear operator $T \in \mathcal{L}(X,Y)$, where $X,Y$ are Banach spaces.
Let also $\mathcal{I},\mathcal{J}$ be an ideals on $\omega$ such that $\mathcal{J}$ is a rapid$^+$ $P^+$-ideal.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ satisfies
\begin{equation}\label{eq:boundedunboundeddefinition}
A \in (c^b(X,\mathcal{I}),c(Y,\mathcal{J}))
\,\,\, \text{ and }\,\,\,
\mathcal{J}\text{-}\lim A\bm{x}=T(\,\mathcal{I}\text{-}\lim \bm{x})
\,\text{ for all }\bm{x} \in c^b(X,\mathcal{I}).
\end{equation}
if and only if
there exists $k_0 \in \omega$ such that \ref{item:T1flat}, \ref{item:T3}, \ref{item:T4}, and \ref{item:T5} hold.
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=f(n)=0$ for all $n\in \omega\setminus J_0$.
\end{thm}
During the proof, we will need an ideal version of the Banach--Steinhaus theorem which has been recently proved in \cite{MR4172859}, see Theorem \ref{thm:uniformJboundedness} below.
Interestingly, the latter result provides a characterization of rapid$^+$ $P^+$-ideals \cite[Theorem 5.1]{MR4172859}, which suggests that Theorem \ref{main:IJREGULARboundedUnbounded} cannot be improved with the current techniques.
We remark that Theorem \ref{main:IJREGULARboundedUnbounded} sheds light on the substantial difference between the classical Silverman--Toeplitz characterization stated in Theorem \ref{thm:SilvermanToeplizsimple}
and its \textquotedblleft multidimensional\textquotedblright\,analogue proved by Robinson \cite{Robinson26} and Hamilton \cite{MR1545904} for double sequences, namely, the weakening of \ref{item:T1} to \ref{item:T1flat}.
For, recall that a double sequence $(x_{m,n}: m,n \in \omega)$ has \emph{Pringsheim limit} $\eta \in X$, shortened as $\mathrm{P}\text{-}\lim_{m,n}x_{m,n}=\eta$, if for all $\varepsilon>0$ there exists $k \in \omega$ such that $\|x_{m,n}-\eta\|<\varepsilon$ for all $m,n \ge k$.
At this point, define the ideal
\begin{equation}\label{eq:definitionIP}
\mathcal{I}_{\mathrm{P}}:=\left\{S\subseteq \omega: \sup\nolimits_{n \in S}\nu_2(n)<\infty\right\},
\end{equation}
where $\nu_2$ is the $2$-adic valution defined by $\nu_2(0):=0$ and $\nu_2(n):=\max\{k \in \omega: 2^k \text{ divides }n\}$ if $n>0$.
Note that the ideal $\mathcal{I}_{\mathrm{P}}$ is countably generated by the sequence of sets $(Q_t: t \in \omega)$, where $Q_t:=\left\{S\subseteq \omega: \sup\nolimits_{n \in S}\nu_2(n)=t\right\}$ for all $t \in \omega$.
Hence $\mathcal{I}_{\mathrm{P}}$ is a rapid$^+$ $P^+$-ideal.
Let also $h: \omega^2\to \omega$ be an arbitrary bijection with the property that $h[\{(m,n) \in \omega^2: \min\{m,n\}=k\}]=Q_k$ for all $k \in \omega$.
Thus, we obtain
\begin{equation}\label{eq:equivalencepringhsheim}
\mathrm{P}\text{-}\lim\nolimits_{m,n}x_{m,n}=\eta
\quad \text{ if and only if }\quad
\mathcal{I}_{\mathrm{P}}\text{-}\lim\nolimits_n x_{h^{-1}(n)}=\eta,
\end{equation}
as it has been observed in \cite[Section 4.2]{MR3955010}, cf. also \cite{MR3671266}.
In other words, $\mathcal{I}_{\mathrm{P}}$ is an isomorphic copy on $\omega$ of the ideal on $\omega^2$ generated by vertical lines and horizontal lines, cf. Remark \ref{rmk:farahisomorphiccountablygeneratedideals} below.
Relying on the equivalence \eqref{eq:equivalencepringhsheim}, the classical definition of \emph{RH-regular matrix} $A$ coincides with \eqref{eq:boundedunboundeddefinition} in the case $X=Y=\mathbf{R}$, $T=I$, and $\mathcal{I}=\mathcal{J}=\mathcal{I}_{\mathrm{P}}$.
With the same notations of Corollary \ref{cor:finitedimensionmain}, we can state the following consequence in the finite dimensional case.
\begin{cor}\label{cor:RHgeneral}
Suppose that $X=\mathbf{R}^d$, $Y=\mathbf{R}^m$, and let $\mathcal{I}$, $\mathcal{J}$ be ideals on $\omega$ such that
$\mathcal{J}$
is countably generated by a sequence of sets $(Q_t: t \in \omega)$.
Then a matrix $A$ satisfies \eqref{eq:boundedunboundeddefinition} if and only if there exists $t_0 \in \omega$ such that \textup{:}
\begin{enumerate}[label={\rm (\textsc{R}\arabic{*})}]
\item \label{item:R1} $\sup_{n \in \omega\setminus Q_{t_0}} \sum_k\sum_{i,j}\left|a_{n,k}(i,j)\right|<\infty$\textup{;}
\item \label{item:R2} $\sum_k\left|a_{n,k}(i,j)\right|<\infty$ for all $n \in Q_{t_0}$, $1\le i\le m$, and $1\le j\le d$\textup{;}
\item \label{item:R4} $\mathcal{J}\text{-}\lim_n \sum_k a_{n,k}(i,j)=t(i,j)$ for all $1\le i \le m$ and $1\le j\le d$\textup{;}
\item \label{item:R6} $\mathcal{J}\text{-}\lim_n \sum_{k\in E}\sum_{i,j}\left|a_{n,k}(i,j)\right|=0$ for all $E \in \mathcal{I}$\textup{.}
\end{enumerate}
\end{cor}
It is clear that, if also $\mathcal{I}$ is countably generated by a sequence $(E_t: t \in \omega)$, which is the case
of RH-regular matrices,
then \ref{item:R6} can be rewritten as:
\begin{enumerate}[label={\rm (\textsc{R}\arabic{*}$^\prime$)}]
\setcounter{enumi}{3}
\item $\mathcal{J}\text{-}\lim_n \sum_{k\in E_t}\sum_{i,j}\left|a_{n,k}(i,j)\right|=0$ for all $t \in \omega$\textup{.}
\end{enumerate}
Another special instance of Corollary \ref{cor:RHgeneral} has been proved in \cite[Theorem 5]{MR3466559} for the case where $X=Y=\mathbf{R}$, $T=I$, $A$ is a RH-regular matrix with nonnegative real entries, $\mathcal{I}$ is a $P$-ideal (that is, if $(S_n)$ is an increasing sequence in $\mathcal{I}$, there exists $S \in \mathcal{I}$ such that $S_n\setminus S \in \mathrm{Fin}$ for all $n$), and $\mathcal{J}=\mathrm{Fin}$.
Other consequences of Theorem \ref{main:IJREGULARboundedUnbounded}, in the same vein of the ones given in Section \ref{subsec:boundedbounded}, may be obtained here, and they are left to the reader.
\begin{rmk}\label{rmk:farahisomorphiccountablygeneratedideals}
An ideal $\mathcal{I}$ on $\omega$ is countably generated if and only if it is isomorphic to one of the following:
\begin{enumerate}[label={\rm (\roman*)}]
\item \label{item:1countablygenerated} $\mathrm{Fin}$;
\item \label{item:2countablygenerated} $\mathrm{Fin}\times \emptyset:=\{S\subseteq \omega^2: \exists n \in \omega, S\subseteq [0,n]\times \omega\}$;
\item \label{item:3countablygenerated} $\mathrm{Fin}\oplus \mathcal{P}(\omega):=\{S\subseteq \{0,1\}\times \omega: |S\cap (\{0\}\times \omega)|<\infty\}$.
\end{enumerate}
(Recall that two ideals $\mathcal{I}_1$ and $\mathcal{I}_2$ on countable sets $H_1$ and $H_2$, respectively, are called \emph{isomorphic}, written as $\mathcal{I}_1 \simeq \mathcal{I}_2$ if there exists a bijection $h: H_1\to H_2$ such that $h[S] \in \mathcal{I}_2$ if and only if $S \in \mathcal{I}_1$ for all $S\subseteq H_1$; accordingly, it is easy to see that $\mathcal{I}_{\mathrm{P}} \simeq \mathrm{Fin}\times \emptyset$, and that the ideals in \ref{item:1countablygenerated}-\ref{item:3countablygenerated} are pairwise nonisomorphic.)
This has been essentially proved in \cite[Proposition 1.2.8]{MR1711328}, however the correct statement appears
in \cite[Section 2]{MR3543775}.
We include its simple proof for the sake of completeness. Suppose that $\mathcal{I}$ is countably generated by a partition $\{Q_j: j \in\omega\}$ of $\omega$, and define $J:=\{j \in \omega: Q_j \in \mathrm{Fin}\}$.
If $J$ is empty then every $Q_j$ is infinite, hence $\mathcal{I}\simeq \mathrm{Fin}\times \emptyset$. If $J$ is nonempty finite then $Q_j$ is infinite for infinitely many $j\in \omega$, hence $\mathcal{I}\simeq \mathcal{J}$, where
$$
\mathcal{J}:=\mathrm{Fin}\oplus (\mathrm{Fin}\times \emptyset):=\{S\subseteq \omega \cup \omega^2: S\cap \omega \in \mathrm{Fin}, S\cap \omega^2 \in \mathrm{Fin}\times \emptyset\}.
$$
However, $\mathcal{J}\simeq \mathrm{Fin}\times \emptyset$, with the witnessing bijection $h: \omega\cup \omega^2\to \omega^2$ defined by $h(a,b)=(a,b+1)$ and $h(a)=(a,0)$ for all $a,b \in \omega$. Hence, let us assume hereafter that $J$ is infinite. If $J^c$ is empty then $\mathcal{I}\simeq \mathrm{Fin}$. If $J^c$ is nonempty finite then $\mathcal{I}\simeq \mathrm{Fin}\oplus \mathcal{P}(\omega)$. Lastly, if also $J^c$ is infinite, then $\mathcal{I}\simeq \mathcal{J}\simeq \mathrm{Fin}\times \emptyset$.
\end{rmk}
\subsection{Unbounded to Bounded case}
In this section, we may assume that $\mathcal{I}\neq \mathrm{Fin}$, otherwise we go back in the previous cases.
Differently from the other results, we are going to show that, quite often, there are no matrices $A$
which satisfy
\begin{equation}\label{eq:boundedunboundeddefinition}
A \in (c(X,\mathcal{I}),c^b(Y,\mathcal{J}))
\,\,\, \text{ and }\,\,\,
\mathcal{J}\text{-}\lim A\bm{x}=T(\,\mathcal{I}\text{-}\lim \bm{x})
\,\text{ for all }\bm{x} \in c(X,\mathcal{I}),
\end{equation}
unless $T$ is the zero operator. To this aim, recall that an ideal $\mathcal{I}$ is said to be \emph{tall} if, for every infinite set $S\subseteq \omega$, there exists an infinite subset $S^\prime \subseteq S$ which belongs to $\mathcal{I}$ (note that countably generated ideals, hence also $\mathrm{Fin}$, are not tall; for a necessary condition in the case of countably generated ideals $\mathcal{J}$ and arbitrary $\mathcal{I}$, cf. Remark \ref{rmk:necessryT5sharp} below).
\begin{thm}\label{main:IJREGULARUnboundedBounded}
Fix a nonzero linear operator $T \in \mathcal{L}(X,Y)$, where $X,Y$ are Banach spaces.
Let also $\mathcal{I},\mathcal{J}$ be an ideals on $\omega$ such that $\mathcal{I}$ is tall.
Then there are no matrices $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ which satisfy \eqref{eq:boundedunboundeddefinition}.
\end{thm}
Of course, if $T=0$, then the zero matrix $A$ (namely, the matrix with $A_{n,k}=0$ for all $n,k$) satisfies \eqref{eq:boundedunboundeddefinition}.
However, this is essentially the unique possibility:
\begin{thm}\label{thm:cIc_0bJ}
Let $X,Y$ be Banach spaces. Let also $\mathcal{I}, \mathcal{J}$ be ideals on $\omega$ such that $\mathcal{I}$ is tall.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ belongs to $(c(X,\mathcal{I}), c_0^b(Y,\mathcal{J}))$ if and only if there exists $k_1 \in \omega$ such that\textup{:}
\begin{enumerate}[label={\rm (\textsc{B}\arabic{*})}]
\item \label{item:B1} $A_{n,k}=0$ for all $n \in \omega$ and $k\ge k_1$\textup{;}
\item \label{item:B2} $\sup_n \|A_{n,k}x\|<\infty$ for all $x \in X$ and $k<k_1$\textup{;}
\item \label{item:B3} $\mathcal{J}\text{-}\lim_n A_{n,k}=0$ for all $k<k_1$\textup{.}
\end{enumerate}
\end{thm}
As it will turn out, condition \ref{item:B1} is satisfied also for all matrices in the larger class $(c_{00}(X,\mathcal{I}), \ell_\infty(Y))$, provided that $\mathcal{I}$ is tall.
\subsection{Unbounded to Unbounded case}
In this last section, we study the analogue condition for a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ to satisfy
\begin{equation}\label{eq:unoundedunboundeddefinition}
A \in (c(X,\mathcal{I}),c(Y,\mathcal{J}))
\,\,\, \text{ and }\,\,\,
\mathcal{J}\text{-}\lim A\bm{x}=T(\,\mathcal{I}\text{-}\lim \bm{x})
\,\text{ for all }\bm{x} \in c(X,\mathcal{I}).
\end{equation}
\begin{rmk}\label{rmk:existence}
In some cases, it is easy to provide examples of matrices which satisfy \eqref{eq:unoundedunboundeddefinition}.
Indeed, suppose that $T\in \mathcal{B}(X,Y)$ and $\mathcal{I}\subseteq \mathcal{J}$. We claim that the matrix $A=(A_{n,k})$ such that $A_{n,k}=T$ if $n=k$ and $A_{n,k}=0$ otherwise has this property.
For, set $\bm{x} \in c(X,\mathcal{I})$ with $\mathcal{I}$-limit $\eta$.
Then
$
\mathcal{I}\text{-}\lim\nolimits_n A_n\bm{x}=\mathcal{I}\text{-}\lim\nolimits_n Tx_n=T(\mathcal{I}\text{-}\lim \bm{x})=T\eta,
$
which implies $\mathcal{J}\text{-}\lim A\bm{x}=T\eta$. Here, we used the fact the $T$ preserves $\mathcal{I}$-convergence: this is clear if $T=0$, otherwise $\{n \in\omega: \|Tx_n-T\eta\|<\varepsilon\}\supseteq \{n \in \omega: \|x_n-\eta\|<\varepsilon/\|T\|\} \in \mathcal{I}^\star$ for all $\varepsilon>0$.
\end{rmk}
In the next results, we will need a further weakening of \ref{item:T1flat} and stronger versions of conditions \ref{item:T3} and \ref{item:T5}, that is,
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^{\flat\flat}$)}]
\setcounter{enumi}{0}
\item \label{item:T1flatflat} For all $n \in \omega$, there exists $f(n) \in \omega$ such that $\|A_{n,\ge f(n)}\|<\infty$\textup{;}
\end{enumerate}
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^{\sharp}$)}]
\setcounter{enumi}{2}
\item \label{item:T3sharp} $\sum\nolimits_kA_{n,k}x_k$ converges in the norm of $Y$ for all $\bm{x} \in c(X,\mathcal{I})$ and $n \in \omega$\textup{;}
\end{enumerate}
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^{\sharp}$)}]
\setcounter{enumi}{4}
\item \label{item:T5sharp} $A \in (c_{00}(X,\mathcal{I}),c_0(Y,\mathcal{J}))$\textup{.}
\end{enumerate}
\begin{thm}\label{main:IJREGULARUnboundedUnbounded}
Fix a linear operator $T \in \mathcal{L}(X,Y)$, where $X,Y$ are Banach spaces.
Let also $\mathcal{I},\mathcal{J}$ be an ideals on $\omega$.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ satisfies \eqref{eq:unoundedunboundeddefinition}
if
there exists $k_0 \in \omega$ such that \ref{item:T1flat}, \ref{item:T3sharp}, \ref{item:T4}, and \ref{item:T5sharp} hold.
Conversely, if $A$ satisfies \eqref{eq:unoundedunboundeddefinition} then
\ref{item:T1flatflat}, \ref{item:T3sharp}, \ref{item:T4}, and \ref{item:T5sharp} hold.
\end{thm}
It turns out that we obtain a complete characterization if $\mathcal{J}$ is a rapid$^+$ $P^+$-ideal:
\begin{thm}\label{thm:cXcYrapid}
Fix a linear operator $T \in \mathcal{L}(X,Y)$, where $X,Y$ are Banach spaces.
Let also $\mathcal{I},\mathcal{J}$ be an ideals on $\omega$ such that $\mathcal{J}$ is a rapid$^+$ $P^+$-ideal.
Then a matrix $A=(A_{n,k})$ of linear operators in $\mathcal{L}(X,Y)$ satisfies \eqref{eq:unoundedunboundeddefinition}
if and only if
there exists $k_0 \in \omega$ such that \ref{item:T1flat}, \ref{item:T3sharp}, \ref{item:T4}, and \ref{item:T5sharp} hold.
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=f(n)=0$ for all
$n \in \omega\setminus J_0$.
\end{thm}
Some additional properties can be obtained in special cases:
\begin{rmk}
Suppose that $\mathcal{I}$ is tall ideal. Then \ref{item:T3sharp} implies, thanks to Lemma \ref{lem:dualc00I}, that $A$ is \emph{row finite}, namely, $\{k \in \omega: A_{n,k}\neq 0\} \in \mathrm{Fin}$ for all $n\in \omega$.
\end{rmk}
\begin{rmk}\label{rmk:necessryT5sharp}
Suppose that $\mathcal{J}$ is a countably generated ideal. Then \ref{item:T5sharp} implies, thanks to Theorem \ref{thm:Xomegaellinfty}, that
for every infinite $E \in \mathcal{I}$ there exists $J \in \mathcal{J}^\star$ such that $\{k \in E: A_{n,k}\neq 0 \text{ for some }n \in J\}$ is finite.
\end{rmk}
\section{Preliminaries}\label{sec:preliminaries}
Unless otherwise stated,
we assume that $X,Y$ are Banach spaces.
We recall the following results on the so-called K\"othe--Toeplitz $\beta$-duals:
\begin{lem}\label{lem:convergenceoperator}
Let $(T_k)$ be a sequence of linear operators in $\mathcal{L}(X,Y)$. Then $\sum_k T_kx_k$ is convergent in the norm of $Y$ for all sequences $\bm{x} \in \ell_\infty(X)$ if and only if\textup{:}
\begin{enumerate}[label={\rm (\textsc{N}\arabic{*})}]
\item \label{item:N2} $\|T_{\ge k_0}\|<\infty$ for some $k_0 \in \omega$\textup{;}
\item \label{item:N3} $\lim_k\|T_{\ge k}\|=0$\textup{.}
\end{enumerate}
In addition, if each $T_k$ is bounded, it is possible to choose $k_0=0$.
\end{lem}
\begin{proof}
See \cite[Proposition 3.1 and Proposition 3.3]{MR568707}.
\end{proof}
In particular, \ref{item:N2} implies that $T_k$ is bounded for all $k\ge k_0$.
We remark that, if $X=\mathbf{R}$ and each linear operator $T_k$ can be written as $T_kx=xy_k$, for some $y_k \in Y$, then the sequence $(T_k)$ of Lemma \ref{lem:convergenceoperator} is also called \textquotedblleft bounded multiplier convergent,\textquotedblright\, see e.g. \cite{MR704294}.
\begin{lem}\label{lem:convergenceoperatorbetadualc(X)}
Let $(T_k)$ be a sequence of linear operators in $\mathcal{L}(X,Y)$. Then $\sum_k T_kx_k$ is norm convergent in $Y$ for all $\bm{x} \in c(X)$ if and only if
\ref{item:N2} holds for some $k_0$, together with\textup{:}
\begin{enumerate}[label={\rm (\textsc{N}\arabic{*}$^\prime$)}]
\setcounter{enumi}{1}
\item \label{item:N2prime} $\sum_kT_k$ converges in the strong operator topology\textup{.}
\end{enumerate}
In addition, if each $T_k$ is bounded, it is possible to choose $k_0=0$.
\end{lem}
\begin{proof}
See \cite[Proposition 3.2]{MR568707}.
\end{proof}
A characterization of the K\"othe--Toeplitz $\beta$-dual of a sequence space which is strictly related to $c(\mathcal{Z})$ can be found in \cite[Theorem 4]{MR261212}.
However, if $X$ is finite dimensional, we have a simpler characterization:
\begin{cor}\label{cor:finitedimensionalVSextremepoints}
Let $(T_k)$ be a sequence of linear operators in $\mathcal{L}(X,Y)$ and assume, in addition, that $X$ is finite dimensional.
Then the following are equivalent:
\begin{enumerate}[label={\rm (\roman*)}]
\item \label{item:1corollaryfinitedim} $\sum_kT_kx_k$ is norm convergent in $Y$ for all sequences $\bm{x} \in \ell_\infty(X)$\textup{;}
\item \label{item:2corollaryfinitedim} $\sum_kT_kx_k$ is norm convergent in $Y$ for all sequences $\bm{x} \in c(X)$\textup{;}
\item \label{item:3corollaryfinitedim} $\|T_\omega\|<\infty$\textup{.}
\end{enumerate}
\end{cor}
\begin{proof}
The implication \ref{item:1corollaryfinitedim} $\implies$ \ref{item:2corollaryfinitedim} is clear and \ref{item:2corollaryfinitedim} $\implies$ \ref{item:3corollaryfinitedim} follows by Lemma \ref{lem:convergenceoperatorbetadualc(X)}. Indeed, since $d:=\mathrm{dim}(X)<\infty$, each $T_{k}$ is bounded.
\ref{item:3corollaryfinitedim} $\implies$ \ref{item:1corollaryfinitedim} It follows by Lemma \ref{lem:convergenceoperator} that it is enough to prove that condition \ref{item:N3} holds, provided that $\|T_\omega\|<\infty$.
To this aim, suppose for the sake of contradiction that there exists $\varepsilon>0$ such that $\limsup_k \|T_{\ge k}\|>\varepsilon$.
Then there exist a sequence $\bm{x}$ taking values in the closed unit ball $B_X$ and a partition $\{I_j: j \in \omega\}$ of $\omega$ in consecutive finite intervals such that
$$
\forall j \in \omega, \quad
\|(T_k: k \in I_j)\|\ge \left\|\sum\nolimits_{k \in I_j}T_kx_k\right\|>\varepsilon.
$$
Assume without loss of generality that $X=\mathbf{R}^d$ and, since every norm is equivalent, endow it with the $1$-norm $\|x\|:=\sum_i|x_i|$.
Define the sequence $\bm{y}$ by $y_j:=\sum\nolimits_{k \in I_j}T_kx_k$ for all $j$.
Let $\{Q_1,\ldots,Q_{2^d}\}$ be the collection of all closed quadrants of $\mathbf{R}^d$.
Since $\bigcup_{i\le 2^d}Q_i=\mathbf{R}^d$, there exist $i_0 \in [1,2^d]$ and an infinite set $J\subseteq \omega$ such that $y_j \in Q_{i_0}$ for all $j \in J$.
It follows that
\begin{displaymath}
\begin{split}
\|T_\omega\|\ge \|(T_k: k \in J)\| &
\ge \sup_{F \subseteq J, F \in \mathrm{Fin}}\left\|\sum\nolimits_{j \in F}y_j\right\| \\
&=\sup_{F \subseteq J, F \in \mathrm{Fin}} \sum\nolimits_{j \in F}\left\|y_j\right\|
\ge \sup_{F \subseteq J, F \in \mathrm{Fin}}|F|\varepsilon =\infty,
\end{split}
\end{displaymath}
which contradicts the standing hypothesis.
\end{proof}
\begin{lem}\label{lem:finitedimensionalnormkjfdhgd}
Let $(T_k)$ be a sequence of linear operators in $\mathcal{L}(X,Y)$ and assume, in addition, that $X$ is finite dimensional and $\mathcal{J}\text{-}\lim_k \|T_kx\|=0$ for all $x \in X$, where $\mathcal{J}$ is an ideal on $\omega$. Then $\mathcal{J}\text{-}\lim_k \|T_k\|=0$.
\end{lem}
\begin{proof}
Since $d:=\mathrm{dim}(X)<\infty$, each $T_k$ is bounded.
Assume without loss of generality that $X=\mathbf{R}^d$ and endow it with the $1$-norm as in the proof of Corollary \ref{cor:finitedimensionalVSextremepoints}.
Note that the set $\mathscr{E}$ of extreme points of the closed unit ball is finite.
Hence, for each $k \in \omega$, there exists $e_k \in \mathscr{E}$ such that $\|T_k\|=\|T_ke_k\|$.
It follows that
$$
\mathcal{J}\text{-}\lim\nolimits_k \|T_k\|\le
\sum\nolimits_{e \in \mathscr{E}}\mathcal{J}\text{-}\lim\nolimits_k \|T_ke\|=0,
$$
which completes the proof.
\end{proof}
As anticipated, we will need an ideal version of the Banach--Steinhaus theorem, which has been recently obtained in \cite{MR4172859}.
\begin{thm}\label{thm:uniformJboundedness}
Let $\mathcal{J}$ be a rapid$^+$ $P^+$-ideal on $\omega$. Also, let $(T_n)$ be a sequence of linear operators in $B(X,Y)$ and suppose that
$$
\forall x\in X, \quad
\mathcal{J}\text{-}\limsup\nolimits_n \|T_nx\|<\infty.
$$
Then $\mathcal{J}\text{-}\limsup_n\|T_n\|<\infty$.
\end{thm}
\begin{proof}
It follows by \cite[Theorem 3.1(b)]{MR4172859}.
\end{proof}
The following result on unbounded operators is due to Lorentz and Macphail \cite{MR52533} in the case $\mathcal{J}=\mathrm{Fin}$, see also \cite[Theorem 4.1]{MR568707} for a textbook exposition.
\begin{thm}\label{thm:lorentzmacphail}
Let $(T_n)$ be a sequence of linear operators in $\mathcal{L}(X,Y)$. Let also $(M_n)$ be a decreasing sequence of closed linear subspaces of $X$ such that each $T_n$ is bounded on $M_n$.
Lastly, fix an ideal $\mathcal{J}$ on $\omega$ and suppose that $(T_nx) \in \ell_\infty(Y,\mathcal{J})$ for all $x \in X$. There there exists $n_0 \in \omega$ such that each $T_n$ is bounded on $M_{n_0}$.
\end{thm}
\begin{proof}
Let us suppose for the sake of contradiction that for every $n_0 \in \omega$ there exists $n \in \omega$ such that $T_n$ is not bounded on $n_0$. Then necessarily $n>n_0$. Passing, if needed to a subsequence, we can suppose that $T_{n+1}$ is not bounded on $M_n$ for all $n \in \omega$. It follows by the proof in \cite[Theorem 4.1]{MR568707} that there exists $x \in X$ such that $\|T_{n+1}x\|\ge n$ for all $n\ge 2$. This contradicts the hypothesis that the sequence $(T_nx)$ is $\mathcal{J}$-bounded.
\end{proof}
\begin{thm}\label{thm:ctoell(X)}
Let $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$.
Let $\mathcal{J}$ be a rapid$^+$ $P^+$-ideal on $\omega$.
Then $A \in (c(X), \ell_\infty(Y, \mathcal{J}))$ only if
there exists $k_0 \in \omega$ which satisfies \ref{item:T1flat}.
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=f(n)=0$ for all
$n \in \omega\setminus J_0$.
\end{thm}
\begin{proof}
Thanks to
Lemma \ref{lem:convergenceoperatorbetadualc(X)}, for each $n \in\omega$, there exists an integer $f(n) \in \omega$ such that $\|A_{n,\ge f(n)}\|<\infty$.
Without loss of generality, we can suppose that the sequence $(f(n): n \in \omega)$ is weakly increasing.
Now, for each $n \in \omega$, define
$$
M_n:=\{\bm{x} \in c(X): x_k=0 \text{ for all }k< f(n)\}.
$$
Then $(M_n)$ is a decreasing sequence of closed linear subspaces of the Banach space $c(X)$.
In addition, for each $n \in \omega$, the linear operator $A_n: c(X)\to Y$ is well defined.
Thanks to the Banach--Steinhaus theorem, $A_n$ is bounded on $M_n$ for all $n \in \omega$.
At this point, it follows by Theorem \ref{thm:lorentzmacphail} that there exists $n_0\in \omega$ such that each $A_n$ is bounded on $M_{n_0}$.
Thanks to Theorem \ref{thm:uniformJboundedness}, we obtain that $\mathcal{J}\text{-}\limsup_n \|A_n\upharpoonright M_{n_0}\|<\infty$, i.e., there exist a constant $\kappa >0$ and $J_0 \in \mathcal{J}^\star$, with $J_0=\omega$ if $\mathcal{J}=\mathrm{Fin}$, such that $\|A_n\bm{x}\|\le \kappa \|\bm{x}\|$ for all $\bm{x} \in M_{n_0}$ and $n \in J_0$.
To complete the proof, observe that
$$
\forall n \in J_0,
\forall \bm{x} \in c_{00}(X) \cap M_{n_0}, \quad
\|A_n\bm{x}\|=\left\|\sum\nolimits_{k\ge f(n_0)}A_{n,k}x_k\right\| \le \kappa \|\bm{x}\|,
$$
which implies that $\|A_{n,\ge f(n_0)}\|\le \kappa$.
Since the upper bound is independent of $n\in J_0$, the claim follows by setting $k_0:=f(n_0)$.
The second part is clear once we observe that it is possible to choose $f(n)=0$ for all $n$.
\end{proof}
In the following results we will need the following weakening of \ref{item:T2}, namely,
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^\flat$)}]
\setcounter{enumi}{1}
\item \label{item:T2flat} $\mathcal{J}\text{-}\limsup_n\|A_{n,k}x\|<\infty$ for all $x \in X$ and $k<k_0$\textup{.}
\end{enumerate}
It is clear that \ref{item:T5} implies \ref{item:T2flat}, which is the reason why it does not appear in Theorem \ref{main:IJREGULARboundedUnbounded}, cf. Remark \ref{rmk:conditionT5}.
\begin{thm}\label{thm:maddoxmain}
Let $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$.
Also, let $\mathcal{J}$ be a rapid$^+$ $P^+$-ideal on $\omega$.
Then $A \in (\ell_\infty(X), \ell_\infty(Y, \mathcal{J}))$
if and only if there exists $k_0 \in \omega$ such that
\ref{item:T1flat}, \ref{item:T2flat}, and \ref{item:T3natural} hold.
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=0$.
\end{thm}
\begin{proof}
\textsc{If part.} Fix $\bm{x}\in \ell_\infty(X)$.
Thanks to Lemma \ref{lem:convergenceoperator}, for each $n$, the sum $A_n\bm{x}=\sum_kA_{n,k}x_k$ is convergent in the norm of $Y$, hence $A\bm{x}$ is well defined.
It follows by \ref{item:T2flat} that there exist $J_1 \in \mathcal{J}^\star$ and $\kappa>0$ such that $\|A_{n,k}x_k\|\le \kappa$ for all $k<k_0$ and $n \in J_1$.
Hence
\begin{equation}\label{eq:boundedness}
\left\|A_n\bm{x}\right\|
\le \sum\nolimits_{k<k_0}\left\|A_{n,k}x_k\right\|+\left\|\sum\nolimits_{k\ge k_0}A_{n,k}x_k\right\|
\le \kappa k_0+\|\bm{x}\|\sup\nolimits_{t\in J_0} \|A_{t, \ge k_0}\|
\end{equation}
for all $n \in J_0\cap J_1 \in \mathcal{J}^\star$, which proves that $A\bm{x} \in \ell_\infty(Y, \mathcal{J})$.
\medskip
\textsc{Only If part.}
The necessity of \ref{item:T1flat} follows by Theorem \ref{thm:ctoell(X)}, with $k_0=0$ if each $A_{n,k}$ is bounded.
Now, if \ref{item:T2flat} does not hold, there would exist $x \in X$ and $k<k_0$ such that $\mathcal{J}\text{-}\limsup_n\|A_{n,k}x\|=\infty$. This contradicts the hypothesis that $A \in (\ell_\infty(X), \ell_\infty(Y,\mathcal{J}))$ by choosing $\bm{x} \in \ell_\infty(X)$ such that $x_t=x$ if $t=k$ and $x_t=0$ otherwise.
Lastly, the necessity of \ref{item:T3natural} follows by Lemma \ref{lem:convergenceoperator}.
\end{proof}
\begin{rmk}\label{rmk:conditionL1tooweak}
The same example given in Remark \ref{rmk:regulardoesnotimplylinftylinfty} proves that, even if $X=Y$, $\mathcal{J}=\mathrm{Fin}$, and each $A_{n,k}$ is bounded (so that \ref{item:T2flat} is void), condition \ref{item:T1flat} (as its stronger version \ref{item:T1}) is not sufficient to characterize the matrix class $(\ell_\infty(X), \ell_\infty(Y,\mathcal{J}))$; cf. Corollary \ref{cor:finitedimensionallinfty} below for the finite dimensional case.
\end{rmk}
\begin{thm}\label{thm:maddoxmaincXellinfty}
Let $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$.
Also, let $\mathcal{J}$ be a rapid$^+$ $P^+$-ideal on $\omega$.
Then $A \in (c(X), \ell_\infty(Y,\mathcal{J}))$ if and only if there exists $k_0 \in \omega$ such that \ref{item:T1flat} and \ref{item:T2flat} hold, together with\textup{:}
\begin{enumerate}[label={\rm (\textsc{T}\arabic{*}$^\flat$)}]
\setcounter{enumi}{3}
\item \label{item:T4flat} $\sum_{k}A_{n,k}$ convergences in the strong operator topology for all $n$\textup{.}
\end{enumerate}
In addition, if each $A_{n,k}$ is bounded, it is possible to choose $k_0=0$.
\end{thm}
\begin{proof}
The proof goes as in Theorem \ref{thm:maddoxmain}, replacing Lemma \ref{lem:convergenceoperator} with Lemma \ref{lem:convergenceoperatorbetadualc(X)}.
\end{proof}
However, if $X$ is finite dimensional, it is possible to simplify the equivalences in Theorem \ref{thm:maddoxmain} and Theorem \ref{thm:maddoxmaincXellinfty}, namely, condition \ref{item:T1flat} is necessary and sufficient in both cases choosing $k_0=0$.
\begin{cor}\label{cor:finitedimensionallinfty}
Let $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$ and assume, in addition, that $X$ is finite dimensional.
Also, let $\mathcal{J}$ be a rapid$^+$ $P^+$-ideal on $\omega$.
Then the following are equivalent\textup{:}
\begin{enumerate}[label={\rm (\roman*)}]
\item \label{item:1corollaryfinitedimellinfty} $A \in (\ell_\infty(X), \ell_\infty(Y,\mathcal{J}))$\textup{;}
\item \label{item:2corollaryfinitedimellinfty} $A \in (c(X), \ell_\infty(Y,\mathcal{J}))$\textup{;}
\item \label{item:3corollaryfinitedimellinfty} \ref{item:T1flat} holds with $k_0=0$\textup{.}
\end{enumerate}
\end{cor}
\begin{proof}
\ref{item:1corollaryfinitedimellinfty} $\implies$ \ref{item:2corollaryfinitedimellinfty} is clear,
\ref{item:2corollaryfinitedimellinfty} $\implies$ \ref{item:3corollaryfinitedimellinfty} follows by Theorem \ref{thm:ctoell(X)}, and
\ref{item:3corollaryfinitedimellinfty} $\implies$ \ref{item:1corollaryfinitedimellinfty} follows by Lemma \ref{lem:convergenceoperator}, Corollary \ref{cor:finitedimensionalVSextremepoints}, Theorem \ref{thm:maddoxmain}, and the fact that each $A_{n,k}$ is bounded.
\end{proof}
\begin{lem}\label{lemma:unboundedcase}
Let $(T_k)$ be a sequence of nonzero linear operators in $\mathcal{L}(X,Y)$. Then there exists a sequence $\bm{x} \in X^\omega$ such that $\sum_kT_kx_k$ is not convergent in the norm of $Y$.
\end{lem}
\begin{proof}
For each $k\in \omega$, pick $y_k \in X$ such that $T_ky_k\neq 0$. Now define the sequence $\bm{x}$ recursively as it follows: $x_0:=y_0$ and, if $x_0,\ldots,x_{n-1}$ are defined for some $n\ge 1$, then $x_n:=\kappa_n y_n$, where $\kappa_n:=(n+\|\sum_{k\le n-1}T_kx_k\|)/\|T_ny_n\|$. Indeed, it follows that
$$
\left\|\sum\nolimits_{k\le n}T_kx_k\right\|\ge \|T_nx_n\|-\left\|\sum\nolimits_{k\le n-1}T_kx_k\right\|=n
$$
for all $n\ge 1$, completing the proof.
\end{proof}
\begin{lem}\label{lem:Itall}
Let $\mathcal{I}$ be an ideal on $\omega$ which is tall. Then there exists a partition $\{E_n: n \in \omega\}$ of $\omega$ such that $E_n \in \mathcal{I}\cap \mathrm{Fin}^+$ for all $n$.
\end{lem}
\begin{proof}
We define such partition recursively, with the property that $\{0,\ldots,n\}\cup E_0\cup \cdots E_n$ for all $n$.
Since $\mathcal{I}$ is tall there exists an infinite set $S_0 \in \mathcal{I}$, and set $E_0:=S_0\cup \{0\}$. Now, suppose that $E_0,\ldots,E_{n-1}$ have been defined, for some $n\ge 1$, so that the set $G_n:=\omega\setminus \bigcup_{i<n}E_i$ belongs to $\mathcal{I}^\star$. Since $\mathcal{I}$ is tall there exists an infinite set $S_n \in \mathcal{I}$ contained in $G_n$. The claim follows defining $E_n:=S_n$ if $n \notin G_n$ and $E_n:=S_n\cup \{n\}$ otherwise.
\end{proof}
\begin{lem}\label{lem:dualc00I}
Let $(T_k)$ be a sequence of linear operators in $\mathcal{L}(X,Y)$. Let also $\mathcal{I}$ be a tall ideal on $\omega$. Then $\sum_kT_kx_k$ is convergent in the norm of $Y$ for all $\bm{x} \in c_{00}(X,\mathcal{I})$ if and only if
$\{k \in \omega: T_k \neq 0\}$ is finite.
\end{lem}
\begin{proof}
Thanks to Lemma \ref{lem:Itall}, there exists a partition $\{E_n: n \in \omega\}$ of $\omega$ such that $E_n \in \mathcal{I}\cap \mathrm{Fin}^+$ for all $n$.
Now, it follows by Lemma \ref{lemma:unboundedcase} that $F_n:=\{k \in E_n: T_k\neq 0\} \in \mathrm{Fin}$.
Define $F:=\bigcup_n F_n$ and suppose for the sake of contradiction that $F\notin \mathrm{Fin}$.
Since $\mathcal{I}$ is tall, there exists an infinite subset $F^\prime \subseteq F$ such that $F^\prime \in \mathcal{I}$.
However, by construction $\{k \in F^\prime: T_k\neq 0\}=F^\prime$, which contradicts Lemma \ref{lemma:unboundedcase}.
\end{proof}
Recalling that a matrix $A$ is said to be row-finite if $\{k \in \omega: A_{n,k}\neq 0\} \in \mathrm{Fin}$ for all $n\in \omega$,
we provide strong necessary conditions on the entries of matrices in $(X^\omega, \ell_\infty(Y,\mathcal{J}))$ and $(c_{00}(X,\mathcal{I}), \ell_\infty(Y))$, where $\mathcal{J}$ is countably generated and $\mathcal{I}$ is tall.
\begin{thm}\label{thm:Xomegaellinfty}
Let $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$ such that $A \in (X^\omega, \ell_\infty(Y,\mathcal{J}))$, where $\mathcal{J}$ is a countably generated ideal on $\omega$.
Then $A$ is row finite and there exist $J \in \mathcal{J}^\star$ and $k_1 \in \omega$ such that $A_{n,k}=0$ for all $n \in J$ and $k\ge k_1$.
\end{thm}
\begin{proof}
First, suppose that $A \in (X^\omega, \ell_\infty(Y,\mathcal{J}))$ and that $\mathcal{J}$ is generated by a sequence of increasing sets $(Q_j)$.
Then $A$ is row finite by Lemma \ref{lemma:unboundedcase}, so that $F_n:=\{k \in\omega: A_{n,k}\neq 0\} \in \mathrm{Fin}$ for all $n \in \omega$.
Suppose for the sake of contradiction that
\begin{equation}\label{eq:contradictionXomegaellinfty}
\forall J \in \mathcal{J}^\star, \forall k_1 \in \omega, \exists k\ge k_1, \exists n \in J, \quad A_{n,k}\neq 0.
\end{equation}
Define a strictly increasing sequence $(s_n)$ in $\omega$ such that $s_0:=\min\{n \in \omega: F_n \neq \emptyset\}$, and, recursively, $s_{n+1}:=\min(S_n\setminus Q_{j_n})$ for all $n\in \omega$, where
$$
S_n:=\{t \in \omega: \max F_t>\max(F_{s_0}\cup\cdots\cup F_{s_n})\}
\,\, \text{ and }\,\,
j_n:=\min\{j \in \omega: s_n \in Q_j\}.
$$
We claim that $S_n\setminus Q_{j_n}$ is nonempty, so that $s_{n+1}$ is well defined. Note that $S_n \in \mathcal{J}^+$: indeed, in the opposite, we would contradict \eqref{eq:contradictionXomegaellinfty} by setting $J=S_n^c$ and $k_1=1+\max(F_{s_0}\cup\cdots\cup F_{s_n})$.
In addition, since $Q_{j_n} \in \mathcal{J}$, it follows that $S_n\setminus Q_{j_n}\in \mathcal{J}^+$ (in particular, it is nonempty).
Since the set $S:=\{s_n: n \in \omega\}$ is not contained in any $Q_j$, it follows that $S \in \mathcal{J}^+$. In addition, set $k_{n}:=\max F_{s_n}$ for all $n \in \omega$. Using a technique similar to the one used in Lemma \ref{lemma:unboundedcase}, we are going to construct a sequence $\bm{x}$ supported on $K:=\{k_n:n \in \omega\}$ such that $\|A_n\bm{x}\|\ge n$ for all $n\in \omega$, so that $A\bm{x}\notin \ell_\infty(Y,\mathcal{J})$.
To this aim, pick an arbitrary sequence $\bm{y} \in X^\omega$ such that $A_{s_n,k_n}y_{n}\neq 0$ for all $n \in \omega$.
Now define $\bm{x}$ recursively so that $x_{k_0}:=y_0$ and, if $x_{k_0},\ldots,x_{k_{n-1}}$ are given for some $n\ge 1$, then
$x_{k_n}:=\kappa_n y_n$,
where $\kappa_n:=(n+\|\sum_{i\le n-1}A_{s_n,k_i}x_{k_i}\|)/\|A_{s_n,k_n}y_n\|$.
Indeed, it follows that
\begin{displaymath}
\begin{split}
\|A_{s_n}\bm{x}\|
=\left\|\sum\nolimits_{k \in F_{s_n}}A_{s_n,k}x_{k}\right\|
&=\left\|\sum\nolimits_{i\le n}A_{s_n,k_i}x_{k_i}\right\|\\
&\ge \kappa_n\|A_{s_n,k_n}y_{k_n}\|-\left\|\sum\nolimits_{i\le n-1}A_{s_n,k_i}x_{k_i}\right\|=n
\end{split}
\end{displaymath}
for all $n\ge 1$. Since $S\in \mathcal{J}^+$, we conclude that $\mathcal{J}\text{-}\limsup_n\|A_n\bm{x}\|\neq 0$.
\end{proof}
\begin{thm}\label{thm:c00ellinfty}
Let $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$ such that $A \in (c_{00}(X,\mathcal{I}), \ell_\infty(Y))$, where $\mathcal{I}$ is a tall ideal on $\omega$.
Then there exists $k_1 \in \omega$ such that $A_{n,k}=0$ for all $n \in \omega$ and $k\ge k_1$.
\end{thm}
\begin{proof}
First, $A$ is row finite by Lemma \ref{lem:dualc00I}. Now, it is enough to repeat the proof of Lemma \ref{lem:dualc00I} replacing Lemma \ref{lemma:unboundedcase} with Theorem \ref{thm:Xomegaellinfty}, with $F_n:=\{k \in \omega: \exists m \in \omega, A_{m,k}\neq 0\}$.
\end{proof}
\section{Key tools}\label{sec:technical}
Let $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$ and note that, for each $n \in \omega$ and sequence $\bm{x}$ taking values on the closed unit ball $B_X$, we have
\begin{equation}\label{eq:inequalitytoverifyequality}
\mathcal{J}\text{-}\limsup\nolimits_n\|A_{n,\omega}\| \ge \mathcal{J}\text{-}\limsup\nolimits_n\|A_n\bm{x}\|
\end{equation}
In this section, we provide sufficient conditions on the matrix $A$ and on the ideal $\mathcal{J}$ for the existence of a sequence $\bm{x}$ such that the above inequality is actually an equality.
The following result is the operator version of the sliding jump argument contained in the proof of \cite[Theorem 1.3]{ConnorLeo} for the one-dimensional case $X=Y=\mathbf{R}$ and $\mathcal{J}=\mathrm{Fin}$.
\begin{thm}\label{thm:key}
Let $\mathcal{J}$ be an ideal on $\omega$ which is countably generated.
Let also $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$ which satisfies conditions \ref{item:T1flat} for some $k_0 \in \omega$, \ref{item:T3natural}, and
\ref{item:T6flat}.
Then there exists a sequence $\bm{x}$ with values on the unit sphere $S_X$ such that
$$
\mathcal{J}\text{-}\limsup\nolimits_{n}\|A_{n,\omega}\|=
\mathcal{J}\text{-}\limsup\nolimits_{n}\|A_n\bm{x}\|.
$$
\end{thm}
\begin{proof}
Suppose that $\mathcal{J}$ is generated by an increasing sequence of sets $(Q_n)$, and define
$
\eta_0:=\mathcal{J}\text{-}\limsup\nolimits_{n}\|A_{n,\omega}\| \in [0,\infty].
$
It follows by \ref{item:T6flat} that
\begin{equation}\label{eq:c0normfinitecolumns}
\forall E \in \mathrm{Fin}, \quad
\mathcal{J}\text{-}\lim\nolimits_n\|A_{n, E}\|=\sum\nolimits_{k \in E}\mathcal{J}\text{-}\lim\nolimits_n\|A_{n, k}\|=0.
\end{equation}
Hence there exists $J_1 \in \mathcal{J}^\star$ such that $\|A_{n, \le k_0}\|\le 1$ for all $n\in J_1$.
Setting $J_2:=J_0\cap J_1 \in \mathcal{J}^\star$, it follows that
$$
\forall n\in J_2, \quad
\|A_{n,\omega}\| \le \|A_{n, \le k_0}\|+\|A_{n,\ge k_0}\| \le 1+\sup\nolimits_{t\in J_0}\|A_{t,\ge k_0}\|,
$$
which is finite by condition \ref{item:T1flat}.
This proves that $\eta_0 \neq \infty$.
Moreover, if $\eta_0=0$, it is enough to let $\bm{x}$ be an arbitrary sequence with values in $S_X$, thanks to \eqref{eq:inequalitytoverifyequality}.
Hence, let us assume hereafter that $\eta_0 \in (0,\infty)$.
At this point, for each $n \in \omega$, define the set
\begin{equation}\label{eq:SdefinitioninJplus}
E_n:=\left\{t \in J_2: \left|\|A_{t,\omega}\|-\eta_0\right|\le \frac{\eta_0}{2^{n}}\right\},
\end{equation}
and note that $(E_n)$ is a decreasing sequence of sets in $\mathcal{J}^+$.
Define also two strictly increasing sequences $(s_n)$ and $(m_n)$ of nonnegative integers, a descreasing sequence $(H_n)$ of sets of $\mathcal{J}^\star$, and a decreasing sequence $(S_n)$ of sets in $\mathcal{J}^+$ as it follows. Set $s_0:=\min S_0$, with $S_0:=E_0$, $H_0:=\omega$, and choose $m_0 \in \omega$ such that $\|A_{s_0, \ge m_0}\|\le \eta_0$, which is possible by \ref{item:T3natural}.
Now, suppose that $s_0,\ldots,s_{n-1},m_0,\ldots,m_{n-1} \in \omega$ and the sets $S_0,\ldots,S_{n-1}\in \mathcal{J}^+$ and $H_0,\ldots,H_{n-1} \in \mathcal{J}^\star$ have been defined for some $n\ge 1$.
\begin{enumerate}[label=(\roman*)]
\item \label{item:condition_setHn} Set $H_n:=H_{n-1} \cap \left\{t \in \omega: \|A_{t,m_{n-1}}\|\le \frac{\eta_0}{2^n}\right\}$, so that $H_n \in \mathcal{J}^\star$ by \eqref{eq:c0normfinitecolumns};
\item \label{item:condition_setSn} Define $S_n:=E_n \cap H_n$,
hence $S_n \in \mathcal{J}^+$ and
\begin{equation}\label{eq:conditionsn}
\forall t\in S_n, \quad \|A_{t,\le m_{n-1}}\|\le \frac{\eta_0}{2^{n}}.
\end{equation}
\item \label{item:condition_sn}
Choose $s_n \in S_{n}\setminus Q_z$, where $k$ is an integer such that $s_{n-1} \in Q_k$. In particular, $s_n>s_{n-1}$. Note that this is possible since $S_{n}\setminus Q_k \in \mathcal{J}^+$.
\item \label{item:condition_mn} Lastly, thanks to \ref{item:T3natural}, choose $m_n>m_{n-1}$ such that
\begin{equation}\label{eq:condition3}
\|A_{s_n, \ge m_n}\| \le \frac{\eta_0}{2^{n}}.
\end{equation}
\end{enumerate}
To conclude the proof, let $\bm{x}=(x_n)$ be a sequence taking values on the unit sphere $S_X$ such that
\begin{equation}\label{eq:condition4}
\forall n \ge 1, \quad
\left\|\sum\nolimits_{k \in M_n}A_{s_n,k}x_{k}\right\|\ge \left\|A_{s_n,M_n}\right\| -\frac{\eta_0}{2^{n}},
\end{equation}
where $M_n:=\{k \in \omega: m_{n-1}<k\le m_n\}$, and $x_n$ is arbitrarily chosen on the unit sphere $S_X$ for all $n \in [0,m_0]$.
It follows by \eqref{eq:SdefinitioninJplus}, \eqref{eq:conditionsn}, \eqref{eq:condition3}, and \eqref{eq:condition4} that
\begin{equation}\label{eq:chaindifficult}
\begin{split}
\forall n\ge 1, \quad
\|A_{s_n}\bm{x}\| &
\ge \left\|\sum\nolimits_{k \in M_n}A_{s_n,k}x_{k}\right\|-\left\|A_{s_n,\le m_{n-1}}\right\|-\left\|A_{s_n,\ge m_{n}}\right\|\\
&\ge \left\|A_{s_n,M_n}\right\|-\left\|A_{s_n,\le m_{n-1}}\right\|-\left\|A_{s_n,\ge m_{n}}\right\|-\frac{\eta_0}{2^{n}} \\
&\ge \left\|A_{s_n,\omega}\right\|-2\left\|A_{s_n,\le m_{n-1}}\right\|-2\left\|A_{s_n,\ge m_{n}}\right\|-\frac{\eta_0}{2^{n}}\\
&\ge \eta_0\left(1-\frac{1}{2^{n}}\right)-\frac{\eta_0}{2^{n-1}}-\frac{\eta_0}{2^{n-1}}-\frac{\eta_0}{2^{n}} >\eta_0\left(1-\frac{1}{2^{n-3}}\right).\\
\end{split}
\end{equation}
It follows by construction that we cannot find an integer $k$ such that $s_n \in Q_k$ for all $n \in \omega$. Therefore $\{s_n: n \in \omega\} \in \mathcal{J}^+$ and $\mathcal{J}\text{-}\limsup_n \|A_n\bm{x}\| \ge \eta_0$.
The converse inequality follows by \eqref{eq:inequalitytoverifyequality}, completing the proof.
\end{proof}
Related results (in the case $X=Y=\mathbf{R}$ and $\mathcal{J}=\mathrm{Fin}$) can be found in \cite[Lemma 3.1]{MR27351} and \cite[Corollary 12]{MR241957}.
The following corollary is immediate (we omit details):
\begin{cor}\label{cor:nontrivialJFin}
With the same hypothesis of Theorem \ref{thm:key}, for each $E\subseteq \omega$, there exists a sequence $\bm{x}$
taking values on $S_X\cup \{0\}$ and supported on $E$ such that
$$
\mathcal{J}\text{-}\limsup\nolimits_{n}\|A_{n,E}\|=
\mathcal{J}\text{-}\limsup\nolimits_{n}\|A_n\bm{x}\|.
$$
\end{cor}
In the finite dimensional case, the statement can be simplified:
\begin{cor}\label{cor:finitedimensionalkey}
With the same notations of Corollary \ref{cor:finitedimensionmain}, suppose that $X=\mathbf{R}^d$, $Y=\mathbf{R}^m$, and that $\mathcal{J}$ is countably generated. Let also $A$ be a matrix which satisfies\textup{:}
\begin{enumerate}[label={\rm (\textsc{K}\arabic{*})}]
\item \label{item:K1} $\sup_{n \in J_0} \sum_k\sum_{i,j}\left|a_{n,k}(i,j)\right|<\infty$ for some $J_0 \in \mathcal{J}^\star$\textup{;}
\item \label{item:K2} $\sum_k\sum_{i,j}\left|a_{n,k}(i,j)\right|<\infty$ for all $n \in \omega$\textup{;}
\item \label{item:K3} $\mathcal{J}\text{-}\lim_n \sum_{i,j}\left|a_{n,k}(i,j)\right|=0$ for all $k \in \omega$\textup{.}
\end{enumerate}
Then, for each $E\subseteq \omega$, there exists a sequence $\bm{x}=(x^{(0)},x^{(1)},\ldots)$
taking values on $S_X\cup \{0\}$ and supported on $E$ such that
$$
\mathcal{J}\text{-}\limsup\nolimits_{n}\sum\nolimits_{k \in E}\sum\nolimits_{i,j}\left|a_{n,k}(i,j)\right|=
\mathcal{J}\text{-}\limsup\nolimits_{n}
\sum\nolimits_i \left|\sum\nolimits_k \sum\nolimits_j a_{n,k}(i,j) x^{(k)}_j \right|.
$$
\end{cor}
\begin{proof}
Conditions \ref{item:K1} and \ref{item:K2} correspond to \ref{item:T1flat}, and \ref{item:K3} corresponds to \ref{item:T6flat}, cf. the proof of Corollary \ref{cor:finitedimensionmain}.
In addition, \ref{item:T1flat} implies \ref{item:T3natural} by Proposition \ref{prop:implications}.\ref{item:5simplification} and Remark \ref{rmk:ont1flat} below.
The claim follows by Corollary \ref{cor:nontrivialJFin}.
\end{proof}
\begin{rmk}\label{rmk:conditionT5insteadofT6flat}
As it has been previously observed in Remark \ref{rmk:evilgroupnorm}, there exists a matrix $A$ which does not satisfy condition \ref{item:T6flat} with $\mathcal{J}=\mathrm{Fin}$ (which corresponds to \ref{item:S3sharp}) and, on other hand, it satisfies \ref{item:T5} with $\mathcal{I}=\mathcal{J}=\mathrm{Fin}$ (which corresponds to \ref{item:S3}).
In addition, it is immediate to see that $A$ satisfies \ref{item:T1} and \ref{item:T3natural} since $A_{n,k}=0$ for all $n\ge 0$ and $k>0$.
However, in this case, the conclusion of Theorem \ref{thm:key} fails: indeed, if $\bm{x}=(x_0,x_1,\ldots)$ is a sequence taking values on the closed unit ball of $\ell_2$ then $\lim_n\|A_n\bm{x}\|=\lim_n\|A_{n,0}x_0\|= 0$, and $\|A_{n,\omega}\|=\|A_{n,0}\|=1$ for all $n\in \omega$.
Hence we cannot replace \ref{item:T6flat} in Theorem \ref{thm:key}
with the weaker pointwise condition $\mathcal{J}\text{-}\lim_nA_{n,k}=0$ for all $k\in \omega$ (namely, \ref{item:T5} with $\mathcal{I}=\mathrm{Fin}$).
\end{rmk}
However, the weaker condition above is sufficient to obtain the same claim if $\eta_0=\infty$;
note that, as it has been shown above, this case is impossible with the hypotheses of Theorem \ref{thm:key}.
\begin{thm}\label{thm:keyfakeinfty}
Let $\mathcal{J}$ be an ideal on $\omega$ which is countably generated. Let also $A=(A_{n,k})$ be a matrix of linear operators in $\mathcal{L}(X,Y)$ which satisfies conditions \ref{item:T3natural} and \ref{item:T5} with $\mathcal{I}=\mathrm{Fin}$. In addition, assume that $\mathcal{J}\text{-}\limsup\nolimits_n \|A_{n,\ge f(n)}\|=\infty$.
Then there exists a sequence $\bm{x}$ with values on $S_X$ such that $\mathcal{J}\text{-}\limsup\nolimits_n\|A_n\bm{x}\|=\infty$.
\end{thm}
\begin{proof}
We proceed with the same strategy of the proof of Theorem \ref{thm:key}.
Note that \ref{item:T3natural} implies \ref{item:T1flatflat}.
Accordingly, define $E_n:=\{t \in \omega: \|A_{t,\ge f(t)}\| \ge n\}$, which belongs to $\mathcal{J}^+$ for each $n \in \omega$.
Set $s_0:=\min S_0$, with $S_0:=E_0$, $H_0:=\omega$, choose $m_0 \in \omega$ such that $\|A_{s_0,\ge m_0}\|\le 1$, which is possible by \ref{item:T3natural}, and pick some arbitrary vectors $x_0,\ldots,x_{m_0} \in S_X$.
Now, suppose that, for some $n\ge 1$, all the integers $s_i, m_i$, sets $S_i, H_i\subseteq \omega$, and vectors $x_j$ have been defined for all $i\le n-1$ and $j\le m_{n-1}$.
Then, define recursively
$$
H_n:=H_{n-1}\cap \left\{t \in\omega: \left\|\sum\nolimits_{k\le m_{n-1}}A_{t,k}x_k\right\|\le 1\right\},
$$
which belongs to $\mathcal{J}^\star$ thanks to \ref{item:T5} with $\mathcal{I}=\mathrm{Fin}$. Define $S_n$ and $s_n$ as in the proof of Theorem \ref{thm:key}, and $m_n>m_{n-1}$ such that $\|A_{s_n,\ge m_n}\|\le 1$, which is possible again by \ref{item:T3natural}. Finally, we choose some vectors $\{x_k: k \in M_n\}$ on the unit sphere $S_X$ such that $\|\sum_{k \in M_n}A_{s_n,k}x_k\| \ge \|A_{s_n,M_n}\|-1$.
(Here, differently from the previous proof, the sequence $\bm{x}$ has been constructed recursively.)
Reasoning as in \eqref{eq:chaindifficult}, we conclude that
$$
\forall n\ge 1, \quad \|A_{s_n}\bm{x}\|\ge n-5
\quad \text{ and }\quad
\{s_t: t \in \omega\} \in \mathcal{J}^+.
$$
Therefore $\mathcal{J}\text{-}\limsup_n \|A_n\bm{x}\|=\infty$.
\end{proof}
As a consequence of the results above, we obtain an ideal version of the Hahn--Schur theorem (where the classical version corresponds to the case $\mathcal{J}=\mathrm{Fin}$):
\begin{thm}\label{thm:hahnschur}
Let $\mathcal{J}$ be a countably generated ideal. Let also $A=(a_{n,k})$ be an infinite real matrix such that
$\sum_k|a_{n,k}|<\infty$ for all $n \in \omega$ and
\begin{equation}\label{eq:hahnschurhypothesis}
\forall E \subseteq \omega, \quad
\mathcal{J}\text{-}\lim\nolimits_n \sum\nolimits_{k \in E}a_{n,k}=0.
\end{equation}
Then $\mathcal{J}\text{-}\lim\nolimits_n \sum\nolimits_{k}|a_{n,k}|=0$.
\end{thm}
\begin{proof}
Set $\eta_0:=\mathcal{J}\text{-}\limsup\nolimits_n \sum_k|a_{n,k}| \in [0,\infty]$. Note also that the standing hypotheses imply \ref{item:T3natural} and \ref{item:T6flat}. In addition, if $\eta_0<\infty$, then \ref{item:T1} holds.
It follows by Theorem \ref{thm:key} and Theorem \ref{thm:keyfakeinfty} that there exists a real sequence $\bm{x}$ taking values in $\{1,-1\}$ such that
\begin{equation}\label{eq:hahnschur}
\mathcal{J}\text{-}\limsup\nolimits_n \sum\nolimits_k a_{n,k}x_k
=\eta_0.
\end{equation}
At this point, define $y_n:=(1+x_n)/2$ for all $n \in\omega$, so that $\bm{y}=\bm{1}_E$, where $E:=\{k \in \omega: x_k=1\}$. It follows by \eqref{eq:hahnschurhypothesis} and \eqref{eq:hahnschur} that
\begin{displaymath}
\begin{split}
\eta_0 &=\mathcal{J}\text{-}\limsup\nolimits_n \left(\sum\nolimits_k a_{n,k}(2y_k-1)\right)\\
&=2\cdot \mathcal{J}\text{-}\limsup\nolimits_n \sum\nolimits_k a_{n,k}y_k - \mathcal{J}\text{-}\lim\nolimits_n \sum\nolimits_k a_{n,k}\\
&=2\cdot \mathcal{J}\text{-}\limsup\nolimits_n \sum\nolimits_{k\in E} a_{n,k} =0.
\end{split}
\end{displaymath}
Therefore $\mathcal{J}\text{-}\lim\nolimits_n \sum\nolimits_{k}|a_{n,k}|=0$.
\end{proof}
Neither \ref{item:T6flat} nor \ref{item:T5} with $\mathcal{I}=\mathrm{Fin}$
will be required for a conclusion as in Theorem \ref{thm:key} in the case of positive linear operators between certain Banach lattices.
\begin{prop}\label{prop:Anonnegativelema}
Let $X$ be an AM-space
with order unit $e>0$ and let $Y$ be a Banach lattice.
Also, let $T=(T_k)$ be a sequence of linear operators in $\mathcal{L}(X,Y)$ which are positive and such that $\sum_k T_kx_k$ is convergent in the norm of $Y$ for all sequences $\bm{x} \in \ell_\infty(X)$.
Then
$$
\forall E\subseteq \omega, \quad
\|(T_k: k \in E)\|=\left\|\sum\nolimits_{k\in E}T_ke\right\|.
$$
\end{prop}
\begin{proof}
Thanks to \cite[Theorem 3.40]{MR2011364},
there exists a (unique, up to homeomorphism) compact Hausdorff space $K$ and a lattice isometry $h: X\to C(K)$ such that $h(e)$ is the constant function $\bm{1}$, where $C(K)$ is the Banach lattice of continuous functions $f:K\to \mathbf{R}$, endowed with the supremum norm.
It follows that the closed unit ball $B_X$ of $X$ is simply the order interval $[-e,e]$, indeed
$$
B_X=h^{-1}(\{f \in C(K): \|f\|\le 1\})=h^{-1}([-\bm{1},\bm{1}])=[-e,e].
$$
At this point, let us fix a nonempty $E\subseteq \omega$,
a positive integer $n$ with $n\le |E|$,
distinct integers $i_1,\ldots,i_n \in E$,
and vectors $x_1,\ldots,x_n \in X$ such that $|x_k|\le e$ for all $k \in \{1,\ldots,n\}$.
By the fact that each $T_{k}$ is a positive linear operator and
\cite[Theorem 1.7(2)]{MR2011364}, it follows that
$$
0\le \left|\sum\nolimits_{k\le n}T_{i_k}x_k\right|
\le \sum\nolimits_{k\le n}\left|T_{i_k}x_k\right|
\le \sum\nolimits_{k\le n}T_{i_k}\left|x_k\right|
\le \sum\nolimits_{k\le n}T_{i_k}e.
$$
By the definition of group norm and the compatibility between the norm and the order structure in $X$, we obtain
$$
\|(T_k: k \in E)\|=\sup\nolimits_{n}\left\|\sum\nolimits_{k\le n, k \in E}T_ke\right\|
=\left\|\sum\nolimits_{k\in E}T_ke\right\|,
$$
which concludes the proof.
\end{proof}
In the same spirit of Corollary \ref{cor:nontrivialJFin}, we obtain the following consequence:
\begin{cor}\label{cor:Anonnegativemain}
Let $X$ be an AM-space with order unit $e>0$ and let $Y$ be a Banach lattice.
Also, let $\mathcal{J}$ be an ideal of $\omega$, and $A=(A_{n,k})$ be a matrix of positive linear operators such that $A \in (\ell_\infty(X), \ell_\infty(Y))$.
The
$$
\forall E\subseteq \omega, \quad
\mathcal{J}\text{-}\limsup\nolimits_n \|A_{n,E}\|=\mathcal{J}\text{-}\limsup\nolimits_n
\|A_n\bm{x}\|,
$$
where $x_n=e$ if $n \in E$ and $x_n=0$ otherwise.
\end{cor}
\begin{proof}
Proposition \ref{prop:Anonnegativelema} implies that $\|A_{n,E}\|=\|\sum\nolimits_{k \in E}A_{n,k}e\|=\|A_n\bm{x}\|$ for all $n$.
\end{proof}
\section{Main proofs}\label{sec:mainproofs}
\begin{proof}[Proof of Theorem \ref{main:IJREGULAR}]
\textsc{If part.} Suppose that \ref{item:T1}-\ref{item:T5} hold and fix a bounded sequence $\bm{x}$ which is $\mathcal{I}$-convergent to $\eta \in X$.
Thanks to \ref{item:T3}, $A_n\bm{x}$ is well defined for each $n \in \omega$. In addition, we obtain by \ref{item:T1} and \ref{item:T2} that Inequality \eqref{eq:boundedness} holds for all $n$ (indeed, in our case $J_0=J_1=\omega$), proving
that $A\bm{x} \in \ell_\infty(Y)$.
First, suppose that $\eta=0$.
Fix $\varepsilon>0$ and define
$$
\delta:=\frac{\varepsilon}{1+\sup_n\|A_{n,\ge k_0}\|}
$$
and $E:=\{n \in \omega: n<k_0 \text{ or }\|x_n\|>\delta\}$.
Note that $E \in \mathcal{I}$ and that, again by \ref{item:T3}, $\left(\sum_{k \in E}A_{n,k}x_k: n \in \omega\right)$ is a well-defined sequence in $Y$.
In addition,
$$
S:=\left\{n \in\omega: \left\|\sum\nolimits_{k \in E}A_{n,k}x_k\right\|>\delta\right\} \in \mathcal{J}
$$
by \ref{item:T5}.
Now, suppose that $\|A_n\bm{x}\|> \varepsilon$ for some $n \in \omega$. It follows that
\begin{equation}\label{eq:ddjyhkjfgf}
\varepsilon<
\left\|\sum\nolimits_{k \in E}A_{n,k}x_k\right\|+ \left\|\sum\nolimits_{k \notin E}A_{n,k}x_k\right\| \le
\left\|\sum\nolimits_{k \in E}A_{n,k}x_k\right\|+ \delta \sup\nolimits_t\|A_{t,\ge k_0}\|.
\end{equation}
By the definition of $\delta$, this implies that
$$
\left\|\sum\nolimits_{k \in E}A_{n,k}x_k\right\| >\varepsilon-\delta\sup\nolimits_t\|A_{t,\ge k_0}\|= \delta,
$$
so that $n \in S$. We conclude that $\{n \in \omega: \|A_n\bm{x}\|>\varepsilon\}\subseteq S \in \mathcal{J}$. By the arbitrariness of $\varepsilon$, we obtain $\mathcal{J}\text{-}\lim A\bm{x}=0$.
At this point, suppose that $\eta \in X$ and define $\bm{y} \in X^\omega$ such that $y_n:=x_n-\eta$ for all $n$. Note that $\bm{y} \in c_0^b(X,\mathcal{I})$, hence by the previous case $A\bm{y}\in c_{0}^b(Y,\mathcal{J})$. It follows by \ref{item:T4} that
$$
\mathcal{J}\text{-}\lim A\bm{x}=\mathcal{J}\text{-}\lim A\bm{y}+\mathcal{J}\text{-}\lim\nolimits_n\sum\nolimits_kA_{n,k}\eta=T\eta,
$$
which proves that $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$.
\medskip
\textsc{Only If part.} Assume that $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$. Hence, the matrix $A$ belongs, in particular, to $(c(X), \ell_\infty(Y))$.
It follows by Theorem \ref{thm:maddoxmaincXellinfty} that conditions \ref{item:T1} and \ref{item:T2} hold.
Also, for each $n \in \omega$, the sum $\sum_kA_{n,k}x_k$ is convergent in the norm of $Y$ for all sequences $\bm{x} \in c^b(X,\mathcal{I})$, hence \ref{item:T3} holds.
Moreover, for each $x \in X$, the constant sequence $(x,x,\ldots)$ has $\mathcal{I}$-limit $x$, hence $\mathcal{J}\text{-}\lim_n\sum\nolimits_kA_{n,k}x=Tx$, which is condition \ref{item:T4}.
Lastly, fix $\bm{x}\in c_{00}^b(X,\mathcal{I})$, so that $\mathcal{I}\text{-}\lim \bm{x}=0$. By the $(\mathcal{I}, \mathcal{J})$-regularity of $A$ with respect to $T$, we obtain that $A\bm{x}$ is well defined and $\mathcal{J}\text{-}\lim A\bm{x}=T(0)=0$. This proves \ref{item:T5}.
\medskip
The second part of the statement follows by Proposition \ref{prop:implications}.\ref{item:4simplification}.
\end{proof}
\begin{proof}
[Proof of Proposition \ref{prop:implications}]
\ref{item:1simplification} Condition \ref{item:T4} implies that $\sum_kA_{n,k}$ is convergent in strong operator topology for each $n \in \omega$. The conclusion follows by Lemma \ref{lem:convergenceoperatorbetadualc(X)}.
\ref{item:2simplification} For all $k \in \omega$ and $x \in X$, if $\lim_n A_{n,k}x=0$ then $\sup_n \|A_{n,k}x\|<\infty$.
\ref{item:3simplification} It follows by Lemma \ref{lem:convergenceoperator}.
\ref{item:4simplification} Since $A$ belongs to $(c(X), \ell_\infty(Y))$, the claim follows by Theorem \ref{thm:ctoell(X)}.
\ref{item:5simplification} It is known that each $A_{n,k}$ is bounded. The second part follows by Lemma \ref{lem:convergenceoperator} and Corollary \ref{cor:finitedimensionalVSextremepoints}. (Note that this is not necessarily true if $\mathrm{dim}(X)=\infty$, cf. Remark \ref{rmk:conditionL1tooweak}.)
\ref{item:6simplification} Given $k \in \omega$ and $x \in X$, the sequence $\bm{x}$ defined by $x_n=x$ if $n=k$ and $x_n=0$ otherwise has $\mathcal{I}$-limit $0$, hence $\mathcal{J}\text{-}\lim_nA_{n,k}x=T(0)=0$. This claim follows by Lemma \ref{lem:finitedimensionalnormkjfdhgd}.
\ref{item:7simplification} For all $n,k \in \omega$, there exists $a_{n,k} \in \mathbf{R}$ such that $A_{n,k}=a_{n,k}A_0$. Note also that
\begin{equation}\label{eq:groupnormlinearcombination}
\forall n \in \omega, \forall E\subseteq \omega, \quad
\|A_{n,E}\|=\|A_0\|\sum\nolimits_{k \in E}|a_{n,k}|.
\end{equation}
Hence condition \ref{item:T1} implies that $A_0$ is necessarily bounded, $k_0=0$ can be chosen by point \ref{item:4simplification} above, and hence $\sup_n\sum\nolimits_k|a_{n,k}|<\infty$. Thanks to \eqref{eq:groupnormlinearcombination}, it is immediate to conclude that $\lim_k\|A_{n,\ge k}\|=\|A_0\|\lim_k \sum_{t\ge k}|a_{n,t}|=0$ for all $n \in \omega$.
\ref{item:8simplification} By point \ref{item:7simplification} above, $A_0$ is bounded. Moreover, if $A_0=0$ the claim is obvious. Otherwise, there exists $x \in X$ such that $A_0x\neq 0$ and by \ref{item:T5} we obtain
$\mathcal{J}\text{-}\lim_n A_{n,k}x=0$, so that $\|A_0x\|\cdot \mathcal{J}\text{-}\lim_n|a_{n,k}|=0$ for all $k \in \omega$. It follows that $\mathcal{J}\text{-}\lim_n\|A_{n,k}\|=\|A_0\|\cdot \mathcal{J}\text{-}\lim_n|a_{n,k}|=0$.
\end{proof}
\begin{rmk}\label{rmk:ont1flat}
As it evident from the above proofs, the statements of Proposition \ref{prop:implications}.\ref{item:1simplification},
\ref{item:3simplification},
\ref{item:5simplification},
\ref{item:7simplification}, and
\ref{item:8simplification}
are correct also replacing \ref{item:T1} with the weaker condition \ref{item:T1flatflat}.
\end{rmk}
\begin{proof}
[Proof of Theorem \ref{thm:JfinIJregularmain}]
First, assume that $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$, with $\mathcal{J}$ countably generated.
Then \ref{item:T1} and \ref{item:T4} holds by Theorem \ref{main:IJREGULAR}.
Now, fix $\bm{x} \in c_{00}^b(X, \mathcal{I})$. Then $\mathcal{I}\text{-}\lim \bm{x}=0$ and, since $A$ is $(\mathcal{I}, \mathcal{J})$-regular, we obtain $\mathcal{J}\text{-}\lim A\bm{x}=0$, namely, $\mathcal{J}\text{-}\lim_n \|A_n\bm{x}\|=0$. Then \ref{item:T6} holds by Corollary \ref{cor:nontrivialJFin}.
Conversely, assume that \ref{item:T1}, \ref{item:T4}, and \ref{item:T6} hold. Then \ref{item:T5} holds (since it is implied by \ref{item:T6}), and conditions \ref{item:T2} and \ref{item:T3} hold by Proposition \ref{prop:implications}.\ref{item:2simplification} and \ref{item:3simplification}, respectively. It follows by Theorem \ref{main:IJREGULAR} that $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$.
\end{proof}
\begin{proof}
[Proof of Theorem \ref{thm:AMspaceregularmain}]
The first proof goes verbatim as in the proof of Theorem \ref{thm:JfinIJregularmain}, replacing Corollary \ref{cor:nontrivialJFin} with Corollary \ref{cor:Anonnegativemain}.
Also the second part proceeds similarly, with the difference that \ref{item:T2} holds by Proposition \ref{prop:implications}.\ref{item:4simplification}: indeed, thanks to \cite[Theorem 4.3]{MR2262133}, each positive linear operator $A_{n,k}$ between Banach lattices is necessarily continuous.
\end{proof}
\begin{proof}
[Proof of Corollary \ref{cor:finitedimensionmain}]
Endow $\mathbf{R}^d$ and $\mathbf{R}^m$ with the corresponding $1$-norm, as in the proof of Corollary \ref{cor:finitedimensionalVSextremepoints}, so that
$\|A_{n,k}\|=\max\nolimits_j\sum\nolimits_i|a_{n,k}(i,j)|$,
cf. e.g. \cite[Example 5.6.4]{MR2978290}.
Also, note that, for all $n \in \omega$ and $E\subseteq \omega$,
\begin{equation}\label{eq:upperlowerboundsnormmatrix}
\frac{1}{d}\sum\nolimits_{k \in E}\sum\nolimits_{i,j}|a_{n,k}(i,j)|\le
\|A_{n,E}\| \le
\sum\nolimits_{k \in E}\sum\nolimits_{i,j}|a_{n,k}(i,j)|.
\end{equation}
Accordingly, conditions \ref{item:F1}, \ref{item:F4}, and \ref{item:F6} are simply a rewriting of \ref{item:T1}, \ref{item:T4}, and \ref{item:T6}, respectively, with $k_0=0$ (which can be chosen thanks to Proposition \ref{prop:implications}.\ref{item:5simplification} and \ref{item:4simplification}; in particular, in following subcases, condition \ref{item:T2} is void).
\medskip
\textsc{Case $\mathcal{I}=\mathrm{Fin}$:} First, assume that $A$ is $(\mathcal{I},\mathcal{J})$-regular with respect to $T$.
Then \ref{item:T1} and \ref{item:T4} hold by Theorem \ref{main:IJREGULAR}.
In addition, by Proposition \ref{prop:implications}.\ref{item:6simplification} also \ref{item:T6flat} holds.
At this point, fix $E \in \mathrm{Fin}$ and note that $\|A_{n,E}\|\le \sum_{k \in E}\|A_{n,k}\|$.
Taking $\mathcal{J}$-limits on both sides, we obtain \ref{item:T6}.
Conversely, assume that \ref{item:T1}, \ref{item:T4} and \ref{item:T6} hold. Then \ref{item:T6} implies \ref{item:T5}. And \ref{item:T3} holds by Proposition \ref{prop:implications}.\ref{item:3simplification} and \ref{item:5simplification}. Hence $A$ is $(\mathcal{I}, \mathcal{J})$-regular with respect to $T$ by Theorem \ref{main:IJREGULAR}.
\medskip
\textsc{Case $\mathcal{J}$ is countably generated:} If $A$ is $(\mathcal{I},\mathcal{J})$-regular with respect to $T$, then \ref{item:T1} and \ref{item:T4} hold by Theorem \ref{main:IJREGULAR}. In addition, \ref{item:T3natural} and \ref{item:T6flat} hold by Proposition \ref{prop:implications}.\ref{item:5simplification} and \ref{item:6simplification}. Then \ref{item:T6} holds by Theorem \ref{thm:JfinIJregularmain}. The converse goes as in the previous case.
\medskip
\textsc{Case $a_{n,k}(i,j)\ge 0$ for all $1\le i\le m$, $1\le j\le d$, and $n,k \in \omega$:} Note that $\mathbf{R}^d$ is an AM-space with order unit $(1,\ldots,1)$. The proof goes on the same lines of the previous case replacing Theorem \ref{thm:JfinIJregularmain} with Theorem \ref{thm:AMspaceregularmain}.
\end{proof}
\begin{proof}
[Proof of Corollary \ref{cor:llinftyc0}]
Let $\mathcal{I}$ and $\mathcal{I}^\prime$ be two maximal ideals such that $\{2\omega, 2\omega+1\}\subseteq \mathcal{I}\cup \mathcal{I}^\prime$.
Since $c^b(\mathbf{R}^d,\mathcal{I})=c^b(\mathbf{R}^d,\mathcal{I}^\prime)=\ell_\infty(\mathbf{R}^d)$, we obtain that \ref{item:F6} holds with both $E=2\omega$ and $E=2\omega+1$, hence it is equivalent to \ref{item:F6prime}.
In turn, \ref{item:F6prime} implies \ref{item:F4} with $T=0$.
The claim follows by Corollary \ref{cor:finitedimensionmain}.
\end{proof}
\begin{proof}
[Proof of Corollary \ref{cor:eachAnkmultiple}]
First, suppose that $A$ is $(\mathcal{I},\mathcal{J})$-regular with respect to $T$.
It follows by Theorem \ref{main:IJREGULAR} that conditions \ref{item:T1}-\ref{item:T5} hold.
Thanks to Equation \eqref{eq:groupnormlinearcombination} above and \ref{item:T1}, there exists $k_0 \in \omega$ such that $\sup_{n}\|A_{n,\ge k_0}\|=\|A_0\|\sup_n\sum_{k\ge k_0}|a_{n,k}|<\infty$ (since $A_0\neq 0$), which implies \ref{item:M0} and \ref{item:M1}. Condition \ref{item:M4} is just a rewriting of \ref{item:T4}. In addition, conditions \ref{item:T3natural} and \ref{item:T6flat} hold by Proposition \ref{prop:implications}.\ref{item:7simplification}-\ref{item:8simplification}.
It follows that \ref{item:M6}, which is just a rewriting of \ref{item:T6}, holds for the same reasons in the proof of Corollary \ref{cor:finitedimensionmain}.
Conversely, assume that conditions \ref{item:M0}-\ref{item:M6} hold. \ref{item:M0} implies that each $A_{n,k}$ is bounded, hence it is possible to choose $k_0=0$ by Proposition \ref{prop:implications}.\ref{item:4simplification}, hence \ref{item:T2} holds.
Accordingly, \ref{item:M1}, \ref{item:M4}, and \ref{item:M6} are just rewritings of \ref{item:T1}, \ref{item:T4}, and \ref{item:T6}, respectively.
Lastly, \ref{item:T3} follows by Proposition \ref{prop:implications}.\ref{item:3simplification} and \ref{item:7simplification}, and \ref{item:T5} is implied by \ref{item:T6}.
To sum up, conditions \ref{item:T1}-\ref{item:T5} hold, and the conclusion follows by Theorem \ref{main:IJREGULAR}.
\end{proof}
\begin{proof}
[Proof of Theorem \ref{main:IJREGULARboundedUnbounded}]
It proceeds on the same lines of the proof of Theorem \ref{main:IJREGULAR} (recalling the \ref{item:T5} implies \ref{item:T2flat}), with the only difference that it is not necessarily true that $J_0=J_1=\omega$ but only $J_0,J_1 \in \mathcal{J}^\star$.
\end{proof}
\begin{proof}
[Proof of Corollary \ref{cor:RHgeneral}]
Proceeding as in the proof of Corollary \ref{cor:finitedimensionmain}, note that conditions \ref{item:R1} and \ref{item:R2} are simply a rewriting of \ref{item:T1flat}, \ref{item:R4} is a rewriting of \ref{item:T4}, and \ref{item:R6} is a rewriting of \ref{item:T6}. Since each $A_{n,k}$ is bounded, we choose $k_0=f(n)=0$ for all $n \in Q_{t_0}$ in the statement of Theorem \ref{main:IJREGULARboundedUnbounded}.
First, assume that $A$ satisfies \eqref{eq:boundedunboundeddefinition}.
It follows by Theorem \ref{main:IJREGULARboundedUnbounded} that \ref{item:T1flat}, \ref{item:T4}, and \ref{item:T5} hold.
In addition, \ref{item:T1flat} implies \ref{item:T3natural} by Lemma \ref{lem:convergenceoperator} and Corollary \ref{cor:finitedimensionalVSextremepoints}; and condition \ref{item:T6flat} holds by Lemma \ref{lem:finitedimensionalnormkjfdhgd}. We conclude by Corollary \ref{cor:nontrivialJFin} that $A$ satisfies \ref{item:T6}.
Conversely, assume that \ref{item:T1flat}, \ref{item:T4}, and \ref{item:T6} hold. Then \ref{item:T6} implies \ref{item:T5}.
As before, \ref{item:T1flat} implies \ref{item:T3natural}, so that $A$ satisfies \ref{item:T3} by Lemma \ref{lem:convergenceoperator}. The conclusion follows by Theorem \ref{main:IJREGULARboundedUnbounded}.
\end{proof}
\begin{proof}
[Proof of Theorem \ref{main:IJREGULARUnboundedBounded}]
Suppose for the sake of contradiction that such a matrix $A$ exists.
Since $A \in (c_{00}(X,\mathcal{I}), \ell_\infty(Y))$, it follows by Theorem \ref{thm:c00ellinfty} that
$$
\exists k_1 \in \omega, \forall k\ge k_1, \forall n \in \omega, \quad A_{n,k}=0.
$$
At this point,
since $A$ is also $(\mathcal{I},\mathcal{J})$-regular with respect to $T$, then \ref{item:T1}-\ref{item:T5} hold by Theorem \ref{main:IJREGULAR}.
In addition, $T$ is a nonzero linear operator, hence there exists (nonzero) $x \in X$ such that $y:=Tx\neq 0$, and pick $\bm{x}:=(x,x,\ldots)$.
Since $A$ satisfies \eqref{eq:boundedunboundeddefinition}, we obtain that
$\mathcal{J}\text{-}\lim A\bm{x}=y$.
However, it follows by condition \ref{item:T5} that
$$
\mathcal{J}\text{-}\lim\nolimits_n A_n\bm{x}=
\mathcal{J}\text{-}\lim\nolimits_n \sum\nolimits_{k<k_1}A_{n,k}x=0,
$$
providing the desired contradiction.
\end{proof}
\begin{proof}
[Proof of Theorem \ref{thm:cIc_0bJ}]
This is an immediate consequence of Theorem \ref{thm:c00ellinfty}.
\end{proof}
\begin{proof}
[Proof of Theorem \ref{main:IJREGULARUnboundedUnbounded}]
First, suppose that \ref{item:T1flat}, \ref{item:T3sharp}, \ref{item:T4}, and \ref{item:T5sharp} hold for some $k_0 \in \omega$.
Fix a sequence $\bm{x}$ such that $\mathcal{I}\text{-}\lim \bm{x}=\eta$.
By \ref{item:T3sharp}, $A\bm{x}$ is well defined. Moreover, since \ref{item:T5sharp} implies \ref{item:T2flat}, we can pick $\kappa$ and $J_1 \in \mathcal{J}^\star$ as in the proof of Theorem \ref{thm:maddoxmain}, and define $E:=\{k \in \omega: \|x_k-\eta\|\ge 1\} \in \mathcal{I}$.
Using \ref{item:T5sharp}, there exists $J_2 \in \mathcal{J}^\star$ such that $\left\|\sum_{k \in E, k\ge k_0}A_{n,k}x_k\right\|\le 1$ for all $n \in J_2$.
At this point, it follows by \ref{item:T1flat} that
\begin{displaymath}
\begin{split}
\|A_n\bm{x}\|&\le \left\|\sum\nolimits_{k<k_0}A_{n,k}x_k\right\|
+ \left\|\sum\nolimits_{k\in E, k\ge k_0}A_{n,k}x_k\right\|+ \left\|\sum\nolimits_{k\notin E, k\ge k_0}A_{n,k}x_k\right\|\\
&\le \kappa k_0+1+\sup\{\|x_k\|: k\notin E\}\cdot \sup\nolimits_{t \in J_0}\left\|A_{t, E^c \setminus [0,k_0)}\right\|\\
&\le \kappa k_0+1+(\|\eta\| +1)\cdot \sup\nolimits_{t \in J_0}\left\|A_{t,\ge k_0}\right\|
\end{split}
\end{displaymath}
for all $n \in J$, where $J:=J_0\cap J_1\cap J_2 \in \mathcal{J}^\star$. Therefore $A\bm{x} \in \ell_\infty(Y,\mathcal{J})$.
We conclude as in the proof of Theorem \ref{main:IJREGULAR} that $A$ satisfies \eqref{eq:unoundedunboundeddefinition}, with the only difference that Inequality \eqref{eq:ddjyhkjfgf} holds for all $n \in J_0$, once we replace $\sup_t\|A_{t,\ge k_0}\|$ with $\sup_{t\in J_0}\|A_{t,\ge k_0}\|$.
\medskip
Conversely, if $A$ satisfies \eqref{eq:unoundedunboundeddefinition}, then
\ref{item:T3sharp} holds because the sequence $A\bm{x}$ is well defined for each $\bm{x} \in c(X,\mathcal{I})$; moreover, $A\in (c^b(X,\mathcal{I}), c(Y,\mathcal{J}))$, and it follows by Theorem \ref{main:IJREGULARboundedUnbounded} that conditions \ref{item:T1flat} and \ref{item:T4} hold.
Lastly, if $\bm{x}$ is a sequence supported on $\mathcal{I}$, then $\mathcal{J}\text{-}\lim A\bm{x}=0$, which proves \ref{item:T5sharp}.
\end{proof}
\begin{proof}
[Proof of Theorem \ref{thm:cXcYrapid}]
Thanks to Theorem \ref{main:IJREGULARUnboundedUnbounded}, we have only to show that if $A$ satisfies \eqref{eq:unoundedunboundeddefinition} then \ref{item:T1flat} holds. This follows by Theorem \ref{thm:ctoell(X)} since $A \in (c(X),\ell_\infty(Y,\mathcal{J}))$, with $k_0=f(n)=0$ for all $n \in \omega\setminus J_0$ if each $A_{n,k}$ is bounded.
\end{proof}
\section{Closing remarks}
We leave as open questions for the reader to check whether Theorem \ref{thm:key} holds for the larger class of rapid$^+$ $P^+$-ideals $\mathcal{J}$, and whether this condition characterizes the latter class of ideals in the same spirit of \cite[Theorem 5.1]{MR4172859}.
In addition, it would be interesting to obtain a characterization of the matrix class $(c(X,\mathcal{Z}), c(Y,\mathcal{Z}))$, analogously to Theorem \ref{thm:cXcYrapid}.
\subsection*{Acknowledgments}
The author is grateful to Simone Cerreia-Vioglio (Università Bocconi, IT) for useful discussions.
\bibliographystyle{amsplain}
|
\section{Introduction}
Random matrix theory has proven over time to be a powerful modern tool in mathematics and physics. With widespread applications in different areas such as engineering, statistical mechanics, probability, number theory, to mention only a few, its theory is rich and has been under intense development in the past thirty or so years. In a sense, much of the success of random matrix theory has been due to its exact solvability, or integrability, turning them into touchstones for predicting and confirming complex phenomenon in nature.
One of the most celebrated results in random matrix theory is the convergence of fluctuations of the largest eigenvalue towards the Tracy-Widom law $F_{\mathrm{GUE}}$. This result was first obtained by Tracy and Widom \cite{TW94} for matrices from the Gaussian Unitary Ensemble (GUE), who also showed that $F_{\mathrm{GUE}}$ is expressible in terms of a particular solution to the Painlevé II equation (shortly PII). Their findings sparked numerous advances in mathematics and physics, which began from the extension to several other matrix models but shortly afterwards widespread beyond the realm of random matrices.
Starting with the celebrated Baik-Deift-Johansson Theorem \cite{baik_deift_johansson}, the distribution $F_{\mathrm{GUE}}$ has been identified as the limiting one-point distribution for the fluctuations of a wide range of different probabilistic models. One of the most ubiquitous of such models is the KPZ equation, introduced in the 1980s by Kardar, Parisi and Zhang. Despite numerous developments surrounding it, exactly solving it remained an outstanding open problem until the early 2010s, where four different groups of researchers \cite{AmirCorwinQuastel2011, SasamotoSpohn2010, Dotsenko2010, CalabreseLeDoussalRosso2010} independently found an exact solution for its so-called narrow wedge solution. Amongst these works, Amir, Corwin and Quastel \cite{AmirCorwinQuastel2011} found the one-point distribution for the height function of the KPZ solution, showing that it relates to a distribution found a little earlier by Johansson in a grand canonical Gaussian-type matrix model \cite{Johansson2007}, and further characterizing it in terms of the integro-differential Painlevé II equation. The latter is an extension of the PII differential equation, and almost as an immediate consequence the authors of \cite{AmirCorwinQuastel2011} also obtained that this one-point distribution, in the large time limit, converges to $F_{\mathrm{GUE}}$ itself.
In much inspired by \cite{baik_deift_johansson,johansson_2000} and later by \cite{AmirCorwinQuastel2011, SasamotoSpohn2010, Dotsenko2010, CalabreseLeDoussalRosso2010}, it has been realized that several stochastic growth models share an inherent connection with statistics of integrable point processes, in what is formally established as an identity between a transformation of the growth model and statistics for the point process. To our knowledge, the very first instance of such relation appears in the work of Borodin \cite{Borodin_2018} which connects higher spin vertex model with Macdonald measures. By taking appropriate limit of such connection, it was later found that the KPZ equation is connected to the Airy$_2$ point process \cite{BorodinGorin2016}, ASEP is related to the discrete Laguerre Ensemble, stochastic six vertex model is connected in the same way to the Meixner ensemble, or yet the Krawtchouk ensemble \cite{BorodinOlshanski2017}.
As a common feature to these connections, the underlying correspondences establish that the so-called $q$-Laplace transform of the associated height function coincides with some multiplicative statistics of the point process. The latter, in turn, admits exact solvability, and it is widely believed that a long list of new insights on the growth models can be obtained by studying the corresponding multiplicative statistics of the point processes.
This program has already been taken to a great start for the KPZ equation, and exploring its connection with the Airy$_2$ point process Corwin and Ghosal \cite{CorwinGhosal2020} were able to obtain bounds for the lower tail of the KPZ equation. Shortly afterwards, such bounds were improved by Cafasso and Claeys \cite{CafassoClaeys2021} with Riemann-Hilbert methods common in random matrix theory.
Our major goal is to take on the program of understanding multiplicative statistics for random particle systems, and carry out its detailed asymptotic analysis for one of the most inspiring models, namely eigenvalues of random matrices.
While fluctuations of linear statistics of the eigenvalues of random matrices has been extensively studied in the past, the study of multiplicative statistics has only been carried over in particular instances, remarkably when dealing with products and quotients of characteristic polynomials, see \cite{DesrosiersLiu2014, BaikDeiftStrahov2003, AkemannStrahovWurfel2020} and the references therein for a complete account.
In this work, we consider the Hermitian matrix model with an arbitrary one-cut regular potential $V$, and associate to it a general family of multiplicative statistics on its eigenvalues, indexed by a function $Q$ satisfying certain regularity conditions. Our findings show that when the number of eigenvalues is large such multiplicative statistics become universal: in the large matrix limit they converge to a multiplicative statistics of the Airy$_2$ point process which is independent of $V$ and $Q$. This limiting statistics admits a characterization in terms of a particular solution to the integro-differential Painlevé II equation, and it is the same quantity that connects the KPZ equation and the Airy$_2$ point process. So, in turn, we find that random matrix theory can recast the narrow wedge solution to KPZ equation for finite time in an universal way.
The random matrix statistics that we study are associated to a deformed orthogonal polynomial ensemble, also indexed by $Q$, which we analyze. As we learn from earlier work of Borodin and Rains \cite{Borodin_Rains_2005} which was recently rediscovered and greatly extended by Claeys and Glesner \cite{ClaeysGlesner2021} (and which we also briefly explain later on), this deformed ensemble is a conditional ensemble of a marked process associated to the original random matrix model. We show that the correlation kernel for this point process converges to a kernel constructed out of the same solution to the integro-differential PII equation that appeared in \cite{AmirCorwinQuastel2011}. This kernel is again universal in both $V$ and $Q$, and turns out to be the kernel of the induced conditional process on the marked Airy$_2$ point process. Naturally, there are orthogonal polynomials and their norming constants and recurrence coefficients associated to this deformed ensemble. With our approach we also obtain similar universality results for such quantities, showing that they are indeed universal in $V$ and $Q$ and also connect to the integro-differential PII in a neat way.
Beyond the concrete results, with this work we also hope to shed light into the rich structure underlying multiplicative statistics for eigenvalues of random matrices. Much of the recent relevance of Painlevé equations is due to its appearance in random matrix theory, see \cite{duits_painleve_kernels} for an overview of several of these connections. There has been a growing recent interest in integro-differential Painlevé-type equations \cite{BothnerCafassoTarricone2021,CafassoClaeysRuzza2021, Bothner2021, Krajenbrink2020, CharlierClaeysRuzza2021,LiechtyWang2020}, and our results place the integro-differential PII as a central universal object in random matrix theory as well.
We scale the multiplicative statistics to produce a critical behavior at a soft edge of the matrix model, and consequently the core of our asymptotic analysis lies within the construction of a local approximation to all the quantities near this critical point. Our main technical tool is the application of the Deift-Zhou nonlinear steepest descent method to the associated Riemann-Hilbert problem (shortly RHP), and the mentioned local approximation is the so-called construction of a local parametrix. In our case, a novel feat is that this local parametrix construction is performed in a two-step way, first with the construction of a model problem varying with large parameter, and second with the asymptotic analysis of this model problem. In the latter, a RHP recently studied by Cafasso and Claeys \cite{CafassoClaeys2021} (see also the subsequent works \cite{CafassoClaeysRuzza2021, CharlierClaeysRuzza2021}) which is related to the lower tail of the KPZ equation shows up, and it is this RHP that ultimately connects all of our considered quantities to the integro-differential PII.
The choice of scaling of our multiplicative statistics is natural, illustrative but not exhaustive. As we point out later, with our approach it becomes clear that other scalings could also be analyzed, say for instance scaling around a bulk point, or yet soft/hard edge points with critical potentials, and indicate that other integrable systems extending the integro-differential PII may emerge.
\section{Statement of main results}\label{sec:StatementResults}
Let $\Lambda^{(n)} \deff(\lambda_1<\ldots <\lambda_n)$ be a $n$-particle system with
distribution
\begin{equation}\label{deff:classicalUE}
\frac{1}{Z_n}\prod_{1\leq j<k\leq n}(\lambda_k-\lambda_j)^2\prod_{j=1}^n \ee^{-nV(x)}\mathrm{d} x,
\end{equation}
where $Z_n$ is the partition function
\begin{equation}\label{deff:nondeformedpartfunction}
Z_n\deff\int_{\mathbb{R}^n} \prod_{1\leq j<k\leq n}(\lambda_k-\lambda_j)^2\prod_{j=1}^n \ee^{-n V(\lambda_j)}\mathrm{d} \lambda_1\cdots \mathrm{d}\lambda_n.
\end{equation}
The distribution \eqref{deff:classicalUE} is the eigenvalue distribution of the unitarily-invariant random matrix model with potential function $V$ \cite{deift_book,mehta_book}.
We associate to $\Lambda^{(n)}$ the multiplicative statistics
\begin{align}
\mathsf L_n^Q(\msf s) \deff& \; \mathbb{E}\left(\prod_{j=1}^n \frac{1}{1+ \ee^{-\msf s -n^{2/3} Q(\lambda_j)}} \right) \nonumber \\
=& \; \frac{1}{Z_n}\int_{\mathbb{R}^n} \prod_{1\leq j<k\leq n}(\lambda_k-\lambda_j)^2\prod_{j=1}^n \sigma_{n}(\lambda_j)\ee^{-n V(\lambda_j)}\mathrm{d} \lambda_1\cdots \mathrm{d}\lambda_n = \frac{\mathsf Z^Q_n(\msf s)}{Z_n},\quad \msf s>0, \label{deff:Ln}
\end{align}
where
$\mathsf Z_n^Q(\msf s)$ is the partition function for the deformed model
\begin{equation}\label{deff:DeformedPartFunction}
\mathsf Z_n^Q(\msf s)\deff \int_{\mathbb{R}^n} \prod_{1\leq j<k\leq n}(\lambda_k-\lambda_j)^2\prod_{j=1}^n \sigma_{n}(\lambda_j)\ee^{-n V(\lambda_j)}\mathrm{d} \lambda_1\cdots \mathrm{d}\lambda_n
\end{equation}
and we denoted
\begin{equation*}
\sigma_{n}(z)=\sigma_{n}(z\mid \msf s)\deff \left(1+ \ee^{-\msf s-n^{2/3} Q(z) } \right)^{-1}.
\end{equation*}
When $Q$ is linear, with a straightforward change of parameters $\mathsf L^Q_n$ reduces to
\begin{equation}\label{eq:qlaplacegeneral}
\mathbb{E}\left(\prod_{x\in \mathcal X}\frac{1}{1+\zeta q^{x}}\right),
\end{equation}
where the expectation is over the set $\mathcal X$ of configurations of points (that is, for us $\mathcal X=\Lambda^{(n)}$) and $q\in (0,1)$ should be viewed as a parameter of the model and, in general, $\zeta\in \mathbb{C}$ is a free parameter. The expression \eqref{eq:qlaplacegeneral} may be viewed as a transformation of the point process, where $\zeta\in \mathbb{C}$ becomes the spectral variable of this transformation, and the matching $\zeta=\ee^{-\msf s}$ motivates the distinguished role of $\msf s$ in \eqref{deff:Ln}. In the context of random particle systems, this particular multiplicative statistics is associated to the notion of a $q$-Laplace transform \cite{BorodinCorwinMacProc,BorodinGorin2016,BorodinOlshanski2017,BorodinCorwinSasamoto2014} that we already mentioned in the Introduction, and it has been one of the key quantities in several outstanding recent progresses in asymptotics for random particle systems \cite{ImamuraSasamoto2019,CorwinGhosal2020,CafassoClaeys2021}.
We work under the following assumptions.
\begin{assumption}\label{MainAssumption}\hfill
\begin{enumerate}[(i)]
\item The potential $V$ is a nonconstant real polynomial of even degree and positive leading coefficient, and its equilibrium measure $\mu_V$ is one-cut regular, we refer to Section~\ref{sec:equilibriummeasure} below for the precise definitions. Performing a shift on the variable, without loss of generality we assume that the right-most endpoint of $\supp\mu_V$ is at the origin, so that
$$
\supp\mu_V=[-a,0],
$$
for some $a>0$.
\item The function $Q$ is real-valued over the real line, and analytic on a neighborhood of the real axis. We also assume that it changes sign at the right-most endpoint of $\supp\mu_V$, with
\begin{equation}\label{eq:signchangeQ}
Q(x)>0 \text{ on }(-\infty,0), \quad Q(x)<0 \text{ on } (0,\infty),
\end{equation}
with $x=0$ being a simple zero of $Q$. A particular role is played by the negative value $Q'(0)$, so we set
\begin{equation}\label{deff:tQprime}
\mathsf t\deff -Q'(0)>0.
\end{equation}
\end{enumerate}
\end{assumption}
Although $\msf t$ in Assumption~\ref{MainAssumption}--(ii) will have the interpretation of time, we stress that in this paper it will be kept fixed within a compact of $(0,+\infty)$ rather than being made large or small.
For our results and throughout the whole work, we also talk about uniformity of several error terms with respect to $\msf t$ in the sense that we now explain. Because $Q$ is analytic on a neighborhood of the real axis, analytic continuation shows that it is completely determined by its derivatives $Q^{(k)}(0)$, $k\geq 0$. When we say that some error is uniform in $\msf t$ within a certain range, we mean uniform when we vary $Q$ as a function of $-Q'(0)=\msf t$ while keeping all other derivatives $Q^{(k)}(0)$, $k\geq 2$, fixed.
The condition in Assumption~\ref{MainAssumption}--(i) is standard in random matrix theory and they are known to hold when, say, $V$ is a convex function \cite{Saff_book}. The one-cut assumption is made just for ease of presentation, as it simplifies the Riemann-Hilbert analysis at the technical level considerably. On the other hand, the regularity condition is used substantially in our arguments, but is standard in Random Matrix Theory literature and holds true generically \cite{KuijlaarsMcLaughlin2000}. Most of our results are of local nature near the right-most endpoint of $\supp\mu_V$ and could be shown to hold true for multi-cut potentials near regular endpoints as well, with appropriate but non-essential modifications.
Assumption~\ref{MainAssumption}--(ii) should be seen as specifying enough regularity on the multiplicative statistics, here indexed by this factor $Q$. Because of condition (ii), we have the pointwise convergence
\begin{equation}\label{eq:limitingpert}
\sigma_n(x)\stackrel{n\to\infty}{\to}
\begin{cases}
0, & x>0, \\
1, & x<0,
\end{cases}
\end{equation}
which means that the introduction of the factor $\sigma_n$ in the original weight $\ee^{-nV}$ has the effect of producing an interpolation between this original weight and its cut-off version $\chi_{(-\infty,0)}\ee^{-nV}$, where from here onward $\chi_J$ is the characteristic version of a set $J$. Comparing the Euler-Lagrange conditions on the equilibrium problem induced by the weights $\ee^{-nV}$ and $\chi_{(-\infty,0)}\ee^{-nV}$, the observation we just made heuristically indicates that the factor $\sigma_n$ does not change the global behavior of eigenvalues. This may also be rigorously confirmed as an immediate consequence of our analysis, but we do not elaborate on this end.
On the other hand, introducing a local coordinate $u=-z/n^{2/3}$ near the origin, the approximation
$$
\ee^{-\msf s-n^{2/3}Q(z)}\approx\ee^{-( \msf t u+\msf s)}
$$
goes through, and we see that there is a competition between the term $\msf s$ and $Q(z)$ that affects the local behavior of the weight at the scale $\mathcal{O}(n^{-2/3})$ near the origin, which is the same scale for nontrivial fluctuations of eigenvalues around the same point. The main results that we are about to state concern obtaining the asymptotic behavior as $n\to \infty$ of several quantities of the model, and in particular they showcase how this term $Q$ affects the local scaling regime of the eigenvalues near the origin and leads to connections with the integro-differential Painlevé II equation as already mentioned.
A central object in this paper is the multiplicative statistics
\begin{equation}\label{deff:MultStatsAiry}
\msf L^{\kern -0.15em \ai}(s,T)\deff\mathbb{E} \left(\prod_{j=1}^\infty \frac{1}{1+\ee^{T^{1/3}(s+\mathfrak{a}_j)}}\right),
\end{equation}
where the expectation is over the Airy$_2$ point process with random configuration of points $\{\mathfrak{a}_j\}$\cite{SpohnPrahofer2002}. The quantity $\msf L^{\kern -0.15em \ai}$ admits two remarkable characterizations, which are also of particular interest to us. The first is the formulation via a Fredholm determinant, namely
$$
\msf L^{\kern -0.15em \ai}(s,T)=\det\left( \mb I-\mb K^{{\rm Ai}}_T \right)\restr{L^2(-s,\infty)},
$$
where $\mb K^{{\rm Ai}}_T$ is the integral operator on $L^2(-s,\infty)$ acting with the finite temperature (or fermi-type) deformation of the Airy kernel $\mathsf K_T^{{\rm Ai}}$, defined by
$$
\mathsf K^{\rm{Ai}}_T(u,v)\deff \int_{-\infty}^\infty \frac{\ee^{T^{1/3}\zeta}}{1+\ee^{T^{1/3}\zeta}}\ai(u+\zeta)\ai(\zeta+v)\mathrm{d} \zeta, \quad u,v\in \mathbb{R}.
$$
The term `temperature' stems from the connection between the KPZ equation and the random polymer models. Despite the name finite temperature, the parameter $T$ here corresponds to the time in the KPZ equation, see \eqref{eq:KPZAiry} below. The Fredholm determinant $\det\left( \mb I-\mb K^{\rm{Ai}}_T \right)$ appeared for the first time in the work of Johansson \cite{Johansson2007} as the limiting process of a grand canonical (that is, when the number of particles/size of matrix is also random) version of a Gaussian random matrix model, and interpolates between the classical Airy kernel when $T\to +\infty$ and the Gumbel distribution when $T\to 0^+$ with $s$ scaled appropriately. In \cite[Remark~1.13]{Johansson2007} Johansson already raises the question on whether a related classical (that is, not grand canonical) matrix model has limiting local statistics that interpolate between Gumbel and Tracy-Widom, as a similar feature to $\det\left( \mb I-\mb K^{\rm{Ai}}_T \right)$. Since then, other works have found $\det\left( \mb I-\mb K^{\rm{Ai}}_T \right)$ to be the limiting distribution for fluctuations around the largest particle of a point process \cite{DeanLeDoussalMajumdarSchehr2015,CundenMezzadriOConnell2018,LiechtyWang2020,BeteaBouttier2019}. In common, these works consider specific models rather than obtaining $\det\left( \mb I-\mb K^{\rm{Ai}}_T \right)$ as the universal limit for a whole family of particle systems.
Another characterization of $\msf L^{\kern -0.15em \ai}$ is via a Tracy-Widom type formula that relates it to the integro-differential PII. It reads
\begin{equation}\label{eq:intdiffRepLAiry}
\log \msf L^{\kern -0.15em \ai}(-\msf S\msf T^{1/3},\msf T^{-2})=-\frac{1}{\msf T}\int_{\msf S}^{\infty}(v-\msf S) \left(\int_{-\infty}^\infty \Phi(r\mid v,\msf T)^2 \frac{\ee^{-r}}{(1+\ee^{-r})^2}\mathrm{d} r - \frac{v}{2}\right) \mathrm{d} v,
\end{equation}
where $\Phi$ solves the integro-differential Painlevé II equation
\begin{equation}\label{eq:intdiffPII}
\partial^2_\msf S \Phi(\xi\mid \msf S,\msf T)=\left(\xi+\frac{\msf S}{\msf T}+\frac{2}{\msf T}\int_{-\infty}^\infty \Phi(r\mid \msf S,\msf T)^2\frac{\ee^{-r}}{(1+\ee^{-r})^2}\mathrm{d} r \right)\Phi(\xi\mid \msf S,\msf T)
\end{equation}
with boundary value
$$
\Phi(\xi\mid \msf S,\msf T)\sim \msf T^{1/6}\ai(\msf T^{2/3}\xi+\msf S\msf T^{-1/3}),\quad \text{as } \xi\to\infty \text{ with } |\arg \xi|<\pi-\delta,
$$
for any $\delta>0$. This characterization has been obtained in the already mentioned work by Amir, Corwin and Quastel \cite{AmirCorwinQuastel2011}, in connection with the narrow wedge solution to the KPZ equation, and following the work \cite{BorodinGorin2016} by Borodin and Gorin has the interpretation that we now describe. For $\mathcal H(X,T)$ being the Hopf-Cole solution to the KPZ equation with narrow wedge initial data at the space-time point $(X,T)$, introduce the rescaled random variable
$$
\Upsilon_T=\frac{\mathcal H(0,2T)+\frac{T}{12}}{T^{1/3}}.
$$
Based on the previous works \cite{AmirCorwinQuastel2011,CalabreseLeDoussalRosso2010,SasamotoSpohn2010,Dotsenko2010}, in \cite{BorodinGorin2016} the identity
\begin{equation}\label{eq:KPZAiry}
\mathbb{E}_{\mathrm{KPZ}}\left(\ee^{-\ee^{T^{1/3}(\Upsilon_T+s)}}\right)=\msf L^{\kern -0.15em \ai}(s,T)
\end{equation}
between the height function of the KPZ equation and the multiplicative statistics $\msf L^{\kern -0.15em \ai}(s,T)$ is identified. This is an instance of matching formulas relating growth processes with determinantal point processes that we already mentioned at the Introduction. One of the key aspects of this representation is that the Airy$_2$ point process is determinantal, and consequently its statistics can be studied using techniques from exactly solvable/integrable models. Indeed Equation~\ref{eq:KPZAiry} is the starting point taken by Cafasso and Claeys \cite{CafassoClaeys2021}, who then connected $\msf L^{\kern -0.15em \ai}(s,T)$ to a RHP that will also play a major role for us. Recently, Cafasso, Claeys and Ruzza \cite{CafassoClaeysRuzza2021} also obtained an independent proof of the representation \eqref{eq:intdiffRepLAiry}, extending it to more general multiplicative statistics of the Airy$_2$ point process. Other proofs and extensions of this integro-differential equation have also been recently found in related contexts \cite{BothnerCafassoTarricone2021, Krajenbrink2020, Bothner2021}. Also, by exploring \eqref{eq:KPZAiry} the tail behavior of the KPZ equation has become rigorously accessible in various asymptotic regimes \cite{CorwinGhosal2020,CafassoClaeys2021,CafassoClaeysRuzza2021,CharlierClaeysRuzza2021}.
As a first result, we prove that the multiplicative statistics $\msf L^{\kern -0.15em \ai}$ is the universal limit of $\mathsf L_n^Q(\msf s)$.
\begin{theorem}\label{thm:asymptoticsqLaplacetransf}
Suppose that $V$ and $Q$ satisfy Assumptions~\ref{MainAssumption} and fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. For a constant $\mathsf c_V>0$ that depends solely on $V$, and any $\nu\in (0,2/3)$, the asymptotic estimate
\begin{equation}\label{eq:AsymptFormLQLAi}
\log \mathsf L^Q_n\left(\msf s\right)=\log\msf L^{\kern -0.15em \ai}\left(-\frac{\msf s\mathsf c_V}{\msf t},\frac{\msf t^3}{\mathsf c_V^3}\right)+\mathcal{O}(n^{-\nu}),\quad n\to\infty
\end{equation}%
holds true uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t \leq 1/\msf t_0$.
\end{theorem}
Findings on random matrix theory surrounding the Tracy-Widom distribution have inspired an enormous development in the KPZ universality theory. For the KPZ equation one of the major developments can be phrased by saying that the fluctuations of the height function for the narrow wedge solution coincide, in the large time limit, with the $\beta=2$ Tracy-Widom law from random matrix theory. Theorem~\ref{thm:asymptoticsqLaplacetransf} is saying that the connection between random matrix theory and the KPZ equation can be recast already at any finite time, and not only for Gaussian models but also universally in $V$ and $Q$. Similar connection exists \cite{BBCW18} between the solution of the KPZ equation in half-space under the Robin boundary condition and Airy$_1$ point process which, in turn, in the large time limit relate this KPZ solution to GOE matrices.
The constant $\mathsf c_V$ is determined from \eqref{eq:spectralcurvehV} and \eqref{eq:derivativepsicpsi} below. It is the first derivative of a conformal map near the origin, which is constructed out of the equilibrium measure for $V$. Ultimately, we make a conformal change of variables of the form $\zeta\approx\mathsf c_V z/n^{2/3}$, which in turn identifies
$$
\frac{1}{1+\ee^{-\msf s-n^{2/3}Q(z)}}\approx \frac{1}{1+\ee^{-\msf s +\msf t \zeta/\mathsf c_V}}.
$$
In light of \eqref{deff:MultStatsAiry}, this explains the evaluation $s=-\msf s \mathsf c_V/\msf t$ and $T=\msf t^3/\mathsf c_V^3$ on the right-hand side of \eqref{eq:AsymptFormLQLAi}.
We emphasize that the error term in \eqref{eq:AsymptFormLQLAi} is not in sharp form. In Section~\ref{sec:issues} we explain how this term arises from our techniques. We do not have indications regarding whether the true optimal error would be $\mathcal{O}(n^{-2/3})$ (or of any polynomial order) or if it should involve, say, logarithmic corrections.
Our next results concern limiting asymptotic formulas for the matrix model underlying $\mathsf L_n^Q$, starting with the partition function $\mathsf Z_n^Q$ from \eqref{deff:nondeformedpartfunction}. For a polynomial $V$, its deformation
\begin{equation}\label{eq:deformedGaussianpotential}
V_t(z)\deff \left(1-\frac{1}{t}\right)z^2+V(z/t),\quad 1\leq t\leq +\infty
\end{equation}
has the property that $V_1=V$ and $V_\infty=z^2$ is Gaussian. Under the assumption that $V_t$ is one-cut regular for every $t\geq 1$, Bleher and Its \cite{BleherIts2005} proved that an expansion of the form
\begin{equation}\label{eq:asymptPartBleherIts}
Z_n=Z_n^{\mathrm{GUE}}\ee^{n^2\bm e_0^V+\bm e_1^V}\left(1+\mathcal{O}(n^{-2})\right),\quad n\to\infty,
\end{equation}%
holds, where $Z_n^{\mathrm{GUE}}$ is the GUE partition function for $V_\infty=z^2$ and $\bm e_0^V$ and $\bm e_1^V$ are functions analytic on the coefficients of $V$. In fact, their result ensures a full asymptotic expansion in inverse powers of $n^2$, see also \cite{bessis_itzykson_zuber,ercolani_mclaughlin_partition_function,BleherEynard2003} for important earlier work obtaining similar results under different conditions, and also the more recent contributions \cite{bleher_deano_partition_function,GuionnetBorot2013}.
As an immediate corollary to Theorem~\ref{thm:asymptoticsqLaplacetransf} we obtain some terms in the asymptotic expansion of the deformed partition function \eqref{deff:DeformedPartFunction}.
\begin{corollary}
Suppose $V$ and $Q$ satisfy Assumptions~\ref{MainAssumption} and fix $\msf s_0>0$ and $\msf t_0>0$. In addition, assume that the potential $V_t$ in \eqref{eq:deformedGaussianpotential} is one-cut regular for any $t\geq 1$. For any $\nu\in (0,2/3)$, the partition function admits an expansion of the form
\begin{equation}\label{eq:DefPartFctionAsympt}
\mathsf Z_n^Q(\msf s)=Z_n^{\mathrm{GUE}}\ee^{n^2 \bm e_1^V +\bm e_0^V}\msf L^{\kern -0.15em \ai}\left(-\frac{\msf s\mathsf c_V}{\msf t},\frac{\msf t^3}{\mathsf c_V^3}\right)\left(1+\mathcal{O}(n^{-\nu})\right),\quad n\to\infty
\end{equation}
which is valid uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$, and the coefficients $\bm e_0^V$ and $\bm e_1^V$ are as in \eqref{eq:asymptPartBleherIts}.
\end{corollary}
\begin{proof}
Follows from the expansion \eqref{eq:asymptPartBleherIts}, Theorem~\ref{thm:asymptoticsqLaplacetransf} and \eqref{deff:Ln}.
\end{proof}
The order of error \eqref{eq:DefPartFctionAsympt} is not $n^{-2}$ as in \eqref{eq:asymptPartBleherIts} but weaker and not sharp. This phenomenon can be traced back to the fact that $\sigma_n$ has infinitely many poles accumulating on the real axis as $n\to \infty$, see the discussion in Section~\ref{sec:issues} below. A similar error order was obtained in \cite[Theorem~9.1 and Equation~(9.68)]{BleherIts2005}, in a transitional regime from a one-cut to two-cut potential, and where the role played here by $\msf L^{\kern -0.15em \ai}$ is replaced by the GUE Tracy-Widom distribution itself.
From the general theory of unitarily invariant random matrix models, it is known that the density appearing in \eqref{deff:Ln} admits a determinantal form. Setting
\begin{equation}\label{def:perturbedweight}
\omega_{n}(z)=\omega^Q_n(z\mid \msf s)\deff\sigma_{n}(z)\ee^{-n V(z)},\quad \text{where we recall}\quad \sigma_{n}(z)=\left(1+ \ee^{-\msf s- n^{2/3}Q(z) } \right)^{-1},
\end{equation}
this means that the identity
$$
\frac{1}{\mathsf Z_n(\msf s)}
\prod_{1\leq j<k\leq n}(\lambda_k-\lambda_j)^2\prod_{j=1}^n \sigma_{n}(\lambda_j)\ee^{-n V(\lambda_j)}
=
\frac{1}{n!}\det\left(\omega_n(\lambda_j)^{1/2}\mathsf K^Q_n(\lambda_j,\lambda_k)\omega_n(\lambda_k)^{1/2}\right)_{j,k=1}^n
$$
holds true for a function of two variables $\mathsf K^Q_n(x,y)$ satisfying certain properties, known as the correlation kernel of the eigenvalue density on the left-hand side. The correlation kernel is not unique, but in the present setup it may be taken to be the Christoffel-Darboux kernel for the orthogonal polynomials for the weight $\omega_n$, as we introduce in detail in \eqref{deff:Kndefweight}, and whenever we talk about $\mathsf K_n^Q$ we mean this Christoffel-Darboux kernel. In particular, $\mathsf K^Q_n=\mathsf K^Q_n(\cdot\mid \msf s)$ does depend on both $Q$ and $\msf s$.
Our second result proves universality of the kernel $\mathsf K_n^Q$, showing that its limit depends solely on $\msf s$ and $\msf t=-Q'(0)$, but not on other aspects on $Q$, and relates to the integro-differential PII. For its statement, it is convenient to introduce the new set of variables
\begin{equation}\label{eq:corr_tad_Tad}
\msf T=\msf t^{-3/2}\quad \text{and}\quad \msf S=\msf s\msf t^{-3/2}.
\end{equation}
With $\Phi(\xi)=\Phi(\xi\mid \msf S,\msf T)$ being the solution to the integro-differential Painlevé II equation in \eqref{eq:intdiffPII} and the variables $\msf s,\msf t$ and $\msf S,\msf T$ related by \eqref{eq:corr_tad_Tad}, we set
$$
\upphi_1(\zeta\mid \msf s,\msf t)=\Phi(\xi(\zeta)\mid \msf S,\msf T),\quad \upphi_2(\zeta\mid \msf s,\msf t)=(\partial_\msf S \Phi)(\xi(\zeta)\mid \msf S,\msf T),\quad \quad \xi(\zeta)\deff -\msf s+\msf t\zeta,
$$
and introduce the kernel
$$
\mathsf K_\infty(u,v\mid \msf s,\msf t)\deff \frac{\upphi_1(v\mid \msf s,\msf t)\upphi_2(u\mid \msf s,\msf t)-\upphi_1(u\mid \msf s,\msf t)\upphi_2(v\mid \msf s,\msf t)}{u-v},\quad u,v\in \mathbb{R}.
$$
\begin{theorem}\label{thm:limitingkernel}
Assume that $V$ and $Q$ satisfy Assumptions~\ref{MainAssumption} and fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. With
\begin{equation}\label{deff:scalingunvn}
u_n\deff\frac{u}{\mathsf c_V n^{2/3}},\quad v_n\deff\frac{v}{\mathsf c_V n^{2/3}},
\end{equation}
the estimate
\begin{equation}\label{eq:LimitKnQFinalForm}
\frac{\ee^{-\frac{n}{2}(V(u_n)+V(v_n))}}{\mathsf c_V n^{2/3}} \mathsf K^Q_n\left(u_n,v_n\mid \msf s\right)= \mathsf K_\infty(u,v\mid \msf s,\msf t/\mathsf c_V)+ \mathcal{O}(n^{-1/3}),\quad n\to\infty,
\end{equation}
holds true uniformly for $u,v$ in compacts of $\mathbb{R}$, and uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{theorem}
In the recent work \cite{ClaeysGlesner2021}, Claeys and Glesner developed a general framework for certain conditional point processes, which in particular yields a probabilistic interpretation of the kernel $\mathsf K_n^Q$ as we now explain. For a point process $\Lambda$, we add a mark $0$ to a point $\lambda\in \Lambda$ with probability $\sigma_n(\lambda)$ and a mark $1$ with complementary probability $1-\sigma_n(\lambda)$. This induces a decomposition of the point process $\Lambda=\Lambda_0\cup \Lambda_1$, where $\Lambda_j$ is the set of eigenvalues with mark $j$. We then consider the induced point process $\widehat\Lambda$ obtained from $\Lambda$ upon conditioning that $\Lambda_1=\emptyset$, that is, that all points have mark $0$.
When applied to the eigenvalue point process $\Lambda=\Lambda^{(n)}$ induced by the distribution \eqref{deff:classicalUE}, the theory developed in \cite{ClaeysGlesner2021} shows that $\widehat\Lambda^{(n)}$ is a determinantal point process with correlation kernel proportional to $\omega_n(x)^{1/2}\omega_n(y)^{1/2}\mathsf K_n^Q(x,y)$ which, in turn, generates the same point process as the left-hand side of \eqref{eq:LimitKnQFinalForm}, see \cite[Sections~4 and 5]{ClaeysGlesner2021}. A comparison of the RHP that characterizes the kernel $\mathsf K_\infty$ (see Section~\ref{sec:KPZRHPKPZ} below, in particular \eqref{eq:transfPhicckernelphicc}) with the discussion in \cite[Section~5.2]{ClaeysGlesner2021} shows that $\mathsf K_\infty$ is a (renormalized) correlation kernel for the marked point process $\widehat{\{\mathfrak a_k\}}$ of the Airy$_2$ point process $\{\mathfrak a_k\}$ with the marking function $(1+\ee^{-\msf s+\msf t \lambda})^{-1}$. So Theorem~\ref{thm:limitingkernel} assures that the conditional process on the marked eigenvalues converges, at the level of rescaled correlation kernels, to the conditional process on the marked Airy$_2$ point process.
We also obtain asymptotics for the norming constant $\upgamma^{(n,Q)}_{n-1}(\msf s)$ for the ${(n-1)}$-th monic orthogonal polynomial for the weight $\omega_n(x\mid \msf s)$ (see \eqref{def:normingdeformed} for the definition), showing that its first correction term depends again solely on $\msf s,\msf t$, and also relates to the integro-differential Painlevé II equation.
\begin{theorem}\label{thm:asymptoticsnormingconstant}
Suppose that $V$ and $Q$ satisfy Assumptions~\ref{MainAssumption} and fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. The norming constant has asymptotic behavior
\begin{equation}\label{eq:asymptformNormingctt}
\upgamma^{(n,Q)}_{n-1}(\msf s)^2=\frac{a}{4\pi }\ee^{-2n\ell_V}\left(\frac{1}{2}-\frac{1}{n^{1/3}}\frac{\mathsf c_V^{1/2}}{\msf t^{1/2}}
\left(\mathsf p(\msf s,\msf t/\mathsf c_V)-\frac{\msf s^2\mathsf c_V^{3/2}}{4\msf t^{3/2}}\right)+\mathcal{O}(n^{-2/3})
\right), \quad n\to \infty,
\end{equation}
uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$, where $\mathsf p(\msf s,\msf t)=\mathsf P(\msf S,\msf T)$, and the function $\mathsf P$ relates to the solution $\Phi$ from \eqref{eq:intdiffPII} via
\begin{equation}\label{eq:intdiffPIIPhip}
\partial_\msf S \mathsf P(\msf S,\msf T)=\frac{\msf S}{2\msf T}+\frac{1}{\msf T}\int_{-\infty}^\infty \Phi(r\mid \msf S,\msf T)^2\frac{\ee^{-r}}{(1+\ee^{-r})^2}\mathrm{d} r.
\end{equation}
\end{theorem}
Our approach also yields asymptotic formulas for the orthogonal polynomials and their recurrence coefficients, and relate them to the integro-differential Painlevé II equation as well, but for the sake of brevity we do not state them.
\section{About our approach: issues and extensions}
\subsection{Issues to be overcome}\label{sec:issues}\hfill
Our main tool for obtaining all of our results is the Fokas-Its-Kitaev \cite{FokasItsKitaev92} Riemann-Hilbert Problem (RHP) for orthogonal polynomials (shortly OPs) that encodes the correlation kernel $\mathsf K_n^Q$, the norming constants $\upgamma^{(n,Q)}_{n-1}(\msf s)^2$ and ultimately also the multiplicative statistics $\mathsf L_n^Q$, and its asymptotic analysis via the Deift-Zhou nonlinear steepest descent method \cite{Deiftetal1999,DeiftZhou93mKdV}. The overall arch of this asymptotic analysis is the usual one, summarized in the diagram in Figure~\ref{fig:diagramRHP}, and we now comment on its major steps.
\begin{figure}[t]
\centering
\begin{tikzpicture}[every text node part/.style={align=center}, node distance=10mm and 2mm,
terminal/.style={
rectangle,minimum size=8mm,rounded corners=3mm,
very thick,draw=black!50,
top color=white,bottom color=black!20},
interm/.style={
rectangle,minimum size=8mm,rounded corners=0mm,
very thick,draw=black!50,
top color=white,bottom color=white}
]
\node (Y) [terminal] {$\bm Y$};
\node (OP) [below=of Y] {RHP for OPs};
\node (g) [above right=of Y] {introduction of \\ $g$-function};
\node (YT) [below=of g]{};
\node (T) [terminal, below right=of g] {$\bm T$};
\node (l) [above right=of T] {opening \\ of lenses};
\node (TS) [below=of l] {};
\node (S) [terminal, below right=of l] {$\bm S$};
\node (lp) [below right=of S] {local parametrix near $z=0$\\ $\bm P(z)=\bm E_n(z)\bm \Phi_n(n^{2/3}\psi(z))$};
\node (SR) [above=of lp]{};
\node (gp) [above right=of S] {global parametrix\\ $\bm G(z)$};
\node (SRg) [below=of gp]{};
\node (SS) [interm, below=of T]{independent step\\ $\bm\Phi_n(\zeta)\to {\bm \Phi}_0(\zeta)$};
\node (R) [terminal, above right=of lp] {$\bm R$};
\path [thick] (Y) edge[->] (T);
\path [thick] (g) edge[-] (YT.center);
\path [thick] (T) edge[->] (S);
\path [thick] (l) edge[-] (TS.center);
\path [thick] (S) edge[->] (R);
\path [thick] (lp) edge[-] (SR.center);
\path [thick] (gp) edge[-] (SRg.center);
\path [thick] (OP) edge[-] (Y);
\path [thick] (SS) edge[->,transform canvas={yshift=-2mm}] (lp);
\end{tikzpicture}
\caption{Schematic diagram for the steps in the asymptotic analysis of the RHP for orthogonal polynomials. There is also an Airy local parametrix used near $z=-a$ which we omit in this diagram.}\label{fig:diagramRHP}
\end{figure}
Starting with the RHP for OPs that we name $\bm Y$, in the first step we transform $\bm Y\mapsto \bm T$ with the introduction of the $g$-function (or, equivalently, the $\phi$-function), and this is done so with the help of the equilibrium measure for $V$ that accounts only for the part $\ee^{-nV}$ of the weight $\omega_n$. In the second step, we open lenses with a transformation $\bm T\mapsto \bm S$ as usual.
The third step is the construction of the global parametrix $\bm G$. In our case, in this construction we also have to account for the perturbation $\sigma_n$ of the weight $\ee^{-nV}$, so a Szeg\"o function-type construction is used.
The fourth step is the construction of local parametrices at the endpoints $z=-a,0$ of $\supp\mu_V$, with the goal of approximating all the quantities locally near these endpoints. This is accomplished by, first, considering a change of variables $z\mapsto \zeta=n^{2/3}\psi(z)$ after the conformal map $\psi$ chosen appropriately for each endpoint and, then, constructing the solution to a model RHP $\bm\Phi(\zeta)$ in the $\zeta$-plane. Following these steps, the local parametrix at the left edge $z=-a$ of $\supp\mu_V$ is standard and utilizes Airy functions.
The construction of the local parametrix at the right edge $z=0$ is, however, a lot more involved. As we mentioned earlier, the factor $\sigma_n$ affects asymptotics of local statistics near the origin. In fact, the weight $\sigma_n$ has singularities precisely at the points of the form
$$
-\frac{\msf s}{n^{2/3}}+\frac{\pi \mathrm{i}(2k+1)}{n^{2/3}},\quad k\in \mathbb{Z}.
$$
This means that for $\msf s\ll n^{2/3}$ there are infinitely many poles of $\sigma_n$ accumulating near the real axis. As such, in this case for large $n$ the perturbed weight $\omega_n$ fails to be analytic in any fixed neighborhood of the origin. If we were to consider only $\msf s\to +\infty$ fast enough, one could still push the standard RHP analysis further with the aid of Airy local parametrices, at the cost of a worse error estimate. However, when $\msf s=\mathcal{O}(1)$ we have poles accumulating too fast to the real axis, and a different asymptotic analysis has to be accomplished, in particular a new local parametrix is needed.
When changing coordinates $z\mapsto \zeta$ near $z=0$, the model problem $\bm\Phi=\bm\Phi_n$ obtained is then $n$-dependent. This is so because the jump of the local parametrix involves $\sigma_n$, and consequently in the process of changing variables the resulting model problem has a jump that involves a transformation of $\sigma_n$ itself. This is in contrast with usual constructions with, say, Airy, Bessel or Painlevé-type parametrices, where the jumps can be turned into piecewise constant in the $z$-plane and, hence, also remain piecewise constant in the $\zeta$-plane. Another feature of the RHP for the model problem $\bm\Phi_n$ is that its jump is not analytic on the whole plane, and instead it is analytic only in a growing (with $n$) disk, and for a fixed $n$ we can only ensure that its jump matrix is $C^\infty$ in a neighborhood of the jump contour.
All in all, this means that carrying out the asymptotic analysis of $\bm \Phi_n(\zeta)$ as $n\to \infty$ is also needed. As we said, the jump for $\bm\Phi_n$ involves a transformation of $\sigma_n$, so ultimately also depends on the function $Q$ from \eqref{def:perturbedweight}. But it turns out that as $n\to\infty$, we have the convergence $\bm \Phi_n\to \bm\Phi_0$ in an appropriate sense, where $\bm \Phi_0$ is independent of $Q$. This limiting $\bm \Phi_0$ is the solution to a RHP that appeared recently in connection with the KPZ equation \cite{CafassoClaeys2021} and which was later shown to connect with the integro-differential PII in the recent work of Claeys, Cafasso and Ruzza \cite{CafassoClaeysRuzza2021,CharlierClaeysRuzza2021}. For this reason we term it the id-PII RHP.
With the construction of the global and local parametrices, the asymptotic analysis is concluded in the usual way, by patching them together and obtaining a new RHP for a matrix function $\bm R$. This matrix $\bm R$, in turn, solves a RHP whose jump is asymptotically close to the identity, and consequently $\bm R$ can be found perturbatively.
After concluding this asymptotic analysis, we undress the transformations $\bm R \mapsto \cdots \mapsto \bm Y$ and obtain asymptotic expressions for the wanted quantities. For the kernel $\mathsf K_n^Q$ and the norming constant $\upgamma^{(n,Q)}_{n-1}(\msf s)^2$, after this undressing Theorems~\ref{thm:limitingkernel} and \ref{thm:asymptoticsnormingconstant} follow in a standard manner.
However, to obtain \eqref{eq:AsymptFormLQLAi} quite some extra work is needed. When dealing with statistics of matrix models via OPs, one of the usual approaches is to extract the needed information via the partition function and its relation with the norming constants via a product formula, see for instance \eqref{eq:PartitionFctionNormingctt} below. Usually this is accomplished via some differential identity or with careful estimate of each term in the product formula, see for instance the works \cite{bleher_deano_partition_function, BleherIts2005, BaikBuckinghamDiFrancoIts2009,Krasovsky07} and their references for explorations along these lines. In virtue of the relation \eqref{deff:Ln} this was in fact our original attempt, but several technical issues arise. Instead, at the end we express $\mathsf L_n^Q$ directly as a weighted double integral of $\mathsf K_n^Q(x,x\mid \msf s)$ in the variables in $x$ and $\msf s$, this is done in Proposition~\ref{prop:intreprLn} below. The $x$-integral takes place over the whole real line, which means that when we undress $\bm R\mapsto \bm Y$ we obtain a formula for $\mathsf L_n^Q$ involving global and all local parametrices. The integral in $\msf s$ extends to $+\infty$, which is one of the main reasons why in our main statements we also keep track of uniformity of errors when $\msf s\to +\infty$. We then have to estimate the double integral, accounting for exponential decays of most of the terms but also exact cancellations of some other terms. Ultimately, the whole analysis leads to a leading contribution coming solely from a portion of the integral that arises from the model problem $\bm \Phi_n$. With a further asymptotic analysis of the later integral we obtain an integral solely of ${\bm \Phi}_0$ which then yields Theorem~\ref{thm:asymptoticsqLaplacetransf}.
The convergence $\bm \Phi_n\to {\bm \Phi}_0$ is treated as a separate issue, and to achieve it we need several information about this id-PII parametrix ${\bm \Phi}_0$. As a final outcome, we obtain that $\bm \Phi_n$ is close to ${\bm \Phi}_0$ with an error term of the form $\mathcal{O}(n^{-\nu})$, for any $\nu\in (0,2/3)$. But, in much due to the non-analyticity of the jump matrix for $\bm \Phi_n$, we are not able to achieve a sharp order $\mathcal{O}(n^{-2/3})$ unless further conditions were placed on $Q$. This non-optimal error explains the appearance of the same error order in \eqref{eq:AsymptFormLQLAi}. In the course of this asymptotic analysis we rely substantially in \cite{CafassoClaeysRuzza2021}. Among other needed info, we also borrow from the same work the connection of ${\bm \Phi}_0$ with the integro-differential PII. In the same work, the authors actually show that ${\bm \Phi}_0$ relates to particular solutions to the KdV equation that reduce to the integro-differential PII. As such, Theorems~\ref{thm:limitingkernel} and \ref{thm:asymptoticsnormingconstant} could be phrased in terms of a solution to the KdV rather than to the integro-differential PII. We opt to phrase them with the latter because this formulation encodes that all self-similarities have already been accounted for.
If we were to assume that the jump matrix for $\bm \Phi_n$ were piecewise analytic on the whole plane and not merely $C^\infty$, we could deform $\bm \Phi_n$ to a family of RHPs considered in \cite{CafassoClaeysRuzza2021}. With this in mind, the analysis of the convergence $\bm \Phi_n\to {\bm \Phi}_0$ is inspired by several aspects in this just mentioned work but, as we already said, here we are forced to work under different conditions on the jump matrix. In particular, one could adapt the methods in \cite{CafassoClaeysRuzza2021} to actually prove that $\bm \Phi_n$ does too relate to an $n$-dependent solution to the integro-differential PII. Consequently, with a careful inspection of our work one could show that Theorems~\ref{thm:asymptoticsqLaplacetransf}, \ref{thm:limitingkernel} and \ref{thm:asymptoticsnormingconstant} admit versions with $n$-dependent leading terms. For instance, relating the norming constant $\upgamma^{(n,Q)}_{n-1}(\msf s)^2$ with the model problem $\bm\Phi_n$ one could obtain an asymptotic formula of the form
$$
\upgamma^{(n,Q)}_{n-1}(\msf s)^2=\frac{a}{4\pi }\ee^{-2n\ell_V}\left(\frac{1}{2}-\frac{1}{n^{1/3}}\frac{\mathsf c_V^{1/2}}{\msf t^{1/2}}
\left(\mathsf p_n(\msf s,\msf t/\mathsf c_V)-\frac{\msf s^2\mathsf c_V^{3/2}}{4\msf t^{3/2}}\right)+\mathcal{O}(n^{-2/3})
\right), \quad n\to \infty,
$$
where the $n$-dependent function $\mathsf p_n$ is obtained from $\bm \Phi_n$ and relates to a $n$-dependent solution $\Phi_n$ to the integro-differential PII. In fact, with standard arguments one could improve the formula above to a full asymptotic expansion in powers of $n^{-1/3}$, with bounded but $n$-dependent coefficients. Underlying our arguments there is the statement that $\mathsf p_n=\mathsf p+\mathcal{O}(n^{-\nu})$ for any $\nu\in (0,2/3)$, which then yields Theorem~\ref{thm:asymptoticsnormingconstant}. But as a drawback, although one could potentially improve \eqref{eq:asymptformNormingctt} and also obtain the term of order $n^{-2/3}$ explicitly, it is not possible to obtain the $\mathcal{O}(n^{-1})$ term in \eqref{eq:asymptformNormingctt} unless one improves the error $\mathcal{O}(n^{-\nu})$ in the convergence $\bm \Phi_n\to {\bm \Phi}_0$ to a sharp error $\mathcal{O}(n^{-2/3})$.
\subsection{Possible extensions}\hfill
Most of our approach may be extended to potentials $V$ for which the equilibrium measure $\mu_V$ is critical, and also under different conditions on $Q$ as we now explain.
Apart from technical adaptations in several steps of the RHP for OPs which are nowadays well understood, our analysis carries over to potentials $V$ for which the equilibrium measure $\mu_V$ is regular but multicut, with the same conditions on $Q$ when $\mu_V$ has the origin as its right-most endpoint.
When, say, the density $\mu_V$ vanishes to a higher power at a soft edge and/or $Q$ changes sign with an arbitrary odd vanishing order at the same soft edge, we need to replace the power $n^{2/3}$ in $\sigma_n$ by another appropriate power to modify the local statistics near this point in a non-trivial critical manner. Once this is done, the asymptotic analysis of the RHP for OPs that we perform carries over mostly with minor modifications, and the only major issue to overcome is in the construction of a new local parametrix $\widetilde{\bm \Phi}_n$ near this soft edge point and its corresponding asymptotic analysis. In this case, we expect that $\widetilde{\bm \Phi}_n\to \widetilde{\bm \Phi}_0$ for a new function $\widetilde{\bm \Phi}_0$. It is relatively simple to write a RHP that should be satisfied by this $\widetilde{\bm \Phi}_0$, and we expect it to be related to the KdV hierarchy \cite{Claeys2012} but with nonstandard initial data. It would be interesting to see if the particular solutions obtained this way reduce to integro-differential hierarchies of Painlevé equations, in the same spirit of the recent works \cite{BothnerCafassoTarricone2021,Krajenbrink2020}.
One could also consider similar statistics to \eqref{deff:Ln} with a $Q$ that vanishes at a bulk point of $\supp\mu_V$. We do expect that most of our work carries through to this situation, at least when we impose $V$ to be again one-cut regular and $Q$ to vanish quadratically at a point inside $\supp\mu_V$. The main issue that should arise is again on the construction of the local parametrix near this point, and its corresponding asymptotic analysis. This model should lead to multiplicative statistics of the Sine kernel (and the higher order generalizations of it). Similar considerations go through to hard-edge models, leading to multiplicative statistics of the Bessel process. To our knowledge, such multiplicative statistics of Bessel and Sine have not been considered in the literature so far. However, finite temperature versions of the Sine and Bessel kernels do have appeared, see for instance \cite{Johansson2007,BeteaBouttier2019,BeteaOccelli2021,CundenMezzadriOConnell2018}.
\subsection{Organization of the paper}\hfill
The paper is organized in two parts. In the first part, we deal with a family of RHPs $\bm \Phi_\tau$ that contains the model RHP $\bm \Phi_n$ needed in the asymptotic analysis of OPs. In Section \ref{sec:modelproblem} we introduce $\bm\Phi_\tau$ formally. In Sections~\ref{sec:KPZRHP} and \ref{sec:boundsKPZRHP} we discuss the RHP ${\bm \Phi}_0$, which is a particular case of $\bm\Phi_\tau$, and review several of its properties, translating results from \cite{CafassoClaeys2021,CafassoClaeysRuzza2021} to our notation and needs. In Section~\ref{sec:asympanalymodelprobl} we prove the convergence $\bm \Phi_\tau\to {\bm \Phi}_0$ and of related quantities in the appropriate sense. The latter section contains all the needed results for the asymptotic analysis of the RHP for OPs, and concludes the first part of this paper.
The second part of the paper is focused on the asymptotic analysis of the RHP for OPs. In Section~\ref{sec:eqmeasuretau} we discuss several aspects that relate to the equilibrium measure. In Section~\ref{sec:rhpapproachops} we introduce the Christoffel-Darboux kernel $\mathsf K_n^Q$ and related quantities, and display how they relate to the RHP for OPs. In particular, in Proposition~\ref{prop:intreprLn} we write $\mathsf L_n^Q$ directly to as an integral of the kernel $\mathsf K_n^Q$, a result which may be of independent interest. In Section~\ref{sec:rhpanalysis} we perform the asymptotic analysis of the RHP for the OPs. In Section~\ref{sec:ProofMainResults} use the conclusions from Sections~\ref{sec:rhpanalysis} and \ref{sec:asympanalymodelprobl} and prove Theorems~\ref{thm:limitingkernel} and \ref{thm:asymptoticsnormingconstant}. Also from the results from Sections~\ref{sec:rhpanalysis} and \ref{sec:asympanalymodelprobl} and assuming additional technical estimates, the proof of Theorem~\ref{thm:asymptoticsqLaplacetransf} is given in Section~\ref{sec:ProofMainResults}. Such remaining technical estimates are also ultimately a consequence of the RHP analysis, but their proofs are rather cumbersome and postponed to Section~\ref{sec:technicalestimates}.
For the remainder of the paper it is convenient to denote
\begin{equation}\label{deff:matrixnot1}
\bm e_1\deff\begin{pmatrix}
1 \\ 0
\end{pmatrix},\quad
\bm e_2\deff\begin{pmatrix}
0 \\ 1
\end{pmatrix},\quad
\bm E_{jk}\deff\bm e_j\bm e_k^\mathrm T,
\end{equation}
so $\bm E_{jk}$ is a $2\times 2$ matrix with the $(j,k)$-entry equals $1$ and all other entries zero. With this notation, the Pauli matrices, for instance, take the form
\begin{equation}\label{deff:matrixnot2}
\bm I\deff\bm E_{11}+\bm E_{22},\quad
\sp_1\deff \bm E_{12}+\bm E_{21}, \quad
\sp_2\deff -\mathrm{i} \bm E_{12}+\mathrm{i} \bm E_{21},\quad
\sp_3\deff\bm E_{11}-\bm E_{22}.
\end{equation}
In particular, for any reasonably regular scalar function $f$, the spectral calculus yields
\begin{equation}\label{deff:matrixnot3}
f(z)^{\sp_3}=
\begin{pmatrix}
f(z) & 0 \\ 0 & 1/f(z)
\end{pmatrix}.
\end{equation}
These notations will be used extensively in the coming sections.
\subsection*{Acknowledgments}\hfill
P. G. wishes to thank Ivan Corwin and Alexandre Krajenbrink for many helpful conversations and Alexei Borodin for comments on the earlier version of this manuscript. G. S. is grateful to Jinho Baik, Tom Claeys, Mattia Cafasso, Lun Zhang and Alfredo Deaño for inspiring conversations, and Dan Betea for pointing us out to relevant references. He also acknowledges his current support by São Paulo Research Foundation under grants \# 2019/16062-1 and \# 2020/02506-2, and by Brazilian National Council for Scientific and Technological Development (CNPq) under grant \# 315256/2020-6.
This work was partially developed while the authors participated in the program {\it Universality and Integrability in Random Matrix Theory and Interacting Particle Systems}, hosted
by the Mathematical Sciences Research Institute in Berkeley, California, during the Fall 2021 semester, supported by the National Science Foundation
under Grant No. DMS-1928930. We are grateful to the organizers for their effort in providing an excellent research atmosphere despite the uncertain times.
\section{A Model Riemann-Hilbert Problem}\label{sec:modelproblem}
In this section we discuss a model Riemann-Hilbert Problem that will be used in the construction of a local parametrix in the asymptotic analysis for the orthogonal polynomials. As such, this model problem plays a central role in obtaining all our major results.
\subsection{The model problem}\hfill
Set
\begin{equation}\label{def:contour_sigma}
\msf \Sigma_0\deff [0,+\infty), \; \msf \Sigma_1\deff [0,\ee^{2\pi \mathrm{i}/3}), \; \msf \Sigma_2\deff (-\infty,0], \; \msf \Sigma_3\deff [0,\ee^{-2\pi \mathrm{i}/3}), \quad \msf \Sigma\deff\bigcup_{j=0}^3 \msf \Sigma_j,
\end{equation}
orienting $\msf \Sigma_0$ from the origin to $\infty$, and the remaining arcs from $\infty$ to the origin, see Figure~\ref{fig:sigmaj}.
\begin{figure}[t]
\centering
\begin{tikzpicture}[scale=1]
\draw [line width=0.4mm,lightgray] (-3,0)--(3,0) node [pos=1,right,black] {$\mathbb{R}$};
\draw [line width=0.4mm,lightgray] (0,-3)--(0,3) node [pos=1,above,black] {$\mathrm{i}\mathbb{R}$};
\draw [thick,postaction={mid arrow={black,scale=1.5}}] (0,0) to (0:3) node [yshift=0pt,xshift=-12pt] (a) {};
\draw [thick,postaction={rmid arrow={black,scale=1.5}}] (0,0) to (120:3) node (b) {};
\draw [thick,postaction={rmid arrow={black,scale=1.5}}] (0,0) to (180:3) node (c) {};
\draw [thick,postaction={rmid arrow={black,scale=1.5}}] (0,0) to (-120:3) node (d) {};
\node at (a) [below left] {$\msf \Sigma_0$};
\node at (b) [above] {$\msf \Sigma_1$};
\node at (c) [below] {$\msf \Sigma_2$};
\node at (d) [below] {$\msf \Sigma_3$};
\node (0:60:0.5) [yshift=12pt,xshift=16pt] {$\scriptstyle \frac{2\pi}{3}$};
\draw[thick] (0.5,0) arc (0:120:0.5);
\end{tikzpicture}
\caption{The contours $\msf \Sigma_0,\msf \Sigma_1,\msf \Sigma_2$ and $\msf \Sigma_3$ in \eqref{def:contour_sigma} that constitute $\msf \Sigma$.}\label{fig:sigmaj}
\end{figure}
The model RHP we are about to introduce depends on a function $\mathsf h:\msf \Sigma\to \mathbb{C}$ used to describe its jump. For the moment we assume
$$\mathsf h\in C^\infty(\msf \Sigma), \quad \mathsf h(z)\in \mathbb{R} \text{ for }z\in \mathbb{R},\quad \text{and}\quad \liminf_{\substack{z\to \infty \\ z\in \msf \Sigma}} \frac{\re \mathsf h(z)}{|z|}>0.
$$
These conditions are present only to ensure the RHP below is well posed and are far from optimal, but enough for our purposes. Later on we will impose more conditions on this function $\mathsf h$, these conditions will be tailored to our later needs regarding the asymptotic analysis of OPs.
The associated RHP asks for finding a $2\times 2$ matrix-valued function $\bm \Phi$ with the following properties.
\begin{enumerate}[\bf $\bm \Phi$-1.]
\item The matrix $\bm \Phi=\bm \Phi(\cdot \mid \mathsf h):\mathbb{C}\setminus \msf \Sigma\to \mathbb{C}^{2\times 2}$ is analytic.
\item Along the interior of the arcs of $\msf \Sigma$ the function $\bm \Phi$ admits continuous boundary values $\bm \Phi_\pm$
related by $\bm \Phi_+(\zeta)=\bm \Phi_-(\zeta)\bm J_{\bm \Phi}(\zeta)$, $\zeta\in \msf \Sigma$, with
\begin{equation}\label{eq:jumpPhimodel}
\bm J_{\bm \Phi}(\zeta)\deff
\begin{dcases}
\bm I+\frac{1}{1+\ee^{-\mathsf h(\zeta)}} \bm E_{12}, & \zeta \in \msf \Sigma_0 , \\
\bm I+(1+\ee^{-\mathsf h(\zeta)})\bm E_{21}, & \zeta \in \msf \Sigma_1\cup\msf \Sigma_3, \\
\frac{1}{1+\ee^{-\mathsf h(\zeta)}}\bm E_{12}-(1+\ee^{-\mathsf h(\zeta)})\bm E_{21}, & \zeta\in\msf \Sigma_2.
\end{dcases}
\end{equation}
\item As $\zeta\to\infty$,
\begin{equation}\label{eq:ModelRHPAsymp}
\bm \Phi(\zeta)=\left(\bm I+\frac{1}{\zeta}\bm \Phi^{(1)}+\mathcal{O}(1/\zeta^{2})\right)\zeta^{\sp_3/4}\bm U_0^{-1}
\ee^{-\frac{2}{3}\zeta^{3/2}\sp_3},
\end{equation}
where
\begin{equation}\label{def:matrixU0modelprobl}
\bm U_0\deff \frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & \mathrm{i} \\ \mathrm{i} & 1
\end{pmatrix}
\end{equation}
and $\bm \Phi^{(1)}=\bm \Phi^{(1)}(\mathsf h)$ is a matrix that depends on the choice of function $\mathsf h$ but it is independent of $\zeta$.
\item The matrix $\bm \Phi$ remains bounded as $\zeta \to 0$.
\end{enumerate}
Given $\mathsf h$, it is not at all obvious that the RHP above has a solution and how to describe it. We study this model problem when $\mathsf h=\mathsf h_\tau$ depends on an additional large parameter $\tau$, in a way that appears naturally in the asymptotic analysis of the orthogonal polynomials mentioned earlier. For large values of $\tau$, we then prove that the solution $\bm \Phi$ exists and is asymptotically close to a model RHP that appeared recently \cite{CafassoClaeys2021} and that we discuss in a moment.
\subsection{The model RHP with admissible data}\hfill
For us, we need to consider the model problem $\bm \Phi=\bm \Phi(\cdot\mid \mathsf h)$ with functions $\mathsf h=\mathsf h_\tau$ satisfying certain properties which are formally introduced in the next definition.
\begin{definition}\label{deff:admissibleh}
We call a function $\mathsf h_\tau:\Sigma\to \mathbb{C}$ {\it admissible} if it is of the form
$$
\mathsf h_\tau(\zeta)=\mathsf h_\tau(\zeta\mid \msf s)=\msf s+\tau\mathsf H\left(\frac{\mathsf \zeta}{\tau}\right),\quad \zeta\in \mathbb{C},\quad \tau>0,\; \msf s\in \mathbb{R},
$$
where $\mathsf H$ is defined on a neighborhood $\mathcal S$ of $\Sigma$ and satisfies the following properties.
\begin{enumerate}[(i)]
\item The function $\mathsf H$ is independent of $\tau$ and $\msf s$, of class $C^\infty$ on $\mathcal S$ and real-valued along $\mathbb{R}$.
\item $\mathsf H$ is analytic on a disk $D_\delta(0)\subset\mathcal S$ centered at the origin, and its unique zero on $D_\delta(0)$ is at $\zeta=0$, with
$$
\msf t\deff -\mathsf H'(0)>0.
$$
\item There exist constants $\eta,\widehat\eta>0$ for which
$$
\re \mathsf H(w)>\eta|w| \quad \text{for } w\in \msf \Sigma_1\cup\msf \Sigma_2\cup\msf \Sigma_3,
$$
and
$$
-\widehat\eta w^{3/2-\epsilon}<\mathsf H(w)<-\eta w \quad \text{for } w\in \msf \Sigma_0,
$$
for some $\epsilon\in (0,1/2]$.
\end{enumerate}
\end{definition}
Conditions (i)--(ii), and also the bounds in (iii) involving $\eta$, are natural in our setup. The bound $\mathsf H(w)>-\widehat\eta w^{3/2-\epsilon}$ is present for technical reasons, and it plays a role only for the proof of Lemma~\ref{lem:estJpsiadcSig0}, allowing us to write certain estimates in a cleaner matter. It could be removed, at the cost of slightly more complicated error terms in the mentioned Lemma. For our purposes, namely to use $\bm \Phi=\bm\Phi(\cdot\mid \mathsf h_\tau)$ as a local parametrix with an appropriate $\mathsf h_{\tau}$, this condition is satisfied anyway (this will be accomplished in Proposition~\ref{prop:ConformExtH}), so we include it in our definition here as well, as it simplifies our analysis.
In the course of the analysis for the RHP for the orthogonal polynomials discussed in Section~\ref{sec:rhpapproachops}, the function $\mathsf H$ will be a transformation of the function $Q$ appearing in the deformed weight \eqref{def:perturbedweight}, and the parameter $\msf t$ that we defined here will play the same role as the one in the definition \eqref{deff:tQprime}.
Given an admissible $\mathsf h_\tau$, we denote
\begin{equation}\label{deff:Phit}
\bm \Phi_\tau(\zeta)\deff \bm \Phi(\zeta\mid \mathsf h=\mathsf h_\tau(\cdot\mid \msf s)).
\end{equation}
We are interested in the asymptotic analysis for $\bm \Phi_\tau$ as $\tau\to +\infty$ and $\msf s\geq -\msf s_0$, for any $\msf s_0>0$, and $\msf t>0$ kept fixed within a compact of the positive axis.
We now explain in an ad hoc manner the appearance of a RHP for the integro-differential equation, which also relates to the KPZ equation. Definition~\ref{deff:admissibleh}--(ii) gives that $\mathsf H$ has an expansion of the form
$$
\mathsf H(\zeta)=-\msf t\zeta(1+\mathcal{O}(\zeta)),\quad |\zeta|\leq \delta,
$$
This means that any admissible function $\mathsf h_\tau$ satisfies
$$
\mathsf h_\tau(\zeta)=\msf s-\msf t\zeta\left(1+\mathcal{O}(\zeta \tau^{-1})\right),\quad |\zeta|\leq \delta \tau.
$$
In particular, the convergence
\begin{equation}\label{deff:h0}
\mathsf h_\tau(\zeta)\to \mathsf h_0(\zeta)=\mathsf h_0(\zeta\mid \msf s,\msf t)\deff \msf s-\msf t\zeta,
\end{equation}
holds true uniformly in compacts as $\tau\to\infty$. This indicates that the solution $\bm \Phi_\tau$ should converge to the solution
\begin{equation}\label{deff:Phicc}
{\bm \Phi}_0\deff \bm\Phi(\cdot\mid \mathsf h=\mathsf h_0)
\end{equation}
of the model problem obtained from $\mathsf h_0$. The RHP-{${\bm \Phi}_0$} relates to the integro-differential PII and is a rescaled version of an RHP that appears in the description of the narrow wedge solution to the KPZ equation, as we discuss in the next section in detail.
\section{The RHP for the integro-differential RHP}\label{sec:KPZRHP}
For the choice
\begin{equation}\label{deff:hkpz}
\msf h^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta)=\msf h^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta\mid s,T)\deff -T^{1/3}(s+\zeta)
\end{equation}
the corresponding solution of the RHP--{$\bm\Phi$}
$$
{\bm \Phi}^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta)={\bm \Phi}^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta\mid s,T)\deff \bm\Phi(\zeta\mid \mathsf h=\msf h^{\mathrm{\scriptscriptstyle (KPZ)}}(\cdot \mid s,T))
$$
appeared for the first time in the work of Cafasso and Clayes \cite{CafassoClaeys2021} (this is the RHP-$\Psi$ in Section~2 therein) in connection with the narrow wedge solution to the KPZ equation as we explain in a moment, in Section~\ref{sec:KPZRHPKPZ}. To avoid confusion with the related quantities that we are about to introduce, we term it the KPZ RHP. In virtue of the identity
$$
\mathsf h_0(\zeta\mid \msf s,\msf t)=\msf h^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta\mid s=-\msf s/\msf t, T=\msf t^3)
$$
which follows from \eqref{deff:h0} and \eqref{deff:hkpz}, we also have the correspondence
\begin{equation}\label{eq:PhiccPhikpz}
{\bm \Phi}_0(\zeta\mid \msf s,\msf t)={\bm \Phi}^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta\mid s=-\msf s/\msf t,T=\msf t^3),
\end{equation}
and we refer to ${\bm \Phi}_0$ as the id-PII RHP. For the record, we state the existence of ${\bm \Phi}_0$ formally as a result.
\begin{prop}\label{prop:existencePhicc}
For any $\msf s\in \mathbb{R}$ and any $\msf t>0$, the solution ${\bm \Phi}_0$ exists and is unique. Furthermore, for any fixed $\msf s_0>0$ and $\msf t_0\in (0,1)$, the solution ${\bm \Phi}_{0,+}(\zeta)$ remains bounded for $\zeta$ in compacts of $\mathbb{R}$ and $\msf s\geq -\msf s_0$, $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{prop}
\begin{proof}
It is a consequence of \cite[Section~2]{CafassoClaeys2021} that the solution ${\bm \Phi}^{\mathrm{\scriptscriptstyle (KPZ)}}(\cdot\mid s,T)$ exists and is unique, for any $s\in \mathbb{R}$ and $T>0$, and from the correspondence \eqref{eq:PhiccPhikpz} the existence and uniqueness of ${\bm \Phi}_0$ is thus granted.
For the boundedness, we start from the representation
$$
{\bm \Phi}_0(\zeta)=\bm I+\frac{1}{2\pi \mathrm{i}}\int_{\Gamma}\frac{{\bm \Phi}_{0,-}(s)(\bm J_{{\bm \Phi}_0}(s)-\bm I)}{s-\zeta}\mathrm{d} s,\quad \zeta\in \mathbb{C}\setminus \Gamma,
$$
which follows from the $L^p$ theory of RHPs (see \cite{DeiftParkCity}). The jump matrix admits an analytic continuation to any neighborhood of the real axis, and this analytic continuation remains bounded in compacts, also uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$ (see for instance \eqref{eq:jumpPhicc} for the exact expression). With these observations in mind, the claimed boundedness follows from standard arguments. We skip additional details, but refer to the proof of Theorem~\ref{thm:PhitauAsympt}, in particular \eqref{eq:maxprincipleconseqPsitauad} {\it et seq.}, for similar arguments in a more involved context.
\end{proof}
In this section we collect several results on ${\bm \Phi}_0$ that were obtained in \cite{CafassoClaeys2021,CafassoClaeysRuzza2021} and which will be needed later.
But before proceeding, a word of caution. As we said, the RHP--{${\bm \Phi}^{\mathrm{\scriptscriptstyle (KPZ)}}$} appeared first in \cite{CafassoClaeys2021}, but was also studied in the subsequent work \cite{CafassoClaeysRuzza2021}. The meanings for the variables $s,x$ and $t$ in these two works are not consistent, but we need results from both of them. Comparing to the work \cite{CafassoClaeys2021} by Cafasso and Claeys, the correspondence is
\begin{equation}\label{notation:cc}
s_{\rm CC}=-\frac{\msf s}{\msf t}\quad \text{and}\quad T_{\rm CC}=\msf t^3.
\end{equation}
This correspondence is consistent with \eqref{eq:PhiccPhikpz}. On the other hand, when comparing to the subsequent work \cite{CafassoClaeysRuzza2021} by Cafasso, Claeys and Ruzza, the correspondence between notations is
\begin{equation}\label{notation:ccr}
t_{\rm CCR}=\frac{1}{\msf t^{3/2}}, \quad x_{\rm CCR}=-\frac{\msf s}{\msf t^{3/2}},\qquad \text{that is}\qquad x_{\rm CCR}=-\msf S,\quad t_{\rm CCR}=\msf T,
\end{equation}
where $\msf T,\msf S$ are as in \eqref{eq:corr_tad_Tad}.
In our asymptotic analysis, the most convenient choice of variables to work with is the choice $(\msf s,\msf t)$ and the correspondence $(\msf S,\msf T)$ from \eqref{eq:corr_tad_Tad} that we have already been using, and which leads to the RHP ${\bm \Phi}_0$ as we introduced. Nevertheless, we will need to collect results from both mentioned works, and when the need arises we refer to the correspondences of variables \eqref{notation:cc}--\eqref{notation:ccr}.
On the other hand, when making correspondence with integrable systems, in particular the integro-differential Painlevé II equation, it is more convenient to work with the variables $\msf S$ and $\msf T$ as in \eqref{eq:corr_tad_Tad}.
\subsection{Properties of the id-PII parametrix}\label{sec:KPZRHPKPZ}\hfill
In this section we describe many of the findings from \cite{CafassoClaeys2021,CafassoClaeysRuzza2021}, in a way suitably adapted to our notation and needs. In particular the connection of ${\bm \Phi}_0$ introduced in \eqref{deff:Phicc} with the integro-differential Painlevé II equation is described in this section.
For
\begin{equation}\label{deff:Dkpz}
{\bm\Delta}^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta)={\bm\Delta}^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta\mid s,T)\deff
\begin{dcases}
\bm I, & \zeta>0, \\
\bm I+(1+\ee^{T^{1/3}(\zeta+s)})\bm E_{21}, & \zeta<0.
\end{dcases}
\end{equation}
the identity
\begin{equation}\label{eq:QccIntRepPhikpz}
\partial_s \log \msf L^{\kern -0.15em \ai}(s,T)=\frac{T^{1/3}}{2\pi \mathrm{i}}\int_{-\infty}^\infty \frac{\ee^{T^{1/3}(x+s)}}{(1+\ee^{T^{1/3}(x+s)})^2}\left[({\bm\Delta}^{\mathrm{\scriptscriptstyle (KPZ)}}(x)^{-1}{\bm \Phi}^{\mathrm{\scriptscriptstyle (KPZ)}}_+(x)^{-1}({\bm \Phi}^{\mathrm{\scriptscriptstyle (KPZ)}}_+ {\bm\Delta}^{\mathrm{\scriptscriptstyle (KPZ)}} )'(x)\right]_{21} \mathrm{d} x,
\end{equation}
was shown in \cite[Theorem~2.1]{CafassoClaeys2021} and will also be useful for us. With \eqref{eq:PhiccPhikpz} we now rewrite this identity in terms of ${\bm \Phi}_0$.
With the principal branch of the argument, set
\begin{equation}\label{deff:Dcc}
{\bm\Delta}_0(\zeta)={\bm\Delta}_0(\zeta\mid \msf s, \msf t)\deff
\begin{cases}
\bm I, & |\arg\zeta|<\frac{2\pi}{3}, \\
\bm I+(1+\ee^{-\msf s+\msf t\zeta})\bm E_{21}, & |\arg\zeta|>\frac{2\pi}{3}.
\end{cases}
\end{equation}
This function relates to ${\bm\Delta}^{\mathrm{\scriptscriptstyle (KPZ)}}$ in \eqref{deff:Dkpz} via
$$
{\bm\Delta}_{0,+}(\zeta)={\bm\Delta}^{\mathrm{\scriptscriptstyle (KPZ)}}(\zeta\mid s=-\msf s/\msf t, T=\msf t^3),\quad \zeta\in \mathbb{R},
$$
and \eqref{eq:QccIntRepPhikpz} rewrites as
\begin{equation}\label{eq:QccIntRepPhicc}
\partial_s \log \msf L^{\kern -0.15em \ai}(s=-\msf s/\msf t,T=\msf t^3)=\frac{\msf t}{2\pi \mathrm{i}}\int_{-\infty}^\infty \frac{\ee^{\msf t x-\msf s}}{(1+\ee^{\msf t x-\msf s})^2}\left[({\bm\Delta}_0(x)^{-1}{\bm \Phi}_{0,+}(x)^{-1}({\bm \Phi}_{0,+} {\bm\Delta}_0 )'(x)\right]_{21} \mathrm{d} x.
\end{equation}
For further reference, it is now convenient to state the RHP for ${\bm \Phi}_0$ explicitly.
\begin{enumerate}[\bf ${\bm \Phi}_0$-1.]
\item The matrix ${\bm \Phi}_0:\mathbb{C}\setminus \msf \Sigma\to \mathbb{C}^{2\times 2}$ is analytic.
\item Along the interior of the arcs of $\msf \Sigma$ the function ${\bm \Phi}_0$ admits continuous boundary values ${\bm \Phi}_{0,\pm}$
related by ${\bm \Phi}_{0,+}(\zeta)={\bm \Phi}_{0,-}(\zeta)\bm J_{{\bm \Phi}_0}(\zeta)$, $\zeta\in \msf \Sigma$, with
\begin{equation}\label{eq:jumpPhicc}
\bm J_{{\bm \Phi}_0}(\zeta)\deff
\begin{dcases}
\bm I+\frac{1}{1+\ee^{-\msf s+\msf t \zeta}} \bm E_{12}, & \zeta \in \msf \Sigma_0 , \\
\bm I+(1+\ee^{-\msf s+\msf t\zeta})\bm E_{21}, & \zeta \in \msf \Sigma_1\cup\msf \Sigma_3, \\
\frac{1}{1+\ee^{-\msf s+\msf t\zeta}}\bm E_{12}-(1+\ee^{-\msf s+\msf t\zeta})\bm E_{21}, & \zeta\in\msf \Sigma_2.
\end{dcases}
\end{equation}
\item As $\zeta\to\infty$,
\begin{equation}\label{eq:AsymptPhicc}
{\bm \Phi}_0(\zeta)=\left(\bm I+\mathcal{O}(1/\zeta)\right)\zeta^{\sp_3/4}\bm U_0^{-1}
\ee^{-\frac{2}{3}\zeta^{3/2}\sigma_3}.
\end{equation}
where we recall that $\bm U_0$ is given in \eqref{def:matrixU0modelprobl}.
\item The matrix ${\bm \Phi}_0$ remains bounded as $\zeta \to 0$.
\end{enumerate}
To compare with \cite{CafassoClaeysRuzza2021} we perform a transformation of this RHP. All the calculations that follow already take into account the correspondence \eqref{notation:ccr} between the notation in the mentioned work and our notation.
With ${\bm\Delta}_0$ as in \eqref{deff:Dcc} and introducing
$$
\xi=\xi(\zeta)=-\msf s+\msf t\zeta,\quad \text{with inverse}\quad \zeta=\zeta(\xi)=\frac{\xi+\msf s}{\msf t},
$$
we transform
\begin{equation}\label{eq:transfPsiccPhicc}
{\bm \Psi}_0(\xi)=\left(\bm I+\frac{\mathrm{i} \msf s^2}{4\msf t^{3/2}}\bm E_{12}\right)\msf t^{\sp_3/4}{\bm \Phi}_0(\zeta(\xi))\times
\begin{cases}
{\bm\Delta}_0(\zeta(\xi)), & \im \xi>0, \; \arg(\zeta(\xi))\neq 2\pi/3,\\
{\bm\Delta}_0(\zeta(\xi))^{-1}, & \im \xi<0, \; \arg(\zeta(\xi))\neq -2\pi/3.
\end{cases}
\end{equation}
Then ${\bm \Psi}_0$ satisfies the following RHP.
\begin{enumerate}[\bf ${\bm \Psi}_0$-1.]
\item The matrix ${\bm \Psi}_0:\mathbb{C}\setminus \mathbb{R}\to \mathbb{C}^{2\times 2}$ is analytic.
\item Along $\mathbb{R}$ the function ${\bm \Psi}_0$ admits continuous boundary values ${\bm \Psi}_{0,\pm}$
related by
%
$$
{\bm \Psi}_{0,+}(\xi)={\bm \Psi}_{0,-}(\xi)
\left(\bm I+\frac{1}{1+\ee^{\xi}} \bm E_{12}\right), \quad \xi\in \mathbb{R}.
$$
\item For any $\delta\in (0,2\pi/3)$, as $\xi\to\infty$ the matrix ${\bm \Psi}_0$ has the following asymptotic behavior,
\begin{equation}\label{eq:RHPAsympPsicc}
{\bm \Psi}_0(\xi)=\left(\bm I+\mathcal{O}(1/\xi)\right)\xi^{\sp_3/4}\bm U_0^{-1}
\ee^{-\msf t^{-3/2}\left(\frac{2}{3}\xi^{3/2}+\msf s \xi^{1/2}\right)\sigma_3}
\times
\begin{cases}
\bm I, & |\arg \xi|\leq \pi-\delta, \\
\bm I\pm \bm E_{21}, & \pi-\delta<\pm \arg \xi <\pi.
\end{cases}
\end{equation}
\end{enumerate}
This RHP is the same RHP considered in \cite[page~1120]{CafassoClaeysRuzza2021}\footnote{The keen reader will notice that there is a sign difference between the last term in the right-hand side of \eqref{eq:ModelRHPAsymp} and the corresponding term in \cite[page~1120]{CafassoClaeysRuzza2021}, but the latter is a typo.} with the choice $\sigma(r)=(1+\ee^{-r})^{-1}$ therein and the correspondence of variables \eqref{notation:ccr}.
As a consequence, and with the change of variables $(\msf s,\msf t)\mapsto (\msf S,\msf T)$ from \eqref{eq:corr_tad_Tad}, we obtain that for some functions $\mathsf Q=\mathsf Q(\msf S,\msf T),\mathsf R=\mathsf R(\msf S,\msf T), \mathsf P=\mathsf P(\msf S,\msf T)$ and
\begin{equation}\label{deff:coeffqrp}
\mathsf q=\mathsf q(\msf s,\msf t)=\mathsf Q(\msf S,\msf T),\quad \mathsf r=\mathsf r(\msf s,\msf t)=\mathsf R(\msf S,\msf T),\quad \mathsf p=\mathsf p(\msf s,\msf t)=\mathsf P(\msf S,\msf T),
\end{equation}
the asymptotic behavior \eqref{eq:RHPAsympPsicc} improves to
\begin{multline}\label{eq:AsympPsiccImprov}
{\bm \Psi}_0(\xi)=\left(\bm I+
\frac{1}{\xi}
\begin{pmatrix}
\mathsf q & -\mathrm{i} \mathsf r \\ \mathrm{i} \mathsf p & -\mathsf q
\end{pmatrix}+
\mathcal{O}(\xi^{-2})\right)\xi^{\sp_3/4}\bm U_0^{-1}\\
\times
\ee^{-\msf t^{-3/2}\left(\frac{2}{3}\xi^{3/2}+\msf s \xi^{1/2}\right)\sigma_3}
\times
\begin{cases}
\bm I, & |\arg \xi|\leq \pi-\delta, \\
\bm I\pm \bm E_{21}, & \pi-\delta<\pm \arg \xi <\pi,
\end{cases}
\quad \xi\to\infty.
\end{multline}
Stressing that the correspondence \eqref{notation:ccr} is in place, the functions $\mathsf P$ and $\mathsf Q$ satisfy the relation \cite[Equation~(3.14)]{CafassoClaeysRuzza2021}
$$
\partial_\msf S \mathsf P(\msf S,\msf T)=2\mathsf Q(\msf S,\msf T)+\mathsf P(\msf S,\msf T)^2.
$$
Furthermore, from \cite[Equations~(3.12),(3.16), Theorem~1.3 and Corollary~1.4]{CafassoClaeysRuzza2021} we see that ${\bm \Psi}_0$ takes the form
\begin{equation}\label{eq:transfPsiccphiad}
{\bm \Psi}_0(\xi\mid \msf s,\msf t)=
\sqrt{2\pi}\ee^{-\frac{\pi\mathrm{i}}{4}\sp_3}\left(\bm I- \mathsf p(\msf s,\msf t)\bm E_{12}\right)
\begin{pmatrix}
-\partial_\msf S \Phi (\xi\mid \msf S,\msf T) & \ast \\
-\Phi(\xi\mid \msf S,\msf T) & \ast
\end{pmatrix}
\ee^{\frac{\pi\mathrm{i}}{4}\sp_3},
\end{equation}
where $\Phi=\Phi(\xi\mid \msf S,\msf T)$ solves the NLS equation with potential $2\partial_\msf S \mathsf P$,
$$
\partial_\msf S^2\Phi(\xi\mid \msf S,\msf T)=(\xi+2\partial_\msf S \mathsf P(\msf S,\msf T))\Phi(\xi\mid \msf S,\msf T).
$$
In addition, $\mathsf P$ and $\Phi$ are related through the identity \eqref{eq:intdiffPIIPhip} which, in turn, implies that $\Phi$ is the solution to the integro-differential Painlevé II equation in \eqref{eq:intdiffPII}.
It is convenient to write some quantities of ${\bm \Phi}_0$ directly in terms of the just introduced functions.
The first identity we need for later is
\begin{multline}\label{eq:transfPhicckernelphicc}
\left[\left({\bm \Phi}_0(\zeta(v)\mid \msf s,\msf t){\bm\Delta}_0(\zeta(v)\mid \msf s,\msf t)\right)^{-1}{\bm \Phi}_0(\zeta(u)\mid \msf s,\msf t){\bm\Delta}_0(\zeta(u)\mid \msf s,\msf t)\right]_{21,+} \\
=-2\pi\mathrm{i} \left(\Phi(u\mid \msf S,\msf T)(\partial_{\msf S}\Phi)(v\mid \msf S,\msf T)-\Phi(v\mid \msf S,\msf T)(\partial_{\msf S}\Phi)(u\mid \msf S,\msf T)\right)
\end{multline}
which follows from \eqref{eq:transfPsiccPhicc} and \eqref{eq:transfPsiccphiad} after a straightforward calculation, accounting also that $\det{\bm \Phi}_0=\det {\bm \Psi}_0\equiv 1$.
The second relation we need is an improvement of the asymptotics of ${\bm \Phi}_0$ in \eqref{eq:AsymptPhicc}. With the coefficients
$$
\mathsf c_1=\mathsf c_1(\msf s,\msf t)\deff -\frac{\msf s^2}{4\msf t^{3/2}},\quad \mathsf c_2=\mathsf c_2(\msf s,\msf t)\deff \frac{\msf s^4}{32\msf t^3},\quad
\mathsf c_3=\mathsf c_3(\msf s,\msf t)\deff -\frac{\msf s^3(\msf s^3-16\msf t^3)}{384\msf t^{9/2}},
$$
and the functions $\mathsf q,\mathsf r,\mathsf p$ in \eqref{deff:coeffqrp}, introduce
\begin{equation}\label{deff:Phikpz_residue}
{\bm \Phi}_0^{(1)}\deff \frac{1}{\msf t}
\begin{pmatrix}
-\dfrac{\msf s}{4}+\mathsf q+\mathsf c_2-\mathsf c_1\mathsf p-\mathsf c_1^2 & \mathrm{i} \msf t^{-1/2}\left(-\mathsf r-2\mathsf q\mathsf c_1+\dfrac{\msf s}{2}\mathsf c_1+\mathsf p\mathsf c_1^2+\mathsf c_1\mathsf c_2-\mathsf c_3\right) \\
\mathrm{i} \msf t^{1/2}(\mathsf p+\mathsf c_1) & \dfrac{\msf s}{4}-\mathsf q+\mathsf p\mathsf c_1+\mathsf c_2
\end{pmatrix}.
\end{equation}
After some cumbersome but straightforward calculations, the asymptotics \eqref{eq:AsympPsiccImprov} improves \eqref{eq:AsymptPhicc} to
$$
{\bm \Phi}_0(\zeta)=\left(\bm I+\frac{{\bm \Phi}_0^{(1)}}{\zeta}+\mathcal{O}(\zeta^{-2})\right)\zeta^{\sp_3/4}\bm U_0^{-1}\ee^{-\frac{2}{3}\zeta^{3/2}\sp_3},\quad \zeta\to \infty.
$$
\section{Bounds on the id-PII RHP}\label{sec:boundsKPZRHP}
We need to obtain certain asymptotic bounds on ${\bm \Phi}_0$ in different regimes. These bounds will be used later to show that the model problem $\bm\Phi_\tau$ converges, as $\tau\to+\infty$, to ${\bm \Phi}_0$ as already indicated in \eqref{deff:Phit} {\it et seq.} We split these necessary estimates in the next subsections, depending on the regime we are.
In what follows, for a matrix-valued function $\bm M=(\bm M_{jk})$ and a contour $\Sigma\subset \mathbb{C}$, we also use the pointwise matrix norm
\begin{equation}\label{deff:matrixnorm}
|\bm M(\zeta)|\deff \max_{j,k} |\bm M_{j,k}(\zeta)|,
\end{equation}
and the matrix $L^p$ norm (possibly also with $p=\infty$)
\begin{equation}\label{deff:matrixLpnorm}
\|\bm M\|_{L^p(\Sigma)}\deff \max_{j,k} \| \bm M_{j,k}\|_{L^p(\Sigma)},
\end{equation}
where the measure is always understood to be the arc-length measure. In particular, for any two given matrices $\bm M_1$ and $\bm M_2$ the inequality
$$
\|\bm M_1\bm M_2\|_{L^\infty(\Sigma)}\leq 2 \|\bm M_1\|_{L^\infty(\Sigma)}\|\bm M_2\|_{L^\infty(\Sigma)}
$$
is satisfied. Similar straightforward inequalities involving $L^1,L^2$ and $L^\infty$ and the pointwise norm \eqref{deff:matrixnorm} also hold, and will be used without further mention. Sometimes we also write
\begin{equation}\label{deff:matrixLpqnorm}
\|\bm M\|_{L^p\cap L^q(\Sigma)}\deff \max \left\{\| \bm M\|_{L^p(\Sigma)}, \| \bm M\|_{L^q(\Sigma)} \right\},
\end{equation}
to identify that possible convergences are taking place in various norms simultaneously. In a similar manner we define norms $\|\cdot\|_{L^{p_1}\cap L^{p_2}\cap L^{p_3}}$ involving three (or possibly more) function spaces.
\subsection{The singular regime}\label{sec:Phiccsingregime}\hfill
The first asymptotic regime we consider is
$$
\msf s\geq \msf s_0 \quad \text{and}\quad \msf t_0\leq \msf t \leq \frac{1}{\msf t_0},
$$
where $\msf t_0\in (0,1)$ is any given value, and $\msf s_0=\msf s_0(\msf t_0)>0$ will be made sufficiently large depending on $\msf t_0>0$, but independent of $\msf t$ within the range above. With \eqref{notation:ccr} in mind, this is a particular case of the {\it singular regime} in \cite{CafassoClaeysRuzza2021}.
For this asymptotic regime, we need the following result.
\begin{prop}\label{prop:EstPhiccSing}
For any $\msf t_0\in (0,1)$ there exists $\msf s_0=\msf s_0(\msf t_0)>0$, $M=M(\msf t_0)>0$ and $\eta=\eta(\msf t_0)>0$ such that the inequalities
\begin{align*}
& \left|{\bm \Phi}_{0,+}(\zeta)\bm E_{12}{\bm \Phi}_{0,+}(\zeta)^{-1}\right|\leq M\ee^{-\eta \re(\zeta^{3/2})}, && \zeta\in \msf \Sigma_0, \\
& \left|{\bm \Phi}_{0,+}(\zeta)\bm E_{21}{\bm \Phi}_{0,+}(\zeta)^{-1}\right|\leq M\ee^{-\eta \re(\zeta^{3/2})}, && \zeta\in \msf \Sigma_1\cup\msf \Sigma_3, \quad \text{and}\\
& \left|{\bm \Phi}_{0,+}(\zeta)\bm E_{22}{\bm \Phi}_{0,+}(\zeta)^{-1}\right|\leq M|\zeta|^{1/2}, && \zeta\in \msf \Sigma_2,
\end{align*}
hold true for any $\msf s\geq \msf s_0$ and any $\msf t\in [\msf t_0,1/\msf t_0]$.
\end{prop}
The proof of Proposition~\ref{prop:EstPhiccNonasymp} is a recollection of the analysis in \cite{CafassoClaeysRuzza2021}, so before going into the details we need to review some further notions from their work.
Introduce
\begin{equation}\label{deff:Phiai}
{\bm \Phi}_{\rm Ai}(\zeta)\deff-\sqrt{2\pi}
\times
\begin{cases}
\begin{pmatrix}
\ai'(\zeta) & -\ee^{2\pi \mathrm{i}/3}\ai'(\ee^{-2\pi \mathrm{i}/3}\zeta ) \\
\mathrm{i} \ai(\zeta) & -\mathrm{i}\ee^{-2\pi \mathrm{i}/3}\ai(\ee^{-2\pi \mathrm{i}/3}\zeta)
\end{pmatrix}, & \im \zeta>0, \\
\begin{pmatrix}
\ai'(\zeta) & \ee^{-2\pi \mathrm{i}/3}\ai'(\ee^{2\pi \mathrm{i}/3}\zeta ) \\
\mathrm{i} \ai(\zeta) & \mathrm{i}\ee^{2\pi \mathrm{i}/3}\ai(\ee^{2\pi \mathrm{i}/3}\zeta)
\end{pmatrix}, & \im \zeta<0.
\end{cases}
\end{equation}
This is the matrix appearing in \cite[Equation~(2.5)]{CafassoClaeysRuzza2021}. With the correspondence of variables \eqref{notation:ccr} in mind, when we combine our identity \eqref{eq:transfPsiccPhicc} with \cite[Equation~(2.8)]{CafassoClaeysRuzza2021}, we obtain the equality
\begin{equation}\label{eq:IdentPhiccPhiaiY}
{\bm \Phi}_0(\zeta)=\bm Y(\zeta){\bm \Phi}_{\rm Ai}(\zeta)
\times
\begin{cases}
\bm I, & -2\pi/3<\arg \zeta<2\pi/3,\\
\bm I\mp (1+\ee^{-\msf s+\msf t \zeta})\bm E_{21}, & 2\pi /3<\pm \arg \zeta <\pi.
\end{cases}
\end{equation}
The exact form of the matrix $\bm Y$ is not important for us, but we can interpret this last equality as a defining identity for $\bm Y$. What is important for us is that $\bm Y$ is analytic off the real axis, with a jump matrix $\bm J_{\bm Y}$ on $\mathbb{R}$ which admits an analytic continuation to a neighborhood of the axis.
The small norm theory for $\bm Y$ in our regime of interest was carried out in \cite[Lemma~5.1 and Section~5.2]{CafassoClaeysRuzza2021}. As a consequence, we obtain that for any $\msf t_0>0$ there exist $M=M(\msf t_0)>0,\msf s_0=\msf s_0(\msf t_0)>0,\eta=\eta(\msf t_0)>0$ such that the inequalities
$$
\|\bm J_{\bm Y}-\bm I\|_{L^{2}\cap L^\infty\cap L^1(\mathbb{R})}\leq M\ee^{-\eta s},\quad \text{and}\quad \|\bm Y_\pm-\bm I\|_{L^{2}\cap L^\infty\cap L^1(\mathbb{R})}\leq M\ee^{-\eta s}
$$
hold for any $\msf s\geq \msf s_0$ and any $\msf t \in [\msf t_0,1/\msf t_0]$. Also as a consequence of the small norm theory, we obtain the expression
$$
\bm Y(\zeta)=\bm I+\frac{1}{2\pi \mathrm{i}}\int_\mathbb{R} \frac{\bm Y_-(x)(\bm J_{\bm Y}(x)-\bm I)}{x-\zeta} \mathrm{d} x,\quad \zeta\in \mathbb{C}\setminus \mathbb{R}.
$$
We combine this last identity with the fact that $\bm J_{\bm Y}$ admits an analytic continuation in a neighborhood of $\mathbb{R}$, and learn that there exists $M=M(\mathsf t_0)>0$ for which
\begin{equation}\label{eq:boundbmY}
|\bm Y(\zeta)^{\pm 1}|\leq M,
\end{equation}
for every $\zeta\in \mathbb{C}$, $\msf s\geq \msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$.
\begin{proof}[Proof of Proposition~\ref{prop:EstPhiccSing}]
For $\zeta\in \msf \Sigma_0=(0,\infty)$, we use \eqref{eq:IdentPhiccPhiaiY} and the definition of ${\bm \Phi}_{\rm Ai}$ in
$$
{\bm \Phi}_{0,+}(\zeta)\bm E_{12}{\bm \Phi}_{0,+}(\zeta)^{-1}=
2\pi \bm Y_+(\zeta)
\begin{pmatrix}
-\mathrm{i} \ai(\zeta)\ai'(\zeta) & \ai'(\zeta)^2 \\
\ai(\zeta)^2 & \mathrm{i} \ai(\zeta)\ai'(\zeta)
\end{pmatrix}
\bm Y_+(\zeta)^{-1}
$$
Using the bound \eqref{eq:boundbmY}, the continuity and the known asymptotics as $\zeta\to \infty$ of the Airy function and its derivative, the claim along $\msf \Sigma_0$ follows.
The claim for $\zeta\in \msf \Sigma_j$ with $j=1,2,3$ follows in exactly the same explicit manner, we skip the details.
\end{proof}
\subsection{The non-asymptotic regime}\label{sec:Phiccnonasymreg}\hfill
In the non-asymptotic regime, we fix {\it any} $\msf t_0\in (0,1)$ and $\msf s_0>0$ and seek for bounds of certain entries of ${\bm \Phi}_0$ which are valid uniformly within the range
$$
|\msf s|\leq \msf s_0\quad \text{and}\quad \msf t_0\leq \msf t\leq \frac{1}{\msf t_0}.
$$
For the next result, we recall the matrix norm introduced in \eqref{deff:matrixnorm}.
\begin{prop}\label{prop:EstPhiccNonasymp}
Fix any values $\msf t_0\in (0,1)$ and $\msf s_0>0$. There exist $M=M(\msf s_0,\msf t_0)>0$ and $\eta=\eta(\msf s_0,\msf t_0)>0$ for which the estimates
\begin{align*}
& \left|{\bm \Phi}_{0,+}(\zeta)\bm E_{12}{\bm \Phi}_{0,+}(\zeta)^{-1}\right|\leq M\ee^{-\eta \re(\zeta^{3/2})}, && \zeta\in \msf \Sigma_0, \\
& \left|{\bm \Phi}_{0,+}(\zeta)\bm E_{21}{\bm \Phi}_{0,+}(\zeta)^{-1}\right|\leq M\ee^{-\eta \re(\zeta^{3/2})}, && \zeta\in \msf \Sigma_1\cup\msf \Sigma_3, \quad \text{and}\\
& \left|{\bm \Phi}_{0,+}(\zeta)\bm E_{22}{\bm \Phi}_{0,+}(\zeta)^{-1}\right|\leq M|\zeta|^{1/2}, && \zeta\in \msf \Sigma_2.
\end{align*}
hold true uniformly for $|\msf s|\leq \msf s_0$ and $\msf t_0\leq \msf t \leq \msf t_0^{-1}$.
\end{prop}
\begin{proof}
The asymptotic behavior as $\zeta\to\infty$ in the {\bf RHP}--{${\bm \Phi}_0$} is valid uniformly up to the boundary values ${\bm \Phi}_{0,\pm}$ as well, and also uniformly when the parameters $\msf s$ and $\msf t$ vary within compact sets, implying that
$$
|{\bm \Phi}_{0,+}(\zeta)\bm E_{12}{\bm \Phi}_{0,+}(\zeta)^{-1}|\leq \ee^{-\frac{4}{3}\re (\zeta^{3/2})} | \zeta^{\sp_3/4}\bm U_0^{-1}\bm E_{12}\bm U_0\zeta^{-\sp_3/4} |(1+\mathcal{O}(\zeta^{-1})),\quad \zeta\to\infty.
$$
Combined with the continuity of the boundary value ${\bm \Phi}_{0,+}$ with respect to both $\zeta$ and also $\msf s,\msf t$, the first estimate follows. The remaining estimates are completely analogous.
\end{proof}
\section{Asymptotic analysis for the model problem with admissible data}\label{sec:asympanalymodelprobl}
We now carry out the asymptotic analysis as $\tau\to +\infty$ of $\bm\Phi_\tau$ introduced in \eqref{deff:Phit}. For that, we fix $\msf s_0>0$ and $\msf t_0\in (0,1)$ and work under the assumption that
\begin{equation}\label{scaling:UTregime}
\tau\to +\infty\quad \text{with}\quad \msf s \geq -\msf s_0 \quad \text{and}\quad \msf t_0\leq \msf t \leq \frac{1}{\msf t_0}.
\end{equation}
During this section, $\mathsf h_\tau$ always denotes an admissible function in the sense of Definition~\ref{deff:admissibleh}, and $\bm\Phi_\tau$ is the solution to the associated RHP.
We also talk about uniformity of error terms in the parameter $\msf t$ ranging on a compact interval $K\subset (0,\infty)$, and by this we mean the following. The solution $\bm\Phi_\tau$ depends on the parameter $\msf t$ via the derivative $\mathsf H'(0)=-\msf t$, see Definition~\ref{deff:admissibleh}. We view $\mathsf H=\mathsf H_\msf t$ as varying with $\msf t$ while keeping all the remaining derivatives $\mathsf H^{(k)}(0)$, $k\neq 1$ fixed. By analyticity this determines $\mathsf H$ uniquely at $D_\delta(0)$, but not outside this disk. We then consider $\mathsf H_\msf t$ outside $D_\delta(0)$ to be any extension from $D_\delta(0)$ that satisfies Definition~\ref{deff:admissibleh} with the additional requirement that the constants $\eta, \widehat\eta$ and $\epsilon$ in (iii) may depend on $K$ but are independent of $\msf t\in K$. Of course, for each $\mathsf H_\msf t$ extended this way there corresponds a solution $\bm\Phi_\tau$ of the associated RHP. By uniformity in $\msf t\in K$ we mean that the error may depend on $K$ and the corresponding values $\eta, \widehat\eta$ and $\epsilon$, but is valid for any $\bm\Phi_\tau$ obtained with an extension $\mathsf H_\msf t$ constructed with the explained requirement.
The asymptotic analysis itself makes use of somewhat standard arguments and objects in the RHP literature. Some consequences of this asymptotic analysis will be needed later, and we now state them.
The first such consequence is the existence of a solution with asymptotic formulas relating quantities of interest with the corresponding quantities in the id-PII RHP.
\begin{theorem}\label{thm:PhitauAsympt}
Fix an admissible function $\mathsf h_\tau$ in the sense of Definition~\ref{deff:admissibleh}. There exists $\tau_0=\tau_0(\msf s_0,\msf t_0)>0$ for which for any $\tau\geq \tau_0$ and any $\msf s,\msf t$ as in \eqref{scaling:UTregime}, the RHP for $\bm\Phi(\cdot\mid \mathsf h_\tau)$ admits a unique solution $\bm\Phi=\bm\Phi_\tau$ as in \eqref{deff:Phit}.
Furthermore, for any $\kappa\in (0,1)$, the coefficient $\bm\Phi^{(1)}=\bm\Phi^{(1)}_\tau $ in the asymptotic condition \eqref{eq:ModelRHPAsymp} satisfies
\begin{equation}\label{eq:asympPhin1Phi01}
\bm\Phi^{(1)}_\tau={\bm \Phi}_0^{(1)}+\mathcal{O}(\tau^{-\kappa}),\quad \tau\to +\infty,
\end{equation}
where $\bm\Phi^{(1)}_0$ is as in \eqref{deff:Phikpz_residue} and the error term is uniform for $\msf s,\msf t$ as in \eqref{scaling:UTregime}. Also, still for $\kappa\in (0,1)$ the asymptotic formula
\begin{equation}\label{eq:asympPhitauPhi0}
\bm\Phi_{\tau,+}(x)=\left(\bm I+\mathcal{O}\left(\frac{1}{\tau^{\kappa}(1+|x|)}\right)\right){\bm \Phi}_{0,+}(x),\quad \tau \to +\infty
\end{equation}
holds true uniformly for $x\in \msf \Sigma$ with $|x|\leq \tau^{(1-\kappa)/2}$, and uniformly for $\msf s,\msf t$ as in \eqref{scaling:UTregime}.
\end{theorem}
The second consequence connects the solution $\bm\Phi_\tau$ directly with the statistics $\mathsf Q$. For its statement, set
$$
\bm\Delta_\tau(x)\deff \bm I+(1+\ee^{-\mathsf h_\tau(x)})\chi_{(-\infty,0)}(x)\bm E_{21},\quad x\in \mathbb{R}.
$$
\begin{theorem}\label{thm:integralPhitauPhicc}
Fix $a,b>0$, $\msf s_0>0$ and $\msf t_0\in (0,1)$. For any $\kappa\in (0,1)$, the estimate
\begin{multline*}
\frac{1}{2\pi \mathrm{i}}
\int_{\msf s}^{\infty}\int_{-\tau a}^{\tau b}\frac{\ee^{\mathsf h_\tau(x\mid u)}}{\left(1+\ee^{\mathsf h_\tau(x\mid u)}\right)^2}\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}\mathrm{d} x \mathrm{d} u
=\\
-\log\msf L^{\kern -0.15em \ai}(-\msf s/\msf t,\msf t^3)
+\mathcal{O}(\tau^{-\kappa})
\end{multline*}
holds as $\tau\to +\infty$, uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{theorem}
For the proof of these results, we compare $\bm\Phi_\tau$ with the solution ${\bm \Phi}_0$ of the id-PII RHP via the Deift-Zhou nonlinear steepest descent method. The required asymptotic analysis itself is carried out Section~\ref{sec:AsymAnalidRHP}, and the proofs of Theorems~\ref{thm:PhitauAsympt} and \ref{thm:integralPhitauPhicc} are completed in Section~\ref{sec:proofadmRHPtakeI}.
\subsection{Asymptotic analysis}\label{sec:AsymAnalidRHP}\hfill
For ${\bm \Phi}_0$ as introduced in \eqref{eq:PhiccPhikpz} and whose properties were discussed in Section~\ref{sec:KPZRHPKPZ}, we perform the transformation
\begin{equation}\label{eq:transfPhitauPsitau}
\bm\Psi_\tau(\zeta)=\bm\Phi_\tau(\zeta){\bm \Phi}_0(\zeta)^{-1},\quad \zeta\in\mathbb{C}\setminus\msf \Sigma.
\end{equation}
Then $\bm\Psi_\tau$ satisfies the following RHP.
\begin{enumerate}[\bf $\bm \Psi_\tau$-1.]
\item The matrix $\bm \Psi_\tau:\mathbb{C}\setminus \msf \Sigma\to \mathbb{C}^{2\times 2}$ is analytic.
\item Along the interior of the arcs of $\msf \Sigma$ the function $\bm \Psi_\tau$ admits continuous boundary values $\bm \Psi_{\tau,\pm}$
related by $\bm \Psi_{\tau,+}(\zeta)=\bm \Psi_{\tau,-}(\zeta)\bm J_{\bm \Psi_\tau}(\zeta)$, $\zeta\in \msf \Sigma$. With
%
$$
\lambda_0(\zeta)\deff \frac{1}{1+\ee^{-\msf h_0(\zeta)}}, \quad \lambda_\tau(\zeta)\deff \frac{1}{1+\ee^{-\mathsf h_\tau(\zeta)}},
$$
%
where $\msf h_0$ is as in \eqref{deff:h0}, the jump matrix $\bm J_{\bm \Psi_\tau}$ is
\begin{equation}\label{eq:jumpPsimodel}
\bm J_{\bm \Psi_\tau}(\zeta)\deff
\begin{dcases}
\bm I+\left(\lambda_\tau(\zeta)-\lambda_0(\zeta)\right) {\bm \Phi}_{0,+}(\zeta)\bm E_{12}{\bm \Phi}_{0,+}(\zeta)^{-1}, & \zeta \in \msf \Sigma_0 , \\
\bm I+\left(\frac{1}{\lambda_\tau(\zeta)}-\frac{1}{\lambda_0(\zeta)}\right){\bm \Phi}_{0,+}(\zeta)\bm E_{21}{\bm \Phi}_{0,+}(\zeta)^{-1}, & \zeta \in \msf \Sigma_1\cup\msf \Sigma_3, \\
\frac{\lambda_\tau(\zeta)}{\lambda_0(\zeta)}{\bm \Phi}_{0,+}(\zeta)\bm E_{11}{\bm \Phi}_{0,+}(\zeta)^{-1}+\frac{\lambda_0(\zeta)}{\lambda_\tau(\zeta)}{\bm \Phi}_{0,+}(\zeta)\bm E_{22}{\bm \Phi}_{0,+}(\zeta)^{-1}, & \zeta\in \msf \Sigma_2.
\end{dcases}
\end{equation}
\item For $\bm \Phi_\tau^{(1)}$ and ${\bm \Phi}_0^{(1)}$ the residues at $\infty$ of $\bm \Phi_\tau$ and ${\bm \Phi}_0$, respectively, the matrix $\bm\Psi_\tau$ has the asymptotic behavior
\begin{equation}\label{eq:Phitauad1Phicc1}
\bm \Psi_\tau(\zeta)=\bm I+\frac{1}{\zeta}(\bm \Phi_\tau^{(1)}-{\bm \Phi}_0^{(1)})+\mathcal{O}(1/\zeta^2)\qquad \text{as}\quad \zeta\to\infty.
\end{equation}
\item The matrix $\bm \Psi_\tau$ remains bounded as $\zeta \to 0$.
\end{enumerate}
The next step is to verify that the jump matrix decays to the identity in the appropriate norms. The terms in the jump that come from ${\bm \Phi}_0$ are precisely the ones we already estimated in Sections~\ref{sec:Phiccsingregime} and \ref{sec:Phiccnonasymreg}, so it remains to estimate the terms involving the $\lambda$-functions. The basic needed estimate is the following lemma.
\begin{lemma}\label{lem:convhtauhcc}
Fix $\nu\in (0,1/2)$ and $\msf t_0\in (0,1)$. The estimate
$$
\ee^{\msf h_0(\zeta)-\mathsf h_\tau(\zeta)}=1+\mathcal{O}(\zeta^2/\tau), \quad \tau\to\infty,
$$
holds true uniformly for $|\zeta|\leq \tau^\nu$ and uniformly for $\msf t_0\leq \msf t\leq 1/\msf t_0$, where the error term is independent of $\msf s\in \mathbb{R}$.
\end{lemma}
\begin{proof}
The Definition~\eqref{deff:admissibleh} of admissibility of $\mathsf h_\tau$ ensures that for $\tau$ sufficiently large, we can expand the term $\mathsf H(\zeta/\tau)$ in power series near the origin and obtain the expansion
$$
\mathsf h_\tau(\zeta)=\msf s -\msf t \zeta +\mathcal{O}(\zeta^2/\tau),\quad \tau\to +\infty,
$$
valid uniformly for $|\zeta|\leq \tau^\nu$, $\msf t_0\leq \msf t\leq 1/\msf t_0$, and with error independent of $\msf s\in \mathbb{R}$. Recalling that $\msf h_0(\zeta)=\msf s-\msf t\zeta$, the proof is complete.
\end{proof}
We are now able to prove the appropriate convergence of $\bm J_{\bm \Psi_\tau}$ to the identity matrix. We split the analysis into three lemmas, corresponding to different pieces of the contour $\msf \Sigma$. In the results that follow we use the matrix norm notations introduced in \eqref{deff:matrixnorm}--\eqref{deff:matrixLpqnorm}.
\begin{lemma}\label{lem:estJpsiadcSig0}
Fix $\msf t_0\in (0,1)$, $\msf s_0>0$ and $\nu\in (0,1/2)$. There exist $\tau_0=\tau_0(\msf t_0,\msf s_0,\nu)>0$, $M=M(\msf t_0,\msf s_0,\nu)>0$ and $\eta=\eta(\msf t_0,\msf s_0,\nu)>0$ for which the inequality
$$
\| \bm J_{\bm\Psi_\tau}-\bm I \|_{L^1\cap L^2\cap L^\infty(\msf \Sigma_0)}\leq M\ee^{-\msf s} \max \left\{ \tau^{-1+2\nu},\ee^{-\eta \tau^{3\nu/2}} \right\}
$$
holds true for any $\tau\geq \tau_0$, $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$.
\end{lemma}
\begin{proof}
Because both $\mathsf h_\tau$ and $\msf h_0$ are real-valued along the real line, the inequality
$$
\left|\lambda_\tau(\zeta)-\lambda_0(\zeta)\right|=\frac{|\ee^{-\mathsf h_\tau(\zeta)}-\ee^{-\msf h_0(\zeta)}|}{(1+\ee^{-\mathsf h_\tau(\zeta)})(1+\ee^{-\msf h_0(\zeta)})}\leq
|\ee^{-\mathsf h_\tau(\zeta)}-\ee^{-\msf h_0(\zeta)}|
$$
is immediate. For $0\leq \zeta\leq \tau^\nu$, we then use Lemma~\ref{lem:convhtauhcc} and the explicit expression for $\msf h_0$ in \eqref{deff:h0} and obtain
\begin{equation}\label{eq:estlambalcccSig01}
\left|\lambda_\tau(\zeta)-\lambda_0(\zeta)\right|=\mathcal{O}\left(\frac{\ee^{-\msf s+\msf t \zeta}}{\tau^{1-2\nu}} \right).
\end{equation}
For $\zeta\geq \tau^\nu$, we instead use that both $\mathsf h_\tau$ and $\msf h_0$ are real-valued along the positive axis and write
$$
|\lambda_\tau(\zeta)-\lambda_0(\zeta)| \leq \left|\lambda_\tau(\zeta)-1 \right| +\left|\lambda_\tau(\zeta)-1 \right| =\frac{\ee^{-\msf s+\msf t \zeta}}{1+\ee^{-\msf s+\msf t\zeta}}+\frac{\ee^{-\msf s-\tau\mathsf H(\zeta/\tau)}}{1+\ee^{-\msf s-\tau\mathsf H(\zeta/\tau)}}\leq \ee^{-\msf s}\left(\ee^{\msf t \zeta}+\ee^{-\tau\mathsf H(\zeta/\tau)}\right).
$$
From Definition~\ref{deff:admissibleh}--(iii) we bound $\ee^{-\tau\mathsf H(\zeta/\tau)}\leq \ee^{\widehat \eta\zeta^{3/2-\epsilon}}$ and simplify the last inequality to
\begin{equation}\label{eq:estlambalcccSig02}
|\lambda_\tau(\zeta)-\lambda_0(\zeta)| \leq \ee^{-\msf s}\ee^{\tilde \eta \zeta^{\alpha}},\quad \zeta\geq \tau^\nu, \; \alpha \deff \max \{1,3/2-\epsilon\}<\frac{3}{2} ,
\end{equation}
for a new value $\tilde \eta>0$.
Recall that $\bm J_{\bm\Psi_\tau}$ was given in \eqref{eq:jumpPsimodel}. We use \eqref{eq:estlambalcccSig01} and \eqref{eq:estlambalcccSig02} in combination with Propositions~\ref{prop:EstPhiccSing} and \ref{prop:EstPhiccNonasymp} to get the existence of a value $\tau_0>0$ for which
\begin{equation}\label{eq:PtwEstPhitauadcSig0}
|\bm J_{{\bm \Psi}_\tau}(\zeta)-\bm I| \leq M \ee^{-\msf s}
\ee^{-\eta \zeta^{3/2}} \left( \chi_{(0,\tau^{\nu})}(\zeta) \frac{\ee^{\msf t \zeta}}{\tau^{1-2\nu}}+\ee^{\tilde \eta \zeta^\alpha}\chi_{(\tau^{\nu},+\infty)}(\zeta)\right),\quad \tau\geq \tau_0,
\end{equation}
where $\eta>0,M>0$ may depend on $\msf s_0,\msf t_0$ and $\nu\in [0,1/2)$, but are independent of $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$. After appropriately changing the values of $\tilde\eta,\eta,M$, and having in mind that $\alpha<3/2$, the result follows from this inequality.
\end{proof}
Next, we prove the equivalent result along the pieces of $\msf \Sigma$ which are not on the real line.
\begin{lemma}\label{lem:estJpsiadcSig13}
Fix $\msf t_0\in (0,1)$, $\msf s_0>0$ and $\nu\in (0,1/2)$. There exist $\tau_0=\tau_0(\msf t_0,\msf s_0,\nu)>0$, $M=M(\msf t_0,\msf s_0,\nu)>0$ and $\eta=\eta(\msf t_0,\msf s_0,\nu)>0$, for which the inequality
$$
\| \bm J_{\bm\Psi_\tau}-\bm I \|_{L^1\cap L^2\cap L^\infty(\msf \Sigma_1\cup \msf \Sigma_3)}\leq M\ee^{-\msf s} \max \left\{ \tau^{-1+2\nu},\ee^{-\eta \tau^{3\nu/2}} \right\}
$$
holds true for any $\tau\geq \tau_0$, $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$.
\end{lemma}
\begin{proof}
Write
$$
\frac{1}{\lambda_\tau(\zeta)}-\frac{1}{\lambda_0(\zeta)}=\ee^{-\mathsf h_\tau(\zeta)}-\ee^{-\msf h_0(\zeta)}=-\ee^{-\msf s +\msf t \zeta}(1-\ee^{\msf h_0(\zeta)-\mathsf h_\tau(\zeta)}).
$$
From Lemma~\ref{lem:convhtauhcc}, we estimate for $0\leq |\zeta|\leq \tau^\nu$,
$$
\frac{1}{\lambda_\tau(\zeta)}-\frac{1}{\lambda_0(\zeta)}=\mathcal{O}\left(\frac{\ee^{-\msf s +\msf t \re \zeta}}{\tau^{1-2\nu}}\right),\quad \tau\to \infty,
$$
where the implicit error term is independent of $\msf s$ and uniform for $\msf t\in [\msf t_0,1/\msf t_0]$. On the other hand, from the explicit form of $\msf h_0$ and Definition~\ref{deff:admissibleh}--(iii),
$$
\left|\frac{1}{\lambda_\tau(\zeta)}-\frac{1}{\lambda_0(\zeta)}\right|
\leq
\ee^{-\msf s}\left(\ee^{\msf t\re\zeta}+\ee^{-\eta|\zeta|}\right).
$$
We combine this inequality with Propositions~\ref{prop:EstPhiccSing} and \ref{prop:EstPhiccNonasymp} and use them on \eqref{eq:jumpPsimodel}. The conclusion is that there exist $M>0$, $\eta_1,\eta_2>0$ and $\tau_0>0$, depending on $\nu,\msf t_0,\msf s_0$, for which the inequality
\begin{equation}\label{eq:PtwEstPhitauadcSig13}
\left| \bm J_{\bm \Psi_\tau}(\zeta)-\bm I\right|\leq M \ee^{-\msf s} \ee^{-\eta_1\re \zeta^{3/2}+\eta_2\re \zeta}
\left( \frac{1}{\tau^{1-2\nu}}\chi_{\{|\zeta|\leq \tau^\nu\}}(\zeta)+\chi_{\{|\zeta|\geq \tau^\nu\}}(\zeta)\right),\quad \zeta\in \msf \Sigma_1\cup \msf \Sigma_3,
\end{equation}
is valid for every $\tau\geq \tau_0,$ $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$. The definition \eqref{def:contour_sigma} of the contours $\msf \Sigma_1$ and $\msf \Sigma_3$ assure us that $\re \zeta^{3/2}>0$ and $\re \zeta<0$ on these contours. After possibly changing the values of the constants $\eta_1,\eta_2$ and $M$, the result follows.
\end{proof}
Finally, we now handle the jump on the negative axis.
\begin{lemma}\label{lem:estJpsiadcSig2}
Fix $\msf t_0\in (0,1)$, $\msf s_0>0$ and $\nu\in (0,1/2)$. There exist $\tau_0=\tau_0(\msf t_0,\msf s_0,\nu)>0, M=M(\msf t_0,\msf s_0,\nu)>0$ and $\eta=\eta(\msf t_0,\msf s_0,\nu)>0$ for which the inequality
$$
\| \bm J_{\bm\Psi_\tau}-\bm I \|_{L^1\cap L^2\cap L^\infty(\msf \Sigma_2)}\leq M \ee^{-\msf s} \max\left\{ \tau^{-1+2\nu},\ee^{-\eta\tau^{\nu}} \right\}
$$
holds true for any $\tau\geq \tau_0$, $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$.
\end{lemma}
\begin{proof}
The initial step is to rewrite the last line of \eqref{eq:jumpPsimodel} as
\begin{equation}\label{eq:estJpsiadcSig1}
\bm J_{\bm\Psi_{\tau}}(\zeta)-\bm I=\left(\frac{\lambda_\tau(\zeta)}{\lambda_0(\zeta)}-1\right)\bm I+\left(\frac{\lambda_0(\zeta)}{\lambda_\tau(\zeta)}-\frac{\lambda_\tau(\zeta)}{\lambda_0(\zeta)}\right){\bm \Phi}_{0,+}(\zeta)\bm E_{22}{\bm \Phi}_{0,+}(\zeta)^{-1},\quad \zeta\in \msf \Sigma_2.
\end{equation}
The identities
$$
\frac{\lambda_\tau(\zeta)}{\lambda_0(\zeta)}-1=\frac{\ee^{-\msf h_0(\zeta)}-\ee^{-\mathsf h_\tau(\zeta)}}{1+\ee^{-\mathsf h_\tau(\zeta)}},\qquad
\frac{\lambda_0(\zeta)}{\lambda_\tau(\zeta)}-1=-\frac{\ee^{-\msf h_0(\zeta)}-\ee^{-\mathsf h_\tau(\zeta)}}{1+\ee^{-\msf h_0(\zeta)}},
$$
are trivial, and because $\msf h_0$ and $\mathsf h_\tau$ are real-valued along $\msf \Sigma_2=(-\infty,0)$, these equalities give
$$
\left|\frac{\lambda_\tau(\zeta)}{\lambda_0(\zeta)}-1\right|+\left| \frac{\lambda_0(\zeta)}{\lambda_\tau(\zeta)}-1\right|
\leq
2\left|\ee^{-\msf h_0(\zeta)}-\ee^{-\mathsf h_\tau(\zeta)}\right|,\quad \zeta<0.
$$
For $|\zeta|\leq \tau^\nu$ we use Lemma~\ref{lem:convhtauhcc} and estimate
$$
\left|\frac{\lambda_\tau(\zeta)}{\lambda_0(\zeta)}-1\right|+\left| \frac{\lambda_0(\zeta)}{\lambda_\tau(\zeta)}-1\right|=\mathcal{O}\left(\frac{\ee^{-\msf s -\msf t |\zeta|}}{\tau^{1-2\nu}}\right),\quad \tau\to +\infty,\quad -\tau^{\nu}\leq \zeta\leq 0,
$$
whereas for $\zeta \leq -\tau^\nu$ we use instead the definition of $\msf h_0$ in \eqref{deff:h0} and Definition~\ref{deff:admissibleh}--(iii) and write
$$
\left|\frac{\lambda_\tau(\zeta)}{\lambda_0(\zeta)}-1\right|+\left| \frac{\lambda_0(\zeta)}{\lambda_\tau(\zeta)}-1\right|
\leq 2\ee^{-\msf s -(\msf t+\eta)|\zeta|}.
$$
We combine these two inequalities with Propositions~\ref{prop:EstPhiccSing} and \ref{prop:EstPhiccNonasymp}, and apply them to \eqref{eq:estJpsiadcSig1}. As a result, we learn that there exist $M>0,\eta>0,\tau_0>0$ for which the estimate
\begin{equation}\label{eq:EstJPsitauadcSig2}
\left| \bm J_{\bm \Psi_\tau}(\zeta)-\bm I \right|\leq M|\zeta|^{1/2}\ee^{-\msf s-\eta |\zeta|} \left(\frac{1}{\tau^{1-2\nu}}\chi_{(-\tau^\nu,0]}(\zeta)+\chi_{(-\infty,-\tau^\nu)}(\zeta)\right),\quad \zeta\leq 0,
\end{equation}
is valid for any $\msf s\geq -\msf s_0,$ $\msf t\in [\msf t_0,1/\msf t_0]$, $\tau\geq \tau_0$. After possibly changing the values of $\eta,M$, the result follows from standard arguments.
\end{proof}
Now that we controlled the asymptotic behavior for the jump matrix $\bm J_{\bm\Psi_\tau}$, we are ready to obtain small norm estimates for $\bm \Psi_{\tau}$ itself. We summarize these estimates in the next result. For that, we recall the matrix norm notations introduced in \eqref{deff:matrixnorm},\eqref{deff:matrixLpnorm},\eqref{deff:matrixLpqnorm}.
\begin{theorem}\label{thm:PsitauSmallnorm}
Fix $\msf t_0\in (0,1)$ and $\msf s_0>0$. There exists $\tau_0=\tau_0(\msf t_0,\msf s_0)>0$ for which the solution $\bm \Psi_\tau$ uniquely exists for any $\tau\geq \tau_0$ and any $\msf s \geq -\msf s_0, \msf t\in [\msf t_0,1/\msf t_0]$. Furthermore, it satisfies the following asymptotic properties.
Its boundary value $\bm \Psi_{\tau,-}$ exists along $\msf \Sigma$, and satisfies the estimate
$$
\|\bm \Psi_{\tau,-}-\bm I \|_{L^2(\msf \Sigma)}=\mathcal{O}\left(\tau^{-\kappa}\right),\quad \tau \to +\infty,
$$
for any $\kappa\in (0,1)$, where the error term, for a given $\kappa$, is uniform for $\msf s \geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$.
For $\tau$ sufficiently large, the solution $\bm \Psi_\tau$ admits the representation
\begin{equation}\label{eq:IntReprPsitau}
\bm \Psi_\tau(\zeta)=\bm I+\frac{1}{2\pi \mathrm{i}}\int_{\msf \Sigma}\frac{\bm \Psi_{\tau,-}(w)(\bm J_{\bm \Psi_\tau}(w)-\bm I)}{w-\zeta}\mathrm{d} w,\quad \zeta\in \mathbb{C}\setminus \msf \Sigma.
\end{equation}
Still for $\tau$ sufficiently large, $\bm \Psi_\tau$ satisfies
\begin{equation}\label{eq:Psitauad1}
\bm \Psi_\tau(\zeta)=\bm I+\bm\Psi_{\tau}^{(1)}\frac{1}{\zeta}+\mathcal{O}(\zeta^{-2}),\quad \zeta\to \infty,\quad \text{with} \quad
\bm\Psi_{\tau}^{(1)}\deff -\frac{1}{2\pi \mathrm{i}}\int_{\msf \Sigma} \bm\Psi_{\tau,-}(\xi)(\bm J_{\bm \Psi_{\tau}}(\xi)-\bm I)\mathrm{d} \xi.
\end{equation}
\end{theorem}
\begin{proof}
The small norm estimates provided by Lemmas~\ref{lem:estJpsiadcSig0}, \ref{lem:estJpsiadcSig13} and \ref{lem:estJpsiadcSig2} allow us to apply the small norm theory for Riemann-Hilbert problems (see for instance \cite{DeiftParkCity, DeiftZhou93mKdV}), and the claims follow with standard methods. We stress that for this statement we only need the $L^2$ and $L^\infty$ estimates from the aforementioned lemmas, but the $L^1$ estimates provided by them will be useful later.
\end{proof}
\subsection{Proof of main results of the section}\label{sec:proofadmRHPtakeI}\hfill
We are ready to prove the main results of this section.
\begin{proof}[Proof of Theorem~\ref{thm:PhitauAsympt}]
During the whole proof we identify $\nu=(1-\kappa)/2$.
The matrix ${\bm \Phi}_0$ always exists, whereas Theorem~\ref{thm:PsitauSmallnorm} provides the existence of $\bm\Psi_{\tau}$ for $\tau$ sufficiently large. From the relation \eqref{eq:transfPhitauPsitau} we obtain the claimed existence of $\bm\Phi_\tau$.
Comparing \eqref{eq:Phitauad1Phicc1} with \eqref{eq:Psitauad1} we obtain the identity
$$
\bm\Phi_\tau^{(1)}={\bm \Phi}_0^{(1)}+\bm\Psi_{\tau}^{(1)}.
$$%
Writing
\begin{equation}\label{eq:EstPsi1tauad}
\bm\Psi_{\tau}^{(1)}=-\frac{1}{2\pi \mathrm{i}}\int_{\msf \Sigma} (\bm\Psi_{\tau,-}(\xi)-\bm I)(\bm J_{\bm \Psi_{\tau}}(\xi)-\bm I)\mathrm{d} \xi -\frac{1}{2\pi \mathrm{i}}\int_{\msf \Sigma} \bm(\bm J_{\bm \Psi_{\tau}}(\xi)-\bm I)\mathrm{d} \xi,
\end{equation}
and using Cauchy-Schwartz,
$$
|\bm\Psi_{\tau}^{(1)}|\leq \|\bm\Psi_{\tau,-}-\bm I\|_{L^2(\msf \Sigma)} \|\bm J_{\bm \Psi_{\tau}}-\bm I\|_{L^2(\msf \Sigma)}+\|\bm J_{\bm \Psi_{\tau}}-\bm I\|_{L^1(\msf \Sigma)},
$$
and from Lemmas~\ref{lem:estJpsiadcSig0}, \ref{lem:estJpsiadcSig13}, \ref{lem:estJpsiadcSig2} and Theorem~\ref{thm:PsitauSmallnorm}, the right-hand side above is $\mathcal{O}(\tau^{-\kappa})$, for any $\kappa\in (0,1)$ and uniformly for $\msf s,\msf t$ as in \eqref{scaling:UTregime}, proving \eqref{eq:asympPhin1Phi01}.
To prove the asymptotic formula \eqref{eq:asympPhitauPhi0} we follow arguments presented in \cite[Theorem~3.1]{Kuijlaars2003lecnotes} and \cite[Lemma~2]{Aptekarev2002sharp}, with minor modifications to handle the uniformity on the unbounded set $|x|\leq \tau^{-\nu}$ as claimed.
First off, the jump matrix $\bm J_{\bm \Psi_\tau}$ is $C^\infty$ on $\msf \Sigma$, in particular H\"older continuous, implying that $\bm \Psi_\tau$ extends continuously to its boundary values $\bm\Psi_{\tau,\pm}$. Accounting also for the behavior of $\bm\Psi_{\tau}$ at $\infty$ and combining with the maximum principle,
\begin{equation}\label{eq:maxprincipleconseqPsitauad}
\|\bm \Psi_{\tau}\|_{L^\infty(\mathbb{C}\setminus \msf \Sigma)}\leq M_\tau\deff \max\left\{ \|\bm\Psi_{\tau,+}\|_{L^\infty(\msf \Sigma)},\|\bm\Psi_{\tau,-}\|_{L^\infty(\msf \Sigma)} \right\},
\end{equation}
where the constant $M_\tau$ is finite.
For a point $x\in \msf \Sigma\setminus \{0\}$ and $\varepsilon>0$, we consider the arcs $C^\pm_\varepsilon(x)$ of the disk centered at $x$ and radius $\varepsilon$ which are on the $\pm$ side of $\msf \Sigma$. We then set
$$
\msf \Sigma^\pm\deff \left(\msf \Sigma\setminus D_\varepsilon(x)\right)\cup C^\pm_\varepsilon(x),
$$
with the orientation induced from $\msf \Sigma$. We deform contour in the integral representation \eqref{eq:IntReprKernel} and then send $z\to x$, obtaining that
\begin{equation}\label{eq:intreprPsitauadpm}
\bm\Psi_{\tau,\pm}(x)=\bm I+\frac{1}{2\pi\mathrm{i}}\int_{\msf \Sigma^\mp}\frac{\bm\Psi_{\tau,-}(s)(\bm J_{\bm \Psi_{\tau}}(s)-\bm I)}{s-x}\mathrm{d} s.
\end{equation}
From standard estimates and using \eqref{eq:maxprincipleconseqPsitauad}, the just written equation yields
$$
|\bm \Psi_{\tau,\pm}(x)|\leq 1+\frac{1}{\pi \varepsilon}\|\bm \Psi_{\tau,-}\|_{L^\infty(\msf \Sigma^\mp)} \|\bm J_{\bm \Psi_\tau}-\bm I\|_{L^1(\msf \Sigma^\mp)}\leq 1+\frac{1}{\pi \varepsilon}M_\tau \|\bm J_{\bm \Psi_\tau}-\bm I\|_{L^1(\msf \Sigma^\mp)},
$$
and therefore
\begin{equation}\label{eq:estimateMtauad}
M_\tau\leq \left(1-\frac{1}{\pi\varepsilon}\|\bm J_{\bm \Psi_\tau}-\bm I\|_{L^1(\msf \Sigma^\mp)}\right)^{-1}.
\end{equation}
Lemmas~\ref{lem:estJpsiadcSig0}, \ref{lem:estJpsiadcSig13} and \ref{lem:estJpsiadcSig2} provide $L^p$ estimates for $\bm J_{\bm \Psi_\tau}-\bm I$ along $\msf \Sigma$. Exploring that $\msf \Sigma^\pm$ is obtained from $\msf \Sigma$ after a small deformation around the point $x=\mathcal{O}(\tau^{-\nu})$, it is straightforward to see that the same estimates hold in $\msf \Sigma^\pm$, which can be summarized as
\begin{equation}\label{eq:estJPsicSigpm}
|\bm J_{\bm \Psi_\tau}(\zeta)-\bm I| \leq M\ee^{-\eta \min\{|\zeta|,|\zeta|^{3/2}\}}\left(\frac{1}{\tau^{\kappa}}\chi_{\{|\zeta|\leq \tau^\nu\}}(\zeta)+\chi_{\{|\zeta|> \tau^\nu\}}(\zeta)\right),\quad \zeta\in \msf \Sigma^\pm,
\end{equation}
for some constants $\eta,M>0$ which may depend on $\msf s_0,\msf t_0,\tau_0$ but are independent of $\msf s\geq \msf s_0,\msf t\in [\msf t_0,1/\msf t_0],\tau\geq \tau_0$, see \eqref{eq:PtwEstPhitauadcSig0}, \eqref{eq:PtwEstPhitauadcSig13} and \eqref{eq:EstJPsitauadcSig2}. Combining with \eqref{eq:estimateMtauad}, we conclude in particular that $M_\tau\leq 2$, for $\msf s,\msf t,\tau$ in the same range of values. Having in mind \eqref{eq:transfPhitauPsitau}, this bound on $M_\tau$ applied to \eqref{eq:intreprPsitauadpm} is enough to ensure that $\bm \Psi_{\tau,\pm}-\bm I=\mathcal{O}(\tau^{-\nu})$, but to obtain the decay in $x$ for the error claimed in \eqref{eq:asympPhitauPhi0} a little more care is needed as follows.
First off, we split the integral in \eqref{eq:intreprPsitauadpm} into two, namely along
$$
J_x\deff \{\zeta\in \msf \Sigma^-\mid |\zeta-x|\geq |x|/2\}\quad \text{and}\quad \msf \Sigma^-\setminus J_x.
$$
For the integral over $J_x$, we estimate as
$$
\left| \int_{J_x}\frac{\bm\Psi_{\tau,-}(s)(\bm J_{\bm \Psi_{\tau}}(s)-\bm I)}{s-x}\mathrm{d} s \right|\leq 2M_\tau \sup_{\zeta\in J_x}\frac{1}{|\zeta-x|} \|\bm J_{\bm \Psi_{\tau}}-\bm I \|_{L^1(J_x)}=\mathcal{O}(\tau^{-\kappa}|x|^{-1}),
$$
where we used \eqref{eq:estJPsicSigpm}, the fact that $x=\mathcal{O}(\tau^{-\kappa})$ and again the bound $M_\tau\leq 2$. Observing that $|\zeta-x|\geq \varepsilon$ along the remaining piece, a similar argument yields
$$
\left| \frac{\bm\Psi_{\tau,-}(s)(\bm J_{\bm \Psi_{\tau}}(s)-\bm I)}{s-x} \right|\leq \frac{4}{\varepsilon} \left|\bm J_{\bm \Psi_{\tau}}(s)-\bm I \right|,\quad s\in \msf \Sigma^-\setminus J_x,
$$
and again from \eqref{eq:estJPsicSigpm} we see that the right-hand side above decays exponentially in $x$ when $x\to \infty$ and is $\mathcal{O}(\tau^{-\kappa})$ when $\tau\to \infty$. From \eqref{eq:intreprPsitauadpm} we thus obtain
$$
\left|\bm\Psi_{\tau,+}(x)-\bm I\right|=\mathcal{O}\left(\frac{1}{(1+|x|)\tau^{\kappa}}\right),\quad \tau\to \infty,
$$
uniformly for $x\in \msf \Sigma, |x|\leq \tau^{-\nu}$, and uniformly for $\msf s\geq -\msf s_0, \msf t_0\leq \msf t\leq 1/\msf t_0$. In virtue of \eqref{eq:transfPhitauPsitau}, this proves \eqref{eq:asympPhitauPhi0}.
\end{proof}
\begin{remark}\label{remark:UnifAsympPhitauad}
For admissible functions $\mathsf h_{\tau}$, the asymptotics \eqref{eq:ModelRHPAsymp} as $\zeta\to \infty$ of $\bm \Phi_\tau$ is valid uniformly in $\tau, \msf t$ and $\msf s$, in the sense that for any $\msf t_0\in (0,1)$ and any $\msf s_0>0$, there exist $K>0$ and $R>0$ such that
$$
\left|
\bm \Phi_\tau(\zeta)\ee^{\frac{2}{3}\zeta^{2/3}\sp_3}\bm U_0\zeta^{\sp_3/4}-\bm I\right|\leq \frac{K}{|\zeta|},\quad \text{whenever } |\zeta|\geq R, \; \msf s\geq -\msf s_0, \; \msf t_0\leq \msf t\leq 1/\msf t_0,
$$
and we emphasize that $K$ and $R$ are independent of $\msf s,\msf t$. To see that this is true, in virtue of \eqref{eq:transfPhitauPsitau} it is enough to show that the asymptotics \eqref{eq:AsymptPhicc} and \eqref{eq:Phitauad1Phicc1} are uniform in the same sense, we indicate the proof for the latter and the former is analogous.
Using the trivial identity $1/(w-\zeta)=s/(s(s-\zeta))-1/\zeta$, we express \eqref{eq:IntReprPsitau} as
$$
\bm\Psi_\tau(\zeta)=\bm I-\frac{1}{2\pi \mathrm{i} \zeta}\int_{\msf \Sigma}\bm \Psi_{\tau,-}(w)(\bm J_{\bm \Psi_\tau}(w)-\bm I)\mathrm{d} w+\frac{1}{2\pi \mathrm{i}\zeta}\int_{\msf \Sigma}\frac{ w\bm \Psi_{\tau,-}(w)(\bm J_{\bm \Psi_\tau}(w)-\bm I)}{w-\zeta}\mathrm{d} w.
$$
Because $\bm \Psi_{\tau,-}\in\bm I+ L^1(\msf \Sigma)$ and $\bm J_{\bm \Psi_{\tau}}-\bm I$ decays pointwise exponentially fast (and uniformly in $\msf s,\msf t$ as claimed), the two integrals can be bounded uniformly in $\zeta,\tau,\msf s$ as claimed, and the uniform decay for $\bm \Psi_{\tau}$ as claimed follows.
\end{remark}
To finish this section, it remains to prove Theorem~\ref{thm:integralPhitauPhicc}, and for that end we first establish a lemma.
\begin{lemma}\label{eq:lemtailsintPhis}
Fix $\msf s\in \mathbb{R}$ and $\msf t_0\in (0,1)$. For any $\nu\in (0,1/2)$, there exists $\eta=\eta(\nu)>0$ independent of $\msf s,\msf t,\tau$ for which the estimates
\begin{equation}\label{eq:EstPhiccinttailest}
\int_{|x|\geq \tau^{\nu}} \frac{\ee^{\msf t x-u}}{(1+\ee^{\msf t x-u})^2}\left|\left[({\bm\Delta}_0(x\mid u,\msf t)^{-1}{\bm \Phi}_{0,+}(x\mid u,\msf t)^{-1}({\bm \Phi}_{0,+} {\bm\Delta}_0 )'(x\mid u,\msf t)\right]_{21} \right|\mathrm{d} x=\mathcal{O}(\ee^{-u-\eta \tau^{\nu}})
\end{equation}
and
\begin{equation}\label{eq:EstPhitauadinttailest}
\int_{|x|\geq \tau^{\nu}} \frac{\ee^{\mathsf h_\tau(x\mid u)}}{\left(1+\ee^{\mathsf h_\tau(x\mid u)}\right)^2}\left|\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}\right|\mathrm{d} x =\mathcal{O}(\ee^{-u-\eta \tau^{\nu}})
\end{equation}
are valid as $\tau\to \infty$, uniformly for $u\geq \msf s$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$. In particular, both integrands are integrable over $\mathbb{R}$.
\end{lemma}
\begin{proof}
We prove \eqref{eq:EstPhitauadinttailest} which is slightly more technical because the integrand depends on $\tau$, the estimate \eqref{eq:EstPhiccinttailest} follows in a similar manner.
Having in mind Remark~\ref{remark:UnifAsympPhitauad}, we use the expansion \eqref{eq:ModelRHPAsymp} to estimate
$$
\bm\Phi_{\tau,+}(x)\bm\Delta_{\tau}(x)=\left(1+\mathcal{O}(x^{-1})\right)x^{\sp_3/4}\bm U_0^{-1}\left(\bm I+\chi_{(-\infty,0)}(x)\ee^{\frac{4}{3}x^{3/2}_+}(1-\ee^{-\mathsf h_{\tau}(x)})\bm E_{21}\right)\ee^{-\frac{2}{3}x_+^{3/2}\sp_3},\quad x\in E_\tau.
$$
Observe that the factor $\chi_{(-\infty,0)}(x)\ee^{\frac{4}{3}x^{3/2}_+}(1-\ee^{-\mathsf h_{\tau}(x)})$ is bounded, thanks to the facts that $x_+^{3/2}\in \mathrm{i}\mathbb{R}$ and $\re H>0$ for $x<0$ (see Definition~\ref{deff:admissibleh}--(iii)). The identity above can be differentiated, and using it we obtain the crude bound
$$
\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}=\ee^{-\frac{4}{3}x_+^{3/2}}\mathcal{O}(|x|^{3}),\quad x\in E_\tau,
$$
which is non-optimal in $x$, but will be enough for the coming estimates, and which is valid uniformly for $u\geq -\msf s_0,\msf t\in [\msf t_0,1/\msf t_0]$ as $\tau\to\infty$. Thus, to conclude the result it is enough to estimate each of the integrals
$$
I_-\deff\int_{-\infty}^{-\tau^\nu}|x|^3\frac{\ee^{\mathsf h_\tau(x\mid u)}}{(1+\ee^{\mathsf h_\tau(x\mid u)})^2}\mathrm{d} x\quad \text{and}\quad
I_+\deff\int_{\tau^{\nu}}^{+\infty}|x|^3\frac{\ee^{-\frac{4}{3}x^{3/2}}\ee^{\mathsf h_\tau(x\mid u)}}{(1+\ee^{\mathsf h_\tau(x\mid u)})^2}\mathrm{d} x.
$$
To estimate $I_-$, we use the inequalities $v/(1+v)\leq 1$ and $1/(1+v)\leq 1/v$, valid for $v>0$, to estimate
$$
\frac{\ee^{\mathsf h_\tau(x\mid u)}}{(1+\ee^{\mathsf h_\tau(x\mid u)})^2}\leq \ee^{-\mathsf h_{\tau}(x\mid u)},
$$
and using now the inequality from Definition~\ref{deff:admissibleh}--(iii) along $\msf \Sigma_2=(-\infty,0)$, we obtain
$$
I_-\leq \ee^{-u}\int_{-\infty}^{-\tau^\nu}|x|^3\ee^{-\eta |x|}\mathrm{d} x=\mathcal{O}(\ee^{-u-\widetilde\eta\tau^{\nu}/2})
$$
for some $\widetilde\eta>0$. In a similar manner, we also obtain that
$$
I_+\leq \ee^{-u} \int_{\tau^\nu}^{+\infty} |x|^3 \ee^{-\frac{4}{3}x^{3/2}+\widehat\eta x^{2/3-\varepsilon} }\mathrm{d} x=\mathcal{O}(\ee^{-u-\tau^{3\nu/2}}),
$$
where now for the last equality we used Definition~\ref{deff:admissibleh}--(iii) along $\msf \Sigma_0=(0,\infty)$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:integralPhitauPhicc}]
As in the previous proof, we identify $\kappa\in (0,1)$ from Theorem~\ref{thm:PhitauAsympt} with $\nu=(1-\kappa)/2\in (0,1/2)$. Thanks to \eqref{eq:EstPhitauadinttailest},
\begin{multline*}
\int_{-a\tau}^{b\tau} \frac{\ee^{\mathsf h_\tau(x\mid u)}}{\left(1+\ee^{\mathsf h_\tau(x\mid u)}\right)^2}\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}\mathrm{d} x\\
=
\int_{-\tau^\nu}^{\tau^\nu} \frac{\ee^{\mathsf h_\tau(x\mid u)}}{\left(1+\ee^{\mathsf h_\tau(x\mid u)}\right)^2}\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}\mathrm{d} x +
\mathcal{O}(\ee^{-u-\eta \tau^{\nu}}),
\end{multline*}
valid as $\tau\to \infty$ and uniformly for $u\geq \msf s$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$. Next, we use Lemma~\ref{lem:convhtauhcc} and \eqref{eq:asympPhitauPhi0} to ensure that
\begin{multline*}
\int_{-\tau^\nu}^{\tau^\nu} \frac{\ee^{\mathsf h_\tau(x\mid u)}}{\left(1+\ee^{\mathsf h_\tau(x\mid u)}\right)^2}\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}\mathrm{d} x=\\
(1+\mathcal{O}(\tau^{-\kappa})) \int_{-\tau^\nu}^{\tau^\nu} \frac{\ee^{\msf h_0(x)}}{(1+\ee^{\msf h_0(x)})^2}\left[({\bm\Delta}_0(x\mid u,\msf t)^{-1}{\bm \Phi}_{0,+}(x\mid u,\msf t)^{-1}({\bm \Phi}_{0,+} {\bm\Delta}_0 )'(x\mid u,\msf t)\right]_{21} \mathrm{d} x.
\end{multline*}
With the help of the calculation
$$
\frac{\ee^{\msf h_0(x)}}{(1+\ee^{\msf h_0(x)})^2}=\frac{1}{(1+\ee^{\msf h_0(x)})(1+\ee^{-\msf h_0(x)})}=\frac{\ee^{-\msf h_0(x)}}{(1+\ee^{-\msf h_0(x)})^2}=\frac{\ee^{\msf t x-u}}{(1+\ee^{\msf t x-u})^2}
$$
we recognize the integrand from \eqref{eq:EstPhiccinttailest}, and then conclude that
\begin{multline*}
\int_{-a\tau}^{b\tau} \frac{\ee^{\mathsf h_\tau(x\mid u)}}{\left(1+\ee^{\mathsf h_\tau(x\mid u)}\right)^2}\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}\mathrm{d} x =\\
(1+\mathcal{O}(\tau^{-\kappa})) \int_{-\infty}^{\infty} \frac{\ee^{\msf t x-u}}{(1+\ee^{\msf t x-u})^2}\left[({\bm\Delta}_0(x\mid u,\msf t)^{-1}{\bm \Phi}_{0,+}(x\mid u,\msf t)^{-1}({\bm \Phi}_{0,+} {\bm\Delta}_0 )'(x\mid u,\msf t)\right]_{21} \mathrm{d} x + \mathcal{O}(\ee^{-u-\eta\tau^{\nu}}).
\end{multline*}
Finally, with arguments very similar to the ones used in the proof of Lemma~\ref{eq:lemtailsintPhis}, we see that this remaining integral on the right-hand side is $\mathcal{O}(\ee^{-u})$. Everything combined, we just proved that
\begin{multline*}
\int_{-\tau a}^{\tau b}\frac{\ee^{\mathsf h_\tau(x\mid u)}}{\left(1+\ee^{\mathsf h_\tau(x\mid u)}\right)^2}\left[ \bm \Delta_\tau(x\mid u)^{-1}\bm \Phi_{\tau,+}(x\mid u)^{-1}\left(\bm \Phi_{\tau,+} \bm\Delta_\tau\right)'(x\mid u)\right]_{21}\mathrm{d} x
=\\
\int_{-\infty}^\infty \frac{\ee^{\msf t x-u}}{(1+\ee^{\msf t x-u})^2}\left[({\bm \Phi}_{0,+}(x\mid u,\msf t){\bm\Delta}_0(x\mid u,\msf t))^{-1}({\bm \Phi}_{0,+} {\bm\Delta}_0 )'(x\mid u,\msf t)\right]_{21} \mathrm{d} x
+\mathcal{O}(\tau^{-\kappa}\ee^{-u}),\quad \tau\to +\infty.
\end{multline*}
We now integrate in $u\in [-\msf s,+\infty)$ and use Equation~\eqref{eq:QccIntRepPhicc} and the limit
\begin{equation*}
\lim_{s\to -\infty}\msf L^{\kern -0.15em \ai}(s,T)=1,
\end{equation*}
which is valid by dominated convergence since $|1+\ee^\alpha|^{-1}\leq 1$ for every $\alpha\in \mathbb{R}$, to conclude the proof.
\end{proof}
\section{The underlying equilibrium measure and related quantities}\label{sec:eqmeasuretau}
As we mentioned earlier, one of the main objects we will analyze is the RHP for the orthogonal polynomials associated to \eqref{deff:Ln}--\eqref{deff:DeformedPartFunction}. In this analysis, the model problem we discussed in the last few sections plays a central role, and another important quantity is the associated equilibrium measure and related objects that we discuss in this section.
\subsection{The equilibrium measure}\label{sec:equilibriummeasure}\hfill
A major role in our calculations is played by the equilibrium measure for the polynomial potential $V$, which is the unique probability measure $\mu_V$ on $\mathbb{R}$ for which the quantity
$$
\iint \log \frac{1}{|x-y|}\mathrm{d} \mu(x) \mathrm{d}\mu(y)+\int V(x)\mathrm{d} \mu(x)
$$
attains its minimum over all Borel probability measures $\mu$ supported on $\mathbb{R}$. Its existence and uniqueness is assured by standard results, see for instance \cite{Saff_book}, and its regularity that we now discuss is of particular relevance.
The measure $\mu_V$ is supported on a finite union of bounded intervals and is absolutely continuous with respect to the Lebesgue measure \cite{deift_kriecherbauer_mclaughlin}. Following Assumption~\ref{MainAssumption}--(i), we assume that $\mu_V$ is one-cut, that is, it is supported on a single interval that we take to be of the form
$$
\supp\mu_V=[-a,0],\quad a>0,
$$
and regular, meaning that the density of $\mu_V$ vanishes as a square-root at a neighborhood of the endpoints, does not vanish on $(-a,0)$, and the Euler-Lagrange equations are valid with strict inequality outside the support,
\begin{equation}\label{eq:eulerlagrange}
\int \log\frac{1}{|x-y|}\mathrm{d} \mu_V(y)+\frac{1}{2} V(x)+\ell_V \;
\begin{cases}
> 0, & x\in \mathbb{R}\setminus \supp\mu_V, \\
=0, & x\in \supp\mu_V,
\end{cases}
\end{equation}
for some constant $\ell_V\in \mathbb{R}$. The notions just introduced are consistent with the notions and notations that we already introduced and used in Section~\ref{sec:StatementResults}.
A transformation of the equilibrium measure of particular interest is its {\it Cauchy transform},
$$
C^{\mu_V}(z)\deff \int \frac{\mathrm{d} \mu_V(x)}{x-z},\quad z\in \mathbb{C}\setminus \supp\mu_V.
$$
Using the Euler-Lagrange identity, it can be shown that $C^{\mu_V}$ satisfies an algebraic equation of the form
\begin{equation}\label{eq:spectralcurvehV}
\left(C^{\mu_V}(z)+\frac{V'(z)}{2}\right)^2=\frac{1}{4}z(z+a)h_V(z)^2,\quad z\in \mathbb{C}\setminus \supp\mu_V,
\end{equation}
for some polynomial $h_V$ which does not vanish on $[-a,0]$ \cite{kuijlaars_silva_s_curves, MartinezOriveRakhmanov2015}.
We also associate to the equilibrium measure its $\phi$ function
\begin{equation}\label{eq:functionphitau}
\phi(z)\deff \int_{0}^z\left(C^{\mu_V}(s)+\frac{1}{2} V'(s)\right)\mathrm{d} s,\quad z\in \mathbb{C}\setminus (-\infty,0].
\end{equation}
The next result summarizes some properties of $\phi$ that will be needed later.
\begin{prop}\label{prop:functionphitau}
The function $\phi$ has the following properties.
\begin{enumerate}[(i)]
\item The function $\phi$ is analytic on $\mathbb{C}\setminus (-\infty,0]$.
\item For $x\in (-a,0)$,
$$
\phi_{+}(x)+\phi_{-}(x)=0, \quad \text{and}\quad \phi_{+}(x)-\phi_{-}(x)=-2\pi\mathrm{i} \mu_V((x,0)).
$$
\item For $x\in (-\infty,-a)$,
$$\phi_{+}(x)-\phi_{-}(x)=-2\pi \mathrm{i}.$$
\item For $x\in \mathbb{R}\setminus [-a,0]$,
$$\re \phi_{+}(x)=\re \phi_{-}(x) >0.$$
\item As $z\to \infty$ and some constant $\phi_\infty$,
$$
\phi(z)=\frac{V(z)}{2}+\ell_V-\log z+\frac{\phi_\infty}{z}+\mathcal{O}(z^{-1}),
$$
where $\ell_V$ is as in \eqref{eq:eulerlagrange}.
\item The function $\phi$ satisfies the estimate
\begin{equation}\label{eq:phi_local_beh}
\phi(z)= \frac{1}{3} h_V(0)a^{1/2} z^{3/2}(1+\mathcal{O}(z)), \quad z\to 0.
\end{equation}
\end{enumerate}
\end{prop}
\begin{proof}
The proof is standard using the properties of the equilibrium measure, see for instance \cite{deift_book}.
\end{proof}
\subsection{The conformal map}\hfill
Finally, using $\phi$ we construct a conformal map $\psi$, introduced formally with the next result.
\begin{prop}\label{prop:conformalmappsi}
The function
$$
\psi(z)\deff \left(\frac{3}{2}\phi(z)\right)^{2/3},
$$
is a conformal map from a neighborhood of the origin to a disk $D_{2\delta}(0)$, with $U_0\deff \psi^{-1}(D_\delta(0))$, and admits an expansion of the form
\begin{equation}\label{eq:derivativepsicpsi}
\psi(z)=\mathsf c_V z(1+\mathcal{O}(z)),\quad z\to 0,\qquad \mathsf c_V\deff 2^{-2/3}h_V(0)^{2/3}a^{1/3}>0.
\end{equation}
\end{prop}
\begin{proof}
The proof is also standard, and follows essentially from Proposition~\ref{prop:functionphitau}-(iv). We omit the details.
\end{proof}
In the previous proposition, the factor $2\delta$ instead of $\delta$ is chosen just for later convenience, for the statement of Proposition~\ref{prop:ConformExtH}. Later on, we use $\psi$ only over the smaller neighborhood $D_\delta(0)$.
As it is customary in RHP analysis, at a later stage we will need to glue the model problem as a local parametrix for the original RHP for orthogonal polynomials. This gluing procedure is done, in our case, using the conformal map $\psi$. In usual situations, the jump matrices of the model local problem are piecewise constant or yet homogeneous, and as such this procedure of using the conformal map does not significantly alter them. However, in our situation the jump involves the function $Q$, and consequently the jump will be altered by the conformal map in a nontrivial way.
With the next result we introduce the necessary quantities needed to keep track of this transformation. Recall the half rays $\msf \Sigma_j$, $j=0,1,2,3,$ which were introduced in \eqref{def:contour_sigma}. For the next statement, we talk about {\it neighborhoods} of $\msf \Sigma_j$, by which we mean open connected sets that contain $\msf \Sigma_j\setminus \{0\}$ in their interior.
\begin{prop}\label{prop:ConformExtH}
There exist neighborhoods $\mathcal S_j$ of $\msf \Sigma_j$ and a function
$$
\mathsf H_Q:\mathcal S\to \mathbb{C},\quad \mathcal S\deff \bigcup_{j=0}^3 \mathcal S_j
$$
with the following properties.
\begin{enumerate}[(i)]
\item For the value $\delta>0$ in Proposition~\ref{prop:conformalmappsi}, the inclusions
$$
D_\delta(0)\subset \mathcal S\quad \text{and}\quad \mathcal S_j\cap \mathcal S_k\subset D_\delta(0),
$$
hold true for any $j\neq k$.
\item The function $\mathsf H_Q$ is $C^\infty$ on $\mathcal S$, and is an extension of $Q\circ \psi^{-1}$ from $D_\delta(0)$, that is,
\begin{equation}\label{eq:RelHQQpsi}
\mathsf H_Q(w)=Q(\psi^{-1}(w)),\quad |w|<\delta.
\end{equation}
\item The function $\mathsf H_Q$ is analytic on $D_\delta(0)$, extends continuously up to the boundary of $D_\delta(0)$ and satisfies
\begin{equation}\label{eq:expansionHQ}
\mathsf H_Q(w)=-\mathsf c_{\mathsf H}w+\mathcal{O}(w),\quad \mathsf c_{\mathsf H}\deff \frac{\msf t}{\mathsf c_V},
\end{equation}
uniformly for $|w|\leq \delta$, where we recall that $\msf t$ and $\mathsf c_V$ are as in \eqref{deff:tQprime} and \eqref{eq:derivativepsicpsi}, respectively.
\item For some constants $\widehat\eta>\eta>0$, the function $\mathsf H_Q$ satisfies the estimates
$$
-\widehat\eta |w|\leq \re \mathsf H_Q(w)\leq -\eta |w| \qquad \text{for every }w\in \mathcal S_0\setminus D_\delta(0),
$$
and
$$
\re \mathsf H_Q(w)\geq \eta |w| \qquad \text{for every } w\in \left(\mathcal S_1\cup\mathcal S_2\cup \mathcal S_3\right)\setminus D_\delta(0).
$$
\end{enumerate}
\end{prop}
\begin{proof}
We construct the set $\mathcal S_j$ as tubular neighborhoods of $\msf \Sigma_j$ away from the origin, and as disks near the origin, namely
$$
\mathcal S_j=\mathcal S_j(\delta')=\left\{ w\in \mathbb{C} \mid \inf_{w'\in \msf \Sigma_j}|w-w'|< \delta' \right\}\cup D_\delta(0).
$$
By choosing $\delta'>0$ sufficiently small, in particular smaller than $\delta>0$, property (i) is immediate.
The function
$$
\mathsf H_Q(w)\deff Q(\psi^{-1}(w)),\quad |w|<\delta,
$$
is obviously analytic on $D_{\delta}(0)$, satisfies (iii) and admits an extension to the larger open set $D_{2\delta}(0)$. A standard argument using partitions of unity allows us to extend it to the sets $\mathcal S_j$'s as claimed by (ii), also making sure that it satisfies (iv).
\end{proof}
\section{Associated orthogonal polynomials}\label{sec:rhpapproachops}
The first and arguably major step towards understanding $\mathsf L^Q_n(\msf s)$ is to study several quantities related to the orthogonal polynomials for the varying weight \eqref{def:perturbedweight}, as we introduce next.
\subsection{Orthogonal polynomials and related quantities}\hfill
Denote by $\mathsf P_k=\mathsf P_k^{(n,\msf s)}$ the monic orthogonal polynomial of degree $k$ for the weight $\omega_n$ in \eqref{def:perturbedweight},
\begin{equation}\label{def:opsdeformed}
\mathsf P_k^{(n,\msf s)}(x)=x^k+\text{(lower degree terms)},\qquad \int_{\mathbb{R}}\mathsf P_k^{(n,\msf s)}(x)x^j\omega_n(x)\mathrm{d} x=0, \quad j=0,\hdots, k-1.
\end{equation}
These polynomials depend on $\omega_n$, so ultimately also on $Q$, but we refrain from stressing this dependence in the notation. We also denote by $\upgamma_k^{(n,Q)}=\upgamma_k^{(n,Q)}(\msf s)>0$ the corresponding norming constant, determined by
\begin{equation}\label{def:normingdeformed}
\frac{1}{\upgamma_k^{(n)}(\msf s)^2}=\int_\mathbb{R} \mathsf P_k^{(n,\msf s)}(x)^2\omega_n(x)\mathrm{d} x.
\end{equation}
We associate to the orthogonal polynomials their Christoffel-Darboux kernel,
\begin{equation}\label{deff:Kndefweight}
\mathsf K^Q_n(x,y)=\mathsf K^Q_n(x,y\mid \msf s)\deff \sum_{k=0}^{n-1} \upgamma_k^{(n)}(\msf s)^2\mathsf P_k^{(n,\msf s)}(x)\mathsf P_k^{(n,\msf s)}(y),
\end{equation}
stressing that we are not including the weight $\omega_n$ in this definition.
In particular, the identity
\begin{equation}\label{eq:IntReprKernel}
\int_{-\infty}^\infty \mathsf K_n(x,x\mid \msf s)\omega_n(x\mid \msf s)\mathrm{d} x=n
\end{equation}
holds true for any $\msf s \in \mathbb{R}$ and follows immediately from \eqref{def:normingdeformed}.
In a similar manner, we introduce the related quantities for the undeformed weight $\ee^{-nV}$. The partition function $Z_n$ already appeared in \eqref{deff:nondeformedpartfunction}, the orthogonal polynomials $P_k=P_k^{(n)}$ are determined by
\begin{equation*}
P_k^{(n)}(x)=x^k+\text{(lower degree terms)},\qquad \int_{\mathbb{R}}P_k^{(n)}(x)x^j\ee^{-nV(x)}\mathrm{d} x=0, \quad j=0,\hdots, k-1,
\end{equation*}
and the norming constants and Christoffel-Darboux kernel are determined from
$$
\frac{1}{(\gamma_k^{(n)})^2}=\int_\mathbb{R} P_k^{(n)}(x)^2\ee^{-nV(x)}\mathrm{d} x,\quad K_n(x,y)\deff \sum_{k=0}^{n-1}(\gamma_k^{(n)})^2P_k^{(n)}(x)P_k^{(n)}(y).
$$
The orthogonal polynomials $\mathsf P_k^{(n,\msf s)}$ vary continuously with $\msf s$, which is a consequence of Heine's formula \cite[Equation~(3.10)]{deift_book}. In particular, when taking the limit $\msf s\to+\infty$ we have that $x^k\omega_n(x)\to x^k\ee^{-nV}$ both uniformly in compacts and also in $L^1$, and $|x|^k\omega_n(x)\leq |x|^k\ee^{-nV(x)}$. Thus, dominated convergence then gives that all the just introduced undeformed quantities are recovered from their deformed versions in the limit $s\to+\infty$. This means, for instance, that the Christoffel-Darboux kernel $K_n$ and the partition function $Z_n$ are recovered via
\begin{equation}\label{eq:KernelContin}
K_n(x,y)=\mathsf K_n(x,y\mid \msf s=+\infty)\quad \text{and}\quad Z_n=\mathsf Z_n(\msf s=+\infty).
\end{equation}
The next result will be key into transforming asymptotics for the orthogonal polynomials to asymptotics for $\mathsf L^Q_n(s)$ itself.
\begin{prop}\label{prop:intreprLn}
The identity
\begin{equation}\label{eq:PartitionFctionNormingctt}
\log\mathsf L_n^Q(\msf s) = -\int_\msf s^{\infty}\int_{-\infty}^{\infty} \mathsf K^Q_n(x,x\mid u)\frac{\omega_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x\; \mathrm{d} u
\end{equation}
holds true for every $\msf s\in \mathbb{R}$.
\end{prop}
\begin{remark}
While we were finishing this manuscript, the work \cite{ClaeysGlesner2021} was posted to the ArXiv. Therein, they also derive the formula \eqref{eq:PartitionFctionNormingctt} in more general terms, using the underlying RHP for IIKS-type integrable operators, see the first displayed formula in page 28 therein. Similar formulas play a fundamental role in the recent works \cite{CafassoClaeys2021,CafassoClaeysRuzza2021}, see for instance \eqref{eq:QccIntRepPhikpz} above. Our proof of \eqref{eq:PartitionFctionNormingctt} relies solely on orthogonality properties, so we decided to present it nevertheless.
\end{remark}
\begin{proof}
The equality
$$
\mathsf Z^Q_n(\msf s)=n!\prod_{k=0}^{n-1}\upgamma_k^{(n)}(\msf s)^{-2}
$$
is standard in random matrix theory. From this identity, \eqref{def:normingdeformed} and the orthogonality relations we derive the deformation formula
\begin{equation*}
\partial_\msf s \log \mathsf Z^Q_n(\msf s) =-\int_\mathbb{R} \partial_\msf s\mathsf K^Q_n(x,x\mid \msf s)\omega_n(x\mid \msf s)\mathrm{d} x,
\end{equation*}
which in fact is valid for general weights depending on an additional parameter $\msf s$. To our knowledge, this last identity was first observed by Krasovsky \cite[Equation~(14)]{Krasovsky07}. We fix constants $L>\msf s>0$ and integrate the identity above,
\begin{equation}\label{eq:propLndef1}
\log\mathsf Z^Q_n(\msf s) =\log\mathsf Z^Q_n(L)+\int_{\msf s}^L \int_{-\infty}^\infty \mathsf \partial_\msf s \mathsf K^Q_n(x,x\mid u)\omega_n(x\mid u)\mathrm{d} x\mathrm{d} u.
\end{equation}
We want to interchange the order of integration in the above. The derivative $\partial_\msf s\mathsf K^Q_n$ is a polynomial of degree at most $2n-2$ in $x$, and from Heine's formula for orthogonal polynomials we see that the polynomial coefficients of $\mathsf K_n$ are continuous functions of $\msf s$. Therefore, for given $\msf s,L$ there exists a constant $M=M(\msf s,L)>0$ for which the pointwise bound
$$
\left|\partial_\msf s \mathsf K^Q_n(x,x\mid u)\right| \leq M \sup_{0\leq k\leq 2n-2}|x|^k,\quad x\in \mathbb{R},
$$
is valid for every $u\in [\msf s,L]$. Together with the inequalities $0\leq \omega_n(x)\leq \ee^{-nV(x)}$, this bound ensures that we can interchange order of integration in \eqref{eq:propLndef1}. After integration by parts, we then obtain
\begin{multline*}
\log\mathsf Z^Q_n(\msf s) =\log\mathsf Z^Q_n(L)+\int_{-\infty}^\infty \mathsf K^Q_n(x,x\mid L)\omega_n(x\mid L)\mathrm{d} x-\int_{-\infty}^\infty \mathsf K^Q_n(x,x\mid \msf s)\omega_n(x\mid \msf s)\mathrm{d} x \\
-\int_{-\infty}^{\infty} \int_\msf s^L \mathsf K^Q_n(x,x\mid u)\frac{\omega_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} u
\end{multline*}
From the identity \eqref{eq:IntReprKernel} the two single integrals cancel one another. The integrand of the double integral is positive, so by Tonelli's Theorem we can interchange order of integration. After this interchange, we take the limit $L\to +\infty$ and use \eqref{eq:KernelContin} and \eqref{deff:Ln} to conclude the proof.
\end{proof}
\subsection{The Riemann-Hilbert Problem for orthogonal polynomials}\hfill
We are ready to introduce the RHP for orthogonal polynomials for the weight $\omega_n$ in \eqref{def:perturbedweight}. During this section, we keep using the matrix notation that was already used in previous sections, recall for instance \eqref{deff:matrixnot1},\eqref{deff:matrixnot2} and \eqref{deff:matrixnot3}.
The RHP for orthogonal polynomials for the weight \eqref{def:perturbedweight} asks for finding a $2\times 2$ matrix-valued function $\bm Y$ with the following properties.
\begin{enumerate}[\bf Y-1.]
\item The matrix $\bm Y:\mathbb{C}\setminus \mathbb{R}\to \mathbb{C}^{2\times 2}$ is analytic.
\item The function $\bm Y$ has continuous boundary values
$$
\bm Y_\pm(x)\deff \lim_{\varepsilon\searrow 0}\bm Y(x\pm \mathrm{i} \varepsilon),\quad x\in \mathbb{R},
$$
which are related by the jump condition $\bm Y_+(x)=\bm Y_-(x)\bm J_{\bm Y}(x)$, $x\in \mathbb{R}$, with
$$
\bm J_{\bm Y}(x)\deff \bm I+\omega_n(x)\bm E_{12}.
$$
\item As $z\to \infty$,
$$
\bm Y(z)=\left(\bm I+\mathcal{O}(z^{-1})\right)z^{n\sp_3}.
$$
\end{enumerate}
Observe that $\bm Y=\bm Y^{(n)}(\cdot \mid \msf s,Q)$ depends on the index $n$ and also on $\msf s$ and $Q$, although we do not make this dependence explicit in our notation. As shown by Fokas, Its and Kitaev \cite{FokasItsKitaev92}, for each $n$ the RHP above has a unique solution, which is explicitly given by
$$
\bm Y(z)=
\begin{pmatrix}
\mathsf P^{(n,\msf s)}_{n}(z) &\displaystyle{ \frac{1}{2\pi \mathrm{i}}\int_{\mathbb{R}} \frac{P_n^{(n,\msf s)}(x)}{x-z}\omega_n(x)\mathrm{d} x }\\[10pt]
-2\pi \mathrm{i} \upgamma^{(n,Q)}_{n-1}(\msf s)^2 \mathsf P^{(n,\msf s)}_{n-1}(z) & \displaystyle{-\upgamma^{(n,Q)}_{n-1}(\msf s)^2\dfrac{1}{2\pi \mathrm{i}}\int_{\mathbb{R}} \frac{\mathsf P^{(n,\msf s)}_{n-1}(x)}{x-z}\omega_n(x)\mathrm{d} x}
\end{pmatrix},
$$
where $\upgamma_k^{(n,Q)}(\msf s)$ and $\mathsf P_k^{(n,\msf s)}$ are as in \eqref{def:opsdeformed} and \eqref{def:normingdeformed}.
In particular, from this identity we obtain the relation
\begin{equation}\label{eq:relationnormingY1}
\upgamma_{n-1}^{(n,Q)}(\msf s)^2=-\frac{1}{2\pi \mathrm{i}}\left(\bm Y^{(n,1)}\right)_{21},
\end{equation}
where $\bm Y^{(n,1)}=\bm Y^{(n,1)}(\msf s,Q)$ is the matrix determined from the more detailed expansion
\begin{equation}\label{eq:ExpYRHPInfinity}
\bm Y(z)=\bm Y^{(n)}(z)=\left(\bm I+\frac{1}{z}\bm Y^{(n,1)}+\frac{1}{z^2}\bm Y^{(n,2)}+\mathcal{O}(z^{-3})\right)z^{n\sp_3},\quad z\to \infty.
\end{equation}
Also, the Christoffel-Darboux kernel \eqref{deff:Kndefweight} can be recast directly from $\bm Y$ from the identity
\begin{equation}\label{eq:KnYxy}
\mathsf K^Q_n(x,y\mid \msf s)=\frac{1}{2\pi\mathrm{i}}\frac{1}{x-y}\bm e_2^\mathrm T \bm Y_+(y)^{-1}\bm Y_+(x) \bm e_1,\quad x,y\in \mathbb{R}, \; x\neq y.
\end{equation}
In the confluent limit $x=y$, this formula yields
\begin{equation}\label{eq:KnYxx}
\mathsf K^Q_n(x,x\mid \msf s)=\frac{1}{2\pi\mathrm{i}}\bm e_2^\mathrm T \bm Y_+(x)^{-1}\bm Y'_+(x) \bm e_1,\quad x\in \mathbb{R}.
\end{equation}
The remainder of this paper is dedicated to applying the Deift-Zhou method for this RHP and collecting its consequences, analysis which will ultimately lead to the proofs of our main results.
\section{The RHP analysis for the orthogonal polynomials}\label{sec:rhpanalysis}
With all the preliminary work completed, we are finally at the stage of performing the asymptotic analysis for the {\bf RHP-$\bm Y$} for orthogonal polynomials that was introduced in Section~\ref{sec:rhpapproachops}. Most of the transformations are standard, so we go over them quickly and without much detail. Care will be taken in the construction of the parametrices, which are the steps where the introduction of the factor $\sigma_n$ plays a major role. We also remind the reader that the function $V$ and $Q$ are always assumed to satisfy Assumptions~\ref{MainAssumption}.
The function $\sigma_n=\sigma_n(z\mid \msf s,Q)$ depends on $\msf s\in \mathbb{R}$ and $Q$, and as such in all the steps below several quantities will also depend on these parameters. Nevertheless, in most of the work that follows the parameter $\msf s$ and the function $Q$ do not play a major role so we omit them in our notations unless when needed to avoid confusion.
\subsection{First transformation: normalization at infinity}\hfill
Recall the function $\phi$ introduced in \eqref{eq:functionphitau}. The first transformation, which has the effect of normalizing the RHP as $z\to \infty$, takes the form
\begin{equation}\label{eq:transfYT}
\bm T(z)\deff \ee^{-n\ell_V \sp_3} \bm Y(z)\ee^{n\left(\phi(z)-\frac{1}{2}V(z)\right)\sp_3},\quad z\in \mathbb{C}\setminus \mathbb{R}.
\end{equation}
From the RHP for $\bm Y$ and the properties from Proposition~\ref{prop:functionphitau}, we obtain that $\bm T$ satisfies the following RHP.
\begin{enumerate}[\bf T-1.]
\item The matrix $\bm T:\mathbb{C}\setminus \mathbb{R}\to \mathbb{C}^{2\times 2}$ is analytic.
\item For $z\in \mathbb{R}$, it satisfies the jump $\bm T_+(z)=\bm T_-(z)\bm J_{\bm T}(z)$, with
$$
\bm J_{\bm T}(z)
\deff
\begin{pmatrix}
\ee^{n(\phi_{+}(z)-\phi_{-}(z))} & \sigma_{n}(z)\ee^{-n(\phi_{+}(z)+\phi_{-}(z))} \\
0 & \ee^{-n(\phi_{+}(z)-\phi_{-}(z))}
\end{pmatrix},\quad z\in \mathbb{R}.
$$
\item As $z\to \infty$,
$$
\bm T(z)=\bm I+\frac{1}{z}\bm T_1+\mathcal{O}(z^{-2}),
$$
where the coefficient $\bm T_1$ is
$$
\bm T_1\deff \ee^{-n\ell_V \sp_3}\bm Y^{(n,1)}\ee^{n\ell_V \sp_3}+\phi_\infty \sp_3,
$$
and we recall that $\bm Y^{(n,1)}$ and $\phi_\infty$ were introduced in \eqref{eq:ExpYRHPInfinity} and in Proposition~\ref{prop:functionphitau}--(v), respectively.
\end{enumerate}
From the properties (ii) of Proposition~\ref{prop:functionphitau}, the jump matrix for $\bm T$ simplifies in convenient ways.
For $-a<z<0$,
\begin{align*}
\bm J_\bm T(z) & =
\begin{pmatrix}
\ee^{2n\phi_{+}(z)} & \sigma_n(z) \\
0 & \ee^{-2n\phi_{+}(z)}
\end{pmatrix} \\
& =\left(\bm I+\frac{1}{\sigma_n(z)} \ee^{-2n\phi_{+}(z)}\bm E_{21}\right)\left(\sigma_n(z)\bm E_{12}-\frac{1}{\sigma_n(z)}\bm E_{21}\right)\left(\bm I+\frac{1}{\sigma_n(z)} \ee^{2n\phi_{+}(z)}\bm E_{21}\right) \\
& =\left(\bm I+\frac{1}{\sigma_n(z)}\ee^{2n\phi_{-}(z)}\bm E_{21}\right)\left(\sigma_n(z)\bm E_{12}-\frac{1}{\sigma_n(z)}\bm E_{21}\right)\left(\bm I+\frac{1}{\sigma_n(z)} \ee^{2n\phi_{+}(z)}\bm E_{21}\right),
\end{align*}
and for $z\in \mathbb{R}\setminus [-a,0]$,
\begin{align*}
\bm J_\bm T(z)=\bm I+\sigma_n(z) \ee^{-2n\phi_{+}(z)}\bm E_{12}.
\end{align*}
\subsection{Second transformation: opening of lenses}\hfill
From the identities just written for $\bm J_{\bm T}$ and Proposition~\ref{prop:functionphitau}-(ii), it follows that the diagonal entries of $\bm J_{\bm T}$ are highly oscillatory on $(-a,0)$ as $n\to\infty$. In the second transformation of the RHP we perform the so-called opening of lenses, which has the effect of moving this oscillatory behavior to a region where it becomes exponentially decaying.
Define regions $\mathcal G^\pm$ on the $\pm$-side of $(-a,0)$ (the lenses, see Figure~\ref{fig:openinglenses}), assuming in addition that for $U_0$ as in Proposition~\ref{prop:conformalmappsi} these regions satisfy
\begin{equation}\label{eq:transformationlenses}
\psi(\partial \mathcal G^\pm \cap U_0 )\subset (0,\ee^{\pm 2\pi \mathrm{i} /3}\infty)\cup (0,\infty),
\end{equation}
which can always be achieved because $\psi:$ is conformal from a neighborhood of $U_0$ to a neighborhood of $D_\delta(0)$.
The function $\sigma_n$ has no zeros and may have singularities, but these are all poles due to the analyticity of $Q$ in a neighborhood of the real axis. Therefore, the fraction $1/\sigma_n$ is analytic on a neighborhood of the real axis. We use this fraction to transform
$$
\bm S(z)\deff
\begin{dcases}
\bm T(z)\left(\bm I\mp \frac{1}{\sigma_n(z)} \ee^{2n\phi(z)}\bm E_{21}\right),& z\in \mathcal G^\pm, \\
\bm T(z), & \text{elsewhere}.
\end{dcases}
$$
\begin{figure}[t]
\centering
\begin{tikzpicture}[scale=1]
\node at (0,0) (a) {};
\node at (5,0) (b) {};
\node at (6,0) (b1) {};
\fill (a) circle[radius=2.5pt] node [above left] {$-a$};
\fill (b1) circle[radius=2.5pt] node [above right] {$0$};
\draw [thick, postaction={vmid arrow={black,scale=1.5}{.7}}] (a.center) to (b.center);
\draw [thick, postaction={vmid arrow={black,scale=1.5}{.6}}] (b.center) to (b1.center);
\draw [thick, postaction={vmid arrow={black,scale=1.5}{.5}}] ($(a)+(-2,0)$) to (a.center);
\draw [thick, postaction={vmid arrow={black,scale=1.5}{.5}}] (b1.center) to ($(b1)+(2,0)$);
\draw [thick, postaction={vmid arrow={black,scale=1.5}{.6}}] (a.center) to [out=45,in=180-45,edge node={node [pos=0.5,below,yshift=-8pt,xshift=-5pt] {$\mathcal G^+$}}] (b1.center);
\draw [thick, postaction={vmid arrow={black,scale=1.5}{.6}}] (a.center) to [out=-45,in=180+45,edge node={node [pos=0.5,above,yshift=8pt,xshift=-5pt] {$\mathcal G^-$}}] (b1.center);
\end{tikzpicture}
\caption{The regions used for the opening of lenses in the transformation $\bm T\mapsto \bm S$.}\label{fig:openinglenses}
\end{figure}
With
$$
\Gamma_{\bm S}\deff\mathbb{R}\cup \partial \mathcal G^+\cup \partial \mathcal G^-,
$$
and using the jump properties of $\phi$ listed in Proposition~\ref{prop:functionphitau}, the matrix $\bm S$ satisfies the following RHP.
\begin{enumerate}[\bf S-1.]
\item The matrix $\bm S:\mathbb{C}\setminus \Gamma_{\bm S}\to \mathbb{C}^{2\times 2}$ is analytic.
\item For $z\in \Gamma_{\bm S}$, it satisfies the jump $\bm S_+(z)=\bm S_-(z)\bm J_{\bm S}(z)$, with
$$
\bm J_{\bm S}(z)\deff
\begin{dcases}
\sigma_n(z)\bm E_{12}-\frac{1}{\sigma_n(z)}\bm E_{21}, & -a<z<0, \\
\bm I+ \frac{1}{\sigma_n(z)}\ee^{2n\phi(z)}\bm E_{21}, & z\in \partial \mathcal G^\pm\setminus \mathbb{R}, \\
\bm I+\sigma_n(z)\ee^{-2n\phi_{+}(z)}\bm E_{12}, & z\in \mathbb{R}\setminus (-a,0).
\end{dcases}
$$
\item As $z\to \infty$,
$$
\bm S(z)=\bm I+\frac{\bm S_1}{z}+\mathcal{O}(z^{-2}),\qquad \text{with}\quad \bm S_1\deff \bm T_1.
$$
\item The matrix $\bm S$ remains bounded near the points $z=-a,0$.
\end{enumerate}
Before moving to the construction of the mentioned parametrices, we conclude this section with the needed estimate for the jump matrix $\bm J_{\bm S}$ away from $[-a,0]$. For that, recall the matrix norm notation introduced in \eqref{deff:matrixnorm},\eqref{deff:matrixLpnorm},\eqref{deff:matrixLpqnorm}.
\begin{prop}\label{prop:decayJS}
For $U_0$ as in Proposition~\ref{prop:conformalmappsi}, introduce the set
$$
\Gamma_\varepsilon\deff \Gamma_{\bm S}\setminus \left([-a,0]\cup U_0\cup D_\varepsilon(-a)\right)
$$
For some $\varepsilon>0$, and possibly reducing $U_0$ if necessary, there is an $\eta>0$ such that
$$
\|\bm J_{\bm S}-\bm I\|_{L^1\cap L^2\cap L^\infty(\Gamma_\varepsilon)}=\mathcal{O}(\ee^{-\eta n}),
$$
as $n\to \infty$.
\end{prop}
\begin{proof}
From Proposition~\ref{prop:functionphitau}-(iv) we obtain that for any $\varepsilon>0$ there is a constant $\eta'>0$ for which
$$
\re \phi_{+}(x)\geq \eta',\quad \text{for every } x\in (-a-\varepsilon,\varepsilon).
$$
On the other hand, the jump conditions on Proposition~\ref{prop:functionphitau}-(ii) combined with Cauchy-Riemann equations imply in a standard way that $\re \phi\leq - \eta$ along the lipses of the lenses and away from the endpoints $-a$ and $0$, as long as the lens stay within a positive but small distance from the interval $[-a,0]$. From these pointwise estimates, the growth of $\re \phi$ as $z\to \pm\infty$ and the fact that $\sigma_n$ remains bounded on $\mathbb{R}$ and $1/\sigma_n$ grows at most with $\mathcal{O}(\ee^{c n^{2/3}})$ on compacts of $\mathbb{C}$, the claimed $L^p$ estimates follow in a standard manner.
\end{proof}
\subsection{Global parametrix}\hfill
The global parametrix problem, obtained after neglecting the jumps of $\bm S$ that are exponentially close to the identity, is the following RHP.
\begin{enumerate}[\bf G-1.]
\item $\bm G:\mathbb{C}\setminus [-a,0] \to \mathbb{C}^{2\times 2}$ is analytic.
\item For $z\in (-a,0)$, it satisfies the jump
$$
\bm G_+(z)=\bm G_-(z)\left(\sigma_n(z)\bm E_{12}-\frac{1}{\sigma_n(z)}\bm E_{21}\right).
$$
\item As $z\to \infty$,
$$
\bm G(z)=\bm I+\mathcal{O}(z^{-1}).
$$
\item $\bm G$ has square-integrable singularities at $z=-a,0$.
\end{enumerate}
The construction of the global parametrix follows standard techniques. First, one introduces a function that we denote $\mathsf q(z)$, with the aim at transforming the RHP for $\bm G$ to a RHP with constant jumps. Then, by diagonalizing the resulting jump matrix, we further reduce the problem to two scalar-valued RHPs. With the help of Plemelj's formula, we then solve these scalar RHPs, and by tracing back all the transformations we recover the matrix $\bm G$ itself.
The procedure just described is standard in RHP literature, see for instance \cite[Appendix~A.1]{BKMMbook}, so we refrain from completing it in detail and instead only describe the final form of the solution.
The function $\sigma_n$ does not vanish and is real and positive over the real axis, so its real logarithm over the real axis is well defined. With this in mind, introduce
$$
\mathsf q(z)\deff \frac{((z+a)z)^{1/2}}{2\pi }\int_{-a}^{0} \frac{\log \sigma_n(x)}{\sqrt{|x|(x+a)}} \frac{\mathrm{d} x}{x-z},\quad z\in \mathbb{C}\setminus [-a,0],
$$
where $(\cdot)^{1/2}$ stands for the principal branch of the square root and $\sqrt{\cdot}$ is reserved for the standard positive real root of positive real numbers.
This function $\mathsf q$ depends on $n$ but we do not make this dependence explicit for ease of notation. It is analytic on $\mathbb{C}\setminus [-a,0]$, and it is chosen to satisfy the jump condition
$$
\mathsf q_+(x)+\mathsf q_-(x)=-\log\sigma_n(x),\quad -a<x<0.
$$
Furthermore, standard calculations show that
$$
\mathsf q(z)=\mathcal{O}(1),\quad z\to -a,0,\qquad \text{and} \qquad \mathsf q(z)=\mathsf q_0+\frac{\mathsf q_1}{z} + \mathcal{O}(z^{-2}), \quad z\to\infty,
$$
with coefficients given by
\begin{equation}\label{eq:def_q0q1}
\mathsf q_0=\mathsf q_0(n)\deff-\frac{1}{2\pi}\int_{-a}^{0} \frac{\log\sigma_n(x)}{\sqrt{|x|(x+a)}}\mathrm{d} x \quad \text{and}\quad
\mathsf q_1=\mathsf q_1(n)\deff-\frac{1}{2\pi}\int_{-a}^{0} \frac{x\log\sigma_n(x)}{\sqrt{|x|(x+a)}}\mathrm{d} x +\frac{a\mathsf q_0}{2}.
\end{equation}
Next, set
\begin{equation}\label{eq:def_U0}
\bm U_0\deff\frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & \mathrm{i} \\ \mathrm{i} & 1
\end{pmatrix},
\quad
\mathsf m(z)\deff \frac{z}{z+a},
\end{equation}
which is consistent with \eqref{def:matrixU0modelprobl}, and introduce
\begin{equation}\label{eq:def_globalparM}
\bm M(z)\deff
\bm U_0
\mathsf m(z)^{\sp_3/4}
\bm U_0^{-1}
=
\frac{1}{2}
\begin{pmatrix}
\mathsf m(z)+\dfrac{1}{\mathsf m(z)} & -\mathrm{i} \left(\mathsf m(z)-\dfrac{1}{\mathsf m(z)}\right) \\
\mathrm{i} \left(\mathsf m(z)-\dfrac{1}{\mathsf m(z)}\right) & \mathsf m(z)+\dfrac{1}{\mathsf m(z)}
\end{pmatrix}
\end{equation}
This matrix $\bm M$ satisfies
$$
\bm M_+(z)=\bm M_-(z)\left(\bm E_{12}-\bm E_{21}\right), \; -a<z<0, \quad \text{and}\quad \bm M(z)=\bm I-\frac{a}{4z}\sp_2+\mathcal{O}(z^{-2}),
$$
where $\sp_2$ is the second Pauli matrix (recall \eqref{deff:matrixnot2}).
Then the solution to the global parametrix {\bf RHP-}{$\bm G$} is
\begin{equation}\label{eq:constructionGparam}
\bm G(z) = \ee^{-\mathsf q_0\sp_3}\bm M(z) \ee^{\mathsf q(z)\sp_3},\quad z\in \mathbb{C}\setminus [-a,0].
\end{equation}
This solution $\bm G$ satisfies
\begin{equation}\label{eq:definitionG1}
\bm G(z)=\mathsf I+\frac{\bm G_1}{z}+\mathcal{O}(z^{-2}),\; z\to \infty,\quad \text{with}\quad \bm G_1\deff
\begin{pmatrix}
\mathsf q_1 & \dfrac{\mathrm{i} a}{4}\ee^{-2\mathsf q_0} \\ -\dfrac{\mathrm{i} a }{4}\ee^{2\mathsf q_0} & -\mathsf q_1
\end{pmatrix}.
\end{equation}
Recall that $U_0$ denotes the neighborhood of the origin given in Proposition~\ref{prop:conformalmappsi}. We will also need some control on $\mathsf q$ inside $U_0$.
For the next result, set
\begin{equation}\label{deff:msfqo1}
\begin{aligned}
& F_{\beta}(\msf s)\deff \int_0^\infty v^{\beta}\log(1+\ee^{-\mathsf s-v})\mathrm{d} v, \\
&\mathsf q_0^{(1)}=\mathsf q_0^{(1)}(\msf s,\msf t)\deff \frac{\msf t^{1/2}}{2\pi a^{1/2}}F_{-1/2}(\msf s), \\
& \mathsf q^{(1)}(z)=\mathsf q^{(1)}(z\mid \msf s,\msf t)\deff \frac{\mathsf t^{1/2}}{2\pi a^{1/2}\mathsf m(z)^{1/2}}F_{-1/2}(\msf s),
\end{aligned}
\end{equation}
which are $n$-independent quantities. The index $\beta$ does not have any specific meaning for what comes later, but it arises naturally from the asymptotic analysis resulting in the following result.
\begin{lemma}\label{lem:estimate_q_q0}
For any fixed $\msf s_0>0$, the estimate
$$
\mathsf q_0=\frac{1}{n^{1/3}}\mathsf q_0^{(1)}+\mathcal{O}(n^{-2/3}),\quad n\to \infty,
$$
is valid uniformly for $\msf s \geq -\msf s_0$. In addition, the estimates
$$
\mathsf q(z)=\frac{1}{n^{1/3}}\mathsf q^{(1)}(z)+\mathcal{O}(n^{-2/3}),\quad \text{and}\quad \mathsf q'(z)=\mathcal{O}(n^{-1/3}),
$$
are valid uniformly for $z$ on compacts of $\mathbb{C}\setminus [-a,0]$ (in particular on $\partial U_0$) and uniformly for $\msf s \geq -\msf s_0$, and carry through to boundary values $\mathsf q_\pm(x)$ for $x$ along $\mathbb{R}\setminus \{-a,0\}$.
Finally,
$$\bm M(z)=\left(\bm I+\frac{1}{n^{1/3}}\left( \mathsf q^{(1)}_0\sp_3 -\mathsf q^{(1)}(z) \bm M(z)\sp_3\bm M(z)^{-1}\right)+ \mathcal{O}(n^{-2/3}) \right)\bm G(z),\quad n\to\infty,$$
uniformly for $z\in \partial U_0$ and $s\geq -\msf s_0$.
\end{lemma}
\begin{proof}
The estimate for $\mathsf q_0(n)$ follows immediately from an application of Proposition~\ref{prop:LaplaceLogIntApp}. The estimates for $\mathsf q(z)$ and $\mathsf q'(z)$ also follow from Proposition~\ref{prop:LaplaceLogIntApp}, once we observe that the integrals defining them can be slightly deformed to the upper/lower half plane in a neighborhood of the unique point in the intersection $\partial U_0\cap (-a,0)$.
Finally, using the first part of the statement and the fact that $\bm M$ is bounded for $z\in\partial U_0$ and independent of $n$, we expand the exponentials in series and write
\begin{align*}
\bm M(z)\bm G(z)^{-1} & =\bm M(z)\ee^{-\mathsf q(z)\sp_3}\bm M(z)^{-1}\ee^{\mathsf q_0\sp_3}\\
& =\left(\bm I-\mathsf q(z)\bm M(z)\sp_3\bm M(z)^{-1}+\mathcal{O}(n^{-2/3})\right)\left(\bm I+\mathsf q_0\sp_3+\mathcal{O}(n^{-2/3})\right),
\end{align*}
and the last claim follows after rearranging the terms in this expansion.
\end{proof}
\subsection{Local Parametrix near $-a$}\hfill
The local parametrix $\bm P=\bm P^{(a)}$ near $z=-a$ is constructed in a neighborhood of $z=-a$ which without loss of generality can be taken to be the disk $D_\delta(-a)$ of radius $\delta$ around $a$, and it is the solution to the following RHP.
\begin{enumerate}[\bf $\bm P^{({a})}$-1.]
\item The matrix $\bm P^{(a)}:D_\delta(-a)\setminus \Gamma_{\bm S}\to \mathbb{C}^{2\times 2}$ is analytic.
\item For $z\in \Gamma_{\bm S}\cap \partial D_\delta(-a)$, it satisfies the jump $\bm P^{(a)}_+(z)=\bm P^{(a)}_-(z)\bm J_{\bm S}(z)$.
\item Uniformly for $z\in \partial D_\delta(-a)$,
$$
\bm P^{(a)}(z)=\left(\bm I+\mathit{o}(1)\right)\bm G(z),\quad n\to\infty.
$$
\item The matrix $\bm P^{(a)}$ remains bounded as $z\to -a$.
\end{enumerate}
The asymptotic condition {\bf $\bm P^{({a})}$-3.} above will be improved to \eqref{eq:decayparametrixa} below.
From the conditions on $Q$, we know that there exists a value $\eta>0$ for which
$$
\re Q(z)\geq 2\eta,\quad |z+a|<\delta.
$$
This value is uniform for $\msf t\in [\msf t_0,1/\msf t_0],$ for any $\msf t_0\in (0,1)$ fixed, and it is independent of $\msf s\in \mathbb{R}$.
In particular, once we fix $\msf s_0>0$ and assume that $\msf s\geq -\msf s_0$, from this inequality we obtain
$$
|\ee^{-\msf s-n^{2/3}Q(z)}|\leq \ee^{-n^{2/3}\eta},\quad |z+a|<\delta,\quad \text{for large enough } n.
$$
This way, for $n>0$ sufficiently large the function $\sigma_n$ admits an analytic continuation to the whole disk $D_\delta(-a)$, and this continuation does not have zeros on the same disk. Thus, a branch of $\log\sigma_n$ is well defined in a neighborhood of $z=-a$, and the just mentioned estimate also shows that
\begin{equation}\label{eq:EstLogSigman}
\log\sigma_n(z)=\mathcal{O}(\ee^{-\eta n^{2/3}}),\quad n\to\infty,
\end{equation}
uniformly for $z$ in a neighborhood of $z=-a$ and $\msf s\geq -\msf s_0$.
With this in mind, the parametrix $\bm P^{(a)}$ can be constructed explicitly out of Airy functions in a standard way, see for instance \cite[Section~7.6]{deift_book}. Since it involves a somewhat nonstandard matching analytic prefactor that accounts for $\sigma_n$, we briefly go over this construction.
Recall the contour $\msf \Sigma$ introduced in \eqref{def:contour_sigma}. With appropriate Airy functions, we construct a $2\times 2$ matrix $\bm\Psi_{\ai}$, which is analytic on $\mathbb{C}\setminus \msf \Sigma$ and satisfies
$$
\bm \Psi_{\ai,+}(\zeta)=\bm \Psi_{\ai,-}(\zeta)\times
\begin{cases}
\bm I-\bm E_{12}, & \zeta \in \msf \Sigma_0, \\
\bm I-\bm E_{21}, & \zeta \in \msf \Sigma_1\cup\msf \Sigma_3, \\
-\bm E_{12}+\bm E_{21}, & \zeta \in \msf \Sigma_2, \\
\end{cases}
$$
and
\begin{equation}\label{eq:AsympAiryPar}
\bm\Psi_{\ai}(\zeta)=\zeta^{\sp_3/4}\bm U_0\left(\bm I+\mathcal{O}(\zeta^{-3/2})\right)\ee^{-\frac{2}{3}\zeta^{3/2}\sp_3},\quad \zeta\to \infty.
\end{equation}
In fact, $\bm\Psi_{\ai}$ can be obtained with a modification of the matrix ${\bm \Phi}_{\rm Ai}$ which we previously used in \eqref{deff:Phiai}. We will not need its explicit form, so we do not write it down explicitly.
Using the properties of $\phi$ we construct a conformal map $\varphi$ from a neighborhood of $-a$ to a neighborhood of the origin, with
$$
\varphi(-a)=0, \quad \varphi'(-a)<0\quad \text{and}\quad \frac{2}{3}\varphi(z)^{3/2}=\phi(z)+2\pi \mathrm{i} \mathbb{Z}, \; z\in D_\delta(-a)\setminus \msf \Sigma_{\bm S}.
$$
With standard arguments (see for instance the proof of Proposition~\ref{prop:estimateL} below for similar arguments), one shows that the matrix
\begin{equation}\label{eq:FFactorRegPar}
\bm F^{(a)}(z)\deff \bm G(z)\ee^{\frac{1}{2}\log\sigma_n(z)\sp_3}\bm U_0^{-1}(n^{2/3}\varphi(z))^{-\sp_3/4}
\end{equation}
is analytic on a neighborhood of $z=-a$. The local parametrix then takes the form
\begin{equation}\label{deff:LocalParRegEdge}
\bm P^{(a)}(z)=\bm F^{(a)}(z)\bm \Psi_{\ai}(n^{2/3}\varphi(z))\ee^{-\frac{1}{2}\log\sigma_n(z)\sp_3}\ee^{n\phi(z)\sp_3},\quad z\in D_\delta(-a)\setminus \Gamma_{\bm S}.
\end{equation}
As a result, the error term in fact takes on the stronger form
\begin{equation}\label{eq:decayparametrixa}
\bm P^{(a)}(z)=\bm G(z)\left(\bm I+\mathcal{O}(n^{-1})\right),\quad n\to\infty,
\end{equation}
which is valid uniformly for $z\in \partial D_\delta(-a)$ and uniformly for $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$, for any $\msf s_0>0$ and $\msf t_0\in (0,1)$ fixed.
\subsection{Local Parametrix near the origin}\hfill
The local parametrix near the origin requires the model problem from Section~\ref{sec:modelproblem}.
Recall the neighborhood $U_0$ of the origin introduced in Proposition~\ref{prop:conformalmappsi}. The initial local parametrix we seek for should be the solution to the following RHP.
\begin{enumerate}[\bf $\bm P^{(0)}$-1.]
\item The matrix $\bm P^{(0)}:U_0\setminus \Gamma_{\bm S}\to \mathbb{C}^{2\times 2}$ is analytic.
\item For $z\in \Gamma_{\bm S}\cap U_0$, it satisfies the jump $\bm P^{(0)}_+(z)=\bm P^{(0)}_-(z)\bm J_{\bm S}(z)$.
\item Uniformly for $z\in \partial U_0$,
$$
\bm P^{(0)}(z)=\left(\bm I+\mathit{o}(1)\right)\bm G(z),\quad n\to\infty.
$$
\item $\bm P^{(0)}$ remains bounded as $z\to 0$.
\end{enumerate}
To construct the solution $\bm P^{(0)}$ required above, some work is needed. Aiming at removing $\phi$ from the jump of $\bm P^{(0)}$, we change this {\bf RHP-$\bm P$} with the transformation
\begin{equation}\label{local_param_LP_rel}
\bm L(z)=\bm P^{(0)}(z)\ee^{-n\phi(z)\sp_3 },\quad z\in U_0\setminus \Gamma_{\bm S}.
\end{equation}
Then the matrix $\bm L$, should it exist, must satisfy the following RHP.
\begin{enumerate}[\bf $\bm L$-1.]
\item The matrix $\bm L:U_0\setminus \Gamma_{\bm S}\to \mathbb{C}^{2\times 2}$ is analytic.
\item For $z\in \Gamma_{\bm S}\cap U_0$, it satisfies the jump $\bm L_+(z)=\bm L_-(z)\bm J_{\bm L}(z)$,
with
\begin{equation}\label{eq:jumpJL}
\bm J_{\bm L}(z)\deff
\begin{dcases}
\sigma_n(z)\bm E_{12}-\frac{1}{\sigma_n(z)}\bm E_{21}, & z\in U_0\cap (-a,0), \\
\bm I+\frac{1}{\sigma_n(z)}\bm E_{21}, & z\in \partial\mathcal G^\pm\cap U_0,\\
\bm I+\sigma_n(z)\bm E_{12}, & z\in U_0\cap (0,\infty).
\end{dcases}
\end{equation}
\item Uniformly for $z\in \partial U_0$,
$$
\bm L(z)=\left(\bm I+\mathit{o}(1)\right)\bm G(z)\ee^{-n\phi(z)\sp_3 },\quad n\to\infty.
$$
\item The matrix $\bm L$ remains bounded as $z\to 0$.
\end{enumerate}
Based on the usual way of matching the local parametrix with a model problem, one is tempted to moving the non-constant part of the jump - namely $\sigma_n$ - to the behavior at $\partial U_0$ as well. This would be done so including a term of the form $\sigma_n^{\sp_3/2}=\ee^{\sp_3\log\sigma_n/2}$ into the transformation $\bm P\mapsto \bm L$, in much the same way we did in \eqref{deff:LocalParRegEdge}. However, as we discussed in Section~\ref{sec:issues}, for any $s\in \mathbb{R}$ fixed there are poles of $\sigma_n$ accumulating too fast near the origin, so $\sigma_n$ fails to be analytic in any small neighborhood of the origin and we have to stick to the non-constant jumps as above.
The {\bf RHP-$\bm L$} has a solution if, and only if, {\bf RHP-$\bm P^{(0)}$} has a solution. Such solutions need not be unique, as one could possibly improve on the asymptotic matching conditions on $\partial U_0$. The goal of the rest of this section is to describe a solution $\bm L$, and consequently a solution $\bm P^{(0)}$ related by \eqref{local_param_LP_rel}, with a more explicit control of the error term in {\bf RHP-$\bm L$-3}. For that, we use the model problem thoroughly studied in Sections~\ref{sec:modelproblem} and \ref{sec:asympanalymodelprobl}.
The construction that follows needs several quantities that appeared before. These are the conformal map $\psi$ appearing in Proposition~\ref{prop:conformalmappsi}, the function $\mathsf H_Q$ introduced with the help of Proposition~\ref{prop:ConformExtH}, the model RHP solution $\bm \Phi=\bm \Phi(\cdot \mid \mathsf h)$ introduced in Section~\ref{sec:modelproblem} and further discussed in Section~\ref{sec:asympanalymodelprobl}, and the constant $\mathsf q_0$ and matrices $\bm U_0$ and $\bm M(z)$ from \eqref{eq:def_q0q1}--\eqref{eq:def_globalparM}. With all these quantities at hand, we set
\begin{equation}\label{local_par_L_Phi}
\begin{aligned}
& \bm L(z)\deff \bm F_n(z)\bm \Phi_n(z),\quad z\in U_0\setminus \msf \Sigma_{\bm S},\qquad \text{with the choices} \\
& \widehat{\bm\Phi}_n(\zeta)=\bm \Phi\left(\zeta \mid \mathsf h=\mathsf h_n\right), \quad \mathsf h_n(\zeta)\deff \msf s+n^{2/3}\mathsf H_Q(\zeta/n^{2/3}), \\
& {\bm \Phi}_n(z)\deff \widehat{\bm \Phi}_n\left(\zeta=n^{2/3}\psi(z)\right),\\
& {\bm F}_n(z)\deff \bm M(z)\bm U_0(n^{2/3}\psi(z))^{-\sp_3/4}=\bm U_0\mathsf m(z)^{\sp_3/4}(n^{2/3}\psi(z))^{-\sp_3/4}.
\end{aligned}
\end{equation}
With the identification $\tau=n^{2/3}$ and thanks to Proposition~\ref{prop:ConformExtH}, the function $\mathsf h_n$ becomes admissible in the sense of Definition~\ref{deff:admissibleh}, so the notation for the corresponding solution $\bm\Phi_n=\bm \Phi(\cdot\mid \mathsf h_n)$ chosen above is consistent with the solution $\bm \Phi_\tau(\zeta)\big|_{\tau=n^{2/3}}$ in \eqref{deff:Phit}. For later reference, we keep track of the expansion
\begin{equation}\label{eq:expansionmsfhnorigin}
\mathsf h_n(\zeta)=\msf s-\mathsf c_{\mathsf H}\zeta+\mathcal{O}(\zeta^{-2}),\quad \zeta\to 0,\quad \text{where we recall that}\quad \mathsf c_{\mathsf H}=\frac{\msf t}{\mathsf c_V},
\end{equation}
compare \eqref{eq:expansionHQ} with Definition~\ref{deff:admissibleh}--(ii).
\begin{prop}\label{prop:propertieshatPhi}
Fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. There exists $n_0=n_0(\msf s_0,\msf t_0)$ for which for any $s\geq -\msf s_0$ and any $\msf t\in [\msf t_0,1/\msf t_0]$ the matrix ${\bm \Phi}_n$ exists for every $n\geq n_0$. This matrix satisfies the jump
$$
{\bm \Phi}_{n,+}(z)={\bm \Phi}_{n,-}(z)\bm J_{\bm L}(z),\quad z\in U_0\cap \Gamma_{\bm S}.
$$
Furthermore, for the matrix
$$
{\bm \Phi}^{(1)}_n\deff \bm \Phi^{(1)}(\mathsf h=\mathsf h_n),\quad \text{with }\bm \Phi^{(1)}(\mathsf h) \text{ as in \eqref{eq:ModelRHPAsymp}},
$$
the asymptotic expansion
\begin{equation*}
{\bm \Phi}_n(z)=\left(\bm I+\frac{1}{n^{2/3}}\frac{1}{\psi(z)}{\bm \Phi}_n^{(1)}+\mathcal{O}(n^{-4/3})\right)(n^{2/3}\psi(z))^{\sp_3/4} \bm U_0^{-1}
\ee^{-n\phi(z)\sp_3} ,\quad n\to \infty,
\end{equation*}
holds true uniformly for $z\in \partial U_0$ and uniformly for $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$.
\end{prop}
\begin{proof}
The existence of ${\bm \Phi}_n$ is granted by the first claim of Theorem~\ref{thm:PhitauAsympt}. The jumps for ${\bm \Phi}_{n,+}$ follow from the jumps in \eqref{eq:jumpPhimodel}, the definition of $\mathsf h_n$ taken in \eqref{local_par_L_Phi}, the correspondence \eqref{eq:RelHQQpsi} and the conformality of the change of variables $\zeta=n^{2/3}\psi(z)$. The asymptotic expansion for ${\bm \Phi}_n$ is immediate from \eqref{eq:ModelRHPAsymp}
\end{proof}
The introduction of the additional notation $\widehat{\bm \Phi}_n$, which plays the role of the local parametrix in the variable $\zeta$, is convenient for later calculations. At that moment, some of its properties will be needed, and we keep track of these properties with the next result. For the formal statement, we recall that $\Phi(\xi\mid \msf S,\msf T)$ is the solution to the integro-differential PII that already appeared in \eqref{eq:intdiffPII} and \eqref{eq:transfPsiccphiad}, and $\bm\Phi_0^{(1)}$ is the residue matrix from \eqref{deff:Phikpz_residue} that collects the functions $\mathsf P(\msf S,\msf T)$ and $\mathsf Q(\msf S,\msf T)$ which, in turn, are related to $\Phi$ as explained in \eqref{deff:coeffqrp} {\it et seq.}.
\begin{prop}
Fix $\msf s_0>0$ and $\msf t_0\in (0,1)$ and $\nu\in (0,2/3)$, and let $\mathsf c_{\mathsf H}$ be as in \eqref{eq:expansionHQ}. The following asymptotic formulas hold true uniformly for $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$.
The matrix $\bm\Phi_n^{(1)}$ from Proposition~\ref{prop:propertieshatPhi} satisfies
\begin{equation}\label{eq:asympPhin1Phi01residue}
\left({\bm \Phi}_n^{(1)}\right)_{21}=\frac{\mathrm{i} }{\mathsf c_{\mathsf H}^{1/2}}\left(\mathsf p(\msf s,\mathsf c_{\mathsf H})-\frac{\msf s^2}{4\mathsf c_{\mathsf H}^{3/2}}\right)+\mathcal{O}(n^{-\nu}),\quad n\to\infty.
\end{equation}
Furthermore, for ${\bm \Phi}_0$ being the solution to the RHP in \eqref{eq:jumpPhicc}--\eqref{eq:AsymptPhicc}, the estimate
\begin{equation}\label{eq:asympPhinPhi0}
\widehat{\bm \Phi}_{n,+}(\zeta)=(\bm I+\mathcal{O}(n^{-\nu})){\bm \Phi}_{0,+}(\zeta\mid \msf s,\mathsf c_{\mathsf H}),\quad n\to\infty,
\end{equation}
holds true uniformly for $\zeta\in \msf \Sigma$ with $|\zeta|\leq n^{\frac{1}{3}-\frac{1}{2}\nu}$.
\end{prop}
\begin{proof}
It is immediate from the identification $\tau=n^{2/3}$, Equations~\eqref{local_par_L_Phi} and Theorem~\ref{thm:PhitauAsympt}.
\end{proof}
Next, we now verify that $\bm L$ given as in \eqref{local_par_L_Phi} indeed solves the {\bf RHP-{$\bm L$}}.
\begin{prop}\label{prop:estimateL}
The matrix $\bm L$ solves the {\bf RHP-{$\bm L$}}.
Furthermore, setting
\begin{equation}\label{eq:AsymptLlocalpar}
\bm L_1(z)\deff \frac{\left({\bm \Phi}^{(1)}_n\right)_{21}}{\psi(z)^{1/2}}\bm M(z)\bm U_0 \bm E_{21} \bm U_0^{-1}\bm M(z)^{-1}, \quad z\in \partial U_0,
\end{equation}
the condition $\bm L${\bf -3.} is improved to
\begin{equation}\label{eq:estimateL_proof1}
\bm L(z)=\left(\bm I+\frac{1}{n^{1/3}}\bm L_1(z) + \mathcal{O}(n^{-2/3})\right)\bm M(z)\ee^{-n\phi(z)\sp_3},\quad n\to \infty,
\end{equation}
uniformly for $z\in \partial U_0$ and $\msf s\geq -\msf s_0$, $\msf t\in [\msf t_0,1/\msf t_0]$, for any $\msf s_0>0,\msf t_0\in (0,1)$.
\end{prop}
\begin{proof}
First we prove that ${\bm F}$ is analytic. For that, notice that a jump for it may come only from the factors $\bm M$ and $\psi^{1/4}$, and therefore only possibly in the interval $(-a,0)\cap U_0$. However, along this interval it is simple to compute that $(\psi^{1/4})_\pm = \ee^{\pm \pi \mathrm{i}/4}|\psi|^{1/4}$, and also that
$$
\bm U_0^{-1}\bm J_{\bm M}(x)\bm U_0=\mathrm{i} \sp_3,
$$
from which we obtain
$$
{\bm F}_-(x)^{-1}{\bm F}_+(x)=n^{\sp_3/6}|\psi|^{\sp_3/4}\ee^{-\pi\mathrm{i}\sp_3/4}\mathrm{i} \sp_3 \ee^{-\pi\mathrm{i}\sp_3/4}|\psi|^{-\sp_3/4}n^{-\sp_3/6}=\bm I,
$$
so ${\bm F}$ is indeed analytic across $(-a,0)\cap U_0$. In principle, ${\bm F}$ may have an isolated singularity at $0$, but because
$$
\bm M(z),\bm \varphi(z)^{1/4}=\mathcal{O}(z^{1/4})\quad \text{as }z\to 0,
$$
we see that $z=0$ is a removable singularity. In virtue of the definition of $\bm L$ in \eqref{local_par_L_Phi} and the jump for $\bm \Phi_n$ from Proposition~\ref{prop:propertieshatPhi}, the analyticity of $\bm F$ is enough to conclude that $\bm L$ satisfies {\bf RHP.}$\bm L${\bf -1}.
Knowing that ${\bm F}$ is analytic, the jump for $\bm L$ is precisely the same as the jump for ${\bm \Phi}_n$, and by Proposition~\ref{prop:propertieshatPhi} we thus have that $\bm L$ satisfies {\bf RHP.}$\bm L${\bf -2}.
Finally, to the asymptotic condition \eqref{eq:AsymptLlocalpar}. For that, we use the asymptotic condition for ${\bm \Phi}_n$ given by Proposition~\ref{prop:propertieshatPhi} and the definition of ${\bm F}$ and write
\begin{multline*}
\bm L(z)=\bm M(z)\bm U_0n^{-\sp_3/6}\psi(z)^{-\sp_3/4}\left(\bm I+ \frac{1}{n^{2/3}\psi(z)}{\bm \Phi}^{(1)}_n+\mathcal{O}(n^{-4/3}) \right) \\
\times
n^{\sp_3/6}\psi(z)^{\sp_3/4}\bm U_0^{-1} \ee^{-n\phi(z)\sp_3},\quad n\to\infty,
\end{multline*}
and where the error is uniform for $z\in \partial U_0$ and $\msf s,\msf t$ as claimed. Since $\partial U_0$ remains within a positive distance from the unique zero $z=0$ of $\psi$, the function $|\psi^{1/4}|$ remains bounded from below away from zero, so the corresponding conjugation of the error by the term $\psi^{\sp_3/4}$ does not change the order of the error. Next, the conjugation by $n^{\sp_3/6}$ contributes at most to an error of order $n^{1/3}$, and only in the $(2,1)$-entry. The remaining term $\bm M$ is bounded along $\partial U_0$, so it can be commuted with the error term above without changing its order, leading to \eqref{eq:estimateL_proof1}. And \eqref{eq:estimateL_proof1} is indeed an improvement of the asymptotic condition {\bf RHP.}$\bm L${\bf -3}, because from Lemma~\ref{lem:estimate_q_q0} we know that $\bm M=(\bm I+\mathcal{O}(n^{-1/3}))\bm G$ uniformly as claimed.
\end{proof}
We now trace back the transformation $\bm L\mapsto \bm P^{(0)}$ and construct the latter as
\begin{equation}\label{deff:ParamOrigin}
\bm P^{(0)}(z)\deff {\bm F}_n(z){\bm \Phi}_n(z)\ee^{n\phi(z)\sp_3},\quad z\in U_0\setminus \msf \Sigma_{\bm S},
\end{equation}
keeping track that ${\bm \Phi}_n$ and $\bm F_n$ are introduced in \eqref{local_par_L_Phi}. With this construction and thanks to Lemma~\ref{lem:estimate_q_q0} and \eqref{eq:estimateL_proof1}, the matching condition {\bf RHP.}$\bm P^{(0)}${\bf -3} is
\begin{equation}\label{eq:errorParOriginfinalform}
\bm P^{(0)}(z) =\left(\bm I+\frac{1}{n^{1/3}}\left(\bm L_1(z)+\mathsf q^{(1)}_0\sp_3 -\mathsf q^{(1)}(z) \bm M(z)\sp_3\bm M(z)^{-1}\right)+\mathcal{O}(n^{-2/3})\right)\bm G(z),\quad n\to \infty,
\end{equation}
and the errors are uniform for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\subsection{Final transformation}\label{sec:finalRops}\hfill
The final transformation combines the local and global parametrices to remove the non-decaying jumps from $\bm S$.
Set
$$
U\deff U_0\cup D_\delta(-a),
$$
orienting $\partial U_0$ and $\partial D_\delta(-a)$ in the clockwise direction. With $\bm P^{(a)}$ being the local parametrix near $-a$, $\bm P^{(0)}$ the local parametrix near the origin and $\bm G$ the global parametrix, we introduce a last parametrix $\bm P$ with unified notation as
$$
\bm P(z)=
\begin{cases}
\bm P^{(a)}(z),& z\in D_\delta(-a)\setminus \Gamma_{\bm S}, \\
\bm P^{(0)}(z), & z\in U_0\setminus \Gamma_{\bm S}, \\
\bm G(z), & \text{elsewhere on } \mathbb{C}\setminus (\Gamma_{\bm S}\cup \partial U).
\end{cases}
$$
The final transformation $\bm S\mapsto \bm R$ is then
$$
\bm R(z)\deff
\bm S(z)\bm P(z)^{-1},\quad z\in \mathbb{C}\setminus (\Gamma_{\bm S}\cup \partial U).
$$
With this transformation, the jumps that $\bm S$ has in common with the parametrices $\bm G$ and $\bm P$ get canceled, and we remain with jumps only away from $[-a,0]$ and $U$. With
$$
\Gamma_{\bm R}\deff \partial U\cup \Gamma_{\bm S}\setminus \left([-a,0]\cup U\right),
$$
the matrix $\bm R$ satisfies the following RHP.
\begin{enumerate}[\bf R-1.]
\item The matrix $\bm R:\mathbb{C}\setminus \Gamma_{\bm R}\to \mathbb{C}^{2\times 2}$ is analytic.
\item For $z\in \Gamma_{\bm R}$, it satisfies the jump $\bm R_+(z)=\bm R_-(z)\bm J_{\bm R}(z)$,
with
\begin{equation}\label{eq:jumpJR}
\bm J_{\bm R}(z)\deff
\begin{dcases}
\bm G(z)\bm J_{\bm S}(z)\bm G(z)^{-1}, & z\in \Gamma_{\bm R}\setminus \partial U, \\
\bm P^{(0)}(z)\bm G(z)^{-1}, & z\in \partial U_0, \\
\bm P^{(a)}(z)\bm G(z)^{-1}, & z\in \partial D_\delta(-a),
\end{dcases}
\end{equation}
\item With $\bm R_1\deff \bm S_1-\bm G_1$,
\begin{equation}\label{eq:expansionR1infinity}
\bm R(z)=\bm I+\frac{1}{z}\bm R_1+\mathcal{O}(1/z^2),\quad z\to\infty.
\end{equation}
\end{enumerate}
To conclude that $\bm R$ is asymptotically close to the identity, we control its jumps. We use the matrix norm notation introduced in \eqref{deff:matrixnorm}--\eqref{deff:matrixLpqnorm}.
\begin{prop}\label{prop:estimateJR}
Fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. There is $\eta>0$ for which the jump matrix for $\bm R$ satisfies the estimates
$$
\|\bm J_{\bm R}-\bm I\|_{L^1\cap L^2\cap L^\infty(\Gamma_{\bm R}\setminus \partial U)}=\mathcal{O}(\ee^{-\eta n})
$$
and
$$
\|\bm J_{\bm R}-\bm I\|_{L^1\cap L^2\cap L^\infty(\partial U_0)}=\mathcal{O}(n^{-1/3}),\quad
\|\bm J_{\bm R}-\bm I\|_{L^1\cap L^2\cap L^\infty(\partial D_\delta(-a))}=\mathcal{O}(n^{-1})
$$
as $n\to\infty$, uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{prop}
\begin{proof}
The function $\bm G$ remains uniformly bounded away from the interval $[-a,0]$, and the first claimed decay then follows from \eqref{eq:jumpJR} and Proposition~\ref{prop:decayJS}.
Using \eqref{eq:decayparametrixa} and \eqref{eq:jumpJR}, we obtain the uniform estimate of the jump along $\partial D_\delta(-a)$ directly. Since $\partial D_\delta(-a)$ is bounded, the $L^1$ and $L^2$ estimates also follow. Finally, estimate for the jump along $\partial U_0$ follows similarly, once we recall \eqref{eq:errorParOriginfinalform}.
\end{proof}
With Proposition~\ref{prop:estimateJR} at hand, the small norm theory of Riemann-Hilbert problems yields
\begin{theorem}\label{thm:Rasymptops}
Fix $\msf s_0>0$ and $\msf t_0>0$. The matrix $\bm R$ satisfies
$$
\bm R(z)= \bm I+\mathcal{O}(n^{-1/3}),\quad \bm R'(z)=\mathcal{O}(n^{-1/3}) \quad \text{and}\quad \bm R(w)^{-1}\bm R(z)=\bm I+\mathcal{O}\left(\frac{z-w}{n^{1/3}}\right),\quad n\to\infty,
$$
and the residue matrix $\bm R_1$ satisfies
\begin{equation}\label{eq:intreprresidueR1}
\bm R_1=-\frac{1}{2\pi \mathrm{i}}\int_{\partial U_0}\bm R_-(s)(\bm J_{\bm R}(s)-\bm I)\mathrm{d} s+\mathcal{O}(n^{-1}),
\end{equation}
where the error terms are all uniform for $z,w$ on the same connected components of $\mathbb{C}\setminus \Gamma_{\bm R}$, and they are also uniform for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{theorem}
\begin{proof}
The arguments are standard, so we only sketch them. The small norm theory for RHPs ensures the representation
\begin{equation}\label{eq:spectralformR}
\bm R(z)=\bm I+\frac{1}{2\pi \mathrm{i}}\int_{\Gamma_{\bm R}}\frac{\bm R_-(s)(\bm J_{\bm R}(s)-\bm I)}{s-z}\mathrm{d} s,\quad z\in \mathbb{C}\setminus \Gamma_{\bm R}.
\end{equation}
This ensures the estimates for $\bm R$ and $\bm R'$ away from $\Gamma_{\bm R}$ and, because the jump matrix is analytic on a neighborhood of $\Gamma_{\bm R}$, we are able to extend these estimates also to $\Gamma_{\bm R}$. To obtain the estimate involving $z$ and $w$, we then write, with the help of Cauchy's formula,
\begin{equation}\label{eq:EstRwz}
\bm R(w)^{-1}\bm R(z)=\bm I+\bm R(w)^{-1}\left(\bm R(w)-\bm R(z)\right)=\bm I+\bm R(w)^{-1}\frac{w-z}{2\pi \mathrm{i}}\oint \frac{\bm R(s)}{(s-w)(s-z)}\mathrm{d} s,
\end{equation}
where the integral is over a contour on $\bm C\setminus \Gamma_{\bm R}$ encircling both $w$ and $z$. The right-hand side is now $\bm I+\mathcal{O}((w-z)n^{-1/3})$ by the estimate on $\bm R$ already proven.
Finally, the estimate for $\bm R_1$ follows expanding \eqref{eq:spectralformR} as $z\to\infty$, and then using that the jump matrix decays to the identity at least as $\mathcal{O}(n^{-1})$ away from $\partial U_0$.
\end{proof}
Later on, we also need the first term in the asymptotic expansion for $\bm R$, we state it as a separate result.
\begin{prop}
Fix $\msf s_0>0$ and $\msf t_0>0$. Setting
\begin{equation}\label{deff:whR1}
\widehat{\bm R}_1(z)\deff \frac{1}{2\pi \mathrm{i}} \int_{\partial U_0}\left(\frac{(\widehat{\bm \Phi}_n^{(1)})_{21}}{\psi(s)^{1/2}\mathsf m(s)^{1/2}}\bm E_{21}-\mathsf q^{(1)}(s)\sp_3\right)\frac{\mathrm{d} s}{s-z}+\mathsf q_0^{(1)}\sp_2,\quad z\in U_0,
\end{equation}
the matrix $\bm R$ has the expansion
\begin{equation}\label{eq:ExpRwhR1}
\bm R(z)= \bm I+\frac{1}{n^{1/3}}\bm U_0 \widehat{\bm R}_1(z)\bm U_0^{-1} + \mathcal{O}(n^{-2/3}),\quad n\to \infty
\end{equation}
where the error term is uniform for $z\in U_0$, and it is also uniform for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{prop}
\begin{proof}
Starting from the representation \eqref{eq:spectralformR}, we write
$$
\bm R(z)-\bm I=\frac{1}{2\pi \mathrm{i}}\int_{\partial U_0}\frac{\bm R_-(s)(\bm J_{\bm R}(s)-\bm I)}{s-z}\mathrm{d} s+\mathcal{O}(n^{-1})=\frac{1}{2\pi \mathrm{i}}\int_{\partial U_0}\frac{\bm J_{\bm R}(s)-\bm I}{s-z}\mathrm{d} s+\mathcal{O}(n^{-2/3}),
$$
and the result follows combining \eqref{eq:jumpJR} and \eqref{eq:errorParOriginfinalform} and performing straightforward calculations.
\end{proof}
All the transformations $\bm Y\mapsto \bm T\mapsto \bm S\mapsto \bm R$ involve only analytic factors in their construction. As such, we can actually slightly deform the contour $\Gamma_{\bm R}$ in all these steps, so that in fact the estimates in the previous result are valid everywhere on $\mathbb{C}$, interpreting them as with boundary values when $z,w\in \Gamma_{\bm R}$. We will use this fact without further warning.
With this theorem at hand, in the next sections we recover the needed asymptotic formulas for the proof of our main results.
\section{Proof of main results}\label{sec:ProofMainResults}
The next step in the direction of concluding our asymptotic analysis is to unravel the transformations $\bm R\mapsto \bm S\mapsto \bm T\mapsto \bm Y$. Introduce
$$
\bm\Lambda_n(x)\deff \bm I+\frac{\ee^{2n\phi_+(x)}}{\sigma_n(x)}\chi_{(-a,0)}(x)\bm E_{21}\quad \text{and}\quad \bm \Delta_n(x)\deff \left(\bm I+\frac{\chi_{(-a,0)}(x)}{\sigma_n(x)}\bm E_{21}\right),\quad x\in \mathbb{R},
$$
which are related by
$$
\bm \Lambda_n(x)\ee^{-n\phi_+(x)\sp_3}=\ee^{-n\phi_+(x)\sp_3}\bm \Delta_n(x).
$$
Then the result of the unfolding of the transformations is
\begin{equation}\label{eq:unfoldtransfYR}
\bm Y_+(z)=\ee^{n\ell_V\sp_3}\bm R_+(x)\bm P_+(x)\bm \Lambda_n(x)\ee^{-n(\phi_+(x)-V(x)/2)\sp_3},\quad x\in \mathbb{R}.
\end{equation}
We split the proofs of our results in the next few sections.
\subsection{Proof of Theorem~\ref{thm:limitingkernel}}\hfill
Thanks to \eqref{eq:unfoldtransfYR} and Theorem~\ref{thm:Rasymptops}, the expression \eqref{eq:KnYxy} reduces to the asymptotic formula
\begin{equation}\label{eq:msfKnunfold}
\ee^{-n(V(x)+V(y))/2}\mathsf K_n^Q(x,y)=\frac{\ee^{-n(\phi_+(x)+\phi_+(y))}}{2\pi\mathrm{i}(x-y)}\bm e^\mathrm T_2\bm\Lambda_{n}(y)^{-1}\bm P_+(y)^{-1}\left(\bm I+\mathcal{O}\left(\frac{x-y}{n^{1/3}}\right)\right)\bm P_+(x)\bm\Lambda_{n}(x)\bm e_1,
\end{equation}
valid as $n\to\infty$, $x,y\in \mathbb{R}$ and uniformly for $\msf s,\msf t$ as in Theorem~\ref{thm:Rasymptops}.
For the value $\mathsf c_V>0$ introduced in \eqref{eq:derivativepsicpsi}, we scale $x=u_n$ and $y=v_n$ as in \eqref{deff:scalingunvn},
where $u,v$ are on any given compact of the real axis. For such values of $u$, for $n$ large enough the points $u_n$ are always on the neighborhood $U_0$ of the origin where $\bm P=\bm P^{(0)}$, and from \eqref{deff:ParamOrigin} we simplify
$$
\bm e_2^\mathrm T \bm \Lambda_{n}^{-1}\bm P_+^{-1}=\ee^{n\phi_+}\bm e_2^\mathrm T \bm \Delta_n^{-1}\bm \Phi_{n,+}^{-1}\bm F_n^{-1}\quad \text{and}\quad
\bm P_+\bm \Lambda_{n}\bm e_1=\bm F_n\bm \Phi_{n,+}\bm \Delta_n\bm e_1\ee^{n\phi_+}.
$$
where all the quantities above are evaluated at $u_n$. The next step is to plug these identities into \eqref{eq:msfKnunfold}, when doing so we also use the estimates
$$
\bm F_n(u_n)=\bm U_0\left(\bm I+\mathcal{O}(n^{-2/3})\right)(\mathsf c_V a)^{-\sp_3/4}n^{-\sp_3/6},\quad n\to\infty,
$$
which follows from \eqref{local_par_L_Phi}, \eqref{eq:def_U0} and \eqref{eq:derivativepsicpsi}, and its immediate consequence
$$
\bm F_n(v_n)^{-1}\bm F_n(u_n)=n^{\sp_3/6}\left(\bm I+\mathcal{O}\left(\frac{u-v}{n^{2/3}}\right)\right)n^{-\sp_3/6}=\bm I+\mathcal{O}\left(\frac{u-v}{n^{1/3}}\right),\quad n\to\infty,
$$
which is obtained with arguments similar to \eqref{eq:EstRwz}. These estimates are uniform for $u,v$ in compacts of $\mathbb{R}$, and are also uniform for $\msf s\geq -\msf s_0$ and $\msf t\in [\msf t_0,1/\msf t_0]$. Equation~\eqref{eq:msfKnunfold} then simplifies to
\begin{multline*}
\ee^{-n(V(u_n)+V(v_n))/2}\mathsf K_n^Q(u_n,v_n)= \\
\frac{\mathsf c_V n^{2/3}}{2\pi \mathrm{i}(u-v)}
\bm e_2^\mathrm T
\bm \Delta_n(v_n)^{-1}
\bm \Phi_{n,+}(v_n)^{-1}
\left(\bm I+\mathcal{O}\left(\frac{u-v}{n^{1/3}}\right)\right)\bm \Phi_{n,+}(u_n)\bm \Delta_n(u_n)\bm e_1.
\end{multline*}
Now, with the definition of $\bm \Phi_n$ in \eqref{local_par_L_Phi} at hand, the aid of \eqref{eq:asympPhinPhi0} and the constant $\mathsf c_{\mathsf H}$ appearing in \eqref{eq:expansionmsfhnorigin}, we get
$$
\bm\Phi_{n,+}(u_n)=\widehat{\bm \Phi}_{n,+}(n^{2/3}\psi(u_n))=\left(\bm I+\mathcal{O}(n^{-\nu})\right){\bm \Phi}_{0,+}(u\mid \msf s,\mathsf c_{\mathsf H}),\quad n\to\infty,
$$
for any $\nu\in (0,2/3)$, with the error being valid uniformly for $u$ in compacts of $\mathbb{R}$, and also uniformly for $\msf s\geq -\msf s_0, \msf t_0\leq \msf t\leq 1/\msf t_0$. Also, thanks to Theorem~\ref{prop:existencePhicc} we know that ${\bm \Phi}_{0,+}(u\mid \msf s,\mathsf c_{\mathsf H})^{\pm 1}$ remain bounded for $u$ in compacts and $\msf s\geq -\msf s_0, \msf t_0\leq \msf t\leq 1/\msf t_0$. Therefore,
$$
\bm \Phi_{n,+}(v_n)^{-1}\left(\bm I+\mathcal{O}\left(\frac{u-v}{n^{1/3}}\right)\right)\bm \Phi_{n,+}(u_n)
=
\left(\bm I+\mathcal{O}(n^{-\nu})\right){\bm \Phi}_{0,+}(v\mid \msf s,\mathsf c_{\mathsf H})^{-1}{\bm \Phi}_{0,+}(u\mid \msf s,\mathsf c_{\mathsf H})+\mathcal{O}\left(\frac{u-v}{n^{1/3}}\right)
$$
as $n\to\infty$.
In addition, the estimate
\begin{equation}\label{eq:estimatesigmanfinalproofkernel}
\frac{1}{\sigma_n(u_n)}=1+\ee^{-\msf s+\msf t u/\mathsf c_V}(1+\mathcal{O}(n^{-2/3}))=1+\ee^{-\msf s+\mathsf c_{\mathsf H} u}(1+\mathcal{O}(n^{-2/3})),\quad n\to\infty,
\end{equation}
is valid uniformly for $u$ in compacts, uniformly for $\msf s\in \mathbb{R}$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$, and is immediate from \eqref{def:perturbedweight}.
Combining everything, and denoting $\chi_+=\chi_{[0,\infty)}$, we obtained the asymptotic formula
\begin{multline*}
\frac{\ee^{-n(V(u_n)+V(v_n))/2}}{\mathsf c_V n^{2/3}}\mathsf K_n^Q(u_n,v_n)=
\frac{1+\mathcal{O}(n^{-\nu})}{2\pi \mathrm{i}(u-v)}
\bm e_2^\mathrm T
\left(\bm I-(1+\ee^{-\msf s+\mathsf c_{\mathsf H} v})\chi_{+}(v)\bm E_{21}\right)\\
\times
{\bm \Phi}_{0,+}(v\mid \msf s,\mathsf c_{\mathsf H})^{-1}{\bm \Phi}_{0,+}(u\mid \msf s,\mathsf c_{\mathsf H})\left(\bm I+(1+\ee^{-\msf s+\mathsf c_{\mathsf H} u})\chi_{+}(u)\bm E_{21}\right)\bm e_1+\mathcal{O}(n^{-1/3}),\quad n\to\infty,
\end{multline*}
which is valid uniformly for $u,v$ in compacts of $\mathbb{R}$, and also uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$. With ${\bm\Delta}_0$ as in \eqref{deff:Dcc}, this identity rewrites as
\begin{multline*}
\frac{\ee^{-n(V(u_n)+V(v_n))/2}}{\mathsf c_V n^{2/3}}\mathsf K_n^Q(u_n,v_n)=
\frac{1+\mathcal{O}(n^{-\nu})}{2\pi \mathrm{i}(u-v)}\\
\times \left[\left({\bm \Phi}_0(v\mid \msf s,\mathsf c_{\mathsf H}){\bm\Delta}_0(v\mid \msf s,\mathsf c_{\mathsf H})\right)^{-1}{\bm \Phi}_0(u\mid \msf s,\mathsf c_{\mathsf H}){\bm\Delta}_0(u\mid \msf s,\mathsf c_{\mathsf H})\right]_{21,+}
+\mathcal{O}(n^{-1/3}),\quad n\to\infty,
\end{multline*}
and the proof of Theorem~\ref{thm:limitingkernel} is now completed using Equation~\eqref{eq:transfPhicckernelphicc}.
\subsection{Proof of Theorem~\ref{thm:asymptoticsnormingconstant}}\hfill
Now that the RHP asymptotic analysis is completed, the proof of Theorem~\ref{thm:asymptoticsnormingconstant} follows standard steps. In our case, there is an additional cancellation that has to be accounted for at a later step, in virtue of the presence of the factor $\mathsf q$ in both the global and local parametrices, see \eqref{eq:constructionGparam} and \eqref{eq:errorParOriginfinalform}. So we opt for presenting the detailed calculation.
Starting from the representation \eqref{eq:relationnormingY1}, we unravel the transformations and obtain
\begin{equation}\label{eq:gammanasympt1}
-2\pi \mathrm{i} \left(\upgamma_{n-1}^{(n,Q)}\right)^2=\ee^{-2n\ell_V}\left(\bm R_1+\bm G_1\right)_{21}=\ee^{-2n\ell_V}\left((\bm R_1)_{21}- \frac{\mathrm{i} a}{4} -\frac{1}{n^{1/3}}\frac{\msf t^{1/2}a^{1/2}}{4\pi \mathrm{i}}F_{-1/2}(\msf s) +\mathcal{O}(n^{-2/3}) \right),
\end{equation}
with $\bm R_1$ as in \eqref{eq:expansionR1infinity}, and where for the second identity we used the definition of $\bm G_1$ in \eqref{eq:definitionG1} and the estimate for $\mathsf q_0$ from Lemma~\ref{lem:estimate_q_q0}.
It remains to estimate $\bm R_1$, which we do so starting from \eqref{eq:intreprresidueR1}. Using Cauchy-Schwartz, Propositions~\ref{prop:estimateJR} and Theorem~\ref{thm:Rasymptops}, we write
$$
\bm R_1=-\frac{1}{2\pi \mathrm{i}}\int_{\partial U_0}(\bm J_{\bm R}(s)-\bm I)\mathrm{d} s+\mathcal{O}(n^{-2/3}),\quad n\to\infty.
$$
The matrix $\bm J_\bm R$ is in \eqref{eq:jumpJR}, and combining its explicit expression along $\partial U_0$ with \eqref{eq:errorParOriginfinalform} and \eqref{eq:def_globalparM}, after a cumbersome but straightforward calculation we arrive at
$$
\left(\bm R_1\right)_{21}=-\frac{1}{2\pi \mathrm{i} n^{1/3}}\int_{\partial U_0}\left(\frac{\mathrm{i} \mathsf q^{(1)}(s)(\mathsf m(s)-1)}{2\mathsf m(s)^{1/2}}+(\bm L_1(s))_{21}\right)\mathrm{d} s+\mathcal{O}(n^{-2/3}),\quad n\to\infty,
$$
where we recall that $\mathsf m$, $\mathsf q^{(1)}$ and $\bm L_1$ are given in \eqref{eq:def_U0}, \eqref{deff:msfqo1} and \eqref{eq:AsymptLlocalpar}, respectively. Using this explicit expression for $\bm L_1$, we see that
$$
(\bm L_1(s))_{21}=\frac{\left({\bm \Phi}^{(1)}_n\right)_{21}}{2\mathsf m(s)^{1/2}\psi(s)^{1/2}}.
$$
Both functions $\mathsf m$ and $\psi$ are analytic in a neighborhood of the origin, and vanish linearly therein. Combining in addition with \eqref{eq:derivativepsicpsi}, we see that the product $\mathsf m(z)^{1/2}\psi(z)^{1/2}$ admits an analytic continuation near the origin, with the expansion
$$
\mathsf m(s)^{1/2}\psi(s)^{1/2}=\frac{\mathsf c_V}{a}s(1+\mathcal{O}(s)),\quad s\to 0.
$$
Thus, computing residues
$$
\int_{\partial U_0}(\bm L_1(s))_{21}\mathrm{d} s=-\pi \mathrm{i} \frac{a }{\mathsf c_V}\left({\bm \Phi}^{(1)}_n\right)_{21}.
$$
Similarly, using now \eqref{deff:msfqo1} we obtain
$$
\int_{\partial_{U_0}}\frac{\mathrm{i} \mathsf q^{(1)}(s)(\mathsf m(s)-1)}{2\mathsf m(s)^{1/2}}\mathrm{d} s=-\mathsf t^{1/2}a^{1/2} F_{-1/2}(\msf s)
$$
Hence,
$$
\left(\bm R_1\right)_{21}=
\frac{1}{n^{1/3}}\left(\frac{a}{2\mathsf c_V}\left(\bm \Phi_n^{(1)}\right)_{21}+\frac{\msf t^{1/2}a^{1/2}}{4\pi\mathrm{i}} F_{-1/2}(\msf s) \right)+\mathcal{O}(n^{-2/3}),
$$
and \eqref{eq:gammanasympt1} updates to
$$
\left(\upgamma_{n-1}^{(n,Q)}(\msf s)\right)^2=\ee^{-2n\ell_V}\left(\frac{a}{8\pi}-\frac{1}{n^{1/3}}\frac{a}{4\pi \mathrm{i}}\left(\bm \Phi_n^{(1)}\right)_{21}+\mathcal{O}(n^{-2/3})\right),\quad n\to\infty,
$$
uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$. We now need only to apply \eqref{eq:asympPhin1Phi01residue} with any $\nu\geq 1/3$ to complete the proof.
\subsection{Proof of Theorem~\ref{thm:asymptoticsqLaplacetransf}}\hfill
Unlike the already proven major results, the proof of Theorem~\ref{thm:asymptoticsqLaplacetransf} is not a straightforward consequence of the steepest descent analysis concluded with Theorem~\ref{thm:Rasymptops}. It does rely substantially on Theorem~\ref{thm:Rasymptops}, but several other inputs are also needed along the way. Equipped with \eqref{eq:KnYxy}, the idea is to account for the different approximations for $\mathsf K_n^Q$ on the different components
$$
\mathbb{R}\setminus (U_0\cup D_\delta(-a)\cup (-a,0)),\quad \mathbb{R}\cup U_0, \quad (-a,-\delta,-a+\delta) \quad \text{and}\quad (-a+\delta,0)\setminus U_0
$$
that arise from the RHP, and integrate each such approximation. With Proposition~\ref{prop:intreprLn} in mind, we are thus able to recover asymptotics for $\mathsf L_n^Q$ itself. As one would expect, it turns out that in this process the terms that arise away from $U_0$ become all exponentially negligible, and only the contribution from $U_0$ survives in the leading contribution. The contribution that arrives this way involves $\bm \Phi_n$, and we further need to split it into different parts and still account for some exact cancellations to arrive at the leading asymptotic contribution. We postpone this analysis to the next section, where it is split into several different lemmas, and summarize the outcome with the next result. For its statement, we recall that $\mathsf h_n$ and $\widehat{\bm \Phi}_n$ were introduced in \eqref{local_par_L_Phi}, and we also set
\begin{equation}\label{deff:Deltan0}
\widehat{\bm \Delta}_n(x)\deff \bm I+ (1+\ee^{-\mathsf h_n(u)})\chi_{(0,+\infty)(x)}\bm E_{21},\quad x\in \mathbb{R}.
\end{equation}
\begin{prop}\label{prop:finalestKnQ}
Fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. There exists a function $\mathsf R(u)=\mathsf R_1(u\mid \msf t)$ satisfying
$$
\int_{\msf s}^\infty \mathsf R(u)\mathrm{d} u=\mathcal{O}(n^{-1/3}),\quad n\to\infty,
$$
uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$, and for which the identity
\begin{multline*}
\int_{-\infty}^{\infty}\mathsf K_n^Q(x,x\mid u)\frac{\omega_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x\\ =\frac{1}{2\pi \mathrm{i}}
\int_{n^{2/3}\zeta_0}^{n^{2/3}\zeta_1}\frac{\ee^{\mathsf h_n(u)}}{\left(1+\ee^{\mathsf h_n(u)}\right)^2}\left[ \widehat{\bm \Delta}_n(x)^{-1}\widehat{\bm \Phi}_{n,+}(x)^{-1}\left(\widehat{\bm \Phi}_{n,+} \widehat{\bm\Delta}_n\right)'(x)\right]_{21}\mathrm{d} x+\mathsf R(u),
\end{multline*}
holds true for every $u\geq -\msf s_0$ and every $\msf t\in [\msf t_0,1/\msf t_0]$.
\end{prop}
In words, Proposition~\ref{prop:finalestKnQ} is saying that the major contribution to $\mathsf L_n^Q$ comes from the neighborhood $U_0$, so from the local parametrix $\widehat{\bm\Phi}_n$. But according to the developments from the previous sections, this local parametrix is close to the id-PII parametrix ${\bm \Phi}_0$, and we are ready to conclude our last major result.
\begin{proof}[Proof of Theorem~\ref{thm:asymptoticsqLaplacetransf}]
We combine Propositions~\ref{prop:intreprLn} and \eqref{prop:finalestKnQ} to obtain
$$
\log\mathsf L_n^Q(\msf s)=-\frac{1}{2\pi \mathrm{i}}\int_{\msf s}^\infty
\int_{n^{2/3}\zeta_0}^{n^{2/3}\zeta_1}\frac{\ee^{\mathsf h_n(u)}}{\left(1+\ee^{\mathsf h_n(u)}\right)^2}\left[ \widehat{\bm \Delta}_n(x)^{-1}\widehat{\bm \Phi}_{n,+}(x)^{-1}\left(\widehat{\bm \Phi}_{n,+} \widehat{\bm\Delta}_n\right)'(x)\right]_{21}\mathrm{d} x\mathrm{d} u+\mathcal{O}(n^{-1/3}),
$$
valid as $n\to\infty$ and uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$. Next, with the identifications \eqref{local_par_L_Phi}--\eqref{eq:expansionmsfhnorigin} in mind, we estimate the integral on the right-hand side with the help of Theorem~\ref{thm:integralPhitauPhicc} and conclude the proof.
\end{proof}
\section{Technical Lemmas}\label{sec:technicalestimates}
It remains to prove Proposition~\ref{prop:finalestKnQ}, analysis which we split into several technical lemmas in this section.
Our starting point is the integral representation for $\mathsf L_n^Q$ from Proposition~\ref{prop:intreprLn} with the asymptotic information for $\mathsf K_n^Q$ provided by the RHP analysis. For $x\in\mathbb{R}$, set
\begin{equation}\label{deff:msfA}
\mathsf A(x)\deff \left[ \bm \Lambda_{n}(x)^{-1}\bm P_+(x)^{-1}\bm R_+(x)^{-1}\bm R_+'(x)\bm P_+(x)\bm\Lambda_n(x)+\bm \Lambda_{n}(x)^{-1}\bm P_+(x)^{-1}\bm P_+'(x)\bm \Lambda_{n}(x) \right]_{21}.
\end{equation}
Recalling \eqref{eq:KnYxx}, the unwrap of the transformations of the RHP yields the identity
$$
\mathsf K_n^Q(x,x\mid u)\frac{\omega_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}=\frac{\sigma_n(x\mid u)\ee^{-2n\phi_+(x)}}{2\pi \mathrm{i} (1+\ee^{u+n^{2/3}Q(x)})}\left(\mathsf A(x)+\left(\frac{\ee^{2n\phi_+(x)}}{\sigma_n(x)}\right)'\chi_{\mathcal G^+}(x)\right)
$$
We now need to integrate each of the terms on the right-hand side above, first in $x\in \mathbb{R}$ and then in $u\in (\msf s,+\infty)$. Each term will be analyzed individually, also depending on whether we integrate $x$ in the bulk, each of the edges or away from the support $[-a,0]$ of the equilibrium measure. To simplify notation, it is convenient to introduce the additional notation for each relevant integral, and for an arbitrary set $J\subset \mathbb{R}$ denote
\begin{equation}\label{deff:I1int}
\mathsf I_1(J)\deff \int_J \frac{\sigma_n(x\mid u)\ee^{-2n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathsf A(x) \mathrm{d} x
\end{equation}
and
\begin{equation}\label{deff:I2int}
\begin{aligned}
\mathsf I_2(J) \deff & \int_J \frac{\sigma_n(x\mid u)\ee^{-2n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\left(\frac{\ee^{2n\phi_+(x)}}{\sigma_n(x)}\right)'\mathrm{d} x \\
=& \int_J \left(2n\phi_{n,+}(x)-\frac{n^{2/3}Q'(x)}{1+\ee^{u+n^{2/3}Q(x)}}\right)\frac{1}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x,
\end{aligned}
\end{equation}
which are functions of $u\in \mathbb{R}$ as well. With $\varepsilon_0,\varepsilon_1>0$ being determined by $(-\varepsilon_0,\varepsilon_1)=U_0\cap \mathbb{R}$, the split
\begin{multline}\label{eq:intKnIsplit}
2\pi \mathrm{i} \int_{-\infty}^{\infty}\mathsf K_n^Q(x,x\mid u)\frac{\omega_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x=
\mathsf I_1(-\infty,-a-\delta)+\mathsf I_1(-a-\delta,a+\delta)
\\ +\mathsf I_1(a+\delta,-\varepsilon_0)+\mathsf I_1(-\varepsilon_0,\varepsilon_1)+\mathsf I_1(\varepsilon_1,+\infty)
+\mathsf I_2(-a,-\varepsilon_0)+\mathsf I_2(-\varepsilon_0,0)
\end{multline}
is immediate, and with the next series of lemmas we estimate each of the terms on the right-hand side.
\begin{lemma}\label{lem:I1inftyI1aI2a}
Fix $\msf s\in \mathbb{R}$ and $\msf t_0\in (0,1)$. There exists $\eta>0$ for which the estimate
$$
\mathsf I_1(-\infty,-a-\delta)+\mathsf I_1(-a+\delta,-\varepsilon_0)+\mathsf I_2(-a,-\varepsilon_0)=\mathcal{O}\left(\ee^{-u}\ee^{-n^{2/3}\eta}\right)
$$
holds true uniformly for $u\geq \msf s$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{lemma}
\begin{proof}
On the intervals $(-\infty,-a-\delta)$ and $(\varepsilon_1,\infty)$ the function $\bm \Lambda_{n}$ is identically the identity matrix. On the interval $(a+\delta,-\varepsilon_0)$ the nontrivial entry of $\bm \Lambda_{n}$ is $\ee^{2n\phi_+}/\sigma_n$ and, because $Q<0$ and $\phi_{+}$ is purely imaginary in this interval, this quotient is bounded. Also, away from the endpoints $-a$ and $0$ we have $\bm P\equiv \bm G$. Both $\bm P$ and $\bm G$, and their $x$-derivatives, decay as $x\to \pm \infty$ and remain bounded as $n\to\infty$, all uniformly in $u$ and $\msf t$ as claimed. All of these facts combined together, we obtain that for some constants $K>0$ and $n_0\geq 1$
\begin{equation}\label{eq:boundmsfA}
|\mathsf A(x)|\leq K,\quad \text{for all}\; x\in \mathbb{R}\setminus \left((-a-\delta,-a+\delta)\cup (-\varepsilon_0,\varepsilon_1)\right),\; n\geq n_0,\; \msf t_0\leq \msf t\leq 1/\msf t_0, \; u\geq \msf s.
\end{equation}
Next, we now use that $Q\geq 0$ on the interval $(-\infty,0)$ to bound
\begin{equation}\label{eq:I1estimate1}
0\leq \frac{\sigma_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\leq \frac{1}{1+\ee^{u+n^{2/3}Q(x)}}\leq \ee^{-u},\quad x\leq 0,
\end{equation}
which is valid for any $n\geq 1, u\in \mathbb{R},\msf t >0$. Thus, combining everything we obtain
$$
|\mathsf I_1(-\infty,-a-\delta)|\leq K\ee^{-u}\int_{-\infty}^{-a-\delta}\ee^{-2n\phi_+(x)}\mathrm{d} x,
$$
and using Proposition~\ref{prop:functionphitau}--(iv),(v), this proves the bound for $\mathsf I_1(-\infty,-a-\delta)$.
For the second integral, we term $\phi_+$ is oscillatory, so to obtain the exponential decay in the $x$-integral we have to argue differently and as follows. The function
$$
v\mapsto \frac{1}{1+\ee^{-v}}\frac{1}{1+\ee^v}=\frac{\ee^{-v}}{(1+\ee^{-v})^2}
$$
is strictly increasing on $(-\infty,0)$ and strictly decreasing on $(0,+\infty)$. Because $Q>0$ on $(-\infty,0)$ and $Q(0)=0$, by reducing $U_0$ if necessary we can assume without loss of generality that $Q(x)\geq Q(-\varepsilon_0)$ for every $x\in [-a,-\varepsilon_0]$. This way, $v=u+n^{2/3}Q(x)\geq v_0\deff u+n^{2/3}Q(-\varepsilon_0)$ and, because $u$ is assumed to be bounded from below, we can make sure that $v_0>0$ for every $u$. Therefore
$$
0\leq \frac{\sigma_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\leq \frac{\ee^{-v_0}}{(1+\ee^{-v_0})^2}\leq \ee^{-u-n^{2/3}Q(-\varepsilon_0)},
$$
and upon integration and using Proposition~\ref{prop:functionphitau}--(ii), we obtain
$$
|\mathsf I_2(-a+\delta,-\varepsilon_0)|\leq (a-\delta+\varepsilon_0)K\ee^{-u-n^{2/3}Q(-\varepsilon_0)}.
$$
for every $n\geq n_0$ and $\msf t,u$ as claimed.
For the estimate for $\mathsf I_2(-a,-\varepsilon_0)$ we use again the last inequality in \eqref{eq:I1estimate1} and also that both $\phi'_{+}$ and $Q'$ are continuous, and hence bounded, on $(-a,-\varepsilon_0)$. This concludes the proof.
\end{proof}
For the integral over $(\varepsilon_1,+\infty)$, it is easier to actually estimate its $u$-integral directly.
\begin{lemma}\label{lem:estI1epsilon1infty}
Fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. There exists $\eta>0$ for which the estimate
$$
\int_{\msf s}^\infty\mathsf I_1(\varepsilon_1,+\infty)\mathrm{d} u=\mathcal{O}\left(\ee^{-n\eta}\right)
$$
holds true uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{lemma}
\begin{proof}
With the bound \eqref{eq:boundmsfA} we see that it is enough to estimate the integral
$$
\int_{\msf s}^\infty \int_{\varepsilon_1}^\infty \frac{\sigma_n(x\mid u)\ee^{-2n\phi(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x\mathrm{d} u=\int_{\varepsilon_1}^\infty\frac{\ee^{-2n\phi_+(x)}}{1+\ee^{\msf s+n^{2/3}Q(x)}}\mathrm{d} x,
$$
where for the equality we used Tonelli's Theorem to interchange the order of integration, and then integrated exactly. The term $(1+\ee^{\msf s+n^{2/3}Q(x)})^{-1}$ is bounded by $1$, and using Proposition~\ref{prop:functionphitau}--(iv),(v) we see that the integral of $\ee^{-2n\phi}$ is $\mathcal{O}(\ee^{-\eta n})$ for some $\eta>0$ independent of $\msf s$ and $\msf t$, which concludes the proof.
\end{proof}
Next, we analyze the contribution coming from a neighborhood of the endpoint $z=a$.
\begin{lemma}\label{lem:I1adelta}
Fix $\msf s\in \mathbb{R}$, $\msf t_0\in (0,1)$. There exists $\eta>0$ such that the estimate
$$
\mathsf I_1(-a-\delta,-a+\delta) =\mathcal{O}(\ee^{-u-\eta n^{2/3}}),\quad n\to\infty,
$$
is valid uniformly for $u\geq \msf s$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{lemma}
\begin{proof}
On the neighborhood $D_\delta(-a)$ the function $\bm P=\bm P^{(a)}$ is the Airy local parametrix \eqref{deff:LocalParRegEdge}, which involves the function $\bm\Phi_{\ai}$ evaluated at the argument $\zeta=n^{2/3}\varphi(z)$. In the $\zeta$-plane, we fix $R>0$ for which the asymptotic expansion \eqref{eq:AsympAiryPar} is valid for $|\zeta|>R$, and split the analysis into two cases, namely for $|n^{2/3}\varphi(z)|\leq R$ and $|n^{2/3}\varphi(z)|\geq R$, and for
$$
A_n\deff \{x\in (-a-\delta,-a+\delta)\mid |n^{2/3}\varphi(x)|\leq R \},\quad B_n\deff (-a-\delta,-a+\delta)\setminus A_n
$$
we write
$$
\mathsf I_1(-a-\delta,-a+\delta)=\mathsf I_1(A_n)+\mathsf I_1(B_n).
$$
and now analyze each integral on the right-hand side separately.
For $|n^{2/3}\varphi(z)|\leq R$, the terms $\bm \Phi_{\ai,+}(\zeta=n^{2/3}\varphi(z))$ and $\bm \Phi'_{\ai,+}(\zeta=n^{2/3}\varphi(z))$ consist of continuous functions evaluated inside the compact interval $[-R,R]$, and therefore they are bounded. By the same reason, the expression \eqref{eq:FFactorRegPar} shows that $\bm F$ is bounded for $|n^{2/3}\varphi(z)|\leq R$. On the other hand, without further analysis we obtain that $\bm F'$ may grow at most as $\mathcal{O}(n^{1/6})$. Finally, combining with \eqref{eq:EstLogSigman} and the fact that the determinant of $\bm P$ is identically $1$, we conclude that for $|n^{2/3}\varphi(x)|\leq R$,
\begin{equation}\label{eq:EstShortRangePplus1}
\bm P_+(x)=\mathcal{O}(1)\ee^{n\phi_+(x)\sp_3},\quad \bm P_+(x)^{-1}=\ee^{-n\phi_+(x)\sp_3}\mathcal{O}(1)\quad \text{and}\quad \bm P_+'(x)=\mathcal{O}(n)\ee^{-n\phi_+(x)\sp_3}
\end{equation}
which is valid as $n\to\infty$ and uniformly for $x\in (-a-\delta,-a+\delta)$, also uniformly for $u\geq \msf s$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
We use \eqref{eq:EstShortRangePplus1} back into \eqref{deff:msfA}, and combined with the fact that $|\sigma_n^{-1}|\leq 2$ for $x\leq (-\infty,0)$ the result is that
$$
\mathsf A(x)=\left[ \bm\Lambda_{n}(x)^{-1}\ee^{-n\phi_+(x)\sp_3}\mathcal{O}(n)\ee^{n\phi_+(x)\sp_3} \bm\Lambda_{n}(x)\right]_{21}=\ee^{2n\phi_+(x)}\mathcal{O}(n),\quad n\to\infty,
$$
where the last error term is a scalar error, valid uniformly in $x,u,\msf t$ as before. Integrating, we obtain that for some absolute constant $K>0$,
\begin{equation}\label{eq:estimateI1An}
|\mathsf I_1(A_n)|\leq Kn \int_{A_n} \frac{\sigma_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x\leq Kn \int_{A_n}\ee^{-u-n^{2/3}Q(x)}\mathrm{d} x=\mathcal{O}(\ee^{-u-\eta n^{2/3}}),\quad n\to \infty,
\end{equation}
where for the second inequality we used that $0<\sigma_n\leq 1$ and for the last estimate we used that $Q$ is strictly positive on $(-a-\delta,-a+\delta)\supset A_n$.
Next, we consider the case $|n^{2/3}\varphi(z)|\geq R$. In such situation, the asymptotics \eqref{eq:AsympAiryPar} take place, and when combined with \eqref{eq:FFactorRegPar} they yield
\begin{equation}\label{eq:EstLongRangePplus1}
\bm P_+(x)=\bm G_+(x)\ee^{-\sp_3\log\sigma_n(x)/2}\left(\bm I+\mathcal{O}(n^{-1})\right)\ee^{\sp_3\log\sigma_n(x)/2}=\bm G_+(x)\left(\bm I+\mathcal{O}(n^{-1})\right).
\end{equation}
where for the last equality we used \eqref{eq:EstLogSigman}.
The matrix $\bm P_+$ is bounded as $x\to 0$, whereas $\bm G$ is not, but the cancellation that leads to this boundedness of $\bm P$ is not captured by this asymptotics. Nevertheless, as we need some uniform control over $x$, we now account for the behavior of $x\to 0$ in a rough manner as follows. We are assuming that $|n^{2/3}\varphi(x)|\geq R$, and because $\varphi$ is conformal with $\varphi(0)=0$ this means that $|x+a|\geq c/n^{2/3}$, for some fixed $c>0$. The function $\bm G$ has a fourth-root singularity at $x=0$, and therefore we arrive at the rough estimate
\begin{equation}\label{eq:EstLongRangePplus2}
\bm P_+(x)=\mathcal{O}(n^{1/6}),\quad n\to\infty.
\end{equation}
The estimate \eqref{eq:EstLongRangePplus1} can be differentiated term by term. With arguments similar to the ones we just applied, we arrive at the rough estimate
\begin{equation}\label{eq:EstLongRangePplus3}
\bm P'_+(x)=\mathcal{O}(n^{7/6}),\quad n\to\infty.
\end{equation}
These latter two estimates are valid uniformly for $u\geq \msf s$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$ as $n\to\infty$.
Finally, on the interval $(-a,0)$ the factor $\phi_+$ is purely imaginary, implying that $\bm\Lambda_{n}^{\pm 1}$ remains bounded therein. All combined, we obtained that
$$
\mathsf A(x)=\mathcal{O}(n),\quad n\to\infty,
$$
uniformly for $x\in B_n$ and $u,\msf t$ as claimed. Proceeding as in \eqref{eq:estimateI1An} we obtain a bound for $I_1(B_n)$ and complete the proof.
\end{proof}
It remains to analyze the two integrals $\mathsf I_1(-\varepsilon_0,\varepsilon_1)$ and $\mathsf I_2(-\varepsilon_0,0)$. The hardest analysis is the integral $\mathsf I_1$ which, as will turn out, contains both the leading contribution, a term to cancel $\mathsf I_2$ and asymptotically negligible terms. For ease of presentation, we explore the expression for $\mathsf A$ in \eqref{deff:msfA} as a sum and split
\begin{equation}\label{eq:splitI1J1J2}
\mathsf I_1(-\varepsilon_0,\varepsilon_1)=\mathsf J_1(-\varepsilon_0,\varepsilon_1)+\mathsf J_2(-\varepsilon_0,\varepsilon_1),
\end{equation}
where, for any measurable set $J\subset \mathbb{R}$,
\begin{equation}\label{deff:J1J2}
\begin{aligned}
\mathsf J_1(J) & \deff\int_J
\left[ \bm \Lambda_{n}(x)^{-1}\bm P_+(x)^{-1}\bm R(x)^{-1}\bm R'(x)\bm P_+(x)\bm \Lambda_{n}(x)\right]_{21}\frac{\sigma_n(x\mid u)\ee^{-2n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}} \mathrm{d} x,\\
\mathsf J_2(J) & \deff\int_J
\left[ \bm \Lambda_{n}(x)^{-1}\bm P_+(x)^{-1}\bm P_+'(x)\bm \Lambda_{n}(x) \right]_{21}\frac{\sigma_n(x\mid u)\ee^{-2n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x,
\end{aligned}
\end{equation}
and analyze each of these terms separately. For the estimation of $\mathsf J_1$, it is also easier to perform the $u$-integral, just as we did in Lemma~\ref{lem:estI1epsilon1infty}.
\begin{lemma}
Fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. The estimate
\begin{equation}
\int_{\msf s}^\infty\mathsf J_1(-\varepsilon_0,\varepsilon_1)\mathrm{d} u=\mathcal{O}(n^{-1/3})
\end{equation}
holds true uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
\end{lemma}
\begin{proof}
The idea is similar to the proof of Lemma~\ref{lem:I1adelta}. We fix a number $R>0$ for which the asymptotic expansion \eqref{eq:ModelRHPAsymp} for $\bm\Phi=\widehat{\bm \Phi}_n$ is valid for $|\zeta|\geq R$, uniformly in $\msf t,\msf s$ as claimed (see Remark~\ref{remark:UnifAsympPhitauad}). Then, introduce
\begin{equation}\label{deff:setsCnDn}
C_n\deff \{ x\in (-\varepsilon_0,\varepsilon_1)\mid |n^{2/3}\psi(x)|\leq R \},\quad D_n\deff (-\varepsilon_0,\varepsilon_1)\setminus C_n.
\end{equation}
We now find bounds for the integrands, with separate arguments for each of $C_n$ and $D_n$.
Recalling \eqref{deff:ParamOrigin}, on the interval $(-\varepsilon_0,\varepsilon_1)$ the parametrix is $\bm P=\bm P^{(0)}=\bm F_n\bm\Phi_n\ee^{n\phi\sp_3}$, with $\bm F_n$ and $\bm \Phi_n$ as in \eqref{local_par_L_Phi}. Using the definition of $\bm F_n$ in \eqref{local_par_L_Phi} and Theorem~\ref{thm:Rasymptops}, we express
\begin{equation}\label{eq:EstRRprimeF}
\bm F_n^{-1}\bm R^{-1}\bm R'\bm F_n=\bm F_n^{-1}(\bm R^{-1}-\bm I)\bm R'\bm F_n-\bm F_n^{-1}\bm R'\bm F_n=-\bm F_n^{-1}\bm R'\bm F_n+\mathcal{O}(n^{-1/3}),\quad n\to\infty,
\end{equation}
valid uniformly when evaluated at $x\in C_n$ and also uniformly for $\msf s,\msf t$ as claimed. Using again the definition of $\bm F_n$, the fact that $\psi/\mathsf m$ remains bounded near $z=0$ and \eqref{eq:ExpRwhR1},
$$
\bm F_n^{-1}\bm R'\bm F_n=\frac{1}{n^{1/3}}\psi^{\sp_3/4}\mathsf m^{-\sp_3/4}n^{\sp_3/6}\widehat{\bm R}_1' n^{-\sp_3/4}\mathsf m^{\sp_3/4}\psi^{-\sp_3/4}+\mathcal{O}(n^{-1/3}).
$$
A careful inspection on \eqref{deff:whR1} shows that $(\widehat{\bm R}_1(z))_{12}$ is independent of $z$, so $R'$ has zero $(1,2)$-entry. Therefore we conclude that
$$
\bm F_n^{-1}\bm R^{-1}\bm R'\bm F_n=\mathcal{O}(n^{-1/3}),
$$
stressing that this is valid for $x\in C_n$. Also, from \eqref{eq:asympPhinPhi0} we learn that $\bm\Phi_n(x)=\mathcal{O}(1)$ uniformly on $C_n$. Everything we have so far combined yields that
$$
\left[\bm \Lambda_{n}^{-1}\bm P_+^{-1}\bm R^{-1}\bm R'\bm P\bm \Lambda_{n}\right]_{21}=\left[\bm \Lambda_{n}^{-1} \ee^{-n\phi\sp_3}\mathcal{O}(n^{-1/3})\ee^{n\phi_3\sp_3} \bm \Lambda_{n}\right]_{21}=\ee^{-2n\phi}\left[\bm \Delta_{n}^{-1}\mathcal{O}(n^{-1/3})\bm \Delta_{n}\right]_{21}.
$$
Along $(-a,0)$ we have $Q>0$ so that $1/\sigma_n$ remains bounded uniformly, and consequently $\bm \Delta_{n}=\mathcal{O}(1)$ in the same interval. Using this information in the last displayed equation, we thus obtain that
\begin{equation}\label{eq:J1Cn}
\mathsf J_1(C_n)=\int_{C_n}\mathcal{O}(n^{-1/3}) \frac{\sigma_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x,\quad n\to\infty,
\end{equation}
uniformly for $u\geq \msf s$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$, and where the error term is uniform for $x\in C_n$.
Next, along $D_n$ we now use the asymptotic expansion \eqref{eq:ModelRHPAsymp} for $\bm\Phi=\widehat{\bm \Phi}_n$ and the definition of $\mathsf F_n$ and obtain that
$$
\bm P(x)=\bm U_0\mathsf m_+(x)^{\sp_3/4}\bm U_0^{-1}\left(1+\mathcal{O}(n^{-1/3})\right)=\bm U_0
\begin{pmatrix}
\mathcal{O}(1) & 0 \\
0 & \mathcal{O}(n^{1/6})
\end{pmatrix}
\bm U_0^{-1}
\left(1+\mathcal{O}(n^{-1/3})\right)
,\quad x\in D_n,
$$
and similarly
$$
\bm P(x)=
\left(1+\mathcal{O}(n^{-1/3})\right)
\bm U_0
\begin{pmatrix}
\mathcal{O}(n^{1/6}) & 0 \\
0 & \mathcal{O}(1)
\end{pmatrix}
\bm U_0^{-1}
,\quad x\in D_n,
$$
all valid as $n\to\infty$, uniformly for $x\in D_n$ and uniformly in $u,\msf t$ as claimed
With the same arguments that we applied in \eqref{eq:EstRRprimeF}, we obtain along $D_n$ as well
$$
\bm P^{-1}\bm R^{-1}\bm R'\bm P=\mathcal{O}(n^{-1/3}),\quad \text{so that}\quad \mathsf J_1(D_n)=\int_{D_n}\mathcal{O}(n^{-1/3}) \frac{\sigma_n(x\mid u)\ee^{-2n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x,\quad n\to\infty.
$$
The factor $\phi_+$ is purely imaginary on $(-a,0)$ and positive on $[0,+\infty)$, so the term $\ee^{-2n\phi_+}$ in the integrand above is bounded by a uniform constant. Combining with \eqref{eq:J1Cn}, we conclude
$$
\mathsf J_1(-\varepsilon_0,\varepsilon_1)=\mathcal{O}(n^{-1/3})\int_{-\varepsilon_0}^{\varepsilon_1}\frac{\sigma_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x
$$
Finally, we now integrate in $u$ and use Tonelli's Theorem to interchange the order of integration, obtaining just like in the proof of Lemma~\ref{lem:estI1epsilon1infty} that
$$
\int_{\msf s}^\infty\int_{-\varepsilon_0}^{\varepsilon_1}\frac{\sigma_n(x\mid u)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x \mathrm{d} u=
\int_{-\varepsilon_0}^{\varepsilon_1}\frac{1}{1+\ee^{\msf s+n^{2/3}Q(x)}}\mathrm{d} x\leq \varepsilon_1+\varepsilon_0,
$$
which concludes the proof.
\end{proof}
Next, in our pursuit of analyzing \eqref{eq:intKnIsplit}, the final missing piece of the puzzle is the term $\mathsf J_2(-\varepsilon_0,\varepsilon_0)$ that arises from $\mathsf I_1(-\varepsilon_0,\varepsilon_0)$. To state the rigorous results, we recall once again that the conformal map $\psi$ was introduced in Proposition~\ref{prop:conformalmappsi}, the function $\mathsf h_n$ is given in \eqref{local_par_L_Phi}, the function $\widehat{\bm \Delta}_n$ is in \eqref{deff:Deltan0} and in addition also set
$$
\zeta_0\deff\psi(-\varepsilon_0)<0,\quad \zeta_1\deff \psi(\varepsilon_1)>0.
$$
\begin{lemma}\label{lem:estJ2}
Fix $\msf s_0>0$ and $\msf t_0\in (0,1)$. There exists a function $\mathsf R_1(u)=\mathsf R_n(u\mid \msf t)$ satisfying the estimate
$$
\int_{\msf s}^\infty \mathsf R(u)\mathrm{d} u=\mathcal{O}(n^{-2/3}),\quad n\to\infty,
$$
uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$, and for which the identity
\begin{multline}\label{eq:I3I1Int}
\mathsf J_2(-\varepsilon_0,\varepsilon_0)=-\mathsf I_2(-\varepsilon_0,0)\\
+\int_{n^{2/3}\zeta_0}^{n^{2/3}\zeta_1}\frac{\ee^{\mathsf h_n(u)}}{\left(1+\ee^{\mathsf h_n(u)}\right)^2}\left[ \widehat{\bm \Delta}_n(x)^{-1}\widehat{\bm \Phi}_{n,+}(x)^{-1}\left(\widehat{\bm \Phi}_{n,+} \widehat{\bm\Delta}_n\right)'(x)\right]_{21}\mathrm{d} x+\mathsf R_1(u),
\end{multline}
holds true for every $u\in \mathbb{R}$ and every $\msf t >0$.
\end{lemma}
\begin{proof}
Recall that $\mathsf J_2$ was introduced in \eqref{deff:J1J2}. In the interval $(-\varepsilon_0,\varepsilon_1)$ the local parametrix $\bm P=\bm P^{(0)}$ coincides with \eqref{deff:ParamOrigin}. A direct calculation from \eqref{local_par_L_Phi} shows that the matrix $\bm F_n$ therein satisfies the identity
$$
\bm F_n'(z)=\frac{1}{4}\left(\frac{\mathsf m'(z)}{\mathsf m(z)}-\frac{\psi'(z)}{\psi(z)}\right)\bm F_n(z)\sp_3.
$$
With the help of this identity, we express
$$
\bm P_+^{-1}\bm P_+'=
\ee^{-n\phi_+\sp_3}\left[\frac{1}{4}\left(\frac{\mathsf m'}{\mathsf m}-\frac{\psi'}{\psi}\right) \bm \Phi_{n,+}^{-1}\sp_3 \bm \Phi_{n,+}+\bm \Phi_{n,+}^{-1}\bm \Phi_{n,+}'+n\phi'_+\sp_3 \right]\ee^{n\phi_+\sp_3},
$$
where both sides are evaluated at $x\in (-\varepsilon_0,\varepsilon_1)$. Thus,
\begin{multline}\label{eq:IntegrI3:1}
\Bigg[\bm\Lambda_{n}^{-1}
\bm P_+^{-1}\bm P_+'
\bm\Lambda_{n}\Bigg]_{21}\\
\begin{aligned}
& = -2n\frac{\phi_+' \chi_0\ee^{2n\phi_+}}{\sigma_n}+\ee^{2n\phi_+}\left[ \bm \Delta_{n} \bm \Phi_{n,+}^{-1}\bm \Phi_{n,+}' \bm\Delta_n\right]_{21}
+\frac{\ee^{2n\phi_+}}{4}\left(\frac{\mathsf m'}{\mathsf m}-\frac{\psi'}{\psi}\right) \left[\bm\Delta_n^{-1}\bm \Phi_{n,+}^{-1}\sp_3 \bm \Phi_{n,+}\bm\Delta_n\right]_{21} \\
&
\begin{multlined}
= -2n\frac{\phi_+' \chi_0\ee^{2n\phi_+}}{\sigma_n}
-\ee^{2n\phi_+}\left[\bm\Delta_n^{-1} \bm\Delta_n'\right]_{21}
+\ee^{2n\phi_+}\left[ \bm \Delta_{n} \bm \Phi_{n,+}^{-1}\left(\bm \Phi_{n,+} \bm\Delta_n\right)'\right]_{21}\\
\phantom{-2n\frac{\phi_+' \chi_0\ee^{2n\phi_+}}{\sigma_n}
-\ee^{2n\phi_+}\left[\bm\Delta_n^{-1} \bm\Delta_n'\right]_{21}}
+\frac{\ee^{2n\phi_+}}{4}\left(\frac{\mathsf m'}{\mathsf m}-\frac{\psi'}{\psi}\right) \left[\bm\Delta_n^{-1}\bm \Phi_{n,+}^{-1}\sp_3 \bm \Phi_{n,+}\bm\Delta_n\right]_{21}
\end{multlined}
\end{aligned}
\end{multline}
We now multiply this last expression by $\ee^{-2n\phi_+}\sigma_n/(1+\ee^{u+n^{2/3}Q})$ and integrate. A direct calculation gives
$$
\frac{\sigma_n}{1+\ee^{u+n^{2/3}Q}}\left(2n\frac{\phi_+' \chi_0}{\sigma_n}+\left[\bm\Delta_n^{-1} \bm\Delta_n'\right]_{21}\right)=
\frac{\sigma_n\chi_0}{1+\ee^{u+n^{2/3}Q}}\left(2n\frac{\phi_+' }{\sigma_n}+\left(\frac{1}{\sigma_n}\right)'\right)=\frac{\chi_0\sigma_n\ee^{-2n\phi_+}}{1+\ee^{u+n^{2/3}Q}}\left(\frac{\ee^{2n\phi_+}}{\sigma_n}\right)',
$$
which is the integrand of $\mathsf I_2(-\varepsilon_0,\varepsilon_1)=\mathsf I_2(-\varepsilon_0,0)$, see \eqref{deff:I2int}. Thus, from \eqref{eq:IntegrI3:1} we obtain
\begin{equation}\label{eq:J2I1intRu}
\mathsf J_2(-\varepsilon_0,\varepsilon_1)=-\mathsf I_1(-\varepsilon_0,0)+\int_{-\varepsilon_0}^{\varepsilon_1} \frac{\sigma_n(x)}{1+\ee^{u+n^{2/3}Q(x)}}\left[\bm\Delta_n(x)^{-1}\bm \Phi_{n,+}(x)^{-1}\left(\bm \Phi_{n,+}(x) \bm\Delta_n(x)\right)'\right]_{21}\mathrm{d} x+\mathsf R_1(u)
\end{equation}
where we have set
$$
\mathsf R_1(u)\deff \frac{1}{4}\int_{-\varepsilon_0}^{\varepsilon_1} \left(\frac{\mathsf m'(x)}{\mathsf m(x)}-\frac{\psi'(x)}{\psi(x)}\right) \frac{\sigma_n(x)}{1+\ee^{u+n^{2/3}Q(x)}}\left[\bm\Delta_n(x)^{-1}\bm \Phi_{n,+}(x)^{-1} \sp_3\bm \Phi_{n,+}(x) \bm\Delta_n(x)\right]_{21}\mathrm{d} x.
$$
It is worth mentioning that both $\mathsf m$ and $\psi$ have a simple zero at $x=0$, so the first term in the integrand of $\mathsf R$ remains bounded near $x=0$.
Recalling \eqref{local_par_L_Phi}, the integrand written explicitly in \eqref{eq:J2I1intRu} is of the form
$$
f_1(n^{2/3}\psi(x))\left[\bm f_2(n^{2/3}\psi(x))\left(\bm f_3(n^{2/3}\psi(x))\right)'\right]_{21}=n^{2/3}\psi'(x)f_1(n^{2/3}\psi(x))\left[\bm f_2(n^{2/3}\psi(x))\bm f_3'(n^{2/3}\psi(x))\right]_{21}
$$
with obvious choices of the functions $f_1,\bm f_2,\bm f_3$, and performing the change of variables $u=n^{2/3}\psi(x)$ we obtain the integral in the right-hand side of \eqref{eq:I3I1Int}.
To conclude, it remains to show the bound for $\mathsf R$. For that, we use again the sets $C_n$ and $D_n$ from \eqref{deff:setsCnDn} and analyze the integral over each of these sets separately.
Along $C_n$, the convergence \eqref{eq:asympPhinPhi0} ensures $\bm \Phi_{n,+}$ remains bounded uniformly, and combined with the boundedness of all the other terms on the whole interval $(-\varepsilon_0,\varepsilon_1)$ we obtain
\begin{multline*}
\int_{C_n}\left(\frac{\mathsf m'(x)}{\mathsf m(x)}-\frac{\psi'(x)}{\psi(x)}\right) \frac{\sigma_n(x)}{1+\ee^{u+n^{2/3}Q(x)}}\left[\bm\Delta_n(x)^{-1}\bm \Phi_{n,+}(x)^{-1} \sp_3\bm \Phi_{n,+}(x) \bm\Delta_n(x)\right]_{21}\mathrm{d} x\\
=\mathcal{O}(1)\int_{C_n} \frac{\sigma_n(x)}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x.
\end{multline*}
The function $\psi$ is conformal, and consequently $|z|\leq cn^{-2/3}$ for $z\in C_n$ and some constant $c$ independent of $\msf t,u,n$. In particular, this ensures that $n^{2/3}Q(x)\leq -\tilde c$ for every $x\in C_n$ and a constant $\tilde c>0$, and therefore
$$
\frac{\sigma_n(x)}{1+\ee^{u+n^{2/3}Q(x)}}\leq \frac{1}{1+\ee^{u-\tilde c}}\leq \ee^{-u+\tilde c}.
$$
Still because $\psi$ is conformal, we are sure that the Lebesgue measure of $C_n$ is $\mathcal{O}(n^{-2/3})$. Everything combined, we conclude the estimate
$$
\int_{C_n}\left(\frac{\mathsf m'(x)}{\mathsf m(x)}-\frac{\psi'(x)}{\psi(x)}\right) \frac{\sigma_n(x)}{1+\ee^{u+n^{2/3}Q(x)}}\left[\bm\Delta_n(x)^{-1}\bm \Phi_{n,+}(x)^{-1} \sp_3\bm \Phi_{n,+}(x) \bm\Delta_n(x)\right]_{21}\mathrm{d} x=\mathcal{O}(\ee^{-u}n^{-2/3}),
$$
as $n\to\infty$, which is valid uniformly for $u,\msf t$ as claimed by the Lemma.
Finally, on $D_n$ we use the expansion \eqref{eq:AsymptPhicc} for $\bm \Phi=\bm\Phi_n$, which provides
$$
\bm\Phi_{n,+}(x)^{-1}\sp_3\bm \Phi_{n,+}(x)=\ee^{n\phi_+(x)\sp_3}\left(\sp_2 +\mathcal{O}(n^{-1/3}) \right)\ee^{-n\phi_+(x)\sp_3},\quad n\to\infty,
$$
which is valid uniformly for $x\in D_n$ and uniformly in the parameters $u,\msf t$ as required. After some straightforward calculations, we thus arrive at
\begin{multline*}
\int_{D_n}\left(\frac{\mathsf m'(x)}{\mathsf m(x)}-\frac{\psi'(x)}{\psi(x)}\right) \frac{\sigma_n(x)}{1+\ee^{u+n^{2/3}Q(x)}}\left[\bm\Delta_n(x)^{-1}\bm \Phi_{n,+}(x)^{-1} \sp_3\bm \Phi_{n,+}(x) \bm\Delta_n(x)\right]_{21}\mathrm{d} x \\
= \int_{D_n}\left(\frac{\mathsf m'(x)}{\mathsf m(x)}-\frac{\psi'(x)}{\psi(x)}\right) \frac{\mathrm{i}\sigma_n(x)\ee^{-2n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}} \left[ 1+\frac{\chi_0(x)}{\sigma_n(x)^2}+\mathcal{O}(n^{-1/3}) \right]\mathrm{d} x,\quad n\to\infty,
\end{multline*}
with, as always, uniform error term in $x\in D_n$, $u,\msf t$. Each of the terms $(\mathsf m'/\mathsf m-\psi'/\psi)$ and $\chi_0/\sigma_n^2$ is bounded on $(-\varepsilon_0,\varepsilon_1)$, so to bound the integral above it is enough to estimate
$$
\int_{-\varepsilon_0}^{\varepsilon_1} \frac{\sigma_n(x)\ee^{-n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x=\int_{-\varepsilon_0}^{0} \frac{\sigma_n(x)\ee^{-n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x+\int_{0}^{\varepsilon_1} \frac{\sigma_n(x)\ee^{-n\phi(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x.
$$
For the integral over $(-\varepsilon_0,0)$, we know that $\re\phi_+=0$, and then using Tonelli's Theorem to integrate first in $u$ we obtain that
$$
\int_{\msf s}^{+\infty} \left|\int_{-\varepsilon_0}^{0} \frac{\sigma_n(x)\ee^{-n\phi_+(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x\right|\mathrm{d} u\leq
\int_{-\varepsilon_0}^0 \frac{1}{1+\ee^{\msf s+n^{2/3}Q(x)}}\mathrm{d} x.
$$
Changing variables $v=n^{2/3}Q$ in this last integral, it then follows that the right-hand side above is $\mathcal{O}(n^{-2/3})$ uniformly for $\msf s\geq -\msf s_0$ and $\msf t_0\leq \msf t\leq 1/\msf t_0$.
Finally, for the integral over $(0,\varepsilon_1)$ we now have that $\phi\geq 0$ in this interval, and it is independent of $u$, so once again interchanging order of integration we obtain
$$
0\leq \int_\msf s^\infty \int_{0}^{\varepsilon_1} \frac{\sigma_n(x)\ee^{-n\phi(x)}}{1+\ee^{u+n^{2/3}Q(x)}}\mathrm{d} x \mathrm{d} u=\int_0^{\varepsilon_1}\frac{\ee^{-n\phi(x)}}{1+\ee^{\msf s+n^{2/3}Q(x)}}\mathrm{d} x\leq \int_0^{\varepsilon_1}\ee^{-n\phi(x)}\mathrm{d} x,
$$
and now changing variables $v=n\phi$ (which is well defined in this interval because of the local behavior \eqref{eq:phi_local_beh}) we see that the integral on the right-most side is $\mathcal{O}(n^{-1})$. This completes the proof.
\end{proof}
To conclude, it remains to prove Proposition~\ref{prop:finalestKnQ}.
\begin{proof}[Proof of Proposition~\ref{prop:finalestKnQ}]
Recalling \eqref{eq:intKnIsplit} and \eqref{eq:splitI1J1J2}, the result is an immediate consequence of Lemmas~\ref{lem:I1inftyI1aI2a}--\ref{lem:estJ2}.
\end{proof}
|
\section{Introduction} \label{section:introduction}
In \cite{berger:79b}, Berger introduced a framework of statistical decision problems under communication constraint.
Inspired by his work, many researchers studied various problems of this kind \cite{AhlCsi86, han:87, ZhaBer88, amari-han:89, amari:89, han-kobayashi:89, ShaPap92};
see \cite{han-amari:98} for a thorough review until 90s. More recently, the problem of communication constrained statistics has regained interest of researchers;
see \cite{Amari:11, ueta-kuzuoka:14, polyanskiy:12, RahWag:12, XiaKim:13, ZhaLai:18, Wat:18, WeiKoc:19, SalWigWan:19, HadLiuPolSha:19, HadSha:19, EscWigZai:20, Bur:20, SreGun:20, SahTya:21}
and references therein.
In this paper, we are interested in the
most basic setting of hypothesis testing in which the sender and the receiver observe correlated sources $X^n$ and $Y^n$, respectively,
and the observation $(X^n,Y^n)$ is distributed according to the product of the null hypothesis $P_{XY}$ or the alternative
hypothesis $Q_{XY}$; the sender transmit a message at rate $R$, and the receiver decide the hypotheses based on the message and its observation.
In \cite{AhlCsi86}, Ahlswede and Csisz\'ar introduced the so-called quantization scheme,
and showed that the quantization scheme is optimal for the testing against independence, i.e.,
the alternative hypothesis is $Q_{XY} = P_X\times P_Y$. In \cite{han:87}, Han introduced an improved
version of the quantization scheme that fully exploit the knowledge of the marginal types.
In \cite{ShiHanAma94}, Shimokawa, Han, and Amari introduced the quantize and binning scheme, which we shall
call the SHA scheme in the following. For a long time, it has not been known if the SHA is optimal or not.
In \cite{RahWag:12}, Rahman and Wagner showed that the SHA scheme is optimal for the testing against conditional
independence.\footnote{In fact, for the testing against conditional independence, it suffices to consider a simpler version of quantize and binning scheme
than the SHA scheme. } More recently, Weinberger and Kochman \cite{WeiKoc:19} (see also \cite{HaiKoch:16}) derived the optimal exponential
trade-off between the type I and type II error probabilities of the quantize and binning scheme.
Since the decision is conducted based on the optimal likelihood decision rule in \cite{WeiKoc:19}, their scheme may potentially provide
a better performance than the SHA scheme; however, since the expression involves complicated optimization over multiple parameters, and a strict improvement has
not been clarified so far.
A main objective of this paper is to investigate if the SHA scheme is optimal or not. In fact, it turns out that the answer is negative,
and there exists a scheme that improves upon the SHA scheme.
Since the general SHA scheme is involved, we focus on the critical rate, i.e., the minimum rate that is required to attain the same type II exponent
as the Stein exponent of the centralized testing. Then, under a certain non-degeneracy condition,
we will prove that the auxiliary random variable in the SHA scheme must coincide with the source $X$ itself, i.e., we must use the binning without quantization.
By leveraging this simplification, for the binary symmetric double sources (BSDS), we derive a closed form expression of the critical rate attainable by the SHA scheme.
Next, we will consider a product of the BSDS such that each component of the null hypothesis and the alternative hypothesis is ``reversely aligned."
Perhaps surprisingly, it turns out that the SHA is sub-optimal for such hypotheses. Our improved scheme is simple, and it uses the SHA scheme
for each component in a sequential manner; however, it should be emphasized that we need to conduct binning of each component separately,
and our scheme is not a naive random binning.
\section{Problem Formulation}
We consider the statistical problem of testing the null hypothesis $\mathtt{H}_0: P_{XY}$ on finite alphabets ${\cal X}\times {\cal Y}$ versus the alternative
hypothesis $\mathtt{H}_1: Q_{XY}$ on the same alphabet. The i.i.d. random variables $(X^n,Y^n)$ distributed according to either $P_{XY}^n$ or $Q_{XY}^n$
are observed by the sender and the receiver; the sender encodes the observation $X^n$ to a message by an encoding function
\begin{align*}
\varphi_n: {\cal X}^n \to {\cal M}_n,
\end{align*}
and the receiver decides whether to accept the null hypothesis or not by a decoding function
\begin{align*}
\psi_n:{\cal M}_n \times {\cal Y}^n \to \{\mathtt{H}_0,\mathtt{H}_1 \}.
\end{align*}
When the block length $n$ is obvious from the context, we omit the subscript $n$.
For a given testing scheme $T_n = (\varphi,\psi)$, the type I error probability is defined by
\begin{align*}
\alpha[T_n] := P\bigg( \psi(\varphi(X^n),Y^n) = \mathtt{H}_1 \bigg)
\end{align*}
and the type II error probability is defined by
\begin{align*}
\beta[T_n] := Q\bigg( \psi(\varphi(X^n),Y^n) = \mathtt{H}_0 \bigg).
\end{align*}
In the following, $P(\cdot)$ (or $Q(\cdot)$) means that $(X^n,Y^n)$ is distributed according to $P_{XY}^n$ (or $Q_{XY}^n$).
In the problem of distributed hypothesis testing, we are interested in the trade-off among the communication rate
and the type I and type II error probabilities. Particularly, we focus on
the optimal trade-off between the communication rate and the exponent of the type II error probability
when the type I error probability is vanishing, which is sometimes called ``Stein regime".
\begin{definition}
A pair $(R,E)$ of the communication rate and the exponent is defined to be achievable if there exists a sequence $\{T_n\}_{n=1}^\infty$
of testing schemes such that
\begin{align*}
\lim_{n\to\infty} \alpha[T_n] = 0
\end{align*}
and
\begin{align*}
\limsup_{n\to\infty} \frac{1}{n} \log |{\cal M}_n| &\le R, \\
\liminf_{n\to\infty} - \frac{1}{n} \log \beta_n[T_n] &\ge E.
\end{align*}
The achievable region ${\cal R}$ is the set of all achievable pair $(R,E)$.
\end{definition}
For a given communication rate $R$, the maximum exponent is denoted by
\begin{align*}
E(R) := \max\{ E : (R,E) \in {\cal R} \}.
\end{align*}
When there is no communication constraint, i.e., $R \ge \log |{\cal X}|$, the sender can send
the full observation without any encoding, and the receiver can use a scheme for the standard
(centralized) hypothesis testing. In such a case, the Stein exponent $D(P_{XY}\|Q_{XY})$ is attainable.
We define the critical rate as the minimum communication rate required to attain the Stein exponent:
\begin{align*}
R_{\mathtt{cr}} := \inf\{ R : E(R) = D(P_{XY}\|Q_{XY}) \}.
\end{align*}
\section{On Evaluation of SHA Scheme}
When the communication rate $R$ is not sufficient to send $X$, the standard approach is to quantize $X$ using
a test channel $P_{U|X}$ for some finite auxiliary alphabet ${\cal U}$. For $P_{UXY} = P_{U|X}P_{XY}$
and $Q_{UXY} = Q_{U|X} Q_{XY}$ with $Q_{U|X} = P_{U|X}$, let
\begin{align*}
\lefteqn{ E(P_{UXY} \| Q_{UXY}) } \\
&:= \min\big\{ D(P_{\tilde{U}\tilde{X}\tilde{Y}} \| Q_{UXY})
: P_{\tilde{U}\tilde{X}} = P_{UX}, P_{\tilde{U}\tilde{Y}}=P_{UY} \big\}.
\end{align*}
In \cite{han:87}, a testing scheme based on quantization was proposed, and the following achievability bound was derived: for any test channel $P_{U|X}$,
\begin{align*}
E(R) \ge E(P_{UXY} \| Q_{UXY}).
\end{align*}
In \cite{ShiHanAma94}, a testing scheme based on quantization and binning was proposed,
and the following achievability bound was derived: for any test channel $P_{U|X}$ such that $R \ge I(U \wedge X|Y)$,
\begin{align*}
E(R) &\ge \min\bigg[ \min_{P_{\tilde{U}\tilde{X}\tilde{Y}} \in{\cal P}_{\mathtt{b}}(P_{U|X})} D(P_{\tilde{U}\tilde{X}\tilde{Y}} \| Q_{UXY}) \\
&~~~ ~~~+ | R - I(U \wedge X|Y)|^+, E(P_{UXY} \| Q_{UXY}) \bigg],
\end{align*}
where $|t|^+ = \max[t,0]$, and
\begin{align*}
{\cal P}_{\mathtt{b}}(P_{U|X}) &:= \big\{ P_{\tilde{U}\tilde{X}\tilde{Y}} : \\
&~~~ P_{\tilde{U}\tilde{X}} = P_{UX}, P_{\tilde{Y}} = P_Y, H(\tilde{U}|\tilde{Y}) \ge H(U|Y) \big\}.
\end{align*}
Particularly, by taking $P_{U|X}$ to be noiseless test channel, i.e., $U=X$, we have
\begin{align} \label{eq:b-X}
E(R) &\ge E_{\mathtt{b}}(R) \\
&:= \min\bigg[ \min_{P_{\tilde{X}\tilde{Y}} \in {\cal P}_{\mathtt{b}}} D(P_{\tilde{X}\tilde{Y}} \| Q_{XY}) + |R - H(X|Y)|^+, \nonumber \\
&~~~~~~ D(P_{XY}\|Q_{XY}) \bigg],
\end{align}
where
\begin{align*}
{\cal P}_{\mathtt{b}} := \big\{ P_{\tilde{X}\tilde{Y}} : P_{\tilde{X}}=P_X, P_{\tilde{Y}}=P_Y, H(\tilde{X}|\tilde{Y}) \ge H(X|Y) \big\}.
\end{align*}
In the following, we focus on the critical rate. Under some mild conditions, in order to attain the Stein exponent,
we must take $U=X$ in the quantization-bining bound, i.e., no quantization. In the rest of this section, we prove this claim.
Suppose that $P_{XY}$ and $Q_{XY}$ have full support. Let $\hat{\Lambda}$ be a function on ${\cal X} \times {\cal Y}$ defined by\footnote{We suppose that
alphabets are ${\cal X}=\{0,\ldots,m_\san{x}\}$ and ${\cal Y}=\{0,\ldots,m_\san{y}\}$; taking $y=0$ as a special symbol in \eqref{eq:LLR}
is just notational convenience.}
\begin{align} \label{eq:LLR}
\hat{\Lambda}(x,y) := \log \frac{P_{XY}(x,y)}{Q_{XY}(x,y)} - \log \frac{P_{XY}(x,0)}{Q_{XY}(x,0)}.
\end{align}
\begin{proposition} \label{prop:no-quantization}
Suppose that, for every $x\neq x^\prime$, rows $\hat{\Lambda}(x,\cdot)$ and $\hat{\Lambda}(x^\prime,\cdot)$ are distinct.
Then, $E(P_{UXY} \| Q_{UXY}) = D(P_{XY} \| Q_{XY})$ only if there does not exist $u\in {\cal U}$ and $x\neq x^\prime$ such that
$P_{U|X}(u|x) P_{U|X}(u|x^\prime) > 0$.
\end{proposition}
\begin{proof}
Since we assumed that $P_{XY}$ and $Q_{XY}$ have full support and since $P_{U|X}=Q_{U|X}$,
$P_{UXY}$ and $Q_{UXY}$ have the same support, which we denote by ${\cal A}$.
Let ${\cal A}^\prime = \mathtt{supp}(P_{UX}) = \mathtt{supp}(Q_{UX})$. Note that $P_{UY}$ and $Q_{UY}$ have full support.
Let ${\cal L} \subseteq {\cal P}({\cal A})$ be the linear family of distributions $P_{\tilde{U}\tilde{X}\tilde{Y}}$ satisfying
\begin{align*}
\sum_{(u,x,y) \in {\cal A}} P_{\tilde{U}\tilde{X}\tilde{Y}}(u,x,y) \delta_{\bar{u}\bar{x}}(u,x) = P_{UX}(\bar{u},\bar{x})
\end{align*}
for every $(\bar{u},\bar{x}) \in {\cal A}^\prime$ and
\begin{align*}
\sum_{(u,x,y) \in {\cal A}} P_{\tilde{U}\tilde{X}\tilde{Y}}(u,x,y) \delta_{\bar{u}\bar{y}}(u,y) = P_{UY}(\bar{u},\bar{y}),
\end{align*}
for every $(\bar{u},\bar{y}) \in {\cal U} \times {\cal Y}$.
Then, we can write $E(P_{UXY}\|Q_{UXY})$ as
\begin{align} \label{eq:minimization-problem}
E(P_{UXY}\|Q_{UXY}) = \min_{P_{\tilde{U}\tilde{X}\tilde{Y}} \in {\cal L}} D(P_{\tilde{U}\tilde{X}\tilde{Y}}\|Q_{UXY}).
\end{align}
If
\begin{align*}
E(P_{UXY} \| Q_{UXY}) = D(P_{XY} \| Q_{XY}) = D(P_{UXY}\|Q_{UXY}),
\end{align*}
then $P_{UXY}$ must be the (unique) minimizer in \eqref{eq:minimization-problem}.
Then, the I-projection theorem (cf.~\cite[Theorem 3.2]{csiszar-shields:04book})
implies that the minimizer $P_{UXY}$ lies in the exponential family generated by
$Q_{UXY}$ and the constraints of ${\cal L}$, i.e., it has of the form
\begin{align*}
\lefteqn{ P_{UXY}(u,x,y) }\\
&\propto Q_{UXY}(u,x,y) \exp\bigg[ \sum_{\bar{u},\bar{x}} \theta_{\bar{u}\bar{x}} \delta_{\bar{u}\bar{x}}(u,x)
+ \sum_{\bar{u},\bar{y}} \theta_{\bar{u}\bar{y}} \delta_{\bar{u}\bar{y}}(u,y) \bigg].
\end{align*}
for some $\theta_{\bar{u}\bar{x}},\theta_{\bar{u}\bar{y}} \in \mathbb{R}$. Equivalently, this condition can be written as
\begin{align*}
\log \frac{P_{UXY}(u,x,y)}{Q_{UXY}(u,x,y)} = g_1(u,x) + g_2(u,y)
\end{align*}
for some functions $g_1$ on ${\cal A}^\prime$ and $g_2$ on ${\cal U}\times {\cal Y}$. If there exist $u\in {\cal U}$ and $x\neq x^\prime$ such that
$P_{U|X}(u|x) P_{U|X}(u|x^\prime) > 0$, then, for every $y\in {\cal Y}$, we have
\begin{align*}
\lefteqn{ \hat{\Lambda}(x,y) - \hat{\Lambda}(x^\prime,y) } \\
&= \bigg[ \log\frac{P_{XY}(x,y)}{Q_{XY}(x,y)} - \log \frac{P_{XY}(x,0)}{Q_{XY}(x,0)} \bigg] \\
&~~~ - \bigg[ \log \frac{P_{XY}(x^\prime,y)}{Q_{XY}(x^\prime,y)} - \log \frac{P_{XY}(x^\prime,0)}{Q_{XY}(x^\prime,0)} \bigg] \\
&= \bigg[ \log\frac{P_{UXY}(u,x,y)}{Q_{UXY}(u,x,y)} - \log \frac{P_{UXY}(u,x,0)}{Q_{UXY}(u,x,0)} \bigg] \\
&~~~ - \bigg[ \log \frac{P_{UXY}(u,x^\prime,y)}{Q_{UXY}(u,x^\prime,y)} - \log \frac{P_{UXY}(u,x^\prime,0)}{Q_{UXY}(u,x^\prime,0)} \bigg] \\
&= \bigg[ g_1(u,x) + g_2(u,y) - g_1(u,x) - g_2(u,0) \bigg] \\
&~~~ - \bigg[ g_1(u,x^\prime) + g_2(u,y) - g_1(u,x^\prime) - g_2(u,0) \bigg] \\
&=0,
\end{align*}
which contradict the assumption that $\hat{\Lambda}(x,\cdot)$ and $\hat{\Lambda}(x^\prime,\cdot)$ are distinct.
\end{proof}
Since the quantization-binning bound is at most $E(P_{UXY}\|Q_{UXY})$, when the assumption of Proposition \ref{prop:no-quantization} is satisfied,
$U=X$ is the only choice of test channel to attain the Stein exponent.\footnote{Even though we can take
a redundant test channel $P_{U|X}$ such that $\mathtt{supp}(P_{U|X}(\cdot|x)) \cap \mathtt{supp}(P_{U|X}(\cdot|x^\prime)) = \emptyset$
for any $x \neq x^\prime$, such a test channel does not improve the attainable exponent.}
In Proposition \ref{prop:no-quantization},
the assumption that every raws of $\hat{\Lambda}(x,y)$ are distinct cannot be replaced by the assumption
that every raws of the log-likelihood function
\begin{align*}
\Lambda(x,y) = \log \frac{P_{XY}(x,y)}{Q_{XY}(x,y)}
\end{align*}
are distinct; for instance, when $P_{XY}$ and $Q_{XY}$ are given by
\begin{align*}
P_{XY} &= \frac{1}{24a} \left[
\begin{array}{ccc}
a & 2 a & 3a \\ 2a & 3a & 3a \\ a & 3a & 6a
\end{array}
\right], \\
Q_{XY} &= \frac{1}{14a}
\left[
\begin{array}{ccc}
2a & a & a \\ a & a & a \\ a & 2a & 4a
\end{array}
\right]
\end{align*}
for some $a >0$, then every raws $\Lambda(x,\cdot)$ are distinct but
$\hat{\Lambda}(1,\cdot) = \hat{\Lambda}(2,\cdot)$. In fact, as is shown in the following Proposition \ref{prop:merged},
the symbols $x=1$ and $x=2$ in this example can be merged while maintaining
$E( P_{UXY} \| Q_{UXY}) = D(P_{XY} \| Q_{XY})$. Even though Proposition \ref{prop:merged} is not used in the rest
of the paper, it may be of independent interest.
\begin{proposition} \label{prop:merged}
Let $\kappa:{\cal X} \to {\cal U}$ be a function such that $\kappa(x)=\kappa(x^\prime)$
if and only if $\hat{\Lambda}(x,\cdot)=\hat{\Lambda}(x^\prime,\cdot)$, and let $U = \kappa(X)$.
Then, for any $P_{\tilde{U}\tilde{X}\tilde{Y}}$ satisfying $P_{\tilde{U}\tilde{X}} = P_{UX}$ and $P_{\tilde{U}\tilde{Y}}=P_{UY}$, we have
\begin{align}
\lefteqn{ D(P_{\tilde{U}\tilde{X}\tilde{Y}} \| Q_{UXY}) } \nonumber \\
&= D(P_{\tilde{U}\tilde{X}\tilde{Y}} \| P_{UXY}) + D(P_{UXY} \| Q_{UXY}). \label{eq:Phytagorean}
\end{align}
Particularly, we have
\begin{align} \label{eq:no-loss}
E( P_{UXY} \| Q_{UXY}) = D(P_{XY} \| Q_{XY}).
\end{align}
\end{proposition}
\begin{proof}
Since \eqref{eq:no-loss} follows from \eqref{eq:Phytagorean}, it suffices to prove \eqref{eq:Phytagorean}.
For each $u$, fix $x_u\in {\cal X}$ such that $\kappa(x_u)=u$. Then, we have
\begin{align*}
\lefteqn{ D(P_{\tilde{U}\tilde{X}\tilde{Y}} \| Q_{UXY} ) - D(P_{\tilde{U}\tilde{X}\tilde{Y}} \| P_{UXY}) - D(P_{UXY} \| Q_{UXY}) } \\
&= \sum_{u,x,y} \big( P_{\tilde{U}\tilde{X}\tilde{Y}}(u,x,y) - P_{UXY}(u,x,y) \big) \log \frac{P_{UXY}(u,x,y)}{Q_{UXY}(u,x,y)} \\
&= \sum_{u,x,y} \big( P_{\tilde{U}\tilde{X}\tilde{Y}}(u,x,y) - P_{UXY}(u,x,y) \big) \log \frac{P_{XY}(x,y)}{Q_{XY}(x,y)} \\
&= \sum_{u,x,y} \big( P_{\tilde{U}\tilde{X}\tilde{Y}}(u,x,y) - P_{UXY}(u,x,y) \big) \hat{\Lambda}(x,y) \\
&= \sum_{u,x,y} \big( P_{\tilde{U}\tilde{X}\tilde{Y}}(u,x,y) - P_{UXY}(u,x,y) \big) \hat{\Lambda}(x_u,y) \\
&= \sum_{u,y} \big( P_{\tilde{U}\tilde{Y}}(u,y) - P_{UY}(u,y) \big) \hat{\Lambda}(x_u,y) \\
&=0,
\end{align*}
where the third equality follows from $P_{\tilde{X}}=P_X$ and the last equality follows from $P_{\tilde{U}\tilde{Y}}=P_{UY}$.
\end{proof}
\begin{remark}
Introducing the function $\Lambda(x,y)$ as in \eqref{eq:LLR} is motivated by the problem of distributed function computing.
Note that, in order to conduct the likelihood ratio test, the receiver need not compute the
log-likelihood ratios $(\Lambda(x_i,y_i) : 1 \le i \le n)$ for a given observation $(\bm{x},\bm{y}) \in {\cal X}^n \times {\cal Y}^n$;
instead, it suffices to compute the summation
\begin{align*}
\Lambda_n(\bm{x},\bm{y}) = \sum_{i=1}^n \Lambda(x_i,y_i).
\end{align*}
In the terminology of distributed computing \cite[Definition 4]{KuzWat16},
we can verify that the function $\Lambda_n$ is
$\overline{{\cal X}}_{\hat{\Lambda}}$-informative.
\end{remark}
\section{Binary Symmetric Double Source (BSDS)}
Let ${\cal X}={\cal Y} = \{0,1\}$, and let $P_{XY}$ and $Q_{XY}$ be given by
\begin{align*}
P_{XY} = \left[
\begin{array}{cc}
\frac{1-p}{2} & \frac{p}{2} \\
\frac{p}{2} & \frac{1-p}{2}
\end{array}
\right],~~~
Q_{XY} = \left[
\begin{array}{cc}
\frac{1-q}{2} & \frac{q}{2} \\
\frac{q}{2} & \frac{1-q}{2}
\end{array}
\right]
\end{align*}
for $0\le p,q \le 1$ and $p \neq q$.
Since $\hat{\Lambda}(0,0)=\hat{\Lambda}(1,0)=0$ and
\begin{align*}
\hat{\Lambda}(0,1) = \log \frac{(1-q)p}{(1-p)q} \neq \log \frac{(1-p)q}{(1-q)p} = \hat{\Lambda}(1,1),
\end{align*}
the assumption of Proposition \ref{prop:no-quantization} is satisfied. Thus, we need to set $U=X$ in the quantization-binning bound to attain the Stein exponent.
In the minimization of \eqref{eq:b-X}, since $P_{\tilde{X}}=P_X$ and $P_{\tilde{Y}}=P_Y$ are the uniform distribution, the only free parameter is the crossover probability
$P_{\tilde{Y}|\tilde{X}}(1|0)=P_{\tilde{Y}|\tilde{X}}(0|1)=\tilde{p}$. Then, we can easily find that the binning bound for BSDS is
\begin{align}
\lefteqn{ E_{\mathtt{b}}(R) } \nonumber \\
&= \left\{
\begin{array}{ll}
\min[ |R - h(p)|^+, D(p\|q) ] & \mbox{ if } h(p) \le h(q) \\
D(p\| q) & \mbox{ if } h(p) > h(q)
\end{array}
\right. \label{eq:b-Exponent-BSDS}
\end{align}
for $R \ge h(p)$. Furthermore, this bound implies that the critical rate attainable by the SHA scheme is
\begin{align} \label{eq:b-critical-BSDS}
R_{\mathtt{cr}} \le \left\{
\begin{array}{ll}
h(p) + D(p\|q) & \mbox{ if } h(p) \le h(q) \\
h(p) & \mbox{ if } h(p) > h(q)
\end{array}
\right..
\end{align}
\section{Product of BSDS} \label{section:product-BSDS}
Next, let us consider the case where $P_{XY} = P_{X_1 Y_1} \times P_{X_2 Y_2}$
and $Q_{XY} = Q_{X_1 Y_1} \times Q_{X_2 Y_2}$, and $P_{X_i Y_i}$ and $Q_{X_i Y_i}$ for $i=1,2$
are DSBSs with parameters $0 \le p_1,p_2,q_1,q_2 \le 1$, respectively. Particularly, we consider the case where $p_1 = q_2$,
$p_2 = q_1$, where $p_1 \neq q_1$. Note that, for the product source, the function defined by \eqref{eq:LLR} can be decomposed as
\begin{align*}
\hat{\Lambda}(x_1x_2,y_1y_2) = \hat{\Lambda}_1(x_1,y_1) + \hat{\Lambda}_2(x_2,y_2),
\end{align*}
where
\begin{align*}
\hat{\Lambda}_i(x_i,y_i) = \log \frac{P_{X_i Y_i}(x_i,y_i)}{Q_{X_iY_i}(x_i,y_i)} - \log \frac{P_{X_i Y_i}(x_i,0)}{Q_{X_iY_i}(x_i,0)}.
\end{align*}
Furthermore, we have $\hat{\Lambda}_1(1,1) = - \hat{\Lambda}_1(0,1)$,
$\hat{\Lambda}_2(0,1)=-\hat{\Lambda}_1(0,1)$, and $\hat{\Lambda}_2(1,1)=\hat{\Lambda}_1(0,1)$. Thus, by denoting $a = \hat{\Lambda}_1(0,1)$, we have
\begin{align*}
\hat{\Lambda}(x_1x_2,y_1y_2) =
\left[
\begin{array}{cccc}
0 & - a & a & 0 \\
0 & a & a & 2 a \\
0 & - a & - a & - 2 a \\
0 & a & - a & 0
\end{array}
\right],
\end{align*}
and the assumption of Proposition \ref{prop:no-quantization} is satisfied. Thus, we need to set $U=X$ in the quantization-binning bound to attain the Stein exponent.
Since $p_1=q_2$ and $p_2=q_1$, the conditional entropies $H_P(X|Y)$ and $H_Q(X|Y)$ with respect to $P_{XY}$ and $Q_{XY}$ satisfy
$H_P(X|Y)=H_Q(X|Y)$. Thus, we find that the inner minimization of the binning bound is
attained by $P_{\tilde{X}\tilde{Y}} = Q_{XY}$, and
\begin{align*}
\lefteqn{ E_{\mathtt{b}}(R) } \\
&= \min\big[ |R- h(p_1) - h(p_2)|^+, D(p_1\|q_1) + D(p_2\|q_2) \big].
\end{align*}
Furthermore, this bound implies that the critical rate attainable by the SHA scheme is
\begin{align} \label{eq:critical-product-SHA}
R_{\mathtt{cr}} \le h(p_1) + h(p_2) + D(p_1\|q_1) + D(p_2 \| q_2).
\end{align}
\section{Sequential Scheme for Product of BSDS} \label{section:sequential}
Now, we describe a modified version of SHA scheme for the product of BSDS.
For $i=1,2$, let $T_{\mathtt{b}}^{[i]} = (\varphi^{[i]},\psi^{[i]})$ be the SHA scheme (without quantization) for $P_{X_iY_i}$ versus $Q_{X_iY_i}$.
For a given rate $R$, we split the rate as $R=R_1 + R_2$, and consider sequential scheme $\tilde{T}_{\mathtt{b}} = (\varphi,\psi)$
constructed from $(T_{\mathtt{b}}^{[1]},T_{\mathtt{b}}^{[2]})$ as follows.
Upon observing $\bm{x} = (\bm{x}_1,\bm{x}_2)$,
the sender transmit $m_1 = \varphi^{[1]}(\bm{x}_1)$ and $m_2 = \varphi^{[2]}(\bm{x}_2)$.\footnote{Note that $m_1$ and $m_2$ include
indices of random binning as well as marginal types of $\bm{x}_1$ and $\bm{x}_2$.}
Then, upon receiving $(m_1,m_2)$ and observing $\bm{y}=(\bm{y}_1,\bm{y}_2)$, the receiver decides
$\psi(m_1,m_2,\bm{y}) = \mathtt{H}_0$ if $\psi^{[1]}(m_1,\bm{y}_1) = \mathtt{H}_0$ and $\psi^{[2]}(m_2,\bm{y}_2) = \mathtt{H}_0$.
The type I and type II error probabilities of this sequential scheme $\tilde{T}_{\mathtt{b}}$ can be evaluated as
\begin{align*}
\lefteqn{ \alpha[\tilde{T}_{\mathtt{b}}] } \\
&= P\bigg( \psi^{[1]}(\varphi^{[1]}(X_1^n),Y_1^n) = \mathtt{H}_1 \vee \psi^{[2]}(\varphi^{[2]}(X_2^n),Y_2^n) = \mathtt{H}_1 \bigg) \nonumber \\
&\le \sum_{i=1}^2 P\bigg( \psi^{[i]}(\varphi^{[i]}(X_i^n),Y_i^n) = \mathtt{H}_1 \bigg)
\end{align*}
by the union bound, and
\begin{align*}
\lefteqn{ \beta[\tilde{T}_{\mathtt{b}}] } \\
&= Q\bigg( \psi^{[1]}(\varphi^{[1]}(X_1^n),Y_1^n) = \mathtt{H}_0,~ \psi^{[2]}(\varphi^{[2]}(X_2^n),Y_2^n) = \mathtt{H}_0 \bigg) \nonumber \\
&= \prod_{i=1}^2 Q\bigg( \psi^{[i]}(\varphi^{[i]}(X_i^n),Y_i^n) = \mathtt{H}_0 \bigg).
\end{align*}
Thus, if the type I error probabilities of each scheme is vanishing, then the type I error probability $\alpha[\tilde{T}_{\mathtt{b}}]$
is also vanishing. On the other hand, the type II exponent of the sequential scheme is the summation of
the type II exponent of each scheme.
As in Section \ref{section:product-BSDS}, suppose that $p_1=q_2$ and $p_2=q_1$. Furthermore, let $h(q_1) < h(p_1)$. Then,
from the above argument,
and \eqref{eq:b-Exponent-BSDS}, we find that the following exponent is attainable by the sequential scheme:
\begin{align}
\tilde{E}_{\mathtt{b}}(R) = D(p_1 \| q_1) + \min[ |R_2 - h(p_2)|^+, D(p_2 \| q_2) ]
\end{align}
for $R_1 \ge h(p_1)$ and $R_2 \ge h(p_2)$.
By setting $R_1 = h(p_1)$ and $R_2 = h(p_2) + D(p_2 \| q_2)$, we can derive the following bound on the critical rate:
\begin{align} \label{eq:critical-sequential-BSDS}
R_{\mathtt{cr}} \le h(p_1) + h(p_2) + D(p_2 \| q_2).
\end{align}
Interestingly, the bound on the critical rate in \eqref{eq:critical-sequential-BSDS} improves upon
the bound in \eqref{eq:critical-product-SHA}.
From an operational point of view, this improvement
can be explained as follows. In the SHA scheme (without quantization), we use the minimum entropy decoder to compute an
estimate $\hat{\bm{x}}$ of $X^n =\bm{x}$; then, if the joint type $\san{t}_{\hat{\bm{x}} \bm{y}}$ is close to $P_{XY}$, we accept the null
hypothesis. When the empirical conditional entropy $H(\bm{x}|\bm{y})$ is such that $H_P(X|Y) \gnsim H(\bm{x}|\bm{y})$,
even if there exists $\hat{\bm{x}}$ satisfying $H(\bm{x}|\bm{y}) \ge H(\hat{\bm{x}}|\bm{y})$, the type II testing error does not occur
though the receiver may erroneously compute $\hat{\bm{x}}\neq \bm{x}$.\footnote{Note that $H_P(X|Y) \gnsim H(\bm{x}|\bm{y}) \ge H(\hat{\bm{x}}|\bm{y})$
implies that $\san{t}_{\hat{\bm{x}} \bm{y}}$ is not close to $P_{XY}$.}
Consequently, the type II testing error may occur
only if $(X^n,Y^n) = (\bm{x},\bm{y})$ satisfies $H(\bm{x}|\bm{y}) \gtrsim H_P(X|Y)$.
For the product of BSDSs, when the types of modulo sums $\bm{x}_1 + \bm{y}_1$ and $\bm{x}_2 + \bm{y}_2$
are $H(\san{t}_{\bm{x}_1 + \bm{y}_1}) \simeq h(\tau_1)$ and $H(\san{t}_{\bm{x}_2+\bm{y}_2}) \simeq h(\tau_2)$,
the type II testing error may occur only if $h(\tau_1) + h(\tau_2) \gtrsim h(p_1) + h(p_2)$; see Fig. \ref{Fig:range-product1}(a).
In the modified SHA scheme, $\bm{x}_1$ and $\bm{x}_2$ are binned and estimated separately. Because of this,
the constraint of type II testing error event become more stringent, i.e., $h(\tau_1) \gtrsim h(p_1)$ and $h(\tau_2) \gtrsim h(p_2)$; see Fig.~Fig. \ref{Fig:range-product1}(b).
\begin{figure}[t]
\centering{
\subfloat[][]{\includegraphics[width=.4\textwidth]{./Range-of-Error-Event-Product1.eps} \label{Fig:range-product1}} \quad
\subfloat[][]{\includegraphics[width=.4\textwidth]{./Range-of-Error-Event-Product2.eps} \label{Fig:range-product2}}
\caption{The meshed area describes the range of $h(\tau_1)$ and $h(\tau_2)$ such that the type II testing error may occur when
$H(\san{t}_{\bm{x}_1+\bm{y}_1}) \simeq h(\tau_1)$ and $H(\san{t}_{\bm{x}_2+\bm{y}_2}) \simeq h(\tau_2)$ for (a) SHA scheme and (b) modified SHA scheme.}
\label{Fig:range-product1}
}
\end{figure}
\section{Discussion}
In this paper, we have developed tools to evaluate the critical rate attainable by the SHA scheme,
and exemplified that the SHA scheme is sub-optimal for a product of BSDSs.
A future problem is to generalize the idea of sequential scheme in Section \ref{section:sequential}
and develop a new achievability scheme for the distributed hypothesis testing.
Another future problem is a potential utility of the scheme in \cite{WeiKoc:19};
as we mentioned in Section \ref{section:introduction}, it has a potential to improve upon the SHA scheme,
but strict improvement is not clear since the expression involves complicated optimization over multiple parameters.
Lastly, it should be pointed out that certain kinds of ``structured" random coding is known to improve upon the naive random coding
in the context of the mismatched decoding problem; eg.~see \cite{Lapidoth96, ScaBarMarFab15} and references therein.
Even though there is no direct connection between the two problems,
it might be interesting to pursue a mathematical connection between these problems.
\section*{Acknowledgment}
The author appreciates Marat Burnashev for valuable discussions at an early stage of the work.
The author also thanks Te Sun Han and Yasutada Oohama for fruitful discussions.
\newpage
\bibliographystyle{../../09-04-17-bibtex/IEEEtran}
|
\section{Introduction}\label{sec:introduction}
A \emph{partially ordered set} (abbreviated \emph{poset}) is an ordered pair $(P,\le_P)$ consisting of a set $P$ and a binary relation $\le_P$ on $P$ such that for all $a,b,c\in P$ we have \begin{itemize}
\item $a\le_P a$,
\item if $a\le_P b$ and $b\le_P a$, then $a=b$, and
\item if $a\le_P b$ and $b\le_P c$, then $a\le_P c$.
\end{itemize}
We will refer to a poset $(P,\le_P)$ by just $P$ if the order $\le_P$ is clear from context. In such cases we may also write $\le$ instead of $\le_P$. We will also write $a<b$ to denote $a\le b$ and $a\ne b$. Finally, we will only work with finite posets unless explicitly stated otherwise. For a finite poset $P$, we denote the cardinality of its underlying set by $\abs{P}$.
For a poset $P$ with elements $a$ and $b$, we say that $a$ and $b$ are \emph{comparable} if $a\le b$ or $b\le a$. Otherwise, we say that $a$ and $b$ are \emph{incomparable}. The simplest example of a poset is one in which all pairs of elements are comparable. We call such a poset a \emph{chain} or \emph{linear order} (also known as a \emph{total order}).
A more complicated example of a poset is the divisibility poset $\D_{[6]}$, which consists of the set $[6]=\{1,2,3,4,5,6\}$ with the relation $a\le b$ if $b$ is divisible by $a$. Note that some pairs of elements are incomparable, such as $5$ and $6$. This poset is depicted below, where $a\le b$ if and only if $a$ is reachable from $b$ by a sequence of downward arrows.
\begin{center}
\includegraphics[scale=0.9,trim=1in 7.65in 1in 1.65in,clip]{diagram1.pdf}
\end{center}
In general, for a poset $P$, we call such a diagram the \emph{Hasse diagram} of $P$. The arrows are defined by \emph{covering relations}: $b$ \emph{covers} $a$ if $a < b$ and there is no $c$ such that $a<c<b$.
Given a poset $P$, a natural question is to ask how complicated $P$ is. A notion such as the set cardinality of $P$ would not be a good answer, since it does not consider the relation $\le_P$. For example, a linear order on $[6]$ is intuitively much simpler than $\D_{[6]}$. A potential approach is to try to understand how complicated the Hasse diagram of $P$ is. Specifically, we could ask for the least $d$ such that the Hasse diagram of $P$ can be viewed ``nicely'' in $d$-dimensional space. For instance, the linear order on $[6]$ can be viewed as a line, while $D_{[6]}$ can't. We formalize this idea of dimension using some more definitions.
Given posets $P$ and $Q$, we say that $P$ \emph{embeds} into $Q$, written $P\into Q$, if $P$ is a subset of $Q$ and $\le_P$ is the restriction of $\le_Q$ to $P$. We also say that $P$ is a \emph{suborder} of $Q$.
This notion of poset containment allows us to relate a poset to $d$-dimensional space. We just need to think of $d$-dimensional space as a poset. For this, we define the product of posets.
Given posets $P$ and $Q$, we define the \emph{product poset} $P\times Q$ to be a poset on the set product $P\times Q$ with the relation $(p_1,q_1)\le (p_2,q_2)$ if and only if $p_1\le_Pp_2$ and $q_1\le_Qq_2$. We can also take the product of several posets by iterating this definition.
Now, we can think of $d$-dimensional space as the poset $\R^d$, where we take the usual linear order on $d$ copies of $\R$ and then take a product. More generally, we can think of $d$-dimensional space as a product of any $d$ chains. Since we will only care about finite posets, we can work with finite chains as well. We are now ready to define the dimension of a poset.
The \emph{Dushnik-Miller dimension} of a poset $P$, denoted $\dim(P)$, is the least $d\in\N$ for which $P$ embeds into a product of $d$ chains. Note that dimension exists for all finite posets since we can always embed $P$ into a product of $\abs{P}$ chains. This notion of dimension was introduced by Dushnik and Miller in 1941 \cite{Dushnik-Miller-paper} and has been extensively studied (see \cite{Trotter-book}). Other notions of dimension exist (see \cite{Nesetril-Pudlak-paper,Novak-paper}), but we will only discuss the Dushnik-Miller dimension of posets.
As an example, any linear order has dimension $1$. A more interesting example is the poset $\D_{[6]}$ from above. One can check that $\dim(\D_{[6]})>1$. Furthermore, $\dim(\D_{[6]})\le2$ by the following embedding into a $4$ by $3$ grid (which is a product of chains of length $4$ and $3$).
\begin{center}
\includegraphics[scale=0.9,trim=1in 7.35in 1in 1.7in,clip]{diagram2.pdf}
\end{center}
A more general example is the boolean hypercube $\Q^n$. We view $\Q^n$ as the poset on subsets of $[n]$, ordered by inclusion (i.e., the ``subset'' relation). Since $\Q^n$ is generally thought of as an $n$-dimensional object, we would expect it to have dimension $n$. In fact, because $\Q^n$ is isomorphic to a product of $n$ chains of length $2$, we have that $\dim(\Q^n)\le n$. We will see in \cref{sec:prelimaries} that indeed $\dim(\Q^n)=n$. Things get more interesting when we consider suborders of $\Q^n$.
Note that $\Q^n$ has a natural partition into ``layers,'' where we partition subsets of $[n]$ by their cardinality. The Hasse diagram of $\Q^n$ can be drawn with each layer being on a horizontal line. So, a natural way to take a suborder of $\Q^n$ is to just consider some subset of its layers. For $I\subseteq [0,n]$, we let $\Q^n_I$ be the suborder of $\Q^n$ on subsets $S$ of $[n]$ for which $\abs{S}\in I$. The dimension of such suborders has been studied in detail.
We will see in \cref{sec:prelimaries} that for $0<k<\ell<n$, $\dim(\Q^n_{\{k,\ell\}})=\dim(\Q^n_{[k,\ell]})$. That is, if we take a suborder on some layers of the hypercube, it only matters what the lowest and highest layers are. In particular, to understand $\dim(\Q^n_I)$ for arbitrary $I$ we just need to understand $\dim(\Q^n_{\{k,\ell\}})$. Several results are known, and we present some of them here. The case $k=1$ was studied first. Dushnik gave an exact formula for $\dim(\Q^n_{[1,\ell]})$ when $\ell>2\sqrt{n}$ in \cite{Dushnik-paper}. In particular, we have the following theorem.
\begin{theorem}[Dushnik]
If $\ell>2\sqrt{n}$, then $\dim(\Q^n_{[1,\ell]})>n-\sqrt{n}.$
\end{theorem}
By using similar techniques, one can obtain the following corollary \cite{Trotter-book}.
\begin{corollary}[Trotter]
If $\ell<\sqrt{n},$ then $\dim(\Q^n_{[1,\ell]})>\frac{\ell^2}{4}.$
\end{corollary}
Together, these results show that
$\dim(\Q^n_{[1,\ell]})$ grows at least quadratically in $\ell$ until it gets close to the upper bound of $\dim(\Q^n_{[1,n]})=n$.
A result by F\"{u}redi and Kahn shows that this quadratic growth is correct up to a $\log n$ factor \cite{Furedi-Khan-paper}. Note that we use $\log$ to denote the natural logarithm throughout this paper.
\begin{theorem}[F\"{u}redi--Kahn]\label{thm:weak-subsets}
For integers $1\le \ell\le n$, we have $\dim(\Q^n_{[1,\ell]})\le(\ell+1)^2\log n.$
\end{theorem}
For the case of general $k$, the lower bounds for $k=1$ also imply that $\dim(\Q^n_{[k,\ell]})$ grows at least quadratically in $(\ell-k)$, by noting that $\Q^n_{[k,\ell]}$ contains a copy of $\Q^{n-k+1}_{[1,\ell-k+1]}$
In 1994, Brightwell, Kierstead, Kostochka, and Trotter gave the following upper bound \cite{BKKT-paper}.
\begin{theorem}[Brightwell--Kierstead--Kostochka--Trotter]\label{thm:subsets}
Let $0\le k< \ell\le n$ be integers. Then we have $\dim(\Q^n_{[k,\ell]})=O((\ell-k)^2\log n).$
\end{theorem}
This shows that $\dim(\Q^n_{[k,\ell]})$ grows quadratically in $(\ell-k)$, up to a $\log n$ factor. These bounds are relatively tight when $\ell-k$ is large compared to $\log n$. For results on the case when $\ell-k$ is smaller, see \cite{Kierstead-paper,Spencer-paper}.
A natural extension of this problem on subsets is to instead consider multisets of $[n]$. Let $\M^n$ be the (infinite) poset on all multisets of $[n]$, ordered by inclusion. As before, for $I\subseteq \N$, let $\M^n_I$ be the suborder of $\M^n$ on multisets $S$ of $[n]$ for which $\abs{S}\in I$ (in this paper $I$ will always be finite). Here $\abs{S}$ counts elements of $S$ with multiplicity. As before, for $0<k<\ell,$ we have $\dim(\M^n_{\{k,\ell\}})=\dim(\M^n_{[k,\ell]})$. We will see why in \cref{sec:prelimaries}.
Our first new result is the following extension of \cref{thm:subsets}.\footnote{We have not optimized the constants in new results in this paper. We also omit some floors and ceilings.}
\begin{theorem}\label{thm:multisets} Let $k,\ell,n$ be positive integers with $k< \ell$. Then we have
$$\dim(\M_{[k,\ell]}^n)\le 34(\ell-k)^2\log n.$$
\end{theorem}
Next, we consider a problem posed by Lewis and Souza in 2021. For a set $R$ of positive integers, let $\D_R$ be the poset on $R$, ordered by divisibility. Earlier we saw an example with $R=[6]$ and showed that $\dim(\D_{[6]})=2$. More generally, one can ask how $\dim(\D_R)$ behaves in terms of $R$. A natural choice for $R$ is the interval $[N]$. Lewis and Souza (with an improvement to the upper bound due to Souza and Versteegen) essentially solved this case \cite{Lewis-Souza-paper, improved-divisibility}.
\begin{theorem}[Lewis--Souza, Souza--Versteegen] For $N\in\N$, we have
$$\left(\frac{1}{16}-o(1)\right)\frac{(\log N)^2}{(\log\log N)^2}\le \dim(\D_{[N]})\le \left(\frac{4}{\log2}+o(1)\right)\frac{(\log N)^2\log\log\log N}{(\log\log N)^2}.$$
\end{theorem}
Another interesting choice of $R$ is $R=[N/\kappa,N]$, where $\kappa> 1$ is a fixed real number and $N$ is allowed to vary. Lewis and Souza noted that $\D_{[1,\kappa]}$ embeds into $\D_{[N/\kappa,N]}$, which gives a lower bound of $$\dim(\D_{[N/\kappa,N]})=\Omega((\log \kappa)^2(\log\log \kappa)^{-2}).$$
Lewis and Souza also showed that $\dim(\D_{[N/\kappa,N]})$ is bounded above by $\kappa (\log\kappa)^{1+o(1)}$, using a general result by Scott and Wood \cite{Scott-Wood-paper}. Note that this is a function of only $\kappa$ and not $N$.
One of our main results in this paper is to significantly improve this upper bound with the following theorem.
\begin{theorem}\label{thm:divisibility} Let $\kappa>1$ be a real number and let $N\in \N$. Then we have
$$\dim(\D_{[N/\kappa,N]})\le \max\left(688\frac{(\log\kappa)^3}{(\log\log \kappa)^2},2\right).$$
\end{theorem}
The $2$ in the $\max$ function handles the case when $\kappa$ is close to $1$. To prove this theorem, we first generalize \cref{thm:multisets}. Note that $\M^n$ is isomorphic to $\Z_{\ge0}^n$, by identifying a multiset $S$ with its tuple of multiplicities $\x(S)$. Additionally, we can identify an integer with the tuple of exponents in its prime factorization. However, when we consider the size of an integer, each prime has a different weight. Thus, to use multisets to capture an interval of integers, we need to weight the elements of $[n]$ in our multisets. So, for a weight vector $\vv\in\R_{>0}^n$, we define the \emph{$\vv$-size} of a multiset $S$ to be $\abs{S}_{\vv}=\x(S)\cdot \vv$. We also define
$$\M_{[k,\ell]}^{n,\vv}=\{S\in\M^n \st \abs{S}_{\vv} \in [k,\ell]\}.$$
Here we allow $k$ and $\ell$ to be real numbers, as $\abs{S}_{\vv}$ may not be an integer. Note that when all entries of $\vv$ are $1$ and $k,\ell$ are integers, $\M_{[k,\ell]}^{n,\vv}=\M_{[k,\ell]}^{n}$.
To state our result on weighted multisets, we first need a technical definition to describe the dependence on $\vv$. For a vector $\vv\in\R_{>0}^n$ and a real $s$, we let $m(\vv,s)$ be the sum of the least $\floor{s}$ coordinates of $\vv$. Then we have the following result for weighted multisets, which generalizes \cref{thm:multisets}.
\begin{theorem}\label{thm:weighted-multisets}
Let $0\le k<\ell$ be real numbers and let $n\in \N$. Let $\vv\in \R_{>0}^n$ be a weight vector. If $r\ge1$ is a real number such that $m(\vv,r)\ge 2(\ell-k)$, then
$$\dim(\M_{[k,\ell]}^{n,\vv})\le 43r^2\log n.$$
\end{theorem}
The key new idea in this paper is a way to deal with arbitrary weight vectors $\vv$ in the proof of \cref{thm:weighted-multisets}. We will prove \cref{thm:divisibility} by applying \cref{thm:weighted-multisets}.
\cref{thm:weighted-multisets} can also be used to analyze other divisibility orders. For example, consider divisibility orders on polynomials. For a prime power $q$ and integers $d_0\ge0$, $\delta>0$, let $P(\F_q)_{[d_0-\delta,d_0]}$ denote the poset on monic polynomials in $\F_q[x]$ with degree in the interval $[d_0-\delta,d_0]$, ordered by divisibility. Using \cref{thm:weighted-multisets}, we obtain the following result.
\begin{theorem}\label{thm:polynomial} Let $q$ be a prime power and let $\delta\in \N$. Then for each nonnegative integer $d_0$,
$$\dim(P(\F_q)_{[d_0-\delta,d_0]})\le \min\left(910\frac{(\delta\log q)^3}{(\log \delta)^2},172\delta^3\log q\right).$$
\end{theorem}
The two bounds in this theorem correspond to the two regimes when $q<\delta$ and $q\ge \delta$.
The paper is organized as follows. In \cref{sec:prelimaries} we will discuss some basic properties of poset dimension that will be useful in the proofs of the main results. In \cref{sec:multisets} we analyze unweighted multiset posets and prove \cref{thm:multisets}. In \cref{sec:weighted-multisets} we extend our arguemnts to weighted multiset posets and prove \cref{thm:weighted-multisets}. In \cref{sec:divisibility} we prove \cref{thm:divisibility} by applying \cref{thm:weighted-multisets}. In \cref{sec:polynomial}, we discuss other divisibility orders and prove \cref{thm:polynomial}. In \cref{sec:further-directions} we discuss further directions.
\section{Preliminaries}\label{sec:prelimaries}
A \emph{linear extension} of a poset $P$ is a total (linear) order on the elements of $P$ that agrees with all relations in $P$. Given a set $\LL$ of linear extensions of $P$, we say that $\LL$ is a \emph{realiser} if for each pair\footnote{We will refer to ordered pairs as pairs when the order is clear.} of incomparable elements ($x,y$), there exists an extension $L\in\LL$ such that $x\ge_L y$. Then $\dim(P)$ is also the minimum size of a realiser of $P$. One way to see this is by embedding $P$ into $\R^d$ and constructing a linear extension from each of the $d$ coordinates.
We now discuss some basic properties of dimension. Given two posets $P$ and $Q$, if $P$ embeds into $Q$, then $\dim(P)\le \dim(Q)$.
An important concept in the study of poset dimensions is the notion of critical pairs. A \emph{critical pair} in a poset $P$ is a pair of incomparable elements $(x,y)$ such that $x$ is minimal among elements incomparable to $y$ and $y$ is maximal among elements incomparable to $x$. We say that a linear extension $L$ of $P$ \emph{reverses} a critical pair $(x,y)$ in $P$ if $x\ge_Ly$.
\begin{proposition}
Let $\LL$ be a set of linear extensions of a (finite) poset $P$. Suppose that each critical pair in $P$ is reversed by some extension in $\LL$. Then $\LL$ is a realiser of $P$.
\end{proposition}
\begin{proof}
Let $(x,y)$ be any pair of incomparable elements in $P$. As long as $(x,y)$ is not a critical pair, we can decrease $x$ or increase $y$ while keeping the pair incomparable. Since $P$ is finite, this means we can find a critical pair $(x',y')$ such that $x\ge_P x'$ and $y\le_P y'$. Now there exists some $L\in\LL$ such that $x'\ge_Ly'$, so we also have $x\ge_Ly$ as desired.
\end{proof}
Using the concept of a critical pair, we can see why we only care about the lowest and highest layers of a suborder of $\Q^n$. Let $0<k<\ell<n$ and consider the poset $\Q^n_{[k,\ell]}$. In this poset the critical pairs are precisely the incomparable pairs of subsets $(X,Y)$ where $\abs{X}=k$ and $\abs{Y}=\ell$. So to construct a realiser for $\Q^n_{[k,\ell]}$ it suffices to construct a realiser for the suborder $\Q^n_{\{k,\ell\}}$. Thus we have that $\dim(\Q^n_{[k,\ell]})\le\dim(\Q^n_{\{k,\ell\}})$. Additionally, since $\Q^n_{\{k,\ell\}}$ is a suborder of $\Q^n_{[k,\ell]}$, we have that $\dim(\Q^n_{\{k,\ell\}})\le\dim(\Q^n_{[k,\ell]})$. Thus $\dim(\Q^n_{\{k,\ell\}})=\dim(\Q^n_{[k,\ell]})$. Similarly, for $0<k<\ell$, we have that $\dim(\M^n_{\{k,\ell\}})=\dim(\M^n_{[k,\ell]})$. Although we will not need this fact in the proofs of new results in this paper, it is helpful to keep in mind.
Critical pairs also explain why $\dim(\Q^n)=n$. The critical pairs in $Q^n$ are precisely the pairs $(\{x\},[n]\setminus\{x\})$ for $x\in [n]$. There are $n$ such critical pairs and a linear extension of $\Q^n$ can reverse at most one of them. Thus $\dim(\Q^n)\ge n$. We already know that $\dim(\Q^n)\le n$, so $\dim(\Q^n)=n$.
Next we consider some basic poset constructions. Given posets $P$ and $Q$, recall the product poset $P\times Q$, where $(p_1,q_1)\le (p_2,q_2)$ if and only if $p_1\le_Pp_2$ and $q_1\le_Qq_2$. Then we have that $\dim(P\times Q)\le \dim(P)+\dim(Q)$ by combining embeddings of $P$ and $Q$ into an embedding of $P\times Q$.
Another poset construction we will use is the disjoint union of two posets. Given posets $P$ and $Q$, let $P\sqcup Q$ be their disjoint union. Here we take the union of a copy of $P$ and a copy of $Q$ with all elements of $P$ being incomparable with all elements of $Q$. As long as $P$ and $Q$ are nonempty, $P\sqcup Q$ cannot be a chain, so $\dim(P\sqcup Q)\ge2$. Next, let $d=\max(\dim(P),\dim(Q),2)$ and note that both $P$ and $Q$ can be embedded into $\R^d$. By appropriately translating these embeddings, we can construct an embedding of $P\sqcup Q$ into $\R^d$. So $\dim(P\sqcup Q)=\max(\dim(P),\dim(Q),2)$.
Next, to illustrate some of the techniques used in the proof of \cref{thm:divisibility}, we will prove the following easier result, which is already a small improvement upon the previous best known bound.
\begin{proposition}\label{prop:easy-divisibility} Let $\kappa>1$ and $N\in \N$. Let $\pi(\kappa)$ denote the number of primes at most $\kappa$. Then we have
$$\dim(\D_{[N/\kappa,N]})\le \max(\pi(\kappa),2)=(1+o(1))\kappa(\log \kappa)^{-1}.$$
\end{proposition}
\begin{proof}
The key idea is to write $\D_{[N/\kappa,N]}$ as a disjoint union of several posets.
Call a prime $p$ \emph{small} if $p\le \kappa$ and \emph{large} if $p>\kappa$. Let $K$ be the set of all positive integers $M\le N$ with no small primes dividing $M$. For each $M\in K$, let $g(M)$ be the set of all positive integers of the form $Mq$, where $q$ has only small prime divisors. Now, note that we have a set partition of $[N/\kappa,N]$ given by $$[N/\kappa,N]=\bigsqcup_{M\in K} g(M)\cap [N/\kappa,N].$$
Furthermore, elements of different sets in this partition are incomparable in $\D_{[N/\kappa,N]}$, because if they
were comparable, their ratio would be divisible by a large prime. Thus $\D_{[N/\kappa,N]}$ is a disjoint union of posets of the form $\D_{g(M)\cap [N/\kappa,N]}$, and
it suffices to show that $\dim(\D_{g(M)\cap [N/\kappa,N]})\le \pi(\kappa)$ for any $M\in K$.
For any given $M\in K$, we can embed $\D_{g(M)\cap [N/\kappa,N]}$ into $\R^{\pi(\kappa)}$ by mapping an integer $m$ to its tuple of exponents of small primes. Applying the prime number theorem completes the proof.
\end{proof}
The main result of this paper is to provide a nontrivial bound on $\dim(\D_{g(M)\cap [N/\kappa,N]})$ using new results on multisets. We first review some notation for multisets.
When referring to multisets, we always count size/cardinality with multiplicity. We denote the set of distinct elements of a multiset $S$ by $\supp(S)$, so that $\dist{S}$ is the number of distinct elements of $S$. For multisets $S$ and $T$, we use $S\setminus T$ to denote the multiset where we subtract the multiplicity of an element in $T$ from its multiplicity in $S$ (without going below $0$). For example, $\supp(S\setminus T)$ is the set of elements which have greater multiplicity in $S$ than in $T$. Also, we write $S\subseteq T$ if the multiplicity of each element in $S$ is at most its multiplicity in $T$.
\section{Results on Multisets}\label{sec:multisets}
In this section we prove \cref{thm:multisets}, which is a modification of the main result of Brightwell et al. \cite{BKKT-paper}.
To prove this theorem, we will construct two sets of linear extensions of $\M_{[k,\ell]}^n$. The first set of extensions, $\LL_1$, will deal with all incomparable pairs of multisets $(S,T)$ for which $\dist{T\setminus S}\le 3(\ell-k).$ The second set of extensions, $\LL_2$, will deal with all incomparable pairs of multisets $(S,T)$ for which $\dist{T\setminus S}>3(\ell-k).$
\begin{lemma}\label{lem:L1}
Let $n$ be a positive integer and let $r$ be a positive real number with $r\ge 1$. Then there exists a set $\LL_1$ of at most $(3r+1)^2\log n$ linear extensions of $\M^n$ such that for every incomparable pair of multisets $(S,T)$ with $\dist{T\setminus S}\le 3r,$ there exists an extension $L\in\LL_1$ where $S>_LT$.
\end{lemma}
\begin{proof}
The proof uses the probabilistic method, with an argument similar to the proof of \cref{thm:weak-subsets} in \cite{Furedi-Khan-paper}.
Note that the result is clear if $n\le 3r+1$, so we will assume $n>3r+1.$
For a total order $\sigma$ on $[n]$, we define the lexicographic linear extension $L_\sigma$ of $\M^n$. Specifically, for $S,T\in \M^n$, we have $S>_{L_\sigma}T$ if $\max_\sigma(\supp(S\setminus T) \cup \supp(T\setminus S)) \in \supp(S\setminus T)$.
For a subset $Y$ of $[n]$ and an element $x\in[n]\setminus Y$, we say $x>_{\sigma}Y$ if $x>_{\sigma}y$ for all $y\in Y$. We would like to construct a set of at most $d=(3r+1)^2\log n$ choices of $\sigma$ such that for every $3r$-element subset $Y\subseteq[n]$ and $x\in[n]\setminus Y$, one of our $d$ choices of $\sigma$ gives $x>_{\sigma}Y$.
To accomplish this, consider sampling $d$ choices of $\sigma$ uniformly and independently at random from all possible total orders. For a given $3r$-element subset $Y\subseteq[n]$ and $x\in[n]\setminus Y$, the probability that $x>_{\sigma}Y$ in a given sample is $\frac{1}{3r+1}$. Thus, in $d$ samples, the probability that $x\not>_{\sigma}Y$ for any of the sampled $\sigma$ is $\left(1-\frac{1}{3r+1}\right)^d.$ The number of possible choices for $(x,Y)$ is at most $n^{3r+1}$. Thus the expected number of pairs $(x,Y)$ for which $x\not>_{\sigma}Y$ for all $d$ choices of $\sigma$ is at most $$n^{3r+1}\left(1-\frac{1}{3r+1}\right)^d<n^{3r+1}e^{-d/(3r+1)}=1 .$$
So, we can choose $d$ total orders $\sigma_1,\dots,\sigma_d$ such that for every choice of $(x,Y)$, $x>_{\sigma_i}Y$ for some chosen $\sigma_i$. Now, let $\LL_1=\{L_{\sigma_i}\colon 1\le i\le d\}$ be the set of lexicographic linear extensions corresponding to these choices of $\sigma$.
Consider an incomparable pair of multisets $(S,T)$ in $\M^n$ with $\dist{T\setminus S}\le 3r$. Choose any $x\in \supp(S\setminus T)$ and any $3r$-element subset $Y$ of $[n]\setminus\{x\}$ containing $\supp(T\setminus S)$. Then we have that $x>_{\sigma_i}Y$ for some chosen $\sigma_i$. In the corresponding linear extension $L_{\sigma_i}$, we will have $S>_{L_{\sigma_i}}T$, as desired.
\end{proof}
Note that \cref{lem:L1} holds for any value of $r$, but we will apply it with $r=\ell-k$. Additionally, note that since $\M_{[k,\ell]}^n$ is a suborder of $\M^n$, we can restrict the linear extensions we obtain from \cref{lem:L1} to linear extensions of $\M_{[k,\ell]}^n$ with the same property.
We construct $\LL_2$ from (ordered) partitions of $[n]$. We can represent a partition of $[n]$ with $a$ parts by a function $f\colon [n]\to [a]$. More generally, we represent a sequence of $t$ such partitions by a function $f\colon [n]\times[t]\to [a]$.
Working with these partitions instead of $[n]$ allows us to have a smaller object to deal with. However, we also need our sequence of partitions to remember some of the structure of $[n]$. This is encoded with the following property.
\begin{definition}[\cite{BKKT-paper}]
A function $f\colon [n]\times[t]\to [a]$ is $(a,b,r,t,n)$-good if for each subset $X\subseteq [n]$ of size $\abs{X}=b$, there exists $\tau\in[t]$ such that $\abs{f(X,\tau)}>r$.
\end{definition}
In other words, for each subset $X$ of $[n]$ with $b$ elements, we need to be able to choose one of our $t$ partitions that divides $X$ into more than $r$ parts. The following lemma guarantees the existence of good functions for appropriately chosen parameters.
\begin{lemma}[\cite{BKKT-paper}]\label{lem:good-func}
Let $a,b,r,t,n$ be positive integers such that $r<b\le n$ and $r<a$. If $$\binom{n}{b}e^{rt}(r/a)^{(b-r)t}<1,$$ then there exists an $(a,b,r,t,n)$-good function.
\end{lemma}
The proof of this lemma is by the probabilistic method; we choose $f$ uniformly at random. See \cite[Lemma 2.2]{BKKT-paper} for a full proof.
We are now ready to construct $\LL_2$. The key idea is to focus on a subset $R\subseteq[n]$ (which we will choose later) and use it to order our multisets by counting only elements of $R$ (with multiplicity). The goal is to find a collection $\mathcal{R}$ of several different subsets of $[n]$ such that for any incomparable pair of multisets $(S,T)$ with $\supp(T\setminus S)> 3(\ell-k)$, there exists an $R\in \mathcal{R}$ inducing an order with $S>T$. We accomplish this by taking several partitions of $[n]$, corresponding to a good function given by \cref{lem:good-func}. The fact that $T$ can have at most $\ell-k$ more elements than $S$ (since $k\le \abs{S},\abs{T}\le \ell$) will allow us to find a suitable collection $\mathcal{R}$ of relatively small size.
We will find $\mathcal{R}$ using a good function. Specifically, for an $(a,b,r,t,n)$-good function $f$ and $\alpha\in[a],\tau\in[t]$, we let $R_{\alpha,\tau}=\{i\in [n] \mid f(i,\tau)=\alpha\}$. That is, $R_{\alpha,\tau}$ denotes part $\alpha$ of partition $\tau$ in the sequence of partitions represented by $f$. We then let $\mathcal{R}=\{R_{\alpha,\tau} \mid \alpha\in[a],\tau\in[t]\}$.
\begin{lemma}\label{lem:L2}
Let $n,k,\ell,r$ be positive integers with $k<\ell$ and $r=\ell-k$. Then there exists a set $\LL_2$ of at most $18r\log n$ linear extensions of $\M_{[k,\ell]}^n$ such that for every incomparable pair of multisets $(S,T)$ with $\dist{T\setminus S}> 3r,$ there exists an extension $L\in\LL_2$ where $S>_LT$.
\end{lemma}
\begin{proof}
If $3r>n$, then $\dim (\M_{[k,\ell]}^n)\le n<18r\log n$. So assume $3r\le n$ and set $a=b=3r$, $t=3\log n$. Note that the condition in \cref{lem:good-func} holds, since
$$\binom{n}{b}e^{rt}(r/a)^{(b-r)t}=\binom{n}{3r}e^{rt}3^{-2rt}<n^{3r}e^{-rt}=1.$$
So, we fix an $(a,b,r,t,n)$-good function $f$. Now we will use $R_{\alpha,\tau}$ as described above.
For each $\alpha\in[a],\tau\in[t],j\in[2]$, we will construct a linear extension $L_{\alpha,\tau,j}$ of $\M_{k,l}^n$. That is, we will construct two extensions for each part in each partition. Our set of extensions $\LL_2$ will consist of these $2at$ extensions.
Fix two linear extensions $M_1$ and $M_2$ of $\M^n$ which both order multisets by size, but order the multisets of a given size in opposite orders. Also let $M_0$ be an arbitrary linear extension of $\M^n$.
For a multiset $S$, let $S_{\alpha,\tau}$ be the multiset obtained by restricting $S$ to the elements of $R_{\alpha,\tau}$, i.e., $S_{\alpha,\tau}=\{i\in S \mid f(i,\tau)=\alpha\}$. Note that $S_{\alpha,\tau}$ keeps the same multiplicities of elements as $S$.
We now construct our linear extensions $L_{\alpha,\tau,j}$. We let $S<_{L_{\alpha,\tau,j}}T$ if $S_{\alpha,\tau}<_{M_j}T_{\alpha,\tau}$. If $S_{\alpha,\tau}=T_{\alpha,\tau}$ then we let $S<_{L_{\alpha,\tau,j}}T$ if and only if $S<_{M_0}T$.
We now show that $\LL_2$ satisfies the desired condition. Suppose that $(S,T)$ is an incomparable pair of multisets in $\M_{[k,\ell]}^n$ with $\dist{T\setminus S}>3r=b$. Since $f$ is $(a,b,r,t,n)$-good, there exists $\tau\in[t]$ such that $\abs{f(T\setminus S,\tau)}>r$\footnote{We let $f(X,\tau)=f(\supp(X),\tau)$ for multisets $X$.}.
If there exists $\alpha\in[a]$ such that $\abs{S_{\alpha,\tau}}>\abs{T_{\alpha,\tau}}$, then $S>T$ in both $L_{\alpha,\tau,1}$ and $L_{\alpha,\tau,2}$. So assume that $\abs{S_{\alpha,\tau}}\le\abs{T_{\alpha,\tau}}$ for all $\alpha\in[a]$. Next, note that
$$\sum_{\alpha\in[a]}\abs{S_{\alpha,\tau}}=\abs{S},\sum_{\alpha\in[a]}\abs{T_{\alpha,\tau}}=\abs{T}.$$
Recall that $k\le \abs{S},\abs{T}\le \ell$, so $\abs{T}-\abs{S}\le \ell-k=r.$ Hence there are at most $r$ values of $\alpha\in[a]$ for which $\abs{S_{\alpha,\tau}}<\abs{T_{\alpha,\tau}}$. Thus there exists $\alpha\in f(T\setminus S,\tau)$ with $\abs{S_{\alpha,\tau}}=\abs{T_{\alpha,\tau}}$. Additionally, we can't have $S_{\alpha,\tau}=T_{\alpha,\tau}$ since $\alpha\in f(T\setminus S,\tau)$. So $S>T$ in either $L_{\alpha,\tau,1}$ or $L_{\alpha,\tau,2}$, and we are done.
\end{proof}
Applying \cref{lem:L1} with $r=\ell-k$ and \cref{lem:L2}, we have that $$\dim(\M_{[k,\ell]}^n)\le (3r+1)^2\log n+18r\log n \le 34(\ell-k)^2\log n.$$
This proves \cref{thm:multisets}.
\section{Weighted Multisets}\label{sec:weighted-multisets}
In this section we prove \cref{thm:weighted-multisets} using the same proof strategy as for \cref{thm:multisets}. This time we are given a value of $r$ which depends somewhat on $\ell-k$. We will again construct two sets of extensions of $\M_{[k,\ell]}^{n,\vv}$. The first set of extensions, $\LL_1$, will deal with all incomparable pairs of multisets $(S,T)$ for which $\dist{T\setminus S}\le 3r$, and the second set of extensions, $\LL_2$, will deal with all incomparable pairs of multisets $(S,T)$ for which $\dist{T\setminus S}>3r.$
To construct $\LL_1$, we apply \cref{lem:L1} and restrict the resulting linear extensions to $\M_{[k,\ell]}^{n,\vv}$. This gives us at most $(3r+1)^2\log n$ linear extensions of $\M_{[k,\ell]}^{n,\vv}$ such that for every incomparable pair of multisets $(S,T)$ with $\dist{T\setminus S}\le 3r,$ there exists an extension $L\in\LL_1$ where $S>_LT$.
To construct $\LL_2$, we will use the same strategy as in the proof of \cref{lem:L2}. However, we will need some new ideas. Previously, for an incomparable pair of multisets $(S,T)$, if a subset $R_{\alpha,\tau}$ of $[n]$ ordered $T$ above $S$, then this would ``cost'' $1$ from $\abs{T}-\abs{S}\le \ell-k$. However, when dealing with $\vv$-size, we can no longer make such a claim, since $\abs{T_{\alpha,\tau}}_{\vv}-\abs{S_{\alpha,\tau}}_{\vv}$ could be very small. Instead, we will find a way to still ``win'' if $\abs{T_{\alpha,\tau}}_{\vv}-\abs{S_{\alpha,\tau}}_{\vv}$ is sufficiently small.
\begin{lemma}\label{lem:slant-L2}
Let $\vv\in\R^n_{>0}$, and $k,\ell,r$ be real numbers with $0<k<\ell$ and $r\ge1$. Suppose that $m(\vv,r)\ge 2(\ell-k)$. Then there exists a set $\LL_2$ of at most $27r\log n$ linear extensions of $\M_{[k,\ell]}^n$ such that for every incomparable pair of multisets $(S,T)$ with $\dist{T\setminus S}> 3r,$ there exists an extension $L\in\LL_2$ where $S>_LT$.
\end{lemma}
\begin{proof}
Set $a=b=3r$, $t=3\log n$, and fix an $(a,b,r,t,n)$-good function $f$ as in the proof of \cref{lem:L2}. Let $R_{\alpha,\tau}=\{i\in [n] \mid f(i,\tau)=\alpha\}$ denote part $\alpha$ of partition $\tau$.
For each $\alpha\in[a]$, $\tau\in[t]$, and $j\in\{0,1,2\}$, we will construct a linear extension $L_{\alpha,\tau,j}$ of $\M_{[k,\ell]}^{n,\vv}$. That is, we will construct three extensions for each part in each partition. Our set of extensions $\LL_2$ will consist of these $3at$ extensions.
For a multiset $S$, define $S_{\alpha,\tau}$ as before. Let $M_1$ and $M_2$ be two linear extensions of $\M^n$ which both order multisets by $\vv$-size, but order the multisets of a given $\vv$-size in opposite orders. Let $M_0$ be an arbitrary linear extension of $\M^n$.
First we define $L_{\alpha,\tau,0}$. For each $\alpha,\tau$, we will have $L_{\alpha,\tau,0}$ order multisets $S$ by $\abs{S_{\alpha,\tau}}_{\vv}$, breaking ties using $M_0$. That is, we have $S<_{L_{\alpha,\tau,0}}T$ if $\abs{ S_{\alpha,\tau} }_{\vv} <\abs{ T_{\alpha,\tau} }_{\vv}$. If $\abs{ S_{\alpha,\tau} }_{\vv} =\abs{ T_{\alpha,\tau} }_{\vv}$, then $S<_{L_{\alpha,\tau,0}}T$ if and only if $S<_{M_0}T$.
We now define $L_{\alpha,\tau,j}$ for $j\in\{1,2\}$. For each $\alpha,\tau$, both extensions will order multisets $S$ by taking into account the $\vv$-size $\abs{S_{\alpha,\tau}}_{\vv}$. Specifically, $L_{\alpha,\tau,j}$ will order multisets $S$ by applying an ordering $K_{\alpha,\tau,j}$ of $\R$ to $\abs{S_{\alpha,\tau}}_{\vv}$ as follows. If $\abs{S_{\alpha,\tau}}_{\vv}<_{K_{\alpha,\tau,j}}\abs{T_{\alpha,\tau}}_{\vv}$, then $S<_{L_{\alpha,\tau,j}}T$. This defines how $L_{\alpha,\tau,j}$ orders multisets $S$ and $T$ except when $\abs{S_{\alpha,\tau}}_{\vv}=\abs{T_{\alpha,\tau}}_{\vv}$. We will deal with this edge case later.
To define $K_{\alpha,\tau,j}$, we first let $\varepsilon=\varepsilon_{\vv,f}(\alpha,\tau)=\min_{i\in R_{\alpha,\tau}}(v_i)$, where $v_i$ is the $i$th entry of $\vv$. Now, consider dividing the real number line into half-open intervals of length $\varepsilon$ in the following two ways.
$$\R=\cdots\cup[0,\varepsilon)\cup[\varepsilon,2\varepsilon)\cup[2\varepsilon,3\varepsilon)\cup\cdots$$
$$\R=\cdots\cup[\varepsilon/2,3\varepsilon/2)\cup[3\varepsilon/2,5\varepsilon/2)\cup[5\varepsilon/2,7\varepsilon/2)\cup\cdots$$
We let $K_{\alpha,\tau,1}$ order the intervals in the first partition of $\R$ in increasing order, but order the elements of each interval in decreasing order. We define $K_{\alpha,\tau,2}$ in the same way from the second partition of $\R$.
Specifically, for reals $x<y$, we have $x<_{K_{\alpha,\tau,1}}y$ if and only if $\floor{x/\varepsilon}<\floor{y/\varepsilon}$, and $x<_{K_{\alpha,\tau,2}}y$ if and only if $\floor{x/\varepsilon-1/2}<\floor{y/\varepsilon-1/2}$.
This in turn defines our linear extensions $L_{\alpha,\tau,1}$ and $L_{\alpha,\tau,2}$. For example, for multisets $S$ and $T$ with $\varepsilon<\abs{S_{\alpha,\tau}}_{\vv}<3\varepsilon/2<\abs{T_{\alpha,\tau}}_{\vv}<2\varepsilon$, we will have $S>_{L_{\alpha,\tau,1}}T$ and $S<_{L_{\alpha,\tau,2}}T$. Additionally, note that if $S$ and $T$ are two multisets satisfying $\abs{T_{\alpha,\tau}}_{\vv}-\varepsilon/2<\abs{S_{\alpha,\tau}}_{\vv}<\abs{T_{\alpha,\tau}}_{\vv}$, then at least one of $L_{\alpha,\tau,1}$ and $L_{\alpha,\tau,2}$ orders $S$ greater than $T$.
Now we deal with the edge case of distinct multisets $S,T$ with $\abs{S_{\alpha,\tau}}_{\vv}=\abs{T_{\alpha,\tau}}_{\vv}$. If $S_{\alpha,\tau}\ne T_{\alpha,\tau}$, then $S<_{L_{\alpha,\tau,j}}T$ if and only if $S_{\alpha,\tau}<_{M_j}T_{\alpha,\tau}$. Finally, if $S_{\alpha,\tau}= T_{\alpha,\tau}$, then $S<_{L_{\alpha,\tau,j}}T$ if and only if $S<_{M_0}T$.
Next, we check that $L_{\alpha,\tau,j}$ is indeed a linear extension of $\M_{[k,\ell]}^{n,\vv}$. Suppose $S,T\in\M_{[k,\ell]}^{n,\vv}$ with $S\subseteq T$. Then $S\le_{M_0}T$. Also, $S_{\alpha,\tau}\subseteq T_{\alpha,\tau}$, so $S_{\alpha,\tau}\le_{M_j}T_{\alpha,\tau}$. Thus $S\le_{L_{\alpha,\tau,j}} T$ if $\abs{S_{\alpha,\tau}}_{\vv}=\abs{T_{\alpha,\tau}}_{\vv}$. Now suppose that $\abs{S_{\alpha,\tau}}_{\vv}<\abs{T_{\alpha,\tau}}_{\vv}$. Pick any $i\in (T\setminus S)_{\alpha,\tau}$. Since $S\subseteq T$, we have $$\abs{T_{\alpha,\tau}}_{\vv}-\abs{S_{\alpha,\tau}}_{\vv}=\abs{(T\setminus S)_{\alpha,\tau}}_{\vv}\ge v_i\ge \varepsilon(\alpha,\tau).$$
Thus $\abs{S_{\alpha,\tau}}_{\vv}$ and $\abs{T_{\alpha,\tau}}_{\vv}$ cannot lie in the same half-open interval of length $\varepsilon(\alpha,\tau)$. So $\abs{S_{\alpha,\tau}}_{\vv}<_{K_{\alpha,\tau,j}}\abs{T_{\alpha,\tau}}_{\vv}$ and $S<_{L_{\alpha,\tau,j}} T$. Therefore $L_{\alpha,\tau,j}$ is indeed a linear extension of $\M_{[k,\ell]}^{n,\vv}$.
We now show that $\LL_2$ satisfies the desired condition. Suppose that $(S,T)$ is an incomparable pair of multisets in $\M_{[k,\ell]}^{n,\vv}$ with $\dist{T\setminus S}>3r=b$. Then $k\le \abs{S}_{\vv},\abs{T}_{\vv}\le \ell$. So $\abs{T}_{\vv}-\abs{S}_{\vv}\le \ell-k.$ Since $f$ is $(a,b,r,t,n)$-good, there exists $\tau\in[t]$ such that $\abs{f(T\setminus S,\tau)}>r$. If there exists $\alpha\in[a]$ such that $\abs{S_{\alpha,\tau}}_{\vv}>\abs{T_{\alpha,\tau}}_{\vv}$, then we are done. So assume that $\abs{S_{\alpha,\tau}}_{\vv}\le\abs{T_{\alpha,\tau}}_{\vv}$ for all $\alpha\in[a]$. Next, consider $\alpha\in f(T\setminus S,\tau)$. If $\abs{S_{\alpha,\tau}}_{\vv}>\abs{T_{\alpha,\tau}}_{\vv}-\varepsilon(\alpha,\tau)/2$, then $S>T$ in $L_{\alpha,\tau,j}$ for at least one $j\in\{1,2\}$. So assume that $\abs{T_{\alpha,\tau}}_{\vv}\ge \abs{S_{\alpha,\tau}}_{\vv}+\varepsilon(\alpha,\tau)/2$ for each $\alpha\in f(T\setminus S,\tau)$. Summing over all such $\alpha$, we obtain
$$\ell-k\ge \abs{T}_{\vv}-\abs{S}_{\vv}\ge \frac{1}{2}\sum_{\alpha\in f(T\setminus S,\tau)}\varepsilon(\alpha,\tau)> \frac{1}{2}m(\vv,r),$$
where the last inequality follows by noting that each $\varepsilon(\alpha,\tau)$ is a distinct entry of $\vv$ and using $\abs{f(T\setminus S,\tau)}>r$. Since $m(\vv,r)\ge 2(\ell-k)$, we have a contradiction. This completes the proof.
\end{proof}
Applying \cref{lem:L1} and \cref{lem:slant-L2}, we have that $$\dim(\M_{[k,\ell]}^{n,\vv})\le (3r+1)^2\log n+27r\log n \le 43r^2\log n.$$
This proves \cref{thm:weighted-multisets}.
\section{Results on Divisibility in $\N$}\label{sec:divisibility}
In this section we use \cref{thm:weighted-multisets} to obtain a significantly better bound on $\dim(\D_{[N/\kappa,N]})$, proving \cref{thm:divisibility}.
We use some of the ideas from the proof of \cref{prop:easy-divisibility}. As before, call a prime $p$ \emph{small} if $p\le \kappa$ and \emph{large} if $p>\kappa$. Let $K$ be the set of all positive integer $M\le N$ with no small primes dividing $M$. For each $M\in K$, let $g(M)$ be the set of all positive integers of the form $Mq$, where $q$ has only small prime divisors. Now, recall that to bound $\dim(\D_{[N/\kappa,N]})$ it suffices to bound $\dim(\D_{g(M)\cap [N/\kappa,N]})$ for each $M\in K$. We will accomplish this by applying \cref{thm:weighted-multisets}.
Let $n=\pi(\kappa)$, and let $\vv=(\log2,\log3,\dots,\log p_n)$, where $p_i$ is the $i$th prime. Fix some $M\in K$ and let $k_M=\log((N/\kappa)/M)$ and $\ell_M=\log(N/M)$, so that $\ell_M-k_M=\log(\kappa)$. Now, by associating an integer with its prime factorization we have an isomorphism between posets $$\M_{[k_M,\ell_M]}^{n,\vv}\cong\D_{g(M)\cap [N/\kappa,N]}.$$
If $\kappa<3$ then $\dim(\M_{[k_M,\ell_M]}^{n,\vv})\le 2$. Otherwise, it can be checked (see \cref{app:prime-bash}) that with $r=4\frac{\log \kappa}{\log \log \kappa}$, we have $m(\vv,r)\ge 2\log(\kappa)$. So by \cref{thm:weighted-multisets}, $\dim(\M_{[k_M,\ell_M]}^{n,\vv})\le 43r^2\log n$. This holds for all choices of $M$, so we have that $$\dim(\D_{[N/\kappa,N]})\le 43r^2\log n\le 688\frac{(\log\kappa)^3}{(\log\log \kappa)^2},$$
which proves \cref{thm:divisibility}.
\section{Other Divisibility Orders}\label{sec:polynomial}
In this section we consider other divisibility orders. In some sense the poset $\M_{[k,\ell]}^{n,\vv}$ captures all possible normed divisibility orders subject to reasonable conditions. Specifically, given a (multiplicative) monoid with a norm and some notion of primes and unique factorization\footnote{We just need commutativity, cancellation, and finitely many elements with norm at most $N$ for each $N$.}, we can construct a divisibility order on the elements with norms lying in an interval $[N/\kappa,N]$. Letting $\vv$ be the vector of the norms of primes with norm at most $\kappa$, we can then apply \cref{thm:weighted-multisets}. If we have an understanding of the primes and the norm we can obtain a concrete bound, as in \cref{thm:divisibility}.
As another example of this principle, we analyze the divisibility order on monic polynomials over a finite field. As in the proof of \cref{prop:easy-divisibility}, $P(\F_q)_{[d_0-\delta,d_0]}$ is isomorphic to a disjoint union of posets of the form $\M^{n,\vv}_{k,\ell},$ where $\ell-k=\delta$, $n$ is the number of monic irreducible polynomials over $\F_q[x]$ with degree at most $\delta$, and $\vv\in \R^n$ is the vector of degrees of these $n$ polynomials.
It can be checked (see \cref{app:polynomial-bash}) that $n\le q^\delta$ and that with $r=4.6\frac{\delta\log q}{\log \delta}$, we have $m(\vv,r)\ge 2\delta$. Additionally, since each entry of $\vv$ is at least $1$, we have $m(v,r)\ge 2\delta$ with $r=2\delta$. These two choices of $r$ handle the regimes when $q<\delta$ and $q\ge \delta$, respectively. By applying \cref{thm:weighted-multisets}, we obtain
$$\dim(P(\F_q)_{[d_0-\delta,d_0]})\le 43\left(\min\left(4.6\frac{\delta\log q}{\log \delta},2\delta\right)\right)^2\log(q^\delta)\le \min\left(910\frac{(\delta\log q)^3}{(\log \delta)^2},172\delta^3\log q\right).$$
This proves \cref{thm:polynomial}.
\section{Further Directions}\label{sec:further-directions}
Considering multiset posets was initially motivated by working on divisibility posets. However, multiset posets are a natural extension of subset posets and interesting on their own. It would be nice to see \cref{thm:weighted-multisets} applied to problems not directly obtained from divisibility orders.
Although \cref{thm:divisibility} is a substantial improvement over previously known bounds, it is not quite tight with the lower bound given by Lewis and Souza. This lower bound cannot be improved without new ideas, since it is obtained by embedding the optimal subset poset of the form $\Q_{[1,\ell]}^n$ into $\D_{[N/\kappa,N]}$ \cite{Lewis-Souza-paper}. On the other hand, there are also reasons to believe that \cref{thm:divisibility} might be close to tight. The limiting factor in the upper bound is the dimension of the subset poset $\Q_{[1,3r]}^n$. In the regime when $r$ is close to $\log n$, Kierstead showed that $\dim(\Q_{[1,3r]}^n)$ is close to $r^3$ \cite{Kierstead-paper}. Unfortunately the proof technique does not extend easily to the divisibility problem, but perhaps it could be modified with new ideas.
\section{Acknowledgements}\label{sec:acknowledgements}
I would like to thank Noah Kravitz for helpful conversations and detailed feedback on earlier versions of this paper. Additionally, I am grateful to Victor Souza for bringing other divisibility orders to my attention, among other helpful comments. I would also like to thank Michael Ren and William Trotter for helpful conversations and Joe Gallian, Carl Schildkraut, and Zach Hunter for comments on earlier versions of this paper. Finally, I want to express my gratitude to Joe Gallian, Amanda Burcroff, Colin Defant, Noah Kravitz, and Yelena Mandelshtam for organizing the Duluth REU and the project suggestion.
This research was conducted at the University of Minnesota Duluth Mathematics REU and was supported, in part, by NSF-DMS Grant 1949884 and NSA Grant H98230-20-1-0009. Additional support was provided
by the CYAN Mathematics Undergraduate Activities Fund.
\bibliographystyle{plain}
|
\section{Methods}
\subsection{Mutual Inductance Set Up}
Coils were formed using 20 $\upmu$m diameter copper wire to limit the total coil diameter to under 0.5 mm to reduce the impact of edge currents. The pickup coil is composed of 400 turns and has dimensions of approximately 0.5 mm x 0.5 mm. The drive coil is composed of 50 turns and has dimensions 0.25 mm x 0.25 mm. Leakage around the sample was measured at 1.5\% using a 100 nm thick aluminum film with lateral dimensions equal to those of the sample (2.5 mm x 5 mm). At this film thickness, there will be no transmission through the film itself, so the mutual inductance measured reflects leakage around the film. Measurements of nickelate samples are calibrated for this leakage by subtraction.
\begin{figure}
\includegraphics[width=6.4 in]{schematic.pdf}
\caption{Schematic of the set up for mutual inductance measurements. a) Drawing showing the key facets of the device. The device follows the design shown in \cite{he2016}, with pogo pins pressing a mechanical wedge into the sample, which thereby presses it into a sapphire plate for thermalization. The coils are positioned directly above and below the sample, and the entire set up is aligned by a copper bracket, which connects it to the mixing chamber. A RuO$_2$ thermometer is glued with silver paint to the sapphire plate as well to provide a local measurement of temperature. b) Photograph of the actual device, affixed to the dilution refrigerator.}
\label{schem}
\end{figure}
\subsection{Measurements}
Measurements were performed with a lock-in amplifier, during recovery of the helium mixture, while the temperature was increased at between 2-5 mK / minute in the temperature range below $T_c/3$ and 10-25 mK/min at the temperature range above. To reduce noise, data is binned in 30 mK increments.
We measure a voltage, $V_{\mathrm{pickup}}= 2 \pi f I_{\mathrm{drive}} M$, where $M$ is the mutual inductance between the two coils. For an infinitely-wide sample, this can be solved
\begin{align}
M(\lambda, \sigma_1) &=\Sigma_{i} \Sigma_{j} \pi \mu_0 r_{\mathrm{dr}}^i r_{\mathrm{pu}}^j \int_0^{\infty} dx \frac{e^{- x D^{i,j}} J_1 (x r_{\mathrm{dr}}^i) J_1(x r_{\mathrm{pu}}^j)}{\cosh(d \chi) + \frac{x^2 + \chi^2}{2 \chi x} \sinh(d \chi)} \\
\chi^2 &= x^2 + 1/\lambda^2 + 2 i \pi \mu_0 f \sigma_1 \nonumber,
\end{align}
where $r_{\mathrm{dr}}^i $ is the radius of the $i^\mathrm{th}$ turn of the drive coil, $r_{\mathrm{pu}}^j $ is the radius of the $j^\mathrm{th}$ turn of the pickup coil, $d$ is the sample thickness, $D^{i,j}$ is the distance between the $i^\mathrm{th}$ turn of the drive coil and the $j^\mathrm{th}$ turn of the pickup coil. The integral term is summed across each combination of a turn of the drive coil and turn of the pickup coil. We calculate this term for a series of penetration depths and for $\lambda=\infty, \sigma_1=0$, which we use to normalize the signal $M_{\mathrm{norm}} = M_x(\lambda, 0)/M_x(\infty, 0)$, thereby reducing the error from uncertainty in geometry. We estimate the error in the absolute value of the penetration depth is about 10$\%$ using this method, dominated by uncertainty in the coil geometry. The error in the temperature dependence of the penetration depth is far lower.
The phase of the measurement is set by auto-phasing the signal at base temperature for our first sample measurement, and keeping the phase constant for the remaining measurements. The phase of the measurement is within 5 degrees of $90^{\circ}$, with the difference most likely occurring due to capacitances in the wiring. The signal can be represented as $ M = M_x + i M_y$. We extract the penetration depth by normalizing $M_x$ by the signal measured directly above $T_c$, $M_{x,n}$, and looking up the result by interpolating a table of $M_{\mathrm{norm}}(\lambda,0)$, calculated as above.
For subsequent measurements, we correct the $M_y$ signal at low temperatures by subtracting an offset from it to set it to 0 at base temperature, which gives the most physical results for the temperature dependence of $M_y(T)$, with $M_y(T)$ never becoming negative. We expect that the small changes in the signal magnitude of about $0.02 M_x$ are due to parasitics in the lines. In the limit where $\sigma_1$ is small, which is true in our experiment, $M_y \propto \sigma_1$, and values of $\sigma_1$ plotted in this paper are equal to $M_y/M_{x,n}$. Due to uncertainty about the exact impact of the wiring, these results are not quantitative, but act as a qualitative guide to the superconducting transition.
\subsection{Transport Measurements}
The samples were contacted using wire-bonded aluminum wires and measured in a Quantum Design Physical Property Measurement System with base temperature of 2 K.
\section{Fit details}
The fit for the nodal gap is derived in \cite{hirschfeld1993}. For a conventional BCS $s$-wave gap that is constant for all values in $k$-space (such as the one shown in Figure 1a), it is straightforward to derive the temperature dependence of the superfluid density,
\begin{equation}
\left(\lambda_0/\lambda(T)\right)^2= 1 - \sqrt{\frac{2 \pi \Delta_0}{\kb T}} \exp(-\Delta_0/\kb T)
\end{equation}
This provides a poor fit to all of the data sets, including the Nd-nickelate. Because the nodal equation (Eq. 2 of main text) has two fit variables, while this equation has only one, this fit is more constrained than the nodal one. Therefore, we use a variant of the BCS equation that has a second fit variable (Eq. 1 of main text) to test for a non-nodal gap. This might be explained physically, for instance, by a multiple gaps or a gap that varies in size while remaining non-nodal.
\section{Fit to data in Arrhenius and log-log plots}
Fits are shown for each of the Arrhenius and power law fits, and the slopes are presented in Table S1. The slope of the Arrhenius plot represents the magnitude of the superconducting gap, $\Delta_0/\kb T_c$. The weak-coupling limit for $s$-wave pairing is $1.76 T_c$. That the slopes of the Arrhenius plot for the Nd-nickelate is far below that is a strong indication that they do not possess conventional $s$-wave pairing. Some pnictide materials possessing $s^{\pm}$ pairing have been measured with gaps of similar magnitude. The La- and Pr-nickelate samples indicate gaps of under $0.2T_c$, suggesting that if there is a non-nodal gap, it has an extremely small minimum.
The slope of the log-log plot, $m$, represents the power-law scaling of the penetration depth, $1-\lambda_0^2/\lambda(T)^2=c T^m$ where $c$ is an arbitrary constant. The slopes of the data sets from the La- and Pr-nickelates are 1.7. If the temperature range from $0.85T_c$ to 3 K is used for Nd-nickelate, the range where magnetism is unlikely to contribute, we find that the data set fits a power law with scaling of 2. However, the significance of this not clear, because this temperature range is mostly above $T_c/3$ and may therefore reflect the temperature dependence of the gap rather than the pairing symmetry of the material.
\begin{figure}
\includegraphics[width=6.4 in]{supplement_arr.pdf}
\caption{Arrhenius plots for all samples. For La-nickelate and Pr-nickelate, the line is a fit to the low temperature data ($T/T_c<0.1$). Nd-nickelate shows distinctly different behavior at low temperatures, most likely due to magnetic effects, so the data is fit to a line between $T/T_c<0.3$ and $T>1.2$ K. The slopes are presented in Table S1.}
\label{arr}
\end{figure}
\begin{figure}
\includegraphics[width=6.4 in]{supplement_pow2.pdf}
\caption{Log-log plots for all samples, labeled with the sample number. For La-nickelate and Pr-nickelate, the line is fit from $0.85T_c$ down to base temperature. The fit-line extends across the straight-line region of the Nd-nickelate, which extends from $0.85 T_c$ to about 3 K. The slopes are presented in Table S1.}
\label{pow}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
R & $T_c$ (K) & Exp. slope & Pow. Slope \\
La & 7.6 & 0.1 (0.005) & 1.7 (0.03) \\
Pr & 10.1 & 0.1 (0.08) & 1.7 (0.04) \\
Nd & 11.1 & 0.5 (0.015) & 2 (0.02) \\
\hline
\end{tabular}
\end{center}
\caption{\label{tab:table1} Slopes of the Arrhenius and Log-log plots, listed as Exp. slope and pow. slope, respectively. Errors in fit parameters are displayed in parentheses.}
\end{table}
|
\section{Conclusions}
In this paper, we addressed the problem of FL model performance degradation in the presence of non-i.i.d. data. We proposed a new concept of dynamic STP collaborative training that is robust against data heterogeneity, and a grouped framework \textsc{FedGSP}~to support dynamic management of the continuously growing client groups. In addition, we proposed ICG~to support efficient group assignment in STP by solving a constrained clustering problem with equal group size constraint, aiming to minimize the data distribution divergence among groups. We experimentally evaluated \textsc{FedGSP}~on LEAF, a widely adopted FL benchmark platform, with the non-i.i.d. FEMNIST dataset. The results showed that \textsc{FedGSP}~significantly outperforms seven state-of-the-art approaches in terms of model accuracy and convergence speed. In addition, \textsc{FedGSP}~is both training- and communication-efficient, making it suitable for practical applications.
\section{Experimental Evaluation}\label{section:experiment}
\subsection{Experiment Setup and Evaluation Metrics}
\textbf{Environment and Hyperparameter Setup.} The experiment platform contains $K=368$ FL clients. The most commonly used FEMNIST\cite{caldas2018leaf} is selected as the benchmark dataset, which is specially designed for non-i.i.d. FL environment and is constructed by dividing 805,263 digit and character samples into 3,550 FL clients in a non-uniform class distribution, with an average of $n=226$ samples per client. For the resource-limited mobile devices, a lightweight neural network composed of 2 convolutional layers and 2 fully connected layers with a total of 6.3 million parameters is adopted as the training model. The standard mini-batch SGD is used by FL clients to train their local models, with the learning rate $\eta=0.01$, the batch size $b=5$ and the local epoch $e=1$. We test \textsc{FedGSP}~for $R=500$ rounds. By default, we set the group sampling rate $\kappa=0.3$, the group number growth function $f=\textsc{Log}$ and the corresponding coefficients $\alpha=2$, $\beta=10$. The values of $\kappa$, $f$, $\alpha$, $\beta$ will be further tuned in the experiment to observe their performance influence.
\textbf{Benchmark Algorithms.} In order to highlight the effect of the proposed STP and ICG~separately, we remove them from \textsc{FedGSP}~to obtain the naive version, NaiveGSP. Then, we compare the performance of the following versions of \textsc{FedGSP}~through ablation studies:
\begin{enumerate}
\item \textit{NaiveGSP}: FL clients are randomly assigned to a fixed number of groups, the clients in the group are trained in sequence and the groups are trained in parallel (e.g., Astraea\cite{duan2019astraea}).
\item \textit{NaiveGSP+ICG}: The ICG~grouping algorithm is adopted in NaiveGSP~to assign FL clients to a fixed number of groups strategically.
\item \textit{NaiveGSP+ICG+STP} (\textsc{FedGSP}): On the basis of NaiveGSP+ICG, FL clients are reassigned to a growing number of groups in each round as required by STP.
\end{enumerate}
In addition, seven state-of-the-art baselines are experimentally compared with \textsc{FedGSP}. They are FedProx \cite{li2018federated}, FedMMD \cite{yao2018two}, FedFusion \cite{yao2019towards}, IDA \cite{yeganeh2020inverse}, and FedAdagrad, FedAdam, FedYogi from \cite{reddi2020adaptive}.
\textbf{Evaluation Metrics.} In addition to the fundamental test accuracy and test loss, we also define the following metrics to assist in performance evaluation.
\textit{Class Probability Distance} (CPD). The maximum mean discrepancy (MMD) distance is a probability measure in the reproducing kernel Hilbert space. We define CPD as the kernel two-sample estimation with Gaussian radial basis kernel $\mathcal{K}$\cite{gretton2012kernel} to measure the difference in class probability (i.e., normalized class distribution) $\mathcal{P}=\mathrm{norm}(\mathcal{V}_{m_1}),\mathcal{Q}=\mathrm{norm}(\mathcal{V}_{m_2})$ between two groups $m_1,m_2$. Generally, the smaller the CPD, the smaller the data heterogeneity between two groups, and therefore the better the grouping strategy.
\begin{align}
\mathrm{CPD}(m_1,m_2)&=\mathrm{MMD}^2(\mathcal{P}, \mathcal{Q}) \\
\nonumber
&= \mathbf{E}_{x,x'\sim\mathcal{P}}\left[\mathcal{K}(x,x')\right]-2\mathbf{E}_{x\sim\mathcal{P},y\sim\mathcal{Q}}\left[\mathcal{K}(x,y)\right]+\mathbf{E}_{y,y'\sim\mathcal{Q}}\left[\mathcal{K}(y,y')\right].
\end{align}
\textit{Computational Time.} We define $T_{\mathrm{comp}}$ in Eq. \eqref{eq:comp-time-metric} to estimate the computational time cost, where the number of floating point operations (FLOPs) is $\mathcal{N}_{\mathrm{calc}}=96$M FLOPs per sample and $\mathcal{N}_{\mathrm{aggr}}=6.3$M FLOPs for global aggregation, and $\mathcal{T}_{\mathrm{FLOPS}}=567$G FLOPs per second is the computing throughput of the Qualcomm Snapdragon 835 smartphone chip equipped with Adreno 540 GPU.
\begin{equation}
\label{eq:comp-time-metric}
T_{\mathrm{comp}}(R)=\sum_{r=1}^{R}\left(\underbrace{\frac{\mathcal{N}_{\mathrm{calc}}}{\mathcal{T}_{\mathrm{FLOPS}}}\cdot\frac{neK}{\min\left\{K,f(r)\right\}}}_{\mathrm{Local~Training}}+\underbrace{\frac{\mathcal{N}_{\mathrm{aggr}}}{\mathcal{T}_{\mathrm{FLOPS}}}\cdot \left[\kappa f(r)-1\right]}_{\mathrm{Global~Aggregation}}\right)~(\mathrm{s}).
\end{equation}
\textit{Communication Time and Traffic.} We define $T_{\mathrm{comm}}$ in Eq. \eqref{eq:comm-time-metric} to estimate the communication time cost and $D_{\mathrm{comm}}$ in Eq. \eqref{eq:comm-traffic-metric} to estimate the total traffic, where the FL model size is $\mathcal{M}=25.2$MB, the inbound and outbound transmission rates are $\mathcal{R}_{\mathrm{in}}=\mathcal{R}_{\mathrm{out}}=567\mathrm{Mbps}$ (tested in the Internet by AWS EC2 r4.large 2 vCPUs with disabled enhanced networking). Eq. \eqref{eq:comm-time-metric} to Eq. \eqref{eq:comm-traffic-metric} consider only the cross-WAN traffic between FL clients and group managers, but the traffic between the top server and group managers is ignored because they are deployed in the same physical machine.
\begin{equation}
\label{eq:comm-time-metric}
T_{\mathrm{comm}}(R)=8\kappa K\mathcal{M}R(\frac{1}{\mathcal{R}_{\mathrm{in}}}+\frac{1}{\mathcal{R}_{\mathrm{out}}})~(\mathrm{s}),
\end{equation}
\begin{equation}
\label{eq:comm-traffic-metric}
D_{\mathrm{comm}}(R)=2\kappa K\mathcal{M}R~(\mathrm{Bytes}).
\end{equation}
Please note that Eq. \eqref{eq:comp-time-metric} to Eq. \eqref{eq:comm-time-metric} are theoretical metrics, which do not consider memory I/O cost, network congestion, and platform configurations such as different versions of CUDNN/MKLDNN libraries.
\subsection{Results and Discussion}
\textbf{The effect of ICG~and STP.} We first compare the CPD of FedAvg\cite{mcmahan2017communication}, NaiveGSP~and NaiveGSP+ICG~in Figure \ref{fig:probability-distance}. These CPDs are calculated between every pair of FL clients. The results show that NaiveGSP+ICG~reduces the median CPD of FedAvg by $82\%$ and NaiveGSP~by $41\%$. We also show their accuracy performance in Figure \ref{fig:acc-curve}. The baseline NaiveGSP~quickly converges but only achieves the accuracy similar to FedAvg. Instead, NaiveGSP+ICG~improves the accuracy by $6\%$. This shows that reducing the data heterogeneity among groups can indeed effectively improve FL performance in the presence of non-i.i.d. data. Although NaiveGSP+ICG~is already very effective, it still has defects. Figure \ref{fig:loss-curve} shows a rise in the loss value of NaiveGSP+ICG, which indicates that it has been overfitted. That is because the training mode of NaiveGSP+ICG~is static, it may learn the client order and forget the previous data. Instead, the dynamic \textsc{FedGSP}~overcomes overfitting and eventually converges to a higher accuracy $85.4\%$, which proves the effectiveness of combining STP and ICG.
\begin{figure}[t]
\centering
\subfloat[CPD]{\includegraphics[width=.33\textwidth]{pic/class-prob-dist.pdf}
\label{fig:probability-distance}}
\subfloat[Accuracy curve]{\includegraphics[width=.33\textwidth]{pic/acc-curve-compare.pdf}
\label{fig:acc-curve}}
\subfloat[Loss curve]{\includegraphics[width=0.34\textwidth]{pic/loss-curve-compare.pdf}
\label{fig:loss-curve}}
\caption{Comparison among FedAvg, NaiveGSP, NaiveGSP+ICG~and \textsc{FedGSP}~in (a) CPD, (b) accuracy curve and (c) loss curve. In subfigure (a), the orange line represents the median value and the green triangle represents the mean value.}
\end{figure}
\textbf{The effect of the growth function $f$ and its coefficients $\alpha,\beta$.} To explore the performance influence of different group number growth functions $f$, we conduct a grid search on $f=\{\textsc{Linear},\textsc{Log},\textsc{Exp}\}$ and $\alpha,\beta$. The test loss heatmap is shown in Figure \ref{fig:loss-heatmap}. The results show that the logarithmic growth function achieves smaller loss 0.453 with $\alpha=2,\beta=10$ among 3 candidate functions. Besides, we found that both lower and higher $\alpha,\beta$ lead to higher loss values. The reasons may be that a slow increase in the number of groups leads to more STM and results in overfitting, while a rapid increase in the number of groups makes \textsc{FedGSP}~degenerate into FedAvg prematurely and suffers the damage of data heterogeneity. Therefore, we recommend $\alpha\cdot\beta$ to be a moderate value, as shown in the green area.
\begin{figure}[t]
\centering
\subfloat[Linear]{\includegraphics[width=0.34\textwidth]{pic/heatmap-linear.pdf}}
\subfloat[Logarithmic]{\includegraphics[width=0.34\textwidth]{pic/heatmap-log.pdf}}
\subfloat[Exponential]{\includegraphics[width=0.34\textwidth]{pic/heatmap-exp.pdf}}\\
\caption{Test loss heatmap of (a) linear, (b) logarithmic and (c) exponential growth functions over different $\alpha$ and $\beta$ settings in \textsc{FedGSP}.}
\label{fig:loss-heatmap}
\end{figure}
\begin{figure}[t]
\centering
\subfloat[Accuracy]{\includegraphics[width=0.33\textwidth]{pic/acc-vs-kappa.pdf}
\label{fig:acc-vs-kappa}}
\subfloat[Computational Time]{\includegraphics[width=0.33\textwidth]{pic/comp-time-vs-kappa.pdf}
\label{fig:comp-time-vs-kappa}}
\subfloat[Communication Time]{\includegraphics[width=0.33\textwidth]{pic/comm-time-vs-kappa.pdf}
\label{fig:comm-time-vs-kappa}}
\caption{Comparison of (a) accuracy and the normalized (b) computational time and (c) communication time over different $\kappa$ settings in \textsc{FedGSP}.}
\label{fig:kappa-comparison}
\end{figure}
\textbf{The effect of the group sampling rate $\kappa$.} $\kappa$ controls the participation rate of groups (also the participation rate of FL clients) in each round. We set $\kappa=\{0.1,0.2,0.3,0.5,1.0\}$ to observe its effect on accuracy and time cost. Figure \ref{fig:acc-vs-kappa} shows the robustness of accuracy to different values of $\kappa$. This is expected because ICG~forces the data of each group to become homogeneous, which enables each group to individually represent the global data. In addition, Figures \ref{fig:comp-time-vs-kappa} and \ref{fig:comm-time-vs-kappa} show that $\kappa$ has a negligible effect on computational time $T_{\mathrm{comp}}$, but a proportional effect on communication time $T_{\mathrm{comm}}$ because a larger $\kappa$ means more model data are involved in data transmission. Therefore, we recommend that only $\kappa\in[0.1,0.3]$ of groups are sampled to participate in FL in each round to reduce the overall time cost. In our experiments, we set $\kappa=0.3$ by default.
\textbf{The performance comparison of \textsc{FedGSP}.} We compare \textsc{FedGSP}~with seven state-of-the-art approaches and summarize their test accuracy, test loss and training rounds (required to reach the accuracy of $80\%$) in Table \ref{table:fedgsp-vs-others}. The results show that \textsc{FedGSP}~achieves $5.3\%$ higher accuracy than FedAvg and reaches the accuracy of $80\%$ within only $34$ rounds. Moreover, \textsc{FedGSP}~outperforms all the comparison approaches, with an average of $3.7\%$ higher accuracy, $0.123$ lower loss and $84\%$ less rounds, which shows its effectiveness to improve FL performance in the presence of non-i.i.d. data.
\textbf{The time and traffic cost of \textsc{FedGSP}.} Figure \ref{fig:6} visualizes the time cost and total traffic of \textsc{FedGSP}~and FedAvg when they reach the accuracy of $80\%$. The time cost consists of computational time $T_{\mathrm{comp}}(R)$ and communication time $T_{\mathrm{comm}}(R)$, of which $T_{\mathrm{comm}}(R)$ accounts for the majority due to the huge data traffic from hundreds of FL clients has exacerbated the bandwidth bottleneck of the cloud server. Figure \ref{fig:6} also shows that \textsc{FedGSP}~spends $93\%$ less time and traffic than FedAvg, which benefits from a cliff-like reduction in the number of training rounds $R$ (only 34 rounds to reach the accuracy of $80\%$). Therefore, \textsc{FedGSP}~is not only accurate, but also training- and communication-efficient.
\vspace{10mm}
\begin{minipage}{\textwidth}
\begin{minipage}[b]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{pic/cost-compare.pdf}\label{fig:cost-compare}
\captionof{figure}{Comparison of time and traffic cost to reach $80\%$ accuracy.}\label{fig:6}
\end{minipage}
\quad
\begin{minipage}[b]{0.5\textwidth}
\centering
\scalebox{0.9}{
\begin{tabular}{c|c|c|c}
\hline
Algorithm & Accuracy & Loss & Rounds \\
\hline\hline
FedAvg & 80.1\% & 0.602 & 470 \\
\hline
FedProx & 78.7\% & 0.633 & $\times$ \\
FedMMD & 81.7\% & 0.587 & 336 \\
FedFusion & 82.4\% & 0.554 & 230 \\
IDA & 82.0\% & 0.567 & 256 \\
FedAdagrad & 81.9\% & 0.582 & 297 \\
FedAdam & 82.1\% & 0.566 & 87 \\
FedYogi & 83.2\% & 0.543 & 93 \\ \hline
\textbf{\textsc{FedGSP}} & \textbf{85.4\%} & \textbf{0.453} & \textbf{34} \\
\hline
\end{tabular}}
\captionof{table}{Comparison of accuracy, loss, and rounds required to reach $80\%$ accuracy.}
\label{table:fedgsp-vs-others}
\end{minipage}
\end{minipage}
\section{Introduction}
Federated learning (FL) \cite{kairouz2019advances}, as a privacy-preserving collaborative paradigm for training machine learning (ML) models with data scattered across a large number of data owners, has attracted increasing attention from both academia and industry. Under FL, data owners (a.k.a. FL clients) submit their local ML models to the FL server for aggregation, while local data remain private. FL has been applied in fields which are highly sensitive to data privacy, including healthcare \cite{xu2021federated}, manufacturing \cite{khan2021federated} and next generation communication networks \cite{lim2020federated}. In practical applications, FL clients' local data distributions can be highly heterogeneous due to diverse usage patterns. This problem is referred to as the non-independent and identically distributed (non-i.i.d.) data challenge, which negatively affects training convergence and the performance of the resulting FL model\cite{zhao2018federated}.
Recently, heterogeneous federated learning approaches have been proposed in an attempt to address this challenge. These works try to make class distributions of different FL clients similar to improve the performance of the resulting FL model. In \cite{zhao2018federated, yao2019federated, yoshida2020hybridfl}, FL clients share a small portion of local data to build a common meta dataset to help correct deviations caused by non-i.i.d. data. In \cite{jeong2018communication, duan2019astraea}, data augmentation is performed for categories with fewer samples to reduce the skew of local datasets. These methods are vulnerable to privacy attacks as misbehaving FL servers or clients can easily compromise the shared private data and the augmentation process. To align client data distributions without exposing the FL process to privacy risks, we group together heterogeneous FL clients so that each group can be perceived as a homogeneous ``client'' to participate in FL. This process does not involve any manipulation of private data itself and is therefore more secure.
An intuitive approach to achieve this goal is to assign FL clients to groups with similar overall class distribution, and use collaborative training to coordinate model training within and among groups. However, designing such an approach is not trivial due to the following two challenges. Firstly, assigning FL clients to a specified number of groups of equal group sizes to minimize the data divergence among groups (which can be reduced from the well-known bin packing problem \cite{garey1978strong}) is an NP-hard problem. Moreover, such group assignment process needs to be performed periodically in a dynamic FL environment, which introduces higher requirements for its effectiveness and execution efficiency. Secondly, even if the data distributions among groups are forced to be homogeneous, the data within each group can still be skewed. Due to the robustness of sequential training mode (STM) to data heterogeneity, some collaborative training approaches (e.g., \cite{duan2019astraea}) adopt STM within a group to train on skewed client data. Then, the typical parallel training mode (PTM) can be applied among homogeneous groups. These methods are promising, but are still limited due to their static properties, which prevents them from adapting to the changing needs of FL at different stages. In FL, STM should be emphasized in the early stage to achieve a rapid increase in accuracy in the presence of non-i.i.d. data, while PTM should be emphasized in the later stage to promote convergence. In the static mode, the above parallelism degree must be carefully designed to realize a proper trade-off between sensitivity to heterogeneous data of PTM and overfitting of STM. Otherwise, the FL model performance may suffer.
To address these challenges, this paper proposes a new concept of dynamic collaborative Sequential-to-Parallel (STP) training to improve FL model performance in the presence of non-i.i.d. data. The core idea of STP is to force STM to be gradually transformed into PTM as FL model training progresses. In this way, STP can better refine unbiased model knowledge in the early stage, and promote convergence while avoiding overfitting in the later stage. To support the proposed STP, we propose a \underline{Fed}erated \underline{G}rouped \underline{S}equential-to-\underline{P}arallel (\textsc{FedGSP}) training framework. \textsc{FedGSP}~allows reassignment of FL clients into more groups in each training round, and introduces group managers to manage the dynamically growing number of groups. It also coordinates model training and transmission within and among groups. In addition, we propose a novel Inter-Cluster Grouping (ICG) method to assign FL clients to a pre-specified number of groups, which uses the centroid equivalence theorem to simplify the original NP-hard grouping problem into a solvable constrained clustering problem with equal group size constraint. ICG~can find an effective solution with high efficiency (with a time complexity of $\mathcal{O}(\frac{K^6\mathcal{F}\tau}{M^2}\log{Kd})$). We evaluate \textsc{FedGSP}~on the most widely adopted non-i.i.d. benchmark dataset FEMNIST\cite{caldas2018leaf} and compare it with seven state-of-the-art approaches including FedProx\cite{li2018federated}, FedMMD\cite{yao2018two}, FedFusion\cite{yao2019towards}, IDA\cite{yeganeh2020inverse}, FedAdam, FedAdagrad and FedYogi\cite{reddi2020adaptive}. The results show that \textsc{FedGSP}~improves model accuracy by 3.7\% on average, and reduces training time and communication overhead by more than 90\%. To the best of our knowledge, \textsc{FedGSP}~is the first dynamic collaborative training approach for FL.
\section{Federated Grouped Sequential-to-Parallel Learning}
In this section, we first describe the concept and design of the STP approach. Then, we present the \textsc{FedGSP}~framework which is used to support STP. Finally, we mathematically formulate the group assignment problem in STP, and present our practical solution ICG.
\subsection{STP: The Sequential-to-Parallel Training Mode}\label{section:stp}
Under our grouped FL setting, FL clients are grouped such that clients in the same group have heterogeneous data but the overall data distributions among the groups are homogeneous. Due to the difference in data heterogeneity, the training modes within and among groups are designed separately. We refer to this jointly designed FL training mode as the ``collaborative training mode''.
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{pic/sequential-to-parallel.pdf}
\caption{An example of STP. The ML model in each group is trained in sequence, while ML models among groups are trained in parallel. In each round $r$, the group number grows according to function $f$, and FL clients are regrouped and shuffled.}
\label{fig:sequential-to-parallel}
\end{figure}
Intuitively, the homogeneous groups can be trained in a simple parallel mode PTM because the heterogeneity of their data has been eliminated by client grouping. Instead, for FL clients in the same group whose local data are still skewed, the sequential mode STM can be useful. In STM, FL clients train the model in a sequential manner. A FL client receives the model from its predecessor client and delivers the local trained model to its successor client to continue training. In the special case of training with only one local epoch (i.e., $e=1$), STM is equivalent to centralized SGD, which gives it robustness against data heterogeneity.
This naive collaborative training mode is static and has limitations. Therefore, we extend it to propose a more dynamic approach STP. As shown in Figure \ref{fig:sequential-to-parallel}, STP reassigns FL clients into $f(r)$ groups and shuffles their order in each round $r$, where $f$ is a pre-specified group number growth function, with the goal to dynamically adjust the degree of parallelism. Then, STP can be smoothly transformed from (full) sequential mode to (full) parallel mode. This design can prevent catastrophic forgetting caused by the long ``chain of clients'' that causes the FL model to forget the data of previous clients and overfit the data of subsequent clients, and can also prevent the FL model from learning interfering information such as the order of clients. Moreover, the growing number of groups improves the parallelism efficiency, which promotes convergence and speeds up training when the global FL model is close to convergence.
\begin{algorithm}[t]
\caption{\colorbox{green}{\textsc{\textbf{Sequential-To-Parallel}}} \textbf{(main)}}\label{alg:stp}
\begin{algorithmic}[1]
\Require All FL clients $\mathcal{C}$, the total number of FL clients $K$, the maximum training rounds $R$, the group number growth function $f$, the group sampling rate $\kappa$.
\Ensure The well-trained global FL model $\omega_{\mathrm{global}}^{R}$.
\vspace{1mm}
\State Initialize the global FL model $\omega_{\mathrm{global}}^{0}$;
\For{each round $r=1,\cdots,R$}
\State Reassign all FL clients $\mathcal{C}$ to $f(r)$ groups to obtain $\mathcal{G}$,\label{code:tab1}\\
\qquad\qquad\qquad\qquad $\mathcal{G}\gets$\Call{\colorbox{yellow}{\textbf{Inter-Cluster-Grouping}}}{$\mathcal{C}, K, f, r$};
\State Randomly sample a subset of groups $\tilde{\mathcal{G}}\subset\mathcal{G}$ with proportion $\kappa$;\label{code:tab2}
\For{each group $\mathcal{G}_{m}$ \textbf{in} $\tilde{\mathcal{G}}$ in parallel}
\State The first FL client $\mathcal{C}_m^1$ in $\mathcal{G}_m$ initializes $\omega_m^1\gets\omega_{\mathrm{global}}^{r-1}$;\label{code:tab3}
\For{each FL client $\mathcal{C}_m^k$ \textbf{in} $\mathcal{G}_m$ in sequence}
\State Train $\omega_m^k$ on local data $\mathcal{D}_m^k$ using mini-batch SGD for one epoch;\label{code:tab4}
\State Send the trained $\omega_m^{k+1}\gets\omega_m^k$ to the next FL client $\mathcal{C}_m^{k+1}$;\label{code:tab5}
\EndFor
\State The last FL client $\mathcal{C}_m^{K/f(r)}$ in $\mathcal{G}_m$ uploads $\omega_m^{K/f(r)}$;\label{code:tab6}
\EndFor
\State Update the global FL model using the aggregation $\omega_{\mathrm{global}}^{r}\gets\frac{\sum_{\forall \mathcal{G}_m\in\tilde{\mathcal{G}}}{(\omega_m^{K/f(r)}})}{f(r)}$;\label{code:tab7}
\EndFor
\State \Return $\omega_{\mathrm{global}}^{R}$;\label{code:tab8}
\end{algorithmic}
\end{algorithm}
The pseudo code of STP is given in Algorithm \ref{alg:stp}. In round $r$, STP divides all FL clients into $f(r)$ groups using the ICG~grouping algorithm (Line \ref{code:tab1}), which will be described in Section \ref{section:grouping}. Due to the similarity of data among groups, each group can independently represent the global distribution, so only a small proportion of $\kappa$ groups are required to participate in each round of training (Line \ref{code:tab2}). The first FL client in each group pulls the global model from the FL server (Line \ref{code:tab3}), and trains its local model using mini-batch SGD for one epoch (Line \ref{code:tab4}). The trained local model is then delivered to the next FL client to continue training (Line \ref{code:tab5}), until the last FL client is reached. The last FL client in each group sends the trained model to the FL server (Line \ref{code:tab6}). Models from all groups are aggregated to update the global FL model (Line \ref{code:tab7}). The above steps repeat until the maximum training round $R$ is reached. Finally, the well-trained global FL model is obtained (Line \ref{code:tab8}).
The choice of the growth function for the number of groups, $f$, is critical for the performance of STP. We give three representative growth functions, including linear (smooth grow), logarithmic (fast first and slow later), and exponential (slow first and fast later) growth functions:
\begin{align}
& \mathrm{Linear\ Growth\ Function:} & f(r)=&\beta\left\lfloor\alpha (r-1)+1\right\rfloor, \\
& \mathrm{Log\ Growth\ Function:} & f(r)=&\beta\left\lfloor\alpha\ln r+1\right\rfloor, \\
& \mathrm{Exp\ Growth\ Function:} & f(r)=&\beta\lfloor(1+\alpha)^{r-1}\rfloor,
\end{align}
where the real number coefficient $\alpha$ controls the growth rate, and the integer coefficient $\beta$ controls the initial number of groups and the growth span. We recommend to initialize $\alpha$, $\beta$ to a moderate value and explore the best setting in an empirical manner.
\subsection{\textsc{FedGSP}: The Grouped FL Framework To Enable STP}\label{section:fedgsp}
In this section, we describe the \textsc{FedGSP}~framework that enables dynamic STP. \textsc{FedGSP}~is generally a grouped FL framework that supports dynamic group management, as shown in Figure \ref{fig:framework}. The basic components include a top server (which acts as an FL server and performs functions related to group assignment) and a large number of FL clients. FL clients can be smart devices with certain available computing and communication capabilities, such as smart phones, laptops, mobile robots and drones. They collect data from the surrounding environment and use the data to train local ML models.
In addition, \textsc{FedGSP}~creates group managers to facilitate the management of the growing number of groups in STP. The group managers can be virtual function nodes deployed in the same machine as the top server. Whenever a new group is built, a new group manager is created to assist the top server to manage this group by performing the following tasks:
\textbf{1. Collect distribution information.} The group manager needs to collect class distributions of FL clients and report them to the top server. These meta information will be used to assign FL clients to $f(r)$ groups via ICG.
\textbf{2. Coordinate model training.} The group manager needs to coordinate the sequential training of FL clients in its group, as well as the parallel training with other groups, according to the rules of STP. Specifically, it needs to shuffle the order of clients and report resulting model to the top server for aggregation.
\textbf{3. Schedule model transmission.} In applications such as Industrial IoT systems, wireless devices can directly communicate with each other through wireless sensor networks (WSNs). However, this cannot be realized in most scenarios. Therefore, the group manager needs to act as a communication relay to schedule the transmission of ML models from one client to another.
\begin{figure}[t]
\centering
\includegraphics[width=.9\textwidth]{pic/framework.pdf}
\caption{An overview of the \textsc{FedGSP}~framework.}
\label{fig:framework}
\end{figure}
\subsection{ICG: The Inter-Cluster Grouping Algorithm\label{section:grouping}}
As required by STP, the equally sized groups containing heterogeneous FL clients should have similar overall class distributions. To achieve this goal, in this section, we first formalize the FL client grouping problem which is NP-hard, and then explain how to simplify to propose the ICG~approach.
\vspace{2mm}
\noindent \textbf{(A) Problem Modeling}
\vspace{2mm}
\noindent Considering an $\mathcal{F}$-class classification task involving $K$ FL clients, STP needs to assign these clients to $M$ groups, where $M$ is determined by the group number growth function $f$ and the current round $r$. Our goal is to find a grouping strategy $\mathbf{x}\in\mathbb{I}^{M\times K}$ in the 0-1 space $\mathbb{I}=\{0,1\}$ to minimize the difference in class distributions of all groups, where $\mathbf{x}_m^k=1$ represents the device $k$ is assigned to the group $m$, $\mathcal{V}\in(\mathbb{Z^+})^{\mathcal{F}\times K}$ is the class distribution matrix composed of $\mathcal{F}$-dimensional class distribution vectors of $K$ FL clients, $\mathcal{V}_m\in(\mathbb{Z}^+)^{\mathcal{F}\times 1}$ represents the overall class distribution of group $m$, and $\left<\cdot,\cdot\right>$ represents the distance between two class distributions. The problem can be formalized as follows:
\begin{align}
\underset{\mathbf{x}}{\mathrm{minimize}}\qquad & z=\sum_{m_1=1}^{M-1}\sum_{m_2=m_1+1}^{M}<\mathcal{V}_{m_1},\mathcal{V}_{m_2}>, \label{eq:objective}\\
\mathrm{s.t.}\qquad & M=f(r), \label{eq:group-number}\\
& \sum_{k=1}^{K}\mathbf{x}_m^k\le \left\lceil \frac{K}{M} \right\rceil \quad \forall m=1,\cdots,M, \label{eq:group-capacity}\\
& \sum_{m=1}^{M}{\mathbf{x}_m^k}=1 \qquad\quad \forall k=1,\cdots,K, \label{eq:client-conflict}\\
& \mathcal{V}_m=\sum_{k=1}^{K}{\mathbf{x}_m^k\mathcal{V}^k}\quad \forall m=1,\cdots,M, \label{eq:overall-dist}\\
& \mathbf{x}_m^k\in\{0,1\}, ~k\in[1,K], ~m\in[1,M].
\label{eq:variable-constraint}
\end{align}
Constraint \eqref{eq:group-number} ensures that the number of groups $M$ meets $f(r)$ required by STP. Constraint \eqref{eq:group-capacity} ensures that the groups have similar or equal size $\left\lceil \frac{K}{M} \right\rceil$. Constraint \eqref{eq:client-conflict} ensures that each client can only be assigned to one group at a time. The overall class distribution $\mathcal{V}_m$ of the group $m$ is defined by Eq. \eqref{eq:overall-dist}, where $\mathcal{V}^k\in\mathcal{V}$ is the class distribution vector of client $k$. Constraint \eqref{eq:variable-constraint} restricts the decision variable $\mathbf{x}$ to only take up a value of 0 or 1.
\begin{proposition}\label{prop:bpp}
The NP-hard bin packing problem (BPP) can be reduced to the grouping problem in Eq. \eqref{eq:objective} to Eq. \eqref{eq:variable-constraint}, making it also an NP-hard problem.
\end{proposition}
\begin{proof}
The problem stated by Eq. \eqref{eq:objective} to Eq. \eqref{eq:variable-constraint} is actually a BPP with additional constraints, where $K$ items with integer weight $\mathcal{V}^k$ and unit volume should be packed into the minimum number of bins of integer capacity $\left\lceil \frac{K}{M} \right\rceil$. The difference is that Eq. \eqref{eq:objective} to Eq. \eqref{eq:variable-constraint} restricts the number of available bins to $M$ instead of unlimited, and the difference in the bin weights not to exceed $\xi$. The input and output of BPP and Eq. \eqref{eq:objective} to Eq. \eqref{eq:variable-constraint} are matched, with only additional $\mathcal{O}(1)$ transformation complexity to set $M$ and $\xi$ to infinity. Therefore, BPP can call the solution of Eq. \eqref{eq:objective} to Eq. \eqref{eq:variable-constraint} in $\mathcal{O}(1)$ time to obtain its solution, which proves that the NP-hard BPP \cite{garey1978strong} can be reduced to the problem stated by Eq. \eqref{eq:objective} to Eq. \eqref{eq:variable-constraint}. Therefore, Eq. \eqref{eq:objective} to Eq. \eqref{eq:variable-constraint} is also an NP-hard problem.
\end{proof}
Therefore, it is almost impossible to find the optimal solution within a polynomial time. To address this issue, we adopt the centroid equivalence theorem to simplify the original problem to a constrained clustering problem.
\vspace{2mm}
\noindent \textbf{(B) Inter-Cluster Grouping (ICG)}
\vspace{2mm}
Consider a constrained clustering problem with $K$ points and $L$ clusters, where the size of all clusters is strictly the same $K/L$.
\begin{assumption}\label{icg-assumption}
We make the following assumptions:
\begin{enumerate}
\item $K$ is divisible by $L$;
\item Take any point $\mathcal{V}^m_l$ from cluster $l$, the squared $l_2$-norm distance $\|\mathcal{V}^m_l-C_l\|_2^2$ between the point $\mathcal{V}^m_l$ and its cluster centroid $C_l$ is bounded by $\sigma_l^2$.
\item Take one point $\mathcal{V}^m_l$ from each of $L$ clusters at random, the sum of deviations of each point from its cluster centroid $\epsilon^m=\sum_{l=1}^{L}(\mathcal{V}^m_l-C_l)$ meets $\mathbf{E}\left[\epsilon^m\right]=0$.
\end{enumerate}
\end{assumption}
\begin{definition}[Group Centroid]\label{def:group-centroid}
Given $L$ clusters of equal size, let group $m$ be constructed from one point randomly sampled from each cluster $\{\mathcal{V}^m_1,\cdots,\mathcal{V}^m_L\}$. Then, the centroid of group $m$ is defined as $C^m=\frac{1}{L}\sum_{l=1}^{L}\mathcal{V}^m_l$.
\end{definition}
\begin{proposition}\label{prop:centroid}
If Assumption \ref{icg-assumption} holds, suppose the centroid of cluster $l$ is $C_l=\frac{L}{K}\sum_{i=1}^{K/L}{\mathcal{V}_l^i}$ and the global centroid is $C_{\mathrm{global}}=\frac{1}{L}\sum_{l=1}^{L}{C_l}$. We have:
\begin{enumerate}
\item The group and global centroids are expected to coincide, $\mathbf{E}[C^m]=C_{\mathrm{global}}$.
\item The error $\|C^m-C_\mathrm{global}\|_2^2$ between the group and global centroids is bounded by $\frac{1}{L^2}\sum_{l=1}^{L}{\sigma_l^2}$.
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{align}
\nonumber
\mathbf{E}[C^m]&=\mathbf{E}[\frac{1}{L}\sum_{l=1}^{L}\mathcal{V}^m_l] =\mathbf{E}[\frac{1}{L}\sum_{l=1}^{L}(\mathcal{V}^m_l-C_l+C_l)] \\
\nonumber
&=\mathbf{E}[\frac{1}{L}\sum_{l=1}^{L}(\mathcal{V}^m_l-C_l)+\frac{1}{L}\sum_{l=1}^{L}{C_l}]=\frac{1}{L}\mathbf{E}[\epsilon^m]+C_{\mathrm{global}}=C_{\mathrm{global}}, \\
\nonumber
\|C^m-C_\mathrm{global}\|_2^2&=\|\frac{1}{L}\sum_{l=1}^{L}{\mathcal{V}_l^m}-\frac{1}{L}\sum_{l=1}^{L}{C_l}\|_2^2=\frac{1}{L^2}\|\sum_{l=1}^{L}{(\mathcal{V}_l^m-C_l)}\|_2^2 \\
\nonumber
&\le\frac{1}{L^2}\sum_{l=1}^{L}\|\mathcal{V}_l^m-C_l\|_2^2=\frac{1}{L^2}\sum_{l=1}^{L}{\sigma_l^2}.
\end{align}
\end{proof}
Proposition \ref{prop:centroid} indicates that there exists a grouping strategy $\tilde{\mathbf{x}}$ and $\mathcal{V}_{m_1}=\sum_{k=1}^{K}\tilde{\mathbf{x}}_{m_1}^k\mathcal{V}^k=LC^{m_1}$, $\mathcal{V}_{m_2}=\sum_{k=1}^{K}\tilde{\mathbf{x}}_{m_2}^k\mathcal{V}^k=LC^{m_2}$ ($\forall m_1\ne m_2$), so that the objective in Eq. \eqref{eq:objective} turns to $z=\sum_{m_1\ne m_2}L<C^{m_1},C^{m_2}>$ and the expectation value reaches 0. This motivates us to use the constrained clustering model to solve $\tilde{\mathbf{x}}$ in the objective Eq. \eqref{eq:objective}. Therefore, we consider the constrained clustering problem below,
\begin{align}
\underset{\mathbf{y}}{\mathrm{minimize}}\qquad & \sum_{k=1}^{K}\sum_{l=1}^{L}\mathbf{y}^k_l\cdot\left(\frac{1}{2}\|\mathcal{V}^k-C_l\|_2^2\right), \label{eq:ccp-objective}\\
\mathrm{s.t.}\qquad & \sum_{k=1}^{K}\mathbf{y}_l^k=\frac{K}{L} \qquad \forall l=1,\cdots,L, \label{eq:ccp-least}\\
& \sum_{l=1}^{L}\mathbf{y}_l^k=1 \qquad\quad \forall k=1,\cdots,K, \label{eq:ccp-client}\\
& \mathbf{y}_l^k\in\{0,1\}, ~k\in[1,K], ~l\in[1,L], \label{eq:ccp-value}
\end{align}
where $\mathbf{y}\in\mathbb{I}^{L\times K}$ is a selector variable, $\mathbf{y}_l^k=1$ means that client $k$ is assigned to cluster $l$ while 0 means not, $C_l$ represents the centroid of cluster $l$. Eq. \eqref{eq:ccp-objective} is the standard clustering objective, which aims to assign $K$ clients to $L$ clusters so that the sum of the squared $l_2$-norm distance between the class distribution vector $\mathcal{V}^k$ and its nearest cluster centroid $C_l$ is minimized. Constraint \eqref{eq:ccp-least} ensures that each cluster has the same size $\frac{K}{L}$. Constraint \eqref{eq:ccp-client} ensures that each client can only be assigned to one cluster at a time. In this simplified problem, Constraint \eqref{eq:client-conflict} is relaxed to $\sum_{m=1}^{M}{\mathbf{x}_m^k}\le 1$ to satisfy the assumption that $K/L$ is divisible.
The above constrained clustering problem can be modeled as a minimum cost flow (MCF) problem and solved by network simplex algorithms \cite{bradley2000constrained}, such as \textsc{SimpleMinCostFlow} in Google OR-Tools. Then, we can alternately perform cluster assignment and cluster update to optimize $\mathbf{y}_l^k$ and $C_l (\forall k,l)$, respectively. Finally, we construct $M$ groups, each group consists of one client randomly sampled from each cluster without replacement, so that their group centroids are expected to coincide with the global centroid. The pseudo code is given in Algorithm \ref{alg:icg}. ICG~has a complexity of $\mathcal{O}(\frac{K^6\mathcal{F}\tau}{M^2}\log{Kd})$, where $d=\max\{\sigma_l^2 | \forall l\in[1,L]\}$, and $K,M,\mathcal{F},\tau$ are the number of clients, groups, categories, and iterations, respectively. In our experiment, ICG~is quite fast, and it can complete group assignment within only 0.1 seconds, with $K=364,M=52,\mathcal{F}=62$ and $\tau=10$.
\begin{algorithm}[t]
\caption{\colorbox{yellow}{\textsc{\textbf{Inter-Cluster-Grouping}}}}\label{alg:icg}
\begin{algorithmic}[1]
\Require All FL clients $\mathcal{C}$ (with attribute $\mathcal{V}^k$), the total number of FL clients $K$, the group number growth function $f$, the current training round $r$.
\Ensure The grouping strategy $\mathcal{G}$.
\vspace{1mm}
\State{Randomly sample $L\cdot\lfloor\frac{K}{L}\rfloor$ clients from $\mathcal{C}$ to meet Assumption \ref{icg-assumption}, where $L=\lfloor\frac{K}{f(r)}\rfloor$;}
\Repeat
\State \textsc{Cluster Assignment:} Fix the cluster centroid $C_l$ and optimize $\mathbf{y}$ in Eq. \eqref{eq:ccp-objective} to Eq. \eqref{eq:ccp-value};
\State \textsc{Cluster Update:} Fix $\mathbf{y}$ and update the cluster centroid $C_l$ as follows, $$C_l\gets\frac{\sum_{k=1}^{K}{\mathbf{y}_l^k\mathcal{V}^k}}{\sum_{k=1}^{K}{\mathbf{y}_l^k}} \quad \forall l=1,\cdots,L;$$
\Until{$C_l$ converges;}
\State \textsc{Group Assignment:} Randomly sample one client from each cluster without replacement to construct group $\mathcal{G}_m (\forall m=1,\cdots,f(r))$;
\State \Return $\mathcal{G}=\{\mathcal{G}_1,\cdots,\mathcal{G}_{f(r)}\}$;
\end{algorithmic}
\end{algorithm}
\section{Related Work}
Existing heterogeneous FL solutions can be divided into three main categories: 1) data augmentation, 2) clustering-based learning, and 3) adaptive optimization.
\textbf{Data Augmentation:} Zhao et al. \cite{zhao2018federated} proved that the FL model accuracy degradation due to heterogeneous local data can be quantified by the earth move distance (EMD) between the client and global data distributions. This result motivates some research works to balance the sample size of each class through data augmentation. Zhao et al. \cite{zhao2018federated} proposed to build a globally shared dataset to expand client data. Jeong et al. \cite{jeong2018communication} used the conditional generative network to generate new samples for categories with fewer samples. Similarly, Duan et al. \cite{duan2019astraea} used augmentation techniques such as random cropping and rotation to expand client data. These methods are effective in improving the FL model accuracy by reducing data skew. However, they involve modifying clients' local data, which can lead to serious privacy risks.
\textbf{Clustering-based Learning:} Another promising way to reduce data heterogeneity is through clustering-based FL. Sattler et al. \cite{sattler2020clustered} groups FL clients with similar class distributions into one cluster, so that FL clients with dissimilar data distributions do not interfere with each other. This method works well in personalized FL \cite{fallah2020personalized} where FL is perform within each cluster and an FL model is produced for each cluster. However, it is not the same as our goal which is to train one shared FL model that can be generalized to all FL clients. Duan et al. \cite{duan2019astraea} makes the KullbackLeibler divergence of class distributions similar among clusters, and proposed a greedy best-fit strategy to assign FL clients.
\textbf{Adaptive Optimization:} Other research explores adaptive methods to better merge and optimize client- and server-side models. On the client side, Li et al. \cite{li2018federated} added a proximal penalty term to the local loss function to constrain the local model to be closer to the global model. Yao et al. \cite{yao2018two} adopted a two-stream framework and used transfer learning to transfer knowledge from the global model to the local model. A feature fusion method has been further proposed to better merge the features of local and global models \cite{yao2019towards}. On the server side, Yeganeh et al. \cite{yeganeh2020inverse} weighed less out-of-distribution models based on inverse distance coefficients during aggregation. Instead, Reddi et al. \cite{reddi2020adaptive} focused on server-side optimization and introduced three advanced adaptive optimizers (Adagrad, Adam and Yogi) to obtain FedAdagrad, FedAdam and FedYogi, respectively. These methods perform well in improving FL model convergence.
Solutions based on data augmentation are at risky due to potential data leakage, while solutions based on adaptive optimization do not solve the problem of class distribution divergence causing FL model performance to degrade. \textsc{FedGSP}~focuses on clustering-based learning. Different from existing research, it takes a novel approach of dynamic collaborative training, which allows dynamic scheduling and reassignment of clients into groups according to the changing needs of FL.
|
\section{Introduction}
Two-component spinors are widely used in four-dimensional Lorentzian geometry.
The seminal books `Spinors and space-time'~\cite{OT,NT} are devoted to such
usage. At a very basic level, two-component spinors arise via the 2--1 covering
of Lie groups
\[{\mathrm{SL}}(2,{\mathbb{C}})\longrightarrow{\mathrm{SO}}^\uparrow(3,1),\]
where ${\mathrm{SO}}^\uparrow(3,1)$ is the identity-connected component of the
Lorentz group. Similarly, in three dimensions, the 2--1 covering
\[{\mathrm{SL}}(2,{\mathbb{R}})\longrightarrow{\mathrm{SO}}^\uparrow(2,1)\]
is responsible for the utility of two-component spinors in three dimensions
(as in Section~\ref{3d_conformal}).
At a very basic level, the two-component spinors in this article arise via
inclusions
\[{\mathrm{SL}}(2,{\mathbb{R}})\hookrightarrow{\mathrm{Sp}}(4,{\mathbb{R}}),\]
where ${\mathrm{Sp}}(4,{\mathbb{R}})$ is the subgroup of
${\mathrm{SL}}(4,{\mathbb{R}})$ consisting of matrices that preserve a fixed
non\-degenerate skew form on ${\mathbb{R}}^4$. There are two such inclusions:
\begin{itemize}\itemsep=0pt
\item by writing ${\mathbb{R}}^4=\bigodot^3{\mathbb{R}}^2$, where $\bigodot$
denotes symmetric tensor product,
\item by writing ${\mathbb{R}}^4={\mathbb{R}}^2\oplus{\mathbb{R}}^2$,
\end{itemize}
giving rise to $G_2$ contact geometry (as in Section~\ref{G2}) and Legendrean
contact geometry (as in Section~\ref{CL}), respectively. These two geometries are
defined on a five-dimensional contact mani\-fold~$M$ as extra structure on the
contact distribution~$H\subset TM$. A contact form $\alpha$ is a $1$-form so
that $H=\ker\alpha$. It gives rise to a nondegenerate skew form, the Levi
form, namely ${\rm d}\alpha|_H$. The extra spin structures in Sections~\ref{G2} and~\ref{CL} are required to be compatible with the Levi form.
Another important theme in \cite{OT,NT} is conformal geometry. It is concerned
with what happens if the (Lorentzian) metric is replaced by a smooth positive
multiple of itself. The resulting formul{\ae}, for
example~(\ref{change_of_spin_connection}), fit very well with spinors. For the
contact geometries in this article, the corresponding freedom is in choosing a
contact form~$\alpha$. The resulting formul{\ae}, for
example~(\ref{change_of_G2_spin_connection}), (\ref{full_transformation_law}),
(\ref{full_barred_transformation_law}), also fit very well with spinors. These
`conformal' structures are examples of {\em parabolic geometries}~\cite{CS}.
In particular, {\em invariant differential operators} play a key part in
parabolic constructions.
Ideally, one would like to approach the natural differential geometric calculus
on these various geometries via invariant differential operators. More
specifically, for a chosen `scale' (a metric in the conformal class or a choice
of contact form) one expects a canonical (partial) connection on all the
natural irreducible vector bundles. This expectation follows from the
\v{C}ap-Slov\'ak theory of {\em Weyl structures and scales} in parabolic
geometry~(see \cite{CSweyl} or \cite[Section~5.1]{CS}). In conformal geometry, it is
just the Levi-Civita connection: see Sections~\ref{LC} and~\ref{conf} for details.
For parabolic contact structures, one expects partial connections defined only
in the contact directions. We shall see in Sections~\ref{G2} and~\ref{CL} that a
suitable collection of invariant operators is, indeed, sufficient for these
purposes. For the five-dimensional Legendrean contact structures
in Section~\ref{CL}, these operators can easily be found (within the Rumin complex,
explained in Section~\ref{rumin}). The two key invariant operators~(\ref{basic_invariant_operators}) in~$G_2$ contact geometry remain somewhat
mysterious: it is shown in Section~\ref{G2} that these operators are, indeed,
invariant but we have not been able to find a~suitable shortcut to their
construction. Nevertheless, we are able to construct, in Section~\ref{saucers}, the
general~$G_2$ contact geometry from a suitable Legendrean contact structure
and, thereby, calculate its basic spinor invariant (a certain septic).
Through this article we shall use Penrose's abstract index notation and other
conventions from~\cite{OT}. Also, to ease the notational burden, we shall not
carefully distinguish between bundles and sections thereof, for example writing
$\omega_a\in\Wedge^1$ instead of $\omega_a\in\Gamma\big({\Wedge}^1\big)$ or even
$\omega_a\in\Gamma\big(M,\Wedge^1\big)$, for a $1$-form~$\omega_a$. Especially as this
article is concerned with {\em local} differential geometry, this should
cause no confusion.
\section{The Levi-Civita connection}\label{LC}
On a general smooth manifold, the exterior derivative and the Lie derivative
are defined independently of local co\"ordinates (and there is little else with
this property~\cite{KMS}). Both of these operations can be defined in terms of
an arbitrary torsion-free affine connection. For the exterior derivative on
$1$-forms, we have (following the conventions of~\cite[equation~(4.3.14)]{OT})
\[\omega_b\longmapsto\nabla_{[a}\omega_{b]}.\]
For the Lie derivative on covariant $2$-tensors, we have
\begin{equation}\label{Lie}
\phi_{bc}\longmapsto({\mathcal{L}}_X\phi)_{bc}\equiv
X^a\nabla_a\phi_{bc}+(\nabla_bX^a)\phi_{ac}+(\nabla_cX^a)\phi_{ba},
\end{equation}
for a vector field~$X^a$ and, irrespective of the usual interpretation of
${\mathcal{L}}_X$ in terms of the flow of~$X^a$, it easy to check that this
expression does not depend on the choice of~$\nabla_a$. It is convenient to
regard the right hand side of~(\ref{Lie}), let's say for a symmetric
tensor~$\phi_{ab}$, as an invariantly defined differential {\em pairing}
\[\textstyle TM\times\bigodot^2\!\Wedge^1\longrightarrow
\bigodot^2\!\Wedge^1.\]
In particular, if $g_{ab}$ is a semi-Riemannian metric, that is to say a
nondegenerate symmetric form, then the Lie derivative of~$g_{ab}$, with a
convenient factor of~$\frac12$ thrown in, can be regarded as a~canonically
defined linear differential operator
\begin{equation}\label{un}
\textstyle TM\ni
X^a\longmapsto
\frac12{\mathcal{L}}_Xg_{bc}\in\bigodot^2\!\Wedge^1.\end{equation}
Of course,
the tensor $g_{ab}$ also defines an isomorphism $TM\cong\Wedge^1$ by
$X^a\mapsto g_{ab}X^b$, and so we have obtained an invariantly defined linear
differential operator
\begin{equation}\label{deux}
\textstyle\Wedge^1\cong TM\longrightarrow\bigodot^2\!\Wedge^1.\end{equation}
In combination with the exterior derivative ${\rm d}\colon \Wedge^1\to\Wedge^2$, we have
obtained
\begin{equation}\label{trois}
\textstyle\Wedge^1\longrightarrow\bigodot^2\!\Wedge^1\oplus\Wedge^2=
\Wedge^1\otimes\Wedge^1\end{equation}
and we claim that this is the Levi-Civita connection defined by~$g_{ab}$. This
is easy to check: if we use the Levi-Civita connection in~(\ref{Lie}), then
\begin{align*}\tfrac12{\mathcal{L}}_Xg_{ab}+\nabla_{[a}X_{b]}
&= \tfrac12\big(X^c\nabla_cg_{ab}+(\nabla_aX^c)g_{cb}+(\nabla_bX^c)g_{ac}\big) +\nabla_{[a}X_{b]}\\
&= \nabla_{(a}X_{b)}+\nabla_{[a}X_{b]}=\nabla_aX_b,
\end{align*}
as required. But we can play the moves (\ref{un}), (\ref{deux}), and
(\ref{trois}) to {\em define} the Levi-Civita connection. An advantage of
this viewpoint is that it is easily modified to define other connections (and
partial connections), as we shall soon see.
\section{Conformal differential geometry}\label{conf}
In semi-Riemannian conformal differential geometry, instead of a
metric~$g_{ab}$, we are given only a~{\em conformal class} of metrics. A
convenient way of expressing this is to say that we are given~$\eta_{ab}$, a
nondegenerate section of $\bigodot^2\!\Wedge^1\otimes L^2$ for some line
bundle~$L$. Thus, if $\sigma$ is a non-vanishing section of~$L$, then
$g_{ab}\equiv\sigma^{-2}\eta_{ab}$ is a genuine metric. We shall refer to
$\sigma$ as a {\em scale} and, if we choose a different section
$\widehat\sigma=\Omega^{-1}\sigma$ for some nowhere vanishing
function~$\Omega$, then we encounter a {\em rescaled} metric
$\widehat{g}_{ab}\equiv\widehat\sigma^{-2}\eta_{ab}=\Omega^2g_{ab}$. For a~given conformal structure~$\eta_{ab}$, let us define~$\eta^{ab}$, a~section of
$\bigodot^2\!TM\otimes L^{-2}$, by $\eta_{ac}\eta^{bc}=\delta_a{}^b$, where
$\delta_a{}^b$ is the invariant (Kronecker delta) pairing
$TM\otimes\Wedge^1\to\Wedge^0$.
On an $n$-dimensional oriented conformal manifold, we may attempt to normalise
a section $\epsilon_{ab\dots d}$ of the bundle $\Wedge^n$, compatible
with the orientation, by insisting that
\begin{equation}\label{normalise}
\epsilon_{ab\dots d}\epsilon_{ef\dots h}\eta^{ae}\eta^{bf}\cdots\eta^{dh}
\equiv n! ,\end{equation}
just as we may do on an $n$-dimensional oriented Riemannian manifold to
normalise and define the volume form. The problem with (\ref{normalise}) is
that the left hand side takes values in the line bundle~$L^{-2n}$. To remedy
this problem, we may insist that~$\epsilon_{ab\dots d}$ take values in
$\Wedge^n\otimes L^n$. The normalisation (\ref{normalise}) now makes good
sense and $\epsilon_{ab\dots d}$ is uniquely determined. Equivalently, we may
view $\epsilon_{ab\dots d}$ as providing a canonical identification
\[L^{-n}=\Wedge^n\qquad
\mbox{given by}\quad\phi\mapsto\phi\epsilon_{ab\dots d}.\]
It it usual to write $L\equiv\Wedge^0[1]$ and refer to sections of
$L^w\equiv\Wedge^0[w]$ as {\em conformal densities of weight}~$w$. In the
presence of a scale $\sigma$, defining a metric $g_{ab}=\sigma^{-2}\eta_{ab}$,
the normalisation (\ref{normalise}) reads
\[(\sigma^n\epsilon_{ab\dots d})(\sigma^n\epsilon_{ef\dots h})
g^{ae}g^{bf}\cdots g^{dh}=n!\]
so $\sigma^n\epsilon_{ab\dots d}$ is the usual volume form for the
metric~$g_{ab}$. This is consistent with the scale $\sigma$ trivialising
all the density bundles $\Wedge^0[w]$. To summarise, in the presence of a
scale~$\sigma$, a conformal density of weight $w$ can be regarded as an ordinary
function but if we change to a new scale $\widehat\sigma=\Omega^{-1}\sigma$,
then the same density is represented by a new function $\widehat{f}=\Omega^wf$.
Once we have decided that $\Wedge^0[1]=L=(\Wedge^n)^{-1/n}$, the Lie
derivative ${\mathcal{L}}_X\colon \Wedge^1\to\Wedge^1$ induces invariantly
defined differential pairings $TM\times B\to B$ for all conformally weighted
tensor bundles~$B$ and, in particular,
\[\textstyle TM\times\bigodot^2\!\Wedge^1[2]\to\bigodot^2\!\Wedge^1[2].\]
Recall that a conformal structure is a nondegenerate section $\eta_{ab}$
of this bundle $\bigodot^2\!\Wedge^1[2]$
and hence we obtain a conformally invariant first-order linear differential
operator
\begin{equation}\label{tricky}\textstyle
TM\ni X^a\longmapsto{\mathcal{L}}_X\eta_{ab}\in\bigodot^2\!\Wedge^1[2].
\end{equation}
But recall that $\epsilon_{ab\dots d}$ is the canonical section of
$\Wedge^n[n]=\Wedge^0$ corresponding to the constant function~$1$. It follows
that ${\mathcal{L}}_X\epsilon_{ab\dots d}=0$ for any vector field $X^a$ and
therefore, from~(\ref{normalise}), that $\eta^{ab}{\mathcal{L}}_X\eta_{ab}=0$.
In addition, we may use the conformal metric $\eta_{ab}$ to lower indices, at
the expense of a conformal weight so that $TM=\Wedge^1[2]$. The Lie derivative
in (\ref{tricky}) thus yields a conformal invariant differential operator
\[\textstyle\Wedge^1[2]\longrightarrow\bigodot_\circ^2\!\Wedge^1[2],\]
where $\circ$ denotes the trace-free part (a manifestly conformally invariant
notion). If $\phi_{b\dots d}$ is an $(n-1)$-form then
\[\epsilon_{ab\dots d}\eta^{be}\cdots\eta^{dg}\phi_{e\dots g}\]
is a $1$-form of conformal weight $n-2(n-1)=2-n$. In other words, the bundles
$\Wedge^{n-1}$ and $\Wedge^1[2-n]$ are canonically isomorphic. Similarly, we
have $\Wedge^n=\Wedge^0[-n]$ and so the exterior derivative
${\rm d}\colon \Wedge^{n-1}\to\Wedge^n$ may be viewed as an invariant differential operator
\[\Wedge^1[2-n]\to\Wedge^0[-n].\]
Together with the exterior derivative ${\rm d}\colon \Wedge^1\to\Wedge^2$, we now have
three conformally invariant linear differential operators defined on variously
weighted $1$-forms, namely
\begin{equation}\label{all_three}\textstyle
\Wedge^1[2]\to\bigodot_\circ^2\!\Wedge^1[2],
\qquad\Wedge^1[2-n]\to\Wedge^0[-n],\qquad\Wedge^1\to\Wedge^2.\end{equation}
As in Section~\ref{LC}, we are now in a position to define a preferred
connection in the presence of a scale $\sigma\in\Gamma(\Wedge^0[1])$.
Specifically, we may drop all the weights in (\ref{all_three}) and combine
them to obtain
\[\textstyle\Wedge^1\xrightarrow{\,\nabla\,}\Wedge^1\otimes\Wedge^1
=\bigodot_\circ^2\!\Wedge^1\oplus\Wedge^0\oplus\Wedge^2.\]
The three bundles on the right constitute the orthogonal decomposition of
$\Wedge^1\otimes\Wedge^1$ into irreducibles with respect to the metric
$g_{ab}=\sigma^{-2}\eta_{ab}$. Of course, it is easily verified that we have
found the Levi-Civita connection for $g_{ab}$ but there are two advantages of
this particular route. One is that we are able to read off the change in the
Levi-Civita connection under conformal rescaling
$\sigma\mapsto\widehat\sigma=\Omega^{-1}\sigma$. The other is that this route
may be employed in other {\em parabolic geometries}. Regarding the
first advantage, another way of pinning down the connection defined by
(\ref{all_three}) and a scale $\sigma\in\Gamma\big({\Wedge}^0[1]\big)$, is to require
that $\nabla_a\sigma=0$. This is consistent with the connections induced on
the line bundles $\Wedge^0[w]$ in the sense that, for
$\phi\in\Gamma\big({\Wedge}^0[w]\big)$, we have $\nabla\phi=\sigma^w{\rm d}(\sigma^{-w}\phi)$.
In particular, for a new scale $\widehat\sigma=\Omega^{-1}\sigma$ and
$\phi\in\Gamma\big({\Wedge}^0[w]\big)$, we have
\[\widehat\nabla\phi=\widehat\sigma^w {\rm d}\big(\widehat\sigma^{-w}\phi\big)=
\Omega^{-w}\sigma^w{\rm d}\big(\Omega^w\sigma^{-w}\phi\big)
=\sigma^w{\rm d}\big(\sigma^{-w}\phi\big)+w\Omega^{-1}({\rm d}\Omega)\phi
=\nabla\phi+w\Upsilon\phi,\]
where $\Upsilon\equiv\Omega^{-1}{\rm d}\Omega$. It follows that, regarding the
change in the Levi-Civita connection on a~$1$-form~$\omega_b$, we have
\[\widehat\nabla_a\omega_b
=\nabla_a\omega_b-\Upsilon_a\omega_b-\Upsilon_b\omega_a
+\Upsilon^c\omega_cg_{ab}\]
since then, if $\omega_b$ has conformal weight~$w$, we deduce that
\[\widehat\nabla_a\omega_b
=\nabla_a\omega_b+(w-1)\Upsilon_a\omega_b-\Upsilon_b\omega_a
+\Upsilon^c\omega_cg_{ab},\]
which is exactly so that the three operators (\ref{all_three}), namely
\[\textstyle
\phi_b\mapsto\nabla_{(a}\phi_{b)}-\frac1n(\nabla^c\phi_c)g_{ab},\qquad
\phi_b\mapsto\nabla^b\phi_b,\qquad\phi_b\mapsto\nabla_{[a}\phi_{b]},\]
are invariantly defined.
Regarding the second advantage, it is convenient to adopt a universal notation
for the natural irreducible vector bundles available on a parabolic geometry,
as detailed in~\cite{BE,CS}. Let's see how this works in $5$-dimensional
conformal geometry. The general irreducible spin bundle is
\[\bthree{a}{b}{c},\qquad\mbox{where}\ a\in{\mathbb{R}}\
\mbox{and}\ b,c\in{\mathbb{Z}}_{\geq 0}\]
and for tensor bundles we restrict $c$ to be even. (Roughly speaking, this
indicates a (complex) irreducible representation of
${\mathbb{R}}_{>0}\times{\mathrm{Spin(5)}}$ with the real number $a$ over the
crossed node recording the action of~${\mathbb{R}}_{>0}$.) The de~Rham complex
is
\begin{equation}\label{deRham5dim}
\bthree{0}{0}{0}\to\bthree{-2}{1}{0}\to\bthree{-3}{0}{2}
\to\bthree{-4}{0}{2}\to\bthree{-5}{1}{0}\to\bthree{-5}{0}{0}\end{equation}
and the three conformally invariant first-order linear differential operators
on weighted $1$-forms are
\[\bthree{0}{1}{0}\to\bthree{-2}{2}{0},\qquad
\bthree{-5}{1}{0}\to\bthree{-5}{0}{0},\qquad
\bthree{-2}{1}{0}\to\bthree{-3}{0}{2}.\]
Notice that $\Wedge^0[w]=\bthree{w}{0}{0}$ and that the identification
$\Wedge^5=\Wedge^0[-5]$ is built into the notation. There are many other
advantages of this seemingly arcane notation: its utility in $5$-dimensional
spin geometry is the subject of the following section.
\section{Five-dimensional conformal spin geometry}\label{5d_conformal}
In this section we discuss five-dimensional conformal geometry by means of
spinors (that is to say four-spinors in this dimension (also known as
{\em twistors})). For simplicity, we shall suppose that the conformal metric
has split signature. There is some elementary linear algebra behind this
discussion as follows. Suppose that ${\mathbb{T}}$ is a four-dimensional real
vector space equipped with a nondegenerate skew form $\epsilon_{\alpha\beta}$.
Let us write $\epsilon^{\alpha\beta}$ for its inverse so that
$\epsilon_{\alpha\beta}\epsilon^{\alpha\gamma}=\delta_\alpha{}^\gamma$. The
vector space $\Wedge^2{\mathbb{T}}$ naturally splits:
\begin{equation}\label{symplectic_decomposition}
\Wedge^2{\mathbb{T}}
=\big\{P_{\alpha\beta}\,|\,\epsilon^{\alpha\beta}P_{\alpha\beta}=0\big\}
\oplus\{P_{\alpha\beta}=\lambda\epsilon_{\alpha\beta}\}
\equiv\Wedge_\perp^2{\mathbb{T}}\oplus{\mathbb{R}}.\end{equation}
Otherwise said, if we let ${\mathrm{Sp}}(4,{\mathbb{R}})$ denote the linear
automorphisms of ${\mathbb{T}}\cong{\mathbb{R}}^4$ preserving the symplectic
form $\epsilon_{\alpha\beta}$, then~(\ref{symplectic_decomposition}) is the
decomposition of $\Wedge^2{\mathbb{T}}$ into
${\mathrm{Sp}}(4,{\mathbb{R}})$-irreducibles. The $5$-dimensional vector space
$\Wedge_\perp^2{\mathbb{T}}$ is acquires a split signature metric
\[\|P_{\alpha\beta}\|^2\equiv
P_{\alpha\beta}P_{\gamma\eta}\epsilon^{\alpha\gamma}\epsilon^{\beta\eta}.\]
Otherwise said, we have constructed the isomorphism
${\mathrm{Sp}}(4,{\mathbb{R}})\cong{\mathrm{Spin}}(3,2)$. According to the
Pl\"ucker relations, the null vectors in $\Wedge_\perp^2{\mathbb{T}}$ are
the decomposable tensors.
With the conventions of~\cite{BE}, the bundle version of this discussion yields
the splitting
\[\Wedge^2\big(\bthree{0}{0}{1}\big)=\bthree{0}{1}{0}\oplus\bthree{1}{0}{0}
=\Wedge^1[2]\oplus\Wedge^0[1].\]
More precisely, we may view a split signature conformal spin structure on a
five-dimensional split signature conformal manifold as a rank $4$ bundle,
denoted by $\bthree{0}{0}{1}$ and equipped with
\begin{itemize}\itemsep=0pt
\item a nondegenerate section of $\Wedge^2\big(\bthree{0}{0}{1}\big)$ with values
in~$\Wedge^0[-1]$,
\item an identification
$\Wedge_\perp^2\big(\bthree{0}{0}{1}\big)=\Wedge^1[2]$.
\end{itemize}
A scale $\sigma\in\Gamma\big({\Wedge}^0[1]\big)$ induces a nondegenerate section of
$\Wedge^2\big(\bthree{0}{0}{1}\big)$ (and, according to the discussion in Section~\ref{conf}, a~compatible nondegenerate (split signature) symmetric form on
$\Wedge^1$).
The first summand in the splitting
\begin{equation}\label{adjoint_bundle}
\operatorname{End}\big(\bthree{0}{0}{1}\big)=\bthree{-1}{0}{2}\oplus\bthree{-1}{1}{0}
\oplus\bthree{0}{0}{0}\end{equation}
captures the trace-free endomorphisms of $\bthree{0}{0}{1}$ that preserve its
conformal skew form \big(and, simultaneously, the second summand in
\[\operatorname{End}\big({\Wedge}^1\big)=\operatorname{End}\big(\, \bthree{-2}{1}{0}\big)
=\bthree{-2}{2}{0}\oplus\bthree{-1}{0}{2}\oplus\bthree{0}{0}{0}\]
plays the same r\^ole with respect to the conformal metric on $\Wedge^1$\big).
{From} this viewpoint, a scale $\sigma\in\Gamma\big({\Wedge}^0[1]\big)$ gives rise to a
connection on the spin bundle $\bthree{0}{0}{1}$ as follows. We firstly insist
that this connection preserve its conformal skew form and also that the induced
connection on~$\Wedge^0[1]$ annihilate~$\sigma$. According to~(\ref{adjoint_bundle}), the freedom in choosing such a connection lies in
\begin{equation}\label{freedom}
\Wedge^1\otimes\bthree{-1}{0}{2}
=\bthree{-2}{1}{0}\otimes\bthree{-1}{0}{2}.\end{equation}
On the other hand, the induced operator $\nabla\colon \Wedge^1\to\Wedge^2$ differs
from the exterior derivative by a homomorphism $\Wedge^1\to\Wedge^2$ (it is the
{\em torsion} of the induced affine connection) and, according
to~(\ref{deRham5dim}), lies in
\begin{equation}\label{torsion}
\operatorname{Hom}\big({\Wedge}^1,\Wedge^2\big)=\bthree{0}{1}{0}\otimes\bthree{-3}{0}{2}.
\end{equation}
Comparing the bundles (\ref{freedom}) and (\ref{torsion}) and noting that they
are canonically isomorphic suggests that the homomorphism $\Wedge^1\to\Wedge^2$
may be precisely eliminated by the allowed freedom. It is easy to check that
this is, indeed, the case.
\section{Three-dimensional conformal spin geometry}\label{3d_conformal}
A split signature three-dimensional conformal spin structure may be viewed as a
rank two `spin bundle' $S$ equipped with an identification
$\bigodot^2\!S=\Wedge^1[2]$ (cf.~\cite{S}). With the Dynkin diagram notation
from~\cite{BE},
\[\textstyle S=\btwo{1}{0},\qquad \bigodot^2\!\btwo{1}{0}=\btwo{2}{0},\qquad
\Wedge^1=\btwo{2}{-2},\qquad\Wedge^0[1]=\btwo{0}{1},\]
and the de~Rham complex is
\[\Wedge^0=\btwo{0}{0}\to\btwo{2}{-2}\to\btwo{2}{-3}\to\btwo{0}{-3}
=\Wedge^3.\]
The conformal structure can now be characterised by decreeing that the simple
spinors in $\btwo{2}{0}$ are the null vectors in $\Wedge^1[2]=\btwo{2}{0}$.
The first summand in
\[\operatorname{End}\big(\btwo{1}{0}\big)=\btwo{2}{-1}\oplus\btwo{0}{0}\]
captures the trace-free endomorphisms of $\btwo{1}{0}$ and it follows that the
freedom in choosing a connection on this spin bundle annihilating a scale,
i.e., a~nowhere vanishing section of $\Wedge^2\big(\btwo{1}{0}\big)=\btwo{0}{1}$, lies
in
\[\Wedge^1\otimes\btwo{2}{-1}=\btwo{2}{-2}\otimes\btwo{2}{-1}.\]
On the other hand the torsion of the induced affine connection lies in
the canonically isomorphic bundle
\[\operatorname{Hom}\big({\Wedge}^1,\Wedge^2\big)=\btwo{2}{0}\otimes\btwo{2}{-3}.\]
It is easy to check that the freedom in choice of connection on $\btwo{1}{0}$
may be used exactly to eliminate this torsion. More precisely, we have the
following:
\begin{prop}\label{one} Given a scale, i.e., a nowhere vanishing
$\sigma\in\Gamma\big(\btwo{0}{1}\big)$, there is a unique connection on $\btwo{1}{0}$
so that
\begin{itemize}\itemsep=0pt
\item the induced connection on $\Wedge^2\big(\btwo{1}{0}\big)=\btwo{0}{1}$
annihilates~$\sigma$,
\item the torsion of the induced connection on $\Wedge^1=\btwo{2}{-2}$
vanishes.
\end{itemize}
\end{prop}
It is straightforward to figure out how this preferred connection changes under
change of scale. To do this, let us adapt the classical two-spinor notation
of~\cite{OT} to write
\[\nabla\colon \ \btwo{1}{0}\longrightarrow\btwo{2}{-2}\otimes\btwo{1}{0}
\quad\mbox{as} \ \phi_C\longmapsto\nabla_{AB}\phi_C.\]
\begin{prop}\label{two} Let us change scale $\sigma\in\Gamma\big({\Wedge}^0[1]\big)$ by
$\widehat\sigma=\Omega^{-1}\sigma$. Then, for $\phi_C\in\Gamma\big(\btwo{1}{0}\big)$,
\begin{equation}\label{change_of_spin_connection}
\widehat\nabla_{AB}\phi_C=\nabla_{AB}\phi_C+\Upsilon_{AB}\phi_C
-\Upsilon_{C(A}\phi_{B)},\qquad\mbox{where} \
\Upsilon_{AB}\equiv\Omega^{-1}\nabla_{AB}\Omega.\end{equation}
\end{prop}
\begin{proof} Given $\nabla_{AB}$ we use (\ref{change_of_spin_connection}) to
define $\widehat\nabla_{AB}$ and then verify that it has the characterising
properties from Proposition~\ref{one} for the scale~$\widehat\sigma$. Firstly,
if $\sigma_{CD}\in\Gamma\big({\Wedge}^2\big(\btwo{1}{0}\big)\big)$, then
\begin{gather}\label{weight1}
\widehat\nabla_{AB}\sigma_{CD}=\nabla_{AB}\sigma_{CD}
+2\Upsilon_{AB}\sigma_{CD}-\Upsilon_{C(A}\sigma_{B)D}
+\Upsilon_{D(A}\sigma_{B)C}=\nabla_{AB}\sigma_{CD}
+\Upsilon_{AB}\sigma_{CD}.\!\!\!\!\!\end{gather}
Therefore $\widehat\nabla_{AB}\widehat\sigma_{CD}
=\nabla_{AB}\big(\Omega^{-1}\sigma_{CD}\big) +\Upsilon_{AB}\big(\Omega^{-1}\sigma_{CD}\big)
=\Omega^{-1}\nabla_{AB}\sigma_{CD}$. Thus, if
$\nabla_{AB}\sigma_{CD}=0$, then $\widehat\nabla_{AB}\widehat\sigma_{CD}=0$,
which accounts for the first condition in Proposition~\ref{one}. Now equation~(\ref{weight1}) may be abbreviated as
$\widehat\nabla_{AB}\sigma=\nabla_{AB}\sigma+\Upsilon_{AB}\sigma$ for
$\sigma\in\Gamma\big(\btwo{0}{1}\big)$ and it follows that
\[\widehat\nabla_{AB}\rho=\nabla_{AB}\rho+w\Upsilon_{AB}\rho,\qquad
\mbox{for}\ \rho\in\Gamma\big(\btwo{0}{w}\big).\]
Combining this with (\ref{change_of_spin_connection}), it follows
that $\widehat\nabla_{AB}\phi_C=\nabla_{AB}\phi_C-\Upsilon_{C(A}\phi_{B)}$,
for $\phi_C\in\Gamma\big(\btwo{1}{-1}\,\big)$ and hence that
\[\widehat\nabla_{AB}\omega_{CD}=\nabla_{AB}\omega_{CD}
-\Upsilon_{C(A}\omega_{B)D}-\Upsilon_{D(A}\omega_{B)C},\]
for $\omega_{CD}\in\Gamma\big(\btwo{2}{-2}\,\big)=\Gamma\big({\Wedge}^1\big)$. It follows that
\[\widehat\nabla_{AB}\omega_{CD}-\widehat\nabla_{CD}\omega_{AB}
=\nabla_{AB}\omega_{CD}-\nabla_{CD}\omega_{AB},\]
which accounts for the second condition in Proposition~\ref{one}.
\end{proof}
\section{The Rumin complex}\label{rumin}
Contact geometry is the geometry of a \textit{maximally non-integrable}
corank one subbundle $H\subset TM$, where $M$ is of dimension $2n+1$. Maximal
non-integrability is to say that locally $H$ is given as the kernel of a
\textit{contact form} $\alpha$ such that $\alpha \wedge ({\rm d}\alpha)^n$ is
non-vanishing.
Alternatively, writing $\Wedge_H^1$ for the bundle of one
forms $\Wedge^1$ restricted naturally to $H$, and $L$ for the annihilator of
$H$, we have a short exact sequence
\begin{equation*
0\to L\to\Wedge^1\to\Wedge_H^1\to 0
\end{equation*}
and, therefore, short exact sequences
\begin{equation}\label{SES2}
0\to \Wedge_H^{k-1}\otimes L\to\Wedge^k\to\Wedge_H^k\to 0\end{equation}
for $k=1,\ldots,2n$.
If we now consider the exterior derivative
\[\begin{array}{ccc}\Wedge_H^1&&\Wedge_H^2\\ \uparrow&&\uparrow\\
\Wedge^1&\xrightarrow{\ {\rm d}\ }\hspace{-14pt}
&\Wedge^2\\ \uparrow&&\uparrow\\
L&&\Wedge_H^1\otimes L,
\end{array}\]
then, by the Leibniz rule, the composition
$L\to\Wedge^1\xrightarrow{\,{\rm d}\,}\Wedge^2\to\Wedge_H^2$ is actually a vector
bundle homomorphism
known as the {\em Levi form}. The maximal non-integrability condition is
equivalent to the Levi form being injective with image consisting of
nondegenerate forms.
If one writes out the de Rham sequence along with the short exact
sequences (\ref{SES2}), one can obtain by diagram chasing, the Rumin
complex~\cite{R}. The Rumin complex is a replacement for the de Rham complex
on any contact manifold in that it computes the de Rham cohomology, but it is
in some sense more efficient in that derivatives are only taken in contact
directions. We are concerned with the case $\dim M = 5$ in which case the
diagram to chase is
\[\begin{array}{@{}ccccccccccc}
&&\Wedge_H^1&&\Wedge_H^2&&\Wedge_H^3&&\Wedge_H^4&&\\
&&\uparrow&&\uparrow&&\uparrow&&\uparrow&&\\
\Wedge^0 & \hspace{3mm}\xrightarrow{\ {\rm d}\ }
\hspace{3mm} & \Wedge^1&\hspace{3mm}\xrightarrow{\ {\rm d}\ }&\Wedge^2
&\xrightarrow{\ {\rm d}\ }&\Wedge^3
&\xrightarrow{\ {\rm d}\ }&\Wedge^4
& \xrightarrow{\ {\rm d}\ }& \hspace{3mm}\Wedge^5\\\
&&\uparrow&&\uparrow&&\uparrow&&\uparrow&&\\
&&L&&\Wedge_H^1\otimes L && \Wedge_H^2\otimes L && \Wedge_H^3\otimes L,&&
\end{array}\]
and, writing $\Wedge_H^2=\Wedge_{H\perp}^2\oplus L$, where $\Wedge_{H\perp}^2$
comprises $2$-forms on $H$ that are trace-free with respect to the Levi form,
one obtains the invariantly defined complex
\[\begin{array}{@{}ccccccccccc}
\Wedge^0 & \hspace{2mm} \xrightarrow{\ {\rm d}_\perp \ }
& \Wedge^1_H&\xrightarrow{\ {\rm d}_\perp \ }
&\Wedge^2_{H\perp}&\xrightarrow{\ {\rm d}_\perp^{(2)} \ }
&\Wedge_{H\perp}^2 \otimes L&\xrightarrow{\ {\rm d}_\perp \ }
&\Wedge^3_H \otimes L & \xrightarrow{\ {\rm d}_\perp \ }
& \hspace{2mm} \Wedge^5.
\end{array}\]
A difference to the de Rham complex is that one obtains
${\rm d}_\perp^{(2)}\colon \Wedge_{H\perp}^2 \to \Wedge_{H\perp}^2 \otimes L$,
which is a~second-order differential operator.
\section[Spinors in G\_2 contact geometry]{Spinors in $\boldsymbol{G_2}$ contact geometry}\label{G2}
A $G_2$ contact geometry is an additional structure on the contact distribution
of a five-dimensional contact manifold. As observed in the previous section, a~contact geometry is naturally equipped with its Levi form $L\to\Wedge_H^2$ and
the contact distribution $H$ thereby inherits a nondegenerate skew form
defined up to scale. This is just what is needed to talk about {\em Legendrean
varieties}~\cite{LM} in the projective bundle ${\mathbb{P}}(H)\to M$. A~{\em $G_2$ contact} structure on $M$ is a field of Legendrean twisted cubics
in ${\mathbb{P}}(H)$. Precisely, this means that, for all $m\in M$, there is a
twisted cubic $C_m\subset{\mathbb{P}}(H_m)$, varying smoothly with $m\in M$,
such that the $2$-planes in $H_m$ covering the tangent lines to the cubic are
null for the Levi form. Equivalently, such a $G_2$ contact structure may be
viewed as a~rank two `spin bundle' S equipped with a `Levi-compatible'
identification $\bigodot^3\!S=\Wedge_H^1[2]$, where $\Wedge^5=\Wedge^0[-3]$.
Levi-compatibility means that the Levi form
\[\textstyle
L[4]\to\Wedge_H^2[4]=\Wedge^2\big({\bigodot}^3\!S\big)
=\big({\bigodot}^4\!S\otimes\Wedge^2S\big)\oplus\big({\Wedge}^2S\big)^3\]
has its range in the second summand and thus provides an identification
$L[4]=\big({\Wedge}^2S\big)^3$. In these circumstances, notice that
\[\textstyle\Wedge_H^4[8]=\Wedge^4\big({\bigodot}^3\!S\big)=\big({\wedge}^2S\big)^6\quad
\Rightarrow\quad
\Wedge^0[9]=\Wedge^5[12]=\Wedge_H^4[8]\otimes L[4]=
\big({\Wedge}^2S\big)^9\]
and, therefore, we find canonical identifications $\Wedge^2S=\Wedge^0[1]$ and
$L=\Wedge^0[-1]$. In any case, the $G_2$ contact structure can now be
characterised by decreeing that the simple spinors in $H=\bigodot^3\!S[-1]$
constitute the cone over the twisted cubic (and there is a clear analogy with
conformal spin geometry in dimension three). More detail can be found in
\cite[Section~5]{EN2} and the flat model is presented in~\cite[Section~4]{EN1}. We should
also point out that the geometry of the rank four bundle~$H$ follows that of
Bryant's `$H_3$-structures' on the tangent bundle in four dimensions~\cite{B}.
The reason for the name `$G_2$ contact structure' is that this geometric data
defines a parabolic geometry of type $(G_2,P)$ where~$G_2$ is the
simply-connected exceptional Lie group of split type~$G_2$ and~$P$ is a~particular parabolic subgroup such that~$G_2/P$ is a contact manifold: see, for
example,~\cite[Section~4.2.8]{CS} (and, in particular, it is explained in~\cite{EN1}
that this particular realisation of the Lie algebra of $G_2$ goes back to
Engel~\cite{E}). With the Dynkin diagram notation from~\cite{BE}, this
motivates our writing $\bigodot^k\!S[w]=\gtwo{k}{w}$ (just another way of
organising the irreducible representations of
${\mathrm{GL}}_+(2,{\mathbb{R}})$) so that
\begin{gather*} \textstyle S=\gtwo{1}{0},\qquad \bigodot^3\!\gtwo{1}{0}=\gtwo{3}{0},\qquad
\Wedge_H^1=\gtwo{3}{-2},\qquad H=\gtwo{3}{-1},\qquad\Wedge^0[1]=\gtwo{0}{1},
\end{gather*}
and the Rumin complex is
\[\gtwo{0}{0}\xrightarrow{\,{\rm d}_\perp\,}\gtwo{3}{-2}
\xrightarrow{\,{\rm d}_\perp\,}\gtwo{4}{-3}
\xrightarrow{\,{\rm d}_\perp^{(2)}\,}\gtwo{4}{-4}
\xrightarrow{\,{\rm d}_\perp\,}\gtwo{3}{-4}\to
\gtwo{0}{-3}\]
(consistent with the basic {\em BGG complex} from~\cite{BE}). Notice that
\[\Wedge_H^2=\Wedge^2\big(\gtwo{3}{-2}\,\big)
=\gtwo{4}{-3}\oplus\gtwo{0}{-1}\]
so that the Levi form $L=\gtwo{0}{-1}\hookrightarrow\Wedge^2\big({\Wedge}_H^1\big)$ is
built into the notation.
Regarding calculus on a contact manifold, it is natural to consider
{\em partial connections}, rather than connections, on vector bundles in which
directional derivatives are defined, in the first instance, only in the contact
directions. More precisely, a {\em partial connection} on a vector bundle~$E$
is a~linear differential operator
\[\nabla_H\colon \ E\to\Wedge_H^1\otimes E\]
satisfying a {\em partial Leibniz rule} $\nabla_H(fs)=f\nabla_Hs+{\rm d}_\perp
f\otimes s$. (In fact, a partial connection can be uniquely promoted
\cite[Proposition~3.5]{EG} to a full connection but we shall not need this
trick.) In analogy with three-dimensional spin geometry, we may construct a~preferred partial connection on~$\gtwo{1}{0}$ in the presence of a~`scale'
$\sigma\in\Gamma\big(\gtwo{0}{1}\big)$.
The construction of this preferred partial connection follows the same route
save for a minor yet crucial distinction. For any contact manifold of dimension
$\geq 5$, a partial connection on $\Wedge_H^1$ gives rise to a linear
differential operator $\nabla_\perp\colon \Wedge_H^1\to\Wedge_{H\perp}^2$ defined as
the composition
\[\Wedge_H^1\xrightarrow{\,\nabla_H\,}
\Wedge_H^1\otimes\Wedge_H^1\xrightarrow{\,\wedge\,}
\Wedge_H^2\to\Wedge_{H\perp}^2\]
with the same symbol as the invariantly defined Rumin operator. It follows
that the difference
\[\nabla_\perp-{\rm d}_\perp\colon \ \Wedge_H^1\to\Wedge_{H\perp}^2\]
is actually a homomorphism of vector bundles. By definition, this is the
{\em partial torsion} of a~partial connection
$\nabla_H\colon \Wedge_H^1\to\Wedge_H^1\otimes\Wedge_H^1$.
In the case of a $G_2$ contact structure, bearing in mind that
$\Wedge^2\big(\gtwo{1}{0}\big)=\gtwo{0}{1}$, a partial connection on the spin bundle
$\gtwo{1}{0}$ induces partial connections on all spin bundles $\gtwo{k}{w}$
and, in particular, on $\Wedge_H^1=\gtwo{3}{-2}$. Thus, we may ask about
its partial torsion, which lies in
\begin{equation}\label{partial_torsion_decomposition}
\operatorname{Hom}\big({\Wedge}_H^1,\Wedge_{H\perp}^2\big)=\operatorname{Hom}\big(\gtwo{3}{-2},\gtwo{4}{-3}\,\big)
=\gtwo{7}{-4}\oplus\gtwo{5}{-3}\oplus\gtwo{3}{-2}\oplus\gtwo{1}{-1}.
\end{equation}
This decomposition is crucial in characterising preferred spin connections as
follows.
\begin{prop}\label{three}
Given a scale, i.e., a nowhere vanishing $\sigma\in\Gamma\big(\gtwo{0}{1}\big)$, there
is a unique partial connection on $\gtwo{1}{0}$ so that
\begin{itemize}\itemsep=0pt
\item the induced partial connection on $\Wedge^2\big(\gtwo{1}{0}\big)=\gtwo{0}{1}$
annihilates~$\sigma$,
\item the partial torsion of the induced partial connection on $\Wedge_H^1=\gtwo{3}{-2}$ lies in $\gtwo{7}{-4}$.
\end{itemize}
\end{prop}
\begin{proof}
The first summand in
\[\operatorname{End}\big(\gtwo{1}{0}\big)=\gtwo{2}{-1}\oplus\gtwo{0}{0}\]
captures the trace-free endomorphisms of $\gtwo{1}{0}$ and it follows that the
freedom in choosing a~partial connection on this spin bundle annihilating
$\sigma$ lies in
\begin{equation}\label{partial_freedom}
\Wedge_H^1\otimes\gtwo{2}{-1}=\gtwo{3}{-2}\otimes\gtwo{2}{-1}
=\gtwo{5}{-3}\oplus\gtwo{3}{-2}\oplus\gtwo{1}{-1}.\end{equation}
Comparison with (\ref{partial_torsion_decomposition}) certainly suggests that
all but the piece in $\gtwo{7}{-4}$ can be uniquely eliminated. We may verify
this using spinors. With the familiar conventions of~\cite{OT}, let us write
$\epsilon_{AB}\in\Gamma\big({\Wedge}^2\big(\gtwo{1}{0}\big)\big)$ rather than
$\sigma\in\Gamma\big(\gtwo{0}{1}\big)$. Then,
choosing a partial connection
\[\phi_D\xrightarrow{\,\nabla_H\,}\nabla_{ABC}\phi_D\qquad\mbox{on} \quad
\gtwo{1}{0},\]
the general partial connection on $\gtwo{1}{0}$ with
$\nabla_{ABC}\epsilon_{DE}=0$ has the form
\[\nabla_{ABC}\phi_D+\Gamma_{ABCD}{}^E\phi_E,\]
where $\Gamma_{ABCDE}=\Gamma_{(ABC)(DE)}$ (i.e., lying in
$\gtwo{3}{-2}\otimes\gtwo{2}{-1}$, as in~(\ref{partial_freedom})). By the
Leibniz rule, the induced
operator $\gtwo{3}{-2}\to\gtwo{4}{-3}$ is
\begin{equation}\label{induced_operator}
\omega_{DEF}\longmapsto\nabla_{(AB}{}^E\omega_{CD)E}
+2\Gamma_{(AB}{}^E{}_C{}^G\omega_{D)EG}
-\Gamma_{E(AB}{}^E{}^G\omega_{CD)G}.\end{equation}
Therefore, according to the decomposition~(\ref{partial_freedom}), we should
now write
\begin{equation}\label{split}\Gamma_{ABC}{}^{DE}=\lambda_{ABC}{}^{DE}
+\mu_{(AB}{}^{(D}\delta_{C)}{}^{E)}
+\nu_{(A}\delta_B{}^D\delta_{C)}{}^E,\end{equation}
where $\lambda_{ABCDE}$ and $\mu_{ABC}$ are symmetric spinors and compute
\[2\Gamma_{(AB}{}^E{}_C{}^G\omega_{D)EG}
-\Gamma_{E(AB}{}^E{}^G\omega_{CD)G}\]
for each term on the right hand side of~(\ref{split}). Clearly, this entails
computing
\begin{equation}\label{entrails}
\Gamma_{(AB}{}^{(E}{}_{C)}{}^{G)}\qquad\mbox{and}\qquad
\Gamma_{EAB}{}^{EG}.\end{equation}
Firstly, if $\Gamma_{ABC}{}^{DE}=\lambda_{ABC}{}^{DE}$, where
$\lambda_{ABCDE}=\lambda_{(ABCDE)}$, then the second term in~(\ref{entrails})
vanishes so for (\ref{induced_operator}) we end up with
\[\omega_{DEF}\longmapsto\nabla_{(AB}{}^E\omega_{CD)E}
+2\lambda_{(ABC}{}^{EG}\omega_{D)EG},\]
which is perfect for eliminating the $\gtwo{5}{-3}$-component of partial
torsion.
Secondly, if $\Gamma_{ABC}{}^{DE}=\mu_{(AB}{}^{(D}\delta_{C)}{}^{E)}$, where
$\mu_{ABC}=\mu_{(ABC)}$, then straightforward spinor computations show that
\[\textstyle\Gamma_{(AB}{}^{(E}{}_{C)}{}^{G)}
=\frac16\mu_{(AB}{}^{(E}\delta_{C)}{}^{G)}\qquad\mbox{and}\qquad
\Gamma_{EAB}{}^{EG}=\frac56\mu_{AB}{}^G\]
so for (\ref{induced_operator}) we end up with
\begin{align*} \omega_{DEF}& \mapsto
\nabla_{(AB}{}^E\omega_{CD)E}
+\tfrac13\mu_{(AB}{}^E\omega_{CD)E}
-\tfrac56\mu_{(AB}{}^G\omega_{CD)G}\\
& =\nabla_{(AB}{}^E\omega_{CD)E}
-\tfrac12\mu_{(AB}{}^E\omega_{CD)E},\end{align*}
which is perfect for eliminating the $\gtwo{3}{-2}$-component of partial
torsion.
Thirdly, if $\Gamma_{ABC}{}^{DE}=\nu_{(A}\delta_B{}^D\delta_{C)}{}^E$, then
straightforward spinor calculations yield
\[\textstyle\Gamma_{(AB}{}^{(E}{}_{C)}{}^{G)}
=-\frac13\nu_{(A}\delta_B{}^E\delta_{C)}{}^G\qquad\mbox{and}\qquad
\Gamma_{EAB}{}^{EG}=\frac43\nu_{(A}\delta_{B)}{}^G\]
so for~(\ref{induced_operator}) we end up with
\[\textstyle\omega_{DEF}\mapsto\nabla_{(AB}{}^E\omega_{CD)E}
-\frac23\nu_{(A}\omega_{BCD)}
-\frac43\nu_{(A}\omega_{BCD)}
=\nabla_{(AB}{}^E\omega_{CD)E}
-2\nu_{(A}\omega_{BCD)},\]
which is perfect for eliminating the $\gtwo{1}{-1}$-component of partial
torsion.
\end{proof}
Several remarks are in order. Firstly, the preferred connection of
Proposition~\ref{three} is constructed by eliminating all but the
$\gtwo{7}{-4}$-component of the partial torsion of the induced partial
connection on $\Wedge_H^1$, decomposed according to
(\ref{partial_torsion_decomposition}). In fact, it is clear from the proof that
the component lying in $\gtwo{7}{-4}$ is the same for any choice of partial
connection on $\gtwo{1}{0}$ and is, therefore, an invariant of the structure.
It is called the {\em torsion} of our $G_2$ contact structure. In the general
theory of parabolic geometry~\cite{CS}, this is the only component of
{\em harmonic curvature} and is therefore the only obstruction to
{\em local flatness}, i.e., to being locally isomorphic to the flat
model~$G_2/P$. Secondly, we should point out that the spinor identities
established by direct calculation in our proof can be avoided by judicious use
of Lie algebra cohomology (as in done in~\cite{CS}). Thirdly, we note that a
{\em scale}, a nowhere vanishing section~$\sigma$ of~$\gtwo{0}{1}$, has a
nice geometric interpretation. Since $\gtwo{0}{-1}=L\hookrightarrow\Wedge^1$
is the bundle of contact forms, we can interpret~$\sigma^{-1}$ as a~choice of
contact form. In other words, the preferred partial connection on the spin
bundle $S=\gtwo{1}{0}$ is obtained in the presence of a contact form.
The transformation law for preferred partial connections in $G_2$ contact
geometry is obtained by analogy with Proposition~\ref{two}. Its proof will
therefore be omitted.
\begin{prop} Let us change scale $\sigma\in\Gamma\big({\Wedge}^0[1]\big)$ by
$\widehat\sigma=\Omega^{-1}\sigma$. Then, for $\phi_D\in\Gamma\big(\gtwo{1}{0}\big)$,
\begin{equation}\label{change_of_G2_spin_connection}
\widehat\nabla_{ABC}\phi_D=\nabla_{ABC}\phi_D+\Upsilon_{ABC}\phi_D
-\Upsilon_{D(AB}\phi_{C)},\qquad\mbox{where} \quad
\Upsilon_{ABC}\equiv\Omega^{-1}\nabla_{ABC}\Omega.\end{equation}
\end{prop}
As an immediate consequence of (\ref{change_of_G2_spin_connection}), if
$\phi_D\in\Gamma\big(\gtwo{1}{0}\big)$, then
$\widehat\nabla_{(ABC}\phi_{D)}=\nabla_{(ABC}\phi_{D)}$. Furthermore, if
$\phi_D\in\Gamma\big(\gtwo{1}{-4/3}\ \big)$, then
\[\textstyle\widehat\nabla_{ABC}\phi_D
=\nabla_{ABC}\phi_D-\frac13\Upsilon_{ABC}\phi_D-\Upsilon_{D(AB}\phi_{C)},\]
whence $\widehat\nabla_{AB}{}^C\phi_C=\nabla_{AB}{}^C\phi_C$. We have found
two {\em invariant operators}
\begin{equation}\label{basic_invariant_operators}
\gtwo{1}{0}\longrightarrow\gtwo{4}{-3}\qquad\mbox{and}\qquad
\gtwo{1}{-4/3}\;\longrightarrow\gtwo{2}{-7/3}\ ,\end{equation}
defined only in terms of the $G_2$ contact structure itself. Conversely, it is
easy to see that the existence of these two operators is sufficient to define
the preferred partial connection on $\gtwo{1}{0}$ associated with a scale and
to capture the transformation law~(\ref{change_of_G2_spin_connection}).
Sadly, we have not been able to manufacture either of the invariant operators
(\ref{basic_invariant_operators}) directly.
In three-dimensional conformal spin geometry, the transformation law
(\ref{change_of_spin_connection}) leads to a pair of basic first-order
invariant differential operators
\[\btwo{1}{0}\longrightarrow\btwo{3}{-2}\qquad\mbox{and}\qquad
\btwo{1}{-3/2}\;\longrightarrow\btwo{1}{-5/2}\ ,\]
given by $\phi_C\mapsto\nabla_{(AB}\phi_{C)}$ and
$\phi_C\mapsto\nabla_{AB}\phi^B$. This suggests that we should refer to the
opera\-tors~(\ref{basic_invariant_operators}) as the `twistor' or `tractor'
operator and `Dirac' operator, respectively, on a $G_2$ contact manifold. Sure
enough, this tractor operator is overdetermined and, in the flat case, has a
$7$-dimensional kernel corresponding to the embedding
$G_2\hookrightarrow{\mathrm{SO}}^\uparrow(4,3)$. The prolongation of this
tractor operator (leading to the {\em standard tractor} bundle) is detailed
in~\cite{Moy}.
\section{Legendrean contact geometry in five dimensions}\label{CL}
A Legendrean contact geometry in five dimensions is a $5$-dimensional contact
manifold equipped with a splitting of the contact distribution as two
rank~$2$ subbundles
\[H=E\oplus F\]
each of which is null with respect to the Levi form
(so that the Levi form reduces to a perfect pairing $E\otimes F\to L^*$). The
flat model naturally arises in \cite[Section~2]{EN1} as the moduli space of flying
saucers in `attacking mode' and \cite[Proposition ~2.5]{EN1} gives
the $15$-dimensional symmetry algebra.
In the spirit of previous sections our aim will be, in the presence of a~scale (equivalently, a~choice of contact form), to construct preferred partial
connections on all the natural irreducible bundles on such a geometry. As
before, these connections can be obtained by means of the canonical
differential operators present on this type of geometry. In fact, we shall
only need to examine the Rumin complex to find sufficiently many canonical
differential operators for these purposes.
A Legendrean contact geometry is a type of parabolic
geometry~\cite[Section~4.2.3]{CS}, specifically \raisebox{1pt}{$\xox{}{}{}$} in the
notation of~\cite{BE}. Then
\[H=\begin{array}{c}E\\ \oplus\\ F\end{array}=
\begin{array}{c}
\xox{1}{1}{-1}\\
\oplus\\
\xox{-1}{1}{1}.
\end{array}\]
The general irreducible bundle has the form $\xox{u}{k}{v}$ for
$k\in{\mathbb{Z}}_{\geq0}$ and $u,v,\in{\mathbb{R}}$. It is convenient to take
$S\equiv\xox{-1}{1}{-1}$ as our basic `spin' bundle. If we also let
$\Wedge^0[u,v]\equiv\xox{u}{0}{v}$, then $L=\Wedge^0[-1,-1]$ and the general
irreducible bundle is $\xox{u}{k}{v}=\bigodot^k\!S[u+k,v+k]$. Notice that
\[E\otimes F
=\xox{1}{1}{-1}\;\otimes\;\xox{-1}{1}{1}=\xox{0}{2}{0}\oplus\xox{1}{0}{1}\]
and the Levi form is built into the notation as projection onto the second
summand. \big(Without the Dynkin diagram notation, it follows from the perfect
pairing $E\otimes F\to L^*$ and the canonical identification $E^*=E\otimes
\det E^*$ that
\[E\otimes(\det E^*)^{1/2}=\big(E\otimes(\det E^*)^{1/2}\big)^*
=F\otimes(\det F^*)^{1/2}\]
and we may take $S\equiv E\otimes(\det E^*)^{1/2}\otimes L^{1/2}
\equiv F\otimes(\det F^*)^{1/2}\otimes L^{1/2}$.\big)
In order to avoid confusion, by default we shall write a section of
$\xox{u}{k}{v}$ with {\em lower} spinor indices
\[\phi_{\mbox{\scriptsize$\underbrace{AB\dots C}_k$}}=\phi_{(AB\dots C)}
\in\Gamma\big(\xox{u}{k}{v}\big)\]
with no special terminology to record the bundle of which it is a section (in
other words, we shall forgo any systematic notion of `weight'). Of course, we
may use the tautological identification $S=S^*\otimes\det S=S^*\otimes L$ to
replace lower spinor indices by upper spinor indices (with an appropriate
change in `weight' if we were to assign one) so there is no loss in using lower
indices by default. As an example of these conventions in action, we may write
the first operator ${\rm d}_\perp\colon \Wedge^0\to\Wedge_H^1$ in the Rumin complex as
\begin{equation}\label{first_rumin}
f\longmapsto {\rm d}_\perp f\equiv\left[\!\begin{array}{c}\nabla_Af\\
\bar\nabla_Af\end{array}\!\right]
\in\begin{array}{c}\xox{-2}{1}{0}\\[-3pt] \oplus\\ \xox{0}{1}{-2}\end{array}
=\begin{array}{c}E^*\\[-3pt] \oplus\\ F^*\end{array}=\Wedge^1_H,
\end{equation}
where $\nabla_A$, respectively~$\bar\nabla_A$, is the directional derivative in
the~$E$, respectively~$F$, direction, both of which are manifestly invariantly
defined. (Although the notion of Legendrean contact geometry pertains in any
odd dimension, it is only in five dimensions that one has the convenience of
spinors and, in particular, that the bundles $E$ and $F$ agree save for a line
bundle factor.) Of course, the directional derivatives $\nabla_Af$ and
$\bar\nabla_Af$ end up as sections of different bundles even though each of
them has a single spinor index.
Soon (as with all parabolic geometries~\cite{CSweyl}), we shall find it
convenient to work in a particular {\em scale}, i.e., with a nowhere vanishing
section $\sigma\in\Gamma\big(\xox{1}{0}{1}\big)$. As with all parabolic contact
structures~\cite[Section~4.2]{CS}, we may interpret the section $\sigma^{-1}$ of
$\,\xox{-1}{0}{-1}\,=L\hookrightarrow\Wedge^1$ as a choice of contact form.
For $5$-dimensional Legendrean contact geometry, however, we may also interpret
a scale as a choice of skew spinor $\epsilon_{AB}$ by dint of
\begin{equation}\label{epsilon}\sigma^{-1}\in\Gamma\big(\,\xox{-1}{0}{-1}\,\big)
=\Gamma\big(\Wedge^2\big(\,\xox{-1}{1}{-1}\,\big)\big)=\Gamma\big(\Wedge^2S\big)
\ni\epsilon_{AB},\end{equation} which we may use to raise and lower spinor
indices with the familiar conventions of~\cite{OT}. In particular, for
$\phi_A\in\Gamma\big(\,\xox{-1}{1}{-1}\,\big)$ and $\psi_A\in\Gamma\big(\xox{0}{1}{0}\big)$, we
may write
\[\textstyle\sigma^{-1}\phi_{[A}\psi_{B]}=\frac12\epsilon_{AB}\psi^C\phi_C
\qquad\mbox{to define }\psi^C\phi_C\in\Gamma\big(\Wedge^0\big),\]
independent of choice of $\sigma$ and identifying
$\big(\,\xox{-1}{1}{-1}\,\big)^*=\xox{0}{1}{0}$, as expected.
Continuing from (\ref{first_rumin}), we may decompose the entire Rumin complex
into its constituent parts via the usual spin-bundle decompositions to obtain
an array of invariantly defined linear differential operators
\begin{gather}\label{array_of_operators}
\raisebox{-40pt}{\setlength{\unitlength}{.95pt}
\begin{picture}(410,120)(0,-25)
\put(0,50){\makebox(0,0){\xox{0}{0}{0}}}
\put(80,70){\makebox(0,0){\xox{-2}{1}{0}}}
\put(80,30){\makebox(0,0){\xox{0}{1}{-2}}}
\put(160,90){\makebox(0,0){\xox{-3}{0}{1}}}
\put(160,50){\makebox(0,0){\xox{-2}{2}{-2}}}
\put(160,10){\makebox(0,0){\xox{1}{0}{-3}}}
\put(240,90){\makebox(0,0){\xox{-4}{0}{0}}}
\put(240,50){\makebox(0,0){\xox{-3}{2}{-3}}}
\put(240,10){\makebox(0,0){\xox{0}{0}{-4}}}
\put(320,70){\makebox(0,0){\xox{-4}{1}{-2}}}
\put(320,30){\makebox(0,0){\xox{-2}{1}{-4}}}
\put(400,50){\makebox(0,0){\xox{-3}{0}{-3}}}
\put(30,49){\vector(2,1){25}}
\put(30,43){\vector(2,-1){25}}
\put(110,69){\vector(2,1){25}}
\put(110,63){\vector(2,-1){25}}
\put(110,29){\vector(2,1){25}}
\put(110,23){\vector(2,-1){25}}
\put(190,86){\vector(1,0){25}}
\put(190,46){\vector(1,0){25}}
\put(190,6){\vector(1,0){25}}
\put(190,83){\vector(3,-4){25}}
\put(190,83){\vector(3,-4){25}}
\put(190,49){\vector(3,4){25}}
\put(190,43){\vector(3,-4){25}}
\put(190,9){\vector(3,4){25}}
\put(270,83){\vector(2,-1){25}}
\put(270,49){\vector(2,1){25}}
\put(270,43){\vector(2,-1){25}}
\put(270,9){\vector(2,1){25}}
\put(350,63){\vector(2,-1){25}}
\put(350,29){\vector(2,1){25}}
\put(202.5,-13){\makebox(0,0){\framebox{second-order operators}}}
\put(202.5,-4){\vector(0,1){7}}
\put(46,-13){\makebox(0,0){\framebox{first-order operators}}}
\put(42.5,-4){\vector(0,1){32}}
\put(100,-13){\line(1,0){20}}
\put(120,-13){\vector(0,1){20}}
\put(359,-13){\makebox(0,0){\framebox{first-order operators}}}
\put(362.5,-4){\vector(0,1){32}}
\put(305,-13){\line(-1,0){20}}
\put(285,-13){\vector(0,1){20}}
\end{picture}}\!\!\!\!\!
\end{gather}
In this diagram we have omitted arrows that correspond to homomorphisms. For
example one can check that the part of the Rumin complex
$\xox{-2}{1}{0}\to\xox{1}{0}{-3}$ is actually a homomorphism, and its vanishing
is equivalent to the integrability of~$F$. As already noted, the operator
\[\nabla_A\colon \ \xox{0}{0}{0}\to\xox{-2}{1}{0}\]
is just the directional derivative on functions in the $E$ direction:
$f \mapsto {\rm d}f|_E$.
More generally, the partial connections we aim to construct naturally split
into a part that differentiates along~$E$ (which we shall denote by~$\nabla_A$)
and a part that differentiates along~$F$ (denoted by~$\bar\nabla_A$). In
particular, amongst the invariant operators in (\ref{array_of_operators}) we
find $\nabla_A\colon \xox{0}{0}{-4}\to\xox{-2}{1}{-4}$ and, therefore, by insisting
on the Leibniz rule, invariantly defined derivatives in the $E$ direction
\[\nabla_A\colon \ \xox{0}{0}{v}\to\xox{-2}{1}{0}\otimes\xox{0}{0}{v},\qquad
\mbox{for all}\ v\in{\mathbb{R}}.\]
Now suppose we are given a nowhere vanishing {\em scale}
$\sigma\in\Gamma\big(\xox{1}{0}{1}\big)$. Then we can define
\begin{align}\label{densities}
\nabla_A\colon \ \xox{u}{0}{v} \to \xox{-2}{1}{0} \otimes \xox{u}{0}{v},\qquad
\mbox{for all}\ u,v\in{\mathbb{R}}
\end{align}
by $\nabla_A(f\sigma^u)\equiv(\nabla_A f)\sigma^u$ for smooth sections $f$ of
$\xox{0}{0}{v-u}$\;. We may compute how this operator changes under a change of
scale $\widehat\sigma=\Omega^{-1}\sigma$, for some nowhere vanishing
smooth function~$\Omega$. Firstly, note that
$\nabla_A(\Omega^{-u}f)=\Omega^{-u}(\nabla_A f-u\Upsilon_Af)$, where
$\Upsilon_A\equiv\Omega^{-1}\nabla_A\Omega$.
Hence,
\begin{align*}
\widehat\nabla_A(f\widehat\sigma^u)& \equiv (\nabla_A f)\widehat\sigma^u=
\Omega^{-u}(\nabla_A f)\sigma^u\\
& = \nabla_A(\Omega^{-u}f)\sigma^u+u\Upsilon_Af\widehat\sigma^u
=\nabla_A(\Omega^{-u}f\sigma^u)+u\Upsilon_Af\widehat\sigma^u
\end{align*}
and, writing $s=f\widehat\sigma^u\in\Gamma\big(\xox{u}{0}{v}\big)$, we obtain
\[\widehat\nabla_A s=\nabla_A s+u\Upsilon_As.\]
In particular, this transformation law records the invariance of $\nabla_A$
when $u=0$. Similarly, starting with
$\bar\nabla_A\colon \xox{-4}{0}{0}\to\xox{-4}{1}{-2}$ from
(\ref{array_of_operators}), in the presence of a scale $\sigma$, we obtain
\[\bar\nabla_A\colon \ \xox{u}{0}{v}\to\xox{0}{1}{-2}\otimes\xox{u}{0}{v},
\qquad\mbox{for all}\ u,v\in{\mathbb{R}}\]
and, under change of scale $\widehat\sigma=\Omega^{-1}\sigma$, we find
that
\[\widehat{\bar\nabla}_As=\bar\nabla_A s+v\bar\Upsilon_As,\]
where $\bar\Upsilon_A\equiv\Omega^{-1}\bar\nabla_A\Omega$.
Referring back to the Rumin complex (\ref{array_of_operators}) we also
have canonical differential operators (in the $E$ direction)
\begin{equation}\label{more_canonical_operators}
\xox{0}{1}{-2} \to \xox{-2}{2}{-2}\qquad\mbox{and}\qquad
\xox{-2}{1}{0} \to \xox{-3}{0}{1},
\end{equation}
which may be combined, via the Leibniz, rule with (\ref{densities}) to obtain,
in the presence of a scale, a~first-order differential operator
\[
\nabla_A\colon \ \xox{u}{1}{v}\to\ \xox{u-2}{2}{v}\;\oplus\ \xox{u-1}{0}{v+1}
\ =\;\xox{-2}{1}{0}\otimes\xox{u}{1}{v}.\]
To express the transformation of this operator under change of scale,
we may split it as
\[\phi_B\mapsto\nabla_A\phi_B=\nabla_{(A}\phi_{B)}+\nabla_{[A}\phi_{B]},\]
and recall that operator (\ref{densities}) on densities
$s\in\Gamma\big(\xox{u}{0}{v}\big)$ transforms as
\begin{equation}\label{transform_on_densities}
\widehat\nabla_A s=\nabla_A+u\Upsilon_A s,\qquad\mbox{where}\quad
\Upsilon_A\equiv\Omega^{-1}\nabla_A\Omega.\end{equation}
\begin{prop} Suppose we change scale $\sigma\in\Gamma\big(\xox{1}{0}{1}\big)$ by
$\widehat\sigma=\Omega^{-1}\sigma$. Then, for $\phi_B$ a~section of
$\xox{u}{1}{v}$ we have
\begin{equation}\label{full_transformation_law}\widehat\nabla_A\phi_B
=\nabla_A\phi_B+(u+1)\Upsilon_A\phi_B-\Upsilon_B\phi_A.\end{equation}
\end{prop}
\begin{proof}
It suffices to note that this transformation law is consistent with the
invariance of the operators~(\ref{more_canonical_operators}), which may be
written as
\[\phi_B\longmapsto\nabla_{(A}\phi_{B)}\qquad\mbox{and}\qquad
\phi_B\longmapsto\nabla_{[A}\phi_{B]}\]
and also with (\ref{transform_on_densities}) on densities.
\end{proof}
Similarly, from the canonical operators
\[\bar\nabla\colon \ \xox{-2}{1}{0}\to\xox{-2}{2}{-2}\qquad\mbox{and}\qquad
\bar\nabla\colon \ \xox{0}{1}{-2}\to\xox{1}{0}{-3},\]
from (\ref{array_of_operators}) we may construct, in the presence of a scale,
\[\Gamma\big(\xox{u}{1}{v}\big)\ni\phi_B\longmapsto\bar\nabla_A\phi_B\]
differentiating in the $F$ direction and transforming by
\begin{equation}\label{full_barred_transformation_law}
\widehat{\bar\nabla}_A\phi_B=\bar\nabla_A\phi_B
+(v+1)\bar\Upsilon_A\phi_B
-\bar\Upsilon_B\phi_A,
\qquad\mbox{where}\quad \bar\Upsilon_A\equiv\Omega^{-1}\bar\nabla_A\Omega.
\end{equation}
Finally, we may combine $\nabla_A$ and $\bar\nabla_A$ to define, in the
presence of a scale $\sigma\in\Gamma\big(\xox{1}{0}{1}\big)$, partial connections
\[\Gamma(\xox{u}{1}{v})\ni\phi_B\to
\left[\!\begin{array}{c}\nabla_A\phi_B\\[3pt]
\bar\nabla_A\phi_B\end{array}\!\right]
\in\begin{array}{c}\xox{-2}{1}{0}\\[-2pt] \oplus\\[-1pt]
\xox{0}{1}{-2}\end{array}\!\!\otimes\xox{u}{1}{v}
=\Wedge_H^1\otimes\xox{u}{1}{v}\]
and, indeed by the Leibniz rule, on all weighted spinor bundles. These partial
connections are generated by $\Wedge_H^1\to\Wedge_H^1\otimes\Wedge_H^1$ and
this basic one is characterised as follows.
\begin{prop}\label{prop6}
Let $\sigma$ be a nowhere vanishing section of $\xox{1}{0}{1}$,
equivalently a choice of contact form. Then, there is a unique partial
connection $\nabla_H\colon \Wedge_H^1\to\Wedge_H^1\otimes\Wedge_H^1$ so that the
induced partial connection on $\Wedge_H^4=\xox{-2}{0}{-2}$ annihilates
$\sigma^{-2}$ and so that~$\nabla_H$ has minimal partial torsion in the sense
that the induced operator $\Wedge_H^1\to\Wedge_{H\perp}^2$ agrees with the
Rumin operator~${\rm d}_\perp$ modulo the homomorphisms that are the obstructions to
integrability.
\end{prop}
\begin{proof} Recall that we constructed this partial connection from the Rumin
complex (\ref{array_of_operators}) modulo the obstructions to integrability.
The only ingredients in this argument not immediately visible in
${\rm d}_\perp\colon \wedge_H^1\to\Wedge_{H\perp}^2$, where the two operators
\[\xox{0}{0}{-4}\to\xox{-2}{1}{-4}\qquad\mbox{and}\qquad
\xox{-4}{0}{0}\to\xox{-4}{1}{-2}\]
coming from further along the Rumin complex. However, it is straightforward to
check that, for example,
\[\xox{0}{0}{-4}=\Wedge^5\otimes\big(\,\xox{-3}{0}{1}\big)^*\longrightarrow
\Wedge^5\otimes\big(\,\xox{-2}{1}{0}\big)^*=\xox{-2}{1}{-4}\]
is the adjoint of $\;\xox{-2}{1}{0}\to\xox{-3}{1}{0}$ so these two hidden
ingredients are also secretly carried by
${\rm d}_\perp\colon \Wedge_H^1\to\Wedge_{H\perp}^2$.
\end{proof}
In fact, this proposition also follows from the general theory \cite{CS}, or by
a more explicit spinor calculation~\cite{Moy}. The prolongation of
$\xox{0}{1}{0}\ni\phi_B\longmapsto
\big(\nabla_{(A}\phi_{B)},\bar\nabla_{(A}\phi_{B)}\big)$ gives an especially convenient
tractor bundle and its (partial) connection~(cf.~\cite{CS,Moy}).
A more familiar way \cite{OT} of saying that $\nabla_A$ annihilates the scale
$\sigma\in\Gamma\big(\xox{1}{0}{1}\big)$ and hence
$\sigma^{-1}\in\Gamma\big(\,\xox{-1}{0}{-1}\,\big)$, is to say that
$\nabla_A\epsilon_{BC}=0$ for the corresponding skew form $\epsilon_{BC}$
under~(\ref{epsilon}). As a final consistency check, we may verify that this
constraint is invariant under~(\ref{full_transformation_law}), as follows. For
$\phi_B$ a section of $\,\xox{-1}{1}{-1}$, the transformation~(\ref{full_transformation_law}) reads
\[\widehat\nabla_A\phi_B=\nabla_A\phi_B-\Upsilon_B\phi_A\]
(cf.~conformal spin geometry in four
dimensions~\cite[equation~(5.6.15)]{OT}) and, therefore, if $\phi_{BC}$ is a
section of $\,\xox{-1}{1}{-1}\,\otimes\,\xox{-1}{1}{-1}\,$, then the Leibniz
rule implies that
\[\widehat\nabla_A\phi_{BC}
=\nabla_A\phi_{BC}-\Upsilon_B\phi_{AC}-\Upsilon_C\phi_{BA}.\]
Hence, if $\phi_{BC}$ is skew, then
$\Upsilon_A\phi_{BC}=\Upsilon_B\phi_{AC}+\Upsilon_C\phi_{BA}$ and it follows
that
\[\widehat\nabla_A\phi_{BC}=\nabla_A\phi_{BC}-\Upsilon_A\phi_{BC}.\]
Finally, when $\widehat\sigma=\Omega^{-1}\sigma$, we find that
$\widehat{\epsilon}_{AB}=\Omega\epsilon_{AB}$ (cf.~\cite[equation~(5.6.2)]{OT})
and, hence, that
\[\widehat\nabla_A\widehat\epsilon_{BC}
=\widehat\nabla_A(\Omega\epsilon_{BC})
=\Omega\big(\widehat\nabla_A\epsilon_{BC}+\Upsilon_A\epsilon_{BC}\big)
=\Omega\nabla_A\epsilon_{BC},\]
as required.
\section{Flying saucers via spinors}\label{saucers}
Distilling the construction in \cite{EN2} down to its key ingredients, we
will explain how to construct a~$G_2$ contact structure starting from a
Legendrean contact structure plus some additional data, a~choice of
appropriately weighted sections of the Legendrean subbundles. We then will
calculate the torsion of the resulting~$G_2$ contact structure in terms of the
input data and the preferred partial connection in the previous section.
Using the Levi form we may identify $F^* = E \otimes L = E[-1,-1]$ and so
can write
\[\Wedge_H^1 \ni \omega_a = \begin{bmatrix} \omega_A \\ \bar{\omega}^A
\end{bmatrix} \in \begin{array}{c}
\hspace{2mm}E^*\\[-3pt] \oplus\\ E[-1,-1]\end{array}. \]
In what follows we will fix a scale $\sigma \in \xox{1}{0}{1}$, or equivalently
a contact form. The expression for the torsion will turn out to be independent
of this choice, but to write it down we will need the distinguished connection
from the previous section. The distinguished partial connection annihilating
the scale can be written on $E^*$ as
\[E^* \ni \omega_A \mapsto \begin{bmatrix} \nabla_A\omega_B \\
\bar{\nabla}^A\omega_B \end{bmatrix} \in \begin{array}{c}
\hspace{2mm}E^*\\[-3pt] \oplus\\ E[-1,-1] \end{array}
\hspace{-2mm} \otimes E^*.
\]
Recall from Section~\ref{CL} that a choice of scale gives rise to a skew spinor
$\epsilon_{AB}$, covariant constant under the distinguished partial connection,
together with its inverse $\epsilon^{AB}$ (such that
$\epsilon_{AB}\epsilon^{AC}=\delta_B{}^C$), which we may use to raise and lower
indices as per \cite{OT}. Covariant constancy implies that we may equate
$\nabla_A \sigma^B = \nabla_A\big(\sigma_C\epsilon^{BC}\big)$ and
$(\nabla_A\sigma_C)\epsilon^{BC}$, as in the usual spinor calculus.
We decompose $\Wedge_{H\perp}^2
= \Wedge^2E^* \oplus (E^* \otimes E)_\circ[-1,-1] \oplus \Wedge^2E[-2,-2]$,
and Proposition~\ref{prop6} then means we can write the Rumin operator as
\[d_\perp
\begin{bmatrix}
\sigma_A \\
\tau^A
\end{bmatrix} = \begin{bmatrix}
-\nabla^A\sigma_A + \Pi_A\tau^A \\
\big(\nabla_A\tau^B - \bar{\nabla}^B\sigma_A\big)_\circ \\
\bar{\nabla}_A\tau^A + \Sigma^A\sigma_A
\end{bmatrix},
\]
where $\Sigma^A$ and $\Pi_A$ are the obstructions to integrability and
$\big(\Xi_A{}^B\big)_\circ = \Xi_A{}^B - \frac{1}{2}\delta_A{}^B\Xi_C{}^C$
for $\Xi_A{}^B \in E \otimes E^*$.
Now suppose we are given sections
$o_A\in E^*[2,-1]$ and $\iota^A\in\ E[-2,1]$ such that
$o_A\iota^A = 1$,
classically known as a {\em spin-frame}~\cite[pp.~110--115]{OT}. Then we
may define a $G_2$ contact geometry as follows. These sections determine an
isomorphism
\[\Wedge^0[0,3] \oplus \Wedge^0[1,2] \oplus \Wedge^0[2,1] \oplus \Wedge^0[3,0]
\cong
E^*[2,2] \oplus E[1,1]\]
given by
\begin{equation}\textstyle\label{defining_isomorphism}
(x,y,z,w)\longmapsto
\big(xo_A-\frac{1}{\sqrt{3}}y\iota_A,w\iota^A-\frac{1}{\sqrt{3}}zo^A\big).
\end{equation}
If we set
$S = \Wedge^0[0,1] \oplus \Wedge^0[1,0]$,
we have an isomorphism
\[\textstyle\bigodot^3\!S\cong E^*[2,2] \oplus E[1,1] = \Wedge_H^1[2,2]\]
as required in the definition of a $G_2$ contact structure. The peculiar
factors here are chosen so that the isomorphism also satisfies the additional
Levi-compatibility condition.
Starting by considering the general connection on $S$ that
annihilates the scale, we can calculate the torsion of the $G_2$ contact
structure by calculating the obstruction to the differential operator
$\Lambda_H^1 \to \Lambda_{H\perp}^2$ induced by the defining isomorphism being
equal to the Rumin operator.
We can write any partial connection on $S$ as
\begin{gather*
\nabla_{a} \begin{bmatrix}
y \\
z
\end{bmatrix}
= \begin{bmatrix}
\nabla_{a} y + \kappa_{a}y + \lambda_{a}z \\
\nabla_{a} z + \mu_{a} y + \nu_{a} z
\end{bmatrix}
\end{gather*}
for appropriately weighted sections $\kappa_{a}$, $\lambda_{a}$, $\mu_{a}$,
$\nu_{a}$, where we will take $\nabla_a$ to be the partial connection
distinguished by the contact form. The connection on $S$ will be compatible
with the contact form when $\kappa_{a} = -\nu_{a}$. Given this, the induced
partial connection on $\bigodot^3\!S \cong \Wedge^0[0,3] \oplus \Wedge^0[1,2]
\oplus \Wedge^0[2,1] \oplus \Wedge^0[3,0] $ is
\[\textstyle
\nabla_{a}
\begin{bmatrix}
x \\
y \\
z \\
w
\end{bmatrix} =
\begin{bmatrix}
\nabla_{a}x \\
\nabla_{a}y \\
\nabla_{a}z \\
\nabla_{a}w
\end{bmatrix} +
\begin{bmatrix}
3\kappa_{a} & \lambda_{a} & 0 & 0 \\
3\mu_{a}& \kappa_{a}& 2\lambda_{a}& 0 \\
0 & 2\mu_{a} & -\kappa_{a} & 3\lambda_{a}\\
0 & 0 & \mu_{a}& -3\kappa_{a}
\end{bmatrix}\begin{bmatrix}
x \\
y \\
z \\
w
\end{bmatrix} \in \Wedge_H^1 \otimes \bigodot^3\!S.
\]
Rewriting the right hand side using the defining isomorphism
(\ref{defining_isomorphism}) yields the differential operator
$\bigodot^3\!S \to \Wedge_H^1 \otimes (E^*[2,2] \oplus E[1,1])$ given by
\[\nabla_{a}
\begin{bmatrix}
x \\
y \\
z \\
w
\end{bmatrix} = \begin{bmatrix}
(\nabla_{a}x + 3x\kappa_{a} +
y\lambda_{a})o_A-\frac{1}{\sqrt{3}}(\nabla_{a}y +
3x\mu_{a} + y\kappa_{a} + 2z\lambda_{a})\iota_A \\[5pt]
-\frac{1}{\sqrt{3}}(\nabla_{a}z +
2y\mu_{a} - z\kappa_{a} + 3w\lambda_{a})o^A + (\nabla_{a}w
+ z\mu_{a} - 3w\kappa_{a})\iota^A
\end{bmatrix}.\]
To calculate the induced operator $\Wedge_H^1 \to \Wedge_{H\perp}^2$ we may use
the canonical identification \mbox{$\Wedge_H^1 = E^* \oplus E[-1,-1]$} and then
project the right hand side above onto the direct sum
\[\Wedge^2E^*[2,2] \oplus (E^* \otimes E)_\circ[1,1] \oplus \Wedge^2E.\]
We write $\kappa_a = \big(\kappa_A, \bar{\kappa}^A\big) \in E^* \oplus E[-1,1]$ and so
on to denote the projections. The induced operator (pulled back via the
isomorphism (\ref{defining_isomorphism})) is therefore
\[\textstyle\bigodot^3\!S[-2,-2] \to \Wedge^2E^*
\oplus (E^* \otimes E)_\circ[-1,-1] \oplus \Wedge^2E[-2,-2]\]
given by
\begin{gather*}
\begin{bmatrix}
x \\
y \\
z \\
w
\end{bmatrix} \mapsto \begin{bmatrix}
\big(\nabla_{A}x+ 3x\kappa_{A}+ y\lambda_{A}\big)o^{A}
- \frac{1}{\sqrt{3}}\big(\nabla_{A}y + 3x\mu_{A}+ y\kappa_{A}
+ 2z\lambda_{A}\big)\iota^{A} \vspace{2mm}\\
\Big( {-}\frac{1}{\sqrt{3}}\big(\nabla_{A}z + 2y\mu_{A} - z\kappa_{A}
+ 3w\lambda_{A}\big)o^B + \big(\nabla_{A}w + z\mu_{A}
- 3w\kappa_{A}\big)\iota^B \vspace{2mm}\\
- \big(\bar{\nabla}^Bx + 3x\bar{\kappa}^B + y\bar{\lambda}^B\big)o_A
+ \frac{1}{\sqrt{3}}\big(\bar{\nabla}^By + 3x\bar{\mu}^B + y\bar{\kappa}^B
+ 2z\bar{\lambda}^B\big)\iota_A \Big)\rule[-5pt]{0pt}{5pt}_{\circ}\vspace{2mm}\\
\frac{1}{\sqrt{3}}\big(\bar{\nabla}^{A}z + 2y\bar{\mu}^{A} - z\bar{ \kappa}^{A}
+ 3w\bar{\lambda}^{A}\big)o_{A} - \big(\bar{\nabla}^{A}w + z\bar{\mu}^{A}
- 3w\bar{\kappa}^{A}\big)\iota_{A}
\end{bmatrix}.
\end{gather*}
This is the operator that we should compare to the Rumin operator (pulled back
via the isomorphism (\ref{defining_isomorphism})), which can be written
\[{\rm d}_\perp \begin{bmatrix} x \\ y \\z \\w \end{bmatrix} =
\begin{bmatrix}
-\nabla^A(xo_A) + \frac{1}{\sqrt{3}}\nabla^A(y\iota_A)
+\Pi_A\big(w\iota^A-\frac{1}{\sqrt{3}}zo^A\big) \vspace{2mm}\\
\big(\nabla_A\big(w\iota^B-\frac{1}{\sqrt{3}}zo^B\big)
- \bar{\nabla}^B\big(xo_A-\frac{1}{\sqrt{3}}y\iota_A\big)\big)_\circ \vspace{2mm}\\
\bar{\nabla}_A\big(w\iota^A\big) - \frac{1}{\sqrt{3}}\bar{\nabla}_A\big(zo^A\big)
+ \Sigma^A\big(xo_A-\frac{1}{\sqrt{3}}y\iota_A\big)
\end{bmatrix}.\]
Insisting that these differential operators are equal (and hence that the
torsion vanishes) we obtain a system of twelve spinor equations
\begin{gather*}
\nabla^Ao_A = 3\kappa^Ao_A - \sqrt{3}\mu^A\iota_A,
\qquad \nabla^A\iota_A = -\sqrt{3}\lambda^Ao_A + \kappa^A\iota_A, \\
\Pi_Ao^A = -2\lambda^A\iota_A \hspace{67pt} \Pi_A\iota^A = 0, \\
\big(\bar{\nabla}^Bo_A\big)_\circ
= \big({-} \sqrt{3}\bar{\mu}^B\iota_A + 3\bar{\kappa}^Bo_A\big)_\circ,\\
\big(\bar{\nabla}^B\iota_A\big)_\circ
= \big({-}2\mu_Ao^B - \sqrt{3}\bar{\lambda}^Bo_A + \bar{\kappa}^B\iota_A\big)_\circ, \\
\big(\nabla_Ao^B\big)_\circ
= \big({-}\kappa_Ao^B - \sqrt{3}\mu_A\iota^B - 2\bar{\lambda}^B\iota_A\big)_\circ, \\
\big(\nabla_A\iota^B\big)_\circ
= \big({-} \sqrt{3}\lambda_Ao^B - 3\kappa_A\iota^B \big)_\circ, \\
\bar{\nabla}_Ao^A
= - \bar{\kappa}_Ao^A - \sqrt{3}\bar{\mu}_A\iota^A,\qquad
\bar{\nabla}_A\iota^A
= -\sqrt{3}\bar{\lambda}_Ao^A - 3\bar{\kappa}_A\iota^A, \\
\Sigma^A\iota_A = 2\bar{\mu}_Ao^A, \qquad \Sigma^Ao_A = 0 .
\end{gather*}
Contracting the middle four equations above with combinations of
$o_A$, $\iota^B$ and their counterparts with indices raised and lowered,
respectively, produces (together with the other eight) a system of twenty
independent linear equations over $\mathbb{R}$ in twelve unknowns $\kappa^A
o_A, \mu^A\iota_A, \dots $ (owing to the trace-free condition, the middle four
equations above yield three independent equations each). This system
is consistent if and only if the following eight obstructions vanish:
\begin{gather*}
\begin{split}
&\psi_0 = \Pi_A\iota^A,\\
&\psi_1 = \textstyle
\Pi_Ao^A - \frac{2}{\sqrt{3}}\iota^A\big(\nabla_A\iota^B\big)\iota_B,\\
&\psi_2 = \iota_A\big(\bar{\nabla}^Ao_B\big)o^B + o_A\big(\bar{\nabla}^Ao_B\big)\iota^B
- \bar{\nabla}_Ao^A + 2o_A\big(\bar{\nabla}^A\iota_B\big)o^B, \\
&\psi_3 = \textstyle\iota_A\big(\bar{\nabla}^A\iota_B\big)o^B
+ o_A\big(\bar{\nabla}^A\iota_B\big)\iota^B - \frac{1}{3}\bar{\nabla}_A\iota^A
+ \frac{2}{3}\iota_A\big(\bar{\nabla}^Ao_B\big)\iota^B
+ \frac{2}{\sqrt{3}}o^A\big(\nabla_Ao^B\big)o_B,\\
&\psi_4 = \textstyle
o^A\big(\nabla_Ao^B\big)\iota_B + \iota^A\big(\nabla_Ao^B\big)o_B - \frac{1}{3}\nabla^Ao_A
+ \frac{2}{3}o^A\big(\nabla_A\iota^B\big)o_B
+ \frac{2}{\sqrt{3}}\iota_A\big(\bar{\nabla}^A\iota_B\big)\iota^B, \\
&\psi_5 =
o^A\big(\nabla_A\iota^B\big)\iota_B + \iota^A\big(\nabla_A\iota^B\big)o_B
- \nabla^A\iota_A + 2\iota^A\big(\nabla_Ao^B\big)\iota_B , \\
&\psi_6 = \textstyle \Sigma^A\iota_A + \frac{2}{\sqrt{3}}o_A\big(\bar{\nabla}^Ao_B\big)o^B,\\
&\psi_7 = \Sigma^Ao_A.
\end{split}
\end{gather*}
The above eight functions vanish with the invariant torsion of the $G_2$
contact structure. One can check using the formul{\ae}
(\ref{full_transformation_law}) and~(\ref{full_barred_transformation_law}) that
these expressions are invariant under change of scale, as they should be.
This construction and resulting formul{\ae} apply in some generality
(locally all~$G_2$ contact structures arise this way \cite{Moy}). In
particular, they generalise \cite[equation~(32)]{EN2} in case that the
spin-frame $o_A$, $\iota^A$ arises from flying saucer
data~\cite[equation~(24)]{EN2} (precisely, with the notation from~\cite{EN2},
this means that $\iota^A=\pi^!\psi$ and $o_A=\Theta^{-1}\pi^!\phi$).
\subsection*{Acknowledgements}
We would like to thank all staff at SIGMA in Kyiv for their extraordinary
courage, continuing their work despite the shocking Russian invasion and
unconscionable aggression.
We would also like to thank the referees for their careful reading of our
manuscript and for their valuable suggestions and corrections.
\pdfbookmark[1]{References}{ref}
|
\section{Introduction}
\label{intro}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/opening.png}
\caption{The value space can be decomposed in a state-wise manner as an intersection of unions of hypersurfaces. Each union corresponds to a state and each hypersurface contains the value functions of policies agreeing on that state.}
\label{fig:opening}
\end{figure}
The space of value functions for stationary policies is a central concept in Reinforcement Learning (RL), since many RL algorithms are essentially navigating this space to find an optimal policy that maximizes the value function, such as policy gradient~\cite{sutton2000policy}, policy iteration~\cite{howard1960dynamic} and evolutionary strategies~\cite{de2005tutorial}.
Characterizing the geometric properties for the space of the value function (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, the value space) would offer insights for RL research. A recent work~\cite{dadashi2019value} shows that the value space for Markov Decision Processes (MDPs) is a possibly non-convex polytope, which inspires new methods in representation learning in RL~\cite{bellemare2019geometric, dabney2021value}.
Compared to MDPs, Robust MDPs (RMDPs) are more general, since they do not assume that the transition dynamics are known exactly but instead may take any value from a given uncertainty set~\cite{xu2006robustness, iyengar2005robust,nilim2005robust,wiesemann2013robust}. This makes RMDPs more suitable for real-world problems where parameters may not be precisely given. Therefore, characterizing the geometric properties of the value space for RMDPs (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, robust value space) is of interest.
However, we find it hard to directly adapt the prior approach~\cite{dadashi2019value} from MDPs to RMDPs. Their method builds upon on a key theorem (the Line Theorem), but we find it difficult to prove a robust counterpart of this theorem (see more discussions in Section~\ref{sec:line-theorem}).
In this work, we introduce a new perspective for investigating the geometry of the space of value functions.
Specifically, we start with revisiting the non-robust case due to its simplicity.
By decomposing the value space in a state-wise manner (as illustrated in Figure~\ref{fig:opening}), we can give an explicit form about the value function polytope.
With this decomposition-based perspective, we show that the robust value space is determined by a set of \emph{conic hypersurfaces}, each of which contains the robust value functions for policies that agree on one state.
Furthermore, from a geometric perspective, we show that the robust value space can be fully determined by a subset of the uncertainty set, which composes of extreme points of the uncertainty set.
As a result, for polyhedral uncertainty set such as $\ell_1$-ball and $\ell_\infty$-ball~\cite{ho2018fast, ho2021partial, behzadian2021fast}, we can replace the infinite uncertainty set with a finite active uncertainty subset, without losing any useful information for policy optimization. Finally, we discuss some other aspects about the robust value space, including policy agreement on more than one state, the non-convexity of the robust value space, and why it is difficult to obtain a Line Theorem for RMDPs.
All proofs and the specifics of MDPs and RMDPs used for illustration can be found in Appendix.
\section{Preliminaries}
\label{sec:prelim}
We introduce backgrounds for MDPs in
Section~\ref{sec:bg-mdp} and for RMDPs in Section~\ref{sec:bg-rmdp}. Importantly, Section~\ref{sec:bg-vfspace} sets up some essential concepts and notations for studying the value space, which will be frequently used in the rest of paper.
\emph{Notations.}
We use $\mathbf{1}$ and $\mathbf{0}$ to denote vectors of all ones and all zeros respectively, and their sizes can be inferred from the context. For vectors and matrices, $<$, $\le$, $>$ and $\ge$ denote element-wise comparisons. Calligraphic letters such as $\mathcal{P}$ are mainly for sets. For an index set $\mathcal{Z}=\{1,\cdots,k\}$, $(x_i)_{i\in\mathcal{Z}}$ denotes a vector $(x_1, x_2, \cdots, x_k)$ if $x_i$ is a scalar, or a matrix $(x_1, x_2, \cdots, x_k)^\top$ if $x_i$ is a vector. $\Delta_\mathcal{U}$ is used to denote the space of probability distributions over a set $\mathcal{U}$. For a non-empty set $\mathcal{U}$, we denote its \emph{polar cone} as $\mathcal{U}^*$\cite{bertsekas2009convex}, given by
\begin{equation}
\mathcal{U}^* \coloneqq \{y \mid \langle y,x\rangle \le 0, \forall\,x\in\mathcal{U}\}.
\end{equation}
We use $\mathbf{conv}(\cdot)$ to denote the convex hull of a set, and $\mathbf{ext}(\cdot)$ to denote the set of extreme points of a non-empty convex set.
\subsection{Markov Decision Processes}
\label{sec:bg-mdp}
We consider an MDP $(\mathcal{S}, \mathcal{A}, P, r, \gamma, p_0)$ with a finite state set $\mathcal{S}$ and a finite action set $\mathcal{A}$.
The number of states $\lvert\mathcal{S}\rvert$ and the number of actions $\lvert\mathcal{A}\rvert$ are denoted with $S$ and $A$, respectively.
The initial state is generated according to the $p_0\in\Delta_\mathcal{S}$.
We use $P_{s,a}\in\Delta_\mathcal{S}$ to specify the probabilities of transiting to new states when taking action $a$ in state $s$, and employ $P\coloneqq(P_{s,a})_{s\in\mathcal{S},a\in\mathcal{A}}\in(\Delta_\mathcal{S})^{\mathcal{S}\times\mathcal{A}}$ as a condensed notation.
An immediate reward $r_{s,a}\in\mathbb{R}$ is given after taking action $a$ in state
$s$, and similarly $r\coloneqq(r_{s,a})_{s\in\mathcal{S},a\in\mathcal{A}}\in\mathbb{R}^{\mathcal{S}\times\mathcal{A}}$ is a condensed notation.
$\gamma\in[0,1)$ is the discount factor.
In addition, we also define $P_s\coloneqq(P_{s,a})_{a\in\mathcal{A}}\in(\Delta_\mathcal{S})^{\mathcal{A}}$ and $r_s\coloneqq(r_{s,a})_{a\in\mathcal{A}}\in\mathbb{R}^{\mathcal{A}}$.
A stationary stochastic policy $\pi\coloneqq(\pi_{s,a})_{s\in\mathcal{S},a\in\mathcal{A}}\in(\Delta_\mathcal{A})^{\mathcal{S}}$ specifies a decision making strategy, where $\pi_{s,a}\in[0,1]$ is the probability of taking some action $a$ in current state $s$. We denote $\pi_s\coloneqq(\pi_{s,a})_{a\in\mathcal{A}}\in\Delta_\mathcal{A}$ as the probability vector over actions. In particular, we use $d_{s,a}\in\Delta_\mathcal{A}$ to represent a deterministic $\pi_s$ that is all-zero except $\pi_{s,a}=1$.
Under a given policy $\pi$, we define the state-to-state transition probability as
\begin{equation}
\begin{aligned}
P^\pi &\coloneqq (P^{\pi_s})_{s\in\mathcal{S}}\in(\Delta_\mathcal{S})^\mathcal{S}, \quad \textrm{where} \\
P^{\pi_s} &\coloneqq P_s\pi_s= \sum_{a\in\mathcal{A}}\pi_{s,a}P_{s,a}\in\Delta_\mathcal{S}.
\end{aligned}
\end{equation}
The reward function under this policy is defined as
\begin{equation}
\begin{aligned}
r^\pi &\coloneqq (r^{\pi_s})_{s\in\mathcal{S}}\in\mathbb{R}^\mathcal{S}, \quad \textrm{where} \\
r^{\pi_s} &\coloneqq r_s^\top\pi_s = \sum_{a\in\mathcal{A}}\pi_{s,a}r_{s,a} \in\mathbb{R}.
\end{aligned}
\end{equation}
The value $V^{\pi,P}\in\mathbb{R}^\mathcal{S}$ is defined to be the expected cumulative reward from starting in a state and acting according to the policy $\pi$ under transition dynamic $P$:
\begin{equation}
V^{\pi,P}(s) \coloneqq \mathbb{E}_{P^\pi} \left[\sum_{t=0}^\infty \gamma^t r_{s_t, a_t}\mid s_0=s\right].
\end{equation}
\subsection{Robust Markov Decision Processes}
\label{sec:bg-rmdp}
Robust Markov Decision Processes (RMDPs) generalize MDPs in that the uncertainty in the transition dynamic $P$ is considered~\cite{iyengar2005robust,nilim2005robust,wiesemann2013robust}. In an RMDP, the transition dynamic $P$ is chosen adversarially from an uncertainty set $\mathcal{P}\subseteq(\Delta_\mathcal{S})^{\mathcal{S}\times \mathcal{A}}$. We assume throughout the paper that the set $\mathcal{P}$ is compact. The robust value function for a policy $\pi$ and the optimal robust value function are defined as
\begin{align}
V^{\pi,\mathcal{P}}(s) &\coloneqq \min_{P\in\mathcal{P}} V^{\pi, P}(s),\\ V^{\star,\mathcal{P}}(s) &\coloneqq \max_{\pi\in\Pi} V^{\pi,\mathcal{P}}(s).
\end{align}
Both the policy evaluation and policy improvement problems are intractable for generic $\mathcal{P}$~\cite{wiesemann2013robust}. However, they become tractable when certain independence assumptions about $\mathcal{P}$ are made. Two common assumptions are \emph{$(s,a)$-rectangularity}~\cite{iyengar2005robust, nilim2005robust} and \emph{$s$-rectangularity}~\cite{wiesemann2013robust}, which we will use in this paper. The $(s,a)$-rectangularity assumes that the adversarial nature selects the worst transition probabilities independently for each state and action. Under $(s,a)$-rectangularity, the uncertainty set $\mathcal{P}$ can be factorized into $\mathcal{P}_{s,a}\subseteq\Delta_\mathcal{S}$ for each state-action pair, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{equation}
\mathcal{P} = \{P\mid P_{s,a}\in \mathcal{P}_{s,a}, \forall\,s\in\mathcal{S}, \forall\,a\in\mathcal{A}\},
\end{equation}
or in short $\mathcal{P}=\underset{(s,a)\in\mathcal{S}\times\mathcal{A}}{\times}\mathcal{P}_{s,a}$ where $\times$ denotes Cartesian product. The \emph{$s$-rectangularity} is less restrictive and assumes the adversarial nature selects the worst transition probabilities independently for each state. Under $s$-rectangularity, the uncertainty set $\mathcal{P}$ can be factorized into $\mathcal{P}_s\subseteq(\Delta_\mathcal{S})^\mathcal{A}$ for each state, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{equation}
\mathcal{P} = \{P\mid P_s\in \mathcal{P}_s, \forall\,s\in\mathcal{S}\},
\end{equation}
or in short $\mathcal{P}=\underset{s\in\mathcal{S}}{\times}\mathcal{P}_s$.
Note that $(s,a)$-rectangularity is a special case of $s$-rectangularity. Below we present a restatement of the remark in~\cite{ho2021partial} that the optimal policy for the robust policy evaluation MDP is deterministic. This restatement will be used later. Under $s$-rectangularity, we have for any $\pi$,
\begin{equation}
\begin{aligned}
\exists\,P\in\mathcal{P}\quad \textrm{s.t.} \quad V^{\pi,P}(s) = V^{\pi,\mathcal{P}}(s),\, \forall s\in\mathcal{S}.
\end{aligned}
\label{eqn:s-rect-exist}
\end{equation}
\subsection{The Space of Value Functions}
\label{sec:bg-vfspace}
The space of value functions (or value space in short) is the set of value functions for all stationary policies. We use $f_P$ and $f_{\mathcal{P}}$ to respectively represent the mapping between a set of policies and their non-robust and robust value functions, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{align}
f_P(U) &\coloneqq \{V^{\pi,P}\mid \pi\in U\}, \\
f_{\mathcal{P}}(U) &\coloneqq \{V^{\pi,\mathcal{P}}\mid \pi\in U\}.
\end{align}
The set of all stationary stochastic policies is denoted as $\Pi=(\Delta_\mathcal{A})^\mathcal{S}$. Then, the non-robust value space for a transition dynamic $P$ and the robust value space for an uncertainty set $\mathcal{P}$ can be respectively expressed as
\begin{align}
\mathcal{V}^P &\coloneqq f_P(\Pi), \label{eqn:def-vf-space}
\\ \mathcal{V}^\mathcal{P} &\coloneqq f_{\mathcal{P}}(\Pi).
\label{eqn:def-robsut-vf-space}
\end{align}
We then introduce some notations that will be frequently used later. We use $Y^{\pi_s}$ to denote the set of policies that agree with $\pi$ on $s$, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{equation}
Y^{\pi_s} \coloneqq \{\pi'\mid \pi'_s = \pi_s\}.
\end{equation}
Note that policy agreement on state $s$ does not imply disagreement on other states. Thus, $\pi$ itself is also in $Y^{\pi_s}$. The row of the matrix $I - \gamma P^\pi$ corresponds to state $s$ is denoted as $L^{\pi_s, P_s}$, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{equation}
L^{\pi_s, P_s} \coloneqq \mathbf{e}_s - \gamma P^{\pi_s} = \mathbf{e}_s - \gamma P_s \pi_s
\end{equation}
where $\mathbf{e}_s\in\mathbb{R}^\mathcal{S}$ is an all-zero vector except the entry corresponding to $s$ being 1.
\section{The Value Function Polytope Revisited}
\label{sec:revisit}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/single-v.png}
\caption{Hyperplanes $H^{\pi_s, P_s}$ corresponding to different $s$ intersect at the value function $V^{\pi,P}$.}
\label{fig:single-v}
\end{figure}
In this section, we revisit the non-robust value space from a new perspective, where the value space is decomposed in a state-wise manner.
This perspective enables us to characterize the polytope shape of the value space in a more straightforward way, and leads to an explicit form of the value polytope.
Our first step is to connect a single value function $V^{\pi,P}$ to a set of hyperplanes, each of which can be expressed as:
\begin{equation}
\label{}
H^{\pi_s, P_s} \coloneqq \{\mathbf{x}\in\mathbb{R}^\mathcal{S}\mid \langle \mathbf{x}, L^{\pi_s,P_s} \rangle = r^{\pi_s} \}.
\end{equation}
As shown in Lemma~3 in \cite{dadashi2019value}, the value functions $f_P(Y^{\pi_s})$ lie in the hyperplane $H^{\pi_s, P_s}$.
Specifically, since $\pi\in Y^{\pi_s}$, we know every hyperplane $H^{\pi_s, P_s}$ passes through $V^{\pi,P}$ (see examples in Figure~\ref{fig:single-v}).
The following lemma states that this intersecting point is unique.
\begin{restatable}{lemma}{singlev}
Consider a policy $\pi$ and a transition dynamic $P$, we have
\begin{equation}
\{V^{\pi,P}\} = \bigcap_{s\in\mathcal{S}} H^{\pi_s, P_s}
\end{equation}
\label{lem:single-v}
\end{restatable}
Lemma~\ref{lem:single-v} bridges between a single value function and the intersection of $S$ different hyperplanes, each of which corresponds to a state $s$.
Then, by definition (Eqn.~\eqref{eqn:def-vf-space}), we can obtain the value space by taking the union over all $\pi\in\Pi$, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{equation}
\label{eqn:vf-space-union}
\mathcal{V}^{P} = \bigcup_{\pi\in\Pi} \bigcap_{s\in\mathcal{S}} H^{\pi_s,P_s},
\end{equation}
as illustrated in Figure~\ref{fig:vf-space}(a).
From Eqn.~\eqref{eqn:vf-space-union}, we observe that the value space $\mathcal{V}^P$ can also be expressed from an alternative perspective (as shown in Figure~\ref{fig:vf-space}(b)):
1) for each state $s\in\mathcal{S}$, taking the union of all hyperplanes corresponding to different $\pi_s\in\Delta_\mathcal{A}$;
2) taking the intersection of the unions obtained in previous step.
The following lemma formalizes this perspective.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figs/vf-space.png}
\caption{Visualization of the value functions for a 2-state 3-action MDP.
\textbf{(a)} For each policy $\pi$, we plot the value function $V^{\pi,P}$ and the corresponding hyperplanes $H^{\pi_s, P_s}$ intersecting at $V^{\pi,P}$.
\textbf{(b)} For each policy $\pi$, the hyperplanes $H^{\pi_s, P_s}$ intersecting at $V^{\pi,P}$ are plotted in different colors for different states.
\textbf{(c)} For each state $s$, the union of $\mathcal{H}^{\pi_s,P_s}_{+}$ and the union of $\mathcal{H}^{\pi_s,P_s}_{-}$ over all $\pi_s\in\Delta_\mathcal{A}$ are highlighted respectively.
\textbf{(d)} For each state $s$, the hyperplanes $H^{d_{s,a},P_s}$ for different actions $a$ are plotted. The union of $\mathcal{H}_{+}^{d_{s,a},P_s}$ and the union of $\mathcal{H}_{-}^{d_{s,a},P_s}$ over all actions $a\in\mathcal{A}$ are highlighted as dashed. The entire value space $\mathcal{V}^P$ is visualized as the purple region.
}
\label{fig:vf-space}
\end{figure*}
\begin{restatable}{lemma}{vfspace}
Consider a transition dynamic $P$, the value space $\mathcal{V}^P$ can be represented as
\begin{equation}
\mathcal{V}^{P} = \bigcup_{\pi\in\Pi} \bigcap_{s\in\mathcal{S}}\, H^{\pi_s,P_s} = \bigcap_{s\in\mathcal{S}}\, \bigcup_{\pi_s\in\Delta_\mathcal{A}} H^{\pi_s,P_s}.
\end{equation}
\label{lem:vf-space}
\end{restatable}
As suggested in Lemma~\ref{lem:vf-space}, the core of this perspective is to decompose the value space in a state-wise manner.
In this way, to study the whole value space, we only need to focus on the union of hyperplanes corresponding to one state.
Specifically, let us denote the two closed half-spaces determined by the hyperplane $H^{\pi_{s},P_s}$ as
\begin{equation}
\begin{aligned}
\mathcal{H}^{\pi_{s},P_s}_{+} &\coloneqq \{\mathbf{x}\in\mathbb{R}^\mathcal{S}\mid \langle \mathbf{x}, L^{\pi_s,P_s} \rangle \ge r^{\pi_s} \}, \\
\mathcal{H}^{\pi_{s},P_s}_{-} &\coloneqq \{\mathbf{x}\in\mathbb{R}^\mathcal{S}\mid \langle \mathbf{x}, L^{\pi_s,P_s} \rangle \le r^{\pi_s} \}.
\end{aligned}
\end{equation}
Then the value space can be expressed in terms of the half-spaces:
\begin{equation}
\label{eqn:vfspace}
\mathcal{V}_P = \bigcap_{s\in\mathcal{S}}\, \bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{H}^{\pi_s,P_s}_{+} \cap \mathcal{H}^{\pi_s,P_s}_{-}.
\end{equation}
Recall that in~\cite{dadashi2019value} a convex polyhedron is defined as a finite intersection of half-spaces, and a polytope is a bounded finite union of convex polyhedra. So our goal is to get rid of this infinite union $\bigcup_{\pi_s\in\Delta_\mathcal{A}}$.
To this end, we first replace the inner union in Eqn.~\eqref{eqn:vfspace} with an intersection of two unions, as illustrated in Figure~\ref{fig:vf-space}(c) and formally stated in the following lemma.
\begin{restatable}{lemma}{vfspacetwo}
Consider a policy $\pi$ and a transition dynamic $P$, we have for all states $s\in\mathcal{S}$,
\begin{multline}
\bigcup_{\pi_s\in\Delta_{\mathcal{A}}} \mathcal{H}^{\pi_s,P_s}_{+}\,\cap\, \mathcal{H}^{\pi_s,P_s}_{-}
= \\
\left[
\bigcup_{\pi_s\in\Delta_{\mathcal{A}}} \mathcal{H}^{\pi_s,P_s}_{+}\right] \cap \left[\bigcup_{\pi_s\in\Delta_{\mathcal{A}}} \mathcal{H}^{\pi_s,P_s}_{-}\right].
\end{multline}
\label{lem:vf-space-2}
\end{restatable}
Although these two unions are still taken over infinite set $\Delta_\mathcal{A}$, the following Lemma~\ref{lem:vf-polyhedron} shows that they actually coincide with the finite unions of half-spaces that correspond to $d_{s,a}$ (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, deterministic $\pi_s$). We can get an intuition by comparing Figure~\ref{fig:vf-space}(c) and Figure~\ref{fig:vf-space}(d).
\begin{restatable}{lemma}{vfpolyhedronlem}
Consider a policy $\pi$ and a transition dynamic $P$, we have for all states $s\in\mathcal{S}$,
\begin{equation}
\bigcup_{\pi_s\in\Delta_{\mathcal{A}}} \mathcal{H}^{\pi_s,P_s}_{\delta} =
\bigcup_{a\in\mathcal{A}} \mathcal{H}^{d_{s,a},P_s}_{\delta}, \quad \forall\,\delta\in\{+,-\}.
\end{equation}
\label{lem:vf-polyhedron}
\end{restatable}
Finally, putting everything together, we are able to represent the value space with finite union and intersection operations on half-spaces, as stated in Theorem~\ref{thm:vf-polyhedron} and illustrated in Figure~\ref{fig:vf-space}(d). Using the distributive law of sets, we can see that the value space $\mathcal{V}^P$ immediately satisfies the definition of polyhedron. Since $\mathcal{V}^P$ is bounded, we can conclude that $\mathcal{V}^P$ is a polytope.
\begin{restatable}{theorem}{vfpolyhedron}
Consider a transition dynamic $P$, the value space $\mathcal{V}^P$ can be represented as
\begin{equation}
\begin{aligned}
\mathcal{V}^{P} & = \bigcap_{s\in \mathcal{S}}\left[ \left[\bigcup_{a\in\mathcal{A}} \mathcal{H}^{d_{s,a},P_s}_{+}\right] \cap \left[\bigcup_{a\in\mathcal{A}} \mathcal{H}^{d_{s,a},P_s}_{-}\right] \right] \\
& = \bigcup_{\mathbf{a}\in\mathcal{A}^{\mathcal{S}}} \bigcup_{\mathbf{a}'\in\mathcal{A}^{\mathcal{S}}} \bigcap_{s\in\mathcal{S}}
\left[\mathcal{H}^{d_{s,\mathbf{a}_s},P_s}_{+} \cap
\mathcal{H}^{d_{s,\mathbf{a}'_s},P_s}_{-}
\right]
\end{aligned}
\end{equation}
where $\mathbf{a}=(\mathbf{a}_s)_{s\in\mathcal{S}}$, $\mathbf{a}'=(\mathbf{a}'_s)_{s\in\mathcal{S}}$, and $\mathbf{a}_s, \mathbf{a}'_s\in\mathcal{A}$.
\label{thm:vf-polyhedron}
\end{restatable}
Compared to the prior approach~\cite{dadashi2019value}, our work gives an explicit form of the value function polytope, showing how the value polytope is formed (\emph{cf}\onedot} \def\Cf{\emph{Cf}\onedot the proof of Proposition~1 in~\cite{dadashi2019value}).
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/robust-v-example.png}
\caption{Visualization of the robust value functions for a 2-state 2-action RMDP with an $s$-rectangular uncertainty set. We consider $\mathcal{P}=\mathcal{P}_{s_1}\times\mathcal{P}_{s_2}$ with $\mathcal{P}_{s_1}=\{P_{s_1}^{(1)},P_{s_1}^{(2)}\}$ and $\mathcal{P}_{s_2}=\{P_{s_2}^{(1)},P_{s_2}^{(2)}\}$. Denote $P^{(ij)}\in\mathcal{P}$ such that $P^{(ij)}_{s_1}=P_{s_1}^{(i)}$ and $P^{(ij)}_{s_2}=P_{s_2}^{(j)}$. We plot with different widths to differentiate overlapping lines. \textbf{(a)} For the same set of policies $Y^{\pi_{s_1}}$, the set of non-robust value functions $f_P(Y^{\pi_{s_1}})$ for different $P\in\mathcal{P}$ and the set of robust value functions $f_\mathcal{P}(Y^{\pi_{s_1}})$ are plotted. \textbf{(b)} For different $P\in\mathcal{P}$, $f_P(Y^{\pi_{s_1}})$ are highlighted in different colors. The hyperplanes corresponding to different $P_{s_1}\in\mathcal{P}_{s_1}$ are plotted.}
\label{fig:robust-v-example}
\end{figure}
\section{Value Space Geometry of RMDPs}
\label{sec:main-sec}
\subsection{Policy Agreement and the Conic Hypersurface}
\label{sec:conic-hypersurface}
Recall that in Section~\ref{sec:revisit}, our new perspective connects the value space to the hyperplanes where $f_P(Y^{\pi_s})$ lies. Thus in order to characterize the robust value space, we start with studying the geometric properties of robust value functions for all policies that agree on one state, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, $f_\mathcal{P}(Y^{\pi_s})$. Unlike the non-robust case, $f_\mathcal{P}(Y^{\pi_s})$ may not lie in a hyperplane, as shown in Figure~\ref{fig:robust-v-example}(a). Nevertheless, it looks like $f_\mathcal{P}(Y^{\pi_s})$ still lies in a hypersurface (also see the example for $\lvert\mathcal{S}\rvert=3$ in the supplementary). In what follows, we are going to characterize this hypersurface.
First, as shown in Figure~\ref{fig:robust-v-example}(b), for different $P\in\mathcal{P}$ that share the same $P_s$, their $f_P(Y^{\pi_s})$ lie in the same hyperplane $H^{\pi_s, P_s}$. Comparing Figure~\ref{fig:robust-v-example}(a) and (b), it seems that the robust value functions $f_{\mathcal{P}}(Y^{\pi_s})$ always lie in the lower half-space $\mathcal{H}^{\pi_s, P_s}_{-}$ for different $P\in\mathcal{P}$.
On the other hand, from Eqn.~\eqref{eqn:s-rect-exist}, we know that there exists $P_s\in\mathcal{P}_s$ such that $V^{\pi,\mathcal{P}}$ lies in the hyperplane $H^{\pi_s, P_s}$.
Putting it together, we have the following lemma about $f_{\mathcal{P}}(Y^{\pi_s})$.
\begin{restatable}{lemma}{robustvboundary}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$ and a policy $\pi$, we have for all states $s\in\mathcal{S}$,
\begin{equation}
f_\mathcal{P}(Y^{\pi_s}) \subseteq \left[ \bigcap_{P_s\in\mathcal{P}_s} \mathcal{H}^{\pi_s, P_s}_{-} \right] \cap \left[ \bigcup_{P_s\in\mathcal{P}_s} H^{\pi_s, P_s} \right].
\label{eqn:robust-halfspaces}
\end{equation}
\label{lem:robust-v-boundary}
\end{restatable}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/conic-hypersurface.png}
\caption{\textbf{(a)} For different $P_{s}\in\mathcal{P}_{s}$, the hyperplanes $H^{\pi_{s}, P_{s}}$ intersect at one point. \textbf{(b)} Illustration of the conic hypersurface in which $f_{\mathcal{P}}(Y^{\pi_s})$ lies.}
\label{fig:conic-hypersurface}
\end{figure}
Note that the right hand side (RHS) of above Eqn.~\eqref{eqn:robust-halfspaces} is essentially the boundary of the intersection of half-spaces $\bigcap_{P_s\in\mathcal{P}_s} \mathcal{H}^{\pi_s, P_s}_{-}$. To further characterize the geometry, we need to know how these half-spaces intersect (equivalently how the hyperplanes intersect). One interesting observation is that when $\mathcal{P}_s$ contains more then 2 elements, the hyperplanes still intersect at one point, as illustrated in Figure~\ref{fig:conic-hypersurface}(a). The following lemma states this property and also gives the intersecting point.
\begin{restatable}{lemma}{intersect}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$ and a policy $\pi$, we have for all states $s\in\mathcal{S}$,
\begin{equation}
\frac{r^{\pi_s}}{1-\gamma}\mathbf{1}\in \bigcap_{P_s\in\mathcal{P}_s} H^{\pi_s, P_s}.
\end{equation}
\label{lem:intersect}
\end{restatable}
Since the hyperplanes intersect at the same point, the intersection of the half-spaces will be a convex cone. We denote
\begin{equation}
\begin{aligned}
\mathcal{C}^{\pi_{s},\mathcal{P}_s}_{+} &= \{\mathbf{x}\in\mathbb{R}^\mathcal{S}\mid \langle \mathbf{x}, L^{\pi_s,P_s} \rangle \ge r^{\pi_s},\, \exists P_s\in\mathcal{P}_s \}, \\
\mathcal{C}^{\pi_{s},\mathcal{P}_s}_{-} &= \{\mathbf{x}\in\mathbb{R}^\mathcal{S}\mid \langle \mathbf{x}, L^{\pi_s,P_s} \rangle \le r^{\pi_s},\, \forall P_s\in\mathcal{P}_s \}.
\end{aligned}
\label{eqn:conic-half-spaces}
\end{equation}
The following corollary characterizes the hypersurface that $f_\mathcal{P}(Y^{\pi_s})$ lies in. Figure~\ref{fig:conic-hypersurface}(b) gives an illustration.
\begin{restatable}{corollary}{conicsurface}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$ and a policy $\pi$, we have for all states $s\in\mathcal{S}$,
\begin{equation}
f_\mathcal{P}(Y^{\pi_s}) \subseteq C^{\pi_s, \mathcal{P}_s}
\end{equation}
where $C^{\pi_s, \mathcal{P}_s}= \mathcal{C}^{\pi_s, \mathcal{P}_s}_{+} \cap \mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$ is a conic hypersurface.
\label{col:conic-surface}
\end{restatable}
\subsection{The Robust Value Space}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/robust-vf-space.png}
\caption{Visualizations of the robust value functions for a 2-state 2-action RMDP with an $s$-rectangular uncertainty set. \textbf{(a)} For a fixed $\pi$, the conic hypersurfaces $C^{\pi_s, \mathcal{P}_s}$ corresponding to different $s$ intersect at the robust value function $V^{\pi,\mathcal{P}}$. \textbf{(b)}
For each policy $\pi$, the robust value function $V^{\pi,\mathcal{P}}$ is plotted, and the corresponding conic hypersurfaces $C^{\pi_{s}, \mathcal{P}_{s}}$ intersecting at $V^{\pi,\mathcal{P}}$ are plotted in different colors for different states.
\textbf{(c)} For each state $s$, the union of $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{+}$ and the union of $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$ over all $\pi_s\in\Delta_\mathcal{A}$ are highlighted respectively. \textbf{(d)} For each state $s$, the union of $\mathcal{C}^{d_{s,a}, \mathcal{P}_s}_{+}$ and the union of $\mathcal{C}^{d_{s,a}, \mathcal{P}_s}_{-}$ over all $a\in\mathcal{A}$ are highlighted respectively.}
\label{fig:robust-vf-space}
\end{figure}
With the knowledge about the geometry of $f_\mathcal{P}(Y^{\pi_s})$, we are now ready to characterize the entire robust value space $\mathcal{V}^{\mathcal{P}}$. Similar to Section~\ref{sec:revisit}, we first connect the single robust value function to the intersection
of $S$ different conic hypersurfaces by the following lemma (see Figure~\ref{fig:robust-vf-space}(a) for an illustration).
\begin{restatable}{lemma}{robustsinglev}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$ and a policy $\pi$, we have
\begin{equation}
\{V^{\pi, \mathcal{P}}\} = \bigcap_{s\in\mathcal{S}} C^{\pi_s, \mathcal{P}_s}.
\end{equation}
\label{lem:robust-single-v}
\end{restatable}
Then from the introduced perspective, we show that the robust value space can also be viewed as an intersection of state-wise unions of conic hypersurfaces, as illustrated in Figure~\ref{fig:robust-vf-space}(b) and formally stated in Lemma~\ref{lem:robust-vf-space}.
\begin{restatable}{lemma}{robustvfspace}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$, the robust value function space $\mathcal{V}^{\mathcal{P}}$ can be represented as
\begin{equation}
\mathcal{V}^\mathcal{P} = \bigcup_{\pi\in\Pi} \bigcap_{s\in\mathcal{S}}\, C^{\pi_s,\mathcal{P}_s} = \bigcap_{s\in\mathcal{S}}\, \bigcup_{\pi_s\in\Delta_\mathcal{A}} C^{\pi_s,\mathcal{P}_s}.
\end{equation}
\label{lem:robust-vf-space}
\end{restatable}
Next, we show the equivalence between each inner union in RHS of the above equation and an intersection of two unions in Lemma~\ref{lem:robust-vf-space-2}. Figure~\ref{fig:robust-vf-space}(c) gives an illustration.
Similar to the non-robust case, Lemma~\ref{lem:robust-vf-space-2} will help us characterize the relationship between the robust value space $\mathcal{V}^\mathcal{P}$ and the conic hypersurfaces corresponding to $d_{s,a}$.
\begin{restatable}{lemma}{robustvfspacetwo}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$, we have for all states $s\in\mathcal{S}$,
\begin{equation}
\bigcup_{\pi_s\in\Delta_\mathcal{A}}\!\!C^{\pi_s,\mathcal{P}_s} =
\left[\bigcup_{\pi_s\in\Delta_\mathcal{A}}\!\! \mathcal{C}^{\pi_s,\mathcal{P}_s}_{+}\right] \cap \left[\bigcup_{\pi_s\in\Delta_\mathcal{A}}\!\! \mathcal{C}^{\pi_s,\mathcal{P}_s}_{-}\right]\!.
\end{equation}
\label{lem:robust-vf-space-2}
\end{restatable}
As shown in Figure~\ref{fig:robust-vf-space}(d), unlike the non-robust case, the infinite union $\bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{C}^{\pi_s,\mathcal{P}_s}_{-}$ does not necessarily coincides with the finite union $\bigcup_{a\in\mathcal{A}} \mathcal{C}^{d_{s,a},\mathcal{P}_s}_{-}$. The following Lemma~\ref{lem:robust-vf-polyhedron} characterizes their relationship.
\begin{restatable}{lemma}{robustvfpolyhedronlem}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$, we have for all states $s\in\mathcal{S}$,
\begin{align}
\bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{C}^{\pi_s,\mathcal{P}_s}_{+} & = \bigcup_{a\in\mathcal{A}} \mathcal{C}^{d_{s,a},\mathcal{P}_s}_{+}, \\ \bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{C}^{\pi_s,\mathcal{P}_s}_{-} & \supseteq \bigcup_{a\in\mathcal{A}} \mathcal{C}^{d_{s,a},\mathcal{P}_s}_{-}, \label{eqn:lower-bound}
\end{align}
where the equality in the second line holds when $\mathcal{P}$ is $(s,a)$-rectangular.
\label{lem:robust-vf-polyhedron}
\end{restatable}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/robust-vf-polyhedron.png}
\caption{The robust value space $\mathcal{V}^\mathcal{P}$ and the conic hypersurfaces $C^{d_{s,a},\mathcal{P}_s}$ under $s$-rectangularity \textbf{(a)} and $(s,a)$-rectangularity \textbf{(b)}.}
\label{fig:robust-vf-polyhedron}
\end{figure}
Putting it together, the robust value space can be characterized in Theorem~\ref{thm:robust-vf-polyhedron}.
Figure~\ref{fig:robust-vf-polyhedron} highlights the difference in the robust value space between $s$-rectangularity and $(s,a)$-rectangularity, by using the same set of probability values (see Appendix~\ref{app:specs}).
Our results also provide a geometric perspective on why the optimal policies under $s$-rectangularity might be stochastic, which is only exemplified in prior works~\cite{wiesemann2013robust}. The robust value functions of deterministic policies always lie in the region defined by RHS of Eqn.~\eqref{eqn:robust-vf-polyhedron} but the optimal value might lie outside.
\begin{restatable}{theorem}{robustvfpolyhedronthm}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$, the robust value function space $\mathcal{V}^{\mathcal{P}}$ satisfies
\begin{equation}
\begin{aligned}
\mathcal{V}^{\mathcal{P}} & = \bigcap_{s\in \mathcal{S}}\left[ \left[\bigcup_{a\in\mathcal{A}} \mathcal{C}^{d_{s,a},\mathcal{P}_s}_{+}\right] \cap \left[\bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{C}^{\pi_s,\mathcal{P}_s}_{-}\right] \right]
\\
& \supseteq \bigcap_{s\in \mathcal{S}}\left[ \left[\bigcup_{a\in\mathcal{A}} \mathcal{C}^{d_{s,a},\mathcal{P}_s}_{+}\right] \cap \left[\bigcup_{a\in\mathcal{A}} \mathcal{C}^{d_{s,a},\mathcal{P}_s}_{-}\right] \right]
\end{aligned}
\label{eqn:robust-vf-polyhedron}
\end{equation}
where the equality in the second line holds when $\mathcal{P}$ is $(s,a)$-rectangular.
\label{thm:robust-vf-polyhedron}
\end{restatable}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/extra-region.png}
\caption{A closer look at the ``extra'' region under $s$-rectangularity. Here $\mathcal{P}_{s_2}=\{P_{s_2}^{(1)},P_{s_2}^{(2)}\}$. We highlight the hyperplanes in \textbf{(a)}, and the upper and lower bounds of the region $\bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{C}^{\pi_s, \mathcal{P}_s}$ in \textbf{(b)}. Note that the black boundaries in \textbf{(b)} are composed by the hyperplanes in \textbf{(a)}.}
\label{fig:extra-region}
\end{figure}
Furthermore, we take a closer look at this ``extra'' region under $s$-rectangularity. Since the space can be decomposed state-wisely, we focus on a single state $s$. Recall the definition of $\mathcal{C}_{-}^{\pi_s, \mathcal{P}_s}$ in Eqn.~\eqref{eqn:conic-half-spaces}, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{equation}
\mathcal{C}_{-}^{\pi_s, \mathcal{P}_s} = \bigcap_{P_s\in\mathcal{P}_s} \mathcal{H}_{-}^{\pi_s,P_s}.
\end{equation}
From our results in Section~\ref{sec:revisit}, we know
\begin{equation}
\mathcal{H}_{-}^{\pi_s,P_s} \subseteq \bigcup_{a\in\mathcal{A}} \mathcal{H}_{-}^{d_{s,a},P_s}.
\end{equation}
Therefore, we can obtain
\begin{equation}
\mathcal{C}_{-}^{\pi_s, \mathcal{P}_s} \subseteq \bigcap_{P_s\in\mathcal{P}_s} \bigcup_{a\in\mathcal{A}} \mathcal{H}_{-}^{d_{s,a},P_s},
\end{equation}
and accordingly
\begin{equation}
\bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{C}_{-}^{\pi_s, \mathcal{P}_s} \subseteq \bigcap_{P_s\in\mathcal{P}_s} \bigcup_{a\in\mathcal{A}} \mathcal{H}_{-}^{d_{s,a},P_s}.
\end{equation}
The RHS of the above equation gives us an upper bound of the region $\bigcup_{\pi_s\in\Delta_\mathcal{A}} \mathcal{C}_{-}^{\pi_s, \mathcal{P}_s}$ while the RHS of Eqn.~\eqref{eqn:lower-bound} provides a lower bound. The ``extra'' region lies within the gap between them. Figure~\ref{fig:extra-region} gives an illustration using the same RMDP example as in Figure~\ref{fig:robust-vf-polyhedron}.
\subsection{Active Uncertainty Subsets}
In above sections, we have shown that the robust value space $\mathcal{V}^\mathcal{P}$ depends on $\mathcal{P}$ in the form of a set of conic hypersurfaces $C^{\pi_s,\mathcal{P}_s}$. In this section, by taking a closer look at how $\mathcal{P}_s$ and $C^{\pi_s,\mathcal{P}_s}$ are related, we will show that only a subset $\mathcal{P}^\dagger\subseteq\mathcal{P}$ is sufficient to determine the robust value space, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot,
\begin{equation}
\mathcal{V}^\mathcal{P} = \mathcal{V}^{\mathcal{P}^\dagger}.
\end{equation}
We term $\mathcal{P}^\dagger$ as \emph{active} uncertainty subset, analogous to active constraints, in the sense that all $P\in\mathcal{P}^\dagger$ are active in determining the shape of the robust value space $\mathcal{V}^\mathcal{P}$.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/polar-cone.png}
\caption{Visualization of the convex cone $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$ for a fixed $\pi_s$ and different $\mathcal{P}_s$. The translation $r^{\pi_s}\mathbf{1}$ is ignored since $\pi_s$ is fixed. \textbf{(a)} We set $\mathcal{P}_s=\{P_{s}^{(1)}, P_{s}^{(2)}\}$. \textbf{(b)} We set $\mathcal{P}_s=\{P_{s}^{(\mu)} \mid P_{s}^{(\mu)}=\mu P_{s}^{(1)} + (1-\mu) P_{s}^{(2)}, 0\le\mu\le 1\}$ and also plot the hyperplanes $H^{\pi_s, P_{s}^{(\mu)}}$ for different $\mu$.}
\label{fig:polar-cone}
\end{figure}
First, let us keep $\pi_s$ fixed, and note that the conic hypersurface $C^{\pi_s, \mathcal{P}_s}$ is uniquely determined by the convex cone $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$. We then focus on the relationship between $\mathcal{P}_s$ and $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$. Denote the set
\begin{equation}
\mathcal{L}^{\pi_s, \mathcal{P}_s} \coloneqq \{L^{\pi_s, P_s}\mid P_s\in\mathcal{P}_s \}.
\end{equation}
From the definition of $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$, we can see $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$ is exactly the polar cone of $\mathcal{L}^{\pi_s, \mathcal{P}_s}$ (plus a translation), denoted with
\begin{equation}
\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-} = (\mathcal{L}^{\pi_s, \mathcal{P}_s})^* + \{r^{\pi_s}\mathbf{1}\}.
\label{eqn:polar-cone}
\end{equation}
Here $+$ denotes the Minkowski addition.
Figure~\ref{fig:polar-cone}(a) gives an illustration. Note that for fixed $\pi_s$, $\mathcal{L}^{\pi_s, \mathcal{P}_s}$ is the image of $\mathcal{P}_s$ under a fixed affine transformation. We denote this affine transformation as $g$, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, $\mathcal{L}^{\pi_s, \mathcal{P}_s}=g(\mathcal{P}_s)$.
Then we are able to obtain the following lemma:
\begin{restatable}{lemma}{activesubsetlem}
Consider a $s$-rectangular uncertainty set $\mathcal{P}$ and a policy $\pi$, we have
\begin{equation}
(\mathcal{L}^{\pi_s, \mathcal{P}_s})^* = (g(\mathcal{P}_s))^* = (g(\mathbf{ext}(\mathbf{conv}(\mathcal{P}_s))))^*.
\end{equation}
\label{lem:activesubsetlem}
\end{restatable}
This lemma implies that, in order to determine the conic hypersurface $C^{\pi_s, \mathcal{P}_s}$, we only need to care about those $P_s\in\mathcal{P}_s$ that are extreme points of the convex hull. Figure~\ref{fig:polar-cone}(b) gives an illustration. We then generalize it to the whole robust value space and present the following theorem:
\begin{restatable}{theorem}{activesubsetthm}
Consider a $s$-rectangular uncertainty set $\mathcal{P}$, we have
\begin{equation}
\mathcal{V}^\mathcal{P} = \mathcal{V}^{\mathcal{P}^\dagger}
\end{equation}
where $\mathcal{P}^\dagger=\mathbf{ext}(\mathbf{conv}(\mathcal{P}))\subseteq\mathcal{P}$.
\end{restatable}
If the $\mathcal{P}$ (or more generally $\mathbf{conv}(\mathcal{P}$)) is polyhedral, such as $\ell_1$-ball and $\ell_\infty$-ball~\cite{ho2018fast, ho2021partial, behzadian2021fast}, then we can reduce $\mathcal{P}$ to a finite set without losing any useful information for policy optimization. In addition, $\mathbf{conv}(\mathcal{P})$ being polyhedral implies that $\mathcal{C}^{\pi_s, \mathcal{P}_s}_{-}$ is a polyhedral cone. Combining with Theorem~\ref{thm:robust-vf-polyhedron}, it means that the robust value space for an $(s,a)$-rectangular uncertainty set will be a polytope.
\section{Discussion}
\label{sec:discuss}
\subsection{Policy Agreement on More States}
\label{sec:sub-polytope}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/sub-polytope.png}
\caption{A spectrum of the spaces of value functions.}
\label{fig:sub-polytope}
\end{figure}
We already know that the value functions for policies that agree on a single state lie in a hyperplane for MDPs~\cite{dadashi2019value}, and a conic hypersurface for $s$-rectangular RMDPs (Section~\ref{sec:conic-hypersurface}).
One natural question is how the space of value functions looks like when we fix the policies at more states.
With our new decomposition-based perspective, the results are immediately available from Lemma~\ref{lem:vf-space} and Lemma~\ref{lem:robust-vf-space}.
In Figure~\ref{fig:sub-polytope}, we show the space of value functions for policies agree on states in $\mathcal{S}^\ast\subseteq\mathcal{S}$, under both non-robust and robust setting.
Moreover, as illustrated in Figure~\ref{fig:sub-polytope}, our perspective reveals a spectrum of the spaces of value functions.
When the policies agree on all states, then it reduces to a single value function.
When the policies are free to vary on all states, then it is the whole value space.
This perspective enables us to characterize every point on this spectrum in an explicit form.
In comparison, for non-robust case, prior works~\cite{dadashi2019value} only prove that the spaces are polytopes without giving a clear characterization.
\subsection{The Non-convexity of the Robust Value Space}
Like the non-robust case, the robust value space $\mathcal{V}^\mathcal{P}$ is also possibly non-convex (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, Figure~\ref{fig:robust-vf-polyhedron}).
Despite the non-convexity, $\mathcal{V}^\mathcal{P}$ exhibits some interesting properties analogous to monotone polygons. As shown in Figure~\ref{fig:non-convexity}(a), for any point in the robust value space $\mathcal{V}^\mathcal{P}$, if we draw an axis-parallel line passing this point, the intersection will be a line segment (or a point in degenerated case).
We formalize this observation in the following corollary.
\begin{restatable}{corollary}{robustvfsegment}
Consider an $s$-rectangular uncertainty set $\mathcal{P}$, if an axis-parallel line intersects with the robust value space $\mathcal{V}^\mathcal{P}$, then the intersection will be a line segment.
\end{restatable}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/non-convexity.png}
\caption{\textbf{(a)} The intersection between the robust value space and axis-parallel lines are line segments. \textbf{(b)} An example showing that the robust value space is not star-convex.}
\label{fig:non-convexity}
\end{figure}
From the examples in Figure~\ref{fig:robust-vf-polyhedron}, one may wonder if the robust value function space $\mathcal{V}^\mathcal{P}$ is a star-convex set. For many randomly generated RMDPs, $\mathcal{V}^\mathcal{P}$ does look like a star-convex set (see Figure~\ref{fig:star-convexity} in Appendix~\ref{app:figure}). However, we show a carefully crafted counter-example in Figure~\ref{fig:non-convexity}(b), which is clearly not star-shaped. Nevertheless, it seems to be a rare case.
One interesting question to explore in the future is, how non-convex the robust value space can be and how likely it exhibits such non-convexity.
If it is nearly convex for most time, then we might be able to design some efficient algorithms tailored for such case.
\subsection{The Line Theorem for RMDPs}
\label{sec:line-theorem}
As mentioned before, one major obstacle that prevents us from adapting the prior method~\cite{dadashi2019value} from MDPs to RMDPs is that deriving a robust counterpart of the Line Theorem is highly challenging. Here we elaborate on this issue, with the help of our findings about the robust value space.
Without loss of generality, suppose the set of policies only differ on $s_1$.
From the discussions in Section~\ref{sec:sub-polytope}, we know the resulting set of robust value functions is
\begin{equation}
\left[\bigcap_{i=2}^S C^{\pi_{s_i},\mathcal{P}_{s_i}} \right] \cap \left[\bigcup_{\pi_{s_1}\in\Delta_\mathcal{A}} C^{\pi_{s_1},\mathcal{P}_{s_1}} \right].
\end{equation}
The first term is an intersection of $S-1$ conic hypersurfaces and the second term is an infinite union of conic hypersurfaces.
Both are hard to further characterize.
For example, though we know the first term could be a curve, it is challenging to give a closed-form expression for it.
In comparison, for MDPs, the first term is just a line and its direction is known (see the proof of Lemma 4 (ii) in ~\cite{dadashi2019value}).
\section{Related Works}
\label{sec:related}
The geometry of the space of value functions has been studied only recently. \citet{dadashi2019value} first investigate it, and establish that for MDPs the value space is a possibly non-convex polytope. Their results provide a geometric perspective to help understand the dynamics of different RL algorithms~\cite{kumar2019generalized, chan2020inverse, harb2020policy, chan2021greedification}, and also inspire new methods in representation learning in RL~\cite{bellemare2019geometric, dabney2021value}. In RMDP literature, some works take advantage of the geometric properties of special uncertainty sets to design efficient algorithms~\cite{ho2018fast, behzadian2021fast, ho2021partial}, but no prior works studies the geometry of the robust value space.
Our work can be viewed as an extension of~\cite{dadashi2019value} to RMDPs. We introduce a new perspective to characterize the geometric properties of the value space for RMDPs. Our approach also leads to a finer characterization of the value function polytope in MDPs setting.
\section{Conclusion and Future Work}
\label{sec:conclusion}
In this work, we characterize the geometry of the space of robust value functions from a new perspective, where the value space is decomposed in a state-wise manner.
We show that the robust value space is determined by a set of conic hypersurfaces. Furthermore, we can reduce the uncertainty set to a subset of extreme points without sacrificing any useful information for policy optimization.
There remain some interesting open questions. As discussed in Section~\ref{sec:discuss}, it is worth studying how non-convex the robust value space can be (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, can it be approximated as a convex set?). A further question is whether the level of non-convexity increases or decreases with the number of states/actions. Another direction is to investigate the geometry for other uncertainty set, such as coupled uncertainty~\cite{mannor2012lightning}, $r$-rectangular sets~\cite{goyal2018robust} or more general ones.
In addition, as in the non-robust case, it is interesting to study the geometry of robust value functions when the state space is very large and some approximation is needed.
We will leave these questions to future works.
\section*{Acknowledgements}
This work was partially supported by the Israel Science Foundation under contract 2199/20.
We appreciate the valuable feedback from ICML anonymous reviewers. We also thank Bingyi Kang and Pengqian Yu for some helpful discussions about RMDPs.
|
\section{More Results}
\subsection{Surface Reconstruction}
We report numerical results of 3D shape reconstruction on a collection of medium-size point sets in Table~\ref{table:qr3}.
The input point clouds contain tens of thousands of points and include holes in the bottom and valleys, as shown in Figure~\ref{fig:shape1} (training samples).
FINN generally outperforms FFN on all testing shapes except ``Bimba'', whose bottom part is empty and reconstruction has large uncertainty. The qualitative results show both FINN and FFN can fit points and interpolate holes well. FINN provides slightly better local details mainly because of the recovering operator. The underlying shapes of input point sets are generally smooth, especially compared with RGB and its hybrid fields. Thus a smaller frequency parameter is enough for reconstruction and is free of noisy artifacts. To further validate the effectiveness of removing noises, we perform 3D reconstruction on noisy point clouds.
\begin{figure*}[t!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{shape1.pdf} }
\end{center}
\caption{3D shape reconstruction from point clouds.}
\label{fig:shape1}
\end{figure*}
\paragraph{Robustness to Noise.}
\begin{figure*}[t!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{shape2.pdf} }
\end{center}
\caption{3D shape reconstruction from noisy point clouds.}
\label{fig:shape2}
\end{figure*}
We randomly perturb the point samples by adding each point with a vector drawn from gaussian noise, $\mathcal{N}(0,\sigma^2),\sigma=0.02$.
And we use the original points and perturbed points as training samples.
FINN can deal with noisy point sets without requiring explicit regularizer terms.
As shown in Figure~\ref{fig:shape2}, FINN reconstructs smooth and clean shapes, while without filtering, FFN overfits noisy points, creating random patches along underlying surfaces.
\subsection{Network Generalization}
\begin{figure*}[t!]
\begin{center}
{\includegraphics[width=0.8\linewidth]{2dgeneralization.pdf} }
\end{center}
\caption{Generalization validation via image super-resolution.
A fitting image (c) is upsampled by 16 via bilinear interpolation (d) or networks (e). While the results of bilinear interpolation are similar, the results of networks are very different. With filtering, FINN consumes high-frequency Fourier features for fitting, similar to FFN but at the same time suppresses high-frequency variations at the neighbors of training samples, thus interpolating pixels with fewer artifacts. Training pixels in (e) are painted in black.}
\label{fig:interpolant}
\end{figure*}
\paragraph{Interpolation Results.}
We fit an image at 128$\times$128 resolution and upsample to 512$\times$512 pixels to validate the learned interpolants, FINN and FFN, equipped with or without filtering. The results in figure~\ref{fig:interpolant} (c) show that FFN and FINN can both fit the training pixels faithfully. Then we upsample the fit images by a factor of 4 along vertical and horizontal directions using bilinear interpolation (d) and the learned interpolants (e). The results based on bilinear interpolation (d) are similar, with very few variations within individual characters. The interpolation via networks (e) without filtering adds many artifacts by randomly deeper and shallower the colors. That indicates when FFN allows ReLU-MLP to fit high-frequency training samples, it also introduces high-frequency noises between them, leading to overfitting. FINN lowers the variations between nearby training pixels, thus generalizing better to unseen pixels. Further, we will compare the interpolations at various sampling rates.
\paragraph{Impact of Sampling Rate.}
\begin{figure*}[t!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{2dsampling.pdf} }
\end{center}
\caption{Impact of the sampling rate.
We keep downsampling an image from 512$\times$512 resolution to 32$\times$32, with sampling rates of $1/4,1/16,1/64,$ and $1/256$. Then we use acquired images to train two networks, one with filtering and the other not, and generate new images all at 512$\times$512 resolution. At the highest resolutions (a), the image quality of the two methods is similar. (b-c) shows FINN interpolates pixels with less noisy artifacts than FFN. When using too fewer training samples, both methods create isolated segments for a single character (d-e). Even though both methods fail in (d-e), we can see that FINN tends to connect isolated training pixels with smooth interpolation, while FFN introduces more randomness between training pixels. Training pixels in (b-e) are painted in black.}
\label{fig:samplerate}
\end{figure*}
To understand the behaviors of interpolation at dense and sparse training sampling, we acquire a set of images by downsampling a given image with decreasing sampling rates, ranging from $100\%$ to $1/256$, as shown in Figure~\ref{fig:samplerate} top row. The networks with or without filtering both interpolate well at unseen pixels (a-c) and fails (d-e) when the distances between neighboring training pixels are too long. Although failed at sparse reconstruction (e), the experiment shows that FINN generates smooth interpolation between training pixels, while FFN overfits individual pixels and experiences randomness at unseen pixels.
\begin{table}[tp!]
\footnotesize
\tabcolsep=0.08cm
\begin{center}
\begin{tabular}{l>{\columncolor{white}}cc>{\columncolor{white}}cc>{\columncolor{white}}cc}\toprule
\rowcolor{white}
&Bunny &Armadillo &Dragon &Fandisk &Bimba &Gargoyle \\ \midrule
FFN & 2.04 & 2.92 & 1.97& 1.77& \textbf{138.11} & 4.58\\
FINN & \textbf{1.90} & \textbf{2.84} & \textbf{1.92} & \textbf{1.69} & 139.06 & \textbf{4.24} \\\bottomrule
\end{tabular}
\caption{Numerical results of 3D shape reconstruction. Chamfer distance is multiplied by $10^5$.}
\label{table:qr3}
\end{center}
\end{table}
\iffalse
\yx{
\subsection{Dual Positional Encoding for NeRF}
for RGB, need high frequency, but density needs low frequency. so in reality, only a number between them could be satisfied. we use a higher frequency than FFN to restore fine RGB, and a lower frequency for recovery to restore small ...
}
\fi
\section{Experimental Details}
\subsection{Networks}
\begin{figure*}[hbp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{networks.pdf} }
\end{center}
\caption{Our networks for three tasks.}
\label{fig:nets}
\end{figure*}
The detailed networks for 2D images, 3D surface, and NeRF are shown in Figure~\ref{fig:nets}.
The source code can be found in the attached file.
Baselines include FFN~\cite{tancik2020fourfeat}, SIREN~\cite{sitzmann2019siren}, SPE~\cite{WangLYT21} and SAPE~\cite{hertz2021sape}. Their codes can be downloaded from the following links,
\begin{itemize}
\item FFN: https://github.com/tancik/fourier-feature-networks.
\item SIREN: https://github.com/vsitzmann/siren.
\item SAPE: https://github.com/amirhertz/SAPE.
\item SPE: https://github.com/microsoft/SplinePosEnc.
\end{itemize}
Simplified NeRF can be found at https://github.com/tancik/ fourier-feature-networks. We use a PyTorch implementation of NeRF from https://github.com/yenchenlin/nerf-pytorch/ and modify it for our network.
Compared to FFN, FINN usually needs more network parameters because of the scaling vector generator. For $d$-dimensional Fourier features and $m$-dimensional hidden layers of MLP, our method need $d\times m$ more parameters, i.e., $512\times 256, 256\times 256, 512\times 256$ for image, shape and NeRF respectively. By increasing the number of hidden layers for FFN, we can set their network capacities to be the same, e.g., we use a 6-layer MLP in FFN and a 4-layer MLP in FINN for the NeRF task.
\iffalse
\begin{itemize}
\item Image: 512$\times$256$\times$2+256$\times$256$\times$2+256$\times$3=1539$\times$256, needs 512$\times$256 more parameters than FFN.
\item Surface: 256$\times$256$\times$2+256$\times$256$\times$3+256$\times$1=1281$\times$256, needs 256$\times$256 more parameters than FFN.
\item NeRF: 512$\times$256$\times$2+256$\times$256$\times$2+256$\times$4=1540$\times$256, needs same number of parameters as FFN.
\end{itemize}
\fi
\subsection{Additional Training Details}
\iffalse
\paragraph{Image Regression.}
The ``natural'' and ``text'' datasets contains 16 images for each.
We use $25\%$ of the pixels from an image for training and the entire set of pixels for testing.
The input coordinates and RGB values are normalized to [0,1]. The evaluation metrics are PSNR and \protect\reflectbox{F}LIP\xspace which can be found at https://github.com/NVlabs/flip.
\fi
\paragraph{Surface Reconstruction.}
The batch size for training large shapes is 200k, which contains 100k on-surface points sampled from the input and 100k off-surface points randomly generated within the box.
For medium-size shapes, we set the batch size to 32$^3$ for training and 256$^3$ for extracting meshes. To compute Chamfer distance, we uniformly sample 64$^3$ surface points.
Loss terms in Equation~\ref{eq4} are weighted as follows,
\begin{equation}
\begin{aligned}
\mathcal{L}_{\text{sdf}} &=\sum_{x\in \Omega}a_1|\|\nabla f_{\theta}(x)\| -1| \\
&+ \sum_{x\in \Omega\setminus \Omega_0}a_2\cdot {\exp(-50|f_{\theta}(x)|)} \\
&+ \sum_{x\in \Omega_0}(a_3|f_{\theta}(x)|+a_4\|\nabla f_{\theta}(x) - n(x)\|)
\end{aligned}
\end{equation}
where $a_1, a_2, a_3$ and $a_4$ are set to $1000, 10, 5$ and $20$ respectively. $\sigma$ is set to $1$ for all examples.
\paragraph{NeRF.}
We use the ``Lego'' and ``Fern'' scene datasets.
Lego contains 100 training images rendered from a virtual scene with black background colors. Fern has 17 training images captured in the real world. The ground truth camera poses are computed via structure from motion (SfM) that optimizes the extrinsic parameters of input images and the corresponding 3D locations of selected pixels. It is common to open source COLMAP~\cite{schoenberger2016sfm} for SfM.
We sample 1024 rays with 128 points along each ray in a batch for training. The training takes around an hour for FFN and FINN for 50k epochs.
\iffalse
\section{Ablation Study}
\subsection{Effects of $F_x$}
In Table~\ref{table:result}, we provides numerical results for the variants of our filtering function, i.e., $F_x = 1, xM^T$, and $\gamma(x)M^T$. By increasing the geometric variation of the recovering function $F_x$, the performance continuously improved. Further increasing the complexity of $F_x$, e.g., using a larger $\sigma$ for positional encoding in $\gamma_{\sigma}(x)M^T$, the fitted function starts to include more high-frequency noises, leading to performance reduction. Note that $\gamma_{\sigma}(x)$ is independent of $\gamma(x)$.
In addition to the qualitative results shown in the main paper, we include more results in Figure~\ref{fig:result}. At the top of the figure, the image contains many small sizes of words that FINN ($\gamma(x)M^T$) can recover most of them while other variants of FINN usually fail. The bottom image contains large homogenous patches (e.g., the polished floor) and complex structures and textures (e.g., the sofa). In the latter example, the one ablating the filtering module from FINN generates noisy artifacts to the result, and other variants of FINN ($F_x = 1, xM^T$) reconstruct blurry patches. In contrast, FINN recovers better details without increasing the noise.
\begin{figure}[t]
\begin{center}
{\includegraphics[width=1.0\linewidth]{comp.pdf} }
\end{center}
\caption{Various filtering effects.}
\label{fig:result}
\end{figure}
\begin{table}[b]
\footnotesize
\tabcolsep=0.14cm
\begin{center}
\begin{tabular}{l|cc|cc}
\hline
& \multicolumn{2}{c|}{PSNR$\uparrow$} & \multicolumn{2}{c}{\protect\reflectbox{F}LIP\xspace$\downarrow$} \\
& Natural & Text & Natural & Text \\ \hline
$F_x = 1$ & {27.91 $\pm$ 4.21} & {31.19 $\pm$ 1.86} & {0.123$\pm$0.036} & {0.085$\pm$0.008} \\\hline
$F_x = {xM^T}$ & {28.01 $\pm$ 4.30} & {33.15 $\pm$ 2.18} & {0.118$\pm$0.041} & {0.051$\pm$0.016} \\\hline
$F_x = \gamma(x)M^T$ & \textbf{28.51 $\pm$ 4.35} & \textbf{33.09 $\pm$ 1.97} & \textbf{0.100$\pm$0.037} & \textbf{0.042$\pm$0.016} \\\hline
$F_x = \gamma_{\sigma_1}(x)M^T$ & {27.94 $\pm$ 4.12} & {32.46 $\pm$ 1.96} & {0.107$\pm$0.037} & {0.046$\pm$0.018} \\\hline
$F_x = \gamma_{\sigma_2}(x)M^T$ & {26.49 $\pm$ 3.70} & {31.13 $\pm$ 1.93} & {0.117$\pm$0.034} & {0.057$\pm$0.026} \\\hline
\end{tabular}
\caption{Quantitative results on image regression task for the variants of FINN. Parameters $\sigma_1$ and $\sigma_2$ in $\gamma_{\sigma_1}(x)$ and $\gamma_{\sigma_2}(x)$ are set to 20 and 30 respectively.}
\vspace{-20pt}
\label{table:result}
\end{center}
\end{table}
\begin{figure}[t]
\begin{center}
{\includegraphics[width=1.0\linewidth]{train_psnr.pdf} }
\end{center}
\caption{Training PSNR curves of FFN and FINN. We compute the average PSNR for all the images from a dataset at each iteration of training.}
\label{fig:psnr}
\end{figure}
So far, all the experiments in the main paper and the appendix use $25\%$ of the pixels from the input image for optimization and reconstructing the full-resolution image. We include more results from the training stage.
In Figure~\ref{fig:psnr} we compare the PSNR curves collected during training FFN and FINN. With filtering modules, FINN can better fit the functions from the training samples and also generalize better to unseen samples.
\begin{figure*}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{result_all.pdf} }
\end{center}
\vspace{-6pt}
\caption{
Qualitative results on image regression task. Better visualized by zooming in.}
\label{fig:results}
\end{figure*}
The qualitative results are shown in Figure~\ref{fig:results}.
\begin{figure*}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{perlin.pdf} }
\end{center}
\caption{Reconstruction of Perlin noise functions. A Perlin noise image (right) is synthesized with $6\times 6$ small patches with varying frequencies. FINN reconstructs the image with high quality at all frequencies. ModSIREN, SAPE, and FFN are struggling to recover high-frequency components. In low-frequency regions, they also generate notable visual artifacts (better visualized by zooming in).}
\label{fig:perlin}
\end{figure*}
\subsubsection{Testing with Wide Frequency Bandwidth Image.}
We demonstrate the capability of our method in reconstructing complex signals containing notably wide frequency bandwidth.
Similar to ModSIREN~\cite{Mehta21}, a testing image is composed of $6\times 6$ grid of Perlin~\cite{Perlin02} noise patches with increasing frequencies along with the horizontal and vertical directions.
As shown in Figure~\ref{fig:perlin}, FINN presents a much better reconstruction compared to ModSIREN, SAPE, and FFN. In particular, FINN handles small structures well that other methods usually fail. Even in low-frequency regions, FINN also outperforms compared methods.
\fi
\section{Conclusion}
We present a novel filtering scheme in MLPs for implicit functions.
The filter improves the reconstruction of complex signals containing a wide range of frequencies.
Our filter combines a smoothing and a recovering operator that first globally over-smooths the result and then adaptively scales local variations to restore fine details smoothed.
Such a filtering scheme helps to fit both high-frequencies and low-frequencies with high fidelity.
We demonstrate the effectiveness of our filter on several tasks and show significant improvement against state-of-the-art methods.
\section{Introduction}
\begin{figure}[t]
\begin{center}
{\includegraphics[width=1.0\linewidth]{network.pdf} }
\end{center}
\caption{
We present a new neural network for implicit functions.
Reconstructing a signal containing low-frequency and high-frequency regions usually suffers from either underfitting or overfitting issues, i.e., losing fine-scale contents or adding noises.
To alleviate such a dilemma, we propose a filtering module for MLPs. Our filter performs smoothing and recovering consecutively at all layers, such that the smoothing operator globally removes high frequencies and the recovering operator restores high frequencies to regions over-smoothed. Two counteractive filtering operators reduce both issues and lead to better training and generalization.
}
\label{fig:network}
\end{figure}
Filtering techniques are widely applied to many applications to enhance the content of interest or remove unexpected artifacts from the data, such as images and 3D shapes.
For the data with many scales of details or a wide range of frequencies, it is challenging to recover the data while maintaining free of unwanted content using neural networks. Therefore, filtering methods are usually necessary for post-processing, otherwise additional strategies, e.g., coarse to fine generation, hierarchical representations, or scale-aware operations, are needed to adapt different scales of details. This paper addresses built-in learnable filters in a fully connected neural network to achieve both reconstruction and filtering purposes for neural implicit representation.
An implicit function is a continuous function that maps spatial coordinates to the corresponding values under varying representations (e.g., RGB, signed distance, etc.).
It can be learned using simple network structures, i.e., Multi-Layer Perceptrons (MLPs) with ReLU activations~\cite{OccNet,park2019deepsdf,chen2018implicit_decoder}. Due to the spectral bias of neural networks~\cite{basri20a,rahaman19a}, conventional MLPs are limited to low-frequency signals, and large networks and much training time are required to fit complex signals.
Recently, MLPs can easily reconstruct high-frequency by encoding the coordinates to Fourier features (FFN)~\cite{tancik2020fourfeat} or replacing ReLU activation functions with sinusoidal functions (SIREN)~\cite{sitzmann2019siren}.
The sinusoidal function increases local variations to the signal by tuning up the frequency parameters and reduces the spectral bias.
\begin{figure*}[t!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{comparison_teaser.pdf} }
\end{center}
\caption{
(a) An image generated by Fourier Features Network (FFN) contains noises (Top) highlighted in the error map from \protect\reflectbox{F}LIP\xspace (Bottom). (b) SAPE smoothes out noises by learning adaptive Fourier features, but at the expense of losing small structures and textures, e.g., clouds and roofs. (c) Our method (FINN) incorporates a filtering module that can effectively remove artifacts while recovering distinct features, containing fewer errors in flat and sharp regions.
We utilize a training image (e) at 256$\times$256 resolution sampled from a reference image (d) and predicts images at 4$\times$ resolutions for all methods.
}
\label{fig:comp_teaser}
\end{figure*}
Despite the impressive progress, FFN and SIREN usually create over-smoothed regions or noise in the results.
It is unavoidable since there is no global effective frequency parameter for both low- and high-frequency regions and each spatial location or region need to be individually adapted.
Towards learning adaptive frequency, based on FFN, SAPE~\cite{hertz2021sape} progressively encodes the input with increasing frequencies for individual spatial coordinates to avoid consuming excessive frequencies at smooth regions. And based on SIREN, ModSIREN~\cite{Mehta21} subdivides the input coordinates into grids and modulates sinusoidal activation functions according to individual grids.
An example is shown in Figure~\ref{fig:comp_teaser}, where the explicit spatial adaptation method (i.e., SAPE) provides results with less noisy artifacts compared to the baseline method (i.e., FFN).
Nevertheless, SAPE tends to over-smooth the result (e.g., the sky and the regions under roofs).
We observe that directly adapting frequency parameters leads to a dilemma – lowering the frequency to remove noisy artifacts is usually at the expense of over smooth while increasing the frequency to reveal fine details may introduce noises.
We show that a general-purpose filter in MLPs can significantly alleviate this dilemma and improve both regression and generalization. Unlike modulating sinusoidal activations (SIREN-based) or adapting frequency parameters of Fourier features (FFN-based), our network (FINN) manipulates intermediate outputs of MLPs with filters, leaving frequency fixed.
\begin{figure}[ht!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{filter_effect_idea.pdf} }
\end{center}
\caption{
Reconstructing a 1D signal in (a) with coordinate-based MLPs could suffer from underfitting (b-1) when the capacity of MLPs is insufficient or overfitting (c-1) when using overly-high frequency for the coordinate embedding of Fourier Features. To overcome individual issues, we present a recovering filter that restores local structures to the underfitting results (b-2) and a smoothing filter that suppresses local high frequencies from the overfitting results (c-2).}
\label{fig:teaser}
\end{figure}
FINN is a simple network function composed of filtering functions and MLPs. Following FFN, we also incorporate the Fourier Features for the representation of high frequencies.
The filtering function is decoupled into two operators, a smoothing operator, and a recovering operator, as shown in Figure~\ref{fig:network}. The two operators play consecutively at all layers of MLPs.
The smoothing operator is a spherical normalization function that constrains the features (i.e., intermediate results) on the hypersphere whose local structures are smooth everywhere, thus encouraging the feature space and final result to be smooth. The recovering operator is a Hadamard product that multiplies the features with learned scaling vectors to manipulate feature spaces, thus enabling it to represent complex local structures.
Thus the inductive bias of the smoothing operator is global smoothing regularization that reduces overfitting, and the recovering operator controls local sharpness that helps to reduce underfitting. Like FiLM~\cite{perez2018film}, the recovering operator modulates features with scaling vectors generated conditioning on coordinates or their Fourier features. It scales the features up or down or negates them to change their shape locally to better fitting the target.
The two operators enable learning adaptive feature embedding at spatial locations containing different scales of variations.
The two filtering operators can play standalone to deal with either overfitting or underfitting issues or combined with the other when both occur.
A 1D toy example in Figure~\ref{fig:teaser} demonstrates the effects of the two filtering modes.
The smoothing operator removes unexpected local bumps on the curve in Figure~\ref{fig:teaser}, (c-1)$\rightarrow$(c-2), and the recovering operator restores missing components and fine details, (b-1)$\rightarrow$(b-2). Note that (b-1) and (b-2) have similar network capacities, and so as (c-1) and (c-2).
For a complex signal containing both high-frequency and low-frequency, applying both filtering modes helps to construct both spectrum bandwidths, in a way that the former first overly smooths the entire regions, and the latter brings fine details to the regions over-smoothed.
To demonstrate the effectiveness of the filtering module, we evaluate our network on several tasks. Experimental results show that our method performs better than STOA methods.
Moreover, our filtering design yields a faster convergence speed and better network stability under SGD optimization.
Our contributions can be summarized as follows.
\begin{itemize
\item We propose a new filtering scheme for MLPs that learns adaptive reconstruction on data containing low and high frequencies. Our filter can remove noisy artifacts while enhancing distinct details that significantly alleviate the dilemma of overfitting and underfitting issues.
\item Our filter design is simple and flexible to control the sharpness and smoothness of the reconstruction. We show the superiority of our method in several tasks against STOA methods.
\end{itemize}
\section{Method}
\label{sec:method}
In this section, we introduce a new fully-connected neural network with filtering modules, as illustrated in Figure~\ref{fig:network}. We first present the neural implicit functions and the positional encoding
Next, we describe the the filtering functions
\subsection{Neural Implicit Functions}
\label{sec:imnet}
An implicit function is a continuous function $f_{\theta}:\mathbb{R}^a \rightarrow \mathbb{R}^n$ which takes as input a coordinate of any query point from the Euclidean space $x \in \mathbb{R}^a$ and predicts a value in the target domain $\mathbb{R}^n$.
Learning $f_{\theta}$ with a neural network requires a set of coordinate samplings as input and corresponding values as output.
Examples of the $f_{\theta}$ include mapping pixels to intensity values for image regression or projecting 3D coordinates to scalar values for 3D surface reconstruction.
It is common to model implicit function $f_{\theta}$ with MLPs. However, ReLU-based MLPs can hardly fit high-frequency signals due to the spectral bias trait, leading to the severe underfitting problem. Hence, to avoid favoring the low-frequency spectrum, FFN projects the input to the frequency domain with a controllable frequency bandwidth. Specifically, the coordinate $x$ is represented by a \textit{d}-dimensional Fourier feature vector, $\gamma(x) \in \mathbb{R}^d$, as follows:
\begin{equation}
\gamma(x) = \frac{s}{\sqrt{d}}[\text{cos}(2\pi x B^T) \| \text{sin}(2\pi x B^T) ]
\label{equ1}
\end{equation}
where $\|$ is the concatenation of two vectors, and $B$ is a $\frac{d}{2} \times a$ matrix randomly sampled from Gaussian distribution with standard deviation $\sigma$. $s$ is a constant for global scaling.
Fourier feature embedding facilitates learning complex functions with wider frequency bandwidths, but at the expense of overfitting issue: introducing unexpected dramatic local variations (e.g., noises, local bumps) to the reconstructed function.
Lowering the global $\sigma$ reduces the noise but towards underfitting. Adapting local $\sigma$ for individual coordinates also leads to over-smoothing (i.e., SAPE). Thus we present a filtering scheme to adaptively influence the reconstruction to alleviate the dilemma of overfitting and underfitting issues.
\iffalse
\begin{figure}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{idea.pdf} }
\end{center}
\caption{
Illustration of smoothing and recovering operations. Input coordinates $x_1,x_2$ are projected to features $p_{x_1},p_{x_2}$ after positional encoding and a fully connected layer.
After smoothing (i.e. spherical normalization), local variations between them are smoothed out, as shown by the segment between $p_{x_1}^{'},p_{x_2}^{'}$ on the hypersphere. To multiple with initial features $F_{x_1},F_{x_2}$, the recovering operation brings $p_{x_1}^{'},p_{x_2}^{'}$ from the hypersphere back to a more complex manifold, from which the interval between the new features $p_{x_1}^{''},p_{x_2}^{''}$ have less variations than the original features $p_{x_1},p_{x_2}$ if $F_{x_1},F_{x_2}$ are chose properly. This process is iterative performed after each fully connected layers.
}
\label{fig:idea}
\end{figure}
\fi
\begin{figure*}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{filter_effect.pdf} }
\end{center}
\caption{
(Top) The reconstruction results reveal a different degree of sharpness depending on scaling vectors $F_x$ (Bottom), which is a constant (a) or generated from the coordinates $x$ (b) or Fourier features $\gamma_s(x)$ (c-e). $F_x$ directly generated from $x$ usually leads to underfitting. In contrast, Fourier features have more flexibility in controlling the fitting capability. In particular, larger $\sigma$ adds more variations to the scaling vectors, resulting in sharper reconstruction, and an overly-high $\sigma$ creates more randomness and noise that reduces generalization. (d) has the best PSNR that balances sharpness and smoothness well.
}
\label{fig:filteringeffect}
\end{figure*}
\subsection{Filtering Functions}
\label{sec:filtering}
Let $\theta_i, i=1,2,...,k$ denote layers of a $k-$layer MLPs, then the coordinate $x$ passing through the MLPs will produce a sequence of outputs, such as $y_i, i=1,2,...,k$. To reconstruct the target function, we apply supervision on the final output function $y_k$. And to control the regional sharpness in $y_k$, we modulate all the intermediate features of MLPs with a filtering function. It is also possible to filter the final output directly, but more sophisticated designs are needed to trade off the data fidelity and the filtering purpose.
The formulation of the new network function can be written in recurrence as follows:
\begin{equation}
\begin{aligned}
F_x &= \gamma_s(x)M^T\\
y_1 &= \theta_1(\gamma(x)) \\
y_i &= o(\theta_i(y_{i-1}))\odot F_x, i=2,...,k-1\\
f_{\theta} &= \theta_k(y_{k-1})
\end{aligned}
\label{equ2}
\end{equation}
where $o(\cdot)$ is the spherical normalization function, such that $o(v)=\frac{v}{\|v\|}$, and $\odot$ is the Hadamard product.
The filter learns to adaptively influence the output by applying normalization and a linear transformation to the MLP’s intermediate features based on spatial locations.
Note that $F_x$ are generated from input Fourier features $\gamma_s(x)$ that can be different to $\gamma(x)$, the input of MLPs. This enables more flexibility in modeling wide frequency bandwidths by setting different $\sigma$ for $\gamma(x)$ and $\gamma_s(x)$. For example, for the color field and signed distance field, $\gamma(x)$ and $\gamma_s(x)$ can be the same, but for mixing fields, like neural radiance field containing color and density field, better to be different.
\iffalse
Inductive Bias of FINN: understanding FINN and potential impacts for deep learning community.
The filtering function composed of spherical normalization and Hadamard product derives smoothing and recovering operator respectively that effectively remove the noisy artifacts while maintaining high fidelity of data reconstruction.
As illustrated in Figure~\ref{fig:idea} (left), dramatic local variations occurs between two nearby feature points $p_{x_1}$ and $p_{x_2}$ in an intermediate feature space. After spherical normalization, the path between the normalized features $p_{x_1}^{'}$ and $p_{x_2}^{'}$ becomes flat. This operation yields an over-smoothed reconstruction, as shown in Figure~\ref{fig:idea2}(b), that some small structures disappear. Nevertheless, the image is still better than the one without applying a smoothing operation, e.g., Figure~\ref{fig:idea2}(a). The smoothing operation indicates that restricting the feature space to hypersphere can still provide both global and local information for data fitting though the results lean toward over-smoothed.
To further restore small structures, the recovering operator is applied, and the variations between the new features $p_{x_1}^{''}$ and $p_{x_2}^{''}$ are increased, as shown in Figure~\ref{fig:idea} (right).
\fi
Figure~\ref{fig:teaser} shows the filtering effects of the two operators on a 1D example.
And a 2D example in Figure~\ref{fig:idea2} also demonstrates the effects of the two operators and their combination. The recovering operator in Figure~\ref{fig:idea2} (c) reconstructs more details than the original FFN (a) but is unable to avoid noise, and the smoothing operator (b) smoothes out noise, but some distinct details are missing. The joint operator (d) recovers fine-grained details without introducing more noise.
As mentioned earlier, the inductive bias of the smoothing and recovering operators provides a way of modulating the feature spaces to influence the output. Specifically, their combination first smooths everywhere and then modulates features in a selective manner, scaling down features to remain smooth in homogenous regions and turning them up or negating them to increase variations in sharp regions.
\paragraph{Filtering Effects}
Besides tuning the frequency parameter in $\gamma(x)$, one can also fix $\gamma(x)$ and modifies $\gamma_s(x)$ to control the sharpness of reconstructions, since $\gamma_s(x)$ controls the complexness of scaling vector $F_x$. For example, smoothing reconstruction requires simpler scaling vector space that can be generated conditioning on a less complex input, e.g., a constant, $x$, or a smaller $\sigma$ in $\gamma_s(x)$.
In Figure~\ref{fig:filteringeffect}, we draw $F_x$ in the bottom row, including $F_x=1, xM^T$, and $\gamma_s(x)M^T$ ($\sigma=1,10$ and $30$). The scaling vectors and their corresponding reconstruction results reveal a different degree of sharpness.
In practice, $\gamma(x)$ and $\gamma_s(x)$ can be different for adapting hybrid fields or the same for single fields.
\iffalse
\subsubsection{Hyperparameters.}
The reconstruction is sensitive to the random Fourier feature embedding, whose parameters are hard to optimize under SGD-based optimization. In Figure~\ref{fig:param} (Right), we show the solution space of a testing image by grid-searching the parameters $\sigma$ (ranging from 1 to 25) and $s$ (ranging from 5 to 160). Specifically, the image is best recovered around $\sigma=10$ and gets over-smoothed or contaminated by noise when using an overly small or large $\sigma$ respectively.
In contrast to $\sigma$, $s$ only slightly change the results when $\sigma$ is fixed. However, $s$ has more impact on the convergence speed, as shown in Figure~\ref{fig:scale}, where for a fixed $\sigma$, faster convergence speed achieves if $s$ grows (from dark to light colors) and becomes stable when $s>50$. Moreover, higher $s$ also improves network stability, as shown in the highlighted box within the figure where curves are also more linear for $s>50$. In all our experiments, including image regression and surface reconstruction tasks, we use a constant number of $s$, i.e., $80$. An overly large value of $s$ brings instability to the network, resulting in performance reduction. Note that in FNN, $s$ does not have similar properties as FINN.
\begin{figure*}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{parameter.pdf} }
\end{center}
\vspace{-6pt}
\caption{Impact of hyperparameters. (Left) Images are reconstructed using different parameters in positional encoding ($\sigma, s$ in Equation~\ref{equ1}). By increasing $\sigma$, image detail becomes clear (a$\rightarrow$b), and noise appears when $\sigma$ is overly high (c). Those images are sampled from the distribution (Right), from which the peak region appears around $\sigma=10$. As deviated from the peak region, the performance decreases monotonically. The other parameter $s$ has less impact on the image quality if $\sigma$ is fixed.}
\label{fig:param}
\end{figure*}
\begin{figure*}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{scale.pdf} }
\end{center}
\vspace{-6pt}
\caption{Convergence speed scales by the magnitude of Fourier feature controlled by parameter $s$. As $s$ increases the convergence speed grows and reaches stable until $s>50$. The curves in the highlighted box indicate higher $s$ also improves network stability.}
\label{fig:scale}
\end{figure*}
\fi
\iffalse
\begin{figure*}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{iteration.pdf} }
\end{center}
\vspace{-6pt}
\caption{
The influence of the number of hidden layers in FINN. In (1), with only one hidden layer, the reconstructed result still contains notable noisy artifacts. The results are improved when more hidden layers are used (2-5). With more than one hidden layer, our method effectively removes the visual artifacts while keeping small patterns, e.g., in the regions of the sky and ocean. The numerical and visual results are merely changed when the number of the hidden layer exceeds one (2-5).}
\label{fig:iteration}
\end{figure*}
\subsubsection{Number of MLP Layers.}
Our filtering function performs iteratively in the hidden layers of MLPs. In Figure~\ref{fig:iteration}, we showcase the performances of FINN using a different number of hidden layers, ranging from one layer to five layers. With only one hidden layer, the noisy artifacts are still notable in the image, and as the number of hidden layers increases, both visual and numerical results are improved. Besides filtering artifacts, FINN also enhances details in blurry regions. The performance of FINN with only one hidden layer (FINN$_{1l}$) is already better than state-of-the-art methods, as shown in Table~\ref{table:result}. Continuous adding more hidden layers to the network does not produce better results when the number of hidden layers already exceeds two.
\fi
\section{Related Work}
\subsubsection{Implicit Neural Representation.}
Deep neural network has been shown to be highly effective for learning implicit function that represents images~\cite{tancik2020fourfeat,sitzmann2019siren,MildenhallSTBRN20,chen2021learning} and 3D objects and scenes~\cite{park2019deepsdf,OccNet,chen2018implicit_decoder,AtzmonL20,GroppYHAL20,SitzmannZW19}.
They incorporate coordinates as inputs to MLPs that could be sampled to an arbitrarily high spatial resolution. Thus such representation can directly be applied to super-resolution tasks. Other applications include view-synthesis~\cite{MildenhallSTBRN20,Martin-BruallaR21,yu2020pixelnerf,Bemana2020xfields}, point cloud based 3D reconstruction~\cite{AtzmonL20,GroppYHAL20,WilliamsTBZ21,WilliamsSSZBP19} and single-image 3D reconstruction~\cite{OccNet,DISN,XuFYS20,saito2019pifu}. Besides generative tasks, implicit representation is also applicable and effective for other tasks, such as feature matching~\cite{ChenFBMG21} and scene understanding~\cite{ZhiICCV2021}
As evidenced by~\cite{basri20a,rahaman19a},
ReLU-based MLP networks are struggling to capture highly detailed signals, due to the spectral bias trait. FFN uses positional encoding to map input coordinates of signals to a high dimensional space using sinusoidal functions.
Meanwhile, SIREN~\cite{sitzmann2019siren} replaces the ReLU activations in the MLP network with sinusoidal functions. They share a similar spirit that manipulates the input or intermediate results in the frequency domain to capture a wider frequency bandwidths in the output. Further, to better fit individual data, the sinusoidal functions in positional encoding or MLP networks are devised to be learnable~\cite{FathonySWK21,Mehta21}.
In contrast to using sinusoidal functions, Spline Positional Encoding (SPE)~\cite{WangLYT21} studies learnable spline functions for coordinate embedding. With sufficient local supports of splines, the SPE can also approximate the signal with high frequencies. However, when using a small number of local supports, the boundaries between patches of the spline become notable that introduces stripe noises and significantly reduce the visual quality.
Reconstructing high-frequency details is commonly at the expense of introducing visual artifacts to the results. Some recent advances have led to structured or hierarchical designs that can further close the gap between generated results and the target function. They divide the complex functions of 3D object and scenes~\cite{ChabraLISSLN20,GenovaCSSF20,JiangSMHNF20} or image~\cite{Mehta21} into regular subregions, and fitting each subregions individually while considering globally consistency.
Region subdivision can hardly take effect when a subregion itself is complex. More recently, spatial adaptation on frequencies appears as an improved solution. SAPE~\cite{hertz2021sape} presents a progressive optimization strategy to encode signals with increasing frequencies at individual spatial locations. The method reduces the noisy artifacts but generates over-smoothed results since it resists overfitting and leans toward underfitting.
This paper seeks an alternative scheme that uses filters to adapt the intermediate outputs of the MLPs.
Existing methods, such as MFN~\cite{FathonySWK21} and BACON~\cite{lindell2021bacon}, pi-GAN~\cite{chanmonteiro2020pi-GAN}, and ModSIREN~\cite{Mehta21} also adapts the intermediate results of MLPs by modulating the activation functions. Particularly, they modulate frequency, phase, or amplification parameters of sinusoidal functions. Tuning those parameters may cause global and possibly large changes in the results, hence those parameters are difficult to optimize and sensitive to initialization. Rather than modulating sinusoidal functions, FINN runs the filtering after each MLP layer, outside sinusoidal functions, thus allowing more local controls on the intermediate results and more stable performance.
\begin{figure*}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{2d-teaser.pdf} }
\end{center}
\caption{
Effects of filtering operators. The target (e) contains homogenous regions everywhere and some tiny-scale texts each referred to low-frequency or high-frequency respectively. (a) is generated by FFN with no filtering and observes random noises (yellow rectangle). With smoothing (b), smoother texts are reconstructed but at the expense of smoothing out some small-scale structures (green rectangle). In contrast, the recovering operator (c) restores fine details but is incapable of reducing noise. Combining two filtering operators enables FINN smooths noises and recovers small structures (d).
}
\label{fig:idea2}
\end{figure*}
\subsubsection{Image Filtering with Deep Neural Networks.}
Image filters usually compute the weighted average of neighboring pixels of the image as output or sometimes leverage regularization terms for image optimization.
Recently, a line of research presents filters with neural networks that can be learned from a large number of datasets for various applications~\cite{JampaniKG16,ZhangPLL017,YoonJYLK15,harbiCBHD17,ChenXK17,LiHAY19}, such as image super-resolution, denoising, and deblurring, etc. They are mainly based on the convolutional network from which the local neighborhood of the query point is determined.
For continuous functions, it is difficult to be aware of the neighbors except to discretize the input domain with prior feature scales or resolutions.
Thus, we attempt to devise a general filtering scheme for MLPs dealing with image functions and other continuous signals and handling different frequency bandwidths.
\section{Experiments}
\label{sec:exp}
We validate the benefits of applying filtering in learning 2D images and 3D shapes, i.e., single fields, and novel synthesis supported by color and density field, i.e., hybrid field.
\subsection{Image Regression and Generalization}
Image functions map a 2D coordinate to RGB color. Our network has 512-dimensional Fourier features ($\sigma=10, s=80$ for $\gamma(x)$ and $\gamma_s(x)$), MLPs with 3 hidden layers, 256 channels, ReLU activation and Sigmoid on output. In addition, our filtering module contains a 512$\times$256 matrix for generating scaling vectors.
We use MSE loss to train the network for 2k epochs with a learning rate of 1e-3.
\begin{table}[t!]
\footnotesize
\tabcolsep=0.07cm
\begin{tabular}{l>{\columncolor{mygray}}ccc>{\columncolor{mygray}}cc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{\centering Model}} & \multicolumn{2}{c}{PSNR$\uparrow$} && \multicolumn{2}{c}{\protect\reflectbox{F}LIP\xspace$\downarrow$} \\ \cmidrule{2-3} \cmidrule{5-6}
\multicolumn{1}{c}{} & \multicolumn{1}{>{\columncolor{white}}c}{Natural} & Text && \multicolumn{1}{>{\columncolor{white}}c}{Natural}& Text \\ \midrule
FFN & 25.57 $\pm$ 4.19 & 30.47 $\pm$ 2.11 && 0.131$\pm$0.041 & 0.096$\pm$0.043 \\
SIREN & 27.03 $\pm$ 4.28 & 30.81 $\pm$ 1.72 && 0.114$\pm$0.040 & 0.070$\pm$0.020 \\
SPE & 26.49 $\pm$ 3.89 & 31.12 $\pm$ 2.18 && 0.130$\pm$0.038 & 0.065$\pm$0.022 \\
SAPE & 28.09 $\pm$ 4.04 & 31.84 $\pm$ 2.15 && 0.118$\pm$0.026 & 0.083$\pm$0.041 \\\midrule
FINN & \textbf{28.51 $\pm$ 4.35} & \textbf{33.09 $\pm$ 1.97} && \textbf{0.100$\pm$0.037} & \textbf{0.042$\pm$0.016} \\\bottomrule
\end{tabular}
\caption{Quantitative results of image generalization.}
\label{table:result}
\end{table}
We compare our method to state-of-the-art methods, including FFN~\cite{tancik2020fourfeat}, SIREN~\cite{sitzmann2019siren}, SPE~\cite{WangLYT21} and SAPE~\cite{hertz2021sape}, on the datasets of ``natural'' and ``text'' images from FFN.
The training pixels are sampled on a regularly-spaced grid containing $25\%$ of the pixels in the image.
Then, we use all the pixels for testing.
We report PSNR and \protect\reflectbox{F}LIP\xspace~\cite{Andersson2020} in Table~\ref{table:result} and qualitative results in Figure~\ref{fig:2dresults}. \protect\reflectbox{F}LIP\xspace provides an error map and a global measurement (i.e., the weighted median of errors on all pixels).
Our method consistently outperforms all compared methods on all datasets and metrics with considerable gains.
The visual results show that all the methods can generate realistic images. Nevertheless, SAPE produces over-smoothed images. FNN and SIREN usually include random noisy artifacts, and SPE contains stripe noise.
FINN reconstructs better in both backgrounds (low-frequency) and small structures (high-frequency), handling over smooth and noise well.
\begin{table}[tp!]
\footnotesize
\tabcolsep=0.08cm
\begin{center}
\begin{tabular}{l>{\columncolor{mygray}}cc>{\columncolor{mygray}}cc}\toprule
\rowcolor{white}
& FINN$_{\text{so}}$ &FINN$_{\text{ro}}$ &FINN$_{\text{in}}$ &FINN$_{\text{fxs}}$ \\ \midrule
Natural & 27.91 $\pm$ 4.21 & {28.09 $\pm$ 4.06} &28.26 $\pm$ 4.35 & \textbf{28.59 $\pm$ 4.39} \\
Text & 31.19 $\pm$ 1.86 & {31.99 $\pm$ 1.73} & 32.22 $\pm$ 2.58 &\textbf{33.17 $\pm$ 2.01} \\\bottomrule
\end{tabular}
\caption{FINN variants on image generalization.}
\label{table:qr2}
\end{center}
\end{table}
\begin{figure}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{2dresults.pdf} }
\end{center}
\caption{Qualitative results of image generalization.}\label{fig:2dresults}
\end{figure}
\subsubsection{Design Choices.}
We validate each component of our filter and show the numerical results of the variants in Table~\ref{table:qr2}.
\textbf{Using smoothing or recovering operator only}. FINN$_{\text{ro}}$ and FINN$_{\text{so}}$ incorporate the recovering and smoothing operator to FFN, respectively. The two variants perform worse than FINN but both outperform FFN. It indicates that either suppressing noises or enhancing details helps to improve the reconstruction. Figure~\ref{fig:idea2} shows the filtering effects of individual operators. Their combination reaches an adaptive reconstruction for different levels of detail.
\textbf{Filtering on inputs instead of features of MLPs}. We apply the filter to input Fourier features with a variant denoted by FINN$_{\text{in}}$. Although FINN$_{\text{in}}$ also learns adaptive feature embedding and achieves better performance than FFN, its performance is inferior to FINN since the no filtered layers of MLPs can still overfit the results, making the smoothing filter less effective.
\textbf{Using layer-wise scaling vectors $F_x^i$ instead of a global $F_x$}. FINN$_{\text{fxs}}$ means scaling vectors passed to different MLP layers are generated independently. Even adding more network parameters, its performance gain over the global $F_x$ is negligible. The main reason is that $F_x$ and all $F_x^i$ are derived directly from the same input Fourier features by a simple linear transformation, which limits their capability of generating diverse scaling vectors. It is possible to use a more complex scaling vector generator instead of a simple linear mapping to boost the performance. But it is not encouraged since a stronger recovering operator will take over the smoothing operator and thus reduce the smoothing effects. In order to control the sharpness, one can manipulate the frequency of Fourier features for scaling vector generator. The filtering effects of various $F_x$ are shown in Figure~\ref{fig:filteringeffect}.
\subsubsection{Network Convergence and Stability.}
Figure~\ref{fig:convergence} shows test PSNR curves of FFN (Left) and FINN (Right), trained on 32 images from the “natural” and “text” datasets for 2k iterations. FINN converges after 300 iterations and keeps stable PSNR scores for individual images since then. In contrast, FFN converges slower and experiences more randomness during optimization, yielding slightly curvy and nonmonotonic PSNR curves. The stability suggests filtering operators serve as a regularization to resist over-fitting during optimization, leading to better generalization.
\begin{figure}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{psnr.pdf} }
\end{center}
\caption{Convergence and stability comparison of networks with and without filters.
}
\label{fig:convergence}
\end{figure}
\subsection{3D Inverse Rendering for View Synthesis}
The view synthesis task receives a set of 2D images with known camera poses and generates images from novel viewpoints. We construct a vector field on 3D space from input images to represent the scene using Neural Radiance Fields (NeRF)~\cite{MildenhallSTBRN20} that utilizes MLPs to map from 3D coordinates to RGB colors and volume densities. Then one can render images from the field with given viewpoints via volume rendering. NeRF is optimized via MSE loss between the input images and their re-rendered images from the predicted field. Following FFN, we use a ``simplified NeRF'' that removes hierarchical sampling and view dependence from the original NeRF framework for evaluation.
We let FINN and FFN have the same number of network parameters: FFN, 512-dimensional Fourier features, MLP with {6} layers, 256 channels, ReLU activation, sigmoid on RGB output; and FINN, similar to FFN, except that a {4} layers MLP and a scaling vector generator with 512$\times$256 parameters. Both models are trained for 50k iterations using the Adam optimizer with a learning rate of 5$\times$10$^{-4}$.
We use two datasets for evaluation, a synthesized scene ``Lego'' and a real scene ``Fern''. The number of training and testing images is (100, 25) for ``Lego'' and (17, 3) for ``Fern''.
FFN sets $\sigma$ to $6.05$ for Lego and $5.0$ for Fern by grid search.
Different from the parameter setting in image regression, we uses a different $\sigma$ for $\gamma(x)$ and $\gamma_s(x)$ in learning NeRF, such that $12,1$ for Lego and $10,1$ for Fern.
The main reason is NeRF contains a density field and an RGB color field that have different complexness and require different frequencies. For example, images (color fields) expect a higher $\sigma$, 10, and signed distance fields (similar to density fields) only use $\sigma\!\!\approx\!\!1$ described in the next section.
Thus a better encoding scheme will use different frequency parameters for individual fields. However, this will need more sophisticated designs and is beyond our research. FINN provides a simple and flexible way to model such hybrid fields by setting {high frequency} in $\gamma(x)$ and {low frequency} in $\gamma_s(x)$), allowing the network to adjust the local variations of the features and final output in a wide range.
The results in Table~\ref{table:nerf} show that FINN performs better than FFN in both training and testing on the two datasets. Figure~\ref{fig:nerf} shows FINN also provides better rendering quality, revealing more distinct details with fewer noises in the results. For sparse training images from Fern, the performance gain and visual improvement of FINN are significant, evidencing that FINN can fit a wider frequency bandwidth and generalizes better to unknown viewpoints.
\begin{figure}[tp!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{nerf.pdf} }
\end{center}
\caption{Qualitative results on novel view synthesis.}
\label{fig:nerf}
\end{figure}
\begin{table}[tp!]
\footnotesize
\tabcolsep=0.07cm
\begin{center}
\begin{tabular}{l>{\columncolor{mygray}}ccc>{\columncolor{mygray}}cc}\toprule
\multicolumn{1}{l}{\multirow{2}{*}{\centering Model}}& \multicolumn{2}{c}{Lego} && \multicolumn{2}{c}{Fern} \\\cmidrule{2-3} \cmidrule{5-6}
\rowcolor{white}
&Train PSNR & Test PSNR &&Train PSNR & Test PSNR \\ \midrule
FFN & 26.25 $\pm$ 1.00 &25.74 $\pm$ 1.06 && 25.06 $\pm$ 0.44 & 24.30 $\pm$ 0.67 \\
FINN & \textbf{26.70 $\pm$ 1.00} & \textbf{26.02 $\pm$ 1.10} && \textbf{25.99 $\pm$ 0.53} & \textbf{25.28 $\pm$ 0.63} \\\bottomrule
\end{tabular}
\caption{Numerical results on novel view synthesis.}
\label{table:nerf}
\end{center}
\end{table}
\iffalse
\begin{table}[h!]
\footnotesize
\tabcolsep=0.2cm
\begin{center}
\begin{tabular}{l|c|c|c}\hline
& FFN (4-layers) & FFN (6-layers) & FINN \\ \hline
Train PSNR & 26.12 $\pm$ 1.05 & 26.25 $\pm$ 1.00 & 26.59 $\pm$ 1.02 \\
Test PSNR & 25.52 $\pm$ 1.08 & 25.74 $\pm$ 1.06 & 26.01 $\pm$ 1.14 \\\hline
\end{tabular}
\caption{NERF.}
\label{table:nerf}
\end{center}
\end{table}
\fi
\subsection{3D Shape Reconstruction}
Further, we evaluate our method in the 3D surface reconstruction task. We consider the 3D surface as a zero level-set of the signed distance field (SDF) from which every 3D query point has a corresponding value telling its distance and whether inside or outside from the surface. We train the network to fit the SDF on point clouds with additional point normals information. In addition to SDFs and normals for supervision, we also apply a regularization term to constrain the gradient of the learned field to be a unit vector at any spatial locations, as suggested by~\cite{AtzmonL20,GroppYHAL20}. The loss function is defined as
\begin{equation}
\begin{aligned}
\mathcal{L}_{\text{sdf}} &=\sum_{x\in \Omega}|\|\nabla f_{\theta}(x)\| -1|
+ \sum_{x\in \Omega\setminus \Omega_0} {\exp(-|f_{\theta}(x)|)} \\
&+ \sum_{x\in \Omega_0}(|f_{\theta}(x)|+\|\nabla f_{\theta}(x) - n(x)\|)
\end{aligned}
\label{eq4}
\end{equation}
where $\Omega$ is the point set and $\Omega_0$ contains only on-surface points. $f_{\theta}(x)$ and $\nabla f(x)$ are the fitted SDFs and gradients, and $n(x)$ is the ground truth point normal. Each loss term is weighted by a constant for optimization.
Our network structure for this task is similar to image regression, except that the output layer is linear, and hidden layers increase from 3 to 4.
SDFs are \emph{scalar} fields, usually less complex than vector fields, e.g., RGB color fields, thus we use a small frequency, $\sigma \in [1,2]$ for FFN and FINN, and reduce the dimension of the Fourier features to 256. Training SDF requires dense samples on the underlying surface and sparse samples randomly scattered in a bounding cube. After trained, we sample a set of regular grid points to obtain the SDFs and use the Marching Cubes algorithm~\cite{lorensen1987marching} to extract the triangular mesh.
We compare our method with FFN and SIREN on large-scale 3D meshes containing tens of millions of triangles. The examples include an object ``Thai Statue'', containing very complex geometric features, and a scene ``Interior Room'', having large flat patches and small shapes, as shown in Figure~\ref{fig:3d}. We set $\sigma$ to $1$ for those two examples for FFN and FINN and use 1600$^3$ grid points to extract high-resolution triangular meshes. The results show that FINN reconstructs both flat (low-frequency) and sharp (high-frequency) surface regions with high fidelity. The results of FFN and SIREN are also plausible, but introduce unexpected local bumps and noises and lose many fine surface details. We will present numerical results in the appendix.
\begin{figure}[t!]
\begin{center}
{\includegraphics[width=1.0\linewidth]{3d.pdf} }
\end{center}
\caption{
Large-scale surface reconstruction. All methods show good reconstructions. With filtering, FINN suppresses high frequencies on homogenous regions and restores sharp geometric features.
}\label{fig:3d}
\end{figure}
{
}
\iffalse
\begin{table}[h!]
\footnotesize
\tabcolsep=0.08cm
\begin{center}
\begin{tabular}{l|c|c|c|c|c|c|c}
&Fandisk & Bimba & Bunny & Gargoyle & Dfaust & Armadillo & Dragon \\ \hline
SPE & \textbf{1.3} & 1.6 & \textbf{1.5} & 2.4 & 9.1 & \textbf{1.3} & 1.8 \\ \hline
FFN & \textbf{1.3} & 1.7 & \textbf{1.5} & 2.6 & 8.6 & \textbf{1.3} & \textbf{1.3} \\ \hline
FINN & \textbf{1.3} & \textbf{1.5} & \textbf{1.5} & \textbf{2.3} & \textbf{8.4} & \textbf{1.3} & \textbf{1.3} \\
\end{tabular}
\caption{Numerical results on surface reconstruction from 7 point sets. The Chamfer distance is multiplied by 10000.}
\label{table:qr}
\end{center}
\end{table}
\fi
|
\section{Introduction}
The realization of the superconductivity (SC) in Sr-doped NdNiO$_2$ films grown on a SrTiO$_3$ substrate~\cite{2019Nature} is the breakthrough in the decade-long pursuit of rare-earth nickelates~\cite{Junjie_review,Botana_review}. Even more fascinating is the successful synthesis of superconducting doped Pr$_{1-x}$Sr$_{x}$NiO$_2$~\cite{Pr}, La$_{1-x}$Ca$_{x}$NiO$_2$~\cite{LaSC}, and undoped quintuple-layer Nd$_6$Ni$_5$O$_{12}$~\cite{Nd6Ni5O12}, which indicates the advent of Ni-based superconducting family. Undoubtedly, the most important question is the pairing symmetry and mechanism in this new family of superconductors and furthermore its similarity and difference from other transition metal based superconductors like Cu- and Fe-based compounds. For example, given that the parent compound of infinite-layer NdNiO$_2$ is discovered not to host the long-range magnetic order~\cite{Pickett2004,Hepting,dft1,dft9,dft16,dft18,dft26,dft25,noLRO} in spite of magnetic correlations~\cite{mag1,mag2,mag3,mag4,mag5}, whether the debated spin-fluctuation scenario for cuprate superconductors also applies for the infinite-layer nickelates is an intriguing question.
Physically, the minimal microscopic model to capture the essential physics of infinite-layer nickelates is a demanding task. Previous numerous electronic structure calculations based on DFT~\cite{Pickett2004,dft1,dft2,dft3,dft4,dft5,dft6,dft7,dft8,dft9,dft10,dft24,dft26} and/or DFT+DMFT~\cite{dft101,dft102,dft11,dft12,dft13,dft14,dft15,dft16,dft17,dft18,dft19,dft20,dft21,dft22,dft23,dft25,Held2020,Held2022,Hanghuireview} methodologies have provided fruitful understanding of the electronic structure of infinite-layer nickelates including R$_{n+1}$Ni$_{n}$O$_{2n+2}$ with R = La, Nd, Pr for varying $n$~\cite{Botana_review,Hanghuireview,Held2022}, whose calculated properties have been partly confirmed by recent experiments for $n=5$~\cite{Nd6Ni5O12}. One distinct feature associated with infinite-layer nickelates lies in the presence of the Nd-derived bands, whose self-doping effect on the magnetic and superconducting properties is highly debated, which is manifested by the opposite arguments supporting the single Ni-$d_{x^2-y^2}$ band description~\cite{oneband,dft101,dft4,dft24} versus the multi-orbital scenarios~\cite{Hepting}, e.g. Ni-$e_g$ with or without other Nd-5d orbitals.
The most significant theoretical reasoning underlying the single-orbital scenario arises from the demonstration of the depletion of the rare-earth bands, namely the extremely low electron density of rare-earth layer, which results in a Fermi surface reconstruction that might render the infinite-layer nickelates essentially akin to cuprates. Despite there have been strong theoretical~\cite{Held2022,dft101,dft4,dft24} and experimental~\cite{oneband} support on this scenario, it is still not completely resolved whether and how the low occupation of rare-earth bands impact the charge~\cite{Pr,Hall1,Hall2}, magnetic~\cite{GuangMing,GuangMing1}, and superconducting properties~\cite{Mi2020,dswave,dft2,dft3,dft101,dft102,spinfluc1,spinfluc2} in the whole phase diagram.
\section{Model and Methodology}
\subsection{Two-orbital $d$-$s$ model}
To provide more insights on the prominent role of the rare-earth bands, we employ a minimal two-orbital $d$-$s$ model, where $d$-orbital mimics Ni-$3d_{x^2-y^2}$. Additionally, we include the rare-earth R-$5d$ orbitals, which can be effectively approximated as an interstitial s-like orbital located at the missing apical Oxygen above and below Ni in the rare-earth layers. The Hamiltonian reads as follows
\begin{align} \label{eq:HM}
\hat{H} = & \sum_{k\sigma} (E^d_k n^d_{k\sigma} +E^s_k n^s_{k\sigma}) + U \sum_i n^d_{i\uparrow}n^d_{i\downarrow}\nonumber\\
& + \sum_{k\sigma} V_k (d^\dagger_{k\sigma}s^{\phantom\dagger}_{k\sigma}+h.c.)
\end{align}
with two orbitals' dispersion and $d$-$s$ hybridization as
\begin{align} \label{dis}
E^d_k = & -2 t_d (\cos k_x+\cos k_y)+4 t'_d \cos k_x\cos k_y -\mu \nonumber\\
E^s_k = & -2 t_s (\cos k_x+\cos k_y) + \epsilon_s -\mu \nonumber\\
V_k = & -2 V (\cos k_x+\cos k_y)
\end{align}
where $d^{\dagger}_{k\sigma}(s^{\dagger}_{k\sigma})$ are creation operators in momentum space for two orbitals. $n^{d(s)}_{i\sigma}$ and $n^{d(s)}_{k\sigma}$ are the associated number operators in the real and momentum spaces separately. The chemical potential $\mu$ can be tuned for a desired total electron density. The on-site energy $\epsilon_s$ of $s$ orbital is a tunable parameter to adjust the relative electron density between two orbitals with the criterion that the $s$ orbital's occupancy is around 6-10\% in an undoped system (total electron density $n=1$) in order to be consistent with the experimental finding that there are approximately 8\% electrons per formula unit for NdNiO$_2$~\cite{Hall1}.
We adopt the convention that the nearest-neighbor hopping between $d$ orbitals $t_d=1$ is set as the energy unit. Following the recent numerical investigation~\cite{TPD_ds}, we choose $t'_d=-0.25, t_s=0.5$ as typical hopping integrals.
Besides, the impact of on-site Coulomb repulsion $U$ of $d$ orbital will be studied in the present work.
The specific dispersion of the $d$-$s$ hybridization $V_k$ might quantitatively affect the detailed evaluation of various physical properties. Nevertheless, as pointed out by Ref.~\cite{TPD_ds}, although the symmetry analysis of the real infinite-layer materials reveals that the inter-layer hoppings between Ni-3d and the rare-earth R-5d orbitals can be negligibly small, the electrons in the rare-earth layer are highly itinerant so that the detailed form of $V_k$ probably only has minor effects. For simplicity, therefore, we focus on the model Hamiltonian where $d$-$s$ has only nearest-neighbor hoppings~\cite{dft25}.
One important issue we will explore is the magnitude of $V$, which is believed to be small $V/t_d \sim 0.1$ in previous studies~\cite{TPD_ds}. Strikingly, however, a recent investigation indicated that the dominant hybridization between Ni-3d orbitals and itinerant electrons of this interstitial s orbital comes from a large inter-cell hopping~\cite{dft25}. Therefore, the present work will not restrict our attention to only small hybridization $V$ but will study the influence of its magnitude. Specifically, we choose both $V/t_d = 0.1$~\cite{TPD_ds} and $V/t_d = 0.6$ estimated from Ref.~\cite{dft25} as two typical values. The focus in this work is to investigate the relevance of the $d$-$s$ model to the infinite-layer nickelates, in particular the most attractive superconducting instabilities in the hole doped systems.
Before proceeding, we remark on some extensively investigated limits of the $d$-$s$ model. On the one hand, at $V=0$, namely the two-dimensional single-band Hubbard model, there have been tremendous theoretical effort for its solutions and more importantly its relevance in strongly correlated quantum materials like cuprate and nickelate superconductors. Regarding its superconducting properties, the consensus is that the $d$-wave pairing is appreciated owing to its retarded nature to minimize the impact of local repulsion $U$~\cite{ScalapinoRMP,Maier06}. On the other hand, at finite hybridization $V$ but $t_d=t'_d=0$, the model recovers the conventional periodic Anderson model (PAM), which is widely treated as the standard model of describing the heavy fermion physics. The deeper connection and moreover the transition or crossover between PAM and $d$-$s$ model will be left for another future independent study.
\subsection{Dynamical cluster approximation}
We adopt the dynamical cluster approximation (DCA)~\cite{Hettler98,Maier05,code} with a continuous time auxilary field (CT-AUX) quantum Monte Carlo (QMC) cluster solver~\cite{GullCTAUX} to
numerically solve the $d$-$s$ model Eq.~\eqref{eq:HM}.
As one of the successful many-body methods, DCA has provided much insight on the strongly correlated physics especially for the Hubbard-type models~\cite{Hettler98,Maier05}. In principle, DCA evaluates the physical properties in the thermodynamic limit via mapping the bulk lattice problem onto a finite cluster embedded in a mean-field self-consistently, where the effective cluster problem is exactly solved by various perturbative or non-perturbative many-body methods~\cite{Hettler98,Maier05}. In practice, DCA is realized by a self-consistent loop for the convergence between the cluster and coarse-grained single-particle Green's function. The truly time-consuming part of the DCA loop lies in the cluster solver, e.g. CT-AUX adopted here. In other words, most uncertainties of DCA method arises from the cluster solver method e.g. quantum Monte Carlo procedure.
To simulate a wide range of doping levels and Hubbard interaction, we stick on the smallest $N_c=2\times2$ DCA cluster to keep the QMC sign problem manageable~\cite{GullCTAUX}
down to the superconducting transition temperatures $T_c$.
In fact, the qualitative trend of our major results can be verified with more expensive simulations employing larger clusters such as $N_c=8$.
To characterize the superconducting instability, we rely on the leading eigenvalues of the Bethe-Salpeter equation (BSE) in the eigen-equation form in the particle-particle channel~\cite{Maier06,Scalapino06}
\begin{align} \label{BSE}
-\frac{T}{N_c}\sum_{K'}
\Gamma^{pp}(K,K')
\bar{\chi}_0^{pp}(K')\phi_\alpha(K') =\lambda_\alpha(T) \phi_\alpha(K)
\end{align}
where $\Gamma^{pp}(K,K')$ denotes the lattice irreducible particle-particle vertex of the effective cluster problem with combining the cluster momenta ${\bf K}$ and Matsubara frequencies $\omega_n=(2n+1)\pi T$ to $K=({\bf K},i\omega)$. One of the key DCA assumptions is that the desired lattice two-particle irreducible vertex $\Gamma$ is approximated by the its cluster counterpart $\Gamma_c$, which can be obtained by
\begin{align} \label{e2}
\chi_{c\sigma\sigma'}(q,K,K') &= \chi^0_{c\sigma\sigma'}(q,K,K') + \chi^0_{c\sigma\sigma''}(q,K,K'') \nonumber \\
& \times \Gamma_{c\sigma''\sigma'''}(q,K'',K''') \chi_{c\sigma'''\sigma'}(q,K''',K')
\end{align}
where the cluster two-particle Green's function
\begin{align} \label{e1}
\chi_{c\sigma\sigma'}(q,K,K') &= \int^{\beta}_0 \int^{\beta}_0 \int^{\beta}_0 \int^{\beta}_0 d\tau_1 d\tau_2 d\tau_3 d\tau_4 \nonumber \\
& \times e^{i[(\omega_n+\nu)\tau_1 -\omega_n\tau_2 +\omega_{n'}\tau_3 -(\omega_{n'}+\nu)\tau_4]} \nonumber \\
\times \langle \mathcal{T} & c^{\dagger}_{K+q,\sigma}(\tau_1) c^{\phantom{\dagger}}_{K\sigma}(\tau_2) c^{\dagger}_{K'\sigma'}(\tau_3) c^{\phantom{\dagger}}_{K'+q,\sigma'}(\tau_4) \rangle
\end{align}
with conventional notation $K=(\mathbf{K},i\omega_n)$, $K'=(\mathbf{K'},i\omega_{n'})$, $q=(\mathbf{q},i\nu)$ and the time-ordering operator $\mathcal{T}$ can be calculated numerically via a DCA cluster solver (CT-AUX in our case). Concurrently, the non-interacting two-particle Green's function $\chi^0_{c\sigma\sigma'}(q,K,K')$ is constructed from the product of a pair of fully dressed single-particle Green's functions. Note the usual convention that the summation is to be made for repeated indices. In this work, we are mostly interested in the even-frequency even-parity (spin singlet) $d$-wave pairing tendency so that $q=(\mathbf{q},i\nu)=0$ is assumed~\cite{Maier06,Scalapino06}, despite that there exists odd-frequency pairing channels at high temperatures that quickly decays with lowering $T$.
In the BSE Eq.~\eqref{BSE}, the coarse-grained bare particle-particle susceptibility
\begin{align}\label{eq:chipp}
\bar{\chi}^{pp}_0(K) = \frac{N_c}{N}\sum_{k'}G(K+k')G(-K-k')
\end{align}
is obtained via the dressed single-particle Green's function $G(k)\equiv G({\bf k},i\omega_n) =
[i\omega_n+\mu-\varepsilon_{\bf k}-\Sigma({\bf K},i\omega_n)]^{-1}$, where $\mathbf{k}$ belongs to the DCA patch surrounding the cluster momentum $\mathbf{K}$, $\mu$ the chemical potential, $\varepsilon_{\bf k}$ the
dispersion relation, and $\Sigma({\bf K},i\omega_n)$ the cluster self-energy.
In practice, we normally choose 16 or more discrete points for both the positive and negative fermionic Matsubara frequency $\omega_n=(2n+1)\pi T$ mesh for measuring the two-particle Green's functions and irreducible vertices $\Gamma$. Therefore, the BSE Eq.~\eqref{BSE} effectively becomes an eigenvalue problem of a matrix of size $(32N_c)\times (32N_c)$, which can be routinely solved straightforwardly.
Physically, the magnitude of the eigenvalue $\lambda_\alpha(T)$ denotes the normal state pairing tendency in the superconducting channel. Correspondingly, the spatial, frequency, and orbital dependence of the eigenvector $\phi_\alpha({\bf K},i\omega_n)$ can be viewed as the normal state analog of the superconducting gap function to reflect the structure of the pairing interaction $\Gamma^{pp}$~\cite{Maier06,Scalapino06}.
The superconducting $T_c$ is extracted via the temperature where the leading eigenvalue of Eq.~\eqref{BSE} $\lambda(T_c)=1$.
Throughout this work, it is found that the leading pairing symmetry occurs for the $d$-wave channel with momentum structure $\cos K_x - \cos K_y$ so that we are only concerned in the leading eigenvalues $\lambda_d$ and associated $\phi_d({\bf K},i\omega_n)$. In other words, the presence of $s$ orbital and its hybridization with $d$ orbital does not affect the $d$-wave pairing symmetry of the conventional Hubbard model with solely $d$ orbital. Its implication on the infinite-layer nickelates including the recent experimental findings will be discussed later on.
\begin{comment}
Note that the above formalism has their counterparts in the magnetic and charge channels. For the Hubbard model, the dominant contribution on the $d$-wave pairing interaction has been shown to arise from the spin-one ($S=1$) particle-hole channel~\cite{Maier06,Scalapino06}.
Therefore, it is requisite to investigate the role of additional $s$ orbital on this spin-fluctuation mediated pairing mechanism in $d$-$s$ model. To this aim, we also extract the irreducible particle-hole vertex $\Gamma^{ph}$ in the magnetic channel of the effective cluster problem from Eq.~\eqref{e2}. Besides, the coarse-grained bare particle-hole susceptibility is given by
\begin{align}\label{eq:chiph}
\bar{\chi}^{ph}_{0}(q,K,K') = \delta_{KK'} \frac{N_c}{N} \sum_{k'} G(K+k') G(K+k'+q)
\end{align}
where only the component at momentum transfer ${\bf Q}=(\pi,\pi)$ and zero frequency transfer ($i\nu=0$) is our focus similar to the particle-particle superconducting channel Eq.~\eqref{BSE}. which can be obtained by the coarse-grained two-particle Green's function $\bar{\chi}^{ph}(q,K,K')$ (instead of cluster quantities that result in cluster susceptibilities), which is in turn calculated via the coarse-grained BSE transformed from Eq.~\eqref{e2} as
\begin{align}\label{}
[\bar{\chi}^{ph}(q,K,K')]^{-1} = [\bar{\chi}^{ph}_{0}(q,K,K')]^{-1} - \Gamma^{ph}(q,K,K')
\end{align}
Then our interested spin susceptibility $\chi_s(q,T)$ can be calculated as
\begin{align}\label{chics}
\chi_{s,c}(q,T) = \frac{T^2}{N_c^2} \sum_{K,K'} \bar{\chi}^{ph}(q,K,K')
\end{align}
\end{comment}
\section{Results}
\begin{figure}
\psfig{figure=fig1a.pdf,height=5.5cm,width=8.5cm,angle=0,clip}
\psfig{figure=fig1b.pdf,height=5.5cm,width=8.5cm,angle=0,clip}
\psfig{figure=fig1c.pdf,height=5.3cm,width=8.9cm,angle=0,clip}
\caption{(a) Temperature dependence of the leading ($d_{x^2-y^2}$-wave) eigenvalue $\lambda_ {d}(T)$; (b) Doping dependence of the $d$-wave superconducting $T_{c}$ extracted from
$\lambda_d(T_c)=1$; (c) Orbital occupancy versus the doping level at $T/t_d=0.08$. The parameters are $U/t_d=7, N_c=4$.}
\label{lambda}
\end{figure}
Fig.~\ref{lambda} summarizes our key results. The panel (a) illustrates the temperature dependence of the leading $d$-wave eigenvalue $\lambda_d(T)$ for different $V$ at two fixed total electron densities $n=0.9, 0.85$, namely the total hole doping $\delta=0.1, 0.15$ respectively.
What one sees is that the presence of the hybridization suppresses $d$-wave pairing regardless of the doping level, which is expected as the additional hybridization tends to affect the original hopping integral between $d$-orbital so that the effective $U/t_d$ will be altered. In fact, even small $V/t_d=0.1$ strongly suppresses $\lambda_d(T)$ at high temperatures while this suppression is gradually diminished at lower $T$. Interestingly, at high $T$, smaller $V/t_d=0.1$ has stronger effects than larger $V/t_d=0.6$ for both $n=0.9, 0.85$. Below a crossover temperature scale $T/t_d \sim 0.2$ this difference is reversed. Although our focus here is the low temperature superconducting instability, we remark that this stronger suppression due to small $V$ above $T/t_d \sim 0.2$, which might be closely related to the promotion of $s$ orbital on SC that will be discussed later, might be worthwhile for further investigation.
Fig.~\ref{lambda}(b) compares the dependence of $T_c$ on the global doping level for the conventional Hubbard model ($V=0$) and $d$-$s$ model at two typical $V$'s. Apparently, the general feature is that the larger $V$ induces lower $T_c$ for all dopings. Nevertheless, this suppression is stronger at low dopings and it effectively diminishes at high dopings. The impact of $V$ on the variation of $T_c(\delta)$ is closely related to the evolution of $d$ and $s$ orbital's occupancies.
As displayed in Fig.~\ref{lambda}(c), we start from the undoped system ($\delta=0$) with self-doped $d$ orbital because of finite but tiny $n_s \sim 0.08$ to be consistent with the experiments~\cite{Hall1}. At small hybridization $V/t_d=0.1$, the doped hole mainly resides on $s$ orbital as evidenced by the rapid decrease of the dashed orange curve with $\delta$. Recall that experimentally the superconductivity dome occurs in the range of $0.12<\delta<0.25$. Hence, in our relevant doping regime, namely $\delta=0.15, 0.2$, the $s$ orbital's concentration is as tiny as only $\sim 0.01$. In this sense, as pointed out by previous work~\cite{TPD_ds,Held2020,Held2022}, the $d$-$s$ model effectively reduces to the single $d$ orbital system. We numerically confirmed this single-band picture in the doping range relevant to the realistic materials. In this case, the decrease of $T_c$ shown in Fig.~\ref{lambda}(b) originates from the decrease of $d$ orbital occupancy.
More interestingly, at stronger $V/t_d=0.6$, both decrease of $n_d$ and $n_s$ with hole doping is smoother than the case of $V/t_d=0.1$. In fact, the doped hole dominantly goes onto $d$ orbital linearly instead, which implies that the less depleted $s$ orbital may play an important role. From the viewpoint of the single-orbital Hubbard model, larger hole doping would rapidly suppress $T_c$. Thus, in the regime of $\delta=0.15, 0.2$, it is striking to see that $T_c$ at $V/t_d=0.6$ conversely approaches to the curves for $V/t_d=0.1$ in Fig.~\ref{lambda}(b) in spite of the increasing reduce of $n_d$ for $V/t_d=0.6$ compared to $V/t_d=0.1$. The only reason originates from the plausible ``enhancing'' effect of the small but finite $n_s$ on inhibiting the suppression of SC. This argument also matches with the elevated $\lambda_d(T)$ for $V/t_d=0.6$ at high temperatures, namely the $s$ orbital's concentration tends to maintain $d$ orbital's SC. This supportive role will be more clear when we discuss the $U$-dependence of $T_c$ in Fig.~\ref{TcU}.
In fact, the $T_c(\delta)$ for three $V$'s seemingly approaches with one another with further hole doping. Unfortunately, due to the limitation of the CT-QMC sign problem, we are not able to push to even higher doping levels and also much larger DCA cluster size to confirm this numerical phenomena to be truly physical.
Putting another way, assuming that the electronic structure calculation~\cite{dft25} leading to $V/t_d \sim 0.6$ is more realistic than the more common belief of the smaller hybridization strength $V/t_d \sim 0.1$, the presence of the electronic density in rare-earth layer can play the vital role for supporting the superconductivity. This constitutes one of our major findings in terms of the constructive role of $d$-$s$ hybridization, which implies that a minimal model accounting for the infinite-layer nickelate SC might need consider the role of the rare-earth layer appropriately in some circumstances.
It is reasonable to argue against the supportive role of $s$ orbital on SC since the $d$-$s$ model has natural connection to PAM, where the $d$-$s$ Kondo screening has destructive effects on SC. Here we emphasize that $s$'s support on SC discussed above might only apply for the systems with tiny $s$ electron density so that the Kondo screening does not have significant impact. We also note that this extreme limit of PAM has rarely been studied before so that it is requisite to explore it in more details in future.
\begin{figure}
\psfig{figure=fig2.pdf,height=5.3cm,width=8.0cm,angle=0,clip}
\\ \caption{The Matsubara frequency dependence of the leading $d$-wave eigenfunction $\phi_d({\bf K},i\omega_n)/\phi_d({\bf K},i\pi T)$ with ${\bf K}=(\pi,0)$ and the normalized cluster spin susceptibility $2\chi_s(\mathbf{Q},i\omega_m)/[\chi_s(\mathbf{Q},0)+\chi_s(\mathbf{Q},i2\pi T]$ for ${\bf Q}=(\pi,\pi)$. The parameters are $U/t_d=7, N_c=4, T/t_d=0.08$ and $n=0.85$.}
\label{eigenvector}
\end{figure}
Now that Fig.~\ref{lambda}(b) has established the notion that the impact of $V$ on the $d$-wave pairing gradually diminishes in the relatively high doping regime of $\delta=0.15-0.2$, it is natural to further explore the behavior of the pairing interaction and moreover the applicability of the spin-fluctuation mediation in $d$-$s$ model. As mentioned earlier, the structure of pairing interaction $\Gamma^{pp}$ is reflected by the BSE eigenvector $\phi_\alpha({\bf K},i\omega_n)$, which is the normal state analog of the superconducting gap function~\cite{Maier06,Scalapino06,Mi2022}.
Fig.~\ref{eigenvector} illustrates the Matsubara frequency dependence of the normalized $\phi_d({\bf K},i\omega_n)$ with ${\bf K}=(\pi,0)$. In addition, the normalized cluster spin susceptibility $\chi_s(\mathbf{Q},i\omega_m)$ for ${\bf Q}=(\pi,\pi)$, which is calculated via $\sum_{K,K'} \chi_{c}(q,K,K')$ in Eq.~\eqref{e1}, is given for comparison to check the role of spin-fluctuation in mediating the $d$-wave pairing as done in conventional Hubbard model~\cite{Maier06,Scalapino06}.
The decrease of $\phi_d$ versus $i\omega_n$ has a characteristic energy scale, which looks increasing with $V$ of a tiny amount. This indicates that the $d$-wave pairing is weakened slightly at larger $V$, which is consistent with $T_c$'s decrease in Fig.~\ref{lambda}(b).
Regarding the antiferromagnetic spin susceptibility, it has been normalized to coincide with $\phi_d(i\omega_n)$ at $\omega_n=\pi T$. Note that the Matsubara frequency entering $\chi_s$ is the bosonic $\omega_m=2m\pi T$.
It can be seen that the antiferromagnetic spin-fluctuation spectrum is characterized by the same energy scale as the dynamics of $\phi_d(i\omega_n)$ reflecting the pairing interaction for $V/t_d=0.1$. Nevertheless, $\chi_s(i\omega_m)$ almost overlap between two cases of $V/t_d=0.1, 0.6$ and thereby the spin-fluctuation and pairing interaction have tiny discrepancy for $V/t_d=0.6$, which might be related to the non-negligible effects of $s$ orbital discuss earlier.
Fig.~\ref{eigenvector} implies that the effectively attractive interaction mediated by the antiferromagnetic fluctuations and the retardation nature of $d$-wave pairing~\cite{Scalapino06} are not qualitatively modified in the presence of $s$ orbital.
\begin{figure}
\psfig{figure=fig3.pdf,height=5.5cm,width=8.5cm,angle=0,clip}
\caption{$T_c$ versus $U$ for $n=0.85, N_c=4$. At stronger $V/t_d=0.6$, the ``optimal'' $U$ for highest $T_c$ is shifted to a higher value.}
\label{TcU}
\end{figure}
To further characterize the SC properties of $d$-$s$ model, Fig.~\ref{TcU} shows the evolution of $T_c$ versus $U$ for different hybridization $V$ at a fixed total density $n=0.85$. Apparently, the small $V$ induces a constant downshift of $T_c$ for all $U$'s, indicating that the extremely tiny concentration of $s$ orbital (shown in Fig.~\ref{lambda}(c)) do not qualitatively modify the superconducting instability. In particular, the existence of an ``optimal'' $U/t_d\sim 6$ of supporting the highest $T_c$ occurs for both single-orbital Hubbard model $V=0$ and $d$-$s$ model at $V/t_d=0.1$. This peak structure is consistent with the notion that it is important to have strong short-range antiferromagnetic correlations, which, however, decreases when $U$ becomes sufficiently large compared to the bandwidth. At strong interaction $U$, the exchange interaction $J \sim 4t^2/U$ decays so that $T_c$ decreases accordingly.
Distinct from the weak hybridization, at relatively strong $V/t_d=0.6$, the alleviated QMC sign problem allows us to simulate the stronger $U/t_d=10$ so that the similar peak structure of $T_c(U)$ can be seen, namely the ``optimal'' $U$ is shifted to a higher value compared to the case of $V/t_d=0.1$. Although we are not able to numerically confirm the trend at even higher $U$ for small $V$, Fig.~\ref{TcU} implies that the additional $s$ orbital and its relatively strong hybridization $V$ with $d$ orbital can have potentially supportive role for an elevated $T_c$ at strong interaction, which further backup our previous arguments discussed for Fig.~\ref{lambda}. For instance, at $U/t_d=9$, $T_c(V/t_d=0.6)>T_c(V/t_d=0.1)$. Unfortunately, it is quite challenging, if not impossible, to determine whether $T_c$ can even exceed the Hubbard model at larger $U$ with the aid of the additional $s$ orbital with appropriate hybridization $V$.
\section{Discussion}
Here we rephrase and remark on some insights on the superconducting properties of the infinite-layer nickelates and more generally other relevant correlated materials.
\begin{description}
\item [$T_c$'s magnitude:]
Fig.~\ref{lambda}(b) reveals the pair-breaking effects of $d$-$s$ hybridization, especially at relatively strong $V/t_d=0.6$. Considering that the cuprate superconductors' optimal hole doping occurs around $\delta \sim 12.5\%$ while the infinite-layer nickelates at higher $\delta \sim 18\%$, $T_c$ has a significant drop regardless of the hybridization strength. This observation is generally consistent with the experimental fact that $T_c$ of infinite-layer nickelates is much lower than the cuprates.
\item [Single-orbital picture:]
The applicability of the single-orbital description depends on the hybridization $V$ strength. Fig.~\ref{lambda} indicates that if $V$ is indeed sufficiently small, it only results in a negligible decrease of $T_c$ corresponding to the tiny $s$ orbital concentration. \\
Nevertheless, if the interlayer hybridization between NiO$_2$ and R layers turns out not to be negligible, a minimal model of the infinite-layer nickelate SC or other similar compounds might need considering the role of the rare-earth layer appropriately even though the R layers have only tiny electron densities.
\item [Undoped system:]
At first glance, the curves in Fig.~\ref{lambda}(b) might hint towards a finite $T_c$ of undoped systems. This obvious discrepancy with the realistic observations probably arises from the small DCA cluster size. Previous expensive DCA calculations with larger $N_c$ but with smaller $U$ has uncovered the dome-shaped $T_c$ versus hole doping~\cite{Jarrell2013}. In this sense, our study here is simply proof-of-principle in terms of the effects of the hybridization $V$.
\item [$V$'s promotion of $T_c$:]
As discussed in Fig.~\ref{lambda} and Fig.~\ref{TcU}, it is striking to observe that the tiny but finite $s$ orbital concentration has promotion effects on SC so that the hybridization strength between NiO$_2$ and rare-earth layers might require more attention, for instance, whether the electronic structure calculation~\cite{dft25} is more realistic.
Even though this picture would be ultimately proved to be irrelevant to the infinite-layer nickelates, it might provide some insights on other strongly correlated compounds.
\item [Optimal $U$:]
Our numerical calculations uncovered some interesting trend of elevated $T_c$ due to relatively large hybridization $V$. Together with the previous point, it might provide some hints for the material design for compounds (not only infinite-layer nickelates) with higher $T_c$.
\item [Heavy fermion physics:]
Despite that our current focus is only the superconducting properties at large hole dopings, the potential heavy fermion physics at or close to the undoped systems deserves further examination. The $d$-$s$ model has natural relation to the Anderson lattice model except for the extremely low conduction electron density. It is worthwhile exploring the potential Kondo screening exhaustion~\cite{exhaustion2019} and its impact on the magnetic properties of the undoped infinite-layer nickelates.
\item [Pairing symmetry:]
Our present work follows the established framework of $d$-wave symmetries based pairing mechanism~\cite{TPD_ds,Held2022,Held2020,Millis2022,Xianxin2020,Adhikary2020,Matianxing2020}. However, the most recent London penetration measurement strongly challenges this scenario by revealing the predominantly nodeless pairing~\cite{nodeless2022}. Specifically, there might exist a 2D-to-3D superconducting states crossover possibly due to the coupling between NiO$_2$ plane and rare-earth spacer layer. This similarity with the iron-based superconductor with nodeless multiband SC is reminiscent of the various studies on the bilayer Hubbard model at strong inter-layer hybridization, where the nodeless $s^{\pm}$-wave pairing is appreciated~\cite{bilayer2011,bilayer2016,bilayer2021}. In this regard, therefore, the $d$-$s$ model as the minimal model seems problematic at least in the regime of relatively small inter-orbital hybridization $V$. This direction may attractive further attention and exploration.
\end{description}
\section{Summary}
In conclusion, we investigated a two-orbital $d$-$s$ model to mimic the interplay between Ni-$d_{x^2-y^2}$ and the effective interstitial s-like orbital located at the missing apical Oxygen above and below Ni in the rare-earth layers to account for the rare-earth 5d orbitals. To accurately approach to the superconducting $T_c$, we employed the dynamic cluster quantum Monte Carlo focussing on the impact of the $d$-$s$ hybridization.
It is found that the $d$-$s$ hybridization suppresses the $d$-wave pairing regardless of the doping level. The suppression of $T_c$ is stronger at low dopings and diminishes with further dopings, which is closely related to the evolution of $d$ and $s$ orbital occupancies. In particular, although the doped hole mainly resides on $s$ orbital at small hybridization $V/t_d=0.1$ so that the $d$-$s$ model effectively reduces to the single $d$ orbital system; at stronger $V/t_d=0.6$, $d$ orbital has considerable concentration even at high doping levels, whose role is manifested by the weakened suppression trend of $V$ on SC. In other words, $T_c$ at $V/t_d=0.6$ approaches to that at $V/t_d=0.1$ at high dopings. It is thus plausible to conjecture about a striking promotion effect of the small but finite $s$ orbital occupancy on SC.
To explore how to promote $T_c$, we further study the impact of Hubbard $U$. Although the weak hybridization $V$ does not qualitatively modify the superconducting properties without $s$ orbital, the relatively large $V/t_d=0.6$ induces that the ``optimal'' $U$ is shifted to a higher value compared to the case of $V/t_d=0.1$. This might imply a potential strategy to enhance the SC instability in general strongly correlated systems including the infinite-layer nickelates via appropriate hybridization with additional conduction electrons.
In addition, in spite of the additional $s$ orbital, the comparison between the Matsubara frequency dependence of the $d$-wave eigenvector (reflecting the pairing interaction) and the antiferromagnetic spin susceptibility supports the spin fluctuation mediation in $d$-wave pairing of $d$ orbital.
Our presented work provides complemental insights on the quantitative effects of the hybridization between NiO$_2$ and rare-earth layers on the superconducting instability relevant to the infinite-layer nickelate superconductors, especially the potentially enhancing effects of rare-earth layers on SC. As a final remark, we believe that the necessity of including the rare-earth layer in a minimal model of infinite-layer nickelates and other relevant strongly correlated compounds can be crucial in the presence of a considerable interlayer hybridization.
\section*{Acknowledgments}
We acknowledge fruitful discussion with Zhicheng Zhong and Guangming Zhang. This work was supported by National Natural Science Foundation of China (NSFC) Grant No. 12174278, the startup fund from Soochow University, and Priority Academic Program Development (PAPD) of Jiangsu Higher Education Institutions.
|
\section{Introduction}
Quantum backflow (QB) is a phenomenon in which the probability density of a quantum mechanical particle may flow in a direction opposite to the momentum \cite{allcock1,Bracken_1994}. One of the most interesting results in this context is that there is a limit on the maximum amount of backflow. It essentially means that although backflow may increase with time, the increase is bounded by a dimensionless number $c_{bm}\approx 0.04$ \cite{Bracken_1994}. Later, improved estimates of the number $c_{bm}$ were found \cite{eveson,penz2005}. QB has been studied in the context of various nonrelativistic systems e.g, optical systems \cite{berry10,Eliezer:20}, elementary quantum systems \cite{yearsley2012}, particle in a constant magnetic field \cite{strange}, correlated quantum states \cite{PhysRevResearch.2.033206}, many-particle systems \cite{Barbier2020:aa}, systems with a scattering potential \cite{PhysRevA.96.012112,cadamuro2017quantum}, noncommutative plane \cite{PhysRevA.102.062218}, open quantum systems \cite{Mousavi2020}, dissipative systems \cite{mousavi:2020aa}. Interestingly, quantum backflow has also been found to be related to nonclassicality \cite{doi:10.1142/S0219749916500325}, superoscillations \cite{muga2000}, decay of quasi stable systems \cite{PhysRevA.100.052101} and classically forbidden probability flow \cite{PhysRevA.99.043626}, transmitting defects \cite{vasconcelos2020quantum}, negative flow of probability \cite{doi:10.1119/10.0000856} etc. Although by nature relativistic quantum systems are more complicated than nonrelativistic ones, nevertheless, the quantum backflow problem has also been studied within the context of relativistic quantum mechanics \cite{Melloy1998:aa,su2018quantum,Ashfaque:2019aa,das2021relativistic}.
In the case of relativistic quantum mechanics it has been shown that, for a free Dirac particle, the probability that flows in a direction opposite to the momentum in a time interval of length $T$ depends on a dimensionless parameter $\epsilon$ which is a function of mass $m$, velocity of light $c$, Planck's constant and $T$ \cite{Melloy1998:aa}. Later, backflow eigenfunctions were determined numerically over a range of values of $\epsilon$ and an explicit expression for relativistic backflow was found \cite{Ashfaque:2019aa}. Also, backflow of a free Dirac particle in a superposition consisting of various energies and helicities \cite{su2018quantum} and in electron wave packets \cite{das2021relativistic} have been studied.
Although there have been formulations of quantum backflow more amenable to experimental verification \cite{Miller2021experimentfriendly,barbier2021experimentfriendly}, this effect whether in the nonrelativistic domain or in the relativistic domain is an as yet unobserved phenomenon. The main reason for this is that the net backflow through a point over a finite time interval is a very small quantity \cite{Melloy1998:aa}. However, it has been shown that, when a nonrelativistic particle is constrained to move on a ring, the chances of observing quantum backflow increases many fold as compared to when the particle moves on a line \cite{Goussev2020:ab}. In the present paper our aim is to apply this strategy to a relativistic quantum mechanical system. More precisely, we shall consider a Dirac fermion on a ring of radius $R$ in the presence as well as in the absence of a constant magnetic field $B$ and for this system we shall compute the quantum backflow as a function of the aforementioned parameter $\epsilon$ and $B$. Finally, we shall also compare our results in the $c\rightarrow \infty$ limit with those of the nonrelativistic backflow problem.
The rest of the paper is organised as follows: in Sec.~\ref{Sec:problemsetup} we set up the problem of the relativistic fermion on a circular ring of radius $R$;
Sec.~\ref{Sec:QBF} discusses the quantum backflow of the relativistic fermion;
Sec.~\ref{Sec:numerical} discusses the full numerical analysis of the corresponding homogeneous Fredholm integral equation to which the quantum backflow problem on the ring is reduced. Finally Sec.~\ref{Sec:conclusions} presents the discussion and conclusions.
\section{Problem setup}
\label{Sec:problemsetup}
The Hamiltonian of a charged Dirac fermion constrained on a plane (a two dimensional system) reads:
\begin{equation}
\label{hamiltonian0}
H= c \,\boldmath{\sigma}\cdot{\boldmath{P}} + mc^2\sigma^3,
\end{equation}
where $\boldmath{\sigma}= (\sigma^1,\sigma^2)$, and $\sigma^i$, ($i=1,2,3$) are the Pauli matrices and the coupling to the electromagnetic field is introduced via the minimal prescription:
\begin{equation}
\boldmath{P} = \boldmath{p} -\frac{q}{c} \boldmath{A}\, ,
\end{equation}
where $\boldmath{p}=-i\hbar\,(\partial_x,\partial_y)$ and $\boldmath{A}$ is the vector potential.
In this work we will consider the dynamics of a charged fermion in a constant magnetic field directed along the $z$-axis, $\boldmath{B}= B\boldmath{e}_z$, perpendicular to the $(x,y)$ plane, which we describe in the so-called symmetric gauge with the following potential:
\begin{equation}
\boldmath{A} = \left(-\frac{By}{2}, \frac{Bx}{2}\right)\,.
\end{equation}
The Hamiltonian in Eq.~\eqref{hamiltonian0} becomes:
\begin{equation}
H = \left( \begin{array}{cc}
mc^2& cP_-\\
cP_+& -mc^2
\end{array}\right),
\end{equation}
where we have defined:
\begin{equation}
P_{\pm}=P_x \pm iP_y \, .
\label{ladderP}
\end{equation}
The operators $P_\pm$ are conveniently written in polar coordinates:
\begin{equation}
P_\pm = - i\, e^{\pm i \varphi}\, \left[\left( \hbar\, \partial_r \pm i\, \frac{\hbar}{r}\,\partial_\varphi\right) \pm \frac{qBr}{2 c}\right].
\end{equation}.
In the following we will be interested specifically in the quantum backflow of the Dirac fermion on the plane when it is constrained to move on a circle of radius $R$. The circle is assumed to be centered at the origin $O=(0,0)$ of the $(x,y)$ coordinate system. Therefore, imposing the constraint $r=R$, the operators $P_\pm$ in Eq.~\eqref{ladderP} are given by:
\begin{equation}
c\,P_\pm = - i\, \frac{\hbar c}{R} e^{\pm i \varphi}\, \left( \pm i\, \partial_\varphi \pm \frac{qB R^2}{2\hbar c}\right) \, .
\end{equation}
We now introduce the dimensionless quantity
\begin{equation}
\label{betafactor}
\beta= \frac{qBR^2}{2 \hbar c}.
\end{equation}
This gives the flux $\Phi$ of the magnetic field across the circle of radius $R$, $\Phi=\pi B R^2$, in units of the flux-quantum $\Phi_0 =hc/e$ (assuming the charge $q$ to be measured in units of $e$). The $P_\pm$ are then written as:
\begin{equation}
c\,P_\pm = \pm \, \frac{\hbar c}{R} e^{\pm i \varphi}\, \left( \partial_\varphi -i \beta \right) \, ,
\end{equation}
so that the Hamiltonian in Eq.~\eqref{hamiltonian0} becomes:
\begin{equation}
\label{hamiltonian1D}
H = \left(\begin{array}{cc}
mc^2& -\frac{\hbar c}{R}e^{-i\varphi}\left( \partial_\varphi -i\beta\right)\\
\frac{\hbar c}{R}e^{+i\varphi} \left(\partial_\varphi -i\beta\right)& -mc^2
\end{array}\right).
\end{equation}
The time dependent Dirac equation with the 2-dimensional Hamiltonian in Eq.~\eqref{hamiltonian0}:
\begin{equation}
\label{time-dep-Dirac}
i\hbar\frac{\partial \psi}{\partial t} = H\psi,
\end{equation}
admits a continuity equation which,
for a generic two-component spinor state $\psi=(\chi_1,\chi_2)^T$, can be written as:
\begin{equation}
\label{continuity}
\frac{\partial \rho}{\partial t} + \boldmath{\nabla} \cdot \boldmath{J}
= 0,
\end{equation}
where the probability density is given by:
\begin{equation}
\label{density}
\rho = \psi^\dagger \, \psi = \chi_1^* \chi_1^{\phantom{*}} + \chi_2^* \chi_2^{\phantom{*}},
\end{equation}
and the current $\boldmath{J}$ is
\begin{equation}
\label{2dimcurrent}
\boldmath{J} = c\, \psi^\dagger \boldmath{\sigma}\, \psi\,.
\end{equation}
The explicit expression of the current in Eq.~\eqref{2dimcurrent} is the usual $U(1)$ current of the Dirac equation minimally coupled to the electromagnetic field. The two-dimensional time dependent Dirac equation in Eq.~\eqref{time-dep-Dirac} is written in manifestly covariant form by introducing a set of two-dimensional gamma matrices $\gamma^\mu$ as $\gamma^0=\sigma^3$, $\gamma^i=\sigma^3 \sigma^i$, ($i=1,2$) which satisfy the Clifford algebra. Then, a straightforward application of Noether's theorem to the global phase symmetry provides the well known $U(1)$ current $j^\mu=\bar{\psi} \gamma^\mu \psi$ with $\bar{\psi} = \psi^\dagger \gamma^0$ which is conserved $\partial_\mu j^\mu=0$. This last condition, when written out explicitly
in terms of its time and space components, reproduces exactly Eqs.~(\ref{continuity},\ref{density},\ref{2dimcurrent}). We have also checked that the same result is obtained by combining the time dependent equations with their adjoints using the explicit form of the Hamiltonian in Eq.~\eqref{hamiltonian1D}.
We
now proceed to compute the explicit form of the continuity equation and of the current for the Dirac fermion on the ring of radius $R$. This is simply done by noting that, in polar coordinates, the gradient operator is $\boldmath{\nabla} = \boldmath{e}_r \partial_r + ({\boldmath{e}_\varphi}/{r})\,\partial_\varphi$. Then the constraint $r=R$ implies that the gradient (for the system constrained on the ring) has no radial component. Thus Eq.~\eqref{continuity} becomes:
\begin{equation}
\frac{\partial\rho}{\partial t} + \frac{c}{R}\, \partial_\varphi\, {\psi}^\dagger \boldmath{e}_\varphi \cdot\boldmath{\sigma}\, \psi =0
\end{equation}
and writing the unit vectors along the $x$ and $y$ axes as linear combination of the radial and longitudinal ones, ($\boldmath{e}_r$, $\boldmath{e}_\varphi$):
\begin{subequations}
\label{versor}
\begin{align}
\label{versora}
\boldmath{e}_x = \boldmath{e}_r\cos\varphi -\boldmath{e}_\varphi \sin\varphi\\
\label{versorb}
\boldmath{e}_y = \boldmath{e}_r\sin\varphi +\boldmath{e}_\varphi \cos\varphi
\end{align}
\end{subequations}
we obtain:
\begin{equation}
\boldmath{e}_\varphi \cdot\boldmath{\sigma} = \sigma^1\,(-\sin\varphi)+ \sigma^2 \cos\varphi=
\left(
\begin{array}{cc}
0 &-i\, e^{-i\varphi}\cr i\,e^{i\varphi}&0
\end{array}
\right)
\end{equation}
and
\begin{equation}
\frac{\partial\rho}{\partial t} + \partial_\varphi J_\varphi=0
\end{equation}
where
\begin{equation}
\label{Eq:current}
J_\varphi=\frac{c}{R} \left[ \chi_1^* (-i) e^{-i\phi} \chi_2^{\phantom{*}} +\chi_2^* (+i) e^{i\phi} \chi_1^{\phantom{*}}\right].
\end{equation}
The spectrum and corresponding eigenfunctions can be easily obtained and are given by (positive energy branch):
\begin{subequations}
\label{spectrum}
\begin{align}
\label{spectruma}
E_\ell &={\sqrt{m^2c^4+ \left(\frac{\hbar c}{R}\right)^2 (\ell-\beta) (\ell -\beta +1)}}\nonumber\\ & = \varepsilon_\ell \, m c^2 \qquad\qquad \ell=0,\pm 1,\pm 2 \cdots \\
\label{spectrumb}\varepsilon_\ell &=\sqrt{1+\left(\frac{\lambdabar_C}{R}\right)^2 (\ell-\beta) (\ell-\beta+1)}
\end{align}
\end{subequations}
with $\lambdabar_C=\hbar/(mc)$ the reduced Compton wavelength of the par\-ti\-cle of mass $m$.
The eigenstates of the Hamiltonian are:
\begin{equation}
\label{Eq:eigenstates}
\psi_\ell = A_\ell \left(\begin{array}{c} e^{i\ell \varphi}\\ \frac{\hbar c/R}{E_\ell +mc^2}\,
i \left(\ell-\beta\right) \, e^{i(\ell+1)\varphi}\end{array}
\right)
\end{equation}
with $A_\ell$ a dimensionless normalization constant which, without loss of generality, can be taken to be real. Imposing the normalization condition:
\begin{equation}
\int_0^{2\pi} d\varphi\, \left| \psi_\ell(\varphi)\right|^2 =1 \end{equation}
for the eigenspinor $\psi_\ell$ gives the normalization constant $A_\ell$ as:
\begin{equation}
\label{coefficientsAL}
A_\ell=\frac{1}{\sqrt{2\pi}} \left[ 1+{\left(\frac{\hbar c}{R}\right)^2 \frac{(\ell-\beta)^2}{(E_\ell+mc^2)^2}}\right]^{-1/2}\,.
\end{equation}
It may be pointed out that the eigenstates (\ref{Eq:eigenstates}) are also eigenstates of $J_z=L_z+\frac{\hbar}{2}\sigma_z$ with eigenvalues $(\ell+\frac{1}{2})$. Note that in the nonrelativistic limit ($c\to \infty$)
we have $E_\ell \to mc^2$ and $\lambdabar_{\text{C}}\to 0$ and $A_\ell \to 1/\sqrt{2\pi}$.
On the other end in the ultra-relativistic limit ($ E_\ell \gg mc^2)$ we obtain:
\begin{equation}
\label{Aellmassless}
A_\ell \longrightarrow \frac{1}{\sqrt{2\pi}} \left(1+\frac{\ell-\beta}{\ell-\beta+1}\right)^{-1/2} \,.
\end{equation}
Each energy level $E_\ell$ is doubly degenerate with eigenfunctions $\psi_\ell$ and $\psi_{(-\ell -1)}$.
The eigenstate $\psi_\ell$ in Eq.~\eqref{Eq:eigenstates} has a constant current:
\begin{equation}
\label{Eq:currentJell}
J_\ell = \frac{| A_\ell |^2}{E_\ell +mc^2} \, \frac{2\hbar c^2}{R^2} \, {\left(\ell-\beta\right)} \,.
\end{equation}
Note that, denoting with $ \lceil x \rceil$ the Ceiling function of $x$, if $\ell \ge \lceil \beta \rceil $ then $J_\ell \ge 0$. The nonrelativistic limit of the current in Eq.~\eqref{Eq:currentJell} reduces to
\begin{equation}
J_\ell \to \frac{\hbar}{2\pi m R^2}\, {\left(\ell-\beta\right)}
\end{equation}
which coincides exactly with the nonrelativistic current discussed in \cite{Goussev2020:ab}.
\section{Relativistic Quantum Backflow on a Ring}
\label{Sec:QBF}
In this section we discuss the quantum backflow of the relativistic fermion constrained on the ring of radius R. We follow closely the approach discussed in \cite{Goussev2020:ab} for the nonrelativistic particle highlighting differences and similarities.
We consider a generic physical state given by a linear combination of the energy eigenstates $\psi_\ell(\varphi,t)$ with $\ell\ge \lceil \beta \rceil$:
\begin{equation}
\label{Eq:state}
\Psi(\varphi, t) = \sum_{\ell=\lceil \beta \rceil}^{+\infty}\, c_\ell \, \psi_\ell (\varphi,t)
\end{equation}
which we assume to be normalized, and thus the coefficients $c_\ell$ satisfy:
\begin{equation}
\label{Eq:constraint_cl}
\sum_{\ell=\lceil \beta \rceil}^{+\infty}\, |c_\ell|^2 \, = 1\,.
\end{equation}
The probability current $J_\Psi(\varphi,t)$ corresponding to the state in Eq.~\eqref{Eq:state} through Eq.~\eqref{Eq:current} reads:
\begin{eqnarray}
\label{Eq:current_psi}
J_\Psi(\varphi,t)&=& \frac{\hbar c^2}{R^2}\!\!\! \sum_{\ell,\ell'=\lceil \beta \rceil}^\infty c_\ell^* A_\ell\left[\frac{{{\ell-\beta}}}{E_\ell +mc^2}+\frac{{\ell'-\beta}}{E_{\ell'} +mc^2} \right] A_{\ell'} c_{\ell'}\nonumber\\
&&\times e^{\frac{i}{\hbar} (E_\ell -E_{\ell'})t} e^{-i(\ell -\ell')\varphi} \,.
\end{eqnarray}
The integrated probability flowing through a fixed point (say $\varphi=0$) in a time window $[-T/2,T/2]$ is :
\begin{equation}
P_\Psi = \int_{-\frac{T}{2}}^{+\frac{T}{2}} \, dt\, J_\Psi(0,t)\, ,
\end{equation}
and
for the state in Eq.~\eqref{Eq:state} given the time dependence of the current in Eq.~\eqref{Eq:current_psi} we just need the time integral:
\begin{subequations}
\begin{align}
\int_{-\frac{T}{2}}^{+\frac{T}{2}} e^{\frac{i}{\hbar}\left(E_{\ell}-E_{\ell^{\prime}}\right) t} d t &= \left\{\begin{array}{ll}
T & \ell=\ell^{\prime} \\
\hbar \left. \displaystyle\frac{e^{{\frac{i}{\hbar}}\left(E_{\ell}-E_{\ell^\prime}\right) t}}{i\left(E_{\ell}-E_{\ell^\prime}\right)}\right|_{-\frac{T}{2}}^{+\frac{T}{2}}
& \ell \ne \ell^{\prime}
\end{array}\right. \nonumber\\
&= T\, \text{sinc} \left[\frac{ T\, \left(E_\ell -E_{\ell^\prime}\right)}{2 \hbar}\right] \nonumber
\end{align}
\end{subequations}
so that we finally obtain $P_\Psi$ in the form:
\begin{equation}
\label{Eq:P_psi2}
P_\Psi = \sum_{\ell,\ell' =\lceil \beta \rceil}^{+\infty} c_\ell^*\, K_{\ell,\ell'}\, c_{\ell'}
\end{equation}
with
\begin{eqnarray}
\label{Eq:kernel1}
K_{\ell,\ell'} &=& \frac{\hbar c^2 T}{R^2}\, A_\ell \, \left( \frac{{{\ell-\beta}}}{E_\ell+mc^2}+ \frac{{{\ell'-\beta}}}{E_{\ell'}+mc^2}\right)A_{\ell'}\times\nonumber\\
&&\phantom{xxxx}\text{sinc}\left[\frac{T(E_\ell-E_{\ell'})}{2\hbar}\right] \,.
\end{eqnarray}
Defining the dimensionless quantity:
\begin{equation}
\label{alpha}
\alpha= \frac{\hbar T}{4 m R^2}
\end{equation}
the kernel $K_{\ell,\ell^\prime}$ can be rewritten:
\begin{eqnarray}
\label{Eq:kernel2}
K_{\ell,\ell'} &=& \left(\frac{\alpha}{\pi}\right)\, 4\pi\, A_\ell \, \left( \frac{\ell-\beta}{1+\varepsilon_\ell}+ \frac{\ell'-\beta}{1+\varepsilon_{\ell'}}\right)A_{\ell'}\times\nonumber\\
&&\phantom{xxxx}\text{sinc}\left[
2 \alpha \left(\frac{R}{\lambdabar_{\text{C}}}\right)^2(\varepsilon_\ell -\varepsilon_{\ell^\prime})\right] \,. \end{eqnarray}
We begin the study of the relativistic backflow on the ring by considering a quantum state with only two eigenstates, $\psi_{\ell_1}$ and $\psi_{\ell_2}$ with $0\le \ell_1 < \ell_2$. Thus in Eq.~\eqref{Eq:state} we parametrize the coefficients $c_\ell$ as:
\begin{equation}
c_{\ell}=\left\{\begin{array}{ll}
\cos \frac{\phi}{2} & \quad \ell=\ell_{1} \geq 0 \\
e^{i \gamma} \sin \frac{\phi}{2} & \quad \ell=\ell_{2}>\ell_{1} \\
0 & \quad \ell \ne \ell_1, \ell_2
\end{array}\right.
\label{Eq:parametrization}
\end{equation}
with $0\le \phi \le \pi $ and $0\le \gamma < 2\pi$. This choice of parameters
will ensure that the normalization condition in Eq.~\eqref{Eq:constraint_cl} is satisfied. Inserting Eq.~\eqref{Eq:parametrization} into Eq.~\eqref{Eq:P_psi2} using either Eq.~\eqref{Eq:kernel1} or Eq.~\eqref{Eq:kernel2} we obtain:
\begin{equation}
P_\Psi^{(\ell_1,\ell_2)} = \frac{\alpha}{\pi} \Big[ A -B \cos\phi +C\, \text{sinc}(D) \cos\gamma\sin\phi\Big]
\end{equation}
where the quantities $A,B,C$ and $D$ are given by:
\begin{equation}
\begin{array}{l}
\displaystyle
A= 4\pi\left(A_{\ell_2}^2\,\frac{ \ell_2 -\beta}{1+\varepsilon_{\ell_2}}+ A_{\ell_1}^2 \frac{\ell_1-\beta}{1+\varepsilon_{\ell_1}}\, \right) \\ \displaystyle B= 4\pi\left(A_{\ell_2}^2\,\frac{ \ell_2 -\beta}{1+\varepsilon_{\ell_2}}- A_{\ell_1}^2 \frac{\ell_1-\beta}{1+\varepsilon_{\ell_1}}\,
\right)\\
\displaystyle
C= 4\pi A_{\ell_2}A_{\ell_1}\left(\frac{\ell_2-\beta}{1+\varepsilon_{\ell_2}}+ \frac{\ell_1-\beta}{1+\varepsilon_{\ell_1}}\right)\\
\displaystyle
D= 2\alpha \left(\frac{R}{\lambdabar_{\text{C}}}\right)^2\left(\varepsilon_{\ell_2}-\varepsilon_{\ell_1}\right) \,.
\end{array}
\end{equation}
We then perform a minimisation procedure over the parameters $\phi, \gamma$:
\begin{equation}
{\cal P}^{(\ell_1,\ell_2)} (\alpha,\beta) = \min\nolimits_{\left\{\phi,\gamma\right\}} P^{(\ell_1,\ell_2)}_\Psi (\phi,\gamma)
\end{equation}
and, similarly to what is found in \cite{Goussev2020:ab}, the result is:
\begin{equation}
\label{Pl1l2}
{\cal P}^{(\ell_1,\ell_2)} (\alpha,\beta) = \frac{\alpha}{\pi} \left(
A -\sqrt{B^2 +C^2\, \text{sinc}^2 (D)}
\right)\,.
\end{equation}
Note however that the coefficients $A, B, C$ and $D$ have a different definitions with respect to that of the nonrelativistic problem.
\begin{figure}[h!]
\includegraphics[width=0.95
\linewidth]{FIG1.pdf}
\caption{Plot of the backflow probability ${\cal P}^{(0,1)}$ at $\varphi=0$ for the relativistic fermion on the ring for $\lambdabar_{\text{C}}/R =\lambda_{\text{C}}/(2\pi R)=0.05$. The solid (blue) line is the backflow probability ${\cal P}^{(0,1)}$ corresponding to a vanishing magnetic field ($\beta=0)$. The orange, green and red lines are the probabilities ${\cal P}^{(0,1)}$ for the relativistic problem corresponding respectively to $\beta= -0.025,-0.05,-0.075$. }
\label{fig:FIG1}
\end{figure}
In Fig.~\ref{fig:FIG1} we show the backflow probability ${\cal P}^{(0,1)}$ for different values of the parameter $\beta$ and for $\lambdabar_{\text{C}}/R=0.05$.
Fig.~\ref{fig:FIG1} shows clearly that the probability can indeed become negative. We see that the lowest value, $\text{min}_{(\alpha,\beta)} {\cal P}^{(0,1)} \approx -0.0508602 $ is attained for $\beta=0$ at $\alpha/\pi\approx 0.195067$.
\subsection{ General Formulation}
Here we shall consider a more general superposition of eigenstates, and study the corresponding problem of maximizing the integrated probability flux through $\varphi=0$, given in Eq.~\eqref{Eq:P_psi2},
i.e. maximize the quantum backflow considering a generic state as in Eq.~\eqref{Eq:state} subject to the normalization constraint Eq.~\eqref{Eq:constraint_cl}.
A comment is in order here about a symmetry of our system. Indeed it can easily be checked that the eigenvalues, c.f. Eq.~\eqref{spectruma}, satisfy the following relation:
\begin{equation}
E_\ell^{(\beta)} = E_{\ell-1}^{(\beta-1)}
\end{equation} and the same is true for the normalization constants, Eq.~\eqref{coefficientsAL}. This in turn implies that the quantity ${\cal P}_\psi$ as given in Eq.~\eqref{Eq:P_psi2} is invariant under the transformation $\beta \to \beta-1$ and $c_\ell \to c_{\ell+1} $ (or equivalently $\beta \to \beta+1$ and $c_\ell \to c_{\ell-1} $). Therefore, it will be enough to consider $\beta $ in the interval:
\begin{equation}
\beta \in (-1, 0] \qquad \text{with} \qquad \lceil\beta\rceil=0\,.
\end{equation}
Hereafter we shall therefore assume $\lceil\beta\rceil=0$.
As also discussed in \cite{Goussev2020:ab} the maximal backflow can be found defining a real-valued functional of the set of coefficients $\left\{ c_\ell \right\}$ by:
\begin{equation}
I\left[c_{\ell}\right]=\sum_{\ell, \ell'=0}^{\infty} c_{\ell}^{*} K_{\ell \ell'} c_{\ell'}-\lambda \sum_{\ell=0}^{\infty} c_{\ell}^{*} c_{\ell}
\end{equation}
and performing an unconstrained minimization
where $\lambda$ plays the role of a Lagrange multiplier. The variation of the functional $I[\left\{ c_\ell \right\}]$ with respect to independent variations $\delta c_\ell$ and $\delta c_\ell^*$ produces the corresponding eigenvalue problem (Euler-Lagrange equations):
\begin{equation}
\label{Eq:eigenvalue}
\sum_{\ell'=0}^{\infty} K_{\ell \ell'} c_{\ell'}=\lambda\, c_{\ell}
\end{equation}
so that the maximal amount of quantum backflow is attained when the value of $P_\Psi$, Eq.~\eqref{Eq:P_psi2}, reaches its infimum and this coincides with the the lowest eigenvalue of the spectrum \{$\lambda$\} of the unconstrained variational problem:
\begin{equation}
\mathcal{P}(\alpha, \beta,\frac{\lambdabar_{\text{C}}}{R}) \equiv \inf _{\Psi} P_{\Psi}=\inf \{\lambda\}
\end{equation}
In the next subsection we will discuss the numerical solution of the eigenvalue equation in Eq.~\eqref{Eq:eigenvalue}, then in the next section we will address the line limit of backflow on a ring which is obviously reached when $R\to \infty$.
\begin{figure*}[t]
\includegraphics[width=0.475
\linewidth]{FIG2a.pdf}\hspace{0.3cm}
\includegraphics[width=0.475
\linewidth]{FIG2b.pdf}
{\includegraphics[width=0.475\linewidth]{FIG2c.pdf}}\hspace{0.3cm}
\includegraphics[width=0.475
\linewidth]{FIG2d.pdf}
\caption{$\mathcal{P}(\alpha,\beta)$ for a fermion of mass $m$, on the ring of radius $R$ for various values of the parameter $\beta$ and for different values of $\lambdabar_{\text{C}}/R=\lambda_{\text{C}}/(2\pi R)$: top left panel (a) $\lambdabar_{\text{C}}/R=0.05$, top right panel (b) $\lambdabar_{\text{C}}/R=0.725$, bottom left panel (c) $\lambdabar_{\text{C}}/R=1$ , bottom right panel (d) $\lambdabar_{\text{C}}/R=2$. We observe a quite different structure depending on the value of the parameter $\lambdabar_{\text{C}}/R$. }
\label{fig:FIG2}
\end{figure*}
\begin{figure}[htb!]
\includegraphics[width=0.95
\linewidth]{FIG3.pdf}
\caption{\label{FIG:3} (Color online) Numerical results of the minimum eigenavule of the Kernel as a function of the parameter $\lambdabar_{\text{C}}/R=\lambdabar_{\text{C}}/(2\pi R)$. We can appreciate an important structure with respect to this parameter. We observe a maximal amount of backflow for $\lambdabar_{\text{C}}/R=0.730$. The dashed line (orange) represent the preliminary result for the limiting value of the backflow as $\lambdabar_{\text{C}}/R\to \infty$ as we have checked values from 20 up 100.000 that are shown in Table~\ref{Table:I}. }
\end{figure}
\subsection{Numerical analysis}
\label{Sec:numerical}
Following the procedure described in \cite{Goussev2020:ab}, we truncate the sum in Eq.~\eqref{Eq:eigenvalue} at a large value $\ell'=N$, and find the minimum eigenvalue $\lambda^{(N)}_{\text{min}}$ of this finite-dimensional problem. By repeating this for a sequence of increasing $N$ we obtain a numerical estimate for $\mathcal{P}(\alpha, \beta)=\lim_{N\to\infty}\lambda^{(N)}_{\text{min}}$ by plotting the data as a function of $1/N$ with a quadratic polynomial and extracting the value at $1/N=0$. The choice of a quadratic polynomial is justified by the values of the sum of squared residuals, with orders generally less or equal than $10^{-10}$. $\mathcal{P}(\alpha, \beta)$ is also a function of the parameter $\lambdabar_{\text{C}}/R$, which is shown in Fig.~\ref{fig:FIG2}, by plotting it as a function of $\alpha$ for different values of $\beta$ and $\lambdabar_{\text{C}}/R$. A detailed numerical analysis reveals that the absolute maximal amount of backflow is found for $\lambdabar_{\text{C}}/R=0.730$:
\begin{equation}
\label{Eq:mainresult}
\left.\inf_{\alpha,\beta}\mathcal{P}\right|_{\lambdabar_{\text{C}}/R = 0.730}=-c_{\text{Dring}},\qquad c_{\text{Dring}}\simeq0.091999
\end{equation}
with the minimum attained for $\beta=0$ and $\alpha/\pi\simeq0.36252$.
\begin{table*}[hbt!]
\caption{\label{Table:I}Numerical estimate of the quantum backflow for large values of the parameter $\lambdabar_{\text{C}}/R$. This corresponds to estimating the QBF for a massless particle since $\lambdabar_{\text{C}}/R \to \infty$ is equivalent to the massless limit ($m\to 0$).}
\begin{ruledtabular}
\begin{tabular}{cccccc}
$\frac{\lambdabar_{\text{C}}}{R} $ & 20 &500&1000&10000&100000 \\
$\text{inf}_{\psi} {\cal P}_\psi$ & -0.074726&-0.074061& -0.074047& -0.074034&-0.074034
\end{tabular}
\end{ruledtabular}
\end{table*}
This was obtained by evaluating the infimum separately for different values of $\lambdabar_{\text{C}}/R$ which results in Fig.~\ref{FIG:3}.
We have estimated the large $\lambdabar_{\text{C}}/R\to \infty$ limit (which corresponds to the massless fermion case ($m\to 0$). While further investigation of the massless case will be the object of a future work, the preliminary results of these estimates are shown in Table~\ref{Table:I} where we can see that the value of the quantum backflow approaches 0.074 for large $\lambdabar_{\text{C}}/R$. This limiting value is shown in Fig.~\ref{FIG:3} as an horizontal (dashed) line.
Using the complete numerical results of the eigensystem in Eq.~\eqref{Eq:eigenvalue}, i.e. the eigenvectors corresponding to the most negative eigenvalue, we can find a numerical approximation to the current corresponding to the backflow-maximizing state for this system. We set $\alpha$, $\beta$ and $\lambdabar_{\text{C}}/R$ to the values corresponding to $\mathcal{P}\simeq-c_{\text{Dring}}$, truncate the sum in Eq.~(\ref{Eq:eigenvalue}) at $N=2000$, and compute the eigenvectors corresponding to the smallest eigenvalue. This will provide the approximate coefficients $c_\ell$ up to $\ell_{\text{max}}=N=2000$ of the backflow-maximizing state, as defined in Eq.~(\ref{Eq:state}). We can then compute the corresponding probability current, Eq.~(\ref{Eq:current_psi}), at $\varphi=0$ as a function of time. It turns out that it is convenient to put the probability current in the form:
\begin{eqnarray}
J_\Psi(0,t)T&=& 4\alpha\!\!\! \sum_{\ell,\ell'=0}^N c_\ell^* A_\ell\left[ \frac{\ell}{1+\varepsilon_\ell}+ \frac{\ell'}{1+\varepsilon_{\ell'}}\right] A_{\ell'} c_{\ell'}\nonumber\\
&&\times e^{i \left(\sqrt{\left(\frac{R}{\lambdabar_{\text{C}}}\right)^2+\ell(\ell+1)}-\sqrt{\left(\frac{R}{\lambdabar_{\text{C}}}\right)^2+\ell'(\ell'+1)}\right)\frac{4\alpha R}{\lambdabar_{\text{C}}}\frac{t}{T}}\nonumber\vspace{0.2cm}\\
\end{eqnarray}
so that it is written in terms of quantities whose values are known for the backflow-maximizing state. The result is shown in Fig.~\ref{FIG:4}a
where it is clear that in the interval $t\in [-T/2, +T/2]$ the current $J_\Psi (0,t)$ corresponding to the maximizing state is always negative. Fig.~\ref{FIG:4}b shows an enlargement of the neighborhood $t/T\approx-1/2$ obtained computing the current. While form Fig.~\ref{FIG:4}a one might think that at $t=-T/2$ there is a discontinuity the zoom proposed in Fig.~\ref{FIG:4}b shows that in reality the current changes without discontinuity. This behaviour of the current is similar to what has been found in other one dimensional systems on the line \cite{Ashfaque:2019aa} as well as in the nonrelativistic problem of a particle on a ring~\cite{Goussev2020:ab}.
\begin{figure*}[htb]
\includegraphics[scale=0.68]{FIG4a.pdf}
\includegraphics[scale=0.69]{FIG4b.pdf}
\caption{\label{FIG:4} (a) Probability current $J_\Psi(0,t)T$ as a function of time $t$, in units of $1/T$, for the backflow-maximizing state. (b) We show a zoom in the neighborhood of $t/T=-0.5$ where the current changes sign. }
\end{figure*}
\section{Line limit}
We discuss here the line limit of the quantum backflow on the ring. Let us consider the limit $R\to\infty$ of the eigenvalue equation, Eq.~\eqref{Eq:eigenvalue}, with $K_{\ell,\ell'}$ given in Eq.~\eqref{Eq:kernel2}. We note that
\begin{eqnarray}
\varepsilon_\ell =\sqrt{1+\frac{\hbar^2}{m^2c^2R^2} \frac{1}{\alpha}\,\sqrt{\alpha}(\ell -\beta)\sqrt{\alpha}(\ell -\beta+1)
}\,.
\end{eqnarray}
The line limit is reached when $\alpha\to 0$ with $\sqrt{\alpha}\beta \to 0$ so that using $\alpha= \hbar T /(4 m R^2)$ we have
\begin{equation}
\varepsilon_\ell \to \sqrt{1+\frac{4\hbar T^{-1}}{mc^2}\, (\sqrt{\alpha}\ell)^2} \end{equation}
and with the definitions:
\begin{equation}
\label{epsilon}
\epsilon= \sqrt{\frac{4\hbar T^{-1}}{mc^2}}\,, \qquad z=\sqrt{\alpha}\, \ell
\end{equation}
the line limit of the quantity $\varepsilon_\ell$ is:
\begin{equation}
\label{gamma}
\varepsilon_{\ell}\to \sqrt{1+\epsilon^2 z^2} \equiv \gamma(z)\,.
\end{equation}
A little algebra shows, with the same reasoning, that the coefficients $A_\ell$ in eq.~\eqref{coefficientsAL} reduce to:
\begin{equation}
\label{coefficientALz}
A_\ell \to\frac{1}{\sqrt{2\pi}} \sqrt{\frac{\gamma(z)+1}{2\gamma(z)}} \,.
\end{equation}
With this in mind,
the eigenvalue equation, Eq.~\eqref{Eq:eigenvalue}, can be first manipulated by distributing $\sqrt{\alpha}$ in the numerators of the two terms in the round brackets in Eq.~\eqref{Eq:kernel2}. We thus find:
\begin{eqnarray}
&&\frac{1}{\pi}\,\sum_{\ell'=0}^{\infty} \, 4\pi\, A_\ell \, \left[ \frac{\sqrt{\alpha}(\ell-\beta)}{1+\varepsilon_\ell}+ \frac{\sqrt{\alpha}(\ell'-\beta)}{1+\varepsilon_{\ell'}}\right]A_{\ell'}\nonumber\times \\&&\phantom{xxxxxxxxxxx} \text{Sinc}\left[
\frac{2}{\epsilon^2}(\varepsilon_\ell -\varepsilon_{\ell^\prime})\right] c_{\ell'} = \lambda \frac{c_\ell}{\sqrt{\alpha}} \end{eqnarray}
so that taking the limit $\alpha\to 0$ by defining the variables $z=\sqrt{\alpha}\,\ell$, $z'=\sqrt{\alpha}\,\ell'$, the function $f(z)=c_\ell/\sqrt{\alpha}$, using the results in Eqs.~(\ref{epsilon},\ref{gamma},\ref{coefficientALz}) and replacing the discrete sum over $\ell'$ by an integral over the variable $z'$ one finds:
\begin{eqnarray}
&&\frac{1}{\pi}\!\int_{0}^{\infty} \,\!\!\! dz'\, \sqrt{\frac{\gamma(z)+1}{\gamma(z)}} \, \left[ \frac{z}{1+\gamma(z)}+ \frac{z'}{1+\gamma(z')}\right]\sqrt{\frac{\gamma(z')+1}{\gamma(z')}}\nonumber\\&&\phantom{xxx}\times \text{sinc}\left[
\frac{2}{\epsilon^2}(\gamma(z) -\gamma(z'))\right]\, f(z') = \lambda\, f(z) \end{eqnarray}
that can be straightforwardly rearranged into
\begin{eqnarray}
&&\frac{1}{\pi}\!\int_{0}^{\infty} \,\!\!\! dz'\,
\frac{z(\gamma(z')+1)+z'(\gamma(z)+1)}{\sqrt{\gamma(z)(\gamma(z)+1)\,\gamma(z')(\gamma(z')+1)}}
\phantom{xxxxxxxx}\nonumber\\&&\phantom{xxxxx}\times \text{sinc}\left[
\frac{2}{\epsilon^2}(\gamma(z) -\gamma(z'))\right]\, f(z') = \lambda\, f(z)
\label{Eq:massive_fermion_line}
\end{eqnarray}
which coincides exactly with the integral eigenvalue equation for a relativistic spin-half fermion on the line discussed in \cite{Melloy1998:aa,Ashfaque:2019aa} (c.f. see their equation 14).
\section{Discussion and Conclusions}
\label{Sec:conclusions}
In this paper we have presented a study of the quantum backflow of a charged relativistic massive fermion in a plane within a perpendicular, constant magnetic field and constrained to move on a ring of radius $R$. The same problem for a charged nonrelativistic spinless particle has been studied in \cite{Goussev2020:ab}. We use the same approach to the problem presented in \cite{Goussev2020:ab} highlighting differences and similarities with the results of the nonrelativistic analysis.
The main finding reported here is that the maximal amount of backflow for the relativistic Dirac fermion on the ring is obtained when the radius of the ring is of the order of the reduced Compton wavelength of the particle, $\lambdabar_{\text{C}}/R=0.730$ (with $\beta=0$ and $\alpha/\pi\simeq0.36252$) and it amounts to $c_{\text{Dring}}=0.092$. We have then discussed the line limit showing that the eigenvalue problem reduces to the integral equation for the quantum backflow of a relativistic particle on a line previously studied in the literature~\cite{Ashfaque:2019aa}.
We have also computed the probability current flowing through the point $\varphi=0$ as a function of time for the state that maximises the net flow in the time interval [$-T/2,+T/2$], finding, as expected, that it is always negative in this interval.
It is also interesting to compare the result in Eq.~\eqref{Eq:mainresult} with the corresponding result obtained in \cite{Goussev2020:ab} for the nonrelativistic problem on the ring $c_{\text{ring non-rel}} = 0.116 816$. We note that the amount of backflow for the Dirac fermion on a ring is about 20\% smaller than that of a nonrelativistic particle but still larger than the backflow value on the line $c_{\text{line}}=0.0389$.
Finally, we provide preliminary estimates of the amount of backflow in the limit $\lambdabar_{\text{C}}/R \to \infty$ which corresponds to the massless fermion case ($m\to 0)$. While we intend to study the massless case comprehensively in a separate work, these preliminary results in our opinion show that there is definitely quantum backflow for a massless Dirac fermion on a ring.
|
\section{Introduction}
The Kohn-Sham formalism of the density functional theory (KS-DFT)
\cite{Kohn1965,Hohenberg1964} provides an efficient approach for calculating
materials properties. However, the required approximation for the exchange and
correlation (XC) energy component of the effective (known as KS) potential imposes
some limitations. For example, the widely used local density (LDA) and the generalized
gradient (GGA) approximations are known to significantly underestimate the band gaps of semiconductors
and insulators, which is a key quantity for many technological applications. It is worth to
mention that the KS-DFT formalism is a ground-state method, and hence it is not supposed to
provide accurate band gaps due to a derivative discontinuity in the XC potential \cite{Godby1988}.
On the other hand, the GW approach \cite{Rinke2005} provides a formal way for
calculating the band gaps but at a significantly higher computational cost, which
makes GW calculations for large systems prohibitive.
In order to provide an improved description of the band gaps within DFT, several
approaches for approximating the XC potential have been devised, such as semilocal
functionals (see for example Refs. \cite{Borlido2020, Borlido2020a, Yang2016}) and
hybrid-functionals \cite{Becke1993,Heyd2003}. Of particular interest here is the modified
Becke-Johnson \cite{Becke2006} exchange-potential combined with LDA correlation (mBJLDA)
\cite{Tran2009}, which has gained popularity over the past decade due to its
ability to provide band gaps with an accuracy comparable to that of the more
advanced GW and hybrid functionals approaches \cite{Tran2019}, at a much lower computational
costs. The mBJLDA potential has been extensively tested on a wide range of materials,
such as the 40-semiconductor test set (SC40) \cite{Ye2015}, the test set of 76
solids \cite{Tran2017}, and 472 materials \cite{Borlido2019}.
The success of the mBJLDA potential to provide highly improved band gaps is attributed to the
contribution of the kinetic energy density term, which affects valence and conduction electrons
differently, leading to a band gap opening \cite{Koller2011,Abu-Farsakh2021}. Despite the
impressive successes of the mBJLDA potential, it is worth to mention that it does not provide
significantly improved band gaps for some materials such as Cu$_2$O \cite{Koller2011} and
ScF$_3$ \cite{Hamed2015}. Moreover, mBJLDA is found \cite{Abu-Farsakh2021,Kim2010} to
overestimate effective masses, to underestimate bandwidths, and to underestimate binding
energies of semicore states. This is not surprising because the mBJLDA potential is optimized
to reproduce the experimental band gaps \cite{Tran2009}.
The pseudopotential plane-waves (PP-PW's) approach is widely used and has proved
to provide a comparable accuracy to all-electron methods \cite{Borlido2020}. It is
worth noting that the mBJLDA potential was optimized using full-potential linearized
augmented plane-waves (FP-LAPW) method \cite{Tran2009}, which is an all electron
(AE) approach, and hence it is not, in principle, supposed in to provide a good description
of the band gaps when used in conjunction with the pseudopotential approach. In spite of
this fact, it has been implemented in several pseudopotential plane waves codes, such as
\textsc{abinit} \cite{Gonze2020,Romero2020}, Quantum Espresso \cite{Germaneau2013},
and \textsc{castep} \cite{Bartok2019}. Very recently, we have thoroughly investigated
\cite{Abu-Farsakh2021} the performance of the mBJLDA pseudopotential (mBJLDA@PP)
approach, as implemented in \textsc{abinit}, for electronic structure calculations
(band gaps, effective masses, band widths, and binding energies of semicore electrons), using
11 carefully selected systems. We found that the improvement in the
mBJLDA@PP band gaps using typical pseudopotentials is not systematic. For some systems the
so-calculated band gaps are significantly underestimated compared to those of the
all-electron mBJLDA (mBJLDA@AE) ones, while accurate band gaps are obtained for the
other considered systems.
Interestingly, we have shown that the band gaps of the former systems can be highly
improved if the uppermost $p$ states are included as valence in the cation PP's.
The so-obtained band gaps and the other considered electronic
structure properties (see above) are found to be in very good agreement with the
corresponding mBJLDA@AE results. Therefore, in accordance with the mBJLDA@AE approach, only
the band gap is well descried by the mBJLDA@PP method \cite{Abu-Farsakh2021,Kim2010}.
The main objective in this work is to further investigate the effects of the number of
electrons treated as valence (NETV) in the PP's of the cations on the mBJLDA@PP band gaps
using a large set of test systems. The considered set contains 83 solids,
which combines the test set of 76 solids used in Ref. \cite{Tran2017} and the
40-semiconductors (SC40) test set of simple and binary semiconductors used in Refs.
\cite{Heyd2005,Ye2015} (see Table S1 in the Supporting Information). These solids
represent a wide range of semiconductors and insulators. The results obtained using
different NETV will be discussed in comparison with the already available mBJLDA@AE
band gaps \cite{Tran2017, Borlido2019, Ye2015} and experimental data \cite{Abu-Farsakh2021}.
\section{Computational details}
\label{Section:comp_details}
As in our previous work \cite{Abu-Farsakh2021}, the LDA and mBJLDA@PP calculations were
performed using the \textsc{abinit} code \cite{Gonze2020,Romero2020} and optimized
norm-conserving Vanderbilt pseudopotentials \cite{Hamann2013}. The generation of the
PP's is also performed as described in Ref. \cite{Abu-Farsakh2021}. To investigate the
effect of the NETV in the PP's of the cations, we have classified the PP's according to the used
NETV to four main categories:
(i) typical pseudopotentials (t-PP's), in which valence and semicore states are
pseudized. The adopted PP's of the anions are of this type.
(ii) PP's in which the outer core $p$ electrons are additionally treated as valence,
referred to as cp-PP's.
(iii) PP's in which the outer core $s$ and $p$ electrons are included in the
valence, referred to as csp-PP's.
(iv) In this work, we encountered a new situation in which the uppermost core states are of
$d$ character. This is the case for Sr, Mo, and Ba atoms
(see Table S2 in the Supporting Information), which belong to the 5\textsuperscript{th}
and 6\textsuperscript{th} periods of the Periodic Table. For these
elements, this fourth category is therefore considered, in which the upper core $d$ electrons
are treated as valence, hereafter referred to as cd-PP's. It is worth noting that for these
three elements the outer core $d$ electrons are also implicitly included as valence in their
so-called cp- and csp-PP's.
The other computational details are as described in Ref. \cite{Abu-Farsakh2021}. We note
that currently the \textsc{abinit} code does not support PP's with non-linear core
correction (NLCC) for meta-GGA's functionals, and therefore the NLCC was not included.
All the generated PP's were checked to be free of ghost states. Convergence tests with
respect to the cutoff energies for the PW's basis were performed for each system and
for each set of PP's. The used cutoff energies and the Monkhorst-Pack grids
\cite{Monkhorst1976} for Brillouin zone samplings produce total energies which are
converged to better than 1 meV/atom. As it is commonly done in band structure
calculations \cite{Tran2017,Tran2009, Jiang2016, Rinke2005},
all the calculations were performed at the experimental lattice parameters,
which are taken from Refs. \cite{Heyd2005,Tran2017}.
\section{Results and discussion}
The calculated band gaps of the 83 semiconductors and insulators are listed in Table S1
of the Supporting Information. The reported band gaps are obtained using LDA and mBJLDA
functionals, and employing the considered PP's with different NETV (see Sec.
\ref{Section:comp_details}). The corresponding mBJLDA@AE \cite{Tran2017,Tran2019,Ye2015}
and experimental (taken from Refs. \cite{Tran2017,Tran2019,Ye2015}) values are also reported.
We note that the LDA band gaps are insensitive to NETV in the
cations' PP's (see Table S1), and hereafter only those of the LDA@t-PP approach are thus
considered and referred to as LDA band gaps.
The comparison with experimental band gaps deserves some comments. Apart from the
experimental uncertainties (of about 0.1 eV), notable differences in the experimental band
gaps of some compounds are observed. These discrepancies can be attributed to
the employed experimental technique, purity of the sample, and temperature \cite{Strehlow1973}.
On the other hand, the electron-phonon interaction is completely neglected in our and in almost
all other mBJLDA, hybrid-functionals, and $GW$ calculations, and thus the
comparison with experimental band gaps should be taken with caution. In the present work we
adopted the experimental band gaps listed in Refs. \cite{Tran2017,Tran2019,Ye2015}, in
which the mBJLDA@AE are reported.
It is worth noting that we have already shown \cite{Abu-Farsakh2021} that the mBJLDA@PP
band gaps are sensitive to NETV in the cations PP's, which is also evident from Table S1.
Therefore, to present and discuss our results we will classify the considered solids
according to the group of the cation, employing the chemical abstracts service (CAS)
numbering scheme of the Periodic Table. This will help us to
identify trends for systems sharing cations from the same group. To analyze our results we shall focus
on the percentage errors in the calculated mBJLDA@PP band gaps compared to experiment.
Finally, the general features and trends will be presented and discussed.
\begin{figure}[t]
\centering
\includegraphics[width=0.65\linewidth]{figs/errors_IA-IB}
\caption{Percentage error in the calculated mBJLDA@PP band gaps for group IA- and
IB-based systems compared to experiment, using different sets of PP's. For
comparison purposes, we also show those of the LDA@t-PP and reference
mBJLDA@AE band gaps. The points are connected as a guide to the eye.}
\label{fig:errorsIA-IB}
\end{figure}
\subsection{Group IA- and IB-based compounds}
The percentage errors in the calculated mBJLDA@PP band gaps of the considered IA- and
IB-based compounds are depicted in Fig. \ref{fig:errorsIA-IB}, using the different
sets of cation PP's (t-PP, cp-PP, csp-PP). Also shown are the percentage errors in
mBJLDA@AE and LDA band gaps. The considered systems are mainly IA-VIIA and
IB-VIIA binary compounds. We note that for the light cations (namely,
Li and Na) only the t-PP's can be generated.
We will start with the considered group IA-based compounds. The three main features
to note from Fig. \ref{fig:errorsIA-IB} are the following: (i) The well-known
systematic underestimation of the band gaps within the LDA approach.
(ii) The only available mBJLDA@t-PP band gaps of the Li- and Na-based compounds are
systematically larger than the corresponding mBJLDA@AE results and the experimental
band gaps, except for LiCl. However, it should be noted that the overall agreement with
experiment is similar to that of the mBJLDA@AE band gaps.
(iii) For the K-based compounds (KF and KCl), including the uppermost core $2p$ or $2s+2p$
electrons as valence in the K pseudopotential (i.e., using respectively cp- and
csp-PP's for K) has relatively small effects on their mBJLDA@PP band gaps -- the band gaps
of both compounds increase slightly by increasing NETV (see Table S1).
Therefore, one can conclude that the mBJLDA@t-PP approach provides a very good
description of the band gaps of compounds involving group IA elements (at least for
Li-, Na- and K-based compounds).
Fig. \ref{fig:errorsIA-IB} shows that the mBJLDA@PP band gaps of the group
IB-based compounds behave differently from the above IA-based ones. The important
features to note are the following.
First, for the considered Cu- and Ag-based compounds,
the mBJLDA@t-PP band gaps are significantly underestimated, and can be even in a worse
agreement with experiment than the LDA ones.
Secondly, including the upper core $p$ or $s+p$ electrons as part of the valence electrons
in the Cu and Ag PP's improves significantly their mBJLDA@PP band gaps. While the best
agreement with the mBJLDA@AE results is obtained when cp-PP's are used, it can be easily
noted that the best agreement with experiment can be achieved when csp-PP's are used for
both Cu and Ag.
Thirdly, the mBJLDA@PP band gaps for Cu-based systems are still significantly underestimated,
which is the case in both the pseudopotential and the all-electron approaches.
For a quantitative analysis we calculate the mean absolute error (MAE) in the
band gaps calculated by employing different approaches, compared to the experimental
data. The MAE in mBJLDA@PP band gaps when t-PP's are used is of 1.62 eV, and it
drops to 0.76 eV when cp-PP's are used whenever applicable\footnote{That is, when
only t-PP's are possible then for the corresponding systems the t-PP's results
are used. This is done so that the different MAE values always include the same number
of systems.}. The MAE decreases further to 0.66 eV when csp-PP's are used whenever applicable.
This clearly indicates that using csp-PP's for group IB-based compounds leads to the best
agreement with experiment. For comparison, the MAE in the corresponding mBJLDA@AE band gaps
is of 0.67 eV. On the other hand, for the group IB-based compounds the mean absolute
difference between mBJLDA@PP and mBJLDA@AE band gaps is of 0.26 eV (0.08 eV)
when the csp-PP's (cp-PP's) are used, which confirms that the best agreement
with the all-electron mBJLDA band gaps is achieved by employing the cp-PP's.
\begin{figure}[t]
\centering
\includegraphics[width=0.65\linewidth]{figs/errors_IIA-IIB}
\caption{As in Fig. \ref{fig:errorsIA-IB} but for group IIA- and IIB-based compounds.}
\label{fig:errorsIIA-IIB}
\end{figure}
\subsection{Group IIA- and IIB-based compounds}
\label{Section:IIA-IIB}
The percentage errors in the calculated mBJLDA@PP band gaps of the considered IIA- and
IIB-based compounds, using the different sets of cations PP's (t-PP, cp-PP, csp-PP), are
depicted in Fig. \ref{fig:errorsIIA-IIB}. Moreover, the obtained results for the Sr- and
Ba-based compounds by using cd-PP's are also shown
(see Sec. \ref{Section:comp_details} and Table S2 in the Supporting Information).
We note that for the light cations (namely, Be and Mg) only the t-PP's can be
generated. We also note that CaTe and SrTe are excluded from this figure
because of the lack of experimental band gaps, but their calculated band gaps are
available in Table S1. Also shown are the percentage errors in mBJLDA@AE and LDA
band gaps.
For the considered group IIA-based compounds, the important features to note from
Fig. \ref{fig:errorsIIA-IIB} are the following.
(i) The best agreement with the mBJLDA@AE band gaps is achieved by employing the
mBJLDA@t-PP approach.
(ii) The best agreement with the experimental band gaps can be achieved when the
upper core electrons are treated as valence in the Ca, Sr, and Ba PP's.
More specifically, for the considered Ca-based compounds the best agreement with
experiment is achieved by using the mBJLDA@csp-PP approach. However, the differences
between the mBJLDA@PP band gaps obtained using cp-PP's and using csp-PP's are relatively
very small. On the other hand, for the Sr- and Ba-based compounds shown in
Fig. \ref{fig:errorsIIA-IIB} the best agreement with experiment is achieved
when cd-PP's are used, except for SrTiO$_3$. Including additional ($p$ or $s+p$) core
electrons as valence leads to slightly smaller mBJLDA@PP band gaps (i.e., increasing
further their underestimation).
(iii) The mBJLDA@PP band gaps of Ca-, Sr-, and Ba-based compounds are still
underestimated compared to experiment, which is also the case for mBJLDA@AE band gaps.
As for group IIB-based compounds, it can be easily seen from
Fig. \ref{fig:errorsIIA-IIB} that mBJLDA@t-PP band gaps are significantly smaller than
the corresponding mBJLDA@AE ones. This behavior is similar to that observed for group
IB-based compounds, see above. However, contrary to group IB-based compounds,
the mBJLDA@t-PP band gaps are better than the corresponding LDA ones, except for ZnO.
Fig. \ref{fig:errorsIIA-IIB} also shows that the best agreement between the mBJLDA@PP
band gaps and the corresponding experimental and mBJLDA@AE band gaps can be achieved
by using the cation cp-PP's, except for ZnS. The additional inclusion of the uppermost
core $s$ electrons (i.e., using csp-PP's) leads to overestimated band gaps compared
to experiment. The resulting small absolute errors in mBJLDA@PP band gaps when cp-PP's
are used (between 0.02-0.26 eV) are close to the uncertainties in the experimental
band gaps \cite{Crowley2016a}.
In summary, for this set of compounds the best overall agreement with the experimental
band gaps is achieved when the electrons in the upper core orbital are treated as
valence in the PP's of the cations whenever possible, except for Ca-based compounds.
That is, by using cd-PP's for Sr and Ba, csp-PP for Ca, and cp-PP's for the remaining cations,
except for Be and Mg where only t-PP's are possible. This can be also seen from
the calculated MAE of the mBJLDA@PP band gaps compared to the experimental data.
The value of the MAE when only t-PP's are used for all these systems is of 0.98 eV.
This value decreases to 0.44 eV when the appropriate cd-, cp-, or csp-PP's are used
(see above). On the other hand, the MAE
increases to 0.54 eV when csp-PP's are consistently used whenever
applicable. For comparison, the MAE of the mBJLDA@AE band gaps for these systems
is of 0.62 eV.
\begin{figure}[t]
\centering
\includegraphics[width=0.65\linewidth]{figs/errors_3B}
\caption{As in Fig. \ref{fig:errorsIA-IB} but for group IIIA-based compounds.}
\label{fig:errorsIIIB}
\end{figure}
\subsection{Group IIIA-based compounds}
The percentage errors in the mBJLDA@PP band gaps of the considered group IIIA-based
compounds (using the different sets of PP's) are shown in Fig. \ref{fig:errorsIIIB},
together with the percentage errors in the mBJLDA@AE and LDA band gaps. We note here
that only t-PP can be generated for B. Additionally, BSb is not included in the
figure due to the lack of experimental band gap, and its calculated band gaps are
listed in Table S1.
The important features to note from Fig. \ref{fig:errorsIIIB} are the following.
(i) The mBJLDA@t-PP band gaps of the considered B- and Al-based compounds are,
generally speaking, in a good agreement with experiment and mBJLDA@AE calculations.
In fact, for the Al-based compounds the best agreement
between mBJLDA@PP and mBJLDA@AE band gaps is achieved when the t-PP of Al is employed,
except for AlP.
On the other hand, using the cp- or csp-PP's of Al leads to a small further opening
of calculated mBJLDA@PP band gaps, which, in turn, leads to a slightly better
agreement with experiment than the mBJLDA@t-PP approach for some compounds, namely AlP
and the wurtzite phase of AlN.
(ii) The situation is different for the considered Ga- and In-based compounds,
where large differences are observed between mBJLDA@PP band gaps obtained using
different sets of PP's. For these systems the mBJLDA@PP band gaps are
underestimated (overestimated) compared to the corresponding experimental data
when t-PP's (csp-PP's) are employed. The best agreement with the experimental band gaps
for the Ga- and In-based compounds is achieved when cp-PP's are used.
In overall, the best agreement with the experimental band gaps for group IIIA-based
compounds with mBJLDA@PP is obtained when cp-PP's are used for the cations
whenever applicable. This is reflected in the MAE in the mBJLDA@PP band gaps,
which decreases from 0.33 eV when only t-PP's are used to 0.22 eV when cp-PP's
are used whenever possible, while it is of 0.36 eV when the csp-PP's are employed
instead of the cp-PP's. We note that the MAE in mBJLDA@AE band gaps for this set
of systems is 0.25 eV.
\begin{figure}[t]
\centering
\includegraphics[width=0.65\linewidth]{figs/errors_TM}
\caption{As in Fig. \ref{fig:errorsIA-IB} but for transition metal compounds.}
\label{fig:errorsTM}
\end{figure}
\subsection{Transition metal compounds}
The percentage errors in the calculated mBJLDA@PP band gaps of the considered
transition metal compounds using the different sets of PP's are shown in
Fig. \ref{fig:errorsTM}. This set of systems contains ScN and transition metal
oxides, in addition to MoS$_2$. Among these systems are antiferromagnetic
transition metal oxides (MnO, FeO, CoO, NiO, Fe$_2$O$_3$, Cr$_2$O$_3$). We note that for
the Mo-based compounds the cd-PP of Mo is also used. Also shown in Fig. \ref{fig:errorsTM}
are the percentage errors in calculated mBJLDA@AE and LDA band gaps.
The three main features to note from Fig. \ref{fig:errorsTM} are the following.
(i) For these highly correlated electron systems, LDA fails badly as expected to describe
their band gaps: the underestimation is close to 100\% for many of them.
(ii) The mBJLDA@t-PP approach provides a significant improvement in the calculated
band gaps, leading generally speaking to a good agreement with the mBJLDA@AE band gaps.
This is particularly the case for systems with light cations (namely, the Sc- and
Ti-based compounds, and VO$_2$).
(iii) Appreciable opening in the mBJLDA@PP band gaps occurs by increasing NETV, except
for FeO, CoO and NiO where small reductions take place. The mBJLDA@PP band gaps of CoO
and NiO are overestimated, and hence increasing NETV leads to a better agreement with
experiment. In the case of MoS$_2$ the dependence on the NETV is very weak ($\sim$0.02 eV).
These effects lead, generally speaking, to a better agreement with experiment especially
when the uppermost ($p$ or $d$) core electrons are included as valence in the cation PP's.
The calculated MAE in the band gaps in this case is 0.40 eV,
compared to 0.58 eV and 0.41 eV when t-PP's and csp-PP's are used, respectively. For
comparison, the MAE in the mBJLDA@AE band gaps is 0.39 eV.
(iv) The mBJLDA@PP band gaps of CoO (of 4.01 and 3.94 eV, using respectively cp-
and csp-PPs) show the largest relative deviations from the adopted experimental value
of 2.5 eV \cite{Elp1991}. It is worth noting that a wide range of experimental band
gaps can be found in literature, ranging from 2.1 eV \cite{Pratt1959} to 5.43 eV
\cite{Kang2007}. Similar variations in the reported experimental band gaps can
be also found for MnO (2.0-4.2 eV) and NiO (3.7-4.3 eV) \cite{Crowley2016a}.
\subsection{Group IVA-based systems, rare gas solids, and Sb$_2$Te$_3$. }
In this subsection we will present and discuss the mBJLDA@PP band gaps of the rest of
the 83 considered systems. Namely, rare gas solids (Ne, Ar, Kr, and Xe), elemental
group IVA solids (C, Si and Ge), some Si- and Sn-based compounds, and
Sb$_2$Te$_3$, which is the only group VA-based compound considered here.
The percentage errors in the calculated mBJLDA@PP band gaps using the different sets
of PP's are shown in Fig. \ref{fig:errorsIVB}.
It is evident from the Fig. \ref{fig:errorsIVB} that the mBJLDA@t-PP approach is
capable of providing a very good description of the band gaps of rare gas solids
and group-IVA elemental solids and compounds. Generally speaking,
increasing the NETV does not lead to improved band gaps. A clear exceptional case is
SnTe where Sn cp-PP provides a better band gap compared to experiment. Moreover, in the
case of Sb$_2$Te$_3$ the use of Sb cp-PP leads to an overestimated band gap. We note
that Sb$_2$Te$_3$ and SnTe have small band gaps (of 0.28 eV and 0.36 eV, respectively), and
hence small differences between the calculated and the experimental
band gaps (e.g., of 0.13 eV and 0.05 eV respectively, when cp-PP's are used)
lead to relatively large percentage errors.
These observations are reflected in the MAE values in
the calculated mBJLDA@PP band gaps, which are respectively of 0.71, 0.74 and 0.79 eV
when t-, cp- and csp-PP's are used. The MAE in mBJLDA@AE band gaps is of 0.48 eV.
A large contribution to this difference in the MAE between the mBJLDA@AE and mBJLDA@PP
band gaps comes from that for Ne, where the absolute error in the mBJLDA@PP band gap is
4.52 eV, compared to 0.85 eV in the mBJLDA@AE value.
\begin{figure}[t]
\centering
\includegraphics[width=0.65\linewidth]{figs/errors_4B5BN}
\caption{As in Fig. \ref{fig:errorsIA-IB} but for group IVA-based systems,
rare gas solids, and Sb$_2$Te$_3$.}
\label{fig:errorsIVB}
\end{figure}
\subsection{General features and trends}
The above discussion shows that the cations of the considered systems can be classified,
according to their optimal PP's to be used in conjunction with the mBJLDA@PP approach, into
two main categories.
(i) Cations where t-PP's are good enough. This set of cations include groups IA and IVA
elements, Sb, rare gases, and the light elements of groups IIA and IIIA
(namely, Be, Mg, B and Al). Including the uppermost core states (whenever applicable) as
valence will not significantly alter the effective potentials seen by the valence and
conduction electrons of their considered solids, leading to only small changes in their band
gaps by increasing NETV.
(ii) Cations where some uppermost core states have to be included as valence. This set
of cations include transition metals, groups IB and IIB, and the remaining (heavier)
elements of groups IIA and IIIA. For these cations including the outer core $p$ or $d$
(in Sr, Ba and Mo atoms) electrons as valence leads to a very good overall agreement between
the mBJLDA@PP and mBJLDA@AE band gaps.
The above different behaviors can be attributed to the degree of overlap between the valence
and core charge densities. If such overlap is insignificant then the cation belongs to category
(i). Otherwise, it belongs to category (ii). Including explicitly
these core electrons as valence alters the effective potential in such a way that the
upward shifts in the eigenvalues of uppermost valence band states are smaller than
those of the lowest conduction band ones, leading to an opening of mBJLDA@PP band gaps.
This is illustrated in details in Ref. \cite{Abu-Farsakh2021} for the case of ZnS.
Now, focusing on the cations of type (ii), it can be easily observed (see
Figs. \ref{fig:errorsIA-IB}, \ref{fig:errorsIIA-IIB} and \ref{fig:errorsIIIB}) that
the opening of the band gap, by including uppermost core electrons as valence, depends
appreciably on the group of the cation.
Staring with group IB-based compounds, the opening of the mBJLDA@PP band gaps
is relatively small and increases with increasing NETV. The band gaps obtained
using both cp- and csp-PP's are smaller than the corresponding experimental band gaps
(see Fig. \ref{fig:errorsIA-IB}), except for the Ag-based compounds. For this reason,
the mBJLDA@PP band gaps of Ca-, Cu- and Ag-based compounds obtained by using csp-PP's
are in better agreement with experiment than those obtained using cp-PP's. Whereas,
the latter band gaps are in better agreement with the mBJLDA@AE ones.
The opening in the mBJLDA@PP band gaps is relatively
larger in the IIB-based compounds (see Fig. \ref{fig:errorsIIA-IIB}), making the ones
obtained using the cp-PP's very close to the experimental data, while those obtained
using the csp-PP's are overestimated.
This trend continues when going to the IIIA-based compounds (see Fig. \ref{fig:errorsIIIB}),
and in this case the mBJLDA@PP band gaps obtained using csp-PP's and some of the ones
obtained using the cp-PP's are overestimated.
In order to investigate the overall performance of the mBJLDA@PP approach for band gap
calculations, we will consider the band gaps obtained by the optimal choices of the
cations PP's (t-, cd-, cp-, or csp-PP's, that lead to the best overall
agreement with experiment, as described above), hereafter referred to as opt-PP's.
In particular, we considered csp-PP's instead of cp-PP's for Ca, Cu, and Ag because they give
better agreement with experiment (see Sec. \ref{Section:IIA-IIB}).
In Fig. \ref{fig:errors_stats} we show histograms of the frequency of
the percentage errors in the calculated mBJLDA@PP band gaps obtained using (a) t-PP's
and (b) opt-PP's whenever applicable, compared to (c) those of the mBJLDA@AE band gaps. Moreover, in
Fig. \ref{fig:copmpare_expt} the calculated mBJLDA@opt-PP and mBJLDA@AE band gaps
are compared to experimental data. For more details on the comparison between the mBJLDA@PP
band gaps and those of mBJLDA@AE and experimental data we refer the interested readers
to Figs. S1, S2, and S3 in the Supporting Information.
It can be easily seen from Fig. \ref{fig:errors_stats} that t-PP's tend to lead
to underestimated band gaps, and this deficiency is highly removed when opt-PP's
are used. The distribution of percentage errors in mBJLDA@opt-PP band gaps
become very close to that for the mBJLDA@AE approach.
One may argue that by treating more core electrons as valence the mBJLDA@PP band gaps
should approach the corresponding all-electron ones. The better agreement with the mBJLDA@AE
band gaps achieved by increasing NETV is, generally speaking, in accordance with this argument.
It is remarkable that an excellent agreement can be achieved by treating only the outermost
core electrons as valence. The small discrepancies between the mBJLDA@opt-PP and mBJLDA@AE
band gaps can be attributed to NETV in both approaches and the pseudization of the
wavefunctions in the former one.
For a more quantitative analysis we show in Table \ref{table:errors} a summary of
the statistical analysis of the errors in the calculated band gaps using different
approaches. Additionally included are the errors in the band gaps obtained with the
hybrid functional HSE06, taken from Refs. \cite{Tran2017,Borlido2019}.
In addition to the MAE, we also show the mean error, ME, and the mean
(absolute) percentage error, M(A)PE. The important features to note are the following.
(i) The ME and and MPE in the LDA band gaps, which
are highly underestimated, are of about $-2$ eV and $-58$\%,
respectively.
(ii) These errors are reduced when the mBJLDA@t-PP's
approach is employed. In this case the MPE is about $-19$\%, which reflects the
tendency of this approach to underestimate the band gaps. The corresponding
MAE and MAPE values (of 0.85 eV and 26\%) are larger than those obtained
with mBJLDA@AE calculations (of 0.49 eV and 15\%).
(iii) When the opt-PP's are used in mBJLDA@PP calculations the errors are substantially
reduced, with MAE and MAPE of 0.46 eV and bout 12\%, respectively. It can be noted that these
errors compare very well to those obtained with mBJLDA@AE calculations.
(iv) The shown ME and MPE values indicate that the mBJLDA@AE approach has more tendency
to underestimate the band gaps than the mBJLDA@opt-PP's method. This is also clear from
Figs. \ref{fig:errors_stats} and \ref{fig:copmpare_expt}. (v) The errors in the
mBJLDA@opt-PP and mBJLDA@AE approaches compare well with those obtained with the HSE06 method.
Finally, we compare the mBJLDA band gaps with those of the $GW$ approximation, which is
the most formal and rigorous approach for band structure calculations. Before doing that,
it worth noting that there are
several technical details that have significant effects on the single-shot $GW$
(or $G_0W_0$) band gaps (see for example Ref. \cite{Golze2019}). Moreover, in
addition to $G_0W_0$, partially \cite{Jiang2016,Shishkin2007} and fully self-consistent
\cite{Grumet2018,Kutepov2017} $GW$ approaches have been developed, and they provide
different levels of agreement with experiment \cite{Golze2019}. Therefore, a comprehensive
comparison between mBJLDA and $GW$ band gaps is not a simple task. As a representative, we
will consider the $GW_0$@PBE band gaps reported by Jiang and Blaha \cite{Jiang2016}, for
a much shorter list of compounds.
The percentage errors in these $GW_0$ band gaps are shown in Fig. \ref{fig:errors_GW},
together with the corresponding ones in the mBJLDA@opt-PP, mBJLDA@AE and HSE06 band gaps.
It can be noted that all these methods generally provide very
good agreement with experiment. The $GW_0$ band gaps have the
lowest ME (MAE) of $-$0.02 (0.21) eV, compared to $-$0.04 (0.25) eV using mBJLDA@opt-PP,
$-$0.24 (0.32) eV using mBJLDA@AE, and $-$0.49 (0.51) eV using HSE06. On the other hand,
the calculated MAPE values in the $GW_0$, mBJLDA@opt-PP, mBJLDA@AE and HSE06 band gaps
are respectively of 7.3\%, 7.1\%, 8.1\% and 9.9\%.
\begin{table}[H]
\centering
\begin{tabular}{lSSSSS}
\toprule
& \multicolumn{1}{c}{LDA} & \multicolumn{3}{c}{mBJLDA} & \multicolumn{1}{c}{HSE06} \\
\cmidrule{3-5}
& & \multicolumn{1}{c}{@t-PP} & \multicolumn{1}{c}{@opt-PP} & \multicolumn{1}{c}{@AE} & \\
\midrule
ME (eV) & -2.15 & -0.41 & 0.01 & -0.33 & -0.69 \\
MAE (eV) & 2.15 & 0.85 & 0.46 & 0.49 & 0.82 \\
MPE (\%) & -57.63 &-18.82 & -1.63 & -6.01 & -7.57 \\
MAPE (\%) & 57.63 & 26.31 & 12.41 & 15.20 & 17.15 \\
\bottomrule
\end{tabular}
\caption{Mean (absolute) error, M(A)E, and mean (absolute) percentage error, M(A)PE, in
the band gaps using different approaches, with respect to experiment. The mBJLDA@AE
and HSE06 band gaps are taken from Refs. \cite{Tran2017,Tran2019,Ye2015} and Refs.
\cite{Tran2017,Borlido2019}, respectively. For more details see Table S1 in the
Supporting Information.}
\label{table:errors}
\end{table}
\begin{figure}[H]
\centering
\includegraphics[width=0.5\linewidth]{figs/stats}
\caption{Histograms showing the frequency of the percentage errors in the
calculated mBJLDA@PP band gaps using typical PP's (a) and optimal PP's (b).
For comparison, in (c) a similar histogram is shown for the percentage
errors in reference mBJLDA@AE band gaps \cite{Tran2017,Tran2019,Ye2015}.}
\label{fig:errors_stats}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{figs/compare_expt_r13}
\caption{mBJLDA@PP band gaps obtained using optimal PP's versus experimental
band gaps. The right panel is an enlarged view of the left panel focusing on
band gaps smaller than 6 eV. For comparison purposes the figure also shows
the LDA@t-PP and reference mBJLDA@AE band gaps \cite{Tran2017,Tran2019,Ye2015}.}
\label{fig:copmpare_expt}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.65\linewidth]{figs/errors_GW0}
\caption{Percentage errors in the $GW_0$@PBE band gaps for selected systems (Ref. \cite{Jiang2016})
together with the corresponding errors in the mBJLDA@opt-PP, mBJLDA@AE (Ref. \cite{Tran2017}) and
HSE06 band gaps (Ref. \cite{Tran2017}). The points are connected as a guide to the eye.}
\label{fig:errors_GW}
\end{figure}
\section{Conclusions}
In this work we performed mBJLDA calculations employing a norm-conserving
pseudopotential plane-waves approach (mBJLDA@PP) for a test set of 83 solids,
representing a wide range of semiconductors and insulators. We investigated
the effect of the number of electrons treated as valence in the PP's of the cations.
The obtained band gaps are discussed in comparison with those obtained using
all-electron mBJLDA (mBJLDA@AE), $GW_0$, HSE06, and experimental
data. The main conclusions based on the results obtained for the above considered
systems can be summarized as follows.
\begin{enumerate}
\item For the group IA-based compounds, the use of typical pseudopotentials
(t-PP's), where valence and semicore states are pseudized, leads to band gaps
that are in very good agreement with the mBJLDA@AE ones and experiment.
\item For the group IIA-based compounds, the best overall agreement between the
mBJLDA@PP and mBJLDA@AE band gaps is obtained by using t-PP's. The effects of including
uppermost core states as valence are quite small. However, including the uppermost $s$
and $p$ (in Ca PP) and $d$ (in the Sr and Ba PP's) core electrons as valence leads to
mBJLDA@PP band gaps in better agreement with experiment. These PP's are referred to
as csp-PP's (cd-PP's).
\item For groups IB-, IIB-, and IIIA-based compounds the best overall
agreement with the mBJLDA@AE band gaps is obtained when t-PP's (for
light elements, namely Be, Mg, B and Al) and cp-PP's (for the heavier elements)
are employed. In the case of Cu- and Ag-based compounds, a better agreement
with experiment is achieved when csp-PP's are employed.
\item For transition-metal based compounds the best overall agreement
with the mBJLDA@AE and experimental band gaps is obtained by using cp-PP's.
\item For group IVA elemental solids and compounds, rare gas solids, and Sb$_2$Te$_3$
the best overall agreement with the mBJLDA@AE and experimental band gaps
are obtained by using t-PP's.
\item The relative opening of the mBJLDA@PP band gaps by including the uppermost core
states as valence increases by going from IB- to IIB- to IIIA-based compounds.
\item The overall accuracy of the mBJLDA@PP approach when the above optimal PP's are employed
is very close to that of $GW_0$ and slightly better than that of the mBJLDA@AE and HSE06 methods.
\end{enumerate}
\section*{Acknowledgments}
The authors would like to thank Mohammed Abu-Jafar for fruitful discussion. The first author would like to
thank Yarmouk University for providing the computational resources, and Prince Sultan University for their support.
\bibliographystyle{elsarticle-num}
\input{manuscript.bbl}
\end{document}
|
\section{Introduction}
The kinetic theory offers an excellent mathematic model for rarefied gases. The celebrated Boltzmann equation \footnote{As usual repeated indices indicated omitted sum symbol.}
\begin{equation} \label{eq:Boltzmann}
\partial_t f^C + \xi^i\, \partial_i f^C= Q^C,
\end{equation}
is widely used in many applications and is still now a challenge for its difficult mathematical questions.
The state of the gas is described by the distribution function $f^C(\mathbf{x}, t, {\boldsymbol \xi})$,
being respectively $\mathbf{x} \equiv(x^i),{\boldsymbol \xi}\equiv(\xi^i),t$ the space coordinates, the microscopic velocity and the time. $Q^C$ denotes the collisional term and $\partial_t = \partial /\partial t$, $\,\, \partial_i = \partial/\partial x_i$, ($i=1,2,3$). There are many results on Boltzmann equation, in particular we quote for the mathematical treatment the books of Cercignani \cite{Cer1,Cer2} who was one of the world leaders that gave fundamental papers on this subject.
The relativistic counterpart of Boltzmann equation is the Boltzmann-Chernikov equation
\cite{BGK,Synge,KC}:
\begin{equation}\label{BoltzR}
p^\alpha \partial_\alpha f = Q,
\end{equation}
in which the relativistic distribution function $f$ depends on $(x^\alpha, p^\beta)$, where $x^\alpha$ are the space-time coordinates, $p^\alpha$ is the four-momentum, $\partial_{\alpha} = \partial/\partial x^\alpha$, $Q$ is the collisional term and $\alpha, \beta =0,1,2,3$.
Formally the relativistic equation converges to the classical one if we take into account the following expressions (see for example \cite{Annals})
\begin{align}\label{limitini}
\begin{split}
&x^0= c t, \quad p^0 = m \, \Gamma \, c \, , \quad p^i = m \, \Gamma \, \xi^i \, , \quad \Gamma = \left( 1- \, \frac{\xi^2}{c^2} \right)^{- \frac{1}{2}} , \\
&\lim_{c \rightarrow \, + \infty} f= \frac{1}{m^3} \, f^C \, , \quad \lim_{c \rightarrow \, + \infty} Q= \frac{1}{m^2} \, Q^C ,
\end{split}
\end{align}
where $c$ denotes the light velocity, $m$ is the particle mass in the rest frame and $\Gamma$ is the Lorentz factor.
The weak point of the Boltzmann equation both in classical and relativistic regimes is that their validity holds only for monatomic gas even if the classical kinetic theory was used in fields very far from gas dynamics like in biological phenomena, socio-economic systems, models of swarming, and many other fields (see, for example, \cite{Preziosi1,Tosin,Carillo} and references therein).
A more realistic case which is important for applications is the kinetic theory of polyatomic gas. In the classical framework was proposed based on two different approaches:
- the description of the internal structure of a polyatomic gas is taken into account by a large number of discrete energy states, so that the gas might be considered as a sort of mixture of monatomic components, which interact by binary collisions with conservation of total energies, but with possible exchange of energy between its kinetic and internal (excitation) forms. The model can be used also in a reactive frame, even in the presence of its self--consistent radiation field \cite{GroppiSpiga}.
\vskip 6pt
- Another approach in the development
of the theory of rarefied polyatomic gases was made by Borgnakke and Larsen \cite{Borgnakke-1975}.
The distribution function is assumed to depend on an additional continuous variable $\cal{I}$
representing the energy of the internal modes of a molecule in order to take into account the exchange
of energy (other than translational one) in binary collisions. This model was initially used
for Monte Carlo simulations of polyatomic gases, and later it has been applied to the derivation
of the generalized Boltzmann equation by Bourgat, Desvillettes, Le Tallec and Perthame \cite{Bourgat-1994}.
In this case the Boltzmann equation \eqref{eq:Boltzmann} has the same form but the
distribution function $f^C(\mathbf{x}, t, {\boldsymbol \xi},\cal{I})$ is
defined on the extended domain ${R}^{3} \times [0,\infty)
\times {R}^{3} \times [0,\infty)$
and the collision integral takes into account
the influence of the internal degrees of freedom through the collisional
cross section. The case of non polytropic gases in which the internal energy is a non-linear function of the temperature was considered by Ruggeri and coworkers in a series of papers \cite{rendmat,RuggeriSpiga,Ruggeri-2020RdM}. A more refined case in which the internal mode is divided into the rotational and vibrational modes was presented by Arima, Ruggeri and Sugiyama \cite{ET7,ET15a}. Concerning the production terms it was used a BGK model \cite{Ruggeri-2020RdM} or an extended one with two or more relaxation times \cite{Struch1,Struch2,ET7,ET15a}.
In the relativistic framework, Pennisi and Ruggeri \cite{Annals} used a similar technique, and they postulate the same Boltzmann-Chernikov equation \eqref{BoltzR} but with a distribution function $f(x^\alpha,p^\alpha,\cal{I})$ depending on the microscopic energy due to the internal mode. The same authors construct a new BGK model both for monatomic and for polyatomic gas in \cite{BGKPR}. The existence and asymptotic behavior of classical solutions for this model when the initial data is sufficiently close to a global equilibrium was the subject of the paper \cite{Koreani}.
Both in the classical and relativistic theory, we can associate macroscopic quantities, called moments, which satisfy an infinite set of balance laws. The choice of the moments is a controversial question in particular in the polyatomic gas.
The closure of moments when the number is finite is the starting point of modern Rational Extended Thermodynamics (RET). The aim of this paper is to discuss first the classical limit of a new hierarchy or relativistic moments and in the particular case of a RET with 15 moments to discuss the qualitative analysis of the solutions. In particular, we prove that these systems, both in relativistic and classical cases, satisfy the conditions of the well-known theorems for the existence of the global smooth solutions for initial data sufficiently small.
In the present paper, first we discuss the relativistic Boltzmann-Chernikov equation for polyatomic gases and after in Sec. 2, we present a brief review on the possible physical moments that take into account the total energy composed by the rest energy and the energy of molecular internal states. Then in Sec. 3, we will study the non-relativistic limit of the system of balance equations for any number of moments. As a particularly interesting case, we summarize in Sec. 4 the result of the RET theory with 15 moments. In Sec. 5, we show that for the relativistic RET with 15 moments and its classical limit hold the theorems of the existence of the global smooth solutions under given sufficiently small initial data.
\section{Moments associated to the Kinetic Equation}
In the classical case and for monatomic gases the moments are:
\begin{equation}
F_{k_1 k_2 \cdots k_j}= m\int_{\mathbb{R}^{3}}{f^C \xi_{k_1} \xi_{k_2} \cdots \xi_{k_j} d{\bm \xi} },\qquad (j=0,1,\dots ), \label{prima}
\end{equation}
$k_1,k_2,\dots =1,2,3$ and
by convention when $j=0$ we have
\begin{equation*}
F = m \int_{\mathbb{R}^{3}}{f d{\bm \xi} }.
\end{equation*}
Due to the Boltzmann equation \eqref{eq:Boltzmann}, the moments satisfy an infinite hierarchy of balance laws in
which the flux in one equation becomes the density in the next one:
\begin{equation} \label{F-Gerarchia}
\partial_t F_A + \partial_i F_{iA}= P_{A}, \qquad (A=0,1,\dots)
\end{equation}
where we introduce the following multi-index notation:
\begin{equation} \label{F-Mom}
F_A = F_{i_1 i_2\dots i_A}, \quad F_{iA} = F_{ii_1 i_2\dots i_A}, \quad P_A = P_{i_1i_2\dots i_A},
\end{equation}
with
\begin{equation}\label{ggg}
P_{k_1 k_2 \cdots k_j}= m \int_{\mathbb{R}^3}{Q \xi_{k_1} \xi_{k_2} \cdots \xi_{k_j} d\boldsymbol{\xi}}.
\end{equation}
The relativistic counterpart of the moment equations for monatomic gas are
\begin{equation}\label{MRel}
\partial_\alpha A^{\alpha \alpha_1 \cdots \alpha_n } = I^{ \alpha_1 \cdots \alpha_n }
\quad \mbox{with} \quad n=0 \, , \,\cdots \, , \,
\end{equation}
with
\begin{align*
\begin{split}
& A^{\alpha \alpha_1 \cdots \alpha_n } = \frac{c}{m^{n-1}} \int_{\mathbb{R}^{3}}
f \, p^\alpha p^{\alpha_1} \cdots p^{\alpha_n} \, \, d \boldsymbol{P}, \\
& I^{\alpha_1 \cdots \alpha_n } = \frac{c}{m^{n-1}} \int_{\mathbb{R}^{3}}
Q \, p^{\alpha_1} \cdots p^{\alpha_n} \, \, d \boldsymbol{P} ,
\end{split}
\end{align*}
where the Greek indices run from $0$ to $4$, and
\begin{equation*}
d \boldsymbol{P} = \frac{dp^1 \, dp^2 \,
dp^3}{p^0} .
\end{equation*}
If we truncate the moments \eqref{MRel} until the index $N$, i.e. $n=0,1,2,\dots,N$, it was proved in \cite{JSP} the following theorem:
\begin{theorem}[Pennisi-Ruggeri \cite{JSP}]\label{CLRM}
- For a prescribed truncation index $N$, for any integer
$0\leq s \leq N$ and for multi-index $0\leq B \leq N-s$, the relativistic moment system for a monatomic gas \eqref{MRel} with $n = 0,1,\dots,N$
converges, when $c\rightarrow \infty$, to the following classical moments system:
\begin{align} \label{HSAM}
\begin{split}
& \partial_t F_{B}+ \partial_i F _{iB} = P _{B}, \quad \text{if} \quad s=0, \quad 0\leq B\leq N, \quad \text{and} \\
& \partial_t F_{j_1 j_1\dots j_s j_s i_1 i_2 \dots i_{N-s}}+ \partial_i F _{i j_1j_1 \dots j_s j_s i_1 i_2 \dots i_{N-s}} = P _{j_1 j_1\dots j_s j_s i_1 i_2 \dots i_{N-s}},
\end{split}
\end{align}
with $1\leq s \leq N$.
The moments $F$'s are given by \eqref{prima} and the productions $P$'s are given by \eqref{ggg}.
In particular, for $s=0$, we have the $ F$'s moments with all free indexes until index of truncation $N$ and for $1\leq s \leq N$ there is a single block of $F$'s moments with increasing number of pairs of contracted indexes. The truncated tensorial index in \eqref{HSAM} is $\bar{N} = 2 N$.
\end{theorem}
This theorem solved the old problem how to choose in an optimal way the moments in the classical case. In fact, there are several degrees of freedom depending on how many indices are saturated in the truncated tensors. For example, in the Grad system for which in \eqref{F-Gerarchia} $\bar{N}=3$ instead of taking all free indices, it was considered two indexes saturated in the triple tensor: $(F,F_i,F_{ij},F_{kki})$.
We remark that, for $N=1$, the system \eqref{MRel} is the Euler relativistic fluid and the classical limit is the Euler classical fluid with moments $(F,F_i,F_{ll})$ of which balance laws correspond to the mass, momentum and energy conservations. While, for $N=2$, the relativistic system \eqref{MRel} is the one proposed by Liu, M\"uller and Ruggeri \cite{LMR} and the classical limit converges to the $14$ moments proposed by Kremer \cite{Kremer}: $(F,F_i,F_{ij},F_{lli},F_{kkjj})$, instead of the Grad system. The Grad $13$ moments don't correspond to any classical limit of a relativistic theory!
\bigskip
How to construct moments in the polyatomic case was an open problem. Starting from the equilibrium case of $5$ moments proposed by Bourgat et al. \cite{Bourgat-1994} a double hierarchy of moments was proposed first at macroscopic level with $14$ field by Arima et al. \cite{dense} and successively at kinetic level in the papers \cite{Pavic-2013,AMR}, (see for more details \cite{newbook}):
\begin{align*
\begin{split}
& F_{i_1 \ldots i_j} = m \int_{\mathbb{R}^{3}}\int_0^{+\infty} f^C \, \xi_{i_1}\cdots \xi_{i_j} \, \, \varphi(\mathcal{I}) \, d\mathcal{I} d {\boldsymbol \xi} , \\
& {G_{lli_1 \ldots i_k}} = 2\int_{\mathbb{R}^{3}}\int_0^{+\infty} f^C \, \left(\frac{m \xi^2}{2} + \mathcal{I}\right) \xi_{i_1}\cdots \xi_{i_k} \, \, \varphi(\mathcal{I}) \, d\mathcal{I} d {\boldsymbol \xi} .
\end{split}
\end{align*}
Here $\varphi(\mathcal{I})$ is the state density corresponding to $\mathcal{I}$, i.e., $\varphi(\mathcal{I}) d\mathcal{I}$ represents the number of internal state between $\mathcal{I}$ and $\mathcal{I} + d\mathcal{I}$. As (for $k=0$) $G_{ll}$ is the energy, except for a factor $2$, we have that the $F's$ are the usual \emph{momentum-like} moments and the $G's$ are \emph{energy-like} moments.
From the Boltzmann equation \eqref{eq:Boltzmann}, we obtain a binary hierarchy of balance equations so called $(F,G)$-hierarchies:
\begin{align}\label{momentiniFG}
\begin{split}
& \partial_t F_{k_1 k_2 \dots k_n} \hspace{0.3cm}+\partial_i F_{k_1 k_2 \dots k_n i} \,\, \,\, = P_{k_1k_2 \dots k_n}, \qquad \quad \,\,\,\, n=0,1,\dots , \bar{N}, \\
& \partial_t G_{ll k_1 k_2 \dots k_m} + \partial_i G_{ll k_1 k_2 \dots k_m i} = Q_{ll k_1k_2 \dots k_m}, \qquad \,\,\, m=0,1,\dots , \bar{M}.
\end{split}
\end{align}
From the requirement of the Galilean invariance and the physically reasonable solutions, it is shown that $\bar{M}=\bar{N}-1$ \cite{AMR}. The case with $\bar{N}=1$ corresponds to the Euler system, and the one with $\bar{N}=2$ corresponds to RET with $14$ moments \cite{dense,Pavic-2013}.
Pennisi and Ruggeri first in \cite{Annals} and then in \cite{JSP} proved that the relativistic theory of moments for polyatomic case contains in the classical limit the $(F,G)$-hierarchy if we consider a system \eqref{MRel} but with the following moments:
\begin{align}
\begin{split}
& A^{\alpha \alpha_1 \cdots \alpha_n } = \frac{1}{m^n c} \int_{\mathbb{R}^{3}}\int_0^{+\infty} f \, p^\alpha p^{\alpha_1} \cdots p^{\alpha_n} \, \left(mc^2 + n \mathcal{I} \right)\,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P} \, , \\
& I^{\alpha_1 \cdots \alpha_n } = \frac{1}{m^{n}c} \int_{\mathbb{R}^{3}}\int_0^{+\infty} Q \, p^{\alpha_1} \cdots p^{\alpha_n} \, \left( mc^2 + n\mathcal{I} \right)\,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P}, \\
\end{split}
\label{relRETold}
\end{align}
where the distribution function $f(x^\alpha, p^\beta,\mathcal{I})$ depends on the extra energy variable $\mathcal{I}$, similar to the classical one.
Pennisi in \cite{Pennisi-2021} noticed first the unphysical situation in which, instead to have the full energy at molecular level, i.e., $mc^2 +\mathcal{I}$ , we have in \eqref{relRETold}$_2$ the term $mc^2 + n\mathcal{I}$ but he observed that $(mc^2)^{n-1}(mc^2 +n \mathcal{I})$ are the first two terms of the Newton binomial formula of $(mc^2 +\mathcal{I})^n/ (mc^2)^{n-1}$.
Therefore he proposed in \cite{Pennisi-2021} to modify, in the relativistic case, the definition of the moments by using the substitution (see also \cite{entropy}):
\begin{equation*
(mc^2)^{n-1}\left( mc^2 + n \mathcal{I} \right) \qquad \text{with \quad } \left( mc^2 + \mathcal{I} \right)^n,
\end{equation*}
i.e., instead of \eqref{relRETold}, the following moments were proposed:
\begin{align} \label{relRET}
& A^{\alpha \alpha_1 \cdots \alpha_n } = \left(\frac{1}{mc}\right)^{2n-1} \int_{\mathbb{R}^{3}}\int_0^{+\infty} f \, p^\alpha p^{\alpha_1} \cdots p^{\alpha_n} \, \left( mc^2 + \mathcal{I} \right)^n\,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P} \, , \nonumber \\
& \\
& I^{\alpha_1 \cdots \alpha_n } = \left(\frac{1}{mc}\right)^{2n-1} \int_{\mathbb{R}^{3}}\int_0^{+\infty} Q \, p^{\alpha_1} \cdots p^{\alpha_n} \, \left( mc^2 + \mathcal{I} \right)^n\,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P}. \nonumber
\end{align}
In the next section we determine what is the classical limit of the truncated system \eqref{MRel} with $n=0,1,\dots,N$ and moments given by \eqref{relRET}.
\section{The non relativistic limit}
In this section we prove the following
\begin{theorem}\label{th2}
For a prescribed truncation integer index $N$ and $0\leq s \leq N$, the relativistic moment system for polyatomic gases \eqref{MRel} (with $n=0,1,\dots,N$) and \eqref{relRET}, converges when $c\rightarrow +\infty$ to the following $N+1$ hierarchies of classical moments:
\begin{align}\label{2.3}
\begin{split}
\partial_t H_{s}^{i_1 \cdots i_h} +& \partial_i \, H_{s}^{i \, i_1 \cdots i_h} = J_s^{ i_1 \cdots i_h } \\
& \mbox{with } s=0 \, , \, \cdots \, , \, N \mbox{ and } h=0 \, , \,\cdots \, , \, N - s
\end{split}
\end{align}
where
\begin{align}\label{2.4}
& H_{s}^{i_1 \cdots i_h} = m \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f^C \, \xi^{i_1} \cdots \xi^{i_h}
\, \left( \frac{2 \, \mathcal{I}}{m} \, + \xi^2 \right)^{s} \, \phi(\mathcal{I}) \, d \, \mathcal{I} \, d{\bm \xi} \, , \nonumber \\
& H_{s}^{i \, i_1 \cdots i_h} = m \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f^C \,\xi^i \xi^{i_1} \cdots \xi^{i_h}
\, \left( \frac{2 \, \mathcal{I}}{m} \, + \xi^2 \right)^{s} \, \phi(\mathcal{I}) \, d \, \mathcal{I} \, d{\bm \xi} \, , \\
& J_s^{ i_1 \cdots i_h } = m \int_{\mathbb{R}^{3}}
\int_0^{+\infty} Q^C \, \xi^{i_1} \cdots \xi^{i_h}
\, \left( \frac{2 \, \mathcal{I}}{m} \, + \xi^2 \right)^{s} \, \phi(\mathcal{I}) \, d \, \mathcal{I} \, d{\bm \xi} \, , \nonumber
\end{align}
$f^C$ and $Q^C$ are the classical limits of $f$ and $Q$ respectively.
In particular, for $s=0$ we have the momentum-like block of equations \eqref{momentiniFG}$_1$, for $s=1$ the energy-like block \eqref{momentiniFG}$_2$ and for $2\leq s \leq N$ there are new blocks never considered before in the literature.
\end{theorem}
\textbf{Proof}: Let us write our equations in 3-dimensional form. Taking into account that $x^ 0=c \, t$ and $\partial_0 = {1}/{c} \, \partial_t $, they become
\begin{equation*}
\frac{1}{c} \, \partial_t A^{0 \alpha_1 \cdots \alpha_n } + \partial_i A^{i \alpha_1 \cdots \alpha_n }= I^{ \alpha_1 \cdots \alpha_n } \, ,
\end{equation*}
or
\begin{align}\label{6a}
\partial_t A^{0 \, {{\scriptsize\stackrel{n-h}{\overbrace{0 \cdots 0}}}} \, i_1 \cdots i_h } + \partial_i \left(c \, A^{i \, {\scriptsize\stackrel{n-h}{\overbrace{0 \cdots 0}}}\, i_1 \cdots i_h } \right)= c \, I^{{\scriptsize\stackrel{n-h}{\overbrace{0 \cdots 0}}}\, i_1 \cdots i_h } \\
\mbox{with} \quad h=0 \, , \, \cdots \, , \, n \, , \quad \mbox{and} \quad n=0 \, , \,\cdots \, , \, N \, . \nonumber
\end{align}
Here $\stackrel{n-h}{\overbrace{0 \cdots 0}}$ represents a set of $n-h$ zeros. From \eqref{limitini} and \eqref{relRET}, we have
\begin{align*}
A^{i \, {\scriptsize\stackrel{n-h}{\overbrace{0 \cdots 0}}}\, i_1 \cdots i_h }= m \, \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f \, c^{n-h} \Gamma^{n+5} \xi^{i} \xi^{i_1} \cdots \xi^{i_h} \, \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right)^n \,
\phi(\mathcal{I}) \,d \, \mathcal{I} \, d{\bm \xi} \, \, ,
\end{align*}
and
\begin{align*}
A^{0 \, {\scriptsize\stackrel{n-h}{\overbrace{0 \cdots 0}}}\, i_1 \cdots i_h }= m \, \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f \, c^{n-h+1} \Gamma^{n+5} \xi^{i_1} \cdots \xi^{i_h} \, \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right)^n \,
\phi(\mathcal{I}) \,d \, \mathcal{I} \, d{\bm \xi} \, \, .
\end{align*}
Eq. \eqref{6a} divided by $c^{n-h+1}$ becomes
\begin{align}\label{2.1}
& \partial_t \tilde{A}_n^{ i_1 \cdots i_h } + \partial_i \, \tilde{A}_n^{i \, i_1 \cdots i_h } = \tilde{I}_n^{ i_1 \cdots i_h } \\
& \mbox{for} \quad h=0 \, , \, \cdots \, , \, n \, , \quad \mbox{and} \quad n=0 \, , \,\cdots \, , \, N \, , \nonumber
\end{align}
with
\begin{eqnarray}\label{2.2}
& \tilde{A}_n^{ i_1 \cdots i_h }= m^4 \, \int_{\mathbb{R}^{3}}\int_0^{+\infty} {f \, \Gamma^{n+5} \xi^{i_1} \cdots \xi^{i_h} \, \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right)^n \,
\phi(\mathcal{I}) \,d \, \mathcal{I} \, d{\bm \xi} \, } \, , \\
& \nonumber \\
& \tilde{A}_n^{ i i_1 \cdots i_h }= m^4 \, \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f \, \Gamma^{n+5} \xi^i \xi^{i_1} \cdots \xi^{i_h} \, \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right)^n \,
\phi(\mathcal{I}) \,d \, \mathcal{I} \, d{\bm \xi} \, \, , \nonumber \\
& \tilde{I}_n^{ i_1 \cdots i_h } = \frac{1}{c^{n-h}} \, I^{ \, i_1 \cdots i_h } =
m^3 \, \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f \, \Gamma^{n+4} \xi^i \xi^{i_1} \cdots \xi^{i_h} \, \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right)^n \,
\phi(\mathcal{I}) \,d \, \mathcal{I} \, d{\bm \xi} \, \nonumber \, .
\end{eqnarray}
We can see that the equations \eqref{2.1} with different n but the same value of $h$ have the same non relativistic limit, so that the number of independent equations is reduced. In order to preserve the number of independent equations, for every fixed value of $h$, we define a new tensor as a linear combination of the $\tilde{A}_n^{ i_1 \cdots i_h }$ from \eqref{2.2}$_1$:
\begin{align}\label{3.2}
& H_{\mbox{rel} , N-n}^{i_1 \cdots i_h} \stackrel{\mbox{def}}{=}\left( 2 \, c^2 \right)^{N-n} \, \sum_{r=0}^{N-n} \, \begin{pmatrix}
\, N-n \, \nonumber\\
\, r \,
\end{pmatrix} \, (-1)^{N-n-r} \tilde{A}_{r+n}^{i_1 \cdots i_h} = \nonumber \\
& m^4 \left( 2 \, c^2 \right)^{N-n} \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f \, \xi^{i_1} \cdots \xi^{i_h} \nonumber \\
& \hspace{1cm} \sum_{r=0}^{N-n} \begin{pmatrix}
\, N-n \, \\
\, r \,
\end{pmatrix}
\, (-1)^{N-n-r} \, \Gamma^{r+n+5} \, \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right)^{r+n}
\phi(\mathcal{I}) \, d \, \mathcal{I} \, d{\bm \xi} \nonumber \\
& = m^4 \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f \, \xi^{i_1} \cdots \xi^{i_h} \, \Gamma^{n+5}
\, \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right)^{n} \nonumber \\
&\hspace{2cm}
\left\{ \underline{2 \, c^2 \, \left[ \Gamma \left( 1 + \frac{\mathcal{I}}{m \, c^2} \right) \, - \, 1 \right]} \right\}^{N-n}
\phi(\mathcal{I}) \, d \, \mathcal{I} \, d{\bm \xi}. \nonumber
\end{align}
The non relativistic limit of the underlined part of the above expression is $2\mathcal{I} /m\, + \xi^2$, so that
\begin{align*}
\begin{split}
& \lim_{c \, \rightarrow \, + \infty} H_{\mbox{rel} , N-n}^{i_1 \cdots i_h} = H_{N-n}^{i_1 \cdots i_h} \quad \mbox{with} \\
& H_{N-n}^{i_1 \cdots i_h} = m \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f^C \, \xi^{i_1} \cdots \xi^{i_h}
\, \left( \frac{2 \, \mathcal{I}}{m} \, + \xi^2 \right)^{N-n} \, \phi(\mathcal{I}) \, d \, \mathcal{I} \, d{\bm \xi} \, ,
\end{split}
\end{align*}
for $h=0 \, , \, \cdots \, , \, n \, , \quad \mbox{and} \quad n=0 \, , \,\cdots \, , \, N$. This set of indexes can be expressed also by the conditions $0 \leq h \leq N \, , \quad 0 \leq n \leq N$ and $h \leq n$. Now we can change index according to the law $N-n=s$ so that
\begin{align*}
H_{s}^{i_1 \cdots i_h} = m \int_{\mathbb{R}^{3}}
\int_0^{+\infty} f^C \, \xi^{i_1} \cdots \xi^{i_h}
\, \left( \frac{2 \, \mathcal{I}}{m} \, + \xi^2 \right)^{s} \, \phi(\mathcal{I}) \, d \, \mathcal{I} \, d{\bm \xi} \, ,
\end{align*}
and the above set of indexes transforms in $0 \leq h \leq N \, , \quad 0 \leq s \leq N$ and $h \leq N-s$ or, equivalently, for $s=0 \, , \, \cdots \, , \, N \, , \quad \mbox{and} \quad h=0 \, , \,\cdots \, , \, N-s$. Eq. $\eqref{2.4}_1$ is proved. \\
The same passages can be followed starting from \eqref{2.2}$_2$, \eqref{2.2}$_3$.
Finally, starting from \eqref{2.1} we can prove our theorem, showing that the non relativistic limit is \eqref{2.3}.
\smallskip
We observe that also in the classical limit now appears in the moments \eqref{2.2} the full energy given by the sum of kinetic energy plus the energy of internal modes: $m \xi^2/2+\cal{I}$.
\subsection{Particular cases}
As example we consider the case $N=1$. The relativistic moments \eqref{MRel} with $n=0,1$ reduces now to
\begin{equation}\label{3.2}
\partial_{\alpha} A^\alpha = 0, \qquad \partial_{\alpha} A^{\alpha \beta} = 0
\end{equation}
that corresponds to the Euler relativistic polyatomic gas. The corresponding limit according with the Theorem \ref{th2} is:
\begin{align*
& s=0 \,\, \| \quad \partial_t H_0^0 +\partial_i H_0^i = 0,\, \quad \leftrightarrow \quad \partial_t F +\partial_i F^i= 0, \quad \,\, \text{(mass)} \nonumber\\
&\quad \quad \quad \,\,\,\,\,\, \partial_t H_0^j +\partial_i H_0^{ij} = 0, \quad \leftrightarrow \quad \partial_t F^j +\partial_i F^{ij}= 0, \quad \text{(momentum)} \\
& s=1 \,\, \| \quad \partial_t H_1^0 +\partial_i H_1^i = 0,\, \quad \leftrightarrow \quad \partial_t G_{ll} +\partial_i G^{lli}= 0, \quad \text{(energy)}, \nonumber
\end{align*}
i.e. the Euler classical polyatomic gas.
In the case $N=2$ the relativistic moments \eqref{MRel} with $n=0,1$ reduce now to the $15$ moments that generalize the LMR theory to the polyatomic gases \cite{entropy}:
\begin{equation}\label{Rpolino}
\partial_{\alpha} A^\alpha = 0, \quad \partial_{\alpha} A^{\alpha \beta} = 0, \quad \partial_{\alpha} A^{\alpha \beta \gamma} = I^{\beta\gamma}
\end{equation}
and the corresponding classical limit is the system:
\begin{align}\label{Cpolino}
& s=0 \,\, \| \quad \partial_t H_0^0 +\partial_i H_0^i = 0,\, \qquad \quad \leftrightarrow \quad \partial_t F +\partial_i F^i= 0, \quad \,\, \nonumber\\
&\quad \quad \quad \,\,\,\,\,\, \partial_t H_0^j +\partial_i H_0^{ij} = 0, \qquad \quad \leftrightarrow \quad \partial_t F^j +\partial_i F^{ij}= 0, \nonumber \\
&\quad \quad \quad \,\,\,\,\,\, \partial_t H_0^{jk} +\partial_i H_0^{ijk} = J_0^{jk}, \quad \leftrightarrow \quad \partial_t F^{jk} +\partial_i F^{ijk}= P^{jk}, \\
& s=1 \,\, \| \quad \partial_t H_1^0 +\partial_i H_1^i = 0,\, \qquad \quad \leftrightarrow \quad \partial_t G_{ll} +\partial_i G^{lli}= 0, \nonumber \\
&\quad \quad \quad \,\,\,\,\,\, \partial_t H_1^{j} +\partial_i H_1^{ij} = J_1^{j}, \,\, \quad \quad \leftrightarrow \quad \partial_t G^{llij} +\partial_i G^{lli}= Q^{llj}, \nonumber \\
& s=2 \,\, \| \quad \partial_t H_2^0 +\partial_i H_2^i = J_2^0,\, \nonumber
\end{align}
where the new scalar moment $H_2^0$ and the corresponding flux $H_2^i$ are
\begin{align}
\left( \begin{array}{l}
H_2^0\\ H_2^{i}
\end{array}\right)
= m
\int_{\mathbb{R}^{3}}\int_0^{+\infty} f \, \left(\xi^2 + 2\frac{\mathcal{I}}{m}\right)^2
\left( \begin{array}{c}
1 \\ \xi_i
\end{array}\right)
\, \varphi(\mathcal{I}) \, d\mathcal{I} d {\boldsymbol \xi} , \nonumber
\end{align}
and the production term
\begin{align}
J_2^0
= m
\int_{\mathbb{R}^{3}}\int_0^{+\infty} Q \, \left(\xi^2 + 2\frac{\mathcal{I}}{m}\right)^2
\, \varphi(\mathcal{I}) \, d\mathcal{I} d {\boldsymbol \xi} .\nonumber
\end{align}
\section{Closure of Moments and RET$_{15}$}
Until now, we discussed the choice of the truncated moments to consider and we proved that for a given relativistic system with truncation index $N+1$, there exists a unique classical limit for the moments. The truncated system are both in relativistic and classic regimes, not closed. The closure procedure belongs in RET theory \cite{ET,RET,book,newbook}. It is expressed by a hyperbolic system of field equations with local constitutive equations. The closure is obtained at the phenomenological level using the universal principle such as the entropy principle, the entropy convexity, and the covariance with respect to the proper group of transformation. Or, at the molecular level, the closure is obtained by using the Maximum Entropy Principle (MEP) introduced in non-equilibrium thermodynamics first by Janes \cite{Janes} and successively developed by M\"uller and Ruggeri that proved as first that the closed system becomes symmetric hyperbolic \cite{ET}.
The closure of polytropic relativistic Euler fluids \eqref{3.2} was given first in the paper \cite{Annals} (see also \cite{JMP,ARMA}), while the closure in the case of $15$ fields (RET$_{15}$) \eqref{Rpolino} (relativistic case) and \eqref{Cpolino} (classical limit) was respectively the subject of the recent papers \cite{entropy} and \cite{IJNM}.
More precisely, in the case $N=2$ the system \eqref{MRel} becomes
\begin{equation}\label{Annalis}
\partial_\alpha A^{\alpha } = 0, \quad \partial_\alpha A^{\alpha \beta} =0, \quad \partial_\alpha A^{\alpha \beta \gamma} = I^{ \beta \gamma }, \qquad \left(\beta,\gamma=0,1,2,3\right).
\end{equation}
with
\begin{align} \label{relRETpol}
\begin{split}
& A^{\alpha } = {mc} \int_{\mathbb{R}^{3}}\int_0^{+\infty} f \, p^\alpha \,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P} \, , \\
&A^{\alpha \beta} = c \int_{\mathbb{R}^{3}}\int_0^{+\infty} f \, p^\alpha p^{\beta} \, \Big( 1 + \frac{\mathcal{I} }{m c^2}\Big)\,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P} \, , \\
& A^{\alpha \beta\gamma } = \frac{c}{m} \int_{\mathbb{R}^{3}}\int_0^{+\infty} f \, p^\alpha p^{\beta} p^{\gamma} \, \Big( 1 + \frac{\mathcal{I}}{mc^2} \Big)^2\,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P} \, , \\
& I^{ \beta \gamma } = \frac{c}{m} \int_{\mathbb{R}^{3}}\int_0^{+\infty} Q \, p^{\beta} p^{\gamma}\, \left(1 + \frac{\mathcal{I}}{mc^2} \right)^2\,
\phi(\mathcal{I}) \, d \mathcal{I} \, d \boldsymbol{P}. \\
\end{split}
\end{align}
Recalling the following decomposition of the particle number vector and the energy-momentum tensor in terms of physical variables:
\begin{align*
A^\alpha =\rho U^\alpha \, , \quad A^{\alpha \beta} = \frac{e}{c^2} \, U^{\alpha } U^\beta + \, \left(p \, + \, \Pi\right)
h^{\alpha \beta} + \frac{1}{c^2} ( U^\alpha q^\beta +U^\beta q^\alpha)+ t^{<\alpha \beta>_3} \, ,
\end{align*}
where $n, \rho = n m, U^\alpha, h^{\alpha\beta},p,e$ are respectively the particle number, the rest mass density, the four-velocity, the projector tensor $(h^{\alpha\beta}= U^\alpha U^\beta/c^2 - g^{\alpha\beta})$, the pressure, the energy. Moreover $g^{\alpha \beta}= \text{diag}(1 \, , \, -1 \, , \, -1 \,, \, -1)$ is the metric tensor, $\Pi$ is the dynamic pressure, $q^\alpha= -h^\alpha_\mu
U_\nu T^{\mu \nu}$ is the heat flux and $t^{<\alpha \beta>_3} = T^{\mu\nu} \left(h^\alpha_\mu h^\beta_\nu - \frac{1}{3}h^{\alpha\beta}h_{\mu\nu}\right)$ is the deviatoric shear viscous stress tensor. We also recall the constraints:
\begin{equation*}
U^\alpha U_\alpha = c^2, \quad q^\alpha U_\alpha = 0, \quad t^{<\alpha \beta>_3} U_\alpha = 0, \quad t^{<\alpha}_{\,\,\,\,\,\ \alpha >_3} =0,
\end{equation*}
and we choose as the $15$th variable:
\begin{align*
\Delta = \frac{4}{c^2} \, U_\alpha U_\beta U_\gamma \, \left( A^{\alpha \beta \gamma} \, - \, A^{\alpha \beta \gamma}_E\right).
\end{align*}
The pressure and the energy
compatible with the equilibrium distribution function are \cite{Annals}:
\begin{align}\label{10}
\begin{split}
& p =\frac{ k_B}{m} \, \rho T \, , \qquad \qquad e=
\rho c^2 \omega(\gamma), \\
& \text{with } \quad \omega(\gamma)= \frac{\int_0^{+\infty} J_{2,2}^* \, \Big( 1 + \frac{\mathcal{I}}{m c^2} \Big) \, \phi(\mathcal{I}) \, d \, \mathcal{I}}{\int_0^{+\infty} J_{2,1}^* \, \phi(\mathcal{I}) \, d \, \mathcal{I}},
\end{split}
\end{align}
\begin{equation}\label{8}
\begin{split}
& J_{m,n}^* = J_{m,n} (\gamma^*), \qquad \gamma^* = \gamma \, \Big( 1
+ \frac{\mathcal{I}}{m \, c^2} \Big), \qquad \gamma = \frac{m \, c^2}{k_B T},
\end{split}
\end{equation}
with $T$ being the temperature and $k_B$ being the Boltzmann constant, and
\begin{equation*}
J_{m,n}(\gamma)= \int_0^{+\infty} e^{-\gamma \cosh s} \sinh^m s \cosh^n s \, d s .
\end{equation*}
To close the system \eqref{relRETpol}, we have adopted in \cite{entropy} the MEP which requires to find the distribution function that maximizes the non-equilibrium entropy density:
\begin{equation}\label{Entropy}
h= h^\alpha U_\alpha \quad \rightarrow \quad \max
\end{equation}
with the entropy four-vector given by
\begin{equation} \label{fourentropy}
h^\alpha = - k_B \, c \, \int_{\mathbb{R}^{3}}\int_0^{+\infty} f
\ln f p^\alpha \phi(\mathcal{I}) \, d\mathcal{I} \, d\boldsymbol{P},
\end{equation}
under the constraints that the temporal part $A^\alpha U_\alpha, A^{\alpha\beta}U_\alpha$ and $A^{\alpha\beta\gamma}U_\alpha$ are prescribed. Proceeding in the usual way as indicated in previous papers of RET (see \cite{RS,Annals}), we obtain:
\begin{align}\label{f15}
\begin{split}
& f= e^{ -1 - \frac{\chi}{k_B}} \, , \quad \mbox{with} \\
& \chi = m \, \lambda \, + \, \lambda_{\mu} \, p^{\mu} \, \left( 1 + \, \frac{\mathcal{I}}{m \, c^2} \right) \, + \, \frac{1}{m} \, \lambda_{\mu \nu} \, p^{\mu} p^{\nu} \, \left( 1 + \, \frac{\mathcal{I}}{m \, c^2} \right)^2,
\end{split}
\end{align}
where $\lambda, \lambda_{\mu}, \lambda_{\mu \nu}$ are the Lagrange multipliers.
In the molecular RET approach, we consider, as usual, the processes near equilibrium. For this reason, we expand \eqref{f15} around an equilibrium state as follows:
\begin{align*
\begin{split}
&f \simeq f_E\left(1-\frac{1}{k_B}\tilde{\chi}\right), \quad \mbox{with} \\
&\tilde{\chi} = m \, (\lambda - \lambda_E) \, + \, (\lambda_{\mu}-\lambda_{\mu_E}) \, p^{\mu} \, \left( 1 + \, \frac{\mathcal{I}}{m \, c^2} \right) \, + \, \frac{1}{m} \, \lambda_{\mu \nu} \, p^{\mu} p^{\nu} \, \left( 1 + \, \frac{\mathcal{I}}{m \, c^2} \right)^2,
\end{split}
\end{align*}
with
\begin{align*
\begin{split}
& {\lambda}_E = - \frac{1}{T}\left(g+ c^2\right), \quad {\lambda}_{\mu_E}= \frac{U_\mu}{T}, \quad \lambda_{\mu \nu_E} =0,
\end{split}
\end{align*}
where $g =\varepsilon +p/\rho -T S $ is the equilibrium chemical potential with $S$ being the equilibrium entropy and $\varepsilon = e/\rho - c^2$.
In \cite{entropy}, it was proved that choosing as collisional term a variant of BGK model proposed in \cite{BGKPR} the triple tensor and the production term have necessarily this closed form:
\begin{align*
\begin{split}
A^{\alpha \beta \gamma} &= \left( \rho\, \theta_{0,2} + \, \frac{1}{4c^4} \, \Delta \right) U^{\alpha } U^{\beta} U^\gamma + \left( \rho \, c^2\,\theta_{1,2} - \, \frac{3}{4c^2}\, \frac{N^\Delta}{D_4} \, \Delta \, - 3 \, \frac{N^\Pi}{D_4} \, \Pi \right)
\, h^{(\alpha\beta} U^{\gamma)} \\
& + \frac{3}{c^2} \frac{N_3}{D_3} \, q^{(\alpha} U^\beta U^{\gamma)} + \frac{3}{5} \frac{N_{31}}{D_3} h^{(\alpha
\beta} q^{\gamma)} + 3 C_5 t^{(<\alpha \beta >_3} U^{\gamma)}
\, ,
\end{split}
\end{align*}
\begin{align*
I^{\beta\gamma} = &\frac{1}{\tau}
\Big\{
- \frac{1}{4c^4 } \Delta \, U^\beta U^{\gamma } + \left( \frac{1}{4c^2} \frac{N^\Delta}{D_4}\Delta +\frac{N^\Pi}{D_4} \Pi \right) h^{\beta\gamma} + \\
&\Big( -\frac{2}{c^2 } \frac{N_3}{D_3} +\frac{\theta_{1,3}}{\theta_{1,2}}\frac{1}{c^2 } \Big)q^{(\beta} U^{ \gamma )} - C_5 t^{<\beta\gamma>_3}
\Big\},
\end{align*}
where all coefficients are explicit functions of
\begin{align}\label{11b}
\theta_{a,b} = \frac{1}{2a+1} \begin{pmatrix}
b+1 \\ 2a
\end{pmatrix} \frac{\int_0^{+\infty} J_{2a+2,b+1-2a}^* \, \Big( 1 + \frac{\mathcal{I}}{m c^2} \Big)^b \, \phi(\mathcal{I}) \, d \, \mathcal{I}}{\int_0^{+\infty} J_{2,1}^* \, \phi(\mathcal{I}) \, d \, \mathcal{I}} \,
\end{align}
that are dimensionless function only of $\gamma$ (i.e. function of the temperature, see \eqref{8}) and depending by recursive formula of the unique function $\omega(\gamma)$ strictly related with the energy trough the relations \eqref{10}.
It was also proved in \cite{entropy} that the classical limit of this model coincides with the corresponding classical RET$_{15}$ studied in \cite{IJNM}.
Both the model of relativistic and classical RET$_{15}$ are very complex, and therefore, in principle, it is hard to discuss the qualitative analysis. Nevertheless, we want to prove that they belong to the systems of balance laws with a convex entropy for which exists general theorems of qualitative analysis as we summarize in the next section.
\section{Qualitative Analysis}
The system \eqref{Annalis} belongs to a general quasi-linear system of $N$ balance laws:
\begin{equation}
\partial _{\alpha }\mathbf{F}^{\alpha }(\mathbf{u})=\mathbf{f}(\mathbf{u}),
\label{r1}
\end{equation}%
compatible with an entropy law
\begin{equation}
\partial _{\alpha }h^{\alpha }(\mathbf{u)}=\Sigma \,(\mathbf{u)}, \qquad \Sigma \geq 0,
\label{e8}
\end{equation}
where $h^\alpha$ and $\Sigma$ are,
respectively, the entropy vector and the entropy production.
For this kind of systems starting from previous results of Godunov \cite{Godunov}, Friedrichs and Lax \cite{FL} and Boillat \cite{Boillat}, Ruggeri and Strumia proved the following theorem \cite{RS}:
\begin{theorem}[Ruggeri-Strumia]
\label{CampoPRIN}
The compatibility between the system of balance laws \eqref{r1} and the supplementary balance law \eqref{e8} with the entropy $h=h^\alpha \xi_\alpha$ being a convex function of $\mathbf{u} \equiv
\mathbf{F}^\alpha \xi_\alpha$, with $\xi_\alpha$ a congruence time-like, implies the existence of the "main field" $ \mathbf{u}^\prime $ that satisfies
\begin{equation*}
dh^\alpha = \mathbf{u}^\prime \cdot d\mathbf{F}^\alpha, \qquad \Sigma =
\mathbf{u}^\prime \cdot \mathbf{f}\,\geq 0.
\end{equation*}
If we choose the components of $ \mathbf{u}^\prime $ as field variables,
we have
\begin{equation}
\mathbf{F}^\alpha = \frac{\partial h^{\prime \alpha}}{\partial \mathbf{u}%
^\prime}, \label{potenzialis}
\end{equation}
and the
original system \textrm{(\ref{r1})} can be rewritten in
a symmetric form with Hessian matrices:
\begin{equation}
\partial_\alpha \left( \frac{\partial h^{\prime \alpha}}{\partial \mathbf{u}%
^\prime } \right)= \mathbf{f} \quad \iff \quad \frac{\partial^2 h^{\prime
\alpha}}{\partial \mathbf{u}^\prime \partial \mathbf{u}^\prime}
\partial_\alpha \mathbf{u}^\prime= \mathbf{f}, \label{simetric}
\end{equation}
where $h^{\prime \alpha}$ is the four-potential defined by
\begin{equation}
h^{\prime \alpha}= \mathbf{u}^\prime \cdot \mathbf{F}^\alpha - h^\alpha.
\label{110}
\end{equation}
The function
\begin{equation*}
h^\prime = h^{\prime \alpha} \xi_\alpha = \mathbf{u}^\prime \cdot \mathbf{u} - h,
\end{equation*}
is the Legendre transformation of $h$ and therefore a convex function of the dual field $\mathbf{u}^\prime$.
\end{theorem}
In the general theory of symmetric hyperbolic balance laws, it is well-known that the system
(\ref{r1}) has a unique local (in time) smooth solution for smooth
initial data \cite{FL,Kawa,fisher}.
However, in a general case, even for arbitrarily small and smooth
initial data, there is no global continuation for these smooth solutions,
which may develop singularities, shocks, or blow{up, in a finite time, see for
instance \cite{maida,dafermos}.
On the other hand, in many physical examples, thanks to the
interplay between the source term and the hyperbolicity, there
exist global smooth solutions for a suitable set of initial data.
In this context, the following \emph{K-condition} \cite{SK} plays an important role:
\begin{definition}[K-condition]
A system (\ref{r1}) satisfies the K-condition if, in
the equilibrium manifold, any right characteristic eigenvectors ${\bf d}$ of (\ref{r1})
are not in the null space of $\nabla {\bf f}$, where $\nabla\equiv \partial/\partial {\bf u}$:
\begin{equation}\label{Kcond}
\left(\nabla {\bf f}\,\, {\bf d}^I\right)_E \neq 0 \quad \forall \, {\bf d}^I, \,\,\,\, I=1,2,\dots N.
\end{equation}
\end{definition}
For dissipative one-dimensional systems (\ref{r1}) satisfying
the K-condition, it is possible to prove the following global
existence theorem by Hanouzet and Natalini \cite{nat}:
\begin{theorem}[Global Existence] \label{global_e}
{Assume that the system} (\ref{r1}) {is strictly dissipative} with a convex entropy
{and that the
K-condition is satisfied. Then there exists $\delta>0$, such that,
if $\left\Vert \mathbf{u}^{\prime }(x,0)\right\Vert _{2}\leq \delta ,$ there is
a unique global smooth solution, which verifies}
\[
\mathbf{u}^\prime \in C^{0}([0,\infty );H^{2}(R)\cap
C^{1}([0,\infty );H^{1}(R)).
\]
\end{theorem}
This global existence theorem was generalized to a higher-dimensional case by Yong \cite{wen} and successively by Bianchini, Hanouzet, and Natalini \cite{bnat}.
Moreover {Ruggeri} and {Serre} \cite{RugSerre} proved that the constant equilibrium
state is stable.
Dafermos showed
the existence and long time behavior of spatially periodic BV solutions \cite{Dafermos-2015}.
The K-condition is
only a sufficient condition for the global existence of smooth solutions.
Lou and Ruggeri \cite{Palermo} observed that there indeed exists a weaker K-condition that is a necessary (but unfortunately not sufficient) condition
for the global existence of smooth solutions. Instead of the condition that the right eigenvectors are not in the null space of $\nabla {\bf f}$, they posed this condition only on the right eigenvectors corresponding to genuine nonlinear eigenvalues.
It was proved that the assumptions of the previous theorems are fulfilled in both
classical \cite{Ruget} and relativistic \cite{RugETR,RUGCHO} RET theories of monatomic gases, and also in the theory of mixtures of gases with multi-temperature \cite{SR}.
In \cite{entropy}, it was proved that at least in a neighborhood of equilibrium, the entropy \eqref{Entropy}
is a convex function of the field $\mathbf{u} = \mathbf{F}^\alpha U_\alpha$, and the entropy principle \eqref{e8} is satisfied, then we need to prove only the K-condition to satisfy the assumptions of previous theorem.
For this aim we first need to calculate the characteristic velocities evaluated in equilibrium.
\subsection{Characteristic velocities in equilibrium}
We recall that in \cite{Boillat-1997} it was proved that in the theory of moments, the main field coincides with the Lagrange multipliers of MEP (see also \cite{BRrel1,newbook}), and therefore \eqref{MRel} taking into account \eqref{potenzialis} and \eqref{simetric}
can be written
\begin{equation}\label{1}
\partial_\alpha \left( \frac{\partial \, h'^\alpha}{\partial \, \lambda_A}\right) = I^A \quad , \quad \mbox{or} \quad \frac{\partial^2 \, h'^\alpha}{\partial \, \lambda_A \partial \, \lambda_B} \, \partial_\alpha \lambda_B = I^A \, ,
\end{equation}
where the multi-index $A$ is used for the Lagrange multipliers in equivalent way of \eqref{F-Mom}:
\begin{equation*}
\lambda_A = \lambda_{\alpha_1 \alpha_2\dots \alpha_A}, \quad (\lambda \,\, \text{when} \,\, A=0),
\end{equation*}
that in the present case of $15$ moments $A=0,1,2$, i.e. $\mathbf{u}^\prime \equiv (\lambda,\lambda_\alpha,\lambda_{\alpha\beta})$.
As it is well-known, the wave equations associated with the system \eqref{1} can be obtained by the following rule:
\begin{equation*}
\partial_\alpha \,\, \rightarrow \,\, \varphi_\alpha\, \delta\,\ , \qquad I^A \,\, \rightarrow \,\, 0,
\end{equation*}
with
\begin{align*}
\varphi_\alpha= \, \frac{V}{c} \, \xi_\alpha + \eta_\alpha \, ,
\end{align*}
where $V$ indicates the characteristic velocity and $\xi_\alpha$ and $ \eta_\alpha$ indicate, respectively, a generic time-like and space-like congruence: $\xi_\alpha \xi^\alpha=1$, $\xi_\alpha \eta^\alpha=0$, $\eta_\alpha\eta^\alpha=-1$.
Therefore we have from \eqref{1}:
\begin{align}\label{4.1}
\varphi_\alpha \, \frac{\partial^2 \, h'^\alpha}{\partial \, \lambda_A \, \partial \, \lambda_B} \, \delta \lambda_B = 0,
\end{align}
where $\delta \lambda_B$ are the right eigenvectors associated to the system \eqref{1}.
For the symmetry of $\frac{\partial^2 h'^\alpha}{\partial \lambda_A \, \partial \lambda_B}$ and the convexity of $h^\prime = h^{\prime\alpha}\xi_\alpha$ with respect to the main field, the quadratic form \footnote{We recall in Mathematical community the entropy is the physical entropy changed by sign and therefore we use still the terms convexity where in reality our function is concave.}
\begin{equation*}
\frac{\partial^2 h^\prime}{\partial \lambda_A \, \partial \lambda_B} \, \delta \lambda_A \, \delta \lambda_B
\end{equation*}
is negative definite $\forall$ time-like 4-vector $\xi_\alpha$, we can deduce that all the eigenvalues $V$ are real and the equations \eqref{4.1} gives a basis of eigenvectors $\delta \lambda_B$; in other words, our field equations, according with Theorem \ref{CampoPRIN}, is symmetric hyperbolic. \\
We have also that the characteristic velocities $ V$ don' t exceed the light speed, i.e., $V^2 \leq c^2$, thanks to theorems proved in \cite{BRrel1} and in the Appendix A of \cite{Hyp}. \\
For the effective evaluation of the wave velocities, we use the same strategy used in a similar problem in \cite{Hyp}. More precisely, by using the definition of the 4-potential \eqref{110}, the expressions \eqref{relRETpol}, $f$ given in \eqref{f15} and the entropy vector \eqref{fourentropy}, we obtain:
\begin{align*}
h'^\alpha = - \, k_B c \, \int_{\mathbb{R}^{3}}\int_0^{+\infty} \, e^{- \, 1 \, - \, \frac{\chi}{k_B}} \, p^\alpha \phi(\mathcal{I}) \, d \, \mathcal{I} \, d \boldsymbol{P} \, ,
\end{align*}
then
\begin{align*}
\frac{\partial \, h'^\alpha }{\partial \, \lambda_A} = c \, \int_{\mathbb{R}^{3}}\int_0^{+\infty} \, e^{- \, 1 \, - \, \frac{\chi}{k_B}} \, \frac{\partial \, \chi }{\partial \, \lambda_A} \, p^\alpha \phi(\mathcal{I}) \, d \, \mathcal{I} \, d \boldsymbol{P} \, .
\end{align*}
Since from \eqref{f15} $\chi$ is linear in the Lagrange multipliers, $\frac{\partial \chi }{\partial \lambda_A}$ does not depend on $\lambda_B$, it follows
\begin{align*}
\frac{\partial^2 \, h'^\alpha }{\partial \, \lambda_B \, \partial \, \lambda_A} = - \, \frac{c}{k_B} \, \int_{\mathbb{R}^{3}}\int_0^{+\infty} \, e^{- \, 1 \, - \, \frac{\chi}{k_B}} \, \frac{\partial \, \chi }{\partial \, \lambda_B} \, \frac{\partial \, \chi }{\partial \, \lambda_A} \, p^\alpha \phi(\mathcal{I}) \, d \, \mathcal{I} \, d \boldsymbol{P} \, .
\end{align*}
By using these results, we can consider the quadratic form
\begin{align*
\delta K= - \, \frac{c}{k_B} \, \varphi_\alpha \, \int_{\mathbb{R}^{3}}\int_0^{+\infty} \, e^{- \, 1 \, - \, \frac{\chi}{k_B}} \, (\delta \, \chi \, )^2 \, p^\alpha \phi(\mathcal{I}) \, d \, \mathcal{I} \, d \boldsymbol{P} \, ,
\end{align*}
and see that the equations \eqref{4.1} for the wave velocities are equivalent to say that the derivatives of $\delta K$ with respect to $ \delta \lambda _A$ are zero.\\
As the closure was obtained only near equilibrium, we rewrite
$ \delta K$ as \begin{align*}
\begin{split}
& \delta K= \varphi_\alpha \, \left[ \frac{\partial^2 \, h'^\alpha }{\partial \, \lambda^2} \, \left( \delta \lambda \right)^2 \, + 2 \, \frac{\partial^2 \, h'^\alpha }{\partial \, \lambda \, \partial \, \lambda_\mu} \, \delta \lambda \, \delta \lambda_\mu + 2 \, \frac{\partial^2 \, h'^\alpha }{\partial \, \lambda \, \partial \, \lambda_{\mu \nu}} \, \delta \lambda \, \delta \lambda_{\mu \nu} + \right. \\
& \left. + \, \frac{\partial^2 \, h'^\alpha }{\partial \, \lambda_\beta \, \partial \, \lambda_\mu} \, \delta \lambda_\beta \, \delta \lambda_\mu + 2 \, \frac{\partial^2 \, h'^\alpha }{\partial \, \lambda_\beta \, \partial \, \lambda_{\mu \nu}} \, \delta \lambda_\beta \, \delta \lambda_{\mu \nu} + \frac{\partial^2 \, h'^\alpha }{\partial \, \lambda_{\beta \gamma} \, \partial \, \lambda_{\mu \nu}} \, \delta \lambda_{\beta \gamma} \, \delta \lambda_{\mu \nu} \right]
\, .
\end{split}
\end{align*}
By calculating the coefficients at equilibrium, it becomes
\begin{align*}
\begin{split}
& \delta K_E= - \, \frac{m}{k_B} \, \varphi_\alpha \, \left[ A_E^\alpha \, \left( \delta \lambda \right)^2 \, + 2 \, A_E^{\alpha \mu} \, \delta \lambda \, \delta \lambda_\mu + 2 \, A_{E}^{\alpha \mu \nu} \, \delta \lambda \, \delta \lambda_{\mu \nu} \, + \right. \\
& \quad \quad + \, A_{E}^{\alpha \beta \delta} \, \delta \lambda_\beta \, \delta \lambda_\delta + \left. 2 \, A_{E}^{\alpha \beta \mu \nu} \, \delta \lambda_\beta \, \delta \lambda_{\mu \nu} + \, A_{E}^{\alpha \beta \gamma \mu \nu} \, \delta \lambda_{\beta \gamma} \, \delta \lambda_{\mu \nu} \right]
\, ,
\end{split}
\end{align*}
where the explicit expressions of the tensors in the right hand side in terms of the $\vartheta_{a,b}$ \eqref{11b} are reported in \cite{entropy}. For the sake of simplicity, we calculate also the coefficients of the differentials in the reference frame where $U^\alpha$ and $\varphi^\alpha$ have the components $U^\alpha \equiv (c \, , \, 0 \, , \, 0 \, , \, 0)$ and $\varphi_\alpha \equiv (\varphi_0 \, , \, \varphi_1 \, , \, 0 \, , \, 0)$; in any case, we can at the end express again all the results in covariant form replacing $\varphi_0$ and $\left( \varphi_1 \right)^2$ with $\varphi_0 = \frac{1}{c} \, \varphi^\alpha U_\alpha$ and $\left( \varphi_1 \right)^2= \varphi_\alpha \varphi_\beta h^{\alpha \beta}$.
After having calculated $ \delta K_E$, we note that a first eigenvalue is
\begin{align}\label{eig1}
\varphi_0=0 \, , \quad \mbox{i.e., } \quad V = -c \, \frac{U^\alpha \eta_\alpha}{U^\gamma \xi_\gamma} \, , \quad \mbox{i.e., } \quad V =0 \, ,
\end{align}
where the last expression holds when $\xi_\gamma = U_\gamma/c$.
We can observe that $\varphi_1 \neq 0$ under the hypothesis that the 2 time-like vectors $\xi_\gamma$ and $U_\gamma$ are oriented both towards the future or both towards the past. After that, if $\varphi_0=0$, the derivatives of $ \delta K_E$ with respect to $ \delta \lambda _A$ give a system whose solution is $ \delta \lambda _1=0$, $ \delta \lambda _{01}=0$, $ \delta \lambda _{12}=0$, $ \delta \lambda _{13}=0$ and the remaining unknowns are linked only by
\begin{align}\label{eig1bis}
\begin{split}
& 6 \, \vartheta_{1,1} \, \delta \lambda \, + \, 2 \, \vartheta_{1,2} \, c \, \delta \lambda_{0} \, + \, \vartheta_{1,3} \, c^2 \, \delta \lambda_{00} \, + \,
2 \, \vartheta_{2,3} \, c^2 \,
\left( 3 \, \delta \lambda_{11} \, + \, \delta \lambda_{22} \, + \, \delta \lambda_{33} \right) =0 \, , \\
& 10 \, \vartheta_{1,2} \, \delta \lambda \, + \, 5 \, \vartheta_{1,3} \, c \, \delta \lambda_{0} \, + \, 3 \, \vartheta_{1,4} \, c^2 \, \delta \lambda_{00} \, + \, 2 \, \vartheta_{2,4} \, c^2 \,
\left( 3 \, \delta \lambda_{11} \, + \, \delta \lambda_{22} \, + \, \delta \lambda_{33} \right) =0 \, , \\
& 5 \, \vartheta_{2,3} \, \delta \lambda_2 \, + \, 2 \, \vartheta_{2,4} \, c \, \delta \lambda_{20} =0 \, , \quad
5 \, \vartheta_{2,3} \, \delta \lambda_3 \, + \, 2 \, \vartheta_{2,4} \, c \, \delta \lambda_{30} =0 \, .
\end{split}
\end{align}
Therefore, we have 7 free unknowns and the eigenvalue \eqref{eig1} has multiplicity 7.
For the research of other eigenvalues we have $\varphi_0 \neq 0$ and we can consider the quadratic form $ - \, \frac{k_B}{m \, \rho \, c \, \varphi_0} \, \delta K_E $. By defining
\noindent
$ \delta \lambda = X_1$, $c \, \delta \lambda_{0} = X_2$, $c \, \delta \lambda_{1} = X_3$, $c^2 \, \delta \lambda_{00} = X_4$, $c^2 \, \delta \lambda_{01} = X_5$, $c^2 \, \delta \lambda_{11} = X_6$, $c^2 \,\left( \delta \lambda_{22} \, + \, \delta \lambda_{33} \right) = X_7$, $c \, \delta \lambda_{2} = Y_1$, $c^2 \, \delta \lambda_{20} = Y_2$, $c^2 \, \delta \lambda_{12} = Y_3$,
$c \, \delta \lambda_{3} = Z_1$, \\
$c^2 \, \delta \lambda_{30} = Z_2$, $c^2 \, \delta \lambda_{13} = Z_3$, $c^2 \delta \lambda_{23} = Y_4$,
$c^2 \left( \delta \lambda_{22} \, - \, \delta \lambda_{33 }\right) = Z_4$,
\noindent
we have
\begin{align*}
- \, \frac{k_B}{m \, \rho \, c \, \varphi_0} \, \delta K_E = \sum_{h,k=1}^{7} \, a_{hk} \, X_h \, X_k \, + \, \sum_{h=1}^{3} \, b_{hk} \, Y_h \, Y_k \, + \, \sum_{h=1}^{3} \, b_{hk} \, Z_h \, Z_k \, + \\
+ \, \frac{4}{15} \, \vartheta_{2,4} \, \left( Y_4 \right)^2 \, + \, \frac{1}{15} \, \vartheta_{2,4} \, \left( Z_4 \right)^2 \, ,
\end{align*}
with
\begin{align*}
\begin{split}
& a_{11} = \vartheta_{0,0} \, , \quad a_{12} = \vartheta_{0,1} \, , \quad a_{13} = \vartheta_{1,1} \, \frac{\varphi_1 }{\varphi_0} \, , \quad a_{14} = \vartheta_{0,2} \, , \quad a_{15} = \frac{2}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} \, , \\
& a_{16} = \frac{1}{3} \, \vartheta_{1,2} \, , \quad a_{17} = a_{16} \, , \quad a_{22} = \vartheta_{0,2} \, , \quad a_{23} = \frac{1}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} \, , \\
& a_{24} = \vartheta_{0,3} \, , \quad a_{25} = \frac{1}{3} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} \, , \quad a_{26} = \frac{1}{6} \, \vartheta_{1,3} \, , \quad a_{27} = a_{26} \, , \quad a_{33} = \frac{1}{3} \, \vartheta_{1,2} \, , \\
& a_{34} = \frac{1}{6} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} \, , \quad a_{35} = \frac{1}{3} \, \vartheta_{1,3} \, , \quad a_{36} = \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} \, , \quad a_{37} = \frac{1}{3} \, \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} \, , \\
& a_{44} = \vartheta_{0,4} \, , \quad a_{45} = \frac{1}{5} \, \vartheta_{1,4} \, \frac{\varphi_1 }{\varphi_0} \, , \quad a_{46} = \frac{1}{10} \, \vartheta_{1,4} \, , \quad a_{47} = a_{46} \, , \quad a_{55} = \frac{2}{5} \, \vartheta_{1,4} \, , \\
& a_{56} = \frac{2}{5} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} \, \quad a_{57} = \frac{2}{15} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} \, , \quad a_{66} = \frac{1}{5} \, \vartheta_{2,4} \, , \quad a_{67} = \frac{1}{15} \, \vartheta_{2,4} \, , \\
& a_{77} = \frac{2}{15} \,
\vartheta_{2,4} \, ,
\end{split}
\end{align*}
\begin{align*}
\begin{split}
& b_{11} = \frac{1}{3} \,\vartheta_{1,2} \, , \quad b_{12} = \frac{1}{3} \, \vartheta_{1,3} \, , \quad b_{13} = \frac{2}{3} \, \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} \, , \\
& b_{22} = \frac{2}{5} \,\vartheta_{1,4} \, , \quad b_{23} = \frac{4}{15} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} \, , \quad b_{33} = \frac{4}{15} \, \vartheta_{2,4} \, .
\end{split}
\end{align*}
From these results it follows that the equations to determine eigenvalues and eigenvectors are
\begin{align}\label{eig2}
\sum_{k=1}^{7} \, a_{hk} \, X_k =0 \, , \quad \sum_{k=1}^{3} \, b_{hk} \, Y_k =0 \, , \quad \sum_{k=1}^{3} \, b_{hk} \, Z_k =0 \, , \quad Y_4=0 \, , \quad Z_4=0 \, .
\end{align}
Here eqs. \eqref{eig2}$_{4,5}$ are $Y_4=0$, $Z_4=0$. The equations \eqref{eig2}$_{2,3}$ shows that 2 eigenvalues with multiplicity 2 are the solution of
\begin{align}\label{eig3}
\det \left\|
\begin{matrix}
\frac{1}{3} \, \vartheta_{1,2} & \frac{1}{3} \, \vartheta_{1,3} & \frac{2}{3} \, \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} \\
&& \\
\frac{1}{3} \, \vartheta_{1,3} & \frac{2}{5} \, \vartheta_{1,4} & \frac{4}{15} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} \\
&& \\
\frac{2}{3} \, \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{4}{15} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} & \frac{4}{15} \, \vartheta_{2,4}
\end{matrix}
\right\| = 0 \, ,
\end{align}
that is,
\begin{align*}
\det \left\|
\begin{matrix}
\vartheta_{1,2} & \vartheta_{1,3} & 2 \, \vartheta_{2,3} \\
&& \\
\vartheta_{1,3} & \frac{6}{5} \, \vartheta_{1,4} & \frac{4}{5} \, \vartheta_{2,4} \\
&& \\
2 \, \vartheta_{2,3} & \frac{4}{5} \, \vartheta_{2,4} & 0
\end{matrix}
\right\| \, \left( \frac{\varphi_1 }{\varphi_0} \right)^2 \, + \, \frac{4}{5} \, \vartheta_{2,4} \, \det \left\| \begin{matrix}
\vartheta_{1,2} & \vartheta_{1,3} \\
& \\
\vartheta_{1,3} & \frac{6}{5} \, \vartheta_{1,4}
\end{matrix} \right\|
= 0 \, ,
\end{align*}
The eigenvalues different from those in \eqref{eig1}, \eqref{eig3} are given by \eqref{eig2}$_{1}$ in the unknowns $X_k$, that is the determinant of the matrix $a_{hk}$ must be zero, i.e.,
\begin{align}\label{eig5}
\det \left\| \begin{matrix}
\vartheta_{0,0} & \vartheta_{0,1} & \vartheta_{1,1} \, \frac{\varphi_1 }{\varphi_0} & \vartheta_{0,2} & \frac{2}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,2} & \frac{1}{3} \, \vartheta_{1,2} \\
&&&&&& \\
\vartheta_{0,1} & \vartheta_{0,2} & \frac{1}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} & \vartheta_{0,3} & \frac{1}{3} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{6} \, \vartheta_{1,3} & \frac{1}{6}\, \vartheta_{1,3} \\
&&&&&& \\
\vartheta_{1,1} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,2} & \frac{1}{6} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,3} & \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} \\
&&&&&& \\
\vartheta_{0,2} & \vartheta_{0,3} & \frac{1}{6} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} & \vartheta_{0,4} & \frac{1}{5} \, \vartheta_{1,4} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{10} \, \vartheta_{1,4} & \frac{1}{10} \, \vartheta_{1,4} \\
&&&&&& \\
\frac{2}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,3} & \frac{1}{5} \, \vartheta_{1,4} \, \frac{\varphi_1 }{\varphi_0} & \frac{2}{5} \, \vartheta_{1,4} & \frac{2}{5} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} & \frac{2}{15} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} \\
&&&&&& \\
\frac{1}{3} \, \vartheta_{1,2} & \frac{1}{6}\, \vartheta_{1,3} & \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{10} \, \vartheta_{1,4} & \frac{2}{5} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{5} \, \vartheta_{2,4} & \frac{1}{15} \, \vartheta_{2,4} \\
&&&&&& \\
\frac{1}{3} \, \vartheta_{1,2} & \frac{1}{6} \, \vartheta_{1,3} & \frac{1}{3} \, \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{10} \, \vartheta_{1,4} & \frac{2}{15} \, \vartheta_{2,4} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{15} \, \vartheta_{2,4} & \frac{2}{15} \vartheta_{2,4}
\end{matrix} \right\| = 0 . \nonumber \\
\nonumber \\
\end{align}
It is easy to prove that this equation depends on $\frac{\varphi_1 }{\varphi_0}$ only through $\left( \frac{\varphi_1 }{\varphi_0} \right)^2= \frac{h^{\alpha \beta} \varphi_\alpha \varphi_\beta}{\left( U^\gamma \varphi_\gamma \right)^2 } \, c^2$ (which is equal to $\left( \frac{c}{V} \right)^2$ if $U^\alpha = c \, \xi^\alpha$) and it is a second degree equation in $\left( \frac{\varphi_1 }{\varphi_0} \right)^2$.
So it gives 4 independent eigenvectors; other 7 come from \eqref{eig1}, other 4 come from \eqref{eig3}. The total is 15, as expected.
\begin{figure}[h!]
\centering
\includegraphics[width=80mm]{CS_ET15-ET5.pdf}
\caption{Dependence of maximum $V/c$ on $\gamma$ for a diatomic gas. The solid line indicates the value of RET$_{15}$ and the dashed line indicates the one of Euler system \cite{ARMA}.}
\label{CS}
\end{figure}
\bigskip
As example we can use the expressions of $\vartheta_{a,b}$ \eqref{11b} given in \cite{entropy} in the case of diatomic gases for which the expression of energy $e$ given in \eqref{10} is explicit because $\omega(\gamma)$ can be written in terms of ratio of modified Bessel functions \cite{ARMA}
\begin{equation*
\omega(\gamma)= \frac{K_0(\gamma)}{K_1(\gamma)} +\frac{3}{\gamma}.
\end{equation*}
As consequence it is easy to plot the maximum characteristic velocity in the rest frame as function of $\gamma$ (see Fig. \ref{CS}).
According with general results for which increasing the number of moments increases the maximum characteristic velocity \cite{Boillat-1997,BRrel1} and
since the relativistic Euler system is a\emph{ principal subsystem} of RET$_{15}$ by the definition given in \cite{BoillatRuggeriARMA}, the sub-characteristic conditions hold and the maximum characteristic velocity of RET$_{15}$ that is obtained from \eqref{eig5} is larger than the one of Euler system which is studied in \cite{ARMA} as evident in Fig. \ref{CS}.
\subsection{K-Condition}
As was proved in \cite{Palermo},
the K-condition \eqref{Kcond} is equivalent to $\delta {\mathbf{f}} \neq 0$
for any characteristic velocities in equilibrium. In the present case this is equivalent to prove that
$\delta I^{\beta \gamma} \neq 0$ at equilibrium.
To prove that it holds in our case, let us suppose by absurd that it does't hold, i.e., that there is at least a characteristic velocity with $\delta I^{\beta \gamma} = 0$.
Now $\delta \, I^{\beta \gamma} = 0$ with the coefficients of the differentials of the independent variables calculated at equilibrium, is equivalent to $ \delta \lambda _{\beta \gamma} = 0$ (because the quadratic form $\Sigma$ is positive defined and, consequently, $I^{\beta \gamma}$ is invertible in $\lambda_{\beta \gamma}$).
If the eigenvalue under consideration is $\varphi_\alpha U^\alpha=0$, this means that, jointly with \eqref{eig1bis}, its expression calculated in $ \delta \lambda _{\beta \gamma} = 0$ holds, i.e.,
\begin{align*}
6 \, \vartheta_{1,1} \, \delta \lambda \, + \, 4 \, \vartheta_{1,2} \, c \, \delta \lambda_{0} =0, \, \, 10 \, \vartheta_{1,2} \, \delta \lambda \, + \, 5 \, \vartheta_{1,3} \, c \, \delta \lambda_{0} =0, \,\, \delta \lambda_2 =0, \, \, \delta \lambda_3 =0.
\end{align*}
This implies $ \delta \lambda _0=0$. Jointly with \eqref{eig1bis} and with the other results written before, we obtain $ \delta \lambda =0$, $ \delta \lambda _\beta=0$, $ \delta \lambda _{\beta \gamma} = 0$. This absurd shows that the K-condition is satisfied for the eigenvalue $\varphi_\alpha U^\alpha=0$.
If the eigenvalue under consideration is one of those in \eqref{eig3}, the absurd hypothesis means that, jointly with \eqref{eig2}, its expression calculated in $ \delta \lambda _{\beta \gamma} = 0$ holds, i.e.,
\begin{align}\label{eige6}
\sum_{k=1}^{3} \, a_{hk} \, X_k =0 \, , \quad b_{h1} \, Y_1 =0 \, , \quad b_{h1} \, Z_1 =0 \, .
\end{align}
The last two of these relations give $ \delta \lambda _2 = 0 $, $ \delta \lambda _3 = 0 $, while the first one gives ( with the results written before \eqref{eig3}) $ \delta \lambda = 0 $, $ \delta \lambda _0 = 0 $, $ \delta \lambda _1 = 0 $. This implies $ \delta \lambda =0$, $ \delta \lambda _\beta=0$, $ \delta \lambda _{\beta \gamma} = 0$. This absurd result shows that the K-condition is satisfied for the eigenvalues \eqref{eig3}.
It remains to prove that it also for the eigenvalues which are solutions of \eqref{eig5}. In this case the absurd hypothesis means that, jointly with \eqref{eig2}, also its expression calculated in $ \delta \lambda _{\beta \gamma} = 0$ holds, i.e., \eqref{eige6}.
The last two of these relations give $ \delta \lambda _2 = 0 $, $ \delta \lambda _3 = 0 $, while the first one says that
\begin{align}\label{eige7}
\left( \begin{matrix}
\vartheta_{0,0} & \vartheta_{0,1} & \vartheta_{1,1} \, \frac{\varphi_1 }{\varphi_0} \\
&& \\
\vartheta_{0,1} & \vartheta_{0,2} & \frac{1}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0}\\
&& \\
\vartheta_{1,1} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,2} \\
&& \\
\vartheta_{0,2} & \vartheta_{0,3} & \frac{1}{6} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} \\
&& \\
\frac{2}{3} \, \vartheta_{1,2} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,3} \, \frac{\varphi_1 }{\varphi_0} & \frac{1}{3} \, \vartheta_{1,3} \\
&& \\
\frac{1}{3} \, \vartheta_{1,2} & \frac{1}{6}\, \vartheta_{1,3} & \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0} \\
&& \\
\frac{1}{3} \, \vartheta_{1,2} & \frac{1}{6} \, \vartheta_{1,3} & \frac{1}{3} \, \vartheta_{2,3} \, \frac{\varphi_1 }{\varphi_0}
\end{matrix} \right)
\left(\begin{matrix}
\delta \lambda \\
c \, \delta \lambda_0 \\
c \, \delta \lambda_1
\end{matrix} \right) = \left(\begin{matrix}
0 \\
0 \\
0
\end{matrix} \right) .
\end{align}
This implies that all the 35 third order minors of the matrix in the left hand side must be zero for the same value of the unknown $\frac{\varphi_1 }{\varphi_0}$. In particular, we may consider eqs. \eqref{eige7}$_{1,2,4}$ (or \eqref{eige7}$_{1,2,6}$, or \eqref{eige7}$_{1,2,7}$, or \eqref{eige7}$_{1,4,6}$, or \eqref{eige7}$_{1,4,7}$, or \eqref{eige7}$_{2,4,6}$, or \eqref{eige7}$_{2,4,7}$, or \eqref{eige7}$_{4,6,7}$) and obtain the result $\frac{\varphi_1 }{\varphi_0}=0$ which is absurd because we have said that $\varphi_1 \neq 0$ and, in any case, the system \eqref{eige7} would give the absurd result
$ \delta \lambda =0$, $ \delta \lambda _0 =0$, $ \delta \lambda _1 =0$.
\bigskip
Similar calculations can be done for the classical limit system \eqref{Cpolino} (see \cite{IJNM}) and it is possible to prove that also in classical case the K-condition is satisfied.
\bigskip
\emph{Therefore we can conclude that both the solutions of the relativistic and classical systems satisfy the theorems before stated, and as a consequence, global smooth solutions exist provided initial data are sufficiently small and not far away from an equilibrium state.}
\bigskip
\bmhead{Acknowledgments}
The work has been partially supported by JSPS KAKENHI Grant Numbers JP18K13471 (TA), by the Italian MIUR through the PRIN2017
project "Multiscale phenomena in Continuum Mechanics:
singular limits, off-equilibrium and transitions" Project Number:
2017YBKNCE (SP) and GNFM/INdAM (MCC, SP and TR).
\section*{Declarations}
\begin{itemize}
\item Funding: JSPS KAKENHI Grant Numbers JP18K13471;
PRIN2017 Number: 2017YBKNCE.
\item Conflict of interest/Competing interests: The authors declare no conflict of interest.
\item Availability of data and materials: Not applicable.
\end{itemize}
|
\section*{Acknowledgments}
We thank CERN for the very successful operation of the LHC, as well as the
support staff from our institutions without whom ATLAS could not be
operated efficiently.
We acknowledge the support of
ANPCyT, Argentina;
YerPhI, Armenia;
ARC, Australia;
BMWFW and FWF, Austria;
ANAS, Azerbaijan;
SSTC, Belarus;
CNPq and FAPESP, Brazil;
NSERC, NRC and CFI, Canada;
CERN;
ANID, Chile;
CAS, MOST and NSFC, China;
Minciencias, Colombia;
MEYS CR, Czech Republic;
DNRF and DNSRC, Denmark;
IN2P3-CNRS and CEA-DRF/IRFU, France;
SRNSFG, Georgia;
BMBF, HGF and MPG, Germany;
GSRI, Greece;
RGC and Hong Kong SAR, China;
ISF and Benoziyo Center, Israel;
INFN, Italy;
MEXT and JSPS, Japan;
CNRST, Morocco;
NWO, Netherlands;
RCN, Norway;
MEiN, Poland;
FCT, Portugal;
MNE/IFA, Romania;
JINR;
MES of Russia and NRC KI, Russian Federation;
MESTD, Serbia;
MSSR, Slovakia;
ARRS and MIZ\v{S}, Slovenia;
DSI/NRF, South Africa;
MICINN, Spain;
SRC and Wallenberg Foundation, Sweden;
SERI, SNSF and Cantons of Bern and Geneva, Switzerland;
MOST, Taiwan;
TAEK, Turkey;
STFC, United Kingdom;
DOE and NSF, United States of America.
In addition, individual groups and members have received support from
BCKDF, CANARIE, Compute Canada and CRC, Canada;
COST, ERC, ERDF, Horizon 2020 and Marie Sk{\l}odowska-Curie Actions, European Union;
Investissements d'Avenir Labex, Investissements d'Avenir Idex and ANR, France;
DFG and AvH Foundation, Germany;
Herakleitos, Thales and Aristeia programmes co-financed by EU-ESF and the Greek NSRF, Greece;
BSF-NSF and GIF, Israel;
Norwegian Financial Mechanism 2014-2021, Norway;
NCN and NAWA, Poland;
La Caixa Banking Foundation, CERCA Programme Generalitat de Catalunya and PROMETEO and GenT Programmes Generalitat Valenciana, Spain;
G\"{o}ran Gustafssons Stiftelse, Sweden;
The Royal Society and Leverhulme Trust, United Kingdom.
The crucial computing support from all WLCG partners is acknowledged gratefully, in particular from CERN, the ATLAS Tier-1 facilities at TRIUMF (Canada), NDGF (Denmark, Norway, Sweden), CC-IN2P3 (France), KIT/GridKA (Germany), INFN-CNAF (Italy), NL-T1 (Netherlands), PIC (Spain), ASGC (Taiwan), RAL (UK) and BNL (USA), the Tier-2 facilities worldwide and large non-WLCG resource providers. Major contributors of computing resources are listed in Ref.~\cite{ATL-SOFT-PUB-2021-003}.
\printbibliography
\clearpage
\FloatBarrier
\clearpage
\input{atlas_authlist}
\end{document}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.